Bump up version to v0.3.2 (#2968)

This version is for more model support. Add support for Gemma models (#2964) and OLMo models (#2832).

.buildkite/run-benchmarks.sh

@@ -0,0 +1,69 @@
+# This script is run by buildkite to run the benchmarks and upload the results to buildkite
+
+set -ex
+set -o pipefail
+
+# cd into parent directory of this file
+cd "$(dirname "${BASH_SOURCE[0]}")/.."
+
+(which wget && which curl) || (apt-get update && apt-get install -y wget curl)
+
+# run python-based benchmarks and upload the result to buildkite
+python3 benchmarks/benchmark_latency.py 2>&1 | tee benchmark_latency.txt
+bench_latency_exit_code=$?
+
+python3 benchmarks/benchmark_throughput.py --input-len 256 --output-len 256 2>&1 | tee benchmark_throughput.txt
+bench_throughput_exit_code=$?
+
+# run server-based benchmarks and upload the result to buildkite
+python3 -m vllm.entrypoints.openai.api_server --model meta-llama/Llama-2-7b-chat-hf &
+server_pid=$!
+wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
+
+# wait for server to start, timeout after 600 seconds
+timeout 600 bash -c 'until curl localhost:8000/v1/models; do sleep 1; done' || exit 1
+python3 benchmarks/benchmark_serving.py \
+    --backend openai \
+    --dataset ./ShareGPT_V3_unfiltered_cleaned_split.json \
+    --model meta-llama/Llama-2-7b-chat-hf \
+    --num-prompts 20 \
+    --endpoint /v1/completions \
+    --tokenizer meta-llama/Llama-2-7b-chat-hf \
+    --save-result \
+    2>&1 | tee benchmark_serving.txt
+bench_serving_exit_code=$?
+kill $server_pid
+
+# write the results into a markdown file
+echo "### Latency Benchmarks" >> benchmark_results.md
+sed -n '1p' benchmark_latency.txt >> benchmark_results.md # first line
+echo "" >> benchmark_results.md
+sed -n '$p' benchmark_latency.txt >> benchmark_results.md # last line
+
+echo "### Throughput Benchmarks" >> benchmark_results.md
+sed -n '1p' benchmark_throughput.txt >> benchmark_results.md # first line
+echo "" >> benchmark_results.md
+sed -n '$p' benchmark_throughput.txt >> benchmark_results.md # last line
+
+echo "### Serving Benchmarks" >> benchmark_results.md
+sed -n '1p' benchmark_serving.txt >> benchmark_results.md # first line
+echo "" >> benchmark_results.md
+tail -n 13 benchmark_serving.txt >> benchmark_results.md # last 13 lines
+
+# upload the results to buildkite
+/workspace/buildkite-agent annotate --style "info" --context "benchmark-results" < benchmark_results.md
+
+# exit with the exit code of the benchmarks
+if [ $bench_latency_exit_code -ne 0 ]; then
+    exit $bench_latency_exit_code
+fi
+
+if [ $bench_throughput_exit_code -ne 0 ]; then
+    exit $bench_throughput_exit_code
+fi
+
+if [ $bench_serving_exit_code -ne 0 ]; then
+    exit $bench_serving_exit_code
+fi
+
+/workspace/buildkite-agent artifact upload openai-*.json

.buildkite/test-pipeline.yaml

@@ -0,0 +1,66 @@
+# In this file, you can add more tests to run either by adding a new step or
+# adding a new command to an existing step. See different options here for examples.
+# This script will be feed into Jinja template in `test-template.j2` to generate
+# the final pipeline yaml file.
+
+steps:
+- label: Regression Test
+  command: pytest -v -s test_regression.py
+  working_dir: "/vllm-workspace/tests" # optional
+
+- label: AsyncEngine Test
+  command: pytest -v -s async_engine
+
+- label: Basic Correctness Test
+  command: pytest -v -s --forked basic_correctness
+
+- label: Distributed Comm Ops Test
+  command: pytest -v -s --forked test_comm_ops.py
+  working_dir: "/vllm-workspace/tests/distributed"
+  num_gpus: 2 # only support 1 or 2 for now.
+
+- label: Distributed Correctness Test
+  command: pytest -v -s --forked test_basic_distributed_correctness.py
+  working_dir: "/vllm-workspace/tests/distributed"
+  num_gpus: 2 # only support 1 or 2 for now.
+
+- label: Engine Test
+  command: pytest -v -s engine
+
+- label: Entrypoints Test
+  command: pytest -v -s entrypoints
+
+- label: Kernels Test
+  command: pytest -v -s kernels
+  soft_fail: true
+
+- label: Models Test
+  commands:
+    - pytest -v -s models --forked
+  soft_fail: true
+
+- label: Prefix Caching Test
+  commands:
+    - pytest -v -s prefix_caching
+
+- label: Samplers Test
+  command: pytest -v -s samplers --forked
+
+- label: Worker Test
+  command: pytest -v -s worker
+
+- label: LoRA Test
+  command: pytest -v -s lora
+
+- label: Benchmarks
+  working_dir: "/vllm-workspace/.buildkite"
+  commands:
+  - pip install aiohttp
+  - bash run-benchmarks.sh
+
+- label: Documentation Build
+  working_dir: "/vllm-workspace/docs"
+  no_gpu: True
+  commands:
+  - pip install -r requirements-docs.txt
+  - SPHINXOPTS=\"-W\" make html

.buildkite/test-template.j2

@@ -0,0 +1,56 @@
+{% set docker_image = "us-central1-docker.pkg.dev/vllm-405802/vllm-ci-test-repo/vllm-test:$BUILDKITE_COMMIT" %}
+{% set default_num_gpu = 1 %}
+{% set default_working_dir = "/vllm-workspace/tests" %}
+
+steps:
+  - label: ":docker: build image"
+    commands:
+      - "docker build --build-arg max_jobs=16 --tag {{ docker_image }} --target test --progress plain ."
+      - "docker push {{ docker_image }}"
+    env:
+      DOCKER_BUILDKIT: "1"
+    retry:
+      automatic:
+        - exit_status: -1  # Agent was lost
+          limit: 5
+  - wait
+
+  {% for step in steps %}
+  - label: "{{ step.label }}"
+    agents:
+      queue: kubernetes
+    soft_fail: {{ step.soft_fail or false }}
+    retry:
+      automatic:
+        - exit_status: -1  # Agent was lost
+          limit: 5
+    plugins:
+      - kubernetes:
+          podSpec:
+            volumes:
+              - name: dshm
+                emptyDir:
+                  medium: Memory
+            containers:
+              - image: "{{ docker_image }}"
+                command: ["bash"]
+                args:
+                - '-c'
+                - "'cd {{ (step.working_dir or default_working_dir) | safe  }} && {{ step.command  or (step.commands | join(' && ')) | safe }}'"
+                {% if not step.no_gpu %}
+                resources:
+                  requests:
+                    nvidia.com/gpu: "{{ step.num_gpus or default_num_gpu }}"
+                  limits:
+                    nvidia.com/gpu: "{{ step.num_gpus or default_num_gpu }}"
+                {% endif %}
+                env:
+                  - name: HF_TOKEN
+                    valueFrom:
+                      secretKeyRef:
+                        name: hf-token-secret
+                        key: token
+                volumeMounts:
+                  - mountPath: /dev/shm
+                    name: dshm
+  {% endfor %}

.dockerignore

@@ -0,0 +1 @@
+vllm/*.so

.github/workflows/publish.yml

@@ -0,0 +1,102 @@
+# This workflow will upload a Python Package to Release asset
+# For more information see: https://help.github.com/en/actions/language-and-framework-guides/using-python-with-github-actions
+
+name: Create Release
+
+on:
+  push:
+    tags:
+      - v*
+
+# Needed to create release and upload assets
+permissions:
+  contents: write
+
+jobs:
+  release:
+    # Retrieve tag and create release
+    name: Create Release
+    runs-on: ubuntu-latest
+    outputs:
+      upload_url: ${{ steps.create_release.outputs.upload_url }}
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v3
+
+      - name: Extract branch info
+        shell: bash
+        run: |
+          echo "release_tag=${GITHUB_REF#refs/*/}" >> $GITHUB_ENV
+
+      - name: Create Release
+        id: create_release
+        uses: "actions/github-script@v6"
+        env:
+          RELEASE_TAG: ${{ env.release_tag }}
+        with:
+          github-token: "${{ secrets.GITHUB_TOKEN }}"
+          script: |
+            const script = require('.github/workflows/scripts/create_release.js')
+            await script(github, context, core)
+
+  wheel:
+    name: Build Wheel
+    runs-on: ${{ matrix.os }}
+    needs: release
+
+    strategy:
+      fail-fast: false
+      matrix:
+          os: ['ubuntu-20.04']
+          python-version: ['3.8', '3.9', '3.10', '3.11']
+          pytorch-version: ['2.1.2']  # Must be the most recent version that meets requirements.txt.
+          cuda-version: ['11.8', '12.1']
+
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v3
+
+      - name: Set up Linux Env
+        if: ${{ runner.os == 'Linux' }}
+        run: |
+          bash -x .github/workflows/scripts/env.sh
+
+      - name: Set up Python
+        uses: actions/setup-python@v4
+        with:
+            python-version: ${{ matrix.python-version }}
+
+      - name: Install CUDA ${{ matrix.cuda-version }}
+        run: |
+          bash -x .github/workflows/scripts/cuda-install.sh ${{ matrix.cuda-version }} ${{ matrix.os }}
+
+      - name: Install PyTorch ${{ matrix.pytorch-version }} with CUDA ${{ matrix.cuda-version }}
+        run: |
+          bash -x .github/workflows/scripts/pytorch-install.sh ${{ matrix.python-version }} ${{ matrix.pytorch-version }} ${{ matrix.cuda-version }}
+
+      - name: Build wheel
+        shell: bash
+        run: |
+          bash -x .github/workflows/scripts/build.sh ${{ matrix.python-version }} ${{ matrix.cuda-version }}
+          wheel_name=$(ls dist/*whl | xargs -n 1 basename)
+          asset_name=${wheel_name//"linux"/"manylinux1"}
+          echo "wheel_name=${wheel_name}" >> $GITHUB_ENV
+          echo "asset_name=${asset_name}" >> $GITHUB_ENV
+
+      - name: Upload Release Asset
+        uses: actions/upload-release-asset@v1
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        with:
+          upload_url: ${{ needs.release.outputs.upload_url }}
+          asset_path: ./dist/${{ env.wheel_name }}
+          asset_name: ${{ env.asset_name }}
+          asset_content_type: application/*
+
+      # (Danielkinz): This last step will publish the .whl to pypi. Warning: untested
+      # - name: Publish package
+      #   uses: pypa/gh-action-pypi-publish@release/v1.8
+      #   with:
+      #     repository-url: https://test.pypi.org/legacy/
+      #     password: ${{ secrets.PYPI_API_TOKEN }}
+      #     skip-existing: true

.github/workflows/ruff.yml

@@ -0,0 +1,31 @@
+name: ruff
+
+on:
+  # Trigger the workflow on push or pull request,
+  # but only for the main branch
+  push:
+    branches:
+      - main
+  pull_request:
+    branches:
+      - main
+
+jobs:
+  ruff:
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        python-version: ["3.10"]
+    steps:
+    - uses: actions/checkout@v2
+    - name: Set up Python ${{ matrix.python-version }}
+      uses: actions/setup-python@v2
+      with:
+        python-version: ${{ matrix.python-version }}
+    - name: Install dependencies
+      run: |
+        python -m pip install --upgrade pip
+        pip install ruff==0.1.5
+    - name: Analysing the code with ruff
+      run: |
+        ruff vllm tests

.github/workflows/scripts/build.sh

@@ -0,0 +1,20 @@
+#!/bin/bash
+
+python_executable=python$1
+cuda_home=/usr/local/cuda-$2
+
+# Update paths
+PATH=${cuda_home}/bin:$PATH
+LD_LIBRARY_PATH=${cuda_home}/lib64:$LD_LIBRARY_PATH
+
+# Install requirements
+$python_executable -m pip install wheel packaging
+$python_executable -m pip install -r requirements.txt
+
+# Limit the number of parallel jobs to avoid OOM
+export MAX_JOBS=1
+# Make sure punica is built for the release (for LoRA)
+export VLLM_INSTALL_PUNICA_KERNELS=1
+
+# Build
+$python_executable setup.py bdist_wheel --dist-dir=dist

.github/workflows/scripts/create_release.js

@@ -0,0 +1,20 @@
+// Uses Github's API to create the release and wait for result.
+// We use a JS script since github CLI doesn't provide a way to wait for the release's creation and returns immediately.
+
+module.exports = async (github, context, core) => {
+	try {
+		const response = await github.rest.repos.createRelease({
+			draft: false,
+			generate_release_notes: true,
+			name: process.env.RELEASE_TAG,
+			owner: context.repo.owner,
+			prerelease: false,
+			repo: context.repo.repo,
+			tag_name: process.env.RELEASE_TAG,
+		});
+
+		core.setOutput('upload_url', response.data.upload_url);
+	} catch (error) {
+		core.setFailed(error.message);
+	}
+}

.github/workflows/scripts/cuda-install.sh

@@ -0,0 +1,23 @@
+#!/bin/bash
+
+# Replace '.' with '-' ex: 11.8 -> 11-8
+cuda_version=$(echo $1 | tr "." "-")
+# Removes '-' and '.' ex: ubuntu-20.04 -> ubuntu2004
+OS=$(echo $2 | tr -d ".\-")
+
+# Installs CUDA
+wget -nv https://developer.download.nvidia.com/compute/cuda/repos/${OS}/x86_64/cuda-keyring_1.1-1_all.deb
+sudo dpkg -i cuda-keyring_1.1-1_all.deb
+rm cuda-keyring_1.1-1_all.deb
+sudo apt -qq update
+sudo apt -y install cuda-${cuda_version} cuda-nvcc-${cuda_version} cuda-libraries-dev-${cuda_version}
+sudo apt clean
+
+# Test nvcc
+PATH=/usr/local/cuda-$1/bin:${PATH}
+nvcc --version
+
+# Log gcc, g++, c++ versions
+gcc --version
+g++ --version
+c++ --version

.github/workflows/scripts/env.sh

@@ -0,0 +1,56 @@
+#!/bin/bash
+
+# This file installs common linux environment tools
+
+export LANG C.UTF-8
+
+# python_version=$1
+
+sudo    apt-get update && \
+sudo    apt-get install -y --no-install-recommends \
+        software-properties-common \
+
+sudo    apt-get install -y --no-install-recommends \
+        build-essential \
+        apt-utils \
+        ca-certificates \
+        wget \
+        git \
+        vim \
+        libssl-dev \
+        curl \
+        unzip \
+        unrar \
+        cmake \
+        net-tools \
+        sudo \
+        autotools-dev \
+        rsync \
+        jq \
+        openssh-server \
+        tmux \
+        screen \
+        htop \
+        pdsh \
+        openssh-client \
+        lshw \
+        dmidecode \
+        util-linux \
+        automake \
+        autoconf \
+        libtool \
+        net-tools \
+        pciutils \
+        libpci-dev \
+        libaio-dev \
+        libcap2 \
+        libtinfo5 \
+        fakeroot \
+        devscripts \
+        debhelper \
+        nfs-common
+
+# Remove github bloat files to free up disk space
+sudo rm -rf "/usr/local/share/boost"
+sudo rm -rf "$AGENT_TOOLSDIRECTORY"
+sudo rm -rf "/usr/share/dotnet"

.github/workflows/scripts/pytorch-install.sh

@@ -0,0 +1,15 @@
+#!/bin/bash
+
+python_executable=python$1
+pytorch_version=$2
+cuda_version=$3
+
+# Install torch
+$python_executable -m pip install numpy pyyaml scipy ipython mkl mkl-include ninja cython typing pandas typing-extensions dataclasses setuptools && conda clean -ya
+$python_executable -m pip install torch==${pytorch_version}+cu${cuda_version//./} --extra-index-url https://download.pytorch.org/whl/cu${cuda_version//./}
+
+# Print version information
+$python_executable --version
+$python_executable -c "import torch; print('PyTorch:', torch.__version__)"
+$python_executable -c "import torch; print('CUDA:', torch.version.cuda)"
+$python_executable -c "from torch.utils import cpp_extension; print (cpp_extension.CUDA_HOME)"

.github/workflows/yapf.yml

@@ -0,0 +1,31 @@
+name: yapf
+
+on:
+  # Trigger the workflow on push or pull request,
+  # but only for the main branch
+  push:
+    branches:
+      - main
+  pull_request:
+    branches:
+      - main
+jobs:
+  yapf:
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        python-version: ["3.10"]
+    steps:
+    - uses: actions/checkout@v2
+    - name: Set up Python ${{ matrix.python-version }}
+      uses: actions/setup-python@v2
+      with:
+        python-version: ${{ matrix.python-version }}
+    - name: Install dependencies
+      run: |
+        python -m pip install --upgrade pip
+        pip install yapf==0.32.0
+        pip install toml==0.10.2
+    - name: Running yapf
+      run: |
+        yapf --diff --recursive .

.gitignore

@@ -1,10 +1,186 @@
-**/*.pyc
-**/__pycache__/
-*.egg-info/
-*.eggs/
-*.so
-build/
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
 
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+cover/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+.pybuilder/
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+#   For a library or package, you might want to ignore these files since the code is
+#   intended to run in multiple environments; otherwise, check them in:
+# .python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# poetry
+#   Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#   https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
+#poetry.lock
+
+# pdm
+#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
+#pdm.lock
+#   pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it
+#   in version control.
+#   https://pdm.fming.dev/#use-with-ide
+.pdm.toml
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/
+
+# PyCharm
+#  JetBrains specific template is maintained in a separate JetBrains.gitignore that can
+#  be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
+#  and can be added to the global gitignore or merged into this file.  For a more nuclear
+#  option (not recommended) you can uncomment the following to ignore the entire idea folder.
+.idea/
+
+# VSCode
+.vscode/
+
+# DS Store
+.DS_Store
+
+# Results
+*.csv
+
+# Python pickle files
 *.pkl
-*.png
-**/log.txt
+
+# Sphinx documentation
+_build/
+
+# vim swap files
+*.swo
+*.swp
+
+# hip files generated by PyTorch
+*.hip
+*_hip*
+
+# Benchmark dataset
+*.json

.readthedocs.yaml

@@ -0,0 +1,21 @@
+# Read the Docs configuration file
+# See https://docs.readthedocs.io/en/stable/config-file/v2.html for details
+
+version: 2
+
+build:
+  os: ubuntu-22.04
+  tools:
+    python: "3.8"
+
+sphinx:
+   configuration: docs/source/conf.py
+
+# If using Sphinx, optionally build your docs in additional formats such as PDF
+formats:
+   - pdf
+
+# Optionally declare the Python requirements required to build your docs
+python:
+   install:
+   - requirements: docs/requirements-docs.txt

CONTRIBUTING.md

@@ -0,0 +1,77 @@
+# Contributing to vLLM
+
+Thank you for your interest in contributing to vLLM!
+Our community is open to everyone and welcomes all kinds of contributions, no matter how small or large.
+There are several ways you can contribute to the project:
+
+- Identify and report any issues or bugs.
+- Request or add a new model.
+- Suggest or implement new features.
+
+However, remember that contributions aren't just about code.
+We believe in the power of community support; thus, answering queries, assisting others, and enhancing the documentation are highly regarded and beneficial contributions.
+
+Finally, one of the most impactful ways to support us is by raising awareness about vLLM.
+Talk about it in your blog posts, highlighting how it's driving your incredible projects.
+Express your support on Twitter if vLLM aids you, or simply offer your appreciation by starring our repository.
+
+
+## Setup for development
+
+### Build from source
+
+```bash
+pip install -r requirements.txt
+pip install -e .  # This may take several minutes.
+```
+
+### Testing
+
+```bash
+pip install -r requirements-dev.txt
+
+# Static type checking
+mypy
+# Unit tests
+pytest tests/
+```
+**Note:** Currently, the repository does not pass the mypy tests.
+
+
+## Contributing Guidelines
+
+### Issue Reporting
+
+If you encounter a bug or have a feature request, please check our issues page first to see if someone else has already reported it.
+If not, please file a new issue, providing as much relevant information as possible.
+
+### Coding Style Guide
+
+In general, we adhere to [Google Python style guide](https://google.github.io/styleguide/pyguide.html) and [Google C++ style guide](https://google.github.io/styleguide/cppguide.html).
+
+We include a formatting script [`format.sh`](./format.sh) to format the code.
+
+### Pull Requests
+
+When submitting a pull request:
+
+1. Make sure your code has been rebased on top of the latest commit on the main branch.
+2. Ensure code is properly formatted by running [`format.sh`](./format.sh).
+3. Include a detailed description of the changes in the pull request.
+Explain why you made the changes you did.
+If your pull request fixes an open issue, please include a reference to it in the description.
+
+### Code Reviews
+
+All submissions, including submissions by project members, require a code review.
+To make the review process as smooth as possible, please:
+
+1. Keep your changes as concise as possible.
+If your pull request involves multiple unrelated changes, consider splitting it into separate pull requests.
+2. Respond to all comments within a reasonable time frame.
+If a comment isn't clear or you disagree with a suggestion, feel free to ask for clarification or discuss the suggestion.
+
+### Thank You
+
+Finally, thank you for taking the time to read these guidelines and for your interest in contributing to vLLM.
+Your contributions make vLLM a great tool for everyone!

Dockerfile

@@ -0,0 +1,105 @@
+# The vLLM Dockerfile is used to construct vLLM image that can be directly used
+# to run the OpenAI compatible server.
+
+#################### BASE BUILD IMAGE ####################
+FROM nvidia/cuda:12.1.0-devel-ubuntu22.04 AS dev
+
+RUN apt-get update -y \
+    && apt-get install -y python3-pip git
+
+# Workaround for https://github.com/openai/triton/issues/2507 and
+# https://github.com/pytorch/pytorch/issues/107960 -- hopefully
+# this won't be needed for future versions of this docker image
+# or future versions of triton.
+RUN ldconfig /usr/local/cuda-12.1/compat/
+
+WORKDIR /workspace
+
+# install build and runtime dependencies
+COPY requirements.txt requirements.txt
+RUN --mount=type=cache,target=/root/.cache/pip \
+    pip install -r requirements.txt
+
+# install development dependencies
+COPY requirements-dev.txt requirements-dev.txt
+RUN --mount=type=cache,target=/root/.cache/pip \
+    pip install -r requirements-dev.txt
+#################### BASE BUILD IMAGE ####################
+
+
+#################### EXTENSION BUILD IMAGE ####################
+FROM dev AS build
+
+# install build dependencies
+COPY requirements-build.txt requirements-build.txt
+RUN --mount=type=cache,target=/root/.cache/pip \
+    pip install -r requirements-build.txt
+
+# copy input files
+COPY csrc csrc
+COPY setup.py setup.py
+COPY requirements.txt requirements.txt
+COPY pyproject.toml pyproject.toml
+COPY vllm/__init__.py vllm/__init__.py
+
+# cuda arch list used by torch
+ARG torch_cuda_arch_list='7.0 7.5 8.0 8.6 8.9 9.0+PTX'
+ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}
+# max jobs used by Ninja to build extensions
+ARG max_jobs=2
+ENV MAX_JOBS=${max_jobs}
+# number of threads used by nvcc
+ARG nvcc_threads=8
+ENV NVCC_THREADS=$nvcc_threads
+# make sure punica kernels are built (for LoRA)
+ENV VLLM_INSTALL_PUNICA_KERNELS=1
+
+RUN python3 setup.py build_ext --inplace
+#################### EXTENSION Build IMAGE ####################
+
+
+#################### TEST IMAGE ####################
+# image to run unit testing suite
+FROM dev AS test
+
+# copy pytorch extensions separately to avoid having to rebuild
+# when python code changes
+WORKDIR /vllm-workspace
+# ADD is used to preserve directory structure
+ADD . /vllm-workspace/
+COPY --from=build /workspace/vllm/*.so /vllm-workspace/vllm/
+# ignore build dependencies installation because we are using pre-complied extensions
+RUN rm pyproject.toml
+RUN --mount=type=cache,target=/root/.cache/pip VLLM_USE_PRECOMPILED=1 pip install . --verbose
+#################### TEST IMAGE ####################
+
+
+#################### RUNTIME BASE IMAGE ####################
+# We used base cuda image because pytorch installs its own cuda libraries.
+# However cupy depends on cuda libraries so we had to switch to the runtime image
+# In the future it would be nice to get a container with pytorch and cuda without duplicating cuda
+FROM nvidia/cuda:12.1.0-runtime-ubuntu22.04 AS vllm-base
+
+# libnccl required for ray
+RUN apt-get update -y \
+    && apt-get install -y python3-pip
+
+WORKDIR /workspace
+COPY requirements.txt requirements.txt
+RUN --mount=type=cache,target=/root/.cache/pip \
+    pip install -r requirements.txt
+#################### RUNTIME BASE IMAGE ####################
+
+
+#################### OPENAI API SERVER ####################
+# openai api server alternative
+FROM vllm-base AS vllm-openai
+# install additional dependencies for openai api server
+RUN --mount=type=cache,target=/root/.cache/pip \
+    pip install accelerate
+
+COPY --from=build /workspace/vllm/*.so /workspace/vllm/
+COPY vllm vllm
+
+ENTRYPOINT ["python3", "-m", "vllm.entrypoints.openai.api_server"]
+#################### OPENAI API SERVER ####################

Dockerfile.rocm

@@ -0,0 +1,95 @@
+# default base image
+ARG BASE_IMAGE="rocm/pytorch:rocm6.0_ubuntu20.04_py3.9_pytorch_2.1.1"
+
+FROM $BASE_IMAGE
+
+ARG BASE_IMAGE="rocm/pytorch:rocm6.0_ubuntu20.04_py3.9_pytorch_2.1.1"
+
+RUN echo "Base image is $BASE_IMAGE"
+
+# BASE_IMAGE for ROCm_5.7: "rocm/pytorch:rocm5.7_ubuntu22.04_py3.10_pytorch_2.0.1"
+# BASE_IMAGE for ROCm_6.0: "rocm/pytorch:rocm6.0_ubuntu20.04_py3.9_pytorch_2.1.1"
+
+
+ARG FA_GFX_ARCHS="gfx90a;gfx942"
+RUN echo "FA_GFX_ARCHS is $FA_GFX_ARCHS"
+
+ARG FA_BRANCH="3d2b6f5"
+RUN echo "FA_BRANCH is $FA_BRANCH"
+
+# whether to build flash-attention
+# if 0, will not build flash attention
+# this is useful for gfx target where flash-attention is not supported
+# In that case, we need to use the python reference attention implementation in vllm
+ARG BUILD_FA="1"
+
+# Install some basic utilities
+RUN apt-get update && apt-get install python3 python3-pip -y
+
+# Install some basic utilities
+RUN apt-get update && apt-get install -y \
+    curl \
+    ca-certificates \
+    sudo \
+    git \
+    bzip2 \
+    libx11-6 \
+    build-essential \
+    wget \
+    unzip \
+    nvidia-cuda-toolkit \
+    tmux \
+ && rm -rf /var/lib/apt/lists/*
+
+### Mount Point ###
+# When launching the container, mount the code directory to /app
+ARG APP_MOUNT=/app
+VOLUME [ ${APP_MOUNT} ]
+WORKDIR ${APP_MOUNT}
+
+RUN python3 -m pip install --upgrade pip
+RUN python3 -m pip install --no-cache-dir fastapi ninja tokenizers pandas
+
+ENV LLVM_SYMBOLIZER_PATH=/opt/rocm/llvm/bin/llvm-symbolizer
+ENV PATH=$PATH:/opt/rocm/bin:/libtorch/bin:
+ENV LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/opt/rocm/lib/:/libtorch/lib:
+ENV CPLUS_INCLUDE_PATH=$CPLUS_INCLUDE_PATH:/libtorch/include:/libtorch/include/torch/csrc/api/include/:/opt/rocm/include/:
+
+# Install ROCm flash-attention
+RUN if [ "$BUILD_FA" = "1" ]; then \
+    mkdir libs \
+    && cd libs \
+    && git clone https://github.com/ROCm/flash-attention.git \
+    && cd flash-attention \
+    && git checkout ${FA_BRANCH} \
+    && git submodule update --init \
+    && export GPU_ARCHS=${FA_GFX_ARCHS} \
+    && if [ "$BASE_IMAGE" = "rocm/pytorch:rocm5.7_ubuntu22.04_py3.10_pytorch_2.0.1" ]; then \
+        patch /opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/utils/hipify/hipify_python.py hipify_patch.patch; fi \
+    && python3 setup.py install \
+    && cd ..; \
+    fi
+
+COPY ./ /app/vllm
+
+RUN python3 -m pip install --upgrade pip
+RUN python3 -m pip install xformers==0.0.23 --no-deps
+
+# Error related to odd state for numpy 1.20.3 where there is no METADATA etc, but an extra LICENSES_bundled.txt.
+# Manually removed it so that later steps of numpy upgrade can continue
+RUN if [ "$BASE_IMAGE" = "rocm/pytorch:rocm6.0_ubuntu20.04_py3.9_pytorch_2.1.1" ]; then \
+    rm -rf /opt/conda/envs/py_3.9/lib/python3.9/site-packages/numpy-1.20.3.dist-info/; fi
+
+RUN cd /app \
+    && cd vllm \
+    && pip install -U -r requirements-rocm.txt \
+    && if [ "$BUILD_FA" = "1" ]; then \
+       bash patch_xformers.rocm.sh; fi \
+    && patch /opt/rocm/include/hip/amd_detail/amd_hip_bf16.h /app/vllm/rocm_patch/rocm_bf16.patch \
+    && python3 setup.py install \
+    && cd ..
+
+RUN python3 -m pip install --upgrade pip
+RUN python3 -m pip install --no-cache-dir ray[all]
+
+CMD ["/bin/bash"]

LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

MANIFEST.in

@@ -0,0 +1,4 @@
+include LICENSE
+include requirements.txt
+
+recursive-include csrc *

README.md

@@ -1,72 +1,110 @@
-# CacheFlow
+<p align="center">
+  <picture>
+    <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/vllm-project/vllm/main/docs/source/assets/logos/vllm-logo-text-dark.png">
+    <img alt="vLLM" src="https://raw.githubusercontent.com/vllm-project/vllm/main/docs/source/assets/logos/vllm-logo-text-light.png" width=55%>
+  </picture>
+</p>
 
-## Installation
+<h3 align="center">
+Easy, fast, and cheap LLM serving for everyone
+</h3>
+
+<p align="center">
+| <a href="https://docs.vllm.ai"><b>Documentation</b></a> | <a href="https://vllm.ai"><b>Blog</b></a> | <a href="https://arxiv.org/abs/2309.06180"><b>Paper</b></a> | <a href="https://discord.gg/jz7wjKhh6g"><b>Discord</b></a> |
+
+</p>
+
+---
+
+*Latest News* 🔥
+- [2024/01] We hosted [the second vLLM meetup](https://lu.ma/ygxbpzhl) in SF! Please find the meetup slides [here](https://docs.google.com/presentation/d/12mI2sKABnUw5RBWXDYY-HtHth4iMSNcEoQ10jDQbxgA/edit?usp=sharing).
+- [2024/01] Added ROCm 6.0 support to vLLM.
+- [2023/12] Added ROCm 5.7 support to vLLM.
+- [2023/10] We hosted [the first vLLM meetup](https://lu.ma/first-vllm-meetup) in SF! Please find the meetup slides [here](https://docs.google.com/presentation/d/1QL-XPFXiFpDBh86DbEegFXBXFXjix4v032GhShbKf3s/edit?usp=sharing).
+- [2023/09] We created our [Discord server](https://discord.gg/jz7wjKhh6g)! Join us to discuss vLLM and LLM serving! We will also post the latest announcements and updates there.
+- [2023/09] We released our [PagedAttention paper](https://arxiv.org/abs/2309.06180) on arXiv!
+- [2023/08] We would like to express our sincere gratitude to [Andreessen Horowitz](https://a16z.com/2023/08/30/supporting-the-open-source-ai-community/) (a16z) for providing a generous grant to support the open-source development and research of vLLM.
+- [2023/07] Added support for LLaMA-2! You can run and serve 7B/13B/70B LLaMA-2s on vLLM with a single command!
+- [2023/06] Serving vLLM On any Cloud with SkyPilot. Check out a 1-click [example](https://github.com/skypilot-org/skypilot/blob/master/llm/vllm) to start the vLLM demo, and the [blog post](https://blog.skypilot.co/serving-llm-24x-faster-on-the-cloud-with-vllm-and-skypilot/) for the story behind vLLM development on the clouds.
+- [2023/06] We officially released vLLM! FastChat-vLLM integration has powered [LMSYS Vicuna and Chatbot Arena](https://chat.lmsys.org) since mid-April. Check out our [blog post](https://vllm.ai).
+
+---
+## About
+vLLM is a fast and easy-to-use library for LLM inference and serving.
+
+vLLM is fast with:
+
+- State-of-the-art serving throughput
+- Efficient management of attention key and value memory with **PagedAttention**
+- Continuous batching of incoming requests
+- Fast model execution with CUDA/HIP graph
+- Quantization: [GPTQ](https://arxiv.org/abs/2210.17323), [AWQ](https://arxiv.org/abs/2306.00978), [SqueezeLLM](https://arxiv.org/abs/2306.07629), FP8 KV Cache
+- Optimized CUDA kernels
+
+vLLM is flexible and easy to use with:
+
+- Seamless integration with popular Hugging Face models
+- High-throughput serving with various decoding algorithms, including *parallel sampling*, *beam search*, and more
+- Tensor parallelism support for distributed inference
+- Streaming outputs
+- OpenAI-compatible API server
+- Support NVIDIA GPUs and AMD GPUs
+- (Experimental) Prefix caching support
+- (Experimental) Multi-lora support
+
+vLLM seamlessly supports many Hugging Face models, including the following architectures:
+
+- Aquila & Aquila2 (`BAAI/AquilaChat2-7B`, `BAAI/AquilaChat2-34B`, `BAAI/Aquila-7B`, `BAAI/AquilaChat-7B`, etc.)
+- Baichuan & Baichuan2 (`baichuan-inc/Baichuan2-13B-Chat`, `baichuan-inc/Baichuan-7B`, etc.)
+- BLOOM (`bigscience/bloom`, `bigscience/bloomz`, etc.)
+- ChatGLM (`THUDM/chatglm2-6b`, `THUDM/chatglm3-6b`, etc.)
+- DeciLM (`Deci/DeciLM-7B`, `Deci/DeciLM-7B-instruct`, etc.)
+- Falcon (`tiiuae/falcon-7b`, `tiiuae/falcon-40b`, `tiiuae/falcon-rw-7b`, etc.)
+- Gemma (`google/gemma-2b`, `google/gemma-7b`, etc.)
+- GPT-2 (`gpt2`, `gpt2-xl`, etc.)
+- GPT BigCode (`bigcode/starcoder`, `bigcode/gpt_bigcode-santacoder`, etc.)
+- GPT-J (`EleutherAI/gpt-j-6b`, `nomic-ai/gpt4all-j`, etc.)
+- GPT-NeoX (`EleutherAI/gpt-neox-20b`, `databricks/dolly-v2-12b`, `stabilityai/stablelm-tuned-alpha-7b`, etc.)
+- InternLM (`internlm/internlm-7b`, `internlm/internlm-chat-7b`, etc.)
+- InternLM2 (`internlm/internlm2-7b`, `internlm/internlm2-chat-7b`, etc.)
+- LLaMA & LLaMA-2 (`meta-llama/Llama-2-70b-hf`, `lmsys/vicuna-13b-v1.3`, `young-geng/koala`, `openlm-research/open_llama_13b`, etc.)
+- Mistral (`mistralai/Mistral-7B-v0.1`, `mistralai/Mistral-7B-Instruct-v0.1`, etc.)
+- Mixtral (`mistralai/Mixtral-8x7B-v0.1`, `mistralai/Mixtral-8x7B-Instruct-v0.1`, etc.)
+- MPT (`mosaicml/mpt-7b`, `mosaicml/mpt-30b`, etc.)
+- OLMo (`allenai/OLMo-1B`, `allenai/OLMo-7B`, etc.)
+- OPT (`facebook/opt-66b`, `facebook/opt-iml-max-30b`, etc.)
+- Phi (`microsoft/phi-1_5`, `microsoft/phi-2`, etc.)
+- Qwen (`Qwen/Qwen-7B`, `Qwen/Qwen-7B-Chat`, etc.)
+- Qwen2 (`Qwen/Qwen2-7B-beta`, `Qwen/Qwen-7B-Chat-beta`, etc.)
+- StableLM(`stabilityai/stablelm-3b-4e1t`, `stabilityai/stablelm-base-alpha-7b-v2`, etc.)
+- Yi (`01-ai/Yi-6B`, `01-ai/Yi-34B`, etc.)
+
+Install vLLM with pip or [from source](https://vllm.readthedocs.io/en/latest/getting_started/installation.html#build-from-source):
 
 ```bash
-pip install psutil numpy ray torch
-pip install git+https://github.com/huggingface/transformers  # Required for LLaMA.
-pip install sentencepiece  # Required for LlamaTokenizer.
-pip install ninja  # To parallelize the compilation of flash-attn.
-pip install flash-attn  # This may take up to 10 mins.
-pip install -e .
+pip install vllm
 ```
 
-## Test simple server
+## Getting Started
 
-```bash
-ray start --head
-python simple_server.py
+Visit our [documentation](https://vllm.readthedocs.io/en/latest/) to get started.
+- [Installation](https://vllm.readthedocs.io/en/latest/getting_started/installation.html)
+- [Quickstart](https://vllm.readthedocs.io/en/latest/getting_started/quickstart.html)
+- [Supported Models](https://vllm.readthedocs.io/en/latest/models/supported_models.html)
+
+## Contributing
+
+We welcome and value any contributions and collaborations.
+Please check out [CONTRIBUTING.md](./CONTRIBUTING.md) for how to get involved.
+
+## Citation
+
+If you use vLLM for your research, please cite our [paper](https://arxiv.org/abs/2309.06180):
+```bibtex
+@inproceedings{kwon2023efficient,
+  title={Efficient Memory Management for Large Language Model Serving with PagedAttention},
+  author={Woosuk Kwon and Zhuohan Li and Siyuan Zhuang and Ying Sheng and Lianmin Zheng and Cody Hao Yu and Joseph E. Gonzalez and Hao Zhang and Ion Stoica},
+  booktitle={Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles},
+  year={2023}
+}
 ```
-
-The detailed arguments for `simple_server.py` can be found by:
-```bash
-python simple_server.py --help
-```
-
-## FastAPI server
-
-Install the following additional dependencies:
-```bash
-pip install fastapi uvicorn
-```
-
-To start the server:
-```bash
-ray start --head
-python -m cacheflow.http_frontend.fastapi_frontend
-```
-
-To test the server:
-```bash
-python -m cacheflow.http_frontend.test_cli_client
-```
-
-## Gradio web server
-
-Install the following additional dependencies:
-```bash
-pip install gradio
-```
-
-Start the server:
-```bash
-python -m cacheflow.http_frontend.fastapi_frontend
-# At another terminal
-python -m cacheflow.http_frontend.gradio_webserver
-```
-
-## Load LLaMA weights
-
-Since LLaMA weight is not fully public, we cannot directly download the LLaMA weights from huggingface. Therefore, you need to follow the following process to load the LLaMA weights.
-
-1. Converting LLaMA weights to huggingface format with [this script](https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/convert_llama_weights_to_hf.py).
-    ```bash
-    python src/transformers/models/llama/convert_llama_weights_to_hf.py \
-        --input_dir /path/to/downloaded/llama/weights --model_size 7B --output_dir /output/path/llama-7b
-    ```
-    Please make sure that `llama` is included in the output directory name.
-2. For all the commands above, specify the model with `--model /output/path/llama-7b` to load the model. For example:
-    ```bash
-    python simple_server.py --model /output/path/llama-7b
-    python -m cacheflow.http_frontend.fastapi_frontend --model /output/path/llama-7b
-    ```

benchmark/benchmark_attention.py

@@ -1,165 +0,0 @@
-import functools
-import random
-import time
-from typing import List
-
-from flash_attn.flash_attn_interface import _flash_attn_forward
-import torch
-
-from cacheflow import attention_ops
-
-
-def benchmark(name, f, num_warmup = 10, num_iters = 100):
-    for _ in range(num_warmup):
-        f()
-    torch.cuda.synchronize()
-
-    start = time.time()
-    for _ in range(num_iters):
-        f()
-    torch.cuda.synchronize()
-    end = time.time()
-    print(f'{name}: {(end - start) / num_iters * 1000:.3f} ms')
-
-
-@torch.inference_mode()
-def benchmark_multi_query_cached_kv_attention(
-    query_lens: List[int],
-    context_lens: List[int],
-    num_heads: int,
-    head_size: int,
-    block_size: int,
-    num_blocks: int,
-    dtype: torch.dtype,
-) -> None:
-    print(f'query_lens: {query_lens}, context_lens: {context_lens}, '
-          f'num_heads: {num_heads}, head_size: {head_size}, block_size: '
-          f'{block_size}, num_blocks: {num_blocks}, dtype: {dtype}')
-    # Create query tensor.
-    num_queries = len(query_lens)
-    cu_query_lens = [0]
-    for query_len in query_lens:
-        cu_query_lens.append(cu_query_lens[-1] + query_len)
-    num_total_tokens = cu_query_lens[-1]
-    qkv = torch.randn(
-        num_total_tokens, 3, num_heads, head_size, dtype=dtype, device='cuda')
-    query, _, _ = qkv.unbind(dim=1)
-
-    # Create key and value cache.
-    x = 16 // torch.tensor([], dtype=dtype).element_size()
-    key_block_shape = (num_heads, head_size // x, block_size, x)
-    key_cache = torch.randn(
-        size=(num_blocks, *key_block_shape), dtype=dtype, device='cuda')
-    value_block_shape = (num_heads, head_size, block_size)
-    value_cache = torch.randn(
-        size=(num_blocks, *value_block_shape), dtype=dtype, device='cuda')
-
-    # Create block tables.
-    max_context_len = max(context_lens)
-    max_num_blocks_per_seq = (max_context_len + block_size - 1) // block_size
-    block_tables = []
-    for _ in range(num_queries):
-        block_table = [
-            random.randint(0, num_blocks - 1)
-            for _ in range(max_num_blocks_per_seq)
-        ]
-        block_tables.append(block_table)
-    block_tables = torch.tensor(block_tables, dtype=torch.int, device='cuda')
-
-    # Create input and output data structures.
-    cu_query_lens = torch.tensor(cu_query_lens, dtype=torch.int, device='cuda')
-    context_len_tensor = torch.tensor(context_lens, dtype=torch.int, device='cuda')
-    scale = float(1.0 / (head_size ** 0.5))
-    output = torch.empty(
-        num_total_tokens, num_heads, head_size, dtype=dtype, device='cuda')
-
-    # Run our implementation.
-    def run_ours():
-        attention_ops.multi_query_cached_kv_attention(
-            cu_query_lens,
-            output,
-            query,
-            key_cache,
-            value_cache,
-            scale,
-            block_tables,
-            context_len_tensor,
-            block_size,
-            max_context_len,
-        )
-    benchmark('Ours', run_ours)
-
-    # Upper bound: Flash attention.
-    # Becuase Flash attention cannot read our own cache,
-    # we make key and value tensors contiguous.
-    num_kv_tokens = sum(context_lens)
-    cu_context_lens = [0]
-    for context_len in context_lens:
-        cu_context_lens.append(cu_context_lens[-1] + context_len)
-    cu_context_lens = torch.tensor(cu_context_lens, dtype=torch.int, device='cuda')
-    qkv = torch.randn(
-        num_kv_tokens, 3, num_heads, head_size, dtype=dtype, device='cuda')
-    _, key, value = qkv.unbind(dim=1)
-    ref_output = torch.empty_like(output)
-
-    # Run Flash attention.
-    def run_flash_attn():
-        _flash_attn_forward(
-            query,
-            key,
-            value,
-            ref_output,
-            cu_query_lens,
-            cu_context_lens,
-            max(query_lens),
-            max_context_len,
-            dropout_p=0.0,
-            softmax_scale=scale,
-            causal=True,
-            return_softmax=False,
-        )
-    benchmark('Flash attention', run_flash_attn)
-
-
-if __name__ == '__main__':
-    BLOCK_SIZE = 8
-    NUM_BLOCKS = 1024
-    DTYPE = torch.half
-
-    # LLaMA-13B and OPT-13B
-    NUM_HEADS = 40
-    HEAD_SIZE = 128
-
-    run_benchmark = functools.partial(
-        benchmark_multi_query_cached_kv_attention,
-        num_heads=NUM_HEADS,
-        head_size=HEAD_SIZE,
-        block_size=BLOCK_SIZE,
-        num_blocks=NUM_BLOCKS,
-        dtype=DTYPE,
-    )
-
-    run_benchmark(
-        query_lens=[64] * 1,
-        context_lens=[64] * 1,
-    )
-    run_benchmark(
-        query_lens=[128] * 1,
-        context_lens=[128] * 1,
-    )
-    run_benchmark(
-        query_lens=[64] * 8,
-        context_lens=[64] * 8,
-    )
-    run_benchmark(
-        query_lens=[128] * 8,
-        context_lens=[128] * 8,
-    )
-    run_benchmark(
-        query_lens=[64, 32, 16],
-        context_lens=[128, 256, 64],
-    )
-    run_benchmark(
-        query_lens=[1024],
-        context_lens=[1024],
-    )

benchmark/benchmark_cache.py

@@ -1,81 +0,0 @@
-import functools
-import random
-import time
-
-import torch
-
-from cacheflow import cache_ops
-
-
-def benchmark(name, f, size: int, num_warmup = 10, num_iters = 100):
-    for _ in range(num_warmup):
-        f()
-    torch.cuda.synchronize()
-
-    start = time.time()
-    for _ in range(num_iters):
-        f()
-    torch.cuda.synchronize()
-    end = time.time()
-    avg_time = (end - start) / num_iters
-    print(f'[Latency] {name}: {avg_time * 1000:.3f} ms')
-    print(f'[Throughput] {name}: {size / avg_time / 2 ** 30:.3f} GB/s')
-
-
-@torch.inference_mode()
-def test_gather_cached_kv(
-    num_tokens: int,
-    num_heads: int,
-    head_size: int,
-    block_size: int,
-    num_blocks: int,
-    dtype: torch.dtype,
-) -> None:
-    print(f'num_tokens: {num_tokens}, num_heads: {num_heads}, '
-          f'head_size: {head_size}, block_size: {block_size}, '
-          f'num_blocks: {num_blocks}, dtype: {dtype}')
-
-    num_slots = block_size * num_blocks
-    slot_mapping = random.sample(range(num_slots), num_tokens)
-    slot_mapping = torch.tensor(slot_mapping, dtype=torch.int, device='cuda')
-
-    qkv = torch.randn(
-        num_tokens, 3, num_heads, head_size, dtype=dtype, device='cuda')
-    _, key, value = qkv.unbind(dim=1)
-
-    x = 16 // torch.tensor([], dtype=dtype).element_size()
-    key_cache_shape = (num_blocks, num_heads, head_size // x, block_size, x)
-    key_cache = torch.randn(size=key_cache_shape, dtype=dtype, device='cuda')
-
-    value_cache_shape = (num_blocks, num_heads, head_size, block_size)
-    value_cache = torch.randn(
-        size=value_cache_shape, dtype=dtype, device='cuda')
-
-    # Run Flash attention.
-    def run():
-        cache_ops.gather_cached_kv(key, value, key_cache, value_cache, slot_mapping)
-
-    benchmark('gather_cached_kv', run,
-              size=num_tokens * num_heads * head_size * 2 * qkv.element_size())
-
-
-if __name__ == '__main__':
-    BLOCK_SIZE = 8
-    NUM_BLOCKS = 1024
-    DTYPE = torch.half
-
-    # LLaMA-13B and OPT-13B
-    NUM_HEADS = 40
-    HEAD_SIZE = 128
-
-    run_benchmark = functools.partial(
-        test_gather_cached_kv,
-        num_heads=NUM_HEADS,
-        head_size=HEAD_SIZE,
-        block_size=BLOCK_SIZE,
-        num_blocks=NUM_BLOCKS,
-        dtype=DTYPE,
-    )
-
-    for i in range(6, 12):
-        run_benchmark(num_tokens=2 ** i)

benchmark/benchmark_latency.py

@@ -1,105 +0,0 @@
-import argparse
-import time
-from typing import List
-
-from tqdm import tqdm
-import numpy as np
-import torch
-
-from cacheflow.master.simple_frontend import SimpleFrontend
-from cacheflow.master.server import (Server, add_server_arguments,
-                                     initialize_ray_cluster)
-from cacheflow.sampling_params import SamplingParams
-from cacheflow.utils import get_gpu_memory, get_cpu_memory
-
-
-def main(args: argparse.Namespace):
-    # TODO(zhuohan): Support pipeline parallelism.
-    assert args.pipeline_parallel_size == 1, (
-        'Pipeline parallelism is not supported yet.')
-
-    (num_nodes, num_devices_per_node, distributed_init_method,
-    all_stage_devices) = (
-        initialize_ray_cluster(
-            address='local',
-            pipeline_parallel_size=args.pipeline_parallel_size,
-            tensor_parallel_size=args.tensor_parallel_size))
-
-    # Create a server.
-    server = Server(
-        model=args.model,
-        model_path=args.model_path,
-        use_dummy_weights=args.use_dummy_weights,
-        pipeline_parallel_size=args.pipeline_parallel_size,
-        tensor_parallel_size=args.tensor_parallel_size,
-        block_size=args.block_size,
-        dtype=args.dtype,
-        seed=args.seed,
-        swap_space=args.swap_space,
-        max_num_batched_tokens=args.max_num_batched_tokens,
-        max_num_sequences=args.max_num_sequences,
-        num_nodes=num_nodes,
-        num_devices_per_node=num_devices_per_node,
-        distributed_init_method=distributed_init_method,
-        all_stage_devices=all_stage_devices,
-        gpu_memory=get_gpu_memory(),
-        cpu_memory=get_cpu_memory(),
-    )
-
-    # Create a frontend.
-    frontend = SimpleFrontend(
-        model_name=args.model,
-        block_size=args.block_size,
-    )
-    sampling_params_dict = {
-        'n': args.n,
-        'temperature': 0.0 if args.use_beam_search else 1.0,
-        'top_p': 1.0,
-        'use_beam_search': args.use_beam_search,
-        'stop_token_ids': set(),
-        'max_num_steps': args.output_len,
-    }
-    sampling_params = SamplingParams.from_dict(sampling_params_dict)
-    print(sampling_params)
-    input_token_ids = [0] * args.input_len
-
-    def profile_step(profile=False):
-        if profile:
-            torch.cuda.cudart().cudaProfilerStart()
-        for _ in range(args.batch_size):
-            frontend._add_query(input_token_ids, sampling_params)
-        server.add_sequence_groups(frontend.get_inputs())
-        start_time = time.time()
-        while True:
-            server.step()
-            if not server.has_unfinished_requests():
-                break
-        end_time = time.time()
-        latency = end_time - start_time
-        if profile:
-            torch.cuda.cudart().cudaProfilerStop()
-        return latency
-
-    print("Warm up step")
-    profile_step()
-
-    # Benchmark.
-    latencies = []
-    for _ in tqdm(range(3), desc="Profile step"):
-        latencies.append(profile_step())
-    print(f'Avg latency: {np.mean(latencies)} seconds')
-
-
-if __name__ == '__main__':
-    parser = argparse.ArgumentParser(description='CacheFlow simple server.')
-    parser = add_server_arguments(parser)
-    parser.add_argument('--input-len', type=int, default=32)
-    parser.add_argument('--output-len', type=int, default=128)
-    parser.add_argument('--batch-size', type=int, default=8)
-    parser.add_argument('--n', type=int, default=1)
-    parser.add_argument('--use-beam-search', action='store_true')
-    args = parser.parse_args()
-    args.max_num_batched_tokens = max(
-        args.max_num_batched_tokens, args.batch_size * args.input_len)
-    print(args)
-    main(args)

benchmark/benchmark_text_completion.py

@@ -1,290 +0,0 @@
-import argparse
-import logging
-import os
-import pickle
-import time
-from typing import List
-
-from tqdm import tqdm
-from transformers import AutoConfig
-
-from benchmark.trace import generate_text_completion_requests
-from cacheflow.master.simple_frontend import SimpleFrontend
-from cacheflow.master.server import (Server, add_server_arguments,
-                                     initialize_ray_cluster)
-from cacheflow.sampling_params import SamplingParams
-from cacheflow.utils import get_gpu_memory, get_cpu_memory
-
-
-logger = logging.getLogger(__name__)
-
-
-def main(args: argparse.Namespace):
-    assert args.pipeline_parallel_size == 1, (
-        'Pipeline parallelism is not supported yet.')
-
-    (num_nodes, num_devices_per_node, distributed_init_method,
-    all_stage_devices) = (
-        initialize_ray_cluster(
-            address='local',
-            pipeline_parallel_size=args.pipeline_parallel_size,
-            tensor_parallel_size=args.tensor_parallel_size))
-
-    # Create a server.
-    server = Server(
-        model=args.model,
-        model_path=args.model_path,
-        use_dummy_weights=args.use_dummy_weights,
-        pipeline_parallel_size=args.pipeline_parallel_size,
-        tensor_parallel_size=args.tensor_parallel_size,
-        block_size=args.block_size,
-        dtype=args.dtype,
-        seed=args.seed,
-        swap_space=args.swap_space,
-        max_num_batched_tokens=args.max_num_batched_tokens,
-        max_num_sequences=args.max_num_sequences,
-        num_nodes=num_nodes,
-        num_devices_per_node=num_devices_per_node,
-        distributed_init_method=distributed_init_method,
-        all_stage_devices=all_stage_devices,
-        gpu_memory=get_gpu_memory(),
-        cpu_memory=get_cpu_memory(),
-        collect_stats=True,
-        do_memory_analysis=args.do_memory_analysis,
-    )
-
-    # Create a frontend.
-    frontend = SimpleFrontend(
-        model_name=args.model,
-        block_size=args.block_size,
-    )
-    # Generate requests.
-    requests = generate_text_completion_requests(
-        args.dataset,
-        args.request_rate,
-        args.duration,
-        args.seed,
-        args.n1,
-        args.n2,
-        args.n3,
-        args.n4,
-        args.n6,
-        args.n2_beam,
-        args.n4_beam,
-        args.n6_beam,
-        args.n8_beam,
-    )
-
-    # Warm up.
-    logger.info('Warming up.')
-    num_warmup_requests = 8
-    warmup_input_len = 8
-    warmup_output_len = 32
-    warmup_sampling_params = SamplingParams(
-        n=1,
-        temperature=1.0,
-        top_p=0.99,
-        max_num_steps=warmup_output_len,
-        use_beam_search=False,
-        stop_token_ids=set(),
-        num_logprobs=0,
-        context_window_size=None,
-    )
-    for _ in range(num_warmup_requests):
-        frontend._add_query([0] * warmup_input_len, warmup_sampling_params)
-    server.add_sequence_groups(frontend.get_inputs())
-    while True:
-        server.step()
-        if not server.has_unfinished_requests():
-            break
-
-    # Start benchmarking.
-    logger.info('Start benchmarking.')
-    # Initialize tqdm.
-    pbar = tqdm(total=len(requests), desc='Finished requests')
-
-    finished = []
-    server.scheduler.reset_stats()
-    start_time = time.time()
-    while True:
-        now = time.time()
-        if args.timeout is not None and now - start_time > args.timeout:
-            logger.info('Timeout. Stop benchmarking.')
-            break
-
-        while requests:
-            if requests[0][0] <= now - start_time:
-                request_time, input_tokens, sampling_params = requests.pop(0)
-                frontend._add_query(
-                    input_tokens, sampling_params, arrival_time=start_time + request_time)
-            else:
-                break
-        server.add_sequence_groups(frontend.get_inputs())
-        updated_seq_groups = server.step()
-
-        now = time.time()
-        for seq_group in updated_seq_groups:
-            if not seq_group.is_finished():
-                continue
-            arrival_time = seq_group.arrival_time
-            finish_time = now
-            for seq in seq_group.get_seqs():
-                seq_len = seq.get_len()
-                output_len = seq_len - seq.prompt_len
-                finished.append({
-                    'group_id': seq_group.group_id,
-                    'seq_id': seq.seq_id,
-                    'arrival_time': arrival_time, 
-                    'finish_time': finish_time,
-                    'prompt_len': seq.prompt_len,
-                    'output_len': output_len,
-                })
-            pbar.update(1)
-
-        if not (requests or server.has_unfinished_requests()):
-            break
-    pbar.close()
-    logger.info('Finish benchmarking. Saving stats.')
-    server.scheduler.save_stats(args.output_dir)
-    with open(os.path.join(args.output_dir, 'sequences.pkl'), 'wb') as f:
-        pickle.dump(finished, f)
-    logger.info('Done.')
-
-
-def get_model_name(model: str) -> str:
-    OPT_MODELS = [
-        'opt-125m',
-        'opt-350m',
-        'opt-1.3b',
-        'opt-2.7b',
-        'opt-6.7b',
-        'opt-13b',
-        'opt-30b',
-        'opt-66b',
-        'opt-175b',
-    ]
-    for opt_model in OPT_MODELS:
-        if opt_model in model:
-            return opt_model
-
-    config = AutoConfig.from_pretrained(model)
-    assert config.model_type == 'llama'
-    hidden_size = config.hidden_size
-    if hidden_size == 4096:
-        return 'llama-7b'
-    elif hidden_size == 5120:
-        return 'llama-13b'
-    elif hidden_size == 6656:
-        return 'llama-30b'
-    elif hidden_size == 8192:
-        return 'llama-65b'
-    else:
-        raise ValueError(f'Unknown model: {model}')
-
-
-def get_dataset_name(dataset: str) -> str:
-    if 'sharegpt' in dataset.lower():
-        return 'sharegpt'
-    elif 'alpaca' in dataset.lower():
-        return 'alpaca'
-    else:
-        raise ValueError(f'Unknown dataset: {dataset}')
-
-
-def get_sampling_dir_name(
-    n1: float,
-    n2: float,
-    n3: float,
-    n4: float,
-    n6: float,
-    n2_beam: float,
-    n4_beam: float,
-    n6_beam: float,
-    n8_beam: float,
-) -> str:
-    method = ''
-    if n1 > 0.0:
-        method = 'n1' if n1 == 1.0 else method + f'n1-{n1}-'
-    if n2 > 0.0:
-        method = 'n2' if n2 == 1.0 else method + f'n2-{n2}-'
-    if n3 > 0.0:
-        method = 'n3' if n3 == 1.0 else method + f'n3-{n3}-'
-    if n4 > 0.0:
-        method = 'n4' if n4 == 1.0 else method + f'n4-{n4}-'
-    if n6 > 0.0:
-        method = 'n6' if n6 == 1.0 else method + f'n6-{n6}-'
-    if n2_beam > 0.0:
-        method = 'n2-beam' if n2_beam == 1.0 else method + f'n2-beam-{n2_beam}-'
-    if n4_beam > 0.0:
-        method = 'n4-beam' if n4_beam == 1.0 else method + f'n4-beam-{n4_beam}-'
-    if n6_beam > 0.0:
-        method = 'n6-beam' if n6_beam == 1.0 else method + f'n6-beam-{n6_beam}-'
-    if n8_beam > 0.0:
-        method = 'n8-beam' if n8_beam == 1.0 else method + f'n8-beam-{n8_beam}-'
-    return method[:-1] if method.endswith('-') else method
-
-
-if __name__ == '__main__':
-    parser = argparse.ArgumentParser(description='CacheFlow simple server.')
-    parser = add_server_arguments(parser) 
-    parser.add_argument('--output-dir', type=str, help='path to output directory', default=None)
-
-    parser.add_argument('--dataset', type=str, help='path to dataset', required=True)
-    parser.add_argument('--request-rate', type=float, help='reqs/sec', required=True)
-    parser.add_argument('--duration', type=int, help='duration in seconds', required=True)
-    parser.add_argument('--do-memory-analysis', action='store_true',
-        help='do memory analysis (This will lower the throughput. Use this only for analysis.)')
-    parser.add_argument('--timeout', type=int, help='time out in seconds', default=None)
-
-    parser.add_argument('--n1', type=float, help='ratio of requests with n=1', default=0.0)
-    parser.add_argument('--n2', type=float, help='ratio of requests with n=2', default=0.0)
-    parser.add_argument('--n3', type=float, help='ratio of requests with n=3', default=0.0)
-    parser.add_argument('--n4', type=float, help='ratio of requests with n=4', default=0.0)
-    parser.add_argument('--n6', type=float, help='ratio of requests with n=6', default=0.0)
-    parser.add_argument('--n2-beam', type=float, help='ratio of requests with n=2 & beam search', default=0.0)
-    parser.add_argument('--n4-beam', type=float, help='ratio of requests with n=4 & beam search', default=0.0)
-    parser.add_argument('--n6-beam', type=float, help='ratio of requests with n=6 & beam search', default=0.0)
-    parser.add_argument('--n8-beam', type=float, help='ratio of requests with n=8 & beam search', default=0.0)
-    args = parser.parse_args()
-    if args.n1 + args.n2 + args.n3 + args.n4 + args.n6 + args.n2_beam + args.n4_beam + args.n6_beam + args.n8_beam != 1.0:
-        raise ValueError('The ratios of requests must sum to 1.')
-
-    model_name = get_model_name(args.model)
-    dataset_name = get_dataset_name(args.dataset)
-    if 'opt' in model_name:
-        if 'opt' not in args.dataset.lower():
-            raise ValueError(f'OPT models can only be used with OPT datasets.')
-    elif 'llama' in model_name:
-        if 'llama' not in args.dataset.lower():
-            raise ValueError(f'Llama models can only be used with Llama datasets.')
-
-    dataset_name = 'sharegpt' if 'sharegpt' in args.dataset else 'alpaca'
-    sample_dir = get_sampling_dir_name(
-        args.n1, args.n2, args.n3, args.n4, args.n6, args.n2_beam, args.n4_beam, args.n6_beam, args.n8_beam)
-    if args.output_dir is None:
-        args.output_dir = os.path.join(
-            '../exp',
-            dataset_name,
-            f'{model_name}-tp{args.tensor_parallel_size}',
-            sample_dir,
-            'cacheflow',
-            f'block{args.block_size}',
-            f'req-rate-{args.request_rate}',
-            f'seed{args.seed}',
-            f'duration-{args.duration}',
-        )
-    os.makedirs(args.output_dir, exist_ok=True)
-
-    # Set up logging.
-    logging.basicConfig(
-        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
-        datefmt="%m/%d/%Y %H:%M:%S",
-        level=logging.INFO,
-        handlers=[
-            logging.StreamHandler(),
-            logging.FileHandler(os.path.join(args.output_dir, 'log.txt')),
-        ],
-    )
-    logger.info(args)
-
-    main(args)

benchmark/trace.py

@@ -1,116 +0,0 @@
-import pickle
-import random
-from typing import List, Tuple
-
-import numpy as np
-
-from cacheflow.sampling_params import SamplingParams
-
-
-def generate_text_completion_requests(
-    dataset: str,
-    request_rate: float,
-    duration: int,
-    seed: int,
-    n1: float = 0.0,
-    n2: float = 0.0,
-    n3: float = 0.0,
-    n4: float = 0.0,
-    n6: float = 0.0,
-    n2_beam: float = 0.0,
-    n4_beam: float = 0.0,
-    n6_beam: float = 0.0,
-    n8_beam: float = 0.0,
-    max_seq_len: int = 2048,
-    time_quantum: int = 10,
-) -> List[Tuple[float, List[int], SamplingParams]]:
-    random.seed(seed)
-    np.random.seed(seed)
-
-    # Generate timestamps for requests using Poisson distribution.
-    lam = request_rate * (time_quantum / 1000)
-    quantums_per_sec = 1000 / time_quantum
-    arrival_times = np.random.poisson(
-        lam=lam, size=int(duration * quantums_per_sec))
-    timestamps = []
-    for i, n in enumerate(arrival_times):
-        timestamps += [i * (time_quantum / 1000)] * n
-
-    # Load and shuffle the dataset.
-    num_requests = len(timestamps)
-    with open(dataset, 'rb') as f:
-        data = pickle.load(f)
-
-    filtered = []
-    for pair in data:
-        input_tokens, output_tokens = pair
-        input_len = len(input_tokens)
-        output_len = len(output_tokens)
-        # Filter out too long sequences.
-        if input_len + output_len < max_seq_len:
-            # Output tokens are not needed for the benchmark.
-            filtered.append((input_tokens, output_len))
-
-    data = []
-    while len(data) < num_requests:
-        data += filtered
-    data = data[:num_requests]
-    # Shuffle the data.
-    assert len(data) == len(timestamps)
-    random.shuffle(data)
-
-    random_sampling_params_dict = {
-        'temperature': 1.0,
-        'top_p': 1.0,
-        'use_beam_search': False,
-        'stop_token_ids': set(),
-        'num_logprobs': 0,
-        'context_window_size': None,
-    }
-    beam_search_params_dict = {
-        'temperature': 0.0,
-        'top_p': 1.0,
-        'use_beam_search': True,
-        'stop_token_ids': set(),
-        'num_logprobs': 0,
-        'context_window_size': None,
-    }
-
-    # Generate requests based on the sampling parameter ratio.
-    requests = []
-    assert n1 + n2 + n3 + n4 + n6 + n2_beam + n4_beam + n6_beam + n8_beam == 1.0
-    cum_sum = 0
-    for timestamp, pair in zip(timestamps, data):
-        input_tokens, output_len = pair
-        if cum_sum < n1 * num_requests:
-            sampling_params = SamplingParams(
-                n=1, max_num_steps=output_len, **random_sampling_params_dict)
-        elif cum_sum < (n1 + n2) * num_requests:
-            sampling_params = SamplingParams(
-                n=2, max_num_steps=output_len, **random_sampling_params_dict)
-        elif cum_sum < (n1 + n2 + n3) * num_requests:
-            sampling_params = SamplingParams(
-                n=3, max_num_steps=output_len, **random_sampling_params_dict)
-        elif cum_sum < (n1 + n2 + n3 + n4) * num_requests:
-            sampling_params = SamplingParams(
-                n=4, max_num_steps=output_len, **random_sampling_params_dict)
-        elif cum_sum < (n1 + n2 + n3 + n4 + n6) * num_requests:
-            sampling_params = SamplingParams(
-                n=6, max_num_steps=output_len, **random_sampling_params_dict)
-        elif cum_sum < (n1 + n2 + n3 + n4 + n6 + n2_beam) * num_requests:
-            sampling_params = SamplingParams(
-                n=2, max_num_steps=output_len, **beam_search_params_dict)
-        elif cum_sum < (n1 + n2 + n3 + n4 + n6 + n2_beam + n4_beam) * num_requests:
-            sampling_params = SamplingParams(
-                n=4, max_num_steps=output_len, **beam_search_params_dict)
-        elif cum_sum < (n1 + n2 + n3 + n4 + n6 + n2_beam + n4_beam + n6_beam) * num_requests:
-            sampling_params = SamplingParams(
-                n=6, max_num_steps=output_len, **beam_search_params_dict)
-        elif cum_sum < (n1 + n2 + n3 + n4 + n6 + n2_beam + n4_beam + n6_beam + n8_beam) * num_requests:
-            sampling_params = SamplingParams(
-                n=8, max_num_steps=output_len, **beam_search_params_dict)
-        else:
-            raise ValueError('Invalid request ratio.')
-        cum_sum += 1
-        requests.append((timestamp, input_tokens, sampling_params))
-    return requests

benchmarks/README.md

@@ -0,0 +1,8 @@
+# Benchmarking vLLM
+
+## Downloading the ShareGPT dataset
+
+You can download the dataset by running:
+```bash
+wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
+```

benchmarks/backend_request_func.py

@@ -0,0 +1,284 @@
+import json
+import os
+import time
+from dataclasses import dataclass
+from typing import Optional
+
+import aiohttp
+from tqdm.asyncio import tqdm
+
+AIOHTTP_TIMEOUT = aiohttp.ClientTimeout(total=6 * 60 * 60)
+
+
+@dataclass
+class RequestFuncInput:
+    prompt: str
+    api_url: str
+    prompt_len: int
+    output_len: int
+    model: str
+    best_of: int = 1
+    use_beam_search: bool = False
+
+
+@dataclass
+class RequestFuncOutput:
+    generated_text: str = ""
+    success: bool = False
+    latency: float = 0
+    ttft: float = 0
+    prompt_len: int = 0
+
+
+async def async_request_tgi(
+    request_func_input: RequestFuncInput,
+    pbar: Optional[tqdm] = None,
+) -> RequestFuncOutput:
+    api_url = request_func_input.api_url
+    assert api_url.endswith("generate_stream")
+
+    async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
+        assert not request_func_input.use_beam_search
+        params = {
+            "best_of": request_func_input.best_of,
+            "max_new_tokens": request_func_input.output_len,
+            "do_sample": True,
+            "temperature": 0.01,  # TGI does not accept 0.0 temperature.
+            "top_p": 0.99,  # TGI does not accept 1.0 top_p.
+        }
+        payload = {
+            "inputs": request_func_input.prompt,
+            "parameters": params,
+        }
+        output = RequestFuncOutput()
+        output.prompt_len = request_func_input.prompt_len
+
+        ttft = 0
+        st = time.perf_counter()
+        try:
+            async with session.post(url=api_url, json=payload) as response:
+                if response.status == 200:
+                    async for data in response.content.iter_any():
+                        if ttft == 0:
+                            ttft = time.perf_counter() - st
+                            output.ttft = ttft
+                    output.latency = time.perf_counter() - st
+
+                    body = data.decode("utf-8").lstrip("data:")
+                    output.generated_text = json.loads(body)["generated_text"]
+                    output.success = True
+                else:
+                    output.success = False
+        except (aiohttp.ClientOSError, aiohttp.ServerDisconnectedError):
+            output.success = False
+
+        if pbar:
+            pbar.update(1)
+        return output
+
+
+async def async_request_vllm(
+    request_func_input: RequestFuncInput,
+    pbar: Optional[tqdm] = None,
+) -> RequestFuncOutput:
+    api_url = request_func_input.api_url
+    assert api_url.endswith("generate")
+
+    async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
+        payload = {
+            "prompt": request_func_input.prompt,
+            "n": 1,
+            "best_of": request_func_input.best_of,
+            "use_beam_search": request_func_input.use_beam_search,
+            "temperature": 0.0 if request_func_input.use_beam_search else 1.0,
+            "top_p": 1.0,
+            "max_tokens": request_func_input.output_len,
+            "ignore_eos": True,
+            "stream": True,
+        }
+        output = RequestFuncOutput()
+        output.prompt_len = request_func_input.prompt_len
+
+        ttft = 0
+        st = time.perf_counter()
+        try:
+            async with session.post(url=api_url, json=payload) as response:
+                if response.status == 200:
+                    async for data in response.content.iter_any():
+                        if ttft == 0:
+                            ttft = time.perf_counter() - st
+                            output.ttft = ttft
+                    output.latency = time.perf_counter() - st
+
+                    # When streaming, '\0' is appended to the end of the response.
+                    body = data.decode("utf-8").strip("\0")
+                    output.generated_text = json.loads(
+                        body)["text"][0][len(request_func_input.prompt):]
+                    output.success = True
+
+                else:
+                    output.success = False
+        except (aiohttp.ClientOSError, aiohttp.ServerDisconnectedError):
+            output.success = False
+
+        if pbar:
+            pbar.update(1)
+        return output
+
+
+async def async_request_trt_llm(
+    request_func_input: RequestFuncInput,
+    pbar: Optional[tqdm] = None,
+) -> RequestFuncOutput:
+    api_url = request_func_input.api_url
+    assert api_url.endswith("generate_stream")
+
+    async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
+        assert not request_func_input.use_beam_search
+        assert request_func_input.best_of == 1
+        payload = {
+            "accumulate_tokens": True,
+            "text_input": request_func_input.prompt,
+            "temperature": 0.0,
+            "top_p": 1.0,
+            "max_tokens": request_func_input.output_len,
+            "stream": True,
+        }
+        output = RequestFuncOutput()
+        output.prompt_len = request_func_input.prompt_len
+        ttft = 0
+
+        st = time.perf_counter()
+        try:
+            async with session.post(url=api_url, json=payload) as resp:
+                if resp.status == 200:
+                    async for data in resp.content.iter_any():
+                        if ttft == 0:
+                            ttft = time.perf_counter() - st
+                            output.ttft = ttft
+                    output.latency = time.perf_counter() - st
+
+                    body = data.decode("utf-8").lstrip("data:")
+                    output.generated_text = json.loads(body)["text_output"]
+                    output.success = True
+
+                else:
+                    output.success = False
+        except (aiohttp.ClientOSError, aiohttp.ServerDisconnectedError):
+            output.success = False
+
+        if pbar:
+            pbar.update(1)
+        return output
+
+
+async def async_request_deepspeed_mii(
+    request_func_input: RequestFuncInput,
+    pbar: Optional[tqdm] = None,
+) -> RequestFuncOutput:
+    async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
+        assert request_func_input.best_of == 1
+        assert not request_func_input.use_beam_search
+
+        payload = {
+            "prompts": request_func_input.prompt,
+            "max_new_tokens": request_func_input.output_len,
+            "ignore_eos": True,
+            "do_sample": True,
+            "temperature":
+            0.01,  # deepspeed-mii does not accept 0.0 temperature.
+            "top_p": 1.0,
+        }
+        output = RequestFuncOutput()
+        output.prompt_len = request_func_input.prompt_len
+
+        # DeepSpeed-MII doesn't support streaming as of Jan 28 2024, will use 0 as placeholder.
+        # https://github.com/microsoft/DeepSpeed-MII/pull/311
+        output.ttft = 0
+
+        st = time.perf_counter()
+        try:
+            async with session.post(url=request_func_input.api_url,
+                                    json=payload) as resp:
+                if resp.status == 200:
+                    parsed_resp = await resp.json()
+                    output.latency = time.perf_counter() - st
+                    output.generated_text = parsed_resp[0]["generated_text"]
+                    output.success = True
+                else:
+                    output.success = False
+        except (aiohttp.ClientOSError, aiohttp.ServerDisconnectedError):
+            output.success = False
+
+        if pbar:
+            pbar.update(1)
+        return output
+
+
+async def async_request_openai_completions(
+    request_func_input: RequestFuncInput,
+    pbar: Optional[tqdm] = None,
+) -> RequestFuncOutput:
+    api_url = request_func_input.api_url
+    assert api_url.endswith("v1/completions")
+
+    async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
+        assert not request_func_input.use_beam_search
+        payload = {
+            "model": request_func_input.model,
+            "prompt": request_func_input.prompt,
+            "temperature": 0.0,
+            "best_of": request_func_input.best_of,
+            "max_tokens": request_func_input.output_len,
+            "stream": True,
+        }
+        headers = {
+            "Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}"
+        }
+
+        output = RequestFuncOutput()
+        output.prompt_len = request_func_input.prompt_len
+
+        generated_text = ""
+        ttft = 0
+        st = time.perf_counter()
+        try:
+            async with session.post(url=api_url, json=payload,
+                                    headers=headers) as response:
+                if response.status == 200:
+                    async for chunk in response.content:
+                        if ttft == 0:
+                            ttft = time.perf_counter() - st
+                            output.ttft = ttft
+
+                        chunk = chunk.strip()
+                        if not chunk:
+                            continue
+
+                        chunk = chunk.decode("utf-8").lstrip("data: ")
+                        if chunk == "[DONE]":
+                            latency = time.perf_counter() - st
+                        else:
+                            body = json.loads(chunk)
+                            generated_text += body["choices"][0]["text"]
+
+                    output.generated_text = generated_text
+                    output.success = True
+                    output.latency = latency
+                else:
+                    output.success = False
+        except (aiohttp.ClientOSError, aiohttp.ServerDisconnectedError):
+            output.success = False
+
+    if pbar:
+        pbar.update(1)
+    return output
+
+
+ASYNC_REQUEST_FUNCS = {
+    "tgi": async_request_tgi,
+    "vllm": async_request_vllm,
+    "deepspeed-mii": async_request_deepspeed_mii,
+    "openai": async_request_openai_completions,
+    "tensorrt-llm": async_request_trt_llm,
+}

benchmarks/benchmark_latency.py

@@ -0,0 +1,149 @@
+"""Benchmark the latency of processing a single batch of requests."""
+import argparse
+import time
+from pathlib import Path
+from typing import Optional
+
+import numpy as np
+import torch
+from tqdm import tqdm
+
+from vllm import LLM, SamplingParams
+
+
+def main(args: argparse.Namespace):
+    print(args)
+
+    # NOTE(woosuk): If the request cannot be processed in a single batch,
+    # the engine will automatically process the request in multiple batches.
+    llm = LLM(
+        model=args.model,
+        tokenizer=args.tokenizer,
+        quantization=args.quantization,
+        tensor_parallel_size=args.tensor_parallel_size,
+        trust_remote_code=args.trust_remote_code,
+        dtype=args.dtype,
+        enforce_eager=args.enforce_eager,
+        kv_cache_dtype=args.kv_cache_dtype,
+        device=args.device,
+    )
+
+    sampling_params = SamplingParams(
+        n=args.n,
+        temperature=0.0 if args.use_beam_search else 1.0,
+        top_p=1.0,
+        use_beam_search=args.use_beam_search,
+        ignore_eos=True,
+        max_tokens=args.output_len,
+    )
+    print(sampling_params)
+    dummy_prompt_token_ids = np.random.randint(10000,
+                                               size=(args.batch_size,
+                                                     args.input_len))
+    dummy_prompt_token_ids = dummy_prompt_token_ids.tolist()
+
+    def run_to_completion(profile_dir: Optional[str] = None):
+        if profile_dir:
+            with torch.profiler.profile(
+                    activities=[
+                        torch.profiler.ProfilerActivity.CPU,
+                        torch.profiler.ProfilerActivity.CUDA,
+                    ],
+                    on_trace_ready=torch.profiler.tensorboard_trace_handler(
+                        str(profile_dir))) as p:
+                llm.generate(prompt_token_ids=dummy_prompt_token_ids,
+                             sampling_params=sampling_params,
+                             use_tqdm=False)
+            print(p.key_averages())
+        else:
+            start_time = time.perf_counter()
+            llm.generate(prompt_token_ids=dummy_prompt_token_ids,
+                         sampling_params=sampling_params,
+                         use_tqdm=False)
+            end_time = time.perf_counter()
+            latency = end_time - start_time
+            return latency
+
+    print("Warming up...")
+    run_to_completion(profile_dir=None)
+
+    if args.profile:
+        profile_dir = args.profile_result_dir
+        if not profile_dir:
+            profile_dir = Path(
+                "."
+            ) / "vllm_benchmark_result" / f"latency_result_{time.time()}"
+        print(f"Profiling (results will be saved to '{profile_dir}')...")
+        run_to_completion(profile_dir=profile_dir)
+        return
+
+    # Benchmark.
+    latencies = []
+    for _ in tqdm(range(args.num_iters), desc="Profiling iterations"):
+        latencies.append(run_to_completion(profile_dir=None))
+    print(f'Avg latency: {np.mean(latencies)} seconds')
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(
+        description='Benchmark the latency of processing a single batch of '
+        'requests till completion.')
+    parser.add_argument('--model', type=str, default='facebook/opt-125m')
+    parser.add_argument('--tokenizer', type=str, default=None)
+    parser.add_argument('--quantization',
+                        '-q',
+                        choices=['awq', 'gptq', 'squeezellm', None],
+                        default=None)
+    parser.add_argument('--tensor-parallel-size', '-tp', type=int, default=1)
+    parser.add_argument('--input-len', type=int, default=32)
+    parser.add_argument('--output-len', type=int, default=128)
+    parser.add_argument('--batch-size', type=int, default=8)
+    parser.add_argument('--n',
+                        type=int,
+                        default=1,
+                        help='Number of generated sequences per prompt.')
+    parser.add_argument('--use-beam-search', action='store_true')
+    parser.add_argument('--num-iters',
+                        type=int,
+                        default=3,
+                        help='Number of iterations to run.')
+    parser.add_argument('--trust-remote-code',
+                        action='store_true',
+                        help='trust remote code from huggingface')
+    parser.add_argument(
+        '--dtype',
+        type=str,
+        default='auto',
+        choices=['auto', 'half', 'float16', 'bfloat16', 'float', 'float32'],
+        help='data type for model weights and activations. '
+        'The "auto" option will use FP16 precision '
+        'for FP32 and FP16 models, and BF16 precision '
+        'for BF16 models.')
+    parser.add_argument('--enforce-eager',
+                        action='store_true',
+                        help='enforce eager mode and disable CUDA graph')
+    parser.add_argument(
+        "--kv-cache-dtype",
+        type=str,
+        choices=['auto', 'fp8_e5m2'],
+        default='auto',
+        help=
+        'Data type for kv cache storage. If "auto", will use model data type.')
+    parser.add_argument(
+        '--profile',
+        action='store_true',
+        help='profile the generation process of a single batch')
+    parser.add_argument(
+        '--profile-result-dir',
+        type=str,
+        default=None,
+        help=('path to save the pytorch profiler output. Can be visualized '
+              'with ui.perfetto.dev or Tensorboard.'))
+    parser.add_argument(
+        "--device",
+        type=str,
+        default="cuda",
+        choices=["cuda"],
+        help='device type for vLLM execution, supporting CUDA only currently.')
+    args = parser.parse_args()
+    main(args)

benchmarks/benchmark_serving.py

@@ -0,0 +1,387 @@
+"""Benchmark online serving throughput.
+
+On the server side, run one of the following commands:
+    (vLLM backend)
+    python -m vllm.entrypoints.api_server \
+        --model <your_model> --swap-space 16 \
+        --disable-log-requests
+
+    (TGI backend)
+    ./launch_hf_server.sh <your_model>
+
+On the client side, run:
+    python benchmarks/benchmark_serving.py \
+        --backend <backend> \
+        --tokenizer <your_model> --dataset <target_dataset> \
+        --request-rate <request_rate>
+"""
+import argparse
+import asyncio
+import json
+import random
+import time
+from dataclasses import dataclass
+from datetime import datetime
+from typing import AsyncGenerator, List, Tuple
+
+import numpy as np
+from tqdm.asyncio import tqdm
+from transformers import PreTrainedTokenizerBase
+from vllm.transformers_utils.tokenizer import get_tokenizer
+
+from backend_request_func import (
+    ASYNC_REQUEST_FUNCS,
+    RequestFuncInput,
+    RequestFuncOutput,
+)
+
+
+@dataclass
+class BenchmarkMetrics:
+    completed: int
+    total_input: int
+    total_output: int
+    request_throughput: float
+    input_throughput: float
+    output_throughput: float
+    mean_ttft_ms: float
+    median_ttft_ms: float
+    p99_ttft_ms: float
+    mean_tpot_ms: float
+    median_tpot_ms: float
+    p99_tpot_ms: float
+
+
+def sample_requests(
+    dataset_path: str,
+    num_requests: int,
+    tokenizer: PreTrainedTokenizerBase,
+) -> List[Tuple[str, int, int]]:
+    # Load the dataset.
+    with open(dataset_path) as f:
+        dataset = json.load(f)
+    # Filter out the conversations with less than 2 turns.
+    dataset = [data for data in dataset if len(data["conversations"]) >= 2]
+    # Only keep the first two turns of each conversation.
+    dataset = [(data["conversations"][0]["value"],
+                data["conversations"][1]["value"]) for data in dataset]
+
+    # some of these will be filtered out, so sample more than we need
+    sampled_indices = random.sample(range(len(dataset)),
+                                    int(num_requests * 1.2))
+    dataset = [dataset[i] for i in sampled_indices]
+
+    # Tokenize the prompts and completions.
+    prompts = [prompt for prompt, _ in dataset]
+    prompt_token_ids = tokenizer(prompts).input_ids
+    completions = [completion for _, completion in dataset]
+    completion_token_ids = tokenizer(completions).input_ids
+    tokenized_dataset = []
+    for i in range(len(dataset)):
+        output_len = len(completion_token_ids[i])
+        tokenized_dataset.append((prompts[i], prompt_token_ids[i], output_len))
+
+    # Filter out too long sequences.
+    filtered_dataset: List[Tuple[str, int, int]] = []
+    for prompt, prompt_token_ids, output_len in tokenized_dataset:
+        prompt_len = len(prompt_token_ids)
+        if prompt_len < 4 or output_len < 4:
+            # Prune too short sequences.
+            # This is because TGI causes errors when the input or output length
+            # is too short.
+            continue
+        if prompt_len > 1024 or prompt_len + output_len > 2048:
+            # Prune too long sequences.
+            continue
+        filtered_dataset.append((prompt, prompt_len, output_len))
+
+    # Sample the requests.
+    sampled_requests = random.sample(filtered_dataset, num_requests)
+    return sampled_requests
+
+
+async def get_request(
+    input_requests: List[Tuple[str, int, int]],
+    request_rate: float,
+) -> AsyncGenerator[Tuple[str, int, int], None]:
+    input_requests = iter(input_requests)
+    for request in input_requests:
+        yield request
+
+        if request_rate == float("inf"):
+            # If the request rate is infinity, then we don't need to wait.
+            continue
+        # Sample the request interval from the exponential distribution.
+        interval = np.random.exponential(1.0 / request_rate)
+        # The next request will be sent after the interval.
+        await asyncio.sleep(interval)
+
+
+def calculate_metrics(
+    input_requests: List[Tuple[str, int, int]],
+    outputs: List[RequestFuncOutput],
+    dur_s: float,
+    tokenizer: PreTrainedTokenizerBase,
+) -> BenchmarkMetrics:
+    total_output = 0
+    total_input = 0
+    completed = 0
+    per_token_latencies = []
+    ttfts = []
+    for i in range(len(outputs)):
+        if outputs[i].success:
+            output_len = len(tokenizer.encode(outputs[i].generated_text))
+            total_output += output_len
+            total_input += input_requests[i][1]
+            per_token_latencies.append(outputs[i].latency / output_len)
+            ttfts.append(outputs[i].ttft)
+            completed += 1
+
+    metrics = BenchmarkMetrics(
+        completed=completed,
+        total_input=total_input,
+        total_output=total_output,
+        request_throughput=completed / dur_s,
+        input_throughput=total_input / dur_s,
+        output_throughput=total_output / dur_s,
+        mean_ttft_ms=np.mean(ttfts) * 1000,
+        median_ttft_ms=np.median(ttfts) * 1000,
+        p99_ttft_ms=np.percentile(ttfts, 99) * 1000,
+        mean_tpot_ms=np.mean(per_token_latencies) * 1000,
+        median_tpot_ms=np.median(per_token_latencies) * 1000,
+        p99_tpot_ms=np.percentile(per_token_latencies, 99) * 1000,
+    )
+
+    return metrics
+
+
+async def benchmark(
+    backend: str,
+    api_url: str,
+    model_id: str,
+    tokenizer: PreTrainedTokenizerBase,
+    input_requests: List[Tuple[str, int, int]],
+    best_of: int,
+    use_beam_search: bool,
+    request_rate: float,
+    disable_tqdm: bool,
+):
+    if backend in ASYNC_REQUEST_FUNCS:
+        request_func = ASYNC_REQUEST_FUNCS.get(backend)
+    else:
+        raise ValueError(f"Unknown backend: {backend}")
+
+    pbar = None if disable_tqdm else tqdm(total=len(input_requests))
+
+    print(f"Traffic request rate: {request_rate}")
+
+    benchmark_start_time = time.perf_counter()
+    tasks = []
+    async for request in get_request(input_requests, request_rate):
+        prompt, prompt_len, output_len = request
+        request_func_input = RequestFuncInput(
+            model=model_id,
+            prompt=prompt,
+            api_url=api_url,
+            prompt_len=prompt_len,
+            output_len=output_len,
+            best_of=best_of,
+            use_beam_search=use_beam_search,
+        )
+        tasks.append(
+            asyncio.create_task(
+                request_func(request_func_input=request_func_input,
+                             pbar=pbar)))
+    outputs = await asyncio.gather(*tasks)
+
+    if not disable_tqdm:
+        pbar.close()
+
+    benchmark_duration = time.perf_counter() - benchmark_start_time
+
+    metrics = calculate_metrics(
+        input_requests=input_requests,
+        outputs=outputs,
+        dur_s=benchmark_duration,
+        tokenizer=tokenizer,
+    )
+
+    print(f"Successful requests: {metrics.completed}")
+    print(f"Benchmark duration: {benchmark_duration:2f} s")
+    print(f"Total input tokens: {metrics.total_input}")
+    print(f"Total generated tokens: {metrics.total_output}")
+    print(f"Request throughput: {metrics.request_throughput:.2f} requests/s")
+    print(f"Input token throughput: {metrics.input_throughput:.2f} tokens/s")
+    print(f"Output token throughput: {metrics.output_throughput:.2f} tokens/s")
+    print(f"Mean TTFT: {metrics.mean_ttft_ms:.2f} ms")
+    print(f"Median TTFT: {metrics.median_ttft_ms:.2f} ms")
+    print(f"P99 TTFT: {metrics.p99_ttft_ms:.2f} ms")
+    print(f"Mean TPOT: {metrics.mean_tpot_ms:.2f} ms")
+    print(f"Median TPOT: {metrics.median_tpot_ms:.2f} ms")
+    print(f"P99 TPOT: {metrics.p99_tpot_ms:.2f} ms")
+
+    result = {
+        "duration": benchmark_duration,
+        "completed": metrics.completed,
+        "total_input_tokens": metrics.total_input,
+        "total_output_tokens": metrics.total_output,
+        "request_inthroughput": metrics.request_throughput,
+        "input_throughput": metrics.input_throughput,
+        "output_throughput": metrics.output_throughput,
+        "mean_ttft_ms": metrics.mean_ttft_ms,
+        "median_ttft_ms": metrics.median_ttft_ms,
+        "p99_ttft_ms": metrics.p99_ttft_ms,
+        "mean_tpot_ms": metrics.mean_tpot_ms,
+        "median_tpot_ms": metrics.median_tpot_ms,
+        "p99_tpot_ms": metrics.p99_tpot_ms
+    }
+    return result
+
+
+def main(args: argparse.Namespace):
+    print(args)
+    random.seed(args.seed)
+    np.random.seed(args.seed)
+
+    backend = args.backend
+    model_id = args.model
+    tokenizer_id = args.tokenizer if args.tokenizer is not None else args.model
+
+    if args.base_url is not None:
+        api_url = f"{args.base_url}{args.endpoint}"
+    else:
+        api_url = f"http://{args.host}:{args.port}{args.endpoint}"
+
+    tokenizer = get_tokenizer(tokenizer_id,
+                              trust_remote_code=args.trust_remote_code)
+    input_requests = sample_requests(args.dataset, args.num_prompts, tokenizer)
+
+    benchmark_result = asyncio.run(
+        benchmark(
+            backend=backend,
+            api_url=api_url,
+            model_id=model_id,
+            tokenizer=tokenizer,
+            input_requests=input_requests,
+            best_of=args.best_of,
+            use_beam_search=args.use_beam_search,
+            request_rate=args.request_rate,
+            disable_tqdm=args.disable_tqdm,
+        ))
+
+    # Save config and results to json
+    if args.save_result:
+        result_json = {}
+
+        # Setup
+        current_dt = datetime.now().strftime("%Y%m%d-%H%M%S")
+        result_json["date"] = current_dt
+        result_json["backend"] = backend
+        result_json["version"] = args.version
+        result_json["model_id"] = model_id
+        result_json["tokenizer_id"] = tokenizer_id
+        result_json["best_of"] = args.best_of
+        result_json["use_beam_search"] = args.use_beam_search
+        result_json["num_prompts"] = args.num_prompts
+
+        # Traffic
+        result_json["request_rate"] = (
+            args.request_rate if args.request_rate < float("inf") else "inf")
+
+        # Merge with benchmark result
+        result_json = {**result_json, **benchmark_result}
+
+        # Save to file
+        base_model_id = model_id.split("/")[-1]
+        file_name = f"{backend}-{args.request_rate}qps-{base_model_id}-{current_dt}.json"
+        with open(file_name, "w") as outfile:
+            json.dump(result_json, outfile)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(
+        description="Benchmark the online serving throughput.")
+    parser.add_argument(
+        "--backend",
+        type=str,
+        default="vllm",
+        choices=list(ASYNC_REQUEST_FUNCS.keys()),
+    )
+    parser.add_argument(
+        "--version",
+        type=str,
+        default="N/A",
+        help="Version of the serving backend/engine.",
+    )
+    parser.add_argument(
+        "--base-url",
+        type=str,
+        default=None,
+        help="Server or API base url if not using http host and port.",
+    )
+    parser.add_argument("--host", type=str, default="localhost")
+    parser.add_argument("--port", type=int, default=8000)
+    parser.add_argument(
+        "--endpoint",
+        type=str,
+        default="/generate",
+        help="API endpoint.",
+    )
+    parser.add_argument("--dataset",
+                        type=str,
+                        required=True,
+                        help="Path to the dataset.")
+    parser.add_argument(
+        "--model",
+        type=str,
+        required=True,
+        help="Name of the model.",
+    )
+    parser.add_argument(
+        "--tokenizer",
+        type=str,
+        help=
+        "Name or path of the tokenizer, if not using the default model tokenizer.",
+    )
+    parser.add_argument(
+        "--best-of",
+        type=int,
+        default=1,
+        help="Generates `best_of` sequences per prompt and "
+        "returns the best one.",
+    )
+    parser.add_argument("--use-beam-search", action="store_true")
+    parser.add_argument(
+        "--num-prompts",
+        type=int,
+        default=1000,
+        help="Number of prompts to process.",
+    )
+    parser.add_argument(
+        "--request-rate",
+        type=float,
+        default=float("inf"),
+        help="Number of requests per second. If this is inf, "
+        "then all the requests are sent at time 0. "
+        "Otherwise, we use Poisson process to synthesize "
+        "the request arrival times.",
+    )
+    parser.add_argument("--seed", type=int, default=0)
+    parser.add_argument(
+        "--trust-remote-code",
+        action="store_true",
+        help="Trust remote code from huggingface",
+    )
+    parser.add_argument(
+        "--disable-tqdm",
+        action="store_true",
+        help="Specify to disbale tqdm progress bar.",
+    )
+    parser.add_argument(
+        "--save-result",
+        action="store_true",
+        help="Specify to save benchmark results to a json file",
+    )
+
+    args = parser.parse_args()
+    main(args)

benchmarks/benchmark_throughput.py

@@ -0,0 +1,336 @@
+"""Benchmark offline inference throughput."""
+import argparse
+import json
+import random
+import time
+from typing import List, Optional, Tuple
+
+import torch
+from transformers import (AutoModelForCausalLM, AutoTokenizer,
+                          PreTrainedTokenizerBase)
+from tqdm import tqdm
+
+
+def sample_requests(
+    dataset_path: str,
+    num_requests: int,
+    tokenizer: PreTrainedTokenizerBase,
+    fixed_output_len: Optional[int],
+) -> List[Tuple[str, int, int]]:
+    if fixed_output_len is not None and fixed_output_len < 4:
+        raise ValueError("output_len too small")
+
+    # Load the dataset.
+    with open(dataset_path) as f:
+        dataset = json.load(f)
+    # Filter out the conversations with less than 2 turns.
+    dataset = [data for data in dataset if len(data["conversations"]) >= 2]
+    # Only keep the first two turns of each conversation.
+    dataset = [(data["conversations"][0]["value"],
+                data["conversations"][1]["value"]) for data in dataset]
+
+    # Tokenize the prompts and completions.
+    prompts = [prompt for prompt, _ in dataset]
+    prompt_token_ids = tokenizer(prompts).input_ids
+    completions = [completion for _, completion in dataset]
+    completion_token_ids = tokenizer(completions).input_ids
+    tokenized_dataset = []
+    for i in range(len(dataset)):
+        output_len = len(completion_token_ids[i])
+        if fixed_output_len is not None:
+            output_len = fixed_output_len
+        tokenized_dataset.append((prompts[i], prompt_token_ids[i], output_len))
+
+    # Filter out too long sequences.
+    filtered_dataset: List[Tuple[str, int, int]] = []
+    for prompt, prompt_token_ids, output_len in tokenized_dataset:
+        prompt_len = len(prompt_token_ids)
+        if prompt_len < 4 or output_len < 4:
+            # Prune too short sequences.
+            continue
+        if prompt_len > 1024 or prompt_len + output_len > 2048:
+            # Prune too long sequences.
+            continue
+        filtered_dataset.append((prompt, prompt_len, output_len))
+
+    # Sample the requests.
+    sampled_requests = random.sample(filtered_dataset, num_requests)
+    return sampled_requests
+
+
+def run_vllm(
+    requests: List[Tuple[str, int, int]],
+    model: str,
+    tokenizer: str,
+    quantization: Optional[str],
+    tensor_parallel_size: int,
+    seed: int,
+    n: int,
+    use_beam_search: bool,
+    trust_remote_code: bool,
+    dtype: str,
+    max_model_len: Optional[int],
+    enforce_eager: bool,
+    kv_cache_dtype: str,
+    device: str,
+) -> float:
+    from vllm import LLM, SamplingParams
+    llm = LLM(
+        model=model,
+        tokenizer=tokenizer,
+        quantization=quantization,
+        tensor_parallel_size=tensor_parallel_size,
+        seed=seed,
+        trust_remote_code=trust_remote_code,
+        dtype=dtype,
+        max_model_len=max_model_len,
+        enforce_eager=enforce_eager,
+        kv_cache_dtype=kv_cache_dtype,
+        device=device,
+    )
+
+    # Add the requests to the engine.
+    for prompt, _, output_len in requests:
+        sampling_params = SamplingParams(
+            n=n,
+            temperature=0.0 if use_beam_search else 1.0,
+            top_p=1.0,
+            use_beam_search=use_beam_search,
+            ignore_eos=True,
+            max_tokens=output_len,
+        )
+        # FIXME(woosuk): Do not use internal method.
+        llm._add_request(
+            prompt=prompt,
+            prompt_token_ids=None,
+            sampling_params=sampling_params,
+        )
+
+    start = time.perf_counter()
+    # FIXME(woosuk): Do not use internal method.
+    llm._run_engine(use_tqdm=True)
+    end = time.perf_counter()
+    return end - start
+
+
+def run_hf(
+    requests: List[Tuple[str, int, int]],
+    model: str,
+    tokenizer: PreTrainedTokenizerBase,
+    n: int,
+    use_beam_search: bool,
+    max_batch_size: int,
+    trust_remote_code: bool,
+) -> float:
+    assert not use_beam_search
+    llm = AutoModelForCausalLM.from_pretrained(
+        model, torch_dtype=torch.float16, trust_remote_code=trust_remote_code)
+    if llm.config.model_type == "llama":
+        # To enable padding in the HF backend.
+        tokenizer.pad_token = tokenizer.eos_token
+    llm = llm.cuda()
+
+    pbar = tqdm(total=len(requests))
+    start = time.perf_counter()
+    batch: List[str] = []
+    max_prompt_len = 0
+    max_output_len = 0
+    for i in range(len(requests)):
+        prompt, prompt_len, output_len = requests[i]
+        # Add the prompt to the batch.
+        batch.append(prompt)
+        max_prompt_len = max(max_prompt_len, prompt_len)
+        max_output_len = max(max_output_len, output_len)
+        if len(batch) < max_batch_size and i != len(requests) - 1:
+            # Check if we can add more requests to the batch.
+            _, next_prompt_len, next_output_len = requests[i + 1]
+            if (max(max_prompt_len, next_prompt_len) +
+                    max(max_output_len, next_output_len)) <= 2048:
+                # We can add more requests to the batch.
+                continue
+
+        # Generate the sequences.
+        input_ids = tokenizer(batch, return_tensors="pt",
+                              padding=True).input_ids
+        llm_outputs = llm.generate(
+            input_ids=input_ids.cuda(),
+            do_sample=not use_beam_search,
+            num_return_sequences=n,
+            temperature=1.0,
+            top_p=1.0,
+            use_cache=True,
+            max_new_tokens=max_output_len,
+        )
+        # Include the decoding time.
+        tokenizer.batch_decode(llm_outputs, skip_special_tokens=True)
+        pbar.update(len(batch))
+
+        # Clear the batch.
+        batch = []
+        max_prompt_len = 0
+        max_output_len = 0
+    end = time.perf_counter()
+    return end - start
+
+
+def run_mii(
+    requests: List[Tuple[str, int, int]],
+    model: str,
+    tensor_parallel_size: int,
+    output_len: int,
+) -> float:
+    from mii import pipeline
+    llm = pipeline(model, tensor_parallel=tensor_parallel_size)
+    prompts = [prompt for prompt, _, _ in requests]
+
+    start = time.perf_counter()
+    llm(prompts, max_new_tokens=output_len)
+    end = time.perf_counter()
+    return end - start
+
+
+def main(args: argparse.Namespace):
+    print(args)
+    random.seed(args.seed)
+
+    # Sample the requests.
+    tokenizer = AutoTokenizer.from_pretrained(
+        args.tokenizer, trust_remote_code=args.trust_remote_code)
+    if args.dataset is None:
+        # Synthesize a prompt with the given input length.
+        prompt = "hi" * (args.input_len - 1)
+        requests = [(prompt, args.input_len, args.output_len)
+                    for _ in range(args.num_prompts)]
+    else:
+        requests = sample_requests(args.dataset, args.num_prompts, tokenizer,
+                                   args.output_len)
+
+    if args.backend == "vllm":
+        elapsed_time = run_vllm(requests, args.model, args.tokenizer,
+                                args.quantization, args.tensor_parallel_size,
+                                args.seed, args.n, args.use_beam_search,
+                                args.trust_remote_code, args.dtype,
+                                args.max_model_len, args.enforce_eager,
+                                args.kv_cache_dtype, args.device)
+    elif args.backend == "hf":
+        assert args.tensor_parallel_size == 1
+        elapsed_time = run_hf(requests, args.model, tokenizer, args.n,
+                              args.use_beam_search, args.hf_max_batch_size,
+                              args.trust_remote_code)
+    elif args.backend == "mii":
+        elapsed_time = run_mii(requests, args.model, args.tensor_parallel_size,
+                               args.output_len)
+    else:
+        raise ValueError(f"Unknown backend: {args.backend}")
+    total_num_tokens = sum(prompt_len + output_len
+                           for _, prompt_len, output_len in requests)
+    print(f"Throughput: {len(requests) / elapsed_time:.2f} requests/s, "
+          f"{total_num_tokens / elapsed_time:.2f} tokens/s")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Benchmark the throughput.")
+    parser.add_argument("--backend",
+                        type=str,
+                        choices=["vllm", "hf", "mii"],
+                        default="vllm")
+    parser.add_argument("--dataset",
+                        type=str,
+                        default=None,
+                        help="Path to the dataset.")
+    parser.add_argument("--input-len",
+                        type=int,
+                        default=None,
+                        help="Input prompt length for each request")
+    parser.add_argument("--output-len",
+                        type=int,
+                        default=None,
+                        help="Output length for each request. Overrides the "
+                        "output length from the dataset.")
+    parser.add_argument("--model", type=str, default="facebook/opt-125m")
+    parser.add_argument("--tokenizer", type=str, default=None)
+    parser.add_argument('--quantization',
+                        '-q',
+                        choices=['awq', 'gptq', 'squeezellm', None],
+                        default=None)
+    parser.add_argument("--tensor-parallel-size", "-tp", type=int, default=1)
+    parser.add_argument("--n",
+                        type=int,
+                        default=1,
+                        help="Number of generated sequences per prompt.")
+    parser.add_argument("--use-beam-search", action="store_true")
+    parser.add_argument("--num-prompts",
+                        type=int,
+                        default=1000,
+                        help="Number of prompts to process.")
+    parser.add_argument("--seed", type=int, default=0)
+    parser.add_argument("--hf-max-batch-size",
+                        type=int,
+                        default=None,
+                        help="Maximum batch size for HF backend.")
+    parser.add_argument('--trust-remote-code',
+                        action='store_true',
+                        help='trust remote code from huggingface')
+    parser.add_argument(
+        '--max-model-len',
+        type=int,
+        default=None,
+        help='Maximum length of a sequence (including prompt and output). '
+        'If None, will be derived from the model.')
+    parser.add_argument(
+        '--dtype',
+        type=str,
+        default='auto',
+        choices=['auto', 'half', 'float16', 'bfloat16', 'float', 'float32'],
+        help='data type for model weights and activations. '
+        'The "auto" option will use FP16 precision '
+        'for FP32 and FP16 models, and BF16 precision '
+        'for BF16 models.')
+    parser.add_argument("--enforce-eager",
+                        action="store_true",
+                        help="enforce eager execution")
+    parser.add_argument(
+        "--kv-cache-dtype",
+        type=str,
+        choices=["auto", "fp8_e5m2"],
+        default="auto",
+        help=
+        'Data type for kv cache storage. If "auto", will use model data type.')
+    parser.add_argument(
+        "--device",
+        type=str,
+        default="cuda",
+        choices=["cuda"],
+        help='device type for vLLM execution, supporting CUDA only currently.')
+    args = parser.parse_args()
+    if args.tokenizer is None:
+        args.tokenizer = args.model
+    if args.dataset is None:
+        assert args.input_len is not None
+        assert args.output_len is not None
+    else:
+        assert args.input_len is None
+
+    if args.backend == "vllm":
+        if args.hf_max_batch_size is not None:
+            raise ValueError("HF max batch size is only for HF backend.")
+    elif args.backend == "hf":
+        if args.hf_max_batch_size is None:
+            raise ValueError("HF max batch size is required for HF backend.")
+        if args.quantization is not None:
+            raise ValueError("Quantization is only for vLLM backend.")
+    elif args.backend == "mii":
+        if args.dtype != "auto":
+            raise ValueError("dtype must be auto for MII backend.")
+        if args.n != 1:
+            raise ValueError("n must be 1 for MII backend.")
+        if args.use_beam_search:
+            raise ValueError("Beam search is not supported for MII backend.")
+        if args.quantization is not None:
+            raise ValueError("Quantization is only for vLLM backend.")
+        if args.hf_max_batch_size is not None:
+            raise ValueError("HF max batch size is only for HF backend.")
+        if args.tokenizer != args.model:
+            raise ValueError("Tokenizer must be the same as the model for MII "
+                             "backend.")
+    main(args)

benchmarks/kernels/benchmark_paged_attention.py

@@ -0,0 +1,205 @@
+from typing import Optional
+import argparse
+import random
+import time
+
+import torch
+
+from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE, create_kv_caches_with_random
+from vllm._C import ops
+
+NUM_BLOCKS = 1024
+PARTITION_SIZE = 512
+
+
+@torch.inference_mode()
+def main(
+    version: str,
+    num_seqs: int,
+    context_len: int,
+    num_query_heads: int,
+    num_kv_heads: int,
+    head_size: int,
+    use_alibi: bool,
+    block_size: int,
+    dtype: torch.dtype,
+    seed: int,
+    do_profile: bool,
+    device: str = "cuda",
+    kv_cache_dtype: Optional[str] = None,
+) -> None:
+    random.seed(seed)
+    torch.random.manual_seed(seed)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed(seed)
+
+    scale = float(1.0 / (head_size**0.5))
+    query = torch.empty(num_seqs,
+                        num_query_heads,
+                        head_size,
+                        dtype=dtype,
+                        device=device)
+    query.uniform_(-scale, scale)
+
+    assert num_query_heads % num_kv_heads == 0
+    alibi_slopes = None
+    if use_alibi:
+        alibi_slopes = torch.randn(num_query_heads,
+                                   dtype=torch.float,
+                                   device=device)
+
+    context_lens = [context_len for _ in range(num_seqs)]
+    max_context_len = max(context_lens)
+    context_lens = torch.tensor(context_lens, dtype=torch.int, device=device)
+
+    # Create the block tables.
+    max_num_blocks_per_seq = (max_context_len + block_size - 1) // block_size
+    block_tables = []
+    for _ in range(num_seqs):
+        block_table = [
+            random.randint(0, NUM_BLOCKS - 1)
+            for _ in range(max_num_blocks_per_seq)
+        ]
+        block_tables.append(block_table)
+    block_tables = torch.tensor(block_tables, dtype=torch.int, device=device)
+
+    # Create the KV cache.
+    key_caches, value_caches = create_kv_caches_with_random(NUM_BLOCKS,
+                                                            block_size,
+                                                            1,
+                                                            num_kv_heads,
+                                                            head_size,
+                                                            kv_cache_dtype,
+                                                            dtype,
+                                                            device=device)
+    key_cache, value_cache = key_caches[0], value_caches[0]
+
+    # Prepare for the paged attention kernel.
+    output = torch.empty_like(query)
+    if version == "v2":
+        num_partitions = ((max_context_len + PARTITION_SIZE - 1) //
+                          PARTITION_SIZE)
+        tmp_output = torch.empty(
+            size=(num_seqs, num_query_heads, num_partitions, head_size),
+            dtype=output.dtype,
+            device=output.device,
+        )
+        exp_sums = torch.empty(
+            size=(num_seqs, num_query_heads, num_partitions),
+            dtype=torch.float32,
+            device=output.device,
+        )
+        max_logits = torch.empty_like(exp_sums)
+
+    def run_cuda_benchmark(num_iters: int, profile: bool = False) -> float:
+        torch.cuda.synchronize()
+        if profile:
+            torch.cuda.cudart().cudaProfilerStart()
+        start_time = time.perf_counter()
+
+        for _ in range(num_iters):
+            if version == "v1":
+                ops.paged_attention_v1(
+                    output,
+                    query,
+                    key_cache,
+                    value_cache,
+                    num_kv_heads,
+                    scale,
+                    block_tables,
+                    context_lens,
+                    block_size,
+                    max_context_len,
+                    alibi_slopes,
+                    kv_cache_dtype,
+                )
+            elif version == "v2":
+                ops.paged_attention_v2(
+                    output,
+                    exp_sums,
+                    max_logits,
+                    tmp_output,
+                    query,
+                    key_cache,
+                    value_cache,
+                    num_kv_heads,
+                    scale,
+                    block_tables,
+                    context_lens,
+                    block_size,
+                    max_context_len,
+                    alibi_slopes,
+                    kv_cache_dtype,
+                )
+            else:
+                raise ValueError(f"Invalid version: {version}")
+        torch.cuda.synchronize()
+
+        end_time = time.perf_counter()
+        if profile:
+            torch.cuda.cudart().cudaProfilerStart()
+        return (end_time - start_time) / num_iters
+
+    # Warmup.
+    print("Warming up...")
+    run_benchmark = run_cuda_benchmark
+    run_benchmark(num_iters=3, profile=False)
+
+    # Benchmark.
+    if do_profile:
+        latency = run_benchmark(num_iters=1, profile=True)
+    else:
+        latency = run_benchmark(num_iters=100, profile=False)
+    print(f"Kernel running time: {latency * 1000000:.3f} us")
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(
+        description="Benchmark the paged attention kernel.")
+    parser.add_argument("--version",
+                        type=str,
+                        choices=["v1", "v2"],
+                        default="v2")
+    parser.add_argument("--batch-size", type=int, default=8)
+    parser.add_argument("--context-len", type=int, default=4096)
+    parser.add_argument("--num-query-heads", type=int, default=64)
+    parser.add_argument("--num-kv-heads", type=int, default=8)
+    parser.add_argument("--head-size",
+                        type=int,
+                        choices=[64, 80, 96, 112, 128, 256],
+                        default=128)
+    parser.add_argument("--block-size", type=int, choices=[16, 32], default=16)
+    parser.add_argument("--use-alibi", action="store_true")
+    parser.add_argument("--dtype",
+                        type=str,
+                        choices=["half", "bfloat16", "float"],
+                        default="half")
+    parser.add_argument("--seed", type=int, default=0)
+    parser.add_argument("--profile", action="store_true")
+    parser.add_argument(
+        "--kv-cache-dtype",
+        type=str,
+        choices=["auto", "fp8_e5m2"],
+        default="auto",
+        help=
+        'Data type for kv cache storage. If "auto", will use model data type.')
+    parser.add_argument("--device", type=str, choices=["cuda"], default="cuda")
+    args = parser.parse_args()
+    print(args)
+
+    if args.num_query_heads % args.num_kv_heads != 0:
+        raise ValueError("num_query_heads must be divisible by num_kv_heads")
+    main(
+        version=args.version,
+        num_seqs=args.batch_size,
+        context_len=args.context_len,
+        num_query_heads=args.num_query_heads,
+        num_kv_heads=args.num_kv_heads,
+        head_size=args.head_size,
+        block_size=args.block_size,
+        use_alibi=args.use_alibi,
+        dtype=STR_DTYPE_TO_TORCH_DTYPE[args.dtype],
+        seed=args.seed,
+        do_profile=args.profile,
+        kv_cache_dtype=args.kv_cache_dtype,
+    )

benchmarks/launch_tgi_server.sh

@@ -0,0 +1,16 @@
+#!/bin/bash
+
+PORT=8000
+MODEL=$1
+TOKENS=$2
+
+docker run --gpus all --shm-size 1g -p $PORT:80 \
+           -v $PWD/data:/data \
+           ghcr.io/huggingface/text-generation-inference:1.4.0 \
+           --model-id $MODEL \
+           --sharded false  \
+           --max-input-length 1024 \
+           --max-total-tokens 2048 \
+           --max-best-of 5 \
+           --max-concurrent-requests 5000 \
+           --max-batch-total-tokens $TOKENS

cacheflow/http_frontend/fastapi_frontend.py

@@ -1,179 +0,0 @@
-import argparse
-import asyncio
-import time
-from typing import List, Dict
-import json
-
-import ray
-from transformers import AutoTokenizer
-from fastapi import FastAPI, Request
-from fastapi.responses import StreamingResponse
-import uvicorn
-
-from cacheflow.sampling_params import SamplingParams
-from cacheflow.sequence import Sequence, SequenceGroup
-from cacheflow.master.server import (Server, add_server_arguments,
-                                     initialize_ray_cluster)
-from cacheflow.worker.controller import DeviceID
-from cacheflow.utils import Counter, get_gpu_memory, get_cpu_memory
-
-TIMEOUT_TO_PREVENT_DEADLOCK = 1 # seconds
-app = FastAPI()
-
-
-class FastAPIFrontend:
-    def __init__(
-        self,
-        model: str,
-        model_path: str,
-        pipeline_parallel_size: int,
-        tensor_parallel_size: int,
-        block_size: int,
-        dtype: str,
-        seed: int,
-        swap_space: int,
-        max_num_batched_tokens: int,
-        num_nodes: int,
-        num_devices_per_node: int,
-        distributed_init_method: str,
-        all_stage_devices: List[List[DeviceID]],
-    ):
-        self.block_size = block_size
-
-        self.tokenizer = AutoTokenizer.from_pretrained(model)
-        self.seq_group_counter = Counter()
-        self.seq_counter = Counter()
-        remote_server_class = ray.remote(num_cpus=0)(Server)
-        self.server = remote_server_class.remote(
-            model=model,
-            model_path=model_path,
-            use_dummy_weights=False,
-            pipeline_parallel_size=pipeline_parallel_size,
-            tensor_parallel_size=tensor_parallel_size,
-            block_size=block_size,
-            dtype=dtype,
-            seed=seed,
-            swap_space=swap_space,
-            max_num_batched_tokens=max_num_batched_tokens,
-            num_nodes=num_nodes,
-            num_devices_per_node=num_devices_per_node,
-            distributed_init_method=distributed_init_method,
-            all_stage_devices=all_stage_devices,
-            gpu_memory=get_gpu_memory(),
-            cpu_memory=get_cpu_memory(),
-        )
-
-        self.running_seq_groups: Dict[int, SequenceGroup] = {}
-        self.sequence_group_events: Dict[int, asyncio.Event] = {}
-        self.is_server_running = False
-
-    async def server_step(self):
-        self.is_server_running = True
-        updated_seq_groups = await self.server.step.remote()
-        self.is_server_running = False
-        # Notify the waiting coroutines that there new outputs ready.
-        for seq_group in updated_seq_groups:
-            group_id = seq_group.group_id
-            self.running_seq_groups[group_id] = seq_group
-            self.sequence_group_events[group_id].set()
-
-    async def generate(self, request_dict: Dict):
-        # Preprocess the request.
-        prompt = request_dict["prompt"]
-        sampling_params = SamplingParams.from_dict(request_dict)
-        sampling_params.stop_token_ids.add(self.tokenizer.eos_token_id)
-        token_ids = self.tokenizer.encode(prompt)
-        seqs: List[Sequence] = []
-        for _ in range(sampling_params.n):
-            seq_id = next(self.seq_counter)
-            seq = Sequence(seq_id, token_ids, block_size=self.block_size)
-            seqs.append(seq)
-
-        arrival_time = time.time()
-        group_id = next(self.seq_group_counter)
-        seq_group = SequenceGroup(group_id, seqs, arrival_time)
-        # Create an event to notify us that there is new output from the
-        # cacheflow server.
-        group_event = asyncio.Event()
-        self.running_seq_groups[group_id] = seq_group
-        self.sequence_group_events[group_id] = group_event
-        # Add the request into the cacheflow server's waiting queue.
-        await self.server.add_sequence_groups.remote([(seq_group, sampling_params)])
-        # The cacheflow server does not have a background loop that keeps
-        # processing incoming requests. Therefore, we need to keep kicking
-        # the server to process the requests.
-        while True:
-            # Kick the server if the server is not running.
-            if not self.is_server_running:
-                await self.server_step()
-            # Wait for new output. The group_event will be set in server_step
-            # when there is new output available for the sequence group.
-            # Added a timeout to prevent deadlock.
-            await asyncio.wait_for(group_event.wait(), timeout=TIMEOUT_TO_PREVENT_DEADLOCK)
-            # Reset the event to wait for the next output.
-            group_event.clear()
-            # Decode and return new outputs
-            seq_group = self.running_seq_groups[group_id]
-            all_outputs = []
-            for seq in seq_group.seqs:
-                token_ids = seq.get_token_ids()
-                output = self.tokenizer.decode(token_ids, skip_special_tokens=True)
-                all_outputs.append(output)
-            ret = {
-                "text": all_outputs,
-                "error": 0,
-            }
-            yield (json.dumps(ret) + "\0").encode("utf-8")
-
-            # Once finished, release the resources of the sequence group.
-            if seq_group.is_finished():
-                del self.running_seq_groups[group_id]
-                del self.sequence_group_events[group_id]
-                # Kick the server if the server is not running. This is to
-                # prevent that there are still requests in server's waiting
-                # queue to be executed.
-                if not self.is_server_running:
-                    await self.server_step()
-                break
-
-
-@app.post("/generate")
-async def generate_stream(request: Request):
-    request_dict = await request.json()
-    return StreamingResponse(frontend.generate(request_dict))
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-    parser.add_argument("--host", type=str, default="localhost")
-    parser.add_argument("--port", type=int, default=10002)
-    parser = add_server_arguments(parser)
-    args = parser.parse_args()
-
-    # TODO(zhuohan): Support pipeline parallelism.
-    assert args.pipeline_parallel_size == 1, (
-        'Pipeline parallelism is not supported yet.')
-
-    (num_nodes, num_devices_per_node, distributed_init_method,
-    all_stage_devices) = (
-        initialize_ray_cluster(
-            pipeline_parallel_size=args.pipeline_parallel_size,
-            tensor_parallel_size=args.tensor_parallel_size))
-
-    frontend = FastAPIFrontend(
-        model=args.model,
-        model_path=args.model_path,
-        pipeline_parallel_size=args.pipeline_parallel_size,
-        tensor_parallel_size=args.tensor_parallel_size,
-        block_size=args.block_size,
-        dtype=args.dtype,
-        seed=args.seed,
-        swap_space=args.swap_space,
-        max_num_batched_tokens=args.max_num_batched_tokens,
-        num_nodes=num_nodes,
-        num_devices_per_node=num_devices_per_node,
-        distributed_init_method=distributed_init_method,
-        all_stage_devices=all_stage_devices,
-    )
-
-    uvicorn.run(app, host=args.host, port=args.port, log_level="info")

cacheflow/http_frontend/gradio_webserver.py

@@ -1,43 +0,0 @@
-import argparse
-import json
-import time
-
-import gradio as gr
-import requests
-
-
-def http_bot(prompt):
-    headers = {"User-Agent": "Cacheflow Client"}
-    pload = {
-        "prompt": prompt,
-        "max_num_steps": 128,
-    }
-    response = requests.post(args.model_url, headers=headers, json=pload, stream=True)
-
-    for chunk in response.iter_lines(chunk_size=8192, decode_unicode=False, delimiter=b"\0"):
-        if chunk:
-            data = json.loads(chunk.decode("utf-8"))
-            output = data["text"][0]
-            yield output
-
-
-def build_demo():
-    with gr.Blocks() as demo:
-        gr.Markdown(
-            "# Cacheflow demo\n"
-        )
-        inputbox = gr.Textbox(label="Input", placeholder="Enter text and press ENTER")# .style(container=False)
-        outputbox = gr.Textbox(label="Output", placeholder="Generated result from the model")
-        inputbox.submit(http_bot, [inputbox], [outputbox])
-    return demo
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-    parser.add_argument("--host", type=str, default="localhost")
-    parser.add_argument("--port", type=int, default=10003)
-    parser.add_argument("--model-url", type=str, default="http://localhost:10002/generate")
-    args = parser.parse_args()
-
-    demo = build_demo()
-    demo.queue(concurrency_count=100).launch(server_name=args.host, server_port=args.port)

cacheflow/http_frontend/test_cli_client.py

@@ -1,23 +0,0 @@
-import requests
-import json
-
-def http_request():
-    prompt = "Ion Stoica is a"
-
-    headers = {"User-Agent": "Test Client"}
-    pload = {
-        "prompt": prompt,
-        "n": 4,
-        "use_beam_search": True,
-        "temperature": 0.0,
-    }
-    response = requests.post("http://localhost:10002/generate", headers=headers, json=pload, stream=True)
-
-    for chunk in response.iter_lines(chunk_size=8192, decode_unicode=False, delimiter=b"\0"):
-        if chunk:
-            data = json.loads(chunk.decode("utf-8"))
-            output = data["text"]
-            yield output
-
-for h in http_request():
-    print(h, flush=True)

cacheflow/master/scheduler.py

@@ -1,529 +0,0 @@
-import enum
-import os
-import pickle
-import time
-from typing import Any, Dict, List, Optional, Tuple
-
-from cacheflow.master.block_manager import BlockSpaceManager
-from cacheflow.master.policy import PolicyFactory
-from cacheflow.sampling_params import SamplingParams
-from cacheflow.sequence import Sequence
-from cacheflow.sequence import SequenceGroup
-from cacheflow.sequence import SequenceGroupInputs
-from cacheflow.sequence import SequenceOutputs
-from cacheflow.sequence import SequenceStatus
-
-
-class PreemptionMode(enum.Enum):
-    """Preemption modes.
-
-    1. Swapping: Swap out the blocks of the preempted sequences to CPU memory
-    and swap them back in when the sequences are resumed.
-    2. Recomputation: Discard the blocks of the preempted sequences and
-    recompute them when the sequences are resumed, treating the sequences as
-    new prompts.
-    """
-    SWAP = enum.auto()
-    RECOMPUTE = enum.auto()
-
-
-class Scheduler:
-
-    def __init__(
-        self,
-        controllers: List,
-        block_size: int,
-        num_gpu_blocks: int,
-        num_cpu_blocks: int,
-        max_num_batched_tokens: int,
-        max_num_sequences: int,
-        collect_stats: bool,
-        do_memory_analysis: bool = False,
-    ) -> None:
-        self.controllers = controllers
-        self.block_size = block_size
-        self.num_gpu_blocks = num_gpu_blocks
-        self.num_cpu_blocks = num_cpu_blocks
-        self.max_num_batched_tokens = max_num_batched_tokens
-        self.max_num_sequences = max_num_sequences
-        self.collect_stats = collect_stats
-        self.do_memory_analysis = do_memory_analysis
-
-        # Instantiate the scheduling policy.
-        self.policy = PolicyFactory.get_policy(policy_name='fcfs')
-        # Create the block space manager.
-        self.block_manager = BlockSpaceManager(
-            block_size=block_size,
-            num_gpu_blocks=num_gpu_blocks,
-            num_cpu_blocks=num_cpu_blocks,
-        )
-
-        # Sequence groups in the WAITING state.
-        self.waiting: List[SequenceGroup] = []
-        # Sequence groups in the RUNNING state.
-        self.running: List[SequenceGroup] = []
-        # Mapping: group_id -> num_steps.
-        self.num_steps: Dict[int, int] = {}
-        # Mapping: group_id -> sampling params.
-        self.sampling_params: Dict[int, SamplingParams] = {}
-        # Sequence groups in the SWAPPED state.
-        self.swapped: List[SequenceGroup] = []
-
-        # Performance-related statistics.
-        self.stats = Stats(num_gpu_blocks, num_cpu_blocks)
-
-    def add_sequence_groups(
-        self,
-        seq_groups: List[Tuple[SequenceGroup, SamplingParams]],
-    ) -> None:
-        # Add sequence groups to the waiting queue.
-        for seq_group, sampling_params in seq_groups:
-            self.waiting.append(seq_group)
-            self.sampling_params[seq_group.group_id] = sampling_params
-
-    def _schedule(
-        self,
-    ) -> Tuple[Dict[int, int], Dict[int, int], Dict[int, List[int]], List[int]]:
-        # Blocks that need to be swaped or copied before model execution.
-        blocks_to_swap_in: Dict[int, int] = {}
-        blocks_to_swap_out: Dict[int, int] = {}
-        blocks_to_copy: Dict[int, List[int]] = {}
-
-        # Fix the current time.
-        now = time.time()
-
-        # NOTE(woosuk): We prioritize the sequence groups in the RUNNING state
-        # in order to minimize the preemption overheads.
-        # Preemption happens only when there is no available slot to keep all
-        # the sequence groups in the RUNNING state.
-        # In this case, the policy is responsible for deciding which sequence
-        # groups to preempt.
-        self.running = self.policy.sort_by_priority(now, self.running)
-
-        # Reserve new token slots for the running sequence groups.
-        running: List[SequenceGroup] = []
-        preempted: List[SequenceGroup] = []
-        while self.running:
-            seq_group = self.running.pop(0)
-            while not self.block_manager.can_append(seq_group):
-                if self.running:
-                    # Preempt the lowest-priority sequence groups.
-                    victim_seq_group = self.running.pop(-1)
-                    self._preempt(victim_seq_group, blocks_to_swap_out)
-                    preempted.append(victim_seq_group)
-                else:
-                    # No other sequence groups can be preempted.
-                    # Preempt the current sequence group.
-                    self._preempt(seq_group, blocks_to_swap_out)
-                    preempted.append(seq_group)
-                    break
-            else:
-                # Append new slots to the sequence group.
-                self._append(seq_group, blocks_to_copy)
-                running.append(seq_group)
-        self.running = running
-
-        # Swap in the sequence groups in the SWAPPED state if possible.
-        self.swapped = self.policy.sort_by_priority(now, self.swapped)
-        # FCFS
-        while self.swapped and not blocks_to_swap_out:
-            seq_group = self.swapped[0]
-            # If the sequence group has been preempted in this step, stop.
-            if seq_group in preempted:
-                break
-            # If the sequence group cannot be swapped in, stop.
-            if not self.block_manager.can_swap_in(seq_group):
-                break
-
-            # The total number of sequences in the RUNNING state should not
-            # exceed the maximum number of sequences.
-            num_seqs = seq_group.num_seqs(status=SequenceStatus.SWAPPED)
-            if len(self.running) + num_seqs > self.max_num_sequences:
-                break
-
-            seq_group = self.swapped.pop(0)
-            self._swap_in(seq_group, blocks_to_swap_in)
-            self._append(seq_group, blocks_to_copy)
-            self.running.append(seq_group)
-
-        num_batched_tokens = sum(
-            seq_group.num_seqs(status=SequenceStatus.RUNNING)
-            for seq_group in self.running
-        )
-
-        # Join waiting sequences if possible.
-        prompt_group_ids: List[int] = []
-        # NOTE(woosuk): The sequence groups in the SWAPPED state are strictly
-        # prioritized over the sequence groups in the WAITING state.
-        # This is because we want to bound the amount of CPU memory taken by
-        # the swapped sequence groups.
-        if not self.swapped:
-            self.waiting = self.policy.sort_by_priority(now, self.waiting)
-            while self.waiting:
-                seq_group = self.waiting[0]
-                # If the sequence group has been preempted in this step, stop.
-                if seq_group in preempted:
-                    break
-                # If the sequence group cannot be allocated, stop.
-                if not self.block_manager.can_allocate(seq_group):
-                    break
-
-                # If the number of batched tokens exceeds the limit, stop.
-                num_prompt_tokens = seq_group.seqs[0].get_len()
-                if (num_batched_tokens + num_prompt_tokens
-                    > self.max_num_batched_tokens):
-                    break
-
-                # The total number of sequences in the RUNNING state should not
-                # exceed the maximum number of sequences.
-                num_seqs = seq_group.num_seqs(status=SequenceStatus.WAITING)
-                if len(self.running) + num_seqs > self.max_num_sequences:
-                    break
-
-                seq_group = self.waiting.pop(0)
-                self._allocate(seq_group)
-                self.running.append(seq_group)
-                num_batched_tokens += num_prompt_tokens
-                prompt_group_ids.append(seq_group.group_id)
-
-        if self.collect_stats:
-            if self.running or blocks_to_swap_in or blocks_to_swap_out:
-                self.stats.timestamps.append(now - self.stats.start_time)
-                self.stats.input_lens.append(num_batched_tokens)
-                self.stats.swap_out_lens.append(len(blocks_to_swap_out) * self.block_size)
-                self.stats.swap_in_lens.append(len(blocks_to_swap_in) * self.block_size)
-                self.stats.num_preemption.append(len(preempted))
-                self.stats.num_swapped.append(len(self.swapped))
-                self.stats.num_running.append(len(self.running))
-                self.stats.num_waiting.append(len(self.waiting))
-
-                num_free_gpu_blocks = self.block_manager.get_num_free_gpu_blocks()
-                num_used_gpu_blocks = self.num_gpu_blocks - num_free_gpu_blocks
-                self.stats.gpu_cache_usage.append(num_used_gpu_blocks / self.num_gpu_blocks)
-                num_free_cpu_blocks = self.block_manager.get_num_free_cpu_blocks()
-                num_used_cpu_blocks = self.num_cpu_blocks - num_free_cpu_blocks
-                self.stats.cpu_cache_usage.append(num_used_cpu_blocks / self.num_cpu_blocks)
-
-                if self.do_memory_analysis:
-                    block_tables = self.block_manager.block_tables
-                    num_logical_blocks = 0
-                    num_logical_tokens = 0
-                    num_physical_blocks = 0
-                    num_physical_tokens = 0
-                    physical_block_numbers = set()
-                    num_reserved_tokens = 0
-                    for seq_group in self.running:
-                        group_id = seq_group.group_id
-                        sampling_params = self.sampling_params[group_id]
-                        max_num_steps = sampling_params.max_num_steps
-                        for seq in seq_group.get_seqs(status=SequenceStatus.RUNNING):
-                            num_logical_blocks += len(seq.logical_token_blocks)
-                            num_logical_tokens += seq.get_len()
-
-                            seq_id = seq.seq_id
-                            block_table = block_tables[seq_id]
-                            for i, block in enumerate(block_table):
-                                if block.block_number in physical_block_numbers:
-                                    continue
-                                physical_block_numbers.add(block.block_number)
-                                num_physical_blocks += 1
-                                num_physical_tokens += seq.logical_token_blocks[i].num_tokens
-                    
-                    assert num_physical_blocks == num_used_gpu_blocks
-                    self.stats.num_logical_blocks.append(num_logical_blocks)
-                    self.stats.num_logical_tokens.append(num_logical_tokens)
-                    self.stats.num_physical_blocks.append(num_physical_blocks)
-                    self.stats.num_physical_tokens.append(num_physical_tokens)
-                    self.stats.num_reserved_tokens.append(num_reserved_tokens)
-
-        return (blocks_to_swap_in,
-                blocks_to_swap_out,
-                blocks_to_copy,
-                prompt_group_ids)
-
-    def step(self) -> List[SequenceGroup]:
-        # Schedule sequence groups.
-        # This function call changes the internal states of the scheduler
-        # such as self.running, self.swapped, and self.waiting.
-        scheduler_output = self._schedule()
-        blocks_to_swap_in = scheduler_output[0]
-        blocks_to_swap_out = scheduler_output[1]
-        blocks_to_copy = scheduler_output[2]
-        prompt_group_ids = scheduler_output[3]
-
-        # Create input data structures.
-        input_seq_groups: List[SequenceGroupInputs] = []
-        updated_seq_groups: List[SequenceGroup] = self.running.copy()
-
-        for seq_group in self.running:
-            group_id = seq_group.group_id
-            is_prompt = group_id in prompt_group_ids
-
-            input_tokens: Dict[int, List[int]] = {}
-            seq_logprobs: Dict[int, float] = {}
-            block_tables: Dict[int, List[int]] = {}
-            for seq in seq_group.get_seqs(status=SequenceStatus.RUNNING):
-                seq_id = seq.seq_id
-                block_tables[seq_id] = self.block_manager.get_block_table(seq)
-                if is_prompt:
-                    input_tokens[seq_id] = seq.get_token_ids()
-                else:
-                    input_tokens[seq_id] = [seq.get_last_token_id()]
-                seq_logprobs[seq_id] = seq.cumulative_logprobs
-                # NOTE(woosuk): Sequences in the same group have the same
-                # sequence length
-                seq_len = seq.get_len()
-
-            input_seq_group = SequenceGroupInputs(
-                group_id=group_id,
-                is_prompt=is_prompt,
-                input_tokens=input_tokens,
-                context_len=seq_len,
-                seq_logprobs=seq_logprobs,
-                sampling_params=self.sampling_params[group_id],
-                block_tables=block_tables,
-            )
-            input_seq_groups.append(input_seq_group)
-
-        # Execute the first stage of the pipeline.
-        if input_seq_groups or blocks_to_swap_in or blocks_to_swap_out:
-            # Swap in and swap out should never happen at the same time.
-            assert not (blocks_to_swap_in and blocks_to_swap_out)
-            self.controllers[0].execute_stage(
-                input_seq_groups,
-                blocks_to_swap_in=blocks_to_swap_in,
-                blocks_to_swap_out=blocks_to_swap_out,
-                blocks_to_copy=blocks_to_copy,
-            )
-
-        return updated_seq_groups
-
-    def post_step(
-        self,
-        seq_outputs: Dict[int, SequenceOutputs],
-    ) -> None:
-        # Update the running sequences and free blocks.
-        for seq_group in self.running:
-            group_id = seq_group.group_id
-            self.num_steps[group_id] += 1
-            stop_token_ids = self.sampling_params[group_id].stop_token_ids
-
-            # Process beam search results before processing the next tokens.
-            for seq in seq_group.seqs:
-                if seq.status == SequenceStatus.FINISHED:
-                    continue
-
-                output = seq_outputs[seq.seq_id]
-                if seq.seq_id != output.parent_seq_id:
-                    # The sequence is a fork of the parent sequence (beam search).
-                    # Free the current sequence.
-                    self.block_manager.free(seq)
-                    # Fork the parent sequence.
-                    parent_seq = seq_group.find(output.parent_seq_id)
-                    parent_seq.fork(seq)
-                    self.block_manager.fork(parent_seq, seq)
-
-            # Process the next tokens.
-            for seq in seq_group.seqs:
-                if seq.status == SequenceStatus.FINISHED:
-                    continue
-
-                # Append a new token to the sequence.
-                output = seq_outputs[seq.seq_id]
-                seq.append(output.output_token, output.logprobs)
-
-                # Check if the sequence has generated a stop token.
-                if output.output_token in stop_token_ids:
-                    self._free_seq(seq)
-                    continue
-
-                # Check if the sequence has reached the maximum number of steps.
-                max_num_steps = self.sampling_params[group_id].max_num_steps
-                if self.num_steps[group_id] == max_num_steps:
-                    self._free_seq(seq)
-                    continue
-
-        # Update the running sequences.
-        running: List[SequenceGroup] = []
-        for seq_group in self.running:
-            if seq_group.is_finished():
-                self._free_seq_group(seq_group)
-            else:
-                running.append(seq_group)
-        self.running = running
-
-    def _allocate(self, seq_group: SequenceGroup) -> None:
-        self.block_manager.allocate(seq_group)
-        for seq in seq_group.seqs:
-            seq.status = SequenceStatus.RUNNING
-        # FIXME(woosuk): Support interactive generation.
-        if seq_group.group_id not in self.num_steps:
-            self.num_steps[seq_group.group_id] = 0
-
-    def _append(
-        self,
-        seq_group: SequenceGroup,
-        blocks_to_copy: Dict[int, List[int]],
-    ) -> None:
-        for seq in seq_group.get_seqs(status=SequenceStatus.RUNNING):
-            ret = self.block_manager.append(seq)
-            if ret is not None:
-                src_block, dst_block = ret
-                if src_block in blocks_to_copy:
-                    blocks_to_copy[src_block].append(dst_block)
-                else:
-                    blocks_to_copy[src_block] = [dst_block]
-
-    def _preempt(
-        self,
-        seq_group: SequenceGroup,
-        blocks_to_swap_out: Dict[int, int],
-        preemption_mode: Optional[PreemptionMode] = None,
-    ) -> None:
-        # If preemption mode is not specified, we determine the mode as follows:
-        # We use recomputation by default since it incurs lower overhead than
-        # swapping. However, when the sequence group has multiple sequences
-        # (e.g., beam search), recomputation is not supported. In such a case,
-        # we use swapping instead.
-        # FIXME(woosuk): This makes our scheduling policy a bit bizarre.
-        # As swapped sequences are prioritized over waiting sequences,
-        # sequence groups with multiple sequences are implicitly prioritized
-        # over sequence groups with a single sequence.
-        # TODO(woosuk): Support recomputation for sequence groups with multiple
-        # sequences. This may require a more sophisticated CUDA kernel.
-        if preemption_mode is None:
-            seqs = seq_group.get_seqs(status=SequenceStatus.RUNNING)
-            if len(seqs) == 1:
-                preemption_mode = PreemptionMode.RECOMPUTE
-            else:
-                preemption_mode = PreemptionMode.SWAP
-        if preemption_mode == PreemptionMode.RECOMPUTE:
-            self._preempt_by_recompute(seq_group)
-        elif preemption_mode == PreemptionMode.SWAP:
-            self._preempt_by_swap(seq_group, blocks_to_swap_out)
-        else:
-            assert False, 'Invalid preemption mode.'
-
-    def _preempt_by_recompute(
-        self,
-        seq_group: SequenceGroup,
-    ) -> None:
-        seqs = seq_group.get_seqs(status=SequenceStatus.RUNNING)
-        assert len(seqs) == 1
-        for seq in seqs:
-            seq.status = SequenceStatus.WAITING
-            self.block_manager.free(seq)
-        self.waiting.append(seq_group)
-
-    def _preempt_by_swap(
-        self,
-        seq_group: SequenceGroup,
-        blocks_to_swap_out: Dict[int, int],
-    ) -> None:
-        seqs = seq_group.get_seqs(status=SequenceStatus.RUNNING)
-        for seq in seqs:
-            seq.status = SequenceStatus.SWAPPED
-        self._swap_out(seq_group, blocks_to_swap_out)
-        self.swapped.append(seq_group)
-
-    def _free_seq(self, seq: Sequence) -> None:
-        seq.status = SequenceStatus.FINISHED
-        self.block_manager.free(seq)
-
-    def _free_seq_group(self, seq_group: SequenceGroup) -> None:
-        group_id = seq_group.group_id
-        del self.num_steps[group_id]
-        del self.sampling_params[group_id]
-
-    def _swap_in(
-        self,
-        seq_group: SequenceGroup,
-        blocks_to_swap_in: Dict[int, int],
-    ) -> None:
-        mapping = self.block_manager.swap_in(seq_group)
-        blocks_to_swap_in.update(mapping)
-        for seq in seq_group.get_seqs(status=SequenceStatus.SWAPPED):
-            seq.status = SequenceStatus.RUNNING
-
-    def _swap_out(
-        self,
-        seq_group: SequenceGroup,
-        blocks_to_swap_out: Dict[int, int],
-    ) -> None:
-        assert self.block_manager.can_swap_out(seq_group)
-        mapping = self.block_manager.swap_out(seq_group)
-        blocks_to_swap_out.update(mapping)
-        for seq in seq_group.get_seqs(status=SequenceStatus.RUNNING):
-            seq.status = SequenceStatus.SWAPPED
-
-    def reset_stats(self) -> None:
-        self.stats.reset(self.num_gpu_blocks, self.num_cpu_blocks)
-
-    def save_stats(
-        self,
-        output_dir: str,
-    ) -> None:
-        assert self.collect_stats, 'Statistics collection is disabled.'
-        self.stats.save(output_dir)
-
-
-class Stats:
-
-    def __init__(
-        self,
-        num_gpu_blocks: int,
-        num_cpu_blocks: int,
-    ) -> None:
-        self.start_time: float = time.time()
-        self.num_gpu_blocks = num_gpu_blocks
-        self.num_cpu_blocks = num_cpu_blocks
-
-        self.timestamps: List[float] = []
-        self.input_lens: List[int] = []
-        self.swap_out_lens: List[int] = []
-        self.swap_in_lens: List[int] = []
-        self.num_preemption: List[int] = []
-        self.num_waiting: List[int] = []
-        self.num_running: List[int] = []
-        self.num_swapped: List[int] = []
-        self.gpu_cache_usage: List[float] = []
-        self.cpu_cache_usage: List[float] = []
-
-        self.num_logical_blocks: List[int] = []
-        self.num_logical_tokens: List[int] = []
-        self.num_physical_blocks: List[int] = []
-        self.num_physical_tokens: List[int] = []
-        self.num_reserved_tokens: List[int] = []
-
-    def reset(
-        self,
-        num_gpu_blocks: int,
-        num_cpu_blocks: int,
-    ) -> None:
-        self.__init__(num_gpu_blocks, num_cpu_blocks)
-
-    def to_dict(self) -> Dict[str, Any]:
-        return {
-            'start_time': self.start_time,
-            'num_gpu_blocks': self.num_gpu_blocks,
-            'num_cpu_blocks': self.num_cpu_blocks,
-            'timestamps': self.timestamps,
-            'input_lens': self.input_lens,
-            'swap_out_lens': self.swap_out_lens,
-            'swap_in_lens': self.swap_in_lens,
-            'num_preemption': self.num_preemption,
-            'num_waiting': self.num_waiting,
-            'num_running': self.num_running,
-            'num_swapped': self.num_swapped,
-            'gpu_cache_usage': self.gpu_cache_usage,
-            'cpu_cache_usage': self.cpu_cache_usage,
-            'num_logical_blocks': self.num_logical_blocks,
-            'num_logical_tokens': self.num_logical_tokens,
-            'num_physical_blocks': self.num_physical_blocks,
-            'num_physical_tokens': self.num_physical_tokens,
-            'num_reserved_tokens': self.num_reserved_tokens,
-        }
-
-    def save(self, output_dir: str) -> None:
-        with open(os.path.join(output_dir, 'stats.pkl'), 'wb') as f:
-            pickle.dump(self.to_dict(), f)

cacheflow/master/server.py

@@ -1,192 +0,0 @@
-import argparse
-from typing import List, Tuple
-import random
-
-import ray
-
-from cacheflow.master.scheduler import Scheduler
-from cacheflow.models import get_memory_analyzer
-from cacheflow.worker.controller import Controller, DeviceID
-from cacheflow.sequence import SequenceGroup
-from cacheflow.sampling_params import SamplingParams
-
-
-class Server:
-    def __init__(
-        self,
-        model: str,
-        model_path: str,
-        use_dummy_weights: bool,
-        pipeline_parallel_size: int,
-        tensor_parallel_size: int,
-        block_size: int,
-        dtype: str,
-        seed: int,
-        swap_space: int,
-        max_num_batched_tokens: int,
-        max_num_sequences: int,
-        num_nodes: int,
-        num_devices_per_node: int,
-        distributed_init_method: str,
-        all_stage_devices: List[List[DeviceID]],
-        gpu_memory: int,
-        cpu_memory: int,
-        collect_stats: bool = False,
-        do_memory_analysis: bool = False,
-    ):
-        self.num_nodes = num_nodes
-        self.num_devices_per_node = num_devices_per_node
-        self.world_size = pipeline_parallel_size * tensor_parallel_size
-
-        self.memory_analyzer = get_memory_analyzer(
-            model_name=model,
-            block_size=block_size,
-            dtype=dtype,
-            gpu_memory=gpu_memory,
-            cpu_memory=cpu_memory,
-            tensor_parallel_size=tensor_parallel_size,
-        )
-        self.num_gpu_blocks = self.memory_analyzer.get_max_num_gpu_blocks(
-            max_num_batched_tokens=max_num_batched_tokens)
-        self.num_cpu_blocks = self.memory_analyzer.get_max_num_cpu_blocks(
-            swap_space=swap_space)
-        print(f'# GPU blocks: {self.num_gpu_blocks}, '
-              f'# CPU blocks: {self.num_cpu_blocks}')
-
-        # Create a controller for each pipeline stage.
-        self.controllers: List[Controller] = []
-        for i in range(pipeline_parallel_size):
-            controller = Controller(
-                stage_id=i,
-                stage_devices=all_stage_devices[i],
-                world_size=self.world_size,
-                pipeline_parallel_size=pipeline_parallel_size,
-                tensor_parallel_size=tensor_parallel_size,
-                distributed_init_method=distributed_init_method,
-                model_name=model,
-                block_size=block_size,
-                num_gpu_blocks=self.num_gpu_blocks,
-                num_cpu_blocks=self.num_cpu_blocks,
-                dtype=dtype,
-                seed=seed,
-                model_path=model_path,
-                use_dummy_weights=use_dummy_weights,
-                max_num_batched_tokens=max_num_batched_tokens,
-            )
-            self.controllers.append(controller)
-
-        # Create a scheduler.
-        self.scheduler = Scheduler(
-            controllers=self.controllers,
-            block_size=block_size,
-            num_gpu_blocks=self.num_gpu_blocks,
-            num_cpu_blocks=self.num_cpu_blocks,
-            max_num_batched_tokens=max_num_batched_tokens,
-            max_num_sequences=max_num_sequences,
-            collect_stats=collect_stats,
-            do_memory_analysis=do_memory_analysis,
-        )
-        # Connect the controllers.
-        for i in range(len(self.controllers) - 1):
-            self.controllers[i].set_next(self.controllers[i + 1])
-        self.controllers[-1].set_next(self.scheduler)
-
-    def add_sequence_groups(
-        self,
-        sequence_groups: List[Tuple[SequenceGroup, SamplingParams]]
-    ):
-        self.scheduler.add_sequence_groups(sequence_groups)
-
-    def step(self):
-        return self.scheduler.step()
-
-    def has_unfinished_requests(self):
-        return (self.scheduler.waiting or self.scheduler.running or
-                self.scheduler.swapped)
-
-
-def initialize_ray_cluster(
-    address: str = 'auto',
-    pipeline_parallel_size: int = 1,
-    tensor_parallel_size: int = 1,
-) -> Tuple[int, int, str, List[List[DeviceID]]]:
-    # Connect to a ray cluster.
-    ray.init(address=address)
-
-    # Assume we have a uniform cluster that each node has the same number of
-    # GPUs for now.
-    valid_node_resources = []
-    num_devices_per_node = None
-    for node in ray.nodes():
-        if (not node['Alive']) or node['Resources']['GPU'] <= 0:
-            continue
-        if num_devices_per_node is None:
-            num_devices_per_node = node['Resources']['GPU']
-        else:
-            assert num_devices_per_node == node['Resources']['GPU'], (
-                "The number of GPUs per node is not uniform.")
-        for key in node['Resources']:
-            if key.startswith('node:'):
-                valid_node_resources.append(key)
-
-    num_nodes = len(valid_node_resources)
-
-    assert (pipeline_parallel_size * tensor_parallel_size
-            <= num_nodes * num_devices_per_node), (
-                "The number of required GPUs exceeds the total number of "
-                "available GPUs.")
-    if tensor_parallel_size >= num_devices_per_node:
-        assert tensor_parallel_size % num_devices_per_node == 0, (
-            "The number of tensor parallelism is not divisible by the "
-            "number of GPUs per node.")
-    else:
-        assert num_devices_per_node % tensor_parallel_size == 0, (
-            "The number of GPUs per node is not divisible by the number "
-            "of tensor parallelism.")
-
-    # Assign GPUs to pipeline stages.
-    rank = 0
-    current_node_id = 0
-    current_device_id = 0
-    distributed_init_method = None
-    all_stage_devices = []
-
-    for i in range(pipeline_parallel_size):
-        stage_devices = []
-        for j in range(tensor_parallel_size):
-            node_resource = valid_node_resources[current_node_id]
-            stage_devices.append((rank, node_resource, current_device_id))
-            if distributed_init_method is None:
-                ip = node_resource.split("node:")[-1]
-                port = random.randint(10000, 20000)
-                distributed_init_method = f"tcp://{ip}:{port}"
-            rank += 1
-            current_device_id += 1
-            if current_device_id >= num_devices_per_node:
-                current_node_id += 1
-                current_device_id = 0
-        all_stage_devices.append(stage_devices)
-
-    return (num_nodes, num_devices_per_node, distributed_init_method,
-            all_stage_devices)
-
-
-def add_server_arguments(parser: argparse.ArgumentParser):
-    # Model arguments
-    parser.add_argument('--model', type=str, default='facebook/opt-125m', help='model name')
-    parser.add_argument('--model-path', type=str, default='~/.cacheflow/model_weights',
-                        help='model path to download and load the weights')
-    # Parallel arguments
-    parser.add_argument('--pipeline-parallel-size', '-pp', type=int, default=1, help='number of pipeline stages')
-    parser.add_argument('--tensor-parallel-size', '-tp', type=int, default=1, help='number of tensor parallel replicas')
-    # KV cache arguments
-    parser.add_argument('--block-size', type=int, default=16, choices=[1, 2, 4, 8, 16, 32, 64, 128, 256], help='token block size')
-    # NOTE(woosuk): If FlashAttention is used, the float data type is not supported.
-    parser.add_argument('--dtype', type=str, default='half', choices=['half'], help='data type')
-    # TODO(woosuk): Support fine-grained seeds (e.g., seed per request).
-    parser.add_argument('--seed', type=int, default=0, help='random seed')
-    parser.add_argument('--swap-space', type=int, default=20, help='CPU swap space size (GiB) per GPU')
-    parser.add_argument('--max-num-batched-tokens', type=int, default=2560, help='maximum number of batched tokens per iteration')
-    parser.add_argument('--max-num-sequences', type=int, default=256, help='maximum number of sequences per iteration')
-    parser.add_argument('--use-dummy-weights', action='store_true', help='use dummy values for model weights')
-    return parser

cacheflow/master/simple_frontend.py

@@ -1,69 +0,0 @@
-import time
-from typing import List, Optional, Set, Tuple
-
-from transformers import AutoTokenizer
-
-from cacheflow.sampling_params import SamplingParams
-from cacheflow.sequence import Sequence, SequenceGroup
-from cacheflow.utils import Counter
-
-
-class SimpleFrontend:
-
-    def __init__(
-        self,
-        model_name: str,
-        block_size: int,
-    ) -> None:
-        self.block_size = block_size
-
-        self.tokenizer = AutoTokenizer.from_pretrained(model_name)
-        self.seq_group_counter = Counter()
-        self.seq_counter = Counter()
-        self.inputs: List[Tuple[SequenceGroup, SamplingParams]] = []
-
-    def add_eos_token(self, sampling_params: SamplingParams) -> SamplingParams:
-        # Stop generation when we see an EOS token.
-        sampling_params.stop_token_ids.add(self.tokenizer.eos_token_id)
-        return sampling_params
-
-    def query(
-        self,
-        prompt: str,
-        sampling_params: SamplingParams,
-    ) -> None:
-        token_ids = self.tokenizer.encode(prompt)
-        self._add_query(token_ids, sampling_params)
-
-    def _add_query(
-        self,
-        token_ids: List[int],
-        sampling_params: SamplingParams,
-        arrival_time: Optional[float] = None,
-    ) -> None:
-        if arrival_time is None:
-            arrival_time = time.time()
-        seqs: List[Sequence] = []
-        for _ in range(sampling_params.n):
-            seq_id = next(self.seq_counter)
-            seq = Sequence(seq_id, token_ids, block_size=self.block_size)
-            seqs.append(seq)
-
-        group_id = next(self.seq_group_counter)
-        seq_group = SequenceGroup(group_id, seqs, arrival_time)
-        self.inputs.append((seq_group, sampling_params))
-
-    def get_inputs(self) -> List[Tuple[SequenceGroup, SamplingParams]]:
-        inputs = self.inputs
-        self.inputs = []
-        return inputs
-
-    def print_response(
-        self,
-        seq_group: SequenceGroup,
-    ) -> None:
-        for seq in seq_group.seqs:
-            token_ids = seq.get_token_ids()
-            output = self.tokenizer.decode(token_ids, skip_special_tokens=True)
-            output = output.strip()
-            print(f'Seq {seq.seq_id}: {output!r}')

cacheflow/models/__init__.py

@@ -1,10 +0,0 @@
-from cacheflow.models.input_metadata import InputMetadata
-from cacheflow.models.model_utils import get_memory_analyzer
-from cacheflow.models.model_utils import get_model
-
-
-__all__ = [
-    'InputMetadata',
-    'get_memory_analyzer',
-    'get_model',
-]

cacheflow/models/activation.py

@@ -1,20 +0,0 @@
-import torch
-import torch.nn as nn
-
-from cacheflow import activation_ops
-
-
-class SiluAndMul(nn.Module):
-
-    def __init__(self):
-        super().__init__()
-
-    def forward(
-        self,
-        x: torch.Tensor,        # (num_tokens, 2 * d)
-    ) -> torch.Tensor:          # (num_tokens, d)
-        num_tokens = x.shape[0]
-        d = x.shape[1] // 2
-        out = torch.empty(num_tokens, d, dtype=x.dtype, device=x.device)
-        activation_ops.silu_and_mul(out, x)
-        return out

cacheflow/models/attention.py

@@ -1,207 +0,0 @@
-from typing import Optional
-
-from flash_attn.flash_attn_interface import _flash_attn_forward
-import torch
-import torch.nn as nn
-
-from cacheflow import attention_ops
-from cacheflow import cache_ops
-from cacheflow import pos_encoding_ops
-from cacheflow.models import InputMetadata
-
-
-class GPTCacheFlowAttention(nn.Module):
-
-    def __init__(self, scale: float) -> None:
-        super().__init__()
-        self.scale = float(scale)
-
-    def multi_query_kv_attention(
-        self,
-        output: torch.Tensor,                   # [num_prompt_tokens, num_heads, head_size]
-        query: torch.Tensor,                    # [num_prompt_tokens, num_heads, head_size]
-        key: torch.Tensor,                      # [num_prompt_tokens, num_heads, head_size]
-        value: torch.Tensor,                    # [num_prompt_tokens, num_heads, head_size]
-        cumulative_prompt_lens: torch.Tensor,   # [num_prompts + 1]
-        max_prompt_len: int,
-    ) -> None:
-        if query.dtype == torch.float:
-            raise ValueError('The float data type is not supported by '
-                             'FlashAttention. Use the half data type instead.')
-        head_size = query.shape[-1]
-        if head_size > 128:
-            raise ValueError('FlashAttention does not support head_size > 128.')
-
-        # Directly call FlashAttention's internal function to avoid allocating
-        # a new tensor for the output.
-        _flash_attn_forward(
-            query,
-            key,
-            value,
-            output,
-            cumulative_prompt_lens,
-            cumulative_prompt_lens,
-            max_prompt_len,
-            max_prompt_len,
-            dropout_p=0.0,
-            softmax_scale=self.scale,
-            causal=True,
-            return_softmax=False,
-        )
-
-    def single_query_cached_kv_attention(
-        self,
-        output: torch.Tensor,           # [num_generation_tokens, num_heads, head_size]
-        query: torch.Tensor,            # [num_generation_tokens, num_heads, head_size]
-        key_cache: torch.Tensor,        # [num_blocks, num_heads, head_size/x, block_size, x]
-        value_cache: torch.Tensor,      # [num_blocks, num_heads, head_size, block_size]
-        input_metadata: InputMetadata,
-    ) -> None:
-        head_size = value_cache.shape[2]
-        supported_head_sizes = [32, 64, 80, 96, 128, 160, 192, 256]
-        if head_size not in supported_head_sizes:
-            raise ValueError(f'head_size ({head_size}) is not supported by '
-                             'the single_query_cached_kv_attention kernel. '
-                             'Use one of the following head sizes: '
-                             f'{supported_head_sizes}.')
-
-        block_size = value_cache.shape[3]
-        attention_ops.single_query_cached_kv_attention(
-            output,
-            query,
-            key_cache,
-            value_cache,
-            self.scale,
-            input_metadata.block_tables,
-            input_metadata.context_lens,
-            block_size,
-            input_metadata.max_context_len,
-        )
-
-    def forward(
-        self,
-        query: torch.Tensor,                    # [num_tokens, num_heads * head_size]
-        key: torch.Tensor,                      # [num_tokens, num_heads * head_size]
-        value: torch.Tensor,                    # [num_tokens, num_heads * head_size]
-        key_cache: torch.Tensor,                # [num_blocks, num_heads, head_size/x, block_size, x]
-        value_cache: torch.Tensor,              # [num_blocks, num_heads, head_size, block_size]
-        input_metadata: InputMetadata,
-        cache_event: Optional[torch.cuda.Event],
-    ) -> torch.Tensor:                          # [num_tokens, num_heads * head_size]
-        # NOTE: The query, key, and value tensors must be sliced from a qkv
-        # tensor of shape [num_tokens, 3 * num_heads * head_size].
-
-        # Reshape the query, key, and value tensors.
-        num_heads = value_cache.shape[1]
-        head_size = value_cache.shape[2]
-        query = query.view(-1, num_heads, head_size)
-        key = key.view(-1, num_heads, head_size)
-        value = value.view(-1, num_heads, head_size)
-
-        # Pre-allocate the output tensor.
-        output = torch.empty_like(query)
-
-        # Compute the attention op for prompts.
-        num_prompt_tokens = input_metadata.num_prompt_tokens
-        if num_prompt_tokens > 0:
-            self.multi_query_kv_attention(
-                output[:num_prompt_tokens],
-                query[:num_prompt_tokens],
-                key[:num_prompt_tokens],
-                value[:num_prompt_tokens],
-                input_metadata.cumulative_prompt_lens,
-                input_metadata.max_prompt_len,
-            )
-
-        # Wait until the cache op is done.
-        if cache_event is not None:
-            cache_event.wait()
-
-        # Reshape the keys and values and store them in the cache.
-        num_valid_tokens = input_metadata.num_valid_tokens
-        if num_valid_tokens > 0:
-            # The stride is 3 because the key and value are sliced from qkv.
-            cache_ops.reshape_and_cache(
-                key[:num_valid_tokens],
-                value[:num_valid_tokens],
-                key_cache,
-                value_cache,
-                input_metadata.slot_mapping,
-            )
-
-        if input_metadata.num_generation_tokens > 0:
-            # Compute the attention op for generation tokens.
-            self.single_query_cached_kv_attention(
-                output[num_prompt_tokens:num_valid_tokens],
-                query[num_prompt_tokens:num_valid_tokens],
-                key_cache,
-                value_cache,
-                input_metadata)
-
-        # Reshape the output tensor.
-        # NOTE(woosuk): The output tensor may include paddings.
-        return output.view(-1, num_heads * head_size)
-
-
-class OPTCacheFlowAttention(GPTCacheFlowAttention):
-    """OPT uses the same attention mechanism as GPT."""
-
-    def __init__(self, scale: float) -> None:
-        super().__init__(scale)
-
-
-class LlamaCacheFlowAttention(GPTCacheFlowAttention):
-    """Llama uses GPT-NeoX style rotary embedding."""
-
-    def __init__(
-        self,
-        scale: float,
-        head_size: int,
-        max_position: int = 8192,
-        base: int = 10000,
-    ) -> None:
-        super().__init__(scale)
-
-        # Create the cos and sin cache.
-        inv_freq = 1.0 / (base ** (torch.arange(0, head_size, 2) / head_size))
-        t = torch.arange(max_position).float()
-        freqs = torch.einsum('i,j -> ij', t, inv_freq.float())
-        cos = freqs.cos()
-        sin = freqs.sin()
-        cache = torch.cat((cos, sin), dim=-1)
-
-        # FIXME(woosuk): This assumes that we configure the default dtype when
-        # initializing the model. Make it more robust.
-        torch_dtype = torch.get_default_dtype()
-        cache = cache.to(torch_dtype)
-        # Embedding size: [max_position, head_size]
-        self.register_buffer('cos_sin_cache', cache, persistent=False)
-
-    def forward(
-        self,
-        positions: torch.LongTensor,            # [num_tokens]
-        query: torch.Tensor,                    # [num_tokens, num_heads * head_size]
-        key: torch.Tensor,                      # [num_tokens, num_heads * head_size]
-        value: torch.Tensor,                    # [num_tokens, num_heads * head_size]
-        key_cache: torch.Tensor,                # [num_blocks, num_heads, head_size/x, block_size, x]
-        value_cache: torch.Tensor,              # [num_blocks, num_heads, head_size, block_size]
-        input_metadata: InputMetadata,
-        cache_event: Optional[torch.cuda.Event],
-    ) -> torch.Tensor:                          # [num_tokens, num_heads * head_size]
-        # Apply rotary embedding to the query and key before passing them
-        # to the attention op.
-        pos_encoding_ops.rotary_embedding_neox(
-            positions,
-            query,
-            key,
-            self.cos_sin_cache,
-        )
-        return super().forward(
-            query,
-            key,
-            value,
-            key_cache,
-            value_cache,
-            input_metadata,
-            cache_event,
-        )

cacheflow/models/input_metadata.py

@@ -1,55 +0,0 @@
-from typing import List, Dict, Tuple
-
-import torch
-
-from cacheflow.sampling_params import SamplingParams
-
-
-class InputMetadata:
-
-    def __init__(
-        self,
-        seq_groups: List[Tuple[List[int], SamplingParams]],
-        seq_logprobs: Dict[int, float],                         # Seq id -> cumulative logprobs.
-        prompt_lens: List[int],
-        cumulative_prompt_lens: torch.Tensor,
-        slot_mapping: torch.Tensor,
-        context_lens: torch.Tensor,
-        max_context_len: int,
-        block_tables: torch.Tensor,
-    ) -> None:
-        self.seq_groups = seq_groups
-        self.seq_logprobs = seq_logprobs
-        self.prompt_lens = prompt_lens
-        self.cumulative_prompt_lens = cumulative_prompt_lens
-        self.slot_mapping = slot_mapping
-        self.context_lens = context_lens
-        self.max_context_len = max_context_len
-        self.block_tables = block_tables
-
-        self.num_prompts = len(prompt_lens)
-        self.num_prompt_tokens = sum(prompt_lens)
-        self.max_prompt_len = max(prompt_lens) if prompt_lens else 0
-        self.num_generation_tokens = context_lens.shape[0]
-        self.num_valid_tokens = slot_mapping.shape[0]
-        if block_tables.numel() > 0:
-            self.max_num_blocks_per_seq = block_tables.shape[1]
-        else:
-            self.max_num_blocks_per_seq = 0
-        assert block_tables.shape[0] == self.num_generation_tokens
-        assert context_lens.shape[0] == self.num_generation_tokens
-
-    def __repr__(self) -> str:
-        return (f'InputMetadata('
-                f'num_prompts={self.num_prompts}, '
-                f'num_prompt_tokens={self.num_prompt_tokens}, '
-                f'max_prompt_len={self.max_prompt_len}, '
-                f'num_generation_tokens={self.num_generation_tokens}, '
-                f'num_valid_tokens={self.num_valid_tokens}, '
-                f'max_num_blocks_per_seq={self.max_num_blocks_per_seq}, '
-                f'max_context_len={self.max_context_len}), '
-                f'prompt_lens={self.prompt_lens}, '
-                f'cumulative_prompt_lens={self.cumulative_prompt_lens}, '
-                f'slot_mapping={self.slot_mapping}, '
-                f'context_lens={self.context_lens}, '
-                f'block_tables={self.block_tables})')

cacheflow/models/layernorm.py

@@ -1,26 +0,0 @@
-import torch
-import torch.nn as nn
-
-from cacheflow import layernorm_ops
-
-
-class RMSNorm(nn.Module):
-
-    def __init__(
-        self,
-        hidden_size: int,
-        eps: float = 1e-6,
-    ) -> None:
-        super().__init__()
-        self.weight = nn.Parameter(torch.ones(hidden_size))
-        self.variance_epsilon = eps
-
-    def forward(self, x: torch.Tensor) -> torch.Tensor:
-        out = torch.empty_like(x)
-        layernorm_ops.rms_norm(
-            out,
-            x,
-            self.weight.data,
-            self.variance_epsilon,
-        )
-        return out

cacheflow/models/llama.py

@@ -1,292 +0,0 @@
-"""1D LLaMA model compatible with HuggingFace weights."""
-import os
-import glob
-import filelock
-from tqdm import tqdm
-from typing import Dict, List, Optional, Tuple
-
-import numpy as np
-import torch
-from torch import nn
-from transformers import LlamaConfig
-
-from cacheflow.models import InputMetadata
-from cacheflow.models.activation import SiluAndMul
-from cacheflow.models.attention import LlamaCacheFlowAttention
-from cacheflow.models.layernorm import RMSNorm
-from cacheflow.models.sample import Sampler
-from cacheflow.parallel_utils.parallel_state import (
-    get_tensor_model_parallel_rank, get_tensor_model_parallel_world_size)
-from cacheflow.parallel_utils.tensor_parallel import (VocabParallelEmbedding,
-                                                      ColumnParallelLinear,
-                                                      RowParallelLinear)
-from cacheflow.sequence import SequenceOutputs
-
-KVCache = Tuple[torch.Tensor, torch.Tensor]
-
-
-class LlamaMLP(nn.Module):
-
-    def __init__(
-        self,
-        hidden_size: int,
-        intermediate_size: int,
-        hidden_act: str,
-    ):
-        super().__init__()
-        self.gate_up_proj = ColumnParallelLinear(hidden_size, 2 * intermediate_size,
-                                                 bias=False, gather_output=False,
-                                                 perform_initialization=False)
-        self.down_proj = RowParallelLinear(intermediate_size, hidden_size,
-                                           bias=False, input_is_parallel=True,
-                                           perform_initialization=False)
-        if hidden_act != 'silu':
-            raise ValueError(f'Unsupported activation: {hidden_act}. '
-                             'Only silu is supported for now.')
-        self.act_fn = SiluAndMul()
-
-    def forward(self, x):
-        gate_up, _ = self.gate_up_proj(x)
-        x = self.act_fn(gate_up)
-        x, _ = self.down_proj(x)
-        return x
-
-
-class LlamaAttention(nn.Module):
-
-    def __init__(
-        self,
-        hidden_size: int,
-        num_heads: int,
-    ):
-        super().__init__()
-        self.hidden_size = hidden_size
-        tensor_model_parallel_world_size = get_tensor_model_parallel_world_size()
-        self.total_num_heads = num_heads
-        assert self.total_num_heads % tensor_model_parallel_world_size == 0
-        self.num_heads = self.total_num_heads // tensor_model_parallel_world_size
-        self.head_dim = hidden_size // self.total_num_heads
-        self.scaling = self.head_dim ** -0.5
-
-        self.qkv_proj = ColumnParallelLinear(
-            hidden_size,
-            3 * self.total_num_heads * self.head_dim,
-            bias=False,
-            gather_output=False,
-            perform_initialization=False,
-        )
-        self.o_proj = RowParallelLinear(
-            self.total_num_heads * self.head_dim,
-            hidden_size,
-            bias=False,
-            input_is_parallel=True,
-            perform_initialization=False,
-        )
-        self.attn = LlamaCacheFlowAttention(self.scaling, self.head_dim)
-
-    def forward(
-        self,
-        positions: torch.LongTensor,
-        hidden_states: torch.Tensor,
-        kv_cache: KVCache,
-        input_metadata: InputMetadata,
-        cache_event: Optional[torch.cuda.Event],
-    ) -> torch.Tensor:
-        qkv, _ = self.qkv_proj(hidden_states)
-        q, k, v = qkv.chunk(chunks=3, dim=-1)
-        k_cache, v_cache = kv_cache
-        attn_output = self.attn(
-            positions, q, k, v, k_cache, v_cache, input_metadata, cache_event)
-        output, _ = self.o_proj(attn_output)
-        return output
-
-
-class LlamaDecoderLayer(nn.Module):
-
-    def __init__(self, config: LlamaConfig):
-        super().__init__()
-        self.hidden_size = config.hidden_size
-        self.self_attn = LlamaAttention(
-            hidden_size=self.hidden_size,
-            num_heads=config.num_attention_heads,
-        )
-        self.mlp = LlamaMLP(
-            hidden_size=self.hidden_size,
-            intermediate_size=config.intermediate_size,
-            hidden_act=config.hidden_act,
-        )
-        self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
-        self.post_attention_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
-
-    def forward(
-        self,
-        positions: torch.LongTensor,
-        hidden_states: torch.Tensor,
-        kv_cache: KVCache,
-        input_metadata: InputMetadata,
-        cache_event: Optional[torch.cuda.Event],
-    ) -> torch.Tensor:
-        # Self Attention
-        residual = hidden_states
-        hidden_states = self.input_layernorm(hidden_states)
-        hidden_states = self.self_attn(
-            positions=positions,
-            hidden_states=hidden_states,
-            kv_cache=kv_cache,
-            input_metadata=input_metadata,
-            cache_event=cache_event,
-        )
-        hidden_states = residual + hidden_states
-
-        # Fully Connected
-        residual = hidden_states
-        hidden_states = self.post_attention_layernorm(hidden_states)
-        hidden_states = self.mlp(hidden_states)
-        hidden_states = residual + hidden_states
-        return hidden_states
-
-
-class LlamaModel(nn.Module):
-
-    def __init__(self, config: LlamaConfig):
-        super().__init__()
-        self.config = config
-        self.padding_idx = config.pad_token_id
-        self.vocab_size = config.vocab_size
-
-        self.embed_tokens = VocabParallelEmbedding(config.vocab_size, config.hidden_size,
-                                                   perform_initialization=False)
-        self.layers = nn.ModuleList([LlamaDecoderLayer(config) for _ in range(config.num_hidden_layers)])
-        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
-
-    def forward(
-        self,
-        input_ids: torch.LongTensor,
-        positions: torch.LongTensor,
-        kv_caches: List[KVCache],
-        input_metadata: InputMetadata,
-        cache_events: Optional[List[torch.cuda.Event]],
-    ) -> torch.Tensor:
-        hidden_states = self.embed_tokens(input_ids)
-        for i in range(len(self.layers)):
-            if cache_events is None:
-                cache_event = None
-            else:
-                cache_event = cache_events[i]
-            layer = self.layers[i]
-            hidden_states = layer(
-                positions,
-                hidden_states,
-                kv_caches[i],
-                input_metadata,
-                cache_event,
-            )
-        hidden_states = self.norm(hidden_states)
-        return hidden_states
-
-
-class LlamaForCausalLM(nn.Module):
-    def __init__(self, config):
-        super().__init__()
-        self.config = config
-        self.model = LlamaModel(config)
-        self.lm_head = ColumnParallelLinear(config.hidden_size,
-                                            config.vocab_size,
-                                            bias=False,
-                                            gather_output=False,
-                                            perform_initialization=False)
-        self.sampler = Sampler()
-
-    def forward(
-        self,
-        input_ids: torch.LongTensor,
-        positions: torch.LongTensor,
-        kv_caches: List[KVCache],
-        input_metadata: InputMetadata,
-        cache_events: Optional[List[torch.cuda.Event]],
-    ) -> Dict[int, SequenceOutputs]:
-        hidden_states = self.model(
-            input_ids, positions, kv_caches, input_metadata, cache_events)
-        next_tokens = self.sampler(
-            self.lm_head.weight, hidden_states, input_metadata)
-        return next_tokens
-
-    _column_parallel_weights = ["embed_tokens.weight", "lm_head.weight",
-                                "qkv_proj.weight", "gate_proj.weight",
-                                "up_proj.weight"]
-    _row_parallel_weights = ["o_proj.weight", "down_proj.weight"]
-
-    def load_weights(self, weights_path: str):
-        tensor_model_parallel_rank = get_tensor_model_parallel_rank()
-        state_dict = self.state_dict()
-        for name, param in state_dict.items():
-            if "qkv_proj" in name or "gate_up_proj" in name:
-                if "qkv_proj" in name:
-                    original_name = "qkv_proj"
-                    weight_names = ["q_proj", "k_proj", "v_proj"]
-                    shard_size = param.shape[0] // 3
-                else:
-                    original_name = "gate_up_proj"
-                    weight_names = ["gate_proj", "up_proj"]
-                    shard_size = param.shape[0] // 2
-                weights_to_concat = []
-                for weight_name in weight_names:
-                    weight = np.load(os.path.join(
-                        weights_path, name.replace(original_name, weight_name)))
-                    weights_to_concat.append(weight[
-                        shard_size * tensor_model_parallel_rank
-                        :shard_size * (tensor_model_parallel_rank + 1)])
-                loaded_weight = torch.from_numpy(
-                    np.concatenate(weights_to_concat, axis=0))
-            else:
-                loaded_weight = torch.from_numpy(
-                    np.load(os.path.join(weights_path, name)))
-                for p in self._column_parallel_weights:
-                    if p in name:
-                        shard_size = param.shape[0]
-                        loaded_weight = loaded_weight[
-                            shard_size * tensor_model_parallel_rank
-                            :shard_size * (tensor_model_parallel_rank + 1)]
-                        break
-                for p in self._row_parallel_weights:
-                    if p in name:
-                        shard_size = param.shape[1]
-                        loaded_weight = loaded_weight[
-                            :,
-                            shard_size * tensor_model_parallel_rank
-                            :shard_size * (tensor_model_parallel_rank + 1)]
-                        break
-
-            assert param.shape == loaded_weight.shape
-            param.data.copy_(loaded_weight)
-
-    @staticmethod
-    def get_weights(model_name: str, path: str):
-        if not os.path.isfile(os.path.join(model_name, "config.json")):
-            raise ValueError("LLaMA model's model_name has to be a path"
-                             "to the huggingface model's directory.")
-        path = os.path.join(model_name, f"np")
-        path = os.path.abspath(os.path.expanduser(path))
-        os.makedirs(path, exist_ok=True)
-        lock_path = os.path.join(path, "file_lock")
-        lock = filelock.FileLock(lock_path)
-
-        with lock:
-            test_weight_path = os.path.join(path, "model.embed_tokens.weight")
-            if os.path.exists(test_weight_path):
-                return path
-
-            bin_files = glob.glob(os.path.join(model_name, "*.bin"))
-
-            for bin_file in tqdm(bin_files, desc="Convert format"):
-                state = torch.load(bin_file, map_location="cpu")
-                for name, param in tqdm(state.items(), leave=False):
-                    param_path = os.path.join(path, name)
-                    with open(param_path, "wb") as f:
-                        np.save(f, param.cpu().detach().numpy())
-
-            return path
-
-    def initialize_dummy_weights(self) -> None:
-        for param in self.state_dict().values():
-            param.data.uniform_(-0.1, 0.1)

cacheflow/models/memory_analyzer.py

@@ -1,240 +0,0 @@
-import torch
-from transformers import AutoConfig
-
-from cacheflow.models.utils import get_dtype_size
-
-_GiB = 1 << 30
-
-
-class CacheFlowMemoryAnalyzer:
-
-    def get_max_num_gpu_blocks(
-        self,
-        max_num_batched_tokens: int,
-        memory_utilization: float,
-    ) -> int:
-        raise NotImplementedError()
-
-    def get_workspace_size(self) -> int:
-        return 1 * _GiB
-
-    def get_cache_block_size(self) -> int:
-        raise NotImplementedError()
-
-    def get_max_num_cpu_blocks(
-        self,
-        swap_space: int,
-    ) -> int:
-        swap_space = swap_space * _GiB
-        cpu_memory = self.cpu_memory
-        if swap_space > 0.8 * cpu_memory:
-            raise ValueError(f'The swap space ({swap_space / _GiB:.2f} GiB) '
-                             'takes more than 80% of the available memory '
-                             f'({cpu_memory / _GiB:.2f} GiB).'
-                             'Please check the swap space size.')
-        if swap_space > 0.5 * cpu_memory:
-            print(f'WARNING: The swap space ({swap_space / _GiB:.2f} GiB) '
-                  'takes more than 50% of the available memory '
-                  f'({cpu_memory / _GiB:.2f} GiB).'
-                  'This may slow the system performance.')
-        max_num_blocks = swap_space // self.get_cache_block_size()
-        return max_num_blocks
-
-
-class OPTMemoryAnalyzer(CacheFlowMemoryAnalyzer):
-
-    def __init__(
-        self,
-        model_name: str,
-        block_size: int,
-        dtype: torch.dtype,
-        gpu_memory: int,
-        cpu_memory: int,
-        tensor_parallel_size: int,
-    ) -> None:
-        self.model_name = model_name
-        self.block_size = block_size
-        self.dtype = dtype
-        self.gpu_memory = gpu_memory
-        self.cpu_memory = cpu_memory
-        self.tensor_parallel_size = tensor_parallel_size
-
-        config = AutoConfig.from_pretrained(model_name)
-        self.num_layers = config.num_hidden_layers
-        self.hidden_size = config.hidden_size
-        self.num_heads = config.num_attention_heads
-        self.head_size = config.hidden_size // self.num_heads
-        self.ffn_size = config.ffn_dim
-        self.embedding_size = config.word_embed_proj_dim
-        self.vocab_size = config.vocab_size
-        self.max_position = config.max_position_embeddings
-
-    def _get_param_size(self) -> int:
-        word_embedding = self.vocab_size * self.embedding_size // self.tensor_parallel_size
-        if self.embedding_size != self.hidden_size:
-            # Project in/out.
-            word_embedding += 2 * self.embedding_size * self.hidden_size
-        position_embedding = self.max_position * self.hidden_size
-
-        ln1 = 2 * self.hidden_size
-        q = self.hidden_size * self.hidden_size // self.tensor_parallel_size + self.hidden_size
-        k = self.hidden_size * self.hidden_size // self.tensor_parallel_size + self.hidden_size
-        v = self.hidden_size * self.hidden_size // self.tensor_parallel_size + self.hidden_size
-        out = self.hidden_size * self.hidden_size // self.tensor_parallel_size + self.hidden_size
-        mha = ln1 + q + k + v + out
-
-        ln2 = 2 * self.hidden_size
-        ffn1 = self.hidden_size * self.ffn_size // self.tensor_parallel_size + self.ffn_size
-        ffn2 = self.ffn_size * self.hidden_size // self.tensor_parallel_size + self.hidden_size
-        ffn = ln2 + ffn1 + ffn2
-
-        total = (word_embedding + position_embedding +
-                 self.num_layers * (mha + ffn))
-        dtype_size = get_dtype_size(self.dtype)
-        return dtype_size * total
-
-    def _get_max_act_size(
-        self,
-        max_num_batched_tokens: int,
-    ) -> int:
-        # NOTE: We approxmiately calculate the maximum activation size by
-        # estimating
-        # 1) the maximum activation tensor size during inference
-        # 2) the residual tensor size during inference
-        # Here, we assume that FlashAttention is used and
-        # thus the attention maps are never materialized in GPU DRAM.
-        residual = max_num_batched_tokens * self.hidden_size
-        qkv = 3 * (max_num_batched_tokens * self.hidden_size) // self.tensor_parallel_size
-        ffn = max_num_batched_tokens * self.ffn_size // self.tensor_parallel_size
-        # Double the activation size for input and output.
-        max_act = 2 * (max(qkv, ffn) + residual)
-        # Size of output logits.
-        output_logits = 2 * (max_num_batched_tokens * self.vocab_size)
-        max_act = max(max_act, output_logits)
-        dtype_size = get_dtype_size(self.dtype)
-        return dtype_size * max_act
-
-    def get_cache_block_size(self) -> int:
-        key_cache_block = self.block_size * self.hidden_size // self.tensor_parallel_size
-        value_cache_block = key_cache_block
-        total = self.num_layers * (key_cache_block + value_cache_block)
-        dtype_size = get_dtype_size(self.dtype)
-        return dtype_size * total
-
-    def get_max_num_gpu_blocks(
-        self,
-        max_num_batched_tokens: int,
-        memory_utilization: float = 0.95,
-    ) -> int:
-        # NOTE(woosuk): This assumes that the machine has homogeneous GPUs.
-        usable_memory = int(memory_utilization * self.gpu_memory)
-
-        param_size = self._get_param_size()
-        act_size = self._get_max_act_size(max_num_batched_tokens)
-        workspace_size = self.get_workspace_size()
-
-        max_cache_size = usable_memory - (param_size + act_size + workspace_size)
-        if max_cache_size <= 0:
-            raise RuntimeError('Not enough GPU memory.')
-        max_num_blocks = max_cache_size // self.get_cache_block_size()
-        return max_num_blocks
-
-
-class LlamaMemoryAnalyzer(CacheFlowMemoryAnalyzer):
-
-    def __init__(
-        self,
-        model_name: str,
-        block_size: int,
-        dtype: torch.dtype,
-        gpu_memory: int,
-        cpu_memory: int,
-        tensor_parallel_size: int,
-    ) -> None:
-        self.model_name = model_name
-        self.block_size = block_size
-        self.dtype = dtype
-        self.gpu_memory = gpu_memory
-        self.cpu_memory = cpu_memory
-        self.tensor_parallel_size = tensor_parallel_size
-
-        config = AutoConfig.from_pretrained(model_name)
-        self.num_layers = config.num_hidden_layers
-        self.hidden_size = config.hidden_size
-        self.num_heads = config.num_attention_heads
-        self.head_size = config.hidden_size // self.num_heads
-        self.ffn_size = config.intermediate_size
-        self.vocab_size = config.vocab_size
-        self.max_position = 8192
-
-    def _get_param_size(self) -> int:
-        word_embedding = self.vocab_size * self.hidden_size // self.tensor_parallel_size
-        position_embedding = self.max_position * self.hidden_size
-
-        # NOTE: LLaMA does not have bias terms.
-        ln1 = self.hidden_size
-        q = self.hidden_size * self.hidden_size // self.tensor_parallel_size
-        k = self.hidden_size * self.hidden_size // self.tensor_parallel_size
-        v = self.hidden_size * self.hidden_size // self.tensor_parallel_size
-        out = self.hidden_size * self.hidden_size // self.tensor_parallel_size
-        # Rotary embedding.
-        # TODO(woosuk): Share the rotary embedding between layers.
-        rot = self.max_position * self.head_size
-        mha = ln1 + q + k + v + out + rot
-
-        ln2 = self.hidden_size
-        gate = self.hidden_size * self.ffn_size // self.tensor_parallel_size
-        down = self.ffn_size * self.hidden_size // self.tensor_parallel_size
-        up = self.hidden_size * self.ffn_size // self.tensor_parallel_size
-        ffn = ln2 + gate + down + up
-
-        total = (word_embedding + position_embedding + self.num_layers * (mha + ffn))
-        dtype_size = get_dtype_size(self.dtype)
-        return dtype_size * total
-
-    def _get_max_act_size(
-        self,
-        max_num_batched_tokens: int,
-    ) -> int:
-        # NOTE: We approxmiately calculate the maximum activation size by
-        # estimating
-        # 1) the maximum activation tensor size during inference
-        # 2) the residual tensor size during inference
-        # Here, we assume that FlashAttention is used and
-        # thus the attention maps are never materialized in GPU DRAM.
-        residual = max_num_batched_tokens * self.hidden_size
-        qkv = 3 * (max_num_batched_tokens * self.hidden_size) // self.tensor_parallel_size
-        ffn = 2 * (max_num_batched_tokens * self.ffn_size) // self.tensor_parallel_size
-        # Double the activation size for input and output.
-        max_act = 2 * (max(qkv, ffn) + residual)
-        # Size of output logits.
-        output_logits = 2 * (max_num_batched_tokens * self.vocab_size)
-        max_act = max(max_act, output_logits)
-        dtype_size = get_dtype_size(self.dtype)
-        return dtype_size * max_act
-
-    def get_cache_block_size(self) -> int:
-        key_cache_block = self.block_size * self.hidden_size // self.tensor_parallel_size
-        value_cache_block = key_cache_block
-        total = self.num_layers * (key_cache_block + value_cache_block)
-        dtype_size = get_dtype_size(self.dtype)
-        return dtype_size * total
-
-    def get_max_num_gpu_blocks(
-        self,
-        max_num_batched_tokens: int,
-        memory_utilization: float = 0.95,
-    ) -> int:
-        # NOTE(woosuk): This assumes that the machine has homogeneous GPUs.
-        gpu_memory = self.gpu_memory
-        usable_memory = int(memory_utilization * gpu_memory)
-
-        param_size = self._get_param_size()
-        act_size = self._get_max_act_size(max_num_batched_tokens)
-        workspace_size = self.get_workspace_size()
-
-        max_cache_size = usable_memory - (param_size + act_size + workspace_size)
-        if max_cache_size <= 0:
-            raise RuntimeError('Not enough GPU memory.')
-        max_num_blocks = max_cache_size // self.get_cache_block_size()
-        return max_num_blocks

cacheflow/models/model_utils.py

@@ -1,72 +0,0 @@
-from typing import Union
-
-import numpy as np
-import torch
-import torch.nn as nn
-from transformers import AutoConfig
-
-from cacheflow.models.memory_analyzer import CacheFlowMemoryAnalyzer
-from cacheflow.models.memory_analyzer import LlamaMemoryAnalyzer
-from cacheflow.models.memory_analyzer import OPTMemoryAnalyzer
-from cacheflow.models.llama import LlamaForCausalLM
-from cacheflow.models.opt import OPTForCausalLM
-from cacheflow.models.utils import get_torch_dtype
-
-
-_MODELS = {
-    'llama': LlamaForCausalLM,
-    'opt': OPTForCausalLM,
-}
-
-_MEMORY_ANALYZERS = {
-    'llama': LlamaMemoryAnalyzer,
-    'opt': OPTMemoryAnalyzer,
-}
-
-
-def get_model(
-    model_name: str,
-    dtype: Union[torch.dtype, str],
-    path: str,
-    use_dummy_weights: bool,
-) -> nn.Module:
-    torch_dtype = get_torch_dtype(dtype)
-    torch.set_default_dtype(torch_dtype)
-    config = AutoConfig.from_pretrained(model_name)
-    for model_class_name, model_class in _MODELS.items():
-        if model_class_name in model_name:
-            if use_dummy_weights:
-                # Create a model instance.
-                # The weights will be initialized as empty tensors.
-                model = model_class(config)
-                model = model.cuda()
-                # NOTE(woosuk): For precise performance evaluation, we assign
-                # random values to the weights. 
-                model.initialize_dummy_weights()
-            else:
-                # Download model weights if it's not cached.
-                weights_dir = model_class.get_weights(model_name, path=path)
-                # Create a model instance.
-                model = model_class(config)
-                # Load the weights from the cached or downloaded files.
-                model.load_weights(weights_dir)
-                model = model.cuda()
-            return model.eval(), torch_dtype
-    raise ValueError(f'Unsupported model name: {model_name}')
-
-
-def get_memory_analyzer(
-    model_name: str,
-    block_size: int,
-    dtype: Union[torch.dtype, str],
-    gpu_memory: int,
-    cpu_memory: int,
-    tensor_parallel_size: int = 1,
-) -> CacheFlowMemoryAnalyzer:
-    torch_dtype = get_torch_dtype(dtype)
-    for model_class, memory_analyzer in _MEMORY_ANALYZERS.items():
-        if model_class in model_name:
-            return memory_analyzer(
-                model_name, block_size, torch_dtype, gpu_memory, cpu_memory,
-                tensor_parallel_size)
-    raise ValueError(f'Unsupported model name: {model_name}')

cacheflow/models/opt.py

@@ -1,330 +0,0 @@
-"""1D OPT model compatible with HuggingFace weights."""
-import os
-import glob
-import filelock
-from tqdm import tqdm
-from typing import Dict, List, Optional, Tuple
-
-import numpy as np
-import torch
-from torch import nn
-from transformers import OPTConfig
-from huggingface_hub import snapshot_download
-
-from cacheflow.models import InputMetadata
-from cacheflow.models.attention import OPTCacheFlowAttention
-from cacheflow.models.sample import Sampler
-from cacheflow.parallel_utils.parallel_state import (
-    get_tensor_model_parallel_rank, get_tensor_model_parallel_world_size)
-from cacheflow.parallel_utils.tensor_parallel import (VocabParallelEmbedding,
-                                                      ColumnParallelLinear,
-                                                      RowParallelLinear)
-from cacheflow.sequence import SequenceOutputs
-
-KVCache = Tuple[torch.Tensor, torch.Tensor]
-
-
-class OPTLearnedPositionalEmbedding(nn.Embedding):
-
-    def __init__(self, num_embeddings: int, embedding_dim: int):
-        # OPT is set up so that if padding_idx is specified then offset the embedding ids by 2
-        # and adjust num_embeddings appropriately. Other models don't have this hack
-        self.offset = 2
-        super().__init__(num_embeddings + self.offset, embedding_dim)
-
-    def forward(self, positions: torch.LongTensor):
-        return super().forward(positions + self.offset)
-
-
-class OPTAttention(nn.Module):
-
-    def __init__(
-        self,
-        embed_dim: int,
-        num_heads: int,
-        bias: bool = True,
-    ) -> None:
-        super().__init__()
-        self.embed_dim = embed_dim
-        tensor_model_parallel_world_size = get_tensor_model_parallel_world_size()
-        total_num_heads = num_heads
-        assert num_heads % tensor_model_parallel_world_size == 0
-        self.num_heads = total_num_heads // tensor_model_parallel_world_size
-        self.head_dim = embed_dim // total_num_heads
-        self.scaling = self.head_dim ** -0.5
-
-        self.qkv_proj = ColumnParallelLinear(embed_dim, 3 * embed_dim, bias=bias,
-                                             gather_output=False,
-                                             perform_initialization=False)
-        self.out_proj = RowParallelLinear(embed_dim, embed_dim, bias=bias,
-                                          input_is_parallel=True,
-                                          perform_initialization=False)
-        self.attn = OPTCacheFlowAttention(scale=self.scaling)
-
-    def forward(
-        self,
-        hidden_states: torch.Tensor,
-        kv_cache: KVCache,
-        input_metadata: InputMetadata,
-        cache_event: Optional[torch.cuda.Event],
-    ) -> torch.Tensor:
-        qkv, _ = self.qkv_proj(hidden_states)
-        q, k, v = qkv.chunk(chunks=3, dim=-1)
-        key_cache, value_cache = kv_cache
-        attn_output = self.attn(
-            q, k, v, key_cache, value_cache, input_metadata, cache_event)
-        output, _ = self.out_proj(attn_output)
-        return output
-
-
-class OPTDecoderLayer(nn.Module):
-
-    def __init__(self, config: OPTConfig):
-        super().__init__()
-        self.config = config
-        self.embed_dim = config.hidden_size
-        self.self_attn = OPTAttention(
-            embed_dim=self.embed_dim,
-            num_heads=config.num_attention_heads,
-            bias=config.enable_bias,
-        )
-        self.do_layer_norm_before = config.do_layer_norm_before
-        assert config.activation_function == 'relu'
-        self.activation_fn = nn.ReLU()
-
-        self.self_attn_layer_norm = nn.LayerNorm(
-            self.embed_dim, elementwise_affine=config.layer_norm_elementwise_affine)
-        self.fc1 = ColumnParallelLinear(self.embed_dim, config.ffn_dim,
-                                        bias=config.enable_bias,
-                                        gather_output=False,
-                                        perform_initialization=False)
-        self.fc2 = RowParallelLinear(config.ffn_dim, self.embed_dim,
-                                     bias=config.enable_bias,
-                                     input_is_parallel=True,
-                                     perform_initialization=False)
-        self.final_layer_norm = nn.LayerNorm(
-            self.embed_dim, elementwise_affine=config.layer_norm_elementwise_affine)
-
-    def forward(
-        self,
-        hidden_states: torch.Tensor,
-        kv_cache: KVCache,
-        input_metadata: InputMetadata,
-        cache_event: Optional[torch.cuda.Event],
-    ) -> torch.Tensor:
-        # Self Attention
-        residual = hidden_states
-        # 125m, 1.7B, ..., 175B applies layer norm BEFORE attention
-        if self.do_layer_norm_before:
-            hidden_states = self.self_attn_layer_norm(hidden_states)
-        hidden_states = self.self_attn(
-            hidden_states=hidden_states,
-            kv_cache=kv_cache,
-            input_metadata=input_metadata,
-            cache_event=cache_event)
-        hidden_states = residual + hidden_states
-        # 350m applies layer norm AFTER attention
-        if not self.do_layer_norm_before:
-            hidden_states = self.self_attn_layer_norm(hidden_states)
-
-        # Fully Connected
-        residual = hidden_states
-        # 125m, 1.7B, ..., 175B applies layer norm BEFORE attention
-        if self.do_layer_norm_before:
-            hidden_states = self.final_layer_norm(hidden_states)
-        hidden_states, _ = self.fc1(hidden_states)
-        hidden_states = self.activation_fn(hidden_states)
-        hidden_states, _ = self.fc2(hidden_states)
-        hidden_states = residual + hidden_states
-        # 350m applies layer norm AFTER attention
-        if not self.do_layer_norm_before:
-            hidden_states = self.final_layer_norm(hidden_states)
-        return hidden_states
-
-
-class OPTDecoder(nn.Module):
-
-    def __init__(self, config: OPTConfig):
-        super().__init__()
-        self.config = config
-        self.padding_idx = config.pad_token_id
-        self.max_target_positions = config.max_position_embeddings
-        self.vocab_size = config.vocab_size
-
-        self.embed_tokens = VocabParallelEmbedding(config.vocab_size,
-                                                   config.word_embed_proj_dim,
-                                                   perform_initialization=False)
-        # Positional embeddings are replicated (not sharded).
-        self.embed_positions = OPTLearnedPositionalEmbedding(
-            config.max_position_embeddings, config.hidden_size)
-
-        # Project out & in will be replicated if they exist.
-        if config.word_embed_proj_dim != config.hidden_size:
-            self.project_out = nn.Linear(config.hidden_size, config.word_embed_proj_dim, bias=False)
-        else:
-            self.project_out = None
-
-        if config.word_embed_proj_dim != config.hidden_size:
-            self.project_in = nn.Linear(config.word_embed_proj_dim, config.hidden_size, bias=False)
-        else:
-            self.project_in = None
-
-        # Note that the only purpose of `config._remove_final_layer_norm` is to keep backward compatibility
-        # with checkpoints that have been fine-tuned before transformers v4.20.1
-        # see https://github.com/facebookresearch/metaseq/pull/164
-        if config.do_layer_norm_before and not config._remove_final_layer_norm:
-            self.final_layer_norm = nn.LayerNorm(
-                config.hidden_size, elementwise_affine=config.layer_norm_elementwise_affine
-            )
-        else:
-            self.final_layer_norm = None
-
-        self.layers = nn.ModuleList([OPTDecoderLayer(config) for _ in range(config.num_hidden_layers)])
-
-    def forward(
-        self,
-        input_ids: torch.LongTensor,
-        positions: torch.LongTensor,
-        kv_caches: List[KVCache],
-        input_metadata: InputMetadata,
-        cache_events: Optional[List[torch.cuda.Event]],
-    ) -> torch.Tensor:
-        inputs_embeds = self.embed_tokens(input_ids)
-        pos_embeds = self.embed_positions(positions)
-        if self.project_in is not None:
-            inputs_embeds = self.project_in(inputs_embeds)
-        hidden_states = inputs_embeds + pos_embeds
-
-        for i in range(len(self.layers)):
-            if cache_events is None:
-                cache_event = None
-            else:
-                cache_event = cache_events[i]
-            layer = self.layers[i]
-            hidden_states = layer(
-                hidden_states, kv_caches[i], input_metadata, cache_event)
-
-        if self.final_layer_norm is not None:
-            hidden_states = self.final_layer_norm(hidden_states)
-        if self.project_out is not None:
-            hidden_states = self.project_out(hidden_states)
-        return hidden_states
-
-
-class OPTModel(nn.Module):
-
-    def __init__(self, config: OPTConfig):
-        super().__init__()
-        self.decoder = OPTDecoder(config)
-
-    def forward(
-        self,
-        input_ids: torch.LongTensor,
-        positions: torch.LongTensor,
-        kv_caches: List[KVCache],
-        input_metadata: InputMetadata,
-        cache_events: Optional[List[torch.cuda.Event]],
-    ) -> torch.Tensor:
-        return self.decoder(
-            input_ids, positions, kv_caches, input_metadata, cache_events)
-
-
-class OPTForCausalLM(nn.Module):
-
-    def __init__(self, config):
-        super().__init__()
-        self.config = config
-        self.model = OPTModel(config)
-        # TODO(zhuohan): create a new weight after implementing pipeline
-        #                parallelism
-        self.lm_head_weight = self.model.decoder.embed_tokens.weight
-        self.sampler = Sampler()
-
-    def forward(
-        self,
-        input_ids: torch.LongTensor,
-        positions: torch.LongTensor,
-        kv_caches: List[KVCache],
-        input_metadata: InputMetadata,
-        cache_events: Optional[List[torch.cuda.Event]],
-    ) -> Dict[int, SequenceOutputs]:
-        hidden_states = self.model(
-            input_ids, positions, kv_caches, input_metadata, cache_events)
-        next_tokens = self.sampler(
-            self.lm_head_weight, hidden_states, input_metadata)
-        return next_tokens
-
-    _column_parallel_weights = ["embed_tokens.weight", "fc1.weight", "fc1.bias"]
-    _row_parallel_weights = ["out_proj.weight", "fc2.weight"]
-
-    def load_weights(self, weights_path: str):
-        tensor_model_parallel_rank = get_tensor_model_parallel_rank()
-        state_dict = self.state_dict()
-        for name, param in state_dict.items():
-            if "lm_head_weight" in name:
-                continue
-            if "qkv_proj" in name:
-                shard_size = param.shape[0] // 3
-                weights_to_concat = []
-                for weight_name in ["q_proj", "k_proj", "v_proj"]:
-                    weight = np.load(os.path.join(
-                        weights_path, name.replace("qkv_proj", weight_name)))
-                    weights_to_concat.append(weight[
-                        shard_size * tensor_model_parallel_rank
-                        :shard_size * (tensor_model_parallel_rank + 1)])
-                loaded_weight = torch.from_numpy(
-                    np.concatenate(weights_to_concat, axis=0))
-            else:
-                loaded_weight = torch.from_numpy(
-                    np.load(os.path.join(weights_path, name)))
-                for p in self._column_parallel_weights:
-                    if p in name:
-                        shard_size = param.shape[0]
-                        loaded_weight = loaded_weight[
-                            shard_size * tensor_model_parallel_rank
-                            :shard_size * (tensor_model_parallel_rank + 1)]
-                        break
-                for p in self._row_parallel_weights:
-                    if p in name:
-                        shard_size = param.shape[1]
-                        loaded_weight = loaded_weight[
-                            :,
-                            shard_size * tensor_model_parallel_rank
-                            :shard_size * (tensor_model_parallel_rank + 1)]
-                        break
-
-            assert param.shape == loaded_weight.shape
-            param.data.copy_(loaded_weight)
-
-    @staticmethod
-    def get_weights(model_name: str, path: str):
-        path = os.path.join(path, f"{model_name}-np")
-        path = os.path.abspath(os.path.expanduser(path))
-        os.makedirs(path, exist_ok=True)
-        lock_path = os.path.join(path, "file_lock")
-        lock = filelock.FileLock(lock_path)
-
-        with lock:
-            test_weight_path = os.path.join(
-                path, "model.decoder.embed_positions.weight")
-            if os.path.exists(test_weight_path):
-                return path
-
-            folder = snapshot_download(model_name, allow_patterns="*.bin",
-                                       cache_dir=os.path.join(path, "cache"))
-            bin_files = glob.glob(os.path.join(folder, "*.bin"))
-
-            for bin_file in tqdm(bin_files, desc="Convert format"):
-                state = torch.load(bin_file, map_location="cpu")
-                for name, param in tqdm(state.items(), leave=False):
-                    if name.startswith("decoder."):
-                        name = "model." + name
-                    param_path = os.path.join(path, name)
-                    with open(param_path, "wb") as f:
-                        np.save(f, param.cpu().detach().numpy())
-
-            return path
-
-    def initialize_dummy_weights(self) -> None:
-        for param in self.state_dict().values():
-            param.data.uniform_(-0.1, 0.1)

cacheflow/models/sample.py

@@ -1,287 +0,0 @@
-from typing import Dict, List, Tuple
-
-import torch
-import torch.nn as nn
-
-from cacheflow.models import InputMetadata
-from cacheflow.sampling_params import SamplingParams
-from cacheflow.sequence import SequenceOutputs
-from cacheflow.parallel_utils.tensor_parallel import gather_from_tensor_model_parallel_region
-
-
-class Sampler(nn.Module):
-
-    def __init__(self) -> None:
-        super().__init__()
-
-    def forward(
-        self,
-        embedding: torch.Tensor,
-        hidden_states: torch.Tensor,
-        input_metadata: InputMetadata,
-    ) -> Dict[int, SequenceOutputs]:
-        # Get the hidden states that we use for sampling.
-        hidden_states = _prune_hidden_states(hidden_states, input_metadata)
-
-        # Get the logits for the next tokens.
-        logits = torch.matmul(hidden_states, embedding.t())
-        logits = gather_from_tensor_model_parallel_region(logits)
-
-        # Apply temperature scaling.
-        temperatures = _get_temperatures(input_metadata)
-        assert len(temperatures) == logits.shape[0]
-        if any(t != 1.0 for t in temperatures):
-            t = torch.tensor(
-                temperatures, dtype=logits.dtype, device=logits.device)
-            # Use in-place division to avoid creating a new tensor.
-            logits.div_(t.unsqueeze(dim=1))
-
-        # We use float32 for probabilities and log probabilities.
-        # Compute the probabilities.
-        probs = torch.softmax(logits, dim=-1, dtype=torch.float)
-        # Compute the log probabilities (before applying top-p).
-        logprobs = torch.log(probs)
-
-        # Apply top-p truncation.
-        top_ps = _get_top_ps(input_metadata)
-        assert len(top_ps) == probs.shape[0]
-        if any(p < 1.0 for p in top_ps):
-            p = torch.tensor(top_ps, dtype=probs.dtype, device=probs.device)
-            probs = _apply_top_p(probs, p)
-
-        # Sample the next tokens.
-        return _sample(probs, logprobs, input_metadata)
-
-
-def _prune_hidden_states(
-    hidden_states: torch.Tensor,
-    input_metadata: InputMetadata,
-) -> torch.Tensor:
-    start_idx = 0
-    last_token_indicies: List[int] = []
-    for prompt_len in input_metadata.prompt_lens:
-        last_token_indicies.append(start_idx + prompt_len - 1)
-        start_idx += prompt_len
-    last_token_indicies.extend(
-        range(start_idx, start_idx + input_metadata.num_generation_tokens))
-    return hidden_states[last_token_indicies]
-
-
-def _get_temperatures(
-    input_metadata: InputMetadata,
-) -> List[float]:
-    # Collect the temperatures for the logits.
-    temperatures: List[float] = []
-    for i, seq_group in enumerate(input_metadata.seq_groups):
-        seq_ids, sampling_params = seq_group
-        temperature = sampling_params.temperature
-        if temperature == 0.0:
-            # NOTE: Zero temperature means deterministic sampling
-            # (i.e., greedy sampling or beam search).
-            # Set the temperature to 1 to avoid division by zero.
-            temperature = 1.0
-
-        if i < input_metadata.num_prompts:
-            # A prompt input.
-            temperatures.append(temperature)
-        else:
-            # A generation token.
-            temperatures += [temperature] * len(seq_ids)
-    return temperatures
-
-
-def _get_top_ps(
-    input_metadata: InputMetadata,
-) -> List[float]:
-    top_ps: List[float] = []
-    for i, seq_group in enumerate(input_metadata.seq_groups):
-        seq_ids, sampling_params = seq_group
-        if i < input_metadata.num_prompts:
-            # A prompt input.
-            top_ps.append(sampling_params.top_p)
-        else:
-            # A generation token.
-            top_ps += [sampling_params.top_p] * len(seq_ids)
-    return top_ps
-
-
-def _apply_top_p(
-    probs: torch.Tensor,
-    p: torch.Tensor,
-) -> torch.Tensor:
-    # TODO(woosuk): Optimize.
-    probs_sort, probs_idx = probs.sort(dim=-1, descending=True)
-    probs_sum = torch.cumsum(probs_sort, dim=-1)
-    mask = (probs_sum - probs_sort) > p.unsqueeze(dim=1)
-    probs_sort[mask] = 0.0
-    probs_sort.div_(probs_sort.sum(dim=-1, keepdim=True))
-    probs = torch.gather(
-        probs_sort, dim=-1, index=torch.argsort(probs_idx, dim=-1))
-    return probs
-
-
-def _get_topk_logprobs(
-    logprobs: torch.Tensor,
-    num_logprobs: int,
-) -> Dict[int, float]:
-    if num_logprobs == 0:
-        return {}
-
-    topk_logprobs, topk_ids = torch.topk(logprobs, num_logprobs)
-    if num_logprobs == 1:
-        topk_logprobs = [topk_logprobs.item()]
-        topk_ids = [topk_ids.item()]
-    else:
-        topk_logprobs = topk_logprobs.tolist()
-        topk_ids = topk_ids.tolist()
-
-    token_to_logprob: Dict[int, float] = {}
-    for token_id, logprob in zip(topk_ids, topk_logprobs):
-        token_to_logprob[token_id] = logprob
-    return token_to_logprob
-
-
-def _sample_from_prompt(
-    prob: torch.Tensor,
-    sampling_params: SamplingParams,
-) -> List[int]:
-    if sampling_params.use_beam_search:
-        # Beam search.
-        beam_width = sampling_params.n
-        _, next_token_ids = torch.topk(prob, beam_width)
-        next_token_ids = next_token_ids.tolist()
-    elif sampling_params.temperature == 0.0:
-        # Greedy sampling.
-        assert sampling_params.n == 1
-        next_token_id = torch.argmax(prob)
-        next_token_ids = [next_token_id.item()]
-    else:
-        # Neucleus sampling.
-        # Sample n tokens for the prompt.
-        n = sampling_params.n
-        next_token_ids = torch.multinomial(
-            prob, num_samples=n, replacement=True)
-        next_token_ids = next_token_ids.tolist()
-    return next_token_ids
-
-
-def _sample_from_generation_tokens(
-    seq_ids: List[int],
-    probs: torch.Tensor,
-    logprobs: torch.Tensor,
-    seq_logprobs: List[float],
-    sampling_params: SamplingParams,
-) -> Tuple[List[int], List[int]]:
-    # NOTE(woosuk): sampling_params.n can be greater than
-    # len(seq_ids) because some sequences in the group might have
-    # been already terminated.
-    if sampling_params.use_beam_search:
-        # Beam search.
-        # Add cumulative logprobs for the sequences in the group.
-        seq_logprobs = torch.tensor(
-            seq_logprobs, dtype=torch.float, device=logprobs.device)
-        logprobs = logprobs + seq_logprobs.unsqueeze(dim=1)
-
-        vocab_size = logprobs.size(-1)
-        beam_width = len(seq_ids)
-        _, topk_ids = torch.topk(logprobs.flatten(), beam_width)
-        topk_ids = topk_ids.tolist()
-        seq_idx = [i // vocab_size for i in topk_ids]
-        beam_seq_ids = [seq_ids[i] for i in seq_idx]
-        token_ids = [i % vocab_size for i in topk_ids]
-
-        beam_outputs: Dict[int, Tuple[int, int]] = {}
-        outstanding_beams: List[Tuple[int, int]] = []
-        # If a beam survives, continue with it.
-        for seq_id, token_id in zip(beam_seq_ids, token_ids):
-            if seq_id not in beam_outputs:
-                beam_outputs[seq_id] = (seq_id, token_id)
-            else:
-                outstanding_beams.append((seq_id, token_id))
-
-        # If a beam is discarded, fork another beam.
-        for seq_id in seq_ids:
-            if seq_id not in beam_outputs:
-                beam_outputs[seq_id] = outstanding_beams.pop()
-        assert not outstanding_beams
-
-        parent_seq_ids = [beam_outputs[seq_id][0] for seq_id in seq_ids]
-        next_token_ids = [beam_outputs[seq_id][1] for seq_id in seq_ids]
-    elif sampling_params.temperature == 0.0:
-        # Greedy sampling.
-        assert len(seq_ids) == 1
-        next_token_id = torch.argmax(probs, dim=-1)
-        next_token_ids = [next_token_id.item()]
-        parent_seq_ids = seq_ids
-    else:
-        # Neucleus sampling.
-        # Sample 1 token for each sequence in the group.
-        next_token_ids = torch.multinomial(
-            probs, num_samples=1, replacement=True)
-        next_token_ids = next_token_ids.squeeze(dim=-1).tolist()
-        parent_seq_ids = seq_ids
-    return parent_seq_ids, next_token_ids
-
-
-def _sample(
-    probs: torch.Tensor,
-    logprobs: torch.Tensor,
-    input_metadata: InputMetadata,
-) -> Dict[int, SequenceOutputs]:
-    seq_outputs: Dict[int, SequenceOutputs] = {}
-
-    # TODO(woosuk): Optimize.
-    idx = 0
-    for i, seq_group in enumerate(input_metadata.seq_groups):
-        seq_ids, sampling_params = seq_group
-        if i < input_metadata.num_prompts:
-            # Generate the next tokens for a prompt input.
-            assert len(seq_ids) == sampling_params.n
-            prob = probs[idx]
-            logprob = logprobs[idx]
-            idx += 1
-
-            # Sample the next tokens.
-            next_token_ids = _sample_from_prompt(prob, sampling_params)
-            # Get top-k log probabilities for the next tokens.
-            next_logprobs = _get_topk_logprobs(
-                logprob, sampling_params.num_logprobs)
-
-            # Build the output.
-            for seq_id, next_token_id in zip(seq_ids, next_token_ids):
-                output_logprobs = next_logprobs.copy()
-                output_logprobs[next_token_id] = logprob[next_token_id].item()
-                seq_outputs[seq_id] = SequenceOutputs(
-                    seq_id, seq_id, next_token_id, output_logprobs)
-        else:
-            # Generate the next tokens for generation tokens.
-            prob = probs[idx:idx + len(seq_ids)]
-            logprob = logprobs[idx:idx + len(seq_ids)]
-            idx += len(seq_ids)
-
-            # Sample the next tokens.
-            seq_logprobs = [
-                input_metadata.seq_logprobs[seq_id] for seq_id in seq_ids]
-            parent_seq_ids, next_token_ids = _sample_from_generation_tokens(
-                seq_ids, prob, logprob, seq_logprobs, sampling_params)
-
-            # Get top-k log probabilities for the next tokens.
-            next_logprobs: Dict[int, Dict[int, float]] = {}
-            for i, seq_id in enumerate(seq_ids):
-                next_logprobs[seq_id] = _get_topk_logprobs(
-                    logprob[i], sampling_params.num_logprobs)
-
-            # Build the output.
-            for seq_id, parent_seq_id, next_token_id in zip(
-                seq_ids, parent_seq_ids, next_token_ids):
-                i = seq_ids.index(parent_seq_id)
-                output_logprobs = next_logprobs[parent_seq_id].copy()
-                output_logprobs[next_token_id] = logprob[i, next_token_id].item()
-                seq_outputs[seq_id] = SequenceOutputs(
-                    seq_id,
-                    parent_seq_id,
-                    next_token_id,
-                    output_logprobs,
-                )
-
-    return seq_outputs

cacheflow/models/utils.py

@@ -1,24 +0,0 @@
-from typing import Union
-
-import torch
-
-_STR_DTYPE_TO_TORCH_DTYPE = {
-    'half': torch.half,
-    'float': torch.float,
-    'float16': torch.float16,
-    'float32': torch.float32,
-}
-
-
-def get_torch_dtype(dtype: Union[torch.dtype, str]) -> torch.dtype:
-    if isinstance(dtype, str):
-        torch_dtype = _STR_DTYPE_TO_TORCH_DTYPE[dtype.lower()]
-    else:
-        torch_dtype = dtype
-    return torch_dtype
-
-
-def get_dtype_size(dtype: Union[torch.dtype, str]) -> int:
-    torch_dtype = get_torch_dtype(dtype)
-    return torch.tensor([], dtype=torch_dtype).element_size()
-

cacheflow/parallel_utils/__init__.py

@@ -1,12 +0,0 @@
-import cacheflow.parallel_utils.parallel_state
-import cacheflow.parallel_utils.tensor_parallel
-import cacheflow.parallel_utils.utils
-
-# Alias parallel_state as mpu, its legacy name
-mpu = parallel_state
-
-__all__ = [
-    "parallel_state",
-    "tensor_parallel",
-    "utils",
-]

cacheflow/parallel_utils/parallel_state.py

@@ -1,593 +0,0 @@
-# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
-
-"""Model and data parallel groups."""
-
-import torch
-from typing import Optional
-
-from .utils import GlobalMemoryBuffer
-
-# Intra-layer model parallel group that the current rank belongs to.
-_TENSOR_MODEL_PARALLEL_GROUP = None
-# Inter-layer model parallel group that the current rank belongs to.
-_PIPELINE_MODEL_PARALLEL_GROUP = None
-# Model parallel group (both intra- and pipeline) that the current rank belongs to.
-_MODEL_PARALLEL_GROUP = None
-# Embedding group.
-_EMBEDDING_GROUP = None
-# Position embedding group.
-_POSITION_EMBEDDING_GROUP = None
-# Data parallel group that the current rank belongs to.
-_DATA_PARALLEL_GROUP = None
-
-_VIRTUAL_PIPELINE_MODEL_PARALLEL_RANK = None
-_VIRTUAL_PIPELINE_MODEL_PARALLEL_WORLD_SIZE = None
-_PIPELINE_MODEL_PARALLEL_SPLIT_RANK = None
-
-# These values enable us to change the mpu sizes on the fly.
-_MPU_TENSOR_MODEL_PARALLEL_WORLD_SIZE = None
-_MPU_PIPELINE_MODEL_PARALLEL_WORLD_SIZE = None
-_MPU_TENSOR_MODEL_PARALLEL_RANK = None
-_MPU_PIPELINE_MODEL_PARALLEL_RANK = None
-
-# A list of ranks that have a copy of the embedding.
-_EMBEDDING_GLOBAL_RANKS = None
-
-# A list of ranks that have a copy of the position embedding.
-_POSITION_EMBEDDING_GLOBAL_RANKS = None
-
-# A list of global ranks for each pipeline group to ease calculation of the source
-# rank when broadcasting from the first or last pipeline stage.
-_PIPELINE_GLOBAL_RANKS = None
-
-# A list of global ranks for each data parallel group to ease calculation of the source
-# rank when broadcasting weights from src to all other data parallel ranks
-_DATA_PARALLEL_GLOBAL_RANKS = None
-
-# Memory buffers to avoid dynamic memory allocation
-_GLOBAL_MEMORY_BUFFER = None
-
-_ALL_REDUCE_LAUNCHER: Optional['GraphAllReduce'] = None
-
-def initialize_model_parallel(
-    tensor_model_parallel_size: int = 1,
-    pipeline_model_parallel_size: int = 1,
-    virtual_pipeline_model_parallel_size: Optional[int] = None,
-    pipeline_model_parallel_split_rank: Optional[int] = None,
-) -> None:
-    """
-    Initialize model data parallel groups.
-
-    Arguments:
-        tensor_model_parallel_size: number of GPUs used for tensor model parallelism.
-        pipeline_model_parallel_size: number of GPUs used for pipeline model parallelism.
-        virtual_pipeline_model_parallel_size: number of virtual stages (interleaved
-                                              pipeline).
-        pipeline_model_parallel_split_rank: for models with both encoder and decoder,
-                                            rank in pipeline with split point.
-
-    Let's say we have a total of 16 GPUs denoted by g0 ... g15 and we
-    use 2 GPUs to parallelize the model tensor, and 4 GPUs to parallelize
-    the model pipeline. The present function will
-    create 8 tensor model-parallel groups, 4 pipeline model-parallel groups
-    and 8 data-parallel groups as:
-        8 data_parallel groups:
-            [g0, g2], [g1, g3], [g4, g6], [g5, g7], [g8, g10], [g9, g11], [g12, g14], [g13, g15]
-        8 tensor model-parallel groups:
-            [g0, g1], [g2, g3], [g4, g5], [g6, g7], [g8, g9], [g10, g11], [g12, g13], [g14, g15]
-        4 pipeline model-parallel groups:
-            [g0, g4, g8, g12], [g1, g5, g9, g13], [g2, g6, g10, g14], [g3, g7, g11, g15]
-    Note that for efficiency, the caller should make sure adjacent ranks
-    are on the same DGX box. For example if we are using 2 DGX-1 boxes
-    with a total of 16 GPUs, rank 0 to 7 belong to the first box and
-    ranks 8 to 15 belong to the second box.
-    """
-    # Get world size and rank. Ensure some consistencies.
-    assert torch.distributed.is_initialized()
-    world_size: int = torch.distributed.get_world_size()
-
-    if world_size % (tensor_model_parallel_size * pipeline_model_parallel_size) != 0:
-        raise RuntimeError(
-            f"world_size ({world_size}) is not divisible by tensor_model_parallel_size "
-            f"({tensor_model_parallel_size}) x pipeline_model_parallel_size ({pipeline_model_parallel_size})"
-        )
-
-    data_parallel_size: int = world_size // (tensor_model_parallel_size *
-                                             pipeline_model_parallel_size)
-
-    num_tensor_model_parallel_groups: int  = world_size // tensor_model_parallel_size
-    num_pipeline_model_parallel_groups: int = world_size // pipeline_model_parallel_size
-    num_data_parallel_groups: int = world_size // data_parallel_size
-
-    if virtual_pipeline_model_parallel_size is not None:
-        if not pipeline_model_parallel_size > 2:
-            raise RuntimeError("pipeline-model-parallel size should be greater than 2 with "
-                               "interleaved schedule")
-        global _VIRTUAL_PIPELINE_MODEL_PARALLEL_RANK
-        global _VIRTUAL_PIPELINE_MODEL_PARALLEL_WORLD_SIZE
-        _VIRTUAL_PIPELINE_MODEL_PARALLEL_RANK = 0
-        _VIRTUAL_PIPELINE_MODEL_PARALLEL_WORLD_SIZE = virtual_pipeline_model_parallel_size
-
-    if pipeline_model_parallel_split_rank is not None:
-        global _PIPELINE_MODEL_PARALLEL_SPLIT_RANK
-        _PIPELINE_MODEL_PARALLEL_SPLIT_RANK = pipeline_model_parallel_split_rank
-
-    rank = torch.distributed.get_rank()
-
-    # Build the data-parallel groups.
-    global _DATA_PARALLEL_GROUP
-    global _DATA_PARALLEL_GLOBAL_RANKS
-    assert _DATA_PARALLEL_GROUP is None, 'data parallel group is already initialized'
-    all_data_parallel_group_ranks = []
-    for i in range(pipeline_model_parallel_size):
-        start_rank = i * num_pipeline_model_parallel_groups
-        end_rank = (i + 1) * num_pipeline_model_parallel_groups
-        for j in range(tensor_model_parallel_size):
-            ranks = range(start_rank + j, end_rank, tensor_model_parallel_size)
-            all_data_parallel_group_ranks.append(list(ranks))
-            group = torch.distributed.new_group(ranks)
-            if rank in ranks:
-                _DATA_PARALLEL_GROUP = group
-                _DATA_PARALLEL_GLOBAL_RANKS = ranks
-
-    # Build the model-parallel groups.
-    global _MODEL_PARALLEL_GROUP
-    assert _MODEL_PARALLEL_GROUP is None, 'model parallel group is already initialized'
-    for i in range(data_parallel_size):
-        ranks = [data_parallel_group_ranks[i]
-                 for data_parallel_group_ranks in all_data_parallel_group_ranks]
-        group = torch.distributed.new_group(ranks)
-        if rank in ranks:
-            _MODEL_PARALLEL_GROUP = group
-
-    # Build the tensor model-parallel groups.
-    global _TENSOR_MODEL_PARALLEL_GROUP
-    assert _TENSOR_MODEL_PARALLEL_GROUP is None, \
-        'tensor model parallel group is already initialized'
-    for i in range(num_tensor_model_parallel_groups):
-        ranks = range(i * tensor_model_parallel_size,
-                      (i + 1) * tensor_model_parallel_size)
-        group = torch.distributed.new_group(ranks)
-        if rank in ranks:
-            _TENSOR_MODEL_PARALLEL_GROUP = group
-
-    # Build the pipeline model-parallel groups and embedding groups
-    # (first and last rank in each pipeline model-parallel group).
-    global _PIPELINE_MODEL_PARALLEL_GROUP
-    global _PIPELINE_GLOBAL_RANKS
-    assert _PIPELINE_MODEL_PARALLEL_GROUP is None, \
-        'pipeline model parallel group is already initialized'
-    global _EMBEDDING_GROUP
-    global _EMBEDDING_GLOBAL_RANKS
-    assert _EMBEDDING_GROUP is None, 'embedding group is already initialized'
-    global _POSITION_EMBEDDING_GROUP
-    global _POSITION_EMBEDDING_GLOBAL_RANKS
-    assert _POSITION_EMBEDDING_GROUP is None, \
-        'position embedding group is already initialized'
-    for i in range(num_pipeline_model_parallel_groups):
-        ranks = range(i, world_size, num_pipeline_model_parallel_groups)
-        group = torch.distributed.new_group(ranks)
-        if rank in ranks:
-            _PIPELINE_MODEL_PARALLEL_GROUP = group
-            _PIPELINE_GLOBAL_RANKS = ranks
-        # Setup embedding group (to exchange gradients between
-        # first and last stages).
-        if len(ranks) > 1:
-            embedding_ranks = [ranks[0], ranks[-1]]
-            position_embedding_ranks = [ranks[0]]
-            if pipeline_model_parallel_split_rank is not None:
-                if ranks[pipeline_model_parallel_split_rank] not in embedding_ranks:
-                    embedding_ranks = [ranks[0],
-                                       ranks[pipeline_model_parallel_split_rank],
-                                       ranks[-1]]
-                if ranks[pipeline_model_parallel_split_rank] not in position_embedding_ranks:
-                    position_embedding_ranks = [ranks[0],
-                                       ranks[pipeline_model_parallel_split_rank]]
-        else:
-            embedding_ranks = ranks
-            position_embedding_ranks = ranks
-
-        group = torch.distributed.new_group(embedding_ranks)
-        if rank in embedding_ranks:
-            _EMBEDDING_GROUP = group
-        if rank in ranks:
-            _EMBEDDING_GLOBAL_RANKS = embedding_ranks
-
-        group = torch.distributed.new_group(position_embedding_ranks)
-        if rank in position_embedding_ranks:
-            _POSITION_EMBEDDING_GROUP = group
-        if rank in ranks:
-            _POSITION_EMBEDDING_GLOBAL_RANKS = position_embedding_ranks
-
-    # Initialize global memory buffer
-    # This isn't really "parallel state" but there isn't another good place to
-    # put this. If we end up with a more generic initialization of megatron-core
-    # we could stick it there
-    _set_global_memory_buffer()
-
-
-def initialize_all_reduce_launcher(
-    max_num_tokens: int,
-    hidden_size: int,
-    dtype: torch.dtype,
-    disable_graph: bool = False,
-) -> None:
-    global _ALL_REDUCE_LAUNCHER
-    _ALL_REDUCE_LAUNCHER = GraphAllReduce(
-        max_num_tokens=max_num_tokens,
-        hidden_size=hidden_size,
-        dtype=dtype,
-        disable_graph=disable_graph,
-    )
-
-def model_parallel_is_initialized():
-    """Check if model and data parallel groups are initialized."""
-    if _TENSOR_MODEL_PARALLEL_GROUP is None or \
-        _PIPELINE_MODEL_PARALLEL_GROUP is None or \
-        _DATA_PARALLEL_GROUP is None:
-        return False
-    return True
-
-
-def get_model_parallel_group():
-    """Get the model parallel group the caller rank belongs to."""
-    assert _MODEL_PARALLEL_GROUP is not None, \
-        'model parallel group is not initialized'
-    return _MODEL_PARALLEL_GROUP
-
-
-def get_tensor_model_parallel_group():
-    """Get the tensor model parallel group the caller rank belongs to."""
-    assert _TENSOR_MODEL_PARALLEL_GROUP is not None, \
-        'intra_layer_model parallel group is not initialized'
-    return _TENSOR_MODEL_PARALLEL_GROUP
-
-
-def get_pipeline_model_parallel_group():
-    """Get the pipeline model parallel group the caller rank belongs to."""
-    assert _PIPELINE_MODEL_PARALLEL_GROUP is not None, \
-        'pipeline_model parallel group is not initialized'
-    return _PIPELINE_MODEL_PARALLEL_GROUP
-
-
-def get_data_parallel_group():
-    """Get the data parallel group the caller rank belongs to."""
-    assert _DATA_PARALLEL_GROUP is not None, \
-        'data parallel group is not initialized'
-    return _DATA_PARALLEL_GROUP
-
-
-def get_embedding_group():
-    """Get the embedding group the caller rank belongs to."""
-    assert _EMBEDDING_GROUP is not None, \
-        'embedding group is not initialized'
-    return _EMBEDDING_GROUP
-
-
-def get_position_embedding_group():
-    """Get the position embedding group the caller rank belongs to."""
-    assert _POSITION_EMBEDDING_GROUP is not None, \
-        'position embedding group is not initialized'
-    return _POSITION_EMBEDDING_GROUP
-
-
-def set_tensor_model_parallel_world_size(world_size):
-    """Set the tensor model parallel size"""
-    global _MPU_TENSOR_MODEL_PARALLEL_WORLD_SIZE
-    _MPU_TENSOR_MODEL_PARALLEL_WORLD_SIZE = world_size
-
-
-def set_pipeline_model_parallel_world_size(world_size):
-    """Set the pipeline model parallel size"""
-    global _MPU_PIPELINE_MODEL_PARALLEL_WORLD_SIZE
-    _MPU_PIPELINE_MODEL_PARALLEL_WORLD_SIZE = world_size
-
-
-def get_tensor_model_parallel_world_size():
-    """Return world size for the tensor model parallel group."""
-    global _MPU_TENSOR_MODEL_PARALLEL_WORLD_SIZE
-    if _MPU_TENSOR_MODEL_PARALLEL_WORLD_SIZE is not None:
-        return _MPU_TENSOR_MODEL_PARALLEL_WORLD_SIZE
-    return torch.distributed.get_world_size(group=get_tensor_model_parallel_group())
-
-
-def get_pipeline_model_parallel_world_size():
-    """Return world size for the pipeline model parallel group."""
-    global _MPU_PIPELINE_MODEL_PARALLEL_WORLD_SIZE
-    if _MPU_PIPELINE_MODEL_PARALLEL_WORLD_SIZE is not None:
-        return _MPU_PIPELINE_MODEL_PARALLEL_WORLD_SIZE
-    return torch.distributed.get_world_size(group=get_pipeline_model_parallel_group())
-
-
-def set_tensor_model_parallel_rank(rank):
-    """Set tensor model parallel rank."""
-    global _MPU_TENSOR_MODEL_PARALLEL_RANK
-    _MPU_TENSOR_MODEL_PARALLEL_RANK = rank
-
-
-def set_pipeline_model_parallel_rank(rank):
-    """Set pipeline model parallel rank."""
-    global _MPU_PIPELINE_MODEL_PARALLEL_RANK
-    _MPU_PIPELINE_MODEL_PARALLEL_RANK = rank
-
-
-def set_pipeline_model_parallel_split_rank(rank):
-    """Set pipeline model parallel split rank."""
-    global _MPU_PIPELINE_MODEL_PARALLEL_SPLIT_RANK
-    _MPU_PIPELINE_MODEL_PARALLEL_SPLIT_RANK = rank
-
-
-def get_tensor_model_parallel_rank():
-    """Return my rank for the tensor model parallel group."""
-    global _MPU_TENSOR_MODEL_PARALLEL_RANK
-    if _MPU_TENSOR_MODEL_PARALLEL_RANK is not None:
-        return _MPU_TENSOR_MODEL_PARALLEL_RANK
-    return torch.distributed.get_rank(group=get_tensor_model_parallel_group())
-
-
-def get_pipeline_model_parallel_rank():
-    """Return my rank for the pipeline model parallel group."""
-    global _MPU_PIPELINE_MODEL_PARALLEL_RANK
-    if _MPU_PIPELINE_MODEL_PARALLEL_RANK is not None:
-        return _MPU_PIPELINE_MODEL_PARALLEL_RANK
-    return torch.distributed.get_rank(group=get_pipeline_model_parallel_group())
-
-
-
-def is_pipeline_first_stage(ignore_virtual=False):
-    """Return True if in the first pipeline model-parallel stage, False otherwise."""
-    if not ignore_virtual:
-        if get_virtual_pipeline_model_parallel_world_size() is not None and \
-            get_virtual_pipeline_model_parallel_rank() != 0:
-            return False
-    return get_pipeline_model_parallel_rank() == 0
-
-
-def is_pipeline_last_stage(ignore_virtual=False):
-    """Return True if in the last pipeline model-parallel stage, False otherwise."""
-    if not ignore_virtual:
-        virtual_pipeline_model_parallel_world_size = \
-            get_virtual_pipeline_model_parallel_world_size()
-        if virtual_pipeline_model_parallel_world_size is not None and \
-            get_virtual_pipeline_model_parallel_rank() != (
-                virtual_pipeline_model_parallel_world_size - 1):
-            return False
-    return get_pipeline_model_parallel_rank() == (
-        get_pipeline_model_parallel_world_size() - 1)
-
-
-def is_rank_in_embedding_group(ignore_virtual=False):
-    """Return true if current rank is in embedding group, False otherwise."""
-    rank = torch.distributed.get_rank()
-    global _EMBEDDING_GLOBAL_RANKS
-    if ignore_virtual:
-        return rank in _EMBEDDING_GLOBAL_RANKS
-    if rank in _EMBEDDING_GLOBAL_RANKS:
-        if rank == _EMBEDDING_GLOBAL_RANKS[0]:
-            return is_pipeline_first_stage(ignore_virtual=False)
-        elif rank == _EMBEDDING_GLOBAL_RANKS[-1]:
-            return is_pipeline_last_stage(ignore_virtual=False)
-        else:
-            return True
-    return False
-
-
-def is_rank_in_position_embedding_group():
-    """Return true if current rank is in position embedding group, False otherwise."""
-    rank = torch.distributed.get_rank()
-    global _POSITION_EMBEDDING_GLOBAL_RANKS
-    return rank in _POSITION_EMBEDDING_GLOBAL_RANKS
-
-
-def is_pipeline_stage_before_split(rank=None):
-    """Return True if pipeline stage executes encoder block for a model
-    with both encoder and decoder."""
-    if get_pipeline_model_parallel_world_size() == 1:
-        return True
-    if rank is None:
-        rank = get_pipeline_model_parallel_rank()
-    global _PIPELINE_MODEL_PARALLEL_SPLIT_RANK
-    if _PIPELINE_MODEL_PARALLEL_SPLIT_RANK is None:
-        return True
-    if rank < _PIPELINE_MODEL_PARALLEL_SPLIT_RANK:
-        return True
-    return False
-
-
-def is_pipeline_stage_after_split(rank=None):
-    """Return True if pipeline stage executes decoder block for a model
-    with both encoder and decoder."""
-    if get_pipeline_model_parallel_world_size() == 1:
-        return True
-    if rank is None:
-        rank = get_pipeline_model_parallel_rank()
-    global _PIPELINE_MODEL_PARALLEL_SPLIT_RANK
-    if _PIPELINE_MODEL_PARALLEL_SPLIT_RANK is None:
-        return True
-    if rank >= _PIPELINE_MODEL_PARALLEL_SPLIT_RANK:
-        return True
-    return False
-
-
-def is_pipeline_stage_at_split():
-    """Return true if pipeline stage executes decoder block and next
-    stage executes encoder block for a model with both encoder and
-    decoder."""
-    rank = get_pipeline_model_parallel_rank()
-    return is_pipeline_stage_before_split(rank) and \
-            is_pipeline_stage_after_split(rank+1)
-
-
-def get_virtual_pipeline_model_parallel_rank():
-    """Return the virtual pipeline-parallel rank."""
-    global _VIRTUAL_PIPELINE_MODEL_PARALLEL_RANK
-    return _VIRTUAL_PIPELINE_MODEL_PARALLEL_RANK
-
-
-def set_virtual_pipeline_model_parallel_rank(rank):
-    """Set the virtual pipeline-parallel rank."""
-    global _VIRTUAL_PIPELINE_MODEL_PARALLEL_RANK
-    _VIRTUAL_PIPELINE_MODEL_PARALLEL_RANK = rank
-
-
-def get_virtual_pipeline_model_parallel_world_size():
-    """Return the virtual pipeline-parallel world size."""
-    global _VIRTUAL_PIPELINE_MODEL_PARALLEL_WORLD_SIZE
-    return _VIRTUAL_PIPELINE_MODEL_PARALLEL_WORLD_SIZE
-
-
-def get_tensor_model_parallel_src_rank():
-    """Calculate the global rank corresponding to the first local rank
-    in the tensor model parallel group."""
-    global_rank = torch.distributed.get_rank()
-    local_world_size = get_tensor_model_parallel_world_size()
-    return (global_rank // local_world_size) * local_world_size
-
-
-def get_data_parallel_src_rank():
-    """Calculate the global rank corresponding to the first local rank
-    in the data parallel group."""
-    assert _DATA_PARALLEL_GLOBAL_RANKS is not None, \
-        "Data parallel group is not initialized"
-    return _DATA_PARALLEL_GLOBAL_RANKS[0]
-
-
-def get_pipeline_model_parallel_first_rank():
-    """Return the global rank of the first process in the pipeline for the
-    current tensor parallel group"""
-    assert _PIPELINE_GLOBAL_RANKS is not None, \
-        "Pipeline parallel group is not initialized"
-    return _PIPELINE_GLOBAL_RANKS[0]
-
-
-def get_pipeline_model_parallel_last_rank():
-    """Return the global rank of the last process in the pipeline for the
-    current tensor parallel group"""
-    assert _PIPELINE_GLOBAL_RANKS is not None, \
-        "Pipeline parallel group is not initialized"
-    last_rank_local = get_pipeline_model_parallel_world_size() - 1
-    return _PIPELINE_GLOBAL_RANKS[last_rank_local]
-
-def get_pipeline_model_parallel_next_rank():
-    """Return the global rank that follows the caller in the pipeline"""
-    assert _PIPELINE_GLOBAL_RANKS is not None, \
-        "Pipeline parallel group is not initialized"
-    rank_in_pipeline = get_pipeline_model_parallel_rank()
-    world_size = get_pipeline_model_parallel_world_size()
-    return _PIPELINE_GLOBAL_RANKS[(rank_in_pipeline + 1) % world_size]
-
-
-def get_pipeline_model_parallel_prev_rank():
-    """Return the global rank that preceeds the caller in the pipeline"""
-    assert _PIPELINE_GLOBAL_RANKS is not None, \
-        "Pipeline parallel group is not initialized"
-    rank_in_pipeline = get_pipeline_model_parallel_rank()
-    world_size = get_pipeline_model_parallel_world_size()
-    return _PIPELINE_GLOBAL_RANKS[(rank_in_pipeline - 1) % world_size]
-
-
-def get_data_parallel_world_size():
-    """Return world size for the data parallel group."""
-    return torch.distributed.get_world_size(group=get_data_parallel_group())
-
-
-def get_data_parallel_rank():
-    """Return my rank for the data parallel group."""
-    return torch.distributed.get_rank(group=get_data_parallel_group())
-
-def _set_global_memory_buffer():
-    """Initialize global buffer"""
-    global _GLOBAL_MEMORY_BUFFER
-    assert _GLOBAL_MEMORY_BUFFER is None, 'global memory buffer is already initialized'
-    _GLOBAL_MEMORY_BUFFER = GlobalMemoryBuffer()
-
-def get_global_memory_buffer():
-    """Return the global GlobalMemoryBuffer object"""
-    assert _GLOBAL_MEMORY_BUFFER is not None, 'global memory buffer is not initialized'
-    return _GLOBAL_MEMORY_BUFFER
-
-def get_all_reduce_launcher() -> 'GraphAllReduce':
-    assert _ALL_REDUCE_LAUNCHER is not None, 'all reduce launcher is not initialized'
-    return _ALL_REDUCE_LAUNCHER
-
-def destroy_model_parallel():
-    """Set the groups to none."""
-    global _MODEL_PARALLEL_GROUP
-    _MODEL_PARALLEL_GROUP = None
-    global _TENSOR_MODEL_PARALLEL_GROUP
-    _TENSOR_MODEL_PARALLEL_GROUP = None
-    global _PIPELINE_MODEL_PARALLEL_GROUP
-    _PIPELINE_MODEL_PARALLEL_GROUP = None
-    global _DATA_PARALLEL_GROUP
-    _DATA_PARALLEL_GROUP = None
-    global _EMBEDDING_GROUP
-    _EMBEDDING_GROUP = None
-    global _POSITION_EMBEDDING_GROUP
-    _POSITION_EMBEDDING_GROUP = None
-    global _VIRTUAL_PIPELINE_MODEL_PARALLEL_RANK
-    _VIRTUAL_PIPELINE_MODEL_PARALLEL_RANK = None
-    global _VIRTUAL_PIPELINE_MODEL_PARALLEL_WORLD_SIZE
-    _VIRTUAL_PIPELINE_MODEL_PARALLEL_WORLD_SIZE = None
-    global _MPU_TENSOR_MODEL_PARALLEL_WORLD_SIZE
-    _MPU_TENSOR_MODEL_PARALLEL_WORLD_SIZE = None
-    global _MPU_PIPELINE_MODEL_PARALLEL_WORLD_SIZE
-    _MPU_PIPELINE_MODEL_PARALLEL_WORLD_SIZE = None
-    global _MPU_TENSOR_MODEL_PARALLEL_RANK
-    _MPU_TENSOR_MODEL_PARALLEL_RANK = None
-    global _MPU_PIPELINE_MODEL_PARALLEL_RANK
-    _MPU_PIPELINE_MODEL_PARALLEL_RANK = None
-    global _GLOBAL_MEMORY_BUFFER
-    _GLOBAL_MEMORY_BUFFER = None
-
-
-class GraphAllReduce:
-
-    def __init__(
-        self,
-        max_num_tokens: int,
-        hidden_size: int,
-        dtype: torch.dtype,
-        disable_graph: bool = False,
-    ) -> None:
-        self.max_num_tokens = max_num_tokens
-        self.hidden_size = hidden_size
-        self.disable_graph = disable_graph
-
-        tp_world_size = get_tensor_model_parallel_world_size()
-        if tp_world_size == 1:
-            return
-
-        self.group = get_tensor_model_parallel_group()
-        self.buffer = torch.empty(
-            size=(max_num_tokens, hidden_size),
-            dtype=dtype,
-            device='cuda',
-        )
-
-        # Build graphs for different number of tokens.
-        if not self.disable_graph:
-            self.graphs = {}
-            for num_tokens in range(8, max_num_tokens + 1, 8):
-                self.graphs[num_tokens] = self._build_graph(num_tokens)
-
-    def _build_graph(self, num_tokens: int) -> torch.cuda.CUDAGraph:
-        # Warm up.
-        torch.distributed.all_reduce(self.buffer[:num_tokens], group=self.group)
-        torch.cuda.synchronize()
-
-        # Build graph.
-        graph = torch.cuda.CUDAGraph()
-        with torch.cuda.graph(graph):
-            torch.distributed.all_reduce(
-                self.buffer[:num_tokens], group=self.group)
-        torch.cuda.synchronize()
-        return graph
-
-    def launch(self, x: torch.Tensor) -> torch.Tensor:
-        # NOTE: x must be a slice of self.buffer.
-        num_tokens = x.shape[0]
-        if self.disable_graph:
-            torch.distributed.all_reduce(x, group=self.group)
-        else:
-            self.graphs[num_tokens].replay()
-        return x

cacheflow/parallel_utils/tensor_parallel/__init__.py

@@ -1,55 +0,0 @@
-from .layers import (
-    ColumnParallelLinear,
-    RowParallelLinear,
-    VocabParallelEmbedding,
-    set_tensor_model_parallel_attributes,
-    set_defaults_if_not_set_tensor_model_parallel_attributes,
-    copy_tensor_model_parallel_attributes,
-    param_is_not_tensor_parallel_duplicate,
-)
-
-from .mappings import (
-    copy_to_tensor_model_parallel_region,
-    gather_from_tensor_model_parallel_region,
-    gather_from_sequence_parallel_region,
-    scatter_to_tensor_model_parallel_region,
-    scatter_to_sequence_parallel_region,
-)
-
-from .random import (
-    checkpoint,
-    get_cuda_rng_tracker,
-    model_parallel_cuda_manual_seed,
-)
-
-from .utils import (
-    split_tensor_along_last_dim,
-    split_tensor_into_1d_equal_chunks,
-    gather_split_1d_tensor,
-)
-
-__all__ = [
-    #layers.py
-    "ColumnParallelLinear",
-    "RowParallelLinear",
-    "VocabParallelEmbedding",
-    "set_tensor_model_parallel_attributes",
-    "set_defaults_if_not_set_tensor_model_parallel_attributes",
-    "copy_tensor_model_parallel_attributes",
-    "param_is_not_tensor_parallel_duplicate",
-    # mappings.py
-    "copy_to_tensor_model_parallel_region",
-    "gather_from_tensor_model_parallel_region",
-    "gather_from_sequence_parallel_region",
-#    "reduce_from_tensor_model_parallel_region",
-    "scatter_to_tensor_model_parallel_region",
-    "scatter_to_sequence_parallel_region",
-    # random.py
-    "checkpoint",
-    "get_cuda_rng_tracker",
-    "model_parallel_cuda_manual_seed",
-    # utils.py
-    "split_tensor_along_last_dim",
-    "split_tensor_into_1d_equal_chunks",
-    "gather_split_1d_tensor",
-]

cacheflow/parallel_utils/tensor_parallel/layers.py

@@ -1,446 +0,0 @@
-# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
-
-# Parts of the code here are adapted from PyTorch
-# repo: https://github.com/pytorch/pytorch
-
-
-import torch
-import torch.nn.functional as F
-import torch.nn.init as init
-from torch.nn.parameter import Parameter
-
-from cacheflow.parallel_utils.parallel_state import (
-    get_tensor_model_parallel_rank,
-    get_tensor_model_parallel_world_size,
-    get_all_reduce_launcher,
-)
-from .mappings import (
-    copy_to_tensor_model_parallel_region,
-    gather_from_tensor_model_parallel_region,
-    reduce_from_tensor_model_parallel_region,
-    scatter_to_tensor_model_parallel_region,
-)
-
-from .random import get_cuda_rng_tracker
-from .utils import (
-    divide,
-    VocabUtility,
-)
-
-_MODEL_PARALLEL_ATTRIBUTE_DEFAULTS = {'tensor_model_parallel': False,
-                                      'partition_dim': -1,
-                                      'partition_stride': 1}
-
-def param_is_not_tensor_parallel_duplicate(param):
-    return (hasattr(param, 'tensor_model_parallel') and
-            param.tensor_model_parallel) or (
-                get_tensor_model_parallel_rank() == 0)
-
-
-def set_tensor_model_parallel_attributes(tensor, is_parallel, dim, stride):
-    # Make sure the attributes are not set.
-    for attribute in _MODEL_PARALLEL_ATTRIBUTE_DEFAULTS:
-        assert not hasattr(tensor, attribute)
-    # Set the attributes.
-    setattr(tensor, 'tensor_model_parallel', is_parallel)
-    setattr(tensor, 'partition_dim', dim)
-    setattr(tensor, 'partition_stride', stride)
-
-
-def set_defaults_if_not_set_tensor_model_parallel_attributes(tensor):
-    def maybe_set(attribute, value):
-        if not hasattr(tensor, attribute):
-            setattr(tensor, attribute, value)
-    for attribute in _MODEL_PARALLEL_ATTRIBUTE_DEFAULTS:
-        maybe_set(attribute, _MODEL_PARALLEL_ATTRIBUTE_DEFAULTS[attribute])
-
-
-def copy_tensor_model_parallel_attributes(destination_tensor, source_tensor):
-    def maybe_copy(attribute):
-        if hasattr(source_tensor, attribute):
-            setattr(destination_tensor, attribute,
-                    getattr(source_tensor, attribute))
-    for attribute in _MODEL_PARALLEL_ATTRIBUTE_DEFAULTS:
-        maybe_copy(attribute)
-
-
-def _initialize_affine_weight_gpu(weight, init_method,
-                                  partition_dim, stride=1):
-    """Initialize affine weight for model parallel on GPU."""
-
-    set_tensor_model_parallel_attributes(tensor=weight,
-                                         is_parallel=True,
-                                         dim=partition_dim,
-                                         stride=stride)
-
-    with get_cuda_rng_tracker().fork():
-        init_method(weight)
-
-
-def _initialize_affine_weight_cpu(weight, output_size, input_size,
-                                  per_partition_size, partition_dim,
-                                  init_method, stride=1,
-                                  return_master_weight=False,
-                                  *, params_dtype=None):
-    """Initialize affine weight for model parallel.
-
-    Build the master weight on all processes and scatter
-    the relevant chunk."""
-
-    set_tensor_model_parallel_attributes(tensor=weight,
-                                         is_parallel=True,
-                                         dim=partition_dim,
-                                         stride=stride)
-
-    if params_dtype is None:
-        params_dtype = torch.get_default_dtype()
-
-    # Initialize master weight
-    master_weight = torch.empty(output_size, input_size,
-                                dtype=torch.float,
-                                requires_grad=False)
-    init_method(master_weight)
-    master_weight = master_weight.to(dtype=params_dtype)
-
-    # Split and copy
-    per_partition_per_stride_size = divide(per_partition_size, stride)
-    weight_list = torch.split(master_weight, per_partition_per_stride_size,
-                              dim=partition_dim)
-    rank = get_tensor_model_parallel_rank()
-    world_size = get_tensor_model_parallel_world_size()
-    my_weight_list = weight_list[rank::world_size]
-
-    with torch.no_grad():
-        torch.cat(my_weight_list, dim=partition_dim, out=weight)
-    if return_master_weight:
-        return master_weight
-    return None
-
-
-class VocabParallelEmbedding(torch.nn.Module):
-    """Embedding parallelized in the vocabulary dimension.
-
-    This is mainly adapted from torch.nn.Embedding and all the default
-    values are kept.
-    Arguments:
-        num_embeddings: vocabulary size.
-        embedding_dim: size of hidden state.
-
-    Keyword Arguments:
-        init_method: method to initialize weights.
-        params_dtype
-        use_cpu_initialization
-        perform_initialization
-    """
-
-    def __init__(self, num_embeddings: int, embedding_dim: int, *,
-                 init_method=init.xavier_normal_,
-                 params_dtype: torch.dtype=None,
-                 use_cpu_initialization: bool=False,
-                 perform_initialization: bool=True):
-        super(VocabParallelEmbedding, self).__init__()
-        # Keep the input dimensions.
-        self.num_embeddings = num_embeddings
-        self.embedding_dim = embedding_dim
-        if params_dtype is None:
-            params_dtype = torch.get_default_dtype()
-
-        # Set the defaults for compatibility.
-        self.padding_idx = None
-        self.max_norm = None
-        self.norm_type = 2.
-        self.scale_grad_by_freq = False
-        self.sparse = False
-        self._weight = None
-        self.tensor_model_parallel_size = get_tensor_model_parallel_world_size()
-        # Divide the weight matrix along the vocaburaly dimension.
-        self.vocab_start_index, self.vocab_end_index = \
-            VocabUtility.vocab_range_from_global_vocab_size(
-                self.num_embeddings, get_tensor_model_parallel_rank(),
-                self.tensor_model_parallel_size)
-        self.num_embeddings_per_partition = self.vocab_end_index - \
-            self.vocab_start_index
-
-        # Allocate weights and initialize.
-        if use_cpu_initialization:
-            self.weight = Parameter(torch.empty(
-                self.num_embeddings_per_partition, self.embedding_dim,
-                dtype=params_dtype))
-            if perform_initialization:
-                _initialize_affine_weight_cpu(
-                    self.weight, self.num_embeddings, self.embedding_dim,
-                    self.num_embeddings_per_partition, 0, init_method,
-                    params_dtype=params_dtype)
-        else:
-            self.weight = Parameter(torch.empty(
-                self.num_embeddings_per_partition, self.embedding_dim,
-                device=torch.cuda.current_device(), dtype=params_dtype))
-            if perform_initialization:
-                _initialize_affine_weight_gpu(self.weight, init_method,
-                                              partition_dim=0, stride=1)
-
-    def forward(self, input_):
-        if self.tensor_model_parallel_size > 1:
-            # Build the mask.
-            input_mask = (input_ < self.vocab_start_index) | \
-                         (input_ >= self.vocab_end_index)
-            # Mask the input.
-            masked_input = input_.clone() - self.vocab_start_index
-            masked_input[input_mask] = 0
-        else:
-            masked_input = input_
-            # Get the embeddings.
-        output_parallel = F.embedding(masked_input, self.weight,
-                                      self.padding_idx, self.max_norm,
-                                      self.norm_type, self.scale_grad_by_freq,
-                                      self.sparse)
-        # Mask the output embedding.
-        if self.tensor_model_parallel_size > 1:
-            output_parallel[input_mask, :] = 0.0
-        # Reduce across all the model parallel GPUs.
-        output = reduce_from_tensor_model_parallel_region(output_parallel)
-        return output
-
-
-class ColumnParallelLinear(torch.nn.Module):
-    """Linear layer with column parallelism.
-
-    The linear layer is defined as Y = XA + b. A is parallelized along
-    its second dimension as A = [A_1, ..., A_p].
-
-    Arguments:
-        input_size: first dimension of matrix A.
-        output_size: second dimension of matrix A.
-
-    Keyword Arguments
-        bias: If true, add bias
-        gather_output: If true, call all-gather on output and make Y available
-                       to all GPUs, otherwise, every GPU will have its output
-                       which is Y_i = XA_i
-        init_method: method to initialize weights. Note that bias is always set
-                     to zero.
-        stride: For the strided linear layers.
-        keep_master_weight_for_test: This was added for testing and should be
-                                     set to False. It returns the master weights
-                                     used for initialization.
-        skip_bias_add: This was added to enable performance optimations where bias
-                       can be fused with other elementwise operations. we skip
-                       adding bias but instead return it.
-        params_dtype:
-        use_cpu_initialization:
-    """
-
-    def __init__(self, input_size, output_size, *,
-                 bias=True, gather_output=True,
-                 init_method=init.xavier_normal_, stride=1,
-                 keep_master_weight_for_test=False,
-                 skip_bias_add=False,
-                 params_dtype=None,
-                 use_cpu_initialization=False,
-                 perform_initialization=True,
-                 ):
-        super(ColumnParallelLinear, self).__init__()
-
-        # Keep input parameters
-        self.input_size = input_size
-        self.output_size = output_size
-        self.gather_output = gather_output
-        # Divide the weight matrix along the last dimension.
-        world_size = get_tensor_model_parallel_world_size()
-        self.output_size_per_partition = divide(output_size, world_size)
-        self.skip_bias_add = skip_bias_add
-
-        if params_dtype is None:
-            params_dtype = torch.get_default_dtype()
-
-        # Parameters.
-        # Note: torch.nn.functional.linear performs XA^T + b and as a result
-        # we allocate the transpose.
-        # Initialize weight.
-        if use_cpu_initialization:
-            self.weight = Parameter(torch.empty(self.output_size_per_partition,
-                                                self.input_size,
-                                                dtype=params_dtype))
-            if perform_initialization:
-                self.master_weight = _initialize_affine_weight_cpu(
-                    self.weight, self.output_size, self.input_size,
-                    self.output_size_per_partition, 0, init_method,
-                    stride=stride, return_master_weight=keep_master_weight_for_test)
-        else:
-            self.weight = Parameter(torch.empty(
-                self.output_size_per_partition, self.input_size,
-                device=torch.cuda.current_device(), dtype=params_dtype))
-            if perform_initialization:
-                _initialize_affine_weight_gpu(self.weight, init_method,
-                                              partition_dim=0, stride=stride)
-
-        if bias:
-            if use_cpu_initialization:
-                self.bias = Parameter(torch.empty(
-                    self.output_size_per_partition, dtype=params_dtype))
-            else:
-                self.bias = Parameter(torch.empty(
-                    self.output_size_per_partition,
-                    device=torch.cuda.current_device(),
-                    dtype=params_dtype))
-            set_tensor_model_parallel_attributes(self.bias, True, 0, stride)
-            # Always initialize bias to zero.
-            with torch.no_grad():
-                self.bias.zero_()
-        else:
-            self.register_parameter('bias', None)
-
-
-    def forward(self, input_):
-        """Forward of ColumnParallelLinear
-
-        Args:
-            input_: 3D tensor whose order of dimension is [sequence, batch, hidden]
-
-        Returns:
-            - output
-            - bias
-        """
-        bias = self.bias if not self.skip_bias_add else None
-
-        input_parallel = copy_to_tensor_model_parallel_region(input_)
-        # Matrix multiply.
-        output_parallel = F.linear(input_parallel, self.weight, bias)
-        if self.gather_output:
-            # All-gather across the partitions.
-            output = gather_from_tensor_model_parallel_region(output_parallel)
-        else:
-            output = output_parallel
-        output_bias = self.bias if self.skip_bias_add else None
-        return output, output_bias
-
-
-class RowParallelLinear(torch.nn.Module):
-    """Linear layer with row parallelism.
-
-    The linear layer is defined as Y = XA + b. A is parallelized along
-    its first dimension and X along its second dimension as:
-               -   -
-              | A_1 |
-              | .   |
-          A = | .   |        X = [X_1, ..., X_p]
-              | .   |
-              | A_p |
-               -   -
-    Arguments:
-        input_size: first dimension of matrix A.
-        output_size: second dimension of matrix A.
-
-    Keyword Arguments:
-        bias: If true, add bias. Note that bias is not parallelized.
-        input_is_parallel: If true, we assume that the input is already
-                           split across the GPUs and we do not split
-                           again.
-        init_method: method to initialize weights. Note that bias is always set
-                     to zero.
-        stride: For the strided linear layers.
-        keep_master_weight_for_test: This was added for testing and should be
-                                     set to False. It returns the master weights
-                                     used for initialization.
-        skip_bias_add: This was added to enable performance optimization where bias
-                       can be fused with other elementwise operations. We skip
-                       adding bias but instead return it.
-        params_dtype:
-        use_cpu_initialization:
-        perform_initialization:
-    """
-
-    def __init__(self, input_size, output_size, *,
-                 bias=True, input_is_parallel=False,
-                 init_method=init.xavier_normal_, stride=1,
-                 keep_master_weight_for_test=False,
-                 skip_bias_add=False,
-                 params_dtype=None,
-                 use_cpu_initialization=False,
-                 perform_initialization=True,
-                 ):
-        super(RowParallelLinear, self).__init__()
-
-        # Keep input parameters
-        self.input_size = input_size
-        self.output_size = output_size
-        self.input_is_parallel = input_is_parallel
-        if params_dtype is None:
-            params_dtype = torch.get_default_dtype()
-
-        # Divide the weight matrix along the last dimension.
-        world_size = get_tensor_model_parallel_world_size()
-        self.input_size_per_partition = divide(input_size, world_size)
-        self.skip_bias_add = skip_bias_add
-
-        # Parameters.
-        # Note: torch.nn.functional.linear performs XA^T + b and as a result
-        # we allocate the transpose.
-        # Initialize weight.
-        if use_cpu_initialization:
-            self.weight = Parameter(torch.empty(self.output_size,
-                                                self.input_size_per_partition,
-                                                dtype=params_dtype))
-            if perform_initialization:
-                self.master_weight = _initialize_affine_weight_cpu(
-                    self.weight, self.output_size, self.input_size,
-                    self.input_size_per_partition, 1, init_method,
-                    stride=stride, return_master_weight=keep_master_weight_for_test,
-                    params_dtype=params_dtype)
-        else:
-            self.weight = Parameter(torch.empty(
-                self.output_size, self.input_size_per_partition,
-                device=torch.cuda.current_device(), dtype=params_dtype))
-            if perform_initialization:
-                _initialize_affine_weight_gpu(self.weight, init_method,
-                                              partition_dim=1, stride=stride)
-        if bias:
-            if use_cpu_initialization:
-                self.bias = Parameter(torch.empty(self.output_size,
-                                                  dtype=params_dtype))
-            else:
-                self.bias = Parameter(torch.empty(
-                    self.output_size, device=torch.cuda.current_device(),
-                    dtype=params_dtype))
-
-            # Always initialize bias to zero.
-            with torch.no_grad():
-                self.bias.zero_()
-        else:
-            self.register_parameter('bias', None)
-        self.weight_t = self.weight.t()
-
-    def forward(self, input_):
-        """Forward of RowParallelLinear
-
-        Args:
-            input_: 3D tensor whose order of dimension is [sequence, batch, hidden]
-
-        Returns:
-            - output
-            - bias
-        """
-        # Set up backprop all-reduce.
-        if self.input_is_parallel:
-            input_parallel = input_
-        else:
-            input_parallel = scatter_to_tensor_model_parallel_region(input_)
-        if get_tensor_model_parallel_world_size() == 1:
-            # Matrix multiply.
-            output_ = F.linear(input_parallel, self.weight)
-        else:
-            # Matrix multiply.
-            all_reduce_launcher = get_all_reduce_launcher()
-            num_tokens = input_parallel.shape[0]
-            output_buffer = all_reduce_launcher.buffer[:num_tokens]
-            torch.matmul(input_parallel, self.weight_t, out=output_buffer)
-            # All-reduce across all the partitions.
-            output_ = all_reduce_launcher.launch(output_buffer)
-
-        if not self.skip_bias_add:
-            output = output_ + self.bias if self.bias is not None else output_
-            output_bias = None
-        else:
-            output = output_
-            output_bias = self.bias
-        return output, output_bias

cacheflow/parallel_utils/tensor_parallel/mappings.py

@@ -1,279 +0,0 @@
-# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
-
-import torch
-
-from cacheflow.parallel_utils.parallel_state import (
-    get_tensor_model_parallel_rank,
-    get_tensor_model_parallel_world_size,
-    get_tensor_model_parallel_group,
-)
-from .utils import split_tensor_along_last_dim
-
-
-def _reduce(input_):
-    """All-reduce the input tensor across model parallel group."""
-
-    # Bypass the function if we are using only 1 GPU.
-    if get_tensor_model_parallel_world_size()==1:
-        return input_
-
-    # All-reduce.
-    torch.distributed.all_reduce(input_, group=get_tensor_model_parallel_group())
-
-    return input_
-
-
-def _split_along_last_dim(input_):
-    """Split the tensor along its last dimension and keep the
-    corresponding slice."""
-
-    world_size = get_tensor_model_parallel_world_size()
-    # Bypass the function if we are using only 1 GPU.
-    if world_size == 1:
-        return input_
-
-    # Split along last dimension.
-    input_list = split_tensor_along_last_dim(input_, world_size)
-
-    # Note: torch.split does not create contiguous tensors by default.
-    rank = get_tensor_model_parallel_rank()
-    output = input_list[rank].contiguous()
-
-    return output
-
-
-def _split_along_first_dim(input_):
-    """Split the tensor along its first dimension and keep the
-    corresponding slice."""
-
-    world_size = get_tensor_model_parallel_world_size()
-    # Bypass the function if we are using only 1 GPU.
-    if world_size == 1:
-        return input_
-
-    # Split along first dimension.
-    dim_size = input_.size()[0]
-    assert dim_size % world_size == 0, \
-        "First dimension of the tensor should be divisible by tensor parallel size"
-    local_dim_size = dim_size // world_size
-    rank = get_tensor_model_parallel_rank()
-    dim_offset = rank * local_dim_size
-
-    output = input_[dim_offset:dim_offset+local_dim_size].contiguous()
-
-    return output
-
-
-def _gather_along_last_dim(input_):
-    """Gather tensors and concatinate along the last dimension."""
-
-    world_size = get_tensor_model_parallel_world_size()
-    # Bypass the function if we are using only 1 GPU.
-    if world_size == 1:
-        return input_
-
-    # Size and dimension.
-    last_dim = input_.dim() - 1
-    rank = get_tensor_model_parallel_rank()
-
-    tensor_list = [torch.empty_like(input_) for _ in range(world_size)]
-    tensor_list[rank] = input_
-    torch.distributed.all_gather(tensor_list, input_, group=get_tensor_model_parallel_group())
-
-    # Note: torch.cat already creates a contiguous tensor.
-    output = torch.cat(tensor_list, dim=last_dim).contiguous()
-
-    return output
-
-
-def _gather_along_first_dim(input_):
-    """Gather tensors and concatinate along the first dimension."""
-
-    world_size = get_tensor_model_parallel_world_size()
-    # Bypass the function if we are using only 1 GPU.
-    if world_size == 1:
-        return input_
-
-    dim_size = list(input_.size())
-    dim_size[0] = dim_size[0] * world_size
-
-    output = torch.empty(dim_size, dtype=input_.dtype,
-                         device=torch.cuda.current_device())
-    torch.distributed._all_gather_base(output, input_.contiguous(),
-                                       group=get_tensor_model_parallel_group())
-
-    return output
-
-def _reduce_scatter_along_first_dim(input_):
-    """Reduce-scatter the input tensor across model parallel group."""
-    world_size = get_tensor_model_parallel_world_size()
-    # Bypass the function if we are using only 1 GPU.
-    if world_size == 1:
-        return input_
-
-    dim_size = list(input_.size())
-    assert dim_size[0] % world_size == 0, \
-        "First dimension of the tensor should be divisible by tensor parallel size"
-
-    dim_size[0] = dim_size[0] // world_size
-
-    output = torch.empty(dim_size, dtype=input_.dtype,
-                         device=torch.cuda.current_device())
-    torch.distributed._reduce_scatter_base(output, input_.contiguous(),
-                                           group=get_tensor_model_parallel_group())
-    return output
-
-
-class _CopyToModelParallelRegion(torch.autograd.Function):
-    """Pass the input to the model parallel region."""
-
-    @staticmethod
-    def symbolic(graph, input_):
-        return input_
-
-    @staticmethod
-    def forward(ctx, input_):
-        return input_
-
-    @staticmethod
-    def backward(ctx, grad_output):
-        return _reduce(grad_output)
-
-
-class _ReduceFromModelParallelRegion(torch.autograd.Function):
-    """All-reduce the input from the model parallel region."""
-
-    @staticmethod
-    def symbolic(graph, input_):
-        return _reduce(input_)
-
-    @staticmethod
-    def forward(ctx, input_):
-        return _reduce(input_)
-
-    @staticmethod
-    def backward(ctx, grad_output):
-        return grad_output
-
-
-class _ScatterToModelParallelRegion(torch.autograd.Function):
-    """Split the input and keep only the corresponding chuck to the rank."""
-
-    @staticmethod
-    def symbolic(graph, input_):
-        return _split_along_last_dim(input_)
-
-    @staticmethod
-    def forward(ctx, input_):
-        return _split_along_last_dim(input_)
-
-    @staticmethod
-    def backward(ctx, grad_output):
-        return _gather_along_last_dim(grad_output)
-
-
-class _GatherFromModelParallelRegion(torch.autograd.Function):
-    """Gather the input from model parallel region and concatinate."""
-
-    @staticmethod
-    def symbolic(graph, input_):
-        return _gather_along_last_dim(input_)
-
-    @staticmethod
-    def forward(ctx, input_):
-        return _gather_along_last_dim(input_)
-
-    @staticmethod
-    def backward(ctx, grad_output):
-        return _split_along_last_dim(grad_output)
-
-
-class _ScatterToSequenceParallelRegion(torch.autograd.Function):
-    """Split the input and keep only the corresponding chuck to the rank."""
-
-    @staticmethod
-    def symbolic(graph, input_):
-        return _split_along_first_dim(input_)
-
-    @staticmethod
-    def forward(ctx, input_):
-        return _split_along_first_dim(input_)
-
-    @staticmethod
-    def backward(ctx, grad_output):
-        return _gather_along_first_dim(grad_output)
-
-
-class _GatherFromSequenceParallelRegion(torch.autograd.Function):
-    """Gather the input from sequence parallel region and concatinate."""
-
-    @staticmethod
-    def symbolic(graph, input_, tensor_parallel_output_grad=True):
-        return _gather_along_first_dim(input_)
-
-    @staticmethod
-    def forward(ctx, input_, tensor_parallel_output_grad=True):
-        ctx.tensor_parallel_output_grad = tensor_parallel_output_grad
-        return _gather_along_first_dim(input_)
-
-    @staticmethod
-    def backward(ctx, grad_output):
-        tensor_parallel_output_grad = ctx.tensor_parallel_output_grad
-
-        # If the computation graph after the gather operation is
-        # in the tensor parallel mode, output gradients need to reduce
-        # scattered and whereas if the computation is duplicated,
-        # output gradients need to be scattered.
-        if tensor_parallel_output_grad:
-            return _reduce_scatter_along_first_dim(grad_output), None
-        else:
-            return _split_along_first_dim(grad_output), None
-
-
-class _ReduceScatterToSequenceParallelRegion(torch.autograd.Function):
-    """Reduce scatter the input from the model parallel region."""
-
-    @staticmethod
-    def symbolic(graph, input_):
-        return _reduce_scatter_along_first_dim(input_)
-
-    @staticmethod
-    def forward(ctx, input_):
-        return _reduce_scatter_along_first_dim(input_)
-
-    @staticmethod
-    def backward(ctx, grad_output):
-        return _gather_along_first_dim(grad_output)
-
-
-# -----------------
-# Helper functions.
-# -----------------
-
-def copy_to_tensor_model_parallel_region(input_):
-    return _CopyToModelParallelRegion.apply(input_)
-
-
-def reduce_from_tensor_model_parallel_region(input_):
-    return _ReduceFromModelParallelRegion.apply(input_)
-
-
-def scatter_to_tensor_model_parallel_region(input_):
-    return _ScatterToModelParallelRegion.apply(input_)
-
-
-def gather_from_tensor_model_parallel_region(input_):
-    return _GatherFromModelParallelRegion.apply(input_)
-
-
-def scatter_to_sequence_parallel_region(input_):
-    return _ScatterToSequenceParallelRegion.apply(input_)
-
-
-def gather_from_sequence_parallel_region(input_, tensor_parallel_output_grad=True):
-    return _GatherFromSequenceParallelRegion.apply(input_, tensor_parallel_output_grad)
-
-
-def reduce_scatter_to_sequence_parallel_region(input_):
-    return _ReduceScatterToSequenceParallelRegion.apply(input_)
-

cacheflow/parallel_utils/tensor_parallel/random.py

@@ -1,253 +0,0 @@
-# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
-
-# Parts of the code here are adapted from PyTorch
-# repo: https://github.com/pytorch/pytorch
-
-import contextlib
-
-import torch
-from torch import _C
-from torch.cuda import _lazy_call, device as device_ctx_manager
-from torch.utils.checkpoint import detach_variable
-
-from cacheflow.parallel_utils.parallel_state import (
-    get_data_parallel_rank,
-    get_tensor_model_parallel_group,
-    get_tensor_model_parallel_rank,
-    get_tensor_model_parallel_world_size,
-)
-
-from .utils import (
-    split_tensor_into_1d_equal_chunks,
-    gather_split_1d_tensor,
-)
-
-from cacheflow.parallel_utils.utils import safely_set_viewless_tensor_data
-
-# Default name for the model parallel rng tracker.
-_MODEL_PARALLEL_RNG_TRACKER_NAME = 'model-parallel-rng'
-
-
-def _set_cuda_rng_state(new_state, device=-1):
-    """Sets the random number generator state of the current GPU.
-
-    Argumentss:
-        new_state (torch.ByteTensor): The desired state
-    This function is adapted from PyTorch repo (torch.cuda.set_rng_state)
-    with a single change: the input state is not cloned. Cloning caused
-    major performance issues for +4 GPU cases.
-    """
-    if hasattr(_C, '_cuda_setRNGState') and callable(_C._cuda_setRNGState):
-        # older PyTorch
-        def cb():
-            with device_ctx_manager(device):
-                _C._cuda_setRNGState(new_state)
-    else:
-        # newer PyTorch
-        if device == -1:
-            device = torch.device('cuda')
-        elif isinstance(device, str):
-            device = torch.device(device)
-        elif isinstance(device, int):
-            device = torch.device('cuda', device)
-
-        def cb():
-            idx = device.index
-            if idx is None:
-                idx = torch.cuda.current_device()
-            default_generator = torch.cuda.default_generators[idx]
-            default_generator.set_state(new_state)
-
-    _lazy_call(cb)
-
-
-
-class CudaRNGStatesTracker:
-    """Tracker for the cuda RNG states.
-
-    Using the `add` method, a cuda rng state is initialized based on
-    the input `seed` and is assigned to `name`. Later, by forking the
-    rng state, we can perform operations and return to our starting
-    cuda state.
-    """
-
-    def __init__(self):
-        # Map from a string name to the cuda rng state.
-        self.states_ = {}
-        # Seeds are just for book keeping and ensure no seed is set twice.
-        self.seeds_ = set()
-
-    def reset(self):
-        """Set to the initial state (no tracker)."""
-        self.states_ = {}
-        self.seeds_ = set()
-
-    def get_states(self):
-        """Get rng states. Copy the dictionary so we have direct
-        pointers to the states, not just a pointer to the dictionary."""
-        states = {}
-        for name in self.states_:
-            states[name] = self.states_[name]
-        return states
-
-    def set_states(self, states):
-        """Set the rng states. For efficiency purposes, we do not check
-        the size of seed for compatibility."""
-        self.states_ = states
-
-    def add(self, name, seed):
-        """Track the rng state."""
-        # Check seed is not already used.
-        if seed in self.seeds_:
-            raise Exception('seed {} already exists'.format(seed))
-        self.seeds_.add(seed)
-        # Check that state is not already defined.
-        if name in self.states_:
-            raise Exception('cuda rng state {} already exists'.format(name))
-        # Get the current rng state.
-        orig_rng_state = torch.cuda.get_rng_state()
-        # Set the new state and store it.
-        torch.cuda.manual_seed(seed)
-        self.states_[name] = torch.cuda.get_rng_state()
-        # Reset rng state to what it was.
-        _set_cuda_rng_state(orig_rng_state)
-
-    @contextlib.contextmanager
-    def fork(self, name=_MODEL_PARALLEL_RNG_TRACKER_NAME):
-        """Fork the cuda rng state, perform operations, and exit with
-        the original state."""
-        # Check if we have added the state
-        if name not in self.states_:
-            raise Exception('cuda rng state {} is not added'.format(name))
-        # Store current rng state.
-        orig_cuda_rng_state = torch.cuda.get_rng_state()
-        # Set rng state to the desired one
-        _set_cuda_rng_state(self.states_[name])
-        # Do the stuff we wanted to do.
-        try:
-            yield
-        finally:
-            # Update the current rng state for later use.
-            self.states_[name] = torch.cuda.get_rng_state()
-            # And set the state to the original state we started with.
-            _set_cuda_rng_state(orig_cuda_rng_state)
-
-
-# RNG tracker object.
-_CUDA_RNG_STATE_TRACKER = CudaRNGStatesTracker()
-
-
-def get_cuda_rng_tracker():
-    """Get cuda rng tracker."""
-    return _CUDA_RNG_STATE_TRACKER
-
-
-def model_parallel_cuda_manual_seed(seed):
-    """Initialize model parallel cuda seed.
-
-    This function should be called after the model parallel is
-    initialized. Also, no torch.cuda.manual_seed should be called
-    after this function. Basically, this is replacement for that
-    function.
-    Two set of RNG states are tracked:
-        default state: This is for data parallelism and is the same among a
-                       set of model parallel GPUs but different across
-                       different model paralle groups. This is used for
-                       example for dropout in the non-tensor-model-parallel regions.
-        tensor-model-parallel state: This state is different among a set of model
-                              parallel GPUs, but the same across data parallel
-                              groups. This is used for example for dropout in
-                              model parallel regions.
-    """
-    # 2718 is just for fun and any POSITIVE value will work.
-    offset = seed + 2718
-    tensor_model_parallel_seed = offset + get_tensor_model_parallel_rank()
-    # Data parallel gets the original seed.
-    data_parallel_seed = seed
-
-    _CUDA_RNG_STATE_TRACKER.reset()
-    # Set the default state.
-    torch.cuda.manual_seed(data_parallel_seed)
-    # and model parallel state.
-    _CUDA_RNG_STATE_TRACKER.add(_MODEL_PARALLEL_RNG_TRACKER_NAME,
-                                tensor_model_parallel_seed)
-
-
-class CheckpointFunction(torch.autograd.Function):
-    """This function is adapted from torch.utils.checkpoint with
-       two main changes:
-           1) torch.cuda.set_rng_state is replaced with `_set_cuda_rng_state`
-           2) the states in the model parallel tracker are also properly
-              tracked/set/reset.
-    """
-    @staticmethod
-    def forward(ctx, run_function, distribute_saved_activations, *args):
-        ctx.run_function = run_function
-        ctx.distribute_saved_activations \
-            = distribute_saved_activations
-
-        # Copy the rng states.
-        ctx.fwd_cpu_rng_state = torch.get_rng_state()
-        ctx.fwd_cuda_rng_state = torch.cuda.get_rng_state()
-        ctx.fwd_cuda_rng_state_tracker = get_cuda_rng_tracker().get_states()
-
-        with torch.no_grad():
-            outputs = run_function(*args)
-
-        # Divide hidden states across model parallel group and only keep
-        # the chunk corresponding to the current rank.
-        if distribute_saved_activations:
-            ctx.input_0_shape = args[0].data.shape
-            safely_set_viewless_tensor_data(
-                args[0],
-                split_tensor_into_1d_equal_chunks(args[0].data, new_buffer=True))
-
-        # Store everything.
-        ctx.save_for_backward(*args)
-
-        return outputs
-
-    @staticmethod
-    def backward(ctx, *args):
-        if not torch.autograd._is_checkpoint_valid():
-            raise RuntimeError("Checkpointing is not compatible with .grad(), "
-                               "please use .backward() if possible")
-        inputs = ctx.saved_tensors
-        if ctx.distribute_saved_activations:
-            safely_set_viewless_tensor_data(
-                inputs[0],
-                gather_split_1d_tensor(inputs[0].data).view(ctx.input_0_shape))
-
-        # Store the current states.
-        bwd_cpu_rng_state = torch.get_rng_state()
-        bwd_cuda_rng_state = torch.cuda.get_rng_state()
-        bwd_cuda_rng_state_tracker = get_cuda_rng_tracker().get_states()
-
-        # Set the states to what it used to be before the forward pass.
-        torch.set_rng_state(ctx.fwd_cpu_rng_state)
-        _set_cuda_rng_state(ctx.fwd_cuda_rng_state)
-        get_cuda_rng_tracker().set_states(ctx.fwd_cuda_rng_state_tracker)
-
-        # Compute the forward pass.
-        detached_inputs = detach_variable(inputs)
-        with torch.enable_grad():
-            outputs = ctx.run_function(*detached_inputs)
-
-        # Set the states back to what it was at the start of this function.
-        torch.set_rng_state(bwd_cpu_rng_state)
-        _set_cuda_rng_state(bwd_cuda_rng_state)
-        get_cuda_rng_tracker().set_states(bwd_cuda_rng_state_tracker)
-
-        if isinstance(outputs, torch.Tensor):
-            outputs = (outputs,)
-        torch.autograd.backward(outputs, args)
-        grads = tuple(inp.grad if isinstance(inp, torch.Tensor) else inp
-                      for inp in detached_inputs)
-        return (None, None) + grads
-
-
-def checkpoint(function, distribute_saved_activations, *args):
-    """Checkpoint a model or part of the model.
-    This has been directly copied from torch.utils.checkpoint."""
-    return CheckpointFunction.apply(function,
-                                    distribute_saved_activations, *args)

cacheflow/parallel_utils/tensor_parallel/utils.py

@@ -1,108 +0,0 @@
-# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
-
-import torch
-from typing import List, Sequence
-
-from cacheflow.parallel_utils.utils import divide
-from cacheflow.parallel_utils import parallel_state
-
-def split_tensor_along_last_dim(
-    tensor: torch.Tensor,
-    num_partitions: int,
-    contiguous_split_chunks: bool = False,
-) -> List[torch.Tensor]:
-    """ Split a tensor along its last dimension.
-
-        Arguments:
-            tensor: input tensor.
-            num_partitions: number of partitions to split the tensor
-            contiguous_split_chunks: If True, make each chunk contiguous
-                                     in memory.
-
-        Returns:
-            A list of Tensors
-    """
-    # Get the size and dimension.
-    last_dim = tensor.dim() - 1
-    last_dim_size = divide(tensor.size()[last_dim], num_partitions)
-    # Split.
-    tensor_list = torch.split(tensor, last_dim_size, dim=last_dim)
-    # Note: torch.split does not create contiguous tensors by default.
-    if contiguous_split_chunks:
-        return tuple(chunk.contiguous() for chunk in tensor_list)
-
-    return tensor_list
-
-def split_tensor_into_1d_equal_chunks(tensor, new_buffer=False):
-    """ Break a tensor into equal 1D chunks across tensor parallel ranks.
-
-        Returns a Tensor or View with this rank's portion of the data.
-
-        Arguments:
-            tensor: The tensor to split
-
-        Keyword Arguments:
-            new_buffer (bool): If True, returns a new Tensor.
-                               If False, returns a view into the existing Tensor.
-                               Default is False
-
-    """
-    partition_size = torch.numel(tensor) // \
-        parallel_state.get_tensor_model_parallel_world_size()
-    start_index = partition_size * parallel_state.get_tensor_model_parallel_rank()
-    end_index = start_index + partition_size
-    if new_buffer:
-        data = torch.empty(partition_size, dtype=tensor.dtype,
-                           device=torch.cuda.current_device(),
-                           requires_grad=False)
-        data.copy_(tensor.view(-1)[start_index:end_index])
-    else:
-        data = tensor.view(-1)[start_index:end_index]
-    return data
-
-
-def gather_split_1d_tensor(tensor):
-    """ Opposite of split_tensor_into_1d_equal_chunks. Gather values from tensor
-        model parallel ranks.
-
-        Returns a new Tensor with the gathered data.
-
-        Arguments:
-            tensor: A Tensor or view of this rank's portion of the data.
-    """
-    numel_gathered = torch.numel(tensor) * \
-        parallel_state.get_tensor_model_parallel_world_size()
-    gathered = torch.empty(numel_gathered, dtype=tensor.dtype,
-                           device=torch.cuda.current_device(),
-                           requires_grad=False)
-    # TODO: This API is experimental in pytorch (as of Feb 2022) and
-    # this might break in future pytorch releases. We chose this API
-    # as opposed to torch.distributed.all_gather for efficiency reasons.
-    # This API calls directly NCCL all-gather versus the former does
-    # internal copies and can potentially cause slow down.
-    torch.distributed._all_gather_base(gathered, tensor,
-                                       group=parallel_state.get_tensor_model_parallel_group())
-    return gathered
-
-
-class VocabUtility:
-    """ Split the vocabulary into `world_size` chunks and return the first
-        and last index of the vocabulary belonging to the `rank`
-        partition: Note that indices in [fist, last)
-
-    """
-
-    @staticmethod
-    def vocab_range_from_per_partition_vocab_size(
-        per_partition_vocab_size: int, rank, world_size: int
-    ) -> Sequence[int]:
-        index_f = rank * per_partition_vocab_size
-        index_l = index_f + per_partition_vocab_size
-        return index_f, index_l
-
-    @staticmethod
-    def vocab_range_from_global_vocab_size(global_vocab_size: int, rank: int, world_size: int) -> Sequence[int]:
-        per_partition_vocab_size = divide(global_vocab_size, world_size)
-        return VocabUtility.vocab_range_from_per_partition_vocab_size(
-            per_partition_vocab_size, rank, world_size
-        )

cacheflow/parallel_utils/utils.py

@@ -1,120 +0,0 @@
-"""Utility functions used throughout Megatron core"""
-from functools import reduce
-import operator
-
-import torch
-
-from cacheflow.parallel_utils import parallel_state
-
-
-def ensure_divisibility(numerator, denominator):
-    """Ensure that numerator is divisible by the denominator."""
-    assert numerator % denominator == 0, "{} is not divisible by {}".format(
-        numerator, denominator
-    )
-
-
-def divide(numerator, denominator):
-    """Ensure that numerator is divisible by the denominator and return
-    the division value."""
-    ensure_divisibility(numerator, denominator)
-    return numerator // denominator
-
-
-class GlobalMemoryBuffer:
-    """Global buffer to avoid dynamic memory allocations.
-    Caller should ensure that buffers of the same name
-    are not used concurrently."""
-
-    def __init__(self):
-        self.buffer = {}
-
-    def get_tensor(self, tensor_shape, dtype, name):
-        required_len = reduce(operator.mul, tensor_shape, 1)
-        if self.buffer.get((name, dtype), None) is None or \
-                self.buffer[(name, dtype)].numel() < required_len:
-            self.buffer[(name, dtype)] = \
-                torch.empty(required_len,
-                            dtype=dtype,
-                            device=torch.cuda.current_device(),
-                            requires_grad=False)
-
-        return self.buffer[(name, dtype)][0:required_len].view(*tensor_shape)
-
-def _kernel_make_viewless_tensor(inp, requires_grad):
-    '''Make a viewless tensor.
-
-    View tensors have the undesirable side-affect of retaining a reference
-    to the originally-viewed tensor, even after manually setting the '.data'
-    field. This method creates a new tensor that links to the old tensor's
-    data, without linking the viewed tensor, referenced via the '._base'
-    field.
-    '''
-    out = torch.empty(
-        (1,),
-        dtype = inp.dtype,
-        device = inp.device,
-        requires_grad = requires_grad,
-    )
-    out.data = inp.data
-    return out
-
-class MakeViewlessTensor(torch.autograd.Function):
-    '''
-    Autograd function to make a viewless tensor.
-
-    This function should be used in cases where the computation graph needs
-    to be propagated, but we only want a viewless tensor (e.g.,
-    ParallelTransformer's hidden_states). Call this function by passing
-    'keep_graph = True' to 'make_viewless_tensor()'.
-    '''
-    @staticmethod
-    def forward(ctx, inp, requires_grad):
-        return _kernel_make_viewless_tensor(inp, requires_grad)
-    @staticmethod
-    def backward(ctx, grad_output):
-        return grad_output, None
-
-def make_viewless_tensor(inp, requires_grad, keep_graph):
-    '''
-    Entry-point for creating viewless tensors.
-
-    This method should be used, rather than calling 'MakeViewlessTensor'
-    or '_kernel_make_viewless_tensor' directly. This method acts as a
-    switch for determining if an autograd function or a regular method
-    should be used to create the tensor.
-    '''
-
-    # return tensor as-is, if not a 'view'
-    if inp._base is None:
-        return inp
-
-    # create viewless tensor
-    if keep_graph:
-        return MakeViewlessTensor.apply(inp, requires_grad)
-    else:
-        return _kernel_make_viewless_tensor(inp, requires_grad)
-
-def assert_viewless_tensor(tensor, extra_msg = None):
-    '''Assert that a tensor is not a view (i.e., its '._base' field is
-    not set).'''
-    if isinstance(tensor, list):
-        [ assert_viewless_tensor(t) for t in tensor ]
-        return tensor
-    if not isinstance(tensor, torch.Tensor):
-        return tensor
-    assert tensor._base is None, (
-        "Ensure tensor._base is None before setting tensor.data or storing "
-        "tensor to memory buffer. Otherwise, a memory leak will occur (and "
-        "likely accumulate over iterations). %s"
-    ) % extra_msg
-    return tensor
-
-def safely_set_viewless_tensor_data(tensor, new_data_tensor):
-    '''Safely set tensor's '.data' field.
-
-    Check first that the tensor is viewless (i.e., '._base' not set). If not,
-    raise an exception.
-    '''
-    assert_viewless_tensor(tensor, extra_msg = "FYI, tensor._base has shape %s, and new_data_tensor has shape %s." % ("--" if tensor._base is None else tensor._base.shape, new_data_tensor.shape))
-    tensor.data = new_data_tensor

cacheflow/sampling_params.py

@@ -1,84 +0,0 @@
-from typing import Optional, Set, Dict
-
-
-class SamplingParams:
-
-    def __init__(
-        self,
-        n: int,
-        temperature: float,
-        top_p: float,
-        use_beam_search: bool,
-        stop_token_ids: Set[int],
-        max_num_steps: int,
-        num_logprobs: int,
-        context_window_size: Optional[int],
-    ) -> None:
-        if n < 1:
-            raise ValueError(f'n must be at least 1, got {n}.')
-        if temperature < 0.0:
-            raise ValueError(
-                f'temperature must be non-negative, got {temperature}.')
-        if not 0.0 < top_p <= 1.0:
-            raise ValueError(f'top_p must be in (0, 1], got {top_p}.')
-        if max_num_steps < 1:
-            raise ValueError(
-                f'max_num_steps must be at least 1, got {max_num_steps}.')
-        if num_logprobs < 0:
-            raise ValueError(
-                f'num_logprobs must be non-negative, got {num_logprobs}.')
-        if context_window_size is not None and context_window_size < 0:
-            raise ValueError(
-                'context_window_size must be non-negative, '
-                f'got {context_window_size}.')
-
-        if use_beam_search:
-            if n == 1:
-                raise ValueError(
-                    'n must be greater than 1 when using beam search.')
-            if temperature > 0.0:
-                raise ValueError(
-                    'temperature must be 0 when using beam search.')
-            if top_p < 1.0:
-                raise ValueError(
-                    'top_p must be 1 when using beam search.')
-        elif temperature == 0.0:
-            # Zero temperature means greedy sampling.
-            if n > 1:
-                raise ValueError(
-                    'n must be 1 when using greedy sampling.')
-            if top_p < 1.0:
-                raise ValueError(
-                    'top_p must be 1 when using greedy sampling.')
-
-        self.n = n
-        self.temperature = temperature
-        self.top_p = top_p
-        self.use_beam_search = use_beam_search
-        self.stop_token_ids = stop_token_ids
-        self.max_num_steps = max_num_steps
-        self.num_logprobs = num_logprobs
-        self.context_window_size = context_window_size
-
-    def __repr__(self) -> str:
-        return (f'SamplingParams(n={self.n}, '
-                f'temperature={self.temperature}, '
-                f'top_p={self.top_p}, '
-                f'use_beam_search={self.use_beam_search}, '
-                f'stop_token_ids={self.stop_token_ids}, '
-                f'max_num_steps={self.max_num_steps}, '
-                f'num_logprobs={self.num_logprobs}, '
-                f'context_window_size={self.context_window_size})')
-
-    @classmethod
-    def from_dict(cls, d: Dict) -> 'SamplingParams':
-        return cls(
-            n=d.get('n', 1),
-            temperature=d.get('temperature', 1.0),
-            top_p=d.get('top_p', 1.0),
-            use_beam_search=d.get('use_beam_search', False),
-            stop_token_ids=set(d.get('stop_token_ids', set())),
-            max_num_steps=d.get('max_num_steps', 16),
-            num_logprobs=d.get('num_logprobs', 0),
-            context_window_size=d.get('context_window_size', None),
-        )

cacheflow/sequence.py

@@ -1,169 +0,0 @@
-import copy
-import enum
-from typing import Dict, List, Optional
-
-from cacheflow.block import LogicalTokenBlock
-from cacheflow.sampling_params import SamplingParams
-
-
-class SequenceStatus(enum.Enum):
-    WAITING = enum.auto()
-    RUNNING = enum.auto()
-    SWAPPED = enum.auto()
-    FINISHED = enum.auto()
-
-
-class Sequence:
-
-    def __init__(
-        self,
-        seq_id: int,
-        token_ids: List[int],
-        block_size: int,
-    ) -> None:
-        self.seq_id = seq_id
-        self.block_size = block_size
-
-        self.logical_token_blocks: List[LogicalTokenBlock] = []
-        # Initialize the logical token blocks with the given token ids.
-        self.add(token_ids)
-
-        self.prompt_len = len(token_ids)
-        self.status = SequenceStatus.WAITING
-        self.output_logprobs: List[Dict[int, float]] = []
-        self.cumulative_logprobs = 0.0
-
-    def add_block(self) -> None:
-        block = LogicalTokenBlock(
-            block_number=len(self.logical_token_blocks),
-            block_size=self.block_size,
-        )
-        self.logical_token_blocks.append(block)
-
-    def add(self, token_ids: List[int]) -> None:
-        while token_ids:
-            if not self.logical_token_blocks:
-                self.add_block()
-
-            last_block = self.logical_token_blocks[-1]
-            if last_block.is_full():
-                self.add_block()
-                last_block = self.logical_token_blocks[-1]
-
-            num_empty_slots = last_block.get_num_empty_slots()
-            last_block.append(token_ids[:num_empty_slots])
-            token_ids = token_ids[num_empty_slots:]
-
-    def append(self, token_id: int, logprobs: Dict[int, float]) -> None:
-        assert token_id in logprobs
-        self.add([token_id])
-        self.output_logprobs.append(logprobs)
-        self.cumulative_logprobs += logprobs[token_id]
-
-    def get_len(self) -> int:
-        return sum(block.num_tokens for block in self.logical_token_blocks)
-
-    def get_token_ids(self) -> List[int]:
-        token_ids: List[int] = []
-        for block in self.logical_token_blocks:
-            token_ids.extend(block.get_token_ids())
-        return token_ids
-
-    def get_last_token_id(self) -> int:
-        return self.logical_token_blocks[-1].get_last_token_id()
-
-    def fork(self, child_seq: 'Sequence') -> 'Sequence':
-        child_seq.logical_token_blocks = copy.deepcopy(self.logical_token_blocks)
-        child_seq.output_logprobs = copy.deepcopy(self.output_logprobs)
-        child_seq.cumulative_logprobs = self.cumulative_logprobs
-
-    def __repr__(self) -> str:
-        return (f'Sequence(seq_id={self.seq_id}, '
-                f'status={self.status.name}, '
-                f'num_blocks={len(self.logical_token_blocks)})')
-
-
-class SequenceGroup:
-
-    def __init__(
-        self,
-        group_id: int,
-        seqs: List[Sequence],
-        arrival_time: float,
-    ) -> None:
-        self.group_id = group_id
-        self.seqs = seqs
-        self.arrival_time = arrival_time
-
-    def get_seqs(
-        self,
-        status: Optional[SequenceStatus] = None,
-    ) -> List[Sequence]:
-        if status is None:
-            return self.seqs
-        else:
-            return [seq for seq in self.seqs if seq.status == status]
-
-    def num_seqs(self, status: Optional[SequenceStatus] = None) -> int:
-        return len(self.get_seqs(status))
-
-    def find(self, seq_id: int) -> Sequence:
-        for seq in self.seqs:
-            if seq.seq_id == seq_id:
-                return seq
-        raise ValueError(f'Sequence {seq_id} not found.')
-
-    def is_finished(self) -> bool:
-        return all(seq.status == SequenceStatus.FINISHED for seq in self.seqs)
-
-    def __repr__(self) -> str:
-        return (f'SequenceGroup(group_id={self.group_id}, '
-                f'num_seqs={len(self.seqs)})')
-
-
-class SequenceGroupInputs:
-
-    def __init__(
-        self,
-        group_id: int,
-        is_prompt: bool,
-        input_tokens: Dict[int, List[int]],     # Seq id -> token ids.
-        context_len: int,
-        seq_logprobs: Dict[int, float],         # Seq id -> cumulative logprobs.
-        sampling_params: SamplingParams,
-        block_tables: Dict[int, List[int]],     # Seq id -> List of physical block numbers.
-    ) -> None:
-        self.group_id = group_id
-        self.is_prompt = is_prompt
-        self.input_tokens = input_tokens
-        self.context_len = context_len
-        self.seq_logprobs = seq_logprobs
-        self.sampling_params = sampling_params
-        self.block_tables = block_tables
-
-
-class SequenceOutputs:
-
-    def __init__(
-        self,
-        seq_id: int,
-        parent_seq_id: int,
-        output_token: int,
-        logprobs: Dict[int, float],         # Token id -> logP(x_i+1 | x_0, ..., x_i).
-    ) -> None:
-        self.seq_id = seq_id
-        self.parent_seq_id = parent_seq_id
-        self.output_token = output_token
-        self.logprobs = logprobs
-
-    def __repr__(self) -> str:
-        return (f'SequenceOutputs(seq_id={self.seq_id}, '
-                f'parent_seq_id={self.parent_seq_id}, '
-                f'output_token={self.output_token}), '
-                f'logprobs={self.logprobs}')
-
-    def __eq__(self, other: 'SequenceOutputs') -> bool:
-        return (self.seq_id == other.seq_id and
-                self.parent_seq_id == other.parent_seq_id and
-                self.output_token == other.output_token and
-                self.logprobs == other.logprobs)

cacheflow/utils.py

@@ -1,47 +0,0 @@
-import enum
-import random
-import psutil
-
-import numpy as np
-import torch
-
-from cacheflow.parallel_utils.parallel_state import model_parallel_is_initialized
-from cacheflow.parallel_utils.tensor_parallel import model_parallel_cuda_manual_seed
-
-
-class Device(enum.Enum):
-    GPU = enum.auto()
-    CPU = enum.auto()
-
-
-class Counter:
-
-    def __init__(self, start: int = 0) -> None:
-        self.counter = start
-
-    def __next__(self) -> int:
-        id = self.counter
-        self.counter += 1
-        return id
-
-    def reset(self) -> None:
-        self.counter = 0
-
-
-def set_random_seed(seed: int):
-    random.seed(seed)
-    np.random.seed(seed)
-    torch.manual_seed(seed)
-    if torch.cuda.is_available():
-        torch.cuda.manual_seed_all(seed)
-
-    if model_parallel_is_initialized():
-        model_parallel_cuda_manual_seed(seed)
-
-
-def get_gpu_memory(gpu: int = 0) -> int:
-    return torch.cuda.get_device_properties(gpu).total_memory
-
-
-def get_cpu_memory() -> int:
-    return psutil.virtual_memory().total

cacheflow/worker/controller.py

@@ -1,101 +0,0 @@
-from typing import Dict, List, Union, Tuple
-
-import ray
-
-from cacheflow.master.scheduler import Scheduler
-from cacheflow.sequence import SequenceGroupInputs
-from cacheflow.worker.worker import Worker
-
-
-DeviceID = Tuple[int, str, int] # rank, node resource (node IP), device id
-
-
-class Controller:
-
-    def __init__(
-        self,
-        stage_id: int,
-        stage_devices: List[DeviceID],
-        world_size: int,
-        tensor_parallel_size: int,
-        pipeline_parallel_size: int,
-        distributed_init_method: str,
-        model_name: str,
-        block_size: int,
-        num_gpu_blocks: int,
-        num_cpu_blocks: int,
-        dtype: str,
-        seed: int,
-        model_path: str,
-        use_dummy_weights: bool,
-        max_num_batched_tokens: int,
-    ) -> None:
-        self.stage_id = stage_id
-        self.stage_devices = stage_devices
-        self.model_name = model_name
-        self.block_size = block_size
-        self.num_gpu_blocks = num_gpu_blocks
-        self.num_cpu_blocks = num_cpu_blocks
-
-        # Which pipeline stage is this node assigned to?
-        self.is_first_stage = stage_id == 0
-        self.is_last_stage = False
-
-        self.workers: List[Worker] = []
-        for rank, node_resource, device_id in stage_devices:
-            worker_cls = ray.remote(num_cpus=0,
-                                    num_gpus=1,
-                                    resources={node_resource: 1e-5})(Worker)
-            worker = worker_cls.remote(
-                model_name=model_name,
-                block_size=block_size,
-                num_gpu_blocks=num_gpu_blocks,
-                num_cpu_blocks=num_cpu_blocks,
-                dtype=dtype,
-                seed=seed,
-                distributed_init_method=distributed_init_method,
-                rank=rank,
-                world_size=world_size,
-                tensor_parallel_size=tensor_parallel_size,
-                pipeline_parallel_size=pipeline_parallel_size,
-                model_path=model_path,
-                use_dummy_weights=use_dummy_weights,
-                max_num_batched_tokens=max_num_batched_tokens,
-            )
-            self.workers.append(worker)
-
-    def set_next(
-        self,
-        next_node: Union['Controller', 'Scheduler'],
-    ) -> None:
-        self.next_node = next_node
-        self.is_last_stage = isinstance(next_node, Scheduler)
-
-    def execute_stage(
-        self,
-        input_seq_groups: List[SequenceGroupInputs],
-        blocks_to_swap_in: Dict[int, int],
-        blocks_to_swap_out: Dict[int, int],
-        blocks_to_copy: Dict[int, List[int]],
-    ) -> None:
-        futures = []
-        for worker in self.workers:
-            future = worker.execute_stage.remote(
-                input_seq_groups,
-                blocks_to_swap_in,
-                blocks_to_swap_out,
-                blocks_to_copy,
-            )
-            futures.append(future)
-
-        all_outputs = ray.get(futures)
-        # Make sure all workers have the same results.
-        output = all_outputs[0]
-        for other_output in all_outputs[1:]:
-            assert output == other_output
-
-        if self.is_last_stage:
-            self.next_node.post_step(output)
-        else:
-            # TODO: Support pipeline parallelism.
-            assert False

cacheflow/worker/worker.py

@@ -1,264 +0,0 @@
-from typing import Dict, List, Tuple
-
-import torch
-
-from cacheflow.models import get_model
-from cacheflow.models import InputMetadata
-from cacheflow.sampling_params import SamplingParams
-from cacheflow.sequence import SequenceGroupInputs
-from cacheflow.sequence import SequenceOutputs
-from cacheflow.worker.cache_engine import CacheEngine
-from cacheflow.parallel_utils.parallel_state import (
-    initialize_model_parallel,
-    initialize_all_reduce_launcher,
-    get_tensor_model_parallel_world_size)
-from cacheflow.utils import set_random_seed
-
-
-class Worker:
-
-    def __init__(
-        self,
-        model_name: str,
-        block_size: int,
-        num_gpu_blocks: int,
-        num_cpu_blocks: int,
-        dtype: str,
-        seed: int,
-        distributed_init_method: str,
-        rank: int,
-        world_size: int,
-        model_path: str,
-        use_dummy_weights: bool,
-        max_num_batched_tokens: int,
-        tensor_parallel_size: int = 1,
-        pipeline_parallel_size: int = 1,
-    ) -> None:
-        self.init_distributed_environment(distributed_init_method,
-                                          rank,
-                                          world_size,
-                                          tensor_parallel_size,
-                                          pipeline_parallel_size)
-        self.worker_id = rank
-        self.block_size = block_size
-        set_random_seed(seed)
-
-        # Initialize the model.
-        self.model, self.dtype = get_model(
-            model_name, dtype=dtype, path=model_path, use_dummy_weights=use_dummy_weights)
-        tensor_model_parallel_world_size = (
-            get_tensor_model_parallel_world_size())
-        initialize_all_reduce_launcher(
-            max_num_batched_tokens, self.model.config.hidden_size, self.dtype)
-        self.num_layers = self.model.config.num_hidden_layers
-        assert self.model.config.num_attention_heads % tensor_model_parallel_world_size == 0
-        self.num_heads = self.model.config.num_attention_heads // tensor_model_parallel_world_size
-        self.head_size = self.model.config.hidden_size // (self.num_heads * tensor_model_parallel_world_size)
-
-        # We reset the seed after initializing the model to ensure that
-        # the random state is not affected by the model initialization.
-        set_random_seed(seed)
-
-        self.cache_engine = CacheEngine(
-            worker_id=self.worker_id,
-            num_layers=self.num_layers,
-            num_heads=self.num_heads,
-            head_size=self.head_size,
-            block_size=block_size,
-            num_gpu_blocks=num_gpu_blocks,
-            num_cpu_blocks=num_cpu_blocks,
-            dtype=self.dtype,
-        )
-        self.cache_events = self.cache_engine.events
-        self.gpu_cache = self.cache_engine.gpu_cache
-
-
-    def init_distributed_environment(self,
-                                     distributed_init_method: str,
-                                     rank: int,
-                                     world_size: int,
-                                     tensor_parallel_size: int = 1,
-                                     pipeline_parallel_size: int = 1) -> None:
-        """Initialize the distributed environment."""
-        torch.distributed.init_process_group(
-            backend='nccl',
-            init_method=distributed_init_method,
-            world_size=world_size,
-            rank=rank,
-        )
-        # A small all_reduce for warmup.
-        torch.distributed.all_reduce(torch.zeros(1).cuda())
-        initialize_model_parallel(tensor_parallel_size,
-                                  pipeline_parallel_size)
-
-
-    def prepare_inputs(
-        self,
-        input_seq_groups: List[SequenceGroupInputs],
-    ) -> Tuple[torch.LongTensor, torch.LongTensor, InputMetadata]:
-        seq_groups: List[Tuple[List[int], SamplingParams]] = []
-        seq_logprobs: Dict[int, float] = {}
-        sampling_params: Dict[int, SamplingParams] = {}
-        input_tokens: List[int] = []
-        input_positions: List[int] = []
-        slot_mapping: List[int] = []
-
-        # Add prompt tokens.
-        prompt_lens: List[int] = []
-        for input_seq_group in input_seq_groups:
-            if not input_seq_group.is_prompt:
-                continue
-
-            seq_ids = list(input_seq_group.input_tokens.keys())
-            sampling_params = input_seq_group.sampling_params
-            seq_groups.append((seq_ids, sampling_params))
-            seq_logprobs.update(input_seq_group.seq_logprobs)
-
-            # Use any sequence in the group.
-            seq_id = seq_ids[0]
-
-            prompt_tokens = input_seq_group.input_tokens[seq_id]
-            prompt_len = len(prompt_tokens)
-            prompt_lens.append(prompt_len)
-
-            input_tokens.extend(prompt_tokens)
-            # NOTE(woosuk): Here we assume that the first token in the prompt
-            # is always the first token in the sequence.
-            input_positions.extend(range(len(prompt_tokens)))
-
-            # Compute the slot mapping.
-            block_table = input_seq_group.block_tables[seq_id]
-            for i in range(prompt_len):
-                block_number = block_table[i // self.block_size]
-                block_offset = i % self.block_size
-                slot = block_number * self.block_size + block_offset
-                slot_mapping.append(slot)
-
-        cumulative_prompt_lens: List[int] = [0]
-        for prompt_len in prompt_lens:
-            cumulative_prompt_lens.append(
-                cumulative_prompt_lens[-1] + prompt_len)
-
-        # Add generation tokens.
-        max_context_len = 0
-        max_num_blocks_per_seq = 0
-        context_lens: List[int] = []
-        generation_block_tables: List[List[int]] = []
-        for input_seq_group in input_seq_groups:
-            if input_seq_group.is_prompt:
-                continue
-
-            seq_ids = list(input_seq_group.input_tokens.keys())
-            sampling_params = input_seq_group.sampling_params
-            seq_groups.append((seq_ids, sampling_params))
-            seq_logprobs.update(input_seq_group.seq_logprobs)
-
-            for seq_id in seq_ids:
-                assert len(input_seq_group.input_tokens[seq_id]) == 1
-                generation_token = input_seq_group.input_tokens[seq_id][0]
-                input_tokens.append(generation_token)
-
-                position = input_seq_group.context_len - 1
-                input_positions.append(position)
-
-                block_table = input_seq_group.block_tables[seq_id]
-                generation_block_tables.append(block_table)
-
-                max_context_len = max(
-                    max_context_len, input_seq_group.context_len)
-                max_num_blocks_per_seq = max(
-                    max_num_blocks_per_seq, len(block_table))
-                context_lens.append(input_seq_group.context_len)
-
-                block_number = block_table[position // self.block_size]
-                block_offset = position % self.block_size
-                slot = block_number * self.block_size + block_offset
-                slot_mapping.append(slot)
-
-        # Optimization: Pad the input length to be a multiple of 8.
-        # This is required for utilizing the Tensor Cores in NVIDIA GPUs.
-        input_tokens = _pad_to_alignment(input_tokens, multiple_of=8)
-        input_positions = _pad_to_alignment(input_positions, multiple_of=8)
-
-        # Convert to tensors.
-        tokens_tensor = torch.tensor(
-            input_tokens, dtype=torch.long, device='cuda')
-        positions_tensor = torch.tensor(
-            input_positions, dtype=torch.long, device='cuda')
-        slot_mapping_tensor = torch.tensor(
-            slot_mapping, dtype=torch.int, device='cuda')
-        context_lens_tensor = torch.tensor(
-            context_lens, dtype=torch.int, device='cuda')
-        padded_block_tables = [
-            _pad_to_max(block_table, max_num_blocks_per_seq)
-            for block_table in generation_block_tables]
-        block_tables_tensor = torch.tensor(
-            padded_block_tables, dtype=torch.int, device='cuda')
-        cumulative_prompt_lens_tensor = torch.tensor(
-            cumulative_prompt_lens, dtype=torch.int, device='cuda')
-
-        input_metadata = InputMetadata(
-            seq_groups=seq_groups,
-            seq_logprobs=seq_logprobs,
-            prompt_lens=prompt_lens,
-            cumulative_prompt_lens=cumulative_prompt_lens_tensor,
-            slot_mapping=slot_mapping_tensor,
-            context_lens=context_lens_tensor,
-            max_context_len=max_context_len,
-            block_tables=block_tables_tensor,
-        )
-        return tokens_tensor, positions_tensor, input_metadata
-
-    @torch.inference_mode()
-    def execute_stage(
-        self,
-        input_seq_groups: List[SequenceGroupInputs],
-        blocks_to_swap_in: Dict[int, int],
-        blocks_to_swap_out: Dict[int, int],
-        blocks_to_copy: Dict[int, List[int]],
-    ) -> Dict[int, SequenceOutputs]:
-        # Issue cache operations.
-        command_issued = False
-        if blocks_to_swap_in:
-            self.cache_engine.swap_in(blocks_to_swap_in)
-            command_issued = True
-        if blocks_to_swap_out:
-            self.cache_engine.swap_out(blocks_to_swap_out)
-            command_issued = True
-        if blocks_to_copy:
-            self.cache_engine.copy(blocks_to_copy)
-            command_issued = True
-
-        if command_issued:
-            cache_events = self.cache_events
-        else:
-            cache_events = None
-
-        # If there is no input, we don't need to execute the model.
-        if not input_seq_groups:
-            if cache_events is not None:
-                for event in cache_events:
-                    event.wait()
-            return {}
-
-        # Prepare input tensors.
-        input_tokens, input_positions, input_metadata = self.prepare_inputs(
-            input_seq_groups)
-
-        # Execute the model.
-        output = self.model(
-            input_ids=input_tokens,
-            positions=input_positions,
-            kv_caches=self.gpu_cache,
-            input_metadata=input_metadata,
-            cache_events=cache_events,
-        )
-        return output
-
-
-def _pad_to_alignment(x: List[int], multiple_of: int) -> List[int]:
-    return x + [0] * ((-len(x)) % multiple_of)
-
-
-def _pad_to_max(x: List[int], max_len: int) -> List[int]:
-    return x + [0] * (max_len - len(x))

csrc/activation.cpp

@@ -1,12 +0,0 @@
-#include <torch/extension.h>
-
-void silu_and_mul(
-  torch::Tensor& out,
-  torch::Tensor& input);
-
-PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
-  m.def(
-    "silu_and_mul",
-    &silu_and_mul,
-    "Activation function used in SwiGLU.");
-}

csrc/activation_kernels.cu

@@ -1,7 +1,11 @@
-#include <torch/extension.h>
 #include <ATen/cuda/CUDAContext.h>
+#include <torch/extension.h>
+#include <c10/cuda/CUDAGuard.h>
 
-namespace cacheflow {
+#include "cuda_compat.h"
+#include "dispatch_utils.h"
+
+namespace vllm {
 
 template<typename T>
 __device__ __forceinline__ T silu(const T& x) {
@@ -11,36 +15,104 @@ __device__ __forceinline__ T silu(const T& x) {
 
 template<typename scalar_t>
 __global__ void silu_and_mul_kernel(
-  scalar_t* __restrict__ out,               // [num_tokens, d]
-  const scalar_t* __restrict__ input,       // [num_tokens, 2, d]
+  scalar_t* __restrict__ out,               // [..., d]
+  const scalar_t* __restrict__ input,       // [..., 2, d]
   const int d) {
-  const int token_idx = blockIdx.x;
-  for (int idx = threadIdx.x; idx < d; idx += blockDim.x) {
-    const scalar_t x = __ldg(&input[token_idx * 2 * d + idx]);
-    const scalar_t y = __ldg(&input[token_idx * 2 * d + d + idx]);
+  const int64_t token_idx = blockIdx.x;
+  for (int64_t idx = threadIdx.x; idx < d; idx += blockDim.x) {
+    const scalar_t x = VLLM_LDG(&input[token_idx * 2 * d + idx]);
+    const scalar_t y = VLLM_LDG(&input[token_idx * 2 * d + d + idx]);
     out[token_idx * d + idx] = silu(x) * y;
   }
 }
 
-} // namespace cacheflow
+} // namespace vllm
 
 void silu_and_mul(
-  torch::Tensor& out,      // [num_tokens, d]
-  torch::Tensor& input)    // [num_tokens, 2 * d]
+  torch::Tensor& out,      // [..., d]
+  torch::Tensor& input)    // [..., 2 * d]
 {
-  int num_tokens = input.size(0);
-  int d = input.size(1) / 2;
+  int64_t num_tokens = input.numel() / input.size(-1);
+  int d = input.size(-1) / 2;
 
   dim3 grid(num_tokens);
   dim3 block(std::min(d, 1024));
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(input));
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
-  AT_DISPATCH_FLOATING_TYPES_AND_HALF(
+  VLLM_DISPATCH_FLOATING_TYPES(
     input.scalar_type(),
     "silu_and_mul_kernel",
     [&] {
-      cacheflow::silu_and_mul_kernel<scalar_t><<<grid, block, 0, stream>>>(
+      vllm::silu_and_mul_kernel<scalar_t><<<grid, block, 0, stream>>>(
         out.data_ptr<scalar_t>(),
         input.data_ptr<scalar_t>(),
         d);
     });
 }
+
+namespace vllm {
+
+// Element-wise activation kernel template.
+template<typename scalar_t, scalar_t (*ACT_FN)(const scalar_t&)>
+__global__ void activation_kernel(
+  scalar_t* __restrict__ out,               // [..., d]
+  const scalar_t* __restrict__ input,       // [..., d]
+  const int d) {
+  const int64_t token_idx = blockIdx.x;
+  for (int64_t idx = threadIdx.x; idx < d; idx += blockDim.x) {
+    const scalar_t x = VLLM_LDG(&input[token_idx * d + idx]);
+    out[token_idx * d + idx] = ACT_FN(x);
+  }
+}
+
+} // namespace vllm
+
+// Launch element-wise activation kernel.
+#define LAUNCH_ACTIVATION_KERNEL(KERNEL)                                                  \
+  int d = input.size(-1);                                                                 \
+  int64_t num_tokens = input.numel() / d;                                                 \
+  dim3 grid(num_tokens);                                                                  \
+  dim3 block(std::min(d, 1024));                                                          \
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(input));                       \
+  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();                           \
+  VLLM_DISPATCH_FLOATING_TYPES(                                                           \
+    input.scalar_type(),                                                                  \
+    "activation_kernel",                                                                  \
+    [&] {                                                                                 \
+      vllm::activation_kernel<scalar_t, KERNEL<scalar_t>><<<grid, block, 0, stream>>>(    \
+        out.data_ptr<scalar_t>(),                                                         \
+        input.data_ptr<scalar_t>(),                                                       \
+        d);                                                                               \
+    });
+
+namespace vllm {
+
+template<typename T>
+__device__ __forceinline__ T gelu_new_kernel(const T& x) {
+  const float x3 = (float) (x * x * x);
+  const T t = (T) tanhf((T) (0.79788456f * (float) (x + (T) (0.044715f * x3))));
+  return ((T) 0.5) * x * (((T) 1.0) + t);
+}
+
+template<typename T>
+__device__ __forceinline__ T gelu_fast_kernel(const T& x) {
+  const float f = (float) x;
+  const T t = (T) tanhf(((T) (f * 0.79788456f)) * (((T) 1.0) + (T) (0.044715f * f) * x));
+  return ((T) 0.5) * x * (((T) 1.0) + t);
+}
+
+} // namespace vllm
+
+void gelu_new(
+  torch::Tensor& out,     // [..., d]
+  torch::Tensor& input)   // [..., d]
+{
+  LAUNCH_ACTIVATION_KERNEL(vllm::gelu_new_kernel);
+}
+
+void gelu_fast(
+  torch::Tensor& out,     // [..., d]
+  torch::Tensor& input)   // [..., d]
+{
+  LAUNCH_ACTIVATION_KERNEL(vllm::gelu_fast_kernel);
+}

csrc/attention.cpp

@@ -1,19 +0,0 @@
-#include <torch/extension.h>
-
-void single_query_cached_kv_attention(
-  torch::Tensor& out,
-  torch::Tensor& query,
-  torch::Tensor& key_cache,
-  torch::Tensor& value_cache,
-  float scale,
-  torch::Tensor& block_tables,
-  torch::Tensor& context_lens,
-  int block_size,
-  int max_context_len);
-
-PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
-  m.def(
-    "single_query_cached_kv_attention",
-    &single_query_cached_kv_attention,
-    "Compute the attention between an input query and the cached key/value tensors");
-}

csrc/attention/attention_dtypes.h

@@ -0,0 +1,7 @@
+#pragma once
+
+#include "attention_generic.cuh"
+#include "dtype_float16.cuh"
+#include "dtype_float32.cuh"
+#include "dtype_bfloat16.cuh"
+#include "dtype_fp8_e5m2.cuh"

csrc/attention/attention_generic.cuh

@@ -0,0 +1,64 @@
+/*
+ * Adapted from https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/kernels/decoder_masked_multihead_attention_utils.h
+ * Copyright (c) 2023, The vLLM team.
+ * Copyright (c) 2020-2023, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#pragma once
+
+#include <stdint.h>
+
+namespace vllm {
+
+// A vector type to store Q, K, V elements.
+template<typename T, int VEC_SIZE>
+struct Vec {};
+
+// A vector type to store FP32 accumulators.
+template<typename T>
+struct FloatVec {};
+
+// Template vector operations.
+template<typename Acc, typename A, typename B>
+inline __device__ Acc mul(A a, B b);
+
+template<typename T>
+inline __device__ float sum(T v);
+
+template<typename T>
+inline __device__ float dot(T a, T b) {
+  return sum(mul<T, T, T>(a, b));
+}
+
+template<typename A, typename T>
+inline __device__ float dot(T a, T b) {
+  return sum(mul<A, T, T>(a, b));
+}
+
+template<typename T>
+inline __device__ void zero(T& dst) {
+  constexpr int WORDS = sizeof(T) / 4;
+  union {
+    T raw;
+    uint32_t words[WORDS];
+  } tmp;
+
+#pragma unroll
+  for (int ii = 0; ii < WORDS; ++ii) {
+    tmp.words[ii] = 0u;
+  }
+  dst = tmp.raw;
+}
+
+} // namespace vllm

csrc/attention/attention_kernels.cu

@@ -0,0 +1,953 @@
+/*
+ * Adapted from https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/kernels/decoder_masked_multihead_attention/decoder_masked_multihead_attention_template.hpp
+ * Copyright (c) 2023, The vLLM team.
+ * Copyright (c) 2020-2023, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifdef USE_ROCM
+#include <hip/hip_runtime.h>
+#endif
+
+#include <torch/extension.h>
+#include <ATen/cuda/CUDAContext.h>
+#include <c10/cuda/CUDAGuard.h>
+
+#include "attention_dtypes.h"
+#include "attention_utils.cuh"
+#ifdef ENABLE_FP8_E5M2
+#include "../quantization/fp8_e5m2_kvcache/quant_utils.cuh"
+#endif
+
+#include <algorithm>
+
+#ifndef USE_ROCM
+#define WARP_SIZE 32
+#else
+#define WARP_SIZE warpSize
+#endif
+#define MAX(a, b) ((a) > (b) ? (a) : (b))
+#define MIN(a, b) ((a) < (b) ? (a) : (b))
+#define DIVIDE_ROUND_UP(a, b) (((a) + (b) - 1) / (b))
+
+namespace vllm {
+
+// Utility function for attention softmax.
+template<int NUM_WARPS>
+inline __device__ float block_sum(float* red_smem, float sum) {
+  // Decompose the thread index into warp / lane.
+  int warp = threadIdx.x / WARP_SIZE;
+  int lane = threadIdx.x % WARP_SIZE;
+
+  // Compute the sum per warp.
+#pragma unroll
+  for (int mask = WARP_SIZE / 2; mask >= 1; mask /= 2) {
+    sum += VLLM_SHFL_XOR_SYNC(sum, mask);
+  }
+
+  // Warp leaders store the data to shared memory.
+  if (lane == 0) {
+    red_smem[warp] = sum;
+  }
+
+  // Make sure the data is in shared memory.
+  __syncthreads();
+
+  // The warps compute the final sums.
+  if (lane < NUM_WARPS) {
+    sum = red_smem[lane];
+  }
+
+  // Parallel reduction inside the warp.
+#pragma unroll
+  for (int mask = NUM_WARPS / 2; mask >= 1; mask /= 2) {
+    sum += VLLM_SHFL_XOR_SYNC(sum, mask);
+  }
+
+  // Broadcast to other threads.
+  return VLLM_SHFL_SYNC(sum, 0);
+}
+
+// TODO(woosuk): Merge the last two dimensions of the grid.
+// Grid: (num_heads, num_seqs, max_num_partitions).
+template<
+  typename scalar_t,
+  typename cache_t,
+  int HEAD_SIZE,
+  int BLOCK_SIZE,
+  int NUM_THREADS,
+  bool IS_FP8_E5M2_KV_CACHE,
+  int PARTITION_SIZE = 0> // Zero means no partitioning.
+__device__ void paged_attention_kernel(
+  float* __restrict__ exp_sums,           // [num_seqs, num_heads, max_num_partitions]
+  float* __restrict__ max_logits,         // [num_seqs, num_heads, max_num_partitions]
+  scalar_t* __restrict__ out,             // [num_seqs, num_heads, max_num_partitions, head_size]
+  const scalar_t* __restrict__ q,         // [num_seqs, num_heads, head_size]
+  const cache_t* __restrict__ k_cache,    // [num_blocks, num_kv_heads, head_size/x, block_size, x]
+  const cache_t* __restrict__ v_cache,    // [num_blocks, num_kv_heads, head_size, block_size]
+  const int num_kv_heads,                 // [num_heads]
+  const float scale,
+  const int* __restrict__ block_tables,   // [num_seqs, max_num_blocks_per_seq]
+  const int* __restrict__ context_lens,   // [num_seqs]
+  const int max_num_blocks_per_seq,
+  const float* __restrict__ alibi_slopes, // [num_heads]
+  const int q_stride,
+  const int kv_block_stride,
+  const int kv_head_stride) {
+  const int seq_idx = blockIdx.y;
+  const int partition_idx = blockIdx.z;
+  const int max_num_partitions = gridDim.z;
+  constexpr bool USE_PARTITIONING = PARTITION_SIZE > 0;
+  const int context_len = context_lens[seq_idx];
+  if (USE_PARTITIONING && partition_idx * PARTITION_SIZE >= context_len) {
+    // No work to do. Terminate the thread block.
+    return;
+  }
+
+  const int num_context_blocks = DIVIDE_ROUND_UP(context_len, BLOCK_SIZE);
+  const int num_blocks_per_partition = USE_PARTITIONING ? PARTITION_SIZE / BLOCK_SIZE : num_context_blocks;
+
+  // [start_block_idx, end_block_idx) is the range of blocks to process.
+  const int start_block_idx = USE_PARTITIONING ? partition_idx * num_blocks_per_partition : 0;
+  const int end_block_idx = MIN(start_block_idx + num_blocks_per_partition, num_context_blocks);
+  const int num_blocks = end_block_idx - start_block_idx;
+
+  // [start_token_idx, end_token_idx) is the range of tokens to process.
+  const int start_token_idx = start_block_idx * BLOCK_SIZE;
+  const int end_token_idx = MIN(start_token_idx + num_blocks * BLOCK_SIZE, context_len);
+  const int num_tokens = end_token_idx - start_token_idx;
+
+  constexpr int THREAD_GROUP_SIZE = MAX(WARP_SIZE / BLOCK_SIZE, 1);
+  constexpr int NUM_THREAD_GROUPS = NUM_THREADS / THREAD_GROUP_SIZE; // Note: This assumes THREAD_GROUP_SIZE divides NUM_THREADS
+  assert(NUM_THREADS % THREAD_GROUP_SIZE == 0);
+  constexpr int NUM_TOKENS_PER_THREAD_GROUP = DIVIDE_ROUND_UP(BLOCK_SIZE, WARP_SIZE);
+  constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
+  const int thread_idx = threadIdx.x;
+  const int warp_idx = thread_idx / WARP_SIZE;
+  const int lane = thread_idx % WARP_SIZE;
+
+  const int head_idx = blockIdx.x;
+  const int num_heads = gridDim.x;
+  const int num_queries_per_kv = num_heads / num_kv_heads;
+  const int kv_head_idx = head_idx / num_queries_per_kv;
+  const float alibi_slope = alibi_slopes == nullptr ? 0.f : alibi_slopes[head_idx];
+
+  // A vector type to store a part of a key or a query.
+  // The vector size is configured in such a way that the threads in a thread group
+  // fetch or compute 16 bytes at a time.
+  // For example, if the size of a thread group is 4 and the data type is half,
+  // then the vector size is 16 / (4 * sizeof(half)) == 2.
+  constexpr int VEC_SIZE = MAX(16 / (THREAD_GROUP_SIZE * sizeof(scalar_t)), 1);
+  using K_vec = typename Vec<scalar_t, VEC_SIZE>::Type;
+  using Q_vec = typename Vec<scalar_t, VEC_SIZE>::Type;
+#ifdef ENABLE_FP8_E5M2
+  using Quant_vec = typename Vec<cache_t, VEC_SIZE>::Type;
+#endif
+
+  constexpr int NUM_ELEMS_PER_THREAD = HEAD_SIZE / THREAD_GROUP_SIZE;
+  constexpr int NUM_VECS_PER_THREAD = NUM_ELEMS_PER_THREAD / VEC_SIZE;
+
+  const int thread_group_idx = thread_idx / THREAD_GROUP_SIZE;
+  const int thread_group_offset = thread_idx % THREAD_GROUP_SIZE;
+
+  // Load the query to registers.
+  // Each thread in a thread group has a different part of the query.
+  // For example, if the the thread group size is 4, then the first thread in the group
+  // has 0, 4, 8, ... th vectors of the query, and the second thread has 1, 5, 9, ...
+  // th vectors of the query, and so on.
+  // NOTE(woosuk): Because q is split from a qkv tensor, it may not be contiguous.
+  const scalar_t* q_ptr = q + seq_idx * q_stride + head_idx * HEAD_SIZE;
+  __shared__ Q_vec q_vecs[THREAD_GROUP_SIZE][NUM_VECS_PER_THREAD];
+#pragma unroll
+  for (int i = thread_group_idx; i < NUM_VECS_PER_THREAD; i += NUM_THREAD_GROUPS) {
+    const int vec_idx = thread_group_offset + i * THREAD_GROUP_SIZE;
+    q_vecs[thread_group_offset][i] = *reinterpret_cast<const Q_vec*>(q_ptr + vec_idx * VEC_SIZE);
+  }
+  __syncthreads(); // TODO(naed90): possible speedup if this is replaced with a memory wall right before we use q_vecs
+
+  // Memory planning.
+  extern __shared__ char shared_mem[];
+  // NOTE(woosuk): We use FP32 for the softmax logits for better accuracy.
+  float* logits = reinterpret_cast<float*>(shared_mem);
+  // Workspace for reduction.
+  __shared__ float red_smem[2 * NUM_WARPS];
+
+  // x == THREAD_GROUP_SIZE * VEC_SIZE
+  // Each thread group fetches x elements from the key at a time.
+  constexpr int x = 16 / sizeof(cache_t);
+  float qk_max = -FLT_MAX;
+
+  // Iterate over the key blocks.
+  // Each warp fetches a block of keys for each iteration.
+  // Each thread group in a warp fetches a key from the block, and computes
+  // dot product with the query.
+  const int* block_table = block_tables + seq_idx * max_num_blocks_per_seq;
+  for (int block_idx = start_block_idx + warp_idx; block_idx < end_block_idx; block_idx += NUM_WARPS) {
+    // NOTE(woosuk): The block number is stored in int32. However, we cast it to int64
+    // because int32 can lead to overflow when this variable is multiplied by large numbers
+    // (e.g., kv_block_stride).
+    const int64_t physical_block_number = static_cast<int64_t>(block_table[block_idx]);
+
+    // Load a key to registers.
+    // Each thread in a thread group has a different part of the key.
+    // For example, if the the thread group size is 4, then the first thread in the group
+    // has 0, 4, 8, ... th vectors of the key, and the second thread has 1, 5, 9, ... th
+    // vectors of the key, and so on.
+    for (int i = 0; i < NUM_TOKENS_PER_THREAD_GROUP; i++) {
+      const int physical_block_offset = (thread_group_idx + i * WARP_SIZE) % BLOCK_SIZE;
+      const int token_idx = block_idx * BLOCK_SIZE + physical_block_offset;
+      K_vec k_vecs[NUM_VECS_PER_THREAD];
+
+#pragma unroll
+      for (int j = 0; j < NUM_VECS_PER_THREAD; j++) {
+        const cache_t* k_ptr = k_cache + physical_block_number * kv_block_stride
+                                       + kv_head_idx * kv_head_stride
+                                       + physical_block_offset * x;
+        const int vec_idx = thread_group_offset + j * THREAD_GROUP_SIZE;
+        const int offset1 = (vec_idx * VEC_SIZE) / x;
+        const int offset2 = (vec_idx * VEC_SIZE) % x;
+        if constexpr (IS_FP8_E5M2_KV_CACHE) {
+#ifdef ENABLE_FP8_E5M2
+          Quant_vec k_vec_quant = *reinterpret_cast<const Quant_vec*>(k_ptr + offset1 * BLOCK_SIZE * x + offset2);
+          // Vector conversion from Quant_vec to K_vec.
+          k_vecs[j] = fp8_e5m2_unscaled::vec_conversion<K_vec, Quant_vec>(k_vec_quant);
+#else
+          assert(false);
+#endif
+        } else {
+          k_vecs[j] = *reinterpret_cast<const K_vec*>(k_ptr + offset1 * BLOCK_SIZE * x + offset2);
+        }
+      }
+
+      // Compute dot product.
+      // This includes a reduction across the threads in the same thread group.
+      float qk = scale * Qk_dot<scalar_t, THREAD_GROUP_SIZE>::dot(q_vecs[thread_group_offset], k_vecs);
+      // Add the ALiBi bias if slopes are given.
+      qk += (alibi_slope != 0) ? alibi_slope * (token_idx - context_len + 1) : 0;
+
+      if (thread_group_offset == 0) {
+        // Store the partial reductions to shared memory.
+        // NOTE(woosuk): It is required to zero out the masked logits.
+        const bool mask = token_idx >= context_len;
+        logits[token_idx - start_token_idx] = mask ? 0.f : qk;
+        // Update the max value.
+        qk_max = mask ? qk_max : fmaxf(qk_max, qk);
+      }
+    }
+  }
+
+  // Perform reduction across the threads in the same warp to get the
+  // max qk value for each "warp" (not across the thread block yet).
+  // The 0-th thread of each thread group already has its max qk value.
+#pragma unroll
+  for (int mask = WARP_SIZE / 2; mask >= THREAD_GROUP_SIZE; mask /= 2) {
+    qk_max = fmaxf(qk_max, VLLM_SHFL_XOR_SYNC(qk_max, mask));
+  }
+  if (lane == 0) {
+    red_smem[warp_idx] = qk_max;
+  }
+  __syncthreads();
+
+  // TODO(woosuk): Refactor this part.
+  // Get the max qk value for the sequence.
+  qk_max = lane < NUM_WARPS ? red_smem[lane] : -FLT_MAX;
+#pragma unroll
+  for (int mask = NUM_WARPS / 2; mask >= 1; mask /= 2) {
+    qk_max = fmaxf(qk_max, VLLM_SHFL_XOR_SYNC(qk_max, mask));
+  }
+  // Broadcast the max qk value to all threads.
+  qk_max = VLLM_SHFL_SYNC(qk_max, 0);
+
+  // Get the sum of the exp values.
+  float exp_sum = 0.f;
+  for (int i = thread_idx; i < num_tokens; i += NUM_THREADS) {
+    float val = __expf(logits[i] - qk_max);
+    logits[i] = val;
+    exp_sum += val;
+  }
+  exp_sum = block_sum<NUM_WARPS>(&red_smem[NUM_WARPS], exp_sum);
+
+  // Compute softmax.
+  const float inv_sum = __fdividef(1.f, exp_sum + 1e-6f);
+  for (int i = thread_idx; i < num_tokens; i += NUM_THREADS) {
+    logits[i] *= inv_sum;
+  }
+  __syncthreads();
+
+  // If partitioning is enabled, store the max logit and exp_sum.
+  if (USE_PARTITIONING && thread_idx == 0) {
+    float* max_logits_ptr = max_logits + seq_idx * num_heads * max_num_partitions
+                                       + head_idx * max_num_partitions
+                                       + partition_idx;
+    *max_logits_ptr = qk_max;
+    float* exp_sums_ptr = exp_sums + seq_idx * num_heads * max_num_partitions
+                                   + head_idx * max_num_partitions
+                                   + partition_idx;
+    *exp_sums_ptr = exp_sum;
+  }
+
+  // Each thread will fetch 16 bytes from the value cache at a time.
+  constexpr int V_VEC_SIZE = MIN(16 / sizeof(scalar_t), BLOCK_SIZE);
+  using V_vec = typename Vec<scalar_t, V_VEC_SIZE>::Type;
+  using L_vec = typename Vec<scalar_t, V_VEC_SIZE>::Type;
+#ifdef ENABLE_FP8_E5M2
+  using V_quant_vec = typename Vec<cache_t, V_VEC_SIZE>::Type;
+#endif
+  using Float_L_vec = typename FloatVec<L_vec>::Type;
+
+  constexpr int NUM_V_VECS_PER_ROW = BLOCK_SIZE / V_VEC_SIZE;
+  constexpr int NUM_ROWS_PER_ITER = WARP_SIZE / NUM_V_VECS_PER_ROW;
+  constexpr int NUM_ROWS_PER_THREAD = DIVIDE_ROUND_UP(HEAD_SIZE, NUM_ROWS_PER_ITER);
+
+  // NOTE(woosuk): We use FP32 for the accumulator for better accuracy.
+  float accs[NUM_ROWS_PER_THREAD];
+#pragma unroll
+  for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
+    accs[i] = 0.f;
+  }
+
+  scalar_t zero_value;
+  zero(zero_value);
+  for (int block_idx = start_block_idx + warp_idx; block_idx < end_block_idx; block_idx += NUM_WARPS) {
+    // NOTE(woosuk): The block number is stored in int32. However, we cast it to int64
+    // because int32 can lead to overflow when this variable is multiplied by large numbers
+    // (e.g., kv_block_stride).
+    const int64_t physical_block_number = static_cast<int64_t>(block_table[block_idx]);
+    const int physical_block_offset = (lane % NUM_V_VECS_PER_ROW) * V_VEC_SIZE;
+    const int token_idx = block_idx * BLOCK_SIZE + physical_block_offset;
+    L_vec logits_vec;
+    from_float(logits_vec, *reinterpret_cast<Float_L_vec*>(logits + token_idx - start_token_idx));
+
+    const cache_t* v_ptr = v_cache + physical_block_number * kv_block_stride
+                                   + kv_head_idx * kv_head_stride;
+#pragma unroll
+    for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
+      const int row_idx = lane / NUM_V_VECS_PER_ROW + i * NUM_ROWS_PER_ITER;
+      if (row_idx < HEAD_SIZE) {
+        const int offset = row_idx * BLOCK_SIZE + physical_block_offset;
+        V_vec v_vec;
+        if constexpr (IS_FP8_E5M2_KV_CACHE) {
+#ifdef ENABLE_FP8_E5M2
+          V_quant_vec v_quant_vec = *reinterpret_cast<const V_quant_vec*>(v_ptr + offset);
+          // Vector conversion from V_quant_vec to V_vec.
+          v_vec = fp8_e5m2_unscaled::vec_conversion<V_vec, V_quant_vec>(v_quant_vec);
+#else
+          assert(false);
+#endif
+        } else {
+          v_vec = *reinterpret_cast<const V_vec*>(v_ptr + offset);
+        }
+        if (block_idx == num_context_blocks - 1) {
+          // NOTE(woosuk): When v_vec contains the tokens that are out of the context,
+          // we should explicitly zero out the values since they may contain NaNs.
+          // See https://github.com/vllm-project/vllm/issues/641#issuecomment-1682544472
+          scalar_t* v_vec_ptr = reinterpret_cast<scalar_t*>(&v_vec);
+#pragma unroll
+          for (int j = 0; j < V_VEC_SIZE; j++) {
+            v_vec_ptr[j] = token_idx + j < context_len ? v_vec_ptr[j] : zero_value;
+          }
+        }
+        accs[i] += dot(logits_vec, v_vec);
+      }
+    }
+  }
+
+  // Perform reduction within each warp.
+#pragma unroll
+  for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
+    float acc = accs[i];
+#pragma unroll
+    for (int mask = NUM_V_VECS_PER_ROW / 2; mask >= 1; mask /= 2) {
+      acc += VLLM_SHFL_XOR_SYNC(acc, mask);
+    }
+    accs[i] = acc;
+  }
+
+  // NOTE(woosuk): A barrier is required because the shared memory space for logits
+  // is reused for the output.
+  __syncthreads();
+
+  // Perform reduction across warps.
+  float* out_smem = reinterpret_cast<float*>(shared_mem);
+#pragma unroll
+  for (int i = NUM_WARPS; i > 1; i /= 2) {
+    int mid = i / 2;
+    // Upper warps write to shared memory.
+    if (warp_idx >= mid && warp_idx < i) {
+      float* dst = &out_smem[(warp_idx - mid) * HEAD_SIZE];
+#pragma unroll
+      for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
+        const int row_idx = lane / NUM_V_VECS_PER_ROW + i * NUM_ROWS_PER_ITER;
+        if (row_idx < HEAD_SIZE && lane % NUM_V_VECS_PER_ROW == 0) {
+          dst[row_idx] = accs[i];
+        }
+      }
+    }
+    __syncthreads();
+
+    // Lower warps update the output.
+    if (warp_idx < mid) {
+      const float* src = &out_smem[warp_idx * HEAD_SIZE];
+#pragma unroll
+      for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
+        const int row_idx = lane / NUM_V_VECS_PER_ROW + i * NUM_ROWS_PER_ITER;
+        if (row_idx < HEAD_SIZE && lane % NUM_V_VECS_PER_ROW == 0) {
+          accs[i] += src[row_idx];
+        }
+      }
+    }
+    __syncthreads();
+  }
+
+  // Write the final output.
+  if (warp_idx == 0) {
+    scalar_t* out_ptr = out + seq_idx * num_heads * max_num_partitions * HEAD_SIZE
+                            + head_idx * max_num_partitions * HEAD_SIZE
+                            + partition_idx * HEAD_SIZE;
+#pragma unroll
+    for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
+      const int row_idx = lane / NUM_V_VECS_PER_ROW + i * NUM_ROWS_PER_ITER;
+      if (row_idx < HEAD_SIZE && lane % NUM_V_VECS_PER_ROW == 0) {
+        from_float(*(out_ptr + row_idx), accs[i]);
+      }
+    }
+  }
+}
+
+// Grid: (num_heads, num_seqs, 1).
+template<
+  typename scalar_t,
+  typename cache_t,
+  int HEAD_SIZE,
+  int BLOCK_SIZE,
+  int NUM_THREADS,
+  bool IS_FP8_E5M2_KV_CACHE>
+__global__ void paged_attention_v1_kernel(
+  scalar_t* __restrict__ out,             // [num_seqs, num_heads, head_size]
+  const scalar_t* __restrict__ q,         // [num_seqs, num_heads, head_size]
+  const cache_t* __restrict__ k_cache,    // [num_blocks, num_kv_heads, head_size/x, block_size, x]
+  const cache_t* __restrict__ v_cache,    // [num_blocks, num_kv_heads, head_size, block_size]
+  const int num_kv_heads,                 // [num_heads]
+  const float scale,
+  const int* __restrict__ block_tables,   // [num_seqs, max_num_blocks_per_seq]
+  const int* __restrict__ context_lens,   // [num_seqs]
+  const int max_num_blocks_per_seq,
+  const float* __restrict__ alibi_slopes, // [num_heads]
+  const int q_stride,
+  const int kv_block_stride,
+  const int kv_head_stride) {
+  paged_attention_kernel<scalar_t, cache_t, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS, IS_FP8_E5M2_KV_CACHE>(
+    /* exp_sums */ nullptr, /* max_logits */ nullptr,
+    out, q, k_cache, v_cache, num_kv_heads, scale, block_tables, context_lens,
+    max_num_blocks_per_seq, alibi_slopes, q_stride, kv_block_stride, kv_head_stride);
+}
+
+// Grid: (num_heads, num_seqs, max_num_partitions).
+template<
+  typename scalar_t,
+  typename cache_t,
+  int HEAD_SIZE,
+  int BLOCK_SIZE,
+  int NUM_THREADS,
+  bool IS_FP8_E5M2_KV_CACHE,
+  int PARTITION_SIZE>
+__global__ void paged_attention_v2_kernel(
+  float* __restrict__ exp_sums,           // [num_seqs, num_heads, max_num_partitions]
+  float* __restrict__ max_logits,         // [num_seqs, num_heads, max_num_partitions]
+  scalar_t* __restrict__ tmp_out,         // [num_seqs, num_heads, max_num_partitions, head_size]
+  const scalar_t* __restrict__ q,         // [num_seqs, num_heads, head_size]
+  const cache_t* __restrict__ k_cache,    // [num_blocks, num_kv_heads, head_size/x, block_size, x]
+  const cache_t* __restrict__ v_cache,    // [num_blocks, num_kv_heads, head_size, block_size]
+  const int num_kv_heads,                 // [num_heads]
+  const float scale,
+  const int* __restrict__ block_tables,   // [num_seqs, max_num_blocks_per_seq]
+  const int* __restrict__ context_lens,   // [num_seqs]
+  const int max_num_blocks_per_seq,
+  const float* __restrict__ alibi_slopes, // [num_heads]
+  const int q_stride,
+  const int kv_block_stride,
+  const int kv_head_stride) {
+  paged_attention_kernel<scalar_t, cache_t, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS, IS_FP8_E5M2_KV_CACHE, PARTITION_SIZE>(
+    exp_sums, max_logits, tmp_out, q, k_cache, v_cache, num_kv_heads, scale,
+    block_tables, context_lens, max_num_blocks_per_seq, alibi_slopes,
+    q_stride, kv_block_stride, kv_head_stride);
+}
+
+// Grid: (num_heads, num_seqs).
+template<
+  typename scalar_t,
+  int HEAD_SIZE,
+  int NUM_THREADS,
+  int PARTITION_SIZE>
+__global__ void paged_attention_v2_reduce_kernel(
+  scalar_t* __restrict__ out,             // [num_seqs, num_heads, head_size]
+  const float* __restrict__ exp_sums,     // [num_seqs, num_heads, max_num_partitions]
+  const float* __restrict__ max_logits,   // [num_seqs, num_heads, max_num_partitions]
+  const scalar_t* __restrict__ tmp_out,   // [num_seqs, num_heads, max_num_partitions, head_size]
+  const int* __restrict__ context_lens,   // [num_seqs]
+  const int max_num_partitions) {
+  const int num_heads = gridDim.x;
+  const int head_idx = blockIdx.x;
+  const int seq_idx = blockIdx.y;
+  const int context_len = context_lens[seq_idx];
+  const int num_partitions = DIVIDE_ROUND_UP(context_len, PARTITION_SIZE);
+  if (num_partitions == 1) {
+    // No need to reduce. Only copy tmp_out to out.
+    scalar_t* out_ptr = out + seq_idx * num_heads * HEAD_SIZE + head_idx * HEAD_SIZE;
+    const scalar_t* tmp_out_ptr = tmp_out + seq_idx * num_heads * max_num_partitions * HEAD_SIZE
+                                          + head_idx * max_num_partitions * HEAD_SIZE;
+    for (int i = threadIdx.x; i < HEAD_SIZE; i += blockDim.x) {
+      out_ptr[i] = tmp_out_ptr[i];
+    }
+    // Terminate the thread block.
+    return;
+  }
+
+  constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
+  const int warp_idx = threadIdx.x / WARP_SIZE;
+  const int lane = threadIdx.x % WARP_SIZE;
+
+  // Size: 2 * num_partitions.
+  extern __shared__ char shared_mem[];
+  // Workspace for reduction.
+  __shared__ float red_smem[2 * NUM_WARPS];
+
+  // Load max logits to shared memory.
+  float* shared_max_logits = reinterpret_cast<float*>(shared_mem);
+  const float* max_logits_ptr = max_logits + seq_idx * num_heads * max_num_partitions
+                                           + head_idx * max_num_partitions;
+  float max_logit = -FLT_MAX;
+  for (int i = threadIdx.x; i < num_partitions; i += blockDim.x) {
+    const float l = max_logits_ptr[i];
+    shared_max_logits[i] = l;
+    max_logit = fmaxf(max_logit, l);
+  }
+  __syncthreads();
+
+  // Get the global max logit.
+  // Reduce within the warp.
+#pragma unroll
+  for (int mask = WARP_SIZE / 2; mask >= 1; mask /= 2) {
+    max_logit = fmaxf(max_logit, VLLM_SHFL_XOR_SYNC(max_logit, mask));
+  }
+  if (lane == 0) {
+    red_smem[warp_idx] = max_logit;
+  }
+  __syncthreads();
+  // Reduce across warps.
+  max_logit = lane < NUM_WARPS ? red_smem[lane] : -FLT_MAX;
+#pragma unroll
+  for (int mask = NUM_WARPS / 2; mask >= 1; mask /= 2) {
+    max_logit = fmaxf(max_logit, VLLM_SHFL_XOR_SYNC(max_logit, mask));
+  }
+  // Broadcast the max value to all threads.
+  max_logit = VLLM_SHFL_SYNC(max_logit, 0);
+
+  // Load rescaled exp sums to shared memory.
+  float* shared_exp_sums = reinterpret_cast<float*>(shared_mem + sizeof(float) * num_partitions);
+  const float* exp_sums_ptr = exp_sums + seq_idx * num_heads * max_num_partitions
+                                       + head_idx * max_num_partitions;
+  float global_exp_sum = 0.0f;
+  for (int i = threadIdx.x; i < num_partitions; i += blockDim.x) {
+    float l = shared_max_logits[i];
+    float rescaled_exp_sum = exp_sums_ptr[i] * expf(l - max_logit);
+    global_exp_sum += rescaled_exp_sum;
+    shared_exp_sums[i] = rescaled_exp_sum;
+  }
+  __syncthreads();
+  global_exp_sum = block_sum<NUM_WARPS>(&red_smem[NUM_WARPS], global_exp_sum);
+  const float inv_global_exp_sum = __fdividef(1.0f, global_exp_sum + 1e-6f);
+
+  // Aggregate tmp_out to out.
+  const scalar_t* tmp_out_ptr = tmp_out + seq_idx * num_heads * max_num_partitions * HEAD_SIZE
+                                        + head_idx * max_num_partitions * HEAD_SIZE;
+  scalar_t* out_ptr = out + seq_idx * num_heads * HEAD_SIZE + head_idx * HEAD_SIZE;
+#pragma unroll
+  for (int i = threadIdx.x; i < HEAD_SIZE; i += NUM_THREADS) {
+    float acc = 0.0f;
+    for (int j = 0; j < num_partitions; ++j) {
+      acc += to_float(tmp_out_ptr[j * HEAD_SIZE + i]) * shared_exp_sums[j] * inv_global_exp_sum;
+    }
+    from_float(out_ptr[i], acc);
+  }
+}
+
+} // namespace vllm
+
+#define LAUNCH_PAGED_ATTENTION_V1(HEAD_SIZE)                                                  \
+  VLLM_DevFuncAttribute_SET_MaxDynamicSharedMemorySize(                                       \
+    ((void*)vllm::paged_attention_v1_kernel<T, CACHE_T, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS,   \
+      IS_FP8_E5M2_KV_CACHE>), shared_mem_size);                                               \
+  vllm::paged_attention_v1_kernel<T, CACHE_T, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS,             \
+  IS_FP8_E5M2_KV_CACHE><<<grid, block, shared_mem_size, stream>>>(                            \
+    out_ptr,                                                                                  \
+    query_ptr,                                                                                \
+    key_cache_ptr,                                                                            \
+    value_cache_ptr,                                                                          \
+    num_kv_heads,                                                                             \
+    scale,                                                                                    \
+    block_tables_ptr,                                                                         \
+    context_lens_ptr,                                                                         \
+    max_num_blocks_per_seq,                                                                   \
+    alibi_slopes_ptr,                                                                         \
+    q_stride,                                                                                 \
+    kv_block_stride,                                                                          \
+    kv_head_stride);
+
+// TODO(woosuk): Tune NUM_THREADS.
+template<
+  typename T,
+  typename CACHE_T,
+  int BLOCK_SIZE,
+  bool IS_FP8_E5M2_KV_CACHE,
+  int NUM_THREADS = 128>
+void paged_attention_v1_launcher(
+  torch::Tensor& out,
+  torch::Tensor& query,
+  torch::Tensor& key_cache,
+  torch::Tensor& value_cache,
+  int num_kv_heads,
+  float scale,
+  torch::Tensor& block_tables,
+  torch::Tensor& context_lens,
+  int max_context_len,
+  const c10::optional<torch::Tensor>& alibi_slopes) {
+  int num_seqs = query.size(0);
+  int num_heads = query.size(1);
+  int head_size = query.size(2);
+  int max_num_blocks_per_seq = block_tables.size(1);
+  int q_stride = query.stride(0);
+  int kv_block_stride = key_cache.stride(0);
+  int kv_head_stride = key_cache.stride(1);
+
+  int thread_group_size = MAX(WARP_SIZE / BLOCK_SIZE, 1);
+  assert(head_size % thread_group_size == 0);
+
+  // NOTE: alibi_slopes is optional.
+  const float* alibi_slopes_ptr = alibi_slopes ?
+    reinterpret_cast<const float*>(alibi_slopes.value().data_ptr())
+    : nullptr;
+
+  T* out_ptr = reinterpret_cast<T*>(out.data_ptr());
+  T* query_ptr = reinterpret_cast<T*>(query.data_ptr());
+  CACHE_T* key_cache_ptr = reinterpret_cast<CACHE_T*>(key_cache.data_ptr());
+  CACHE_T* value_cache_ptr = reinterpret_cast<CACHE_T*>(value_cache.data_ptr());
+  int* block_tables_ptr = block_tables.data_ptr<int>();
+  int* context_lens_ptr = context_lens.data_ptr<int>();
+
+  constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
+  int padded_max_context_len = DIVIDE_ROUND_UP(max_context_len, BLOCK_SIZE) * BLOCK_SIZE;
+  int logits_size = padded_max_context_len * sizeof(float);
+  int outputs_size = (NUM_WARPS / 2) * head_size * sizeof(float);
+  // Python-side check in vllm.worker.worker._check_if_can_support_max_seq_len
+  // Keep that in sync with the logic here!
+  int shared_mem_size = std::max(logits_size, outputs_size);
+
+  dim3 grid(num_heads, num_seqs, 1);
+  dim3 block(NUM_THREADS);
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(query));
+  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+  switch (head_size) {
+    // NOTE(woosuk): To reduce the compilation time, we only compile for the
+    // head sizes that we use in the model. However, we can easily extend this
+    // to support any head size which is a multiple of 16.
+    case 64:
+      LAUNCH_PAGED_ATTENTION_V1(64);
+      break;
+    case 80:
+      LAUNCH_PAGED_ATTENTION_V1(80);
+      break;
+    case 96:
+      LAUNCH_PAGED_ATTENTION_V1(96);
+      break;
+    case 112:
+      LAUNCH_PAGED_ATTENTION_V1(112);
+      break;
+    case 128:
+      LAUNCH_PAGED_ATTENTION_V1(128);
+      break;
+    case 256:
+      LAUNCH_PAGED_ATTENTION_V1(256);
+      break;
+    default:
+      TORCH_CHECK(false, "Unsupported head size: ", head_size);
+      break;
+  }
+}
+
+#define CALL_V1_LAUNCHER(T, CACHE_T, BLOCK_SIZE, IS_FP8_E5M2_KV_CACHE)       \
+  paged_attention_v1_launcher<T, CACHE_T, BLOCK_SIZE, IS_FP8_E5M2_KV_CACHE>( \
+    out,                                                                     \
+    query,                                                                   \
+    key_cache,                                                               \
+    value_cache,                                                             \
+    num_kv_heads,                                                            \
+    scale,                                                                   \
+    block_tables,                                                            \
+    context_lens,                                                            \
+    max_context_len,                                                         \
+    alibi_slopes);
+
+// NOTE(woosuk): To reduce the compilation time, we omitted block sizes
+// 1, 2, 4, 64, 128, 256.
+#define CALL_V1_LAUNCHER_BLOCK_SIZE(T, CACHE_T, IS_FP8_E5M2_KV_CACHE) \
+  switch (block_size) {                                               \
+    case 8:                                                           \
+      CALL_V1_LAUNCHER(T, CACHE_T, 8, IS_FP8_E5M2_KV_CACHE);          \
+      break;                                                          \
+    case 16:                                                          \
+      CALL_V1_LAUNCHER(T, CACHE_T, 16, IS_FP8_E5M2_KV_CACHE);         \
+      break;                                                          \
+    case 32:                                                          \
+      CALL_V1_LAUNCHER(T, CACHE_T, 32, IS_FP8_E5M2_KV_CACHE);         \
+      break;                                                          \
+    default:                                                          \
+      TORCH_CHECK(false, "Unsupported block size: ", block_size);     \
+      break;                                                          \
+  }
+
+void paged_attention_v1(
+  torch::Tensor& out,             // [num_seqs, num_heads, head_size]
+  torch::Tensor& query,           // [num_seqs, num_heads, head_size]
+  torch::Tensor& key_cache,       // [num_blocks, num_heads, head_size/x, block_size, x]
+  torch::Tensor& value_cache,     // [num_blocks, num_heads, head_size, block_size]
+  int num_kv_heads,               // [num_heads]
+  float scale,
+  torch::Tensor& block_tables,    // [num_seqs, max_num_blocks_per_seq]
+  torch::Tensor& context_lens,    // [num_seqs]
+  int block_size,
+  int max_context_len,
+  const c10::optional<torch::Tensor>& alibi_slopes,
+  const std::string& kv_cache_dtype) {
+  if (kv_cache_dtype == "auto") {
+    if (query.dtype() == at::ScalarType::Float) {
+      CALL_V1_LAUNCHER_BLOCK_SIZE(float, float, false);
+    } else if (query.dtype() == at::ScalarType::Half) {
+      CALL_V1_LAUNCHER_BLOCK_SIZE(uint16_t, uint16_t, false);
+    } else if (query.dtype() == at::ScalarType::BFloat16) {
+      CALL_V1_LAUNCHER_BLOCK_SIZE(__nv_bfloat16, __nv_bfloat16, false);
+    } else {
+      TORCH_CHECK(false, "Unsupported data type: ", query.dtype());
+    }
+  } else if (kv_cache_dtype == "fp8_e5m2") {
+    if (query.dtype() == at::ScalarType::Float) {
+      CALL_V1_LAUNCHER_BLOCK_SIZE(float, uint8_t, true);
+    } else if (query.dtype() == at::ScalarType::Half) {
+      CALL_V1_LAUNCHER_BLOCK_SIZE(uint16_t, uint8_t, true);
+    } else if (query.dtype() == at::ScalarType::BFloat16) {
+      CALL_V1_LAUNCHER_BLOCK_SIZE(__nv_bfloat16, uint8_t, true);
+    } else {
+      TORCH_CHECK(false, "Unsupported data type: ", query.dtype());
+    }
+  } else {
+    TORCH_CHECK(false, "Unsupported data type of kv cache: ", kv_cache_dtype);
+  }
+}
+
+#define LAUNCH_PAGED_ATTENTION_V2(HEAD_SIZE)                                                  \
+  vllm::paged_attention_v2_kernel<T, CACHE_T, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS,             \
+  IS_FP8_E5M2_KV_CACHE, PARTITION_SIZE>                                                       \
+  <<<grid, block, shared_mem_size, stream>>>(                                                 \
+    exp_sums_ptr,                                                                             \
+    max_logits_ptr,                                                                           \
+    tmp_out_ptr,                                                                              \
+    query_ptr,                                                                                \
+    key_cache_ptr,                                                                            \
+    value_cache_ptr,                                                                          \
+    num_kv_heads,                                                                             \
+    scale,                                                                                    \
+    block_tables_ptr,                                                                         \
+    context_lens_ptr,                                                                         \
+    max_num_blocks_per_seq,                                                                   \
+    alibi_slopes_ptr,                                                                         \
+    q_stride,                                                                                 \
+    kv_block_stride,                                                                          \
+    kv_head_stride);                                                                          \
+  vllm::paged_attention_v2_reduce_kernel<T, HEAD_SIZE, NUM_THREADS, PARTITION_SIZE>           \
+  <<<reduce_grid, block, reduce_shared_mem_size, stream>>>(                                   \
+    out_ptr,                                                                                  \
+    exp_sums_ptr,                                                                             \
+    max_logits_ptr,                                                                           \
+    tmp_out_ptr,                                                                              \
+    context_lens_ptr,                                                                         \
+    max_num_partitions);
+
+template<
+  typename T,
+  typename CACHE_T,
+  int BLOCK_SIZE,
+  bool IS_FP8_E5M2_KV_CACHE,
+  int NUM_THREADS = 128,
+  int PARTITION_SIZE = 512>
+void paged_attention_v2_launcher(
+  torch::Tensor& out,
+  torch::Tensor& exp_sums,
+  torch::Tensor& max_logits,
+  torch::Tensor& tmp_out,
+  torch::Tensor& query,
+  torch::Tensor& key_cache,
+  torch::Tensor& value_cache,
+  int num_kv_heads,
+  float scale,
+  torch::Tensor& block_tables,
+  torch::Tensor& context_lens,
+  int max_context_len,
+  const c10::optional<torch::Tensor>& alibi_slopes) {
+  int num_seqs = query.size(0);
+  int num_heads = query.size(1);
+  int head_size = query.size(2);
+  int max_num_blocks_per_seq = block_tables.size(1);
+  int q_stride = query.stride(0);
+  int kv_block_stride = key_cache.stride(0);
+  int kv_head_stride = key_cache.stride(1);
+
+  int thread_group_size = MAX(WARP_SIZE / BLOCK_SIZE, 1);
+  assert(head_size % thread_group_size == 0);
+
+  // NOTE: alibi_slopes is optional.
+  const float* alibi_slopes_ptr = alibi_slopes ?
+    reinterpret_cast<const float*>(alibi_slopes.value().data_ptr())
+    : nullptr;
+
+  T* out_ptr = reinterpret_cast<T*>(out.data_ptr());
+  float* exp_sums_ptr = reinterpret_cast<float*>(exp_sums.data_ptr());
+  float* max_logits_ptr = reinterpret_cast<float*>(max_logits.data_ptr());
+  T* tmp_out_ptr = reinterpret_cast<T*>(tmp_out.data_ptr());
+  T* query_ptr = reinterpret_cast<T*>(query.data_ptr());
+  CACHE_T* key_cache_ptr = reinterpret_cast<CACHE_T*>(key_cache.data_ptr());
+  CACHE_T* value_cache_ptr = reinterpret_cast<CACHE_T*>(value_cache.data_ptr());
+  int* block_tables_ptr = block_tables.data_ptr<int>();
+  int* context_lens_ptr = context_lens.data_ptr<int>();
+
+  constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
+  int max_num_partitions = DIVIDE_ROUND_UP(max_context_len, PARTITION_SIZE);
+  int logits_size = PARTITION_SIZE * sizeof(float);
+  int outputs_size = (NUM_WARPS / 2) * head_size * sizeof(float);
+
+  // For paged attention v2 kernel.
+  dim3 grid(num_heads, num_seqs, max_num_partitions);
+  int shared_mem_size = std::max(logits_size, outputs_size);
+  // For paged attention v2 reduce kernel.
+  dim3 reduce_grid(num_heads, num_seqs);
+  int reduce_shared_mem_size = 2 * max_num_partitions * sizeof(float);
+
+  dim3 block(NUM_THREADS);
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(query));
+  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+  switch (head_size) {
+    // NOTE(woosuk): To reduce the compilation time, we only compile for the
+    // head sizes that we use in the model. However, we can easily extend this
+    // to support any head size which is a multiple of 16.
+    case 64:
+      LAUNCH_PAGED_ATTENTION_V2(64);
+      break;
+    case 80:
+      LAUNCH_PAGED_ATTENTION_V2(80);
+      break;
+    case 96:
+      LAUNCH_PAGED_ATTENTION_V2(96);
+      break;
+    case 112:
+      LAUNCH_PAGED_ATTENTION_V2(112);
+      break;
+    case 128:
+      LAUNCH_PAGED_ATTENTION_V2(128);
+      break;
+    case 256:
+      LAUNCH_PAGED_ATTENTION_V2(256);
+      break;
+    default:
+      TORCH_CHECK(false, "Unsupported head size: ", head_size);
+      break;
+  }
+}
+
+#define CALL_V2_LAUNCHER(T, CACHE_T, BLOCK_SIZE, IS_FP8_E5M2_KV_CACHE)           \
+  paged_attention_v2_launcher<T, CACHE_T, BLOCK_SIZE, IS_FP8_E5M2_KV_CACHE>(     \
+    out,                                                                         \
+    exp_sums,                                                                    \
+    max_logits,                                                                  \
+    tmp_out,                                                                     \
+    query,                                                                       \
+    key_cache,                                                                   \
+    value_cache,                                                                 \
+    num_kv_heads,                                                                \
+    scale,                                                                       \
+    block_tables,                                                                \
+    context_lens,                                                                \
+    max_context_len,                                                             \
+    alibi_slopes);
+
+// NOTE(woosuk): To reduce the compilation time, we omitted block sizes
+// 1, 2, 4, 64, 128, 256.
+#define CALL_V2_LAUNCHER_BLOCK_SIZE(T, CACHE_T, IS_FP8_E5M2_KV_CACHE)       \
+  switch (block_size) {                                                     \
+    case 8:                                                                 \
+      CALL_V2_LAUNCHER(T, CACHE_T, 8, IS_FP8_E5M2_KV_CACHE);                \
+      break;                                                                \
+    case 16:                                                                \
+      CALL_V2_LAUNCHER(T, CACHE_T, 16, IS_FP8_E5M2_KV_CACHE);               \
+      break;                                                                \
+    case 32:                                                                \
+      CALL_V2_LAUNCHER(T, CACHE_T, 32, IS_FP8_E5M2_KV_CACHE);               \
+      break;                                                                \
+    default:                                                                \
+      TORCH_CHECK(false, "Unsupported block size: ", block_size);           \
+      break;                                                                \
+  }
+
+void paged_attention_v2(
+  torch::Tensor& out,             // [num_seqs, num_heads, head_size]
+  torch::Tensor& exp_sums,        // [num_seqs, num_heads, max_num_partitions]
+  torch::Tensor& max_logits,      // [num_seqs, num_heads, max_num_partitions]
+  torch::Tensor& tmp_out,         // [num_seqs, num_heads, max_num_partitions, head_size]
+  torch::Tensor& query,           // [num_seqs, num_heads, head_size]
+  torch::Tensor& key_cache,       // [num_blocks, num_heads, head_size/x, block_size, x]
+  torch::Tensor& value_cache,     // [num_blocks, num_heads, head_size, block_size]
+  int num_kv_heads,               // [num_heads]
+  float scale,
+  torch::Tensor& block_tables,    // [num_seqs, max_num_blocks_per_seq]
+  torch::Tensor& context_lens,    // [num_seqs]
+  int block_size,
+  int max_context_len,
+  const c10::optional<torch::Tensor>& alibi_slopes,
+  const std::string& kv_cache_dtype) {
+  if (kv_cache_dtype == "auto") {
+    if (query.dtype() == at::ScalarType::Float) {
+      CALL_V2_LAUNCHER_BLOCK_SIZE(float, float, false);
+    } else if (query.dtype() == at::ScalarType::Half) {
+      CALL_V2_LAUNCHER_BLOCK_SIZE(uint16_t, uint16_t, false);
+    } else if (query.dtype() == at::ScalarType::BFloat16) {
+      CALL_V2_LAUNCHER_BLOCK_SIZE(__nv_bfloat16, __nv_bfloat16, false);
+    } else {
+      TORCH_CHECK(false, "Unsupported data type: ", query.dtype());
+    }
+  } else if (kv_cache_dtype == "fp8_e5m2") {
+    if (query.dtype() == at::ScalarType::Float) {
+      CALL_V2_LAUNCHER_BLOCK_SIZE(float, uint8_t, true);
+    } else if (query.dtype() == at::ScalarType::Half) {
+      CALL_V2_LAUNCHER_BLOCK_SIZE(uint16_t, uint8_t, true);
+    } else if (query.dtype() == at::ScalarType::BFloat16) {
+      CALL_V2_LAUNCHER_BLOCK_SIZE(__nv_bfloat16, uint8_t, true);
+    } else {
+      TORCH_CHECK(false, "Unsupported data type: ", query.dtype());
+    }
+  } else {
+    TORCH_CHECK(false, "Unsupported data type of kv cache: ", kv_cache_dtype);
+  }
+}
+
+#undef WARP_SIZE
+#undef MAX
+#undef MIN
+#undef DIVIDE_ROUND_UP

csrc/attention/attention_utils.cuh

@@ -0,0 +1,56 @@
+/*
+ * Adapted from https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/kernels/decoder_masked_multihead_attention/decoder_masked_multihead_attention_template.hpp
+ * Copyright (c) 2023, The vLLM team.
+ * Copyright (c) 2020-2023, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#pragma once
+
+#include "../cuda_compat.h"
+#include "attention_dtypes.h"
+
+#include <float.h>
+#include <type_traits>
+
+namespace vllm {
+
+// Q*K^T operation.
+template<int THREAD_GROUP_SIZE, typename Vec, int N>
+inline __device__ float qk_dot_(const Vec (&q)[N], const Vec (&k)[N]) {
+  using A_vec = typename FloatVec<Vec>::Type;
+  // Compute the parallel products for Q*K^T (treat vector lanes separately).
+  A_vec qk_vec = mul<A_vec, Vec, Vec>(q[0], k[0]);
+#pragma unroll
+  for (int ii = 1; ii < N; ++ii) {
+    qk_vec = fma(q[ii], k[ii], qk_vec);
+  }
+
+  // Finalize the reduction across lanes.
+  float qk = sum(qk_vec);
+#pragma unroll
+  for (int mask = THREAD_GROUP_SIZE / 2; mask >= 1; mask /= 2) {
+    qk += VLLM_SHFL_XOR_SYNC(qk, mask);
+  }
+  return qk;
+}
+
+template<typename T, int THREAD_GROUP_SIZE>
+struct Qk_dot {
+  template<typename Vec, int N>
+  static inline __device__ float dot(const Vec (&q)[N], const Vec (&k)[N]) {
+    return qk_dot_<THREAD_GROUP_SIZE>(q, k);
+  }
+};
+
+} // namespace vllm

csrc/attention/dtype_bfloat16.cuh

@@ -0,0 +1,451 @@
+/*
+ * Adapted from https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/kernels/decoder_masked_multihead_attention/decoder_masked_multihead_attention_template.hpp
+ * and https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/kernels/decoder_masked_multihead_attention_utils.h
+ * Copyright (c) 2023, The vLLM team.
+ * Copyright (c) 2020-2023, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#pragma once
+
+#include "attention_generic.cuh"
+#include "dtype_float32.cuh"
+
+#ifndef USE_ROCM
+  #include <cuda_bf16.h>
+  #include <cuda_fp16.h>
+#else
+  #include <hip/hip_bf16.h>
+  #include <hip/hip_fp16.h>
+
+  typedef __hip_bfloat162 __nv_bfloat162;
+  typedef __hip_bfloat16 __nv_bfloat16;
+#endif
+
+#include <stdint.h>
+
+namespace vllm {
+
+// Define custom BF16 vector data types.
+struct bf16_4_t {
+  __nv_bfloat162 x;
+  __nv_bfloat162 y;
+};
+
+struct bf16_8_t {
+  __nv_bfloat162 x;
+  __nv_bfloat162 y;
+  __nv_bfloat162 z;
+  __nv_bfloat162 w;
+};
+
+// BF16 vector types for Q, K, V.
+template<>
+struct Vec<__nv_bfloat16, 1> {
+  using Type = __nv_bfloat16;
+};
+template<>
+struct Vec<__nv_bfloat16, 2> {
+  using Type = __nv_bfloat162;
+};
+template<>
+struct Vec<__nv_bfloat16, 4> {
+  using Type = bf16_4_t;
+};
+template<>
+struct Vec<__nv_bfloat16, 8> {
+  using Type = bf16_8_t;
+};
+
+// FP32 accumulator vector types corresponding to Vec.
+template<>
+struct FloatVec<__nv_bfloat16> {
+  using Type = float;
+};
+template<>
+struct FloatVec<__nv_bfloat162> {
+  using Type = float2;
+};
+template<>
+struct FloatVec<bf16_4_t> {
+  using Type = Float4_;
+};
+template<>
+struct FloatVec<bf16_8_t> {
+  using Type = Float8_;
+};
+
+// Utility functions for type conversions.
+inline __device__ float2 bf1622float2(const __nv_bfloat162 val) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
+  assert(false);
+#else
+  return __bfloat1622float2(val);
+#endif
+}
+
+inline __device__ __nv_bfloat162 bf162bf162(const __nv_bfloat16 val) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
+  assert(false);
+#else
+  return __bfloat162bfloat162(val);
+#endif
+}
+
+// Vector addition.
+inline __device__ __nv_bfloat16 add(__nv_bfloat16 a, __nv_bfloat16 b) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
+  assert(false);
+#else
+  #ifndef USE_ROCM
+    return a + b;
+  #else
+    return __hadd(a, b);
+  #endif
+#endif
+}
+
+inline __device__ __nv_bfloat162 add(__nv_bfloat162 a, __nv_bfloat162 b) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
+  assert(false);
+#else
+  return __hadd2(a, b);
+#endif
+}
+
+inline __device__ bf16_4_t add(bf16_4_t a, bf16_4_t b) {
+  bf16_4_t c;
+  c.x = add(a.x, b.x);
+  c.y = add(a.y, b.y);
+  return c;
+}
+
+inline __device__ bf16_8_t add(bf16_8_t a, bf16_8_t b) {
+  bf16_8_t c;
+  c.x = add(a.x, b.x);
+  c.y = add(a.y, b.y);
+  c.z = add(a.z, b.z);
+  c.w = add(a.w, b.w);
+  return c;
+}
+
+inline __device__ float2 add(__nv_bfloat162 a, float2 fb) {
+  float2 fa = bf1622float2(a);
+  return add(fa, fb);
+}
+
+inline __device__ Float4_ add(bf16_4_t a, Float4_ fb) {
+  Float4_ fc;
+  fc.x = add(a.x, fb.x);
+  fc.y = add(a.y, fb.y);
+  return fc;
+}
+
+inline __device__ Float8_ add(bf16_8_t a, Float8_ fb) {
+  Float8_ fc;
+  fc.x = add(a.x, fb.x);
+  fc.y = add(a.y, fb.y);
+  fc.z = add(a.z, fb.z);
+  fc.w = add(a.w, fb.w);
+  return fc;
+}
+
+// Vector multiplication.
+template<>
+inline __device__ __nv_bfloat16 mul(__nv_bfloat16 a, __nv_bfloat16 b) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
+  assert(false);
+#else
+  return __hmul(a, b);
+#endif
+}
+
+template<>
+inline __device__ __nv_bfloat162 mul(__nv_bfloat162 a, __nv_bfloat162 b) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
+  assert(false);
+#else
+  return __hmul2(a, b);
+#endif
+}
+
+template<>
+inline __device__ __nv_bfloat162 mul(__nv_bfloat16 a, __nv_bfloat162 b) {
+  return mul<__nv_bfloat162, __nv_bfloat162, __nv_bfloat162>(bf162bf162(a), b);
+}
+
+template<>
+inline __device__ bf16_4_t mul(bf16_4_t a, bf16_4_t b) {
+  bf16_4_t c;
+  c.x = mul<__nv_bfloat162, __nv_bfloat162, __nv_bfloat162>(a.x, b.x);
+  c.y = mul<__nv_bfloat162, __nv_bfloat162, __nv_bfloat162>(a.y, b.y);
+  return c;
+}
+
+template<>
+inline __device__ bf16_4_t mul(__nv_bfloat16 a, bf16_4_t b) {
+  __nv_bfloat162 s = bf162bf162(a);
+  bf16_4_t c;
+  c.x = mul<__nv_bfloat162, __nv_bfloat162, __nv_bfloat162>(s, b.x);
+  c.y = mul<__nv_bfloat162, __nv_bfloat162, __nv_bfloat162>(s, b.y);
+  return c;
+}
+
+template<>
+inline __device__ bf16_8_t mul(bf16_8_t a, bf16_8_t b) {
+  bf16_8_t c;
+  c.x = mul<__nv_bfloat162, __nv_bfloat162, __nv_bfloat162>(a.x, b.x);
+  c.y = mul<__nv_bfloat162, __nv_bfloat162, __nv_bfloat162>(a.y, b.y);
+  c.z = mul<__nv_bfloat162, __nv_bfloat162, __nv_bfloat162>(a.z, b.z);
+  c.w = mul<__nv_bfloat162, __nv_bfloat162, __nv_bfloat162>(a.w, b.w);
+  return c;
+}
+
+template<>
+inline __device__ bf16_8_t mul(__nv_bfloat16 a, bf16_8_t b) {
+  __nv_bfloat162 s = bf162bf162(a);
+  bf16_8_t c;
+  c.x = mul<__nv_bfloat162, __nv_bfloat162, __nv_bfloat162>(s, b.x);
+  c.y = mul<__nv_bfloat162, __nv_bfloat162, __nv_bfloat162>(s, b.y);
+  c.z = mul<__nv_bfloat162, __nv_bfloat162, __nv_bfloat162>(s, b.z);
+  c.w = mul<__nv_bfloat162, __nv_bfloat162, __nv_bfloat162>(s, b.w);
+  return c;
+}
+
+template<>
+inline __device__ float mul(__nv_bfloat16 a, __nv_bfloat16 b) {
+  float fa = __bfloat162float(a);
+  float fb = __bfloat162float(b);
+  return fa * fb;
+}
+
+template<>
+inline __device__ float2 mul(__nv_bfloat162 a, __nv_bfloat162 b) {
+  float2 fa = bf1622float2(a);
+  float2 fb = bf1622float2(b);
+  return mul<float2, float2, float2>(fa, fb);
+}
+
+template<>
+inline __device__ float2 mul(__nv_bfloat16 a, __nv_bfloat162 b) {
+  return mul<float2, __nv_bfloat162, __nv_bfloat162>(bf162bf162(a), b);
+}
+
+template<>
+inline __device__ Float4_ mul(bf16_4_t a, bf16_4_t b) {
+  Float4_ fc;
+  fc.x = mul<float2, __nv_bfloat162, __nv_bfloat162>(a.x, b.x);
+  fc.y = mul<float2, __nv_bfloat162, __nv_bfloat162>(a.y, b.y);
+  return fc;
+}
+
+template<>
+inline __device__ Float4_ mul(__nv_bfloat16 a, bf16_4_t b) {
+  __nv_bfloat162 s = bf162bf162(a);
+  Float4_ fc;
+  fc.x = mul<float2, __nv_bfloat162, __nv_bfloat162>(s, b.x);
+  fc.y = mul<float2, __nv_bfloat162, __nv_bfloat162>(s, b.y);
+  return fc;
+}
+
+template<>
+inline __device__ Float8_ mul(bf16_8_t a, bf16_8_t b) {
+  Float8_ fc;
+  fc.x = mul<float2, __nv_bfloat162, __nv_bfloat162>(a.x, b.x);
+  fc.y = mul<float2, __nv_bfloat162, __nv_bfloat162>(a.y, b.y);
+  fc.z = mul<float2, __nv_bfloat162, __nv_bfloat162>(a.z, b.z);
+  fc.w = mul<float2, __nv_bfloat162, __nv_bfloat162>(a.w, b.w);
+  return fc;
+}
+
+template<>
+inline __device__ Float8_ mul(__nv_bfloat16 a, bf16_8_t b) {
+  __nv_bfloat162 s = bf162bf162(a);
+  Float8_ fc;
+  fc.x = mul<float2, __nv_bfloat162, __nv_bfloat162>(s, b.x);
+  fc.y = mul<float2, __nv_bfloat162, __nv_bfloat162>(s, b.y);
+  fc.z = mul<float2, __nv_bfloat162, __nv_bfloat162>(s, b.z);
+  fc.w = mul<float2, __nv_bfloat162, __nv_bfloat162>(s, b.w);
+  return fc;
+}
+
+// Vector fused multiply-add.
+inline __device__ __nv_bfloat162 fma(__nv_bfloat162 a, __nv_bfloat162 b, __nv_bfloat162 c) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
+  assert(false);
+#else
+  return __hfma2(a, b, c);
+#endif
+}
+
+inline __device__ __nv_bfloat162 fma(__nv_bfloat16 a, __nv_bfloat162 b, __nv_bfloat162 c) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
+  assert(false);
+#else
+  return __hfma2(bf162bf162(a), b, c);
+#endif
+}
+
+inline __device__ bf16_4_t fma(bf16_4_t a, bf16_4_t b, bf16_4_t c) {
+  bf16_4_t d;
+  d.x = fma(a.x, b.x, c.x);
+  d.y = fma(a.y, b.y, c.y);
+  return d;
+}
+
+inline __device__ bf16_4_t fma(__nv_bfloat16 a, bf16_4_t b, bf16_4_t c) {
+  __nv_bfloat162 s = bf162bf162(a);
+  bf16_4_t d;
+  d.x = fma(s, b.x, c.x);
+  d.y = fma(s, b.y, c.y);
+  return d;
+}
+
+inline __device__ bf16_8_t fma(bf16_8_t a, bf16_8_t b, bf16_8_t c) {
+  bf16_8_t d;
+  d.x = fma(a.x, b.x, c.x);
+  d.y = fma(a.y, b.y, c.y);
+  d.z = fma(a.z, b.z, c.z);
+  d.w = fma(a.w, b.w, c.w);
+  return d;
+}
+
+inline __device__ bf16_8_t fma(__nv_bfloat16 a, bf16_8_t b, bf16_8_t c) {
+  __nv_bfloat162 s = bf162bf162(a);
+  bf16_8_t d;
+  d.x = fma(s, b.x, c.x);
+  d.y = fma(s, b.y, c.y);
+  d.z = fma(s, b.z, c.z);
+  d.w = fma(s, b.w, c.w);
+  return d;
+}
+
+inline __device__ float fma(__nv_bfloat16 a, __nv_bfloat16 b, float fc) {
+  return __bfloat162float(a) * __bfloat162float(b) + fc;
+}
+
+inline __device__ float2 fma(__nv_bfloat162 a, __nv_bfloat162 b, float2 fc) {
+  float2 fa = bf1622float2(a);
+  float2 fb = bf1622float2(b);
+  return fma(fa, fb, fc);
+}
+
+inline __device__ float2 fma(__nv_bfloat16 a, __nv_bfloat162 b, float2 fc) {
+  return fma(bf162bf162(a), b, fc);
+}
+
+inline __device__ Float4_ fma(bf16_4_t a, bf16_4_t b, Float4_ fc) {
+  Float4_ fd;
+  fd.x = fma(a.x, b.x, fc.x);
+  fd.y = fma(a.y, b.y, fc.y);
+  return fd;
+}
+
+inline __device__ Float4_ fma(__nv_bfloat16 a, bf16_4_t b, Float4_ fc) {
+  __nv_bfloat162 s = bf162bf162(a);
+  Float4_ fd;
+  fd.x = fma(s, b.x, fc.x);
+  fd.y = fma(s, b.y, fc.y);
+  return fd;
+}
+
+inline __device__ Float8_ fma(bf16_8_t a, bf16_8_t b, Float8_ fc) {
+  Float8_ fd;
+  fd.x = fma(a.x, b.x, fc.x);
+  fd.y = fma(a.y, b.y, fc.y);
+  fd.z = fma(a.z, b.z, fc.z);
+  fd.w = fma(a.w, b.w, fc.w);
+  return fd;
+}
+
+inline __device__ Float8_ fma(__nv_bfloat16 a, bf16_8_t b, Float8_ fc) {
+  __nv_bfloat162 s = bf162bf162(a);
+  Float8_ fd;
+  fd.x = fma(s, b.x, fc.x);
+  fd.y = fma(s, b.y, fc.y);
+  fd.z = fma(s, b.z, fc.z);
+  fd.w = fma(s, b.w, fc.w);
+  return fd;
+}
+
+// Vector sum.
+template<>
+inline __device__ float sum(__nv_bfloat16 v) {
+  return __bfloat162float(v);
+}
+
+template<>
+inline __device__ float sum(__nv_bfloat162 v) {
+  float2 vf = bf1622float2(v);
+  return vf.x + vf.y;
+}
+
+template<>
+inline __device__ float sum(bf16_4_t v) {
+  return sum(v.x) + sum(v.y);
+}
+
+template<>
+inline __device__ float sum(bf16_8_t v) {
+  return sum(v.x) + sum(v.y) + sum(v.z) + sum(v.w);
+}
+
+// From float32 to bfloat16.
+inline __device__ void from_float(__nv_bfloat16& dst, float src) {
+  dst = __float2bfloat16(src);
+}
+
+inline __device__ void from_float(__nv_bfloat162& dst, float2 src) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
+  assert(false);
+#else
+  dst = __float22bfloat162_rn(src);
+#endif
+}
+
+inline __device__ void from_float(bf16_4_t& dst, Float4_ src) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
+  assert(false);
+#else
+  dst.x = __float22bfloat162_rn(src.x);
+  dst.y = __float22bfloat162_rn(src.y);
+#endif
+}
+
+inline __device__ void from_float(bf16_8_t& dst, Float8_ src) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
+  assert(false);
+#else
+  dst.x = __float22bfloat162_rn(src.x);
+  dst.y = __float22bfloat162_rn(src.y);
+  dst.z = __float22bfloat162_rn(src.z);
+  dst.w = __float22bfloat162_rn(src.w);
+#endif
+}
+
+// From bfloat16 to float32.
+inline __device__ float to_float(__nv_bfloat16 u) {
+  return __bfloat162float(u);
+}
+
+// Zero-out a variable.
+inline __device__ void zero(__nv_bfloat16& dst) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
+  assert(false);
+#else
+  // Same as CUDART_ZERO_BF16 introduced in CUDA 12.2.
+  dst = __ushort_as_bfloat16((unsigned short)0x0000U);
+#endif
+}
+
+} // namespace vllm

csrc/attention/dtype_float16.cuh

@@ -0,0 +1,502 @@
+/*
+ * Adapted from https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/kernels/decoder_masked_multihead_attention/decoder_masked_multihead_attention_template.hpp
+ * and https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/kernels/decoder_masked_multihead_attention_utils.h
+ * Copyright (c) 2023, The vLLM team.
+ * Copyright (c) 2020-2023, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#pragma once
+
+#include "attention_generic.cuh"
+#include "dtype_float32.cuh"
+
+#ifdef USE_ROCM
+  #include <hip/hip_fp16.h>
+#endif
+
+#include <stdint.h>
+
+namespace vllm {
+
+// FP16 vector types for Q, K, V.
+template<>
+struct Vec<uint16_t, 1> {
+  using Type = uint16_t;
+};
+template<>
+struct Vec<uint16_t, 2> {
+  using Type = uint32_t;
+};
+template<>
+struct Vec<uint16_t, 4> {
+  using Type = uint2;
+};
+template<>
+struct Vec<uint16_t, 8> {
+  using Type = uint4;
+};
+
+// FP32 accumulator vector types corresponding to Vec.
+template<>
+struct FloatVec<uint16_t> {
+  using Type = float;
+};
+template<>
+struct FloatVec<uint32_t> {
+  using Type = float2;
+};
+template<>
+struct FloatVec<uint2> {
+  using Type = Float4_;
+};
+template<>
+struct FloatVec<uint4> {
+  using Type = Float8_;
+};
+
+// Utility functions for type conversions.
+inline __device__ uint32_t h0_h0(uint16_t a) {
+#ifndef USE_ROCM
+  uint32_t b;
+  asm volatile("mov.b32 %0, {%1, %1};" : "=r"(b) : "h"(a));
+  return b;
+#else
+  union {
+   uint32_t u32;
+   uint16_t u16[2];
+  } tmp;
+  tmp.u16[0] = a;
+  tmp.u16[1] = a;
+  return tmp.u32;
+#endif
+}
+
+inline __device__ float half_to_float(uint16_t h) {
+  float f;
+#ifndef USE_ROCM
+  asm volatile("cvt.f32.f16 %0, %1;\n" : "=f"(f) : "h"(h));
+#else
+  asm volatile("v_cvt_f32_f16 %0, %1;" : "=v"(f) : "v"(h));
+#endif
+  return f;
+}
+
+inline __device__ float2 half2_to_float2(uint32_t v) {
+#ifndef USE_ROCM
+  uint16_t lo, hi;
+  asm volatile("mov.b32 {%0, %1}, %2;\n" : "=h"(lo), "=h"(hi) : "r"(v));
+  return make_float2(half_to_float(lo), half_to_float(hi));
+#else
+  union {
+    uint32_t u32;
+    uint16_t u16[2];
+  } tmp;
+  tmp.u32 = v;
+  float2 ret;
+  ret.x = half_to_float(tmp.u16[0]);
+  ret.y = half_to_float(tmp.u16[1]);
+  return ret;
+#endif
+}
+
+inline __device__ uint16_t float_to_half(float f) {
+  union {
+    uint32_t u32;
+    uint16_t u16[2];
+  } tmp;
+#ifndef USE_ROCM
+  asm volatile("cvt.rn.f16.f32 %0, %1;\n" : "=h"(tmp.u16[0]) : "f"(f));
+#else
+  asm volatile("v_cvt_f16_f32 %0, %1;\n" : "=v"(tmp.u32) : "v"(f));
+#endif
+  return tmp.u16[0];
+}
+
+inline __device__ uint32_t float2_to_half2(float2 f) {
+  union {
+    uint32_t u32;
+    uint16_t u16[2];
+  } tmp;
+#ifndef USE_ROCM
+  #if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800
+    asm volatile("cvt.rn.f16x2.f32 %0, %1, %2;\n" : "=r"(tmp.u32) : "f"(f.y), "f"(f.x));
+  #else
+    asm volatile("cvt.rn.f16.f32 %0, %1;\n" : "=h"(tmp.u16[0]) : "f"(f.x));
+    asm volatile("cvt.rn.f16.f32 %0, %1;\n" : "=h"(tmp.u16[1]) : "f"(f.y));
+  #endif
+#else
+  tmp.u16[0] = float_to_half(f.x);
+  tmp.u16[1] = float_to_half(f.y);
+#endif
+  return tmp.u32;
+}
+
+// Vector addition.
+inline __device__ uint16_t add(uint16_t a, uint16_t b) {
+  uint16_t c;
+#ifndef USE_ROCM
+  asm volatile("add.f16 %0, %1, %2;\n" : "=h"(c) : "h"(a), "h"(b));
+#else
+  asm volatile("v_add_f16 %0, %1, %2;\n" : "=v"(c) : "v"(a), "v"(b));
+#endif
+  return c;
+}
+
+inline __device__ uint32_t add(uint32_t a, uint32_t b) {
+  uint32_t c;
+#ifndef USE_ROCM
+  asm volatile("add.f16x2 %0, %1, %2;\n" : "=r"(c) : "r"(a), "r"(b));
+#else
+  asm volatile("v_pk_add_f16 %0, %1, %2;\n" : "=v"(c) : "v"(a), "v"(b));
+#endif
+  return c;
+}
+
+inline __device__ uint2 add(uint2 a, uint2 b) {
+  uint2 c;
+  c.x = add(a.x, b.x);
+  c.y = add(a.y, b.y);
+  return c;
+}
+
+inline __device__ uint4 add(uint4 a, uint4 b) {
+  uint4 c;
+  c.x = add(a.x, b.x);
+  c.y = add(a.y, b.y);
+  c.z = add(a.z, b.z);
+  c.w = add(a.w, b.w);
+  return c;
+}
+
+inline __device__ float2 add(uint32_t a, float2 fb) {
+  float2 fa = half2_to_float2(a);
+  return add(fa, fb);
+}
+
+inline __device__ Float4_ add(uint2 a, Float4_ fb) {
+  Float4_ fc;
+  fc.x = add(a.x, fb.x);
+  fc.y = add(a.y, fb.y);
+  return fc;
+}
+
+inline __device__ Float8_ add(uint4 a, Float8_ fb) {
+  Float8_ fc;
+  fc.x = add(a.x, fb.x);
+  fc.y = add(a.y, fb.y);
+  fc.z = add(a.z, fb.z);
+  fc.w = add(a.w, fb.w);
+  return fc;
+}
+
+// Vector multiplication.
+template<>
+inline __device__ uint16_t mul(uint16_t a, uint16_t b) {
+  uint16_t c;
+#ifndef USE_ROCM
+  asm volatile("mul.f16 %0, %1, %2;\n" : "=h"(c) : "h"(a), "h"(b));
+#else
+  asm volatile("v_mul_f16 %0, %1, %2;\n" : "=v"(c) : "v"(a), "v"(b));
+#endif
+  return c;
+}
+
+template<>
+inline __device__ uint32_t mul(uint32_t a, uint32_t b) {
+  uint32_t c;
+#ifndef USE_ROCM
+  asm volatile("mul.f16x2 %0, %1, %2;\n" : "=r"(c) : "r"(a), "r"(b));
+#else
+  asm volatile("v_pk_mul_f16 %0, %1, %2;\n" : "=v"(c) : "v"(a), "v"(b));
+#endif
+  return c;
+}
+
+template<>
+inline __device__ uint32_t mul(uint16_t a, uint32_t b) {
+  return mul<uint32_t, uint32_t, uint32_t>(h0_h0(a), b);
+}
+
+template<>
+inline __device__ uint2 mul(uint2 a, uint2 b) {
+  uint2 c;
+  c.x = mul<uint32_t, uint32_t, uint32_t>(a.x, b.x);
+  c.y = mul<uint32_t, uint32_t, uint32_t>(a.y, b.y);
+  return c;
+}
+
+template<>
+inline __device__ uint2 mul(uint16_t a, uint2 b) {
+  uint32_t s = h0_h0(a);
+  uint2 c;
+  c.x = mul<uint32_t, uint32_t, uint32_t>(s, b.x);
+  c.y = mul<uint32_t, uint32_t, uint32_t>(s, b.y);
+  return c;
+}
+
+template<>
+inline __device__ uint4 mul(uint4 a, uint4 b) {
+  uint4 c;
+  c.x = mul<uint32_t, uint32_t, uint32_t>(a.x, b.x);
+  c.y = mul<uint32_t, uint32_t, uint32_t>(a.y, b.y);
+  c.z = mul<uint32_t, uint32_t, uint32_t>(a.z, b.z);
+  c.w = mul<uint32_t, uint32_t, uint32_t>(a.w, b.w);
+  return c;
+}
+
+template<>
+inline __device__ uint4 mul(uint16_t a, uint4 b) {
+  uint32_t s = h0_h0(a);
+  uint4 c;
+  c.x = mul<uint32_t, uint32_t, uint32_t>(s, b.x);
+  c.y = mul<uint32_t, uint32_t, uint32_t>(s, b.y);
+  c.z = mul<uint32_t, uint32_t, uint32_t>(s, b.z);
+  c.w = mul<uint32_t, uint32_t, uint32_t>(s, b.w);
+  return c;
+}
+
+template<>
+inline __device__ float mul(uint16_t a, uint16_t b) {
+  float fa = half_to_float(a);
+  float fb = half_to_float(b);
+  return fa * fb;
+}
+
+template<>
+inline __device__ float2 mul(uint32_t a, uint32_t b) {
+  float2 fa = half2_to_float2(a);
+  float2 fb = half2_to_float2(b);
+  return mul<float2, float2, float2>(fa, fb);
+}
+
+template<>
+inline __device__ float2 mul(uint16_t a, uint32_t b) {
+  return mul<float2, uint32_t, uint32_t>(h0_h0(a), b);
+}
+
+template<>
+inline __device__ Float4_ mul(uint2 a, uint2 b) {
+  Float4_ fc;
+  fc.x = mul<float2, uint32_t, uint32_t>(a.x, b.x);
+  fc.y = mul<float2, uint32_t, uint32_t>(a.y, b.y);
+  return fc;
+}
+
+template<>
+inline __device__ Float4_ mul(uint16_t a, uint2 b) {
+  uint32_t s = h0_h0(a);
+  Float4_ fc;
+  fc.x = mul<float2, uint32_t, uint32_t>(s, b.x);
+  fc.y = mul<float2, uint32_t, uint32_t>(s, b.y);
+  return fc;
+}
+
+template<>
+inline __device__ Float8_ mul(uint4 a, uint4 b) {
+  Float8_ fc;
+  fc.x = mul<float2, uint32_t, uint32_t>(a.x, b.x);
+  fc.y = mul<float2, uint32_t, uint32_t>(a.y, b.y);
+  fc.z = mul<float2, uint32_t, uint32_t>(a.z, b.z);
+  fc.w = mul<float2, uint32_t, uint32_t>(a.w, b.w);
+  return fc;
+}
+
+template<>
+inline __device__ Float8_ mul(uint16_t a, uint4 b) {
+  uint32_t s = h0_h0(a);
+  Float8_ fc;
+  fc.x = mul<float2, uint32_t, uint32_t>(s, b.x);
+  fc.y = mul<float2, uint32_t, uint32_t>(s, b.y);
+  fc.z = mul<float2, uint32_t, uint32_t>(s, b.z);
+  fc.w = mul<float2, uint32_t, uint32_t>(s, b.w);
+  return fc;
+}
+
+// Vector fused multiply-add.
+inline __device__ uint32_t fma(uint32_t a, uint32_t b, uint32_t c) {
+  uint32_t d;
+#ifndef USE_ROCM
+  asm volatile("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(d) : "r"(a), "r"(b), "r"(c));
+#else
+  asm volatile("v_pk_fma_f16 %0, %1, %2, %3;\n" : "=v"(d) : "v"(a), "v"(b), "v"(c));
+#endif
+  return d;
+}
+
+inline __device__ uint32_t fma(uint16_t a, uint32_t b, uint32_t c) {
+  return fma(h0_h0(a), b, c);
+}
+
+inline __device__ uint2 fma(uint2 a, uint2 b, uint2 c) {
+  uint2 d;
+  d.x = fma(a.x, b.x, c.x);
+  d.y = fma(a.y, b.y, c.y);
+  return d;
+}
+
+inline __device__ uint2 fma(uint16_t a, uint2 b, uint2 c) {
+  uint32_t s = h0_h0(a);
+  uint2 d;
+  d.x = fma(s, b.x, c.x);
+  d.y = fma(s, b.y, c.y);
+  return d;
+}
+
+inline __device__ uint4 fma(uint4 a, uint4 b, uint4 c) {
+  uint4 d;
+  d.x = fma(a.x, b.x, c.x);
+  d.y = fma(a.y, b.y, c.y);
+  d.z = fma(a.z, b.z, c.z);
+  d.w = fma(a.w, b.w, c.w);
+  return d;
+}
+
+inline __device__ uint4 fma(uint16_t a, uint4 b, uint4 c) {
+  uint32_t s = h0_h0(a);
+  uint4 d;
+  d.x = fma(s, b.x, c.x);
+  d.y = fma(s, b.y, c.y);
+  d.z = fma(s, b.z, c.z);
+  d.w = fma(s, b.w, c.w);
+  return d;
+}
+
+inline __device__ float fma(uint16_t a, uint16_t b, float fc) {
+  float fa = half_to_float(a);
+  float fb = half_to_float(b);
+  return fa * fb + fc;
+}
+
+inline __device__ float2 fma(uint32_t a, uint32_t b, float2 fc) {
+  float2 fa = half2_to_float2(a);
+  float2 fb = half2_to_float2(b);
+  return fma(fa, fb, fc);
+}
+
+inline __device__ float2 fma(uint16_t a, uint32_t b, float2 fc) {
+  return fma(h0_h0(a), b, fc);
+}
+
+inline __device__ Float4_ fma(uint2 a, uint2 b, Float4_ fc) {
+  Float4_ fd;
+  fd.x = fma(a.x, b.x, fc.x);
+  fd.y = fma(a.y, b.y, fc.y);
+  return fd;
+}
+
+inline __device__ Float4_ fma(uint16_t a, uint2 b, Float4_ fc) {
+  uint32_t s = h0_h0(a);
+  Float4_ fd;
+  fd.x = fma(s, b.x, fc.x);
+  fd.y = fma(s, b.y, fc.y);
+  return fd;
+}
+
+inline __device__ Float8_ fma(uint4 a, uint4 b, Float8_ fc) {
+  Float8_ fd;
+  fd.x = fma(a.x, b.x, fc.x);
+  fd.y = fma(a.y, b.y, fc.y);
+  fd.z = fma(a.z, b.z, fc.z);
+  fd.w = fma(a.w, b.w, fc.w);
+  return fd;
+}
+
+inline __device__ Float8_ fma(uint16_t a, uint4 b, Float8_ fc) {
+  uint32_t s = h0_h0(a);
+  Float8_ fd;
+  fd.x = fma(s, b.x, fc.x);
+  fd.y = fma(s, b.y, fc.y);
+  fd.z = fma(s, b.z, fc.z);
+  fd.w = fma(s, b.w, fc.w);
+  return fd;
+}
+
+// Vector sum.
+template<>
+inline __device__ float sum(uint16_t v) {
+  return half_to_float(v);
+}
+
+template<>
+inline __device__ float sum(uint32_t v) {
+  float2 tmp = half2_to_float2(v);
+  return tmp.x + tmp.y;
+}
+
+template<>
+inline __device__ float sum(uint2 v) {
+  uint32_t c = add(v.x, v.y);
+  return sum(c);
+}
+
+template<>
+inline __device__ float sum(uint4 v) {
+  uint32_t c = add(v.x, v.y);
+  c = add(c, v.z);
+  c = add(c, v.w);
+  return sum(c);
+}
+
+// From float32 to float16.
+inline __device__ void from_float(uint16_t& dst, float src) {
+  dst = float_to_half(src);
+}
+
+inline __device__ void from_float(uint32_t& dst, float2 src) {
+  dst = float2_to_half2(src);
+}
+
+inline __device__ void from_float(uint2& dst, Float4_ src) {
+  dst.x = float2_to_half2(src.x);
+  dst.y = float2_to_half2(src.y);
+}
+
+inline __device__ void from_float(uint4& dst, Float8_ src) {
+  dst.x = float2_to_half2(src.x);
+  dst.y = float2_to_half2(src.y);
+  dst.z = float2_to_half2(src.z);
+  dst.w = float2_to_half2(src.w);
+}
+
+// From float16 to float32.
+inline __device__ float to_float(uint16_t u) {
+  return half_to_float(u);
+}
+
+inline __device__ float2 to_float(uint32_t u) {
+  return half2_to_float2(u);
+}
+
+inline __device__ Float4_ to_float(uint2 u) {
+  Float4_ tmp;
+  tmp.x = half2_to_float2(u.x);
+  tmp.y = half2_to_float2(u.y);
+  return tmp;
+}
+
+inline __device__ Float8_ to_float(uint4 u) {
+  Float8_ tmp;
+  tmp.x = half2_to_float2(u.x);
+  tmp.y = half2_to_float2(u.y);
+  tmp.z = half2_to_float2(u.z);
+  tmp.w = half2_to_float2(u.w);
+  return tmp;
+}
+
+// Zero-out a variable.
+inline __device__ void zero(uint16_t& dst) {
+  dst = uint16_t(0);
+}
+
+} // namespace vllm

csrc/attention/dtype_float32.cuh

@@ -0,0 +1,273 @@
+/*
+ * Adapted from https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/kernels/decoder_masked_multihead_attention/decoder_masked_multihead_attention_template.hpp
+ * and https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/kernels/decoder_masked_multihead_attention_utils.h
+ * Copyright (c) 2023, The vLLM team.
+ * Copyright (c) 2020-2023, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#pragma once
+
+#include "attention_generic.cuh"
+
+#include <stdint.h>
+
+namespace vllm {
+
+// Define custom FP32 vector data types.
+struct Float4_ {
+  float2 x;
+  float2 y;
+};
+
+struct Float8_ {
+  float2 x;
+  float2 y;
+  float2 z;
+  float2 w;
+};
+
+// FP32 vector types for Q, K, V.
+template<>
+struct Vec<float, 1> {
+  using Type = float;
+};
+template<>
+struct Vec<float, 2> {
+  using Type = float2;
+};
+template<>
+struct Vec<float, 4> {
+  using Type = float4;
+};
+
+// FP32 accumulator vector types corresponding to Vec.
+template<>
+struct FloatVec<float> {
+  using Type = float;
+};
+template<>
+struct FloatVec<float2> {
+  using Type = float2;
+};
+template<>
+struct FloatVec<float4> {
+  using Type = float4;
+};
+
+// Vector addition.
+inline __device__ float add(float a, float b) {
+  return a + b;
+}
+
+inline __device__ float2 add(float2 a, float2 b) {
+  float2 c;
+  c.x = add(a.x, b.x);
+  c.y = add(a.y, b.y);
+  return c;
+}
+
+inline __device__ float4 add(float4 a, float4 b) {
+  float4 c;
+  c.x = add(a.x, b.x);
+  c.y = add(a.y, b.y);
+  c.z = add(a.z, b.z);
+  c.w = add(a.w, b.w);
+  return c;
+}
+
+// Vector multiplication.
+template<>
+inline __device__ float mul<float, float>(float a, float b) {
+  return a * b;
+}
+
+template<>
+inline __device__ float2 mul(float2 a, float2 b) {
+  float2 c;
+  c.x = a.x * b.x;
+  c.y = a.y * b.y;
+  return c;
+}
+
+template<>
+inline __device__ float2 mul(float a, float2 b) {
+  float2 c;
+  c.x = a * b.x;
+  c.y = a * b.y;
+  return c;
+}
+
+template<>
+inline __device__ float4 mul(float4 a, float4 b) {
+  float4 c;
+  c.x = a.x * b.x;
+  c.y = a.y * b.y;
+  c.z = a.z * b.z;
+  c.w = a.w * b.w;
+  return c;
+}
+
+template<>
+inline __device__ float4 mul(float a, float4 b) {
+  float4 c;
+  c.x = a * b.x;
+  c.y = a * b.y;
+  c.z = a * b.z;
+  c.w = a * b.w;
+  return c;
+}
+
+// Vector fused multiply-add.
+inline __device__ float fma(float a, float b, float c) {
+  return a * b + c;
+}
+
+inline __device__ float2 fma(float2 a, float2 b, float2 c) {
+  float2 d;
+  d.x = fma(a.x, b.x, c.x);
+  d.y = fma(a.y, b.y, c.y);
+  return d;
+}
+
+inline __device__ float2 fma(float a, float2 b, float2 c) {
+  float2 d;
+  d.x = fma(a, b.x, c.x);
+  d.y = fma(a, b.y, c.y);
+  return d;
+}
+
+inline __device__ float4 fma(float4 a, float4 b, float4 c) {
+  float4 d;
+  d.x = fma(a.x, b.x, c.x);
+  d.y = fma(a.y, b.y, c.y);
+  d.z = fma(a.z, b.z, c.z);
+  d.w = fma(a.w, b.w, c.w);
+  return d;
+}
+
+inline __device__ float4 fma(float a, float4 b, float4 c) {
+  float4 d;
+  d.x = fma(a, b.x, c.x);
+  d.y = fma(a, b.y, c.y);
+  d.z = fma(a, b.z, c.z);
+  d.w = fma(a, b.w, c.w);
+  return d;
+}
+
+inline __device__ Float4_ fma(float a, Float4_ b, Float4_ c) {
+  Float4_ d;
+  d.x = fma(a, b.x, c.x);
+  d.y = fma(a, b.y, c.y);
+  return d;
+}
+
+inline __device__ Float8_ fma(float a, Float8_ b, Float8_ c) {
+  Float8_ d;
+  d.x = fma(a, b.x, c.x);
+  d.y = fma(a, b.y, c.y);
+  d.z = fma(a, b.z, c.z);
+  d.w = fma(a, b.w, c.w);
+  return d;
+}
+
+// Vector sum.
+template<>
+inline __device__ float sum(float v) {
+  return v;
+}
+
+template<>
+inline __device__ float sum(float2 v) {
+  return v.x + v.y;
+}
+
+template<>
+inline __device__ float sum(float4 v) {
+  return v.x + v.y + v.z + v.w;
+}
+
+template<>
+inline __device__ float sum(Float4_ v) {
+  return v.x.x + v.x.y + v.y.x + v.y.y;
+}
+
+template<>
+inline __device__ float sum(Float8_ v) {
+  return v.x.x + v.x.y + v.y.x + v.y.y + v.z.x + v.z.y + v.w.x + v.w.y;
+}
+
+// Vector dot product.
+inline __device__ float dot(float a, float b) {
+  return a * b;
+}
+
+inline __device__ float dot(float2 a, float2 b) {
+  float2 c = mul<float2, float2, float2>(a, b);
+  return c.x + c.y;
+}
+
+inline __device__ float dot(Float4_ a, Float4_ b) {
+  float2 acc = mul<float2, float2, float2>(a.x, b.x);
+  acc = fma(a.y, b.y, acc);
+  return acc.x + acc.y;
+}
+
+inline __device__ float dot(Float8_ a, Float8_ b) {
+  float2 acc = mul<float2, float2, float2>(a.x, b.x);
+  acc = fma(a.y, b.y, acc);
+  acc = fma(a.z, b.z, acc);
+  acc = fma(a.w, b.w, acc);
+  return acc.x + acc.y;
+}
+
+// From float to float.
+inline __device__ void from_float(float& dst, float src) {
+  dst = src;
+}
+
+inline __device__ void from_float(float2& dst, float2 src) {
+  dst = src;
+}
+
+inline __device__ void from_float(float4& dst, float4 src) {
+  dst = src;
+}
+
+// From float to float.
+inline __device__ float to_float(float u) {
+  return u;
+}
+
+inline __device__ float2 to_float(float2 u) {
+  return u;
+}
+
+inline __device__ float4 to_float(float4 u) {
+  return u;
+}
+
+inline __device__ Float4_ to_float(Float4_ u) {
+  return u;
+}
+
+inline __device__ Float8_ to_float(Float8_ u) {
+  return u;
+}
+
+// Zero-out a variable.
+inline __device__ void zero(float& dst) {
+  dst = 0.f;
+}
+
+} // namespace vllm

csrc/attention/dtype_fp8_e5m2.cuh

@@ -0,0 +1,35 @@
+#pragma once
+
+#include "attention_generic.cuh"
+
+#include <stdint.h>
+#ifdef ENABLE_FP8_E5M2
+#include <cuda_fp8.h>
+#endif
+
+namespace vllm {
+#ifdef ENABLE_FP8_E5M2
+// fp8 vector types for quantization of kv cache
+
+template<>
+struct Vec<uint8_t, 1> {
+    using Type = uint8_t;
+};
+
+template<>
+struct Vec<uint8_t, 2> {
+    using Type = uint16_t;
+};
+
+template<>
+struct Vec<uint8_t, 4> {
+    using Type = uint32_t;
+};
+
+template<>
+struct Vec<uint8_t, 8> {
+    using Type = uint2;
+};
+#endif // ENABLE_FP8_E5M2
+
+} // namespace vllm

csrc/attention_kernels.cu

@@ -1,896 +0,0 @@
-#include <torch/extension.h>
-#include <ATen/cuda/CUDAContext.h>
-
-#include "attention_utils.h"
-#include "cuda_primitives.h"
-#include "reduction_utils.h"
-
-#include <algorithm>
-
-#define WARP_SIZE 32
-#define MAX(a, b) ((a) > (b) ? (a) : (b))
-#define MIN(a, b) ((a) < (b) ? (a) : (b))
-
-namespace cacheflow {
-
-// Grid: (num_heads, num_seqs).
-template<
-  typename scalar_t,
-  int HEAD_SIZE,
-  int BLOCK_SIZE,
-  int NUM_THREADS>
-__global__ void single_query_cached_kv_attention_kernel(
-  scalar_t* __restrict__ out,             // [num_seqs, num_heads, head_size]
-  const scalar_t* __restrict__ q,         // [num_seqs, num_heads, head_size]
-  const scalar_t* __restrict__ k_cache,   // [num_blocks, num_heads, head_size/x, block_size, x]
-  const scalar_t* __restrict__ v_cache,   // [num_blocks, num_heads, head_size, block_size]
-  const float scale,
-  const int* __restrict__ block_tables,   // [num_seqs, max_num_blocks_per_seq]
-  const int* __restrict__ context_lens,   // [num_seqs]
-  const int max_num_blocks_per_seq,
-  const int q_stride) {
-  constexpr int THREAD_GROUP_SIZE = MAX(WARP_SIZE / BLOCK_SIZE, 1);
-  constexpr int NUM_TOKENS_PER_THREAD_GROUP = (BLOCK_SIZE + WARP_SIZE - 1) / WARP_SIZE;
-  constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
-  const int thread_idx = threadIdx.x;
-  const int warp_idx = thread_idx / WARP_SIZE;
-  const int lane = thread_idx % WARP_SIZE;
-
-  const int head_idx = blockIdx.x;
-  const int num_heads = gridDim.x;
-  const int seq_idx = blockIdx.y;
-
-  // A vector type to store a part of a key or a query.
-  // The vector size is configured in such a way that the threads in a thread group
-  // fetch or compute 16 bytes at a time.
-  // For example, if the size of a thread group is 4 and the data type is half,
-  // then the vector size is 16 / (4 * sizeof(half)) == 2.
-  constexpr int VEC_SIZE = MAX(16 / (THREAD_GROUP_SIZE * sizeof(scalar_t)), 1);
-  using K_vec = typename Vec<scalar_t, VEC_SIZE>::Type;
-  using Q_vec = typename Vec<scalar_t, VEC_SIZE>::Type;
-
-  constexpr int NUM_ELEMS_PER_THREAD = HEAD_SIZE / THREAD_GROUP_SIZE;
-  constexpr int NUM_VECS_PER_THREAD = NUM_ELEMS_PER_THREAD / VEC_SIZE;
-
-  const int thread_group_idx = thread_idx / THREAD_GROUP_SIZE;
-  const int thread_group_offset = thread_idx % THREAD_GROUP_SIZE;
-
-  // Load the query to registers.
-  // Each thread in a thread group has a different part of the query.
-  // For example, if the the thread group size is 4, then the first thread in the group
-  // has 0, 4, 8, ... th vectors of the query, and the second thread has 1, 5, 9, ...
-  // th vectors of the query, and so on.
-  // NOTE(woosuk): Because q is split from a qkv tensor, it may not be contiguous.
-  const scalar_t* q_ptr = q + seq_idx * q_stride + head_idx * HEAD_SIZE;
-  Q_vec q_vecs[NUM_VECS_PER_THREAD];
-#pragma unroll
-  for (int i = 0; i < NUM_VECS_PER_THREAD; i++) {
-    const int vec_idx = thread_group_offset + i * THREAD_GROUP_SIZE;
-    q_vecs[i] = *reinterpret_cast<const Q_vec*>(q_ptr + vec_idx * VEC_SIZE);
-  }
-
-  // Memory planning.
-  extern __shared__ char shared_mem[];
-  // NOTE(woosuk): We use FP32 logits and accumulation.
-  float *logits = reinterpret_cast<float*>(shared_mem);
-  // Workspace for reduction.
-  __shared__ float red_smem[2 * NUM_WARPS];
-
-  // x == THREAD_GROUP_SIZE * VEC_SIZE
-  // Each thread group fetches x elements from the key at a time.
-  constexpr int x = 16 / sizeof(scalar_t);
-  float qk_max = -FLT_MAX;
-
-  const int* block_table = block_tables + seq_idx * max_num_blocks_per_seq;
-  const int context_len = context_lens[seq_idx];
-  const int num_blocks = (context_len + BLOCK_SIZE - 1) / BLOCK_SIZE;
-
-  // Iterate over the key blocks.
-  // Each warp fetches a block of keys for each iteration.
-  // Each thread group in a warp fetches a key from the block, and computes
-  // dot product with the query.
-  for (int block_idx = warp_idx; block_idx < num_blocks; block_idx += NUM_WARPS) {
-    const int physical_block_number = block_table[block_idx];
-
-    // Load a key to registers.
-    // Each thread in a thread group has a different part of the key.
-    // For example, if the the thread group size is 4, then the first thread in the group
-    // has 0, 4, 8, ... th vectors of the key, and the second thread has 1, 5, 9, ... th
-    // vectors of the key, and so on.
-    for (int i = 0; i < NUM_TOKENS_PER_THREAD_GROUP; i++) {
-      const int physical_block_offset = (thread_group_idx + i * WARP_SIZE) % BLOCK_SIZE;
-      const int token_idx = block_idx * BLOCK_SIZE + physical_block_offset;
-      K_vec k_vecs[NUM_VECS_PER_THREAD];
-
-#pragma unroll
-      for (int j = 0; j < NUM_VECS_PER_THREAD; j++) {
-        const scalar_t* k_ptr = k_cache + physical_block_number * num_heads * HEAD_SIZE * BLOCK_SIZE
-                                        + head_idx * HEAD_SIZE * BLOCK_SIZE
-                                        + physical_block_offset * x;
-        const int vec_idx = thread_group_offset + j * THREAD_GROUP_SIZE;
-        const int offset1 = (vec_idx * VEC_SIZE) / x;
-        const int offset2 = (vec_idx * VEC_SIZE) % x;
-        k_vecs[j] = *reinterpret_cast<const K_vec*>(k_ptr + offset1 * BLOCK_SIZE * x + offset2);
-      }
-
-      // Compute dot product.
-      // This includes a reduction across the threads in the same thread group.
-      const float qk = scale * Qk_dot<scalar_t, THREAD_GROUP_SIZE>::dot(q_vecs, k_vecs);
-      const bool mask = token_idx >= context_len;
-    
-      if (thread_group_offset == 0) {
-        // Store the partial reductions to shared memory.
-        // NOTE(woosuk): It is required to zero out the masked logits.
-        logits[token_idx] = mask ? 0.f : qk;
-        // Update the max value.
-        qk_max = mask ? qk_max : fmaxf(qk_max, qk);
-      }
-    }
-  }
-
-  // Perform reduction across the threads in the same warp to get the
-  // max qk value for each "warp" (not across the thread block yet).
-  // The 0-th thread of each thread group already has its max qk value.
-#pragma unroll
-  for (int mask = WARP_SIZE / 2; mask >= THREAD_GROUP_SIZE; mask /= 2) {
-    qk_max = fmaxf(qk_max, __shfl_xor_sync(uint32_t(-1), qk_max, mask));
-  }
-  if (lane == 0) {
-    red_smem[warp_idx] = qk_max;
-  }
-  __syncthreads();
-
-  // TODO(woosuk): Refactor this part.
-  // Get the max qk value for the sequence.
-  qk_max = lane < NUM_WARPS ? red_smem[lane] : -FLT_MAX;
-#pragma unroll
-  for (int mask = NUM_WARPS / 2; mask >= 1; mask /= 2) {
-      qk_max = fmaxf(qk_max, __shfl_xor_sync(uint32_t(-1), qk_max, mask));
-  }
-  // Broadcast the max qk value to all threads.
-  qk_max = __shfl_sync(uint32_t(-1), qk_max, 0);
-
-  // Get the sum of the exp values.
-  float exp_sum = 0.f;
-  for (int i = thread_idx; i < context_len; i += NUM_THREADS) {
-    float val = __expf(logits[i] - qk_max);
-    logits[i] = val;
-    exp_sum += val;
-  }
-  exp_sum = block_sum<NUM_WARPS>(&red_smem[NUM_WARPS], exp_sum);
-
-  // Compute softmax.
-  const float inv_sum = __fdividef(1.f, exp_sum + 1e-6f);
-  for (int i = thread_idx; i < context_len; i += NUM_THREADS) {
-    logits[i] *= inv_sum;
-  }
-  __syncthreads();
-
-  // Each thread will fetch 16 bytes from the value cache at a time.
-  constexpr int V_VEC_SIZE = MIN(16 / sizeof(scalar_t), BLOCK_SIZE);
-  using V_vec = typename Vec<scalar_t, V_VEC_SIZE>::Type;
-  using L_vec = typename FloatVec<V_vec>::Type;
-
-  constexpr int NUM_V_VECS_PER_ROW = BLOCK_SIZE / V_VEC_SIZE;
-  constexpr int NUM_ROWS_PER_ITER = WARP_SIZE / NUM_V_VECS_PER_ROW;
-  constexpr int NUM_ROWS_PER_THREAD = (HEAD_SIZE + NUM_ROWS_PER_ITER - 1) / NUM_ROWS_PER_ITER;
-
-  float accs[NUM_ROWS_PER_THREAD];
-#pragma unroll
-  for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
-    accs[i] = 0.f;
-  }
-
-  for (int block_idx = warp_idx; block_idx < num_blocks; block_idx += NUM_WARPS) {
-    const int physical_block_number = block_table[block_idx];
-    const int physical_block_offset = (lane % NUM_V_VECS_PER_ROW) * V_VEC_SIZE;
-    const int token_idx = block_idx * BLOCK_SIZE + physical_block_offset;
-    L_vec logits_vec = *reinterpret_cast<L_vec*>(logits + token_idx);
-
-    const scalar_t* v_ptr = v_cache + physical_block_number * num_heads * HEAD_SIZE * BLOCK_SIZE
-                                    + head_idx * HEAD_SIZE * BLOCK_SIZE;
-#pragma unroll
-    for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
-      const int row_idx = lane / NUM_V_VECS_PER_ROW + i * NUM_ROWS_PER_ITER;
-      if (row_idx < HEAD_SIZE) {
-        const int offset = row_idx * BLOCK_SIZE + physical_block_offset;
-        V_vec v_vec = *reinterpret_cast<const V_vec*>(v_ptr + offset);
-        accs[i] += dot(logits_vec, cast_to_float(v_vec));
-      }
-    }
-  }
-
-  // Perform reduction within each warp.
-#pragma unroll
-  for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
-    float acc = accs[i];
-#pragma unroll
-    for (int mask = NUM_V_VECS_PER_ROW / 2; mask >= 1; mask /= 2) {
-      acc += __shfl_xor_sync(uint32_t(-1), acc, mask);
-    }
-    accs[i] = acc;
-  }
-
-  // NOTE(woosuk): A barrier is required because the shared memory space for logits
-  // is reused for the output.
-  __syncthreads();
-
-  // Perform reduction across warps.
-  float* out_smem = reinterpret_cast<float*>(shared_mem);
-#pragma unroll
-  for (int i = NUM_WARPS; i > 1; i /= 2) {
-    int mid = i / 2;
-    // Upper warps write to shared memory.
-    if (warp_idx >= mid && warp_idx < i) {
-      float* dst = &out_smem[(warp_idx - mid) * HEAD_SIZE];
-#pragma unroll
-      for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
-        const int row_idx = lane / NUM_V_VECS_PER_ROW + i * NUM_ROWS_PER_ITER;
-        if (row_idx < HEAD_SIZE && lane % NUM_V_VECS_PER_ROW == 0) {
-          dst[row_idx] = accs[i];
-        }
-      }
-    }
-    __syncthreads();
-
-    // Lower warps update the output.
-    if (warp_idx < mid) {
-      const float* src = &out_smem[warp_idx * HEAD_SIZE];
-#pragma unroll
-      for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
-        const int row_idx = lane / NUM_V_VECS_PER_ROW + i * NUM_ROWS_PER_ITER;
-        if (row_idx < HEAD_SIZE && lane % NUM_V_VECS_PER_ROW == 0) {
-          accs[i] += src[row_idx];
-        }
-      }
-    }
-    __syncthreads();
-  }
-
-  // Write the final output.
-  if (warp_idx == 0) {
-    scalar_t* out_ptr = out + seq_idx * num_heads * HEAD_SIZE + head_idx * HEAD_SIZE;
-#pragma unroll
-    for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
-      const int row_idx = lane / NUM_V_VECS_PER_ROW + i * NUM_ROWS_PER_ITER;
-      if (row_idx < HEAD_SIZE && lane % NUM_V_VECS_PER_ROW == 0) {
-        convert_from_float(*(out_ptr + row_idx), accs[i]);
-      }
-    }
-  }
-}
-
-} // namespace cacheflow
-
-#define LAUNCH_ATTENTION_KERNEL(T, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS)                        \
-  cacheflow::single_query_cached_kv_attention_kernel<T, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS>   \
-  <<<grid, block, shared_mem_size, stream>>>(                                                 \
-    out_ptr,                                                                                  \
-    query_ptr,                                                                                \
-    key_cache_ptr,                                                                            \
-    value_cache_ptr,                                                                          \
-    scale,                                                                                    \
-    block_tables_ptr,                                                                         \
-    context_lens_ptr,                                                                         \
-    max_num_blocks_per_seq,                                                                   \
-    query_stride);
-
-// TODO(woosuk): Tune NUM_THREADS.
-template<
-  typename T,
-  int BLOCK_SIZE,
-  int NUM_THREADS = 128>
-void single_query_cached_kv_attention_launcher(
-  torch::Tensor& out,
-  torch::Tensor& query,
-  torch::Tensor& key_cache,
-  torch::Tensor& value_cache,
-  float scale,
-  torch::Tensor& block_tables,
-  torch::Tensor& context_lens,
-  int max_context_len) {
-  int num_seqs = query.size(0);
-  int num_heads = query.size(1);
-  int head_size = query.size(2);
-  int max_num_blocks_per_seq = block_tables.size(1);
-  int query_stride = query.stride(0);
-
-  int thread_group_size = MAX(WARP_SIZE / BLOCK_SIZE, 1);
-  assert(head_size % thread_group_size == 0);
-
-  T* out_ptr = reinterpret_cast<T*>(out.data_ptr());
-  T* query_ptr = reinterpret_cast<T*>(query.data_ptr());
-  T* key_cache_ptr = reinterpret_cast<T*>(key_cache.data_ptr());
-  T* value_cache_ptr = reinterpret_cast<T*>(value_cache.data_ptr());
-  int* block_tables_ptr = block_tables.data_ptr<int>();
-  int* context_lens_ptr = context_lens.data_ptr<int>();
-
-  constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
-  int padded_max_context_len = ((max_context_len + BLOCK_SIZE - 1) / BLOCK_SIZE) * BLOCK_SIZE;
-  int logits_size = padded_max_context_len * sizeof(float);
-  int outputs_size = (NUM_WARPS / 2) * head_size * sizeof(float);
-  int shared_mem_size = std::max(logits_size, outputs_size);
-
-  dim3 grid(num_heads, num_seqs);
-  dim3 block(NUM_THREADS);
-  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
-  switch (head_size) {
-    case 32:
-      LAUNCH_ATTENTION_KERNEL(T, 32, BLOCK_SIZE, NUM_THREADS);
-      break;
-    case 64:
-      LAUNCH_ATTENTION_KERNEL(T, 64, BLOCK_SIZE, NUM_THREADS);
-      break;
-    case 80:
-      LAUNCH_ATTENTION_KERNEL(T, 80, BLOCK_SIZE, NUM_THREADS);
-      break;
-    case 96:
-      LAUNCH_ATTENTION_KERNEL(T, 96, BLOCK_SIZE, NUM_THREADS);
-      break;
-    case 128:
-      LAUNCH_ATTENTION_KERNEL(T, 128, BLOCK_SIZE, NUM_THREADS);
-      break;
-    case 160:
-      LAUNCH_ATTENTION_KERNEL(T, 160, BLOCK_SIZE, NUM_THREADS);
-      break;
-    case 192:
-      LAUNCH_ATTENTION_KERNEL(T, 192, BLOCK_SIZE, NUM_THREADS);
-      break;
-    case 256:
-      LAUNCH_ATTENTION_KERNEL(T, 256, BLOCK_SIZE, NUM_THREADS);
-      break;
-    default:
-      assert(false);
-      break;
-  }
-}
-
-#define CALL_KERNEL_LAUNCHER(T, BLOCK_SIZE)                         \
-  single_query_cached_kv_attention_launcher<T, BLOCK_SIZE>(         \ 
-        out,                                                        \
-        query,                                                      \
-        key_cache,                                                  \
-        value_cache,                                                \
-        scale,                                                      \
-        block_tables,                                               \
-        context_lens,                                               \
-        max_context_len);
-
-void single_query_cached_kv_attention(
-  torch::Tensor& out,             // [num_seqs, num_heads, head_size]
-  torch::Tensor& query,           // [num_seqs, num_heads, head_size]
-  torch::Tensor& key_cache,       // [num_blocks, num_heads, head_size/x, block_size, x]
-  torch::Tensor& value_cache,     // [num_blocks, num_heads, head_size, block_size]
-  float scale,
-  torch::Tensor& block_tables,    // [num_seqs, max_num_blocks_per_seq]
-  torch::Tensor& context_lens,    // [num_seqs]
-  int block_size,
-  int max_context_len) {
-  // TODO(woosuk): Support BF16.
-  if (query.element_size() == 2) {
-    // Half.
-    if (block_size == 1) {
-      CALL_KERNEL_LAUNCHER(uint16_t, 1);
-    } else if (block_size == 2) {
-      CALL_KERNEL_LAUNCHER(uint16_t, 2);
-    } else if (block_size == 4) {
-      CALL_KERNEL_LAUNCHER(uint16_t, 4);
-    } else if (block_size == 8) {
-      CALL_KERNEL_LAUNCHER(uint16_t, 8);
-    } else if (block_size == 16) {
-      CALL_KERNEL_LAUNCHER(uint16_t, 16);
-    } else if (block_size == 32) {
-      CALL_KERNEL_LAUNCHER(uint16_t, 32);
-    } else if (block_size == 64) {
-      CALL_KERNEL_LAUNCHER(uint16_t, 64);
-    } else if (block_size == 128) {
-      CALL_KERNEL_LAUNCHER(uint16_t, 128);
-    } else if (block_size == 256) {
-      CALL_KERNEL_LAUNCHER(uint16_t, 256);
-    } else {
-      assert(false);
-    }
-  } else {
-    // Float.
-    assert(false);
-  }
-}
-
-// namespace cacheflow {
-
-// // Grid: (num_heads, num_query_tokens).
-// template<
-//   typename scalar_t,
-//   int HEAD_SIZE,
-//   int BLOCK_SIZE,
-//   int NUM_THREADS>
-// __device__ void multi_query_cached_kv_attention_kernel_unoptimized_(
-//   scalar_t* __restrict__ out,             // [num_seqs, num_heads, head_size]
-//   const scalar_t* __restrict__ q,         // [num_seqs, num_heads, head_size]
-//   const int seq_start_idx,
-//   const int seq_len,
-//   const scalar_t* __restrict__ k_cache,   // [num_blocks, num_heads, head_size/x, block_size, x]
-//   const scalar_t* __restrict__ v_cache,   // [num_blocks, num_heads, head_size, block_size]
-//   const float scale,
-//   const int* __restrict__ block_table,   // [num_seqs, max_num_blocks_per_seq]
-//   const int context_len,
-//   const int max_num_blocks_per_seq,
-//   const int q_stride) {
-//   constexpr int THREAD_GROUP_SIZE = WARP_SIZE / BLOCK_SIZE;
-//   constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
-//   const int thread_idx = threadIdx.x;
-//   const int warp_idx = thread_idx / WARP_SIZE;
-//   const int lane = thread_idx % WARP_SIZE;
-
-//   const int head_idx = blockIdx.x;
-//   const int num_heads = gridDim.x;
-//   const int seq_idx = blockIdx.y;
-
-//   // A vector type to store a part of a key or a query.
-//   // The vector size is configured in such a way that the threads in a thread group
-//   // fetch or comput 16 bytes at a time.
-//   // For example, if the size of a thread group is 4 and the data type is half,
-//   // then the vector size is 16 / (4 * sizeof(half)) == 2.
-//   constexpr int VEC_SIZE = 16 / (THREAD_GROUP_SIZE * sizeof(scalar_t));
-//   using K_vec = typename Vec<scalar_t, VEC_SIZE>::Type;
-//   using Q_vec = typename Vec<scalar_t, VEC_SIZE>::Type;
-
-//   constexpr int NUM_ELEMS_PER_THREAD = HEAD_SIZE / THREAD_GROUP_SIZE;
-//   constexpr int NUM_VECS_PER_THREAD = NUM_ELEMS_PER_THREAD / VEC_SIZE;
-
-//   const int thread_group_idx = thread_idx / THREAD_GROUP_SIZE;
-//   const int thread_group_offset = thread_idx % THREAD_GROUP_SIZE;
-
-//   // Load the query to registers.
-//   // Each thread in a thread group has a different part of the query.
-//   // For example, if the the thread group size is 4, then the first thread in the group
-//   // has 0, 4, 8, ... th vectors of the query, and the second thread has 1, 5, 9, ...
-//   // th vectors of the query, and so on.
-//   // NOTE(woosuk): Because q is split from a qkv tensor, it may not be contiguous.
-//   const scalar_t* q_ptr = q + seq_idx * q_stride + head_idx * HEAD_SIZE;
-//   Q_vec q_vecs[NUM_VECS_PER_THREAD];
-// #pragma unroll
-//   for (int i = 0; i < NUM_VECS_PER_THREAD; i++) {
-//     const int vec_idx = thread_group_offset + i * THREAD_GROUP_SIZE;
-//     q_vecs[i] = *reinterpret_cast<const Q_vec*>(q_ptr + vec_idx * VEC_SIZE);
-//   }
-
-//   // Memory planning.
-//   extern __shared__ char shared_mem[];
-//   // NOTE(woosuk): We use FP32 logits and accumulation.
-//   float *logits = reinterpret_cast<float*>(shared_mem);
-//   // Workspace for reduction.
-//   __shared__ float red_smem[2 * NUM_WARPS];
-
-//   // x == THREAD_GROUP_SIZE * VEC_SIZE
-//   // Each thread group fetches x elements from the key at a time.
-//   constexpr int x = 16 / sizeof(scalar_t);
-//   float qk_max = -FLT_MAX;
-
-//   const int num_blocks = (context_len + BLOCK_SIZE - 1) / BLOCK_SIZE;
-//   const int mask_boundary = context_len - seq_len + 1 + (seq_idx - seq_start_idx);
-
-//   // Iterate over the key blocks.
-//   // Each warp fetches a block of keys for each iteration.
-//   // Each thread group in a warp fetches a key from the block, and computes
-//   // dot product with the query.
-//   for (int block_idx = warp_idx; block_idx < num_blocks; block_idx += NUM_WARPS) {
-//     const int physical_block_number = block_table[block_idx];
-//     const int physical_block_offset = thread_group_idx % BLOCK_SIZE;
-//     const int token_idx = block_idx * BLOCK_SIZE + physical_block_offset;
-
-//     // Load a key to registers.
-//     // Each thread in a thread group has a different part of the key.
-//     // For example, if the the thread group size is 4, then the first thread in the group
-//     // has 0, 4, 8, ... th vectors of the key, and the second thread has 1, 5, 9, ... th
-//     // vectors of the key, and so on.
-//     K_vec k_vecs[NUM_VECS_PER_THREAD];
-// #pragma unroll
-//     for (int i = 0; i < NUM_VECS_PER_THREAD; i++) {
-//       const scalar_t* k_ptr = k_cache + physical_block_number * num_heads * HEAD_SIZE * BLOCK_SIZE
-//                                       + head_idx * HEAD_SIZE * BLOCK_SIZE
-//                                       + physical_block_offset * x;
-//       const int vec_idx = thread_group_offset + i * THREAD_GROUP_SIZE;
-//       const int offset1 = (vec_idx * VEC_SIZE) / x;
-//       const int offset2 = (vec_idx * VEC_SIZE) % x;
-//       k_vecs[i] = *reinterpret_cast<const K_vec*>(k_ptr + offset1 * BLOCK_SIZE * x + offset2);
-//     }
-
-//     // Compute dot product.
-//     // This includes a reduction across the threads in the same thread group.
-//     const float qk = scale * Qk_dot<scalar_t, THREAD_GROUP_SIZE>::dot(q_vecs, k_vecs);
-//     const bool mask = token_idx >= mask_boundary;
-
-//     if (thread_group_offset == 0) {
-//       // Store the partial reductions to shared memory.
-//       // NOTE(woosuk): It is required to zero out the masked logits.
-//       logits[token_idx] = mask ? 0.f : qk;
-//       // Update the max value.
-//       qk_max = mask ? qk_max : fmaxf(qk_max, qk);
-//     }
-//   }
-
-//   // Perform reduction across the threads in the same warp to get the
-//   // max qk value for each "warp" (not across the thread block yet).
-//   // The 0-th thread of each thread group already has its max qk value.
-// #pragma unroll
-//   for (int mask = WARP_SIZE / 2; mask >= THREAD_GROUP_SIZE; mask /= 2) {
-//     qk_max = fmaxf(qk_max, __shfl_xor_sync(uint32_t(-1), qk_max, mask));
-//   }
-//   if (lane == 0) {
-//     red_smem[warp_idx] = qk_max;
-//   }
-//   __syncthreads();
-
-//   // TODO(woosuk): Refactor this part.
-//   // Get the max qk value for the sequence.
-//   qk_max = lane < NUM_WARPS ? red_smem[lane] : -FLT_MAX;
-// #pragma unroll
-//   for (int mask = NUM_WARPS / 2; mask >= 1; mask /= 2) {
-//       qk_max = fmaxf(qk_max, __shfl_xor_sync(uint32_t(-1), qk_max, mask));
-//   }
-//   // Broadcast the max qk value to all threads.
-//   qk_max = __shfl_sync(uint32_t(-1), qk_max, 0);
-
-//   // Get the sum of the exp values.
-//   float exp_sum = 0.f;
-//   for (int i = thread_idx; i < mask_boundary; i += NUM_THREADS) {
-//     float val = __expf(logits[i] - qk_max);
-//     logits[i] = val;
-//     exp_sum += val;
-//   }
-//   exp_sum = block_sum<NUM_WARPS>(&red_smem[NUM_WARPS], exp_sum);
-
-//   // Compute softmax.
-//   const float inv_sum = __fdividef(1.f, exp_sum + 1e-6f);
-//   for (int i = thread_idx; i < context_len; i += NUM_THREADS) {
-//     logits[i] *= inv_sum;
-//   }
-//   __syncthreads();
-
-//   // Each thread will fetch 16 bytes from the value cache at a time.
-//   constexpr int V_VEC_SIZE = 16 / sizeof(scalar_t);
-//   using V_vec = typename Vec<scalar_t, V_VEC_SIZE>::Type;
-//   using L_vec = typename FloatVec<V_vec>::Type;
-
-//   constexpr int NUM_V_VECS_PER_ROW = BLOCK_SIZE / V_VEC_SIZE;
-//   constexpr int NUM_ROWS_PER_ITER = WARP_SIZE / NUM_V_VECS_PER_ROW;
-//   constexpr int NUM_ROWS_PER_THREAD = (HEAD_SIZE + NUM_ROWS_PER_ITER - 1) / NUM_ROWS_PER_ITER;
-
-//   float accs[NUM_ROWS_PER_THREAD];
-// #pragma unroll
-//   for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
-//     accs[i] = 0.f;
-//   }
-
-//   for (int block_idx = warp_idx; block_idx < num_blocks; block_idx += NUM_WARPS) {
-//     const int physical_block_number = block_table[block_idx];
-//     const int physical_block_offset = (lane % NUM_V_VECS_PER_ROW) * V_VEC_SIZE;
-//     const int token_idx = block_idx * BLOCK_SIZE + physical_block_offset;
-//     L_vec logits_vec = *reinterpret_cast<L_vec*>(logits + token_idx);
-
-//     const scalar_t* v_ptr = v_cache + physical_block_number * num_heads * HEAD_SIZE * BLOCK_SIZE
-//                                     + head_idx * HEAD_SIZE * BLOCK_SIZE;
-// #pragma unroll
-//     for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
-//       const int row_idx = lane / NUM_V_VECS_PER_ROW + i * NUM_ROWS_PER_ITER;
-//       if (row_idx < HEAD_SIZE) {
-//         const int offset = row_idx * BLOCK_SIZE + physical_block_offset;
-//         V_vec v_vec = *reinterpret_cast<const V_vec*>(v_ptr + offset);
-//         accs[i] += dot(logits_vec, cast_to_float(v_vec));
-//       }
-//     }
-//   }
-
-//   // Perform reduction within each warp.
-// #pragma unroll
-//   for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
-//     float acc = accs[i];
-// #pragma unroll
-//     for (int mask = NUM_V_VECS_PER_ROW / 2; mask >= 1; mask /= 2) {
-//       acc += __shfl_xor_sync(uint32_t(-1), acc, mask);
-//     }
-//     accs[i] = acc;
-//   }
-
-//   // NOTE(woosuk): A barrier is required because the shared memory space for logits
-//   // is reused for the output.
-//   __syncthreads();
-
-//   // Perform reduction across warps.
-//   float* out_smem = reinterpret_cast<float*>(shared_mem);
-// #pragma unroll
-//   for (int i = NUM_WARPS; i > 1; i /= 2) {
-//     int mid = i / 2;
-//     // Upper warps write to shared memory.
-//     if (warp_idx >= mid && warp_idx < i) {
-//       float* dst = &out_smem[(warp_idx - mid) * HEAD_SIZE];
-// #pragma unroll
-//       for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
-//         const int row_idx = lane / NUM_V_VECS_PER_ROW + i * NUM_ROWS_PER_ITER;
-//         if (row_idx < HEAD_SIZE && lane % NUM_V_VECS_PER_ROW == 0) {
-//           dst[row_idx] = accs[i];
-//         }
-//       }
-//     }
-//     __syncthreads();
-
-//     // Lower warps update the output.
-//     if (warp_idx < mid) {
-//       const float* src = &out_smem[warp_idx * HEAD_SIZE];
-// #pragma unroll
-//       for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
-//         const int row_idx = lane / NUM_V_VECS_PER_ROW + i * NUM_ROWS_PER_ITER;
-//         if (row_idx < HEAD_SIZE && lane % NUM_V_VECS_PER_ROW == 0) {
-//           accs[i] += src[row_idx];
-//         }
-//       }
-//     }
-//     __syncthreads();
-//   }
-
-//   // Write the final output.
-//   if (warp_idx == 0) {
-//     scalar_t* out_ptr = out + seq_idx * num_heads * HEAD_SIZE + head_idx * HEAD_SIZE;
-// #pragma unroll
-//     for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
-//       const int row_idx = lane / NUM_V_VECS_PER_ROW + i * NUM_ROWS_PER_ITER;
-//       if (row_idx < HEAD_SIZE && lane % NUM_V_VECS_PER_ROW == 0) {
-//         convert_from_float(*(out_ptr + row_idx), accs[i]);
-//       }
-//     }
-//   }
-// }
-
-
-// // Grid: (num_heads, num_query_tokens).
-// template<
-//   typename scalar_t,
-//   int HEAD_SIZE,
-//   int BLOCK_SIZE,
-//   int NUM_THREADS>
-// __global__ void multi_query_cached_kv_attention_kernel(
-//   const int* cu_query_lens,               // [num_prompts+1]
-//   const int* seq_prompt_mapping,          // [num_seqs] mapping from seq_idx to prompt_idx
-//   scalar_t* __restrict__ out,             // [num_seqs, num_heads, head_size]
-//   const scalar_t* __restrict__ q,         // [num_seqs, num_heads, head_size]
-//   const scalar_t* __restrict__ k_cache,   // [num_blocks, num_heads, head_size/x, block_size, x]
-//   const scalar_t* __restrict__ v_cache,   // [num_blocks, num_heads, head_size, block_size]
-//   const float scale,
-//   const int* __restrict__ block_tables,   // [num_prompts, max_num_blocks_per_seq]
-//   const int* __restrict__ context_lens,   // [num_prompts]
-//   const int max_num_blocks_per_seq,
-//   const int q_stride) {
-//     const int seq_idx = blockIdx.y;
-//     const int prompt_idx = seq_prompt_mapping[seq_idx];
-//     const int seq_start_idx = cu_query_lens[prompt_idx];
-//     const int seq_len = cu_query_lens[prompt_idx + 1] - seq_start_idx;
-//     const int* block_table = block_tables + prompt_idx * max_num_blocks_per_seq;
-//     const int context_len = context_lens[prompt_idx];
-//     multi_query_cached_kv_attention_kernel_unoptimized_<
-//         scalar_t, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS>(
-//           out,
-//           q,
-//           seq_start_idx,
-//           seq_len,
-//           k_cache,
-//           v_cache,
-//           scale,
-//           block_table,
-//           context_len,
-//           max_num_blocks_per_seq,
-//           q_stride);
-// }
-
-// } // namespace cacheflow
-
-// #define LAUNCH_MULTI_ATTENTION_KERNEL(T, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS)                  \
-//   cacheflow::multi_query_cached_kv_attention_kernel<T, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS>    \
-//   <<<grid, block, shared_mem_size, stream>>>(                                                 \
-//     cu_query_lens_ptr,                                                                        \
-//     seq_prompt_mapping_ptr,                                                                   \
-//     out_ptr,                                                                                  \
-//     query_ptr,                                                                                \
-//     key_cache_ptr,                                                                            \
-//     value_cache_ptr,                                                                          \
-//     scale,                                                                                    \
-//     block_tables_ptr,                                                                         \
-//     context_lens_ptr,                                                                         \
-//     max_num_blocks_per_seq,                                                                   \
-//     query_stride);
-
-
-// // TODO(woosuk): Tune NUM_THREADS.
-// template<
-//   typename T,
-//   int BLOCK_SIZE,
-//   int NUM_THREADS = 128>
-// void multi_query_cached_kv_attention_launcher(
-//   torch::Tensor& cu_query_lens,
-//   torch::Tensor& seq_prompt_mapping,
-//   torch::Tensor& out,
-//   torch::Tensor& query,
-//   torch::Tensor& key_cache,
-//   torch::Tensor& value_cache,
-//   float scale,
-//   torch::Tensor& block_tables,
-//   torch::Tensor& context_lens,
-//   int max_context_len) {
-//   int num_seqs = query.size(0);
-//   int num_heads = query.size(1);
-//   int head_size = query.size(2);
-//   int max_num_blocks_per_seq = block_tables.size(1);
-//   int query_stride = query.stride(0);
-
-//   int* cu_query_lens_ptr = cu_query_lens.data_ptr<int>();
-//   int* seq_prompt_mapping_ptr = seq_prompt_mapping.data_ptr<int>();
-//   T* out_ptr = reinterpret_cast<T*>(out.data_ptr());
-//   T* query_ptr = reinterpret_cast<T*>(query.data_ptr());
-//   T* key_cache_ptr = reinterpret_cast<T*>(key_cache.data_ptr());
-//   T* value_cache_ptr = reinterpret_cast<T*>(value_cache.data_ptr());
-//   int* block_tables_ptr = block_tables.data_ptr<int>();
-//   int* context_lens_ptr = context_lens.data_ptr<int>();
-
-//   constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
-//   int padded_max_context_len = ((max_context_len + BLOCK_SIZE - 1) / BLOCK_SIZE) * BLOCK_SIZE;
-//   int logits_size = padded_max_context_len * sizeof(float);
-//   int outputs_size = (NUM_WARPS / 2) * head_size * sizeof(float);
-//   int shared_mem_size = std::max(logits_size, outputs_size);
-
-//   dim3 grid(num_heads, num_seqs);
-//   dim3 block(NUM_THREADS);
-//   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
-//   switch (head_size) {
-//     case 32:
-//       LAUNCH_MULTI_ATTENTION_KERNEL(T, 32, BLOCK_SIZE, NUM_THREADS);
-//       break;
-//     case 64:
-//       LAUNCH_MULTI_ATTENTION_KERNEL(T, 64, BLOCK_SIZE, NUM_THREADS);
-//       break;
-//     case 80:
-//       LAUNCH_MULTI_ATTENTION_KERNEL(T, 80, BLOCK_SIZE, NUM_THREADS);
-//       break;
-//     case 96:
-//       LAUNCH_MULTI_ATTENTION_KERNEL(T, 96, BLOCK_SIZE, NUM_THREADS);
-//       break;
-//     case 128:
-//       LAUNCH_MULTI_ATTENTION_KERNEL(T, 128, BLOCK_SIZE, NUM_THREADS);
-//       break;
-//     case 160:
-//       LAUNCH_MULTI_ATTENTION_KERNEL(T, 160, BLOCK_SIZE, NUM_THREADS);
-//       break;
-//     case 192:
-//       LAUNCH_MULTI_ATTENTION_KERNEL(T, 192, BLOCK_SIZE, NUM_THREADS);
-//       break;
-//     case 256:
-//       LAUNCH_MULTI_ATTENTION_KERNEL(T, 256, BLOCK_SIZE, NUM_THREADS);
-//       break;
-//     default:
-//       assert(false);
-//       break;
-//   }
-// }
-
-// void multi_query_cached_kv_attention(
-//   torch::Tensor& cu_query_lens,
-//   torch::Tensor& out,
-//   torch::Tensor& query,
-//   torch::Tensor& key_cache,
-//   torch::Tensor& value_cache,
-//   float scale,
-//   torch::Tensor& block_tables,
-//   torch::Tensor& context_lens,
-//   int block_size,
-//   int max_context_len) {
-
-//   torch::Tensor query_lens = cu_query_lens.to(torch::kCPU);
-  
-//   int num_queries = query_lens.size(0) - 1;
-//   const int* query_lens_ptr = query_lens.data_ptr<int>();
-//   int num_seqs = query.size(0);
-
-//   torch::Tensor cpu_tensor = torch::empty({num_seqs}, torch::dtype(torch::kInt32));
-//   auto accessor = cpu_tensor.accessor<int32_t, 1>();
-//   for (int i = 0, query_cursor = 0; i < num_seqs; ++i) {
-//     if (i >= query_lens_ptr[query_cursor + 1]) {
-//       ++query_cursor; 
-//     }
-//     accessor[i] = query_cursor;
-//   }
-
-//   // TODO(suquark): This can be slow, as it to(torch::kCPU) and to(torch::kCUDA)
-//   // implicitly synchronizes the CPU and GPU. And we can avoid this issue by giving
-//   // the mapping as an input parameter. Let's do this optimization in a later PR.
-//   torch::Tensor seq_prompt_mapping = cpu_tensor.to(torch::kCUDA);
-
-//   // TODO(woosuk): Support BF16.
-//   if (query.element_size() == 2) {
-//     // Half.
-//     if (block_size == 8) {
-//       multi_query_cached_kv_attention_launcher<uint16_t, 8>(
-//         cu_query_lens,
-//         seq_prompt_mapping,
-//         out,
-//         query,
-//         key_cache,
-//         value_cache,
-//         scale,
-//         block_tables,
-//         context_lens,
-//         max_context_len);
-//     } else if (block_size == 16) {
-//       multi_query_cached_kv_attention_launcher<uint16_t, 16>(
-//         cu_query_lens,
-//         seq_prompt_mapping,
-//         out,
-//         query,
-//         key_cache,
-//         value_cache,
-//         scale,
-//         block_tables,
-//         context_lens,
-//         max_context_len);
-//     } else if (block_size == 32) {
-//       multi_query_cached_kv_attention_launcher<uint16_t, 32>(
-//         cu_query_lens,
-//         seq_prompt_mapping,
-//         out,
-//         query,
-//         key_cache,
-//         value_cache,
-//         scale,
-//         block_tables,
-//         context_lens,
-//         max_context_len);
-//     } else {
-//       assert(false);
-//     }
-//   } else if (query.element_size() == 4) {
-//     // Float.
-//     if (block_size == 8) {
-//       multi_query_cached_kv_attention_launcher<float, 8>(
-//         cu_query_lens,
-//         seq_prompt_mapping,
-//         out,
-//         query,
-//         key_cache,
-//         value_cache,
-//         scale,
-//         block_tables,
-//         context_lens,
-//         max_context_len);
-//     } else if (block_size == 16) {
-//       multi_query_cached_kv_attention_launcher<float, 16>(
-//         cu_query_lens,
-//         seq_prompt_mapping,
-//         out,
-//         query,
-//         key_cache,
-//         value_cache,
-//         scale,
-//         block_tables,
-//         context_lens,
-//         max_context_len);
-//     } else if (block_size == 32) {
-//       multi_query_cached_kv_attention_launcher<float, 32>(
-//         cu_query_lens,
-//         seq_prompt_mapping,
-//         out,
-//         query,
-//         key_cache,
-//         value_cache,
-//         scale,
-//         block_tables,
-//         context_lens,
-//         max_context_len);
-//     } else {
-//       assert(false);
-//     }
-//   } else {
-//     assert(false);
-//   }
-// }
-
-#undef WARP_SIZE
-#undef MAX
-#undef MIN

csrc/attention_utils.h

@@ -1,165 +0,0 @@
-#pragma once
-
-#include "cuda_primitives.h"
-
-#include <float.h>
-#include <type_traits>
-
-#define MMHA_USE_FP32_ACUM_FOR_FMA
-#define MMHA_USE_FP32_ACUM_FOR_OUT
-
-namespace cacheflow {
-
-// A vector type to store Q, K, V elements.
-template<typename T, int VEC_SIZE>
-struct Vec {};
-template<>
-struct Vec<float, 1> {
-    using Type = float;
-};
-template<>
-struct Vec<float, 2> {
-    using Type = float2;
-};
-template<>
-struct Vec<float, 4> {
-    using Type = float4;
-};
-template<>
-struct Vec<uint16_t, 1> {
-    using Type = uint16_t;
-};
-template<>
-struct Vec<uint16_t, 2> {
-    using Type = uint32_t;
-};
-template<>
-struct Vec<uint16_t, 4> {
-    using Type = uint2;
-};
-template<>
-struct Vec<uint16_t, 8> {
-    using Type = uint4;
-};
-
-template<typename T>
-struct FloatVec {};
-template<>
-struct FloatVec<float> {
-    using Type = float;
-};
-template<>
-struct FloatVec<float2> {
-    using Type = float2;
-};
-template<>
-struct FloatVec<float4> {
-    using Type = float4;
-};
-template<>
-struct FloatVec<uint16_t> {
-    using Type = float;
-};
-template<>
-struct FloatVec<uint32_t> {
-    using Type = float2;
-};
-template<>
-struct FloatVec<uint2> {
-    using Type = Float4_;
-};
-template<>
-struct FloatVec<uint4> {
-    using Type = Float8_;
-};
-
-template<int THREADS_PER_KEY, typename K_vec, int N>
-inline __device__ float qk_dot_(const K_vec (&q)[N], const K_vec (&k)[N])
-{
-    using K_vec_acum = typename FloatVec<K_vec>::Type;
-    // Compute the parallel products for Q*K^T (treat vector lanes separately).
-    K_vec_acum qk_vec = mul<K_vec_acum, K_vec, K_vec>(q[0], k[0]);
-#pragma unroll
-    for (int ii = 1; ii < N; ++ii) {
-        qk_vec = fma(q[ii], k[ii], qk_vec);
-    }
-
-    // Finalize the reduction across lanes.
-    float qk = sum(qk_vec);
-#pragma unroll
-    for (int mask = THREADS_PER_KEY / 2; mask >= 1; mask /= 2) {
-        qk += __shfl_xor_sync(uint32_t(-1), qk, mask);
-    }
-    return qk;
-}
-
-////////////////////////////////////////////////////////////////////////////////////////////////////
-
-template<typename T, int THREADS_PER_KEY>
-struct Qk_dot {
-    template<typename K_vec, int N>
-    static inline __device__ float dot(const K_vec (&q)[N], const K_vec (&k)[N])
-    {
-        return qk_dot_<THREADS_PER_KEY>(q, k);
-    }
-};
-
-////////////////////////////////////////////////////////////////////////////////////////////////////
-
-inline __device__ float4 hmma_fp32(const uint2& a, uint32_t b)
-{
-    float4 c;
-    float zero = 0.f;
-    asm volatile("mma.sync.aligned.m16n8k8.row.col.f32.f16.f16.f32 \n"
-                 "    {%0, %1, %2, %3}, \n"
-                 "    {%4, %5}, \n"
-                 "    {%6}, \n"
-                 "    {%7, %7, %7, %7}; \n"
-
-                 : "=f"(c.x), "=f"(c.y), "=f"(c.z), "=f"(c.w)
-                 : "r"(a.x) "r"(a.y), "r"(b), "f"(zero));
-    return c;
-}
-
-////////////////////////////////////////////////////////////////////////////////////////////////////
-
-template<int N>
-inline __device__ float qk_hmma_dot_(const uint32_t (&q)[N], const uint32_t (&k)[N])
-{
-#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 750
-    using K_vec_acum = typename FloatVec<uint32_t>::Type;
-    K_vec_acum qk_vec = mul<K_vec_acum, uint32_t, uint32_t>(q[0], k[0]);
-#pragma unroll
-    for (int ii = 1; ii < N; ++ii) {
-        qk_vec = fma(q[ii], k[ii], qk_vec);
-    }
-#ifdef MMHA_USE_FP32_ACUM_FOR_FMA
-    uint32_t qk_vec_ = float2_to_half2(qk_vec);
-    return hmma_fp32(make_uint2(qk_vec_, 0u), 0x3c003c00u).x;
-#else
-    return hmma_fp32(make_uint2(qk_vec, 0u), 0x3c003c00u).x;
-#endif
-#else
-    return 0.f;
-#endif
-}
-
-////////////////////////////////////////////////////////////////////////////////////////////////////
-
-template<>
-struct Qk_dot<uint16_t, 4> {
-    template<int N>
-    static inline __device__ float dot(const uint32_t (&q)[N], const uint32_t (&k)[N])
-    {
-#if __CUDA_ARCH__ >= 750 && defined(MMHA_USE_HMMA_FOR_REDUCTION)
-        return qk_hmma_dot_(q, k);
-#else
-        return qk_dot_<4>(q, k);
-#endif  // defined MMHA_USE_HMMA_FOR_REDUCTION
-    }
-};
-
-} // namespace cacheflow
-
-#undef MMHA_USE_FP32_ACUM_FOR_FMA
-#undef MMHA_USE_FP32_ACUM_FOR_OUT

csrc/cache.h

@@ -1,3 +1,5 @@
+#pragma once
+
 #include <torch/extension.h>
 
 #include <map>
@@ -18,7 +20,8 @@ void reshape_and_cache(
   torch::Tensor& value,
   torch::Tensor& key_cache,
   torch::Tensor& value_cache,
-  torch::Tensor& slot_mapping);
+  torch::Tensor& slot_mapping,
+  const std::string& kv_cache_dtype);
 
 void gather_cached_kv(
   torch::Tensor& key,
@@ -27,21 +30,7 @@ void gather_cached_kv(
   torch::Tensor& value_cache,
   torch::Tensor& slot_mapping);
 
-PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
-  m.def(
-    "swap_blocks",
-    &swap_blocks,
-    "Swap in (out) the cache blocks from src to dst");
-  m.def(
-    "copy_blocks",
-    &copy_blocks,
-    "Copy the cache blocks from src to dst");
-  m.def(
-    "reshape_and_cache",
-    &reshape_and_cache,
-    "Reshape the key and value tensors and cache them");
-  m.def(
-    "gather_cached_kv",
-    &gather_cached_kv,
-    "Gather key and value from the cache into contiguous QKV tensors");
-}
+// Just for unittest
+void convert_fp8_e5m2(
+  torch::Tensor& src_cache,
+  torch::Tensor& dst_cache);

csrc/cache_kernels.cu

@@ -1,11 +1,23 @@
 #include <torch/extension.h>
 #include <ATen/cuda/CUDAContext.h>
+#include <c10/cuda/CUDAGuard.h>
+
+#include "cuda_compat.h"
+#include "dispatch_utils.h"
+#ifdef ENABLE_FP8_E5M2
+#include "quantization/fp8_e5m2_kvcache/quant_utils.cuh"
+#endif
 
 #include <algorithm>
 #include <cassert>
 #include <map>
 #include <vector>
 
+#ifdef USE_ROCM
+  #include <hip/hip_bf16.h>
+  typedef __hip_bfloat16 __nv_bfloat16;
+#endif
+
 void swap_blocks(
   torch::Tensor& src,
   torch::Tensor& dst,
@@ -14,21 +26,25 @@ void swap_blocks(
   torch::Device dst_device = dst.device();
   cudaMemcpyKind memcpy_type;
   if (src_device.is_cuda() && dst_device.is_cuda()) {
-    assert(src_device.index() == dst_device.index());
+    TORCH_CHECK(
+      src_device.index() == dst_device.index(),
+      "src and dst must be on the same GPU");
     memcpy_type = cudaMemcpyDeviceToDevice;
   } else if (src_device.is_cuda() && dst_device.is_cpu()) {
     memcpy_type = cudaMemcpyDeviceToHost;
   } else if (src_device.is_cpu() && dst_device.is_cuda()) {
     memcpy_type = cudaMemcpyHostToDevice;
   } else {
-    assert(false);
+    TORCH_CHECK(false, "Invalid device combination");
   }
 
-  void *src_ptr = src.data_ptr();
-  void *dst_ptr = dst.data_ptr();
+  char *src_ptr = static_cast<char*>(src.data_ptr());
+  char *dst_ptr = static_cast<char*>(dst.data_ptr());
 
   const int64_t block_size_in_bytes = src.element_size() * src[0].numel();
+  const at::cuda::OptionalCUDAGuard device_guard(src_device.is_cuda() ? src_device : dst_device);
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+  // NOTE(woosuk): This can be slow if the number of blocks is large.
   for (const auto& pair : block_mapping) {
     int64_t src_block_number = pair.first;
     int64_t dst_block_number = pair.second;
@@ -43,38 +59,38 @@ void swap_blocks(
   }
 }
 
-namespace cacheflow {
+namespace vllm {
 
 // Grid: (num_layers, num_pairs)
 template<typename scalar_t>
 __global__ void copy_blocks_kernel(
   int64_t* key_cache_ptrs,
   int64_t* value_cache_ptrs,
-  const int* __restrict__ block_mapping,
+  const int64_t* __restrict__ block_mapping,
   const int numel_per_block) {
   const int layer_idx = blockIdx.x;
   const int pair_idx = blockIdx.y;
 
   scalar_t* key_cache = reinterpret_cast<scalar_t*>(key_cache_ptrs[layer_idx]);
   scalar_t* value_cache = reinterpret_cast<scalar_t*>(value_cache_ptrs[layer_idx]);
-  int src_block_number = block_mapping[2 * pair_idx];
-  int dst_block_number = block_mapping[2 * pair_idx + 1];
+  int64_t src_block_number = block_mapping[2 * pair_idx];
+  int64_t dst_block_number = block_mapping[2 * pair_idx + 1];
 
-  const int src_block_offset = src_block_number * numel_per_block;
-  const int dst_block_offset = dst_block_number * numel_per_block;
+  const int64_t src_block_offset = src_block_number * numel_per_block;
+  const int64_t dst_block_offset = dst_block_number * numel_per_block;
   for (int i = threadIdx.x; i < numel_per_block; i += blockDim.x) {
-    int src_offset = src_block_offset + i;
-    int dst_offset = dst_block_offset + i;
+    int64_t src_offset = src_block_offset + i;
+    int64_t dst_offset = dst_block_offset + i;
     key_cache[dst_offset] = key_cache[src_offset];
   }
   for (int i = threadIdx.x; i < numel_per_block; i += blockDim.x) {
-    int src_offset = src_block_offset + i;
-    int dst_offset = dst_block_offset + i;
+    int64_t src_offset = src_block_offset + i;
+    int64_t dst_offset = dst_block_offset + i;
     value_cache[dst_offset] = value_cache[src_offset];
   }
 }
 
-} // namespace cacheflow
+} // namespace vllm
 
 void copy_blocks(
   std::vector<torch::Tensor>& key_caches,
@@ -97,15 +113,15 @@ void copy_blocks(
     value_cache_ptrs[layer_idx] = reinterpret_cast<int64_t>(value_caches[layer_idx].data_ptr());
   }
   // Create block mapping array.
-  std::vector<int> block_mapping_vec;
+  std::vector<int64_t> block_mapping_vec;
   for (const auto& pair : block_mapping) {
-    int src_block_number = pair.first;
-    for (int dst_block_number : pair.second) {
+    int64_t src_block_number = pair.first;
+    for (int64_t dst_block_number : pair.second) {
       block_mapping_vec.push_back(src_block_number);
       block_mapping_vec.push_back(dst_block_number);
     }
   }
-  int* block_mapping_array = block_mapping_vec.data();
+  int64_t* block_mapping_array = block_mapping_vec.data();
   int num_pairs = block_mapping_vec.size() / 2;
 
   // Move the data structures to the GPU.
@@ -115,67 +131,144 @@ void copy_blocks(
   torch::Tensor value_cache_ptrs_tensor = torch::from_blob(
     value_cache_ptrs, {num_layers}, torch::kInt64).to(cache_device);
   torch::Tensor block_mapping_tensor = torch::from_blob(
-    block_mapping_array, {2 * num_pairs}, torch::kInt).to(cache_device);
+    block_mapping_array, {2 * num_pairs}, torch::kInt64).to(cache_device);
 
   // Launch the kernel.
   const int numel_per_block = key_caches[0][0].numel();
   dim3 grid(num_layers, num_pairs);
   dim3 block(std::min(1024, numel_per_block));
+  const at::cuda::OptionalCUDAGuard device_guard(cache_device);
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
-  AT_DISPATCH_FLOATING_TYPES_AND_HALF(
+  VLLM_DISPATCH_FLOATING_AND_BYTE_TYPES(
     key_caches[0].scalar_type(), "copy_blocks_kernel", ([&] {
-      cacheflow::copy_blocks_kernel<scalar_t><<<grid, block, 0, stream>>>(
+      vllm::copy_blocks_kernel<scalar_t><<<grid, block, 0, stream>>>(
         key_cache_ptrs_tensor.data_ptr<int64_t>(),
         value_cache_ptrs_tensor.data_ptr<int64_t>(),
-        block_mapping_tensor.data_ptr<int>(),
+        block_mapping_tensor.data_ptr<int64_t>(),
         numel_per_block);
     }));
 }
 
-namespace cacheflow {
+namespace vllm {
 
-template<typename scalar_t>
+template<typename scalar_t, typename cache_t, bool is_fp8_e5m2_kv_cache>
 __global__ void reshape_and_cache_kernel(
-  const scalar_t* __restrict__ key,     // [num_tokens, num_heads, head_size]
-  const scalar_t* __restrict__ value,   // [num_tokens, num_heads, head_size]
-  scalar_t* __restrict__ key_cache,     // [num_blocks, num_heads, head_size/x, block_size, x]
-  scalar_t* __restrict__ value_cache,   // [num_blocks, num_heads, head_size, block_size]
-  const int* __restrict__ slot_mapping, // [num_tokens]
+  const scalar_t* __restrict__ key,           // [num_tokens, num_heads, head_size]
+  const scalar_t* __restrict__ value,         // [num_tokens, num_heads, head_size]
+  cache_t* __restrict__ key_cache,            // [num_blocks, num_heads, head_size/x, block_size, x]
+  cache_t* __restrict__ value_cache,          // [num_blocks, num_heads, head_size, block_size]
+  const int64_t* __restrict__ slot_mapping,   // [num_tokens]
   const int key_stride,
   const int value_stride,
   const int num_heads,
   const int head_size,
   const int block_size,
   const int x) {
-  const int token_idx = blockIdx.x;
-  const int slot_idx = slot_mapping[token_idx];
-  const int block_idx = slot_idx / block_size;
-  const int block_offset = slot_idx % block_size;
+  const int64_t token_idx = blockIdx.x;
+  const int64_t slot_idx = slot_mapping[token_idx];
+  if (slot_idx < 0) {
+    // Padding token that should be ignored.
+    return;
+  }
+
+  const int64_t block_idx = slot_idx / block_size;
+  const int64_t block_offset = slot_idx % block_size;
 
   const int n = num_heads * head_size;
   for (int i = threadIdx.x; i < n; i += blockDim.x) {
-    const int src_key_idx = token_idx * key_stride + i;
-    const int src_value_idx = token_idx * value_stride + i;
+    const int64_t src_key_idx = token_idx * key_stride + i;
+    const int64_t src_value_idx = token_idx * value_stride + i;
 
     const int head_idx = i / head_size;
     const int head_offset = i % head_size;
     const int x_idx = head_offset / x;
     const int x_offset = head_offset % x;
 
-    const int tgt_key_idx = block_idx * num_heads * (head_size / x) * block_size * x
-                            + head_idx * (head_size / x) * block_size * x
-                            + x_idx * block_size * x
-                            + block_offset * x
-                            + x_offset;
-    const int tgt_value_idx = block_idx * num_heads * head_size * block_size
-                              + head_idx * head_size * block_size
-                              + head_offset * block_size
-                              + block_offset;
-    key_cache[tgt_key_idx] = __ldg(&key[src_key_idx]);
-    value_cache[tgt_value_idx] = __ldg(&value[src_value_idx]);
+    const int64_t tgt_key_idx = block_idx * num_heads * (head_size / x) * block_size * x
+                                + head_idx * (head_size / x) * block_size * x
+                                + x_idx * block_size * x
+                                + block_offset * x
+                                + x_offset;
+    const int64_t tgt_value_idx = block_idx * num_heads * head_size * block_size
+                                  + head_idx * head_size * block_size
+                                  + head_offset * block_size
+                                  + block_offset;
+    scalar_t tgt_key = key[src_key_idx];
+    scalar_t tgt_value = value[src_value_idx];
+    if constexpr (is_fp8_e5m2_kv_cache) {
+#ifdef ENABLE_FP8_E5M2
+      key_cache[tgt_key_idx] = fp8_e5m2_unscaled::vec_conversion<uint8_t, scalar_t>(tgt_key);
+      value_cache[tgt_value_idx] = fp8_e5m2_unscaled::vec_conversion<uint8_t, scalar_t>(tgt_value);
+#else
+      assert(false);
+#endif
+    } else {
+      key_cache[tgt_key_idx] = tgt_key;
+      value_cache[tgt_value_idx] = tgt_value;
+    }
   }
 }
 
+} // namespace vllm
+
+#define CALL_RESHAPE_AND_CACHE(KV_T, CACHE_T, IS_FP8_E5M2_KV_CACHE)                                \
+  vllm::reshape_and_cache_kernel<KV_T, CACHE_T, IS_FP8_E5M2_KV_CACHE><<<grid, block, 0, stream>>>( \
+    reinterpret_cast<KV_T*>(key.data_ptr()),                                                       \
+    reinterpret_cast<KV_T*>(value.data_ptr()),                                                     \
+    reinterpret_cast<CACHE_T*>(key_cache.data_ptr()),                                              \
+    reinterpret_cast<CACHE_T*>(value_cache.data_ptr()),                                            \
+    slot_mapping.data_ptr<int64_t>(),                                                              \
+    key_stride,                                                                                    \
+    value_stride,                                                                                  \
+    num_heads,                                                                                     \
+    head_size,                                                                                     \
+    block_size,                                                                                    \
+    x);
+
+void reshape_and_cache(
+  torch::Tensor& key,           // [num_tokens, num_heads, head_size]
+  torch::Tensor& value,         // [num_tokens, num_heads, head_size]
+  torch::Tensor& key_cache,     // [num_blocks, num_heads, head_size/x, block_size, x]
+  torch::Tensor& value_cache,   // [num_blocks, num_heads, head_size, block_size]
+  torch::Tensor& slot_mapping,  // [num_tokens]
+  const std::string& kv_cache_dtype)
+{
+  int num_tokens = key.size(0);
+  int num_heads = key.size(1);
+  int head_size = key.size(2);
+  int block_size = key_cache.size(3);
+  int x = key_cache.size(4);
+
+  int key_stride = key.stride(0);
+  int value_stride = value.stride(0);
+
+  dim3 grid(num_tokens);
+  dim3 block(std::min(num_heads * head_size, 512));
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(key));
+  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+  if (kv_cache_dtype == "auto") {
+    if (key.dtype() == at::ScalarType::Float) {
+      CALL_RESHAPE_AND_CACHE(float, float, false);
+    } else if (key.dtype() == at::ScalarType::Half) {
+      CALL_RESHAPE_AND_CACHE(uint16_t, uint16_t, false);
+    } else if (key.dtype() == at::ScalarType::BFloat16) {
+      CALL_RESHAPE_AND_CACHE(__nv_bfloat16, __nv_bfloat16, false);
+    }
+  } else if (kv_cache_dtype == "fp8_e5m2") {
+    if (key.dtype() == at::ScalarType::Float) {
+      CALL_RESHAPE_AND_CACHE(float, uint8_t, true);
+    } else if (key.dtype() == at::ScalarType::Half) {
+      CALL_RESHAPE_AND_CACHE(uint16_t, uint8_t, true);
+    } else if (key.dtype() == at::ScalarType::BFloat16) {
+      CALL_RESHAPE_AND_CACHE(__nv_bfloat16, uint8_t, true);
+    }
+  } else {
+    TORCH_CHECK(false, "Unsupported data type of kv cache: ", kv_cache_dtype);
+  }
+}
+
+namespace vllm {
+
 // Grid: (num_blocks, block_size).
 template<typename scalar_t>
 __global__ void gather_cached_kv_kernel(
@@ -199,12 +292,12 @@ __global__ void gather_cached_kv_kernel(
     for (int i = threadIdx.x; i < num_tokens; i += blockDim.x) {
       const int tgt_key_idx = token_idx * key_stride + i;
       const int tgt_value_idx = token_idx * value_stride + i;
-  
+
       const int head_idx = i / head_size;
       const int head_offset = i % head_size;
       const int x_idx = head_offset / x;  // the offset of the [head_size/x] dimension
       const int x_offset = head_offset % x;
-  
+
       const int src_key_idx = block_idx * num_heads * (head_size / x) * block_size * x
                               + head_idx * (head_size / x) * block_size * x
                               + x_idx * block_size * x
@@ -215,8 +308,8 @@ __global__ void gather_cached_kv_kernel(
                                 + head_offset * block_size
                                 + block_offset;
 
-      key[tgt_key_idx] = __ldg(&key_cache[src_key_idx]);
-      value[tgt_value_idx] = __ldg(&value_cache[src_value_idx]);
+      key[tgt_key_idx] = VLLM_LDG(&key_cache[src_key_idx]);
+      value[tgt_value_idx] = VLLM_LDG(&value_cache[src_value_idx]);
     }
 }
 
@@ -281,8 +374,8 @@ __global__ void gather_cached_kv_kernel_optimized(
             src_key_indices[j] = src_key_idx;
             src_value_indices[j] = src_value_idx;
 
-            keys_to_store[j] = __ldg(&key_cache[src_key_idx]);
-            values_to_store[j] = __ldg(&value_cache[src_value_idx]);
+            keys_to_store[j] = VLLM_LDG(&key_cache[src_key_idx]);
+            values_to_store[j] = VLLM_LDG(&value_cache[src_value_idx]);
         }
 
         #pragma unroll
@@ -294,46 +387,7 @@ __global__ void gather_cached_kv_kernel_optimized(
     }
 }
 
-} // namespace cacheflow
-
-void reshape_and_cache(
-  torch::Tensor& key,           // [num_tokens, num_heads, head_size]
-  torch::Tensor& value,         // [num_tokens, num_heads, head_size]
-  torch::Tensor& key_cache,     // [num_blocks, num_heads, head_size/x, block_size, x]
-  torch::Tensor& value_cache,   // [num_blocks, num_heads, head_size, block_size]
-  torch::Tensor& slot_mapping)  // [num_tokens]
-{
-  int num_tokens = key.size(0);
-  int num_heads = key.size(1);
-  int head_size = key.size(2);
-  int block_size = key_cache.size(3);
-  int x = key_cache.size(4);
-
-  int key_stride = key.stride(0);
-  int value_stride = value.stride(0);
-
-  dim3 grid(num_tokens);
-  dim3 block(std::min(num_heads * head_size, 512));
-  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
-  AT_DISPATCH_FLOATING_TYPES_AND_HALF(
-    key.scalar_type(),
-    "reshape_and_cache_kernel",
-    [&] {
-      cacheflow::reshape_and_cache_kernel<scalar_t><<<grid, block, 0, stream>>>(
-        key.data_ptr<scalar_t>(),
-        value.data_ptr<scalar_t>(),
-        key_cache.data_ptr<scalar_t>(),
-        value_cache.data_ptr<scalar_t>(),
-        slot_mapping.data_ptr<int>(),
-        key_stride,
-        value_stride,
-        num_heads,
-        head_size,
-        block_size,
-        x);
-    });
-}
-
+} // namespace vllm
 
 void gather_cached_kv(
   torch::Tensor& key,           // [out] [num_tokens, num_heads, head_size]
@@ -353,12 +407,13 @@ void gather_cached_kv(
 
   dim3 grid(num_tokens);
   dim3 block(std::min(num_heads * head_size, 512));
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(key));
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
-  AT_DISPATCH_FLOATING_TYPES_AND_HALF(
+  VLLM_DISPATCH_FLOATING_AND_BYTE_TYPES(
     key.scalar_type(),
     "gather_cached_kv_kernel_optimized",
     [&] {
-      cacheflow::gather_cached_kv_kernel_optimized<scalar_t><<<grid, block, 0, stream>>>(
+      vllm::gather_cached_kv_kernel_optimized<scalar_t><<<grid, block, 0, stream>>>(
         key.data_ptr<scalar_t>(),
         value.data_ptr<scalar_t>(),
         key_cache.data_ptr<scalar_t>(),
@@ -372,3 +427,55 @@ void gather_cached_kv(
         x);
     });
 }
+
+namespace vllm {
+
+template<typename Tout, typename Tin>
+__global__ void convert_fp8_e5m2_kernel(
+  const Tin* __restrict__ src_cache,
+  Tout* __restrict__ dst_cache,
+  const int64_t block_stride) {
+  const int64_t block_idx = blockIdx.x;
+  for (int i = threadIdx.x; i < block_stride; i += blockDim.x) {
+    int64_t idx = block_idx * block_stride + i;
+#ifdef ENABLE_FP8_E5M2
+    dst_cache[idx] = fp8_e5m2_unscaled::vec_conversion<Tout, Tin>(src_cache[idx]);
+#else
+    assert(false);
+#endif
+  }
+}
+
+} // namespace vllm
+
+#define CALL_CONVERT_FP8_E5M2(Tout, Tin)                                 \
+  vllm::convert_fp8_e5m2_kernel<Tout, Tin><<<grid, block, 0, stream>>>(  \
+    reinterpret_cast<Tin*>(src_cache.data_ptr()),                        \
+    reinterpret_cast<Tout*>(dst_cache.data_ptr()),                       \
+    block_stride);
+
+void convert_fp8_e5m2(
+  torch::Tensor& src_cache,
+  torch::Tensor& dst_cache)
+{
+  int64_t num_blocks = src_cache.size(0);
+  int64_t block_stride = src_cache.stride(0);
+
+  dim3 grid(num_blocks);
+  dim3 block(std::min(block_stride, int64_t(512)));
+  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+
+  if (src_cache.dtype() == at::ScalarType::Float) {
+    CALL_CONVERT_FP8_E5M2(uint8_t, float);
+  } else if (src_cache.dtype() == at::ScalarType::Half) {
+    CALL_CONVERT_FP8_E5M2(uint8_t, uint16_t);
+  } else if (src_cache.dtype() == at::ScalarType::BFloat16) {
+    CALL_CONVERT_FP8_E5M2(uint8_t, __nv_bfloat16);
+  } else if (dst_cache.dtype() == at::ScalarType::Float) {
+    CALL_CONVERT_FP8_E5M2(float, uint8_t);
+  } else if (dst_cache.dtype() == at::ScalarType::Half) {
+    CALL_CONVERT_FP8_E5M2(uint16_t, uint8_t);
+  } else if (dst_cache.dtype() == at::ScalarType::BFloat16) {
+    CALL_CONVERT_FP8_E5M2(__nv_bfloat16, uint8_t);
+  }
+}

csrc/cuda_compat.h

@@ -0,0 +1,28 @@
+#pragma once
+
+#ifndef USE_ROCM
+  #define VLLM_LDG(arg) __ldg(arg)
+#else
+  #define VLLM_LDG(arg) *(arg)
+#endif
+
+#ifndef USE_ROCM
+  #define VLLM_SHFL_XOR_SYNC(var, lane_mask) __shfl_xor_sync(uint32_t(-1), var, lane_mask)
+#else
+  #define VLLM_SHFL_XOR_SYNC(var, lane_mask) __shfl_xor(var, lane_mask)
+#endif
+
+#ifndef USE_ROCM
+  #define VLLM_SHFL_SYNC(var, src_lane) __shfl_sync(uint32_t(-1), var, src_lane)
+#else
+  #define VLLM_SHFL_SYNC(var, src_lane) __shfl(var, src_lane)
+#endif
+
+#ifndef USE_ROCM
+  #define VLLM_DevFuncAttribute_SET_MaxDynamicSharedMemorySize(FUNC, VAL) \
+    cudaFuncSetAttribute(FUNC, cudaFuncAttributeMaxDynamicSharedMemorySize, VAL)
+#else
+  #define VLLM_DevFuncAttribute_SET_MaxDynamicSharedMemorySize(FUNC, VAL) \
+    hipFuncSetAttribute(FUNC, hipFuncAttributeMaxDynamicSharedMemorySize, VAL)
+#endif
+

csrc/cuda_utils.h

@@ -0,0 +1,10 @@
+#pragma once
+
+#include <torch/extension.h>
+
+int get_device_attribute(
+    int attribute,
+    int device_id);
+
+int get_max_shared_memory_per_block_device_attribute(
+    int device_id);

csrc/cuda_utils_kernels.cu

@@ -0,0 +1,35 @@
+#ifdef USE_ROCM
+  #include <hip/hip_runtime.h>
+  #include <hip/hip_runtime_api.h>
+#endif
+int get_device_attribute(
+    int attribute,
+    int device_id)
+{
+    int device, value;
+    if (device_id < 0) {
+        cudaGetDevice(&device);
+    }
+    else {
+        device = device_id;
+    }
+    cudaDeviceGetAttribute(&value, static_cast<cudaDeviceAttr>(attribute), device);
+    return value;
+}
+
+
+int get_max_shared_memory_per_block_device_attribute(
+    int device_id)
+{
+int attribute;    
+// https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__TYPES.html
+// cudaDevAttrMaxSharedMemoryPerBlockOptin = 97 if not is_hip() else 74
+
+#ifdef USE_ROCM
+    attribute = hipDeviceAttributeMaxSharedMemoryPerBlock;
+#else
+    attribute = cudaDevAttrMaxSharedMemoryPerBlockOptin;
+#endif
+
+    return get_device_attribute(attribute, device_id);
+}

csrc/custom_all_reduce.cu

@@ -0,0 +1,148 @@
+#include <ATen/cuda/Exceptions.h>
+#include <c10/cuda/CUDAGuard.h>
+#include <c10/cuda/CUDAStream.h>
+#include <torch/extension.h>
+
+#include "custom_all_reduce.cuh"
+
+// fake pointer type
+using fptr_t = uint64_t;
+static_assert(sizeof(void *) == sizeof(fptr_t));
+
+fptr_t init_custom_ar(torch::Tensor &meta, torch::Tensor &rank_data,
+                      const std::vector<std::string> &handles,
+                      const std::vector<int64_t> &offsets, int rank,
+                      bool full_nvlink) {
+  int world_size = offsets.size();
+  if (world_size > 8)
+    throw std::invalid_argument("world size > 8 is not supported");
+  if (world_size % 2 != 0)
+    throw std::invalid_argument("Odd num gpus is not supported for now");
+  if (world_size != handles.size())
+    throw std::invalid_argument(
+        "handles length should equal to offsets length");
+  if (rank < 0 || rank >= world_size)
+    throw std::invalid_argument("invalid rank passed in");
+
+  cudaIpcMemHandle_t ipc_handles[8];
+  for (int i = 0; i < world_size; i++) {
+    std::memcpy(&ipc_handles[i], handles[i].data(), sizeof(cudaIpcMemHandle_t));
+  }
+  return (fptr_t) new vllm::CustomAllreduce(
+      reinterpret_cast<vllm::Metadata *>(meta.data_ptr()), rank_data.data_ptr(),
+      rank_data.numel(), ipc_handles, offsets, rank, full_nvlink);
+}
+
+/**
+ * Make sure tensor t's data lies completely within ((char)t.data_ptr()) +
+ * t.numel() * t.element_size(). This is slightly weaker than t.is_contiguous()
+ * because it allows transpose of contiguous slice (i.e. slicing the first
+ * dimension). Currently, we require this because stride information is not
+ * passed into the kernels and we treat input tensors as flat.
+ *
+ * Examples
+ * A = torch.zeros(3, 3, 3)
+ * 1. A: OK
+ * 2. A[1:]: OK
+ * 3. A.permute(2, 0, 1): OK
+ * 4. A[1:].permute(2, 0, 1): OK
+ * 5. A[None].expand(2, -1, -1, -1): Not OK
+ * 6. A[:, 1:, 1:]: Not OK
+ */
+bool _is_weak_contiguous(torch::Tensor &t) {
+  return t.is_contiguous() ||
+         (t.storage().nbytes() - t.storage_offset() * t.element_size() ==
+          t.numel() * t.element_size());
+}
+
+bool should_custom_ar(torch::Tensor &inp, int max_size, int world_size,
+                      bool full_nvlink) {
+  auto inp_size = inp.numel() * inp.element_size();
+  // custom allreduce requires input byte size to be multiples of 16
+  if (inp_size % 16 != 0) return false;
+  if (!_is_weak_contiguous(inp)) return false;
+  if (world_size == 2 || full_nvlink) return inp_size <= max_size;
+  // 4 PCIE GPUs use 2 stage allreduce, and is only faster than NCCL when size
+  // <= 512k
+  return world_size <= 4 && inp_size <= 512 * 1024;
+}
+
+void _all_reduce(fptr_t _fa, torch::Tensor &inp, torch::Tensor &out,
+                 cudaStream_t stream) {
+  auto fa = reinterpret_cast<vllm::CustomAllreduce *>(_fa);
+  TORCH_CHECK(_is_weak_contiguous(out));
+  switch (out.scalar_type()) {
+    case at::ScalarType::Float: {
+      fa->allreduce<float>(stream, reinterpret_cast<float *>(inp.data_ptr()),
+                           reinterpret_cast<float *>(out.data_ptr()),
+                           out.numel());
+      break;
+    }
+    case at::ScalarType::Half: {
+      fa->allreduce<half>(stream, reinterpret_cast<half *>(inp.data_ptr()),
+                          reinterpret_cast<half *>(out.data_ptr()),
+                          out.numel());
+      break;
+    }
+#if (__CUDA_ARCH__ >= 800 || !defined(__CUDA_ARCH__))
+    case at::ScalarType::BFloat16: {
+      fa->allreduce<nv_bfloat16>(
+          stream, reinterpret_cast<nv_bfloat16 *>(inp.data_ptr()),
+          reinterpret_cast<nv_bfloat16 *>(out.data_ptr()), out.numel());
+      break;
+    }
+#endif
+    default:
+      throw std::runtime_error(
+          "custom allreduce only supports float32, float16 and bfloat16");
+  }
+}
+
+void all_reduce_reg(fptr_t _fa, torch::Tensor &inp, torch::Tensor &out) {
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(inp));
+  auto stream = c10::cuda::getCurrentCUDAStream().stream();
+  TORCH_CHECK_EQ(inp.scalar_type(), out.scalar_type());
+  TORCH_CHECK_EQ(inp.numel(), out.numel());
+  _all_reduce(_fa, inp, out, stream);
+}
+
+void all_reduce_unreg(fptr_t _fa, torch::Tensor &inp, torch::Tensor &reg_buffer,
+                      torch::Tensor &out) {
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(inp));
+  auto stream = c10::cuda::getCurrentCUDAStream().stream();
+
+  auto input_size = inp.numel() * inp.element_size();
+  TORCH_CHECK_EQ(inp.scalar_type(), out.scalar_type());
+  TORCH_CHECK_EQ(inp.numel(), out.numel());
+  TORCH_CHECK(input_size <= reg_buffer.numel() * reg_buffer.element_size(),
+              "registered buffer is too small to contain the input");
+  AT_CUDA_CHECK(cudaMemcpyAsync(reg_buffer.data_ptr(), inp.data_ptr(),
+                                input_size, cudaMemcpyDeviceToDevice, stream));
+  _all_reduce(_fa, reg_buffer, out, stream);
+}
+
+void dispose(fptr_t _fa) {
+  auto fa = reinterpret_cast<vllm::CustomAllreduce *>(_fa);
+  delete fa;
+}
+
+int meta_size() { return sizeof(vllm::Metadata); }
+
+void register_buffer(fptr_t _fa, torch::Tensor &t,
+                     const std::vector<std::string> &handles,
+                     const std::vector<int64_t> &offsets) {
+  auto fa = reinterpret_cast<vllm::CustomAllreduce *>(_fa);
+  fa->register_buffer(handles, offsets, t.data_ptr());
+}
+
+std::pair<std::vector<uint8_t>, std::vector<int64_t>> get_graph_buffer_ipc_meta(
+    fptr_t _fa) {
+  auto fa = reinterpret_cast<vllm::CustomAllreduce *>(_fa);
+  return fa->get_graph_buffer_ipc_meta();
+}
+
+void register_graph_buffers(fptr_t _fa, const std::vector<std::string> &handles,
+                            const std::vector<std::vector<int64_t>> &offsets) {
+  auto fa = reinterpret_cast<vllm::CustomAllreduce *>(_fa);
+  fa->register_graph_buffers(handles, offsets);
+}

csrc/custom_all_reduce.cuh

@@ -0,0 +1,562 @@
+#pragma once
+
+#include <cuda.h>
+#include <cuda_bf16.h>
+#include <cuda_fp16.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <limits>
+#include <map>
+#include <unordered_map>
+#include <vector>
+
+#define CUDACHECK(cmd)                                              \
+  do {                                                              \
+    cudaError_t e = cmd;                                            \
+    if (e != cudaSuccess) {                                         \
+      printf("Failed: Cuda error %s:%d '%s'\n", __FILE__, __LINE__, \
+             cudaGetErrorString(e));                                \
+      exit(EXIT_FAILURE);                                           \
+    }                                                               \
+  } while (0)
+
+namespace vllm {
+
+struct Signal {
+  alignas(64) union {
+    uint64_t flag;
+    unsigned char data[8];
+  } start;
+  alignas(64) union {
+    uint64_t flag;
+    unsigned char data[8];
+  } end;
+};
+
+struct Metadata {
+  alignas(128) Signal sg;
+  alignas(128) int counter;
+};
+static_assert(offsetof(Metadata, counter) == 128);
+static_assert(sizeof(Metadata) == 256);
+
+struct __align__(16) RankData { const void *__restrict__ ptrs[8]; };
+
+struct RankSignals {
+  volatile Signal *signals[8];
+};
+
+// like std::array, but aligned
+template <typename T, int sz>
+struct __align__(alignof(T) * sz) array_t {
+  T data[sz];
+  using type = T;
+  static constexpr int size = sz;
+};
+
+// use packed type to maximize memory efficiency
+// goal: generate ld.128 and st.128 instructions
+template <typename T>
+struct packed_t {
+  // the (P)acked type for load/store
+  using P = array_t<T, 16 / sizeof(T)>;
+  // the (A)ccumulator type for reduction
+  using A = array_t<float, 16 / sizeof(T)>;
+};
+
+#define DINLINE __device__ __forceinline__
+
+// scalar cast functions
+DINLINE float upcast_s(half val) { return __half2float(val); }
+
+template <typename T>
+DINLINE T downcast_s(float val);
+template <>
+DINLINE half downcast_s(float val) {
+  return __float2half(val);
+}
+
+// scalar add functions
+// for some reason when compiling with Pytorch, the + operator for half and
+// bfloat is disabled so we call the intrinsics directly
+DINLINE half &assign_add(half &a, half b) {
+  a = __hadd(a, b);
+  return a;
+}
+DINLINE float &assign_add(float &a, float b) { return a += b; }
+
+#if (__CUDA_ARCH__ >= 800 || !defined(__CUDA_ARCH__))
+DINLINE float upcast_s(nv_bfloat16 val) { return __bfloat162float(val); }
+template <>
+DINLINE nv_bfloat16 downcast_s(float val) {
+  return __float2bfloat16(val);
+}
+DINLINE nv_bfloat16 &assign_add(nv_bfloat16 &a, nv_bfloat16 b) {
+  a = __hadd(a, b);
+  return a;
+}
+#endif
+
+template <typename T, int N>
+DINLINE array_t<T, N> &packed_assign_add(array_t<T, N> &a, array_t<T, N> b) {
+#pragma unroll
+  for (int i = 0; i < N; i++) {
+    assign_add(a.data[i], b.data[i]);
+  }
+  return a;
+}
+
+template <typename T, int N>
+DINLINE array_t<float, N> upcast(array_t<T, N> val) {
+  if constexpr (std::is_same<T, float>::value) {
+    return val;
+  } else {
+    array_t<float, N> out;
+#pragma unroll
+    for (int i = 0; i < N; i++) {
+      out.data[i] = upcast_s(val.data[i]);
+    }
+    return out;
+  }
+}
+
+template <typename O>
+DINLINE O downcast(array_t<float, O::size> val) {
+  if constexpr (std::is_same<typename O::type, float>::value) {
+    return val;
+  } else {
+    O out;
+#pragma unroll
+    for (int i = 0; i < O::size; i++) {
+      out.data[i] = downcast_s<typename O::type>(val.data[i]);
+    }
+    return out;
+  }
+}
+
+// compute flag at compile time
+__host__ __device__ constexpr uint64_t compute_flag(int ngpus) {
+  auto m = std::numeric_limits<uint64_t>::max();
+  return m >> ((8 - ngpus) * 8);
+}
+
+template <int ngpus>
+DINLINE void start_sync(const RankSignals &sg, volatile Metadata *meta,
+                        int rank) {
+  constexpr auto FLAG = compute_flag(ngpus);
+  if (blockIdx.x == 0) {
+    if (threadIdx.x < ngpus)
+      // simultaneously write to the corresponding byte to all other ranks.
+      // Latency = 1 p2p write
+      sg.signals[threadIdx.x]->start.data[rank] = 255;
+    else if (threadIdx.x == 32)
+      // reset
+      meta->sg.end.flag = 0;
+  }
+  if (threadIdx.x == 0) {
+    while (meta->sg.start.flag != FLAG)
+      ;
+  }
+  __syncthreads();
+}
+
+template <int ngpus, bool final_sync = false>
+DINLINE void end_sync(const RankSignals &sg, volatile Metadata *meta,
+                      int rank) {
+  constexpr auto FLAG = compute_flag(ngpus);
+  __syncthreads();
+  __shared__ int num;
+  if (threadIdx.x == 0) num = atomicAdd((int *)&meta->counter, 1);
+  __syncthreads();
+
+  // Only the last completing block can perform the end synchronization
+  // This can ensures when the final busy wait ends, all ranks must have
+  // finished reading each other's buffer.
+  if (num == gridDim.x - 1) {
+    if (threadIdx.x == 32) {
+      // reset in a different warp
+      meta->counter = 0;
+      meta->sg.start.flag = 0;
+    } else if (threadIdx.x < ngpus) {
+      // simultaneously write to the corresponding byte to all other ranks.
+      // Latency = 1 p2p write
+      sg.signals[threadIdx.x]->end.data[rank] = 255;
+    }
+    // if this is the final sync, only one block needs it
+    // because kernel exit can serve as sync
+    if constexpr (final_sync) {
+      if (threadIdx.x == 0) {
+        while (meta->sg.end.flag != FLAG)
+          ;
+      }
+    }
+  }
+  if constexpr (!final_sync) {
+    if (threadIdx.x == 0) {
+      while (meta->sg.end.flag != FLAG)
+        ;
+    }
+    __syncthreads();
+  }
+}
+
+template <typename P, int ngpus, typename A>
+DINLINE P packed_reduce(const P *ptrs[], int idx) {
+  A tmp = upcast(ptrs[0][idx]);
+#pragma unroll
+  for (int i = 1; i < ngpus; i++) {
+    packed_assign_add(tmp, upcast(ptrs[i][idx]));
+  }
+  return downcast<P>(tmp);
+}
+
+template <typename T, int ngpus>
+__global__ void __launch_bounds__(512, 1)
+    cross_device_reduce_1stage(RankData *_dp, RankSignals sg,
+                               volatile Metadata *meta, T *__restrict__ result,
+                               int rank, int size) {
+  using P = typename packed_t<T>::P;
+  using A = typename packed_t<T>::A;
+  // note: we don't reorder the address so the accumulation order is the same
+  // for all ranks, ensuring bitwise identical results
+  auto dp = *_dp;
+  start_sync<ngpus>(sg, meta, rank);
+  // do the actual reduction
+  for (int idx = blockIdx.x * blockDim.x + threadIdx.x; idx < size;
+       idx += gridDim.x * blockDim.x) {
+    ((P *)result)[idx] =
+        packed_reduce<P, ngpus, A>((const P **)&dp.ptrs[0], idx);
+  }
+  end_sync<ngpus, true>(sg, meta, rank);
+}
+
+template <typename P>
+DINLINE P *get_tmp_buf(volatile Signal *sg) {
+  return (P *)(((Metadata *)sg) + 1);
+}
+
+template <typename T, int ngpus>
+__global__ void __launch_bounds__(512, 1)
+    cross_device_reduce_2stage(RankData *_dp, RankSignals sg,
+                               volatile Metadata *meta, T *__restrict__ result,
+                               int rank, int size) {
+  int tid = blockIdx.x * blockDim.x + threadIdx.x;
+  int stride = gridDim.x * blockDim.x;
+  using P = typename packed_t<T>::P;
+  using A = typename packed_t<T>::A;
+  int part = size / ngpus;
+  int start = rank * part;
+  int end = rank == ngpus - 1 ? size : start + part;
+  const P *ptrs[ngpus];
+  P *tmps[ngpus];
+#pragma unroll
+  for (int i = 0; i < ngpus; i++) {
+    int target = (rank + i) % ngpus;
+    ptrs[i] = (const P *)_dp->ptrs[target];
+    tmps[i] = get_tmp_buf<P>(sg.signals[target]);
+  }
+  auto tmp_out = tmps[0];
+  start_sync<ngpus>(sg, meta, rank);
+  // stage 1: reduce scatter
+  for (int idx = start + tid; idx < end; idx += stride) {
+    tmp_out[idx - start] = packed_reduce<P, ngpus, A>(ptrs, idx);
+  }
+  // Maybe TODO: replace this with per-block release-acquire
+  // can save about 1-2us (not a lot though)
+  end_sync<ngpus>(sg, meta, rank);
+
+  // stage 2: allgather
+  for (int idx = tid; idx < part; idx += stride) {
+#pragma unroll
+    for (int i = 0; i < ngpus; i++) {
+      int dst_idx = ((rank + i) % ngpus) * part + idx;
+      ((P *)result)[dst_idx] = tmps[i][idx];
+    }
+  }
+  // process the last larger partition
+  int remaining = size - part * ngpus;
+  if (tid < remaining) {
+    int dst_idx = tid + part * ngpus;
+    ((P *)result)[dst_idx] = get_tmp_buf<P>(sg.signals[ngpus - 1])[part + tid];
+  }
+
+  // faster than this
+  // for (int idx = tid; idx < size; idx += stride) {
+  //   int target_rank = idx / part;
+  //   if (target_rank == ngpus) target_rank -= 1;
+  //   ((P *)result)[idx] = tmps[target_rank][idx - target_rank * part];
+  // }
+}
+
+template <typename T, int ngpus>
+__global__ void __launch_bounds__(512, 1)
+    cross_device_reduce_half_butterfly(RankData *_dp, RankSignals sg,
+                                       volatile Metadata *meta,
+                                       T *__restrict__ result, int rank,
+                                       int size) {
+  int tid = blockIdx.x * blockDim.x + threadIdx.x;
+  int stride = gridDim.x * blockDim.x;
+  using P = typename packed_t<T>::P;
+  using A = typename packed_t<T>::A;
+  auto tmp_out = get_tmp_buf<P>(sg.signals[rank]);
+  constexpr int hg = ngpus / 2;
+  // Actually not quite half butterfly.
+  // This is an all-to-all within each group containing half of the ranks
+  // followed by cross-group add. Equivalent to half butterfly when there
+  // are 4 GPUs, a common case for PCIe cards like T4 and A10.
+  const P *ptrs[hg];
+  {
+    int start = rank - rank % hg;
+#pragma unroll
+    for (int i = 0; i < hg; i++) {
+      ptrs[i] = (const P *)_dp->ptrs[i + start];
+    }
+  }
+  start_sync<ngpus>(sg, meta, rank);
+  for (int idx = tid; idx < size; idx += stride) {
+    tmp_out[idx] = packed_reduce<P, hg, A>(ptrs, idx);
+  }
+  end_sync<ngpus>(sg, meta, rank);
+
+  auto src = get_tmp_buf<P>(sg.signals[(ngpus - 1) - rank % ngpus]);
+  // do the cross group reduction
+  for (int idx = tid; idx < size; idx += stride) {
+    auto tmp = tmp_out[idx];
+    packed_assign_add(tmp, src[idx]);
+    ((P *)result)[idx] = tmp;
+  }
+}
+
+using IPC_KEY = std::array<uint8_t, sizeof(cudaIpcMemHandle_t)>;
+static_assert(sizeof(IPC_KEY) == sizeof(cudaIpcMemHandle_t));
+static_assert(alignof(IPC_KEY) == alignof(cudaIpcMemHandle_t));
+
+class CustomAllreduce {
+ public:
+  int rank_;
+  int world_size_;
+  bool full_nvlink_;
+
+  // below are device pointers
+  RankSignals sg_;
+  std::unordered_map<void *, RankData *> buffers_;
+  Metadata *meta_;
+
+  // stores the registered device pointers from all ranks
+  RankData *d_rank_data_base_, *d_rank_data_end_;
+  std::vector<void *> graph_unreg_buffers_;
+  // a map from IPC handles to opened IPC pointers
+  std::map<IPC_KEY, char *> ipc_handles_;
+
+  /**
+   * meta is a pointer to device metadata and temporary buffer for allreduce.
+   *
+   * There's a total of sizeof(Metadata) of prefix before the actual data,
+   * so meta + 1 points to actual temporary buffer.
+   *
+   * note: this class does not own any device memory. Any required buffers
+   * are passed in from the constructor
+   */
+  CustomAllreduce(Metadata *meta, void *rank_data, size_t rank_data_sz,
+                  const cudaIpcMemHandle_t *handles,
+                  const std::vector<int64_t> &offsets, int rank,
+                  bool full_nvlink = true)
+      : rank_(rank),
+        world_size_(offsets.size()),
+        full_nvlink_(full_nvlink),
+        meta_(meta),
+        d_rank_data_base_(reinterpret_cast<RankData *>(rank_data)),
+        d_rank_data_end_(d_rank_data_base_ + rank_data_sz / sizeof(RankData)) {
+    for (int i = 0; i < world_size_; i++) {
+      Metadata *rank_meta;
+      if (i != rank_) {
+        char *handle = open_ipc_handle(&handles[i]);
+        handle += offsets[i];
+        rank_meta = (Metadata *)handle;
+      } else {
+        rank_meta = meta_;
+      }
+      sg_.signals[i] = &rank_meta->sg;
+    }
+  }
+
+  char *open_ipc_handle(const void *ipc_handle) {
+    auto [it, new_handle] =
+        ipc_handles_.insert({*((IPC_KEY *)ipc_handle), nullptr});
+    if (new_handle) {
+      char *ipc_ptr;
+      CUDACHECK(cudaIpcOpenMemHandle((void **)&ipc_ptr,
+                                     *((const cudaIpcMemHandle_t *)ipc_handle),
+                                     cudaIpcMemLazyEnablePeerAccess));
+      it->second = ipc_ptr;
+    }
+    return it->second;
+  }
+
+  std::pair<std::vector<uint8_t>, std::vector<int64_t>>
+  get_graph_buffer_ipc_meta() {
+    auto num_buffers = graph_unreg_buffers_.size();
+    auto handle_sz = sizeof(cudaIpcMemHandle_t);
+    std::vector<uint8_t> handles(handle_sz * num_buffers, 0);
+    std::vector<int64_t> offsets(num_buffers);
+    for (int i = 0; i < num_buffers; i++) {
+      auto ptr = graph_unreg_buffers_[i];
+      void *base_ptr;
+      // note: must share the base address of each allocation, or we get wrong
+      // address
+      if (cuPointerGetAttribute(&base_ptr,
+                                CU_POINTER_ATTRIBUTE_RANGE_START_ADDR,
+                                (CUdeviceptr)ptr) != CUDA_SUCCESS)
+        throw std::runtime_error("failed to get pointer attr");
+      CUDACHECK(cudaIpcGetMemHandle(
+          (cudaIpcMemHandle_t *)&handles[i * handle_sz], base_ptr));
+      offsets[i] = ((char *)ptr) - ((char *)base_ptr);
+    }
+    return std::make_pair(handles, offsets);
+  }
+
+  void check_rank_data_capacity(size_t num = 1) {
+    if (d_rank_data_base_ + num > d_rank_data_end_)
+      throw std::runtime_error(
+          "Rank data buffer is overflowed by " +
+          std::to_string(d_rank_data_base_ + num - d_rank_data_end_));
+  }
+
+  void register_buffer(const std::vector<std::string> &handles,
+                       const std::vector<int64_t> &offsets, void *self) {
+    check_rank_data_capacity();
+    RankData data;
+    for (int i = 0; i < world_size_; i++) {
+      if (i != rank_) {
+        char *handle = open_ipc_handle(handles[i].data());
+        handle += offsets[i];
+        data.ptrs[i] = handle;
+      } else {
+        data.ptrs[i] = self;
+      }
+    }
+    auto d_data = d_rank_data_base_++;
+    CUDACHECK(
+        cudaMemcpy(d_data, &data, sizeof(RankData), cudaMemcpyHostToDevice));
+    buffers_[self] = d_data;
+  }
+
+  // note: when registering graph buffers, we intentionally choose to not
+  // deduplicate the addresses. That means if the allocator reuses some
+  // addresses, they will be registered again. This is to account for the remote
+  // possibility of different allocation patterns between ranks. For example,
+  // rank 1 may get the same input address for the second allreduce, but rank 2
+  // got a different address. IPC handles have internal reference counting
+  // mechanism so overhead should be small.
+  void register_graph_buffers(
+      const std::vector<std::string> &handles,
+      const std::vector<std::vector<int64_t>> &offsets) {
+    auto num_buffers = graph_unreg_buffers_.size();
+    check_rank_data_capacity(num_buffers);
+    std::vector<RankData> rank_data(num_buffers);
+    for (int i = 0; i < num_buffers; i++) {
+      auto self_ptr = graph_unreg_buffers_[i];
+      auto &rd = rank_data[i];
+      for (int j = 0; j < world_size_; j++) {
+        if (j != rank_) {
+          char *handle =
+              open_ipc_handle(&handles[j][i * sizeof(cudaIpcMemHandle_t)]);
+          handle += offsets[j][i];
+          rd.ptrs[j] = handle;
+        } else {
+          rd.ptrs[j] = self_ptr;
+        }
+      }
+    }
+    CUDACHECK(cudaMemcpy(d_rank_data_base_, rank_data.data(),
+                         sizeof(RankData) * num_buffers,
+                         cudaMemcpyHostToDevice));
+    d_rank_data_base_ += num_buffers;
+    graph_unreg_buffers_.clear();
+  }
+
+  /**
+   * This is the result after careful grid search. Using 36 blocks give the best
+   * or close to the best runtime on the devices I tried: A100, A10, A30, T4,
+   * V100. You'll notice that NCCL kernels also only take a small amount of SMs.
+   * Not quite sure the underlying reason, but my guess is that too many SMs
+   * will cause contention on NVLink bus.
+   */
+  template <typename T>
+  void allreduce(cudaStream_t stream, T *input, T *output, int size,
+                 int threads = 512, int block_limit = 36) {
+    auto d = packed_t<T>::P::size;
+    if (size % d != 0)
+      throw std::runtime_error(
+          "custom allreduce currently requires input length to be multiple "
+          "of " +
+          std::to_string(d));
+
+    RankData *ptrs;
+    cudaStreamCaptureStatus status;
+    CUDACHECK(cudaStreamIsCapturing(stream, &status));
+    if (status == cudaStreamCaptureStatusActive) {
+      ptrs = d_rank_data_base_ + graph_unreg_buffers_.size();
+      graph_unreg_buffers_.push_back(input);
+    } else {
+      auto it = buffers_.find(input);
+      if (it == buffers_.end())
+        throw std::runtime_error(
+            "buffer address " +
+            std::to_string(reinterpret_cast<uint64_t>(input)) +
+            " is not registered!");
+      ptrs = it->second;
+    }
+
+    size /= d;
+    auto bytes = size * sizeof(typename packed_t<T>::P);
+    int blocks = std::min(block_limit, (size + threads - 1) / threads);
+#define KL(ngpus, name) \
+  name<T, ngpus>        \
+      <<<blocks, threads, 0, stream>>>(ptrs, sg_, meta_, output, rank_, size);
+#define REDUCE_CASE(ngpus)                            \
+  case ngpus: {                                       \
+    if (world_size_ == 2) {                           \
+      KL(ngpus, cross_device_reduce_1stage);          \
+    } else if (full_nvlink_) {                        \
+      if ((world_size_ <= 4 && bytes < 512 * 1024) || \
+          (world_size_ <= 8 && bytes < 256 * 1024)) { \
+        KL(ngpus, cross_device_reduce_1stage);        \
+      } else {                                        \
+        KL(ngpus, cross_device_reduce_2stage);        \
+      }                                               \
+    } else {                                          \
+      KL(ngpus, cross_device_reduce_half_butterfly);  \
+    }                                                 \
+    break;                                            \
+  }
+
+    switch (world_size_) {
+      REDUCE_CASE(2)
+      REDUCE_CASE(4)
+      REDUCE_CASE(6)
+      REDUCE_CASE(8)
+      default:
+        throw std::runtime_error(
+            "custom allreduce only supports num gpus in (2,4,6,8). Actual num "
+            "gpus = " +
+            std::to_string(world_size_));
+    }
+#undef REDUCE_CASE
+#undef KL
+  }
+
+  ~CustomAllreduce() {
+    for (auto [_, ptr] : ipc_handles_) {
+      CUDACHECK(cudaIpcCloseMemHandle(ptr));
+    }
+  }
+};
+/**
+ * To inspect PTX/SASS, copy paste this header file to compiler explorer and add
+ a template instantiation:
+ * template void CustomAllreduce::allreduce<half>(cudaStream_t, half *, half *,
+ int, int, int);
+*/
+}  // namespace vllm

csrc/custom_all_reduce_test.cu

@@ -0,0 +1,284 @@
+/**
+ * This is a standalone test for custom allreduce.
+ * To compile, make sure you have MPI and NCCL installed in your system.
+ * export MPI_HOME=XXX
+ * nvcc -O2 -arch=native -std=c++17 custom_all_reduce_test.cu -o
+ * custom_all_reduce_test -lnccl -I${MPI_HOME}/include -lmpi
+ *
+ * Warning: this C++ test is not designed to be very readable and was used
+ * during the rapid prototyping process.
+ *
+ * To run:
+ * mpirun -np 8 ./custom_all_reduce_test
+ */
+#include <cuda.h>
+#include <curand_kernel.h>
+#include <stdio.h>
+#include <stdlib.h>
+
+#include <limits>
+#include <vector>
+
+#include "cuda_profiler_api.h"
+#include "custom_all_reduce.cuh"
+#include "mpi.h"
+#include "nccl.h"
+
+#define MPICHECK(cmd)                                                  \
+  do {                                                                 \
+    int e = cmd;                                                       \
+    if (e != MPI_SUCCESS) {                                            \
+      printf("Failed: MPI error %s:%d '%d'\n", __FILE__, __LINE__, e); \
+      exit(EXIT_FAILURE);                                              \
+    }                                                                  \
+  } while (0)
+
+#define NCCLCHECK(cmd)                                              \
+  do {                                                              \
+    ncclResult_t r = cmd;                                           \
+    if (r != ncclSuccess) {                                         \
+      printf("Failed, NCCL error %s:%d '%s'\n", __FILE__, __LINE__, \
+             ncclGetErrorString(r));                                \
+      exit(EXIT_FAILURE);                                           \
+    }                                                               \
+  } while (0)
+
+__global__ void dummy_kernel() {
+  for (int i = 0; i < 100; i++) __nanosleep(1000000);  // 100ms
+}
+
+template <typename T>
+__global__ void set_data(T *data, int size, int myRank) {
+  for (int idx = blockIdx.x * blockDim.x + threadIdx.x; idx < size;
+       idx += gridDim.x * blockDim.x) {
+    data[idx] = myRank * 0.11f;
+  }
+}
+
+template <typename T>
+__global__ void convert_data(const T *data1, const T *data2, double *fdata1,
+                             double *fdata2, int size) {
+  for (int idx = blockIdx.x * blockDim.x + threadIdx.x; idx < size;
+       idx += gridDim.x * blockDim.x) {
+    fdata1[idx] = data1[idx];
+    fdata2[idx] = data2[idx];
+  }
+}
+
+__global__ void init_rand(curandState_t *state, int size, int nRanks) {
+  for (int idx = blockIdx.x * blockDim.x + threadIdx.x; idx < size;
+       idx += gridDim.x * blockDim.x) {
+    for (int i = 0; i < nRanks; i++) {
+      curand_init(i + 1, idx, 0, &state[idx * nRanks + i]);
+    }
+  }
+}
+
+template <typename T>
+__global__ void gen_data(curandState_t *state, T *data, double *ground_truth,
+                         int myRank, int nRanks, int size) {
+  for (int idx = blockIdx.x * blockDim.x + threadIdx.x; idx < size;
+       idx += gridDim.x * blockDim.x) {
+    double sum = 0.0;
+    for (int i = 0; i < nRanks; i++) {
+      double val = curand_uniform_double(&state[idx * nRanks + i]) * 4;
+      T hval = val;  // downcast first
+      sum += static_cast<double>(hval);
+      if (i == myRank) data[idx] = hval;
+    }
+    ground_truth[idx] = sum;
+  }
+}
+
+template <typename T>
+void run(int myRank, int nRanks, ncclComm_t &comm, int threads, int block_limit,
+         int data_size) {
+  T *result;
+  cudaStream_t stream;
+  CUDACHECK(cudaStreamCreateWithFlags(&stream, cudaStreamNonBlocking));
+  CUDACHECK(cudaMalloc(&result, data_size * sizeof(T)));
+  CUDACHECK(cudaMemset(result, 0, data_size * sizeof(T)));
+
+  cudaIpcMemHandle_t self_data_handle;
+  cudaIpcMemHandle_t data_handles[8];
+  vllm::Metadata *buffer;
+  T *self_data_copy;
+  /**
+   * Allocate IPC buffer
+   *
+   * The first section is a temporary buffer for storing intermediate allreduce
+   * results, if a particular algorithm requires it. The second section is for
+   * the input to the allreduce. The actual API takes the input pointer as an
+   * argument (that is, they can and usually should be allocated separately).
+   * But since the input pointers and the temporary buffer all require IPC
+   * registration, they are allocated and registered together in the test for
+   * convenience.
+   */
+  CUDACHECK(
+      cudaMalloc(&buffer, 2 * data_size * sizeof(T) + sizeof(vllm::Metadata)));
+  CUDACHECK(cudaMemset(buffer, 0,
+                       2 * data_size * sizeof(T) + sizeof(vllm::Metadata)));
+  CUDACHECK(cudaMalloc(&self_data_copy, data_size * sizeof(T)));
+  CUDACHECK(cudaIpcGetMemHandle(&self_data_handle, buffer));
+
+  MPICHECK(MPI_Allgather(&self_data_handle, sizeof(cudaIpcMemHandle_t),
+                         MPI_BYTE, data_handles, sizeof(cudaIpcMemHandle_t),
+                         MPI_BYTE, MPI_COMM_WORLD));
+
+  void *rank_data;
+  size_t rank_data_sz = 16 * 1024 * 1024;
+  CUDACHECK(cudaMalloc(&rank_data, rank_data_sz));
+  std::vector<int64_t> offsets(nRanks, 0);
+  vllm::CustomAllreduce fa(buffer, rank_data, rank_data_sz, data_handles,
+                           offsets, myRank);
+  auto *self_data =
+      reinterpret_cast<T *>(reinterpret_cast<char *>(buffer) +
+                            sizeof(vllm::Metadata) + data_size * sizeof(T));
+  // hack buffer registration
+  {
+    std::vector<std::string> handles;
+    handles.reserve(nRanks);
+    for (int i = 0; i < nRanks; i++) {
+      char *begin = (char *)&data_handles[i];
+      char *end = (char *)&data_handles[i + 1];
+      handles.emplace_back(begin, end);
+    }
+    std::vector<int64_t> offsets(
+        nRanks, sizeof(vllm::Metadata) + data_size * sizeof(T));
+    fa.register_buffer(handles, offsets, self_data);
+  }
+
+  double *ground_truth;
+  CUDACHECK(cudaMallocHost(&ground_truth, data_size * sizeof(double)));
+  curandState_t *states;
+  CUDACHECK(cudaMalloc(&states, sizeof(curandState_t) * nRanks * data_size));
+  init_rand<<<108, 1024, 0, stream>>>(states, data_size, nRanks);
+  gen_data<T><<<108, 1024, 0, stream>>>(states, self_data, ground_truth, myRank,
+                                        nRanks, data_size);
+  CUDACHECK(cudaMemcpyAsync(self_data_copy, self_data, data_size * sizeof(T),
+                            cudaMemcpyDeviceToDevice, stream));
+  cudaEvent_t start, stop;
+  CUDACHECK(cudaEventCreate(&start));
+  CUDACHECK(cudaEventCreate(&stop));
+
+  ncclDataType_t ncclDtype;
+  if (std::is_same<T, half>::value) {
+    ncclDtype = ncclFloat16;
+  } else if (std::is_same<T, nv_bfloat16>::value) {
+    ncclDtype = ncclBfloat16;
+  } else {
+    ncclDtype = ncclFloat;
+  }
+
+  dummy_kernel<<<1, 1, 0, stream>>>();
+  constexpr int warmup_iters = 5;
+  constexpr int num_iters = 25;
+  // warmup
+  for (int i = 0; i < warmup_iters; i++) {
+    NCCLCHECK(ncclAllReduce(result, result, data_size, ncclDtype, ncclSum, comm,
+                            stream));
+  }
+  CUDACHECK(cudaEventRecord(start, stream));
+  for (int i = 0; i < num_iters; i++) {
+    NCCLCHECK(ncclAllReduce(result, result, data_size, ncclDtype, ncclSum, comm,
+                            stream));
+  }
+  CUDACHECK(cudaEventRecord(stop, stream));
+  CUDACHECK(cudaStreamSynchronize(stream));
+  float allreduce_ms = 0;
+  cudaEventElapsedTime(&allreduce_ms, start, stop);
+
+  // if (myRank == 1) dummy_kernel<<<1, 1, 0, stream>>>();
+  // set_data<T><<<16, 1024, 0, stream>>>(self_data, data_size, myRank);
+
+  dummy_kernel<<<1, 1, 0, stream>>>();
+  // warm up
+  for (int i = 0; i < warmup_iters; i++) {
+    fa.allreduce<T>(stream, self_data, result, data_size, threads, block_limit);
+  }
+  CUDACHECK(cudaEventRecord(start, stream));
+  for (int i = 0; i < num_iters; i++) {
+    fa.allreduce<T>(stream, self_data, result, data_size, threads, block_limit);
+  }
+  CUDACHECK(cudaEventRecord(stop, stream));
+  CUDACHECK(cudaStreamSynchronize(stream));
+
+  float duration_ms = 0;
+  cudaEventElapsedTime(&duration_ms, start, stop);
+  if (myRank == 0)
+    printf(
+        "Rank %d done, nGPUs:%d, sz (kb): %d, %d, %d, my time:%.2fus, nccl "
+        "time:%.2fus\n",
+        myRank, nRanks, data_size * sizeof(T) / 1024, threads, block_limit,
+        duration_ms * 1e3 / num_iters, allreduce_ms * 1e3 / num_iters);
+
+  // And wait for all the queued up work to complete
+  CUDACHECK(cudaStreamSynchronize(stream));
+
+  NCCLCHECK(ncclAllReduce(self_data_copy, self_data, data_size, ncclDtype,
+                          ncclSum, comm, stream));
+
+  double *nccl_result, *my_result;
+  CUDACHECK(cudaMallocHost(&nccl_result, data_size * sizeof(double)));
+  CUDACHECK(cudaMallocHost(&my_result, data_size * sizeof(double)));
+
+  convert_data<T><<<108, 1024, 0, stream>>>(self_data, result, nccl_result,
+                                            my_result, data_size);
+  CUDACHECK(cudaStreamSynchronize(stream));
+
+  for (unsigned long j = 0; j < data_size; j++) {
+    auto diff = abs(nccl_result[j] - my_result[j]);
+    if (diff >= 1e-2) {
+      printf("Rank %d: Verification mismatch at %lld: %f != (my) %f, gt=%f\n",
+             myRank, j, nccl_result[j], my_result[j], ground_truth[j]);
+      break;
+    }
+  }
+
+  long double nccl_diffs = 0.0;
+  long double my_diffs = 0.0;
+  for (int j = 0; j < data_size; j++) {
+    nccl_diffs += abs(nccl_result[j] - ground_truth[j]);
+    my_diffs += abs(my_result[j] - ground_truth[j]);
+  }
+  if (myRank == 0)
+    std::cout << "average abs diffs: nccl: " << nccl_diffs / data_size
+              << " me: " << my_diffs / data_size << std::endl;
+
+  CUDACHECK(cudaFree(result));
+  CUDACHECK(cudaFree(self_data_copy));
+  CUDACHECK(cudaFree(rank_data));
+  CUDACHECK(cudaFree(buffer));
+  CUDACHECK(cudaFree(states));
+  CUDACHECK(cudaFreeHost(ground_truth));
+  CUDACHECK(cudaFreeHost(nccl_result));
+  CUDACHECK(cudaFreeHost(my_result));
+  CUDACHECK(cudaStreamDestroy(stream));
+}
+
+int main(int argc, char **argv) {
+  int nRanks, myRank;
+  MPICHECK(MPI_Init(&argc, &argv));
+  MPICHECK(MPI_Comm_rank(MPI_COMM_WORLD, &myRank));
+  MPICHECK(MPI_Comm_size(MPI_COMM_WORLD, &nRanks));
+  CUDACHECK(cudaSetDevice(myRank));
+  ncclUniqueId id;
+  ncclComm_t comm;
+  if (myRank == 0) ncclGetUniqueId(&id);
+  MPICHECK(MPI_Bcast(static_cast<void *>(&id), sizeof(id), MPI_BYTE, 0,
+                     MPI_COMM_WORLD));
+  NCCLCHECK(ncclCommInitRank(&comm, nRanks, id, myRank));
+
+  cudaProfilerStart();
+  // for (int threads : {256, 512}) {
+  //   for (int block_limit = 16; block_limit < 112; block_limit += 4) {
+  //     run<half>(myRank, nRanks, comm, threads, block_limit, 4096 * 1024);
+  //   }
+  // }
+  for (int sz = 512; sz <= (32 << 20); sz *= 2) {
+    run<half>(myRank, nRanks, comm, 512, 36, sz + 8 * 50);
+  }
+
+  cudaProfilerStop();
+  return EXIT_SUCCESS;
+}

csrc/dispatch_utils.h

@@ -0,0 +1,37 @@
+/*
+ * Adapted from
+ * https://github.com/pytorch/pytorch/blob/v2.0.1/aten/src/ATen/Dispatch.h
+ */
+#pragma once
+
+#include <torch/extension.h>
+
+#define VLLM_DISPATCH_CASE_FLOATING_TYPES(...)              \
+  AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__)      \
+  AT_DISPATCH_CASE(at::ScalarType::Half, __VA_ARGS__)       \
+  AT_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__)
+
+#define VLLM_DISPATCH_FLOATING_TYPES(TYPE, NAME, ...)             \
+  AT_DISPATCH_SWITCH(                                             \
+    TYPE, NAME, VLLM_DISPATCH_CASE_FLOATING_TYPES(__VA_ARGS__))
+
+#define VLLM_DISPATCH_CASE_FLOATING_AND_BYTE_TYPES(...)     \
+  AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__)      \
+  AT_DISPATCH_CASE(at::ScalarType::Half, __VA_ARGS__)       \
+  AT_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__)   \
+  AT_DISPATCH_CASE(at::ScalarType::Byte, __VA_ARGS__)
+
+#define VLLM_DISPATCH_FLOATING_AND_BYTE_TYPES(TYPE, NAME, ...)           \
+  AT_DISPATCH_SWITCH(                                                    \
+    TYPE, NAME, VLLM_DISPATCH_CASE_FLOATING_AND_BYTE_TYPES(__VA_ARGS__))
+    
+#define VLLM_DISPATCH_CASE_INTEGRAL_TYPES(...)             \
+  AT_DISPATCH_CASE(at::ScalarType::Byte, __VA_ARGS__)      \
+  AT_DISPATCH_CASE(at::ScalarType::Char, __VA_ARGS__)      \
+  AT_DISPATCH_CASE(at::ScalarType::Short, __VA_ARGS__)     \
+  AT_DISPATCH_CASE(at::ScalarType::Int, __VA_ARGS__)       \
+  AT_DISPATCH_CASE(at::ScalarType::Long, __VA_ARGS__)
+
+#define VLLM_DISPATCH_INTEGRAL_TYPES(TYPE, NAME, ...)             \
+  AT_DISPATCH_SWITCH(                                             \
+    TYPE, NAME, VLLM_DISPATCH_CASE_INTEGRAL_TYPES(__VA_ARGS__))

csrc/layernorm.cpp

@@ -1,14 +0,0 @@
-#include <torch/extension.h>
-
-void rms_norm(
-  torch::Tensor& out,
-  torch::Tensor& input,
-  torch::Tensor& weight,
-  float epsilon);
-
-PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
-  m.def(
-    "rms_norm",
-    &rms_norm,
-    "Apply Root Mean Square (RMS) Normalization to the input tensor.");
-}

csrc/layernorm_kernels.cu

@@ -1,15 +1,17 @@
 #include <torch/extension.h>
 #include <ATen/cuda/CUDAContext.h>
+#include <c10/cuda/CUDAGuard.h>
 
-#include "reduction_utils.h"
+#include "dispatch_utils.h"
+#include "reduction_utils.cuh"
 
-namespace cacheflow {
+namespace vllm {
 
 // TODO(woosuk): Further optimize this kernel.
 template<typename scalar_t>
 __global__ void rms_norm_kernel(
-  scalar_t* __restrict__ out,             // [num_tokens, hidden_size]
-  const scalar_t* __restrict__ input,     // [num_tokens, hidden_size]
+  scalar_t* __restrict__ out,             // [..., hidden_size]
+  const scalar_t* __restrict__ input,     // [..., hidden_size]
   const scalar_t* __restrict__ weight,    // [hidden_size]
   const float epsilon,
   const int num_tokens,
@@ -33,24 +35,55 @@ __global__ void rms_norm_kernel(
   }
 }
 
-} // namespace cacheflow
+// TODO: Further optimize this kernel.
+template<typename scalar_t>
+__global__ void fused_add_rms_norm_kernel(
+  scalar_t* __restrict__ input,           // [..., hidden_size]
+  scalar_t* __restrict__ residual,        // [..., hidden_size]
+  const scalar_t* __restrict__ weight,    // [hidden_size]
+  const float epsilon,
+  const int num_tokens,
+  const int hidden_size) {
+  __shared__ float s_variance;
+  float variance = 0.0f;
+
+  for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
+    float x = (float) input[blockIdx.x * hidden_size + idx];
+    x += (float) residual[blockIdx.x * hidden_size + idx];
+    variance += x * x;
+    residual[blockIdx.x * hidden_size + idx] = (scalar_t) x;
+  }
+  variance = blockReduceSum<float>(variance);
+  if (threadIdx.x == 0) {
+    s_variance = rsqrtf(variance / hidden_size + epsilon);
+  }
+  __syncthreads();
+
+  for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
+    float x = (float) residual[blockIdx.x * hidden_size + idx];
+    input[blockIdx.x * hidden_size + idx] = ((scalar_t) (x * s_variance)) * weight[idx];
+  }
+}
+
+} // namespace vllm
 
 void rms_norm(
-  torch::Tensor& out,      // [num_tokens, hidden_size]
-  torch::Tensor& input,    // [num_tokens, hidden_size]
+  torch::Tensor& out,      // [..., hidden_size]
+  torch::Tensor& input,    // [..., hidden_size]
   torch::Tensor& weight,   // [hidden_size]
   float epsilon) {
-  int num_tokens = input.size(0);
-  int hidden_size = input.size(1);
+  int hidden_size = input.size(-1);
+  int num_tokens = input.numel() / hidden_size;
 
   dim3 grid(num_tokens);
   dim3 block(std::min(hidden_size, 1024));
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(input));
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
-  AT_DISPATCH_FLOATING_TYPES_AND_HALF(
+  VLLM_DISPATCH_FLOATING_TYPES(
     input.scalar_type(),
     "rms_norm_kernel",
     [&] {
-      cacheflow::rms_norm_kernel<scalar_t><<<grid, block, 0, stream>>>(
+      vllm::rms_norm_kernel<scalar_t><<<grid, block, 0, stream>>>(
         out.data_ptr<scalar_t>(),
         input.data_ptr<scalar_t>(),
         weight.data_ptr<scalar_t>(),
@@ -59,3 +92,29 @@ void rms_norm(
         hidden_size);
     });
 }
+
+void fused_add_rms_norm(
+  torch::Tensor& input,    // [..., hidden_size]
+  torch::Tensor& residual, // [..., hidden_size]
+  torch::Tensor& weight,   // [hidden_size]
+  float epsilon) {
+  int hidden_size = input.size(-1);
+  int num_tokens = input.numel() / hidden_size;
+
+  dim3 grid(num_tokens);
+  dim3 block(std::min(hidden_size, 1024));
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(input));
+  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+  VLLM_DISPATCH_FLOATING_TYPES(
+    input.scalar_type(),
+    "fused_add_rms_norm_kernel",
+    [&] {
+      vllm::fused_add_rms_norm_kernel<scalar_t><<<grid, block, 0, stream>>>(
+        input.data_ptr<scalar_t>(),
+        residual.data_ptr<scalar_t>(),
+        weight.data_ptr<scalar_t>(),
+        epsilon,
+        num_tokens,
+        hidden_size);
+    });
+}

csrc/moe/moe_ops.cpp

@@ -0,0 +1,7 @@
+#include "moe_ops.h"
+
+#include <torch/extension.h>
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+  m.def("topk_softmax", &topk_softmax, "Apply topk softmax to the gating outputs.");
+}

csrc/moe/moe_ops.h

@@ -0,0 +1,9 @@
+#pragma once
+
+#include <torch/extension.h>
+
+void topk_softmax(
+  torch::Tensor& topk_weights,
+  torch::Tensor& topk_indices,
+  torch::Tensor& token_expert_indices,
+  torch::Tensor& gating_output);

csrc/moe/topk_softmax_kernels.cu

@@ -0,0 +1,499 @@
+/*
+ * Adapted from https://github.com/NVIDIA/TensorRT-LLM/blob/v0.7.1/cpp/tensorrt_llm/kernels/mixtureOfExperts/moe_kernels.cu
+ * Copyright (c) 2024, The vLLM team.
+ * SPDX-FileCopyrightText: Copyright (c) 1993-2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include <torch/extension.h>
+#include <ATen/cuda/CUDAContext.h>
+#include <c10/cuda/CUDAGuard.h>
+
+#include <cub/cub.cuh>
+#include <cub/util_type.cuh>
+
+namespace vllm {
+namespace moe {
+
+static constexpr int WARP_SIZE = 32;
+
+/// Aligned array type
+template <
+    typename T,
+    /// Number of elements in the array
+    int N,
+    /// Alignment requirement in bytes
+    int Alignment = sizeof(T) * N
+>
+class alignas(Alignment) AlignedArray {
+    float data[N];
+};
+
+// ====================== Softmax things ===============================
+// We have our own implementation of softmax here so we can support transposing the output
+// in the softmax kernel when we extend this module to support expert-choice routing.
+template <int TPB>
+__launch_bounds__(TPB) __global__
+    void moeSoftmax(const float* input, const bool* finished, float* output, const int num_cols)
+{
+    using BlockReduce = cub::BlockReduce<float, TPB>;
+    __shared__ typename BlockReduce::TempStorage tmpStorage;
+
+    __shared__ float normalizing_factor;
+    __shared__ float float_max;
+
+    const int thread_row_offset = blockIdx.x * num_cols;
+
+    cub::Sum sum;
+    float threadData(-FLT_MAX);
+
+    // Don't touch finished rows.
+    if ((finished != nullptr) && finished[blockIdx.x])
+    {
+        return;
+    }
+
+    for (int ii = threadIdx.x; ii < num_cols; ii += TPB)
+    {
+        const int idx = thread_row_offset + ii;
+        threadData = max(static_cast<float>(input[idx]), threadData);
+    }
+
+    const float maxElem = BlockReduce(tmpStorage).Reduce(threadData, cub::Max());
+    if (threadIdx.x == 0)
+    {
+        float_max = maxElem;
+    }
+    __syncthreads();
+
+    threadData = 0;
+
+    for (int ii = threadIdx.x; ii < num_cols; ii += TPB)
+    {
+        const int idx = thread_row_offset + ii;
+        threadData += exp((static_cast<float>(input[idx]) - float_max));
+    }
+
+    const auto Z = BlockReduce(tmpStorage).Reduce(threadData, sum);
+
+    if (threadIdx.x == 0)
+    {
+        normalizing_factor = 1.f / Z;
+    }
+    __syncthreads();
+
+    for (int ii = threadIdx.x; ii < num_cols; ii += TPB)
+    {
+        const int idx = thread_row_offset + ii;
+        const float val = exp((static_cast<float>(input[idx]) - float_max)) * normalizing_factor;
+        output[idx] = val;
+    }
+}
+
+template <int TPB>
+__launch_bounds__(TPB) __global__ void moeTopK(const float* inputs_after_softmax, const bool* finished, float* output,
+    int* indices, int* source_rows, const int num_experts, const int k, const int start_expert, const int end_expert)
+{
+
+    using cub_kvp = cub::KeyValuePair<int, float>;
+    using BlockReduce = cub::BlockReduce<cub_kvp, TPB>;
+    __shared__ typename BlockReduce::TempStorage tmpStorage;
+
+    cub_kvp thread_kvp;
+    cub::ArgMax arg_max;
+
+    const int num_rows = gridDim.x;
+    const int block_row = blockIdx.x;
+
+    const bool row_is_active = finished ? !finished[block_row] : true;
+    const int thread_read_offset = blockIdx.x * num_experts;
+    for (int k_idx = 0; k_idx < k; ++k_idx)
+    {
+        thread_kvp.key = 0;
+        thread_kvp.value = -1.f; // This is OK because inputs are probabilities
+
+        cub_kvp inp_kvp;
+        for (int expert = threadIdx.x; expert < num_experts; expert += TPB)
+        {
+            const int idx = thread_read_offset + expert;
+            inp_kvp.key = expert;
+            inp_kvp.value = inputs_after_softmax[idx];
+
+            for (int prior_k = 0; prior_k < k_idx; ++prior_k)
+            {
+                const int prior_winning_expert = indices[k * block_row + prior_k];
+
+                if (prior_winning_expert == expert)
+                {
+                    inp_kvp = thread_kvp;
+                }
+            }
+
+            thread_kvp = arg_max(inp_kvp, thread_kvp);
+        }
+
+        const cub_kvp result_kvp = BlockReduce(tmpStorage).Reduce(thread_kvp, arg_max);
+        if (threadIdx.x == 0)
+        {
+            // Ignore experts the node isn't responsible for with expert parallelism
+            const int expert = result_kvp.key;
+            const bool node_uses_expert = expert >= start_expert && expert < end_expert;
+            const bool should_process_row = row_is_active && node_uses_expert;
+
+            const int idx = k * block_row + k_idx;
+            output[idx] = result_kvp.value;
+            indices[idx] = should_process_row ? (expert - start_expert) : num_experts;
+            assert(indices[idx] >= 0);
+            source_rows[idx] = k_idx * num_rows + block_row;
+        }
+        __syncthreads();
+    }
+}
+
+// ====================== TopK softmax things ===============================
+
+/*
+  A Top-K gating softmax written to exploit when the number of experts in the MoE layers
+  are a small power of 2. This allows us to cleanly share the rows among the threads in
+  a single warp and eliminate communication between warps (so no need to use shared mem).
+
+  It fuses the softmax, max and argmax into a single kernel.
+
+  Limitations:
+  1) This implementation is intended for when the number of experts is a small power of 2.
+  2) This implementation assumes k is small, but will work for any k.
+*/
+
+template <int VPT, int NUM_EXPERTS, int WARPS_PER_CTA, int BYTES_PER_LDG>
+__launch_bounds__(WARPS_PER_CTA* WARP_SIZE) __global__
+    void topkGatingSoftmax(const float* input, const bool* finished, float* output, const int num_rows, int* indices,
+        int* source_rows, const int k, const int start_expert, const int end_expert)
+{
+    // We begin by enforcing compile time assertions and setting up compile time constants.
+    static_assert(VPT == (VPT & -VPT), "VPT must be power of 2");
+    static_assert(NUM_EXPERTS == (NUM_EXPERTS & -NUM_EXPERTS), "NUM_EXPERTS must be power of 2");
+    static_assert(BYTES_PER_LDG == (BYTES_PER_LDG & -BYTES_PER_LDG), "BYTES_PER_LDG must be power of 2");
+    static_assert(BYTES_PER_LDG <= 16, "BYTES_PER_LDG must be leq 16");
+
+    // Number of bytes each thread pulls in per load
+    static constexpr int ELTS_PER_LDG = BYTES_PER_LDG / sizeof(float);
+    static constexpr int ELTS_PER_ROW = NUM_EXPERTS;
+    static constexpr int THREADS_PER_ROW = ELTS_PER_ROW / VPT;
+    static constexpr int LDG_PER_THREAD = VPT / ELTS_PER_LDG;
+
+    // Restrictions based on previous section.
+    static_assert(VPT % ELTS_PER_LDG == 0, "The elements per thread must be a multiple of the elements per ldg");
+    static_assert(WARP_SIZE % THREADS_PER_ROW == 0, "The threads per row must cleanly divide the threads per warp");
+    static_assert(THREADS_PER_ROW == (THREADS_PER_ROW & -THREADS_PER_ROW), "THREADS_PER_ROW must be power of 2");
+    static_assert(THREADS_PER_ROW <= WARP_SIZE, "THREADS_PER_ROW can be at most warp size");
+
+    // We have NUM_EXPERTS elements per row. We specialize for small #experts
+    static constexpr int ELTS_PER_WARP = WARP_SIZE * VPT;
+    static constexpr int ROWS_PER_WARP = ELTS_PER_WARP / ELTS_PER_ROW;
+    static constexpr int ROWS_PER_CTA = WARPS_PER_CTA * ROWS_PER_WARP;
+
+    // Restrictions for previous section.
+    static_assert(ELTS_PER_WARP % ELTS_PER_ROW == 0, "The elts per row must cleanly divide the total elt per warp");
+
+    // ===================== From this point, we finally start computing run-time variables. ========================
+
+    // Compute CTA and warp rows. We pack multiple rows into a single warp, and a block contains WARPS_PER_CTA warps.
+    // This, each block processes a chunk of rows. We start by computing the start row for each block.
+    const int cta_base_row = blockIdx.x * ROWS_PER_CTA;
+
+    // Now, using the base row per thread block, we compute the base row per warp.
+    const int warp_base_row = cta_base_row + threadIdx.y * ROWS_PER_WARP;
+
+    // The threads in a warp are split into sub-groups that will work on a row.
+    // We compute row offset for each thread sub-group
+    const int thread_row_in_warp = threadIdx.x / THREADS_PER_ROW;
+    const int thread_row = warp_base_row + thread_row_in_warp;
+
+    // Threads with indices out of bounds should early exit here.
+    if (thread_row >= num_rows)
+    {
+        return;
+    }
+    const bool row_is_active = finished ? !finished[thread_row] : true;
+
+    // We finally start setting up the read pointers for each thread. First, each thread jumps to the start of the
+    // row it will read.
+    const float* thread_row_ptr = input + thread_row * ELTS_PER_ROW;
+
+    // Now, we compute the group each thread belong to in order to determine the first column to start loads.
+    const int thread_group_idx = threadIdx.x % THREADS_PER_ROW;
+    const int first_elt_read_by_thread = thread_group_idx * ELTS_PER_LDG;
+    const float* thread_read_ptr = thread_row_ptr + first_elt_read_by_thread;
+
+    // Determine the pointer type to use to read in the data depending on the BYTES_PER_LDG template param. In theory,
+    // this can support all powers of 2 up to 16.
+    // NOTE(woosuk): The original implementation uses CUTLASS aligned array here.
+    // We defined our own aligned array and use it here to avoid the dependency on CUTLASS.
+    using AccessType = AlignedArray<float, ELTS_PER_LDG>;
+
+    // Finally, we pull in the data from global mem
+    float row_chunk[VPT];
+    AccessType* row_chunk_vec_ptr = reinterpret_cast<AccessType*>(&row_chunk);
+    const AccessType* vec_thread_read_ptr = reinterpret_cast<const AccessType*>(thread_read_ptr);
+#pragma unroll
+    for (int ii = 0; ii < LDG_PER_THREAD; ++ii)
+    {
+        row_chunk_vec_ptr[ii] = vec_thread_read_ptr[ii * THREADS_PER_ROW];
+    }
+
+    // First, we perform a max reduce within the thread. We can do the max in fp16 safely (I think) and just
+    // convert to float afterwards for the exp + sum reduction.
+    float thread_max = row_chunk[0];
+#pragma unroll
+    for (int ii = 1; ii < VPT; ++ii)
+    {
+        thread_max = max(thread_max, row_chunk[ii]);
+    }
+
+// Now, we find the max within the thread group and distribute among the threads. We use a butterfly reduce.
+#pragma unroll
+    for (int mask = THREADS_PER_ROW / 2; mask > 0; mask /= 2)
+    {
+        thread_max = max(thread_max, __shfl_xor_sync(0xFFFFFFFF, thread_max, mask, THREADS_PER_ROW));
+    }
+
+    // From this point, thread max in all the threads have the max within the row.
+    // Now, we subtract the max from each element in the thread and take the exp. We also compute the thread local sum.
+    float row_sum = 0;
+#pragma unroll
+    for (int ii = 0; ii < VPT; ++ii)
+    {
+        row_chunk[ii] = expf(row_chunk[ii] - thread_max);
+        row_sum += row_chunk[ii];
+    }
+
+// Now, we perform the sum reduce within each thread group. Similar to the max reduce, we use a bufferfly pattern.
+#pragma unroll
+    for (int mask = THREADS_PER_ROW / 2; mask > 0; mask /= 2)
+    {
+        row_sum += __shfl_xor_sync(0xFFFFFFFF, row_sum, mask, THREADS_PER_ROW);
+    }
+
+    // From this point, all threads have the max and the sum for their rows in the thread_max and thread_sum variables
+    // respectively. Finally, we can scale the rows for the softmax. Technically, for top-k gating we don't need to
+    // compute the entire softmax row. We can likely look at the maxes and only compute for the top-k values in the row.
+    // However, this kernel will likely not be a bottle neck and it seems better to closer match torch and find the
+    // argmax after computing the softmax.
+    const float reciprocal_row_sum = 1.f / row_sum;
+
+#pragma unroll
+    for (int ii = 0; ii < VPT; ++ii)
+    {
+        row_chunk[ii] = row_chunk[ii] * reciprocal_row_sum;
+    }
+
+    // Now, softmax_res contains the softmax of the row chunk. Now, I want to find the topk elements in each row, along
+    // with the max index.
+    int start_col = first_elt_read_by_thread;
+    static constexpr int COLS_PER_GROUP_LDG = ELTS_PER_LDG * THREADS_PER_ROW;
+
+    for (int k_idx = 0; k_idx < k; ++k_idx)
+    {
+        // First, each thread does the local argmax
+        float max_val = row_chunk[0];
+        int expert = start_col;
+#pragma unroll
+        for (int ldg = 0, col = start_col; ldg < LDG_PER_THREAD; ++ldg, col += COLS_PER_GROUP_LDG)
+        {
+#pragma unroll
+            for (int ii = 0; ii < ELTS_PER_LDG; ++ii)
+            {
+                float val = row_chunk[ldg * ELTS_PER_LDG + ii];
+
+                // No check on the experts here since columns with the smallest index are processed first and only
+                // updated if > (not >=)
+                if (val > max_val)
+                {
+                    max_val = val;
+                    expert = col + ii;
+                }
+            }
+        }
+
+// Now, we perform the argmax reduce. We use the butterfly pattern so threads reach consensus about the max.
+// This will be useful for K > 1 so that the threads can agree on "who" had the max value. That thread can
+// then blank out their max with -inf and the warp can run more iterations...
+#pragma unroll
+        for (int mask = THREADS_PER_ROW / 2; mask > 0; mask /= 2)
+        {
+            float other_max = __shfl_xor_sync(0xFFFFFFFF, max_val, mask, THREADS_PER_ROW);
+            int other_expert = __shfl_xor_sync(0xFFFFFFFF, expert, mask, THREADS_PER_ROW);
+
+            // We want lower indices to "win" in every thread so we break ties this way
+            if (other_max > max_val || (other_max == max_val && other_expert < expert))
+            {
+                max_val = other_max;
+                expert = other_expert;
+            }
+        }
+
+        // Write the max for this k iteration to global memory.
+        if (thread_group_idx == 0)
+        {
+            // Add a guard to ignore experts not included by this node
+            const bool node_uses_expert = expert >= start_expert && expert < end_expert;
+            const bool should_process_row = row_is_active && node_uses_expert;
+
+            // The lead thread from each sub-group will write out the final results to global memory. (This will be a
+            // single) thread per row of the input/output matrices.
+            const int idx = k * thread_row + k_idx;
+            output[idx] = max_val;
+            indices[idx] = should_process_row ? (expert - start_expert) : NUM_EXPERTS;
+            source_rows[idx] = k_idx * num_rows + thread_row;
+        }
+
+        // Finally, we clear the value in the thread with the current max if there is another iteration to run.
+        if (k_idx + 1 < k)
+        {
+            const int ldg_group_for_expert = expert / COLS_PER_GROUP_LDG;
+            const int thread_to_clear_in_group = (expert / ELTS_PER_LDG) % THREADS_PER_ROW;
+
+            // Only the thread in the group which produced the max will reset the "winning" value to -inf.
+            if (thread_group_idx == thread_to_clear_in_group)
+            {
+                const int offset_for_expert = expert % ELTS_PER_LDG;
+                // Safe to set to any negative value since row_chunk values must be between 0 and 1.
+                row_chunk[ldg_group_for_expert * ELTS_PER_LDG + offset_for_expert] = -10000.f;
+            }
+        }
+    }
+}
+
+namespace detail
+{
+// Constructs some constants needed to partition the work across threads at compile time.
+template <int EXPERTS, int BYTES_PER_LDG>
+struct TopkConstants
+{
+    static constexpr int ELTS_PER_LDG = BYTES_PER_LDG / sizeof(float);
+    static_assert(EXPERTS / (ELTS_PER_LDG * WARP_SIZE) == 0 || EXPERTS % (ELTS_PER_LDG * WARP_SIZE) == 0, "");
+    static constexpr int VECs_PER_THREAD = std::max(1, EXPERTS / (ELTS_PER_LDG * WARP_SIZE));
+    static constexpr int VPT = VECs_PER_THREAD * ELTS_PER_LDG;
+    static constexpr int THREADS_PER_ROW = EXPERTS / VPT;
+    static constexpr int ROWS_PER_WARP = WARP_SIZE / THREADS_PER_ROW;
+};
+} // namespace detail
+
+template <int EXPERTS, int WARPS_PER_TB>
+void topkGatingSoftmaxLauncherHelper(const float* input, const bool* finished, float* output, int* indices,
+    int* source_row, const int num_rows, const int k, const int start_expert, const int end_expert, cudaStream_t stream)
+{
+    static constexpr std::size_t MAX_BYTES_PER_LDG = 16;
+
+    static constexpr int BYTES_PER_LDG = std::min(MAX_BYTES_PER_LDG, sizeof(float) * EXPERTS);
+    using Constants = detail::TopkConstants<EXPERTS, BYTES_PER_LDG>;
+    static constexpr int VPT = Constants::VPT;
+    static constexpr int ROWS_PER_WARP = Constants::ROWS_PER_WARP;
+    const int num_warps = (num_rows + ROWS_PER_WARP - 1) / ROWS_PER_WARP;
+    const int num_blocks = (num_warps + WARPS_PER_TB - 1) / WARPS_PER_TB;
+
+    dim3 block_dim(WARP_SIZE, WARPS_PER_TB);
+    topkGatingSoftmax<VPT, EXPERTS, WARPS_PER_TB, BYTES_PER_LDG><<<num_blocks, block_dim, 0, stream>>>(
+        input, finished, output, num_rows, indices, source_row, k, start_expert, end_expert);
+}
+
+#define LAUNCH_SOFTMAX(NUM_EXPERTS, WARPS_PER_TB)                       \
+    topkGatingSoftmaxLauncherHelper<NUM_EXPERTS, WARPS_PER_TB>(         \
+        gating_output, nullptr, topk_weights, topk_indicies,            \
+        token_expert_indices, num_tokens, topk, 0, num_experts,         \
+        stream);
+
+void topkGatingSoftmaxKernelLauncher(
+    const float* gating_output,
+    float* topk_weights,
+    int* topk_indicies,
+    int* token_expert_indices,
+    float* softmax_workspace,
+    const int num_tokens,
+    const int num_experts,
+    const int topk,
+    cudaStream_t stream) {
+    static constexpr int WARPS_PER_TB = 4;
+    switch (num_experts) {
+        case 1:
+            LAUNCH_SOFTMAX(1, WARPS_PER_TB);
+            break;
+        case 2:
+            LAUNCH_SOFTMAX(2, WARPS_PER_TB);
+            break;
+        case 4:
+            LAUNCH_SOFTMAX(4, WARPS_PER_TB);
+            break;
+        case 8:
+            LAUNCH_SOFTMAX(8, WARPS_PER_TB);
+            break;
+        case 16:
+            LAUNCH_SOFTMAX(16, WARPS_PER_TB);
+            break;
+        case 32:
+            LAUNCH_SOFTMAX(32, WARPS_PER_TB);
+            break;
+        case 64:
+            LAUNCH_SOFTMAX(64, WARPS_PER_TB);
+            break;
+        case 128:
+            LAUNCH_SOFTMAX(128, WARPS_PER_TB);
+            break;
+        case 256:
+            LAUNCH_SOFTMAX(256, WARPS_PER_TB);
+            break;
+        default: {
+            TORCH_CHECK(softmax_workspace != nullptr,
+                "softmax_workspace must be provided for num_experts that are not a power of 2.");
+            static constexpr int TPB = 256;
+            moeSoftmax<TPB><<<num_tokens, TPB, 0, stream>>>(
+                gating_output, nullptr, softmax_workspace, num_experts);
+            moeTopK<TPB><<<num_tokens, TPB, 0, stream>>>(
+                softmax_workspace, nullptr, topk_weights, topk_indicies, token_expert_indices,
+                num_experts, topk, 0, num_experts);
+        }
+    }
+}
+
+} // namespace moe
+} // namespace vllm
+
+void topk_softmax(
+    torch::Tensor& topk_weights,                // [num_tokens, topk]
+    torch::Tensor& topk_indices,                // [num_tokens, topk]
+    torch::Tensor& token_expert_indices,        // [num_tokens, topk]
+    torch::Tensor& gating_output)               // [num_tokens, num_experts]
+{
+    const int num_experts = gating_output.size(-1);
+    const int num_tokens = gating_output.numel() / num_experts;
+    const int topk = topk_weights.size(-1);
+
+    const bool is_pow_2 = (num_experts != 0) && ((num_experts & (num_experts - 1)) == 0);
+    const bool needs_workspace = !is_pow_2 || num_experts > 256;
+    const int64_t workspace_size = needs_workspace ? num_tokens * num_experts : 0;
+
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(gating_output));
+    const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+    torch::Tensor softmax_workspace = torch::empty({workspace_size}, gating_output.options());
+    vllm::moe::topkGatingSoftmaxKernelLauncher(
+        gating_output.data_ptr<float>(),
+        topk_weights.data_ptr<float>(),
+        topk_indices.data_ptr<int>(),
+        token_expert_indices.data_ptr<int>(),
+        softmax_workspace.data_ptr<float>(),
+        num_tokens,
+        num_experts,
+        topk,
+        stream);
+}

csrc/moe_align_block_size_kernels.cu

@@ -0,0 +1,108 @@
+#include <torch/extension.h>
+#include <ATen/cuda/CUDAContext.h>
+
+#include <ATen/ATen.h>
+#include <THC/THCAtomics.cuh>
+
+#include "cuda_compat.h"
+#include "dispatch_utils.h"
+
+const static size_t NUM_MAX_EXPERTS = 64;
+#define CEILDIV(x,y) (((x) + (y) - 1) / (y))
+
+namespace vllm {
+template <typename scalar_t>
+__global__ void moe_align_block_size_kernel(scalar_t *__restrict__ topk_ids, 
+                                int32_t *sorted_token_ids, 
+                                int32_t *expert_ids, 
+                                int32_t *total_tokens_post_pad,
+                                int32_t num_experts, 
+                                int32_t block_size, 
+                                size_t numel) {
+    const size_t tokens_per_thread = CEILDIV(numel, blockDim.x);
+    const size_t start_idx = threadIdx.x * tokens_per_thread;
+    __shared__ int32_t tokens_cnts[NUM_MAX_EXPERTS + 1][NUM_MAX_EXPERTS];
+    __shared__ int32_t cumsum[NUM_MAX_EXPERTS + 1];
+    for (int i = 0; i < num_experts; ++i) {
+        tokens_cnts[threadIdx.x + 1][i] = 0;
+    }
+
+    /**
+    * In the first step we compute token_cnts[thread_index + 1][expert_index],
+    * which counts how many tokens in the token shard of thread_index are assigned
+    * to expert expert_index.
+    */
+    for (int i = start_idx; i < numel && i < start_idx + tokens_per_thread; ++i) {
+        ++tokens_cnts[threadIdx.x + 1][topk_ids[i]]; 
+    }
+
+    __syncthreads();
+
+    // For each expert we accumulate the token counts from the different threads.
+    tokens_cnts[0][threadIdx.x] = 0;
+    for (int i = 1; i <= blockDim.x; ++i) {
+        tokens_cnts[i][threadIdx.x] += tokens_cnts[i-1][threadIdx.x];
+    }
+
+    __syncthreads();
+    
+    // We accumulate the token counts of all experts in thread 0.
+    if (threadIdx.x == 0) {
+        cumsum[0] = 0;
+        for (int i = 1; i <= num_experts; ++i) {
+            cumsum[i] = cumsum[i-1] + CEILDIV(tokens_cnts[blockDim.x][i - 1], block_size) * block_size;
+        }
+        *total_tokens_post_pad = cumsum[num_experts];
+    }
+
+    __syncthreads();
+
+    /**
+    * For each expert, each thread processes the tokens of the corresponding blocks
+    * and stores the corresponding expert_id for each block.
+    */
+    for (int i = cumsum[threadIdx.x];i < cumsum[threadIdx.x + 1];i += block_size) {
+        expert_ids[i / block_size] = threadIdx.x;
+    }
+    
+    /**
+    * Each thread processes a token shard, calculating the index of each token after
+    * sorting by expert number. Given the example topk_ids = [0,1,2,1,2,3,0,3,4] and
+    * block_size = 4, then the output would be [0, 6, *, *, 1, 3, *, *, 2, 4, *, *, 5, 7, *, *, 8, *, *, *],
+    * where * represents a padding value(preset in python).
+    */
+    for (int i = start_idx; i < numel && i < start_idx + tokens_per_thread; ++i) {
+        int32_t expert_id = topk_ids[i];
+        /** The cumsum[expert_id] stores the starting index of the tokens that the
+        * expert with expert_id needs to process, and tokens_cnts[threadIdx.x][expert_id]
+        * stores the indices of the tokens processed by the expert with expert_id within
+        * the current thread's token shard.
+        */
+        int32_t rank_post_pad = tokens_cnts[threadIdx.x][expert_id] + cumsum[expert_id];
+        sorted_token_ids[rank_post_pad] = i;
+        ++tokens_cnts[threadIdx.x][expert_id];
+    }
+}
+}
+
+void moe_align_block_size(
+    torch::Tensor topk_ids,
+    int num_experts,
+    int block_size,
+    torch::Tensor sorted_token_ids,
+    torch::Tensor experts_ids,
+    torch::Tensor num_tokens_post_pad) {
+    const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+    assert(num_experts <= NUM_MAX_EXPERTS);
+    VLLM_DISPATCH_INTEGRAL_TYPES(
+        topk_ids.scalar_type(), "moe_align_block_size_kernel", [&] {
+        vllm::moe_align_block_size_kernel<scalar_t><<<1, num_experts, 0, stream>>>(
+            topk_ids.data_ptr<scalar_t>(), 
+            sorted_token_ids.data_ptr<int32_t>(), 
+            experts_ids.data_ptr<int32_t>(), 
+            num_tokens_post_pad.data_ptr<int32_t>(), 
+            num_experts,
+            block_size,
+            topk_ids.numel());
+    });
+}

csrc/ops.h

@@ -0,0 +1,130 @@
+#pragma once
+
+#include <torch/extension.h>
+
+void paged_attention_v1(
+  torch::Tensor& out,
+  torch::Tensor& query,
+  torch::Tensor& key_cache,
+  torch::Tensor& value_cache,
+  int num_kv_heads,
+  float scale,
+  torch::Tensor& block_tables,
+  torch::Tensor& context_lens,
+  int block_size,
+  int max_context_len,
+  const c10::optional<torch::Tensor>& alibi_slopes,
+  const std::string& kv_cache_dtype);
+
+void paged_attention_v2(
+  torch::Tensor& out,
+  torch::Tensor& exp_sums,
+  torch::Tensor& max_logits,
+  torch::Tensor& tmp_out,
+  torch::Tensor& query,
+  torch::Tensor& key_cache,
+  torch::Tensor& value_cache,
+  int num_kv_heads,
+  float scale,
+  torch::Tensor& block_tables,
+  torch::Tensor& context_lens,
+  int block_size,
+  int max_context_len,
+  const c10::optional<torch::Tensor>& alibi_slopes,
+  const std::string& kv_cache_dtype);
+
+void rms_norm(
+  torch::Tensor& out,
+  torch::Tensor& input,
+  torch::Tensor& weight,
+  float epsilon);
+
+void fused_add_rms_norm(
+  torch::Tensor& input,
+  torch::Tensor& residual,
+  torch::Tensor& weight,
+  float epsilon);
+
+void rotary_embedding(
+  torch::Tensor& positions,
+  torch::Tensor& query,
+  torch::Tensor& key,
+  int head_size,
+  torch::Tensor& cos_sin_cache,
+  bool is_neox);
+
+void silu_and_mul(
+  torch::Tensor& out,
+  torch::Tensor& input);
+
+void gelu_new(
+  torch::Tensor& out,
+  torch::Tensor& input);
+
+void gelu_fast(
+  torch::Tensor& out,
+  torch::Tensor& input);
+
+#ifndef USE_ROCM
+torch::Tensor awq_gemm(
+  torch::Tensor _in_feats,
+  torch::Tensor _kernel,
+  torch::Tensor _scaling_factors,
+  torch::Tensor _zeros,
+  int split_k_iters);
+
+torch::Tensor awq_dequantize(
+    torch::Tensor _kernel,
+    torch::Tensor _scaling_factors,
+    torch::Tensor _zeros,
+    int split_k_iters,
+    int thx,
+    int thy);
+#endif
+
+void squeezellm_gemm(
+  torch::Tensor vec,
+  torch::Tensor mat,
+  torch::Tensor mul,
+  torch::Tensor lookup_table);
+
+torch::Tensor gptq_gemm(
+  torch::Tensor a,
+  torch::Tensor b_q_weight,
+  torch::Tensor b_gptq_qzeros,
+  torch::Tensor b_gptq_scales,
+  torch::Tensor b_g_idx,
+  bool use_exllama);
+
+void gptq_shuffle(
+  torch::Tensor q_weight,
+  torch::Tensor q_perm);
+
+void moe_align_block_size(
+  torch::Tensor topk_ids,
+  int num_experts,
+  int block_size,
+  torch::Tensor sorted_token_ids,
+  torch::Tensor experts_ids,
+  torch::Tensor num_tokens_post_pad);
+
+#ifndef USE_ROCM
+using fptr_t = uint64_t;
+fptr_t init_custom_ar(torch::Tensor &meta, torch::Tensor &rank_data,
+                    const std::vector<std::string> &handles,
+                    const std::vector<int64_t> &offsets, int rank,
+                    bool full_nvlink);
+bool should_custom_ar(torch::Tensor &inp, int max_size, int world_size,
+                      bool full_nvlink);
+void all_reduce_reg(fptr_t _fa, torch::Tensor &inp, torch::Tensor &out);
+void all_reduce_unreg(fptr_t _fa, torch::Tensor &inp, torch::Tensor &reg_buffer,
+                      torch::Tensor &out);
+void dispose(fptr_t _fa);
+int meta_size();
+void register_buffer(fptr_t _fa, torch::Tensor &t,
+                     const std::vector<std::string> &handles,
+                     const std::vector<int64_t> &offsets);
+std::pair<std::vector<uint8_t>, std::vector<int64_t>> get_graph_buffer_ipc_meta(fptr_t _fa);
+void register_graph_buffers(fptr_t _fa, const std::vector<std::string> &handles,
+                            const std::vector<std::vector<int64_t>> &offsets);
+#endif

csrc/pos_encoding.cpp

@@ -1,14 +0,0 @@
-#include <torch/extension.h>
-
-void rotary_embedding_neox(
-  torch::Tensor& positions,
-  torch::Tensor& query,
-  torch::Tensor& key,
-  torch::Tensor& cos_sin_cache);
-
-PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
-  m.def(
-    "rotary_embedding_neox",
-    &rotary_embedding_neox,
-    "Apply GPT-NeoX style rotary embedding to query and key");
-}

csrc/pos_encoding_kernels.cu

@@ -1,78 +1,130 @@
 #include <torch/extension.h>
 #include <ATen/cuda/CUDAContext.h>
+#include <c10/cuda/CUDAGuard.h>
 
-namespace cacheflow {
+#include "cuda_compat.h"
+#include "dispatch_utils.h"
 
-template<typename scalar_t>
-__global__ void rotary_embedding_neox_kernel(
-  const int64_t* __restrict__ positions,        // [num_tokens]
-  scalar_t* __restrict__ query,                 // [num_tokens, num_heads, head_size]
-  scalar_t* __restrict__ key,                   // [num_tokens, num_heads, head_size]
-  const scalar_t* __restrict__ cos_sin_cache,   // [max_position, 2, head_size // 2]
-  const int stride,
+namespace vllm {
+
+template<typename scalar_t, bool IS_NEOX>
+inline __device__ void apply_rotary_embedding(
+  scalar_t* __restrict__ arr,
+  const scalar_t* __restrict__ cos_ptr,
+  const scalar_t* __restrict__ sin_ptr,
+  int rot_offset,
+  int embed_dim)
+{
+  int x_index, y_index;
+  scalar_t cos, sin;
+  if (IS_NEOX) {
+    // GPT-NeoX style rotary embedding.
+    x_index = rot_offset;
+    y_index = embed_dim + rot_offset;
+    cos = VLLM_LDG(cos_ptr + x_index);
+    sin = VLLM_LDG(sin_ptr + x_index);
+  } else {
+    // GPT-J style rotary embedding.
+    x_index = 2 * rot_offset;
+    y_index = 2 * rot_offset + 1;
+    cos = VLLM_LDG(cos_ptr + x_index / 2);
+    sin = VLLM_LDG(sin_ptr + x_index / 2);
+  }
+
+  const scalar_t x = arr[x_index];
+  const scalar_t y = arr[y_index];
+  arr[x_index] = x * cos - y * sin;
+  arr[y_index] = y * cos + x * sin;
+}
+
+template<typename scalar_t, bool IS_NEOX>
+__global__ void rotary_embedding_kernel(
+  const int64_t* __restrict__ positions,        // [batch_size, seq_len] or [num_tokens]
+  scalar_t* __restrict__ query,                 // [batch_size, seq_len, num_heads, head_size] or [num_tokens, num_heads, head_size]
+  scalar_t* __restrict__ key,                   // [batch_size, seq_len, num_kv_heads, head_size] or [num_tokens, num_kv_heads, head_size]
+  const scalar_t* __restrict__ cos_sin_cache,   // [max_position, 2, rot_dim // 2]
+  const int rot_dim,
+  const int64_t query_stride,
+  const int64_t key_stride,
   const int num_heads,
+  const int num_kv_heads,
   const int head_size) {
   // Each thread block is responsible for one token.
   const int token_idx = blockIdx.x;
   int64_t pos = positions[token_idx];
-  const scalar_t* cache_ptr = cos_sin_cache + pos * head_size;
+  const scalar_t* cache_ptr = cos_sin_cache + pos * rot_dim;
 
-  const int embed_dim = head_size / 2;
-  const int n = num_heads * embed_dim;
-  for (int i = threadIdx.x; i < n; i += blockDim.x) {
+  const int embed_dim = rot_dim / 2;
+  const scalar_t* cos_ptr = cache_ptr;
+  const scalar_t* sin_ptr = cache_ptr + embed_dim;
+
+  const int nq = num_heads * embed_dim;
+  for (int i = threadIdx.x; i < nq; i += blockDim.x) {
     const int head_idx = i / embed_dim;
-    const int token_head = token_idx * stride + head_idx * head_size;
-
+    const int64_t token_head = token_idx * query_stride + head_idx * head_size;
     const int rot_offset = i % embed_dim;
-    const int x_index = rot_offset;
-    const int y_index = embed_dim + rot_offset;
+    apply_rotary_embedding<scalar_t, IS_NEOX>(query + token_head, cos_ptr,
+                                              sin_ptr, rot_offset, embed_dim);
+  }
 
-    const int out_x = token_idx * stride + head_idx * head_size + x_index;
-    const int out_y = token_idx * stride + head_idx * head_size + y_index;
-
-    const scalar_t cos = __ldg(cache_ptr + x_index);
-    const scalar_t sin = __ldg(cache_ptr + y_index);
-
-    const scalar_t q_x = query[token_head + x_index];
-    const scalar_t q_y = query[token_head + y_index];
-    query[out_x] = q_x * cos - q_y * sin;
-    query[out_y] = q_y * cos + q_x * sin;
-
-    const scalar_t k_x = key[token_head + x_index];
-    const scalar_t k_y = key[token_head + y_index];
-    key[out_x] = k_x * cos - k_y * sin;
-    key[out_y] = k_y * cos + k_x * sin;
+  const int nk = num_kv_heads * embed_dim;
+  for (int i = threadIdx.x; i < nk; i += blockDim.x) {
+    const int head_idx = i / embed_dim;
+    const int64_t token_head = token_idx * key_stride + head_idx * head_size;
+    const int rot_offset = i % embed_dim;
+    apply_rotary_embedding<scalar_t, IS_NEOX>(key + token_head, cos_ptr,
+                                              sin_ptr, rot_offset, embed_dim);
   }
 }
 
-} // namespace cacheflow
+} // namespace vllm
 
-void rotary_embedding_neox(
-  torch::Tensor& positions,         // [num_tokens]
-  torch::Tensor& query,             // [num_tokens, num_heads * head_size]
-  torch::Tensor& key,               // [num_tokens, num_heads * head_size]
-  torch::Tensor& cos_sin_cache)     // [max_position, head_size]
-{
-  int num_tokens = query.size(0);
-  int head_size = cos_sin_cache.size(1);
-  int num_heads = query.size(1) / head_size;
-  int stride = query.stride(0);
-  TORCH_CHECK(stride == key.stride(0));
+void rotary_embedding(
+  torch::Tensor& positions,         // [batch_size, seq_len] or [num_tokens]
+  torch::Tensor& query,             // [batch_size, seq_len, num_heads * head_size] or [num_tokens, num_heads * head_size]
+  torch::Tensor& key,               // [batch_size, seq_len, num_kv_heads * head_size] or [num_tokens, num_kv_heads * head_size]
+  int head_size,
+  torch::Tensor& cos_sin_cache,     // [max_position, rot_dim]
+  bool is_neox) {
+  int64_t num_tokens = query.numel() / query.size(-1);
+  int rot_dim = cos_sin_cache.size(1);
+  int num_heads = query.size(-1) / head_size;
+  int num_kv_heads = key.size(-1) / head_size;
+  int64_t query_stride = query.stride(-2);
+  int64_t key_stride = key.stride(-2);
 
   dim3 grid(num_tokens);
-  dim3 block(std::min(num_heads * head_size / 2, 512));
+  dim3 block(std::min(num_heads * rot_dim / 2, 512));
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(query));
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
-  AT_DISPATCH_FLOATING_TYPES_AND_HALF(
+  VLLM_DISPATCH_FLOATING_TYPES(
     query.scalar_type(),
-    "rotary_embedding_neox",
+    "rotary_embedding",
     [&] {
-      cacheflow::rotary_embedding_neox_kernel<scalar_t><<<grid, block, 0, stream>>>(
-        positions.data_ptr<int64_t>(),
-        query.data_ptr<scalar_t>(),
-        key.data_ptr<scalar_t>(),
-        cos_sin_cache.data_ptr<scalar_t>(),
-        stride,
-        num_heads,
-        head_size);
+      if (is_neox) {
+        vllm::rotary_embedding_kernel<scalar_t, true><<<grid, block, 0, stream>>>(
+          positions.data_ptr<int64_t>(),
+          query.data_ptr<scalar_t>(),
+          key.data_ptr<scalar_t>(),
+          cos_sin_cache.data_ptr<scalar_t>(),
+          rot_dim,
+          query_stride,
+          key_stride,
+          num_heads,
+          num_kv_heads,
+          head_size);
+      } else {
+        vllm::rotary_embedding_kernel<scalar_t, false><<<grid, block, 0, stream>>>(
+          positions.data_ptr<int64_t>(),
+          query.data_ptr<scalar_t>(),
+          key.data_ptr<scalar_t>(),
+          cos_sin_cache.data_ptr<scalar_t>(),
+          rot_dim,
+          query_stride,
+          key_stride,
+          num_heads,
+          num_kv_heads,
+          head_size);
+      }
     });
 }

csrc/punica/LICENSE

@@ -0,0 +1,217 @@
+Contains code from https://github.com/punica-ai/punica
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "{}"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright {yyyy} {name of copyright owner}
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+
+------------------------------------------------------------------------------------
+
+This product bundles various third-party components under other open source licenses.
+This section summarizes those components and their licenses. See licenses/
+for text of these licenses.
+
+
+Apache-2.0
+* third_party/nvbench (with LLVM exception)
+* third_party/flashinfer
+
+BSD-3-Clause:
+* third_party/cutlass

csrc/punica/bgmv/bgmv_bf16_bf16_bf16.cu

@@ -0,0 +1,4 @@
+#include "bgmv_config.h"
+#include "bgmv_impl.cuh"
+
+FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_bfloat16, nv_bfloat16, nv_bfloat16)

csrc/punica/bgmv/bgmv_bf16_bf16_fp16.cu

@@ -0,0 +1,4 @@
+#include "bgmv_config.h"
+#include "bgmv_impl.cuh"
+
+FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_bfloat16, nv_bfloat16, nv_half)

csrc/punica/bgmv/bgmv_bf16_fp16_bf16.cu

@@ -0,0 +1,4 @@
+#include "bgmv_config.h"
+#include "bgmv_impl.cuh"
+
+FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_bfloat16, nv_half, nv_bfloat16)

csrc/punica/bgmv/bgmv_bf16_fp16_fp16.cu

@@ -0,0 +1,4 @@
+#include "bgmv_config.h"
+#include "bgmv_impl.cuh"
+
+FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_bfloat16, nv_half, nv_half)

csrc/punica/bgmv/bgmv_bf16_fp32_bf16.cu

@@ -0,0 +1,4 @@
+#include "bgmv_config.h"
+#include "bgmv_impl.cuh"
+
+FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_bfloat16, float, nv_bfloat16)