Bump up to v0.2.3 (#1903)

Co-authored-by: Antoni Baum <antoni.baum@protonmail.com>

.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: ['2.1.0']
+          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

@@ -1,4 +1,4 @@
-name: pylint
+name: ruff
 
 on:
   # Trigger the workflow on push or pull request,
@@ -11,7 +11,7 @@ on:
       - main
 
 jobs:
-  pylint:
+  ruff:
     runs-on: ubuntu-latest
     strategy:
       matrix:
@@ -25,7 +25,7 @@ jobs:
     - name: Install dependencies
       run: |
         python -m pip install --upgrade pip
-        pip install pylint==2.8.2
-    - name: Analysing the code with pylint
+        pip install ruff==0.1.5
+    - name: Analysing the code with ruff
       run: |
-        pylint vllm
+        ruff vllm tests

.github/workflows/scripts/build.sh

@@ -0,0 +1,18 @@
+#!/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
+$python_executable -m pip install -r requirements.txt
+
+# Limit the number of parallel jobs to avoid OOM
+export MAX_JOBS=1
+
+# Build
+$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

@@ -28,4 +28,4 @@ jobs:
         pip install toml==0.10.2
     - name: Running yapf
       run: |
-        yapf --diff --recursive vllm --exclude 'vllm/model_executor/parallel_utils/**'
+        yapf --diff --recursive vllm tests

.gitignore

@@ -173,3 +173,7 @@ cython_debug/
 
 # Sphinx documentation
 _build/
+
+# vim swap files
+*.swo
+*.swp

.pylintrc

@@ -1,434 +0,0 @@
-# This Pylint rcfile contains a best-effort configuration to uphold the
-# best-practices and style described in the Google Python style guide:
-#   https://google.github.io/styleguide/pyguide.html
-#
-# Its canonical open-source location is:
-#   https://google.github.io/styleguide/pylintrc
-
-[MASTER]
-
-# Files or directories to be skipped. They should be base names, not paths.
-ignore=docs,parallel_utils
-
-# Files or directories matching the regex patterns are skipped. The regex
-# matches against base names, not paths.
-ignore-patterns=
-
-# Pickle collected data for later comparisons.
-persistent=no
-
-# List of plugins (as comma separated values of python modules names) to load,
-# usually to register additional checkers.
-load-plugins=
-
-# Use multiple processes to speed up Pylint.
-jobs=4
-
-# Allow loading of arbitrary C extensions. Extensions are imported into the
-# active Python interpreter and may run arbitrary code.
-unsafe-load-any-extension=no
-
-
-[MESSAGES CONTROL]
-
-# Only show warnings with the listed confidence levels. Leave empty to show
-# all. Valid levels: HIGH, INFERENCE, INFERENCE_FAILURE, UNDEFINED
-confidence=
-
-# Enable the message, report, category or checker with the given id(s). You can
-# either give multiple identifier separated by comma (,) or put this option
-# multiple time (only on the command line, not in the configuration file where
-# it should appear only once). See also the "--disable" option for examples.
-#enable=
-
-# Disable the message, report, category or checker with the given id(s). You
-# can either give multiple identifiers separated by comma (,) or put this
-# option multiple times (only on the command line, not in the configuration
-# file where it should appear only once).You can also use "--disable=all" to
-# disable everything first and then reenable specific checks. For example, if
-# you want to run only the similarities checker, you can use "--disable=all
-# --enable=similarities". If you want to run only the classes checker, but have
-# no Warning level messages displayed, use"--disable=all --enable=classes
-# --disable=W"
-disable=abstract-method,
-        apply-builtin,
-        arguments-differ,
-        attribute-defined-outside-init,
-        backtick,
-        bad-option-value,
-        basestring-builtin,
-        buffer-builtin,
-        c-extension-no-member,
-        consider-using-enumerate,
-        cmp-builtin,
-        cmp-method,
-        coerce-builtin,
-        coerce-method,
-        delslice-method,
-        div-method,
-        duplicate-code,
-        eq-without-hash,
-        execfile-builtin,
-        file-builtin,
-        filter-builtin-not-iterating,
-        fixme,
-        getslice-method,
-        global-statement,
-        hex-method,
-        idiv-method,
-        implicit-str-concat-in-sequence,
-        import-error,
-        import-self,
-        import-star-module-level,
-        inconsistent-return-statements,
-        input-builtin,
-        intern-builtin,
-        invalid-str-codec,
-        locally-disabled,
-        logging-fstring-interpolation,  # added by vLLM
-        logging-not-lazy,  # added by vLLM
-        long-builtin,
-        long-suffix,
-        map-builtin-not-iterating,
-        misplaced-comparison-constant,
-        missing-class-docstring,  # TODO (vLLM): enable
-        missing-function-docstring,
-        missing-module-docstring,  # TODO (vLLM): enable
-        metaclass-assignment,
-        next-method-called,
-        next-method-defined,
-        no-absolute-import,
-        no-else-break,
-        no-else-continue,
-        no-else-raise,
-        no-else-return,
-        no-init,  # added
-        no-member,
-        no-name-in-module,
-        no-self-use,
-        nonzero-method,
-        oct-method,
-        old-division,
-        old-ne-operator,
-        old-octal-literal,
-        old-raise-syntax,
-        parameter-unpacking,
-        print-statement,
-        raising-string,
-        range-builtin-not-iterating,
-        raw_input-builtin,
-        rdiv-method,
-        reduce-builtin,
-        relative-import,
-        reload-builtin,
-        round-builtin,
-        setslice-method,
-        signature-differs,
-        standarderror-builtin,
-        suppressed-message,
-        sys-max-int,
-        too-few-public-methods,
-        too-many-ancestors,
-        too-many-arguments,
-        too-many-boolean-expressions,
-        too-many-branches,
-        too-many-instance-attributes,
-        too-many-locals,
-        too-many-nested-blocks,
-        too-many-public-methods,
-        too-many-return-statements,
-        too-many-statements,
-        trailing-newlines,
-        unichr-builtin,
-        unicode-builtin,
-        unnecessary-pass,
-        unpacking-in-except,
-        unspecified-encoding,
-        useless-else-on-loop,
-        useless-object-inheritance,
-        useless-suppression,
-        using-cmp-argument,
-        wrong-import-order,
-        xrange-builtin,
-        zip-builtin-not-iterating,
-
-
-[REPORTS]
-
-# Set the output format. Available formats are text, parseable, colorized, msvs
-# (visual studio) and html. You can also give a reporter class, eg
-# mypackage.mymodule.MyReporterClass.
-output-format=text
-
-# Tells whether to display a full report or only the messages
-reports=no
-
-# Python expression which should return a note less than 10 (10 is the highest
-# note). You have access to the variables errors warning, statement which
-# respectively contain the number of errors / warnings messages and the total
-# number of statements analyzed. This is used by the global evaluation report
-# (RP0004).
-evaluation=10.0 - ((float(5 * error + warning + refactor + convention) / statement) * 10)
-
-# Template used to display messages. This is a python new-style format string
-# used to format the message information. See doc for all details
-#msg-template=
-
-
-[BASIC]
-
-# Good variable names which should always be accepted, separated by a comma
-good-names=main,_
-
-# Bad variable names which should always be refused, separated by a comma
-bad-names=
-
-# Colon-delimited sets of names that determine each other's naming style when
-# the name regexes allow several styles.
-name-group=
-
-# Include a hint for the correct naming format with invalid-name
-include-naming-hint=no
-
-# List of decorators that produce properties, such as abc.abstractproperty. Add
-# to this list to register other decorators that produce valid properties.
-property-classes=abc.abstractproperty,cached_property.cached_property,cached_property.threaded_cached_property,cached_property.cached_property_with_ttl,cached_property.threaded_cached_property_with_ttl
-
-# Regular expression matching correct function names
-function-rgx=^(?:(?P<exempt>setUp|tearDown|setUpModule|tearDownModule)|(?P<camel_case>_?[A-Z][a-zA-Z0-9]*)|(?P<snake_case>_?[a-z][a-z0-9_]*))$
-
-# Regular expression matching correct variable names
-variable-rgx=^[a-z][a-z0-9_]*$
-
-# Regular expression matching correct constant names
-const-rgx=^(_?[A-Z][A-Z0-9_]*|__[a-z0-9_]+__|_?[a-z][a-z0-9_]*)$
-
-# Regular expression matching correct attribute names
-attr-rgx=^_{0,2}[a-z][a-z0-9_]*$
-
-# Regular expression matching correct argument names
-argument-rgx=^[a-z][a-z0-9_]*$
-
-# Regular expression matching correct class attribute names
-class-attribute-rgx=^(_?[A-Z][A-Z0-9_]*|__[a-z0-9_]+__|_?[a-z][a-z0-9_]*)$
-
-# Regular expression matching correct inline iteration names
-inlinevar-rgx=^[a-z][a-z0-9_]*$
-
-# Regular expression matching correct class names
-class-rgx=^_?[A-Z][a-zA-Z0-9]*$
-
-# Regular expression matching correct module names
-module-rgx=^(_?[a-z][a-z0-9_]*|__init__)$
-
-# Regular expression matching correct method names
-method-rgx=(?x)^(?:(?P<exempt>_[a-z0-9_]+__|runTest|setUp|tearDown|setUpTestCase|tearDownTestCase|setupSelf|tearDownClass|setUpClass|(test|assert)_*[A-Z0-9][a-zA-Z0-9_]*|next)|(?P<camel_case>_{0,2}[A-Z][a-zA-Z0-9_]*)|(?P<snake_case>_{0,2}[a-z][a-z0-9_]*))$
-
-# Regular expression which should only match function or class names that do
-# not require a docstring.
-no-docstring-rgx=(__.*__|main|test.*|.*test|.*Test)$
-
-# Minimum line length for functions/classes that require docstrings, shorter
-# ones are exempt.
-docstring-min-length=10
-
-
-[TYPECHECK]
-
-# List of decorators that produce context managers, such as
-# contextlib.contextmanager. Add to this list to register other decorators that
-# produce valid context managers.
-contextmanager-decorators=contextlib.contextmanager,contextlib2.contextmanager
-
-# Tells whether missing members accessed in mixin class should be ignored. A
-# mixin class is detected if its name ends with "mixin" (case insensitive).
-ignore-mixin-members=yes
-
-# List of module names for which member attributes should not be checked
-# (useful for modules/projects where namespaces are manipulated during runtime
-# and thus existing member attributes cannot be deduced by static analysis. It
-# supports qualified module names, as well as Unix pattern matching.
-ignored-modules=
-
-# List of class names for which member attributes should not be checked (useful
-# for classes with dynamically set attributes). This supports the use of
-# qualified names.
-ignored-classes=optparse.Values,thread._local,_thread._local
-
-# List of members which are set dynamically and missed by pylint inference
-# system, and so shouldn't trigger E1101 when accessed. Python regular
-# expressions are accepted.
-generated-members=
-
-
-[FORMAT]
-
-# Maximum number of characters on a single line.
-max-line-length=80
-
-# TODO(https://github.com/PyCQA/pylint/issues/3352): Direct pylint to exempt
-# lines made too long by directives to pytype.
-
-# Regexp for a line that is allowed to be longer than the limit.
-ignore-long-lines=(?x)(
-  ^\s*(\#\ )?<?https?://\S+>?$|
-  ^\s*(from\s+\S+\s+)?import\s+.+$)
-
-# Allow the body of an if to be on the same line as the test if there is no
-# else.
-single-line-if-stmt=yes
-
-# Maximum number of lines in a module
-max-module-lines=99999
-
-# String used as indentation unit.  The internal Google style guide mandates 2
-# spaces.  Google's externaly-published style guide says 4, consistent with
-# PEP 8.  Here, we use 2 spaces, for conformity with many open-sourced Google
-# projects (like TensorFlow).
-indent-string='    '
-
-# Number of spaces of indent required inside a hanging  or continued line.
-indent-after-paren=4
-
-# Expected format of line ending, e.g. empty (any line ending), LF or CRLF.
-expected-line-ending-format=
-
-
-[MISCELLANEOUS]
-
-# List of note tags to take in consideration, separated by a comma.
-notes=TODO
-
-
-[STRING]
-
-# This flag controls whether inconsistent-quotes generates a warning when the
-# character used as a quote delimiter is used inconsistently within a module.
-check-quote-consistency=yes
-
-
-[VARIABLES]
-
-# Tells whether we should check for unused import in __init__ files.
-init-import=no
-
-# A regular expression matching the name of dummy variables (i.e. expectedly
-# not used).
-dummy-variables-rgx=^\*{0,2}(_$|unused_|dummy_)
-
-# List of additional names supposed to be defined in builtins. Remember that
-# you should avoid to define new builtins when possible.
-additional-builtins=
-
-# List of strings which can identify a callback function by name. A callback
-# name must start or end with one of those strings.
-callbacks=cb_,_cb
-
-# List of qualified module names which can have objects that can redefine
-# builtins.
-redefining-builtins-modules=six,six.moves,past.builtins,future.builtins,functools
-
-
-[LOGGING]
-
-# Logging modules to check that the string format arguments are in logging
-# function parameter format
-logging-modules=logging,absl.logging,tensorflow.io.logging
-
-
-[SIMILARITIES]
-
-# Minimum lines number of a similarity.
-min-similarity-lines=4
-
-# Ignore comments when computing similarities.
-ignore-comments=yes
-
-# Ignore docstrings when computing similarities.
-ignore-docstrings=yes
-
-# Ignore imports when computing similarities.
-ignore-imports=no
-
-
-[SPELLING]
-
-# Spelling dictionary name. Available dictionaries: none. To make it working
-# install python-enchant package.
-spelling-dict=
-
-# List of comma separated words that should not be checked.
-spelling-ignore-words=
-
-# A path to a file that contains private dictionary; one word per line.
-spelling-private-dict-file=
-
-# Tells whether to store unknown words to indicated private dictionary in
-# --spelling-private-dict-file option instead of raising a message.
-spelling-store-unknown-words=no
-
-
-[IMPORTS]
-
-# Deprecated modules which should not be used, separated by a comma
-deprecated-modules=regsub,
-                   TERMIOS,
-                   Bastion,
-                   rexec,
-                   sets
-
-# Create a graph of every (i.e. internal and external) dependencies in the
-# given file (report RP0402 must not be disabled)
-import-graph=
-
-# Create a graph of external dependencies in the given file (report RP0402 must
-# not be disabled)
-ext-import-graph=
-
-# Create a graph of internal dependencies in the given file (report RP0402 must
-# not be disabled)
-int-import-graph=
-
-# Force import order to recognize a module as part of the standard
-# compatibility libraries.
-known-standard-library=
-
-# Force import order to recognize a module as part of a third party library.
-known-third-party=enchant, absl
-
-# Analyse import fallback blocks. This can be used to support both Python 2 and
-# 3 compatible code, which means that the block might have code that exists
-# only in one or another interpreter, leading to false positives when analysed.
-analyse-fallback-blocks=no
-
-
-[CLASSES]
-
-# List of method names used to declare (i.e. assign) instance attributes.
-defining-attr-methods=__init__,
-                      __new__,
-                      setUp
-
-# List of member names, which should be excluded from the protected access
-# warning.
-exclude-protected=_asdict,
-                  _fields,
-                  _replace,
-                  _source,
-                  _make
-
-# List of valid names for the first argument in a class method.
-valid-classmethod-first-arg=cls,
-                            class_
-
-# List of valid names for the first argument in a metaclass class method.
-valid-metaclass-classmethod-first-arg=mcs
-
-
-[EXCEPTIONS]
-
-# Exceptions that will emit a warning when being caught. Defaults to
-# "Exception"
-overgeneral-exceptions=StandardError,
-                       Exception,
-                       BaseException

Dockerfile

@@ -0,0 +1,77 @@
+FROM nvidia/cuda:12.1.0-devel-ubuntu22.04 AS dev
+
+RUN apt-get update -y \
+    && apt-get install -y python3-pip
+
+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
+
+# image to build pytorch extensions
+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
+
+# max jobs used by Ninja to build extensions
+ENV MAX_JOBS=$max_jobs
+RUN python3 setup.py build_ext --inplace
+
+# image to run unit testing suite
+FROM dev AS test
+
+# copy pytorch extensions separately to avoid having to rebuild
+# when python code changes
+COPY --from=build /workspace/vllm/*.so /workspace/vllm/
+COPY tests tests
+COPY vllm vllm
+
+ENTRYPOINT ["python3", "-m", "pytest", "tests"]
+
+# use CUDA base as CUDA runtime dependencies are already installed via pip
+FROM nvidia/cuda:12.1.0-base-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
+
+FROM vllm-base AS vllm
+COPY --from=build /workspace/vllm/*.so /workspace/vllm/
+COPY vllm vllm
+
+EXPOSE 8000
+ENTRYPOINT ["python3", "-m", "vllm.entrypoints.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 fschat
+
+COPY --from=build /workspace/vllm/*.so /workspace/vllm/
+COPY vllm vllm
+
+ENTRYPOINT ["python3", "-m", "vllm.entrypoints.openai.api_server"]
+

README.md

@@ -10,13 +10,18 @@ Easy, fast, and cheap LLM serving for everyone
 </h3>
 
 <p align="center">
-| <a href="https://vllm.readthedocs.io/en/latest/"><b>Documentation</b></a> | <a href="https://vllm.ai"><b>Blog</b></a> | <a href="https://github.com/vllm-project/vllm/discussions"><b>Discussions</b></a> |
+| <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* 🔥
+- [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).
 
@@ -33,21 +38,31 @@ vLLM is fast with:
 
 vLLM is flexible and easy to use with:
 
-- Seamless integration with popular HuggingFace models
+- 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
 
-vLLM seamlessly supports many Huggingface models, including the following architectures:
+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.)
+- Falcon (`tiiuae/falcon-7b`, `tiiuae/falcon-40b`, `tiiuae/falcon-rw-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.)
-- LLaMA (`lmsys/vicuna-13b-v1.3`, `young-geng/koala`, `openlm-research/open_llama_13b`, etc.)
+- InternLM (`internlm/internlm-7b`, `internlm/internlm-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.)
 - MPT (`mosaicml/mpt-7b`, `mosaicml/mpt-30b`, etc.)
 - OPT (`facebook/opt-66b`, `facebook/opt-iml-max-30b`, etc.)
+- Phi-1.5 (`microsoft/phi-1_5`, etc.)
+- Qwen (`Qwen/Qwen-7B`, `Qwen/Qwen-7B-Chat`, 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):
 
@@ -62,37 +77,19 @@ Visit our [documentation](https://vllm.readthedocs.io/en/latest/) to get started
 - [Quickstart](https://vllm.readthedocs.io/en/latest/getting_started/quickstart.html)
 - [Supported Models](https://vllm.readthedocs.io/en/latest/models/supported_models.html)
 
-## Performance
-
-vLLM outperforms HuggingFace Transformers (HF) by up to 24x and Text Generation Inference (TGI) by up to 3.5x, in terms of throughput.
-For details, check out our [blog post](https://vllm.ai).
-
-<p align="center">
-  <picture>
-  <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/vllm-project/vllm/main/docs/source/assets/figures/perf_a10g_n1_dark.png">
-  <img src="https://raw.githubusercontent.com/vllm-project/vllm/main/docs/source/assets/figures/perf_a10g_n1_light.png" width="45%">
-  </picture>
-  <picture>
-  <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/vllm-project/vllm/main/docs/source/assets/figures/perf_a100_n1_dark.png">
-  <img src="https://raw.githubusercontent.com/vllm-project/vllm/main/docs/source/assets/figures/perf_a100_n1_light.png" width="45%">
-  </picture>
-  <br>
-  <em> Serving throughput when each request asks for 1 output completion. </em>
-</p>
-
-<p align="center">
-  <picture>
-  <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/vllm-project/vllm/main/docs/source/assets/figures/perf_a10g_n3_dark.png">
-  <img src="https://raw.githubusercontent.com/vllm-project/vllm/main/docs/source/assets/figures/perf_a10g_n3_light.png" width="45%">
-  </picture>
-  <picture>
-  <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/vllm-project/vllm/main/docs/source/assets/figures/perf_a100_n3_dark.png">
-  <img src="https://raw.githubusercontent.com/vllm-project/vllm/main/docs/source/assets/figures/perf_a100_n3_light.png" width="45%">
-  </picture>  <br>
-  <em> Serving throughput when each request asks for 3 output completions. </em>
-</p>
-
 ## 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}
+}
+```

benchmarks/benchmark_latency.py

@@ -12,15 +12,15 @@ from vllm import LLM, SamplingParams
 def main(args: argparse.Namespace):
     print(args)
 
-    # Process all the requests in a single batch if possible.
     # 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,
-        max_num_seqs=args.batch_size,
-        max_num_batched_tokens=args.batch_size * args.input_len,
+        trust_remote_code=args.trust_remote_code,
+        dtype=args.dtype,
     )
 
     sampling_params = SamplingParams(
@@ -36,22 +36,31 @@ def main(args: argparse.Namespace):
 
     def run_to_completion(profile: bool = False):
         if profile:
-            torch.cuda.cudart().cudaProfilerStart()
-        start_time = time.time()
-
-        llm.generate(prompt_token_ids=dummy_prompt_token_ids,
-                     sampling_params=sampling_params,
-                     use_tqdm=False)
-
-        end_time = time.time()
-        latency = end_time - start_time
-        if profile:
-            torch.cuda.cudart().cudaProfilerStop()
-        return latency
+            with torch.profiler.profile(activities=[
+                    torch.profiler.ProfilerActivity.CPU,
+                    torch.profiler.ProfilerActivity.CUDA,
+            ]) 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=False)
 
+    if args.profile:
+        print("Profiling...")
+        run_to_completion(profile=True)
+        return
+
     # Benchmark.
     latencies = []
     for _ in tqdm(range(args.num_iters), desc="Profiling iterations"):
@@ -62,17 +71,41 @@ def main(args: argparse.Namespace):
 if __name__ == '__main__':
     parser = argparse.ArgumentParser(
         description='Benchmark the latency of processing a single batch of '
-                    'requests till completion.')
+        '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', '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,
+    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,
+    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(
+        '--profile',
+        action='store_true',
+        help='profile the generation process of a single batch')
     args = parser.parse_args()
     main(args)

benchmarks/benchmark_serving.py

@@ -105,7 +105,7 @@ async def send_request(
     best_of: int,
     use_beam_search: bool,
 ) -> None:
-    request_start_time = time.time()
+    request_start_time = time.perf_counter()
 
     headers = {"User-Agent": "Benchmark Client"}
     if backend == "vllm":
@@ -148,7 +148,7 @@ async def send_request(
             if "error" not in output:
                 break
 
-    request_end_time = time.time()
+    request_end_time = time.perf_counter()
     request_latency = request_end_time - request_start_time
     REQUEST_LATENCY.append((prompt_len, output_len, request_latency))
 
@@ -177,13 +177,13 @@ def main(args: argparse.Namespace):
     np.random.seed(args.seed)
 
     api_url = f"http://{args.host}:{args.port}/generate"
-    tokenizer = get_tokenizer(args.tokenizer)
+    tokenizer = get_tokenizer(args.tokenizer, trust_remote_code=args.trust_remote_code)
     input_requests = sample_requests(args.dataset, args.num_prompts, tokenizer)
 
-    benchmark_start_time = time.time()
+    benchmark_start_time = time.perf_counter()
     asyncio.run(benchmark(args.backend, api_url, input_requests, args.best_of,
                           args.use_beam_search, args.request_rate))
-    benchmark_end_time = time.time()
+    benchmark_end_time = time.perf_counter()
     benchmark_time = benchmark_end_time - benchmark_start_time
     print(f"Total time: {benchmark_time:.2f} s")
     print(f"Throughput: {args.num_prompts / benchmark_time:.2f} requests/s")
@@ -227,5 +227,7 @@ if __name__ == "__main__":
                              "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')
     args = parser.parse_args()
     main(args)

benchmarks/benchmark_throughput.py

@@ -3,34 +3,31 @@ import argparse
 import json
 import random
 import time
-from typing import List, Tuple
+from typing import List, Optional, Tuple
 
 import torch
-from transformers import AutoModelForCausalLM, PreTrainedTokenizerBase
+from transformers import (AutoModelForCausalLM, AutoTokenizer,
+                          PreTrainedTokenizerBase)
 from tqdm import tqdm
 
-from vllm import LLM, SamplingParams
-from vllm.transformers_utils.tokenizer import get_tokenizer
-
 
 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
-    ]
+    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
-    ]
+    dataset = [(data["conversations"][0]["value"],
+                data["conversations"][1]["value"]) for data in dataset]
 
     # Tokenize the prompts and completions.
     prompts = [prompt for prompt, _ in dataset]
@@ -40,6 +37,8 @@ def sample_requests(
     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.
@@ -63,16 +62,25 @@ 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] = None,
 ) -> 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,
     )
 
     # Add the requests to the engine.
@@ -92,10 +100,10 @@ def run_vllm(
             sampling_params=sampling_params,
         )
 
-    start = time.time()
-    # FIXME(woosuk): Do use internal method.
+    start = time.perf_counter()
+    # FIXME(woosuk): Do not use internal method.
     llm._run_engine(use_tqdm=True)
-    end = time.time()
+    end = time.perf_counter()
     return end - start
 
 
@@ -106,16 +114,18 @@ def run_hf(
     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)
+    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.time()
+    start = time.perf_counter()
     batch: List[str] = []
     max_prompt_len = 0
     max_output_len = 0
@@ -128,13 +138,14 @@ def run_hf(
         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:
+            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
+        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,
@@ -152,7 +163,23 @@ def run_hf(
         batch = []
         max_prompt_len = 0
         max_output_len = 0
-    end = time.time()
+    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
 
 
@@ -161,45 +188,105 @@ def main(args: argparse.Namespace):
     random.seed(args.seed)
 
     # Sample the requests.
-    tokenizer = get_tokenizer(args.tokenizer)
-    requests = sample_requests(args.dataset, args.num_prompts, tokenizer)
+    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.tensor_parallel_size,
-            args.seed, args.n, args.use_beam_search)
+        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)
     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.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
-    )
+    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"],
+    parser.add_argument("--backend",
+                        type=str,
+                        choices=["vllm", "hf", "mii"],
                         default="vllm")
-    parser.add_argument("--dataset", type=str, required=True,
+    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', 'squeezellm', None],
+                        default=None)
     parser.add_argument("--tensor-parallel-size", "-tp", type=int, default=1)
-    parser.add_argument("--n", 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,
+    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,
+    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.')
     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:
@@ -207,7 +294,20 @@ if __name__ == "__main__":
     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.tokenizer is None:
-        args.tokenizer = args.model
-
+        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 +1,197 @@
+import argparse
+import random
+import time
+
+import torch
+
+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,
+) -> None:
+    random.seed(seed)
+    torch.random.manual_seed(seed)
+    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="cuda")
+    query.uniform_(-scale, scale)
+
+    assert num_query_heads % num_kv_heads == 0
+    num_queries_per_kv = num_query_heads // num_kv_heads
+    head_mapping = torch.repeat_interleave(
+        torch.arange(num_kv_heads, dtype=torch.int32, device="cuda"),
+        num_queries_per_kv)
+    alibi_slopes = None
+    if use_alibi:
+        alibi_slopes = torch.randn(num_query_heads,
+                                   dtype=torch.float,
+                                   device="cuda")
+
+    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="cuda")
+
+    # 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="cuda")
+
+    # Create the KV cache.
+    x = 16 // torch.tensor([], dtype=dtype).element_size()
+    key_cache_shape = (NUM_BLOCKS, num_kv_heads, head_size // x, block_size, x)
+    key_cache = torch.empty(size=key_cache_shape, dtype=dtype, device="cuda")
+    key_cache.uniform_(-scale, scale)
+    value_cache_shape = (NUM_BLOCKS, num_kv_heads, head_size, block_size)
+    value_cache = torch.empty(size=value_cache_shape,
+                              dtype=dtype,
+                              device="cuda")
+    value_cache.uniform_(-scale, scale)
+
+    # 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_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,
+                    head_mapping,
+                    scale,
+                    block_tables,
+                    context_lens,
+                    block_size,
+                    max_context_len,
+                    alibi_slopes,
+                )
+            elif version == "v2":
+                ops.paged_attention_v2(
+                    output,
+                    exp_sums,
+                    max_logits,
+                    tmp_output,
+                    query,
+                    key_cache,
+                    value_cache,
+                    head_mapping,
+                    scale,
+                    block_tables,
+                    context_lens,
+                    block_size,
+                    max_context_len,
+                    alibi_slopes,
+                )
+            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(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")
+    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")
+    dtype_to_torch_dtype = {
+        "half": torch.half,
+        "bfloat16": torch.bfloat16,
+        "float": torch.float,
+    }
+    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=dtype_to_torch_dtype[args.dtype],
+        seed=args.seed,
+        do_profile=args.profile,
+    )

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

@@ -1,6 +1,8 @@
 #include <torch/extension.h>
 #include <ATen/cuda/CUDAContext.h>
 
+#include "dispatch_utils.h"
+
 namespace vllm {
 
 template<typename T>
@@ -11,11 +13,11 @@ __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 int64_t token_idx = blockIdx.x;
+  for (int64_t 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]);
     out[token_idx * d + idx] = silu(x) * y;
@@ -25,18 +27,16 @@ __global__ void silu_and_mul_kernel(
 } // 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 cudaStream_t stream = at::cuda::getCurrentCUDAStream();
-  AT_DISPATCH_FLOATING_TYPES_AND2(
-    at::ScalarType::Half,
-    at::ScalarType::BFloat16,
+  VLLM_DISPATCH_FLOATING_TYPES(
     input.scalar_type(),
     "silu_and_mul_kernel",
     [&] {
@@ -46,3 +46,69 @@ void silu_and_mul(
         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 = __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 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

@@ -1,21 +0,0 @@
-#include <torch/extension.h>
-#include <c10/util/Optional.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,
-  const c10::optional<torch::Tensor>& alibi_slopes);
-
-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_kernels.cu

@@ -26,6 +26,7 @@
 #define WARP_SIZE 32
 #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 {
 
@@ -65,25 +66,57 @@ inline __device__ float block_sum(float* red_smem, float sum) {
   return __shfl_sync(uint32_t(-1), sum, 0);
 }
 
-// Grid: (num_heads, num_seqs).
+// TODO(woosuk): Merge the last two dimensions of the grid.
+// Grid: (num_heads, num_seqs, max_num_partitions).
 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]
+  int NUM_THREADS,
+  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 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 scalar_t* __restrict__ k_cache,   // [num_blocks, num_kv_heads, head_size/x, block_size, x]
+  const scalar_t* __restrict__ v_cache,   // [num_blocks, num_kv_heads, head_size, block_size]
+  const int* __restrict__ head_mapping,   // [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 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_TOKENS_PER_THREAD_GROUP = (BLOCK_SIZE + WARP_SIZE - 1) / WARP_SIZE;
+  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;
@@ -91,7 +124,7 @@ __global__ void single_query_cached_kv_attention_kernel(
 
   const int head_idx = blockIdx.x;
   const int num_heads = gridDim.x;
-  const int seq_idx = blockIdx.y;
+  const int kv_head_idx = head_mapping[head_idx];
   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.
@@ -116,12 +149,13 @@ __global__ void single_query_cached_kv_attention_kernel(
   // 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];
+  __shared__ Q_vec q_vecs[THREAD_GROUP_SIZE][NUM_VECS_PER_THREAD];
 #pragma unroll
-  for (int i = 0; i < NUM_VECS_PER_THREAD; i++) {
+  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[i] = *reinterpret_cast<const Q_vec*>(q_ptr + vec_idx * VEC_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[];
@@ -135,16 +169,16 @@ __global__ void single_query_cached_kv_attention_kernel(
   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];
+  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.
@@ -158,8 +192,8 @@ __global__ void single_query_cached_kv_attention_kernel(
 
 #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
+        const scalar_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;
@@ -169,15 +203,15 @@ __global__ void single_query_cached_kv_attention_kernel(
 
       // 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, k_vecs);
+      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) : 0;
+      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] = mask ? 0.f : qk;
+        logits[token_idx - start_token_idx] = mask ? 0.f : qk;
         // Update the max value.
         qk_max = mask ? qk_max : fmaxf(qk_max, qk);
       }
@@ -208,7 +242,7 @@ __global__ void single_query_cached_kv_attention_kernel(
 
   // Get the sum of the exp values.
   float exp_sum = 0.f;
-  for (int i = thread_idx; i < context_len; i += NUM_THREADS) {
+  for (int i = thread_idx; i < num_tokens; i += NUM_THREADS) {
     float val = __expf(logits[i] - qk_max);
     logits[i] = val;
     exp_sum += val;
@@ -217,11 +251,23 @@ __global__ void single_query_cached_kv_attention_kernel(
 
   // Compute softmax.
   const float inv_sum = __fdividef(1.f, exp_sum + 1e-6f);
-  for (int i = thread_idx; i < context_len; i += NUM_THREADS) {
+  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;
@@ -230,7 +276,7 @@ __global__ void single_query_cached_kv_attention_kernel(
 
   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;
+  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];
@@ -239,21 +285,36 @@ __global__ void single_query_cached_kv_attention_kernel(
     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];
+  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));
+    from_float(logits_vec, *reinterpret_cast<Float_L_vec*>(logits + token_idx - start_token_idx));
 
-    const scalar_t* v_ptr = v_cache + physical_block_number * num_heads * HEAD_SIZE * BLOCK_SIZE
-                                    + head_idx * HEAD_SIZE * BLOCK_SIZE;
+    const scalar_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 = *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);
       }
     }
@@ -308,7 +369,9 @@ __global__ void single_query_cached_kv_attention_kernel(
 
   // Write the final output.
   if (warp_idx == 0) {
-    scalar_t* out_ptr = out + seq_idx * num_heads * HEAD_SIZE + head_idx * HEAD_SIZE;
+    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;
@@ -319,32 +382,193 @@ __global__ void single_query_cached_kv_attention_kernel(
   }
 }
 
+// Grid: (num_heads, num_seqs, 1).
+template<
+  typename scalar_t,
+  int HEAD_SIZE,
+  int BLOCK_SIZE,
+  int NUM_THREADS>
+__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 scalar_t* __restrict__ k_cache,   // [num_blocks, num_kv_heads, head_size/x, block_size, x]
+  const scalar_t* __restrict__ v_cache,   // [num_blocks, num_kv_heads, head_size, block_size]
+  const int* __restrict__ head_mapping,   // [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, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS>(
+    /* exp_sums */ nullptr, /* max_logits */ nullptr,
+    out, q, k_cache, v_cache, head_mapping, 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,
+  int HEAD_SIZE,
+  int BLOCK_SIZE,
+  int NUM_THREADS,
+  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 scalar_t* __restrict__ k_cache,   // [num_blocks, num_kv_heads, head_size/x, block_size, x]
+  const scalar_t* __restrict__ v_cache,   // [num_blocks, num_kv_heads, head_size, block_size]
+  const int* __restrict__ head_mapping,   // [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, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS, PARTITION_SIZE>(
+    exp_sums, max_logits, tmp_out, q, k_cache, v_cache, head_mapping, 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, __shfl_xor_sync(uint32_t(-1), 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, __shfl_xor_sync(uint32_t(-1), max_logit, mask));
+  }
+  // Broadcast the max value to all threads.
+  max_logit = __shfl_sync(uint32_t(-1), 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_ATTENTION_KERNEL(T, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS)                        \
-  vllm::single_query_cached_kv_attention_kernel<T, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS>        \
+#define LAUNCH_PAGED_ATTENTION_V1(HEAD_SIZE)                                                  \
+  cudaFuncSetAttribute(                                                                       \
+    vllm::paged_attention_v1_kernel<T, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS>,                   \
+    cudaFuncAttributeMaxDynamicSharedMemorySize, shared_mem_size);                            \
+  vllm::paged_attention_v1_kernel<T, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS>                      \
   <<<grid, block, shared_mem_size, stream>>>(                                                 \
     out_ptr,                                                                                  \
     query_ptr,                                                                                \
     key_cache_ptr,                                                                            \
     value_cache_ptr,                                                                          \
+    head_mapping_ptr,                                                                         \
     scale,                                                                                    \
     block_tables_ptr,                                                                         \
     context_lens_ptr,                                                                         \
     max_num_blocks_per_seq,                                                                   \
     alibi_slopes_ptr,                                                                         \
-    query_stride);
+    q_stride,                                                                                 \
+    kv_block_stride,                                                                          \
+    kv_head_stride);
 
 // TODO(woosuk): Tune NUM_THREADS.
 template<
   typename T,
   int BLOCK_SIZE,
   int NUM_THREADS = 128>
-void single_query_cached_kv_attention_launcher(
+void paged_attention_v1_launcher(
   torch::Tensor& out,
   torch::Tensor& query,
   torch::Tensor& key_cache,
   torch::Tensor& value_cache,
+  torch::Tensor& head_mapping,
   float scale,
   torch::Tensor& block_tables,
   torch::Tensor& context_lens,
@@ -354,7 +578,9 @@ void single_query_cached_kv_attention_launcher(
   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 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);
@@ -368,60 +594,56 @@ void single_query_cached_kv_attention_launcher(
   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* head_mapping_ptr = reinterpret_cast<int*>(head_mapping.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 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);
+  dim3 grid(num_heads, num_seqs, 1);
   dim3 block(NUM_THREADS);
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
   switch (head_size) {
-    // NOTE(woosuk): To reduce the compilation time, we omitted head sizes
-    // 32, 160, 192, 256.
-    // case 32:
-    //   LAUNCH_ATTENTION_KERNEL(T, 32, BLOCK_SIZE, NUM_THREADS);
-    //   break;
+    // 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_ATTENTION_KERNEL(T, 64, BLOCK_SIZE, NUM_THREADS);
+      LAUNCH_PAGED_ATTENTION_V1(64);
       break;
     case 80:
-      LAUNCH_ATTENTION_KERNEL(T, 80, BLOCK_SIZE, NUM_THREADS);
+      LAUNCH_PAGED_ATTENTION_V1(80);
       break;
     case 96:
-      LAUNCH_ATTENTION_KERNEL(T, 96, BLOCK_SIZE, NUM_THREADS);
+      LAUNCH_PAGED_ATTENTION_V1(96);
       break;
     case 112:
-      LAUNCH_ATTENTION_KERNEL(T, 112, BLOCK_SIZE, NUM_THREADS);
+      LAUNCH_PAGED_ATTENTION_V1(112);
       break;
     case 128:
-      LAUNCH_ATTENTION_KERNEL(T, 128, BLOCK_SIZE, NUM_THREADS);
+      LAUNCH_PAGED_ATTENTION_V1(128);
+      break;
+    case 256:
+      LAUNCH_PAGED_ATTENTION_V1(256);
       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:
       TORCH_CHECK(false, "Unsupported head size: ", head_size);
       break;
   }
 }
 
-#define CALL_KERNEL_LAUNCHER(T, BLOCK_SIZE)                         \
-  single_query_cached_kv_attention_launcher<T, BLOCK_SIZE>(         \
+#define CALL_V1_LAUNCHER(T, BLOCK_SIZE)                             \
+  paged_attention_v1_launcher<T, BLOCK_SIZE>(                       \
     out,                                                            \
     query,                                                          \
     key_cache,                                                      \
     value_cache,                                                    \
+    head_mapping,                                                   \
     scale,                                                          \
     block_tables,                                                   \
     context_lens,                                                   \
@@ -430,45 +652,28 @@ void single_query_cached_kv_attention_launcher(
 
 // NOTE(woosuk): To reduce the compilation time, we omitted block sizes
 // 1, 2, 4, 64, 128, 256.
-#define CALL_KERNEL_LAUNCHER_BLOCK_SIZE(T)                          \
+#define CALL_V1_LAUNCHER_BLOCK_SIZE(T)                              \
   switch (block_size) {                                             \
-    /* case 1:                         */                           \
-    /*   CALL_KERNEL_LAUNCHER(T, 1);   */                           \
-    /*   break;                        */                           \
-    /* case 2:                         */                           \
-    /*   CALL_KERNEL_LAUNCHER(T, 2);   */                           \
-    /*   break;                        */                           \
-    /* case 4:                         */                           \
-    /*   CALL_KERNEL_LAUNCHER(T, 4);   */                           \
-    /*   break;                        */                           \
     case 8:                                                         \
-      CALL_KERNEL_LAUNCHER(T, 8);                                   \
+      CALL_V1_LAUNCHER(T, 8);                                       \
       break;                                                        \
     case 16:                                                        \
-      CALL_KERNEL_LAUNCHER(T, 16);                                  \
+      CALL_V1_LAUNCHER(T, 16);                                      \
       break;                                                        \
     case 32:                                                        \
-      CALL_KERNEL_LAUNCHER(T, 32);                                  \
+      CALL_V1_LAUNCHER(T, 32);                                      \
       break;                                                        \
-    /* case 64:                        */                           \
-    /*   CALL_KERNEL_LAUNCHER(T, 64);  */                           \
-    /*   break;                        */                           \
-    /* case 128:                       */                           \
-    /*   CALL_KERNEL_LAUNCHER(T, 128); */                           \
-    /*   break;                        */                           \
-    /* case 256:                       */                           \
-    /*   CALL_KERNEL_LAUNCHER(T, 256); */                           \
-    /*   break;                        */                           \
     default:                                                        \
       TORCH_CHECK(false, "Unsupported block size: ", block_size);   \
       break;                                                        \
   }
 
-void single_query_cached_kv_attention(
+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]
+  torch::Tensor& head_mapping,    // [num_heads]
   float scale,
   torch::Tensor& block_tables,    // [num_seqs, max_num_blocks_per_seq]
   torch::Tensor& context_lens,    // [num_seqs]
@@ -476,11 +681,186 @@ void single_query_cached_kv_attention(
   int max_context_len,
   const c10::optional<torch::Tensor>& alibi_slopes) {
   if (query.dtype() == at::ScalarType::Float) {
-    CALL_KERNEL_LAUNCHER_BLOCK_SIZE(float);
+    CALL_V1_LAUNCHER_BLOCK_SIZE(float);
   } else if (query.dtype() == at::ScalarType::Half) {
-    CALL_KERNEL_LAUNCHER_BLOCK_SIZE(uint16_t);
+    CALL_V1_LAUNCHER_BLOCK_SIZE(uint16_t);
   } else if (query.dtype() == at::ScalarType::BFloat16) {
-    CALL_KERNEL_LAUNCHER_BLOCK_SIZE(__nv_bfloat16);
+    CALL_V1_LAUNCHER_BLOCK_SIZE(__nv_bfloat16);
+  } else {
+    TORCH_CHECK(false, "Unsupported data type: ", query.dtype());
+  }
+}
+
+#define LAUNCH_PAGED_ATTENTION_V2(HEAD_SIZE)                                                  \
+  vllm::paged_attention_v2_kernel<T, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS, 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,                                                                          \
+    head_mapping_ptr,                                                                         \
+    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,
+  int BLOCK_SIZE,
+  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,
+  torch::Tensor& head_mapping,
+  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());
+  T* key_cache_ptr = reinterpret_cast<T*>(key_cache.data_ptr());
+  T* value_cache_ptr = reinterpret_cast<T*>(value_cache.data_ptr());
+  int* head_mapping_ptr = reinterpret_cast<int*>(head_mapping.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 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, BLOCK_SIZE)                             \
+  paged_attention_v2_launcher<T, BLOCK_SIZE>(                       \
+    out,                                                            \
+    exp_sums,                                                       \
+    max_logits,                                                     \
+    tmp_out,                                                        \
+    query,                                                          \
+    key_cache,                                                      \
+    value_cache,                                                    \
+    head_mapping,                                                   \
+    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)                              \
+  switch (block_size) {                                             \
+    case 8:                                                         \
+      CALL_V2_LAUNCHER(T, 8);                                       \
+      break;                                                        \
+    case 16:                                                        \
+      CALL_V2_LAUNCHER(T, 16);                                      \
+      break;                                                        \
+    case 32:                                                        \
+      CALL_V2_LAUNCHER(T, 32);                                      \
+      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]
+  torch::Tensor& head_mapping,    // [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) {
+  if (query.dtype() == at::ScalarType::Float) {
+    CALL_V2_LAUNCHER_BLOCK_SIZE(float);
+  } else if (query.dtype() == at::ScalarType::Half) {
+    CALL_V2_LAUNCHER_BLOCK_SIZE(uint16_t);
+  } else if (query.dtype() == at::ScalarType::BFloat16) {
+    CALL_V2_LAUNCHER_BLOCK_SIZE(__nv_bfloat16);
   } else {
     TORCH_CHECK(false, "Unsupported data type: ", query.dtype());
   }
@@ -489,3 +869,4 @@ void single_query_cached_kv_attention(
 #undef WARP_SIZE
 #undef MAX
 #undef MIN
+#undef DIVIDE_ROUND_UP

csrc/attention/dtype_bfloat16.cuh

@@ -420,4 +420,19 @@ inline __device__ void from_float(bf16_8_t& dst, Float8_ src) {
 #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

@@ -390,11 +390,6 @@ inline __device__ float sum(uint4 v) {
   return sum(c);
 }
 
-// Zero-out a vector.
-inline __device__ void zero(uint16_t& dst) {
-  dst = uint16_t(0);
-}
-
 // From float32 to float16.
 inline __device__ void from_float(uint16_t& dst, float src) {
   dst = float_to_half(src);
@@ -441,4 +436,9 @@ inline __device__ Float8_ to_float(uint4 u) {
   return tmp;
 }
 
+// Zero-out a variable.
+inline __device__ void zero(uint16_t& dst) {
+  dst = uint16_t(0);
+}
+
 } // namespace vllm

csrc/attention/dtype_float32.cuh

@@ -265,4 +265,9 @@ 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/cache.h

@@ -26,22 +26,3 @@ void gather_cached_kv(
   torch::Tensor& key_cache,
   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");
-}

csrc/cache_kernels.cu

@@ -1,6 +1,8 @@
 #include <torch/extension.h>
 #include <ATen/cuda/CUDAContext.h>
 
+#include "dispatch_utils.h"
+
 #include <algorithm>
 #include <cassert>
 #include <map>
@@ -53,26 +55,26 @@ 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];
   }
 }
@@ -100,15 +102,15 @@ 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.
@@ -118,21 +120,19 @@ 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 cudaStream_t stream = at::cuda::getCurrentCUDAStream();
-  AT_DISPATCH_FLOATING_TYPES_AND2(
-    at::ScalarType::Half,
-    at::ScalarType::BFloat16,
+  VLLM_DISPATCH_FLOATING_TYPES(
     key_caches[0].scalar_type(), "copy_blocks_kernel", ([&] {
       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);
     }));
 }
@@ -141,43 +141,48 @@ namespace vllm {
 
 template<typename scalar_t>
 __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]
+  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 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;
+    key_cache[tgt_key_idx] = key[src_key_idx];
+    value_cache[tgt_value_idx] = value[src_value_idx];
   }
 }
 
@@ -202,9 +207,7 @@ void reshape_and_cache(
   dim3 grid(num_tokens);
   dim3 block(std::min(num_heads * head_size, 512));
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
-  AT_DISPATCH_FLOATING_TYPES_AND2(
-    at::ScalarType::Half,
-    at::ScalarType::BFloat16,
+  VLLM_DISPATCH_FLOATING_TYPES(
     key.scalar_type(),
     "reshape_and_cache_kernel",
     [&] {
@@ -213,7 +216,7 @@ void reshape_and_cache(
         value.data_ptr<scalar_t>(),
         key_cache.data_ptr<scalar_t>(),
         value_cache.data_ptr<scalar_t>(),
-        slot_mapping.data_ptr<int>(),
+        slot_mapping.data_ptr<int64_t>(),
         key_stride,
         value_stride,
         num_heads,
@@ -364,9 +367,7 @@ void gather_cached_kv(
   dim3 grid(num_tokens);
   dim3 block(std::min(num_heads * head_size, 512));
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
-  AT_DISPATCH_FLOATING_TYPES_AND2(
-    at::ScalarType::Half,
-    at::ScalarType::BFloat16,
+  VLLM_DISPATCH_FLOATING_TYPES(
     key.scalar_type(),
     "gather_cached_kv_kernel_optimized",
     [&] {

csrc/cuda_utils.h

@@ -0,0 +1,5 @@
+#include <torch/extension.h>
+
+int get_device_attribute(
+    int attribute,
+    int device_id);

csrc/cuda_utils_kernels.cu

@@ -0,0 +1,14 @@
+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;
+}

csrc/dispatch_utils.h

@@ -0,0 +1,14 @@
+/*
+ * Adapted from
+ * https://github.com/pytorch/pytorch/blob/v2.0.1/aten/src/ATen/Dispatch.h
+ */
+#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__))

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

@@ -1,6 +1,7 @@
 #include <torch/extension.h>
 #include <ATen/cuda/CUDAContext.h>
 
+#include "dispatch_utils.h"
 #include "reduction_utils.cuh"
 
 namespace vllm {
@@ -8,8 +9,8 @@ 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,
@@ -33,22 +34,50 @@ __global__ void rms_norm_kernel(
   }
 }
 
+// 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 cudaStream_t stream = at::cuda::getCurrentCUDAStream();
-  AT_DISPATCH_FLOATING_TYPES_AND2(
-    at::ScalarType::Half,
-    at::ScalarType::BFloat16,
+  VLLM_DISPATCH_FLOATING_TYPES(
     input.scalar_type(),
     "rms_norm_kernel",
     [&] {
@@ -61,3 +90,28 @@ 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 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/ops.h

@@ -0,0 +1,75 @@
+#include <torch/extension.h>
+
+void paged_attention_v1(
+  torch::Tensor& out,
+  torch::Tensor& query,
+  torch::Tensor& key_cache,
+  torch::Tensor& value_cache,
+  torch::Tensor& head_mapping,
+  float scale,
+  torch::Tensor& block_tables,
+  torch::Tensor& context_lens,
+  int block_size,
+  int max_context_len,
+  const c10::optional<torch::Tensor>& alibi_slopes);
+
+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,
+  torch::Tensor& head_mapping,
+  float scale,
+  torch::Tensor& block_tables,
+  torch::Tensor& context_lens,
+  int block_size,
+  int max_context_len,
+  const c10::optional<torch::Tensor>& alibi_slopes);
+
+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);
+
+torch::Tensor awq_gemm(
+  torch::Tensor _in_feats,
+  torch::Tensor _kernel,
+  torch::Tensor _scaling_factors,
+  torch::Tensor _zeros,
+  int split_k_iters);
+
+void squeezellm_gemm(
+  torch::Tensor vec,
+  torch::Tensor mat,
+  torch::Tensor mul,
+  torch::Tensor lookup_table);

csrc/pos_encoding.cpp

@@ -1,15 +0,0 @@
-#include <torch/extension.h>
-
-void rotary_embedding_neox(
-  torch::Tensor& positions,
-  torch::Tensor& query,
-  torch::Tensor& key,
-  int head_size,
-  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

@@ -1,17 +1,51 @@
 #include <torch/extension.h>
 #include <ATen/cuda/CUDAContext.h>
 
+#include "dispatch_utils.h"
+
 namespace vllm {
 
-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]
+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 = __ldg(cos_ptr + x_index);
+    sin = __ldg(sin_ptr + x_index);
+  } else {
+    // GPT-J style rotary embedding.
+    x_index = 2 * rot_offset;
+    y_index = 2 * rot_offset + 1;
+    cos = __ldg(cos_ptr + x_index / 2);
+    sin = __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 int stride,
+  const int query_stride,
+  const int 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;
@@ -19,65 +53,75 @@ __global__ void rotary_embedding_neox_kernel(
   const scalar_t* cache_ptr = cos_sin_cache + pos * rot_dim;
 
   const int embed_dim = rot_dim / 2;
-  const int n = num_heads * embed_dim;
-  for (int i = threadIdx.x; i < n; i += blockDim.x) {
+  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 int 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 int 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 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]
+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]
-{
-  int num_tokens = query.size(0);
+  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 stride = query.stride(0);
-  TORCH_CHECK(stride == key.stride(0));
+  int num_heads = query.size(-1) / head_size;
+  int num_kv_heads = key.size(-1) / head_size;
+  int query_stride = query.stride(-2);
+  int key_stride = key.stride(-2);
 
   dim3 grid(num_tokens);
   dim3 block(std::min(num_heads * rot_dim / 2, 512));
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
-  AT_DISPATCH_FLOATING_TYPES_AND2(
-    at::ScalarType::Half,
-    at::ScalarType::BFloat16,
+  VLLM_DISPATCH_FLOATING_TYPES(
     query.scalar_type(),
-    "rotary_embedding_neox",
+    "rotary_embedding",
     [&] {
-      vllm::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>(),
-        rot_dim,
-        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/pybind.cpp

@@ -0,0 +1,80 @@
+#include "cache.h"
+#include "cuda_utils.h"
+#include "ops.h"
+#include <torch/extension.h>
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+  // vLLM custom ops
+  pybind11::module ops = m.def_submodule("ops", "vLLM custom operators");
+
+  // Attention ops
+  ops.def(
+    "paged_attention_v1",
+    &paged_attention_v1,
+    "Compute the attention between an input query and the cached keys/values using PagedAttention.");
+  ops.def(
+    "paged_attention_v2",
+    &paged_attention_v2,
+    "PagedAttention V2.");
+
+  // Activation ops
+  ops.def(
+    "silu_and_mul",
+    &silu_and_mul,
+    "Activation function used in SwiGLU.");
+  ops.def(
+    "gelu_new",
+    &gelu_new,
+    "GELU implementation used in GPT-2.");
+  ops.def(
+    "gelu_fast",
+    &gelu_fast,
+    "Approximate GELU implementation.");
+
+  // Layernorm
+  ops.def(
+    "rms_norm",
+    &rms_norm,
+    "Apply Root Mean Square (RMS) Normalization to the input tensor.");
+
+  ops.def(
+    "fused_add_rms_norm",
+    &fused_add_rms_norm,
+    "In-place fused Add and RMS Normalization");
+
+  // Rotary embedding
+  ops.def(
+    "rotary_embedding",
+    &rotary_embedding,
+    "Apply GPT-NeoX or GPT-J style rotary embedding to query and key");
+
+  // Quantization ops
+  ops.def("awq_gemm", &awq_gemm, "Quantized GEMM for AWQ");
+  ops.def("squeezellm_gemm", &squeezellm_gemm, "Quantized GEMM for SqueezeLLM");
+
+  // Cache ops
+  pybind11::module cache_ops = m.def_submodule("cache_ops", "vLLM cache ops");
+  cache_ops.def(
+    "swap_blocks",
+    &swap_blocks,
+    "Swap in (out) the cache blocks from src to dst");
+  cache_ops.def(
+    "copy_blocks",
+    &copy_blocks,
+    "Copy the cache blocks from src to dst");
+  cache_ops.def(
+    "reshape_and_cache",
+    &reshape_and_cache,
+    "Reshape the key and value tensors and cache them");
+  cache_ops.def(
+    "gather_cached_kv",
+    &gather_cached_kv,
+    "Gather key and value from the cache into contiguous QKV tensors");
+
+  // Cuda utils
+  pybind11::module cuda_utils = m.def_submodule("cuda_utils", "vLLM cuda utils");
+  cuda_utils.def(
+    "get_device_attribute",
+    &get_device_attribute,
+    "Gets the specified device attribute.");
+}

csrc/quantization/awq/dequantize.cuh

@@ -0,0 +1,87 @@
+/*
+Adapted from https://github.com/mit-han-lab/llm-awq
+Modified from NVIDIA FasterTransformer: https://github.com/NVIDIA/FasterTransformer/blob/main/src/fastertransformer/cutlass_extensions/include/cutlass_extensions/interleaved_numeric_conversion.h
+@article{lin2023awq,
+  title={AWQ: Activation-aware Weight Quantization for LLM Compression and Acceleration},
+  author={Lin, Ji and Tang, Jiaming and Tang, Haotian and Yang, Shang and Dang, Xingyu and Han, Song},
+  journal={arXiv},
+  year={2023}
+}
+*/
+
+#pragma once
+
+namespace vllm {
+namespace awq {
+
+__device__ uint4 dequantize_s4_to_fp16x2(uint32_t const& source)
+{
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 750
+  assert(false);
+#else
+    uint4 result;
+
+    uint32_t*      h   = reinterpret_cast<uint32_t*>(&result);
+    uint32_t const i4s = reinterpret_cast<uint32_t const&>(source);
+
+    // First, we extract the i4s and construct an intermediate fp16 number.
+    static constexpr uint32_t immLut                = (0xf0 & 0xcc) | 0xaa;
+    static constexpr uint32_t BOTTOM_MASK           = 0x000f000f;
+    static constexpr uint32_t TOP_MASK              = 0x00f000f0;
+    static constexpr uint32_t I4s_TO_F16s_MAGIC_NUM = 0x64006400;
+
+    // Note that the entire sequence only requires 1 shift instruction. This is thanks to the register packing
+    // format and the fact that we force our integers to be unsigned, and account for this in the fp16 subtractions.
+    // In addition, I exploit the fact that sub and fma have the same throughput in order to convert elt_23 and
+    // elt_67 to fp16 without having to shift them to the bottom bits before hand.
+
+    // Shift right by 8 to now consider elt_45 and elt_67. Issue first to hide RAW dependency if we issue
+    // immediately before required.
+    const uint32_t top_i4s = i4s >> 8;
+    // Extract elt_01 - (i4s & 0x000f000f) | 0x64006400
+    asm volatile("lop3.b32 %0, %1, %2, %3, %4;\n"
+                    : "=r"(h[0])
+                    : "r"(i4s), "n"(BOTTOM_MASK), "n"(I4s_TO_F16s_MAGIC_NUM), "n"(immLut));
+    // Extract elt_23 (i4s & 0x00f000f0) | 0x64006400
+    asm volatile("lop3.b32 %0, %1, %2, %3, %4;\n"
+                    : "=r"(h[1])
+                    : "r"(i4s), "n"(TOP_MASK), "n"(I4s_TO_F16s_MAGIC_NUM), "n"(immLut));
+    // Extract elt_45 (top_i4s & 0x000f000f) | 0x64006400
+    asm volatile("lop3.b32 %0, %1, %2, %3, %4;\n"
+                    : "=r"(h[2])
+                    : "r"(top_i4s), "n"(BOTTOM_MASK), "n"(I4s_TO_F16s_MAGIC_NUM), "n"(immLut));
+    // Extract elt_67 (top_i4s & 0x00f000f0) | 0x64006400
+    asm volatile("lop3.b32 %0, %1, %2, %3, %4;\n"
+                    : "=r"(h[3])
+                    : "r"(top_i4s), "n"(TOP_MASK), "n"(I4s_TO_F16s_MAGIC_NUM), "n"(immLut));
+
+    // I use inline PTX below because I am not sure if the compiler will emit float2half instructions if I use the
+    // half2 ctor. In this case, I chose performance reliability over code readability.
+
+    // This is the half2 {1032, 1032} represented as an integer.
+    // static constexpr uint32_t FP16_TOP_MAGIC_NUM = 0x64086408;
+    // Haotian: subtract {1024, 1024} instead, we do not need to map to [-8, 7]
+    static constexpr uint32_t FP16_TOP_MAGIC_NUM = 0x64006400;
+    // This is the half2 {1 / 16, 1 / 16} represented as an integer.
+    static constexpr uint32_t ONE_SIXTEENTH = 0x2c002c00;
+    // This is the half2 {-72, -72} represented as an integer.
+    // static constexpr uint32_t NEG_72 = 0xd480d480;
+    // Haotian: Let's use {-64, -64}.
+    static constexpr uint32_t NEG_64 = 0xd400d400;
+
+    // Finally, we construct the output numbers.
+    // Convert elt_01
+    asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(h[0]) : "r"(h[0]), "r"(FP16_TOP_MAGIC_NUM));
+    // Convert elt_23
+    asm volatile("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(h[1]) : "r"(h[1]), "r"(ONE_SIXTEENTH), "r"(NEG_64));
+    // Convert elt_45
+    asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(h[2]) : "r"(h[2]), "r"(FP16_TOP_MAGIC_NUM));
+    // Convert elt_67
+    asm volatile("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(h[3]) : "r"(h[3]), "r"(ONE_SIXTEENTH), "r"(NEG_64));
+
+    return result;
+#endif
+}
+
+} // namespace awq
+} // namespace vllm

csrc/quantization/awq/gemm_kernels.cu

@@ -0,0 +1,560 @@
+/*
+Adapted from https://github.com/mit-han-lab/llm-awq
+@article{lin2023awq,
+  title={AWQ: Activation-aware Weight Quantization for LLM Compression and Acceleration},
+  author={Lin, Ji and Tang, Jiaming and Tang, Haotian and Yang, Shang and Dang, Xingyu and Han, Song},
+  journal={arXiv},
+  year={2023}
+}
+ */
+
+
+#include <torch/extension.h>
+#include <c10/cuda/CUDAGuard.h>
+
+#include "dequantize.cuh"
+
+#include <cuda_fp16.h>
+
+namespace vllm {
+namespace awq {
+
+// Pack two half values.
+static inline __device__ __host__ unsigned
+__pack_half2(const half x, const half y) {
+  unsigned v0 = *((unsigned short *)&x);
+  unsigned v1 = *((unsigned short *)&y);
+  return (v1 << 16) | v0;
+}
+
+__global__ void __launch_bounds__(64) gemm_forward_4bit_cuda_m16n128k32(int G, int split_k_iters, half* __restrict__ A, int* __restrict__ B, half* __restrict__ scaling_factors, int* __restrict__ zeros, int M, int IC, int OC, half* __restrict__ C) 
+{
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 750
+  assert(false);
+#else
+  static constexpr uint32_t ZERO = 0x0;
+  float C_warp[32];
+  __shared__ half A_shared[16 * (32 + 8)];
+  __shared__ half B_shared[32 * (128 + 8)];
+  
+  __shared__ half scaling_factors_shared[128];
+  __shared__ half zeros_shared[128];
+
+  int j_factors1 = ((OC + 128 - 1) / 128);
+  int blockIdx_x = 0;
+  int blockIdx_y = blockIdx.x % ((M + 16 - 1) / 16 * j_factors1);
+  int blockIdx_z = blockIdx.x / ((M + 16 - 1) / 16 * j_factors1);
+
+  half A_shared_warp[8];
+  half B_shared_warp[32];
+  for (int j_0_4_init = 0; j_0_4_init < 4; ++j_0_4_init) {
+    for (int i = 0; i < 8; ++i) {
+      C_warp[(j_0_4_init * 8) + i] = 0.0;
+    }
+  }
+
+  static constexpr int row_stride_warp = 32 * 8 / 32;
+  static constexpr int row_stride = 2 * 32 * 8 / 128;
+  bool ld_zero_flag = (threadIdx.y * 32 + threadIdx.x) * 8 < 128;
+  // TODO: Haotian: blockIdx_y / j_factors1 in A loading to support bsz > 16
+  bool ld_A_flag = (blockIdx_y / j_factors1 * 16 + threadIdx.y * row_stride_warp + threadIdx.x * 8 / 32) < M;     // threadIdx.y is warp_id
+  // bool wb_C_flag = (threadIdx.x / 4) < M;
+
+  half* A_ptr = A 
+                + (((int)blockIdx_y) / j_factors1 * 16 + (((int)threadIdx.y) * row_stride_warp) + ((int)threadIdx.x) / (32 / 8)) * IC
+                + (((int)threadIdx.x) % (32 / 8)) * 8;
+  
+  int* B_ptr = B
+            + ((int)threadIdx.y) * (OC / 8) * 2
+            + (((int)threadIdx.x) / (128 / 8)) * (OC / 8)
+            + (((int)blockIdx_y) % j_factors1) * (128 / 8)
+            + (((int)threadIdx.x) % (128 / 8)) * 1;
+// Why * 1 in the above line?
+                        
+  half* A_shared_ptr = A_shared 
+                    + ((int)threadIdx.y) * row_stride_warp * (32 + 8) 
+                    + (((int)threadIdx.x) / (32 / 8)) * (32 + 8)
+                    + (((int)threadIdx.x) % (32 / 8) ) * 8;
+
+  half* B_shared_ptr = B_shared
+                    + ((int)threadIdx.y) * (row_stride / 2) * (128 + 8)
+                    + (((int)threadIdx.x) / (128 / 8)) * (128 + 8)
+                    + (((int)threadIdx.x) % (128 / 8)) * 8;
+  
+  int* zeros_ptr = zeros
+                + (((int)blockIdx_y) % j_factors1) * (128 / 8)
+                + ((int)threadIdx.x) % (128 / 8);
+  
+  half* scaling_factors_ptr = scaling_factors
+                            + (((int)blockIdx_y) % j_factors1) * (128) 
+                            + (((int)threadIdx.x) % (128 / 8)) * 8;
+
+  half* C_ptr = C 
+              + static_cast<long long>(blockIdx_z) * M * OC        // blockIdz.x -> split_k dim
+              + (((int)blockIdx_y) % j_factors1) * 128
+              + ((int)threadIdx.y) * 64
+              + (((int)threadIdx.x) % 4) * 2;
+
+  // preload s.f. and zeros
+  int k_bound = (IC / 32 + split_k_iters - 1) / split_k_iters;
+  if ((k_bound - 1) * split_k_iters * 32 + blockIdx_z * 32 >= IC) k_bound -= 1;
+  for (int _k_0_0 = 0; _k_0_0 < k_bound; ++_k_0_0) {
+    int k_0_0 = _k_0_0 * split_k_iters + blockIdx_z;
+    __syncthreads();
+    // TODO: Haotian: blockIdx_y / j_factors1 in A loading to support bsz > 16
+    if (ld_A_flag)
+    {
+      *(uint4*)(A_shared_ptr) = *(uint4*)(A_ptr + (k_0_0 * 32));
+    }
+    else
+    {
+      *(uint4*)(A_shared_ptr) = make_uint4(0, 0, 0, 0);
+    }
+
+    // for (int ax0_ax1_fused_0 = 0; ax0_ax1_fused_0 < 2; ++ax0_ax1_fused_0) {
+    uint32_t zeros_loaded = *(uint32_t*)(zeros_ptr + k_0_0 * 32 / G * (OC / 8));
+    uint4 B_loaded_zero = dequantize_s4_to_fp16x2(zeros_loaded);
+    uint4 B_loaded_scale = *(uint4*)(scaling_factors_ptr + k_0_0 * 32 / G * (OC));
+    /*
+    if (blockIdx_z == 0 && blockIdx_y == 0 && k_0_0 == 0 && threadIdx.x == 0 && threadIdx.y == 0){
+      printf("%x %x %x %x %x %x %x %x\n", B_loaded_scale.x, B_loaded_scale.y, B_loaded_scale.z, B_loaded_scale.w, B_loaded_zero.x, B_loaded_zero.y, B_loaded_zero.z, B_loaded_zero.w);
+    }
+    */
+    // uint4 B_loaded_scale = make_uint4(0, 0, 0, 0);
+    int* B_ptr_local = B_ptr + k_0_0 * 32 * (OC / 8);
+
+    for (int ax0_ax1_fused_0 = 0; ax0_ax1_fused_0 < 8; ++ax0_ax1_fused_0) {
+
+      // B: 32 x 136 (128+8) float16
+      // each warp: 32 x 4
+      // each thr: read 32 bit -> convert to 8xFP16 (a UINT4) -> scale and minus zero -> WB UINT4
+      // *(uint4*)(B_shared + ((((ax0_ax1_fused_0 * 544) + (((int)threadIdx.y) * 272)) + ((((int)threadIdx.x) >> 4) * 136)) + ((((int)threadIdx.x) & 15) * 8))) = *(uint4*)(B + ((((((k_0_0 * 163840) + (ax0_ax1_fused_0 * 20480)) + (((int)threadIdx.y) * 10240)) + ((((int)threadIdx.x) >> 4) * 5120)) + (((int)blockIdx_y) * 128)) + ((((int)threadIdx.x) & 15) * 8)));
+      // row stride in shared memory: (NWARPS * 32 * 8 / cta_N) 
+      uint32_t B_loaded = *(uint32_t*)(B_ptr_local + ax0_ax1_fused_0 * row_stride * (OC / 8));
+      uint4 B_loaded_fp16 = dequantize_s4_to_fp16x2(B_loaded);
+      //uint4 B_loaded_zero = *(uint4*)(zeros_shared + (threadIdx.x % (cta_N / 8)) * 8);
+
+      // uint4 B_loaded_scale = *(uint4*)(scaling_factors_shared + (threadIdx.x % (cta_N / 8)) * 8);
+      // - zero and * scale
+      // TODO (Haotian): can save 4 assembly instructions if sormulate as deq = q * scale - zero * scale.
+      asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(B_loaded_fp16.x) : "r"(B_loaded_fp16.x), "r"(B_loaded_zero.x));
+      asm volatile("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(B_loaded_fp16.x) : "r"(B_loaded_fp16.x), "r"(B_loaded_scale.x), "r"(ZERO));
+      asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(B_loaded_fp16.y) : "r"(B_loaded_fp16.y), "r"(B_loaded_zero.y));
+      asm volatile("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(B_loaded_fp16.y) : "r"(B_loaded_fp16.y), "r"(B_loaded_scale.y), "r"(ZERO));
+      asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(B_loaded_fp16.z) : "r"(B_loaded_fp16.z), "r"(B_loaded_zero.z));
+      asm volatile("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(B_loaded_fp16.z) : "r"(B_loaded_fp16.z), "r"(B_loaded_scale.z), "r"(ZERO));
+      asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(B_loaded_fp16.w) : "r"(B_loaded_fp16.w), "r"(B_loaded_zero.w));
+      asm volatile("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(B_loaded_fp16.w) : "r"(B_loaded_fp16.w), "r"(B_loaded_scale.w), "r"(ZERO));
+      /*
+      if (ax0_ax1_fused_0 == 0 && blockIdx_z == 0 && blockIdx_y == 0 && k_0_0 == 0 && threadIdx.x == 17 && threadIdx.y == 0){
+        printf("[x] %X %X %X %X\n", B_loaded_fp16.x, B_loaded_fp16.y, B_loaded_fp16.z, B_loaded_fp16.w);
+      }
+      */
+
+      // write back
+      *(uint4*)(B_shared_ptr + ax0_ax1_fused_0 * row_stride * (128 + 8)) = B_loaded_fp16;
+    }
+    __syncthreads();
+
+    for (int k_0_1 = 0; k_0_1 < 2; ++k_0_1) {
+      {
+        unsigned int addr;
+        __asm__ __volatile__(
+          "{ .reg .u64 addr; cvta.to.shared.u64 addr, %1; cvt.u32.u64 %0, addr; }\n"
+          : "=r"(addr)
+          : "l"((void *)((&(A_shared[(k_0_1 * 16)])) + (((((int)threadIdx.x) & 15) * 40) + ((((int)threadIdx.x) >> 4) * 8))))
+        );
+
+
+        __asm__ __volatile__(
+          "ldmatrix.sync.aligned.m8n8.x4.shared.b16"
+          "{%0, %1, %2, %3}, [%4];\n"
+          : "=r"(((unsigned *)(A_shared_warp + 0))[0]), "=r"(((unsigned *)(A_shared_warp + 0))[1]), "=r"(((unsigned *)(A_shared_warp + 0))[2]), "=r"(((unsigned *)(A_shared_warp + 0))[3])
+          : "r"(addr)
+        );
+      }
+
+      for (int ax1_0 = 0; ax1_0 < 4; ++ax1_0) {
+        {
+          unsigned int addr;
+          __asm__ __volatile__(
+            "{ .reg .u64 addr; cvta.to.shared.u64 addr, %1; cvt.u32.u64 %0, addr; }\n"
+            : "=r"(addr)
+            : "l"((void *)((&(B_shared[(((k_0_1 * 2176) + (((int)threadIdx.y) * 64)) + (ax1_0 * 16))])) + (((((int)threadIdx.x) & 15) * 136) + ((((int)threadIdx.x) >> 4) * 8))))
+          );
+          __asm__ __volatile__(
+            "ldmatrix.sync.aligned.m8n8.x4.trans.shared.b16"
+            "{%0, %1, %2, %3}, [%4];\n"
+            : "=r"(((unsigned *)(B_shared_warp + (ax1_0 * 8)))[0]), "=r"(((unsigned *)(B_shared_warp + (ax1_0 * 8)))[1]), "=r"(((unsigned *)(B_shared_warp + (ax1_0 * 8)))[2]), "=r"(((unsigned *)(B_shared_warp + (ax1_0 * 8)))[3])
+            : "r"(addr)
+          );
+        }
+      }
+      for (int j_0_4 = 0; j_0_4 < 4; ++j_0_4) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ == 750
+        {
+          __asm__ __volatile__(
+            "mma.sync.aligned.m16n8k8.row.col.f32.f16.f16.f32"
+            "{%0, %1, %2, %3}, {%4, %5}, {%6}, {%7, %8, %9, %10};\n"
+            :  "=f"(((float *)(C_warp + (j_0_4 * 8)))[0]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[1]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[2]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[3])
+            : "r"(((unsigned *)(A_shared_warp + 0))[0]), "r"(((unsigned *)(A_shared_warp + 0))[1]), "r"(((unsigned *)(B_shared_warp + (j_0_4 * 8)))[0]), "f"(((float *)(C_warp + (j_0_4 * 8)))[0]), "f"(((float *)(C_warp + (j_0_4 * 8)))[1]), "f"(((float *)(C_warp + (j_0_4 * 8)))[2]), "f"(((float *)(C_warp + (j_0_4 * 8)))[3]));
+        }
+
+        {
+          __asm__ __volatile__(
+            "mma.sync.aligned.m16n8k8.row.col.f32.f16.f16.f32"
+            "{%0, %1, %2, %3}, {%4, %5}, {%6}, {%7, %8, %9, %10};\n"
+            :  "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[0]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[1]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[2]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[3])
+            : "r"(((unsigned *)(A_shared_warp + 0))[0]), "r"(((unsigned *)(A_shared_warp + 0))[1]), "r"(((unsigned *)(B_shared_warp + ((j_0_4 * 8) + 4)))[0]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[0]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[1]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[2]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[3]));
+        }
+
+        {
+          __asm__ __volatile__(
+            "mma.sync.aligned.m16n8k8.row.col.f32.f16.f16.f32"
+            "{%0, %1, %2, %3}, {%4, %5}, {%6}, {%7, %8, %9, %10};\n"
+            :  "=f"(((float *)(C_warp + (j_0_4 * 8)))[0]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[1]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[2]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[3])
+            : "r"(((unsigned *)(A_shared_warp + 0))[2]), "r"(((unsigned *)(A_shared_warp + 0))[3]), "r"(((unsigned *)(B_shared_warp + (j_0_4 * 8)))[1]), "f"(((float *)(C_warp + (j_0_4 * 8)))[0]), "f"(((float *)(C_warp + (j_0_4 * 8)))[1]), "f"(((float *)(C_warp + (j_0_4 * 8)))[2]), "f"(((float *)(C_warp + (j_0_4 * 8)))[3]));
+        }
+
+        {
+          __asm__ __volatile__(
+            "mma.sync.aligned.m16n8k8.row.col.f32.f16.f16.f32"
+            "{%0, %1, %2, %3}, {%4, %5}, {%6}, {%7, %8, %9, %10};\n"
+            :  "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[0]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[1]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[2]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[3])
+            : "r"(((unsigned *)(A_shared_warp + 0))[2]), "r"(((unsigned *)(A_shared_warp + 0))[3]), "r"(((unsigned *)(B_shared_warp + ((j_0_4 * 8) + 4)))[1]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[0]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[1]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[2]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[3]));
+        }
+#else
+        {
+          __asm__ __volatile__(
+            "mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32"
+            "{%0, %1, %2, %3}, {%4, %5, %6, %7}, {%8, %9}, {%10, %11, %12, %13};\n"
+            :  "=f"(((float *)(C_warp + (j_0_4 * 8)))[0]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[1]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[2]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[3])
+            : "r"(((unsigned *)(A_shared_warp + 0))[0]), "r"(((unsigned *)(A_shared_warp + 0))[1]), "r"(((unsigned *)(A_shared_warp + 0))[2]), "r"(((unsigned *)(A_shared_warp + 0))[3]), "r"(((unsigned *)(B_shared_warp + (j_0_4 * 8)))[0]), "r"(((unsigned *)(B_shared_warp + (j_0_4 * 8)))[1]), "f"(((float *)(C_warp + (j_0_4 * 8)))[0]), "f"(((float *)(C_warp + (j_0_4 * 8)))[1]), "f"(((float *)(C_warp + (j_0_4 * 8)))[2]), "f"(((float *)(C_warp + (j_0_4 * 8)))[3]));
+        }
+
+        {
+          __asm__ __volatile__(
+            "mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32"
+            "{%0, %1, %2, %3}, {%4, %5, %6, %7}, {%8, %9}, {%10, %11, %12, %13};\n"
+            :  "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[0]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[1]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[2]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[3])
+            : "r"(((unsigned *)(A_shared_warp + 0))[0]), "r"(((unsigned *)(A_shared_warp + 0))[1]), "r"(((unsigned *)(A_shared_warp + 0))[2]), "r"(((unsigned *)(A_shared_warp + 0))[3]), "r"(((unsigned *)(B_shared_warp + ((j_0_4 * 8) + 4)))[0]), "r"(((unsigned *)(B_shared_warp + ((j_0_4 * 8) + 4)))[1]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[0]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[1]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[2]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[3]));
+        }
+
+#endif
+      }
+    }
+  }
+
+// TODO: Shang: Hoist loop invariance.
+  for (int ax1_0_1 = 0; ax1_0_1 < 4; ++ax1_0_1) {
+    for (int local_id = 0; local_id < 8; ++local_id) {
+      int row_offset = (((int)blockIdx_y) / j_factors1) * 16 + ((int)threadIdx.x) / 4 + (local_id % 4) / 2 * 8;
+      if (row_offset < M)
+      {
+        *(C_ptr + ax1_0_1 * 16 + row_offset * OC + (local_id / 4) * 8 + local_id % 2) = __float2half(C_warp[(ax1_0_1 * 8) + local_id]);
+      }
+    }
+  }
+#endif
+}
+
+
+__global__ void __launch_bounds__(64) gemm_forward_4bit_cuda_m16n64k32(int G, int split_k_iters, half* __restrict__ A, int* __restrict__ B, half* __restrict__ scaling_factors, int* __restrict__ zeros, int M, int IC, int OC, half* __restrict__ C) 
+{
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 750
+  assert(false);
+#else
+  static constexpr uint32_t ZERO = 0x0;
+  float C_warp[32];
+  __shared__ half A_shared[16 * (32 + 8)];
+  __shared__ half B_shared[32 * (64 + 8)];
+  
+  __shared__ half scaling_factors_shared[64];
+  __shared__ half zeros_shared[64];
+
+  int j_factors1 = ((OC + 64 - 1) / 64);
+
+  int blockIdx_x = 0;
+  int blockIdx_y = blockIdx.x % ((M + 16 - 1) / 16 * j_factors1);
+  int blockIdx_z = blockIdx.x / ((M + 16 - 1) / 16 * j_factors1);
+
+  half A_shared_warp[8];
+  half B_shared_warp[16];
+  for (int j_0_4_init = 0; j_0_4_init < 2; ++j_0_4_init) {
+    for (int i = 0; i < 8; ++i) {
+      C_warp[(j_0_4_init * 8) + i] = 0.0;
+    }
+  }
+
+  static constexpr int row_stride_warp = 32 * 8 / 32;
+  static constexpr int row_stride = 2 * 32 * 8 / 64;
+  bool ld_zero_flag = (threadIdx.y * 32 + threadIdx.x) * 8 < 64;
+  // TODO: Haotian: blockIdx_y / j_factors1 in A loading to support bsz > 16
+  bool ld_A_flag = (blockIdx_y / j_factors1 * 16 + threadIdx.y * row_stride_warp + threadIdx.x * 8 / 32) < M;     // threadIdx.y is warp_id
+  // bool wb_C_flag = (threadIdx.x / 4) < M;
+
+  half* A_ptr = A 
+                + (((int)blockIdx_y) / j_factors1 * 16 + (((int)threadIdx.y) * row_stride_warp) + ((int)threadIdx.x) / (32 / 8)) * IC
+                + (((int)threadIdx.x) % (32 / 8)) * 8;
+  
+  int* B_ptr = B
+            + ((int)threadIdx.y) * (OC / 8) * 4
+            + (((int)threadIdx.x) / (64 / 8)) * (OC / 8)
+            + (((int)blockIdx_y) % j_factors1) * (64 / 8)
+            + (((int)threadIdx.x) % (64 / 8)) * 1;
+// Why * 1 in the above line?
+                        
+  half* A_shared_ptr = A_shared 
+                    + ((int)threadIdx.y) * row_stride_warp * (32 + 8) 
+                    + (((int)threadIdx.x) / (32 / 8)) * (32 + 8)
+                    + (((int)threadIdx.x) % (32 / 8) ) * 8;
+
+  half* B_shared_ptr = B_shared
+                    + ((int)threadIdx.y) * (row_stride / 2) * (64 + 8)
+                    + (((int)threadIdx.x) / (64 / 8)) * (64 + 8)
+                    + (((int)threadIdx.x) % (64 / 8)) * 8;
+  
+  int* zeros_ptr = zeros
+                + (((int)blockIdx_y) % j_factors1) * (64 / 8)
+                + ((int)threadIdx.x) % (64 / 8);
+  
+  half* scaling_factors_ptr = scaling_factors
+                            + (((int)blockIdx_y) % j_factors1) * (64) 
+                            + (((int)threadIdx.x) % (64 / 8)) * 8;
+
+  half* C_ptr = C 
+              + static_cast<long long>(blockIdx_z) * M * OC        // blockIdz.x -> split_k dim
+              + (((int)blockIdx_y) % j_factors1) * 64
+              + ((int)threadIdx.y) * 32
+              + (((int)threadIdx.x) % 4) * 2;
+
+  // preload s.f. and zeros
+  int k_bound = (IC / 32 + split_k_iters - 1) / split_k_iters;
+  if ((k_bound - 1) * split_k_iters * 32 + blockIdx_z * 32 >= IC) k_bound -= 1;
+  for (int _k_0_0 = 0; _k_0_0 < k_bound; ++_k_0_0) {
+    int k_0_0 = _k_0_0 * split_k_iters + blockIdx_z;
+    __syncthreads();
+    // TODO: Haotian: blockIdx_y / j_factors1 in A loading to support bsz > 16
+    if (ld_A_flag)
+    {
+      *(uint4*)(A_shared_ptr) = *(uint4*)(A_ptr + (k_0_0 * 32));
+    }
+    else
+    {
+      *(uint4*)(A_shared_ptr) = make_uint4(0, 0, 0, 0);
+    }
+
+    // for (int ax0_ax1_fused_0 = 0; ax0_ax1_fused_0 < 2; ++ax0_ax1_fused_0) {
+    uint32_t zeros_loaded = *(uint32_t*)(zeros_ptr + k_0_0 * 32 / G * (OC / 8));
+    uint4 B_loaded_zero = dequantize_s4_to_fp16x2(zeros_loaded);
+    uint4 B_loaded_scale = *(uint4*)(scaling_factors_ptr + k_0_0 * 32 / G * (OC));
+    /*
+    if (blockIdx_z == 0 && blockIdx_y == 0 && k_0_0 == 0 && threadIdx.x == 0 && threadIdx.y == 0){
+      printf("%x %x %x %x %x %x %x %x\n", B_loaded_scale.x, B_loaded_scale.y, B_loaded_scale.z, B_loaded_scale.w, B_loaded_zero.x, B_loaded_zero.y, B_loaded_zero.z, B_loaded_zero.w);
+    }
+    */
+    // uint4 B_loaded_scale = make_uint4(0, 0, 0, 0);
+    int* B_ptr_local = B_ptr + k_0_0 * 32 * (OC / 8);
+
+    for (int ax0_ax1_fused_0 = 0; ax0_ax1_fused_0 < 4; ++ax0_ax1_fused_0) {
+
+      // B: 32 x 136 (128+8) float16
+      // each warp: 32 x 4
+      // each thr: read 32 bit -> convert to 8xFP16 (a UINT4) -> scale and minus zero -> WB UINT4
+      // *(uint4*)(B_shared + ((((ax0_ax1_fused_0 * 544) + (((int)threadIdx.y) * 272)) + ((((int)threadIdx.x) >> 4) * 136)) + ((((int)threadIdx.x) & 15) * 8))) = *(uint4*)(B + ((((((k_0_0 * 163840) + (ax0_ax1_fused_0 * 20480)) + (((int)threadIdx.y) * 10240)) + ((((int)threadIdx.x) >> 4) * 5120)) + (((int)blockIdx_y) * 128)) + ((((int)threadIdx.x) & 15) * 8)));
+      // row stride in shared memory: (NWARPS * 32 * 8 / cta_N) 
+      uint32_t B_loaded = *(uint32_t*)(B_ptr_local + ax0_ax1_fused_0 * row_stride * (OC / 8));
+      uint4 B_loaded_fp16 = dequantize_s4_to_fp16x2(B_loaded);
+      //uint4 B_loaded_zero = *(uint4*)(zeros_shared + (threadIdx.x % (cta_N / 8)) * 8);
+
+      // uint4 B_loaded_scale = *(uint4*)(scaling_factors_shared + (threadIdx.x % (cta_N / 8)) * 8);
+      // - zero and * scale
+      // TODO (Haotian): can save 4 assembly instructions if sormulate as deq = q * scale - zero * scale.
+      asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(B_loaded_fp16.x) : "r"(B_loaded_fp16.x), "r"(B_loaded_zero.x));
+      asm volatile("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(B_loaded_fp16.x) : "r"(B_loaded_fp16.x), "r"(B_loaded_scale.x), "r"(ZERO));
+      asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(B_loaded_fp16.y) : "r"(B_loaded_fp16.y), "r"(B_loaded_zero.y));
+      asm volatile("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(B_loaded_fp16.y) : "r"(B_loaded_fp16.y), "r"(B_loaded_scale.y), "r"(ZERO));
+      asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(B_loaded_fp16.z) : "r"(B_loaded_fp16.z), "r"(B_loaded_zero.z));
+      asm volatile("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(B_loaded_fp16.z) : "r"(B_loaded_fp16.z), "r"(B_loaded_scale.z), "r"(ZERO));
+      asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(B_loaded_fp16.w) : "r"(B_loaded_fp16.w), "r"(B_loaded_zero.w));
+      asm volatile("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(B_loaded_fp16.w) : "r"(B_loaded_fp16.w), "r"(B_loaded_scale.w), "r"(ZERO));
+      /*
+      if (ax0_ax1_fused_0 == 0 && blockIdx_z == 0 && blockIdx_y == 0 && k_0_0 == 0 && threadIdx.x == 17 && threadIdx.y == 0){
+        printf("[x] %X %X %X %X\n", B_loaded_fp16.x, B_loaded_fp16.y, B_loaded_fp16.z, B_loaded_fp16.w);
+      }
+      */
+
+      // write back
+      *(uint4*)(B_shared_ptr + ax0_ax1_fused_0 * row_stride * (64 + 8)) = B_loaded_fp16;
+    }
+    __syncthreads();
+
+    for (int k_0_1 = 0; k_0_1 < 2; ++k_0_1) 
+    {
+      {
+        unsigned int addr;
+        __asm__ __volatile__(
+          "{ .reg .u64 addr; cvta.to.shared.u64 addr, %1; cvt.u32.u64 %0, addr; }\n"
+          : "=r"(addr)
+          : "l"((void *)((&(A_shared[(k_0_1 * 16)])) + (((((int)threadIdx.x) & 15) * 40) + ((((int)threadIdx.x) >> 4) * 8))))
+        );
+        __asm__ __volatile__(
+          "ldmatrix.sync.aligned.m8n8.x4.shared.b16"
+          "{%0, %1, %2, %3}, [%4];\n"
+          : "=r"(((unsigned *)(A_shared_warp + 0))[0]), "=r"(((unsigned *)(A_shared_warp + 0))[1]), "=r"(((unsigned *)(A_shared_warp + 0))[2]), "=r"(((unsigned *)(A_shared_warp + 0))[3])
+          : "r"(addr)
+        );
+      }
+        
+
+      for (int ax1_0 = 0; ax1_0 < 2; ++ax1_0) 
+      {
+        {
+          unsigned int addr;
+          __asm__ __volatile__(
+            "{ .reg .u64 addr; cvta.to.shared.u64 addr, %1; cvt.u32.u64 %0, addr; }\n"
+            : "=r"(addr)
+            : "l"((void *)((&(B_shared[(((k_0_1 * 1152) + (((int)threadIdx.y) * 32)) + (ax1_0 * 16))])) + (((((int)threadIdx.x) & 15) * 72) + ((((int)threadIdx.x) >> 4) * 8))))
+          );
+          __asm__ __volatile__(
+            "ldmatrix.sync.aligned.m8n8.x4.trans.shared.b16"
+            "{%0, %1, %2, %3}, [%4];\n"
+            : "=r"(((unsigned *)(B_shared_warp + (ax1_0 * 8)))[0]), "=r"(((unsigned *)(B_shared_warp + (ax1_0 * 8)))[1]), "=r"(((unsigned *)(B_shared_warp + (ax1_0 * 8)))[2]), "=r"(((unsigned *)(B_shared_warp + (ax1_0 * 8)))[3])
+            : "r"(addr)
+          );
+        }
+      }
+      
+      for (int j_0_4 = 0; j_0_4 < 2; ++j_0_4) 
+      {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ == 750
+        {
+          __asm__ __volatile__(
+            "mma.sync.aligned.m16n8k8.row.col.f32.f16.f16.f32"
+            "{%0, %1, %2, %3}, {%4, %5}, {%6}, {%7, %8, %9, %10};\n"
+            :  "=f"(((float *)(C_warp + (j_0_4 * 8)))[0]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[1]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[2]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[3])
+            : "r"(((unsigned *)(A_shared_warp + 0))[0]), "r"(((unsigned *)(A_shared_warp + 0))[1]), "r"(((unsigned *)(B_shared_warp + (j_0_4 * 8)))[0]), "f"(((float *)(C_warp + (j_0_4 * 8)))[0]), "f"(((float *)(C_warp + (j_0_4 * 8)))[1]), "f"(((float *)(C_warp + (j_0_4 * 8)))[2]), "f"(((float *)(C_warp + (j_0_4 * 8)))[3]));
+        }
+
+        {
+          __asm__ __volatile__(
+            "mma.sync.aligned.m16n8k8.row.col.f32.f16.f16.f32"
+            "{%0, %1, %2, %3}, {%4, %5}, {%6}, {%7, %8, %9, %10};\n"
+            :  "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[0]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[1]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[2]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[3])
+            : "r"(((unsigned *)(A_shared_warp + 0))[0]), "r"(((unsigned *)(A_shared_warp + 0))[1]), "r"(((unsigned *)(B_shared_warp + ((j_0_4 * 8) + 4)))[0]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[0]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[1]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[2]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[3]));
+        }
+
+        {
+          __asm__ __volatile__(
+            "mma.sync.aligned.m16n8k8.row.col.f32.f16.f16.f32"
+            "{%0, %1, %2, %3}, {%4, %5}, {%6}, {%7, %8, %9, %10};\n"
+            :  "=f"(((float *)(C_warp + (j_0_4 * 8)))[0]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[1]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[2]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[3])
+            : "r"(((unsigned *)(A_shared_warp + 0))[2]), "r"(((unsigned *)(A_shared_warp + 0))[3]), "r"(((unsigned *)(B_shared_warp + (j_0_4 * 8)))[1]), "f"(((float *)(C_warp + (j_0_4 * 8)))[0]), "f"(((float *)(C_warp + (j_0_4 * 8)))[1]), "f"(((float *)(C_warp + (j_0_4 * 8)))[2]), "f"(((float *)(C_warp + (j_0_4 * 8)))[3]));
+        }
+
+        {
+          __asm__ __volatile__(
+            "mma.sync.aligned.m16n8k8.row.col.f32.f16.f16.f32"
+            "{%0, %1, %2, %3}, {%4, %5}, {%6}, {%7, %8, %9, %10};\n"
+            :  "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[0]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[1]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[2]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[3])
+            : "r"(((unsigned *)(A_shared_warp + 0))[2]), "r"(((unsigned *)(A_shared_warp + 0))[3]), "r"(((unsigned *)(B_shared_warp + ((j_0_4 * 8) + 4)))[1]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[0]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[1]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[2]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[3]));
+        }
+#else
+        {
+          __asm__ __volatile__(
+            "mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32"
+            "{%0, %1, %2, %3}, {%4, %5, %6, %7}, {%8, %9}, {%10, %11, %12, %13};\n"
+            :  "=f"(((float *)(C_warp + (j_0_4 * 8)))[0]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[1]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[2]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[3])
+            : "r"(((unsigned *)(A_shared_warp + 0))[0]), "r"(((unsigned *)(A_shared_warp + 0))[1]), "r"(((unsigned *)(A_shared_warp + 0))[2]), "r"(((unsigned *)(A_shared_warp + 0))[3]), "r"(((unsigned *)(B_shared_warp + (j_0_4 * 8)))[0]), "r"(((unsigned *)(B_shared_warp + (j_0_4 * 8)))[1]), "f"(((float *)(C_warp + (j_0_4 * 8)))[0]), "f"(((float *)(C_warp + (j_0_4 * 8)))[1]), "f"(((float *)(C_warp + (j_0_4 * 8)))[2]), "f"(((float *)(C_warp + (j_0_4 * 8)))[3]));
+        }
+
+        {
+          __asm__ __volatile__(
+            "mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32"
+            "{%0, %1, %2, %3}, {%4, %5, %6, %7}, {%8, %9}, {%10, %11, %12, %13};\n"
+            :  "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[0]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[1]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[2]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[3])
+            : "r"(((unsigned *)(A_shared_warp + 0))[0]), "r"(((unsigned *)(A_shared_warp + 0))[1]), "r"(((unsigned *)(A_shared_warp + 0))[2]), "r"(((unsigned *)(A_shared_warp + 0))[3]), "r"(((unsigned *)(B_shared_warp + ((j_0_4 * 8) + 4)))[0]), "r"(((unsigned *)(B_shared_warp + ((j_0_4 * 8) + 4)))[1]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[0]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[1]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[2]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[3]));
+        }
+#endif
+      }
+    }
+  }
+
+// TODO: Shang: Hoist loop invariance.
+  for (int ax1_0_1 = 0; ax1_0_1 < 2; ++ax1_0_1) {
+    for (int local_id = 0; local_id < 8; ++local_id) {
+      int row_offset = (((int)blockIdx_y) / j_factors1) * 16 + ((int)threadIdx.x) / 4 + (local_id % 4) / 2 * 8;
+      if (row_offset < M)
+      {
+        *(C_ptr + ax1_0_1 * 16 + row_offset * OC + (local_id / 4) * 8 + local_id % 2) = __float2half(C_warp[(ax1_0_1 * 8) + local_id]);
+      }
+    }
+  }
+#endif
+}
+
+} // namespace awq
+} // namespace vllm
+
+// in_feats: M, IC [float16]
+// kernel: IC, OC // 8 [int32] -> cast to IC, OC [uint4b]
+// scaling_factors: IC // G, OC [float16]
+// zeros: IC // G, OC // 8 [int32] -> cast to IC // G, OC [uint4b]
+// assume that batch_size < 16 for now
+
+torch::Tensor awq_gemm(
+    torch::Tensor _in_feats,
+    torch::Tensor _kernel,
+    torch::Tensor _scaling_factors,
+    torch::Tensor _zeros,
+    int split_k_iters)
+{
+    int num_in_feats = _in_feats.size(0);
+    int num_in_channels = _in_feats.size(1);
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(_in_feats));
+
+    auto options = torch::TensorOptions().dtype(_in_feats.dtype()).device(_in_feats.device());
+    at::Tensor _out_feats = torch::empty({split_k_iters, num_in_feats, _kernel.size(1) * 8}, options);
+    int num_out_feats = _out_feats.size(-2);
+    int num_out_channels = _out_feats.size(-1);
+
+    auto in_feats = reinterpret_cast<half*>(_in_feats.data_ptr<at::Half>());
+    auto kernel = reinterpret_cast<int*>(_kernel.data_ptr<int>());
+    auto out_feats = reinterpret_cast<half*>(_out_feats.data_ptr<at::Half>());
+    auto scaling_factors = reinterpret_cast<half*>(_scaling_factors.data_ptr<at::Half>());
+    auto zeros = reinterpret_cast<int*>(_zeros.data_ptr<int>());
+    int group_size = num_in_channels / _scaling_factors.size(0);
+
+    if (num_out_channels % 64 != 0)
+        throw std::invalid_argument("OC is not multiple of cta_N = 64");
+    if (num_out_channels % 8 != 0)
+        throw std::invalid_argument("OC is not multiple of pack_num = 8");
+    if (group_size % 32 != 0)
+	      throw std::invalid_argument("Group size should be a multiple of 32");
+    if (num_out_channels % group_size != 0)
+        throw std::invalid_argument("OC is not multiple of Group size");
+
+    const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+    if (num_out_channels % 128 == 0)
+    {
+        int j_factors1 = num_out_channels / 128 / 1;
+        dim3 num_blocks((num_out_feats + 16 - 1) / 16 * j_factors1 * split_k_iters);
+        // threadIdx.x: 32
+        // threadIdx.y: i_factors[2] * j_factors[2]
+        dim3 threads_per_block(32, 2);
+        vllm::awq::gemm_forward_4bit_cuda_m16n128k32<<<num_blocks, threads_per_block, 0, stream>>>(
+            group_size, split_k_iters, in_feats, kernel, scaling_factors, zeros, num_in_feats, num_in_channels, num_out_channels, out_feats);
+    }
+    else if (num_out_channels % 64 == 0)
+    {
+        int j_factors1 = num_out_channels / 64 / 1;
+        dim3 num_blocks(1 * (num_out_feats + 16 - 1) / 16 * j_factors1 * split_k_iters);
+    
+        // threadIdx.x: 32
+        // threadIdx.y: i_factors[2] * j_factors[2]
+        dim3 threads_per_block(32, 2);
+        vllm::awq::gemm_forward_4bit_cuda_m16n64k32<<<num_blocks, threads_per_block, 0, stream>>>(
+            group_size, split_k_iters, in_feats, kernel, scaling_factors, zeros, num_in_feats, num_in_channels, num_out_channels, out_feats);
+    }
+    return _out_feats.sum(0);
+}

csrc/quantization/squeezellm/quant_cuda_kernel.cu

@@ -0,0 +1,148 @@
+#include <torch/all.h>
+#include <torch/python.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+#include <cuda_fp16.h>
+
+// half-tensor
+#include <c10/cuda/CUDAStream.h>
+#include <ATen/cuda/CUDATensorMethods.cuh>
+
+#define BLOCKWIDTH 128
+#define BLOCKHEIGHT4 16
+
+namespace vllm {
+namespace squeezellm {
+
+__device__ inline unsigned int as_unsigned(int i) {
+  return *reinterpret_cast<unsigned int*>(&i);
+}
+
+// 4-bit matvec kernel (LUT-based)
+__global__ void NUQ4MatMulKernel(
+    const  half2* __restrict__ vec,
+    const    int* __restrict__ mat,
+           half2* __restrict__ mul,
+    const  __half* __restrict__ lookup_table,
+    int height,
+    int width,
+    int batch,
+    int vec_height
+) {
+
+  const int blockwidth2 = BLOCKWIDTH / 2;
+
+  int row = BLOCKHEIGHT4 * blockIdx.x;
+  int col =  BLOCKWIDTH * blockIdx.y + threadIdx.x;
+
+  __shared__ half2 blockvec[blockwidth2];
+
+  __shared__ __half deq2[16][BLOCKWIDTH];
+  int off = threadIdx.x;
+  int column_offset = col * 16;
+  for (int val = 0; val < 16; val += 1) {
+    int lut_index = column_offset + val;
+    deq2[val][off] = lookup_table[lut_index];
+  }
+
+  __half res;
+  half2 res2;
+  half2 tmp2;
+
+  int i;
+  int k;
+
+  unsigned int tmp1;
+  unsigned int lut_index1, lut_index2;
+
+  for (int b = 0; b < batch; ++b){
+    i = width * row + col;
+    res = __int2half_rd(0);
+    k = 0;
+
+    __syncthreads();
+    if (threadIdx.x < blockwidth2)
+      blockvec[threadIdx.x] = vec[b * vec_height / 2 + (row / BLOCKHEIGHT4) * blockwidth2 + threadIdx.x];
+    __syncthreads();
+
+    while (k < blockwidth2) {
+      tmp1 = as_unsigned(mat[i]);
+
+      res2 = {};
+      tmp2 = {};
+
+      lut_index1 = tmp1 & 0xF;
+      lut_index2 = (tmp1 >> 4) & 0xF;
+      tmp2.x = deq2[lut_index1][off];
+      tmp2.y = deq2[lut_index2][off];
+      res2 = __hfma2(tmp2, blockvec[k + 0], res2);
+
+      lut_index1 = (tmp1 >> 8) & 0xF;
+      lut_index2 = (tmp1 >> 12) & 0xF;
+      tmp2.x = deq2[lut_index1][off];
+      tmp2.y = deq2[lut_index2][off];
+      res2 = __hfma2(tmp2, blockvec[k + 1], res2);
+
+      lut_index1 = (tmp1 >> 16) & 0xF;
+      lut_index2 = (tmp1 >> 20) & 0xF;
+      tmp2.x = deq2[lut_index1][off];
+      tmp2.y = deq2[lut_index2][off];
+      res2 = __hfma2(tmp2, blockvec[k + 2], res2);
+
+      lut_index1 = (tmp1 >> 24) & 0xF;
+      lut_index2 = (tmp1 >> 28) & 0xF;
+      tmp2.x = deq2[lut_index1][off];
+      tmp2.y = deq2[lut_index2][off];
+      res2 = __hfma2(tmp2, blockvec[k + 3], res2);
+
+      res = __hadd(__hadd(res2.x, res2.y), res);
+
+      i += width;
+      k += 4;
+    }
+
+    // col%2 -> only set one of the two values
+    half2 res3 = {};
+    if (col % 2 == 0) {
+      res3.x = res;
+    } else {
+      res3.y = res;
+    }
+
+    atomicAdd(&mul[b * width / 2 + col / 2], res3);
+  }
+}
+
+} // namespace squeezellm
+} // namespace vllm
+
+// 4-bit matvec kernel (LUT-based)
+void squeezellm_gemm(
+  torch::Tensor vec,
+  torch::Tensor mat,
+  torch::Tensor mul,
+  torch::Tensor lookup_table
+) {
+  int height = mat.size(0);
+  int width = mat.size(1);
+
+  int batch = vec.size(0);
+  int vec_height = vec.size(1);
+
+  dim3 blocks(
+    (height + BLOCKHEIGHT4 - 1) / BLOCKHEIGHT4,
+    (width + BLOCKWIDTH - 1) / BLOCKWIDTH
+  );
+  dim3 threads(BLOCKWIDTH);
+
+  vllm::squeezellm::NUQ4MatMulKernel<<<blocks, threads>>>(
+    (half2*) vec.data<at::Half>(),
+    mat.data_ptr<int>(),
+    (half2*) mul.data<at::Half>(),
+    (__half*) lookup_table.data<at::Half>(),
+    height, width, batch, vec_height
+  );
+}
+
+#undef BLOCKWIDTH
+#undef BLOCKHEIGHT4

docs/source/getting_started/installation.rst

@@ -3,30 +3,14 @@
 Installation
 ============
 
-vLLM is a Python library that also contains some C++ and CUDA code.
-This additional code requires compilation on the user's machine.
+vLLM is a Python library that also contains pre-compiled C++ and CUDA (12.1) binaries.
 
 Requirements
 ------------
 
 * OS: Linux
-* Python: 3.8 or higher
-* CUDA: 11.0 -- 11.8
-* GPU: compute capability 7.0 or higher (e.g., V100, T4, RTX20xx, A100, L4, etc.)
-
-.. note::
-    As of now, vLLM does not support CUDA 12.
-    If you are using Hopper or Lovelace GPUs, please use CUDA 11.8 instead of CUDA 12.
-
-.. tip::
-    If you have trouble installing vLLM, we recommend using the NVIDIA PyTorch Docker image.
-
-    .. code-block:: console
-
-        $ # Pull the Docker image with CUDA 11.8.
-        $ docker run --gpus all -it --rm --shm-size=8g nvcr.io/nvidia/pytorch:22.12-py3
-
-    Inside the Docker container, please execute :code:`pip uninstall torch` before installing vLLM.
+* Python: 3.8 -- 3.11
+* GPU: compute capability 7.0 or higher (e.g., V100, T4, RTX20xx, A100, L4, H100, etc.)
 
 Install with pip
 ----------------
@@ -39,8 +23,23 @@ You can install vLLM using pip:
     $ conda create -n myenv python=3.8 -y
     $ conda activate myenv
 
-    $ # Install vLLM.
-    $ pip install vllm  # This may take 5-10 minutes.
+    $ # Install vLLM with CUDA 12.1.
+    $ pip install vllm
+
+.. note::
+
+    As of now, vLLM's binaries are compiled on CUDA 12.1 by default.
+    However, you can install vLLM with CUDA 11.8 by running:
+
+    .. code-block:: console
+
+        $ # Install vLLM with CUDA 11.8.
+        $ # Replace `cp310` with your Python version (e.g., `cp38`, `cp39`, `cp311`).
+        $ pip install https://github.com/vllm-project/vllm/releases/download/v0.2.2/vllm-0.2.2+cu118-cp310-cp310-manylinux1_x86_64.whl
+
+        $ # Re-install PyTorch with CUDA 11.8.
+        $ pip uninstall torch -y
+        $ pip install torch --upgrade --index-url https://download.pytorch.org/whl/cu118
 
 
 .. _build_from_source:
@@ -55,3 +54,11 @@ You can also build and install vLLM from source:
     $ git clone https://github.com/vllm-project/vllm.git
     $ cd vllm
     $ pip install -e .  # This may take 5-10 minutes.
+
+.. tip::
+    If you have trouble building vLLM, we recommend using the NVIDIA PyTorch Docker image.
+
+    .. code-block:: console
+
+        $ # Use `--ipc=host` to make sure the shared memory is large enough.
+        $ docker run --gpus all -it --rm --ipc=host nvcr.io/nvidia/pytorch:23.10-py3

docs/source/getting_started/quickstart.rst

@@ -40,6 +40,16 @@ Initialize vLLM's engine for offline inference with the ``LLM`` class and the `O
 
     llm = LLM(model="facebook/opt-125m")
 
+Use model from www.modelscope.cn
+
+.. code-block:: shell
+
+    export VLLM_USE_MODELSCOPE=True
+
+.. code-block:: python
+
+    llm = LLM(model="qwen/Qwen-7B-Chat", revision="v1.1.8", trust_remote_code=True)
+
 Call ``llm.generate`` to generate the outputs. It adds the input prompts to vLLM engine's waiting queue and executes the vLLM engine to generate the outputs with high throughput. The outputs are returned as a list of ``RequestOutput`` objects, which include all the output tokens.
 
 .. code-block:: python
@@ -67,6 +77,16 @@ Start the server:
 
     $ python -m vllm.entrypoints.api_server
 
+Use model from www.modelscope.cn
+
+.. code-block:: console
+
+    $ VLLM_USE_MODELSCOPE=True python -m vllm.entrypoints.api_server \
+    $    --model="qwen/Qwen-7B-Chat" \
+    $    --revision="v1.1.8" \
+    $    --trust-remote-code
+
+
 By default, this command starts the server at ``http://localhost:8000`` with the OPT-125M model.
 
 Query the model in shell:
@@ -87,6 +107,7 @@ OpenAI-Compatible Server
 ------------------------
 
 vLLM can be deployed as a server that mimics the OpenAI API protocol. This allows vLLM to be used as a drop-in replacement for applications using OpenAI API.
+By default, it starts the server at ``http://localhost:8000``. You can specify the address with ``--host`` and ``--port`` arguments. The server currently hosts one model at a time (OPT-125M in the above command) and implements `list models <https://platform.openai.com/docs/api-reference/models/list>`_, `create chat completion <https://platform.openai.com/docs/api-reference/chat/completions/create>`_, and `create completion <https://platform.openai.com/docs/api-reference/completions/create>`_ endpoints. We are actively adding support for more endpoints.
 
 Start the server:
 
@@ -95,7 +116,20 @@ Start the server:
     $ python -m vllm.entrypoints.openai.api_server \
     $     --model facebook/opt-125m
 
-By default, it starts the server at ``http://localhost:8000``. You can specify the address with ``--host`` and ``--port`` arguments. The server currently hosts one model at a time (OPT-125M in the above command) and implements `list models <https://platform.openai.com/docs/api-reference/models/list>`_ and `create completion <https://platform.openai.com/docs/api-reference/completions/create>`_ endpoints. We are actively adding support for more endpoints.
+Use model from www.modelscope.cn
+
+.. code-block:: console
+
+    $ VLLM_USE_MODELSCOPE=True python -m vllm.entrypoints.openai.api_server \
+    $     --model="qwen/Qwen-7B-Chat" --revision="v1.1.8" --trust-remote-code
+
+By default, the server uses a predefined chat template stored in the tokenizer. You can override this template by using the ``--chat-template`` argument:
+
+.. code-block:: console
+
+   $ python -m vllm.entrypoints.openai.api_server \
+   $     --model facebook/opt-125m \
+   $     --chat-template ./examples/template_chatml.json
 
 This server can be queried in the same format as OpenAI API. For example, list the models:
 
@@ -103,6 +137,9 @@ This server can be queried in the same format as OpenAI API. For example, list t
 
     $ curl http://localhost:8000/v1/models
 
+Using OpenAI Completions API with vLLM
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
 Query the model with input prompts:
 
 .. code-block:: console
@@ -120,12 +157,65 @@ Since this server is compatible with OpenAI API, you can use it as a drop-in rep
 
 .. code-block:: python
 
-    import openai
+    from openai import OpenAI
+
     # Modify OpenAI's API key and API base to use vLLM's API server.
-    openai.api_key = "EMPTY"
-    openai.api_base = "http://localhost:8000/v1"
-    completion = openai.Completion.create(model="facebook/opt-125m",
+    openai_api_key = "EMPTY"
+    openai_api_base = "http://localhost:8000/v1"
+    client = OpenAI(
+        api_key=openai_api_key,
+        base_url=openai_api_base,
+    )
+    completion = client.completions.create(model="facebook/opt-125m",
                                           prompt="San Francisco is a")
     print("Completion result:", completion)
 
-For a more detailed client example, refer to `examples/openai_client.py <https://github.com/vllm-project/vllm/blob/main/examples/openai_client.py>`_.
+For a more detailed client example, refer to `examples/openai_completion_client.py <https://github.com/vllm-project/vllm/blob/main/examples/openai_completion_client.py>`_.
+
+Using OpenAI Chat API with vLLM
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+The vLLM server is designed to support the OpenAI Chat API, allowing you to engage in dynamic conversations with the model. The chat interface is a more interactive way to communicate with the model, allowing back-and-forth exchanges that can be stored in the chat history. This is useful for tasks that require context or more detailed explanations.
+
+Querying the model using OpenAI Chat API:
+
+You can use the `create chat completion <https://platform.openai.com/docs/api-reference/chat/completions/create>`_ endpoint to communicate with the model in a chat-like interface:
+
+.. code-block:: console
+
+    $ curl http://localhost:8000/v1/chat/completions \
+    $     -H "Content-Type: application/json" \
+    $     -d '{
+    $         "model": "facebook/opt-125m",
+    $         "messages": [
+    $             {"role": "system", "content": "You are a helpful assistant."},
+    $             {"role": "user", "content": "Who won the world series in 2020?"}
+    $         ]
+    $     }'
+
+Python Client Example:
+
+Using the `openai` python package, you can also communicate with the model in a chat-like manner:
+
+.. code-block:: python
+
+    from openai import OpenAI
+    # Set OpenAI's API key and API base to use vLLM's API server.
+    openai_api_key = "EMPTY"
+    openai_api_base = "http://localhost:8000/v1"
+
+    client = OpenAI(
+        api_key=openai_api_key,
+        base_url=openai_api_base,
+    )
+
+    chat_response = client.chat.completions.create(
+        model="facebook/opt-125m",
+        messages=[
+            {"role": "system", "content": "You are a helpful assistant."},
+            {"role": "user", "content": "Tell me a joke."},
+        ]
+    )
+    print("Chat response:", chat_response)
+
+For more in-depth examples and advanced features of the chat API, you can refer to the official OpenAI documentation.

docs/source/index.rst

@@ -43,6 +43,7 @@ vLLM is flexible and easy to use with:
 For more information, check out the following:
 
 * `vLLM announcing blog post <https://vllm.ai>`_ (intro to PagedAttention)
+* `vLLM paper <https://arxiv.org/abs/2309.06180>`_ (SOSP 2023)
 * `How continuous batching enables 23x throughput in LLM inference while reducing p50 latency <https://www.anyscale.com/blog/continuous-batching-llm-inference>`_ by Cade Daniel et al.
 
 
@@ -62,6 +63,11 @@ Documentation
    :caption: Serving
 
    serving/distributed_serving
+   serving/run_on_sky
+   serving/deploying_with_triton
+   serving/deploying_with_docker
+   serving/serving_with_langchain
+   serving/metrics
 
 .. toctree::
    :maxdepth: 1
@@ -69,3 +75,10 @@ Documentation
 
    models/supported_models
    models/adding_model
+   models/engine_args
+
+.. toctree::
+   :maxdepth: 1
+   :caption: Quantization
+
+   quantization/auto_awq

docs/source/models/adding_model.rst

@@ -18,7 +18,7 @@ This document provides a high-level guide on integrating a `HuggingFace Transfor
 0. Fork the vLLM repository
 --------------------------------
 
-Start by forking our `GitHub <https://github.com/vllm-project/vllm/>`_ repository and then :ref:`build it from source <build_from_source>`.
+Start by forking our `GitHub`_ repository and then :ref:`build it from source <build_from_source>`.
 This gives you the ability to modify the codebase and test your model.
 
 
@@ -59,34 +59,37 @@ Next, you need to rewrite the :code:`forward` methods of your model by following
     +    kv_caches: List[KVCache],
     +    input_metadata: InputMetadata,
     +    cache_events: Optional[List[torch.cuda.Event]],
-    +) -> Dict[int, SequenceOutputs]:
+    +) -> SamplerOutput:
 
 3. Update the code by considering that :code:`input_ids` and :code:`positions` are now flattened tensors.
-4. Replace the attention operation with either :code:`GPTPagedAttention` or :code:`GPTNeoXPagedAttention`, depending on the model's architecture.
+4. Replace the attention operation with either :code:`PagedAttention`, :code:`PagedAttentionWithRoPE`, or :code:`PagedAttentionWithALiBi` depending on the model's architecture.
 
 .. note::
     Currently, vLLM supports the basic multi-head attention mechanism and its variant with rotary positional embeddings.
     If your model employs a different attention mechanism, you will need to implement a new attention layer in vLLM.
 
 
-3. (Optional) Implement tensor parallelism support
---------------------------------------------------
+3. (Optional) Implement tensor parallelism and quantization support
+-------------------------------------------------------------------
 
 If your model is too large to fit into a single GPU, you can use tensor parallelism to manage it.
 To do this, substitute your model's linear and embedding layers with their tensor-parallel versions.
-For the embedding layer, you can simply replace :code:`nn.Embedding` with :code:`VocabParallelEmbedding`.
-When it comes to the linear layers, you should use either :code:`RowParallelLinear` or :code:`ColumnParallelLinear`.
-Typically, :code:`ColumnParallelLinear` is used for QKV linear layers and the first linear layers of the MLP blocks.
-For the remaining linear layers, :code:`RowParallelLinear` is used.
+For the embedding layer, you can simply replace :code:`nn.Embedding` with :code:`VocabParallelEmbedding`. For the output LM head, you can use :code:`ParallelLMHead`.
+When it comes to the linear layers, we provide the following options to parallelize them:
 
+* :code:`ReplicatedLinear`: Replicates the inputs and weights across multiple GPUs. No memory saving.
+* :code:`RowParallelLinear`: The input tensor is partitioned along the hidden dimension. The weight matrix is partitioned along the rows (input dimension). An *all-reduce* operation is performed after the matrix multiplication to reduce the results. Typically used for the second FFN layer and the output linear transformation of the attention layer.
+* :code:`ColumnParallelLinear`: The input tensor is replicated. The weight matrix is partitioned along the columns (output dimension). The result is partitioned along the column dimension. Typically used for the first FFN layer and the separated QKV transformation of the attention layer in the original Transformer.
+* :code:`MergedColumnParallelLinear`: Column-parallel linear that merges multiple `ColumnParallelLinear` operators. Typically used for the first FFN layer with weighted activation functions (e.g., SiLU). This class handles the sharded weight loading logic of multiple weight matrices.
+* :code:`QKVParallelLinear`: Parallel linear layer for the query, key, and value projections of the multi-head and grouped-query attention mechanisms. When number of key/value heads are less than the world size, this class replicates the key/value heads properly. This class handles the weight loading and replication of the weight matrices.
+
+Note that all the linear layers above take `linear_method` as an input. vLLM will set this parameter according to different quantization schemes to support weight quantization.
 
 4. Implement the weight loading logic
 -------------------------------------
 
 You now need to implement the :code:`load_weights` method in your :code:`*ForCausalLM` class.
-This method should load the weights from the HuggingFace's checkpoint file and assign them to the corresponding layers in your model.
-While the process is straightforward for most layers, the tensor-parallel layers necessitate some additional care as their weights should be partitioned to multiple GPUs.
-
+This method should load the weights from the HuggingFace's checkpoint file and assign them to the corresponding layers in your model. Specifically, for `MergedColumnParallelLinear` and `QKVParallelLinear` layers, if the original model has separated weight matrices, you need to load the different parts separately.
 
 5. Register your model
 ----------------------

docs/source/models/engine_args.rst

@@ -0,0 +1,114 @@
+.. _engine_args:
+
+Engine Arguments
+================
+
+Below, you can find an explanation of every engine argument for vLLM:
+
+.. option:: --model <model_name_or_path>
+
+    Name or path of the huggingface model to use.
+
+.. option:: --tokenizer <tokenizer_name_or_path>
+
+    Name or path of the huggingface tokenizer to use.
+
+.. option:: --revision <revision>
+
+    The specific model version to use. It can be a branch name, a tag name, or a commit id. If unspecified, will use the default version.
+
+.. option:: --tokenizer-revision <revision>
+
+    The specific tokenizer version to use. It can be a branch name, a tag name, or a commit id. If unspecified, will use the default version.
+
+.. option:: --tokenizer-mode {auto,slow}
+
+    The tokenizer mode.
+    
+    * "auto" will use the fast tokenizer if available.
+    * "slow" will always use the slow tokenizer.
+
+.. option:: --trust-remote-code
+
+    Trust remote code from huggingface.
+
+.. option:: --download-dir <directory>
+
+    Directory to download and load the weights, default to the default cache dir of huggingface.
+
+.. option:: --load-format {auto,pt,safetensors,npcache,dummy}
+
+    The format of the model weights to load.
+
+    * "auto" will try to load the weights in the safetensors format and fall back to the pytorch bin format if safetensors format is not available.
+    * "pt" will load the weights in the pytorch bin format.
+    * "safetensors" will load the weights in the safetensors format.
+    * "npcache" will load the weights in pytorch format and store a numpy cache to speed up the loading.
+    * "dummy" will initialize the weights with random values, mainly for profiling.
+
+.. option:: --dtype {auto,half,float16,bfloat16,float,float32}
+
+    Data type for model weights and activations.
+
+    * "auto" will use FP16 precision for FP32 and FP16 models, and BF16 precision for BF16 models.
+    * "half" for FP16. Recommended for AWQ quantization.
+    * "float16" is the same as "half".
+    * "bfloat16" for a balance between precision and range.
+    * "float" is shorthand for FP32 precision.
+    * "float32" for FP32 precision.
+
+.. option:: --max-model-len <length>
+
+    Model context length. If unspecified, will be automatically derived from the model config.
+
+.. option:: --worker-use-ray
+
+    Use Ray for distributed serving, will be automatically set when using more than 1 GPU.
+
+.. option:: --pipeline-parallel-size (-pp) <size>
+
+    Number of pipeline stages.
+
+.. option:: --tensor-parallel-size (-tp) <size>
+
+    Number of tensor parallel replicas.
+
+.. option:: --max-parallel-loading-workers <workers>
+
+    Load model sequentially in multiple batches, to avoid RAM OOM when using tensor parallel and large models.
+
+.. option:: --block-size {8,16,32}
+
+    Token block size for contiguous chunks of tokens.
+
+.. option:: --seed <seed>
+
+    Random seed for operations.
+
+.. option:: --swap-space <size>
+
+    CPU swap space size (GiB) per GPU.
+
+.. option:: --gpu-memory-utilization <percentage>
+
+    The percentage of GPU memory to be used for the model executor.
+
+.. option:: --max-num-batched-tokens <tokens>
+
+    Maximum number of batched tokens per iteration.
+
+.. option:: --max-num-seqs <sequences>
+
+    Maximum number of sequences per iteration.
+
+.. option:: --max-paddings <paddings>
+
+    Maximum number of paddings in a batch.
+
+.. option:: --disable-log-stats
+
+    Disable logging statistics.
+
+.. option:: --quantization (-q) {awq,squeezellm,None}
+
+    Method used to quantize the weights.

docs/source/models/supported_models.rst

@@ -14,27 +14,57 @@ Alongside each architecture, we include some popular models that use it.
   * - Architecture
     - Models
     - Example HuggingFace Models
+  * - :code:`AquilaForCausalLM`
+    - Aquila
+    - :code:`BAAI/Aquila-7B`, :code:`BAAI/AquilaChat-7B`, etc.
+  * - :code:`BaiChuanForCausalLM`
+    - Baichuan
+    - :code:`baichuan-inc/Baichuan2-13B-Chat`, :code:`baichuan-inc/Baichuan-7B`, etc.
+  * - :code:`ChatGLMModel`
+    - ChatGLM
+    - :code:`THUDM/chatglm2-6b`, :code:`THUDM/chatglm3-6b`, etc.
   * - :code:`BloomForCausalLM`
     - BLOOM, BLOOMZ, BLOOMChat
     - :code:`bigscience/bloom`, :code:`bigscience/bloomz`, etc.
+  * - :code:`FalconForCausalLM`
+    - Falcon
+    - :code:`tiiuae/falcon-7b`, :code:`tiiuae/falcon-40b`, :code:`tiiuae/falcon-rw-7b`, etc.
   * - :code:`GPT2LMHeadModel`
     - GPT-2
     - :code:`gpt2`, :code:`gpt2-xl`, etc.
   * - :code:`GPTBigCodeForCausalLM`
     - StarCoder, SantaCoder, WizardCoder
     - :code:`bigcode/starcoder`, :code:`bigcode/gpt_bigcode-santacoder`, :code:`WizardLM/WizardCoder-15B-V1.0`, etc.
+  * - :code:`GPTJForCausalLM`
+    - GPT-J
+    - :code:`EleutherAI/gpt-j-6b`, :code:`nomic-ai/gpt4all-j`, etc.
   * - :code:`GPTNeoXForCausalLM`
     - GPT-NeoX, Pythia, OpenAssistant, Dolly V2, StableLM
     - :code:`EleutherAI/gpt-neox-20b`, :code:`EleutherAI/pythia-12b`, :code:`OpenAssistant/oasst-sft-4-pythia-12b-epoch-3.5`, :code:`databricks/dolly-v2-12b`, :code:`stabilityai/stablelm-tuned-alpha-7b`, etc.
+  * - :code:`InternLMForCausalLM`
+    - InternLM
+    - :code:`internlm/internlm-7b`, :code:`internlm/internlm-chat-7b`, etc.
   * - :code:`LlamaForCausalLM`
-    - LLaMA, Vicuna, Alpaca, Koala, Guanaco
-    - :code:`openlm-research/open_llama_13b`, :code:`lmsys/vicuna-13b-v1.3`, :code:`young-geng/koala`, :code:`JosephusCheung/Guanaco`, etc.
+    - LLaMA, LLaMA-2, Vicuna, Alpaca, Koala, Guanaco
+    - :code:`meta-llama/Llama-2-13b-hf`, :code:`meta-llama/Llama-2-70b-hf`, :code:`openlm-research/open_llama_13b`, :code:`lmsys/vicuna-13b-v1.3`, :code:`young-geng/koala`, etc.
+  * - :code:`MistralForCausalLM`
+    - Mistral, Mistral-Instruct
+    - :code:`mistralai/Mistral-7B-v0.1`, :code:`mistralai/Mistral-7B-Instruct-v0.1`, etc.
   * - :code:`MPTForCausalLM`
     - MPT, MPT-Instruct, MPT-Chat, MPT-StoryWriter
     - :code:`mosaicml/mpt-7b`, :code:`mosaicml/mpt-7b-storywriter`, :code:`mosaicml/mpt-30b`, etc.
   * - :code:`OPTForCausalLM`
     - OPT, OPT-IML
     - :code:`facebook/opt-66b`, :code:`facebook/opt-iml-max-30b`, etc.
+  * - :code:`PhiForCausalLM`
+    - Phi-1.5
+    - :code:`microsoft/phi-1_5`, etc.
+  * - :code:`QWenLMHeadModel`
+    - Qwen
+    - :code:`Qwen/Qwen-7B`, :code:`Qwen/Qwen-7B-Chat`, etc.
+  * - :code:`YiForCausalLM`
+    - Yi
+    - :code:`01-ai/Yi-6B`, :code:`01-ai/Yi-34B`, etc.
 
 If your model uses one of the above model architectures, you can seamlessly run your model with vLLM.
 Otherwise, please refer to :ref:`Adding a New Model <adding_a_new_model>` for instructions on how to implement support for your model.
@@ -51,4 +81,18 @@ Alternatively, you can raise an issue on our `GitHub <https://github.com/vllm-pr
         output = llm.generate("Hello, my name is")
         print(output)
 
+    To use model from www.modelscope.cn
+
+    .. code-block:: shell
+
+       $ export VLLM_USE_MODELSCOPE=True
+
+    .. code-block:: python
+
+        from vllm import LLM
+
+        llm = LLM(model=..., revision=..., trust_remote_code=True)  # Name or path of your model
+        output = llm.generate("Hello, my name is")
+        print(output)
+
     If vLLM successfully generates text, it indicates that your model is supported.

docs/source/quantization/auto_awq.rst

@@ -0,0 +1,75 @@
+.. _auto_awq:
+
+AutoAWQ
+==================
+
+.. warning::
+
+   Please note that AWQ support in vLLM is under-optimized at the moment. We would recommend using the unquantized version of the model for better
+   accuracy and higher throughput. Currently, you can use AWQ as a way to reduce memory footprint. As of now, it is more suitable for low latency
+   inference with small number of concurrent requests. vLLM's AWQ implementation have lower throughput than unquantized version.
+
+To create a new 4-bit quantized model, you can leverage `AutoAWQ <https://github.com/casper-hansen/AutoAWQ>`_. 
+Quantizing reduces the model's precision from FP16 to INT4 which effectively reduces the file size by ~70%.
+The main benefits are lower latency and memory usage.
+
+You can quantize your own models by installing AutoAWQ or picking one of the `400+ models on Huggingface <https://huggingface.co/models?sort=trending&search=awq>`_. 
+
+.. code-block:: console
+
+    $ pip install autoawq
+
+After installing AutoAWQ, you are ready to quantize a model. Here is an example of how to quantize Vicuna 7B v1.5:
+
+.. code-block:: python
+
+    from awq import AutoAWQForCausalLM
+    from transformers import AutoTokenizer
+
+    model_path = 'lmsys/vicuna-7b-v1.5'
+    quant_path = 'vicuna-7b-v1.5-awq'
+    quant_config = { "zero_point": True, "q_group_size": 128, "w_bit": 4, "version": "GEMM" }
+
+    # Load model
+    model = AutoAWQForCausalLM.from_pretrained(model_path, **{"low_cpu_mem_usage": True})
+    tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+
+    # Quantize
+    model.quantize(tokenizer, quant_config=quant_config)
+
+    # Save quantized model
+    model.save_quantized(quant_path)
+    tokenizer.save_pretrained(quant_path)
+
+To run an AWQ model with vLLM, you can use `TheBloke/Llama-2-7b-Chat-AWQ <https://huggingface.co/TheBloke/Llama-2-7b-Chat-AWQ>`_ with the following command:
+
+.. code-block:: console
+
+    $ python examples/llm_engine_example.py --model TheBloke/Llama-2-7b-Chat-AWQ --quantization awq
+
+AWQ models are also supported directly through the LLM entrypoint:
+
+.. code-block:: python
+
+    from vllm import LLM, SamplingParams
+
+    # Sample prompts.
+    prompts = [
+        "Hello, my name is",
+        "The president of the United States is",
+        "The capital of France is",
+        "The future of AI is",
+    ]
+    # Create a sampling params object.
+    sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
+
+    # Create an LLM.
+    llm = LLM(model="TheBloke/Llama-2-7b-Chat-AWQ", quantization="AWQ")
+    # Generate texts from the prompts. The output is a list of RequestOutput objects
+    # that contain the prompt, generated text, and other information.
+    outputs = llm.generate(prompts, sampling_params)
+    # Print the outputs.
+    for output in outputs:
+        prompt = output.prompt
+        generated_text = output.outputs[0].text
+        print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")

docs/source/serving/deploying_with_docker.rst

@@ -0,0 +1,43 @@
+.. _deploying_with_docker:
+
+Deploying with Docker
+============================
+
+vLLM offers official docker image for deployment.
+The image can be used to run OpenAI compatible server.
+The image is available on Docker Hub as `vllm/vllm-openai <https://hub.docker.com/r/vllm/vllm-openai/tags>`_.
+
+.. code-block:: console
+
+    $ docker run --runtime nvidia --gpus all \
+        -v ~/.cache/huggingface:/root/.cache/huggingface \
+        --env "HUGGING_FACE_HUB_TOKEN=<secret>" \
+        -p 8000:8000 \
+        --ipc=host \
+        vllm/vllm-openai:latest \
+        --model mistralai/Mistral-7B-v0.1
+
+
+.. note::
+
+        You can either use the ``ipc=host`` flag or ``--shm-size`` flag to allow the
+        container to access the host's shared memory. vLLM uses PyTorch, which uses shared
+        memory to share data between processes under the hood, particularly for tensor parallel inference.
+
+
+You can build and run vLLM from source via the provided dockerfile. To build vLLM:
+
+.. code-block:: console
+
+    $ DOCKER_BUILDKIT=1 docker build . --target vllm-openai --tag vllm/vllm-openai --build-arg max_jobs=8
+
+To run vLLM:
+
+.. code-block:: console
+
+    $ docker run --runtime nvidia --gpus all \
+        -v ~/.cache/huggingface:/root/.cache/huggingface \
+        -p 8000:8000 \
+        --env "HUGGING_FACE_HUB_TOKEN=<secret>" \
+        vllm/vllm-openai <args...>
+

docs/source/serving/deploying_with_triton.rst

@@ -0,0 +1,6 @@
+.. _deploying_with_triton:
+
+Deploying with NVIDIA Triton
+============================
+
+The `Triton Inference Server <https://github.com/triton-inference-server>`_ hosts a tutorial demonstrating how to quickly deploy a simple `facebook/opt-125m <https://huggingface.co/facebook/opt-125m>`_ model using vLLM. Please see `Deploying a vLLM model in Triton <https://github.com/triton-inference-server/tutorials/blob/main/Quick_Deploy/vLLM/README.md#deploying-a-vllm-model-in-triton>`_ for more details.

docs/source/serving/metrics.rst

@@ -0,0 +1,13 @@
+Production Metrics
+==================
+
+vLLM exposes a number of metrics that can be used to monitor the health of the
+system. These metrics are exposed via the `/metrics` endpoint on the vLLM
+OpenAI compatible API server.
+
+The following metrics are exposed:
+
+.. literalinclude:: ../../../vllm/engine/metrics.py
+    :language: python
+    :start-after: begin-metrics-definitions
+    :end-before: end-metrics-definitions

docs/source/serving/run_on_sky.rst

@@ -0,0 +1,69 @@
+.. _on_cloud:
+
+Running on clouds with SkyPilot
+===============================
+
+.. raw:: html
+
+    <p align="center">
+        <img src="https://imgur.com/yxtzPEu.png" alt="vLLM"/>
+    </p>
+
+vLLM can be run on the cloud to scale to multiple GPUs with `SkyPilot <https://github.com/skypilot-org/skypilot>`__, an open-source framework for running LLMs on any cloud.
+
+To install SkyPilot and setup your cloud credentials, run:
+
+.. code-block:: console
+
+    $ pip install skypilot
+    $ sky check
+
+See the vLLM SkyPilot YAML for serving, `serving.yaml <https://github.com/skypilot-org/skypilot/blob/master/llm/vllm/serve.yaml>`__.
+
+.. code-block:: yaml
+
+    resources:
+        accelerators: A100
+
+    envs:
+        MODEL_NAME: decapoda-research/llama-13b-hf
+        TOKENIZER: hf-internal-testing/llama-tokenizer
+
+    setup: |
+        conda create -n vllm python=3.9 -y
+        conda activate vllm
+        git clone https://github.com/vllm-project/vllm.git
+        cd vllm
+        pip install .
+        pip install gradio
+
+    run: |
+        conda activate vllm
+        echo 'Starting vllm api server...'
+        python -u -m vllm.entrypoints.api_server \
+                        --model $MODEL_NAME \
+                        --tensor-parallel-size $SKYPILOT_NUM_GPUS_PER_NODE \
+                        --tokenizer $TOKENIZER 2>&1 | tee api_server.log &
+        echo 'Waiting for vllm api server to start...'
+        while ! `cat api_server.log | grep -q 'Uvicorn running on'`; do sleep 1; done
+        echo 'Starting gradio server...'
+        python vllm/examples/gradio_webserver.py
+
+Start the serving the LLaMA-13B model on an A100 GPU:
+
+.. code-block:: console
+
+    $ sky launch serving.yaml
+
+Check the output of the command. There will be a sharable gradio link (like the last line of the following). Open it in your browser to use the LLaMA model to do the text completion.
+
+.. code-block:: console
+
+    (task, pid=7431) Running on public URL: https://<gradio-hash>.gradio.live
+
+**Optional**: Serve the 65B model instead of the default 13B and use more GPU:
+
+.. code-block:: console
+
+    sky launch -c vllm-serve-new -s serve.yaml --gpus A100:8 --env MODEL_NAME=decapoda-research/llama-65b-hf
+

docs/source/serving/serving_with_langchain.rst

@@ -0,0 +1,31 @@
+.. _run_on_langchain:
+
+Serving with Langchain
+============================
+
+vLLM is also available via `Langchain <https://github.com/langchain-ai/langchain>`_ .
+
+To install langchain, run
+
+.. code-block:: console
+
+    $ pip install langchain -q
+
+To run inference on a single or multiple GPUs, use ``VLLM`` class from ``langchain``.
+
+.. code-block:: python
+
+    from langchain.llms import VLLM
+
+    llm = VLLM(model="mosaicml/mpt-7b",
+               trust_remote_code=True,  # mandatory for hf models
+               max_new_tokens=128,
+               top_k=10,
+               top_p=0.95,
+               temperature=0.8,
+               # tensor_parallel_size=... # for distributed inference
+    )
+
+    print(llm("What is the capital of France ?"))
+
+Please refer to this `Tutorial <https://github.com/langchain-ai/langchain/blob/master/docs/extras/integrations/llms/vllm.ipynb>`_ for more details.

examples/gradio_webserver.py

@@ -39,7 +39,7 @@ def build_demo():
 
 if __name__ == "__main__":
     parser = argparse.ArgumentParser()
-    parser.add_argument("--host", type=str, default="localhost")
+    parser.add_argument("--host", type=str, default=None)
     parser.add_argument("--port", type=int, default=8001)
     parser.add_argument("--model-url",
                         type=str,

examples/llm_engine_example.py

@@ -1,16 +1,14 @@
 import argparse
+from typing import List, Tuple
 
-from vllm import EngineArgs, LLMEngine, SamplingParams
+from vllm import EngineArgs, LLMEngine, SamplingParams, RequestOutput
 
 
-def main(args: argparse.Namespace):
-    # Parse the CLI argument and initialize the engine.
-    engine_args = EngineArgs.from_cli_args(args)
-    engine = LLMEngine.from_engine_args(engine_args)
-
-    # Test the following prompts.
-    test_prompts = [
-        ("A robot may not injure a human being", SamplingParams()),
+def create_test_prompts() -> List[Tuple[str, SamplingParams]]:
+    """Create a list of test prompts with their sampling parameters."""
+    return [
+        ("A robot may not injure a human being",
+         SamplingParams(temperature=0.0, logprobs=1, prompt_logprobs=1)),
         ("To be or not to be,",
          SamplingParams(temperature=0.8, top_k=5, presence_penalty=0.2)),
         ("What is the meaning of life?",
@@ -24,22 +22,36 @@ def main(args: argparse.Namespace):
                         temperature=0.0)),
     ]
 
-    # Run the engine by calling `engine.step()` manually.
+
+def process_requests(engine: LLMEngine,
+                     test_prompts: List[Tuple[str, SamplingParams]]):
+    """Continuously process a list of prompts and handle the outputs."""
     request_id = 0
-    while True:
-        # To test iteration-level scheduling, we add one request at each step.
+
+    while test_prompts or engine.has_unfinished_requests():
         if test_prompts:
             prompt, sampling_params = test_prompts.pop(0)
             engine.add_request(str(request_id), prompt, sampling_params)
             request_id += 1
 
-        request_outputs = engine.step()
+        request_outputs: List[RequestOutput] = engine.step()
+
         for request_output in request_outputs:
             if request_output.finished:
                 print(request_output)
 
-        if not (engine.has_unfinished_requests() or test_prompts):
-            break
+
+def initialize_engine(args: argparse.Namespace) -> LLMEngine:
+    """Initialize the LLMEngine from the command line arguments."""
+    engine_args = EngineArgs.from_cli_args(args)
+    return LLMEngine.from_engine_args(engine_args)
+
+
+def main(args: argparse.Namespace):
+    """Main function that sets up and runs the prompt processing."""
+    engine = initialize_engine(args)
+    test_prompts = create_test_prompts()
+    process_requests(engine, test_prompts)
 
 
 if __name__ == '__main__':

examples/openai_chatcompletion_client.py

@@ -0,0 +1,37 @@
+from openai import OpenAI
+
+# Modify OpenAI's API key and API base to use vLLM's API server.
+openai_api_key = "EMPTY"
+openai_api_base = "http://localhost:8000/v1"
+
+client = OpenAI(
+    # defaults to os.environ.get("OPENAI_API_KEY")
+    api_key=openai_api_key,
+    base_url=openai_api_base,
+)
+
+models = client.models.list()
+model = models.data[0].id
+
+chat_completion = client.chat.completions.create(
+    messages=[{
+        "role": "system",
+        "content": "You are a helpful assistant."
+    }, {
+        "role": "user",
+        "content": "Who won the world series in 2020?"
+    }, {
+        "role":
+        "assistant",
+        "content":
+        "The Los Angeles Dodgers won the World Series in 2020."
+    }, {
+        "role": "user",
+        "content": "Where was it played?"
+    }],
+    model=model,
+)
+
+
+print("Chat completion results:")
+print(chat_completion)

examples/openai_client.py

@@ -1,28 +0,0 @@
-import openai
-
-# Modify OpenAI's API key and API base to use vLLM's API server.
-openai.api_key = "EMPTY"
-openai.api_base = "http://localhost:8000/v1"
-model = "facebook/opt-125m"
-
-# Test list models API
-models = openai.Model.list()
-print("Models:", models)
-
-# Test completion API
-stream = True
-completion = openai.Completion.create(
-    model=model,
-    prompt="A robot may not injure a human being",
-    echo=False,
-    n=2,
-    best_of=3,
-    stream=stream,
-    logprobs=3)
-
-# print the completion
-if stream:
-    for c in completion:
-        print(c)
-else:
-    print("Completion result:", completion)

examples/openai_completion_client.py

@@ -0,0 +1,32 @@
+from openai import OpenAI
+
+# Modify OpenAI's API key and API base to use vLLM's API server.
+openai_api_key = "EMPTY"
+openai_api_base = "http://localhost:8000/v1"
+
+client = OpenAI(
+    # defaults to os.environ.get("OPENAI_API_KEY")
+    api_key=openai_api_key,
+    base_url=openai_api_base,
+)
+
+models = client.models.list()
+model = models.data[0].id
+
+# Completion API
+stream = False
+completion = client.completions.create(
+    model=model,
+    prompt="A robot may not injure a human being",
+    echo=False,
+    n=2,
+    stream=stream,
+    logprobs=3
+)
+
+print("Completion results:")
+if stream:
+    for c in completion:
+        print(c)
+else:
+    print(completion)

examples/template_alpaca.jinja

@@ -0,0 +1,29 @@
+{{ (messages|selectattr('role', 'equalto', 'system')|list|last).content|trim if (messages|selectattr('role', 'equalto', 'system')|list) else '' }}
+
+{% for message in messages %}
+{% if message['role'] == 'user' %}
+### Instruction:
+{{ message['content']|trim -}}
+{% if not loop.last %}
+
+
+{% endif %}
+{% elif message['role'] == 'assistant' %}
+### Response:
+{{ message['content']|trim -}}
+{% if not loop.last %}
+
+
+{% endif %}
+{% elif message['role'] == 'user_context' %}
+### Input:
+{{ message['content']|trim -}}
+{% if not loop.last %}
+
+
+{% endif %}
+{% endif %}
+{% endfor %}
+{% if add_generation_prompt and messages[-1]['role'] != 'assistant' %}
+### Response:
+{% endif %}

examples/template_chatml.jinja

@@ -0,0 +1,2 @@
+{% for message in messages %}{{'<|im_start|>' + message['role'] + '\n' + message['content']}}{% if (loop.last and add_generation_prompt) or not loop.last %}{{ '<|im_end|>' + '\n'}}{% endif %}{% endfor %}
+{% if add_generation_prompt and messages[-1]['role'] != 'assistant' %}{{ '<|im_start|>assistant\n' }}{% endif %}

examples/template_inkbot.jinja

@@ -0,0 +1,30 @@
+<#meta#>
+- Date: {{ (messages|selectattr('role', 'equalto', 'meta-current_date')|list|last).content|trim if (messages|selectattr('role', 'equalto', 'meta-current_date')|list) else '' }}
+- Task: {{ (messages|selectattr('role', 'equalto', 'meta-task_name')|list|last).content|trim if (messages|selectattr('role', 'equalto', 'meta-task_name')|list) else '' }}
+<#system#>
+{{ (messages|selectattr('role', 'equalto', 'system')|list|last).content|trim if (messages|selectattr('role', 'equalto', 'system')|list) else '' }}
+<#chat#>
+{% for message in messages %}
+{% if message['role'] == 'user' %}
+<#user#>
+{{ message['content']|trim -}}
+{% if not loop.last %}
+
+{% endif %}
+{% elif message['role'] == 'assistant' %}
+<#bot#>
+{{ message['content']|trim -}}
+{% if not loop.last %}
+
+{% endif %}
+{% elif message['role'] == 'user_context' %}
+<#user_context#>
+{{ message['content']|trim -}}
+{% if not loop.last %}
+
+{% endif %}
+{% endif %}
+{% endfor %}
+{% if add_generation_prompt and messages[-1]['role'] != 'assistant' %}
+<#bot#>
+{% endif %}

format.sh

@@ -7,7 +7,7 @@
 #    # Format files that differ from origin/main.
 #    bash format.sh
 
-#    # Commit changed files with message 'Run yapf and pylint'
+#    # Commit changed files with message 'Run yapf and ruff'
 #
 #
 # YAPF + Clang formatter (if installed). This script formats all changed files from the last mergebase.
@@ -22,7 +22,7 @@ ROOT="$(git rev-parse --show-toplevel)"
 builtin cd "$ROOT" || exit 1
 
 YAPF_VERSION=$(yapf --version | awk '{print $2}')
-PYLINT_VERSION=$(pylint --version | head -n 1 | awk '{print $2}')
+RUFF_VERSION=$(ruff --version | awk '{print $2}')
 MYPY_VERSION=$(mypy --version | awk '{print $2}')
 
 # # params: tool name, tool version, required version
@@ -34,7 +34,7 @@ tool_version_check() {
 }
 
 tool_version_check "yapf" $YAPF_VERSION "$(grep yapf requirements-dev.txt | cut -d'=' -f3)"
-tool_version_check "pylint" $PYLINT_VERSION "$(grep "pylint==" requirements-dev.txt | cut -d'=' -f3)"
+tool_version_check "ruff" $RUFF_VERSION "$(grep "ruff==" requirements-dev.txt | cut -d'=' -f3)"
 tool_version_check "mypy" "$MYPY_VERSION" "$(grep mypy requirements-dev.txt | cut -d'=' -f3)"
 
 YAPF_FLAGS=(
@@ -44,7 +44,6 @@ YAPF_FLAGS=(
 
 YAPF_EXCLUDES=(
     '--exclude' 'build/**'
-    '--exclude' 'vllm/model_executor/parallel_utils/**'
 )
 
 # Format specified files
@@ -72,7 +71,7 @@ format_changed() {
 
 # Format all files
 format_all() {
-    yapf --in-place "${YAPF_FLAGS[@]}" "${YAPF_EXCLUDES[@]}" vllm
+    yapf --in-place "${YAPF_FLAGS[@]}" "${YAPF_EXCLUDES[@]}" vllm tests
 }
 
 ## This flag formats individual files. --files *must* be the first command line
@@ -94,9 +93,43 @@ echo 'vLLM yapf: Done'
 # echo 'vLLM mypy:'
 # mypy
 
-# Run Pylint
-echo 'vLLM Pylint:'
-pylint vllm
+# Lint specified files
+lint() {
+    ruff "$@"
+}
+
+# Lint files that differ from main branch. Ignores dirs that are not slated
+# for autolint yet.
+lint_changed() {
+    # The `if` guard ensures that the list of filenames is not empty, which
+    # could cause ruff to receive 0 positional arguments, making it hang
+    # waiting for STDIN.
+    #
+    # `diff-filter=ACM` and $MERGEBASE is to ensure we only lint files that
+    # exist on both branches.
+    MERGEBASE="$(git merge-base origin/main HEAD)"
+
+    if ! git diff --diff-filter=ACM --quiet --exit-code "$MERGEBASE" -- '*.py' '*.pyi' &>/dev/null; then
+        git diff --name-only --diff-filter=ACM "$MERGEBASE" -- '*.py' '*.pyi' | xargs \
+             ruff
+    fi
+
+}
+
+# Run Ruff
+echo 'vLLM Ruff:'
+## This flag lints individual files. --files *must* be the first command line
+## arg to use this option.
+if [[ "$1" == '--files' ]]; then
+   lint "${@:2}"
+   # If `--all` is passed, then any further arguments are ignored and the
+   # entire python directory is linted.
+elif [[ "$1" == '--all' ]]; then
+   lint vllm tests
+else
+   # Format only the files that changed in last commit.
+   lint_changed
+fi
 
 if ! git diff --quiet &>/dev/null; then
     echo 'Reformatted files. Please review and stage the changes.'

pyproject.toml

@@ -1,9 +1,34 @@
 [build-system]
+# Should be mirrored in requirements-build.txt
 requires = [
     "ninja",
     "packaging",
-    "setuptools",
-    "torch >= 2.0.0",
+    "setuptools >= 49.4.0",
+    "torch >= 2.1.0",
     "wheel",
 ]
 build-backend = "setuptools.build_meta"
+
+[tool.ruff.lint]
+select = [
+    # pycodestyle
+    "E",
+    # Pyflakes
+    "F",
+    # pyupgrade
+    # "UP",
+    # flake8-bugbear
+    "B",
+    # flake8-simplify
+    "SIM",
+    # isort
+    # "I",
+]
+ignore = [
+    # star imports
+    "F405", "F403",
+    # lambda expression assignment
+    "E731",
+    # line too long, handled by black formatting
+    "E501",
+]

requirements-build.txt

@@ -0,0 +1,6 @@
+# Should be mirrored in pyproject.toml
+ninja
+packaging
+setuptools>=49.4.0
+torch>=2.1.0
+wheel

requirements-dev.txt

@@ -1,6 +1,6 @@
 # formatting
 yapf==0.32.0
-pylint==2.8.2
+ruff==0.1.5
 
 # type checking
 mypy==0.991
@@ -10,3 +10,6 @@ types-setuptools
 
 # testing
 pytest
+pytest-forked
+pytest-asyncio
+

requirements.txt

@@ -1,12 +1,15 @@
 ninja  # For faster builds.
 psutil
-ray
+ray >= 2.5.1
+pandas  # Required for Ray data.
+pyarrow  # Required for Ray data.
 sentencepiece  # Required for LLaMA tokenizer.
 numpy
-torch >= 2.0.0
-transformers >= 4.28.0  # Required for LLaMA.
-xformers >= 0.0.19
+einops  # Required for phi-1_5
+torch >= 2.1.0
+transformers >= 4.34.0  # Required for Mistral.
+xformers >= 0.0.22.post7  # Required for CUDA 12.1.
 fastapi
-uvicorn
-pydantic  # Required for OpenAI server.
-fschat  # Required for OpenAI ChatCompletion Endpoint.
+uvicorn[standard]
+pydantic == 1.10.13  # Required for OpenAI server.
+aioprometheus[starlette]

setup.py

@@ -3,6 +3,7 @@ import os
 import re
 import subprocess
 from typing import List, Set
+import warnings
 
 from packaging.version import parse, Version
 import setuptools
@@ -11,6 +12,11 @@ from torch.utils.cpp_extension import BuildExtension, CUDAExtension, CUDA_HOME
 
 ROOT_DIR = os.path.dirname(__file__)
 
+MAIN_CUDA_VERSION = "12.1"
+
+# Supported NVIDIA GPU architectures.
+SUPPORTED_ARCHS = {"7.0", "7.5", "8.0", "8.6", "8.9", "9.0"}
+
 # Compiler flags.
 CXX_FLAGS = ["-g", "-O2", "-std=c++17"]
 # TODO(woosuk): Should we use -O3?
@@ -22,7 +28,7 @@ NVCC_FLAGS += [f"-D_GLIBCXX_USE_CXX11_ABI={ABI}"]
 
 if CUDA_HOME is None:
     raise RuntimeError(
-        f"Cannot find CUDA_HOME. CUDA must be available in order to build the package.")
+        "Cannot find CUDA_HOME. CUDA must be available to build the package.")
 
 
 def get_nvcc_cuda_version(cuda_dir: str) -> Version:
@@ -38,32 +44,97 @@ def get_nvcc_cuda_version(cuda_dir: str) -> Version:
     return nvcc_cuda_version
 
 
-# Collect the compute capabilities of all available GPUs.
-device_count = torch.cuda.device_count()
-compute_capabilities: Set[int] = set()
-for i in range(device_count):
-    major, minor = torch.cuda.get_device_capability(i)
-    if major < 7:
+def get_torch_arch_list() -> Set[str]:
+    # TORCH_CUDA_ARCH_LIST can have one or more architectures,
+    # e.g. "8.0" or "7.5,8.0,8.6+PTX". Here, the "8.6+PTX" option asks the
+    # compiler to additionally include PTX code that can be runtime-compiled
+    # and executed on the 8.6 or newer architectures. While the PTX code will
+    # not give the best performance on the newer architectures, it provides
+    # forward compatibility.
+    env_arch_list = os.environ.get("TORCH_CUDA_ARCH_LIST", None)
+    if env_arch_list is None:
+        return set()
+
+    # List are separated by ; or space.
+    torch_arch_list = set(env_arch_list.replace(" ", ";").split(";"))
+    if not torch_arch_list:
+        return set()
+
+    # Filter out the invalid architectures and print a warning.
+    valid_archs = SUPPORTED_ARCHS.union({s + "+PTX" for s in SUPPORTED_ARCHS})
+    arch_list = torch_arch_list.intersection(valid_archs)
+    # If none of the specified architectures are valid, raise an error.
+    if not arch_list:
         raise RuntimeError(
-            "GPUs with compute capability less than 7.0 are not supported.")
-    compute_capabilities.add(major * 10 + minor)
-# If no GPU is available, add all supported compute capabilities.
+            "None of the CUDA architectures in `TORCH_CUDA_ARCH_LIST` env "
+            f"variable ({env_arch_list}) is supported. "
+            f"Supported CUDA architectures are: {valid_archs}.")
+    invalid_arch_list = torch_arch_list - valid_archs
+    if invalid_arch_list:
+        warnings.warn(
+            f"Unsupported CUDA architectures ({invalid_arch_list}) are "
+            "excluded from the `TORCH_CUDA_ARCH_LIST` env variable "
+            f"({env_arch_list}). Supported CUDA architectures are: "
+            f"{valid_archs}.",
+            stacklevel=2)
+    return arch_list
+
+
+# First, check the TORCH_CUDA_ARCH_LIST environment variable.
+compute_capabilities = get_torch_arch_list()
 if not compute_capabilities:
-    compute_capabilities = {70, 75, 80, 86, 90}
-# Add target compute capabilities to NVCC flags.
-for capability in compute_capabilities:
-    NVCC_FLAGS += ["-gencode", f"arch=compute_{capability},code=sm_{capability}"]
+    # If TORCH_CUDA_ARCH_LIST is not defined or empty, target all available
+    # GPUs on the current machine.
+    device_count = torch.cuda.device_count()
+    for i in range(device_count):
+        major, minor = torch.cuda.get_device_capability(i)
+        if major < 7:
+            raise RuntimeError(
+                "GPUs with compute capability below 7.0 are not supported.")
+        compute_capabilities.add(f"{major}.{minor}")
+
+nvcc_cuda_version = get_nvcc_cuda_version(CUDA_HOME)
+if not compute_capabilities:
+    # If no GPU is specified nor available, add all supported architectures
+    # based on the NVCC CUDA version.
+    compute_capabilities = SUPPORTED_ARCHS.copy()
+    if nvcc_cuda_version < Version("11.1"):
+        compute_capabilities.remove("8.6")
+    if nvcc_cuda_version < Version("11.8"):
+        compute_capabilities.remove("8.9")
+        compute_capabilities.remove("9.0")
 
 # Validate the NVCC CUDA version.
-nvcc_cuda_version = get_nvcc_cuda_version(CUDA_HOME)
 if nvcc_cuda_version < Version("11.0"):
     raise RuntimeError("CUDA 11.0 or higher is required to build the package.")
-if 86 in compute_capabilities and nvcc_cuda_version < Version("11.1"):
+if (nvcc_cuda_version < Version("11.1")
+        and any(cc.startswith("8.6") for cc in compute_capabilities)):
     raise RuntimeError(
-        "CUDA 11.1 or higher is required for GPUs with compute capability 8.6.")
-if 90 in compute_capabilities and nvcc_cuda_version < Version("11.8"):
-    raise RuntimeError(
-        "CUDA 11.8 or higher is required for GPUs with compute capability 9.0.")
+        "CUDA 11.1 or higher is required for compute capability 8.6.")
+if nvcc_cuda_version < Version("11.8"):
+    if any(cc.startswith("8.9") for cc in compute_capabilities):
+        # CUDA 11.8 is required to generate the code targeting compute capability 8.9.
+        # However, GPUs with compute capability 8.9 can also run the code generated by
+        # the previous versions of CUDA 11 and targeting compute capability 8.0.
+        # Therefore, if CUDA 11.8 is not available, we target compute capability 8.0
+        # instead of 8.9.
+        warnings.warn(
+            "CUDA 11.8 or higher is required for compute capability 8.9. "
+            "Targeting compute capability 8.0 instead.",
+            stacklevel=2)
+        compute_capabilities = set(cc for cc in compute_capabilities
+                                   if not cc.startswith("8.9"))
+        compute_capabilities.add("8.0+PTX")
+    if any(cc.startswith("9.0") for cc in compute_capabilities):
+        raise RuntimeError(
+            "CUDA 11.8 or higher is required for compute capability 9.0.")
+
+# Add target compute capabilities to NVCC flags.
+for capability in compute_capabilities:
+    num = capability[0] + capability[2]
+    NVCC_FLAGS += ["-gencode", f"arch=compute_{num},code=sm_{num}"]
+    if capability.endswith("+PTX"):
+        NVCC_FLAGS += ["-gencode", f"arch=compute_{num},code=compute_{num}"]
 
 # Use NVCC threads to parallelize the build.
 if nvcc_cuda_version >= Version("11.2"):
@@ -71,68 +142,60 @@ if nvcc_cuda_version >= Version("11.2"):
     NVCC_FLAGS += ["--threads", str(num_threads)]
 
 ext_modules = []
-
-# Cache operations.
-cache_extension = CUDAExtension(
-    name="vllm.cache_ops",
-    sources=["csrc/cache.cpp", "csrc/cache_kernels.cu"],
-    extra_compile_args={"cxx": CXX_FLAGS, "nvcc": NVCC_FLAGS},
+vllm_extension = CUDAExtension(
+    name="vllm._C",
+    sources=[
+        "csrc/cache_kernels.cu",
+        "csrc/attention/attention_kernels.cu",
+        "csrc/pos_encoding_kernels.cu",
+        "csrc/activation_kernels.cu",
+        "csrc/layernorm_kernels.cu",
+        "csrc/quantization/awq/gemm_kernels.cu",
+        "csrc/quantization/squeezellm/quant_cuda_kernel.cu",
+        "csrc/cuda_utils_kernels.cu",
+        "csrc/pybind.cpp",
+    ],
+    extra_compile_args={
+        "cxx": CXX_FLAGS,
+        "nvcc": NVCC_FLAGS,
+    },
 )
-ext_modules.append(cache_extension)
-
-# Attention kernels.
-attention_extension = CUDAExtension(
-    name="vllm.attention_ops",
-    sources=["csrc/attention.cpp", "csrc/attention/attention_kernels.cu"],
-    extra_compile_args={"cxx": CXX_FLAGS, "nvcc": NVCC_FLAGS},
-)
-ext_modules.append(attention_extension)
-
-# Positional encoding kernels.
-positional_encoding_extension = CUDAExtension(
-    name="vllm.pos_encoding_ops",
-    sources=["csrc/pos_encoding.cpp", "csrc/pos_encoding_kernels.cu"],
-    extra_compile_args={"cxx": CXX_FLAGS, "nvcc": NVCC_FLAGS},
-)
-ext_modules.append(positional_encoding_extension)
-
-# Layer normalization kernels.
-layernorm_extension = CUDAExtension(
-    name="vllm.layernorm_ops",
-    sources=["csrc/layernorm.cpp", "csrc/layernorm_kernels.cu"],
-    extra_compile_args={"cxx": CXX_FLAGS, "nvcc": NVCC_FLAGS},
-)
-ext_modules.append(layernorm_extension)
-
-# Activation kernels.
-activation_extension = CUDAExtension(
-    name="vllm.activation_ops",
-    sources=["csrc/activation.cpp", "csrc/activation_kernels.cu"],
-    extra_compile_args={"cxx": CXX_FLAGS, "nvcc": NVCC_FLAGS},
-)
-ext_modules.append(activation_extension)
+ext_modules.append(vllm_extension)
 
 
 def get_path(*filepath) -> str:
     return os.path.join(ROOT_DIR, *filepath)
 
 
-def find_version(filepath: str):
+def find_version(filepath: str) -> str:
     """Extract version information from the given filepath.
 
     Adapted from https://github.com/ray-project/ray/blob/0b190ee1160eeca9796bc091e07eaebf4c85b511/python/setup.py
     """
     with open(filepath) as fp:
-        version_match = re.search(
-            r"^__version__ = ['\"]([^'\"]*)['\"]", fp.read(), re.M)
+        version_match = re.search(r"^__version__ = ['\"]([^'\"]*)['\"]",
+                                  fp.read(), re.M)
         if version_match:
             return version_match.group(1)
         raise RuntimeError("Unable to find version string.")
 
 
+def get_vllm_version() -> str:
+    version = find_version(get_path("vllm", "__init__.py"))
+    cuda_version = str(nvcc_cuda_version)
+    if cuda_version != MAIN_CUDA_VERSION:
+        cuda_version_str = cuda_version.replace(".", "")[:3]
+        version += f"+cu{cuda_version_str}"
+    return version
+
+
 def read_readme() -> str:
-    """Read the README file."""
-    return io.open(get_path("README.md"), "r", encoding="utf-8").read()
+    """Read the README file if present."""
+    p = get_path("README.md")
+    if os.path.isfile(p):
+        return io.open(get_path("README.md"), "r", encoding="utf-8").read()
+    else:
+        return ""
 
 
 def get_requirements() -> List[str]:
@@ -144,10 +207,11 @@ def get_requirements() -> List[str]:
 
 setuptools.setup(
     name="vllm",
-    version=find_version(get_path("vllm", "__init__.py")),
+    version=get_vllm_version(),
     author="vLLM Team",
     license="Apache 2.0",
-    description="A high-throughput and memory-efficient inference and serving engine for LLMs",
+    description=("A high-throughput and memory-efficient inference and "
+                 "serving engine for LLMs"),
     long_description=read_readme(),
     long_description_content_type="text/markdown",
     url="https://github.com/vllm-project/vllm",
@@ -159,13 +223,15 @@ setuptools.setup(
         "Programming Language :: Python :: 3.8",
         "Programming Language :: Python :: 3.9",
         "Programming Language :: Python :: 3.10",
+        "Programming Language :: Python :: 3.11",
         "License :: OSI Approved :: Apache Software License",
         "Topic :: Scientific/Engineering :: Artificial Intelligence",
     ],
-    packages=setuptools.find_packages(
-        exclude=("assets", "benchmarks", "csrc", "docs", "examples", "tests")),
+    packages=setuptools.find_packages(exclude=("benchmarks", "csrc", "docs",
+                                               "examples", "tests")),
     python_requires=">=3.8",
     install_requires=get_requirements(),
     ext_modules=ext_modules,
     cmdclass={"build_ext": BuildExtension},
+    package_data={"vllm": ["py.typed"]},
 )

tests/async_engine/api_server_async_engine.py

@@ -0,0 +1,50 @@
+"""vllm.entrypoints.api_server with some extra logging for testing."""
+import argparse
+from typing import Any, Dict
+
+import uvicorn
+from fastapi.responses import JSONResponse, Response
+
+import vllm.entrypoints.api_server
+from vllm.engine.arg_utils import AsyncEngineArgs
+from vllm.engine.async_llm_engine import AsyncLLMEngine
+
+app = vllm.entrypoints.api_server.app
+
+
+class AsyncLLMEngineWithStats(AsyncLLMEngine):
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self._num_aborts = 0
+
+    async def abort(self, request_id: str) -> None:
+        await super().abort(request_id)
+        self._num_aborts += 1
+
+    def testing_stats(self) -> Dict[str, Any]:
+        return {"num_aborted_requests": self._num_aborts}
+
+
+@app.get("/stats")
+def stats() -> Response:
+    """Get the statistics of the engine."""
+    return JSONResponse(engine.testing_stats())
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--host", type=str, default="localhost")
+    parser.add_argument("--port", type=int, default=8000)
+    parser = AsyncEngineArgs.add_cli_args(parser)
+    args = parser.parse_args()
+
+    engine_args = AsyncEngineArgs.from_cli_args(args)
+    engine = AsyncLLMEngineWithStats.from_engine_args(engine_args)
+    vllm.entrypoints.api_server.engine = engine
+    uvicorn.run(
+        app,
+        host=args.host,
+        port=args.port,
+        log_level="debug",
+        timeout_keep_alive=vllm.entrypoints.api_server.TIMEOUT_KEEP_ALIVE)

tests/async_engine/test_api_server.py

@@ -0,0 +1,86 @@
+import subprocess
+import sys
+import time
+from multiprocessing import Pool
+from pathlib import Path
+
+import pytest
+import requests
+
+
+def _query_server(prompt: str) -> dict:
+    response = requests.post("http://localhost:8000/generate",
+                             json={
+                                 "prompt": prompt,
+                                 "max_tokens": 100,
+                                 "temperature": 0,
+                                 "ignore_eos": True
+                             })
+    response.raise_for_status()
+    return response.json()
+
+
+@pytest.fixture
+def api_server():
+    script_path = Path(__file__).parent.joinpath(
+        "api_server_async_engine.py").absolute()
+    uvicorn_process = subprocess.Popen([
+        sys.executable, "-u",
+        str(script_path), "--model", "facebook/opt-125m"
+    ])
+    yield
+    uvicorn_process.terminate()
+
+
+def test_api_server(api_server):
+    """
+    Run the API server and test it.
+
+    We run both the server and requests in separate processes.
+
+    We test that the server can handle incoming requests, including
+    multiple requests at the same time, and that it can handle requests
+    being cancelled without crashing.
+    """
+    with Pool(32) as pool:
+        # Wait until the server is ready
+        prompts = ["Hello world"] * 1
+        result = None
+        while not result:
+            try:
+                for _ in pool.map(_query_server, prompts):
+                    break
+            except Exception:
+                time.sleep(1)
+
+        # Actual tests start here
+        # Try with 1 prompt
+        for result in pool.map(_query_server, prompts):
+            assert result
+
+        num_aborted_requests = requests.get(
+            "http://localhost:8000/stats").json()["num_aborted_requests"]
+        assert num_aborted_requests == 0
+
+        # Try with 100 prompts
+        prompts = ["Hello world"] * 100
+        for result in pool.map(_query_server, prompts):
+            assert result
+
+        # Cancel requests
+        pool.map_async(_query_server, prompts)
+        time.sleep(0.01)
+        pool.terminate()
+        pool.join()
+
+        # check cancellation stats
+        num_aborted_requests = requests.get(
+            "http://localhost:8000/stats").json()["num_aborted_requests"]
+        assert num_aborted_requests > 0
+
+    # check that server still runs after cancellations
+    with Pool(32) as pool:
+        # Try with 100 prompts
+        prompts = ["Hello world"] * 100
+        for result in pool.map(_query_server, prompts):
+            assert result

tests/async_engine/test_async_llm_engine.py

@@ -0,0 +1,80 @@
+import asyncio
+from dataclasses import dataclass
+
+import pytest
+
+from vllm.engine.async_llm_engine import AsyncLLMEngine
+
+
+@dataclass
+class RequestOutput:
+    request_id: int
+    finished: bool = False
+
+
+class MockEngine:
+
+    def __init__(self):
+        self.step_calls = 0
+        self.add_request_calls = 0
+        self.abort_request_calls = 0
+        self.request_id = None
+
+    async def step_async(self):
+        self.step_calls += 1
+        return [RequestOutput(
+            request_id=self.request_id)] if self.request_id else []
+
+    def generate(self, request_id):
+        self.request_id = request_id
+
+    def stop_generating(self):
+        self.request_id = None
+
+    def add_request(self, **kwargs):
+        del kwargs  # Unused
+        self.add_request_calls += 1
+
+    def abort_request(self, request_id):
+        del request_id  # Unused
+        self.abort_request_calls += 1
+
+
+class MockAsyncLLMEngine(AsyncLLMEngine):
+
+    def _init_engine(self, *args, **kwargs):
+        return MockEngine()
+
+
+@pytest.mark.asyncio
+async def test_new_requests_event():
+    engine = MockAsyncLLMEngine(worker_use_ray=False, engine_use_ray=False)
+    engine.start_background_loop()
+    await asyncio.sleep(0.01)
+    assert engine.engine.step_calls == 0
+
+    await engine.add_request("1", "", None)
+    await asyncio.sleep(0.01)
+    assert engine.engine.add_request_calls == 1
+    assert engine.engine.step_calls == 1
+
+    await engine.add_request("2", "", None)
+    engine.engine.generate("2")
+    await asyncio.sleep(0)
+    assert engine.engine.add_request_calls == 2
+    assert engine.engine.step_calls == 2
+    await asyncio.sleep(0)
+    assert engine.engine.step_calls == 3
+    engine.engine.stop_generating()
+    await asyncio.sleep(0)
+    assert engine.engine.step_calls == 4
+    await asyncio.sleep(0)
+    assert engine.engine.step_calls == 4
+
+    await engine.add_request("3", "", None)
+    await asyncio.sleep(0.01)
+    assert engine.engine.add_request_calls == 3
+    assert engine.engine.step_calls == 5
+    await asyncio.sleep(0.01)
+    assert engine.engine.add_request_calls == 3
+    assert engine.engine.step_calls == 5

tests/async_engine/test_openai_server.py

@@ -0,0 +1,119 @@
+from argparse import Namespace
+from dataclasses import dataclass
+
+import pytest
+from fastapi.testclient import TestClient
+
+from vllm.entrypoints.openai.api_server import *
+
+# Define models, templates, and their corresponding expected outputs
+MODEL_TEMPLATE_GENERATON_OUTPUT = [
+    ("facebook/opt-125m", None, True,
+     "Hello</s>Hi there!</s>What is the capital of</s>"),
+    ("facebook/opt-125m", None, False,
+     "Hello</s>Hi there!</s>What is the capital of</s>"),
+    ("facebook/opt-125m", "../../examples/template_chatml.jinja", True,
+     """<|im_start|>user
+Hello<|im_end|>
+<|im_start|>assistant
+Hi there!<|im_end|>
+<|im_start|>user
+What is the capital of<|im_end|>
+<|im_start|>assistant
+"""),
+    ("facebook/opt-125m", "../../examples/template_chatml.jinja", False,
+     """<|im_start|>user
+Hello<|im_end|>
+<|im_start|>assistant
+Hi there!<|im_end|>
+<|im_start|>user
+What is the capital of""")
+]
+
+TEST_MESSAGES = [
+    {
+        'role': 'user',
+        'content': 'Hello'
+    },
+    {
+        'role': 'assistant',
+        'content': 'Hi there!'
+    },
+    {
+        'role': 'user',
+        'content': 'What is the capital of'
+    },
+]
+client = TestClient(app)
+
+
+@dataclass
+class MockTokenizer:
+    chat_template = None
+
+
+def test_load_chat_template():
+    # Testing chatml template
+    template = "../../examples/template_chatml.jinja"
+    mock_args = Namespace(chat_template=template)
+    tokenizer = MockTokenizer()
+
+    # Call the function with the mocked args
+    load_chat_template(mock_args, tokenizer)
+
+    template_content = tokenizer.chat_template
+
+    # Test assertions
+    assert template_content is not None
+    # Hard coded value for template_chatml.jinja
+    assert template_content == """{% for message in messages %}{{'<|im_start|>' + message['role'] + '\\n' + message['content']}}{% if (loop.last and add_generation_prompt) or not loop.last %}{{ '<|im_end|>' + '\\n'}}{% endif %}{% endfor %}
+{% if add_generation_prompt and messages[-1]['role'] != 'assistant' %}{{ '<|im_start|>assistant\\n' }}{% endif %}"""
+
+
+def test_no_load_chat_template():
+    # Testing chatml template
+    template = "../../examples/does_not_exist"
+    mock_args = Namespace(chat_template=template)
+    tokenizer = MockTokenizer()
+
+    # Call the function with the mocked args
+    load_chat_template(mock_args, tokenizer=tokenizer)
+    template_content = tokenizer.chat_template
+
+    # Test assertions
+    assert template_content is not None
+    # Hard coded value for template_chatml.jinja
+    assert template_content == """../../examples/does_not_exist"""
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "model,template,add_generation_prompt,expected_output",
+    MODEL_TEMPLATE_GENERATON_OUTPUT)
+async def test_get_gen_prompt(model, template, add_generation_prompt,
+                              expected_output):
+    # Initialize the tokenizer
+    tokenizer = get_tokenizer(tokenizer_name=model)
+
+    mock_args = Namespace(chat_template=template)
+    load_chat_template(mock_args, tokenizer)
+
+    # Create a mock request object using keyword arguments
+    mock_request = ChatCompletionRequest(
+        model=model,
+        messages=TEST_MESSAGES,
+        add_generation_prompt=add_generation_prompt)
+
+    # Call the function and get the result
+    result = tokenizer.apply_chat_template(
+        conversation=mock_request.messages,
+        tokenize=False,
+        add_generation_prompt=mock_request.add_generation_prompt)
+
+    # Test assertion
+    assert result == expected_output, f"The generated prompt does not match the expected output for model {model} and template {template}"
+
+
+def test_health_endpoint():
+    response = client.get("/health")
+    assert response.status_code == 200

tests/async_engine/test_request_tracker.py

@@ -0,0 +1,75 @@
+import pytest
+
+from vllm.engine.async_llm_engine import RequestTracker
+from vllm.outputs import RequestOutput
+
+
+class DummyEvent:
+
+    def __init__(self):
+        self.flag = False
+
+    def set(self):
+        self.flag = True
+
+    def clear(self):
+        self.flag = False
+
+
+def test_request_tracker():
+    tracker = RequestTracker()
+    tracker.new_requests_event = DummyEvent()
+    stream_1 = tracker.add_request("1")
+    assert tracker.new_requests_event.flag
+    new, finished = tracker.get_new_and_finished_requests()
+    assert not tracker.new_requests_event.flag
+    assert len(new) == 1
+    assert new[0]["request_id"] == "1"
+    assert not finished
+    assert not stream_1.finished
+
+    stream_2 = tracker.add_request("2")
+    stream_3 = tracker.add_request("3")
+    assert tracker.new_requests_event.flag
+    new, finished = tracker.get_new_and_finished_requests()
+    assert not tracker.new_requests_event.flag
+    assert len(new) == 2
+    assert new[0]["request_id"] == "2"
+    assert new[1]["request_id"] == "3"
+    assert not finished
+    assert not stream_2.finished
+    assert not stream_3.finished
+
+    # request_ids must be unique
+    with pytest.raises(KeyError):
+        tracker.add_request("1")
+    assert not tracker.new_requests_event.flag
+
+    tracker.abort_request("1")
+    new, finished = tracker.get_new_and_finished_requests()
+    assert len(finished) == 1
+    assert "1" in finished
+    assert not new
+    assert stream_1.finished
+
+    stream_4 = tracker.add_request("4")
+    tracker.abort_request("4")
+    assert tracker.new_requests_event.flag
+    new, finished = tracker.get_new_and_finished_requests()
+    assert len(finished) == 1
+    assert "4" in finished
+    assert not new
+    assert stream_4.finished
+
+    stream_5 = tracker.add_request("5")
+    assert tracker.new_requests_event.flag
+    tracker.process_request_output(
+        RequestOutput("2", "output", [], [], [], finished=True))
+    new, finished = tracker.get_new_and_finished_requests()
+    assert not tracker.new_requests_event.flag
+    assert len(finished) == 1
+    assert "2" in finished
+    assert len(new) == 1
+    assert new[0]["request_id"] == "5"
+    assert stream_2.finished
+    assert not stream_5.finished

tests/conftest.py

@@ -0,0 +1,211 @@
+from typing import List, Optional, Tuple
+
+import pytest
+import torch
+from transformers import AutoModelForCausalLM
+
+from vllm import LLM, SamplingParams
+from vllm.transformers_utils.tokenizer import get_tokenizer
+
+_TEST_PROMPTS = [
+    "vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs.",
+    "Briefly describe the major milestones in the development of artificial intelligence from 1950 to 2020.",
+    "Compare and contrast artificial intelligence with human intelligence in terms of processing information.",
+    "Describe the basic components of a neural network and how it can be trained.",
+    "Write a short story about a robot that dreams for the first time.",
+    "Analyze the impact of the COVID-19 pandemic on global economic structures and future business models.",
+    "Explain the cultural significance of the Mona Lisa painting, and how its perception might vary in Western versus Eastern societies.",
+    "Translate the following English sentence into Japanese, French, and Swahili: 'The early bird catches the worm.'",
+]
+
+
+@pytest.fixture
+def example_prompts() -> List[str]:
+    return _TEST_PROMPTS
+
+
+_STR_DTYPE_TO_TORCH_DTYPE = {
+    "half": torch.half,
+    "bfloat16": torch.bfloat16,
+    "float": torch.float,
+}
+
+
+class HfRunner:
+
+    def __init__(
+        self,
+        model_name: str,
+        tokenizer_name: Optional[str] = None,
+        dtype: str = "half",
+    ) -> None:
+        assert dtype in _STR_DTYPE_TO_TORCH_DTYPE
+        torch_dtype = _STR_DTYPE_TO_TORCH_DTYPE[dtype]
+        self.model = AutoModelForCausalLM.from_pretrained(
+            model_name,
+            torch_dtype=torch_dtype,
+            trust_remote_code=True,
+        ).cuda()
+        if tokenizer_name is None:
+            tokenizer_name = model_name
+        self.tokenizer = get_tokenizer(tokenizer_name, trust_remote_code=True)
+
+    def generate(
+        self,
+        prompts: List[str],
+        **kwargs,
+    ) -> List[Tuple[List[int], str]]:
+        outputs: List[Tuple[List[int], str]] = []
+        for prompt in prompts:
+            input_ids = self.tokenizer(prompt, return_tensors="pt").input_ids
+            output_ids = self.model.generate(
+                input_ids.cuda(),
+                use_cache=True,
+                **kwargs,
+            )
+            output_str = self.tokenizer.batch_decode(
+                output_ids,
+                skip_special_tokens=True,
+                clean_up_tokenization_spaces=False,
+            )
+            output_ids = output_ids.cpu().tolist()
+            outputs.append((output_ids, output_str))
+        return outputs
+
+    def generate_greedy(
+        self,
+        prompts: List[str],
+        max_tokens: int,
+    ) -> List[Tuple[List[int], str]]:
+        outputs = self.generate(prompts,
+                                do_sample=False,
+                                max_new_tokens=max_tokens)
+        for i in range(len(outputs)):
+            output_ids, output_str = outputs[i]
+            outputs[i] = (output_ids[0], output_str[0])
+        return outputs
+
+    def generate_beam_search(
+        self,
+        prompts: List[str],
+        beam_width: int,
+        max_tokens: int,
+    ) -> List[Tuple[List[int], str]]:
+        outputs = self.generate(prompts,
+                                do_sample=False,
+                                max_new_tokens=max_tokens,
+                                num_beams=beam_width,
+                                num_return_sequences=beam_width)
+        for i in range(len(outputs)):
+            output_ids, output_str = outputs[i]
+            for j in range(len(output_ids)):
+                output_ids[j] = [
+                    x for x in output_ids[j]
+                    if x != self.tokenizer.pad_token_id
+                ]
+            outputs[i] = (output_ids, output_str)
+        return outputs
+
+    def generate_greedy_logprobs(
+        self,
+        prompts: List[str],
+        max_tokens: int,
+    ) -> List[List[torch.Tensor]]:
+        all_logprobs = []
+        for prompt in prompts:
+            input_ids = self.tokenizer(prompt, return_tensors="pt").input_ids
+            output = self.model.generate(
+                input_ids.cuda(),
+                use_cache=True,
+                do_sample=False,
+                max_new_tokens=max_tokens,
+                output_hidden_states=True,
+                return_dict_in_generate=True,
+            )
+            seq_logprobs = []
+            for hidden_states in output.hidden_states:
+                last_hidden_states = hidden_states[-1][0]
+                logits = torch.matmul(
+                    last_hidden_states,
+                    self.model.get_output_embeddings().weight.t(),
+                )
+                if self.model.get_output_embeddings().bias is not None:
+                    logits += self.model.get_output_embeddings(
+                    ).bias.unsqueeze(0)
+                logprobs = torch.nn.functional.log_softmax(logits,
+                                                           dim=-1,
+                                                           dtype=torch.float32)
+                seq_logprobs.append(logprobs)
+            all_logprobs.append(seq_logprobs)
+        return all_logprobs
+
+
+@pytest.fixture
+def hf_runner():
+    return HfRunner
+
+
+class VllmRunner:
+
+    def __init__(
+        self,
+        model_name: str,
+        tokenizer_name: Optional[str] = None,
+        dtype: str = "half",
+    ) -> None:
+        self.model = LLM(
+            model=model_name,
+            tokenizer=tokenizer_name,
+            trust_remote_code=True,
+            dtype=dtype,
+            swap_space=0,
+        )
+
+    def generate(
+        self,
+        prompts: List[str],
+        sampling_params: SamplingParams,
+    ) -> List[Tuple[List[int], str]]:
+        req_outputs = self.model.generate(prompts,
+                                          sampling_params=sampling_params)
+        outputs = []
+        for req_output in req_outputs:
+            prompt_str = req_output.prompt
+            prompt_ids = req_output.prompt_token_ids
+            req_sample_output_ids = []
+            req_sample_output_strs = []
+            for sample in req_output.outputs:
+                output_str = sample.text
+                output_ids = sample.token_ids
+                req_sample_output_ids.append(prompt_ids + output_ids)
+                req_sample_output_strs.append(prompt_str + output_str)
+            outputs.append((req_sample_output_ids, req_sample_output_strs))
+        return outputs
+
+    def generate_greedy(
+        self,
+        prompts: List[str],
+        max_tokens: int,
+    ) -> List[Tuple[List[int], str]]:
+        greedy_params = SamplingParams(temperature=0.0, max_tokens=max_tokens)
+        outputs = self.generate(prompts, greedy_params)
+        return [(output_ids[0], output_str[0])
+                for output_ids, output_str in outputs]
+
+    def generate_beam_search(
+        self,
+        prompts: List[str],
+        beam_width: int,
+        max_tokens: int,
+    ) -> List[Tuple[List[int], str]]:
+        beam_search_params = SamplingParams(n=beam_width,
+                                            use_beam_search=True,
+                                            temperature=0.0,
+                                            max_tokens=max_tokens)
+        outputs = self.generate(prompts, beam_search_params)
+        return outputs
+
+
+@pytest.fixture
+def vllm_runner():
+    return VllmRunner

tests/distributed/test_comm_ops.py

@@ -0,0 +1,83 @@
+"""Test the communication operators.
+
+Run `pytest tests/distributed/test_comm_ops.py --forked`.
+"""
+from multiprocessing import Process, set_start_method
+
+import pytest
+import torch
+
+from vllm.config import ParallelConfig
+from vllm.engine.ray_utils import get_open_port
+from vllm.model_executor.parallel_utils.communication_op import (
+    tensor_model_parallel_all_reduce,
+    tensor_model_parallel_all_gather,
+)
+from vllm.worker.worker import _init_distributed_environment
+
+
+def init_test_distributed_environment(pipeline_parallel_size: int,
+                                      tensor_parallel_size: int, rank: int,
+                                      distributed_init_port: str):
+    parallel_config = ParallelConfig(pipeline_parallel_size,
+                                     tensor_parallel_size,
+                                     worker_use_ray=True)
+    distributed_init_method = f"tcp://localhost:{distributed_init_port}"
+    torch.cuda.set_device(rank)
+    _init_distributed_environment(parallel_config, rank,
+                                  distributed_init_method)
+
+
+def all_reduce_test_worker(tensor_parallel_size: int, rank: int,
+                           distributed_init_port: str):
+    init_test_distributed_environment(1, tensor_parallel_size, rank,
+                                      distributed_init_port)
+    num_elements = 8
+    all_tensors = [
+        torch.arange(num_elements, dtype=torch.float32, device="cuda") *
+        (r + 1) for r in range(tensor_parallel_size)
+    ]
+    expected = torch.sum(torch.stack(all_tensors, dim=0), dim=0)
+    t = all_tensors[rank]
+    t = tensor_model_parallel_all_reduce(t)
+    assert torch.allclose(t, expected)
+
+
+def all_gather_test_worker(tensor_parallel_size: int, rank: int,
+                           distributed_init_port: str):
+    init_test_distributed_environment(1, tensor_parallel_size, rank,
+                                      distributed_init_port)
+    num_dimensions = 3
+    tensor_size = list(range(2, num_dimensions + 2))
+    total_size = 1
+    for s in tensor_size:
+        total_size *= s
+    for all_gather_dimension in range(num_dimensions):
+        all_tensors = [
+            torch.arange(total_size, dtype=torch.float32,
+                         device="cuda").reshape(tensor_size) * (r + 1)
+            for r in range(tensor_parallel_size)
+        ]
+        expected = torch.cat(all_tensors, dim=all_gather_dimension)
+        t = all_tensors[rank]
+        t = tensor_model_parallel_all_gather(t, all_gather_dimension)
+        assert torch.allclose(t, expected)
+
+
+@pytest.mark.skipif(torch.cuda.device_count() < 2,
+                    reason="Need at least 2 GPUs to run the test.")
+@pytest.mark.parametrize("tensor_parallel_size", [2])
+@pytest.mark.parametrize("test_target",
+                         [all_reduce_test_worker, all_gather_test_worker])
+def test_multi_process_tensor_parallel(tensor_parallel_size, test_target):
+    set_start_method("spawn", force=True)
+    distributed_init_port = get_open_port()
+    processes = []
+    for rank in range(tensor_parallel_size):
+        p = Process(target=test_target,
+                    args=(tensor_parallel_size, rank, distributed_init_port))
+        p.start()
+        processes.append(p)
+    for p in processes:
+        p.join()
+    assert all(p.exitcode == 0 for p in processes)

tests/engine/test_detokenize.py

@@ -0,0 +1,62 @@
+import pytest
+
+from transformers import AutoTokenizer
+
+from vllm.transformers_utils.tokenizer import detokenize_incrementally
+
+TRUTH = [
+    "Hello here, this is a simple test",  # noqa: E501
+    "vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs. It is designed to be used in production environments, where inference and serving",  # noqa: E501
+    "我很感谢你的热情"  # noqa: E501
+]
+TOKENIZERS = [
+    "facebook/opt-125m",
+    "gpt2",
+    "bigcode/tiny_starcoder_py",
+    "EleutherAI/gpt-j-6b",
+    "EleutherAI/pythia-70m",
+    "bigscience/bloom-560m",
+    "mosaicml/mpt-7b",
+    "tiiuae/falcon-7b",
+    "meta-llama/Llama-2-7b-hf",
+    "codellama/CodeLlama-7b-hf",
+]
+
+
+def _run_incremental_decode(tokenizer, all_input_ids,
+                            skip_special_tokens: bool):
+    decoded_text = ""
+    offset = 0
+    token_offset = 0
+    prev_tokens = None
+    for i in range(len(all_input_ids)):
+        new_tokens, text, offset, token_offset = detokenize_incrementally(
+            tokenizer,
+            all_input_ids[:i + 1],
+            prev_tokens,
+            offset,
+            token_offset,
+            skip_special_tokens=skip_special_tokens)
+        decoded_text += text
+        if prev_tokens is None:
+            prev_tokens = new_tokens
+        else:
+            prev_tokens += new_tokens
+    return decoded_text
+
+
+@pytest.mark.parametrize("truth", TRUTH)
+@pytest.mark.parametrize("tokenizer_id", TOKENIZERS)
+@pytest.mark.parametrize("skip_special_tokens", (True, False))
+def test_decode_streaming(tokenizer_id, truth, skip_special_tokens):
+    tokenizer = AutoTokenizer.from_pretrained(tokenizer_id)
+    all_input_ids = tokenizer(truth, add_special_tokens=False)["input_ids"]
+    if skip_special_tokens:
+        all_input_ids = ([tokenizer.bos_token_id]
+                         if tokenizer.bos_token_id is not None else
+                         []) + all_input_ids + [tokenizer.eos_token_id]
+
+    decoded_text = _run_incremental_decode(
+        tokenizer, all_input_ids, skip_special_tokens=skip_special_tokens)
+
+    assert decoded_text == truth

tests/kernels/conftest.py

@@ -0,0 +1,43 @@
+from typing import List, Tuple
+
+import pytest
+import torch
+
+
+def create_kv_caches(
+    num_blocks: int,
+    block_size: int,
+    num_layers: int,
+    num_heads: int,
+    head_size: int,
+    dtype: torch.dtype,
+    seed: int,
+) -> Tuple[List[torch.Tensor], List[torch.Tensor]]:
+    torch.random.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
+
+    scale = head_size**-0.5
+    x = 16 // torch.tensor([], dtype=dtype).element_size()
+    key_cache_shape = (num_blocks, num_heads, head_size // x, block_size, x)
+    key_caches = []
+    for _ in range(num_layers):
+        key_cache = torch.empty(size=key_cache_shape,
+                                dtype=dtype,
+                                device='cuda')
+        key_cache.uniform_(-scale, scale)
+        key_caches.append(key_cache)
+
+    value_cache_shape = (num_blocks, num_heads, head_size, block_size)
+    value_caches = []
+    for _ in range(num_layers):
+        value_cache = torch.empty(size=value_cache_shape,
+                                  dtype=dtype,
+                                  device='cuda')
+        value_cache.uniform_(-scale, scale)
+        value_caches.append(value_cache)
+    return key_caches, value_caches
+
+
+@pytest.fixture()
+def kv_cache_factory():
+    return create_kv_caches

tests/kernels/test_activation.py

@@ -1,30 +1,68 @@
+import pytest
 import torch
-import torch.nn.functional as F
 
-from vllm import activation_ops
-
-
-def ref_silu_and_mul(x: torch.Tensor) -> torch.Tensor:
-    x1, x2 = x.chunk(chunks=2, dim=1)
-    return F.silu(x1) * x2
+from vllm.model_executor.layers.activation import FastGELU, NewGELU, SiluAndMul
+
+DTYPES = [torch.half, torch.bfloat16, torch.float]
+NUM_TOKENS = [7, 83, 2048]  # Arbitrary values for testing
+D = [512, 4096, 5120, 13824]  # Arbitrary values for testing
+SEEDS = [0]
 
 
+@pytest.mark.parametrize("num_tokens", NUM_TOKENS)
+@pytest.mark.parametrize("d", D)
+@pytest.mark.parametrize("dtype", DTYPES)
+@pytest.mark.parametrize("seed", SEEDS)
 @torch.inference_mode()
-def run_silu_and_mul(
+def test_silu_and_mul(
     num_tokens: int,
     d: int,
     dtype: torch.dtype,
+    seed: int,
 ) -> None:
-    x = torch.randn(num_tokens, 2 * d, dtype=dtype, device='cuda')
-    out = torch.empty(num_tokens, d, dtype=dtype, device='cuda')
-    activation_ops.silu_and_mul(out, x)
-    ref_out = ref_silu_and_mul(x)
+    torch.random.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
+    x = torch.randn(num_tokens, 2 * d, dtype=dtype, device="cuda")
+    layer = SiluAndMul()
+    out = layer(x)
+    ref_out = layer._forward(x)
     assert torch.allclose(out, ref_out, atol=1e-5, rtol=1e-5)
 
 
-def test_silu_and_mul() -> None:
-    for dtype in [torch.half, torch.bfloat16, torch.float]:
-        for num_tokens in [7, 83, 2048]:
-            for d in [512, 4096, 5120, 13824]:
-                print(f'Testing dtype={dtype}, num_tokens={num_tokens}, d={d}')
-                run_silu_and_mul(num_tokens, d, dtype)
+@pytest.mark.parametrize("num_tokens", NUM_TOKENS)
+@pytest.mark.parametrize("d", D)
+@pytest.mark.parametrize("dtype", DTYPES)
+@pytest.mark.parametrize("seed", SEEDS)
+@torch.inference_mode()
+def test_gelu_new(
+    num_tokens: int,
+    d: int,
+    dtype: torch.dtype,
+    seed: int,
+) -> None:
+    torch.random.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
+    x = torch.randn(num_tokens, d, dtype=dtype, device="cuda")
+    layer = NewGELU()
+    out = layer(x)
+    ref_out = layer._forward(x)
+    assert torch.allclose(out, ref_out, atol=1e-5, rtol=1e-5)
+
+
+@pytest.mark.parametrize("num_tokens", NUM_TOKENS)
+@pytest.mark.parametrize("d", D)
+@pytest.mark.parametrize("dtype", DTYPES)
+@pytest.mark.parametrize("seed", SEEDS)
+def test_gelu_fast(
+    num_tokens: int,
+    d: int,
+    dtype: torch.dtype,
+    seed: int,
+) -> None:
+    torch.random.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
+    x = torch.randn(num_tokens, d, dtype=dtype, device="cuda")
+    layer = FastGELU()
+    out = layer(x)
+    ref_out = layer._forward(x)
+    assert torch.allclose(out, ref_out, atol=1e-5, rtol=1e-5)

tests/kernels/test_attention.py

@@ -1,14 +1,29 @@
 import random
-from typing import List, Optional
+from typing import List, Optional, Tuple
 
+import pytest
 import torch
 from xformers import ops as xops
 from xformers.ops.fmha.attn_bias import BlockDiagonalCausalMask
 
-from vllm import attention_ops
+from vllm._C import ops
+from vllm.utils import get_max_shared_memory_bytes
 
-MAX_SEQ_LEN = 4096
-TEST_SEED = 0
+FLOAT32_BYTES = torch.finfo(torch.float).bits // 8
+# This will change depending on the compute capability.
+# - 512 as a buffer
+MAX_SEQ_LEN = get_max_shared_memory_bytes() // FLOAT32_BYTES - 512
+NUM_BLOCKS = 40000  # Arbitrary values for testing
+PARTITION_SIZE = 512
+
+DTYPES = [torch.half, torch.bfloat16, torch.float]
+NUM_GEN_SEQS = [7]  # Arbitrary values for testing
+NUM_PREFILL_SEQS = [3]  # Arbitrary values for testing
+NUM_HEADS = [(40, 40), (64, 8)]  # Arbitrary values for testing
+HEAD_SIZES = [64, 80, 96, 112, 128, 256]
+BLOCK_SIZES = [16, 32]
+USE_ALIBI = [False, True]
+SEEDS = [0]
 
 
 def ref_masked_attention(
@@ -18,29 +33,34 @@ def ref_masked_attention(
     scale: float,
     attn_mask: Optional[torch.Tensor] = None,
 ) -> torch.Tensor:
-    query = query * scale
-    attn = torch.einsum('qhd,khd->hqk', query, key)
+    attn_weights = scale * torch.einsum("qhd,khd->hqk", query, key).float()
     if attn_mask is not None:
-        attn = attn + attn_mask
-    attn = torch.softmax(attn, dim=-1)
-    out = torch.einsum('hqk,khd->qhd', attn, value)
+        attn_weights = attn_weights + attn_mask.float()
+    attn_weights = torch.softmax(attn_weights, dim=-1).to(value.dtype)
+    out = torch.einsum("hqk,khd->qhd", attn_weights, value)
     return out
 
 
 def ref_single_query_cached_kv_attention(
     output: torch.Tensor,
     query: torch.Tensor,
+    num_queries_per_kv: int,
     key_cache: torch.Tensor,
     value_cache: torch.Tensor,
     block_tables: torch.Tensor,
     context_lens: torch.Tensor,
+    scale: float,
+    alibi_slopes: Optional[torch.Tensor],
 ) -> None:
-    num_heads = value_cache.shape[1]
+    num_query_heads = query.shape[1]
+    num_kv_heads = value_cache.shape[1]
     head_size = value_cache.shape[2]
     block_size = value_cache.shape[3]
+    num_seqs = query.shape[0]
 
-    num_input_tokens = query.shape[0]
-    for i in range(num_input_tokens):
+    block_tables = block_tables.cpu().tolist()
+    context_lens = context_lens.cpu().tolist()
+    for i in range(num_seqs):
         q = query[i].unsqueeze(0)
         block_table = block_tables[i]
         context_len = int(context_lens[i])
@@ -52,30 +72,175 @@ def ref_single_query_cached_kv_attention(
             block_offset = j % block_size
 
             k = key_cache[block_number, :, :, block_offset, :]
-            k = k.reshape(num_heads, head_size)
+            k = k.reshape(num_kv_heads, head_size)
             keys.append(k)
 
             v = value_cache[block_number, :, :, block_offset]
             values.append(v)
         keys = torch.stack(keys, dim=0)
         values = torch.stack(values, dim=0)
+        if num_queries_per_kv > 1:
+            # Handle MQA and GQA
+            keys = torch.repeat_interleave(keys, num_queries_per_kv, dim=1)
+            values = torch.repeat_interleave(values, num_queries_per_kv, dim=1)
 
-        scale = 1.0 / (head_size**0.5)
-        out = ref_masked_attention(q, keys, values, scale)
-        out = out.view(num_heads, head_size)
+        alibi_bias = None
+        if alibi_slopes is not None:
+            # Create the ALiBi bias used in the paged attention kernel.
+            position_ids = torch.arange(context_len, device="cuda").int()
+            alibi_bias = (position_ids - context_len + 1).float()
+            alibi_bias = alibi_slopes.view(-1, 1, 1) * alibi_bias.view(
+                1, 1, -1)
+
+        out = ref_masked_attention(q, keys, values, scale, alibi_bias)
+        out = out.view(num_query_heads, head_size)
         output[i].copy_(out, non_blocking=True)
 
 
+@pytest.mark.parametrize("version", ["v1", "v2"])
+@pytest.mark.parametrize("num_seqs", NUM_GEN_SEQS)
+@pytest.mark.parametrize("num_heads", NUM_HEADS)
+@pytest.mark.parametrize("head_size", HEAD_SIZES)
+@pytest.mark.parametrize("use_alibi", USE_ALIBI)
+@pytest.mark.parametrize("block_size", BLOCK_SIZES)
+@pytest.mark.parametrize("dtype", DTYPES)
+@pytest.mark.parametrize("seed", SEEDS)
+def test_paged_attention(
+    kv_cache_factory,
+    version: str,
+    num_seqs: int,
+    num_heads: Tuple[int, int],
+    head_size: int,
+    use_alibi: bool,
+    block_size: int,
+    dtype: torch.dtype,
+    seed: int,
+) -> None:
+    random.seed(seed)
+    torch.random.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
+
+    scale = float(1.0 / (head_size**0.5))
+    num_query_heads, num_kv_heads = num_heads
+    query = torch.empty(num_seqs,
+                        num_query_heads,
+                        head_size,
+                        dtype=dtype,
+                        device="cuda")
+    query.uniform_(-scale, scale)
+
+    assert num_query_heads % num_kv_heads == 0
+    num_queries_per_kv = num_query_heads // num_kv_heads
+    head_mapping = torch.repeat_interleave(
+        torch.arange(num_kv_heads, dtype=torch.int32, device="cuda"),
+        num_queries_per_kv)
+    alibi_slopes = None
+    if use_alibi:
+        alibi_slopes = torch.randn(num_query_heads,
+                                   dtype=torch.float,
+                                   device="cuda")
+
+    context_lens = [random.randint(1, MAX_SEQ_LEN) for _ in range(num_seqs)]
+    context_lens[-1] = MAX_SEQ_LEN
+    max_context_len = max(context_lens)
+    context_lens = torch.tensor(context_lens, dtype=torch.int, device="cuda")
+
+    # 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="cuda")
+
+    # Create the KV caches.
+    key_caches, value_caches = kv_cache_factory(NUM_BLOCKS, block_size, 1,
+                                                num_kv_heads, head_size, dtype,
+                                                seed)
+    key_cache, value_cache = key_caches[0], value_caches[0]
+
+    # Call the paged attention kernel.
+    output = torch.empty_like(query)
+    if version == "v1":
+        ops.paged_attention_v1(
+            output,
+            query,
+            key_cache,
+            value_cache,
+            head_mapping,
+            scale,
+            block_tables,
+            context_lens,
+            block_size,
+            max_context_len,
+            alibi_slopes,
+        )
+    elif version == "v2":
+        num_partitions = ((max_context_len + PARTITION_SIZE - 1) //
+                          PARTITION_SIZE)
+        assert PARTITION_SIZE % block_size == 0
+        num_seqs, num_heads, head_size = output.shape
+        tmp_output = torch.empty(
+            size=(num_seqs, num_heads, num_partitions, head_size),
+            dtype=output.dtype,
+            device=output.device,
+        )
+        exp_sums = torch.empty(
+            size=(num_seqs, num_heads, num_partitions),
+            dtype=torch.float32,
+            device=output.device,
+        )
+        max_logits = torch.empty_like(exp_sums)
+        ops.paged_attention_v2(
+            output,
+            exp_sums,
+            max_logits,
+            tmp_output,
+            query,
+            key_cache,
+            value_cache,
+            head_mapping,
+            scale,
+            block_tables,
+            context_lens,
+            block_size,
+            max_context_len,
+            alibi_slopes,
+        )
+    else:
+        raise AssertionError(f"Unknown version: {version}")
+
+    # Run the reference implementation.
+    ref_output = torch.empty_like(query)
+    ref_single_query_cached_kv_attention(
+        ref_output,
+        query,
+        num_queries_per_kv,
+        key_cache,
+        value_cache,
+        block_tables,
+        context_lens,
+        scale,
+        alibi_slopes,
+    )
+
+    # NOTE(woosuk): Due to the kernel-level differences in the two
+    # implementations, there is a small numerical difference in the two
+    # outputs. Thus, we use a relaxed tolerance for the test.
+    assert torch.allclose(output, ref_output, atol=1e-3, rtol=1e-5)
+
+
 def ref_multi_query_kv_attention(
     cu_seq_lens: List[int],
     query: torch.Tensor,
     key: torch.Tensor,
     value: torch.Tensor,
+    scale: float,
     dtype: torch.dtype,
 ) -> torch.Tensor:
-    head_size = query.shape[-1]
-    scale = 1.0 / (head_size**0.5)
-
     num_seqs = len(cu_seq_lens) - 1
     ref_outputs = []
     for i in range(num_seqs):
@@ -87,7 +252,7 @@ def ref_multi_query_kv_attention(
         attn_mask = torch.triu(torch.ones(seq_len, seq_len, dtype=dtype),
                                diagonal=1)
         attn_mask = attn_mask * torch.finfo(dtype).min
-        attn_mask = attn_mask.to(dtype=dtype, device='cuda')
+        attn_mask = attn_mask.to(dtype=dtype, device="cuda")
 
         ref_output = ref_masked_attention(
             query[start_idx:end_idx],
@@ -101,161 +266,47 @@ def ref_multi_query_kv_attention(
     return ref_output
 
 
-def ref_multi_query_cached_kv_attention(
-    cu_query_lens: List[int],
-    query: torch.Tensor,
-    key_cache: torch.Tensor,
-    value_cache: torch.Tensor,
-    block_tables: torch.Tensor,
-    context_lens: torch.Tensor,
-    dtype: torch.dtype,
-) -> torch.Tensor:
-    num_heads = value_cache.shape[1]
-    head_size = value_cache.shape[2]
-    block_size = value_cache.shape[3]
-    scale = 1.0 / (head_size**0.5)
-
-    num_queries = len(cu_query_lens) - 1
-    ref_outputs = []
-    for i in range(num_queries):
-        start_idx = cu_query_lens[i]
-        end_idx = cu_query_lens[i + 1]
-        query_len = end_idx - start_idx
-        context_len = int(context_lens[i])
-        block_table = block_tables[i]
-
-        # Create attention mask
-        attn_mask = torch.triu(torch.ones(query_len, context_len),
-                               diagonal=context_len - query_len + 1) * -1e5
-        attn_mask = attn_mask.to(dtype=dtype, device='cuda')
-
-        keys = []
-        values = []
-        for j in range(context_len):
-            block_number = int(block_table[j // block_size])
-            block_offset = j % block_size
-
-            k = key_cache[block_number, :, :, block_offset, :]
-            k = k.reshape(num_heads, head_size)
-            keys.append(k)
-
-            v = value_cache[block_number, :, :, block_offset]
-            values.append(v)
-        keys = torch.stack(keys, dim=0)
-        values = torch.stack(values, dim=0)
-
-        ref_output = ref_masked_attention(
-            query[start_idx:end_idx],
-            keys,
-            values,
-            scale,
-            attn_mask=attn_mask,
-        )
-        ref_outputs.append(ref_output)
-    ref_output = torch.cat(ref_outputs, dim=0)
-    return ref_output
-
-
+# TODO(woosuk): Add tests for USE_ALIBI=True.
+@pytest.mark.parametrize("num_seqs", NUM_PREFILL_SEQS)
+@pytest.mark.parametrize("num_heads", NUM_HEADS)
+@pytest.mark.parametrize("head_size", HEAD_SIZES)
+@pytest.mark.parametrize("dtype", DTYPES)
+@pytest.mark.parametrize("seed", SEEDS)
 @torch.inference_mode()
-def run_single_query_cached_kv_attention(
-    num_tokens: int,
-    num_heads: int,
-    head_size: int,
-    block_size: int,
-    num_blocks: int,
-    dtype: torch.dtype,
-) -> None:
-    qkv = torch.empty(num_tokens,
-                      3,
-                      num_heads,
-                      head_size,
-                      dtype=dtype,
-                      device='cuda')
-    qkv.uniform_(-1e-3, 1e-3)
-    query, _, _ = qkv.unbind(dim=1)
-
-    x = 16 // torch.tensor([], dtype=dtype).element_size()
-    key_block_shape = (num_heads, head_size // x, block_size, x)
-    key_cache = torch.empty(size=(num_blocks, *key_block_shape),
-                            dtype=dtype,
-                            device='cuda')
-    key_cache.uniform_(-1e-3, 1e-3)
-    value_block_shape = (num_heads, head_size, block_size)
-    value_cache = torch.empty(size=(num_blocks, *value_block_shape),
-                              dtype=dtype,
-                              device='cuda')
-    value_cache.uniform_(-1e-3, 1e-3)
-
-    context_lens = [random.randint(1, MAX_SEQ_LEN) for _ in range(num_tokens)]
-    max_context_len = max(context_lens)
-    context_lens = torch.tensor(context_lens, dtype=torch.int, device='cuda')
-
-    max_num_blocks_per_seq = (max_context_len + block_size - 1) // block_size
-    block_tables = []
-    for _ in range(num_tokens):
-        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')
-
-    scale = float(1.0 / (head_size**0.5))
-    output = torch.empty(num_tokens,
-                         num_heads,
-                         head_size,
-                         dtype=dtype,
-                         device='cuda')
-    attention_ops.single_query_cached_kv_attention(
-        output,
-        query,
-        key_cache,
-        value_cache,
-        scale,
-        block_tables,
-        context_lens,
-        block_size,
-        max_context_len,
-        None,  # ALiBi slopes.
-    )
-
-    ref_output = torch.empty_like(query)
-    ref_single_query_cached_kv_attention(
-        ref_output,
-        query,
-        key_cache,
-        value_cache,
-        block_tables,
-        context_lens,
-    )
-    # NOTE(woosuk): Due to the difference in the data types the two
-    # implementations use for attention softmax logits and accumulation,
-    # there is a small difference in the final outputs.
-    # We should use a relaxed tolerance for the test.
-    assert torch.allclose(output, ref_output, atol=1e-3, rtol=1e-5)
-
-
-@torch.inference_mode()
-def run_multi_query_kv_attention(
+def test_multi_query_kv_attention(
     num_seqs: int,
-    num_heads: int,
+    num_heads: Tuple[int, int],
     head_size: int,
     dtype: torch.dtype,
+    seed: int,
 ) -> None:
-    seq_lens = random.sample(range(1, MAX_SEQ_LEN), num_seqs)
+    random.seed(seed)
+    torch.random.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
+
+    # MAX_SEQ_LEN sometimes causes OOM in the reference implementation.
+    # As the xformers library is already tested with its own tests, we can use
+    # a smaller MAX_SEQ_LEN here.
+    max_len = min(MAX_SEQ_LEN, 4096)
+    seq_lens = random.sample(range(1, max_len), num_seqs)
     num_tokens = sum(seq_lens)
 
     scale = float(1.0 / (head_size**0.5))
+    num_query_heads, num_kv_heads = num_heads
     qkv = torch.empty(num_tokens,
-                      3,
-                      num_heads,
+                      num_query_heads + 2 * num_kv_heads,
                       head_size,
                       dtype=dtype,
-                      device='cuda')
-    qkv.uniform_(-1e-3, 1e-3)
-    query, key, value = qkv.unbind(dim=1)
+                      device="cuda")
+    qkv.uniform_(-scale, scale)
+    query, key, value = qkv.split(
+        [num_query_heads, num_kv_heads, num_kv_heads], dim=1)
 
-    attn_op = xops.fmha.cutlass.FwOp()
+    num_queries_per_kv = num_query_heads // num_kv_heads
+    if num_queries_per_kv > 1:
+        # Handle MQA and GQA
+        key = torch.repeat_interleave(key, num_queries_per_kv, dim=1)
+        value = torch.repeat_interleave(value, num_queries_per_kv, dim=1)
     attn_bias = BlockDiagonalCausalMask.from_seqlens(seq_lens)
     output = xops.memory_efficient_attention_forward(
         query.unsqueeze(0),
@@ -264,7 +315,6 @@ def run_multi_query_kv_attention(
         attn_bias=attn_bias,
         p=0.0,
         scale=scale,
-        op=attn_op,
     )
     output = output.squeeze(0)
 
@@ -276,40 +326,7 @@ def run_multi_query_kv_attention(
         query,
         key,
         value,
+        scale,
         dtype,
     )
     assert torch.allclose(output, ref_output, atol=1e-3, rtol=1e-5)
-
-
-def test_single_query_cached_kv_attention() -> None:
-    torch.random.manual_seed(TEST_SEED)
-    torch.cuda.manual_seed(TEST_SEED)
-    for dtype in [torch.half, torch.bfloat16, torch.float]:
-        for block_size in [8, 16, 32]:
-            for head_size in [64, 80, 96, 128]:
-                print(f'Testing single_query_cached_kv_attention with '
-                      f'dtype={dtype}, block_size={block_size}, '
-                      f'head_size={head_size}')
-                run_single_query_cached_kv_attention(
-                    num_tokens=37,
-                    num_heads=3,
-                    head_size=head_size,
-                    block_size=block_size,
-                    num_blocks=1024,
-                    dtype=dtype,
-                )
-
-
-def test_multi_query_kv_attention() -> None:
-    torch.random.manual_seed(TEST_SEED)
-    torch.cuda.manual_seed(TEST_SEED)
-    for dtype in [torch.half, torch.bfloat16, torch.float]:
-        for head_size in [64, 80, 96, 128]:
-            print(f'Testing multi_query_kv_attention with dtype={dtype}, '
-                  f'head_size={head_size}')
-            run_multi_query_kv_attention(
-                num_seqs=5,
-                num_heads=3,
-                head_size=head_size,
-                dtype=dtype,
-            )

tests/kernels/test_cache.py

@@ -1,12 +1,32 @@
 import random
 
+import pytest
 import torch
 
-from vllm import cache_ops
+from vllm._C import cache_ops
+
+DTYPES = [torch.half, torch.bfloat16, torch.float]
+NUM_TOKENS = [83]  # Arbitrary values for testing
+NUM_LAYERS = [1]  # Arbitrary values for testing
+NUM_HEADS = [8]  # Arbitrary values for testing
+HEAD_SIZES = [64, 80, 96, 112, 128, 256]
+BLOCK_SIZES = [8, 16, 32]
+NUM_BLOCKS = [1024, 36000]  # Arbitrary values for testing
+NUM_MAPPINGS = [256]  # Arbitrary values for testing
+SEEDS = [0]
 
 
+@pytest.mark.parametrize("num_mappings", NUM_MAPPINGS)
+@pytest.mark.parametrize("num_layers", NUM_LAYERS)
+@pytest.mark.parametrize("num_heads", NUM_HEADS)
+@pytest.mark.parametrize("head_size", HEAD_SIZES)
+@pytest.mark.parametrize("block_size", BLOCK_SIZES)
+@pytest.mark.parametrize("num_blocks", NUM_BLOCKS)
+@pytest.mark.parametrize("dtype", DTYPES)
+@pytest.mark.parametrize("seed", SEEDS)
 @torch.inference_mode()
-def run_copy_blocks(
+def test_copy_blocks(
+    kv_cache_factory,
     num_mappings: int,
     num_layers: int,
     num_heads: int,
@@ -14,49 +34,44 @@ def run_copy_blocks(
     block_size: int,
     num_blocks: int,
     dtype: torch.dtype,
+    seed: int,
 ) -> None:
-    # Generate random block mappings.
+    random.seed(seed)
+    torch.random.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
+
+    # Generate random block mappings where each source block is mapped to two
+    # destination blocks.
+    assert 2 * num_mappings <= num_blocks
     src_blocks = random.sample(range(num_blocks), num_mappings)
     remainig_blocks = list(set(range(num_blocks)) - set(src_blocks))
-    dst_blocks = random.sample(remainig_blocks, num_mappings)
-    block_mapping = {src: [dst] for src, dst in zip(src_blocks, dst_blocks)}
+    dst_blocks = random.sample(remainig_blocks, 2 * num_mappings)
+    block_mapping = {}
+    for i in range(num_mappings):
+        src = src_blocks[i]
+        dst1 = dst_blocks[2 * i]
+        dst2 = dst_blocks[2 * i + 1]
+        block_mapping[src] = [dst1, dst2]
 
-    # Create the KV cache.
-    x = 16 // torch.tensor([], dtype=dtype).element_size()
-    key_cache_shape = (num_blocks, num_heads, head_size // x, block_size, x)
-    key_caches = []
-    for _ in range(num_layers):
-        key_cache = torch.randn(size=key_cache_shape,
-                                dtype=dtype,
-                                device='cuda')
-        key_caches.append(key_cache)
-    cloned_key_caches = []
-    for key_cache in key_caches:
-        cloned_key_caches.append(key_cache.clone())
+    # Create the KV caches.
+    key_caches, value_caches = kv_cache_factory(num_blocks, block_size,
+                                                num_layers, num_heads,
+                                                head_size, dtype, seed)
 
-    value_cache_shape = (num_blocks, num_heads, head_size, block_size)
-    value_caches = []
-    for _ in range(num_layers):
-        value_cache = torch.randn(size=value_cache_shape,
-                                  dtype=dtype,
-                                  device='cuda')
-        value_caches.append(value_cache)
-    cloned_value_caches = []
-    for value_cache in value_caches:
-        cloned_value_caches.append(value_cache.clone())
+    # Clone the KV caches.
+    cloned_key_caches = [key_cache.clone() for key_cache in key_caches]
+    cloned_value_caches = [value_cache.clone() for value_cache in value_caches]
 
     # Call the copy blocks kernel.
     cache_ops.copy_blocks(key_caches, value_caches, block_mapping)
 
-    # Reference implementation.
+    # Run the reference implementation.
     for src, dsts in block_mapping.items():
         for dst in dsts:
-            for key_cache, cloned_key_cache in zip(key_caches,
-                                                   cloned_key_caches):
-                cloned_key_cache[dst] = cloned_key_cache[src]
-            for value_cache, cloned_value_cache in zip(value_caches,
-                                                       cloned_value_caches):
-                cloned_value_cache[dst] = cloned_value_cache[src]
+            for cloned_key_cache in cloned_key_caches:
+                cloned_key_cache[dst].copy_(cloned_key_cache[src])
+            for cloned_value_cache in cloned_value_caches:
+                cloned_value_cache[dst].copy_(cloned_value_cache[src])
 
     # Compare the results.
     for key_cache, cloned_key_cache in zip(key_caches, cloned_key_caches):
@@ -66,131 +81,66 @@ def run_copy_blocks(
         assert torch.allclose(value_cache, cloned_value_cache)
 
 
+@pytest.mark.parametrize("num_tokens", NUM_TOKENS)
+@pytest.mark.parametrize("num_heads", NUM_HEADS)
+@pytest.mark.parametrize("head_size", HEAD_SIZES)
+@pytest.mark.parametrize("block_size", BLOCK_SIZES)
+@pytest.mark.parametrize("num_blocks", NUM_BLOCKS)
+@pytest.mark.parametrize("dtype", DTYPES)
+@pytest.mark.parametrize("seed", SEEDS)
 @torch.inference_mode()
-def run_reshape_and_cache(
+def test_reshape_and_cache(
+    kv_cache_factory,
     num_tokens: int,
     num_heads: int,
     head_size: int,
     block_size: int,
     num_blocks: int,
     dtype: torch.dtype,
+    seed: int,
 ) -> None:
+    random.seed(seed)
+    torch.random.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
+
+    # Create a random slot mapping.
     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')
+    slot_mapping = torch.tensor(slot_mapping, dtype=torch.long, device="cuda")
 
     qkv = torch.randn(num_tokens,
                       3,
                       num_heads,
                       head_size,
                       dtype=dtype,
-                      device='cuda')
+                      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')
-    cloned_key_cache = key_cache.clone()
+    # Create the KV caches.
+    key_caches, value_caches = kv_cache_factory(num_blocks, block_size, 1,
+                                                num_heads, head_size, dtype,
+                                                seed)
+    key_cache, value_cache = key_caches[0], value_caches[0]
 
-    value_cache_shape = (num_blocks, num_heads, head_size, block_size)
-    value_cache = torch.randn(size=value_cache_shape,
-                              dtype=dtype,
-                              device='cuda')
+    # Clone the KV caches.
+    cloned_key_cache = key_cache.clone()
     cloned_value_cache = value_cache.clone()
 
+    # Call the reshape_and_cache kernel.
     cache_ops.reshape_and_cache(key, value, key_cache, value_cache,
                                 slot_mapping)
 
+    # Run the reference implementation.
+    reshaped_key = key.reshape(num_tokens, *key_cache[0, :, :, 0, :].shape)
+    block_indicies = torch.div(slot_mapping, block_size, rounding_mode="floor")
+    block_indicies = block_indicies.cpu().tolist()
+    block_offsets = slot_mapping % block_size
+    block_offsets = block_offsets.cpu().tolist()
     for i in range(num_tokens):
-        reshaped_key = key.reshape(num_tokens, num_heads, head_size // x, x)
-        block_idx = torch.div(slot_mapping[i],
-                              block_size,
-                              rounding_mode='floor')
-        block_offset = slot_mapping[i] % block_size
+        block_idx = block_indicies[i]
+        block_offset = block_offsets[i]
         cloned_key_cache[block_idx, :, :, block_offset, :] = reshaped_key[i]
         cloned_value_cache[block_idx, :, :, block_offset] = value[i]
 
     assert torch.allclose(key_cache, cloned_key_cache)
     assert torch.allclose(value_cache, cloned_value_cache)
-
-
-@torch.inference_mode()
-def run_gather_cached_kv(
-    num_tokens: int,
-    num_heads: int,
-    head_size: int,
-    block_size: int,
-    num_blocks: int,
-    dtype: torch.dtype,
-) -> None:
-    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)
-
-    qkv_clone = qkv.clone()
-    _, cloned_key, cloned_value = qkv_clone.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')
-
-    cache_ops.gather_cached_kv(key, value, key_cache, value_cache,
-                               slot_mapping)
-
-    # Reference implementation.
-    for i in range(num_tokens):
-        reshaped_key = cloned_key.reshape(num_tokens, num_heads,
-                                          head_size // x, x)
-        block_idx = torch.div(slot_mapping[i],
-                              block_size,
-                              rounding_mode='floor')
-        block_offset = slot_mapping[i] % block_size
-        reshaped_key[i] = key_cache[block_idx, :, :, block_offset, :]
-        cloned_value[i] = value_cache[block_idx, :, :, block_offset]
-
-    assert torch.allclose(key, cloned_key)
-    assert torch.allclose(value, cloned_value)
-
-
-def test_copy_blocks() -> None:
-    for dtype in [torch.half, torch.bfloat16, torch.float]:
-        run_copy_blocks(num_mappings=23,
-                        num_layers=7,
-                        num_heads=17,
-                        head_size=16,
-                        block_size=8,
-                        num_blocks=1024,
-                        dtype=dtype)
-
-
-def test_reshape_and_cache() -> None:
-    for dtype in [torch.half, torch.bfloat16, torch.float]:
-        run_reshape_and_cache(num_tokens=3,
-                              num_heads=2,
-                              head_size=16,
-                              block_size=8,
-                              num_blocks=2,
-                              dtype=dtype)
-
-
-def test_gather_cached_kv() -> None:
-    for dtype in [torch.half, torch.bfloat16, torch.float]:
-        run_gather_cached_kv(num_tokens=3,
-                             num_heads=2,
-                             head_size=16,
-                             block_size=8,
-                             num_blocks=2,
-                             dtype=dtype)

tests/kernels/test_layernorm.py

@@ -1,56 +1,47 @@
+import pytest
 import torch
-import torch.nn as nn
 
-from vllm import layernorm_ops
-
-
-class RefRMSNorm(nn.Module):
-
-    def __init__(self, hidden_size, eps=1e-6):
-        super().__init__()
-        weight = torch.empty(hidden_size)
-        weight.uniform_(-1e-3, 1e-3)
-        self.weight = nn.Parameter(weight)
-        self.variance_epsilon = eps
-
-    def forward(self, hidden_states):
-        variance = hidden_states.to(torch.float32).pow(2).mean(-1,
-                                                               keepdim=True)
-        hidden_states = hidden_states * torch.rsqrt(variance +
-                                                    self.variance_epsilon)
-        if self.weight.dtype in [torch.half, torch.float16, torch.bfloat16]:
-            hidden_states = hidden_states.to(self.weight.dtype)
-        return self.weight * hidden_states
+from vllm.model_executor.layers.layernorm import RMSNorm
+
+DTYPES = [torch.half, torch.bfloat16, torch.float]
+NUM_TOKENS = [7, 83, 4096]  # Arbitrary values for testing
+HIDDEN_SIZES = [768, 5120, 8192]  # Arbitrary values for testing
+ADD_RESIDUAL = [False, True]
+SEEDS = [0]
 
 
+@pytest.mark.parametrize("num_tokens", NUM_TOKENS)
+@pytest.mark.parametrize("hidden_size", HIDDEN_SIZES)
+@pytest.mark.parametrize("add_residual", ADD_RESIDUAL)
+@pytest.mark.parametrize("dtype", DTYPES)
+@pytest.mark.parametrize("seed", SEEDS)
 @torch.inference_mode()
-def run_rms_norm(
+def test_rms_norm(
     num_tokens: int,
     hidden_size: int,
+    add_residual: bool,
     dtype: torch.dtype,
+    seed: int,
 ) -> None:
-    x = torch.randn(num_tokens, hidden_size, dtype=dtype, device='cuda')
-    ref = RefRMSNorm(hidden_size).to(dtype).cuda()
+    torch.random.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
 
-    out = torch.empty_like(x)
-    layernorm_ops.rms_norm(
-        out,
-        x,
-        ref.weight.data,
-        ref.variance_epsilon,
-    )
-    ref_out = ref(x)
-    assert torch.allclose(out, ref_out, atol=1e-3, rtol=1e-5)
+    layer = RMSNorm(hidden_size).to(dtype).cuda()
+    layer.weight.data.normal_(mean=1.0, std=0.1)
+    scale = 1 / (2 * hidden_size)
+    x = torch.randn(num_tokens, hidden_size, dtype=dtype, device="cuda")
+    x *= scale
+    residual = torch.randn_like(x) * scale if add_residual else None
 
-
-def test_rms_norm() -> None:
-    for dtype in [torch.half, torch.bfloat16, torch.float]:
-        for num_tokens in [7, 128, 2048]:
-            for hidden_size in [13, 64, 1024, 5120]:
-                print(f'Testing RMS kernel with dtype={dtype}, num_tokens='
-                      f'{num_tokens}, hidden_size={hidden_size}')
-                run_rms_norm(
-                    num_tokens=num_tokens,
-                    hidden_size=hidden_size,
-                    dtype=dtype,
-                )
+    # NOTE(woosuk): The reference implementation should be executed first
+    # because the custom kernel is in-place.
+    ref_out = layer._forward(x, residual)
+    out = layer(x, residual)
+    # NOTE(woosuk): LayerNorm operators (including RMS) typically have larger
+    # numerical errors than other operators because they involve reductions.
+    # Therefore, we use a larger tolerance.
+    if add_residual:
+        assert torch.allclose(out[0], ref_out[0], atol=1e-2, rtol=1e-2)
+        assert torch.allclose(out[1], ref_out[1], atol=1e-2, rtol=1e-2)
+    else:
+        assert torch.allclose(out, ref_out, atol=1e-2, rtol=1e-2)

tests/kernels/test_pos_encoding.py

@@ -1,147 +1,65 @@
-from typing import Tuple
+from typing import Optional
 
+import pytest
 import torch
-import torch.nn as nn
-import torch.nn.functional as F
 
-from vllm import pos_encoding_ops
-
-
-def rotate_half(x: torch.Tensor) -> torch.Tensor:
-    x1 = x[..., :x.shape[-1] // 2]
-    x2 = x[..., x.shape[-1] // 2:]
-    return torch.cat((-x2, x1), dim=-1)
-
-
-def apply_rotary_pos_emb(
-    q: torch.Tensor,
-    k: torch.Tensor,
-    cos: torch.Tensor,
-    sin: torch.Tensor,
-) -> Tuple[torch.Tensor, torch.Tensor]:
-    q_embed = (q * cos) + (rotate_half(q) * sin)
-    k_embed = (k * cos) + (rotate_half(k) * sin)
-    return q_embed, k_embed
-
-
-class RefRotaryEmbeddingNeox(nn.Module):
-    """Reference implementation of the GPT-NeoX style rotary embedding."""
-
-    def __init__(
-        self,
-        dim: int,
-        max_position_embeddings: int = 2048,
-        base: int = 10000,
-    ) -> None:
-        super().__init__()
-        self.rotary_dim = dim
-        self.max_position_embeddings = max_position_embeddings
-
-        # Create cos and sin embeddings.
-        inv_freq = 1.0 / (base**(torch.arange(0, dim, 2) / dim))
-        t = torch.arange(max_position_embeddings).float()
-        freqs = torch.einsum("i,j->ij", t, inv_freq.float())
-        emb = torch.cat((freqs, freqs), dim=-1)
-        cos = emb.cos().to(dtype=inv_freq.dtype)
-        sin = emb.sin().to(dtype=inv_freq.dtype)
-        self.register_buffer("cos_cached", cos, persistent=False)
-        self.register_buffer("sin_cached", sin, persistent=False)
-
-    def forward(
-        self,
-        positions: torch.Tensor,  # [num_tokens]
-        query: torch.Tensor,  # [num_tokens, num_heads, head_size]
-        key: torch.Tensor,  # [num_tokens, num_heads, head_size]
-    ) -> Tuple[torch.Tensor, torch.Tensor]:
-
-        query_rot = query[..., :self.rotary_dim]
-        query_pass = query[..., self.rotary_dim:]
-        key_rot = key[..., :self.rotary_dim]
-        key_pass = key[..., self.rotary_dim:]
-
-        query_rot = query_rot.transpose(0, 1)
-        key_rot = key_rot.transpose(0, 1)
-        cos = F.embedding(positions, self.cos_cached)
-        sin = F.embedding(positions, self.sin_cached)
-        query_rot, key_rot = apply_rotary_pos_emb(query_rot, key_rot, cos, sin)
-        query_rot = query_rot.transpose(0, 1).contiguous()
-        key_rot = key_rot.transpose(0, 1).contiguous()
-
-        query = torch.cat((query_rot, query_pass), dim=-1)
-        key = torch.cat((key_rot, key_pass), dim=-1)
-
-        # Output query/key shape: [num_tokens, num_tokens, head_size]
-        return query, key
+from vllm.model_executor.layers.rotary_embedding import get_rope
+
+IS_NEOX_STYLE = [True, False]
+DTYPES = [torch.half, torch.bfloat16, torch.float]
+HEAD_SIZES = [64, 80, 96, 112, 128, 256]
+ROTARY_DIMS = [None, 32]  # None means rotary dim == head size
+NUM_HEADS = [7, 17]  # Arbitrary values for testing
+BATCH_SIZES = [1, 5]  # Arbitrary values for testing
+SEQ_LENS = [11, 8192]  # Arbitrary values for testing
+SEEDS = [0]
 
 
+@pytest.mark.parametrize("is_neox_style", IS_NEOX_STYLE)
+@pytest.mark.parametrize("batch_size", BATCH_SIZES)
+@pytest.mark.parametrize("seq_len", SEQ_LENS)
+@pytest.mark.parametrize("num_heads", NUM_HEADS)
+@pytest.mark.parametrize("head_size", HEAD_SIZES)
+@pytest.mark.parametrize("rotary_dim", ROTARY_DIMS)
+@pytest.mark.parametrize("dtype", DTYPES)
+@pytest.mark.parametrize("seed", SEEDS)
 @torch.inference_mode()
-def run_rotary_embedding_neox(
-    num_tokens: int,
+def test_rotary_embedding(
+    is_neox_style: bool,
+    batch_size: int,
+    seq_len: int,
     num_heads: int,
     head_size: int,
-    max_position: int,
-    rotary_dim: int,
+    rotary_dim: Optional[int],
     dtype: torch.dtype,
+    seed: int,
+    max_position: int = 8192,
     base: int = 10000,
 ) -> None:
-    positions = torch.randint(0, max_position, (num_tokens, ), device='cuda')
-    query = torch.randn(num_tokens,
+    if rotary_dim is None:
+        rotary_dim = head_size
+    torch.random.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
+
+    if rotary_dim is None:
+        rotary_dim = head_size
+    rope = get_rope(head_size, rotary_dim, max_position, base, is_neox_style)
+    rope = rope.to(dtype).cuda()
+
+    positions = torch.randint(0,
+                              max_position, (batch_size, seq_len),
+                              device="cuda")
+    query = torch.randn(batch_size,
+                        seq_len,
                         num_heads * head_size,
                         dtype=dtype,
-                        device='cuda')
-    key = torch.randn(num_tokens,
-                      num_heads * head_size,
-                      dtype=dtype,
-                      device='cuda')
-
-    # Create the rotary embedding.
-    inv_freq = 1.0 / (base**(torch.arange(0, rotary_dim, 2) / rotary_dim))
-    t = torch.arange(max_position).float()
-    freqs = torch.einsum('i,j -> ij', t, inv_freq.float())
-    cos = freqs.cos()
-    sin = freqs.sin()
-    cos_sin_cache = torch.cat((cos, sin), dim=-1)
-    cos_sin_cache = cos_sin_cache.to(dtype=dtype, device='cuda')
-
-    # Run the kernel. The kernel is in-place, so we need to clone the inputs.
-    out_query = query.clone()
-    out_key = key.clone()
-    pos_encoding_ops.rotary_embedding_neox(
-        positions,
-        out_query,
-        out_key,
-        head_size,
-        cos_sin_cache,
-    )
-
-    # Run the reference implementation.
-    ref_rotary_embedding = RefRotaryEmbeddingNeox(
-        dim=rotary_dim,
-        max_position_embeddings=max_position,
-        base=base,
-    ).to(dtype=dtype, device='cuda')
-    ref_query, ref_key = ref_rotary_embedding(
-        positions,
-        query.view(num_tokens, num_heads, head_size),
-        key.view(num_tokens, num_heads, head_size),
-    )
-    ref_query = ref_query.view(num_tokens, num_heads * head_size)
-    ref_key = ref_key.view(num_tokens, num_heads * head_size)
+                        device="cuda")
+    key = torch.randn_like(query)
 
+    # NOTE(woosuk): The reference implementation should be executed first
+    # because the custom kernel is in-place.
+    ref_query, ref_key = rope._forward(positions, query, key)
+    out_query, out_key = rope.forward(positions, query, key)
     # Compare the results.
-    assert torch.allclose(out_query, ref_query, atol=1e-3, rtol=1e-5)
-    assert torch.allclose(out_key, ref_key, atol=1e-3, rtol=1e-5)
-
-
-def test_rotary_embedding_neox() -> None:
-    for dtype in [torch.half, torch.bfloat16, torch.float]:
-        for head_size in [32, 64, 80, 96, 128, 160, 192, 256]:
-            print(f'Running tests for head_size={head_size} and dtype={dtype}')
-            run_rotary_embedding_neox(
-                num_tokens=2145,
-                num_heads=5,
-                head_size=head_size,
-                max_position=8192,
-                rotary_dim=head_size,
-                dtype=dtype,
-            )
+    assert torch.allclose(out_query, ref_query, atol=1e-5, rtol=1e-5)
+    assert torch.allclose(out_key, ref_key, atol=1e-5, rtol=1e-5)

tests/models/test_models.py

@@ -0,0 +1,47 @@
+"""Compare the outputs of HF and vLLM when using greedy sampling.
+
+Run `pytest tests/models/test_models.py --forked`.
+"""
+import pytest
+
+MODELS = [
+    "facebook/opt-125m",
+    "meta-llama/Llama-2-7b-hf",
+    "mistralai/Mistral-7B-v0.1",
+    "tiiuae/falcon-7b",
+    "gpt2",
+    "bigcode/tiny_starcoder_py",
+    "EleutherAI/gpt-j-6b",
+    "EleutherAI/pythia-70m",
+    "bigscience/bloom-560m",
+    "mosaicml/mpt-7b",
+    "microsoft/phi-1_5",
+]
+
+
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("dtype", ["half"])
+@pytest.mark.parametrize("max_tokens", [128])
+def test_models(
+    hf_runner,
+    vllm_runner,
+    example_prompts,
+    model: str,
+    dtype: str,
+    max_tokens: int,
+) -> None:
+    hf_model = hf_runner(model, dtype=dtype)
+    hf_outputs = hf_model.generate_greedy(example_prompts, max_tokens)
+    del hf_model
+
+    vllm_model = vllm_runner(model, dtype=dtype)
+    vllm_outputs = vllm_model.generate_greedy(example_prompts, max_tokens)
+    del vllm_model
+
+    for i in range(len(example_prompts)):
+        hf_output_ids, hf_output_str = hf_outputs[i]
+        vllm_output_ids, vllm_output_str = vllm_outputs[i]
+        assert hf_output_str == vllm_output_str, (
+            f"Test{i}:\nHF: {hf_output_str!r}\nvLLM: {vllm_output_str!r}")
+        assert hf_output_ids == vllm_output_ids, (
+            f"Test{i}:\nHF: {hf_output_ids}\nvLLM: {vllm_output_ids}")

tests/samplers/test_beam_search.py

@@ -0,0 +1,46 @@
+"""Compare the outputs of HF and vLLM when using beam search.
+
+Run `pytest tests/samplers/test_beam_search.py --forked`.
+"""
+import pytest
+
+# FIXME(zhuohan): The test can not pass if we:
+#   1. Increase max_tokens to 256.
+#   2. Increase beam_width to 8.
+#   3. Use the model "huggyllama/llama-7b".
+MAX_TOKENS = [128]
+BEAM_WIDTHS = [4]
+MODELS = ["facebook/opt-125m"]
+
+
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("dtype", ["half"])
+@pytest.mark.parametrize("max_tokens", MAX_TOKENS)
+@pytest.mark.parametrize("beam_width", BEAM_WIDTHS)
+def test_beam_search_single_input(
+    hf_runner,
+    vllm_runner,
+    example_prompts,
+    model: str,
+    dtype: str,
+    max_tokens: int,
+    beam_width: int,
+) -> None:
+    hf_model = hf_runner(model, dtype=dtype)
+    hf_outputs = hf_model.generate_beam_search(example_prompts, beam_width,
+                                               max_tokens)
+    del hf_model
+
+    vllm_model = vllm_runner(model, dtype=dtype)
+    vllm_outputs = vllm_model.generate_beam_search(example_prompts, beam_width,
+                                                   max_tokens)
+    del vllm_model
+
+    for i in range(len(example_prompts)):
+        hf_output_ids, _ = hf_outputs[i]
+        vllm_output_ids, _ = vllm_outputs[i]
+        assert len(hf_output_ids) == len(vllm_output_ids)
+        for j in range(len(hf_output_ids)):
+            assert hf_output_ids[j] == vllm_output_ids[j], (
+                f"Test{i} output{j}:\nHF: {hf_output_ids}\n"
+                f"vLLM: {vllm_output_ids}")

tests/samplers/test_logprobs.py

@@ -0,0 +1,55 @@
+import pytest
+import torch
+
+from vllm import SamplingParams
+
+MODELS = ["facebook/opt-125m"]
+
+
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("dtype", ["half"])
+def test_get_prompt_logprobs(
+    hf_runner,
+    vllm_runner,
+    model,
+    dtype,
+    example_prompts,
+):
+    max_tokens = 5
+    hf_model = hf_runner(model, dtype=dtype)
+    hf_logprobs = hf_model.generate_greedy_logprobs(
+        example_prompts,
+        max_tokens=max_tokens,
+    )
+    del hf_model
+
+    vllm_model = vllm_runner(model, dtype=dtype)
+    vllm_sampling_params = SamplingParams(max_tokens=max_tokens,
+                                          logprobs=5,
+                                          prompt_logprobs=5,
+                                          temperature=0.0)
+    vllm_results = vllm_model.model.generate(
+        example_prompts, sampling_params=vllm_sampling_params)
+
+    # Test whether logprobs are included in the results.
+    for result in vllm_results:
+        assert result.prompt_logprobs is not None
+        assert result.outputs[0].logprobs is not None
+
+    # Test whether prompt logprobs are consistent with HF
+    for vllm_result, hf_logprob in zip(vllm_results, hf_logprobs):
+        # Check prompt logprobs
+        vllm_prompt_logprobs = vllm_result.prompt_logprobs[1:]
+        for i, vllm_prompt_logprob_dict in enumerate(vllm_prompt_logprobs):
+            for token_id, logprob in vllm_prompt_logprob_dict.items():
+                torch.testing.assert_close(logprob,
+                                           hf_logprob[0][i][token_id].item(),
+                                           atol=1e-2,
+                                           rtol=1e-2)
+        vllm_sample_logprobs = vllm_result.outputs[0].logprobs
+        for i, vllm_sample_logprob_dict in enumerate(vllm_sample_logprobs):
+            for token_id, logprob in vllm_sample_logprob_dict.items():
+                torch.testing.assert_close(logprob,
+                                           hf_logprob[i][-1][token_id].item(),
+                                           atol=1e-2,
+                                           rtol=1e-2)

tests/samplers/test_sampler.py

@@ -0,0 +1,235 @@
+import random
+from typing import Tuple
+from unittest.mock import patch
+
+import pytest
+import torch
+
+from vllm.model_executor.layers.sampler import Sampler
+from vllm.model_executor.utils import set_random_seed
+from vllm.sequence import SamplingParams, SequenceData, SequenceGroupMetadata
+from vllm.worker.model_runner import ModelRunner
+
+
+class MockLogitsSampler(Sampler):
+
+    def __init__(self, vocab_size: int, fake_logits: torch.Tensor):
+        super().__init__(vocab_size=vocab_size)
+        self.fake_logits = fake_logits
+
+    def forward(self, *args, **kwargs):
+        with patch("vllm.model_executor.layers.sampler._prune_hidden_states",
+                   lambda x, y: x), patch(
+                       "vllm.model_executor.layers.sampler._get_logits",
+                       lambda *args, **kwargs: self.fake_logits):
+            return super().forward(*args, **kwargs)
+
+
+def _prepare_test(
+    batch_size: int
+) -> Tuple[torch.Tensor, torch.Tensor, MockLogitsSampler, ModelRunner]:
+    vocab_size = 32000
+    input_tensor = torch.rand((batch_size, 1024),
+                              device="cuda",
+                              dtype=torch.float16)
+    fake_logits = torch.full((batch_size, vocab_size),
+                             1e-2,
+                             device=input_tensor.device,
+                             dtype=input_tensor.dtype)
+    sampler = MockLogitsSampler(32000, fake_logits)
+    model_runner = ModelRunner(None, None, None)
+    return input_tensor, fake_logits, sampler, model_runner
+
+
+RANDOM_SEEDS = list(range(128))
+
+
+@pytest.mark.parametrize("seed", RANDOM_SEEDS)
+def test_sampler_all_greedy(seed: int):
+    set_random_seed(seed)
+    batch_size = random.randint(1, 256)
+    input_tensor, fake_logits, sampler, model_runner = _prepare_test(
+        batch_size)
+
+    seq_group_metadata_list = []
+    prompt_lens = []
+    for i in range(batch_size):
+        seq_group_metadata_list.append(
+            SequenceGroupMetadata(
+                request_id=f"test_{i}",
+                is_prompt=True,
+                seq_data={0: SequenceData([1, 2, 3])},
+                sampling_params=SamplingParams(temperature=0, ),
+                block_tables={0: [1]},
+            ))
+        prompt_lens.append(seq_group_metadata_list[-1].seq_data[0].get_len())
+
+    sampling_metadata = model_runner._prepare_sample(seq_group_metadata_list,
+                                                     prompt_lens)
+    sampler_output = sampler(embedding=None,
+                             hidden_states=input_tensor,
+                             sampling_metadata=sampling_metadata)
+    expected = torch.argmax(fake_logits, dim=-1)
+    for i, sequence_output in enumerate(sampler_output):
+        for nth_output in sequence_output.samples:
+            assert nth_output.output_token == expected[i].item()
+
+
+@pytest.mark.parametrize("seed", RANDOM_SEEDS)
+def test_sampler_all_random(seed: int):
+    set_random_seed(seed)
+    batch_size = random.randint(1, 256)
+    input_tensor, fake_logits, sampler, model_runner = _prepare_test(
+        batch_size)
+
+    for i in range(batch_size):
+        fake_logits[i, i] = 1e2
+
+    seq_group_metadata_list = []
+    prompt_lens = []
+    for i in range(batch_size):
+        seq_group_metadata_list.append(
+            SequenceGroupMetadata(
+                request_id=f"test_{i}",
+                is_prompt=True,
+                seq_data={0: SequenceData([1, 2, 3])},
+                sampling_params=SamplingParams(
+                    temperature=1.0,
+                    n=random.randint(1, 10),
+                ),
+                block_tables={0: [1]},
+            ))
+        prompt_lens.append(seq_group_metadata_list[-1].seq_data[0].get_len())
+
+    sampling_metadata = model_runner._prepare_sample(seq_group_metadata_list,
+                                                     prompt_lens)
+    sampler_output = sampler(embedding=None,
+                             hidden_states=input_tensor,
+                             sampling_metadata=sampling_metadata)
+    for i, sequence_output in enumerate(sampler_output):
+        for nth_output in sequence_output.samples:
+            assert nth_output.output_token == i
+
+
+@pytest.mark.parametrize("seed", RANDOM_SEEDS)
+def test_sampler_all_beam(seed: int):
+    set_random_seed(seed)
+    batch_size = random.randint(1, 256)
+    input_tensor, _, sampler, model_runner = _prepare_test(batch_size)
+
+    seq_group_metadata_list = []
+    prompt_lens = []
+    for i in range(batch_size):
+        seq_group_metadata_list.append(
+            SequenceGroupMetadata(
+                request_id=f"test_{i}",
+                is_prompt=True,
+                seq_data={0: SequenceData([1, 2, 3])},
+                sampling_params=SamplingParams(
+                    temperature=0,
+                    best_of=2,
+                    use_beam_search=True,
+                ),
+                block_tables={0: [1]},
+            ))
+        prompt_lens.append(seq_group_metadata_list[-1].seq_data[0].get_len())
+
+    sampling_metadata = model_runner._prepare_sample(seq_group_metadata_list,
+                                                     prompt_lens)
+    sampler(embedding=None,
+            hidden_states=input_tensor,
+            sampling_metadata=sampling_metadata)
+    # no assertion here as I am not sure how to determine whether
+    # the outputs are expected - in other words, this just tests
+    # whether there are no exceptions in the sampler
+    # when handling an all-beam search case.
+
+
+@pytest.mark.parametrize("seed", RANDOM_SEEDS)
+def test_sampler_mixed(seed: int):
+    set_random_seed(seed)
+    batch_size = random.randint(1, 256)
+    input_tensor, fake_logits, sampler, model_runner = _prepare_test(
+        batch_size)
+
+    seq_group_metadata_list = []
+    expected_tokens = []
+    prompt_lens = []
+    for i in range(batch_size):
+        n = 1
+        sampling_type = random.randint(0, 2)
+        if sampling_type == 0:
+            sampling_params = SamplingParams(temperature=0)
+        elif sampling_type == 1:
+            n = random.randint(1, 10)
+            sampling_params = SamplingParams(
+                temperature=random.random() + 0.1,
+                top_p=min(random.random() + 0.1, 1),
+                top_k=random.randint(0, 10) or -1,
+                n=n,
+                presence_penalty=random.randint(0, 1),
+            )
+        else:
+            sampling_params = SamplingParams(temperature=0,
+                                             use_beam_search=True,
+                                             best_of=2)
+        for idx in range(n):
+            fake_logits[i, i + idx] = 1e2
+            expected_tokens.append(i + idx)
+        seq_group_metadata_list.append(
+            SequenceGroupMetadata(
+                request_id=f"test_{i}",
+                is_prompt=True,
+                seq_data={0: SequenceData([1, 2, 3])},
+                sampling_params=sampling_params,
+                block_tables={0: [1]},
+            ))
+        prompt_lens.append(seq_group_metadata_list[-1].seq_data[0].get_len())
+
+    sampling_metadata = model_runner._prepare_sample(seq_group_metadata_list,
+                                                     prompt_lens)
+    sampler_output = sampler(embedding=None,
+                             hidden_states=input_tensor,
+                             sampling_metadata=sampling_metadata)
+    for i, sequence_output in enumerate(sampler_output):
+        if seq_group_metadata_list[i].sampling_params.use_beam_search:
+            continue
+        for nth_output in sequence_output.samples:
+            assert nth_output.output_token in expected_tokens
+
+
+@pytest.mark.parametrize("seed", RANDOM_SEEDS)
+def test_sampler_logits_processors(seed: int):
+    set_random_seed(seed)
+    batch_size = random.randint(1, 256)
+    input_tensor, _, sampler, model_runner = _prepare_test(batch_size)
+
+    # This sample logits processor gives infinite score to the i-th token,
+    # where i is the length of the input sequence.
+    # We therefore expect the output token sequence to be [0, 1, 2, ...]
+    def pick_ith(token_ids, logits):
+        logits[len(token_ids)] = float("inf")
+        return logits
+
+    seq_group_metadata_list = []
+    prompt_lens = []
+    for i in range(batch_size):
+        seq_group_metadata_list.append(
+            SequenceGroupMetadata(
+                request_id=f"test_{i}",
+                is_prompt=True,
+                seq_data={0: SequenceData([1, 2, 3])},
+                sampling_params=SamplingParams(temperature=0,
+                                               logits_processors=[pick_ith]),
+                block_tables={0: [1]},
+            ))
+        prompt_lens.append(seq_group_metadata_list[-1].seq_data[0].get_len())
+
+    sampling_metadata = model_runner._prepare_sample(seq_group_metadata_list,
+                                                     prompt_lens)
+    sampler_output = sampler(embedding=None,
+                             hidden_states=input_tensor,
+                             sampling_metadata=sampling_metadata)
+    for _, sequence_output in enumerate(sampler_output):
+        for idx, nth_output in enumerate(sequence_output.samples):
+            assert nth_output.output_token == idx

tests/test_regression.py

@@ -0,0 +1,27 @@
+"""Containing tests that check for regressions in vLLM's behavior.
+
+It should include tests that are reported by users and making sure they
+will never happen again.
+
+"""
+from vllm import LLM, SamplingParams
+
+
+def test_duplicated_ignored_sequence_group():
+    """https://github.com/vllm-project/vllm/issues/1655"""
+
+    sampling_params = SamplingParams(temperature=0.01,
+                                     top_p=0.1,
+                                     max_tokens=256)
+    llm = LLM(model="facebook/opt-125m",
+              max_num_batched_tokens=4096,
+              tensor_parallel_size=1)
+    prompts = ["This is a short prompt", "This is a very long prompt " * 1000]
+    outputs = llm.generate(prompts, sampling_params=sampling_params)
+
+    assert len(prompts) == len(outputs)
+
+
+if __name__ == "__main__":
+    import pytest
+    pytest.main([__file__])

tests/worker/test_model_runner.py

@@ -0,0 +1,48 @@
+import random
+import torch
+
+from vllm.sequence import SamplingParams, SequenceData, SequenceGroupMetadata
+from vllm.worker.model_runner import ModelRunner
+
+
+def test_prepare_prompt():
+    model_runner = ModelRunner(None, None, None)
+    model_runner.set_block_size(16)
+
+    batch_size = random.randint(1, 256)
+    prompt_lens = []
+    seq_group_metadata_list = []
+    for i in range(batch_size):
+        # make sure all tokens fit into one block
+        prompt_len = i % (model_runner.block_size - 1) + 1
+        prompt_lens.append(prompt_len)
+        seq_data = list(range(prompt_len))
+        seq_group_metadata_list.append(
+            SequenceGroupMetadata(
+                request_id=f"test_{i}",
+                is_prompt=True,
+                seq_data={0: SequenceData(seq_data)},
+                sampling_params=SamplingParams(temperature=0),
+                block_tables={0: [1]},
+            ))
+
+    expected_selected_token_indices = []
+    selected_token_start_idx = 0
+    max_seq_len = max(prompt_lens)
+    for prompt_len in prompt_lens:
+        expected_selected_token_indices.append(selected_token_start_idx +
+                                               prompt_len - 1)
+        selected_token_start_idx += max_seq_len
+    input_tokens, input_positions, _ = model_runner._prepare_prompt(
+        seq_group_metadata_list)
+    sampling_metadata = model_runner._prepare_sample(seq_group_metadata_list,
+                                                     prompt_lens)
+    assert input_tokens.shape == (batch_size, max_seq_len)
+    assert input_positions.shape == (batch_size, max_seq_len)
+    torch.testing.assert_close(input_tokens, input_positions)
+
+    actual = sampling_metadata.selected_token_indices
+    expected = torch.tensor(expected_selected_token_indices,
+                            device=actual.device,
+                            dtype=actual.dtype)
+    torch.testing.assert_close(actual, expected)

vllm/__init__.py

@@ -8,7 +8,7 @@ from vllm.entrypoints.llm import LLM
 from vllm.outputs import CompletionOutput, RequestOutput
 from vllm.sampling_params import SamplingParams
 
-__version__ = "0.1.2"
+__version__ = "0.2.3"
 
 __all__ = [
     "LLM",

vllm/config.py

@@ -1,4 +1,5 @@
-from typing import Optional
+from typing import Optional, Union
+import os
 
 import torch
 from transformers import PretrainedConfig
@@ -20,15 +21,34 @@ class ModelConfig:
         tokenizer: Name or path of the huggingface tokenizer to use.
         tokenizer_mode: Tokenizer mode. "auto" will use the fast tokenizer if
             available, and "slow" will always use the slow tokenizer.
+        trust_remote_code: Trust remote code (e.g., from HuggingFace) when
+            downloading the model and tokenizer.
         download_dir: Directory to download and load the weights, default to the
             default cache directory of huggingface.
-        use_np_weights: Save a numpy copy of model weights for faster loading.
-            This can increase the disk usage by up to 2x.
-        use_dummy_weights: Use dummy values for model weights (for profiling).
+        load_format: The format of the model weights to load:
+            "auto" will try to load the weights in the safetensors format and
+                fall back to the pytorch bin format if safetensors format is
+                not available.
+            "pt" will load the weights in the pytorch bin format.
+            "safetensors" will load the weights in the safetensors format.
+            "npcache" will load the weights in pytorch format and store
+                a numpy cache to speed up the loading.
+            "dummy" will initialize the weights with random values, which is
+                mainly for profiling.
         dtype: 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.
         seed: Random seed for reproducibility.
+        revision: The specific model version to use. It can be a branch name,
+            a tag name, or a commit id. If unspecified, will use the default
+            version.
+        tokenizer_revision: The specific tokenizer version to use. It can be a
+            branch name, a tag name, or a commit id. If unspecified, will use
+            the default version.
+        max_model_len: Maximum length of a sequence (including prompt and
+            output). If None, will be derived from the model.
+        quantization: Quantization method that was used to quantize the model
+            weights. If None, we assume the model weights are not quantized.
     """
 
     def __init__(
@@ -36,23 +56,55 @@ class ModelConfig:
         model: str,
         tokenizer: str,
         tokenizer_mode: str,
+        trust_remote_code: bool,
         download_dir: Optional[str],
-        use_np_weights: bool,
-        use_dummy_weights: bool,
-        dtype: str,
+        load_format: str,
+        dtype: Union[str, torch.dtype],
         seed: int,
+        revision: Optional[str] = None,
+        tokenizer_revision: Optional[str] = None,
+        max_model_len: Optional[int] = None,
+        quantization: Optional[str] = None,
     ) -> None:
         self.model = model
         self.tokenizer = tokenizer
         self.tokenizer_mode = tokenizer_mode
+        self.trust_remote_code = trust_remote_code
         self.download_dir = download_dir
-        self.use_np_weights = use_np_weights
-        self.use_dummy_weights = use_dummy_weights
+        self.load_format = load_format
         self.seed = seed
+        self.revision = revision
+        self.tokenizer_revision = tokenizer_revision
+        self.quantization = quantization
 
-        self.hf_config = get_config(model)
+        if os.environ.get("VLLM_USE_MODELSCOPE", "False").lower() == "true":
+            # download model from ModelScope hub,
+            # lazy import so that modelscope is not required for normal use.
+            from modelscope.hub.snapshot_download import snapshot_download  # pylint: disable=C
+            model_path = snapshot_download(model_id=model,
+                                           cache_dir=download_dir,
+                                           revision=revision)
+            self.model = model_path
+            self.download_dir = model_path
+            self.tokenizer = model_path
+
+        self.hf_config = get_config(self.model, trust_remote_code, revision)
         self.dtype = _get_and_verify_dtype(self.hf_config, dtype)
+        self.max_model_len = _get_and_verify_max_len(self.hf_config,
+                                                     max_model_len)
+        self._verify_load_format()
         self._verify_tokenizer_mode()
+        self._verify_quantization()
+
+    def _verify_load_format(self) -> None:
+        load_format = self.load_format.lower()
+        if load_format not in [
+                "auto", "pt", "safetensors", "npcache", "dummy"
+        ]:
+            raise ValueError(
+                f"Unknown load format: {self.load_format}. Must be one of "
+                "'auto', 'pt', 'safetensors', 'npcache', or 'dummy'.")
+        self.load_format = load_format
 
     def _verify_tokenizer_mode(self) -> None:
         tokenizer_mode = self.tokenizer_mode.lower()
@@ -62,6 +114,33 @@ class ModelConfig:
                 "either 'auto' or 'slow'.")
         self.tokenizer_mode = tokenizer_mode
 
+    def _verify_quantization(self) -> None:
+        supported_quantization = ["awq", "squeezellm"]
+        if self.quantization is not None:
+            self.quantization = self.quantization.lower()
+
+        # Parse quantization method from the HF model config, if available.
+        hf_quant_config = getattr(self.hf_config, "quantization_config", None)
+        if hf_quant_config is not None:
+            hf_quant_method = str(hf_quant_config["quant_method"]).lower()
+            if self.quantization is None:
+                self.quantization = hf_quant_method
+            elif self.quantization != hf_quant_method:
+                raise ValueError(
+                    "Quantization method specified in the model config "
+                    f"({hf_quant_method}) does not match the quantization "
+                    f"method specified in the `quantization` argument "
+                    f"({self.quantization}).")
+
+        if self.quantization is not None:
+            if self.quantization not in supported_quantization:
+                raise ValueError(
+                    f"Unknown quantization method: {self.quantization}. Must "
+                    f"be one of {supported_quantization}.")
+            logger.warning(f"{self.quantization} quantization is not fully "
+                           "optimized yet. The speed can be slower than "
+                           "non-quantized models.")
+
     def verify_with_parallel_config(
         self,
         parallel_config: "ParallelConfig",
@@ -82,6 +161,12 @@ class ModelConfig:
                 "must be divisible by pipeline parallel size "
                 f"({pipeline_parallel_size}).")
 
+    def get_sliding_window(self) -> Optional[int]:
+        return getattr(self.hf_config, "sliding_window", None)
+
+    def get_vocab_size(self) -> int:
+        return self.hf_config.vocab_size
+
     def get_hidden_size(self) -> int:
         return self.hf_config.hidden_size
 
@@ -89,9 +174,49 @@ class ModelConfig:
         # FIXME(woosuk): This may not be true for all models.
         return self.hf_config.hidden_size // self.hf_config.num_attention_heads
 
-    def get_num_heads(self, parallel_config: "ParallelConfig") -> int:
-        total_num_attention_heads = self.hf_config.num_attention_heads
-        return total_num_attention_heads // parallel_config.tensor_parallel_size
+    def get_total_num_kv_heads(self) -> int:
+        """Returns the total number of KV heads."""
+        # For GPTBigCode & Falcon:
+        # NOTE: for falcon, when new_decoder_architecture is True, the
+        # multi_query flag is ignored and we use n_head_kv for the number of
+        # KV heads.
+        falcon_model_types = ["falcon", "RefinedWeb", "RefinedWebModel"]
+        new_decoder_arch_falcon = (
+            self.hf_config.model_type in falcon_model_types
+            and getattr(self.hf_config, "new_decoder_architecture", False))
+        if not new_decoder_arch_falcon and getattr(self.hf_config,
+                                                   "multi_query", False):
+            # Multi-query attention, only one KV head.
+            # Currently, tensor parallelism is not supported in this case.
+            return 1
+
+        attributes = [
+            # For Falcon:
+            "n_head_kv",
+            "num_kv_heads",
+            # For LLaMA-2:
+            "num_key_value_heads",
+            # For ChatGLM:
+            "multi_query_group_num",
+        ]
+        for attr in attributes:
+            num_kv_heads = getattr(self.hf_config, attr, None)
+            if num_kv_heads is not None:
+                return num_kv_heads
+
+        # For non-grouped-query attention models, the number of KV heads is
+        # equal to the number of attention heads.
+        return self.hf_config.num_attention_heads
+
+    def get_num_kv_heads(self, parallel_config: "ParallelConfig") -> int:
+        """Returns the number of KV heads per GPU."""
+        total_num_kv_heads = self.get_total_num_kv_heads()
+        # If tensor parallelism is used, we divide the number of KV heads by
+        # the tensor parallel size. We will replicate the KV heads in the
+        # case where the number of KV heads is smaller than the tensor
+        # parallel size so each GPU has at least one KV head.
+        return max(1,
+                   total_num_kv_heads // parallel_config.tensor_parallel_size)
 
     def get_num_layers(self, parallel_config: "ParallelConfig") -> int:
         total_num_hidden_layers = self.hf_config.num_hidden_layers
@@ -113,10 +238,12 @@ class CacheConfig:
         block_size: int,
         gpu_memory_utilization: float,
         swap_space: int,
+        sliding_window: Optional[int] = None,
     ) -> None:
         self.block_size = block_size
         self.gpu_memory_utilization = gpu_memory_utilization
         self.swap_space_bytes = swap_space * _GB
+        self.sliding_window = sliding_window
         self._verify_args()
 
         # Will be set after profiling.
@@ -164,10 +291,12 @@ class ParallelConfig:
         pipeline_parallel_size: int,
         tensor_parallel_size: int,
         worker_use_ray: bool,
+        max_parallel_loading_workers: Optional[int] = None,
     ) -> None:
         self.pipeline_parallel_size = pipeline_parallel_size
         self.tensor_parallel_size = tensor_parallel_size
         self.worker_use_ray = worker_use_ray
+        self.max_parallel_loading_workers = max_parallel_loading_workers
 
         self.world_size = pipeline_parallel_size * tensor_parallel_size
         if self.world_size > 1:
@@ -188,15 +317,43 @@ class SchedulerConfig:
             a single iteration.
         max_num_seqs: Maximum number of sequences to be processed in a single
             iteration.
-        max_seq_len: Maximum length of a sequence (including prompt
+        max_model_len: Maximum length of a sequence (including prompt
             and generated text).
+        max_paddings: Maximum number of paddings to be added to a batch.
     """
 
-    def __init__(self, max_num_batched_tokens: int, max_num_seqs: int,
-                 max_seq_len: int) -> None:
-        self.max_num_batched_tokens = max_num_batched_tokens
+    def __init__(
+        self,
+        max_num_batched_tokens: Optional[int],
+        max_num_seqs: int,
+        max_model_len: int,
+        max_paddings: int,
+    ) -> None:
+        if max_num_batched_tokens is not None:
+            self.max_num_batched_tokens = max_num_batched_tokens
+        else:
+            # If max_model_len is too short, use 2048 as the default value for
+            # higher throughput.
+            self.max_num_batched_tokens = max(max_model_len, 2048)
         self.max_num_seqs = max_num_seqs
-        self.max_seq_len = max_seq_len
+        self.max_model_len = max_model_len
+        self.max_paddings = max_paddings
+        self._verify_args()
+
+    def _verify_args(self) -> None:
+        if self.max_num_batched_tokens < self.max_model_len:
+            raise ValueError(
+                f"max_num_batched_tokens ({self.max_num_batched_tokens}) is "
+                f"smaller than max_model_len ({self.max_model_len}). "
+                "This effectively limits the maximum sequence length to "
+                "max_num_batched_tokens and makes vLLM reject longer "
+                "sequences. Please increase max_num_batched_tokens or "
+                "decrease max_model_len.")
+        if self.max_num_batched_tokens < self.max_num_seqs:
+            raise ValueError(
+                f"max_num_batched_tokens ({self.max_num_batched_tokens}) must "
+                "be greater than or equal to max_num_seqs "
+                f"({self.max_num_seqs}).")
 
 
 _STR_DTYPE_TO_TORCH_DTYPE = {
@@ -210,7 +367,7 @@ _STR_DTYPE_TO_TORCH_DTYPE = {
 
 def _get_and_verify_dtype(
     config: PretrainedConfig,
-    dtype: str,
+    dtype: Union[str, torch.dtype],
 ) -> torch.dtype:
     # NOTE: getattr(config, "torch_dtype", torch.float32) is not correct
     # because config.torch_dtype can be None.
@@ -218,17 +375,23 @@ def _get_and_verify_dtype(
     if config_dtype is None:
         config_dtype = torch.float32
 
-    dtype = dtype.lower()
-    if dtype == "auto":
-        if config_dtype == torch.float32:
-            # Following the common practice, we use float16 for float32 models.
-            torch_dtype = torch.float16
+    if isinstance(dtype, str):
+        dtype = dtype.lower()
+        if dtype == "auto":
+            if config_dtype == torch.float32:
+                # Following the common practice, we use float16 for float32
+                # models.
+                torch_dtype = torch.float16
+            else:
+                torch_dtype = config_dtype
         else:
-            torch_dtype = config_dtype
+            if dtype not in _STR_DTYPE_TO_TORCH_DTYPE:
+                raise ValueError(f"Unknown dtype: {dtype}")
+            torch_dtype = _STR_DTYPE_TO_TORCH_DTYPE[dtype]
+    elif isinstance(dtype, torch.dtype):
+        torch_dtype = dtype
     else:
-        if dtype not in _STR_DTYPE_TO_TORCH_DTYPE:
-            raise ValueError(f"Unknown dtype: {dtype}")
-        torch_dtype = _STR_DTYPE_TO_TORCH_DTYPE[dtype]
+        raise ValueError(f"Unknown dtype: {dtype}")
 
     # Verify the dtype.
     if torch_dtype != config_dtype:
@@ -242,13 +405,62 @@ def _get_and_verify_dtype(
             # Casting between float16 and bfloat16 is allowed with a warning.
             logger.warning(f"Casting {config_dtype} to {torch_dtype}.")
 
-    # Check if the GPU supports the dtype.
-    if torch_dtype == torch.bfloat16:
-        compute_capability = torch.cuda.get_device_capability()
-        if compute_capability[0] < 8:
-            gpu_name = torch.cuda.get_device_name()
-            raise ValueError(
-                "Bfloat16 is only supported on GPUs with compute capability "
-                f"of at least 8.0. Your {gpu_name} GPU has compute capability "
-                f"{compute_capability[0]}.{compute_capability[1]}.")
     return torch_dtype
+
+
+def _get_and_verify_max_len(
+    hf_config: PretrainedConfig,
+    max_model_len: Optional[int],
+) -> int:
+    """Get and verify the model's maximum length."""
+    derived_max_model_len = float("inf")
+    possible_keys = [
+        # OPT
+        "max_position_embeddings",
+        # GPT-2
+        "n_positions",
+        # MPT
+        "max_seq_len",
+        # ChatGLM2
+        "seq_length",
+        # Others
+        "max_sequence_length",
+        "max_seq_length",
+        "seq_len",
+    ]
+    for key in possible_keys:
+        max_len_key = getattr(hf_config, key, None)
+        if max_len_key is not None:
+            derived_max_model_len = min(derived_max_model_len, max_len_key)
+    if derived_max_model_len == float("inf"):
+        if max_model_len is not None:
+            # If max_model_len is specified, we use it.
+            return max_model_len
+
+        default_max_len = 2048
+        logger.warning(
+            "The model's config.json does not contain any of the following "
+            "keys to determine the original maximum length of the model: "
+            f"{possible_keys}. Assuming the model's maximum length is "
+            f"{default_max_len}.")
+        derived_max_model_len = default_max_len
+
+    rope_scaling = getattr(hf_config, "rope_scaling", None)
+    if rope_scaling is not None:
+        assert "factor" in rope_scaling
+        scaling_factor = rope_scaling["factor"]
+        if rope_scaling["type"] == "yarn":
+            derived_max_model_len = rope_scaling[
+                "original_max_position_embeddings"]
+        derived_max_model_len *= scaling_factor
+
+    if max_model_len is None:
+        max_model_len = derived_max_model_len
+    elif max_model_len > derived_max_model_len:
+        raise ValueError(
+            f"User-specified max_model_len ({max_model_len}) is greater than "
+            f"the derived max_model_len ({max_len_key}={derived_max_model_len}"
+            " in model's config.json). This may lead to incorrect model "
+            "outputs or CUDA errors. Make sure the value is correct and "
+            "within the model context size.")
+    return int(max_model_len)

vllm/core/block_manager.py

@@ -1,10 +1,14 @@
 """A block manager that manages token blocks."""
+import enum
 from typing import Dict, List, Optional, Set, Tuple
 
 from vllm.block import PhysicalTokenBlock
 from vllm.sequence import Sequence, SequenceGroup, SequenceStatus
 from vllm.utils import Device
 
+# Mapping: logical block number -> physical block.
+BlockTable = List[PhysicalTokenBlock]
+
 
 class BlockAllocator:
     """Manages free physical token blocks for a device.
@@ -25,7 +29,7 @@ class BlockAllocator:
         self.num_blocks = num_blocks
 
         # Initialize the free blocks.
-        self.free_blocks: List[PhysicalTokenBlock] = []
+        self.free_blocks: BlockTable = []
         for i in range(num_blocks):
             block = PhysicalTokenBlock(device=device,
                                        block_number=i,
@@ -50,8 +54,18 @@ class BlockAllocator:
         return len(self.free_blocks)
 
 
-# Mapping: logical block number -> physical block.
-BlockTable = List[PhysicalTokenBlock]
+class AllocStatus(enum.Enum):
+    """Result for BlockSpaceManager.can_allocate
+
+    1. Ok: seq_group can be allocated now.
+    2. Later: seq_group cannot be allocated.
+      The capacity of allocator is larger than seq_group required.
+    3. Never: seq_group can never be allocated.
+      The seq_group is too large to allocated in GPU.
+    """
+    OK = enum.auto()
+    LATER = enum.auto()
+    NEVER = enum.auto()
 
 
 class BlockSpaceManager:
@@ -63,10 +77,18 @@ class BlockSpaceManager:
         num_gpu_blocks: int,
         num_cpu_blocks: int,
         watermark: float = 0.01,
+        sliding_window: Optional[int] = None,
     ) -> None:
         self.block_size = block_size
         self.num_total_gpu_blocks = num_gpu_blocks
         self.num_total_cpu_blocks = num_cpu_blocks
+
+        self.block_sliding_window = None
+        if sliding_window is not None:
+            assert sliding_window % block_size == 0, (sliding_window,
+                                                      block_size)
+            self.block_sliding_window = sliding_window // block_size
+
         self.watermark = watermark
         assert watermark >= 0.0
 
@@ -78,15 +100,24 @@ class BlockSpaceManager:
         # Mapping: seq_id -> BlockTable.
         self.block_tables: Dict[int, BlockTable] = {}
 
-    def can_allocate(self, seq_group: SequenceGroup) -> bool:
+    def can_allocate(self, seq_group: SequenceGroup) -> AllocStatus:
         # FIXME(woosuk): Here we assume that all sequences in the group share
         # the same prompt. This may not be true for preempted sequences.
         seq = seq_group.get_seqs()[0]
         num_required_blocks = len(seq.logical_token_blocks)
+        if self.block_sliding_window is not None:
+            num_required_blocks = min(num_required_blocks,
+                                      self.block_sliding_window)
         num_free_gpu_blocks = self.gpu_allocator.get_num_free_blocks()
+
         # Use watermark to avoid frequent cache eviction.
-        return (num_free_gpu_blocks - num_required_blocks >=
-                self.watermark_blocks)
+        if (self.num_total_gpu_blocks - num_required_blocks <
+                self.watermark_blocks):
+            return AllocStatus.NEVER
+        if num_free_gpu_blocks - num_required_blocks >= self.watermark_blocks:
+            return AllocStatus.OK
+        else:
+            return AllocStatus.LATER
 
     def allocate(self, seq_group: SequenceGroup) -> None:
         # NOTE: Here we assume that all sequences in the group have the same
@@ -95,8 +126,12 @@ class BlockSpaceManager:
 
         # Allocate new physical token blocks that will store the prompt tokens.
         block_table: BlockTable = []
-        for _ in range(len(seq.logical_token_blocks)):
-            block = self.gpu_allocator.allocate()
+        for logical_idx in range(len(seq.logical_token_blocks)):
+            if (self.block_sliding_window is not None
+                    and logical_idx >= self.block_sliding_window):
+                block = block_table[logical_idx % self.block_sliding_window]
+            else:
+                block = self.gpu_allocator.allocate()
             # Set the reference counts of the token blocks.
             block.ref_count = seq_group.num_seqs()
             block_table.append(block)
@@ -118,11 +153,17 @@ class BlockSpaceManager:
         block_table = self.block_tables[seq.seq_id]
 
         if len(block_table) < len(logical_blocks):
-            # The sequence has a new logical block.
-            # Allocate a new physical block.
-            block = self.gpu_allocator.allocate()
-            block_table.append(block)
-            return None
+            if (self.block_sliding_window
+                    and len(block_table) >= self.block_sliding_window):
+                # re-use a block
+                block_table.append(block_table[len(block_table) %
+                                               self.block_sliding_window])
+            else:
+                # The sequence has a new logical block.
+                # Allocate a new physical block.
+                block = self.gpu_allocator.allocate()
+                block_table.append(block)
+                return None
 
         # We want to append the token to the last physical block.
         last_block = block_table[-1]
@@ -154,9 +195,7 @@ class BlockSpaceManager:
         for seq in seq_group.get_seqs():
             if seq.is_finished():
                 continue
-            block_table = self.block_tables[seq.seq_id]
-            for block in block_table:
-                blocks.add(block)
+            blocks.update(self.block_tables[seq.seq_id])
         return list(blocks)
 
     def can_swap_in(self, seq_group: SequenceGroup) -> bool:
@@ -172,9 +211,7 @@ class BlockSpaceManager:
     def swap_in(self, seq_group: SequenceGroup) -> Dict[int, int]:
         # CPU block -> GPU block.
         mapping: Dict[PhysicalTokenBlock, PhysicalTokenBlock] = {}
-        for seq in seq_group.get_seqs():
-            if seq.is_finished():
-                continue
+        for seq in seq_group.get_seqs(status=SequenceStatus.SWAPPED):
             new_block_table: BlockTable = []
             block_table = self.block_tables[seq.seq_id]
 
@@ -203,9 +240,7 @@ class BlockSpaceManager:
     def swap_out(self, seq_group: SequenceGroup) -> Dict[int, int]:
         # GPU block -> CPU block.
         mapping: Dict[PhysicalTokenBlock, PhysicalTokenBlock] = {}
-        for seq in seq_group.get_seqs():
-            if seq.is_finished():
-                continue
+        for seq in seq_group.get_seqs(status=SequenceStatus.RUNNING):
             new_block_table: BlockTable = []
             block_table = self.block_tables[seq.seq_id]
 
@@ -228,7 +263,7 @@ class BlockSpaceManager:
         return block_number_mapping
 
     def _free_block_table(self, block_table: BlockTable) -> None:
-        for block in block_table:
+        for block in set(block_table):
             if block.device == Device.GPU:
                 self.gpu_allocator.free(block)
             else:

vllm/core/scheduler.py

@@ -1,19 +1,16 @@
 import enum
 import time
-from typing import Dict, List, Optional, Tuple
+from typing import Dict, Iterable, List, Optional, Tuple, Union
 
 from vllm.config import CacheConfig, SchedulerConfig
-from vllm.core.block_manager import BlockSpaceManager
+from vllm.core.block_manager import AllocStatus, BlockSpaceManager
 from vllm.core.policy import PolicyFactory
 from vllm.logger import init_logger
 from vllm.sequence import (Sequence, SequenceData, SequenceGroup,
-                           SequenceGroupMetadata, SequenceOutputs,
-                           SequenceStatus)
+                           SequenceGroupMetadata, SequenceStatus)
 
 logger = init_logger(__name__)
 
-_LOGGING_INTERVAL_SEC = 5
-
 
 class PreemptionMode(enum.Enum):
     """Preemption modes.
@@ -32,19 +29,28 @@ class SchedulerOutputs:
 
     def __init__(
         self,
+        scheduled_seq_groups: List[SequenceGroup],
+        prompt_run: bool,
+        num_batched_tokens: int,
         blocks_to_swap_in: Dict[int, int],
         blocks_to_swap_out: Dict[int, int],
         blocks_to_copy: Dict[int, List[int]],
+        ignored_seq_groups: List[SequenceGroup],
     ) -> None:
+        self.scheduled_seq_groups = scheduled_seq_groups
+        self.prompt_run = prompt_run
+        self.num_batched_tokens = num_batched_tokens
         self.blocks_to_swap_in = blocks_to_swap_in
         self.blocks_to_swap_out = blocks_to_swap_out
         self.blocks_to_copy = blocks_to_copy
         # Swap in and swap out should never happen at the same time.
         assert not (blocks_to_swap_in and blocks_to_swap_out)
+        self.ignored_seq_groups = ignored_seq_groups
 
     def is_empty(self) -> bool:
-        return (not self.blocks_to_swap_in and not self.blocks_to_swap_out
-                and not self.blocks_to_copy)
+        # NOTE: We do not consider the ignored sequence groups.
+        return (not self.scheduled_seq_groups and not self.blocks_to_swap_in
+                and not self.blocks_to_swap_out and not self.blocks_to_copy)
 
 
 class Scheduler:
@@ -53,11 +59,12 @@ class Scheduler:
         self,
         scheduler_config: SchedulerConfig,
         cache_config: CacheConfig,
-        log_stats: bool,
     ) -> None:
         self.scheduler_config = scheduler_config
         self.cache_config = cache_config
-        self.log_stats = log_stats
+
+        self.prompt_limit = min(self.scheduler_config.max_model_len,
+                                self.scheduler_config.max_num_batched_tokens)
 
         # Instantiate the scheduling policy.
         self.policy = PolicyFactory.get_policy(policy_name="fcfs")
@@ -66,8 +73,9 @@ class Scheduler:
             block_size=self.cache_config.block_size,
             num_gpu_blocks=self.cache_config.num_gpu_blocks,
             num_cpu_blocks=self.cache_config.num_cpu_blocks,
-        )
+            sliding_window=self.cache_config.sliding_window)
 
+        # TODO(zhuohan): Use deque instead of list for better performance.
         # Sequence groups in the WAITING state.
         self.waiting: List[SequenceGroup] = []
         # Sequence groups in the RUNNING state.
@@ -75,25 +83,30 @@ class Scheduler:
         # Sequence groups in the SWAPPED state.
         self.swapped: List[SequenceGroup] = []
 
-        self.last_logging_time: float = 0.0
-        # List[timestamp, num_tokens]
-        self.num_input_tokens: List[Tuple[float, int]] = []
-
     def add_seq_group(self, seq_group: SequenceGroup) -> None:
         # Add sequence groups to the waiting queue.
         self.waiting.append(seq_group)
 
-    def abort_seq_group(self, request_id: str) -> None:
+    def abort_seq_group(self, request_id: Union[str, Iterable[str]]) -> None:
+        if isinstance(request_id, str):
+            request_id = (request_id, )
+        request_ids = set(request_id)
         for state_queue in [self.waiting, self.running, self.swapped]:
-            for seq_group in state_queue:
-                if seq_group.request_id == request_id:
+            # We need to reverse the list as we are removing elements
+            # from it as we iterate over it. If we don't do it,
+            # indices will get messed up and we will skip over elements.
+            for seq_group in reversed(state_queue):
+                if seq_group.request_id in request_ids:
                     # Remove the sequence group from the state queue.
                     state_queue.remove(seq_group)
-                    for seq in seq_group.seqs:
+                    for seq in seq_group.get_seqs():
                         if seq.is_finished():
                             continue
-                        self.free_seq(seq, SequenceStatus.FINISHED_ABORTED)
-                    return
+                        seq.status = SequenceStatus.FINISHED_ABORTED
+                        self.free_seq(seq)
+                    request_ids.remove(seq_group.request_id)
+                    if not request_ids:
+                        return
 
     def has_unfinished_seqs(self) -> bool:
         return self.waiting or self.running or self.swapped
@@ -101,21 +114,99 @@ class Scheduler:
     def get_num_unfinished_seq_groups(self) -> int:
         return len(self.waiting) + len(self.running) + len(self.swapped)
 
-    def _schedule(
-            self) -> Tuple[SchedulerOutputs, List[str], List[SequenceGroup]]:
+    def _schedule(self) -> SchedulerOutputs:
         # 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]] = {}
-        ignored_seq_groups: List[SequenceGroup] = []
 
         # Fix the current time.
-        now = time.time()
+        now = time.monotonic()
 
-        # 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.
+        # Join waiting sequences if possible.
+        if not self.swapped:
+            ignored_seq_groups: List[SequenceGroup] = []
+            scheduled: List[SequenceGroup] = []
+            # The total number of sequences on the fly, including the
+            # requests in the generation phase.
+            num_curr_seqs = sum(seq_group.get_max_num_running_seqs()
+                                for seq_group in self.running)
+            seq_lens: List[int] = []
+
+            # Optimization: We do not sort the waiting queue since the preempted
+            # sequence groups are added to the front and the new sequence groups
+            # are added to the back.
+            while self.waiting:
+                seq_group = self.waiting[0]
+
+                assert seq_group.num_seqs() == 1, (
+                    "Waiting sequence group should have only one prompt "
+                    "sequence.")
+                num_prompt_tokens = seq_group.get_seqs()[0].get_len()
+                if num_prompt_tokens > self.prompt_limit:
+                    logger.warning(
+                        f"Input prompt ({num_prompt_tokens} tokens) is too long"
+                        f" and exceeds limit of {self.prompt_limit}")
+                    for seq in seq_group.get_seqs():
+                        seq.status = SequenceStatus.FINISHED_IGNORED
+                    ignored_seq_groups.append(seq_group)
+                    self.waiting.pop(0)
+                    continue
+
+                # If the sequence group cannot be allocated, stop.
+                can_allocate = self.block_manager.can_allocate(seq_group)
+                if can_allocate == AllocStatus.LATER:
+                    break
+                elif can_allocate == AllocStatus.NEVER:
+                    logger.warning(
+                        f"Input prompt ({num_prompt_tokens} tokens) is too long"
+                        f" and exceeds the capacity of block_manager")
+                    for seq in seq_group.get_seqs():
+                        seq.status = SequenceStatus.FINISHED_IGNORED
+                    ignored_seq_groups.append(seq_group)
+                    self.waiting.pop(0)
+                    continue
+
+                # If the number of batched tokens exceeds the limit, stop.
+                new_seq_lens = seq_lens + [num_prompt_tokens]
+                num_batched_tokens = len(new_seq_lens) * max(new_seq_lens)
+                if (num_batched_tokens >
+                        self.scheduler_config.max_num_batched_tokens):
+                    break
+
+                # The total number of sequences in the RUNNING state should not
+                # exceed the maximum number of sequences.
+                num_new_seqs = seq_group.get_max_num_running_seqs()
+                if (num_curr_seqs + num_new_seqs >
+                        self.scheduler_config.max_num_seqs):
+                    break
+
+                num_paddings = num_batched_tokens - sum(new_seq_lens)
+                if num_paddings > self.scheduler_config.max_paddings:
+                    break
+                seq_lens = new_seq_lens
+
+                seq_group = self.waiting.pop(0)
+                self._allocate(seq_group)
+                self.running.append(seq_group)
+                num_curr_seqs += num_new_seqs
+                scheduled.append(seq_group)
+
+            if scheduled or ignored_seq_groups:
+                scheduler_outputs = SchedulerOutputs(
+                    scheduled_seq_groups=scheduled,
+                    prompt_run=True,
+                    num_batched_tokens=len(seq_lens) *
+                    max(seq_lens) if seq_lens else 0,
+                    blocks_to_swap_in=blocks_to_swap_in,
+                    blocks_to_swap_out=blocks_to_swap_out,
+                    blocks_to_copy=blocks_to_copy,
+                    ignored_seq_groups=ignored_seq_groups,
+                )
+                return scheduler_outputs
+
+        # NOTE(woosuk): 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)
@@ -145,151 +236,57 @@ class Scheduler:
 
         # Swap in the sequence groups in the SWAPPED state if possible.
         self.swapped = self.policy.sort_by_priority(now, self.swapped)
-        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
+        if not preempted:
+            num_curr_seqs = sum(seq_group.get_max_num_running_seqs()
+                                for seq_group in self.running)
 
-            # The total number of sequences in the RUNNING state should not
-            # exceed the maximum number of sequences.
-            num_new_seqs = seq_group.num_seqs(status=SequenceStatus.SWAPPED)
-            num_curr_seqs = sum(
-                seq_group.num_seqs(status=SequenceStatus.RUNNING)
-                for seq_group in self.running)
-            if (num_curr_seqs + num_new_seqs >
-                    self.scheduler_config.max_num_seqs):
-                break
-
-            seq_group = self.swapped.pop(0)
-            self._swap_in(seq_group, blocks_to_swap_in)
-            self._append_slot(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[str] = []
-        # 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:
-            # Optimization: We do not sort the waiting queue since the preempted
-            # sequence groups are added to the front and the new sequence groups
-            # are added to the back.
-            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
-
-                num_prompt_tokens = seq_group.get_seqs()[0].get_len()
-                if num_prompt_tokens >= self.scheduler_config.max_seq_len:
-                    logger.warning(
-                        f"Input prompt ({num_prompt_tokens} tokens) is too long"
-                        " and exceeds limit of "
-                        f"{self.scheduler_config.max_seq_len}")
-                    for seq in seq_group.get_seqs():
-                        seq.status = SequenceStatus.FINISHED_IGNORED
-                    ignored_seq_groups.append(seq_group)
-                    self.waiting.pop(0)
-                    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.
-                if (num_batched_tokens + num_prompt_tokens >
-                        self.scheduler_config.max_num_batched_tokens):
+            while self.swapped:
+                seq_group = self.swapped[0]
+                # 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_new_seqs = seq_group.num_seqs(
-                    status=SequenceStatus.WAITING)
-                num_curr_seqs = sum(
-                    seq_group.num_seqs(status=SequenceStatus.RUNNING)
-                    for seq_group in self.running)
+                num_new_seqs = seq_group.get_max_num_running_seqs()
                 if (num_curr_seqs + num_new_seqs >
                         self.scheduler_config.max_num_seqs):
                     break
 
-                seq_group = self.waiting.pop(0)
-                self._allocate(seq_group)
+                seq_group = self.swapped.pop(0)
+                self._swap_in(seq_group, blocks_to_swap_in)
+                self._append_slot(seq_group, blocks_to_copy)
+                num_curr_seqs += num_new_seqs
                 self.running.append(seq_group)
-                num_batched_tokens += num_prompt_tokens
-                prompt_group_ids.append(seq_group.request_id)
+
+        # Each sequence in the generation phase only takes one token slot.
+        # Therefore, the number of batched tokens is equal to the number of
+        # sequences in the RUNNING state.
+        num_batched_tokens = sum(
+            seq_group.num_seqs(status=SequenceStatus.RUNNING)
+            for seq_group in self.running)
 
         scheduler_outputs = SchedulerOutputs(
+            scheduled_seq_groups=self.running,
+            prompt_run=False,
+            num_batched_tokens=num_batched_tokens,
             blocks_to_swap_in=blocks_to_swap_in,
             blocks_to_swap_out=blocks_to_swap_out,
             blocks_to_copy=blocks_to_copy,
+            ignored_seq_groups=[],
         )
-        if not self.log_stats:
-            return scheduler_outputs, prompt_group_ids, ignored_seq_groups
+        return scheduler_outputs
 
-        # TODO(woosuk): Move the below code to the engine.
-        now = time.time()
-        if num_batched_tokens > 0:
-            self.num_input_tokens.append((now, num_batched_tokens))
-        elapsed_time = now - self.last_logging_time
-        if elapsed_time > _LOGGING_INTERVAL_SEC:
-            self.last_logging_time = now
-            self.num_input_tokens = [(t, n) for t, n in self.num_input_tokens
-                                     if now - t < _LOGGING_INTERVAL_SEC]
-            if len(self.num_input_tokens) > 1:
-                total_num_tokens = sum(n
-                                       for _, n in self.num_input_tokens[:-1])
-                window = now - self.num_input_tokens[0][0]
-                avg_throughput = total_num_tokens / window
-            else:
-                avg_throughput = 0.0
-
-            total_num_gpu_blocks = self.cache_config.num_gpu_blocks
-            num_free_gpu_blocks = self.block_manager.get_num_free_gpu_blocks()
-            num_used_gpu_blocks = total_num_gpu_blocks - num_free_gpu_blocks
-            gpu_cache_usage = num_used_gpu_blocks / total_num_gpu_blocks
-
-            total_num_cpu_blocks = self.cache_config.num_cpu_blocks
-            if total_num_cpu_blocks > 0:
-                num_free_cpu_blocks = (
-                    self.block_manager.get_num_free_cpu_blocks())
-                num_used_cpu_blocks = total_num_cpu_blocks - num_free_cpu_blocks
-                cpu_cache_usage = num_used_cpu_blocks / total_num_cpu_blocks
-            else:
-                cpu_cache_usage = 0.0
-
-            logger.info(f"Throughput: {avg_throughput:.1f} tokens/s, "
-                        f"Running: {len(self.running)} reqs, "
-                        f"Swapped: {len(self.swapped)} reqs, "
-                        f"Pending: {len(self.waiting)} reqs, "
-                        f"GPU KV cache usage: {gpu_cache_usage * 100:.1f}%, "
-                        f"CPU KV cache usage: {cpu_cache_usage * 100:.1f}%")
-        return scheduler_outputs, prompt_group_ids, ignored_seq_groups
-
-    def schedule(
-        self
-    ) -> Tuple[List[SequenceGroupMetadata], SchedulerOutputs,
-               List[SequenceGroup]]:
+    def schedule(self) -> Tuple[List[SequenceGroupMetadata], SchedulerOutputs]:
         # Schedule sequence groups.
         # This function call changes the internal states of the scheduler
         # such as self.running, self.swapped, and self.waiting.
-        (scheduler_outputs, prompt_group_ids,
-         ignored_seq_groups) = self._schedule()
+        scheduler_outputs = self._schedule()
 
         # Create input data structures.
         seq_group_metadata_list: List[SequenceGroupMetadata] = []
-        for seq_group in self.running:
-            is_prompt = seq_group.request_id in prompt_group_ids
-
-            seq_data: Dict[int, List[SequenceData]] = {}
+        for seq_group in scheduler_outputs.scheduled_seq_groups:
+            seq_data: Dict[int, SequenceData] = {}
             block_tables: Dict[int, List[int]] = {}
             for seq in seq_group.get_seqs(status=SequenceStatus.RUNNING):
                 seq_id = seq.seq_id
@@ -298,43 +295,18 @@ class Scheduler:
 
             seq_group_metadata = SequenceGroupMetadata(
                 request_id=seq_group.request_id,
-                is_prompt=is_prompt,
+                is_prompt=scheduler_outputs.prompt_run,
                 seq_data=seq_data,
                 sampling_params=seq_group.sampling_params,
                 block_tables=block_tables,
             )
             seq_group_metadata_list.append(seq_group_metadata)
-        return seq_group_metadata_list, scheduler_outputs, ignored_seq_groups
+        return seq_group_metadata_list, scheduler_outputs
 
-    def update(
-        self,
-        seq_outputs: Dict[int, SequenceOutputs],
-    ) -> List[SequenceGroup]:
-        # Update the running sequences and free blocks.
-        for seq_group in self.running:
-            # Process beam search results before processing the new tokens.
-            for seq in seq_group.get_seqs(status=SequenceStatus.RUNNING):
-                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)
+    def fork_seq(self, parent_seq: Sequence, child_seq: Sequence) -> None:
+        self.block_manager.fork(parent_seq, child_seq)
 
-            # Process the new tokens.
-            for seq in seq_group.get_seqs(status=SequenceStatus.RUNNING):
-                # Append a new token to the sequence.
-                output = seq_outputs[seq.seq_id]
-                seq.append_token_id(output.output_token, output.logprobs)
-        # Return a shallow copy of the running queue to prevent the queue
-        # from being modified by the caller.
-        return self.running.copy()
-
-    def free_seq(self, seq: Sequence, finish_status: SequenceStatus) -> None:
-        seq.status = finish_status
+    def free_seq(self, seq: Sequence) -> None:
         self.block_manager.free(seq)
 
     def free_finished_seq_groups(self) -> None:
@@ -371,8 +343,8 @@ class Scheduler:
         # 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.
+        # (e.g., beam search), recomputation is not currently 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
@@ -380,8 +352,7 @@ class Scheduler:
         # 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:
+            if seq_group.get_max_num_running_seqs() == 1:
                 preemption_mode = PreemptionMode.RECOMPUTE
             else:
                 preemption_mode = PreemptionMode.SWAP
@@ -390,7 +361,7 @@ class Scheduler:
         elif preemption_mode == PreemptionMode.SWAP:
             self._preempt_by_swap(seq_group, blocks_to_swap_out)
         else:
-            assert False, "Invalid preemption mode."
+            raise AssertionError("Invalid preemption mode.")
 
     def _preempt_by_recompute(
         self,
@@ -410,9 +381,6 @@ class Scheduler:
         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)
 

vllm/engine/arg_utils.py

@@ -13,30 +13,39 @@ class EngineArgs:
     model: str
     tokenizer: Optional[str] = None
     tokenizer_mode: str = 'auto'
+    trust_remote_code: bool = False
     download_dir: Optional[str] = None
-    use_np_weights: bool = False
-    use_dummy_weights: bool = False
+    load_format: str = 'auto'
     dtype: str = 'auto'
     seed: int = 0
+    max_model_len: Optional[int] = None
     worker_use_ray: bool = False
     pipeline_parallel_size: int = 1
     tensor_parallel_size: int = 1
+    max_parallel_loading_workers: Optional[int] = None
     block_size: int = 16
     swap_space: int = 4  # GiB
     gpu_memory_utilization: float = 0.90
-    max_num_batched_tokens: int = 2560
+    max_num_batched_tokens: Optional[int] = None
     max_num_seqs: int = 256
+    max_paddings: int = 256
     disable_log_stats: bool = False
+    revision: Optional[str] = None
+    tokenizer_revision: Optional[str] = None
+    quantization: Optional[str] = None
 
     def __post_init__(self):
         if self.tokenizer is None:
             self.tokenizer = self.model
-        self.max_num_seqs = min(self.max_num_seqs, self.max_num_batched_tokens)
 
     @staticmethod
     def add_cli_args(
             parser: argparse.ArgumentParser) -> argparse.ArgumentParser:
         """Shared CLI arguments for vLLM engine."""
+
+        # NOTE: If you update any of the arguments below, please also
+        # make sure to update docs/source/models/engine_args.rst
+
         # Model arguments
         parser.add_argument(
             '--model',
@@ -48,6 +57,20 @@ class EngineArgs:
             type=str,
             default=EngineArgs.tokenizer,
             help='name or path of the huggingface tokenizer to use')
+        parser.add_argument(
+            '--revision',
+            type=str,
+            default=None,
+            help='the specific model version to use. It can be a branch '
+            'name, a tag name, or a commit id. If unspecified, will use '
+            'the default version.')
+        parser.add_argument(
+            '--tokenizer-revision',
+            type=str,
+            default=None,
+            help='the specific tokenizer version to use. It can be a branch '
+            'name, a tag name, or a commit id. If unspecified, will use '
+            'the default version.')
         parser.add_argument('--tokenizer-mode',
                             type=str,
                             default=EngineArgs.tokenizer_mode,
@@ -55,30 +78,46 @@ class EngineArgs:
                             help='tokenizer mode. "auto" will use the fast '
                             'tokenizer if available, and "slow" will '
                             'always use the slow tokenizer.')
+        parser.add_argument('--trust-remote-code',
+                            action='store_true',
+                            help='trust remote code from huggingface')
         parser.add_argument('--download-dir',
                             type=str,
                             default=EngineArgs.download_dir,
                             help='directory to download and load the weights, '
                             'default to the default cache dir of '
                             'huggingface')
-        parser.add_argument('--use-np-weights',
-                            action='store_true',
-                            help='save a numpy copy of model weights for '
-                            'faster loading. This can increase the disk '
-                            'usage by up to 2x.')
-        parser.add_argument('--use-dummy-weights',
-                            action='store_true',
-                            help='use dummy values for model weights')
-        # TODO(woosuk): Support FP32.
+        parser.add_argument(
+            '--load-format',
+            type=str,
+            default=EngineArgs.load_format,
+            choices=['auto', 'pt', 'safetensors', 'npcache', 'dummy'],
+            help='The format of the model weights to load. '
+            '"auto" will try to load the weights in the safetensors format '
+            'and fall back to the pytorch bin format if safetensors format '
+            'is not available. '
+            '"pt" will load the weights in the pytorch bin format. '
+            '"safetensors" will load the weights in the safetensors format. '
+            '"npcache" will load the weights in pytorch format and store '
+            'a numpy cache to speed up the loading. '
+            '"dummy" will initialize the weights with random values, '
+            'which is mainly for profiling.')
         parser.add_argument(
             '--dtype',
             type=str,
             default=EngineArgs.dtype,
-            choices=['auto', 'half', 'bfloat16', 'float'],
+            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('--max-model-len',
+                            type=int,
+                            default=None,
+                            help='model context length. If unspecified, '
+                            'will be automatically derived from the model.')
         # Parallel arguments
         parser.add_argument('--worker-use-ray',
                             action='store_true',
@@ -94,6 +133,12 @@ class EngineArgs:
                             type=int,
                             default=EngineArgs.tensor_parallel_size,
                             help='number of tensor parallel replicas')
+        parser.add_argument(
+            '--max-parallel-loading-workers',
+            type=int,
+            help='load model sequentially in multiple batches, '
+            'to avoid RAM OOM when using tensor '
+            'parallel and large models')
         # KV cache arguments
         parser.add_argument('--block-size',
                             type=int,
@@ -123,9 +168,20 @@ class EngineArgs:
                             type=int,
                             default=EngineArgs.max_num_seqs,
                             help='maximum number of sequences per iteration')
+        parser.add_argument('--max-paddings',
+                            type=int,
+                            default=EngineArgs.max_paddings,
+                            help='maximum number of paddings in a batch')
         parser.add_argument('--disable-log-stats',
                             action='store_true',
                             help='disable logging statistics')
+        # Quantization settings.
+        parser.add_argument('--quantization',
+                            '-q',
+                            type=str,
+                            choices=['awq', 'squeezellm', None],
+                            default=None,
+                            help='Method used to quantize the weights')
         return parser
 
     @classmethod
@@ -139,22 +195,24 @@ class EngineArgs:
     def create_engine_configs(
         self,
     ) -> Tuple[ModelConfig, CacheConfig, ParallelConfig, SchedulerConfig]:
-        # Initialize the configs.
         model_config = ModelConfig(self.model, self.tokenizer,
-                                   self.tokenizer_mode, self.download_dir,
-                                   self.use_np_weights, self.use_dummy_weights,
-                                   self.dtype, self.seed)
+                                   self.tokenizer_mode, self.trust_remote_code,
+                                   self.download_dir, self.load_format,
+                                   self.dtype, self.seed, self.revision,
+                                   self.tokenizer_revision, self.max_model_len,
+                                   self.quantization)
         cache_config = CacheConfig(self.block_size,
                                    self.gpu_memory_utilization,
-                                   self.swap_space)
+                                   self.swap_space,
+                                   model_config.get_sliding_window())
         parallel_config = ParallelConfig(self.pipeline_parallel_size,
                                          self.tensor_parallel_size,
-                                         self.worker_use_ray)
-        model_max_len = getattr(model_config.hf_config,
-                                'max_position_embeddings', float('inf'))
-        max_seq_len = min(self.max_num_batched_tokens, model_max_len)
+                                         self.worker_use_ray,
+                                         self.max_parallel_loading_workers)
         scheduler_config = SchedulerConfig(self.max_num_batched_tokens,
-                                           self.max_num_seqs, max_seq_len)
+                                           self.max_num_seqs,
+                                           model_config.max_model_len,
+                                           self.max_paddings)
         return model_config, cache_config, parallel_config, scheduler_config
 
 
@@ -163,6 +221,7 @@ class AsyncEngineArgs(EngineArgs):
     """Arguments for asynchronous vLLM engine."""
     engine_use_ray: bool = False
     disable_log_requests: bool = False
+    max_log_len: Optional[int] = None
 
     @staticmethod
     def add_cli_args(
@@ -175,4 +234,10 @@ class AsyncEngineArgs(EngineArgs):
         parser.add_argument('--disable-log-requests',
                             action='store_true',
                             help='disable logging requests')
+        parser.add_argument('--max-log-len',
+                            type=int,
+                            default=None,
+                            help='max number of prompt characters or prompt '
+                            'ID numbers being printed in log. '
+                            'Default: unlimited.')
         return parser

vllm/engine/async_llm_engine.py

@@ -1,6 +1,8 @@
 import asyncio
 import time
-from typing import Dict, List, Optional
+from functools import partial
+from typing import (Any, Dict, Iterable, List, Optional, Set, Tuple, Type,
+                    Union)
 
 from vllm.config import ModelConfig
 from vllm.engine.arg_utils import AsyncEngineArgs
@@ -12,7 +14,218 @@ from vllm.sampling_params import SamplingParams
 
 logger = init_logger(__name__)
 
-TIMEOUT_TO_PREVENT_DEADLOCK = 1  # seconds
+
+class AsyncEngineDeadError(RuntimeError):
+    pass
+
+
+def _raise_exception_on_finish(task: asyncio.Task,
+                               request_tracker: "RequestTracker") -> None:
+    msg = ("Task finished unexpectedly. This should never happen! "
+           "Please open an issue on Github.")
+    try:
+        try:
+            task.result()
+        except asyncio.CancelledError:
+            return
+        except Exception as exc:
+            raise AsyncEngineDeadError(
+                msg + " See stack trace above for the actual cause.") from exc
+        raise AsyncEngineDeadError(msg)
+    except Exception as exc:
+        request_tracker.propagate_exception(exc)
+        raise exc
+
+
+class AsyncStream:
+    """A stream of RequestOutputs for a request that can be
+    iterated over asynchronously."""
+
+    def __init__(self, request_id: str) -> None:
+        self.request_id = request_id
+        self._queue = asyncio.Queue()
+        self._finished = False
+
+    def put(self, item: RequestOutput) -> None:
+        if self._finished:
+            return
+        self._queue.put_nowait(item)
+
+    def finish(self) -> None:
+        self._queue.put_nowait(StopIteration)
+        self._finished = True
+
+    @property
+    def finished(self) -> bool:
+        return self._finished
+
+    def __aiter__(self):
+        return self
+
+    async def __anext__(self) -> RequestOutput:
+        result = await self._queue.get()
+        if result is StopIteration:
+            raise StopAsyncIteration
+        elif isinstance(result, Exception):
+            raise result
+        return result
+
+
+class RequestTracker:
+    """Synchronous abstraction for tracking requests."""
+
+    def __init__(self) -> None:
+        self._request_streams: Dict[str, AsyncStream] = {}
+        self._finished_requests: asyncio.Queue[str] = asyncio.Queue()
+        self._new_requests: asyncio.Queue[Tuple[AsyncStream,
+                                                dict]] = asyncio.Queue()
+        self.new_requests_event = None
+
+    def __contains__(self, item):
+        return item in self._request_streams
+
+    def init_event(self):
+        self.new_requests_event = asyncio.Event()
+
+    def propagate_exception(self,
+                            exc: Exception,
+                            request_id: Optional[str] = None) -> None:
+        """Propagate an exception to request streams
+        (all if request_id is None)."""
+        if request_id is not None:
+            self._request_streams[request_id].put(exc)
+        else:
+            for stream in self._request_streams.values():
+                stream.put(exc)
+
+    def process_request_output(self,
+                               request_output: RequestOutput,
+                               *,
+                               verbose: bool = False) -> None:
+        """Process a request output from the engine."""
+        request_id = request_output.request_id
+
+        self._request_streams[request_id].put(request_output)
+        if request_output.finished:
+            if verbose:
+                logger.info(f"Finished request {request_id}.")
+            self.abort_request(request_id)
+
+    def add_request(self, request_id: str,
+                    **engine_add_request_kwargs) -> AsyncStream:
+        """Add a request to be sent to the engine on the next background
+        loop iteration."""
+        if request_id in self._request_streams:
+            raise KeyError(f"Request {request_id} already exists.")
+
+        stream = AsyncStream(request_id)
+        self._new_requests.put_nowait((stream, {
+            "request_id": request_id,
+            **engine_add_request_kwargs
+        }))
+
+        self.new_requests_event.set()
+
+        return stream
+
+    def abort_request(self, request_id: str, *, verbose: bool = False) -> None:
+        """Abort a request during next background loop iteration."""
+        if verbose:
+            logger.info(f"Aborted request {request_id}.")
+
+        self._finished_requests.put_nowait(request_id)
+
+        if request_id not in self._request_streams or self._request_streams[
+                request_id].finished:
+            # The request has already finished or been aborted.
+            return
+
+        self._request_streams[request_id].finish()
+
+    def get_new_and_finished_requests(self) -> Tuple[List[Dict], Set[str]]:
+        """Get the new requests and finished requests to be
+        sent to the engine."""
+        new_requests: List[Dict] = []
+        finished_requests: Set[str] = set()
+
+        while not self._finished_requests.empty():
+            request_id = self._finished_requests.get_nowait()
+            finished_requests.add(request_id)
+            self._request_streams.pop(request_id, None)
+
+        while not self._new_requests.empty():
+            stream, new_request = self._new_requests.get_nowait()
+            if stream.request_id in finished_requests:
+                # The request has already been aborted.
+                stream.finish()
+                continue
+            self._request_streams[stream.request_id] = stream
+            new_requests.append(new_request)
+
+        self.new_requests_event.clear()
+
+        return new_requests, finished_requests
+
+    async def wait_for_new_requests(self):
+        await self.new_requests_event.wait()
+
+
+class _AsyncLLMEngine(LLMEngine):
+    """Extension of LLMEngine to add async methods."""
+
+    async def step_async(self) -> List[RequestOutput]:
+        """Performs one decoding iteration and returns newly generated results.
+        The workers are ran asynchronously if possible.
+
+        This function performs one decoding iteration of the engine. It first
+        schedules the sequences to be executed in the next iteration and the
+        token blocks to be swapped in/out/copy. Then, it executes the model
+        and updates the scheduler with the model outputs. Finally, it decodes
+        the sequences and returns the newly generated results.
+        """
+        seq_group_metadata_list, scheduler_outputs, ignored = self._schedule()
+        if scheduler_outputs.is_empty():
+            return ignored
+
+        # Execute the model.
+        output = await self._run_workers_async(
+            "execute_model",
+            seq_group_metadata_list=seq_group_metadata_list,
+            blocks_to_swap_in=scheduler_outputs.blocks_to_swap_in,
+            blocks_to_swap_out=scheduler_outputs.blocks_to_swap_out,
+            blocks_to_copy=scheduler_outputs.blocks_to_copy,
+        )
+
+        return self._process_model_outputs(output, scheduler_outputs) + ignored
+
+    async def _run_workers_async(
+        self,
+        method: str,
+        *args,
+        get_all_outputs: bool = False,
+        **kwargs,
+    ) -> Any:
+        """Runs the given method on all workers."""
+        coros = []
+        for worker in self.workers:
+            if self.parallel_config.worker_use_ray:
+                coros.append(
+                    worker.execute_method.remote(method, *args, **kwargs))
+            else:
+                executor = getattr(worker, method)
+                coros.append(asyncio.get_event_loop().run_in_executor(
+                    None, partial(executor, *args, **kwargs)))
+
+        all_outputs = await asyncio.gather(*coros)
+
+        if get_all_outputs:
+            return all_outputs
+
+        # Make sure all workers have the same results.
+        output = all_outputs[0]
+        for other_output in all_outputs[1:]:
+            assert output == other_output
+        return output
 
 
 class AsyncLLMEngine:
@@ -34,52 +247,158 @@ class AsyncLLMEngine:
             async frontend will be executed in a separate process as the
             model workers.
         log_requests: Whether to log the requests.
+        start_engine_loop: If True, the background task to run the engine
+            will be automatically started in the generate call.
         *args, *kwargs: Arguments for LLMEngine.
     """
 
+    _engine_class: Type[_AsyncLLMEngine] = _AsyncLLMEngine
+
     def __init__(self,
                  worker_use_ray: bool,
                  engine_use_ray: bool,
                  *args,
                  log_requests: bool = True,
+                 max_log_len: Optional[int] = None,
+                 start_engine_loop: bool = True,
                  **kwargs) -> None:
         self.worker_use_ray = worker_use_ray
         self.engine_use_ray = engine_use_ray
         self.log_requests = log_requests
-        if not self.engine_use_ray:
-            engine_class = LLMEngine
-        elif self.worker_use_ray:
-            engine_class = ray.remote(num_cpus=0)(LLMEngine).remote
-        else:
-            engine_class = ray.remote(num_gpus=1)(LLMEngine).remote
-        self.engine = engine_class(*args, **kwargs)
-        # Request id -> request output.
-        self.request_outputs: Dict[str, RequestOutput] = {}
-        # Request id -> event to notify that there is new output.
-        self.request_events: Dict[str, asyncio.Event] = {}
-        self.is_engine_running = False
-        self.kicking_request_id: Optional[str] = None
+        self.max_log_len = max_log_len
+        self.engine = self._init_engine(*args, **kwargs)
+
+        self.background_loop = None
+        # We need to keep a reference to unshielded
+        # task as well to prevent it from being garbage
+        # collected
+        self._background_loop_unshielded = None
+        self.start_engine_loop = start_engine_loop
+        self._request_tracker = RequestTracker()
+
+    @property
+    def is_running(self) -> bool:
+        return (self.background_loop is not None
+                and not self.background_loop.done())
+
+    def start_background_loop(self) -> None:
+        """Start the background loop."""
+        if self.is_running:
+            raise RuntimeError("Background loop is already running.")
+        self._request_tracker.init_event()
+
+        self._background_loop_unshielded = asyncio.get_event_loop(
+        ).create_task(self.run_engine_loop())
+        self._background_loop_unshielded.add_done_callback(
+            partial(_raise_exception_on_finish,
+                    request_tracker=self._request_tracker))
+        self.background_loop = asyncio.shield(self._background_loop_unshielded)
+
+    def _init_engine(self, *args,
+                     **kwargs) -> Union[_AsyncLLMEngine, "ray.ObjectRef"]:
+        if not self.engine_use_ray:
+            engine_class = self._engine_class
+        elif self.worker_use_ray:
+            engine_class = ray.remote(num_cpus=0)(self._engine_class).remote
+        else:
+            # FIXME(woosuk): This is a bit hacky. Be careful when changing the
+            # order of the arguments.
+            cache_config = args[1]
+            parallel_config = args[2]
+            if parallel_config.tensor_parallel_size == 1:
+                num_gpus = cache_config.gpu_memory_utilization
+            else:
+                num_gpus = 1
+            engine_class = ray.remote(num_gpus=num_gpus)(
+                self._engine_class).remote
+        return engine_class(*args, **kwargs)
+
+    async def engine_step(self) -> bool:
+        """Kick the engine to process the waiting requests.
+
+        Returns True if there are in-progress requests."""
+
+        new_requests, finished_requests = (
+            self._request_tracker.get_new_and_finished_requests())
+
+        for new_request in new_requests:
+            # Add the request into the vLLM engine's waiting queue.
+            # TODO: Maybe add add_request_batch to reduce Ray overhead
+            if self.engine_use_ray:
+                await self.engine.add_request.remote(**new_request)
+            else:
+                self.engine.add_request(**new_request)
+
+        if finished_requests:
+            await self._engine_abort(finished_requests)
 
-    async def engine_step(self, kicking_request_id: Optional[str] = None):
-        """Kick the engine to process the waiting requests."""
-        self.is_engine_running = True
-        self.kicking_request_id = kicking_request_id
         if self.engine_use_ray:
             request_outputs = await self.engine.step.remote()
         else:
-            # Yield to the event loop to allow other coroutines to run
-            # while is_engine_running is True. This let the engine to add new
-            # requests into the queue.
-            await asyncio.sleep(0)
-            request_outputs = self.engine.step()
-        self.is_engine_running = False
-        self.kicking_request_id = None
+            request_outputs = await self.engine.step_async()
 
-        # Notify the waiting coroutines that there are new outputs ready.
+        # Put the outputs into the corresponding streams.
         for request_output in request_outputs:
-            request_id = request_output.request_id
-            self.request_outputs[request_id] = request_output
-            self.request_events[request_id].set()
+            self._request_tracker.process_request_output(
+                request_output, verbose=self.log_requests)
+
+        return len(request_outputs) > 0
+
+    async def _engine_abort(self, request_ids: Iterable[str]):
+        if self.engine_use_ray:
+            await self.engine.abort_request.remote(request_ids)
+        else:
+            self.engine.abort_request(request_ids)
+
+    async def run_engine_loop(self):
+        # Initialize the RequestTracker here so it uses the right event loop.
+        has_requests_in_progress = False
+        while True:
+            if not has_requests_in_progress:
+                await self._request_tracker.wait_for_new_requests()
+            has_requests_in_progress = await self.engine_step()
+            await asyncio.sleep(0)
+
+    async def add_request(
+        self,
+        request_id: str,
+        prompt: Optional[str],
+        sampling_params: SamplingParams,
+        prompt_token_ids: Optional[List[int]] = None,
+        arrival_time: Optional[float] = None,
+    ) -> AsyncStream:
+        if self.log_requests:
+            shortened_prompt = prompt
+            shortened_token_ids = prompt_token_ids
+            if self.max_log_len is not None:
+                if shortened_prompt is not None:
+                    shortened_prompt = shortened_prompt[:self.max_log_len]
+                if shortened_token_ids is not None:
+                    shortened_token_ids = shortened_token_ids[:self.
+                                                              max_log_len]
+            logger.info(f"Received request {request_id}: "
+                        f"prompt: {shortened_prompt!r}, "
+                        f"sampling params: {sampling_params}, "
+                        f"prompt token ids: {shortened_token_ids}.")
+
+        if not self.is_running:
+            if self.start_engine_loop:
+                self.start_background_loop()
+            else:
+                raise AsyncEngineDeadError(
+                    "Background loop is not running. If it was running, "
+                    "inspect the output to find the stacktrace of the "
+                    "error that caused the background loop to stop "
+                    "(AsyncEngineDeadError).")
+
+        stream = self._request_tracker.add_request(
+            request_id,
+            prompt=prompt,
+            sampling_params=sampling_params,
+            prompt_token_ids=prompt_token_ids,
+            arrival_time=arrival_time)
+
+        return stream
 
     async def generate(
             self,
@@ -106,78 +425,23 @@ class AsyncLLMEngine:
             request.
         """
         # Preprocess the request.
-        arrival_time = time.time()
+        # This should not be used for logging, as it is monotonic time.
+        arrival_time = time.monotonic()
 
-        # Create an event to notify us that there is new output from the
-        # vLLM engine.
-        request_event = asyncio.Event()
-        self.request_events[request_id] = request_event
+        try:
+            stream = await self.add_request(request_id,
+                                            prompt,
+                                            sampling_params,
+                                            prompt_token_ids=prompt_token_ids,
+                                            arrival_time=arrival_time)
 
-        if self.log_requests:
-            logger.info(f"Received request {request_id}: "
-                        f"prompt: {prompt!r}, "
-                        f"sampling params: {sampling_params}, "
-                        f"prompt token ids: {prompt_token_ids}.")
-
-        # Add the request into the vLLM engine's waiting queue.
-        if self.engine_use_ray:
-            await self.engine.add_request.remote(
-                request_id,
-                prompt,
-                sampling_params,
-                prompt_token_ids=prompt_token_ids,
-                arrival_time=arrival_time)
-        else:
-            self.engine.add_request(request_id,
-                                    prompt,
-                                    sampling_params,
-                                    prompt_token_ids=prompt_token_ids,
-                                    arrival_time=arrival_time)
-
-        # The vLLM engine does not have a background loop that keeps
-        # processing incoming requests. Therefore, we need to keep kicking
-        # the engine to process the requests.
-        while True:
-            if request_id not in self.request_events:
-                # The request has been aborted.
-                return
-
-            # Kick the engine if the engine is not running.
-            if not self.is_engine_running:
-                try:
-                    await self.engine_step(request_id)
-                except RuntimeError as e:
-                    await self.abort(request_id)
-                    raise e
-
-            # Wait for new output. The group_event will be set in engine_step
-            # when there is new output available for the sequence group.
-            # Added a timeout to prevent deadlock.
-            try:
-                await asyncio.wait_for(request_event.wait(),
-                                       timeout=TIMEOUT_TO_PREVENT_DEADLOCK)
-            except asyncio.TimeoutError:
-                continue
-            # Reset the event to wait for the next output.
-            request_event.clear()
-
-            # Decode and return new outputs.
-            request_output = self.request_outputs[request_id]
-            yield request_output
-
-            # Once finished, release the resources of the sequence group.
-            if request_output.finished:
-                if self.log_requests:
-                    logger.info(f"Finished request {request_id}.")
-
-                del self.request_outputs[request_id]
-                del self.request_events[request_id]
-                # Kick the engine if the engine is not running. This is to
-                # prevent that there are still requests in engine's waiting
-                # queue to be executed.
-                if not self.is_engine_running:
-                    await self.engine_step()
-                break
+            async for request_output in stream:
+                yield request_output
+        except (Exception, asyncio.CancelledError) as e:
+            # If there is an exception or coroutine is cancelled, abort the
+            # request.
+            self._abort(request_id)
+            raise e
 
     async def abort(self, request_id: str) -> None:
         """Abort a request.
@@ -188,28 +452,26 @@ class AsyncLLMEngine:
         Args:
             request_id: The unique id of the request.
         """
-        if request_id not in self.request_events:
-            # The request has already finished or been aborted.
-            return
+        if not self.is_running:
+            raise AsyncEngineDeadError(
+                "Background loop is not running. If it was running, "
+                "inspect the output to find the stacktrace of the "
+                "error that caused the background loop to stop "
+                "(AsyncEngineDeadError).")
 
-        if self.log_requests:
-            logger.info(f"Aborted request {request_id}.")
+        return self._abort(request_id)
 
-        if self.engine_use_ray:
-            await self.engine.abort_request.remote(request_id)
-        else:
-            self.engine.abort_request(request_id)
+    def _abort(self, request_id: str) -> None:
+        """Abort a request.
 
-        if request_id in self.request_events:
-            del self.request_events[request_id]
-        if request_id in self.request_outputs:
-            del self.request_outputs[request_id]
+        Abort a submitted request. If the request is finished or not found,
+        this method will be a no-op.
 
-        # To prevent deadlock when a request is aborted while the engine is
-        # running.
-        if self.kicking_request_id == request_id:
-            self.is_engine_running = False
-            self.kicking_request_id = None
+        Args:
+            request_id: The unique id of the request.
+        """
+        self._request_tracker.abort_request(request_id,
+                                            verbose=self.log_requests)
 
     async def get_model_config(self) -> ModelConfig:
         """Get the model configuration of the vLLM engine."""
@@ -220,20 +482,23 @@ class AsyncLLMEngine:
 
     @classmethod
     def from_engine_args(cls,
-                         engine_args: AsyncEngineArgs) -> "AsyncLLMEngine":
+                         engine_args: AsyncEngineArgs,
+                         start_engine_loop: bool = True) -> "AsyncLLMEngine":
         """Creates an async LLM engine from the engine arguments."""
         # Create the engine configs.
         engine_configs = engine_args.create_engine_configs()
         parallel_config = engine_configs[2]
         # Initialize the cluster.
-        distributed_init_method, devices = initialize_cluster(
+        distributed_init_method, placement_group = initialize_cluster(
             parallel_config, engine_args.engine_use_ray)
         # Create the async LLM engine.
-        engine = cls(engine_args.worker_use_ray,
+        engine = cls(parallel_config.worker_use_ray,
                      engine_args.engine_use_ray,
                      *engine_configs,
                      distributed_init_method,
-                     devices,
+                     placement_group,
                      log_requests=not engine_args.disable_log_requests,
-                     log_stats=not engine_args.disable_log_stats)
+                     log_stats=not engine_args.disable_log_stats,
+                     max_log_len=engine_args.max_log_len,
+                     start_engine_loop=start_engine_loop)
         return engine

vllm/engine/llm_engine.py

@@ -1,22 +1,35 @@
+import copy
 import time
-from typing import Any, List, Optional
+from functools import partial
+from typing import TYPE_CHECKING, Any, Iterable, List, Optional, Tuple, Union
 
 from vllm.config import (CacheConfig, ModelConfig, ParallelConfig,
                          SchedulerConfig)
-from vllm.core.scheduler import Scheduler
+from vllm.core.scheduler import Scheduler, SchedulerOutputs
 from vllm.engine.arg_utils import EngineArgs
-from vllm.engine.ray_utils import DeviceID, initialize_cluster, ray
+from vllm.engine.metrics import record_metrics
+from vllm.engine.ray_utils import RayWorkerVllm, initialize_cluster, ray
 from vllm.logger import init_logger
 from vllm.outputs import RequestOutput
 from vllm.sampling_params import SamplingParams
-from vllm.sequence import Sequence, SequenceGroup, SequenceStatus
+from vllm.sequence import (SamplerOutput, Sequence, SequenceGroup,
+                           SequenceGroupMetadata, SequenceGroupOutput,
+                           SequenceOutput, SequenceStatus)
 from vllm.transformers_utils.tokenizer import (detokenize_incrementally,
                                                get_tokenizer)
 from vllm.utils import Counter
-from vllm.worker.worker import Worker
+
+if ray:
+    from ray.air.util.torch_dist import init_torch_dist_process_group
+    from ray.util.scheduling_strategies import PlacementGroupSchedulingStrategy
+
+if TYPE_CHECKING:
+    from ray.util.placement_group import PlacementGroup
 
 logger = init_logger(__name__)
 
+_LOGGING_INTERVAL_SEC = 5
+
 
 class LLMEngine:
     """An LLM engine that receives requests and generates texts.
@@ -42,8 +55,8 @@ class LLMEngine:
         scheduler_config: The configuration related to the request scheduler.
         distributed_init_method: The initialization method for distributed
             execution. See `torch.distributed.init_process_group` for details.
-        stage_devices: The list of devices for each stage. Each stage is a list
-            of (rank, node_resource, device) tuples.
+        placement_group: Ray placement group for distributed execution.
+            Required for distributed execution.
         log_stats: Whether to log statistics.
     """
 
@@ -54,7 +67,7 @@ class LLMEngine:
         parallel_config: ParallelConfig,
         scheduler_config: SchedulerConfig,
         distributed_init_method: str,
-        stage_devices: List[List[DeviceID]],
+        placement_group: Optional["PlacementGroup"],
         log_stats: bool,
     ) -> None:
         logger.info(
@@ -62,11 +75,15 @@ class LLMEngine:
             f"model={model_config.model!r}, "
             f"tokenizer={model_config.tokenizer!r}, "
             f"tokenizer_mode={model_config.tokenizer_mode}, "
+            f"revision={model_config.revision}, "
+            f"tokenizer_revision={model_config.tokenizer_revision}, "
+            f"trust_remote_code={model_config.trust_remote_code}, "
             f"dtype={model_config.dtype}, "
-            f"use_dummy_weights={model_config.use_dummy_weights}, "
+            f"max_seq_len={model_config.max_model_len}, "
             f"download_dir={model_config.download_dir!r}, "
-            f"use_np_weights={model_config.use_np_weights}, "
+            f"load_format={model_config.load_format}, "
             f"tensor_parallel_size={parallel_config.tensor_parallel_size}, "
+            f"quantization={model_config.quantization}, "
             f"seed={model_config.seed})")
         # TODO(woosuk): Print more configs in debug mode.
 
@@ -78,34 +95,108 @@ class LLMEngine:
         self._verify_args()
 
         self.tokenizer = get_tokenizer(
-            model_config.tokenizer, tokenizer_mode=model_config.tokenizer_mode)
+            model_config.tokenizer,
+            tokenizer_mode=model_config.tokenizer_mode,
+            trust_remote_code=model_config.trust_remote_code,
+            tokenizer_revision=model_config.tokenizer_revision,
+            revision=model_config.revision)
         self.seq_counter = Counter()
 
         # Create the parallel GPU workers.
-        self.workers: List[Worker] = []
-        assert len(stage_devices) == 1, "Only support one stage for now."
-        for rank, node_resource, _ in stage_devices[0]:
-            worker_cls = Worker
-            if self.parallel_config.worker_use_ray:
-                worker_cls = ray.remote(
-                    num_cpus=0,
-                    num_gpus=1,
-                    resources={node_resource: 1e-3},
-                )(worker_cls).remote
+        if self.parallel_config.worker_use_ray:
+            self._init_workers_ray(placement_group)
+        else:
+            self._init_workers(distributed_init_method)
 
-            worker = worker_cls(
-                model_config,
-                parallel_config,
-                scheduler_config,
-                rank,
-                distributed_init_method,
-            )
-            self.workers.append(worker)
         # Profile the memory usage and initialize the cache.
         self._init_cache()
 
         # Create the scheduler.
-        self.scheduler = Scheduler(scheduler_config, cache_config, log_stats)
+        self.scheduler = Scheduler(scheduler_config, cache_config)
+
+        # Logging.
+        self.last_logging_time = 0.0
+        # List of (timestamp, num_tokens)
+        self.num_prompt_tokens: List[Tuple[float, int]] = []
+        # List of (timestamp, num_tokens)
+        self.num_generation_tokens: List[Tuple[float, int]] = []
+
+    def _init_workers(self, distributed_init_method: str):
+        # Lazy import the Worker to avoid importing torch.cuda/xformers
+        # before CUDA_VISIBLE_DEVICES is set in the Worker
+        from vllm.worker.worker import Worker
+
+        assert self.parallel_config.world_size == 1, (
+            "Ray is required if parallel_config.world_size > 1.")
+
+        self.workers: List[Worker] = []
+        worker = Worker(
+            self.model_config,
+            self.parallel_config,
+            self.scheduler_config,
+            0,
+            distributed_init_method,
+        )
+        self.workers.append(worker)
+        self._run_workers(
+            "init_model",
+            get_all_outputs=True,
+        )
+        self._run_workers(
+            "load_model",
+            get_all_outputs=True,
+            max_concurrent_workers=self.parallel_config.
+            max_parallel_loading_workers,
+        )
+
+    def _init_workers_ray(self, placement_group: "PlacementGroup",
+                          **ray_remote_kwargs):
+        # Lazy import the Worker to avoid importing torch.cuda/xformers
+        # before CUDA_VISIBLE_DEVICES is set in the Worker
+        from vllm.worker.worker import Worker
+
+        self.workers: List[Worker] = []
+        for bundle in placement_group.bundle_specs:
+            if not bundle.get("GPU", 0):
+                continue
+            if self.parallel_config.tensor_parallel_size == 1:
+                num_gpus = self.cache_config.gpu_memory_utilization
+            else:
+                num_gpus = 1
+            worker = ray.remote(
+                num_cpus=0,
+                num_gpus=num_gpus,
+                scheduling_strategy=PlacementGroupSchedulingStrategy(
+                    placement_group=placement_group,
+                    placement_group_capture_child_tasks=True),
+                **ray_remote_kwargs,
+            )(RayWorkerVllm).remote(self.model_config.trust_remote_code)
+            self.workers.append(worker)
+
+        # Initialize torch distributed process group for the workers.
+        init_torch_dist_process_group(self.workers, backend="nccl")
+        model_config = copy.deepcopy(self.model_config)
+        parallel_config = copy.deepcopy(self.parallel_config)
+        scheduler_config = copy.deepcopy(self.scheduler_config)
+        self._run_workers("init_worker",
+                          get_all_outputs=True,
+                          worker_init_fn=lambda: Worker(
+                              model_config,
+                              parallel_config,
+                              scheduler_config,
+                              None,
+                              None,
+                          ))
+        self._run_workers(
+            "init_model",
+            get_all_outputs=True,
+        )
+        self._run_workers(
+            "load_model",
+            get_all_outputs=True,
+            max_concurrent_workers=self.parallel_config.
+            max_parallel_loading_workers,
+        )
 
     def _verify_args(self) -> None:
         self.model_config.verify_with_parallel_config(self.parallel_config)
@@ -149,11 +240,12 @@ class LLMEngine:
         engine_configs = engine_args.create_engine_configs()
         parallel_config = engine_configs[2]
         # Initialize the cluster.
-        distributed_init_method, devices = initialize_cluster(parallel_config)
+        distributed_init_method, placement_group = initialize_cluster(
+            parallel_config)
         # Create the LLM engine.
         engine = cls(*engine_configs,
                      distributed_init_method,
-                     devices,
+                     placement_group,
                      log_stats=not engine_args.disable_log_stats)
         return engine
 
@@ -179,34 +271,31 @@ class LLMEngine:
             prompt_token_ids: The token IDs of the prompt. If None, we
                 use the tokenizer to convert the prompts to token IDs.
             arrival_time: The arrival time of the request. If None, we use
-                the current time.
+                the current monotonic time.
         """
         if arrival_time is None:
-            arrival_time = time.time()
+            arrival_time = time.monotonic()
         if prompt_token_ids is None:
             assert prompt is not None
             prompt_token_ids = self.tokenizer.encode(prompt)
 
         # Create the sequences.
         block_size = self.cache_config.block_size
-        seqs: List[Sequence] = []
-        for _ in range(sampling_params.best_of):
-            seq_id = next(self.seq_counter)
-            seq = Sequence(seq_id, prompt, prompt_token_ids, block_size)
-            seqs.append(seq)
+        seq_id = next(self.seq_counter)
+        seq = Sequence(seq_id, prompt, prompt_token_ids, block_size)
 
         # Create the sequence group.
-        seq_group = SequenceGroup(request_id, seqs, sampling_params,
+        seq_group = SequenceGroup(request_id, [seq], sampling_params,
                                   arrival_time)
 
         # Add the sequence group to the scheduler.
         self.scheduler.add_seq_group(seq_group)
 
-    def abort_request(self, request_id: str) -> None:
-        """Aborts a request with the given ID.
+    def abort_request(self, request_id: Union[str, Iterable[str]]) -> None:
+        """Aborts a request(s) with the given ID.
 
         Args:
-            request_id: The ID of the request to abort.
+            request_id: The ID(s) of the request to abort.
         """
         self.scheduler.abort_seq_group(request_id)
 
@@ -222,6 +311,255 @@ class LLMEngine:
         """Returns True if there are unfinished requests."""
         return self.scheduler.has_unfinished_seqs()
 
+    def _schedule(
+        self
+    ) -> Tuple[List[SequenceGroupMetadata], SchedulerOutputs,
+               List[RequestOutput]]:
+        seq_group_metadata_list, scheduler_outputs = self.scheduler.schedule()
+        return seq_group_metadata_list, scheduler_outputs, [
+            RequestOutput.from_seq_group(seq_group)
+            for seq_group in scheduler_outputs.ignored_seq_groups
+        ]
+
+    def _check_beam_search_early_stopping(
+        self,
+        early_stopping: Union[bool, str],
+        sampling_params: SamplingParams,
+        best_running_seq: Sequence,
+        current_worst_seq: Sequence,
+    ) -> bool:
+        assert sampling_params.use_beam_search
+        length_penalty = sampling_params.length_penalty
+        if early_stopping is True:
+            return True
+
+        current_worst_score = (current_worst_seq.get_beam_search_score(
+            length_penalty=length_penalty,
+            eos_token_id=self.tokenizer.eos_token_id))
+        if early_stopping is False:
+            highest_attainable_score = (best_running_seq.get_beam_search_score(
+                length_penalty=length_penalty,
+                eos_token_id=self.tokenizer.eos_token_id))
+        else:
+            assert early_stopping == "never"
+            if length_penalty > 0.0:
+                # If length_penalty > 0.0, beam search will prefer longer
+                # sequences. The highest attainable score calculation is
+                # based on the longest possible sequence length in this case.
+                max_possible_length = max(
+                    best_running_seq.get_prompt_len() +
+                    sampling_params.max_tokens,
+                    self.scheduler_config.max_model_len)
+                highest_attainable_score = (
+                    best_running_seq.get_beam_search_score(
+                        length_penalty=length_penalty,
+                        eos_token_id=self.tokenizer.eos_token_id,
+                        seq_len=max_possible_length))
+            else:
+                # Otherwise, beam search will prefer shorter sequences. The
+                # highest attainable score calculation is based on the current
+                # sequence length.
+                highest_attainable_score = (
+                    best_running_seq.get_beam_search_score(
+                        length_penalty=length_penalty,
+                        eos_token_id=self.tokenizer.eos_token_id))
+        return current_worst_score >= highest_attainable_score
+
+    def _process_sequence_group_outputs(self, seq_group: SequenceGroup,
+                                        outputs: SequenceGroupOutput) -> None:
+        # Process prompt logprobs
+        prompt_logprobs = outputs.prompt_logprobs
+        if prompt_logprobs is not None:
+            seq_group.prompt_logprobs = prompt_logprobs
+
+        # Process samples
+        samples = outputs.samples
+        parent_seqs = seq_group.get_seqs(status=SequenceStatus.RUNNING)
+        existing_finished_seqs = seq_group.get_finished_seqs()
+        parent_child_dict = {
+            parent_seq.seq_id: []
+            for parent_seq in parent_seqs
+        }
+        for sample in samples:
+            parent_child_dict[sample.parent_seq_id].append(sample)
+        # List of (child, parent)
+        child_seqs: List[Tuple[Sequence, Sequence]] = []
+
+        # Process the child samples for each parent sequence
+        for parent in parent_seqs:
+            child_samples: List[SequenceOutput] = parent_child_dict[
+                parent.seq_id]
+            if len(child_samples) == 0:
+                # This parent sequence has no children samples. Remove
+                # the parent sequence from the sequence group since it will
+                # not be used in the future iterations.
+                parent.status = SequenceStatus.FINISHED_ABORTED
+                seq_group.remove(parent.seq_id)
+                self.scheduler.free_seq(parent)
+                continue
+            # Fork the parent sequence if there are multiple child samples.
+            for child_sample in child_samples[:-1]:
+                new_child_seq_id = next(self.seq_counter)
+                child = parent.fork(new_child_seq_id)
+                child.append_token_id(child_sample.output_token,
+                                      child_sample.logprobs)
+                child_seqs.append((child, parent))
+            # Continue the parent sequence for the last child sample.
+            # We reuse the parent sequence here to reduce redundant memory
+            # copies, especially when using non-beam search sampling methods.
+            last_child_sample = child_samples[-1]
+            parent.append_token_id(last_child_sample.output_token,
+                                   last_child_sample.logprobs)
+            child_seqs.append((parent, parent))
+
+        for seq, _ in child_seqs:
+            self._decode_sequence(seq, seq_group.sampling_params)
+            self._check_stop(seq, seq_group.sampling_params)
+
+        # Non-beam search case
+        if not seq_group.sampling_params.use_beam_search:
+            # For newly created child sequences, add them to the sequence group
+            # and fork them in block manager if they are not finished.
+            for seq, parent in child_seqs:
+                if seq is not parent:
+                    seq_group.add(seq)
+                    if not seq.is_finished():
+                        self.scheduler.fork_seq(parent, seq)
+
+            # Free the finished and selected parent sequences' memory in block
+            # manager. Keep them in the sequence group as candidate output.
+            # NOTE: we need to fork the new sequences before freeing the
+            # old sequences.
+            for seq, parent in child_seqs:
+                if seq is parent and seq.is_finished():
+                    self.scheduler.free_seq(seq)
+            return
+
+        # Beam search case
+        # Select the child sequences to keep in the sequence group.
+        selected_child_seqs = []
+        unselected_child_seqs = []
+        beam_width = seq_group.sampling_params.best_of
+        length_penalty = seq_group.sampling_params.length_penalty
+
+        # Select the newly finished sequences with the highest scores
+        # to replace existing finished sequences.
+        # Tuple of (seq, parent, is_new)
+        existing_finished_seqs = [(seq, None, False)
+                                  for seq in existing_finished_seqs]
+        new_finished_seqs = [(seq, parent, True) for seq, parent in child_seqs
+                             if seq.is_finished()]
+        all_finished_seqs = existing_finished_seqs + new_finished_seqs
+        # Sort the finished sequences by their scores.
+        all_finished_seqs.sort(key=lambda x: x[0].get_beam_search_score(
+            length_penalty=length_penalty,
+            eos_token_id=self.tokenizer.eos_token_id),
+                               reverse=True)
+        for seq, parent, is_new in all_finished_seqs[:beam_width]:
+            if is_new:
+                # A newly generated child sequence finishes and has a high
+                # score, so we will add it into the sequence group.
+                selected_child_seqs.append((seq, parent))
+        for seq, parent, is_new in all_finished_seqs[beam_width:]:
+            if is_new:
+                # A newly generated child sequence finishes but has a low
+                # score, so we will not add it into the sequence group.
+                # Additionally, if this sequence is a continuation of a
+                # parent sequence, we will need remove the parent sequence
+                # from the sequence group.
+                unselected_child_seqs.append((seq, parent))
+            else:
+                # An existing finished sequence has a low score, so we will
+                # remove it from the sequence group.
+                seq_group.remove(seq.seq_id)
+
+        # select the top beam_width sequences from the running
+        # sequences for the next iteration to continue the beam
+        # search.
+        running_child_seqs = [(seq, parent) for seq, parent in child_seqs
+                              if not seq.is_finished()]
+        # Sort the running sequences by their scores.
+        running_child_seqs.sort(key=lambda x: x[0].get_beam_search_score(
+            length_penalty=length_penalty,
+            eos_token_id=self.tokenizer.eos_token_id),
+                                reverse=True)
+
+        # Check if we can stop the beam search.
+        if len(running_child_seqs) == 0:
+            # No running sequences, stop the beam search.
+            stop_beam_search = True
+        elif len(all_finished_seqs) < beam_width:
+            # Not enough finished sequences, continue the beam search.
+            stop_beam_search = False
+        else:
+            # Check the early stopping criteria
+            best_running_seq = running_child_seqs[0][0]
+            current_worst_seq = all_finished_seqs[beam_width - 1][0]
+            stop_beam_search = self._check_beam_search_early_stopping(
+                seq_group.sampling_params.early_stopping,
+                seq_group.sampling_params, best_running_seq, current_worst_seq)
+
+        if stop_beam_search:
+            # Stop the beam search and remove all the running sequences from
+            # the sequence group.
+            unselected_child_seqs.extend(running_child_seqs)
+        else:
+            # Continue the beam search and select the top beam_width sequences
+            # to continue the beam search.
+            selected_child_seqs.extend(running_child_seqs[:beam_width])
+            # The remaining running sequences will not be used in the next
+            # iteration. Again, if these sequences are continuations of
+            # parent sequences, we will need to remove the parent sequences
+            # from the sequence group.
+            unselected_child_seqs.extend(running_child_seqs[beam_width:])
+
+        # For newly created child sequences, add them to the sequence group
+        # and fork them in block manager if they are not finished.
+        for seq, parent in selected_child_seqs:
+            if seq is not parent:
+                seq_group.add(seq)
+                if not seq.is_finished():
+                    self.scheduler.fork_seq(parent, seq)
+
+        # Free the finished and selected parent sequences' memory in block
+        # manager. Keep them in the sequence group as candidate output.
+        for seq, parent in selected_child_seqs:
+            if seq is parent and seq.is_finished():
+                self.scheduler.free_seq(seq)
+
+        # Remove the unselected parent sequences from the sequence group and
+        # free their memory in block manager.
+        for seq, parent in unselected_child_seqs:
+            if seq is parent:
+                # Remove the parent sequence if it is not selected for next
+                # iteration
+                seq_group.remove(seq.seq_id)
+                self.scheduler.free_seq(seq)
+
+    def _process_model_outputs(
+            self, output: SamplerOutput,
+            scheduler_outputs: SchedulerOutputs) -> List[RequestOutput]:
+        # Update the scheduled sequence groups with the model outputs.
+        scheduled_seq_groups = scheduler_outputs.scheduled_seq_groups
+        for seq_group, outputs in zip(scheduled_seq_groups, output):
+            self._process_sequence_group_outputs(seq_group, outputs)
+
+        # Free the finished sequence groups.
+        self.scheduler.free_finished_seq_groups()
+
+        # Create the outputs.
+        request_outputs: List[RequestOutput] = []
+        for seq_group in (scheduled_seq_groups +
+                          scheduler_outputs.ignored_seq_groups):
+            request_output = RequestOutput.from_seq_group(seq_group)
+            request_outputs.append(request_output)
+
+        if self.log_stats:
+            # Log the system stats.
+            self._log_system_stats(scheduler_outputs.prompt_run,
+                                   scheduler_outputs.num_batched_tokens)
+        return request_outputs
+
     def step(self) -> List[RequestOutput]:
         """Performs one decoding iteration and returns newly generated results.
 
@@ -231,12 +569,9 @@ class LLMEngine:
         and updates the scheduler with the model outputs. Finally, it decodes
         the sequences and returns the newly generated results.
         """
-        (seq_group_metadata_list, scheduler_outputs,
-         ignored_seq_groups) = self.scheduler.schedule()
-        if ((not seq_group_metadata_list) and scheduler_outputs.is_empty()
-                and (not ignored_seq_groups)):
-            # Nothing to do.
-            return []
+        seq_group_metadata_list, scheduler_outputs, ignored = self._schedule()
+        if scheduler_outputs.is_empty():
+            return ignored
 
         # Execute the model.
         output = self._run_workers(
@@ -246,92 +581,173 @@ class LLMEngine:
             blocks_to_swap_out=scheduler_outputs.blocks_to_swap_out,
             blocks_to_copy=scheduler_outputs.blocks_to_copy,
         )
-        # Update the scheduler with the model outputs.
-        seq_groups = self.scheduler.update(output)
 
-        # Decode the sequences.
-        self._decode_sequences(seq_groups)
-        # Stop the sequences that meet the stopping criteria.
-        self._stop_sequences(seq_groups)
-        # Free the finished sequence groups.
-        self.scheduler.free_finished_seq_groups()
+        return self._process_model_outputs(output, scheduler_outputs)
 
-        # Create the outputs.
-        request_outputs: List[RequestOutput] = []
-        for seq_group in seq_groups + ignored_seq_groups:
-            request_output = RequestOutput.from_seq_group(seq_group)
-            request_outputs.append(request_output)
-        return request_outputs
+    def _log_system_stats(
+        self,
+        prompt_run: bool,
+        num_batched_tokens: int,
+    ) -> None:
+        now = time.monotonic()
+        # Log the number of batched input tokens.
+        if prompt_run:
+            self.num_prompt_tokens.append((now, num_batched_tokens))
+        else:
+            self.num_generation_tokens.append((now, num_batched_tokens))
 
-    def _decode_sequences(self, seq_groups: List[SequenceGroup]) -> None:
-        """Decodes the sequence outputs."""
-        for seq_group in seq_groups:
-            for seq in seq_group.get_seqs(status=SequenceStatus.RUNNING):
-                new_token, new_output_text = detokenize_incrementally(
-                    self.tokenizer,
-                    seq.output_tokens,
-                    seq.get_last_token_id(),
-                    skip_special_tokens=True,
-                )
-                seq.output_tokens.append(new_token)
-                seq.output_text = new_output_text
+        should_log = now - self.last_logging_time >= _LOGGING_INTERVAL_SEC
+        if not should_log:
+            return
 
-    def _stop_sequences(self, seq_groups: List[SequenceGroup]) -> None:
+        # Discard the old stats.
+        self.num_prompt_tokens = [(t, n) for t, n in self.num_prompt_tokens
+                                  if now - t < _LOGGING_INTERVAL_SEC]
+        self.num_generation_tokens = [(t, n)
+                                      for t, n in self.num_generation_tokens
+                                      if now - t < _LOGGING_INTERVAL_SEC]
+
+        if len(self.num_prompt_tokens) > 1:
+            total_num_tokens = sum(n for _, n in self.num_prompt_tokens[:-1])
+            window = now - self.num_prompt_tokens[0][0]
+            avg_prompt_throughput = total_num_tokens / window
+        else:
+            avg_prompt_throughput = 0.0
+        if len(self.num_generation_tokens) > 1:
+            total_num_tokens = sum(n
+                                   for _, n in self.num_generation_tokens[:-1])
+            window = now - self.num_generation_tokens[0][0]
+            avg_generation_throughput = total_num_tokens / window
+        else:
+            avg_generation_throughput = 0.0
+
+        total_num_gpu_blocks = self.cache_config.num_gpu_blocks
+        num_free_gpu_blocks = (
+            self.scheduler.block_manager.get_num_free_gpu_blocks())
+        num_used_gpu_blocks = total_num_gpu_blocks - num_free_gpu_blocks
+        gpu_cache_usage = num_used_gpu_blocks / total_num_gpu_blocks
+
+        total_num_cpu_blocks = self.cache_config.num_cpu_blocks
+        if total_num_cpu_blocks > 0:
+            num_free_cpu_blocks = (
+                self.scheduler.block_manager.get_num_free_cpu_blocks())
+            num_used_cpu_blocks = total_num_cpu_blocks - num_free_cpu_blocks
+            cpu_cache_usage = num_used_cpu_blocks / total_num_cpu_blocks
+        else:
+            cpu_cache_usage = 0.0
+
+        record_metrics(
+            avg_prompt_throughput=avg_prompt_throughput,
+            avg_generation_throughput=avg_generation_throughput,
+            scheduler_running=len(self.scheduler.running),
+            scheduler_swapped=len(self.scheduler.swapped),
+            scheduler_waiting=len(self.scheduler.waiting),
+            gpu_cache_usage=gpu_cache_usage,
+            cpu_cache_usage=cpu_cache_usage,
+        )
+
+        logger.info("Avg prompt throughput: "
+                    f"{avg_prompt_throughput:.1f} tokens/s, "
+                    "Avg generation throughput: "
+                    f"{avg_generation_throughput:.1f} tokens/s, "
+                    f"Running: {len(self.scheduler.running)} reqs, "
+                    f"Swapped: {len(self.scheduler.swapped)} reqs, "
+                    f"Pending: {len(self.scheduler.waiting)} reqs, "
+                    f"GPU KV cache usage: {gpu_cache_usage * 100:.1f}%, "
+                    f"CPU KV cache usage: {cpu_cache_usage * 100:.1f}%")
+        self.last_logging_time = now
+
+    def _decode_sequence(self, seq: Sequence, prms: SamplingParams) -> None:
+        """Decodes the new token for a sequence."""
+        (new_tokens, new_output_text, prefix_offset,
+         read_offset) = detokenize_incrementally(
+             self.tokenizer,
+             all_input_ids=seq.get_token_ids(),
+             prev_tokens=seq.tokens,
+             prefix_offset=seq.prefix_offset,
+             read_offset=seq.read_offset,
+             skip_special_tokens=prms.skip_special_tokens,
+             spaces_between_special_tokens=prms.spaces_between_special_tokens,
+         )
+        if seq.tokens is None:
+            seq.tokens = new_tokens
+        else:
+            seq.tokens.extend(new_tokens)
+        seq.prefix_offset = prefix_offset
+        seq.read_offset = read_offset
+        seq.output_text += new_output_text
+
+    def _check_stop(self, seq: Sequence,
+                    sampling_params: SamplingParams) -> None:
         """Stop the finished sequences."""
-        for seq_group in seq_groups:
-            sampling_params = seq_group.sampling_params
-            for seq in seq_group.get_seqs(status=SequenceStatus.RUNNING):
-                # Check if the sequence has generated a stop string.
-                stopped = False
-                for stop_str in sampling_params.stop:
-                    if seq.output_text.endswith(stop_str):
-                        # Truncate the output text so that the stop string is
-                        # not included in the output.
-                        seq.output_text = seq.output_text[:-len(stop_str)]
-                        self.scheduler.free_seq(
-                            seq, SequenceStatus.FINISHED_STOPPED)
-                        stopped = True
-                        break
-                if stopped:
-                    continue
+        for stop_str in sampling_params.stop:
+            if seq.output_text.endswith(stop_str):
+                # Truncate the output text so that the stop string is
+                # not included in the output.
+                seq.output_text = seq.output_text[:-len(stop_str)]
+                seq.status = SequenceStatus.FINISHED_STOPPED
+                return
+        if seq.get_last_token_id() in sampling_params.stop_token_ids:
+            seq.status = SequenceStatus.FINISHED_STOPPED
+            return
 
-                # Check if the sequence has reached max_seq_len.
-                if (seq.get_len() >=
-                        self.scheduler.scheduler_config.max_seq_len):
-                    self.scheduler.free_seq(
-                        seq, SequenceStatus.FINISHED_LENGTH_CAPPED)
-                    continue
-                # Check if the sequence has reached max_tokens.
-                if seq.get_output_len() == sampling_params.max_tokens:
-                    self.scheduler.free_seq(
-                        seq, SequenceStatus.FINISHED_LENGTH_CAPPED)
-                    continue
-                # Check if the sequence has generated the EOS token.
-                if not sampling_params.ignore_eos:
-                    if seq.get_last_token_id() == self.tokenizer.eos_token_id:
-                        self.scheduler.free_seq(
-                            seq, SequenceStatus.FINISHED_STOPPED)
-                        continue
+        # Check if the sequence has reached max_model_len.
+        if seq.get_len() > self.scheduler_config.max_model_len:
+            seq.status = SequenceStatus.FINISHED_LENGTH_CAPPED
+            return
+
+        # Check if the sequence has reached max_tokens.
+        if seq.get_output_len() == sampling_params.max_tokens:
+            seq.status = SequenceStatus.FINISHED_LENGTH_CAPPED
+            return
+
+        # Check if the sequence has generated the EOS token.
+        if ((not sampling_params.ignore_eos)
+                and seq.get_last_token_id() == self.tokenizer.eos_token_id):
+            seq.status = SequenceStatus.FINISHED_STOPPED
+            return
+
+    def _run_workers_in_batch(
+        self,
+        workers,
+        method: str,
+        *args,
+        **kwargs,
+    ):
+        all_outputs = []
+        for worker in workers:
+            if self.parallel_config.worker_use_ray:
+                executor = partial(worker.execute_method.remote, method)
+            else:
+                executor = getattr(worker, method)
+
+            output = executor(*args, **kwargs)
+            all_outputs.append(output)
+        if self.parallel_config.worker_use_ray:
+            all_outputs = ray.get(all_outputs)
+        return all_outputs
 
     def _run_workers(
         self,
         method: str,
         *args,
         get_all_outputs: bool = False,
+        max_concurrent_workers: Optional[int] = None,
         **kwargs,
     ) -> Any:
         """Runs the given method on all workers."""
         all_outputs = []
-        for worker in self.workers:
-            executor = getattr(worker, method)
-            if self.parallel_config.worker_use_ray:
-                executor = executor.remote
+        if max_concurrent_workers:
+            work_groups = [
+                self.workers[i:i + max_concurrent_workers]
+                for i in range(0, len(self.workers), max_concurrent_workers)
+            ]
+        else:
+            work_groups = [self.workers]
 
-            output = executor(*args, **kwargs)
-            all_outputs.append(output)
-
-        if self.parallel_config.worker_use_ray:
-            all_outputs = ray.get(all_outputs)
+        for workers in work_groups:
+            all_outputs.extend(
+                self._run_workers_in_batch(workers, method, *args, **kwargs))
 
         if get_all_outputs:
             return all_outputs

vllm/engine/metrics.py

@@ -0,0 +1,51 @@
+from aioprometheus import Gauge
+
+# The begin-* and end* here are used by the documentation generator
+# to extract the metrics definitions.
+
+# begin-metrics-definitions
+gauge_avg_prompt_throughput = Gauge("vllm:avg_prompt_throughput_toks_per_s",
+                                    "Average prefill throughput in tokens/s.")
+gauge_avg_generation_throughput = Gauge(
+    "vllm:avg_generation_throughput_toks_per_s",
+    "Average generation throughput in tokens/s.")
+
+gauge_scheduler_running = Gauge(
+    "vllm:num_requests_running",
+    "Number of requests that is currently running for inference.")
+gauge_scheduler_swapped = Gauge("vllm:num_requests_swapped",
+                                "Number requests swapped to CPU.")
+gauge_scheduler_waiting = Gauge("vllm:num_requests_waiting",
+                                "Number of requests waiting to be processed.")
+
+gauge_gpu_cache_usage = Gauge(
+    "vllm:gpu_cache_usage_perc",
+    "GPU KV-cache usage. 1 means 100 percent usage.")
+gauge_cpu_cache_usage = Gauge(
+    "vllm:cpu_cache_usage_perc",
+    "CPU KV-cache usage. 1 means 100 percent usage.")
+# end-metrics-definitions
+
+labels = {}
+
+
+def add_global_metrics_labels(**kwargs):
+    labels.update(kwargs)
+
+
+def record_metrics(
+    avg_prompt_throughput: float,
+    avg_generation_throughput: float,
+    scheduler_running: int,
+    scheduler_swapped: int,
+    scheduler_waiting: int,
+    gpu_cache_usage: float,
+    cpu_cache_usage: float,
+):
+    gauge_avg_prompt_throughput.set(labels, avg_prompt_throughput)
+    gauge_avg_generation_throughput.set(labels, avg_generation_throughput)
+    gauge_scheduler_running.set(labels, scheduler_running)
+    gauge_scheduler_swapped.set(labels, scheduler_swapped)
+    gauge_scheduler_waiting.set(labels, scheduler_waiting)
+    gauge_gpu_cache_usage.set(labels, gpu_cache_usage)
+    gauge_cpu_cache_usage.set(labels, cpu_cache_usage)