Bump up the version to v0.2.1 (#1355)

Co-authored-by: Yunmo Chen <16273544+wanmok@users.noreply.github.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: ['2.0.1']
+          cuda-version: ['11.8'] # Github runner can't build anything older than 11.8
+
+    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/pylint.yml

@@ -0,0 +1,31 @@
+name: pylint
+
+on:
+  # Trigger the workflow on push or pull request,
+  # but only for the main branch
+  push:
+    branches:
+      - main
+  pull_request:
+    branches:
+      - main
+
+jobs:
+  pylint:
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        python-version: ["3.10"]
+    steps:
+    - uses: actions/checkout@v2
+    - name: Set up Python ${{ matrix.python-version }}
+      uses: actions/setup-python@v2
+      with:
+        python-version: ${{ matrix.python-version }}
+    - name: Install dependencies
+      run: |
+        python -m pip install --upgrade pip
+        pip install pylint==2.8.2
+    - name: Analysing the code with pylint
+      run: |
+        pylint vllm tests

.github/workflows/scripts/build.sh

@@ -0,0 +1,15 @@
+#!/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
+$python_executable -m pip install -r requirements.txt
+
+# Build
+$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 +1,18 @@
+#!/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

.github/workflows/scripts/env.sh

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

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

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

.github/workflows/yapf.yml

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

.gitignore

@@ -170,3 +170,10 @@ cython_debug/
 
 # Python pickle files
 *.pkl
+
+# Sphinx documentation
+_build/
+
+# vim swap files
+*.swo
+*.swp

.pylintrc

@@ -0,0 +1,434 @@
+# 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
+
+# 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

CONTRIBUTING.md

@@ -49,12 +49,15 @@ If not, please file a new issue, providing as much relevant information as possi
 
 In general, we adhere to [Google Python style guide](https://google.github.io/styleguide/pyguide.html) and [Google C++ style guide](https://google.github.io/styleguide/cppguide.html).
 
+We include a formatting script [`format.sh`](./format.sh) to format the code.
+
 ### Pull Requests
 
 When submitting a pull request:
 
 1. Make sure your code has been rebased on top of the latest commit on the main branch.
-2. Include a detailed description of the changes in the pull request.
+2. Ensure code is properly formatted by running [`format.sh`](./format.sh).
+3. Include a detailed description of the changes in the pull request.
 Explain why you made the changes you did.
 If your pull request fixes an open issue, please include a reference to it in the description.
 

README.md

@@ -10,15 +10,20 @@ 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://vllm.readthedocs.io/en/latest/"><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* 🔥
-
-- [2023/06] We officially released vLLM! vLLM has powered [LMSYS Vicuna and Chatbot Arena](https://chat.lmsys.org) since mid April. Check out our [blog post](https://vllm.ai).
+- [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).
 
 ---
 
@@ -28,23 +33,33 @@ vLLM is fast with:
 
 - State-of-the-art serving throughput
 - Efficient management of attention key and value memory with **PagedAttention**
-- Dynamic batching of incoming requests
+- Continuous batching of incoming requests
 - Optimized CUDA kernels
 
 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 (`baichuan-inc/Baichuan-7B`, `baichuan-inc/Baichuan-13B-Chat`, etc.)
+- BLOOM (`bigscience/bloom`, `bigscience/bloomz`, etc.)
+- Falcon (`tiiuae/falcon-7b`, `tiiuae/falcon-40b`, `tiiuae/falcon-rw-7b`, etc.)
 - GPT-2 (`gpt2`, `gpt2-xl`, etc.)
-- GPTNeoX (`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.)
+- GPT BigCode (`bigcode/starcoder`, `bigcode/gpt_bigcode-santacoder`, etc.)
+- GPT-J (`EleutherAI/gpt-j-6b`, `nomic-ai/gpt4all-j`, etc.)
+- GPT-NeoX (`EleutherAI/gpt-neox-20b`, `databricks/dolly-v2-12b`, `stabilityai/stablelm-tuned-alpha-7b`, etc.)
+- InternLM (`internlm/internlm-7b`, `internlm/internlm-chat-7b`, etc.)
+- 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.)
+- Qwen (`Qwen/Qwen-7B`, `Qwen/Qwen-7B-Chat`, etc.)
 
 Install vLLM with pip or [from source](https://vllm.readthedocs.io/en/latest/getting_started/installation.html#build-from-source):
 
@@ -59,37 +74,19 @@ 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

@@ -17,9 +17,13 @@ def main(args: argparse.Namespace):
     # 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(
@@ -36,13 +40,13 @@ def main(args: argparse.Namespace):
     def run_to_completion(profile: bool = False):
         if profile:
             torch.cuda.cudart().cudaProfilerStart()
-        start_time = time.time()
+        start_time = time.perf_counter()
 
         llm.generate(prompt_token_ids=dummy_prompt_token_ids,
                      sampling_params=sampling_params,
                      use_tqdm=False)
 
-        end_time = time.time()
+        end_time = time.perf_counter()
         latency = end_time - start_time
         if profile:
             torch.cuda.cudart().cudaProfilerStop()
@@ -61,16 +65,37 @@ 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', 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.')
     args = parser.parse_args()
     main(args)

benchmarks/benchmark_serving.py

@@ -24,20 +24,13 @@ from typing import AsyncGenerator, List, Tuple
 
 import aiohttp
 import numpy as np
-from transformers import AutoConfig, AutoTokenizer, PreTrainedTokenizerBase
+from transformers import PreTrainedTokenizerBase
+from vllm.transformers_utils.tokenizer import get_tokenizer
 
 # (prompt len, output len, latency)
 REQUEST_LATENCY: List[Tuple[int, int, float]] = []
 
 
-def get_tokenizer(model_name: str) -> PreTrainedTokenizerBase:
-    config = AutoConfig.from_pretrained(model_name)
-    if config.model_type == "llama":
-        # A workaround for potential protobuf errors.
-        model_name = "hf-internal-testing/llama-tokenizer"
-    return AutoTokenizer.from_pretrained(model_name)
-
-
 def sample_requests(
     dataset_path: str,
     num_requests: int,
@@ -112,7 +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":
@@ -155,7 +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))
 
@@ -184,13 +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")
@@ -217,7 +210,7 @@ if __name__ == "__main__":
     parser.add_argument("--backend", type=str, default="vllm",
                         choices=["vllm", "tgi"])
     parser.add_argument("--host", type=str, default="localhost")
-    parser.add_argument("--port", type=int, default=8001)
+    parser.add_argument("--port", type=int, default=8000)
     parser.add_argument("--dataset", type=str, required=True,
                         help="Path to the dataset.")
     parser.add_argument("--tokenizer", type=str, required=True,
@@ -234,5 +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

@@ -3,26 +3,14 @@ import argparse
 import json
 import random
 import time
-from typing import List, Tuple
+from typing import List, Optional, Tuple
 
 import torch
-from transformers import (AutoConfig, AutoTokenizer, AutoModelForCausalLM,
-                          PreTrainedTokenizerBase)
+from transformers import AutoModelForCausalLM, PreTrainedTokenizerBase
 from tqdm import tqdm
 
 from vllm import LLM, SamplingParams
-
-
-def get_tokenizer(model_name: str) -> PreTrainedTokenizerBase:
-    config = AutoConfig.from_pretrained(model_name)
-    if config.model_type == "llama":
-        # A workaround for potential protobuf errors.
-        model_name = "hf-internal-testing/llama-tokenizer"
-        tokenizer = AutoTokenizer.from_pretrained(model_name)
-        # To enable padding in the HF backend.
-        tokenizer.pad_token = tokenizer.eos_token
-        return tokenizer
-    return AutoTokenizer.from_pretrained(model_name)
+from vllm.transformers_utils.tokenizer import get_tokenizer
 
 
 def sample_requests(
@@ -34,15 +22,10 @@ def sample_requests(
     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]
@@ -74,15 +57,23 @@ def sample_requests(
 def run_vllm(
     requests: List[Tuple[str, int, int]],
     model: str,
+    tokenizer: str,
+    quantization: Optional[str],
     tensor_parallel_size: int,
     seed: int,
     n: int,
     use_beam_search: bool,
+    trust_remote_code: bool,
+    dtype: str,
 ) -> float:
     llm = LLM(
         model=model,
+        tokenizer=tokenizer,
+        quantization=quantization,
         tensor_parallel_size=tensor_parallel_size,
         seed=seed,
+        trust_remote_code=trust_remote_code,
+        dtype=dtype,
     )
 
     # Add the requests to the engine.
@@ -102,10 +93,10 @@ def run_vllm(
             sampling_params=sampling_params,
         )
 
-    start = time.time()
+    start = time.perf_counter()
     # FIXME(woosuk): Do use internal method.
     llm._run_engine(use_tqdm=True)
-    end = time.time()
+    end = time.perf_counter()
     return end - start
 
 
@@ -116,15 +107,18 @@ def run_hf(
     n: int,
     use_beam_search: bool,
     max_batch_size: int,
+    trust_remote_code: bool,
 ) -> float:
     assert not use_beam_search
-    tokenizer = get_tokenizer(model)
     llm = AutoModelForCausalLM.from_pretrained(
-        model, torch_dtype=torch.float16)
+        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
@@ -137,13 +131,14 @@ 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,
@@ -161,7 +156,7 @@ def run_hf(
         batch = []
         max_prompt_len = 0
         max_output_len = 0
-    end = time.time()
+    end = time.perf_counter()
     return end - start
 
 
@@ -170,49 +165,82 @@ def main(args: argparse.Namespace):
     random.seed(args.seed)
 
     # Sample the requests.
-    tokenizer = get_tokenizer(args.model)
+    tokenizer = get_tokenizer(args.tokenizer,
+                              trust_remote_code=args.trust_remote_code)
     requests = sample_requests(args.dataset, args.num_prompts, tokenizer)
 
     if args.backend == "vllm":
-        elapsed_time = run_vllm(
-            requests, args.model, 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)
     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)
     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"],
                         default="vllm")
-    parser.add_argument("--dataset", type=str, required=True,
+    parser.add_argument("--dataset",
+                        type=str,
+                        required=True,
                         help="Path to 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', 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(
+        '--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.backend == "vllm":
         if args.hf_max_batch_size is not None:
             raise ValueError("HF max batch size is only for HF backend.")
     elif args.backend == "hf":
         if args.hf_max_batch_size is None:
             raise ValueError("HF max batch size is required for HF backend.")
+        if args.quantization is not None:
+            raise ValueError("Quantization is only for vLLM backend.")
+    if args.tokenizer is None:
+        args.tokenizer = args.model
 
     main(args)

benchmarks/kernels/benchmark_paged_attention.py

@@ -0,0 +1,197 @@
+import argparse
+import random
+import time
+
+import torch
+
+from vllm import attention_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":
+                attention_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":
+                attention_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,
+    )

benchmarks/launch_tgi_server.sh

@@ -1,6 +1,6 @@
 #!/bin/bash
 
-PORT=8001
+PORT=8000
 MODEL=$1
 TOKENS=$2
 

csrc/activation.cpp

@@ -4,9 +4,25 @@ 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);
+
 PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
   m.def(
     "silu_and_mul",
     &silu_and_mul,
     "Activation function used in SwiGLU.");
+  m.def(
+    "gelu_new",
+    &gelu_new,
+    "GELU implementation used in GPT-2.");
+  m.def(
+    "gelu_fast",
+    &gelu_fast,
+    "Approximate GELU implementation.");
 }

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>
@@ -34,9 +36,7 @@ void silu_and_mul(
   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,               // [num_tokens, d]
+  const scalar_t* __restrict__ input,       // [num_tokens, d]
+  const int d) {
+  const int token_idx = blockIdx.x;
+  for (int idx = threadIdx.x; idx < d; idx += blockDim.x) {
+    const scalar_t x = __ldg(&input[token_idx * d + idx]);
+    out[token_idx * d + idx] = ACT_FN(x);
+  }
+}
+
+} // namespace vllm
+
+// Launch element-wise activation kernel.
+#define LAUNCH_ACTIVATION_KERNEL(KERNEL)                                                  \
+  int num_tokens = input.size(0);                                                         \
+  int d = input.size(1);                                                                  \
+  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,     // [num_tokens, d]
+  torch::Tensor& input)   // [num_tokens, d]
+{
+  LAUNCH_ACTIVATION_KERNEL(vllm::gelu_new_kernel);
+}
+
+void gelu_fast(
+  torch::Tensor& out,     // [num_tokens, d]
+  torch::Tensor& input)   // [num_tokens, d]
+{
+  LAUNCH_ACTIVATION_KERNEL(vllm::gelu_fast_kernel);
+}

csrc/attention.cpp

@@ -1,19 +1,42 @@
 #include <torch/extension.h>
+#include <c10/util/Optional.h>
 
-void single_query_cached_kv_attention(
+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);
+  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);
 
 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");
+    "paged_attention_v1",
+    &paged_attention_v1,
+    "Compute the attention between an input query and the cached keys/values using PagedAttention.");
+  m.def(
+    "paged_attention_v2",
+    &paged_attention_v2,
+    "PagedAttention V2.");
 }

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 {
 
@@ -65,24 +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 int q_stride) {
+  const float* __restrict__ alibi_slopes, // [num_heads]
+  const int q_stride,
+  const int kv_block_stride,
+  const int kv_head_stride) {
+  const int seq_idx = blockIdx.y;
+  const int partition_idx = blockIdx.z;
+  const int max_num_partitions = gridDim.z;
+  constexpr bool USE_PARTITIONING = PARTITION_SIZE > 0;
+  const int context_len = context_lens[seq_idx];
+  if (USE_PARTITIONING && partition_idx * PARTITION_SIZE >= context_len) {
+    // No work to do. Terminate the thread block.
+    return;
+  }
+
+  const int num_context_blocks = DIVIDE_ROUND_UP(context_len, BLOCK_SIZE);
+  const int num_blocks_per_partition = USE_PARTITIONING ? PARTITION_SIZE / BLOCK_SIZE : num_context_blocks;
+
+  // [start_block_idx, end_block_idx) is the range of blocks to process.
+  const int start_block_idx = USE_PARTITIONING ? partition_idx * num_blocks_per_partition : 0;
+  const int end_block_idx = MIN(start_block_idx + num_blocks_per_partition, num_context_blocks);
+  const int num_blocks = end_block_idx - start_block_idx;
+
+  // [start_token_idx, end_token_idx) is the range of tokens to process.
+  const int start_token_idx = start_block_idx * BLOCK_SIZE;
+  const int end_token_idx = MIN(start_token_idx + num_blocks * BLOCK_SIZE, context_len);
+  const int num_tokens = end_token_idx - start_token_idx;
+
   constexpr int THREAD_GROUP_SIZE = MAX(WARP_SIZE / BLOCK_SIZE, 1);
-  constexpr int NUM_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;
@@ -90,7 +124,8 @@ __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.
   // The vector size is configured in such a way that the threads in a thread group
@@ -114,12 +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[];
@@ -133,15 +169,12 @@ __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* 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) {
     const int physical_block_number = block_table[block_idx];
 
     // Load a key to registers.
@@ -156,8 +189,8 @@ __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;
@@ -167,13 +200,15 @@ __global__ void single_query_cached_kv_attention_kernel(
 
       // Compute dot product.
       // This includes a reduction across the threads in the same thread group.
-      const float qk = scale * Qk_dot<scalar_t, THREAD_GROUP_SIZE>::dot(q_vecs, k_vecs);
-      const bool mask = token_idx >= context_len;
-    
+      float qk = scale * Qk_dot<scalar_t, THREAD_GROUP_SIZE>::dot(q_vecs[thread_group_offset], k_vecs);
+      // Add the ALiBi bias if slopes are given.
+      qk += (alibi_slope != 0) ? alibi_slope * (token_idx - context_len + 1) : 0;
+
       if (thread_group_offset == 0) {
         // Store the partial reductions to shared memory.
         // NOTE(woosuk): It is required to zero out the masked logits.
-        logits[token_idx] = mask ? 0.f : qk;
+        const bool mask = token_idx >= context_len;
+        logits[token_idx - start_token_idx] = mask ? 0.f : qk;
         // Update the max value.
         qk_max = mask ? qk_max : fmaxf(qk_max, qk);
       }
@@ -204,7 +239,7 @@ __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;
@@ -213,11 +248,23 @@ __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;
@@ -226,7 +273,7 @@ __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];
@@ -235,21 +282,33 @@ __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) {
+  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) {
     const int physical_block_number = block_table[block_idx];
     const int physical_block_offset = (lane % NUM_V_VECS_PER_ROW) * V_VEC_SIZE;
     const int token_idx = block_idx * BLOCK_SIZE + physical_block_offset;
     L_vec logits_vec;
-    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);
       }
     }
@@ -304,7 +363,9 @@ __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;
@@ -315,156 +376,485 @@ __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,                                                                   \
-    query_stride);
+    alibi_slopes_ptr,                                                                         \
+    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,
-  int max_context_len) {
+  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 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);
 
+  // NOTE: alibi_slopes is optional.
+  const float* alibi_slopes_ptr = alibi_slopes ?
+    reinterpret_cast<const float*>(alibi_slopes.value().data_ptr())
+    : nullptr;
+
   T* out_ptr = reinterpret_cast<T*>(out.data_ptr());
   T* query_ptr = reinterpret_cast<T*>(query.data_ptr());
   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_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,                                                   \
-    max_context_len);
+    max_context_len,                                                \
+    alibi_slopes);
 
 // 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]
   int block_size,
-  int max_context_len) {
+  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());
   }
@@ -473,3 +863,4 @@ 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_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>
@@ -125,9 +127,7 @@ void copy_blocks(
   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>(),
@@ -202,9 +202,7 @@ 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",
     [&] {
@@ -364,9 +362,7 @@ 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.cpp

@@ -0,0 +1,13 @@
+#include <torch/extension.h>
+
+int get_device_attribute(
+    int attribute,
+    int device_id);
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+  m.def(
+    "get_device_attribute",
+    &get_device_attribute,
+    "Gets the specified device attribute.");
+}
+

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_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 {
@@ -46,9 +47,7 @@ void rms_norm(
   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",
     [&] {

csrc/pos_encoding.cpp

@@ -1,15 +1,16 @@
 #include <torch/extension.h>
 
-void rotary_embedding_neox(
+void rotary_embedding(
   torch::Tensor& positions,
   torch::Tensor& query,
   torch::Tensor& key,
   int head_size,
-  torch::Tensor& cos_sin_cache);
+  torch::Tensor& cos_sin_cache,
+  bool is_neox);
 
 PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
   m.def(
-    "rotary_embedding_neox",
-    &rotary_embedding_neox,
-    "Apply GPT-NeoX style rotary embedding to query and key");
+    "rotary_embedding",
+    &rotary_embedding,
+    "Apply GPT-NeoX or GPT-J 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(
+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,        // [num_tokens]
   scalar_t* __restrict__ query,                 // [num_tokens, num_heads, head_size]
-  scalar_t* __restrict__ key,                   // [num_tokens, num_heads, head_size]
+  scalar_t* __restrict__ key,                   // [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(
+void rotary_embedding(
   torch::Tensor& positions,         // [num_tokens]
   torch::Tensor& query,             // [num_tokens, num_heads * head_size]
-  torch::Tensor& key,               // [num_tokens, num_heads * head_size]
+  torch::Tensor& key,               // [num_tokens, num_kv_heads * head_size]
   int head_size,
-  torch::Tensor& cos_sin_cache)     // [max_position, rot_dim]
-{
+  torch::Tensor& cos_sin_cache,     // [max_position, rot_dim]
+  bool is_neox) {
   int num_tokens = query.size(0);
   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_kv_heads = key.size(1) / head_size;
+  int query_stride = query.stride(0);
+  int key_stride = key.stride(0);
 
   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/quantization.cpp

@@ -0,0 +1,15 @@
+#include <torch/extension.h>
+
+torch::Tensor awq_gemm(
+  torch::Tensor _in_feats,
+  torch::Tensor _kernel,
+  torch::Tensor _scaling_factors,
+  torch::Tensor _zeros,
+  int split_k_iters);
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+  m.def(
+    "awq_gemm",
+    &awq_gemm,
+    "Quantized GEMM for AWQ");
+}

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);
+}

docs/README.md

@@ -4,14 +4,14 @@
 
 ```bash
 # Install dependencies.
-pip -r requirements-docs.txt
+pip install -r requirements-docs.txt
 
 # Build the docs.
 make clean
 make html
 ```
 
-## Open the docs with your brower
+## Open the docs with your browser
 
 ```bash
 python -m http.server -d build/html/

docs/source/getting_started/installation.rst

@@ -3,31 +3,15 @@
 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 (11.8) binaries.
 
 Requirements
 ------------
 
 * OS: Linux
-* Python: 3.8 or higher
-* CUDA: 11.0 -- 11.8
+* Python: 3.8 -- 3.11
 * 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.
-
 Install with pip
 ----------------
 
@@ -40,7 +24,7 @@ You can install vLLM using pip:
     $ conda activate myenv
 
     $ # Install vLLM.
-    $ pip install vllm  # This may take 5-10 minutes.
+    $ pip install vllm
 
 
 .. _build_from_source:
@@ -55,3 +39,12 @@ 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
+
+        $ # Pull the Docker image with CUDA 11.8.
+        $ # Use `--ipc=host` to make sure the shared memory is large enough.
+        $ docker run --gpus all -it --rm --ipc=host nvcr.io/nvidia/pytorch:22.12-py3

docs/source/getting_started/quickstart.rst

@@ -128,4 +128,4 @@ Since this server is compatible with OpenAI API, you can use it as a drop-in rep
                                           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>`_.

docs/source/index.rst

@@ -29,7 +29,7 @@ vLLM is fast with:
 
 * State-of-the-art serving throughput
 * Efficient management of attention key and value memory with **PagedAttention**
-* Dynamic batching of incoming requests
+* Continuous batching of incoming requests
 * Optimized CUDA kernels
 
 vLLM is flexible and easy to use with:
@@ -40,7 +40,12 @@ vLLM is flexible and easy to use with:
 * Streaming outputs
 * OpenAI-compatible API server
 
-For more information, please refer to our `blog post <https://vllm.ai>`_.
+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.
+
 
 
 Documentation
@@ -53,6 +58,14 @@ Documentation
    getting_started/installation
    getting_started/quickstart
 
+.. toctree::
+   :maxdepth: 1
+   :caption: Serving
+
+   serving/distributed_serving
+   serving/run_on_sky
+   serving/deploying_with_triton
+
 .. toctree::
    :maxdepth: 1
    :caption: Models

docs/source/models/adding_model.rst

@@ -59,7 +59,7 @@ 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.

docs/source/models/supported_models.rst

@@ -14,18 +14,48 @@ 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/Baichuan-7B`, :code:`baichuan-inc/Baichuan-13B-Chat`, 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:`QWenLMHeadModel`
+    - Qwen
+    - :code:`Qwen/Qwen-7B`, :code:`Qwen/Qwen-7B-Chat`, 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.

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/distributed_serving.rst

@@ -0,0 +1,38 @@
+.. _distributed_serving:
+
+Distributed Inference and Serving
+=================================
+
+vLLM supports distributed tensor-parallel inference and serving. Currently, we support `Megatron-LM's tensor parallel algorithm <https://arxiv.org/pdf/1909.08053.pdf>`_. We manage the distributed runtime with `Ray <https://github.com/ray-project/ray>`_. To run distributed inference, install Ray with:
+
+.. code-block:: console
+
+    $ pip install ray
+
+To run multi-GPU inference with the :code:`LLM` class, set the :code:`tensor_parallel_size` argument to the number of GPUs you want to use. For example, to run inference on 4 GPUs:
+
+.. code-block:: python
+
+    from vllm import LLM
+    llm = LLM("facebook/opt-13b", tensor_parallel_size=4)
+    output = llm.generate("San Franciso is a")
+
+To run multi-GPU serving, pass in the :code:`--tensor-parallel-size` argument when starting the server. For example, to run API server on 4 GPUs:
+
+.. code-block:: console
+
+    $ python -m vllm.entrypoints.api_server \
+    $     --model facebook/opt-13b \
+    $     --tensor-parallel-size 4
+
+To scale vLLM beyond a single machine, start a `Ray runtime <https://docs.ray.io/en/latest/ray-core/starting-ray.html>`_ via CLI before running vLLM:
+
+.. code-block:: console
+
+    $ # On head node
+    $ ray start --head
+
+    $ # On worker nodes
+    $ ray start --address=<ray-head-address>
+
+After that, you can run inference and serving on multiple machines by launching the vLLM process on the head node by setting :code:`tensor_parallel_size` to the number of GPUs to be the total number of GPUs across all machines.

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
+

examples/api_client.py

@@ -14,7 +14,9 @@ def clear_line(n: int = 1) -> None:
         print(LINE_UP, end=LINE_CLEAR, flush=True)
 
 
-def post_http_request(prompt: str, api_url: str, n: int = 1,
+def post_http_request(prompt: str,
+                      api_url: str,
+                      n: int = 1,
                       stream: bool = False) -> requests.Response:
     headers = {"User-Agent": "Test Client"}
     pload = {
@@ -30,7 +32,8 @@ def post_http_request(prompt: str, api_url: str, n: int = 1,
 
 
 def get_streaming_response(response: requests.Response) -> Iterable[List[str]]:
-    for chunk in response.iter_lines(chunk_size=8192, decode_unicode=False,
+    for chunk in response.iter_lines(chunk_size=8192,
+                                     decode_unicode=False,
                                      delimiter=b"\0"):
         if chunk:
             data = json.loads(chunk.decode("utf-8"))

examples/gradio_webserver.py

@@ -12,9 +12,14 @@ def http_bot(prompt):
         "stream": True,
         "max_tokens": 128,
     }
-    response = requests.post(args.model_url, headers=headers, json=pload, stream=True)
+    response = requests.post(args.model_url,
+                             headers=headers,
+                             json=pload,
+                             stream=True)
 
-    for chunk in response.iter_lines(chunk_size=8192, decode_unicode=False, delimiter=b"\0"):
+    for chunk in response.iter_lines(chunk_size=8192,
+                                     decode_unicode=False,
+                                     delimiter=b"\0"):
         if chunk:
             data = json.loads(chunk.decode("utf-8"))
             output = data["text"][0]
@@ -23,20 +28,22 @@ def http_bot(prompt):
 
 def build_demo():
     with gr.Blocks() as demo:
-        gr.Markdown(
-            "# vLLM text completion demo\n"
-        )
-        inputbox = gr.Textbox(label="Input", placeholder="Enter text and press ENTER")
-        outputbox = gr.Textbox(label="Output", placeholder="Generated result from the model")
+        gr.Markdown("# vLLM text completion demo\n")
+        inputbox = gr.Textbox(label="Input",
+                              placeholder="Enter text and press ENTER")
+        outputbox = gr.Textbox(label="Output",
+                               placeholder="Generated result from the model")
         inputbox.submit(http_bot, [inputbox], [outputbox])
     return demo
 
 
 if __name__ == "__main__":
     parser = argparse.ArgumentParser()
-    parser.add_argument("--host", type=str, default="localhost")
+    parser.add_argument("--host", type=str, default=None)
     parser.add_argument("--port", type=int, default=8001)
-    parser.add_argument("--model-url", type=str, default="http://localhost:8000/generate")
+    parser.add_argument("--model-url",
+                        type=str,
+                        default="http://localhost:8000/generate")
     args = parser.parse_args()
 
     demo = build_demo()

examples/llm_engine_example.py

@@ -10,19 +10,25 @@ def main(args: argparse.Namespace):
 
     # Test the following prompts.
     test_prompts = [
-        ("A robot may not injure a human being", SamplingParams()),
+        ("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?",
-         SamplingParams(n=2, best_of=5, temperature=0.8, top_p=0.95, frequency_penalty=0.1)),
+         SamplingParams(n=2,
+                        best_of=5,
+                        temperature=0.8,
+                        top_p=0.95,
+                        frequency_penalty=0.1)),
         ("It is only with the heart that one can see rightly",
-         SamplingParams(n=3, best_of=3, use_beam_search=True, temperature=0.0)),
+         SamplingParams(n=3, best_of=3, use_beam_search=True,
+                        temperature=0.0)),
     ]
 
     # Run the engine by calling `engine.step()` manually.
     request_id = 0
     while True:
-        # To test iteration-level scheduling, we add one request at each step.
+        # To test continuous batching, we add one request at each step.
         if test_prompts:
             prompt, sampling_params = test_prompts.pop(0)
             engine.add_request(str(request_id), prompt, sampling_params)
@@ -30,7 +36,7 @@ def main(args: argparse.Namespace):
 
         request_outputs = engine.step()
         for request_output in request_outputs:
-            if request_output.finished():
+            if request_output.finished:
                 print(request_output)
 
         if not (engine.has_unfinished_requests() or test_prompts):

examples/offline_inference.py

@@ -1,6 +1,5 @@
 from vllm import LLM, SamplingParams
 
-
 # Sample prompts.
 prompts = [
     "Hello, my name is",

examples/openai_chatcompletion_client.py

@@ -0,0 +1,33 @@
+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"
+
+# List models API
+models = openai.Model.list()
+print("Models:", models)
+
+model = models["data"][0]["id"]
+
+# Chat completion API
+chat_completion = openai.ChatCompletion.create(
+    model=model,
+    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?"
+    }])
+
+print("Chat completion results:")
+print(chat_completion)

examples/openai_completion_client.py

@@ -3,21 +3,26 @@ 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
+# 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)
+model = models["data"][0]["id"]
 
-# print the completion
+# Completion API
+stream = False
+completion = openai.Completion.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 result:", completion)
+    print(completion)

format.sh

@@ -0,0 +1,107 @@
+#!/usr/bin/env bash
+# YAPF formatter, adapted from ray and skypilot.
+#
+# Usage:
+#    # Do work and commit your work.
+
+#    # Format files that differ from origin/main.
+#    bash format.sh
+
+#    # Commit changed files with message 'Run yapf and pylint'
+#
+#
+# YAPF + Clang formatter (if installed). This script formats all changed files from the last mergebase.
+# You are encouraged to run this locally before pushing changes for review.
+
+# Cause the script to exit if a single command fails
+set -eo pipefail
+
+# this stops git rev-parse from failing if we run this from the .git directory
+builtin cd "$(dirname "${BASH_SOURCE:-$0}")"
+ROOT="$(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 $2}')
+MYPY_VERSION=$(mypy --version | awk '{print $2}')
+
+# # params: tool name, tool version, required version
+tool_version_check() {
+    if [[ $2 != $3 ]]; then
+        echo "Wrong $1 version installed: $3 is required, not $2."
+        exit 1
+    fi
+}
+
+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 "mypy" "$MYPY_VERSION" "$(grep mypy requirements-dev.txt | cut -d'=' -f3)"
+
+YAPF_FLAGS=(
+    '--recursive'
+    '--parallel'
+)
+
+YAPF_EXCLUDES=(
+    '--exclude' 'build/**'
+)
+
+# Format specified files
+format() {
+    yapf --in-place "${YAPF_FLAGS[@]}" "$@"
+}
+
+# Format files that differ from main branch. Ignores dirs that are not slated
+# for autoformat yet.
+format_changed() {
+    # The `if` guard ensures that the list of filenames is not empty, which
+    # could cause yapf to receive 0 positional arguments, making it hang
+    # waiting for STDIN.
+    #
+    # `diff-filter=ACM` and $MERGEBASE is to ensure we only format files that
+    # exist on both branches.
+    MERGEBASE="$(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 -P 5 \
+             yapf --in-place "${YAPF_EXCLUDES[@]}" "${YAPF_FLAGS[@]}"
+    fi
+
+}
+
+# Format all files
+format_all() {
+    yapf --in-place "${YAPF_FLAGS[@]}" "${YAPF_EXCLUDES[@]}" vllm tests
+}
+
+## This flag formats individual files. --files *must* be the first command line
+## arg to use this option.
+if [[ "$1" == '--files' ]]; then
+   format "${@:2}"
+   # If `--all` is passed, then any further arguments are ignored and the
+   # entire python directory is formatted.
+elif [[ "$1" == '--all' ]]; then
+   format_all
+else
+   # Format only the files that changed in last commit.
+   format_changed
+fi
+echo 'vLLM yapf: Done'
+
+# Run mypy
+# TODO(zhuohan): Enable mypy
+# echo 'vLLM mypy:'
+# mypy
+
+# Run Pylint
+echo 'vLLM Pylint:'
+pylint vllm tests
+
+if ! git diff --quiet &>/dev/null; then
+    echo 'Reformatted files. Please review and stage the changes.'
+    echo 'Changes not staged for commit:'
+    echo
+    git --no-pager diff --name-only
+
+    exit 1
+fi

pyproject.toml

@@ -3,7 +3,7 @@ requires = [
     "ninja",
     "packaging",
     "setuptools",
-    "torch >= 2.0.0",
+    "torch == 2.0.1",
     "wheel",
 ]
 build-backend = "setuptools.build_meta"

requirements-dev.txt

@@ -1,2 +1,14 @@
-mypy
+# formatting
+yapf==0.32.0
+pylint==2.8.2
+
+# type checking
+mypy==0.991
+types-PyYAML
+types-requests
+types-setuptools
+
+# testing
 pytest
+pytest-forked
+pytest-asyncio

requirements.txt

@@ -1,11 +1,13 @@
 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
+torch == 2.0.1
+transformers >= 4.34.0  # Required for Mistral.
+xformers == 0.0.22  # Required for Mistral.
 fastapi
-uvicorn
-pydantic  # Required for OpenAI server.
+uvicorn[standard]
+pydantic < 2  # Required for OpenAI server.

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 +12,9 @@ from torch.utils.cpp_extension import BuildExtension, CUDAExtension, CUDA_HOME
 
 ROOT_DIR = os.path.dirname(__file__)
 
+# 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?
@@ -20,10 +24,9 @@ ABI = 1 if torch._C._GLIBCXX_USE_CXX11_ABI else 0
 CXX_FLAGS += [f"-D_GLIBCXX_USE_CXX11_ABI={ABI}"]
 NVCC_FLAGS += [f"-D_GLIBCXX_USE_CXX11_ABI={ABI}"]
 
-if not torch.cuda.is_available():
+if CUDA_HOME is None:
     raise RuntimeError(
-        f"Cannot find CUDA at CUDA_HOME: {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:
@@ -39,32 +42,95 @@ 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}.")
+    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"):
-    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.")
+if nvcc_cuda_version < Version("11.1"):
+    if any(cc.startswith("8.6") for cc in compute_capabilities):
+        raise RuntimeError(
+            "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.")
+        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"):
@@ -77,7 +143,10 @@ ext_modules = []
 cache_extension = CUDAExtension(
     name="vllm.cache_ops",
     sources=["csrc/cache.cpp", "csrc/cache_kernels.cu"],
-    extra_compile_args={"cxx": CXX_FLAGS, "nvcc": NVCC_FLAGS},
+    extra_compile_args={
+        "cxx": CXX_FLAGS,
+        "nvcc": NVCC_FLAGS,
+    },
 )
 ext_modules.append(cache_extension)
 
@@ -85,7 +154,10 @@ ext_modules.append(cache_extension)
 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},
+    extra_compile_args={
+        "cxx": CXX_FLAGS,
+        "nvcc": NVCC_FLAGS,
+    },
 )
 ext_modules.append(attention_extension)
 
@@ -93,7 +165,10 @@ ext_modules.append(attention_extension)
 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},
+    extra_compile_args={
+        "cxx": CXX_FLAGS,
+        "nvcc": NVCC_FLAGS,
+    },
 )
 ext_modules.append(positional_encoding_extension)
 
@@ -101,7 +176,10 @@ ext_modules.append(positional_encoding_extension)
 layernorm_extension = CUDAExtension(
     name="vllm.layernorm_ops",
     sources=["csrc/layernorm.cpp", "csrc/layernorm_kernels.cu"],
-    extra_compile_args={"cxx": CXX_FLAGS, "nvcc": NVCC_FLAGS},
+    extra_compile_args={
+        "cxx": CXX_FLAGS,
+        "nvcc": NVCC_FLAGS,
+    },
 )
 ext_modules.append(layernorm_extension)
 
@@ -109,10 +187,38 @@ ext_modules.append(layernorm_extension)
 activation_extension = CUDAExtension(
     name="vllm.activation_ops",
     sources=["csrc/activation.cpp", "csrc/activation_kernels.cu"],
-    extra_compile_args={"cxx": CXX_FLAGS, "nvcc": NVCC_FLAGS},
+    extra_compile_args={
+        "cxx": CXX_FLAGS,
+        "nvcc": NVCC_FLAGS,
+    },
 )
 ext_modules.append(activation_extension)
 
+# Quantization kernels.
+quantization_extension = CUDAExtension(
+    name="vllm.quantization_ops",
+    sources=[
+        "csrc/quantization.cpp",
+        "csrc/quantization/awq/gemm_kernels.cu",
+    ],
+    extra_compile_args={
+        "cxx": CXX_FLAGS,
+        "nvcc": NVCC_FLAGS,
+    },
+)
+ext_modules.append(quantization_extension)
+
+# Misc. CUDA utils.
+cuda_utils_extension = CUDAExtension(
+    name="vllm.cuda_utils",
+    sources=["csrc/cuda_utils.cpp", "csrc/cuda_utils_kernels.cu"],
+    extra_compile_args={
+        "cxx": CXX_FLAGS,
+        "nvcc": NVCC_FLAGS,
+    },
+)
+ext_modules.append(cuda_utils_extension)
+
 
 def get_path(*filepath) -> str:
     return os.path.join(ROOT_DIR, *filepath)
@@ -124,8 +230,8 @@ def find_version(filepath: str):
     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.")
@@ -148,7 +254,8 @@ setuptools.setup(
     version=find_version(get_path("vllm", "__init__.py")),
     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",
@@ -160,11 +267,12 @@ 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,

tests/async_engine/api_server_async_engine.py

@@ -0,0 +1,51 @@
+"""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):
+
+    # pylint: disable=redefined-outer-name
+    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 +1,89 @@
+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()
+    # pylint: disable=consider-using-with
+    uvicorn_process = subprocess.Popen([
+        sys.executable, "-u",
+        str(script_path), "--model", "facebook/opt-125m"
+    ])
+    yield
+    uvicorn_process.terminate()
+
+
+# pylint: disable=redefined-outer-name, unused-argument
+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:
+            # pylint: disable=bare-except
+            try:
+                for result in pool.map(_query_server, prompts):
+                    break
+            except:
+                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_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 +1,212 @@
+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 = [
+    # pylint: disable=line-too-long
+    "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 +1,82 @@
+"""Test the communication operators.
+
+Run `pytest tests/distributed/test_comm_ops.py --forked`.
+"""
+from multiprocessing import Process
+
+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):
+    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 +1,63 @@
+import pytest
+
+from transformers import AutoTokenizer
+
+from vllm.transformers_utils.tokenizer import detokenize_incrementally
+
+TRUTH = [
+    # pylint: disable=line-too-long
+    "Hello here, this is a simple test",
+    "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",
+    "我很感谢你的热情"
+]
+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 +1,75 @@
+import pytest
 import torch
 import torch.nn.functional as F
+from transformers.activations import get_activation
 
 from vllm import activation_ops
 
+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]
+
 
 def ref_silu_and_mul(x: torch.Tensor) -> torch.Tensor:
     x1, x2 = x.chunk(chunks=2, dim=1)
     return F.silu(x1) * x2
 
 
+@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')
+    torch.random.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
+    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)
     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")
+    out = torch.empty(num_tokens, d, dtype=dtype, device="cuda")
+    activation_ops.gelu_new(out, x)
+    ref_out = get_activation("gelu_new")(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")
+    out = torch.empty(num_tokens, d, dtype=dtype, device="cuda")
+    activation_ops.gelu_fast(out, x)
+    ref_out = get_activation("gelu_fast")(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.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 = 128  # 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":
+        attention_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)
+        attention_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:
+        assert False, 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):
@@ -84,10 +249,10 @@ def ref_multi_query_kv_attention(
         seq_len = end_idx - start_idx
 
         # Create attention mask.
-        attn_mask = torch.triu(
-            torch.ones(seq_len, seq_len, dtype=dtype), diagonal=1)
+        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],
@@ -101,147 +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,
-    )
-
-    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))
-    qkv = torch.empty(
-        num_tokens, 3, num_heads, head_size, dtype=dtype, device='cuda')
-    qkv.uniform_(-1e-3, 1e-3)
-    query, key, value = qkv.unbind(dim=1)
+    scale = float(1.0 / (head_size**0.5))
+    num_query_heads, num_kv_heads = num_heads
+    qkv = torch.empty(num_tokens,
+                      num_query_heads + 2 * num_kv_heads,
+                      head_size,
+                      dtype=dtype,
+                      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),
@@ -250,7 +315,6 @@ def run_multi_query_kv_attention(
         attn_bias=attn_bias,
         p=0.0,
         scale=scale,
-        op=attn_op,
     )
     output = output.squeeze(0)
 
@@ -262,40 +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
 
+DTYPES = [torch.half, torch.bfloat16, torch.float]
+NUM_TOKENS = [7, 83, 2048]  # Arbitrary values for testing
+NUM_LAYERS = [5]  # 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]  # Arbitrary values for testing
+NUM_MAPPINGS = [32, 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,
@@ -14,151 +34,113 @@ 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):
+            for cloned_key_cache in cloned_key_caches:
                 cloned_key_cache[dst] = cloned_key_cache[src]
-            for value_cache, cloned_value_cache in zip(value_caches, cloned_value_caches):
+            for cloned_value_cache in cloned_value_caches:
                 cloned_value_cache[dst] = cloned_value_cache[src]
 
     # Compare the results.
     for key_cache, cloned_key_cache in zip(key_caches, cloned_key_caches):
         assert torch.allclose(key_cache, cloned_key_cache)
-    for value_cache, cloned_value_cache in zip(value_caches, cloned_value_caches):
+    for value_cache, cloned_value_cache in zip(value_caches,
+                                               cloned_value_caches):
         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.int, device="cuda")
 
-    qkv = torch.randn(
-        num_tokens, 3, num_heads, head_size, dtype=dtype, device='cuda')
+    qkv = torch.randn(num_tokens,
+                      3,
+                      num_heads,
+                      head_size,
+                      dtype=dtype,
+                      device="cuda")
     _, key, value = qkv.unbind(dim=1)
 
-    x = 16 // torch.tensor([], dtype=dtype).element_size()
-    key_cache_shape = (num_blocks, num_heads, head_size // x, block_size, x)
-    key_cache = torch.randn(size=key_cache_shape, dtype=dtype, device='cuda')
-    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()
 
-    cache_ops.reshape_and_cache(key, value, key_cache, value_cache, slot_mapping)
+    # 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

@@ -1,33 +1,50 @@
+import pytest
 import torch
 import torch.nn as nn
 
 from vllm import layernorm_ops
 
+DTYPES = [torch.half, torch.bfloat16, torch.float]
+HIDDEN_SIZES = [67, 768, 2048, 5120, 8192]  # Arbitrary values for testing
+NUM_TOKENS = [7, 83, 4096]  # Arbitrary values for testing
+SEEDS = [0]
+
 
 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)
+        weight.normal_(mean=1.0, std=0.1)
         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
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance +
+                                                    self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
 
 
+@pytest.mark.parametrize("num_tokens", NUM_TOKENS)
+@pytest.mark.parametrize("hidden_size", HIDDEN_SIZES)
+@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,
     dtype: torch.dtype,
+    seed: int,
 ) -> None:
-    x = torch.randn(num_tokens, hidden_size, dtype=dtype, device='cuda')
+    torch.random.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
+
+    scale = float(hidden_size**-0.5)
+    x = torch.empty(num_tokens, hidden_size, dtype=dtype, device="cuda")
+    x.uniform_(-scale, scale)
     ref = RefRMSNorm(hidden_size).to(dtype).cuda()
 
     out = torch.empty_like(x)
@@ -38,17 +55,4 @@ def run_rms_norm(
         ref.variance_epsilon,
     )
     ref_out = ref(x)
-    assert torch.allclose(out, ref_out, atol=1e-3, rtol=1e-5)
-
-
-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,
-                )
+    assert torch.allclose(out, ref_out, atol=1e-2, rtol=1e-5)

tests/kernels/test_pos_encoding.py

@@ -1,47 +1,70 @@
-from typing import Tuple
+from typing import Optional, Tuple
 
+import pytest
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
 
 from vllm import pos_encoding_ops
 
+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, 12, 40, 52]  # Arbitrary values for testing
+NUM_TOKENS = [11, 83, 2048]  # Arbitrary values for testing
+SEEDS = [0]
 
-def rotate_half(x: torch.Tensor) -> torch.Tensor:
-    x1 = x[..., : x.shape[-1] // 2]
-    x2 = x[..., x.shape[-1] // 2 :]
+
+def rotate_neox(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(
+def rotate_gptj(x: torch.Tensor) -> torch.Tensor:
+    x1 = x[..., ::2]
+    x2 = x[..., 1::2]
+    x = torch.stack((-x2, x1), dim=-1)
+    return x.flatten(-2)
+
+
+def apply_rope(
     q: torch.Tensor,
     k: torch.Tensor,
     cos: torch.Tensor,
     sin: torch.Tensor,
+    is_neox_style: bool,
 ) -> Tuple[torch.Tensor, torch.Tensor]:
-    q_embed = (q * cos) + (rotate_half(q) * sin)
-    k_embed = (k * cos) + (rotate_half(k) * sin)
+    rotate_fn = rotate_neox if is_neox_style else rotate_gptj
+    q_embed = (q * cos) + (rotate_fn(q) * sin)
+    k_embed = (k * cos) + (rotate_fn(k) * sin)
     return q_embed, k_embed
 
 
-class RefRotaryEmbeddingNeox(nn.Module):
-    """Reference implementation of the GPT-NeoX style rotary embedding."""
+class RefRotaryEmbedding(nn.Module):
+    """Reference implementation of rotary embedding."""
 
     def __init__(
         self,
         dim: int,
-        max_position_embeddings: int = 2048,
+        is_neox_style: bool,
+        max_position_embeddings: int = 8192,
         base: int = 10000,
     ) -> None:
         super().__init__()
         self.rotary_dim = dim
+        self.is_neox_style = is_neox_style
         self.max_position_embeddings = max_position_embeddings
 
         # Create cos and sin embeddings.
-        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2) / dim))
+        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)
+        if is_neox_style:
+            emb = torch.cat((freqs, freqs), dim=-1)
+        else:
+            emb = torch.repeat_interleave(freqs, 2, -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)
@@ -49,22 +72,22 @@ class RefRotaryEmbeddingNeox(nn.Module):
 
     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]
+        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[..., :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, key_rot = apply_rope(query_rot, key_rot, cos, sin,
+                                        self.is_neox_style)
         query_rot = query_rot.transpose(0, 1).contiguous()
         key_rot = key_rot.transpose(0, 1).contiguous()
 
@@ -75,46 +98,69 @@ class RefRotaryEmbeddingNeox(nn.Module):
         return query, key
 
 
+@pytest.mark.parametrize("is_neox_style", IS_NEOX_STYLE)
+@pytest.mark.parametrize("num_tokens", NUM_TOKENS)
+@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(
+def test_rotary_embedding(
+    is_neox_style: bool,
     num_tokens: 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, num_heads * head_size, dtype=dtype, device='cuda')
-    key = torch.randn(num_tokens, num_heads * head_size, dtype=dtype, device='cuda')
+    if rotary_dim is None:
+        rotary_dim = head_size
+    torch.random.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
+
+    positions = torch.randint(0, max_position, (num_tokens, ), device="cuda")
+    query = torch.randn(num_tokens,
+                        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))
+    inv_freq = 1.0 / (base**(
+        torch.arange(0, rotary_dim, 2, dtype=torch.float) / rotary_dim))
     t = torch.arange(max_position).float()
-    freqs = torch.einsum('i,j -> ij', t, inv_freq.float())
+    freqs = torch.einsum("i,j -> ij", t, inv_freq)
     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')
+    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(
+    pos_encoding_ops.rotary_embedding(
         positions,
         out_query,
         out_key,
         head_size,
         cos_sin_cache,
+        is_neox_style,
     )
 
     # Run the reference implementation.
-    ref_rotary_embedding = RefRotaryEmbeddingNeox(
+    ref_rotary_embedding = RefRotaryEmbedding(
         dim=rotary_dim,
+        is_neox_style=is_neox_style,
         max_position_embeddings=max_position,
         base=base,
-    ).to(dtype=dtype, device='cuda')
+    ).to(dtype=dtype, device="cuda")
     ref_query, ref_key = ref_rotary_embedding(
         positions,
         query.view(num_tokens, num_heads, head_size),
@@ -124,19 +170,5 @@ def run_rotary_embedding_neox(
     ref_key = ref_key.view(num_tokens, num_heads * head_size)
 
     # 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 +1,45 @@
+"""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",
+    "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",
+]
+
+
+@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 +1,185 @@
+# pylint: disable=protected-access
+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.worker import Worker
+
+
+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):
+            with 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, Worker]:
+    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)
+    worker = Worker(None, None, None)
+    worker.block_size = 16
+    return input_tensor, fake_logits, sampler, worker
+
+
+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, worker = _prepare_test(batch_size)
+
+    seq_group_metadata_list = []
+    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]},
+            ))
+
+    _, _, input_metadata = worker._prepare_inputs(seq_group_metadata_list)
+    sampler_output = sampler(embedding=None,
+                             hidden_states=input_tensor,
+                             input_metadata=input_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, worker = _prepare_test(batch_size)
+
+    for i in range(batch_size):
+        fake_logits[i, i] = 1e2
+
+    seq_group_metadata_list = []
+    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]},
+            ))
+
+    _, _, input_metadata = worker._prepare_inputs(seq_group_metadata_list)
+    sampler_output = sampler(embedding=None,
+                             hidden_states=input_tensor,
+                             input_metadata=input_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, worker = _prepare_test(batch_size)
+
+    seq_group_metadata_list = []
+    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]},
+            ))
+
+    _, _, input_metadata = worker._prepare_inputs(seq_group_metadata_list)
+    sampler(embedding=None,
+            hidden_states=input_tensor,
+            input_metadata=input_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, worker = _prepare_test(batch_size)
+
+    seq_group_metadata_list = []
+    expected_tokens = []
+    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]},
+            ))
+
+    _, _, input_metadata = worker._prepare_inputs(seq_group_metadata_list)
+    sampler_output = sampler(embedding=None,
+                             hidden_states=input_tensor,
+                             input_metadata=input_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

vllm/__init__.py

@@ -1,3 +1,5 @@
+"""vLLM: a high-throughput and memory-efficient inference engine for LLMs"""
+
 from vllm.engine.arg_utils import AsyncEngineArgs, EngineArgs
 from vllm.engine.async_llm_engine import AsyncLLMEngine
 from vllm.engine.llm_engine import LLMEngine
@@ -6,7 +8,7 @@ from vllm.entrypoints.llm import LLM
 from vllm.outputs import CompletionOutput, RequestOutput
 from vllm.sampling_params import SamplingParams
 
-__version__ = "0.1.0"
+__version__ = "0.2.1"
 
 __all__ = [
     "LLM",

vllm/block.py

@@ -35,7 +35,8 @@ class LogicalTokenBlock:
 
     def append_tokens(self, token_ids: List[int]) -> None:
         assert len(token_ids) <= self.get_num_empty_slots()
-        self.token_ids[self.num_tokens:self.num_tokens + len(token_ids)] = token_ids
+        curr_idx = self.num_tokens
+        self.token_ids[curr_idx:curr_idx + len(token_ids)] = token_ids
         self.num_tokens += len(token_ids)
 
     def get_token_ids(self) -> List[int]:

vllm/config.py

@@ -1,14 +1,15 @@
 from typing import Optional
 
 import torch
-from transformers import AutoConfig, PretrainedConfig
+from transformers import PretrainedConfig
 
 from vllm.logger import init_logger
+from vllm.transformers_utils.config import get_config
 from vllm.utils import get_cpu_memory
 
 logger = init_logger(__name__)
 
-_GiB = 1 << 30
+_GB = 1 << 30
 
 
 class ModelConfig:
@@ -16,34 +17,101 @@ class ModelConfig:
 
     Args:
         model: Name or path of the huggingface model to use.
+        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__(
         self,
         model: str,
+        tokenizer: str,
+        tokenizer_mode: str,
+        trust_remote_code: bool,
         download_dir: Optional[str],
-        use_np_weights: bool,
-        use_dummy_weights: bool,
+        load_format: str,
         dtype: str,
         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: PretrainedConfig = AutoConfig.from_pretrained(model)
+        self.hf_config = get_config(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()
+        if tokenizer_mode not in ["auto", "slow"]:
+            raise ValueError(
+                f"Unknown tokenizer mode: {self.tokenizer_mode}. Must be "
+                "either 'auto' or 'slow'.")
+        self.tokenizer_mode = tokenizer_mode
+
+    def _verify_quantization(self) -> None:
+        supported_quantization = ["awq"]
+        if self.quantization is None:
+            return
+        quantization = self.quantization.lower()
+        if quantization not in supported_quantization:
+            raise ValueError(
+                f"Unknown quantization: {self.quantization}. Must be one of "
+                f"{supported_quantization}.")
+        self.quantization = quantization
 
     def verify_with_parallel_config(
         self,
@@ -72,7 +140,32 @@ 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:
+    def get_num_kv_heads(self, parallel_config: "ParallelConfig") -> int:
+        """Returns the number of KV heads per GPU worker."""
+        # 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
+        # For Falcon:
+        if getattr(self.hf_config, "n_head_kv", None) is not None:
+            return (self.hf_config.n_head_kv //
+                    parallel_config.tensor_parallel_size)
+        if getattr(self.hf_config, "num_kv_heads", None) is not None:
+            return (self.hf_config.num_kv_heads //
+                    parallel_config.tensor_parallel_size)
+        # For LLaMA-2:
+        if getattr(self.hf_config, "num_key_value_heads", None) is not None:
+            return (self.hf_config.num_key_value_heads //
+                    parallel_config.tensor_parallel_size)
         total_num_attention_heads = self.hf_config.num_attention_heads
         return total_num_attention_heads // parallel_config.tensor_parallel_size
 
@@ -90,15 +183,18 @@ class CacheConfig:
             vLLM execution.
         swap_space: Size of the CPU swap space per GPU (in GiB).
     """
+
     def __init__(
         self,
         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 * _GiB
+        self.swap_space_bytes = swap_space * _GB
+        self.sliding_window = sliding_window
         self._verify_args()
 
         # Will be set after profiling.
@@ -121,14 +217,13 @@ class CacheConfig:
         num_gpus_per_node = parallel_config.tensor_parallel_size
         cpu_memory_usage = self.swap_space_bytes * num_gpus_per_node
 
-        msg = (
-            f"{cpu_memory_usage / _GiB:.2f} GiB out of "
-            f"the {total_cpu_memory / _GiB:.2f} GiB total CPU memory is "
-            "allocated for the swap space.")
+        msg = (f"{cpu_memory_usage / _GB:.2f} GiB out of "
+               f"the {total_cpu_memory / _GB:.2f} GiB total CPU memory is "
+               "allocated for the swap space.")
         if cpu_memory_usage > 0.7 * total_cpu_memory:
             raise ValueError("Too large swap space. " + msg)
         elif cpu_memory_usage > 0.4 * total_cpu_memory:
-            logger.warn("Possibly too large swap space. " + msg)
+            logger.warning("Possibly too large swap space. " + msg)
 
 
 class ParallelConfig:
@@ -141,6 +236,7 @@ class ParallelConfig:
             True if either pipeline_parallel_size or tensor_parallel_size is
             greater than 1.
     """
+
     def __init__(
         self,
         pipeline_parallel_size: int,
@@ -170,14 +266,40 @@ class SchedulerConfig:
             a single iteration.
         max_num_seqs: Maximum number of sequences to be processed in a single
             iteration.
+        max_model_len: Maximum length of a sequence (including prompt
+            and generated text).
     """
+
     def __init__(
         self,
-        max_num_batched_tokens: int,
+        max_num_batched_tokens: Optional[int],
         max_num_seqs: int,
+        max_model_len: int,
     ) -> None:
-        self.max_num_batched_tokens = max_num_batched_tokens
+        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_model_len = max_model_len
+        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 = {
@@ -221,15 +343,59 @@ def _get_and_verify_dtype(
             pass
         else:
             # Casting between float16 and bfloat16 is allowed with a warning.
-            logger.warn(f"Casting {config_dtype} to {torch_dtype}.")
+            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",
+        # 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"]
+        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

@@ -27,8 +27,9 @@ class BlockAllocator:
         # Initialize the free blocks.
         self.free_blocks: List[PhysicalTokenBlock] = []
         for i in range(num_blocks):
-            block = PhysicalTokenBlock(
-                device=device, block_number=i, block_size=block_size)
+            block = PhysicalTokenBlock(device=device,
+                                       block_number=i,
+                                       block_size=block_size)
             self.free_blocks.append(block)
 
     def allocate(self) -> PhysicalTokenBlock:
@@ -62,10 +63,18 @@ 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
 
@@ -82,18 +91,27 @@ class BlockSpaceManager:
         # 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
+        return (num_free_gpu_blocks - num_required_blocks >=
+                self.watermark_blocks)
 
     def allocate(self, seq_group: SequenceGroup) -> None:
-        # NOTE: Here we assume that all sequences in the group have the same prompt.
+        # NOTE: Here we assume that all sequences in the group have the same
+        # prompt.
         seq = seq_group.get_seqs()[0]
 
         # 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)
@@ -115,11 +133,17 @@ 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]
@@ -143,16 +167,15 @@ class BlockSpaceManager:
         for block in src_block_table:
             block.ref_count += 1
 
-    def _get_physical_blocks(self, seq_group: SequenceGroup) -> List[PhysicalTokenBlock]:
+    def _get_physical_blocks(
+            self, seq_group: SequenceGroup) -> List[PhysicalTokenBlock]:
         # NOTE: Here, we assume that the physical blocks are only shared by
         # the sequences in the same group.
         blocks: Set[PhysicalTokenBlock] = set()
         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:
@@ -168,9 +191,7 @@ 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]
 
@@ -199,9 +220,7 @@ 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]
 
@@ -224,7 +243,7 @@ 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.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.
@@ -32,20 +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:
@@ -54,21 +59,23 @@ 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')
+        self.policy = PolicyFactory.get_policy(policy_name="fcfs")
         # Create the block space manager.
         self.block_manager = BlockSpaceManager(
             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.
@@ -76,25 +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
@@ -102,19 +114,81 @@ 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]]:
+    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]] = {}
 
         # 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)
+            num_batched_tokens = 0
+            # 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.
+                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):
+                    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
+
+                seq_group = self.waiting.pop(0)
+                self._allocate(seq_group)
+                self.running.append(seq_group)
+                num_batched_tokens += num_prompt_tokens
+                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=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=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)
@@ -144,129 +218,56 @@ 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 = len(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
-                # If the sequence group cannot be allocated, stop.
-                if not self.block_manager.can_allocate(seq_group):
-                    break
-
-                # If the number of batched tokens exceeds the limit, stop.
-                num_prompt_tokens = seq_group.get_seqs()[0].get_len()
-                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 = len(self.running)
-                if num_curr_seqs + num_new_seqs > self.scheduler_config.max_num_seqs:
+                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
-
-        # 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
+        return scheduler_outputs
 
     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 = 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
-
+        for seq_group in scheduler_outputs.scheduled_seq_groups:
             seq_data: Dict[int, List[SequenceData]] = {}
             block_tables: Dict[int, List[int]] = {}
             for seq in seq_group.get_seqs(status=SequenceStatus.RUNNING):
@@ -276,7 +277,7 @@ 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,
@@ -284,35 +285,10 @@ class Scheduler:
             seq_group_metadata_list.append(seq_group_metadata)
         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:
@@ -349,8 +325,8 @@ 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
@@ -358,8 +334,7 @@ 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
@@ -368,7 +343,7 @@ class Scheduler:
         elif preemption_mode == PreemptionMode.SWAP:
             self._preempt_by_swap(seq_group, blocks_to_swap_out)
         else:
-            assert False, 'Invalid preemption mode.'
+            assert False, "Invalid preemption mode."
 
     def _preempt_by_recompute(
         self,
@@ -388,9 +363,6 @@ 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

@@ -11,88 +11,165 @@ from vllm.config import (CacheConfig, ModelConfig, ParallelConfig,
 class EngineArgs:
     """Arguments for vLLM engine."""
     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
-    dtype: str = "auto"
+    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
     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
     disable_log_stats: bool = False
+    revision: Optional[str] = None
+    tokenizer_revision: Optional[str] = None
+    quantization: Optional[str] = None
 
     def __post_init__(self):
-        self.max_num_seqs = min(self.max_num_seqs, self.max_num_batched_tokens)
+        if self.tokenizer is None:
+            self.tokenizer = self.model
 
     @staticmethod
     def add_cli_args(
-        parser: argparse.ArgumentParser,
-    ) -> argparse.ArgumentParser:
+            parser: argparse.ArgumentParser) -> argparse.ArgumentParser:
         """Shared CLI arguments for vLLM engine."""
         # Model arguments
-        parser.add_argument('--model', type=str, default='facebook/opt-125m',
-                            help='name or path of the huggingface model to use')
-        parser.add_argument('--download-dir', type=str,
+        parser.add_argument(
+            '--model',
+            type=str,
+            default='facebook/opt-125m',
+            help='name or path of the huggingface model to use')
+        parser.add_argument(
+            '--tokenizer',
+            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,
+                            choices=['auto', 'slow'],
+                            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('--dtype', type=str, default=EngineArgs.dtype,
-                            choices=['auto', 'half', 'bfloat16', 'float'],
-                            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.')
+                            'default to the default cache dir of '
+                            'huggingface')
+        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', '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',
+        parser.add_argument('--worker-use-ray',
+                            action='store_true',
                             help='use Ray for distributed serving, will be '
-                                 'automatically set when using more than 1 GPU')
-        parser.add_argument('--pipeline-parallel-size', '-pp', type=int,
+                            'automatically set when using more than 1 GPU')
+        parser.add_argument('--pipeline-parallel-size',
+                            '-pp',
+                            type=int,
                             default=EngineArgs.pipeline_parallel_size,
                             help='number of pipeline stages')
-        parser.add_argument('--tensor-parallel-size', '-tp', type=int,
+        parser.add_argument('--tensor-parallel-size',
+                            '-tp',
+                            type=int,
                             default=EngineArgs.tensor_parallel_size,
                             help='number of tensor parallel replicas')
         # KV cache arguments
-        parser.add_argument('--block-size', type=int,
+        parser.add_argument('--block-size',
+                            type=int,
                             default=EngineArgs.block_size,
                             choices=[8, 16, 32],
                             help='token block size')
         # TODO(woosuk): Support fine-grained seeds (e.g., seed per request).
-        parser.add_argument('--seed', type=int, default=EngineArgs.seed,
+        parser.add_argument('--seed',
+                            type=int,
+                            default=EngineArgs.seed,
                             help='random seed')
-        parser.add_argument('--swap-space', type=int,
+        parser.add_argument('--swap-space',
+                            type=int,
                             default=EngineArgs.swap_space,
                             help='CPU swap space size (GiB) per GPU')
-        parser.add_argument('--gpu-memory-utilization', type=float,
+        parser.add_argument('--gpu-memory-utilization',
+                            type=float,
                             default=EngineArgs.gpu_memory_utilization,
                             help='the percentage of GPU memory to be used for'
-                                 'the model executor')
-        parser.add_argument('--max-num-batched-tokens', type=int,
+                            'the model executor')
+        parser.add_argument('--max-num-batched-tokens',
+                            type=int,
                             default=EngineArgs.max_num_batched_tokens,
                             help='maximum number of batched tokens per '
-                                 'iteration')
-        parser.add_argument('--max-num-seqs', type=int,
+                            'iteration')
+        parser.add_argument('--max-num-seqs',
+                            type=int,
                             default=EngineArgs.max_num_seqs,
                             help='maximum number of sequences per iteration')
-        parser.add_argument('--disable-log-stats', action='store_true',
+        parser.add_argument('--disable-log-stats',
+                            action='store_true',
                             help='disable logging statistics')
+        # Quantization settings.
+        parser.add_argument('--quantization',
+                            '-q',
+                            type=str,
+                            choices=['awq', None],
+                            default=None,
+                            help='Method used to quantize the weights')
         return parser
 
     @classmethod
-    def from_cli_args(cls, args: argparse.Namespace) -> "EngineArgs":
+    def from_cli_args(cls, args: argparse.Namespace) -> 'EngineArgs':
         # Get the list of attributes of this dataclass.
         attrs = [attr.name for attr in dataclasses.fields(cls)]
         # Set the attributes from the parsed arguments.
@@ -102,17 +179,21 @@ class EngineArgs:
     def create_engine_configs(
         self,
     ) -> Tuple[ModelConfig, CacheConfig, ParallelConfig, SchedulerConfig]:
-        # Initialize the configs.
-        model_config = ModelConfig(
-            self.model, self.download_dir, self.use_np_weights,
-            self.use_dummy_weights, self.dtype, self.seed)
-        cache_config = CacheConfig(self.block_size, self.gpu_memory_utilization,
-                                   self.swap_space)
+        model_config = ModelConfig(self.model, self.tokenizer,
+                                   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,
+            getattr(model_config.hf_config, 'sliding_window', None))
         parallel_config = ParallelConfig(self.pipeline_parallel_size,
                                          self.tensor_parallel_size,
                                          self.worker_use_ray)
         scheduler_config = SchedulerConfig(self.max_num_batched_tokens,
-                                           self.max_num_seqs)
+                                           self.max_num_seqs,
+                                           model_config.max_model_len)
         return model_config, cache_config, parallel_config, scheduler_config
 
 
@@ -121,15 +202,23 @@ 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(
-        parser: argparse.ArgumentParser,
-    ) -> argparse.ArgumentParser:
+            parser: argparse.ArgumentParser) -> argparse.ArgumentParser:
         parser = EngineArgs.add_cli_args(parser)
-        parser.add_argument('--engine-use-ray', action='store_true',
+        parser.add_argument('--engine-use-ray',
+                            action='store_true',
                             help='use Ray to start the LLM engine in a '
-                                 'separate process as the server process.')
-        parser.add_argument('--disable-log-requests', action='store_true',
+                            'separate process as the server process.')
+        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

@@ -1,7 +1,10 @@
 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
 from vllm.engine.llm_engine import LLMEngine
 from vllm.engine.ray_utils import initialize_cluster, ray
@@ -11,7 +14,219 @@ 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."""
+        all_outputs = []
+        for worker in self.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 = await asyncio.gather(*all_outputs)
+
+        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:
@@ -33,55 +248,156 @@ 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.
     """
-    def __init__(self, worker_use_ray: bool, engine_use_ray: bool,
-                 log_requests: bool = True, *args, **kwargs) -> None:
+
+    _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:
+            engine_class = ray.remote(num_gpus=1)(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)
 
-    async def generate(
+        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,
-        request_id: str,
-        prompt_token_ids: Optional[List[int]] = None
-    ) -> RequestOutput:
+        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,
+            prompt: Optional[str],
+            sampling_params: SamplingParams,
+            request_id: str,
+            prompt_token_ids: Optional[List[int]] = None) -> RequestOutput:
         """Generate outputs for a request.
 
         Generate outputs for a request. This method is a coroutine. It adds the
@@ -101,71 +417,23 @@ 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:
-                await self.engine_step(request_id)
-
-            # 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.
@@ -176,43 +444,53 @@ 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)
 
+    def _abort(self, request_id: str) -> None:
+        """Abort a request.
+
+        Abort a submitted request. If the request is finished or not found,
+        this method will be a no-op.
+
+        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."""
         if self.engine_use_ray:
-            await self.engine.abort_request.remote(request_id)
+            return await self.engine.get_model_config.remote()
         else:
-            self.engine.abort_request(request_id)
-
-        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]
-
-        # 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
+            return self.engine.get_model_config()
 
     @classmethod
-    def from_engine_args(cls, engine_args: AsyncEngineArgs) -> "AsyncLLMEngine":
+    def from_engine_args(cls,
+                         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_args.engine_use_ray,
-                     not engine_args.disable_log_requests,
                      *engine_configs,
-                     distributed_init_method, devices,
-                     log_stats=not engine_args.disable_log_stats)
+                     distributed_init_method,
+                     placement_group,
+                     log_requests=not engine_args.disable_log_requests,
+                     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

@@ -1,21 +1,34 @@
+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.tokenizer_utils import detokenize_incrementally, get_tokenizer
+from vllm.engine.ray_utils import RayWorker, 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, SequenceGroupOutputs,
+                           SequenceOutputs, 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.
@@ -41,8 +54,8 @@ 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.
     """
 
@@ -53,56 +66,122 @@ 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(
             "Initializing an LLM engine with config: "
             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"seed={model_config.seed})"
-        )
+            f"quantization={model_config.quantization}, "
+            f"seed={model_config.seed})")
         # TODO(woosuk): Print more configs in debug mode.
 
         self.model_config = model_config
         self.cache_config = cache_config
+        assert self.cache_config.sliding_window == getattr(
+            self.model_config.hf_config, "sliding_window", None)
         self.parallel_config = parallel_config
         self.scheduler_config = scheduler_config
         self.log_stats = log_stats
         self._verify_args()
 
-        self.tokenizer = get_tokenizer(model_config.model)
+        self.tokenizer = get_tokenizer(
+            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-5},
-                )(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  # pylint: disable=import-outside-toplevel
+
+        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,
+        )
+
+    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  # pylint: disable=import-outside-toplevel
+
+        self.workers: List[Worker] = []
+        for bundle in placement_group.bundle_specs:
+            if not bundle.get("GPU", 0):
+                continue
+            worker = ray.remote(
+                num_cpus=0,
+                num_gpus=1,
+                scheduling_strategy=PlacementGroupSchedulingStrategy(
+                    placement_group=placement_group,
+                    placement_group_capture_child_tasks=True),
+                **ray_remote_kwargs,
+            )(RayWorker).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,
+        )
 
     def _verify_args(self) -> None:
         self.model_config.verify_with_parallel_config(self.parallel_config)
@@ -125,8 +204,14 @@ class LLMEngine:
         num_gpu_blocks = min(b[0] for b in num_blocks)
         num_cpu_blocks = min(b[1] for b in num_blocks)
         # FIXME(woosuk): Change to debug log.
-        logger.info(f'# GPU blocks: {num_gpu_blocks}, '
-                    f'# CPU blocks: {num_cpu_blocks}')
+        logger.info(f"# GPU blocks: {num_gpu_blocks}, "
+                    f"# CPU blocks: {num_cpu_blocks}")
+
+        if num_gpu_blocks <= 0:
+            raise ValueError("No available memory for the cache blocks. "
+                             "Try increasing `gpu_memory_utilization` when "
+                             "initializing the engine.")
+
         self.cache_config.num_gpu_blocks = num_gpu_blocks
         self.cache_config.num_cpu_blocks = num_cpu_blocks
 
@@ -140,9 +225,12 @@ 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,
+        engine = cls(*engine_configs,
+                     distributed_init_method,
+                     placement_group,
                      log_stats=not engine_args.disable_log_stats)
         return engine
 
@@ -168,37 +256,38 @@ 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)
 
+    def get_model_config(self) -> ModelConfig:
+        """Gets the model configuration."""
+        return self.model_config
+
     def get_num_unfinished_requests(self) -> int:
         """Gets the number of unfinished requests."""
         return self.scheduler.get_num_unfinished_seq_groups()
@@ -207,6 +296,255 @@ 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: SequenceGroupOutputs) -> 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[SequenceOutputs] = 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.
 
@@ -216,10 +554,9 @@ 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 = self.scheduler.schedule()
-        if (not seq_group_metadata_list) and scheduler_outputs.is_empty():
-            # 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(
@@ -229,80 +566,136 @@ 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) + ignored
 
-        # Create the outputs.
-        request_outputs: List[RequestOutput] = []
-        for seq_group in 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
+        elapsed_time = now - self.last_logging_time
+        if elapsed_time < _LOGGING_INTERVAL_SEC:
+            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
+
+        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,
+                         sampling_params: 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=sampling_params.skip_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_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(
         self,
         method: str,
-        get_all_outputs: bool = False,
         *args,
+        get_all_outputs: bool = False,
         **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
+                executor = partial(worker.execute_method.remote, method)
+            else:
+                executor = getattr(worker, method)
 
             output = executor(*args, **kwargs)
             all_outputs.append(output)

vllm/engine/ray_utils.py

@@ -1,21 +1,59 @@
-import random
-from typing import List, Optional, Tuple
+import socket
+from typing import Optional, Tuple, TYPE_CHECKING
+
+from vllm.config import ParallelConfig
+from vllm.logger import init_logger
+
+logger = init_logger(__name__)
 
 try:
     import ray
-except ImportError:
+    from ray.air.util.torch_dist import TorchDistributedWorker
+
+    class RayWorker(TorchDistributedWorker):
+        """Ray wrapper for vllm.worker.Worker, allowing Worker to be
+        lazliy initialized after Ray sets CUDA_VISIBLE_DEVICES."""
+
+        def __init__(self, init_cached_hf_modules=False) -> None:
+            if init_cached_hf_modules:
+                # pylint: disable=import-outside-toplevel
+                from transformers.dynamic_module_utils import init_hf_modules
+                init_hf_modules()
+            self.worker = None
+
+        def init_worker(self, worker_init_fn):
+            self.worker = worker_init_fn()
+
+        def __getattr__(self, name):
+            return getattr(self.worker, name)
+
+        def execute_method(self, method, *args, **kwargs):
+            executor = getattr(self, method)
+            return executor(*args, **kwargs)
+
+except ImportError as e:
+    logger.warning(f"Failed to import Ray with {e!r}. "
+                   "For distributed inference, please install Ray with "
+                   "`pip install ray pandas pyarrow`.")
     ray = None
+    TorchDistributedWorker = None
+    RayWorker = None  # pylint: disable=invalid-name
 
-from vllm.config import ParallelConfig
+if TYPE_CHECKING:
+    from ray.util.placement_group import PlacementGroup
 
-DeviceID = Tuple[int, Optional[str], int]  # rank, node resource (node IP), device id
+
+def get_open_port():
+    with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
+        s.bind(("", 0))
+        return s.getsockname()[1]
 
 
 def initialize_cluster(
     parallel_config: ParallelConfig,
     engine_use_ray: bool = False,
     ray_address: Optional[str] = None,
-) -> Tuple[str, List[List[DeviceID]]]:
+) -> Tuple[str, Optional["PlacementGroup"]]:
     """Initialize the distributed cluster probably with Ray.
 
     Args:
@@ -25,11 +63,10 @@ def initialize_cluster(
             the default Ray cluster address.
 
     Returns:
-        A tuple of (`distributed_init_method`, `all_stage_devices`). The
+        A tuple of (`distributed_init_method`, `placement_group`). The
         `distributed_init_method` is the address for initializing the
-        distributed backend. `all_stage_devices` includes device IDs for
-        each worker in each pipeline stage. Each device ID is a tuple of
-        (rank, node resource, device id).
+        distributed backend. `placement_group` includes the specification
+        of the resources for each distributed worker.
     """
     if parallel_config.worker_use_ray or engine_use_ray:
         if ray is None:
@@ -37,71 +74,46 @@ def initialize_cluster(
                 "Ray is not installed. Please install Ray to use distributed "
                 "serving.")
         # Connect to a ray cluster.
-        ray.init(address=ray_address)
+        ray.init(address=ray_address, ignore_reinit_error=True)
 
     if not parallel_config.worker_use_ray:
         # Initialize cluster locally.
-        port = random.randint(10000, 20000)
+        port = get_open_port()
         # We need to setup the distributed init method to make sure
         # the distributed megatron code (e.g., get world size) works correctly.
         distributed_init_method = f"tcp://localhost:{port}"
-        all_stage_devices = [[(0, None, 0)]]
-        return distributed_init_method, all_stage_devices
+        return distributed_init_method, None
 
-    # Assume we have a uniform cluster that each node has the same number of
-    # GPUs for now.
-    valid_node_resources = []
-    num_devices_per_node = None
-    for node in ray.nodes():
-        if (not node['Alive']) or node['Resources']['GPU'] <= 0:
-            continue
-        if num_devices_per_node is None:
-            num_devices_per_node = node['Resources']['GPU']
-        else:
-            assert num_devices_per_node == node['Resources']['GPU'], (
-                "The number of GPUs per node is not uniform.")
-        for key in node['Resources']:
-            if key.startswith('node:'):
-                valid_node_resources.append(key)
-
-    # Verify the parallel config.
-    num_nodes = len(valid_node_resources)
-    if parallel_config.world_size > num_nodes * num_devices_per_node:
-        raise ValueError(
-            "The number of required GPUs exceeds the total number of "
-            "available GPUs.")
-    if parallel_config.tensor_parallel_size >= num_devices_per_node:
-        if parallel_config.tensor_parallel_size % num_devices_per_node != 0:
+    current_placement_group = ray.util.get_current_placement_group()
+    if current_placement_group:
+        # We are in a placement group
+        bundles = current_placement_group.bundle_specs
+        # Verify that we can use the placement group.
+        gpu_bundles = 0
+        for bundle in bundles:
+            bundle_gpus = bundle.get("GPU", 0)
+            if bundle_gpus > 1:
+                raise ValueError(
+                    "Placement group bundle cannot have more than 1 GPU.")
+            if bundle_gpus:
+                gpu_bundles += 1
+        if parallel_config.world_size > gpu_bundles:
             raise ValueError(
-                "The number of tensor parallelism is not divisible by the "
-                "number of GPUs per node.")
+                "The number of required GPUs exceeds the total number of "
+                "available GPUs in the placement group.")
     else:
-        if num_devices_per_node % parallel_config.tensor_parallel_size != 0:
+        num_gpus_in_cluster = ray.cluster_resources().get("GPU", 0)
+        if parallel_config.world_size > num_gpus_in_cluster:
             raise ValueError(
-                "The number of GPUs per node is not divisible by the number "
-                "of tensor parallelism.")
+                "The number of required GPUs exceeds the total number of "
+                "available GPUs in the cluster.")
+        # Create a new placement group
+        current_placement_group = ray.util.placement_group([{
+            "GPU": 1
+        }] * parallel_config.world_size)
+        # Wait until PG is ready - this will block until all
+        # requested resources are available, and will timeout
+        # if they cannot be provisioned.
+        ray.get(current_placement_group.ready(), timeout=1800)
 
-    # Assign GPUs to pipeline stages.
-    rank = 0
-    current_node_id = 0
-    current_device_id = 0
-    distributed_init_method = None
-    all_stage_devices = []
-
-    for _ in range(parallel_config.pipeline_parallel_size):
-        stage_devices = []
-        for _ in range(parallel_config.tensor_parallel_size):
-            node_resource = valid_node_resources[current_node_id]
-            stage_devices.append((rank, node_resource, current_device_id))
-            if distributed_init_method is None:
-                ip = node_resource.split("node:")[-1]
-                port = random.randint(10000, 20000)
-                distributed_init_method = f"tcp://{ip}:{port}"
-            rank += 1
-            current_device_id += 1
-            if current_device_id >= num_devices_per_node:
-                current_node_id += 1
-                current_device_id = 0
-        all_stage_devices.append(stage_devices)
-
-    return distributed_init_method, all_stage_devices
+    return None, current_placement_group

vllm/engine/tokenizer_utils.py

@@ -1,92 +0,0 @@
-from typing import List, Tuple, Union
-
-from transformers import (AutoConfig, AutoTokenizer, PreTrainedTokenizer,
-                          PreTrainedTokenizerFast)
-
-from vllm.logger import init_logger
-
-logger = init_logger(__name__)
-
-_MODEL_TYPES_WITH_SLOW_TOKENIZER = []
-
-
-def get_tokenizer(
-    model_name: str,
-    *args,
-    **kwargs,
-) -> Union[PreTrainedTokenizer, PreTrainedTokenizerFast]:
-    """Gets a tokenizer for the given model name via Huggingface."""
-    config = AutoConfig.from_pretrained(model_name)
-    if "open_llama" in model_name:
-        kwargs["use_fast"] = False
-        logger.info(
-            "OpenLLaMA models do not support the fast tokenizer. "
-            "Using the slow tokenizer instead.")
-    elif config.model_type == "llama" and getattr(kwargs, "use_fast", True):
-        # LLaMA fast tokenizer causes protobuf errors in some environments.
-        # However, we found that the below LLaMA fast tokenizer works well in
-        # most environments.
-        model_name = "hf-internal-testing/llama-tokenizer"
-        logger.info(
-            f"Using the LLaMA fast tokenizer in '{model_name}' to avoid "
-            "potential protobuf errors.")
-    elif config.model_type in _MODEL_TYPES_WITH_SLOW_TOKENIZER:
-        if getattr(kwargs, "use_fast", False) == True:
-            raise ValueError(
-                f"Cannot use the fast tokenizer for {config.model_type} due to "
-                "bugs in the fast tokenizer.")
-        logger.info(
-            f"Using the slow tokenizer for {config.model_type} due to bugs in "
-            "the fast tokenizer. This could potentially lead to performance "
-            "degradation.")
-        kwargs["use_fast"] = False
-    return AutoTokenizer.from_pretrained(model_name, *args, **kwargs)
-
-
-def detokenize_incrementally(
-    tokenizer: Union[PreTrainedTokenizer, PreTrainedTokenizerFast],
-    prev_output_tokens: List[str],
-    new_token_id: int,
-    skip_special_tokens: bool,
-) -> Tuple[str, str]:
-    """Detokenizes the new token in conjuction with the previous output tokens.
-
-    NOTE: This function does not update prev_output_tokens.
-
-    Returns:
-        new_token: The new token as a string.
-        output_text: The new output text as a string.
-    """
-    new_token = tokenizer.convert_ids_to_tokens(
-        new_token_id, skip_special_tokens=skip_special_tokens)
-    output_tokens = prev_output_tokens + [new_token]
-
-    # Convert the tokens to a string.
-    # Optimization: If the tokenizer does not have `added_tokens_encoder`,
-    # then we can directly use `convert_tokens_to_string`.
-    if not getattr(tokenizer, "added_tokens_encoder", {}):
-        output_text = tokenizer.convert_tokens_to_string(output_tokens)
-        return new_token, output_text
-
-    # Adapted from https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/tokenization_utils.py#L921
-    # NOTE(woosuk): The following code is slow because it runs a for loop over
-    # the output_tokens. In Python, running a for loop over a list can be slow
-    # even when the loop body is very simple.
-    sub_texts = []
-    current_sub_text = []
-    for token in output_tokens:
-        if skip_special_tokens and token in tokenizer.all_special_ids:
-            continue
-        if token in tokenizer.added_tokens_encoder:
-            if current_sub_text:
-                sub_text = tokenizer.convert_tokens_to_string(current_sub_text)
-                sub_texts.append(sub_text)
-                current_sub_text = []
-            sub_texts.append(token)
-        else:
-            current_sub_text.append(token)
-    if current_sub_text:
-        sub_text = tokenizer.convert_tokens_to_string(current_sub_text)
-        sub_texts.append(sub_text)
-    output_text = " ".join(sub_texts)
-    return new_token, output_text

vllm/entrypoints/api_server.py

@@ -2,8 +2,8 @@ import argparse
 import json
 from typing import AsyncGenerator
 
-from fastapi import BackgroundTasks, FastAPI, Request
-from fastapi.responses import Response, StreamingResponse
+from fastapi import FastAPI, Request
+from fastapi.responses import JSONResponse, Response, StreamingResponse
 import uvicorn
 
 from vllm.engine.arg_utils import AsyncEngineArgs
@@ -11,9 +11,10 @@ from vllm.engine.async_llm_engine import AsyncLLMEngine
 from vllm.sampling_params import SamplingParams
 from vllm.utils import random_uuid
 
-TIMEOUT_KEEP_ALIVE = 5 # seconds.
-TIMEOUT_TO_PREVENT_DEADLOCK = 1 # seconds
+TIMEOUT_KEEP_ALIVE = 5  # seconds.
+TIMEOUT_TO_PREVENT_DEADLOCK = 1  # seconds.
 app = FastAPI()
+engine = None
 
 
 @app.post("/generate")
@@ -30,6 +31,7 @@ async def generate(request: Request) -> Response:
     stream = request_dict.pop("stream", False)
     sampling_params = SamplingParams(**request_dict)
     request_id = random_uuid()
+
     results_generator = engine.generate(prompt, sampling_params, request_id)
 
     # Streaming case
@@ -37,20 +39,13 @@ async def generate(request: Request) -> Response:
         async for request_output in results_generator:
             prompt = request_output.prompt
             text_outputs = [
-                prompt + output.text
-                for output in request_output.outputs
+                prompt + output.text for output in request_output.outputs
             ]
             ret = {"text": text_outputs}
             yield (json.dumps(ret) + "\0").encode("utf-8")
 
-    async def abort_request() -> None:
-        await engine.abort(request_id)
-
     if stream:
-        background_tasks = BackgroundTasks()
-        # Abort the request if the client disconnects.
-        background_tasks.add_task(abort_request)
-        return StreamingResponse(stream_results(), background=background_tasks)
+        return StreamingResponse(stream_results())
 
     # Non-streaming case
     final_output = None
@@ -63,17 +58,14 @@ async def generate(request: Request) -> Response:
 
     assert final_output is not None
     prompt = final_output.prompt
-    text_outputs = [
-        prompt + output.text
-        for output in final_output.outputs
-    ]
+    text_outputs = [prompt + output.text for output in final_output.outputs]
     ret = {"text": text_outputs}
-    return Response(content=json.dumps(ret))
+    return JSONResponse(ret)
 
 
 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=8000)
     parser = AsyncEngineArgs.add_cli_args(parser)
     args = parser.parse_args()
@@ -81,5 +73,8 @@ if __name__ == "__main__":
     engine_args = AsyncEngineArgs.from_cli_args(args)
     engine = AsyncLLMEngine.from_engine_args(engine_args)
 
-    uvicorn.run(app, host=args.host, port=args.port, log_level="debug",
+    uvicorn.run(app,
+                host=args.host,
+                port=args.port,
+                log_level="debug",
                 timeout_keep_alive=TIMEOUT_KEEP_ALIVE)

vllm/entrypoints/llm.py

@@ -25,6 +25,11 @@ class LLM:
 
     Args:
         model: The name or path of a HuggingFace Transformers model.
+        tokenizer: The name or path of a HuggingFace Transformers tokenizer.
+        tokenizer_mode: The 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.
         tensor_parallel_size: The number of GPUs to use for distributed
             execution with tensor parallelism.
         dtype: The data type for the model weights and activations. Currently,
@@ -32,34 +37,72 @@ class LLM:
             the `torch_dtype` attribute specified in the model config file.
             However, if the `torch_dtype` in the config is `float32`, we will
             use `float16` instead.
+        quantization: The method used to quantize the model weights. Currently,
+            we support "awq". If None, we assume the model weights are not
+            quantized and use `dtype` to determine the data type of the weights.
+        revision: The specific model version to use. It can be a branch name,
+            a tag name, or a commit id.
+        tokenizer_revision: The specific tokenizer version to use. It can be a
+            branch name, a tag name, or a commit id.
         seed: The seed to initialize the random number generator for sampling.
+        gpu_memory_utilization: The ratio (between 0 and 1) of GPU memory to
+            reserve for the model weights, activations, and KV cache. Higher
+            values will increase the KV cache size and thus improve the model's
+            throughput. However, if the value is too high, it may cause out-of-
+            memory (OOM) errors.
+        swap_space: The size (GiB) of CPU memory per GPU to use as swap space.
+            This can be used for temporarily storing the states of the requests
+            when their `best_of` sampling parameters are larger than 1. If all
+            requests will have `best_of=1`, you can safely set this to 0.
+            Otherwise, too small values may cause out-of-memory (OOM) errors.
     """
 
     def __init__(
         self,
         model: str,
+        tokenizer: Optional[str] = None,
+        tokenizer_mode: str = "auto",
+        trust_remote_code: bool = False,
         tensor_parallel_size: int = 1,
         dtype: str = "auto",
+        quantization: Optional[str] = None,
+        revision: Optional[str] = None,
+        tokenizer_revision: Optional[str] = None,
         seed: int = 0,
+        gpu_memory_utilization: float = 0.9,
+        swap_space: int = 4,
         **kwargs,
     ) -> None:
         if "disable_log_stats" not in kwargs:
             kwargs["disable_log_stats"] = True
         engine_args = EngineArgs(
             model=model,
+            tokenizer=tokenizer,
+            tokenizer_mode=tokenizer_mode,
+            trust_remote_code=trust_remote_code,
             tensor_parallel_size=tensor_parallel_size,
             dtype=dtype,
+            quantization=quantization,
+            revision=revision,
+            tokenizer_revision=tokenizer_revision,
             seed=seed,
+            gpu_memory_utilization=gpu_memory_utilization,
+            swap_space=swap_space,
             **kwargs,
         )
         self.llm_engine = LLMEngine.from_engine_args(engine_args)
         self.request_counter = Counter()
 
     def get_tokenizer(
-        self,
-    ) -> Union[PreTrainedTokenizer, PreTrainedTokenizerFast]:
+            self) -> Union[PreTrainedTokenizer, PreTrainedTokenizerFast]:
         return self.llm_engine.tokenizer
 
+    def set_tokenizer(
+        self,
+        tokenizer: Union[PreTrainedTokenizer, PreTrainedTokenizerFast],
+    ) -> None:
+        self.llm_engine.tokenizer = tokenizer
+
     def generate(
         self,
         prompts: Optional[Union[str, List[str]]] = None,
@@ -133,10 +176,14 @@ class LLM:
         while self.llm_engine.has_unfinished_requests():
             step_outputs = self.llm_engine.step()
             for output in step_outputs:
-                if output.finished():
+                if output.finished:
                     outputs.append(output)
                     if use_tqdm:
                         pbar.update(1)
         if use_tqdm:
             pbar.close()
+        # Sort the outputs by request ID.
+        # This is necessary because some requests may be finished earlier than
+        # its previous requests.
+        outputs = sorted(outputs, key=lambda x: int(x.request_id))
         return outputs

vllm/entrypoints/openai/api_server.py

@@ -1,47 +1,61 @@
-# Adapted from https://github.com/lm-sys/FastChat/blob/168ccc29d3f7edc50823016105c024fe2282732a/fastchat/serve/openai_api_server.py
+# Adapted from
+# https://github.com/lm-sys/FastChat/blob/168ccc29d3f7edc50823016105c024fe2282732a/fastchat/serve/openai_api_server.py
 
 import argparse
-from http import HTTPStatus
+import asyncio
 import json
 import time
-from typing import AsyncGenerator, Dict, List, Optional
+from http import HTTPStatus
+from typing import AsyncGenerator, Dict, List, Optional, Tuple, Union
 
 import fastapi
-from fastapi import BackgroundTasks, Request
+import uvicorn
+from fastapi import Request
 from fastapi.exceptions import RequestValidationError
 from fastapi.middleware.cors import CORSMiddleware
 from fastapi.responses import JSONResponse, StreamingResponse
-import uvicorn
+from packaging import version
 
 from vllm.engine.arg_utils import AsyncEngineArgs
 from vllm.engine.async_llm_engine import AsyncLLMEngine
-from vllm.engine.tokenizer_utils import get_tokenizer
 from vllm.entrypoints.openai.protocol import (
     CompletionRequest, CompletionResponse, CompletionResponseChoice,
-    CompletionResponseStreamChoice, CompletionStreamResponse, ErrorResponse,
+    CompletionResponseStreamChoice, CompletionStreamResponse,
+    ChatCompletionRequest, ChatCompletionResponse,
+    ChatCompletionResponseChoice, ChatCompletionResponseStreamChoice,
+    ChatCompletionStreamResponse, ChatMessage, DeltaMessage, ErrorResponse,
     LogProbs, ModelCard, ModelList, ModelPermission, UsageInfo)
 from vllm.logger import init_logger
 from vllm.outputs import RequestOutput
 from vllm.sampling_params import SamplingParams
+from vllm.transformers_utils.tokenizer import get_tokenizer
 from vllm.utils import random_uuid
 
-TIMEOUT_KEEP_ALIVE = 5 # seconds
+try:
+    import fastchat
+    from fastchat.conversation import Conversation, SeparatorStyle
+    from fastchat.model.model_adapter import get_conversation_template
+    _fastchat_available = True
+except ImportError:
+    _fastchat_available = False
+
+TIMEOUT_KEEP_ALIVE = 5  # seconds
 
 logger = init_logger(__name__)
 served_model = None
 app = fastapi.FastAPI()
+engine = None
 
 
 def create_error_response(status_code: HTTPStatus,
                           message: str) -> JSONResponse:
-    return JSONResponse(
-        ErrorResponse(message=message, type="invalid_request_error").dict(),
-        status_code=status_code.value
-    )
+    return JSONResponse(ErrorResponse(message=message,
+                                      type="invalid_request_error").dict(),
+                        status_code=status_code.value)
 
 
 @app.exception_handler(RequestValidationError)
-async def validation_exception_handler(request, exc):
+async def validation_exception_handler(request, exc):  # pylint: disable=unused-argument
     return create_error_response(HTTPStatus.BAD_REQUEST, str(exc))
 
 
@@ -55,11 +69,90 @@ async def check_model(request) -> Optional[JSONResponse]:
     return ret
 
 
+async def get_gen_prompt(request) -> str:
+    if not _fastchat_available:
+        raise ModuleNotFoundError(
+            "fastchat is not installed. Please install fastchat to use "
+            "the chat completion and conversation APIs: `$ pip install fschat`"
+        )
+    if version.parse(fastchat.__version__) < version.parse("0.2.23"):
+        raise ImportError(
+            f"fastchat version is low. Current version: {fastchat.__version__} "
+            "Please upgrade fastchat to use: `$ pip install -U fschat`")
+
+    conv = get_conversation_template(request.model)
+    conv = Conversation(
+        name=conv.name,
+        system_template=conv.system_template,
+        system_message=conv.system_message,
+        roles=conv.roles,
+        messages=list(conv.messages),  # prevent in-place modification
+        offset=conv.offset,
+        sep_style=SeparatorStyle(conv.sep_style),
+        sep=conv.sep,
+        sep2=conv.sep2,
+        stop_str=conv.stop_str,
+        stop_token_ids=conv.stop_token_ids,
+    )
+
+    if isinstance(request.messages, str):
+        prompt = request.messages
+    else:
+        for message in request.messages:
+            msg_role = message["role"]
+            if msg_role == "system":
+                conv.system_message = message["content"]
+            elif msg_role == "user":
+                conv.append_message(conv.roles[0], message["content"])
+            elif msg_role == "assistant":
+                conv.append_message(conv.roles[1], message["content"])
+            else:
+                raise ValueError(f"Unknown role: {msg_role}")
+
+        # Add a blank message for the assistant.
+        conv.append_message(conv.roles[1], None)
+        prompt = conv.get_prompt()
+
+    return prompt
+
+
+async def check_length(
+    request: Union[ChatCompletionRequest, CompletionRequest],
+    prompt: Optional[str] = None,
+    prompt_ids: Optional[List[int]] = None
+) -> Tuple[List[int], Optional[JSONResponse]]:
+    assert (not (prompt is None and prompt_ids is None)
+            and not (prompt is not None and prompt_ids is not None)
+            ), "Either prompt or prompt_ids should be provided."
+    if prompt_ids is not None:
+        input_ids = prompt_ids
+    else:
+        input_ids = tokenizer(prompt).input_ids
+    token_num = len(input_ids)
+
+    if request.max_tokens is None:
+        request.max_tokens = max_model_len - token_num
+    if token_num + request.max_tokens > max_model_len:
+        return input_ids, create_error_response(
+            HTTPStatus.BAD_REQUEST,
+            f"This model's maximum context length is {max_model_len} tokens. "
+            f"However, you requested {request.max_tokens + token_num} tokens "
+            f"({token_num} in the messages, "
+            f"{request.max_tokens} in the completion). "
+            f"Please reduce the length of the messages or completion.",
+        )
+    else:
+        return input_ids, None
+
+
 @app.get("/v1/models")
 async def show_available_models():
     """Show available models. Right now we only have one model."""
-    model_cards = [ModelCard(id=served_model, root=served_model,
-                             permission=[ModelPermission()])]
+    model_cards = [
+        ModelCard(id=served_model,
+                  root=served_model,
+                  permission=[ModelPermission()])
+    ]
     return ModelList(data=model_cards)
 
 
@@ -76,17 +169,184 @@ def create_logprobs(token_ids: List[int],
         if len(logprobs.text_offset) == 0:
             logprobs.text_offset.append(initial_text_offset)
         else:
-            logprobs.text_offset.append(logprobs.text_offset[-1] + last_token_len)
+            logprobs.text_offset.append(logprobs.text_offset[-1] +
+                                        last_token_len)
         last_token_len = len(token)
 
-        logprobs.top_logprobs.append(
-            {tokenizer.convert_ids_to_tokens(i): p
-             for i, p in id_logprob.items()})
+        logprobs.top_logprobs.append({
+            tokenizer.convert_ids_to_tokens(i): p
+            for i, p in id_logprob.items()
+        })
     return logprobs
 
 
+@app.post("/v1/chat/completions")
+async def create_chat_completion(request: ChatCompletionRequest,
+                                 raw_request: Request):
+    """Completion API similar to OpenAI's API.
+
+    See  https://platform.openai.com/docs/api-reference/chat/create
+    for the API specification. This API mimics the OpenAI ChatCompletion API.
+
+    NOTE: Currently we do not support the following features:
+        - function_call (Users should implement this by themselves)
+        - logit_bias (to be supported by vLLM engine)
+    """
+    logger.info(f"Received chat completion request: {request}")
+
+    error_check_ret = await check_model(request)
+    if error_check_ret is not None:
+        return error_check_ret
+
+    if request.logit_bias is not None and len(request.logit_bias) > 0:
+        # TODO: support logit_bias in vLLM engine.
+        return create_error_response(HTTPStatus.BAD_REQUEST,
+                                     "logit_bias is not currently supported")
+
+    prompt = await get_gen_prompt(request)
+    token_ids, error_check_ret = await check_length(request, prompt=prompt)
+    if error_check_ret is not None:
+        return error_check_ret
+
+    model_name = request.model
+    request_id = f"cmpl-{random_uuid()}"
+    created_time = int(time.monotonic())
+    try:
+        sampling_params = SamplingParams(
+            n=request.n,
+            presence_penalty=request.presence_penalty,
+            frequency_penalty=request.frequency_penalty,
+            temperature=request.temperature,
+            top_p=request.top_p,
+            stop=request.stop,
+            stop_token_ids=request.stop_token_ids,
+            max_tokens=request.max_tokens,
+            best_of=request.best_of,
+            top_k=request.top_k,
+            ignore_eos=request.ignore_eos,
+            use_beam_search=request.use_beam_search,
+            skip_special_tokens=request.skip_special_tokens,
+        )
+    except ValueError as e:
+        return create_error_response(HTTPStatus.BAD_REQUEST, str(e))
+
+    result_generator = engine.generate(prompt, sampling_params, request_id,
+                                       token_ids)
+
+    def create_stream_response_json(
+        index: int,
+        text: str,
+        finish_reason: Optional[str] = None,
+    ) -> str:
+        choice_data = ChatCompletionResponseStreamChoice(
+            index=index,
+            delta=DeltaMessage(content=text),
+            finish_reason=finish_reason,
+        )
+        response = ChatCompletionStreamResponse(
+            id=request_id,
+            created=created_time,
+            model=model_name,
+            choices=[choice_data],
+        )
+        response_json = response.json(ensure_ascii=False)
+
+        return response_json
+
+    async def completion_stream_generator() -> AsyncGenerator[str, None]:
+        # First chunk with role
+        for i in range(request.n):
+            choice_data = ChatCompletionResponseStreamChoice(
+                index=i,
+                delta=DeltaMessage(role="assistant"),
+                finish_reason=None,
+            )
+            chunk = ChatCompletionStreamResponse(id=request_id,
+                                                 choices=[choice_data],
+                                                 model=model_name)
+            data = chunk.json(exclude_unset=True, ensure_ascii=False)
+            yield f"data: {data}\n\n"
+
+        previous_texts = [""] * request.n
+        previous_num_tokens = [0] * request.n
+        async for res in result_generator:
+            res: RequestOutput
+            for output in res.outputs:
+                i = output.index
+                delta_text = output.text[len(previous_texts[i]):]
+                previous_texts[i] = output.text
+                previous_num_tokens[i] = len(output.token_ids)
+                response_json = create_stream_response_json(
+                    index=i,
+                    text=delta_text,
+                )
+                yield f"data: {response_json}\n\n"
+                if output.finish_reason is not None:
+                    response_json = create_stream_response_json(
+                        index=i,
+                        text="",
+                        finish_reason=output.finish_reason,
+                    )
+                    yield f"data: {response_json}\n\n"
+        yield "data: [DONE]\n\n"
+
+    # Streaming response
+    if request.stream:
+        return StreamingResponse(completion_stream_generator(),
+                                 media_type="text/event-stream")
+
+    # Non-streaming response
+    final_res: RequestOutput = None
+    async for res in result_generator:
+        if await raw_request.is_disconnected():
+            # Abort the request if the client disconnects.
+            await engine.abort(request_id)
+            return create_error_response(HTTPStatus.BAD_REQUEST,
+                                         "Client disconnected")
+        final_res = res
+    assert final_res is not None
+    choices = []
+    for output in final_res.outputs:
+        choice_data = ChatCompletionResponseChoice(
+            index=output.index,
+            message=ChatMessage(role="assistant", content=output.text),
+            finish_reason=output.finish_reason,
+        )
+        choices.append(choice_data)
+
+    num_prompt_tokens = len(final_res.prompt_token_ids)
+    num_generated_tokens = sum(
+        len(output.token_ids) for output in final_res.outputs)
+    usage = UsageInfo(
+        prompt_tokens=num_prompt_tokens,
+        completion_tokens=num_generated_tokens,
+        total_tokens=num_prompt_tokens + num_generated_tokens,
+    )
+    response = ChatCompletionResponse(
+        id=request_id,
+        created=created_time,
+        model=model_name,
+        choices=choices,
+        usage=usage,
+    )
+
+    if request.stream:
+        # When user requests streaming but we don't stream, we still need to
+        # return a streaming response with a single event.
+        response_json = response.json(ensure_ascii=False)
+
+        async def fake_stream_generator() -> AsyncGenerator[str, None]:
+            yield f"data: {response_json}\n\n"
+            yield "data: [DONE]\n\n"
+
+        return StreamingResponse(fake_stream_generator(),
+                                 media_type="text/event-stream")
+
+    return response
+
+
 @app.post("/v1/completions")
-async def create_completion(raw_request: Request):
+async def create_completion(request: CompletionRequest, raw_request: Request):
     """Completion API similar to OpenAI's API.
 
     See https://platform.openai.com/docs/api-reference/completions/create
@@ -99,7 +359,6 @@ async def create_completion(raw_request: Request):
           suffix)
         - logit_bias (to be supported by vLLM engine)
     """
-    request = CompletionRequest(**await raw_request.json())
     logger.info(f"Received completion request: {request}")
 
     error_check_ret = await check_model(request)
@@ -115,17 +374,44 @@ async def create_completion(raw_request: Request):
     if request.suffix is not None:
         # The language models we currently support do not support suffix.
         return create_error_response(HTTPStatus.BAD_REQUEST,
-                                    "suffix is not currently supported")
+                                     "suffix is not currently supported")
 
-    if request.logit_bias is not None:
+    if request.logit_bias is not None and len(request.logit_bias) > 0:
         # TODO: support logit_bias in vLLM engine.
         return create_error_response(HTTPStatus.BAD_REQUEST,
                                      "logit_bias is not currently supported")
 
     model_name = request.model
     request_id = f"cmpl-{random_uuid()}"
-    prompt = request.prompt
-    created_time = int(time.time())
+
+    use_token_ids = False
+    if isinstance(request.prompt, list):
+        if len(request.prompt) == 0:
+            return create_error_response(HTTPStatus.BAD_REQUEST,
+                                         "please provide at least one prompt")
+        first_element = request.prompt[0]
+        if isinstance(first_element, int):
+            use_token_ids = True
+            prompt = request.prompt
+        elif isinstance(first_element, (str, list)):
+            # TODO: handles multiple prompt case in list[list[int]]
+            if len(request.prompt) > 1:
+                return create_error_response(
+                    HTTPStatus.BAD_REQUEST,
+                    "multiple prompts in a batch is not currently supported")
+            use_token_ids = not isinstance(first_element, str)
+            prompt = request.prompt[0]
+    else:
+        prompt = request.prompt
+
+    if use_token_ids:
+        _, error_check_ret = await check_length(request, prompt_ids=prompt)
+    else:
+        token_ids, error_check_ret = await check_length(request, prompt=prompt)
+    if error_check_ret is not None:
+        return error_check_ret
+
+    created_time = int(time.monotonic())
     try:
         sampling_params = SamplingParams(
             n=request.n,
@@ -136,30 +422,37 @@ async def create_completion(raw_request: Request):
             top_p=request.top_p,
             top_k=request.top_k,
             stop=request.stop,
+            stop_token_ids=request.stop_token_ids,
             ignore_eos=request.ignore_eos,
             max_tokens=request.max_tokens,
             logprobs=request.logprobs,
             use_beam_search=request.use_beam_search,
+            skip_special_tokens=request.skip_special_tokens,
         )
     except ValueError as e:
         return create_error_response(HTTPStatus.BAD_REQUEST, str(e))
 
-    result_generator = engine.generate(prompt, sampling_params,
-                                       request_id)
+    if use_token_ids:
+        result_generator = engine.generate(None,
+                                           sampling_params,
+                                           request_id,
+                                           prompt_token_ids=prompt)
+    else:
+        result_generator = engine.generate(prompt, sampling_params, request_id,
+                                           token_ids)
 
     # Similar to the OpenAI API, when n != best_of, we do not stream the
     # results. In addition, we do not stream the results when use beam search.
-    stream = (request.stream and
-              (request.best_of is None or request.n == request.best_of) and
-              not request.use_beam_search)
+    stream = (request.stream
+              and (request.best_of is None or request.n == request.best_of)
+              and not request.use_beam_search)
 
-    async def abort_request() -> None:
-        await engine.abort(request_id)
-
-    def create_stream_response_json(index: int,
-                                    text: str,
-                                    logprobs: Optional[LogProbs] = None,
-                                    finish_reason: Optional[str] = None) -> str:
+    def create_stream_response_json(
+        index: int,
+        text: str,
+        logprobs: Optional[LogProbs] = None,
+        finish_reason: Optional[str] = None,
+    ) -> str:
         choice_data = CompletionResponseStreamChoice(
             index=index,
             text=text,
@@ -200,7 +493,8 @@ async def create_completion(raw_request: Request):
                 )
                 yield f"data: {response_json}\n\n"
                 if output.finish_reason is not None:
-                    logprobs = LogProbs() if request.logprobs is not None else None
+                    logprobs = (LogProbs()
+                                if request.logprobs is not None else None)
                     response_json = create_stream_response_json(
                         index=i,
                         text="",
@@ -208,23 +502,19 @@ async def create_completion(raw_request: Request):
                         finish_reason=output.finish_reason,
                     )
                     yield f"data: {response_json}\n\n"
-            yield "data: [DONE]\n\n"
+        yield "data: [DONE]\n\n"
 
     # Streaming response
     if stream:
-        background_tasks = BackgroundTasks()
-        # Abort the request if the client disconnects.
-        background_tasks.add_task(abort_request)
         return StreamingResponse(completion_stream_generator(),
-                                 media_type="text/event-stream",
-                                 background=background_tasks)
+                                 media_type="text/event-stream")
 
     # Non-streaming response
     final_res: RequestOutput = None
     async for res in result_generator:
         if await raw_request.is_disconnected():
             # Abort the request if the client disconnects.
-            await abort_request()
+            await engine.abort(request_id)
             return create_error_response(HTTPStatus.BAD_REQUEST,
                                          "Client disconnected")
         final_res = res
@@ -244,8 +534,8 @@ async def create_completion(raw_request: Request):
         choices.append(choice_data)
 
     num_prompt_tokens = len(final_res.prompt_token_ids)
-    num_generated_tokens = sum(len(output.token_ids)
-                               for output in final_res.outputs)
+    num_generated_tokens = sum(
+        len(output.token_ids) for output in final_res.outputs)
     usage = UsageInfo(
         prompt_tokens=num_prompt_tokens,
         completion_tokens=num_generated_tokens,
@@ -263,9 +553,11 @@ async def create_completion(raw_request: Request):
         # When user requests streaming but we don't stream, we still need to
         # return a streaming response with a single event.
         response_json = response.json(ensure_ascii=False)
+
         async def fake_stream_generator() -> AsyncGenerator[str, None]:
             yield f"data: {response_json}\n\n"
             yield "data: [DONE]\n\n"
+
         return StreamingResponse(fake_stream_generator(),
                                  media_type="text/event-stream")
 
@@ -274,26 +566,31 @@ async def create_completion(raw_request: Request):
 
 if __name__ == "__main__":
     parser = argparse.ArgumentParser(
-        description="vLLM OpenAI-Compatible RESTful API server."
-    )
-    parser.add_argument("--host", type=str, default="localhost", help="host name")
+        description="vLLM OpenAI-Compatible RESTful API server.")
+    parser.add_argument("--host", type=str, default=None, help="host name")
     parser.add_argument("--port", type=int, default=8000, help="port number")
-    parser.add_argument(
-        "--allow-credentials", action="store_true", help="allow credentials"
-    )
-    parser.add_argument(
-        "--allowed-origins", type=json.loads, default=["*"], help="allowed origins"
-    )
-    parser.add_argument(
-        "--allowed-methods", type=json.loads, default=["*"], help="allowed methods"
-    )
-    parser.add_argument(
-        "--allowed-headers", type=json.loads, default=["*"], help="allowed headers"
-    )
-    parser.add_argument("--served-model-name", type=str, default=None,
-                        help="The model name used in the API. If not specified, "
-                             "the model name will be the same as the "
-                             "huggingface name.")
+    parser.add_argument("--allow-credentials",
+                        action="store_true",
+                        help="allow credentials")
+    parser.add_argument("--allowed-origins",
+                        type=json.loads,
+                        default=["*"],
+                        help="allowed origins")
+    parser.add_argument("--allowed-methods",
+                        type=json.loads,
+                        default=["*"],
+                        help="allowed methods")
+    parser.add_argument("--allowed-headers",
+                        type=json.loads,
+                        default=["*"],
+                        help="allowed headers")
+    parser.add_argument("--served-model-name",
+                        type=str,
+                        default=None,
+                        help="The model name used in the API. If not "
+                        "specified, the model name will be the same as "
+                        "the huggingface name.")
+
     parser = AsyncEngineArgs.add_cli_args(parser)
     args = parser.parse_args()
 
@@ -307,13 +604,23 @@ if __name__ == "__main__":
 
     logger.info(f"args: {args}")
 
-    served_model = args.served_model_name or args.model
+    if args.served_model_name is not None:
+        served_model = args.served_model_name
+    else:
+        served_model = args.model
 
     engine_args = AsyncEngineArgs.from_cli_args(args)
     engine = AsyncLLMEngine.from_engine_args(engine_args)
+    engine_model_config = asyncio.run(engine.get_model_config())
+    max_model_len = engine_model_config.max_model_len
 
     # A separate tokenizer to map token IDs to strings.
-    tokenizer = get_tokenizer(args.model)
+    tokenizer = get_tokenizer(engine_args.tokenizer,
+                              tokenizer_mode=engine_args.tokenizer_mode,
+                              trust_remote_code=engine_args.trust_remote_code)
 
-    uvicorn.run(app, host=args.host, port=args.port, log_level="info",
+    uvicorn.run(app,
+                host=args.host,
+                port=args.port,
+                log_level="info",
                 timeout_keep_alive=TIMEOUT_KEEP_ALIVE)

vllm/entrypoints/openai/protocol.py

@@ -1,4 +1,5 @@
-# Adapted from https://github.com/lm-sys/FastChat/blob/168ccc29d3f7edc50823016105c024fe2282732a/fastchat/protocol/openai_api_protocol.py
+# Adapted from
+# https://github.com/lm-sys/FastChat/blob/168ccc29d3f7edc50823016105c024fe2282732a/fastchat/protocol/openai_api_protocol.py
 import time
 from typing import Dict, List, Literal, Optional, Union
 
@@ -53,21 +54,30 @@ class UsageInfo(BaseModel):
 
 class ChatCompletionRequest(BaseModel):
     model: str
-    messages: List[Dict[str, str]]
+    messages: Union[str, List[Dict[str, str]]]
     temperature: Optional[float] = 0.7
     top_p: Optional[float] = 1.0
     n: Optional[int] = 1
     max_tokens: Optional[int] = None
-    stop: Optional[Union[str, List[str]]] = None
+    stop: Optional[Union[str, List[str]]] = Field(default_factory=list)
     stream: Optional[bool] = False
     presence_penalty: Optional[float] = 0.0
     frequency_penalty: Optional[float] = 0.0
+    logit_bias: Optional[Dict[str, float]] = None
     user: Optional[str] = None
+    # Additional parameters supported by vLLM
+    best_of: Optional[int] = None
+    top_k: Optional[int] = -1
+    ignore_eos: Optional[bool] = False
+    use_beam_search: Optional[bool] = False
+    stop_token_ids: Optional[List[int]] = Field(default_factory=list)
+    skip_special_tokens: Optional[bool] = True
 
 
 class CompletionRequest(BaseModel):
     model: str
-    prompt: str
+    # a string, array of strings, array of tokens, or array of token arrays
+    prompt: Union[List[int], List[List[int]], str, List[str]]
     suffix: Optional[str] = None
     max_tokens: Optional[int] = 16
     temperature: Optional[float] = 1.0
@@ -86,13 +96,16 @@ class CompletionRequest(BaseModel):
     top_k: Optional[int] = -1
     ignore_eos: Optional[bool] = False
     use_beam_search: Optional[bool] = False
+    stop_token_ids: Optional[List[int]] = Field(default_factory=list)
+    skip_special_tokens: Optional[bool] = True
 
 
 class LogProbs(BaseModel):
     text_offset: List[int] = Field(default_factory=list)
     token_logprobs: List[Optional[float]] = Field(default_factory=list)
     tokens: List[str] = Field(default_factory=list)
-    top_logprobs: List[Optional[Dict[str, float]]] = Field(default_factory=list)
+    top_logprobs: List[Optional[Dict[str,
+                                     float]]] = Field(default_factory=list)
 
 
 class CompletionResponseChoice(BaseModel):
@@ -124,3 +137,42 @@ class CompletionStreamResponse(BaseModel):
     created: int = Field(default_factory=lambda: int(time.time()))
     model: str
     choices: List[CompletionResponseStreamChoice]
+
+
+class ChatMessage(BaseModel):
+    role: str
+    content: str
+
+
+class ChatCompletionResponseChoice(BaseModel):
+    index: int
+    message: ChatMessage
+    finish_reason: Optional[Literal["stop", "length"]] = None
+
+
+class ChatCompletionResponse(BaseModel):
+    id: str = Field(default_factory=lambda: f"chatcmpl-{random_uuid()}")
+    object: str = "chat.completion"
+    created: int = Field(default_factory=lambda: int(time.time()))
+    model: str
+    choices: List[ChatCompletionResponseChoice]
+    usage: UsageInfo
+
+
+class DeltaMessage(BaseModel):
+    role: Optional[str] = None
+    content: Optional[str] = None
+
+
+class ChatCompletionResponseStreamChoice(BaseModel):
+    index: int
+    delta: DeltaMessage
+    finish_reason: Optional[Literal["stop", "length"]] = None
+
+
+class ChatCompletionStreamResponse(BaseModel):
+    id: str = Field(default_factory=lambda: f"chatcmpl-{random_uuid()}")
+    object: str = "chat.completion.chunk"
+    created: int = Field(default_factory=lambda: int(time.time()))
+    model: str
+    choices: List[ChatCompletionResponseStreamChoice]

vllm/logger.py

@@ -1,9 +1,9 @@
-# Adapted from https://github.com/skypilot-org/skypilot/blob/86dc0f6283a335e4aa37b3c10716f90999f48ab6/sky/sky_logging.py
-
+# Adapted from
+# https://github.com/skypilot-org/skypilot/blob/86dc0f6283a335e4aa37b3c10716f90999f48ab6/sky/sky_logging.py
+"""Logging configuration for vLLM."""
 import logging
 import sys
 
-
 _FORMAT = "%(levelname)s %(asctime)s %(filename)s:%(lineno)d] %(message)s"
 _DATE_FORMAT = "%m-%d %H:%M:%S"
 

vllm/model_executor/__init__.py

@@ -2,7 +2,6 @@ from vllm.model_executor.input_metadata import InputMetadata
 from vllm.model_executor.model_loader import get_model
 from vllm.model_executor.utils import set_random_seed
 
-
 __all__ = [
     "InputMetadata",
     "get_model",

vllm/model_executor/input_metadata.py

@@ -1,23 +1,35 @@
-from typing import Dict, List, Tuple
+from typing import Dict, List, Optional, Tuple
 
 import torch
-from xformers.ops.fmha.attn_bias import BlockDiagonalCausalMask
+from xformers.ops import AttentionBias
 
 from vllm.sampling_params import SamplingParams
 from vllm.sequence import SequenceData
 
 
 class InputMetadata:
+    """Metadata for input sequences. Used for PagedAttention.
+
+    Args:
+        seq_groups: List of (seq_ids, sampling_params).
+        seq_data: Seq_id -> SequenceData.
+        prompt_lens: Lengths of prompts.
+        slot_mapping: The address to write the new KV to of each token.
+        context_lens: the length of attention context for each generation token.
+        max_context_len: The maximum context length.
+        block_tables: The block tables. (Seq id -> list of physical block)
+    """
 
     def __init__(
         self,
-        seq_groups: List[Tuple[List[int], SamplingParams]],     # List of (seq_ids, sampling_params).
-        seq_data: Dict[int, SequenceData],                      # Seq_id -> SequenceData.
+        seq_groups: List[Tuple[List[int], SamplingParams]],
+        seq_data: Dict[int, SequenceData],
         prompt_lens: List[int],
         slot_mapping: torch.Tensor,
         context_lens: torch.Tensor,
         max_context_len: int,
         block_tables: torch.Tensor,
+        sliding_window: Optional[int] = None,
     ) -> None:
         self.seq_groups = seq_groups
         self.seq_data = seq_data
@@ -27,7 +39,24 @@ class InputMetadata:
         self.max_context_len = max_context_len
         self.block_tables = block_tables
 
-        self.attn_bias = BlockDiagonalCausalMask.from_seqlens(prompt_lens)
+        self.to_cache = None
+        if sliding_window is not None:
+            # We need to keep the positions of sliding windows within
+            # the key / value tables, this is helpful to know which
+            # elements we need to cache and where
+            to_cache, start_idx = [], 0
+            for prompt_len in self.prompt_lens:
+                to_cache.extend(
+                    range(
+                        start_idx + max(0, prompt_len - sliding_window),
+                        start_idx + prompt_len,
+                    ))
+                start_idx += prompt_len
+            to_cache.extend(range(start_idx, slot_mapping.shape[0]))
+            self.to_cache = torch.tensor(to_cache,
+                                         dtype=torch.int32,
+                                         device=self.slot_mapping.device)
+
         self.num_prompts = len(prompt_lens)
         self.num_prompt_tokens = sum(prompt_lens)
         self.num_generation_tokens = context_lens.shape[0]
@@ -39,6 +68,9 @@ class InputMetadata:
         assert block_tables.shape[0] == self.num_generation_tokens
         assert context_lens.shape[0] == self.num_generation_tokens
 
+        # Set during the execution of the first attention op.
+        self.attn_bias: List[AttentionBias] = []
+
     def __repr__(self) -> str:
         # Print only useful metadata.
         return (f'InputMetadata('

vllm/model_executor/layers/activation.py

@@ -4,10 +4,50 @@ import torch.nn as nn
 
 from vllm import activation_ops
 
+
+class SiluAndMul(nn.Module):
+    """An activation function for SwiGLU.
+
+    The function computes x -> silu(x[:d]) * x[d:] where d = x.shape[1] // 2.
+
+    Shapes:
+        x: (num_tokens, 2 * d)
+        return: (num_tokens, d)
+    """
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        num_tokens = x.shape[0]
+        d = x.shape[1] // 2
+        out = torch.empty(num_tokens, d, dtype=x.dtype, device=x.device)
+        activation_ops.silu_and_mul(out, x)
+        return out
+
+
+class NewGELU(nn.Module):
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        num_tokens = x.shape[0]
+        d = x.shape[1]
+        out = torch.empty(num_tokens, d, dtype=x.dtype, device=x.device)
+        activation_ops.gelu_new(out, x)
+        return out
+
+
+class FastGELU(nn.Module):
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        num_tokens = x.shape[0]
+        d = x.shape[1]
+        out = torch.empty(num_tokens, d, dtype=x.dtype, device=x.device)
+        activation_ops.gelu_fast(out, x)
+        return out
+
+
 _ACTIVATION_REGISTRY = {
     "gelu": nn.GELU(),
-    "gelu_new": nn.GELU(approximate="tanh"),   # NOTE: This may introduce small rounding errors.
-    "gelu_fast": nn.GELU(approximate="tanh"),  # NOTE: This may introduce small rounding errors.
+    "gelu_fast": FastGELU(),
+    "gelu_new": NewGELU(),
+    "gelu_pytorch_tanh": nn.GELU(approximate="tanh"),
     "relu": nn.ReLU(),
 }
 
@@ -18,23 +58,3 @@ def get_act_fn(act_fn: str) -> nn.Module:
     if act_fn in _ACTIVATION_REGISTRY:
         return _ACTIVATION_REGISTRY[act_fn]
     raise ValueError(f"Activation function {act_fn!r} is not supported.")
-
-
-class SiluAndMul(nn.Module):
-    """An activation function for SwiGLU.
-
-    The function computes x -> silu(x[:d]) * x[d:] where d = x.shape[1] // 2.
-    """
-
-    def __init__(self):
-        super().__init__()
-
-    def forward(
-        self,
-        x: torch.Tensor,        # (num_tokens, 2 * d)
-    ) -> torch.Tensor:          # (num_tokens, d)
-        num_tokens = x.shape[0]
-        d = x.shape[1] // 2
-        out = torch.empty(num_tokens, d, dtype=x.dtype, device=x.device)
-        activation_ops.silu_and_mul(out, x)
-        return out

vllm/model_executor/layers/attention.py

@@ -1,28 +1,43 @@
 """Multi-head attention."""
-from typing import Optional
+from typing import Any, Dict, List, Optional
 
 import torch
 import torch.nn as nn
 from xformers import ops as xops
+from xformers.ops.fmha.attn_bias import (BlockDiagonalCausalMask,
+                                         LowerTriangularMaskWithTensorBias)
 
 from vllm import attention_ops
 from vllm import cache_ops
-from vllm import pos_encoding_ops
 from vllm.model_executor.input_metadata import InputMetadata
+from vllm.model_executor.layers.rotary_embedding import (
+    DynamicNTKScalingRotaryEmbedding, LinearScalingRotaryEmbedding,
+    RotaryEmbedding)
 
-_SUPPORTED_HEAD_SIZES = [64, 80, 96, 128]
+_SUPPORTED_HEAD_SIZES = [64, 80, 96, 112, 128, 256]
+# Should be the same as PARTITION_SIZE in `paged_attention_v2_launcher`.
+_PARTITION_SIZE = 512
 
 
 class PagedAttention(nn.Module):
+    # pylint: disable=line-too-long
     """GPT-style multi-head PagedAttention.
 
     This class takes flattened 1D query, key, and value tensors as input. The
-    input 1D tensors can be split into three parts: the prompt tokens, the
-    generation tokens, and the paddings.
+    input 1D tensors can either contain prompt tokens or generation tokens, in
+    addition to paddings.
 
-    |<------------------------------------- num_valid_tokens ------------------------------------->|
-    |<--------------- num_prompt_tokens -------------->|<------- num_generation_tokens (M) ------->|
-    |<--prompt_0-->|<--prompt_1-->|...|<--prompt_N-1-->|<--generation_0-->|...|<--generation_M-1-->|<--padding-->|
+    If the input tensors contain prompt tokens, the layout is as follows:
+
+    |<---------------------- num_valid_tokens ---------------------->|
+    |<--------------- num_prompt_tokens -------------->|
+    |<--prompt_0-->|<--prompt_1-->|...|<--prompt_N-1-->|<--padding-->|
+
+    Otherwise, the layout is as follows:
+
+    |<------------------ num_valid_tokens ------------------->|
+    |<------- num_generation_tokens (M) ------->|
+    |<--generation_0-->|...|<--generation_M-1-->|<--padding-->|
 
     The prompts might have different lengths, while the generation tokens always
     have length 1. The paddings are appended to make the input length a multiple
@@ -41,77 +56,203 @@ class PagedAttention(nn.Module):
     5. Output a flattened 1D tensor.
     """
 
-    def __init__(self, num_heads: int, head_size: int, scale: float) -> None:
+    def __init__(self,
+                 num_heads: int,
+                 head_size: int,
+                 scale: float,
+                 num_kv_heads: Optional[int] = None,
+                 sliding_window: Optional[int] = None) -> None:
         super().__init__()
         self.num_heads = num_heads
         self.head_size = head_size
         self.scale = float(scale)
-        self.attn_op = xops.fmha.cutlass.FwOp()
+        self.num_kv_heads = num_heads if num_kv_heads is None else num_kv_heads
+        self.sliding_window = sliding_window
+
+        assert self.num_heads % self.num_kv_heads == 0
+        self.num_queries_per_kv = self.num_heads // self.num_kv_heads
+        self.head_mapping = torch.repeat_interleave(
+            torch.arange(self.num_kv_heads, dtype=torch.int32, device="cuda"),
+            self.num_queries_per_kv)
 
         if self.head_size not in _SUPPORTED_HEAD_SIZES:
             raise ValueError(f"head_size ({self.head_size}) is not supported. "
                              f"Supported head sizes: {_SUPPORTED_HEAD_SIZES}.")
 
+    def set_attn_bias(
+        self,
+        input_metadata: InputMetadata,
+        dtype: torch.dtype,
+    ) -> None:
+        del dtype  # Unused.
+        if input_metadata.attn_bias:
+            # Already set by a previous layer.
+            return
+        prompt_lens = input_metadata.prompt_lens
+        attn_bias = BlockDiagonalCausalMask.from_seqlens(prompt_lens)
+        if self.sliding_window is not None:
+            attn_bias = attn_bias.make_local_attention(self.sliding_window)
+        input_metadata.attn_bias.append(attn_bias)
+
     def multi_query_kv_attention(
         self,
-        output: torch.Tensor,                   # [num_prompt_tokens, num_heads, head_size]
-        query: torch.Tensor,                    # [num_prompt_tokens, num_heads, head_size]
-        key: torch.Tensor,                      # [num_prompt_tokens, num_heads, head_size]
-        value: torch.Tensor,                    # [num_prompt_tokens, num_heads, head_size]
-        attn_bias: xops.AttentionBias,
+        output: torch.Tensor,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        input_metadata: InputMetadata,
     ) -> torch.Tensor:
+        """Normal attention for the prompt tokens.
+
+        Args:
+            output: shape = [num_prompt_tokens, num_heads, head_size]
+            query: shape = [num_prompt_tokens, num_heads, head_size]
+            key: shape = [num_prompt_tokens, num_kv_heads, head_size]
+            value: shape = [num_prompt_tokens, num_kv_heads, head_size]
+            input_metadata: metadata for paged attention.
+        """
+
+        if self.num_kv_heads != self.num_heads:
+            # Project the key and value tensors to the desired number of heads.
+            key = torch.repeat_interleave(key, self.num_queries_per_kv, dim=1)
+            value = torch.repeat_interleave(value,
+                                            self.num_queries_per_kv,
+                                            dim=1)
+
         # TODO(woosuk): The unsqueeze op may incur some CPU overhead. Optimize.
         out = xops.memory_efficient_attention_forward(
             query.unsqueeze(0),
             key.unsqueeze(0),
             value.unsqueeze(0),
-            attn_bias=attn_bias,
+            attn_bias=input_metadata.attn_bias[0],
             p=0.0,
             scale=self.scale,
-            op=self.attn_op,
         )
         # TODO(woosuk): Unnecessary copy. Optimize.
         output.copy_(out.squeeze(0))
         return output
 
+    def get_alibi_slopes(self) -> Optional[torch.Tensor]:
+        """Returns the slopes for the alibi attention bias.
+
+        Returns:
+            slopes: shape = [num_heads]
+        """
+        return None
+
     def single_query_cached_kv_attention(
         self,
-        output: torch.Tensor,           # [num_generation_tokens, num_heads, head_size]
-        query: torch.Tensor,            # [num_generation_tokens, num_heads, head_size]
-        key_cache: torch.Tensor,        # [num_blocks, num_heads, head_size/x, block_size, x]
-        value_cache: torch.Tensor,      # [num_blocks, num_heads, head_size, block_size]
+        output: torch.Tensor,
+        query: torch.Tensor,
+        key_cache: torch.Tensor,
+        value_cache: torch.Tensor,
         input_metadata: InputMetadata,
+        alibi_slopes: Optional[torch.Tensor],
     ) -> None:
+        """PagedAttention for the generation tokens.
+
+        Args:
+            output: shape = [num_generation_tokens, num_heads, head_size]
+            query: shape = [num_generation_tokens, num_heads, head_size]
+            key_cache: shape = [num_blocks, num_kv_heads, head_size/x,
+                block_size, x]
+            value_cache: shape = [num_blocks, num_kv_heads, head_size,
+                block_size]
+            input_metadata: metadata for paged attention.
+            alibi_slopes: shape = [num_heads]
+        """
         block_size = value_cache.shape[3]
-        attention_ops.single_query_cached_kv_attention(
-            output,
-            query,
-            key_cache,
-            value_cache,
-            self.scale,
-            input_metadata.block_tables,
-            input_metadata.context_lens,
-            block_size,
-            input_metadata.max_context_len,
-        )
+        num_seqs, num_heads, head_size = query.shape
+        max_num_partitions = (
+            (input_metadata.max_context_len + _PARTITION_SIZE - 1) //
+            _PARTITION_SIZE)
+        # NOTE(woosuk): We use a simple heuristic to decide whether to use
+        # PagedAttention V1 or V2. If the number of partitions is 1, we use
+        # V1 to avoid the overhead of reduction. Also, if the number of
+        # sequences or heads is large, we use V1 since there is enough work
+        # to parallelize.
+        # TODO(woosuk): Tune this heuristic.
+        use_v1 = max_num_partitions == 1 or num_seqs * num_heads > 512
+        if use_v1:
+            # Run PagedAttention V1.
+            attention_ops.paged_attention_v1(
+                output,
+                query,
+                key_cache,
+                value_cache,
+                self.head_mapping,
+                self.scale,
+                input_metadata.block_tables,
+                input_metadata.context_lens,
+                block_size,
+                input_metadata.max_context_len,
+                alibi_slopes,
+            )
+        else:
+            # Run PagedAttention V2.
+            assert _PARTITION_SIZE % block_size == 0
+            tmp_output = torch.empty(
+                size=(num_seqs, num_heads, max_num_partitions, head_size),
+                dtype=output.dtype,
+                device=output.device,
+            )
+            exp_sums = torch.empty(
+                size=(num_seqs, num_heads, max_num_partitions),
+                dtype=torch.float32,
+                device=output.device,
+            )
+            max_logits = torch.empty_like(exp_sums)
+            attention_ops.paged_attention_v2(
+                output,
+                exp_sums,
+                max_logits,
+                tmp_output,
+                query,
+                key_cache,
+                value_cache,
+                self.head_mapping,
+                self.scale,
+                input_metadata.block_tables,
+                input_metadata.context_lens,
+                block_size,
+                input_metadata.max_context_len,
+                alibi_slopes,
+            )
 
     def forward(
         self,
-        query: torch.Tensor,                    # [num_tokens, num_heads * head_size]
-        key: torch.Tensor,                      # [num_tokens, num_heads * head_size]
-        value: torch.Tensor,                    # [num_tokens, num_heads * head_size]
-        key_cache: Optional[torch.Tensor],      # [num_blocks, num_heads, head_size/x, block_size, x]
-        value_cache: Optional[torch.Tensor],    # [num_blocks, num_heads, head_size, block_size]
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        key_cache: Optional[torch.Tensor],
+        value_cache: Optional[torch.Tensor],
         input_metadata: InputMetadata,
         cache_event: Optional[torch.cuda.Event],
-    ) -> torch.Tensor:                          # [num_tokens, num_heads * head_size]
-        # NOTE: The query, key, and value tensors must be sliced from a qkv
-        # tensor of shape [num_tokens, 3 * num_heads * head_size].
+    ) -> torch.Tensor:
+        """PagedAttention forward pass.
+
+        NOTE: The query, key, and value tensors must be sliced from a qkv
+        tensor of shape [num_tokens, 3 * num_heads * head_size].
+
+        Args:
+            query: shape = [num_tokens, num_heads * head_size]
+            key: shape = [num_tokens, num_kv_heads * head_size]
+            value: shape = [num_tokens, num_kv_heads * head_size]
+            key_cache: shape = [num_blocks, num_kv_heads, head_size/x,
+                block_size, x]
+            value_cache: shape = [num_blocks, num_kv_heads, head_size,
+                block_size]
+            input_metadata: metadata for paged attention.
+            cache_event: event to wait for the cache operations to finish.
+
+        Returns:
+            shape = [num_tokens, num_heads * head_size]
+        """
 
         # Reshape the query, key, and value tensors.
         query = query.view(-1, self.num_heads, self.head_size)
-        key = key.view(-1, self.num_heads, self.head_size)
-        value = value.view(-1, self.num_heads, self.head_size)
+        key = key.view(-1, self.num_kv_heads, self.head_size)
+        value = value.view(-1, self.num_kv_heads, self.head_size)
 
         # Pre-allocate the output tensor.
         output = torch.empty_like(query)
@@ -119,12 +260,15 @@ class PagedAttention(nn.Module):
         # Compute the attention op for prompts.
         num_prompt_tokens = input_metadata.num_prompt_tokens
         if num_prompt_tokens > 0:
+            # Prompt run.
+            assert input_metadata.num_generation_tokens == 0
+            self.set_attn_bias(input_metadata, dtype=query.dtype)
             self.multi_query_kv_attention(
                 output[:num_prompt_tokens],
                 query[:num_prompt_tokens],
                 key[:num_prompt_tokens],
                 value[:num_prompt_tokens],
-                input_metadata.attn_bias,
+                input_metadata,
             )
 
         # Wait until the cache op is done.
@@ -136,28 +280,35 @@ class PagedAttention(nn.Module):
         # and value vectors will not be cached.
         num_valid_tokens = input_metadata.num_valid_tokens
         if (num_valid_tokens > 0 and key_cache is not None
-            and value_cache is not None):
+                and value_cache is not None):
             # The stride is 3 because the key and value are sliced from qkv.
+            key_to_cache = key[:num_valid_tokens]
+            value_to_cache = value[:num_valid_tokens]
+            slot_mapping = input_metadata.slot_mapping
+            if input_metadata.to_cache is not None:
+                key_to_cache = key_to_cache[input_metadata.to_cache]
+                value_to_cache = value_to_cache[input_metadata.to_cache]
+                slot_mapping = slot_mapping[input_metadata.to_cache]
+
             cache_ops.reshape_and_cache(
-                key[:num_valid_tokens],
-                value[:num_valid_tokens],
+                key_to_cache,
+                value_to_cache,
                 key_cache,
                 value_cache,
-                input_metadata.slot_mapping,
+                slot_mapping,
             )
 
         if input_metadata.num_generation_tokens > 0:
+            # Decoding run.
+            assert input_metadata.num_prompt_tokens == 0
             assert key_cache is not None and value_cache is not None, (
                 "key_cache and value_cache must be provided when "
-                "generating tokens."
-            )
+                "generating tokens.")
             # Compute the attention op for generation tokens.
             self.single_query_cached_kv_attention(
                 output[num_prompt_tokens:num_valid_tokens],
-                query[num_prompt_tokens:num_valid_tokens],
-                key_cache,
-                value_cache,
-                input_metadata)
+                query[num_prompt_tokens:num_valid_tokens], key_cache,
+                value_cache, input_metadata, self.get_alibi_slopes())
 
         # Reshape the output tensor.
         # NOTE(woosuk): The output tensor may include paddings.
@@ -165,7 +316,7 @@ class PagedAttention(nn.Module):
 
 
 class PagedAttentionWithRoPE(PagedAttention):
-    """PagedAttention with GPT-NeoX style rotary embedding."""
+    """PagedAttention with rotary positional embedding."""
 
     def __init__(
         self,
@@ -175,44 +326,66 @@ class PagedAttentionWithRoPE(PagedAttention):
         rotary_dim: int,
         max_position: int = 8192,
         base: int = 10000,
+        num_kv_heads: Optional[int] = None,
+        is_neox_style: bool = True,
+        rope_scaling: Optional[Dict[str, Any]] = None,
+        sliding_window: Optional[int] = None,
     ) -> None:
-        super().__init__(num_heads, head_size, scale)
-
-        # Create the cos and sin cache.
-        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()
-        cache = torch.cat((cos, sin), dim=-1)
-
-        # FIXME(woosuk): This assumes that we configure the default dtype when
-        # initializing the model. Make it more robust.
-        torch_dtype = torch.get_default_dtype()
-        cache = cache.to(torch_dtype)
-        # Embedding size: [max_position, rotary_dim]
-        self.register_buffer("cos_sin_cache", cache, persistent=False)
+        super().__init__(num_heads,
+                         head_size,
+                         scale,
+                         num_kv_heads,
+                         sliding_window=sliding_window)
+        if rope_scaling is None:
+            self.rotary_emb = RotaryEmbedding(head_size, rotary_dim,
+                                              max_position, base,
+                                              is_neox_style)
+        else:
+            scaling_type = rope_scaling["type"]
+            scaling_factor = rope_scaling["factor"]
+            if scaling_type == "linear":
+                self.rotary_emb = LinearScalingRotaryEmbedding(
+                    head_size, rotary_dim, max_position, base, is_neox_style,
+                    scaling_factor)
+            elif scaling_type == "dynamic":
+                self.rotary_emb = DynamicNTKScalingRotaryEmbedding(
+                    head_size, rotary_dim, max_position, base, is_neox_style,
+                    scaling_factor)
+            else:
+                raise ValueError(f"Unknown RoPE scaling type {scaling_type}")
 
     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]
-        value: torch.Tensor,                    # [num_tokens, num_heads * head_size]
-        key_cache: torch.Tensor,                # [num_blocks, num_heads, head_size/x, block_size, x]
-        value_cache: torch.Tensor,              # [num_blocks, num_heads, head_size, block_size]
+        positions: torch.Tensor,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        key_cache: torch.Tensor,
+        value_cache: torch.Tensor,
         input_metadata: InputMetadata,
         cache_event: Optional[torch.cuda.Event],
-    ) -> torch.Tensor:                          # [num_tokens, num_heads * head_size]
+    ) -> torch.Tensor:
+        """ PagedAttention forward pass with rotary embedding.
+
+        Args:
+            positions: shape = [num_tokens]
+            query: shape = [num_tokens, num_heads * head_size]
+            key: shape = [num_tokens, num_kv_heads * head_size]
+            value: shape = [num_tokens, num_kv_heads * head_size]
+            key_cache: shape = [num_blocks, num_kv_heads, head_size/x,
+                block_size, x]
+            value_cache: shape = [num_blocks, num_kv_heads, head_size,
+                block_size]
+            input_metadata: metadata for paged attention.
+            cache_event: event to wait for the cache operations to finish.
+
+        Returns:
+            shape = [num_tokens, num_heads * head_size]
+        """
+
         # Apply rotary embedding to the query and key before passing them
         # to the attention op.
-        pos_encoding_ops.rotary_embedding_neox(
-            positions,
-            query,
-            key,
-            self.head_size,
-            self.cos_sin_cache,
-        )
+        query, key = self.rotary_emb(positions, query, key)
         return super().forward(
             query,
             key,
@@ -222,3 +395,96 @@ class PagedAttentionWithRoPE(PagedAttention):
             input_metadata,
             cache_event,
         )
+
+
+class PagedAttentionWithALiBi(PagedAttention):
+    """PagedAttention with ALiBi attention bias."""
+
+    def __init__(self,
+                 num_heads: int,
+                 head_size: int,
+                 scale: float,
+                 slopes: List[float],
+                 num_kv_heads: Optional[int] = None) -> None:
+        super().__init__(num_heads, head_size, scale, num_kv_heads)
+        assert len(slopes) == num_heads
+
+        slopes = torch.tensor(slopes, dtype=torch.float32)
+        self.register_buffer("alibi_slopes", slopes, persistent=False)
+
+    def set_attn_bias(self, input_metadata: InputMetadata,
+                      dtype: torch.dtype) -> None:
+        if input_metadata.attn_bias:
+            # Already set by a previous layer.
+            return
+        # Generates ALiBi mask for each prompt.
+        for prompt_len in input_metadata.prompt_lens:
+            bias = torch.arange(prompt_len, dtype=dtype)
+            # NOTE(zhuohan): HF uses
+            #     `bias = bias[None, :].repeat(prompt_len, 1)`
+            # here. We find that both biases give the same results, but
+            # the bias below more accurately follows the original ALiBi
+            # paper.
+            bias = bias[None, :] - bias[:, None]
+            bias = bias.to(self.alibi_slopes.device)
+
+            # When using custom attention bias, xformers requires the bias to
+            # be sliced from a tensor whose length is a multiple of 8.
+            padded_len = (prompt_len + 7) // 8 * 8
+            bias = torch.empty(
+                1,  # batch_size
+                self.num_heads,
+                prompt_len,
+                padded_len,
+                device=self.alibi_slopes.device,
+                dtype=dtype,
+            )[:, :, :, :prompt_len].copy_(bias)
+            bias.mul_(self.alibi_slopes[:, None, None])
+            attn_bias = LowerTriangularMaskWithTensorBias(bias)
+            input_metadata.attn_bias.append(attn_bias)
+
+    def multi_query_kv_attention(
+        self,
+        output: torch.Tensor,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        """Attention with ALiBi bias for the prompt tokens.
+
+        Args:
+            output: shape = [num_prompt_tokens, num_heads, head_size]
+            query: shape = [num_prompt_tokens, num_heads, head_size]
+            key: shape = [num_prompt_tokens, num_kv_heads, head_size]
+            value: shape = [num_prompt_tokens, num_kv_heads, head_size]
+            input_metadata: metadata for paged attention.
+        """
+        if self.num_kv_heads != self.num_heads:
+            # Project the key and value tensors to the desired number of heads.
+            key = torch.repeat_interleave(key, self.num_queries_per_kv, dim=1)
+            value = torch.repeat_interleave(value,
+                                            self.num_queries_per_kv,
+                                            dim=1)
+
+        # FIXME(woosuk): Because xformers does not support dynamic sequence
+        # lengths with custom attention bias, we process each prompt one by
+        # one. This is inefficient, especially when we have many short prompts.
+        start = 0
+        for i, prompt_len in enumerate(input_metadata.prompt_lens):
+            end = start + prompt_len
+            out = xops.memory_efficient_attention_forward(
+                query[None, start:end],
+                key[None, start:end],
+                value[None, start:end],
+                attn_bias=input_metadata.attn_bias[i],
+                p=0.0,
+                scale=self.scale,
+            )
+            # TODO(woosuk): Unnecessary copy. Optimize.
+            output[start:end].copy_(out.squeeze(0))
+            start += prompt_len
+        return output
+
+    def get_alibi_slopes(self) -> Optional[torch.Tensor]:
+        return self.alibi_slopes

vllm/model_executor/layers/quantized_linear/__init__.py

@@ -0,0 +1,37 @@
+from vllm.model_executor.layers.quantized_linear.awq import (
+    AWQColumnParallelLinear, AWQRowParallelLinear)
+from vllm.model_executor.parallel_utils.layers import (ColumnParallelLinear,
+                                                       RowParallelLinear)
+
+_QUANTIZED_LINEAR_REGISTRY = {
+    "awq": (AWQColumnParallelLinear, AWQRowParallelLinear),
+}
+
+
+class ParallelLinear:
+
+    @classmethod
+    def column(cls, *args, **kwargs) -> ColumnParallelLinear:
+        quant_config = kwargs.get("quant_config", None)
+        if quant_config is None:
+            return ColumnParallelLinear(*args, **kwargs)
+
+        name = quant_config.get_name()
+        if name not in _QUANTIZED_LINEAR_REGISTRY:
+            raise ValueError(f"No quantized linear is found for {name}")
+
+        quant_linear_cls = _QUANTIZED_LINEAR_REGISTRY[name][0]
+        return quant_linear_cls(*args, **kwargs)
+
+    @classmethod
+    def row(cls, *args, **kwargs) -> RowParallelLinear:
+        quant_config = kwargs.get("quant_config", None)
+        if quant_config is None:
+            return RowParallelLinear(*args, **kwargs)
+
+        name = quant_config.get_name()
+        if name not in _QUANTIZED_LINEAR_REGISTRY:
+            raise ValueError(f"No quantized linear is found for {name}")
+
+        quant_linear_cls = _QUANTIZED_LINEAR_REGISTRY[name][1]
+        return quant_linear_cls(*args, **kwargs)

vllm/model_executor/layers/quantized_linear/awq.py

@@ -0,0 +1,102 @@
+from typing import Optional
+
+import torch
+from torch.nn.parameter import Parameter
+
+from vllm import quantization_ops
+from vllm.model_executor.parallel_utils.layers import (ColumnParallelLinear,
+                                                       RowParallelLinear)
+
+
+class AWQColumnParallelLinear(ColumnParallelLinear):
+
+    def create_weights(self, dtype: torch.dtype) -> None:
+        assert self.input_size % self.quant_config.weight_bits == 0
+        assert (self.output_size_per_partition %
+                self.quant_config.pack_factor == 0)
+        self.qweight = Parameter(
+            torch.empty(
+                self.input_size,
+                self.output_size_per_partition //
+                self.quant_config.pack_factor,
+                device="cuda",
+                dtype=torch.int32,
+            ),
+            requires_grad=False,
+        )
+        self.qzeros = Parameter(
+            torch.empty(
+                self.input_size // self.quant_config.group_size,
+                self.output_size_per_partition //
+                self.quant_config.pack_factor,
+                device="cuda",
+                dtype=torch.int32,
+            ),
+            requires_grad=False,
+        )
+        self.scales = Parameter(
+            torch.empty(
+                self.input_size // self.quant_config.group_size,
+                self.output_size_per_partition,
+                device="cuda",
+                dtype=dtype,
+            ),
+            requires_grad=False,
+        )
+
+    def apply_weights(
+        self,
+        x: torch.Tensor,
+        bias: Optional[torch.Tensor],
+    ) -> torch.Tensor:
+        pack_factor = self.quant_config.pack_factor
+        out_shape = (x.shape[-2], self.qweight.shape[-1] * pack_factor)
+        reshaped_x = x.reshape(-1, x.shape[-1])
+        out = quantization_ops.awq_gemm(reshaped_x, self.qweight, self.scales,
+                                        self.qzeros, pack_factor)
+        if bias is not None:
+            out = out + bias
+        return out.reshape(out_shape)
+
+
+class AWQRowParallelLinear(RowParallelLinear):
+
+    def create_weights(self, dtype: torch.dtype) -> None:
+        assert (self.input_size_per_partition %
+                self.quant_config.weight_bits == 0)
+        assert self.output_size % self.quant_config.pack_factor == 0
+        self.qweight = Parameter(
+            torch.empty(
+                self.input_size_per_partition,
+                self.output_size // self.quant_config.pack_factor,
+                device="cuda",
+                dtype=torch.int32,
+            ),
+            requires_grad=False,
+        )
+        self.qzeros = Parameter(
+            torch.empty(
+                self.input_size_per_partition // self.quant_config.group_size,
+                self.output_size // self.quant_config.pack_factor,
+                device="cuda",
+                dtype=torch.int32,
+            ),
+            requires_grad=False,
+        )
+        self.scales = Parameter(
+            torch.empty(
+                self.input_size_per_partition // self.quant_config.group_size,
+                self.output_size,
+                device="cuda",
+                dtype=dtype,
+            ),
+            requires_grad=False,
+        )
+
+    def apply_weights(self, x: torch.Tensor) -> torch.Tensor:
+        pack_factor = self.quant_config.pack_factor
+        out_shape = (x.shape[-2], self.qweight.shape[-1] * pack_factor)
+        reshaped_x = x.reshape(-1, x.shape[-1])
+        out = quantization_ops.awq_gemm(reshaped_x, self.qweight, self.scales,
+                                        self.qzeros, pack_factor)
+        return out.reshape(out_shape)