add custom vllm source code

server/vllm/examples/offline_inference.py

+from vllm import LLM, SamplingParams
+
+# Sample prompts.
+prompts = [
+    "Hello, my name is",
+    "The president of the United States is",
+    "The capital of France is",
+    "The future of AI is",
+]
+# Create a sampling params object.
+sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
+
+# Create an LLM.
+llm = LLM(model="facebook/opt-125m")
+# Generate texts from the prompts. The output is a list of RequestOutput objects
+# that contain the prompt, generated text, and other information.
+outputs = llm.generate(prompts, sampling_params)
+# Print the outputs.
+for output in outputs:
+    prompt = output.prompt
+    generated_text = output.outputs[0].text
+    print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")

server/vllm/examples/openai_chatcompletion_client.py

+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)

server/vllm/examples/openai_completion_client.py

+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"]
+
+# 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)

server/vllm/format.sh

+#!/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

server/vllm/mypy.ini

+[mypy]
+python_version = 3.8
+
+ignore_missing_imports = True
+
+files = vllm
+# TODO(woosuk): Include the code from Megatron and HuggingFace.
+exclude = vllm/model_executor/parallel_utils/|vllm/model_executor/models/

server/vllm/pyproject.toml

+[build-system]
+requires = [
+    "ninja",
+    "packaging",
+    "setuptools",
+    "torch == 2.0.1",
+    "wheel",
+]
+build-backend = "setuptools.build_meta"

server/vllm/requirements-dev.txt

+# 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

server/vllm/requirements.txt

+ninja  # For faster builds.
+psutil
+pandas  # Required for Ray data.
+pyarrow  # Required for Ray data.
+sentencepiece  # Required for LLaMA tokenizer.
+numpy
+transformers==4.33.2  # match TGI
+uvicorn[standard]
+pydantic < 2  # Required for OpenAI server.

server/vllm/setup.py

+import io
+import os
+import re
+import subprocess
+from typing import List, Set
+import warnings
+
+from packaging.version import parse, Version
+import setuptools
+import torch
+from torch.utils.cpp_extension import BuildExtension, CUDAExtension, CUDA_HOME, ROCM_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?
+NVCC_FLAGS = ["-O2", "-std=c++17","--gpu-max-threads-per-block=1024"]
+
+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 torch.version.hip:
+    if ROCM_HOME is not None:
+        NVCC_FLAGS += [f"-DUSE_ROCM"]
+
+if not torch.version.hip:
+    if CUDA_HOME is None:
+        raise RuntimeError(
+            "Cannot find CUDA_HOME. CUDA must be available to build the package.")
+
+def get_nvcc_cuda_version(cuda_dir: str) -> Version:
+    """Get the CUDA version from nvcc.
+
+    Adapted from https://github.com/NVIDIA/apex/blob/8b7a1ff183741dd8f9b87e7bafd04cfde99cea28/setup.py
+    """
+    nvcc_output = subprocess.check_output([cuda_dir + "/bin/nvcc", "-V"],
+                                          universal_newlines=True)
+    output = nvcc_output.split()
+    release_idx = output.index("release") + 1
+    nvcc_cuda_version = parse(output[release_idx].split(",")[0])
+    return nvcc_cuda_version
+
+
+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(
+            "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
+
+def get_cuda_compute_capabilities(nvcc_cuda_version):
+    # First, check the TORCH_CUDA_ARCH_LIST environment variable.
+    compute_capabilities = get_torch_arch_list()
+    if not compute_capabilities:
+        # 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}")
+
+    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")
+
+    return compute_capabilities
+
+def validate_nvcc_cuda_version(nvcc_cuda_version, compute_capabilities):
+    if nvcc_cuda_version < Version("11.0"):
+        raise RuntimeError("CUDA 11.0 or higher is required to build the package.")
+    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.")
+
+if not torch.version.hip:
+    nvcc_cuda_version = get_nvcc_cuda_version(CUDA_HOME)
+    compute_capabilities = get_cuda_compute_capabilities(nvcc_cuda_version)
+    validate_nvcc_cuda_version(nvcc_cuda_version, compute_capabilities)
+
+    # 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"):
+        num_threads = min(os.cpu_count(), 8)
+        NVCC_FLAGS += ["--threads", str(num_threads)]
+
+ext_modules = []
+
+# Cache operations.
+cache_extension = CUDAExtension(
+    name="vllm.cache_ops",
+    sources=["csrc/cache.cpp", "csrc/cache_kernels.cu"],
+    extra_compile_args={
+        "cxx": CXX_FLAGS,
+        "nvcc": NVCC_FLAGS,
+    },
+)
+ext_modules.append(cache_extension)
+
+# Attention kernels.
+attention_extension = CUDAExtension(
+    name="vllm.attention_ops",
+    sources=["csrc/attention.cpp", "csrc/attention/attention_kernels.cu"],
+    extra_compile_args={
+        "cxx": CXX_FLAGS,
+        "nvcc": NVCC_FLAGS,
+    },
+)
+ext_modules.append(attention_extension)
+
+# Positional encoding kernels.
+positional_encoding_extension = CUDAExtension(
+    name="vllm.pos_encoding_ops",
+    sources=["csrc/pos_encoding.cpp", "csrc/pos_encoding_kernels.cu"],
+    extra_compile_args={
+        "cxx": CXX_FLAGS,
+        "nvcc": NVCC_FLAGS,
+    },
+)
+ext_modules.append(positional_encoding_extension)
+
+# Layer normalization kernels.
+layernorm_extension = CUDAExtension(
+    name="vllm.layernorm_ops",
+    sources=["csrc/layernorm.cpp", "csrc/layernorm_kernels.cu"],
+    extra_compile_args={
+        "cxx": CXX_FLAGS,
+        "nvcc": NVCC_FLAGS,
+    },
+)
+ext_modules.append(layernorm_extension)
+
+# Activation kernels.
+activation_extension = CUDAExtension(
+    name="vllm.activation_ops",
+    sources=["csrc/activation.cpp", "csrc/activation_kernels.cu"],
+    extra_compile_args={
+        "cxx": CXX_FLAGS,
+        "nvcc": NVCC_FLAGS,
+    },
+)
+ext_modules.append(activation_extension)
+
+# 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) #ROCM also need this
+# if not torch.version.hip:
+#     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)
+
+
+def find_version(filepath: str):
+    """Extract version information from the given filepath.
+
+    Adapted from https://github.com/ray-project/ray/blob/0b190ee1160eeca9796bc091e07eaebf4c85b511/python/setup.py
+    """
+    with open(filepath) as fp:
+        version_match = re.search(r"^__version__ = ['\"]([^'\"]*)['\"]",
+                                  fp.read(), re.M)
+        if version_match:
+            return version_match.group(1)
+        raise RuntimeError("Unable to find version string.")
+
+
+def read_readme() -> str:
+    """Read the README file."""
+    return io.open(get_path("README.md"), "r", encoding="utf-8").read()
+
+
+def get_requirements() -> List[str]:
+    """Get Python package dependencies from requirements.txt."""
+    with open(get_path("requirements.txt")) as f:
+        requirements = f.read().strip().split("\n")
+    return requirements
+
+
+setuptools.setup(
+    name="vllm",
+    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"),
+    long_description=read_readme(),
+    long_description_content_type="text/markdown",
+    url="https://github.com/vllm-project/vllm",
+    project_urls={
+        "Homepage": "https://github.com/vllm-project/vllm",
+        "Documentation": "https://vllm.readthedocs.io/en/latest/",
+    },
+    classifiers=[
+        "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=("benchmarks", "csrc", "docs",
+                                               "examples", "tests")),
+    python_requires=">=3.8",
+    install_requires=get_requirements(),
+    ext_modules=ext_modules,
+    cmdclass={"build_ext": BuildExtension},
+)

server/vllm/tests/async_engine/api_server_async_engine.py

+"""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)

server/vllm/tests/async_engine/test_api_server.py

+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

server/vllm/tests/async_engine/test_async_llm_engine.py

+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

server/vllm/tests/async_engine/test_request_tracker.py

+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

server/vllm/tests/conftest.py

+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

server/vllm/tests/distributed/test_comm_ops.py

+"""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)

server/vllm/tests/engine/test_detokenize.py

+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

server/vllm/tests/kernels/conftest.py

+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

server/vllm/tests/kernels/test_activation.py

+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 test_silu_and_mul(
+    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, 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)
+
+
+@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)

server/vllm/tests/kernels/test_attention.py

+import random
+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
+
+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(
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    scale: float,
+    attn_mask: Optional[torch.Tensor] = None,
+) -> torch.Tensor:
+    attn_weights = scale * torch.einsum("qhd,khd->hqk", query, key).float()
+    if attn_mask is not None:
+        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_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]
+
+    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])
+
+        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_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)
+
+        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:
+    num_seqs = len(cu_seq_lens) - 1
+    ref_outputs = []
+    for i in range(num_seqs):
+        start_idx = cu_seq_lens[i]
+        end_idx = cu_seq_lens[i + 1]
+        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 = attn_mask * torch.finfo(dtype).min
+        attn_mask = attn_mask.to(dtype=dtype, device="cuda")
+
+        ref_output = ref_masked_attention(
+            query[start_idx:end_idx],
+            key[start_idx:end_idx],
+            value[start_idx:end_idx],
+            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 test_multi_query_kv_attention(
+    num_seqs: int,
+    num_heads: Tuple[int, int],
+    head_size: int,
+    dtype: torch.dtype,
+    seed: int,
+) -> None:
+    random.seed(seed)
+    torch.random.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
+
+    # MAX_SEQ_LEN sometimes causes OOM in the reference implementation.
+    # As the xformers library is already tested with its own tests, we can use
+    # a smaller MAX_SEQ_LEN here.
+    max_len = min(MAX_SEQ_LEN, 4096)
+    seq_lens = random.sample(range(1, max_len), num_seqs)
+    num_tokens = sum(seq_lens)
+
+    scale = float(1.0 / (head_size**0.5))
+    num_query_heads, num_kv_heads = num_heads
+    qkv = torch.empty(num_tokens,
+                      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)
+
+    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),
+        key.unsqueeze(0),
+        value.unsqueeze(0),
+        attn_bias=attn_bias,
+        p=0.0,
+        scale=scale,
+    )
+    output = output.squeeze(0)
+
+    cu_seq_lens = [0]
+    for seq_len in seq_lens:
+        cu_seq_lens.append(cu_seq_lens[-1] + seq_len)
+    ref_output = ref_multi_query_kv_attention(
+        cu_seq_lens,
+        query,
+        key,
+        value,
+        scale,
+        dtype,
+    )
+    assert torch.allclose(output, ref_output, atol=1e-3, rtol=1e-5)

server/vllm/tests/kernels/test_cache.py

+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 test_copy_blocks(
+    kv_cache_factory,
+    num_mappings: int,
+    num_layers: 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)
+
+    # 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, 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 caches.
+    key_caches, value_caches = kv_cache_factory(num_blocks, block_size,
+                                                num_layers, num_heads,
+                                                head_size, dtype, seed)
+
+    # 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)
+
+    # Run the reference implementation.
+    for src, dsts in block_mapping.items():
+        for dst in dsts:
+            for cloned_key_cache in cloned_key_caches:
+                cloned_key_cache[dst] = cloned_key_cache[src]
+            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):
+        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 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")
+
+    qkv = torch.randn(num_tokens,
+                      3,
+                      num_heads,
+                      head_size,
+                      dtype=dtype,
+                      device="cuda")
+    _, key, value = qkv.unbind(dim=1)
+
+    # 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]
+
+    # Clone the KV caches.
+    cloned_key_cache = key_cache.clone()
+    cloned_value_cache = value_cache.clone()
+
+    # Call the reshape_and_cache kernel.
+    cache_ops.reshape_and_cache(key, value, key_cache, value_cache,
+                                slot_mapping)
+
+    # Run the reference implementation.
+    reshaped_key = key.reshape(num_tokens, *key_cache[0, :, :, 0, :].shape)
+    block_indicies = torch.div(slot_mapping, block_size, rounding_mode="floor")
+    block_indicies = block_indicies.cpu().tolist()
+    block_offsets = slot_mapping % block_size
+    block_offsets = block_offsets.cpu().tolist()
+    for i in range(num_tokens):
+        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)