Merge remote-tracking branch 'upstream/main' into control-refactor-2

.buildkite/run-benchmarks.sh

@@ -54,12 +54,12 @@ tail -n 24 benchmark_serving.txt >> benchmark_results.md # last 24 lines
 echo '```' >> benchmark_results.md
 
 # if the agent binary is not found, skip uploading the results, exit 0
-if [ ! -f /workspace/buildkite-agent ]; then
+if [ ! -f /usr/bin/buildkite-agent ]; then
     exit 0
 fi
 
 # upload the results to buildkite
-/workspace/buildkite-agent annotate --style "info" --context "benchmark-results" < benchmark_results.md
+buildkite-agent annotate --style "info" --context "benchmark-results" < benchmark_results.md
 
 # exit with the exit code of the benchmarks
 if [ $bench_latency_exit_code -ne 0 ]; then
@@ -75,4 +75,4 @@ if [ $bench_serving_exit_code -ne 0 ]; then
 fi
 
 rm ShareGPT_V3_unfiltered_cleaned_split.json
-/workspace/buildkite-agent artifact upload "*.json"
+buildkite-agent artifact upload "*.json"

.buildkite/test-pipeline.yaml

@@ -37,6 +37,7 @@ steps:
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
   commands:
+  - VLLM_TEST_SAME_HOST=1 torchrun --nproc-per-node=4 distributed/test_same_node.py
   - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_basic_distributed_correctness.py
   - TEST_DIST_MODEL=meta-llama/Llama-2-7b-hf DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_basic_distributed_correctness.py
   - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_chunked_prefill_distributed.py

.buildkite/test-template-aws.j2

@@ -19,17 +19,47 @@ steps:
           limit: 5
   - wait
 
+  - group: "AMD Tests"
+    depends_on: ~
+    steps:
+    {% for step in steps %}
+    {% if step.mirror_hardwares and "amd" in step.mirror_hardwares %}
+      - label: "AMD: {{ step.label }}"
+        agents:
+          queue: amd
+        command: bash .buildkite/run-amd-test.sh "cd {{ (step.working_dir or default_working_dir) | safe  }} ; {{ step.command  or (step.commands | join(" ; ")) | safe }}"
+        env:
+          DOCKER_BUILDKIT: "1"
+        soft_fail: true
+    {% endif %}
+    {% endfor %}
+
+  - label: "Neuron Test"
+    depends_on: ~
+    agents:
+      queue: neuron
+    command: bash .buildkite/run-neuron-test.sh
+    soft_fail: false
+
+  - label: "Intel Test"
+    depends_on: ~
+    agents:
+      queue: intel
+    command: bash .buildkite/run-cpu-test.sh
+
   {% for step in steps %}
   - label: "{{ step.label }}"
     agents:
-      {% if step.no_gpu %}
+      {% if step.label == "Documentation Build" %}
+      queue: small_cpu_queue
+      {% elif step.no_gpu %}
       queue: cpu_queue
       {% elif step.num_gpus == 2 or step.num_gpus == 4 %}
       queue: gpu_4_queue
       {% else %}
       queue: gpu_1_queue
       {% endif %}
-    soft_fail: true
+    soft_fail: {{ step.soft_fail or false }}
     {% if step.parallelism %}
     parallelism: {{ step.parallelism }}
     {% endif %}
@@ -47,6 +77,9 @@ steps:
           {% if not step.no_gpu %}
           gpus: all
           {% endif %}
+          {% if step.label == "Benchmarks" %}
+          mount-buildkite-agent: true
+          {% endif %}
           command: ["bash", "-c", "cd {{ (step.working_dir or default_working_dir) | safe  }} && {{ step.command  or (step.commands | join(' && ')) | safe }}"]
           environment:
             - VLLM_USAGE_SOURCE=ci-test

.github/workflows/ruff.yml

@@ -25,7 +25,7 @@ jobs:
     - name: Install dependencies
       run: |
         python -m pip install --upgrade pip
-        pip install ruff==0.1.5 codespell==2.2.6 tomli==2.0.1 isort==5.13.2
+        pip install ruff==0.1.5 codespell==2.3.0 tomli==2.0.1 isort==5.13.2
     - name: Analysing the code with ruff
       run: |
         ruff .

CMakeLists.txt

@@ -66,19 +66,6 @@ endif()
 #
 find_package(Torch REQUIRED)
 
-#
-# Normally `torch.utils.cpp_extension.CUDAExtension` would add
-# `libtorch_python.so` for linking against an extension. Torch's cmake
-# configuration does not include this library (presumably since the cmake
-# config is used for standalone C++ binaries that link against torch).
-# The `libtorch_python.so` library defines some of the glue code between
-# torch/python via pybind and is required by VLLM extensions for this
-# reason. So, add it by manually with `find_library` using torch's
-# installed library path.
-#
-find_library(torch_python_LIBRARY torch_python PATHS
-  "${TORCH_INSTALL_PREFIX}/lib")
-
 #
 # Forward the non-CUDA device extensions to external CMake scripts.
 #
@@ -171,7 +158,7 @@ set(VLLM_EXT_SRC
   "csrc/quantization/fp8/common.cu"
   "csrc/cuda_utils_kernels.cu"
   "csrc/moe_align_block_size_kernels.cu"
-  "csrc/pybind.cpp")
+  "csrc/torch_bindings.cpp")
 
 if(VLLM_GPU_LANG STREQUAL "CUDA")
   include(FetchContent)
@@ -218,14 +205,15 @@ define_gpu_extension_target(
   COMPILE_FLAGS ${VLLM_GPU_FLAGS}
   ARCHITECTURES ${VLLM_GPU_ARCHES}
   INCLUDE_DIRECTORIES ${CUTLASS_INCLUDE_DIR};${CUTLASS_TOOLS_UTIL_INCLUDE_DIR}
+  USE_SABI 3
   WITH_SOABI)
 
 #
 # _moe_C extension
 #
 
 set(VLLM_MOE_EXT_SRC
-  "csrc/moe/moe_ops.cpp"
+  "csrc/moe/torch_bindings.cpp"
   "csrc/moe/topk_softmax_kernels.cu")
 
 define_gpu_extension_target(
@@ -235,6 +223,7 @@ define_gpu_extension_target(
   SOURCES ${VLLM_MOE_EXT_SRC}
   COMPILE_FLAGS ${VLLM_GPU_FLAGS}
   ARCHITECTURES ${VLLM_GPU_ARCHES}
+  USE_SABI 3
   WITH_SOABI)
 
 #
@@ -249,7 +238,7 @@ set(VLLM_PUNICA_EXT_SRC
   "csrc/punica/bgmv/bgmv_fp32_bf16_bf16.cu"
   "csrc/punica/bgmv/bgmv_fp32_fp16_fp16.cu"
   "csrc/punica/punica_ops.cu"
-  "csrc/punica/punica_pybind.cpp")
+  "csrc/punica/torch_bindings.cpp")
 
 #
 # Copy GPU compilation flags+update for punica
@@ -286,6 +275,7 @@ if (VLLM_PUNICA_GPU_ARCHES)
     SOURCES ${VLLM_PUNICA_EXT_SRC}
     COMPILE_FLAGS ${VLLM_PUNICA_GPU_FLAGS}
     ARCHITECTURES ${VLLM_PUNICA_GPU_ARCHES}
+    USE_SABI 3
     WITH_SOABI)
 else()
   message(WARNING "Unable to create _punica_C target because none of the "

Dockerfile.neuron

@@ -28,7 +28,7 @@ COPY ./requirements-neuron.txt /app/vllm/requirements-neuron.txt
 RUN cd /app/vllm \
     && python3 -m pip install -U -r requirements-neuron.txt
 
-ENV VLLM_BUILD_WITH_NEURON 1
+ENV VLLM_TARGET_DEVICE neuron
 RUN cd /app/vllm \
     && pip install -e . \
     && cd ..

Dockerfile.rocm

@@ -106,9 +106,9 @@ RUN --mount=type=cache,target=/root/.cache/pip \
     pip install -U -r requirements-rocm.txt \
     && patch /opt/rocm/include/hip/amd_detail/amd_hip_bf16.h ./rocm_patch/rocm_bf16.patch \
     && python3 setup.py install \
-    && cp build/lib.linux-x86_64-cpython-39/vllm/_C.cpython-39-x86_64-linux-gnu.so vllm/ \
-    && cp build/lib.linux-x86_64-cpython-39/vllm/_punica_C.cpython-39-x86_64-linux-gnu.so vllm/ \
-    && cp build/lib.linux-x86_64-cpython-39/vllm/_moe_C.cpython-39-x86_64-linux-gnu.so vllm/ \
+    && cp build/lib.linux-x86_64-cpython-39/vllm/_C.abi3.so vllm/ \
+    && cp build/lib.linux-x86_64-cpython-39/vllm/_punica_C.abi3.so vllm/ \
+    && cp build/lib.linux-x86_64-cpython-39/vllm/_moe_C.abi3.so vllm/ \
     && cd ..
 
 

Dockerfile.tpu

@@ -0,0 +1,19 @@
+ARG NIGHTLY_DATE="20240601"
+ARG BASE_IMAGE="us-central1-docker.pkg.dev/tpu-pytorch-releases/docker/xla:nightly_3.10_tpuvm_$NIGHTLY_DATE"
+
+FROM $BASE_IMAGE
+
+WORKDIR /workspace
+COPY . /workspace/vllm
+
+ENV VLLM_TARGET_DEVICE="tpu"
+# Install aiohttp separately to avoid build errors.
+RUN pip install aiohttp
+# Install the TPU and Pallas dependencies.
+RUN pip install torch_xla[tpu] -f https://storage.googleapis.com/libtpu-releases/index.html
+RUN pip install torch_xla[pallas] -f https://storage.googleapis.com/jax-releases/jax_nightly_releases.html -f https://storage.googleapis.com/jax-releases/jaxlib_nightly_releases.html
+
+# Build vLLM.
+RUN cd /workspace/vllm && python setup.py develop
+
+CMD ["/bin/bash"]

README.md

@@ -111,10 +111,10 @@ vLLM is a community project. Our compute resources for development and testing a
 - Dropbox
 - Lambda Lab
 - NVIDIA
-- Sequoia Capital
 - Replicate
 - Roblox
 - RunPod
+- Sequoia Capital
 - Trainy
 - UC Berkeley
 - UC San Diego

benchmarks/benchmark_latency.py

@@ -36,7 +36,8 @@ def main(args: argparse.Namespace):
               enable_chunked_prefill=args.enable_chunked_prefill,
               download_dir=args.download_dir,
               block_size=args.block_size,
-              gpu_memory_utilization=args.gpu_memory_utilization)
+              gpu_memory_utilization=args.gpu_memory_utilization,
+              distributed_executor_backend=args.distributed_executor_backend)
 
     sampling_params = SamplingParams(
         n=args.n,
@@ -188,7 +189,7 @@ def run_to_completion(profile_dir: Optional[str] = None):
         "--device",
         type=str,
         default="cuda",
-        choices=["cuda", "cpu"],
+        choices=["cuda", "cpu", "tpu"],
         help='device type for vLLM execution, supporting CUDA and CPU.')
     parser.add_argument('--block-size',
                         type=int,
@@ -221,5 +222,12 @@ def run_to_completion(profile_dir: Optional[str] = None):
                         help='the fraction of GPU memory to be used for '
                         'the model executor, which can range from 0 to 1.'
                         'If unspecified, will use the default value of 0.9.')
+    parser.add_argument(
+        '--distributed-executor-backend',
+        choices=['ray', 'mp'],
+        default=None,
+        help='Backend to use for distributed serving. When more than 1 GPU '
+        'is used, will be automatically set to "ray" if installed '
+        'or "mp" (multiprocessing) otherwise.')
     args = parser.parse_args()
     main(args)

benchmarks/benchmark_throughput.py

@@ -78,6 +78,7 @@ def run_vllm(
     enable_prefix_caching: bool,
     enable_chunked_prefill: bool,
     max_num_batched_tokens: int,
+    distributed_executor_backend: Optional[str],
     gpu_memory_utilization: float = 0.9,
     download_dir: Optional[str] = None,
 ) -> float:
@@ -100,6 +101,7 @@ def run_vllm(
         download_dir=download_dir,
         enable_chunked_prefill=enable_chunked_prefill,
         max_num_batched_tokens=max_num_batched_tokens,
+        distributed_executor_backend=distributed_executor_backend,
     )
 
     # Add the requests to the engine.
@@ -225,8 +227,8 @@ def main(args: argparse.Namespace):
             args.enforce_eager, args.kv_cache_dtype,
             args.quantization_param_path, args.device,
             args.enable_prefix_caching, args.enable_chunked_prefill,
-            args.max_num_batched_tokens, args.gpu_memory_utilization,
-            args.download_dir)
+            args.max_num_batched_tokens, args.distributed_executor_backend,
+            args.gpu_memory_utilization, args.download_dir)
     elif args.backend == "hf":
         assert args.tensor_parallel_size == 1
         elapsed_time = run_hf(requests, args.model, tokenizer, args.n,
@@ -344,7 +346,7 @@ def main(args: argparse.Namespace):
         "--device",
         type=str,
         default="cuda",
-        choices=["cuda", "cpu"],
+        choices=["cuda", "cpu", "tpu"],
         help='device type for vLLM execution, supporting CUDA and CPU.')
     parser.add_argument(
         "--enable-prefix-caching",
@@ -368,6 +370,13 @@ def main(args: argparse.Namespace):
         type=str,
         default=None,
         help='Path to save the throughput results in JSON format.')
+    parser.add_argument(
+        '--distributed-executor-backend',
+        choices=['ray', 'mp'],
+        default=None,
+        help='Backend to use for distributed serving. When more than 1 GPU '
+        'is used, will be automatically set to "ray" if installed '
+        'or "mp" (multiprocessing) otherwise.')
     args = parser.parse_args()
     if args.tokenizer is None:
         args.tokenizer = args.model

cmake/cpu_extension.cmake

@@ -12,7 +12,7 @@ include_directories("${CMAKE_SOURCE_DIR}/csrc")
 #
 # Check the compile flags
 #
-list(APPEND CXX_COMPILE_FLAGS 
+list(APPEND CXX_COMPILE_FLAGS
     "-fopenmp"
     "-DVLLM_CPU_EXTENSION")
 
@@ -44,8 +44,8 @@ if (AVX512_FOUND)
 
     find_isa(${CPUINFO} "avx512_bf16" AVX512BF16_FOUND)
     if (AVX512BF16_FOUND OR ENABLE_AVX512BF16)
-        if (CMAKE_CXX_COMPILER_ID STREQUAL "GNU" AND 
-            CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 12.3) 
+        if (CMAKE_CXX_COMPILER_ID STREQUAL "GNU" AND
+            CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 12.3)
             list(APPEND CXX_COMPILE_FLAGS "-mavx512bf16")
         else()
             message(WARNING "Disable AVX512-BF16 ISA support, requires gcc/g++ >= 12.3")
@@ -73,18 +73,18 @@ set(VLLM_EXT_SRC
     "csrc/cpu/cache.cpp"
     "csrc/cpu/layernorm.cpp"
     "csrc/cpu/pos_encoding.cpp"
-    "csrc/cpu/pybind.cpp")
+    "csrc/cpu/torch_bindings.cpp")
 
 define_gpu_extension_target(
     _C
     DESTINATION vllm
     LANGUAGE CXX
     SOURCES ${VLLM_EXT_SRC}
     COMPILE_FLAGS ${CXX_COMPILE_FLAGS}
-    WITH_SOABI 
+    USE_SABI 3
+    WITH_SOABI
 )
 
 add_custom_target(default)
 message(STATUS "Enabling C extension.")
 add_dependencies(default _C)
-

cmake/utils.cmake

@@ -5,7 +5,7 @@
 macro (find_python_from_executable EXECUTABLE SUPPORTED_VERSIONS)
   file(REAL_PATH ${EXECUTABLE} EXECUTABLE)
   set(Python_EXECUTABLE ${EXECUTABLE})
-  find_package(Python COMPONENTS Interpreter Development.Module)
+  find_package(Python COMPONENTS Interpreter Development.Module Development.SABIModule)
   if (NOT Python_FOUND)
     message(FATAL_ERROR "Unable to find python matching: ${EXECUTABLE}.")
   endif()
@@ -294,14 +294,15 @@ endmacro()
 # INCLUDE_DIRECTORIES <dirs> - Extra include directories.
 # LIBRARIES <libraries>      - Extra link libraries.
 # WITH_SOABI                 - Generate library with python SOABI suffix name.
+# USE_SABI <version>         - Use python stable api <version>
 #
 # Note: optimization level/debug info is set via cmake build type.
 #
 function (define_gpu_extension_target GPU_MOD_NAME)
   cmake_parse_arguments(PARSE_ARGV 1
     GPU
     "WITH_SOABI"
-    "DESTINATION;LANGUAGE"
+    "DESTINATION;LANGUAGE;USE_SABI"
     "SOURCES;ARCHITECTURES;COMPILE_FLAGS;INCLUDE_DIRECTORIES;LIBRARIES")
 
   # Add hipify preprocessing step when building with HIP/ROCm.
@@ -315,7 +316,11 @@ function (define_gpu_extension_target GPU_MOD_NAME)
     set(GPU_WITH_SOABI)
   endif()
 
-  Python_add_library(${GPU_MOD_NAME} MODULE "${GPU_SOURCES}" ${GPU_WITH_SOABI})
+  if (GPU_USE_SABI)
+    Python_add_library(${GPU_MOD_NAME} MODULE USE_SABI ${GPU_USE_SABI} ${GPU_WITH_SOABI} "${GPU_SOURCES}")
+  else()
+    Python_add_library(${GPU_MOD_NAME} MODULE ${GPU_WITH_SOABI} "${GPU_SOURCES}")
+  endif()
 
   if (GPU_LANGUAGE STREQUAL "HIP")
     # Make this target dependent on the hipify preprocessor step.

csrc/activation_kernels.cu

@@ -1,5 +1,5 @@
 #include <ATen/cuda/CUDAContext.h>
-#include <torch/extension.h>
+#include <torch/all.h>
 #include <c10/cuda/CUDAGuard.h>
 
 #include <cmath>