Sync with upstream@v0.4.3-60-gbaa15a9e

.buildkite/test-pipeline.yaml

@@ -46,6 +46,7 @@ steps:
   - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=mp pytest -v -s distributed/test_chunked_prefill_distributed.py
   - TEST_DIST_MODEL=meta-llama/Llama-2-7b-hf DISTRIBUTED_EXECUTOR_BACKEND=mp pytest -v -s distributed/test_chunked_prefill_distributed.py
   - pytest -v -s spec_decode/e2e/test_integration_dist.py
+  - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s test_sharded_state_loader.py
 
 - label: Distributed Tests (Multiple Groups)
   #mirror_hardwares: [amd]
@@ -138,14 +139,7 @@ steps:
   num_gpus: 4
   # This test runs llama 13B, so it is required to run on 4 GPUs.
   commands:
-    # Temporarily run this way because we cannot clean up GPU mem usage
-    # for multi GPU tests.
-    # TODO(sang): Fix it.
-    - pytest -v -s lora/test_long_context.py::test_rotary_emb_replaced
-    - pytest -v -s lora/test_long_context.py::test_batched_rope_kernel
-    - pytest -v -s lora/test_long_context.py::test_self_consistency
-    - pytest -v -s lora/test_long_context.py::test_quality
-    - pytest -v -s lora/test_long_context.py::test_max_len
+    - pytest -v -s -x lora/test_long_context.py
 
 - label: Tensorizer Test
   #mirror_hardwares: [amd]

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.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 ..
 
 

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,
@@ -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,
@@ -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>

csrc/attention/attention_kernels.cu

@@ -17,7 +17,7 @@
  * limitations under the License.
  */
 
-#include <torch/extension.h>
+#include <torch/all.h>
 #include <ATen/cuda/CUDAContext.h>
 #include <c10/cuda/CUDAGuard.h>
 #include <algorithm>
@@ -808,16 +808,17 @@ void paged_attention_v1(
     torch::Tensor&
         key_cache,  // [num_blocks, num_heads, head_size/x, block_size, x]
     torch::Tensor&
-        value_cache,   // [num_blocks, num_heads, head_size, block_size]
-    int num_kv_heads,  // [num_heads]
-    float scale,
+        value_cache,       // [num_blocks, num_heads, head_size, block_size]
+    int64_t num_kv_heads,  // [num_heads]
+    double scale,
     torch::Tensor& block_tables,  // [num_seqs, max_num_blocks_per_seq]
     torch::Tensor& seq_lens,      // [num_seqs]
-    int block_size, int max_seq_len,
+    int64_t block_size, int64_t max_seq_len,
     const c10::optional<torch::Tensor>& alibi_slopes,
-    const std::string& kv_cache_dtype, float kv_scale, const int tp_rank,
-    const int blocksparse_local_blocks, const int blocksparse_vert_stride,
-    const int blocksparse_block_size, const int blocksparse_head_sliding_step) {
+    const std::string& kv_cache_dtype, double kv_scale, const int64_t tp_rank,
+    const int64_t blocksparse_local_blocks,
+    const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
+    const int64_t blocksparse_head_sliding_step) {
   const bool is_block_sparse = (blocksparse_vert_stride > 1);
 
   DISPATCH_BY_KV_CACHE_DTYPE(query.dtype(), kv_cache_dtype,
@@ -972,16 +973,17 @@ void paged_attention_v2(
     torch::Tensor&
         key_cache,  // [num_blocks, num_heads, head_size/x, block_size, x]
     torch::Tensor&
-        value_cache,   // [num_blocks, num_heads, head_size, block_size]
-    int num_kv_heads,  // [num_heads]
-    float scale,
+        value_cache,       // [num_blocks, num_heads, head_size, block_size]
+    int64_t num_kv_heads,  // [num_heads]
+    double scale,
     torch::Tensor& block_tables,  // [num_seqs, max_num_blocks_per_seq]
     torch::Tensor& seq_lens,      // [num_seqs]
-    int block_size, int max_seq_len,
+    int64_t block_size, int64_t max_seq_len,
     const c10::optional<torch::Tensor>& alibi_slopes,
-    const std::string& kv_cache_dtype, float kv_scale, const int tp_rank,
-    const int blocksparse_local_blocks, const int blocksparse_vert_stride,
-    const int blocksparse_block_size, const int blocksparse_head_sliding_step) {
+    const std::string& kv_cache_dtype, double kv_scale, const int64_t tp_rank,
+    const int64_t blocksparse_local_blocks,
+    const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
+    const int64_t blocksparse_head_sliding_step) {
   const bool is_block_sparse = (blocksparse_vert_stride > 1);
   DISPATCH_BY_KV_CACHE_DTYPE(query.dtype(), kv_cache_dtype,
                              CALL_V2_LAUNCHER_BLOCK_SIZE)
@@ -990,4 +992,4 @@ void paged_attention_v2(
 #undef WARP_SIZE
 #undef MAX
 #undef MIN
-#undef DIVIDE_ROUND_UP
+#undef DIVIDE_ROUND_UP

csrc/cache.h

@@ -1,21 +1,25 @@
 #pragma once
 
-#include <torch/extension.h>
+#include <torch/all.h>
 
 #include <map>
 #include <vector>
 
 void swap_blocks(torch::Tensor& src, torch::Tensor& dst,
                  const torch::Tensor& block_mapping);
 
-void copy_blocks(std::vector<torch::Tensor>& key_caches,
-                 std::vector<torch::Tensor>& value_caches,
+// Note: the key_caches and value_caches vectors are constant but
+// not the Tensors they contain. The vectors need to be const refs
+// in order to satisfy pytorch's C++ operator registration code.
+void copy_blocks(std::vector<torch::Tensor> const& key_caches,
+                 std::vector<torch::Tensor> const& value_caches,
                  const torch::Tensor& block_mapping);
 
 void reshape_and_cache(torch::Tensor& key, torch::Tensor& value,
                        torch::Tensor& key_cache, torch::Tensor& value_cache,
                        torch::Tensor& slot_mapping,
-                       const std::string& kv_cache_dtype, const float kv_scale);
+                       const std::string& kv_cache_dtype,
+                       const double kv_scale);
 
 void reshape_and_cache_flash(torch::Tensor& key, torch::Tensor& value,
                              torch::Tensor& key_cache,
@@ -25,4 +29,4 @@ void reshape_and_cache_flash(torch::Tensor& key, torch::Tensor& value,
 
 // Just for unittest
 void convert_fp8(torch::Tensor& dst_cache, torch::Tensor& src_cache,
-                 const float scale, const std::string& kv_cache_dtype);
+                 const double scale, const std::string& kv_cache_dtype);

csrc/cache_kernels.cu

@@ -1,4 +1,4 @@
-#include <torch/extension.h>
+#include <torch/all.h>
 #include <ATen/cuda/CUDAContext.h>
 #include <c10/cuda/CUDAGuard.h>
 
@@ -95,8 +95,11 @@ __global__ void copy_blocks_kernel(int64_t* key_cache_ptrs,
 
 }  // namespace vllm
 
-void copy_blocks(std::vector<torch::Tensor>& key_caches,
-                 std::vector<torch::Tensor>& value_caches,
+// Note: the key_caches and value_caches vectors are constant but
+// not the Tensors they contain. The vectors need to be const refs
+// in order to satisfy pytorch's C++ operator registration code.
+void copy_blocks(std::vector<torch::Tensor> const& key_caches,
+                 std::vector<torch::Tensor> const& value_caches,
                  const torch::Tensor& block_mapping) {
   int num_layers = key_caches.size();
   TORCH_CHECK(num_layers == value_caches.size());
@@ -255,7 +258,7 @@ void reshape_and_cache(
     torch::Tensor&
         value_cache,  // [num_blocks, num_heads, head_size, block_size]
     torch::Tensor& slot_mapping,  // [num_tokens]
-    const std::string& kv_cache_dtype, const float kv_scale) {
+    const std::string& kv_cache_dtype, const double kv_scale) {
   int num_tokens = key.size(0);
   int num_heads = key.size(1);
   int head_size = key.size(2);
@@ -334,7 +337,7 @@ __global__ void convert_fp8_kernel(const Tin* __restrict__ src_cache,
 
 // Only for testing.
 void convert_fp8(torch::Tensor& dst_cache, torch::Tensor& src_cache,
-                 const float kv_scale, const std::string& kv_cache_dtype) {
+                 const double kv_scale, const std::string& kv_cache_dtype) {
   torch::Device src_device = src_cache.device();
   torch::Device dst_device = dst_cache.device();
   TORCH_CHECK(src_device.is_cuda(), "src must be on a GPU")