Merge pull request ROCm#538 from ROCm/upstream_merge_2025_05_06

Upstream merge 2025 05 06

Author: gshtras

.buildkite/release-pipeline.yaml

@@ -57,6 +57,7 @@ steps:
     agents:
       queue: tpu_queue_postmerge
     commands:
+      - "yes | docker system prune -a"
       - "git fetch --all"
       - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --tag vllm/vllm-tpu:nightly --tag vllm/vllm-tpu:$BUILDKITE_COMMIT --progress plain -f docker/Dockerfile.tpu ."
       - "docker push vllm/vllm-tpu:nightly"

.buildkite/test-pipeline.yaml

@@ -39,7 +39,7 @@ steps:
   - pip install -r ../../requirements/docs.txt
   - SPHINXOPTS=\"-W\" make html
   # Check API reference (if it fails, you may have missing mock imports)
-  - grep \"sig sig-object py\" build/html/api/inference_params.html
+  - grep \"sig sig-object py\" build/html/api/vllm/vllm.sampling_params.html
 
 - label: Async Engine, Inputs, Utils, Worker Test # 24min
   source_file_dependencies:

.gitignore

@@ -80,6 +80,7 @@ instance/
 # Sphinx documentation
 docs/_build/
 docs/source/getting_started/examples/
+docs/source/api/vllm
 
 # PyBuilder
 .pybuilder/

.pre-commit-config.yaml

@@ -46,7 +46,7 @@ repos:
   rev: 0.6.17
   hooks:
     - id: pip-compile
-      args: [requirements/test.in, -o, requirements/test.txt, --index-strategy, unsafe-best-match]
+      args: [requirements/test.in, -o, requirements/test.txt, --index-strategy, unsafe-best-match, --torch-backend, cu128]
       files: ^requirements/test\.(in|txt)$
 - repo: local
   hooks:
@@ -101,8 +101,8 @@ repos:
     args:
       - -c
       - |
-        if ! grep -q "^Signed-off-by: $(git config user.name) <$(git config user.email)>" .git/COMMIT_EDITMSG; then
-          printf "\nSigned-off-by: $(git config user.name) <$(git config user.email)>\n" >> .git/COMMIT_EDITMSG
+        if ! grep -q "^Signed-off-by: $(git config user.name) <$(git config user.email)>" "$(git rev-parse --git-path COMMIT_EDITMSG)"; then
+          printf "\nSigned-off-by: $(git config user.name) <$(git config user.email)>\n" >> "$(git rev-parse --git-path COMMIT_EDITMSG)"
         fi
     language: system
     verbose: true
@@ -125,8 +125,6 @@ repos:
     name: Update Dockerfile dependency graph
     entry: tools/update-dockerfile-graph.sh
     language: script
-    files: ^docker/Dockerfile$
-    pass_filenames: false
   # Keep `suggestion` last
   - id: suggestion
     name: Suggestion

CMakeLists.txt

@@ -15,7 +15,6 @@ project(vllm_extensions LANGUAGES CXX)
 
 # CUDA by default, can be overridden by using -DVLLM_TARGET_DEVICE=... (used by setup.py)
 set(VLLM_TARGET_DEVICE "cuda" CACHE STRING "Target device backend for vLLM")
-
 message(STATUS "Build type: ${CMAKE_BUILD_TYPE}")
 message(STATUS "Target device: ${VLLM_TARGET_DEVICE}")
 
@@ -251,9 +250,8 @@ set(VLLM_EXT_SRC
 if(VLLM_GPU_LANG STREQUAL "CUDA")
   SET(CUTLASS_ENABLE_HEADERS_ONLY ON CACHE BOOL "Enable only the header library")
 
-  # Set CUTLASS_REVISION manually -- its revision detection doesn't work in this case.
-  # Please keep this in sync with FetchContent_Declare line below.
-  set(CUTLASS_REVISION "v3.9.0" CACHE STRING "CUTLASS revision to use")
+  # Set CUTLASS_REVISION. Used for FetchContent. Also fixes some bogus messages when building.
+  set(CUTLASS_REVISION "v3.9.2" CACHE STRING "CUTLASS revision to use")
 
   # Use the specified CUTLASS source directory for compilation if VLLM_CUTLASS_SRC_DIR is provided
   if (DEFINED ENV{VLLM_CUTLASS_SRC_DIR})
@@ -271,7 +269,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
         cutlass
         GIT_REPOSITORY https://github.com/nvidia/cutlass.git
         # Please keep this in sync with CUTLASS_REVISION line above.
-        GIT_TAG v3.9.0
+        GIT_TAG ${CUTLASS_REVISION}
         GIT_PROGRESS TRUE
 
         # Speed up CUTLASS download by retrieving only the specified GIT_TAG instead of the history.
@@ -304,8 +302,52 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # are not supported by Machete yet.
   cuda_archs_loose_intersection(MARLIN_ARCHS "8.0;8.6;8.7;8.9;9.0;10.0;10.1;12.0" "${CUDA_ARCHS}")
   if (MARLIN_ARCHS)
+
+    #
+    # For the Marlin kernels we automatically generate sources for various
+    # preselected input type pairs and schedules.
+    # Generate sources:
+    set(MARLIN_GEN_SCRIPT
+      ${CMAKE_CURRENT_SOURCE_DIR}/csrc/quantization/gptq_marlin/generate_kernels.py)
+    file(MD5 ${MARLIN_GEN_SCRIPT} MARLIN_GEN_SCRIPT_HASH)
+
+    message(STATUS "Marlin generation script hash: ${MARLIN_GEN_SCRIPT_HASH}")
+    message(STATUS "Last run Marlin generate script hash: $CACHE{MARLIN_GEN_SCRIPT_HASH}")
+
+    if (NOT DEFINED CACHE{MARLIN_GEN_SCRIPT_HASH}
+        OR NOT $CACHE{MARLIN_GEN_SCRIPT_HASH} STREQUAL ${MARLIN_GEN_SCRIPT_HASH})
+      execute_process(
+        COMMAND ${CMAKE_COMMAND} -E env
+        PYTHONPATH=$PYTHONPATH
+          ${Python_EXECUTABLE} ${MARLIN_GEN_SCRIPT}
+        RESULT_VARIABLE marlin_generation_result
+        OUTPUT_VARIABLE marlin_generation_result
+        OUTPUT_FILE ${CMAKE_CURRENT_BINARY_DIR}/marlin_generation.log
+        ERROR_FILE ${CMAKE_CURRENT_BINARY_DIR}/marlin_generation.log
+      )
+
+      if (NOT marlin_generation_result EQUAL 0)
+        message(FATAL_ERROR "Marlin generation failed."
+                            " Result: \"${marlin_generation_result}\""
+                            "\nCheck the log for details: "
+                            "${CMAKE_CURRENT_BINARY_DIR}/marlin_generation.log")
+      else()
+        set(MARLIN_GEN_SCRIPT_HASH ${MARLIN_GEN_SCRIPT_HASH}
+            CACHE STRING "Last run Marlin generate script hash" FORCE)
+        message(STATUS "Marlin generation completed successfully.")
+      endif()
+    else()
+      message(STATUS "Marlin generation script has not changed, skipping generation.")
+    endif()
+
+    file(GLOB MARLIN_TEMPLATE_KERNEL_SRC "csrc/quantization/gptq_marlin/kernel_*.cu")
+    set_gencode_flags_for_srcs(
+      SRCS "${MARLIN_TEMPLATE_KERNEL_SRC}"
+      CUDA_ARCHS "${MARLIN_ARCHS}")
+
+    list(APPEND VLLM_EXT_SRC ${MARLIN_TEMPLATE_KERNEL_SRC})
+
     set(MARLIN_SRCS
-       "csrc/quantization/fp8/fp8_marlin.cu"
        "csrc/quantization/marlin/dense/marlin_cuda_kernel.cu"
        "csrc/quantization/marlin/sparse/marlin_24_cuda_kernel.cu"
        "csrc/quantization/marlin/qqq/marlin_qqq_gemm_kernel.cu"
@@ -647,7 +689,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
         OR NOT $CACHE{MOE_MARLIN_GEN_SCRIPT_HASH} STREQUAL ${MOE_MARLIN_GEN_SCRIPT_HASH})
       execute_process(
         COMMAND ${CMAKE_COMMAND} -E env
-        PYTHONPATH=${CMAKE_CURRENT_SOURCE_DIR}/csrc/cutlass_extensions/:${CUTLASS_DIR}/python/:${VLLM_PYTHON_PATH}:$PYTHONPATH
+        PYTHONPATH=$PYTHONPATH
           ${Python_EXECUTABLE} ${MOE_MARLIN_GEN_SCRIPT}
         RESULT_VARIABLE moe_marlin_generation_result
         OUTPUT_VARIABLE moe_marlin_generation_output
@@ -683,6 +725,17 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   endif()
 endif()
 
+if(VLLM_GPU_LANG STREQUAL "CUDA")
+  set(MOE_PERMUTE_SRC
+      "csrc/moe/permute_unpermute_kernels/moe_permute_unpermute_kernel.cu"
+      "csrc/moe/moe_permute_unpermute_op.cu")
+
+  set_gencode_flags_for_srcs(
+    SRCS "${MARLIN_PERMUTE_SRC}"
+    CUDA_ARCHS "${MOE_PERMUTE_ARCHS}")
+
+  list(APPEND VLLM_MOE_EXT_SRC "${MOE_PERMUTE_SRC}")
+endif()
 message(STATUS "Enabling moe extension.")
 define_gpu_extension_target(
   _moe_C
@@ -691,6 +744,8 @@ define_gpu_extension_target(
   SOURCES ${VLLM_MOE_EXT_SRC}
   COMPILE_FLAGS ${VLLM_GPU_FLAGS}
   ARCHITECTURES ${VLLM_GPU_ARCHES}
+  INCLUDE_DIRECTORIES ${CUTLASS_INCLUDE_DIR}
+  INCLUDE_DIRECTORIES ${CUTLASS_TOOLS_UTIL_INCLUDE_DIR}
   USE_SABI 3
   WITH_SOABI)
 

benchmarks/benchmark_dataset.py

@@ -315,13 +315,15 @@ def sample(
         )
 
         vocab_size = tokenizer.vocab_size
+        num_special_tokens = tokenizer.num_special_tokens_to_add()
+        real_input_len = input_len - num_special_tokens
 
         prefix_token_ids = (np.random.randint(
             0, vocab_size, size=prefix_len).tolist() if prefix_len > 0 else [])
 
         # New sampling logic: [X * (1 - b), X * (1 + b)]
-        input_low = int(input_len * (1 - range_ratio))
-        input_high = int(input_len * (1 + range_ratio))
+        input_low = int(real_input_len * (1 - range_ratio))
+        input_high = int(real_input_len * (1 + range_ratio))
         output_low = int(output_len * (1 - range_ratio))
         output_high = int(output_len * (1 + range_ratio))
 
@@ -344,6 +346,17 @@ def sample(
                          vocab_size).tolist()
             token_sequence = prefix_token_ids + inner_seq
             prompt = tokenizer.decode(token_sequence)
+            # After decoding the prompt we have to encode and decode it again.
+            # This is done because in some cases N consecutive tokens
+            # give a string tokenized into != N number of tokens.
+            # For example for GPT2Tokenizer:
+            # [6880, 6881] -> ['Ġcalls', 'here'] ->
+            # [1650, 939, 486] -> ['Ġcall', 'sh', 'ere']
+            # To avoid uncontrolled change of the prompt length,
+            # the encoded sequence is truncated before being decode again.
+            re_encoded_sequence = tokenizer.encode(
+                prompt, add_special_tokens=False)[:input_lens[i]]
+            prompt = tokenizer.decode(re_encoded_sequence)
             total_input_len = prefix_len + int(input_lens[i])
             requests.append(
                 SampleRequest(

benchmarks/benchmark_serving_structured_output.py

@@ -414,7 +414,6 @@ async def benchmark(
     ignore_eos: bool,
     max_concurrency: Optional[int],
     structured_output_ratio: float,
-    structured_output_backend: str,
     goodput_config_dict: Optional[dict[str, float]] = None,
 ):
     if backend in ASYNC_REQUEST_FUNCS:
@@ -426,8 +425,6 @@ def prepare_extra_body(request) -> dict:
         extra_body = {}
         # Add the schema to the extra_body
         extra_body[request.structure_type] = request.schema
-        # Add the specific structured_output_backend
-        extra_body["guided_decoding_backend"] = structured_output_backend
         return extra_body
 
     print("Starting initial single prompt test run...")
@@ -785,7 +782,6 @@ def main(args: argparse.Namespace):
             ignore_eos=args.ignore_eos,
             max_concurrency=args.max_concurrency,
             structured_output_ratio=args.structured_output_ratio,
-            structured_output_backend=args.structured_output_backend,
             goodput_config_dict=goodput_config_dict,
         ))
 
@@ -1000,14 +996,6 @@ def main(args: argparse.Namespace):
                         type=float,
                         default=1.0,
                         help="Ratio of Structured Outputs requests")
-    parser.add_argument("--structured-output-backend",
-                        type=str,
-                        choices=[
-                            "outlines", "lm-format-enforcer", "xgrammar",
-                            "guidance", "auto"
-                        ],
-                        default="auto",
-                        help="Backend to use for structured outputs")
 
     args = parser.parse_args()
     main(args)

benchmarks/kernels/benchmark_grouped_gemm_cutlass.py

@@ -90,7 +90,8 @@ def bench_run(results: list[benchmark.Measurement], model: str,
 
     score = torch.randn((m, num_experts), device="cuda", dtype=dtype)
 
-    topk_weights, topk_ids = fused_topk(a, score, topk, renormalize=False)
+    topk_weights, topk_ids, token_expert_indices = fused_topk(
+        a, score, topk, renormalize=False)
 
     def run_triton_moe(a: torch.Tensor, w1: torch.Tensor, w2: torch.Tensor,
                        topk_weights: torch.Tensor, topk_ids: torch.Tensor,

benchmarks/kernels/benchmark_moe.py

@@ -10,12 +10,12 @@
 
 import ray
 import torch
-import triton
 from ray.experimental.tqdm_ray import tqdm
 from transformers import AutoConfig
 
 from vllm.model_executor.layers.fused_moe.fused_moe import *
 from vllm.platforms import current_platform
+from vllm.triton_utils import triton
 from vllm.utils import FlexibleArgumentParser
 
 FP8_DTYPE = current_platform.fp8_dtype()
@@ -115,8 +115,8 @@ def run():
         from vllm.model_executor.layers.fused_moe import override_config
         with override_config(config):
             if use_deep_gemm:
-                topk_weights, topk_ids = fused_topk(x, input_gating, topk,
-                                                    False)
+                topk_weights, topk_ids, token_expert_indices = fused_topk(
+                    x, input_gating, topk, False)
                 return fused_experts(
                     x,
                     w1,
@@ -442,8 +442,14 @@ def tune(
                                                    hidden_size, search_space,
                                                    is_fp16, topk)
 
-        with torch.cuda.device(self.device_id) if current_platform.is_rocm(
-        ) else nullcontext():
+        need_device_guard = False
+        if current_platform.is_rocm():
+            visible_device = os.environ.get("ROCR_VISIBLE_DEVICES", None)
+            if visible_device != f"{self.device_id}":
+                need_device_guard = True
+
+        with torch.cuda.device(
+                self.device_id) if need_device_guard else nullcontext():
             for config in tqdm(search_space):
                 try:
                     kernel_time = benchmark_config(
@@ -578,6 +584,15 @@ def main(args: argparse.Namespace):
 
     use_deep_gemm = bool(args.use_deep_gemm)
 
+    if current_platform.is_rocm() and "HIP_VISIBLE_DEVICES" in os.environ:
+        # Ray will set ROCR_VISIBLE_DEVICES for device visibility
+        logger.warning(
+            "Ray uses ROCR_VISIBLE_DEVICES to control device accessibility."
+            "Replacing HIP_VISIBLE_DEVICES with ROCR_VISIBLE_DEVICES.")
+        val = os.environ["HIP_VISIBLE_DEVICES"]
+        os.environ["ROCR_VISIBLE_DEVICES"] = val
+        del os.environ["HIP_VISIBLE_DEVICES"]
+
     ray.init()
     num_gpus = int(ray.available_resources()["GPU"])
     workers = [BenchmarkWorker.remote(args.seed) for _ in range(num_gpus)]