[ROCm][CI] Pin TorchCodec to v0.10.0 for ROCm compatibility (#34447)

Signed-off-by: Andreas Karatzas <akaratza@amd.com>
(cherry picked from commit 4c078fa546)
(cherry picked from commit a976961fb77d38129abf69edd4952101731f2421)

.buildkite/image_build/image_build.yaml

@@ -3,7 +3,6 @@ steps:
   - label: ":docker: Build image"
     key: image-build
     depends_on: []
-    timeout_in_minutes: 600
     commands:
     - if [[ "$BUILDKITE_BRANCH" != "main" ]]; then .buildkite/image_build/image_build.sh $REGISTRY $REPO $BUILDKITE_COMMIT $BRANCH $VLLM_USE_PRECOMPILED $VLLM_MERGE_BASE_COMMIT $IMAGE_TAG; fi
     - if [[ "$BUILDKITE_BRANCH" == "main" ]]; then .buildkite/image_build/image_build.sh $REGISTRY $REPO $BUILDKITE_COMMIT $BRANCH $VLLM_USE_PRECOMPILED $VLLM_MERGE_BASE_COMMIT $IMAGE_TAG $IMAGE_TAG_LATEST; fi
@@ -42,7 +41,7 @@ steps:
           limit: 2
         - exit_status: -10  # Agent was lost
           limit: 2
-
+  
   - label: ":docker: Build CPU arm64 image"
     key: cpu-arm64-image-build
     depends_on: []

.buildkite/image_build/image_build_cpu_arm64.sh

@@ -14,7 +14,7 @@ BUILDKITE_COMMIT=$3
 aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin $REGISTRY
 
 # skip build if image already exists
-if [[ -z $(docker manifest inspect $REGISTRY/$REPO:$BUILDKITE_COMMIT-cpu) ]]; then
+if [[ -z $(docker manifest inspect "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-arm64-cpu) ]]; then
   echo "Image not found, proceeding with build..."
 else
   echo "Image found"
@@ -24,10 +24,10 @@ fi
 # build
 docker build --file docker/Dockerfile.cpu \
   --build-arg max_jobs=16 \
-  --build-arg buildkite_commit=$BUILDKITE_COMMIT \
-  --tag $REGISTRY/$REPO:$BUILDKITE_COMMIT-cpu \
+  --build-arg buildkite_commit="$BUILDKITE_COMMIT" \
+  --tag "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-arm64-cpu \
   --target vllm-test \
   --progress plain .
 
 # push
-docker push $REGISTRY/$REPO:$BUILDKITE_COMMIT-cpu
+docker push "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-arm64-cpu

.buildkite/release-pipeline.yaml

@@ -248,8 +248,8 @@ steps:
           DOCKER_BUILDKIT: "1"
           DOCKERHUB_USERNAME: "vllmbot"
 
-  - group: "Publish wheels"
-    key: "publish-wheels"
+  - group: "Publish release artifacts"
+    key: "publish-release-artifacts"
     steps:
       - block: "Confirm update release wheels to PyPI (experimental, use with caution)?"
         key: block-upload-release-wheels
@@ -265,6 +265,27 @@ steps:
           queue: small_cpu_queue_postmerge
         commands:
           - "bash .buildkite/scripts/upload-release-wheels-pypi.sh"
+      
+      - block: "Confirm update release images to DockerHub"
+        key: block-update-release-images-dockerhub
+        depends_on:
+          - input-release-version
+          - annotate-release-workflow
+      
+      - label: "Publish release images to DockerHub"
+        depends_on:
+          - block-update-release-images-dockerhub
+        agents:
+          queue: small_cpu_queue_postmerge
+        commands:
+          - "bash .buildkite/scripts/push-release-images-dockerhub.sh"
+        plugins:
+          - docker-login#v3.0.0:
+              username: vllmbot
+              password-env: DOCKERHUB_TOKEN
+        env:
+          DOCKER_BUILDKIT: "1"
+          DOCKERHUB_USERNAME: "vllmbot"
 
   # =============================================================================
   # ROCm Release Pipeline (x86_64 only)

.buildkite/scripts/push-release-images-dockerhub.sh

@@ -0,0 +1,98 @@
+#!/bin/bash
+
+set -ex
+
+RELEASE_VERSION=$(buildkite-agent meta-data get release-version 2>/dev/null | sed 's/^v//')
+if [ -z "${RELEASE_VERSION}" ]; then
+    echo "RELEASE_VERSION is not set"
+    exit 1
+fi
+
+aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7
+
+# Download images:
+
+docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-x86_64
+docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-aarch64
+docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-x86_64-cu130
+docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-aarch64-cu130
+docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-rocm-base
+docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-rocm
+docker pull public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:v${RELEASE_VERSION}
+docker pull public.ecr.aws/q9t5s3a7/vllm-arm64-cpu-release-repo:v${RELEASE_VERSION}
+
+# Tag and push images:
+
+## CUDA
+
+docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-x86_64 vllm/vllm-openai:x86_64
+docker tag vllm/vllm-openai:x86_64 vllm/vllm-openai:latest-x86_64
+docker tag vllm/vllm-openai:x86_64 vllm/vllm-openai:v${RELEASE_VERSION}-x86_64
+docker push vllm/vllm-openai:latest-x86_64
+docker push vllm/vllm-openai:v${RELEASE_VERSION}-x86_64
+
+docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-x86_64-cu130 vllm/vllm-openai:x86_64-cu130
+docker tag vllm/vllm-openai:x86_64-cu130 vllm/vllm-openai:latest-x86_64-cu130
+docker tag vllm/vllm-openai:x86_64-cu130 vllm/vllm-openai:v${RELEASE_VERSION}-x86_64-cu130
+docker push vllm/vllm-openai:latest-x86_64-cu130
+docker push vllm/vllm-openai:v${RELEASE_VERSION}-x86_64-cu130
+
+docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-aarch64 vllm/vllm-openai:aarch64
+docker tag vllm/vllm-openai:aarch64 vllm/vllm-openai:latest-aarch64
+docker tag vllm/vllm-openai:aarch64 vllm/vllm-openai:v${RELEASE_VERSION}-aarch64
+docker push vllm/vllm-openai:latest-aarch64
+docker push vllm/vllm-openai:v${RELEASE_VERSION}-aarch64
+
+docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-aarch64-cu130 vllm/vllm-openai:aarch64-cu130
+docker tag vllm/vllm-openai:aarch64-cu130 vllm/vllm-openai:latest-aarch64-cu130
+docker tag vllm/vllm-openai:aarch64-cu130 vllm/vllm-openai:v${RELEASE_VERSION}-aarch64-cu130
+docker push vllm/vllm-openai:latest-aarch64-cu130
+docker push vllm/vllm-openai:v${RELEASE_VERSION}-aarch64-cu130
+
+## ROCm
+
+docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-rocm vllm/vllm-openai-rocm:${BUILDKITE_COMMIT}
+docker tag vllm/vllm-openai-rocm:${BUILDKITE_COMMIT} vllm/vllm-openai-rocm:latest
+docker tag vllm/vllm-openai-rocm:${BUILDKITE_COMMIT} vllm/vllm-openai-rocm:v${RELEASE_VERSION}
+docker push vllm/vllm-openai-rocm:latest
+docker push vllm/vllm-openai-rocm:v${RELEASE_VERSION}
+
+docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-rocm-base vllm/vllm-openai-rocm:${BUILDKITE_COMMIT}-base
+docker tag vllm/vllm-openai-rocm:${BUILDKITE_COMMIT}-base vllm/vllm-openai-rocm:latest-base
+docker tag vllm/vllm-openai-rocm:${BUILDKITE_COMMIT}-base vllm/vllm-openai-rocm:v${RELEASE_VERSION}-base
+docker push vllm/vllm-openai-rocm:latest-base
+docker push vllm/vllm-openai-rocm:v${RELEASE_VERSION}-base
+
+## CPU
+
+docker tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:v${RELEASE_VERSION} vllm/vllm-openai-cpu:x86_64
+docker tag vllm/vllm-openai-cpu:x86_64 vllm/vllm-openai-cpu:latest-x86_64
+docker tag vllm/vllm-openai-cpu:x86_64 vllm/vllm-openai-cpu:v${RELEASE_VERSION}-x86_64
+docker push vllm/vllm-openai-cpu:latest-x86_64
+docker push vllm/vllm-openai-cpu:v${RELEASE_VERSION}-x86_64
+
+docker tag public.ecr.aws/q9t5s3a7/vllm-arm64-cpu-release-repo:v${RELEASE_VERSION} vllm/vllm-openai-cpu:arm64
+docker tag vllm/vllm-openai-cpu:arm64 vllm/vllm-openai-cpu:latest-arm64
+docker tag vllm/vllm-openai-cpu:arm64 vllm/vllm-openai-cpu:v${RELEASE_VERSION}-arm64
+docker push vllm/vllm-openai-cpu:latest-arm64
+docker push vllm/vllm-openai-cpu:v${RELEASE_VERSION}-arm64
+
+# Create multi-arch manifest:
+
+docker manifest rm vllm/vllm-openai:latest
+docker manifest create vllm/vllm-openai:latest vllm/vllm-openai:latest-x86_64 vllm/vllm-openai:latest-aarch64
+docker manifest create vllm/vllm-openai:v${RELEASE_VERSION} vllm/vllm-openai:v${RELEASE_VERSION}-x86_64 vllm/vllm-openai:v${RELEASE_VERSION}-aarch64
+docker manifest push vllm/vllm-openai:latest
+docker manifest push vllm/vllm-openai:v${RELEASE_VERSION}
+
+docker manifest rm vllm/vllm-openai:latest-cu130
+docker manifest create vllm/vllm-openai:latest-cu130 vllm/vllm-openai:latest-x86_64-cu130 vllm/vllm-openai:latest-aarch64-cu130
+docker manifest create vllm/vllm-openai:v${RELEASE_VERSION}-cu130 vllm/vllm-openai:v${RELEASE_VERSION}-x86_64-cu130 vllm/vllm-openai:v${RELEASE_VERSION}-aarch64-cu130
+docker manifest push vllm/vllm-openai:latest-cu130
+docker manifest push vllm/vllm-openai:v${RELEASE_VERSION}-cu130
+
+docker manifest rm vllm/vllm-openai-cpu:latest || true
+docker manifest create vllm/vllm-openai-cpu:latest vllm/vllm-openai-cpu:latest-x86_64 vllm/vllm-openai-cpu:latest-arm64
+docker manifest create vllm/vllm-openai-cpu:v${RELEASE_VERSION} vllm/vllm-openai-cpu:v${RELEASE_VERSION}-x86_64 vllm/vllm-openai-cpu:v${RELEASE_VERSION}-arm64
+docker manifest push vllm/vllm-openai-cpu:latest
+docker manifest push vllm/vllm-openai-cpu:v${RELEASE_VERSION}

CMakeLists.txt

@@ -56,8 +56,8 @@ endif()
 # requirements.txt files and should be kept consistent.  The ROCm torch
 # versions are derived from docker/Dockerfile.rocm
 #
-set(TORCH_SUPPORTED_VERSION_CUDA "2.10.0")
-set(TORCH_SUPPORTED_VERSION_ROCM "2.10.0")
+set(TORCH_SUPPORTED_VERSION_CUDA "2.9.1")
+set(TORCH_SUPPORTED_VERSION_ROCM "2.9.1")
 
 #
 # Try to find python package with an executable that exactly matches
@@ -433,7 +433,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
       list(APPEND VLLM_EXT_SRC ${MARLIN_TEMPLATE_BF16_KERNEL_SRC})
     endif()
 
-    if (MARLIN_SM75_ARCHS)
+    if (MARLIN_SM75_ARCHS) 
       file(GLOB MARLIN_TEMPLATE_SM75_KERNEL_SRC "csrc/quantization/marlin/sm75_kernel_*.cu")
       set_gencode_flags_for_srcs(
         SRCS "${MARLIN_TEMPLATE_SM75_KERNEL_SRC}"
@@ -445,7 +445,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
       list(APPEND VLLM_EXT_SRC ${MARLIN_TEMPLATE_SM75_KERNEL_SRC})
     endif()
 
-    if (MARLIN_FP8_ARCHS)
+    if (MARLIN_FP8_ARCHS) 
       file(GLOB MARLIN_TEMPLATE_FP8_KERNEL_SRC "csrc/quantization/marlin/sm89_kernel_*.cu")
       set_gencode_flags_for_srcs(
         SRCS "${MARLIN_TEMPLATE_FP8_KERNEL_SRC}"
@@ -1042,7 +1042,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
       list(APPEND VLLM_MOE_EXT_SRC ${MARLIN_MOE_SRC})
     endif()
 
-    if (MARLIN_MOE_SM75_ARCHS)
+    if (MARLIN_MOE_SM75_ARCHS) 
       file(GLOB MARLIN_MOE_SM75_SRC "csrc/moe/marlin_moe_wna16/sm75_kernel_*.cu")
       set_gencode_flags_for_srcs(
         SRCS "${MARLIN_MOE_SM75_SRC}"

benchmarks/kernels/benchmark_moe.py

@@ -686,6 +686,7 @@ def get_model_params(config):
         "DeepseekV2ForCausalLM",
         "DeepseekV3ForCausalLM",
         "DeepseekV32ForCausalLM",
+        "GlmMoeDsaForCausalLM",
         "Glm4MoeForCausalLM",
         "Glm4MoeLiteForCausalLM",
         "NemotronHForCausalLM",

cmake/external_projects/triton_kernels.cmake

@@ -1,9 +1,9 @@
 # Install OpenAI triton_kernels from https://github.com/triton-lang/triton/tree/main/python/triton_kernels
 
-set(DEFAULT_TRITON_KERNELS_TAG "v3.6.0")
+set(DEFAULT_TRITON_KERNELS_TAG "v3.5.0")
 
 # Set TRITON_KERNELS_SRC_DIR for use with local development with vLLM. We expect TRITON_KERNELS_SRC_DIR to
-# be directly set to the triton_kernels python directory.
+# be directly set to the triton_kernels python directory. 
 if (DEFINED ENV{TRITON_KERNELS_SRC_DIR})
   message(STATUS "[triton_kernels] Fetch from $ENV{TRITON_KERNELS_SRC_DIR}")
   FetchContent_Declare(
@@ -24,7 +24,7 @@ else()
   )
 endif()
 
-# Fetch content
+# Fetch content 
 FetchContent_MakeAvailable(triton_kernels)
 
 if (NOT triton_kernels_SOURCE_DIR)
@@ -47,7 +47,7 @@ install(CODE "file(MAKE_DIRECTORY \"\${CMAKE_INSTALL_PREFIX}/vllm/third_party/tr
 ## Copy .py files to install directory.
 install(DIRECTORY
         ${TRITON_KERNELS_PYTHON_DIR}
-        DESTINATION
+        DESTINATION 
         vllm/third_party/triton_kernels/
         COMPONENT triton_kernels
         FILES_MATCHING PATTERN "*.py")

pyproject.toml

@@ -6,7 +6,7 @@ requires = [
     "packaging>=24.2",
     "setuptools>=77.0.3,<81.0.0",
     "setuptools-scm>=8.0",
-    "torch == 2.10.0",
+    "torch == 2.9.1",
     "wheel",
     "jinja2",
     "grpcio-tools==1.78.0",

requirements/build.txt

@@ -4,7 +4,7 @@ ninja
 packaging>=24.2
 setuptools>=77.0.3,<81.0.0
 setuptools-scm>=8
-torch==2.10.0
+torch==2.9.1
 wheel
 jinja2>=3.1.6
 regex

requirements/cuda.txt

@@ -5,9 +5,9 @@ numba == 0.61.2 # Required for N-gram speculative decoding
 
 # Dependencies for NVIDIA GPUs
 ray[cgraph]>=2.48.0
-torch==2.10.0
-torchaudio==2.10.0
+torch==2.9.1
+torchaudio==2.9.1
 # These must be updated alongside torch
-torchvision==0.25.0 # Required for phi3v processor. See https://github.com/pytorch/vision?tab=readme-ov-file#installation for corresponding version
+torchvision==0.24.1 # Required for phi3v processor. See https://github.com/pytorch/vision?tab=readme-ov-file#installation for corresponding version
 # FlashInfer should be updated together with the Dockerfile
 flashinfer-python==0.6.3

requirements/nightly_torch_test.txt

@@ -43,5 +43,5 @@ tritonclient>=2.51.0
 numba == 0.61.2 # Required for N-gram speculative decoding
 numpy
 runai-model-streamer[s3,gcs]==0.15.3
-fastsafetensors>=0.1.10
+fastsafetensors>=0.2.2
 pydantic>=2.12 # 2.11 leads to error on python 3.13

requirements/rocm-build.txt

@@ -1,11 +1,12 @@
 # Common dependencies
 -r common.txt
 
---extra-index-url https://download.pytorch.org/whl/test/rocm7.0
-torch==2.10.0
-torchvision==0.25.0
-torchaudio==2.10.0
-triton==3.6.0
+--extra-index-url https://download.pytorch.org/whl/rocm6.4
+torch==2.9.1
+torchvision==0.24.1
+torchaudio==2.9.1
+
+triton==3.5.1
 cmake>=3.26.1,<4
 packaging>=24.2
 setuptools>=77.0.3,<80.0.0

requirements/rocm.txt

@@ -1,6 +1,11 @@
 # Common dependencies
 -r common.txt
 
+# The version of gRPC libraries should be consistent with each other
+grpcio==1.78.0
+grpcio-reflection==1.78.0
+grpcio-tools==1.78.0
+
 numba == 0.61.2 # Required for N-gram speculative decoding
 
 # Dependencies for AMD GPUs
@@ -14,5 +19,4 @@ setuptools>=77.0.3,<80.0.0
 setuptools-scm>=8
 runai-model-streamer[s3,gcs]==0.15.3
 conch-triton-kernels==1.2.1
-timm>=1.0.17
-grpcio-tools==1.78.0 # Should match `build.txt`
+timm>=1.0.17

requirements/test.in

@@ -24,10 +24,10 @@ sentence-transformers>=5.2.0 # required for embedding tests
 soundfile # required for audio tests
 jiwer # required for audio tests
 tblib # for pickling test exceptions
-timm >=1.0.17 # required for internvl and gemma3n-mm test
-torch==2.10.0
-torchaudio==2.10.0
-torchvision==0.25.0
+timm==1.0.17 # required for internvl and gemma3n-mm test
+torch==2.9.1
+torchaudio==2.9.1
+torchvision==0.24.1
 transformers_stream_generator # required for qwen-vl test
 matplotlib # required for qwen-vl test
 mistral_common[image,audio] >= 1.9.0 # required for voxtral test
@@ -48,12 +48,16 @@ buildkite-test-collector==0.1.9
 genai_perf>=0.0.8
 tritonclient>=2.51.0
 
-grpcio-tools==1.78.0 # Should match `build.txt`
+# The version of gRPC libraries should be consistent with each other
+grpcio==1.78.0
+grpcio-reflection==1.78.0
+grpcio-tools==1.78.0
+
 arctic-inference == 0.1.1 # Required for suffix decoding test
 numba == 0.61.2 # Required for N-gram speculative decoding
 numpy
 runai-model-streamer[s3,gcs]==0.15.3
-fastsafetensors>=0.1.10
+fastsafetensors>=0.2.2 # 0.2.2 contains important fixes for multi-GPU mem usage
 pydantic>=2.12 # 2.11 leads to error on python 3.13
 decord==0.6.0
 terratorch @ git+https://github.com/IBM/terratorch.git@1.1.rc3 # required for PrithviMAE test

requirements/test.txt

@@ -155,10 +155,6 @@ coverage==7.10.6
     # via pytest-cov
 cramjam==2.9.0
     # via fastparquet
-cuda-bindings==12.9.4
-    # via torch
-cuda-pathfinder==1.3.3
-    # via cuda-bindings
 cupy-cuda12x==13.6.0
     # via ray
 cycler==0.12.1
@@ -224,7 +220,7 @@ fastparquet==2024.11.0
     # via genai-perf
 fastrlock==0.8.2
     # via cupy-cuda12x
-fastsafetensors==0.1.10
+fastsafetensors==0.2.2
     # via -r requirements/test.in
 filelock==3.16.1
     # via
@@ -309,8 +305,13 @@ greenlet==3.2.3
     # via sqlalchemy
 grpcio==1.78.0
     # via
+    #   -r requirements/test.in
+    #   grpcio-reflection
     #   grpcio-tools
     #   ray
+    #   tensorboard
+grpcio-reflection==1.78.0
+    # via -r requirements/test.in
 grpcio-tools==1.78.0
     # via -r requirements/test.in
 gunicorn==23.0.0
@@ -635,7 +636,7 @@ nvidia-nvjitlink-cu12==12.9.86
     #   nvidia-cusolver-cu12
     #   nvidia-cusparse-cu12
     #   torch
-nvidia-nvshmem-cu12==3.4.5
+nvidia-nvshmem-cu12==3.3.20
     # via torch
 nvidia-nvtx-cu12==12.9.79
     # via torch
@@ -785,6 +786,7 @@ protobuf==6.33.2
     # via
     #   google-api-core
     #   googleapis-common-protos
+    #   grpcio-reflection
     #   grpcio-tools
     #   mlflow-skinny
     #   opentelemetry-proto
@@ -1167,14 +1169,13 @@ tomli==2.2.1
     # via schemathesis
 tomli-w==1.2.0
     # via schemathesis
-torch==2.10.0+cu129
+torch==2.9.1+cu129
     # via
     #   -r requirements/test.in
     #   accelerate
     #   bitsandbytes
     #   efficientnet-pytorch
     #   encodec
-    #   fastsafetensors
     #   kornia
     #   lightly
     #   lightning
@@ -1196,7 +1197,7 @@ torch==2.10.0+cu129
     #   torchvision
     #   vector-quantize-pytorch
     #   vocos
-torchaudio==2.10.0+cu129
+torchaudio==2.9.1+cu129
     # via
     #   -r requirements/test.in
     #   encodec
@@ -1209,7 +1210,7 @@ torchmetrics==1.7.4
     #   pytorch-lightning
     #   terratorch
     #   torchgeo
-torchvision==0.25.0+cu129
+torchvision==0.24.1+cu129
     # via
     #   -r requirements/test.in
     #   lightly
@@ -1251,7 +1252,7 @@ transformers==4.57.5
     #   transformers-stream-generator
 transformers-stream-generator==0.0.5
     # via -r requirements/test.in
-triton==3.6.0
+triton==3.5.1
     # via torch
 tritonclient==2.64.0
     # via -r requirements/test.in

setup.py

@@ -1035,7 +1035,7 @@ setup(
     extras_require={
         "bench": ["pandas", "matplotlib", "seaborn", "datasets", "scipy"],
         "tensorizer": ["tensorizer==2.10.1"],
-        "fastsafetensors": ["fastsafetensors >= 0.1.10"],
+        "fastsafetensors": ["fastsafetensors >= 0.2.2"],
         "runai": ["runai-model-streamer[s3,gcs] >= 0.15.3"],
         "audio": [
             "librosa",

tests/compile/test_aot_compile.py

@@ -90,7 +90,9 @@ def use_vllm_config(vllm_config: VllmConfig):
         yield
 
 
-@pytest.mark.skipif(not is_torch_equal_or_newer("2.10.0"), reason="requires torch 2.10")
+@pytest.mark.skipif(
+    not is_torch_equal_or_newer("2.10.0.dev"), reason="requires torch 2.10"
+)
 def test_no_dynamo_cache_entry(monkeypatch: pytest.MonkeyPatch):
     with monkeypatch.context() as m:
         vllm_config = make_vllm_config()
@@ -114,7 +116,9 @@ def test_no_dynamo_cache_entry(monkeypatch: pytest.MonkeyPatch):
             assert torch.allclose(actual, expected)
 
 
-@pytest.mark.skipif(not is_torch_equal_or_newer("2.10.0"), reason="requires torch 2.10")
+@pytest.mark.skipif(
+    not is_torch_equal_or_newer("2.10.0.dev"), reason="requires torch 2.10"
+)
 def test_force_aot_load(monkeypatch: pytest.MonkeyPatch):
     with tempfile.TemporaryDirectory() as tmpdirname, monkeypatch.context() as m:
         args = (torch.randn(10, 10),)
@@ -128,7 +132,9 @@ def test_force_aot_load(monkeypatch: pytest.MonkeyPatch):
             CompiledMod(vllm_config=vllm_config)(*args)
 
 
-@pytest.mark.skipif(not is_torch_equal_or_newer("2.10.0"), reason="requires torch 2.10")
+@pytest.mark.skipif(
+    not is_torch_equal_or_newer("2.10.0.dev"), reason="requires torch 2.10"
+)
 def test_save_and_load(monkeypatch: pytest.MonkeyPatch):
     with monkeypatch.context() as m:
         args = (torch.randn(10, 10),)
@@ -156,7 +162,9 @@ def test_save_and_load(monkeypatch: pytest.MonkeyPatch):
             assert torch.allclose(ret, expected)
 
 
-@pytest.mark.skipif(not is_torch_equal_or_newer("2.10.0"), reason="requires torch 2.10")
+@pytest.mark.skipif(
+    not is_torch_equal_or_newer("2.10.0.dev"), reason="requires torch 2.10"
+)
 def test_cache_load_returns_tuple_consistency(monkeypatch: pytest.MonkeyPatch):
     """
     Test that cache loading correctly handles the returns_tuple logic.
@@ -215,7 +223,9 @@ def test_cache_load_returns_tuple_consistency(monkeypatch: pytest.MonkeyPatch):
             )
 
 
-@pytest.mark.skipif(not is_torch_equal_or_newer("2.10.0"), reason="requires torch 2.10")
+@pytest.mark.skipif(
+    not is_torch_equal_or_newer("2.10.0.dev"), reason="requires torch 2.10"
+)
 def test_cache_load_returns_tuple_consistency_tuple_output(
     monkeypatch: pytest.MonkeyPatch,
 ):
@@ -284,7 +294,9 @@ def test_cache_load_returns_tuple_consistency_tuple_output(
             )
 
 
-@pytest.mark.skipif(not is_torch_equal_or_newer("2.10.0"), reason="requires torch 2.10")
+@pytest.mark.skipif(
+    not is_torch_equal_or_newer("2.10.0.dev"), reason="requires torch 2.10"
+)
 def test_shape_env(monkeypatch: pytest.MonkeyPatch):
     """
     Test that the shape environment is correctly serialized and preserved
@@ -321,7 +333,9 @@ def test_shape_env(monkeypatch: pytest.MonkeyPatch):
                 assert guards_string == " - s77 <= 42\n - Eq(Mod(s77, 2), 0)"
 
 
-@pytest.mark.skipif(not is_torch_equal_or_newer("2.10.0"), reason="requires torch 2.10")
+@pytest.mark.skipif(
+    not is_torch_equal_or_newer("2.10.0.dev"), reason="requires torch 2.10"
+)
 def test_partition_wrapper_applied_on_aot_load(
     monkeypatch: pytest.MonkeyPatch, vllm_tmp_cache: Path, mocker
 ):
@@ -412,7 +426,9 @@ def test_partition_wrapper_applied_on_aot_load(
         )
 
 
-@pytest.mark.skipif(not is_torch_equal_or_newer("2.10.0"), reason="requires torch 2.10")
+@pytest.mark.skipif(
+    not is_torch_equal_or_newer("2.10.0.dev"), reason="requires torch 2.10"
+)
 @create_new_process_for_each_test("spawn")
 def test_gpt2_cache_hit(monkeypatch: pytest.MonkeyPatch):
     """
@@ -476,7 +492,9 @@ def test_gpt2_cache_hit(monkeypatch: pytest.MonkeyPatch):
         symbolic_shapes_module.make_symbol = original_make_symbol
 
 
-@pytest.mark.skipif(not is_torch_equal_or_newer("2.10.0"), reason="requires torch 2.10")
+@pytest.mark.skipif(
+    not is_torch_equal_or_newer("2.10.0.dev"), reason="requires torch 2.10"
+)
 class TestStandaloneCompiledArtifacts:
     def test_init(self):
         cache = StandaloneCompiledArtifacts()
@@ -650,7 +668,9 @@ class TestStandaloneCompiledArtifacts:
         assert len(restored_cache.loaded_submodule_store) == 0
 
 
-@pytest.mark.skipif(not is_torch_equal_or_newer("2.10.0"), reason="requires torch 2.10")
+@pytest.mark.skipif(
+    not is_torch_equal_or_newer("2.10.0.dev"), reason="requires torch 2.10"
+)
 class TestStandaloneCompiledArtifactsIntegration:
     def test_add_pickle_unpickle(self):
         cache = StandaloneCompiledArtifacts()

tests/compile/test_dynamic_shapes_compilation.py

@@ -39,7 +39,9 @@ def get_test_models():
 @pytest.mark.parametrize("use_aot_compile", ["0", "1"])
 @pytest.mark.parametrize("use_bytecode_hook", [True, False])
 @pytest.mark.parametrize("evaluate_guards", [False, True])
-@pytest.mark.skipif(not is_torch_equal_or_newer("2.10.0"), reason="requires torch 2.10")
+@pytest.mark.skipif(
+    not is_torch_equal_or_newer("2.10.0.dev"), reason="requires torch 2.10"
+)
 def test_dynamic_shapes_compilation(
     monkeypatch,
     model_name,

tests/entrypoints/openai/test_realtime_validation.py

@@ -129,5 +129,5 @@ async def test_multi_chunk_streaming(
                 " First words I spoke in the original phonograph."
                 " A little piece of practical poetry. Mary had a little lamb,"
                 " it sleeps with quite a flow, and everywhere that Mary went,"
-                " the lamb was sure to go"
+                " the lamb was sure to go."
             )

tests/kernels/moe/test_shared_fused_moe_routed_transform.py

@@ -14,7 +14,6 @@ import torch.nn as nn
 from vllm.config import VllmConfig, set_current_vllm_config
 from vllm.forward_context import set_forward_context
 from vllm.model_executor.layers.fused_moe.shared_fused_moe import SharedFusedMoE
-from vllm.utils.torch_utils import is_torch_equal_or_newer
 
 
 class SimpleLinear(nn.Module):
@@ -61,10 +60,6 @@ def setup_cuda():
 @pytest.mark.parametrize("num_tokens", [1, 32])
 @pytest.mark.parametrize("hidden_size,latent_size", [(256, 128), (128, 64)])
 @pytest.mark.parametrize("dtype", [torch.bfloat16])
-@pytest.mark.skipif(
-    is_torch_equal_or_newer("2.10.0"),
-    reason="Test fails with PyTorch 2.10.0 see: https://github.com/vllm-project/vllm/issues/33995",
-)
 def test_routed_input_transform_inside_vs_outside(
     num_tokens: int,
     hidden_size: int,

tests/models/registry.py

@@ -275,6 +275,9 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
         "zai-org/GLM-4.7-Flash",
         min_transformers_version="5.0.0",
     ),
+    "GlmMoeDsaForCausalLM": _HfExamplesInfo(
+        "zai-org/GLM-5", min_transformers_version="5.0.1", is_available_online=False
+    ),
     "GPT2LMHeadModel": _HfExamplesInfo("openai-community/gpt2", {"alias": "gpt2"}),
     "GPTBigCodeForCausalLM": _HfExamplesInfo(
         "bigcode/starcoder",

tests/models/test_initialization.py

@@ -97,7 +97,7 @@ def can_initialize(
             "pickle error when loading `transformers.models.auto.CONFIG_MAPPING`"
         )
 
-    if model_arch == "DeepseekV32ForCausalLM":
+    if model_arch in ["DeepseekV32ForCausalLM", "GlmMoeDsaForCausalLM"]:
         from vllm.platforms import current_platform
 
         capability = current_platform.get_device_capability()

tools/install_torchcodec_rocm.sh

@@ -7,7 +7,8 @@
 set -e
 
 TORCHCODEC_REPO="${TORCHCODEC_REPO:-https://github.com/pytorch/torchcodec.git}"
-TORCHCODEC_BRANCH="${TORCHCODEC_BRANCH:-main}"
+# Pin to a specific release for reproducibility; update as needed.
+TORCHCODEC_BRANCH="${TORCHCODEC_BRANCH:-v0.10.0}"
 
 echo "=== TorchCodec Installation Script ==="
 

vllm/compilation/compiler_interface.py

@@ -233,7 +233,7 @@ class InductorStandaloneAdaptor(CompilerInterface):
 
         from torch._inductor import standalone_compile
 
-        supports_aot = is_torch_equal_or_newer("2.10.0")
+        supports_aot = is_torch_equal_or_newer("2.10.0.dev")
 
         if not supports_aot and envs.VLLM_USE_MEGA_AOT_ARTIFACT:
             logger.error(

vllm/compilation/decorators.py

@@ -333,7 +333,7 @@ def _support_torch_compile(
     ) -> None:
         def mark_dynamic(arg: torch.Tensor, dims: list[int]) -> None:
             if ds_type == DynamicShapesType.UNBACKED:
-                if is_torch_equal_or_newer("2.10.0"):
+                if is_torch_equal_or_newer("2.10.0.dev"):
                     for dim in dims:
                         torch._dynamo.decorators.mark_unbacked(
                             arg, dim, hint_override=arg.size()[dim]
@@ -373,7 +373,7 @@ def _support_torch_compile(
                     if isinstance(arg, torch.Tensor):
                         # In case dims is specified with negative indexing
                         dims = [arg.ndim + dim if dim < 0 else dim for dim in dims]
-                        if is_torch_equal_or_newer("2.10.0"):
+                        if is_torch_equal_or_newer("2.10.0.dev"):
                             for dim in dims:
                                 torch._dynamo.decorators.mark_unbacked(
                                     arg, dim, hint_override=arg.size()[dim]
@@ -525,9 +525,9 @@ def _support_torch_compile(
             fx_config_patches["backed_size_oblivious"] = True
 
         # Prepare inductor config patches
-        # assume_32bit_indexing is only available in torch 2.10.0+
+        # assume_32bit_indexing is only available in torch 2.10.0.dev+
         inductor_config_patches = {}
-        if is_torch_equal_or_newer("2.10.0"):
+        if is_torch_equal_or_newer("2.10.0.dev"):
             inductor_config_patches["assume_32bit_indexing"] = (
                 self.compilation_config.dynamic_shapes_config.assume_32_bit_indexing
             )

vllm/config/speculative.py

@@ -181,7 +181,7 @@ class SpeculativeConfig:
     @staticmethod
     def hf_config_override(hf_config: PretrainedConfig) -> PretrainedConfig:
         initial_architecture = hf_config.architectures[0]
-        if hf_config.model_type in ("deepseek_v3", "deepseek_v32"):
+        if hf_config.model_type in ("deepseek_v3", "deepseek_v32", "glm_moe_dsa"):
             hf_config.model_type = "deepseek_mtp"
         if hf_config.model_type == "deepseek_mtp":
             n_predict = getattr(hf_config, "num_nextn_predict_layers", None)

vllm/entrypoints/openai/realtime/connection.py

@@ -48,7 +48,6 @@ class RealtimeConnection:
         self.generation_task: asyncio.Task | None = None
 
         self._is_connected = False
-        self._is_input_finished = False
         self._is_model_validated = False
 
         self._max_audio_filesize_mb = envs.VLLM_MAX_AUDIO_CLIP_FILESIZE_MB
@@ -145,7 +144,7 @@ class RealtimeConnection:
             commit_event = InputAudioBufferCommit(**event)
             # final signals that the audio is finished
             if commit_event.final:
-                self._is_input_finished = True
+                self.audio_queue.put_nowait(None)
             else:
                 await self.start_generation()
         else:
@@ -239,11 +238,6 @@ class RealtimeConnection:
                     # finish because websocket connection was killed
                     break
 
-                if self.audio_queue.empty() and self._is_input_finished:
-                    # finish because client signals that audio input
-                    # is finished
-                    break
-
             usage = UsageInfo(
                 prompt_tokens=prompt_token_ids_len,
                 completion_tokens=completion_tokens_len,

vllm/envs.py

@@ -271,7 +271,7 @@ def use_aot_compile() -> bool:
 
     default_value = (
         "1"
-        if is_torch_equal_or_newer("2.11.0.dev") and not disable_compile_cache()
+        if is_torch_equal_or_newer("2.10.0.dev") and not disable_compile_cache()
         else "0"
     )
 

vllm/model_executor/layers/batch_invariant.py

@@ -974,7 +974,7 @@ def enable_batch_invariant_mode():
     )
 
     reduced_precision_val = (
-        (False, False) if is_torch_equal_or_newer("2.10.0") else False
+        (False, False) if is_torch_equal_or_newer("2.10.0.dev") else False
     )
     torch.backends.cuda.matmul.allow_fp16_reduced_precision_reduction = (
         reduced_precision_val

vllm/model_executor/layers/fused_moe/gpt_oss_triton_kernels_moe.py

@@ -19,42 +19,17 @@ from vllm.model_executor.layers.fused_moe.utils import _resize_cache
 from vllm.model_executor.layers.quantization.utils.quant_utils import (
     QuantKey,
 )
-from vllm.platforms import current_platform
 from vllm.triton_utils import tl, triton
 from vllm.utils.import_utils import has_triton_kernels
 
 logger = init_logger(__name__)
 
-use_legacy_triton_kernels = False
-
 if has_triton_kernels():
     try:
         import triton_kernels.swiglu
-        from triton_kernels.matmul_ogs import (
-            FnSpecs,
-            FusedActivation,
-            GatherIndx,
-            RoutingData,
-            ScatterIndx,
-            matmul_ogs,
-        )
-        from triton_kernels.tensor import (
-            BIT,
-            Bitmatrix,
-        )
-        from triton_kernels.topk import topk
-
-        try:
-            from triton_kernels.tensor import (
-                SparseMatrix,
-                make_ragged_tensor_metadata,
-            )
-        except ImportError:
-            if current_platform.is_rocm():
-                logger.warning_once("Using legacy triton_kernels on ROCm")
-                use_legacy_triton_kernels = True
-            else:
-                raise
+        from triton_kernels.matmul_ogs import FnSpecs, FusedActivation, matmul_ogs
+        from triton_kernels.routing import RoutingData, routing, routing_from_bitmatrix
+        from triton_kernels.tensor import Bitmatrix
     except (AttributeError, ImportError) as e:
         logger.error(
             "Failed to import Triton kernels. Please make sure your triton "
@@ -103,68 +78,6 @@ def pack_bitmatrix(
         tl.store(bitmatrix_ptrs, y, mask=offsets_m[:, None] < n_rows)
 
 
-def legacy_routing_from_bitmatrix(
-    bitmatrix: "Bitmatrix",
-    expt_scal: torch.Tensor,
-    expt_indx: torch.Tensor,
-    n_expts_tot: int,
-    n_expts_act: int,
-) -> tuple["RoutingData", "GatherIndx", "ScatterIndx"]:
-    """
-    Replacement for the removed triton_kernels.routing.routing_from_bitmatrix.
-    Creates routing data from a bitmatrix representation.
-    """
-    if use_legacy_triton_kernels:
-        from triton_kernels.routing import routing_from_bitmatrix
-
-        return routing_from_bitmatrix(
-            bitmatrix, expt_scal, expt_indx, n_expts_tot, n_expts_act
-        )
-    sparse_logits = SparseMatrix(indx=expt_indx, vals=expt_scal, mask=bitmatrix)
-    dispatch_indx = sparse_logits.mask_metadata.row_sorted_indx
-    combine_indx = sparse_logits.mask_metadata.col_sorted_indx
-    ragged_batch_metadata = make_ragged_tensor_metadata(
-        sparse_logits.mask_metadata.col_sum,
-        dispatch_indx.shape[0],
-    )
-    gate_scal = sparse_logits.vals.flatten()[combine_indx]
-    routing_data = RoutingData(
-        gate_scal,
-        ragged_batch_metadata.block_sizes,
-        n_expts_tot,
-        n_expts_act,
-        ragged_batch_metadata,
-    )
-    gather_idx = GatherIndx(combine_indx, dispatch_indx)
-    scatter_idx = ScatterIndx(dispatch_indx, combine_indx)
-    return routing_data, gather_idx, scatter_idx
-
-
-def legacy_routing(
-    logits: torch.Tensor,
-    n_expts_act: int,
-    sm_first: bool = False,
-) -> tuple["RoutingData", "GatherIndx", "ScatterIndx"]:
-    """
-    Replacement for the removed triton_kernels.routing.routing function.
-    Computes routing data from gating logits.
-    """
-    if use_legacy_triton_kernels:
-        from triton_kernels.routing import routing
-
-        return routing(logits, n_expts_act, sm_first=sm_first)
-    if sm_first:
-        logits = torch.softmax(logits, dim=-1)
-    sparse_logits = topk(logits, n_expts_act, apply_softmax=not sm_first)
-    return legacy_routing_from_bitmatrix(
-        sparse_logits.mask,
-        sparse_logits.vals,
-        sparse_logits.indx,
-        logits.shape[-1],
-        n_expts_act,
-    )
-
-
 def triton_kernel_moe_forward(
     hidden_states: torch.Tensor,
     w1,  # Tensor or triton_kernels.Tensor
@@ -178,7 +91,7 @@ def triton_kernel_moe_forward(
     global_num_experts: int = -1,
     expert_map: torch.Tensor | None = None,
 ) -> torch.Tensor:
-    routing_data, gather_idx, scatter_idx = legacy_routing(
+    routing_data, gather_idx, scatter_idx = routing(
         gating_output, topk, sm_first=not renormalize
     )
 
@@ -254,22 +167,10 @@ def triton_kernel_fused_experts(
     )
     output_tensor = _resize_cache(output_tensor, (batch_dim, M, K))
 
-    act = (
-        FusedActivation(
-            FnSpecs(
-                "swiglu",
-                triton_kernels.swiglu.swiglu_fn,
-                ("alpha", "limit"),
-                reduction_n=2,
-            ),
-            (swiglu_alpha, swiglu_limit),
-        )
-        if not use_legacy_triton_kernels
-        else FusedActivation(
-            FnSpecs("swiglu", triton_kernels.swiglu.swiglu_fn, ("alpha", "limit")),
-            (swiglu_alpha, swiglu_limit),
-            2,
-        )
+    act = FusedActivation(
+        FnSpecs("swiglu", triton_kernels.swiglu.swiglu_fn, ("alpha", "limit")),
+        (swiglu_alpha, swiglu_limit),
+        2,
     )
     gammas = routing_data.gate_scal if routing_data else None
 
@@ -330,22 +231,13 @@ def make_routing_data(
 
     bitmatrix_shape = [n_rows, bm_cols * 32]
     bitmatrix_shape_max = [n_rows, None]
-    bitmatrix = (
-        Bitmatrix(
-            bitmatrix, dtype=BIT, shape=bitmatrix_shape, shape_max=bitmatrix_shape_max
-        )
-        if not use_legacy_triton_kernels
-        else Bitmatrix(
-            bitmatrix,
-            shape=bitmatrix_shape,
-            shape_max=bitmatrix_shape_max,
-            scratchpad=None,
-        )
+    bitmatrix = Bitmatrix(
+        bitmatrix, shape=bitmatrix_shape, shape_max=bitmatrix_shape_max, scratchpad=None
     )
 
     # matmul_ogs expects invalid topk_weights to be -1s
     topk_weights = torch.where(topk_ids == -1, -1.0, topk_weights)
-    routing_data, gather_indx, scatter_indx = legacy_routing_from_bitmatrix(
+    routing_data, gather_indx, scatter_indx = routing_from_bitmatrix(
         bitmatrix, topk_weights, topk_ids, num_local_experts, num_topk
     )
 

vllm/model_executor/models/deepseek_v2.py

@@ -836,7 +836,7 @@ class DeepseekV2MLAAttention(nn.Module):
                 qk_rope_head_dim,
                 max_position=max_position_embeddings,
                 rope_parameters=config.rope_parameters,
-                is_neox_style=True,
+                is_neox_style=not getattr(config, "indexer_rope_interleave", False),
             )
             self.indexer = Indexer(
                 vllm_config,
@@ -1499,6 +1499,10 @@ class DeepseekV3ForCausalLM(DeepseekV2ForCausalLM):
     pass
 
 
+class GlmMoeDsaForCausalLM(DeepseekV2ForCausalLM):
+    pass
+
+
 # Compatibility with
 # https://huggingface.co/deepseek-ai/DeepSeek-V3-Base/blob/main/configuration_deepseek.py
 def get_spec_layer_idx_from_weight_name(

vllm/model_executor/models/registry.py

@@ -114,6 +114,7 @@ _TEXT_GENERATION_MODELS = {
     "Glm4ForCausalLM": ("glm4", "Glm4ForCausalLM"),
     "Glm4MoeForCausalLM": ("glm4_moe", "Glm4MoeForCausalLM"),
     "Glm4MoeLiteForCausalLM": ("glm4_moe_lite", "Glm4MoeLiteForCausalLM"),
+    "GlmMoeDsaForCausalLM": ("deepseek_v2", "GlmMoeDsaForCausalLM"),
     "GptOssForCausalLM": ("gpt_oss", "GptOssForCausalLM"),
     "GPT2LMHeadModel": ("gpt2", "GPT2LMHeadModel"),
     "GPTBigCodeForCausalLM": ("gpt_bigcode", "GPTBigCodeForCausalLM"),

vllm/transformers_utils/model_arch_config_convertor.py

@@ -237,6 +237,7 @@ class ModelArchConfigConvertorBase:
             "deepseek_v3",
             "deepseek_v32",
             "deepseek_mtp",
+            "glm_moe_dsa",
             "glm4_moe_lite",
             "glm4_moe_lite_mtp",
             "kimi_k2",

vllm/v1/spec_decode/eagle.py

@@ -1503,6 +1503,24 @@ class SpecDecodeBaseProposer:
                 del self.model.lm_head
             self.model.lm_head = target_language_model.lm_head
 
+            # MTP models call compute_logits via shared_head.head (a
+            # ParallelLMHead inside each MTP layer), not self.model.lm_head.
+            # If the checkpoint omits a copy of the lm_head weights at the
+            # MTP layer path, shared_head.head stays uninitialised and
+            # produces NaN logits. Always share it explicitly.
+            inner = getattr(self.model, "model", None)
+            layers = getattr(inner, "layers", None) if inner else None
+            if layers is not None:
+                items = layers.values() if isinstance(layers, nn.ModuleDict) else layers
+                for layer in items:
+                    sh = getattr(layer, "shared_head", None)
+                    if sh is not None and hasattr(sh, "head"):
+                        del sh.head
+                        sh.head = target_language_model.lm_head
+                        logger.info(
+                            "Shared target model lm_head with MTP shared_head.head."
+                        )
+
     @torch.inference_mode()
     def dummy_run(
         self,