From 4f9a41ac94beb940249d9712a6dd2996cf37da61 Mon Sep 17 00:00:00 2001 From: Tyler Michael Smith Date: Thu, 17 Sep 2026 11:05:31 -0400 Subject: [PATCH] [CI/Build][Hardware][NVIDIA] Add Rubin CUDA 13.4 nightly images (#55953) Signed-off-by: Tyler Michael Smith Co-authored-by: OpenAI Codex Co-authored-by: OpenAI Codex Co-authored-by: Misha Goin Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com> --- .buildkite/release-pipeline.yaml | 92 +++++++++++++++++++ .buildkite/scripts/cleanup-nightly-builds.sh | 4 +- docker/Dockerfile | 5 + tests/v1/engine/test_output_processor.py | 3 +- .../kv_connector/unit/test_nixl_connector.py | 3 +- vllm/models/qwen4_exp/nvidia/ops/qsa.py | 3 +- 6 files changed, 105 insertions(+), 5 deletions(-) diff --git a/.buildkite/release-pipeline.yaml b/.buildkite/release-pipeline.yaml index 4af4eaf56e..2b90fef820 100644 --- a/.buildkite/release-pipeline.yaml +++ b/.buildkite/release-pipeline.yaml @@ -316,6 +316,66 @@ steps: - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu129" - 'bash .buildkite/scripts/annotate-build-artifact.sh "$$BUILDKITE_LABEL" "public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu129" release-images' + - label: "Build release image - x86_64 - Rubin CUDA 13.4" + depends_on: ~ + id: build-release-image-x86-rubin-cuda-13-4 + if: build.env("NIGHTLY") == "1" + agents: + queue: cpu_queue_release + commands: + - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7" + - | + DOCKER_BUILDKIT=1 docker buildx build --load \ + $(bash .buildkite/scripts/docker-build-metadata-args.sh cu134) \ + --platform linux/amd64 \ + --build-arg max_jobs=16 \ + --build-arg USE_SCCACHE=1 \ + --build-arg GIT_REPO_CHECK=1 \ + --build-arg CUDA_VERSION=13.4 \ + --build-arg torch_cuda_arch_list="${CUDA_ARCH_X86}" \ + --build-arg INSTALL_KV_CONNECTORS=true \ + --build-arg INSTALL_RUBIN_PRERELEASE=true \ + --build-arg TRITON_INSTALL_FROM_SOURCE_REPO=https://github.com/triton-lang/triton.git \ + --build-arg TRITON_INSTALL_FROM_SOURCE_REVISION=3f6e41132b5edf639bfb872ad73d4688765e08b8 \ + --build-arg RUN_WHEEL_CHECK=false \ + --build-arg BUILD_BASE_IMAGE=pytorch/manylinux2_28-builder:cuda13.4 \ + --build-arg FINAL_BASE_IMAGE=nvcr.io/nvidia/cuda-dl-base:26.08-cuda13.4-devel-ubuntu24.04 \ + --target vllm-openai \ + --progress plain \ + -f docker/Dockerfile . + - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu134" + - 'bash .buildkite/scripts/annotate-build-artifact.sh "$$BUILDKITE_LABEL" "public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu134" release-images' + + - label: "Build release image - aarch64 - Rubin CUDA 13.4" + depends_on: ~ + id: build-release-image-arm64-rubin-cuda-13-4 + if: build.env("NIGHTLY") == "1" + agents: + queue: arm64_cpu_queue_release + commands: + - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7" + - | + DOCKER_BUILDKIT=1 docker buildx build --load \ + $(bash .buildkite/scripts/docker-build-metadata-args.sh cu134) \ + --platform linux/arm64 \ + --build-arg max_jobs=16 \ + --build-arg USE_SCCACHE=1 \ + --build-arg GIT_REPO_CHECK=1 \ + --build-arg CUDA_VERSION=13.4 \ + --build-arg torch_cuda_arch_list="${CUDA_ARCH_AARCH64}" \ + --build-arg INSTALL_KV_CONNECTORS=true \ + --build-arg INSTALL_RUBIN_PRERELEASE=true \ + --build-arg TRITON_INSTALL_FROM_SOURCE_REPO=https://github.com/triton-lang/triton.git \ + --build-arg TRITON_INSTALL_FROM_SOURCE_REVISION=3f6e41132b5edf639bfb872ad73d4688765e08b8 \ + --build-arg RUN_WHEEL_CHECK=false \ + --build-arg BUILD_BASE_IMAGE=pytorch/manylinuxaarch64-builder:cuda13.4 \ + --build-arg FINAL_BASE_IMAGE=nvcr.io/nvidia/cuda-dl-base:26.08-cuda13.4-devel-ubuntu24.04 \ + --target vllm-openai \ + --progress plain \ + -f docker/Dockerfile . + - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu134" + - 'bash .buildkite/scripts/annotate-build-artifact.sh "$$BUILDKITE_LABEL" "public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu134" release-images' + - label: "Build release image - x86_64 - CUDA 13.0 - Ubuntu 24.04" depends_on: ~ id: build-release-image-x86-ubuntu2404 @@ -513,6 +573,20 @@ steps: - "docker manifest push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu129" - 'bash .buildkite/scripts/annotate-build-artifact.sh "Manifest: CUDA 12.9" "public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu129" release-manifests' + - label: "Create multi-arch manifest - Rubin CUDA 13.4" + depends_on: + - build-release-image-x86-rubin-cuda-13-4 + - build-release-image-arm64-rubin-cuda-13-4 + id: create-multi-arch-manifest-rubin-cuda-13-4 + if: build.env("NIGHTLY") == "1" + agents: + queue: small_cpu_queue_release + commands: + - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7" + - "docker manifest create public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu134 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-x86_64-cu134 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-aarch64-cu134 --amend" + - "docker manifest push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu134" + - 'bash .buildkite/scripts/annotate-build-artifact.sh "Manifest: Rubin CUDA 13.4" "public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu134" release-manifests' + - label: "Create multi-arch manifest - CUDA 13.0 - Ubuntu 24.04" depends_on: - build-release-image-x86-ubuntu2404 @@ -585,6 +659,24 @@ steps: DOCKER_BUILDKIT: "1" DOCKERHUB_USERNAME: "vllmbot" + - label: "Publish nightly multi-arch Rubin image to DockerHub - CUDA 13.4" + depends_on: + - create-multi-arch-manifest-rubin-cuda-13-4 + if: build.env("NIGHTLY") == "1" + agents: + queue: small_cpu_queue_release + commands: + - "bash .buildkite/scripts/push-nightly-builds.sh cu134" + # Clean up old nightly builds (keep only last 14) + - "bash .buildkite/scripts/cleanup-nightly-builds.sh cu134-nightly-" + plugins: + - docker-login#v3.0.0: + username: vllmbot + password-env: DOCKERHUB_TOKEN + env: + DOCKER_BUILDKIT: "1" + DOCKERHUB_USERNAME: "vllmbot" + - label: "Publish nightly CPU multi-arch image to DockerHub" depends_on: - build-cpu-release-image-x86 diff --git a/.buildkite/scripts/cleanup-nightly-builds.sh b/.buildkite/scripts/cleanup-nightly-builds.sh index b324778b1f..41b04996e4 100755 --- a/.buildkite/scripts/cleanup-nightly-builds.sh +++ b/.buildkite/scripts/cleanup-nightly-builds.sh @@ -56,8 +56,8 @@ get_all_tags() { set -x # Get both last_updated timestamp and tag name, separated by | - # Exclude nightly-dev tags from cleanup - local tags=$(echo "$response" | jq -r --arg prefix "$TAG_PREFIX" '.results[] | select(.name | startswith($prefix)) | select(.name | startswith("nightly-dev") | not) | "\(.last_updated)|\(.name)"') + # Exclude mutable architecture tags and nightly-dev tags from cleanup + local tags=$(echo "$response" | jq -r --arg prefix "$TAG_PREFIX" '.results[] | select(.name | startswith($prefix)) | select(.name != ($prefix + "x86_64") and .name != ($prefix + "aarch64")) | select(.name | startswith("nightly-dev") | not) | "\(.last_updated)|\(.name)"') if [ -z "$tags" ]; then break diff --git a/docker/Dockerfile b/docker/Dockerfile index d1a7afcbfb..ad010285b1 100644 --- a/docker/Dockerfile +++ b/docker/Dockerfile @@ -231,10 +231,12 @@ RUN --mount=type=cache,target=/opt/uv/cache \ echo "Installing Rubin prerelease requirements..." \ && /opt/venv/bin/python3 use_existing_torch.py --prefix \ && uv pip install --python /opt/venv/bin/python3 \ + --prerelease=allow --index-strategy unsafe-first-match \ -r /tmp/vllm-requirements/rubin-prerelease.txt \ && uv pip install --python /opt/venv/bin/python3 \ -r requirements/cuda.txt \ && uv pip install --python /opt/venv/bin/python3 \ + --prerelease=allow --index-strategy unsafe-first-match \ -r /tmp/vllm-requirements/rubin-prerelease.txt; \ elif [ "${PYTORCH_NIGHTLY}" = "1" ]; then \ echo "Installing torch nightly..." \ @@ -941,11 +943,13 @@ RUN --mount=type=cache,target=/opt/uv/cache \ fi && \ if [ "${INSTALL_RUBIN_PRERELEASE}" = "true" ]; then \ uv pip install --system \ + --prerelease=allow --index-strategy unsafe-first-match \ -r /tmp/vllm-requirements/rubin-prerelease.txt \ && sed -Ei '/^[[:space:]]*(torch|torchvision|torchaudio)[[:space:]]*==/Id' \ /tmp/requirements-cuda.txt \ && uv pip install --system -r /tmp/requirements-cuda.txt \ && uv pip install --system \ + --prerelease=allow --index-strategy unsafe-first-match \ -r /tmp/vllm-requirements/rubin-prerelease.txt \ && printf '%s\n' \ "/usr/local/lib/python${PYTHON_VERSION}/dist-packages/nvpl/lib" \ @@ -1262,6 +1266,7 @@ RUN --mount=type=cache,target=/opt/uv/cache \ if [ "${INSTALL_RUBIN_PRERELEASE}" = "true" ]; then \ echo "Installing Rubin prerelease requirements..." \ && uv pip install --system \ + --prerelease=allow --index-strategy unsafe-first-match \ -r /tmp/vllm-requirements/rubin-prerelease.txt; \ fi diff --git a/tests/v1/engine/test_output_processor.py b/tests/v1/engine/test_output_processor.py index 9df43088e6..69128f55b7 100644 --- a/tests/v1/engine/test_output_processor.py +++ b/tests/v1/engine/test_output_processor.py @@ -177,7 +177,8 @@ def test_incremental_detokenization( def test_remote_prefill_cached_tokens_override(do_remote_prefill: bool): """P/D disaggregation: num_cached_tokens should report the P worker's cache hits (passed via kv_transfer_params) instead of the local count, - which sees the KVs pulled from the remote prefill as a ~100% hit.""" + which sees the KVs pulled from the remote prefill as a ~100% hit. + """ output_processor = OutputProcessor(tokenizer=None, log_stats=False) prompt_tokens = [1, 2, 3, 4, 5, 6, 7, 8] diff --git a/tests/v1/kv_connector/unit/test_nixl_connector.py b/tests/v1/kv_connector/unit/test_nixl_connector.py index 2185ba62d2..9ab4b0ae95 100644 --- a/tests/v1/kv_connector/unit/test_nixl_connector.py +++ b/tests/v1/kv_connector/unit/test_nixl_connector.py @@ -350,7 +350,8 @@ def test_prefill_exports_cached_tokens_in_kv_transfer_params(): """The P worker reports its own prefix-cache hits in the returned kv_transfer_params so the D worker can surface them in prompt_tokens_details instead of the ~100% local hit it measures - when pulling the KVs from the remote.""" + when pulling the KVs from the remote. + """ vllm_config = create_vllm_config() scheduler = create_scheduler(vllm_config) diff --git a/vllm/models/qwen4_exp/nvidia/ops/qsa.py b/vllm/models/qwen4_exp/nvidia/ops/qsa.py index 816ccdd498..401048e02c 100644 --- a/vllm/models/qwen4_exp/nvidia/ops/qsa.py +++ b/vllm/models/qwen4_exp/nvidia/ops/qsa.py @@ -482,7 +482,8 @@ def _select_sm120_config( """(block_n, target_splits, num_warps) retuned on sm_120 (RTX PRO 6000 Blackwell), split by cache dtype: bf16 and fp8 favour different configs on sm_120, most visibly on the large-prefill region. Each entry is the fastest - config that stays correct on its own path; the main win is more warps.""" + config that stays correct on its own path; the main win is more warps. + """ if is_fp8: if base_programs > 2048: return (32, 2, 1) if use_prefill_config else (64, 1, 2)