[NVIDIA] Add CUDA 13.4 container for initial Rubin support (#36233)

.github/workflows/release-docker-cu134-nightly.yml

@@ -0,0 +1,104 @@
+name: Release Docker Images Nightly (CUDA 13.4 Rubin)
+#
+# Builds and publishes nightly CUDA 13.4 / Rubin (sm_107) images from
+# docker/Dockerfile.cu134, aarch64 only:
+#   - lmsysorg/sglang:nightly-cu134-{date}-{sha}
+#   - lmsysorg/sglang:nightly-cu134
+#
+
+on:
+  workflow_dispatch:
+    inputs:
+      image_repo:
+        description: "Docker Hub repo to push to. Use lmsysorg/sglang-staging for testing."
+        required: false
+        default: "lmsysorg/sglang"
+      docker_target:
+        description: "Dockerfile stage to build and publish."
+        required: false
+        default: "runtime"
+      build_only:
+        description: "Build and validate without logging in or pushing."
+        required: false
+        type: boolean
+        default: false
+  schedule:
+    # Offset from the other nightly docker builds; this one runs long.
+    - cron: "0 10 * * *"
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  build-arm64:
+    if: github.repository == 'sgl-project/sglang'
+    runs-on: arm-docker-build-node
+    environment: ${{ (github.event_name == 'schedule' || !inputs.build_only) && 'prod' || null }}
+    # Cold builds recompile three CUDA wheels from source.
+    timeout-minutes: 360
+
+    steps:
+      - name: Delete huge unnecessary tools folder
+        run: rm -rf /opt/hostedtoolcache
+
+      - name: Cleanup workspace (remove root-owned files from prior runs)
+        run: sudo rm -rf "$GITHUB_WORKSPACE"/* || true
+
+      - name: Checkout repository
+        uses: actions/checkout@v4
+
+      - name: Compute image metadata
+        id: meta
+        run: |
+          set -euo pipefail
+          date_tag="$(date +%Y%m%d)"
+          commit_hash="$(git rev-parse --short=7 HEAD)"
+          image_repo="${{ inputs.image_repo || 'lmsysorg/sglang' }}"
+          echo "image_repo=${image_repo}"                                  >> "$GITHUB_OUTPUT"
+          echo "tag=nightly-cu134-${date_tag}-${commit_hash}"              >> "$GITHUB_OUTPUT"
+          echo "tag_rolling=nightly-cu134"                                 >> "$GITHUB_OUTPUT"
+
+      - name: Login to Docker Hub
+        if: github.event_name == 'schedule' || !inputs.build_only
+        uses: docker/login-action@v2
+        with:
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_TOKEN }}
+
+      - name: Build image
+        run: |
+          set -euo pipefail
+          docker buildx build \
+            --builder default \
+            --progress=plain \
+            --platform linux/arm64 \
+            --target "${{ inputs.docker_target || 'runtime' }}" \
+            --build-arg BRANCH_TYPE=local \
+            --build-arg BUILD_TYPE=all \
+            --build-arg BUILD_JOBS="$(nproc)" \
+            -f docker/Dockerfile.cu134 \
+            -t "${{ steps.meta.outputs.image_repo }}:${{ steps.meta.outputs.tag }}" \
+            -t "${{ steps.meta.outputs.image_repo }}:${{ steps.meta.outputs.tag_rolling }}" \
+            .
+
+      - name: Push image
+        if: github.event_name == 'schedule' || !inputs.build_only
+        run: |
+          set -euo pipefail
+          push_with_retry() {
+            for i in 1 2 3 4 5; do
+              docker push "$1" && return 0
+              echo "push failed (attempt $i), retrying in 30s"
+              sleep 30
+            done
+            return 1
+          }
+          push_with_retry "${{ steps.meta.outputs.image_repo }}:${{ steps.meta.outputs.tag }}"
+          push_with_retry "${{ steps.meta.outputs.image_repo }}:${{ steps.meta.outputs.tag_rolling }}"
+
+      - name: Cleanup local images
+        if: always()
+        run: |
+          docker rmi -f "${{ steps.meta.outputs.image_repo }}:${{ steps.meta.outputs.tag }}" || true
+          docker rmi -f "${{ steps.meta.outputs.image_repo }}:${{ steps.meta.outputs.tag_rolling }}" || true

python/sglang/kernels/aot/CMakeLists.txt

@@ -20,7 +20,8 @@ set(GITHUB_ARTIFACTORY "github.com" CACHE STRING "GitHub mirror URL")
 find_package(Python COMPONENTS Interpreter Development.Module ${SKBUILD_SABI_COMPONENT} REQUIRED)
 
 # CXX
-set(CMAKE_CXX_STANDARD 17)
+set(SGL_KERNEL_CXX_STANDARD "17" CACHE STRING "C++ standard used for host and CUDA compilation")
+set(CMAKE_CXX_STANDARD ${SGL_KERNEL_CXX_STANDARD})
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O3")
 
 # CUDA
@@ -125,7 +126,7 @@ set(SGL_KERNEL_CUDA_FLAGS
     "-Xcompiler"
     "-fPIC"
     "-gencode=arch=compute_90,code=sm_90"
-    "-std=c++17"
+    "-std=c++${SGL_KERNEL_CXX_STANDARD}"
     "-DFLASHINFER_ENABLE_F16"
     "-DCUTE_USE_PACKED_TUPLE=1"
     "-DCUTLASS_ENABLE_TENSOR_CORE_MMA=1"
@@ -399,7 +400,7 @@ if (SGL_KERNEL_ENABLE_FA3)
         "-Xcompiler"
         "-fPIC"
         "-gencode=arch=compute_90a,code=sm_90a"
-        "-std=c++17"
+        "-std=c++${SGL_KERNEL_CXX_STANDARD}"
         "-DCUTE_USE_PACKED_TUPLE=1"
         "-DCUTLASS_ENABLE_TENSOR_CORE_MMA=1"
         "-DCUTLASS_VERSIONS_GENERATED"
@@ -493,7 +494,7 @@ endif()
 set(INFLLM_FLASH_CUDA_FLAGS
     "-DNDEBUG"
     "-O3"
-    "-std=c++17"
+    "-std=c++${SGL_KERNEL_CXX_STANDARD}"
     "-Xcompiler"
     "-fPIC"
     "-U__CUDA_NO_HALF_OPERATORS__"

python/sglang/kernels/aot/cmake/flashmla.cmake

@@ -2,8 +2,8 @@
 # sm90 dense decode HEAD_DIM_K=512 support (sgl-project/FlashMLA#9, merged).
 FetchContent_Declare(
     repo-flashmla
-    URL      https://${GITHUB_ARTIFACTORY}/sgl-project/FlashMLA/archive/05e26647fe840b8baedae486c2d86d5ce4efeb7c.tar.gz
-    URL_HASH SHA256=ce369489bbfc42cdfbba9aa949de0270e64469d530748dea9f4f60b3c69dea9b
+    URL      https://${GITHUB_ARTIFACTORY}/sgl-project/FlashMLA/archive/c1dee569a494b184811a08171a690ece21420262.tar.gz
+    URL_HASH SHA256=77d3f1714b5903dc8f7a99fbc3a5d9a2b886e449f994b6c4b1d2feeacb467b1e
 )
 FetchContent_Populate(repo-flashmla)