[Deps] Upgrade CUDA PyTorch stack to 2.13 (#28836)

Co-authored-by: Brayden Zhong <b8zhong@uwaterloo.ca>

.github/workflows/_docker-build-and-publish.yml

@@ -120,7 +120,7 @@ jobs:
             --platform linux/amd64 \
             --output type=image,name=${{ inputs.image_repo }},push-by-digest=true,name-canonical=true,push=true \
             -f docker/Dockerfile \
-            --build-arg CUDA_VERSION=12.9.1 \
+            --build-arg CUDA_VERSION=12.9.2 \
             --build-arg BUILD_TYPE=all \
             --build-arg GRACE_BLACKWELL=0 \
             --build-arg INSTALL_FLASHINFER_JIT_CACHE=1 \
@@ -149,7 +149,7 @@ jobs:
             --platform linux/amd64 \
             --output type=image,name=${{ inputs.image_repo }},push-by-digest=true,name-canonical=true,push=true \
             -f docker/Dockerfile \
-            --build-arg CUDA_VERSION=13.0.1 \
+            --build-arg CUDA_VERSION=13.0.3 \
             --build-arg BUILD_TYPE=all \
             --build-arg GRACE_BLACKWELL=0 \
             --build-arg INSTALL_FLASHINFER_JIT_CACHE=1 \
@@ -232,7 +232,7 @@ jobs:
             --platform linux/arm64 \
             --output type=image,name=${{ inputs.image_repo }},push-by-digest=true,name-canonical=true,push=true \
             -f docker/Dockerfile \
-            --build-arg CUDA_VERSION=12.9.1 \
+            --build-arg CUDA_VERSION=12.9.2 \
             --build-arg BUILD_TYPE=all \
             --build-arg GRACE_BLACKWELL=1 \
             --build-arg INSTALL_FLASHINFER_JIT_CACHE=1 \
@@ -261,7 +261,7 @@ jobs:
             --platform linux/arm64 \
             --output type=image,name=${{ inputs.image_repo }},push-by-digest=true,name-canonical=true,push=true \
             -f docker/Dockerfile \
-            --build-arg CUDA_VERSION=13.0.1 \
+            --build-arg CUDA_VERSION=13.0.3 \
             --build-arg BUILD_TYPE=all \
             --build-arg GRACE_BLACKWELL=1 \
             --build-arg INSTALL_FLASHINFER_JIT_CACHE=1 \

.github/workflows/_pr-test-stage.yml

@@ -68,6 +68,7 @@ on:
 # defines must be redeclared here for the called job to see the same context.
 env:
   SGLANG_IS_IN_CI: true
+  NCCL_NVLS_ENABLE: "0"
   # base-a runs the fwd-occupancy sanity kit; async-assert probes add GPU work
   # and skew the occupancy measurement, so keep them off in that stage.
   SGLANG_ENABLE_ASYNC_ASSERT: ${{ startsWith(inputs.self_name, 'base-a') && 'false' || 'true' }}

.github/workflows/nightly-72-gpu-gb200.yml

@@ -195,7 +195,7 @@ jobs:
             --output type=image,name=${IMAGE_REF},push=true \
             --target framework_final \
             -f docker/Dockerfile \
-            --build-arg CUDA_VERSION=13.0.1 \
+            --build-arg CUDA_VERSION=13.0.3 \
             --build-arg BUILD_TYPE=all \
             --build-arg CMAKE_BUILD_PARALLEL_LEVEL=$(nproc) \
             --build-arg GRACE_BLACKWELL=1 \

.github/workflows/pr-test-jit-kernel.yml

@@ -36,6 +36,7 @@ on:
 # The github context (including github.event_name) IS inherited from the caller.
 env:
   SGLANG_IS_IN_CI: true
+  NCCL_NVLS_ENABLE: "0"
   SGLANG_CUDA_COREDUMP: "1"
   SGLANG_JIT_DEEPGEMM_FAST_WARMUP: true
   PR_TEST_BYPASS_MAINTENANCE_ON_MAIN: ${{ github.ref == 'refs/heads/main' && 'true' || 'false' }}

.github/workflows/pr-test-multimodal-gen.yml

@@ -44,6 +44,7 @@ on:
 # The github context (including github.event_name) IS inherited from the caller.
 env:
   SGLANG_IS_IN_CI: true
+  NCCL_NVLS_ENABLE: "0"
   SGLANG_CUDA_COREDUMP: "1"
   PR_TEST_BYPASS_MAINTENANCE_ON_MAIN: ${{ github.ref == 'refs/heads/main' && 'true' || 'false' }}
   SKIP_PR_TEST_HEALTH_CHECK: ${{ inputs.skip_pr_test_health_check == 'true' }}

.github/workflows/pr-test.yml

@@ -59,6 +59,8 @@ env:
   SGLANG_CUDA_COREDUMP: "1"
   SGLANG_JIT_DEEPGEMM_FAST_WARMUP: true
   SKIP_PR_TEST_HEALTH_CHECK: ${{ (inputs.skip_pr_test_health_check == true || inputs.test_parallel_dispatch == true || inputs.run_all_tests == true) && 'true' || 'false' }}
+  # TEMP: force DeepEP rebuilds during the initial Torch 2.13 rollout so every CI runner replaces its Torch 2.11 build; remove after the runner fleet has cycled.
+  FORCE_REBUILD_DEEPEP: '1'
   # Schedule / main-branch dispatch / workflow_call from main use refs/heads/main; PR events use refs/pull/*/merge
   PR_TEST_BYPASS_MAINTENANCE_ON_MAIN: ${{ github.ref == 'refs/heads/main' && 'true' || 'false' }}
   USE_VENV: false

.github/workflows/release-docker-dev.yml

@@ -176,7 +176,7 @@ jobs:
             --load \
             -t "$IMAGE" \
             -f docker/Dockerfile \
-            --build-arg CUDA_VERSION=13.0.1 \
+            --build-arg CUDA_VERSION=13.0.3 \
             --build-arg BUILD_TYPE=all \
             --build-arg GRACE_BLACKWELL=0 \
             --build-arg INSTALL_FLASHINFER_JIT_CACHE=1 \

.github/workflows/release-whl-deepgemm.yml

@@ -27,9 +27,10 @@ concurrency:
   cancel-in-progress: true
 
 env:
+  NCCL_NVLS_ENABLE: "0"
   # Must match TORCH_VER in docker/sgl-deep-gemm.Dockerfile — the wheel's
   # pre-compiled _C.so links against this torch ABI.
-  TORCH_VER: "2.11.0"
+  TORCH_VER: "2.13.0"
 
 jobs:
   build-cu129-matrix:

.github/workflows/rerun-test.yml

@@ -62,7 +62,7 @@ env:
   SGLANG_ENABLE_ASYNC_ASSERT: true
   SGLANG_CUDA_COREDUMP: "1"
   SGLANG_JIT_DEEPGEMM_FAST_WARMUP: true
-  # TEMP: rebuild deepep against the new torch for torch-211-merge PR only — revert before merging to main.
+  # TEMP: force DeepEP rebuilds during the initial Torch 2.13 rollout so every CI runner replaces its Torch 2.11 build; remove after the runner fleet has cycled.
   FORCE_REBUILD_DEEPEP: '1'
 
 permissions: