update ci (#857)

Author: zhyncs

.github/workflows/release_wheel.yml

@@ -27,9 +27,11 @@ jobs:
       matrix:
         cuda: ["11.8", "12.1", "12.4"]
         torch: ["2.4", "2.5", "2.6"]
-        exclude: # We use release_wheel_sglang.yml for faster release and verification. If everything is okay, then we trigger release_wheel.yml. This combination (cuda 12.4 + torch 2.5) is already handled in release_wheel_sglang.yml
+        exclude: # We use release_wheel_sglang.yml for faster release and verification. If everything is okay, then we trigger release_wheel.yml. This combination (cuda 12.4 or 11.8 + torch 2.5) is already handled in release_wheel_sglang.yml
           - cuda: "12.4"
             torch: "2.5"
+          - cuda: "11.8"
+            torch: "2.5"
 
     runs-on: [self-hosted]
     steps:

.github/workflows/release_wheel_sglang.yml

@@ -22,7 +22,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        cuda: ["12.4"]
+        cuda: ["11.8", "12.4"]
         torch: ["2.5"]
 
     runs-on: [self-hosted]
@@ -41,6 +41,7 @@ jobs:
               -e FLASHINFER_CI_CUDA_VERSION=${{ matrix.cuda }} \
               -e FLASHINFER_CI_TORCH_VERSION=${{ matrix.torch }} \
               -e FLASHINFER_CI_PYTHON_VERSION=3.10 \
+              -e FLASHINFER_HEAD_DIMS="64,128,256" \
               -e TORCH_CUDA_ARCH_LIST="$TORCH_CUDA_ARCH_LIST" \
               -e MAX_JOBS=128 \
               --user $CI_UID:$CI_GID \