Remove extra copy in deepseek forward absorb (#5578)

Co-authored-by: saienduri <saimanas.enduri@amd.com>
This commit is contained in:
Ke Bao
2025-04-22 10:33:21 +08:00
committed by GitHub
parent b9c87e781d
commit 11b23ae97b
3 changed files with 18 additions and 21 deletions

View File

@@ -38,12 +38,12 @@ jobs:
else
DEVICE_FLAG="--device /dev/dri"
fi
docker pull lmsysorg/sglang:v0.4.5-rocm630
docker pull lmsysorg/sglang:v0.4.5.post2-rocm630
docker run -dt --user root --device=/dev/kfd $DEVICE_FLAG \
-v ${{ github.workspace }}:/sglang-checkout --ipc=host --group-add video \
--cap-add=SYS_PTRACE -e HF_TOKEN=${HF_TOKEN} --security-opt seccomp=unconfined \
-w /sglang-checkout --name ci_sglang \
lmsysorg/sglang:v0.4.5-rocm630
lmsysorg/sglang:v0.4.5.post2-rocm630
- name: Install dependencies
run: |
@@ -82,12 +82,12 @@ jobs:
else
DEVICE_FLAG="--device /dev/dri"
fi
docker pull lmsysorg/sglang:v0.4.5-rocm630
docker pull lmsysorg/sglang:v0.4.5.post2-rocm630
docker run -dt --user root --device=/dev/kfd $DEVICE_FLAG \
-v ${{ github.workspace }}:/sglang-checkout --ipc=host --group-add video \
--cap-add=SYS_PTRACE -e HF_TOKEN=${{ secrets.AMD_HF_TOKEN }} --security-opt seccomp=unconfined \
-w /sglang-checkout --name ci_sglang \
lmsysorg/sglang:v0.4.5-rocm630
lmsysorg/sglang:v0.4.5.post2-rocm630
- name: Install dependencies
run: |
@@ -120,12 +120,12 @@ jobs:
else
DEVICE_FLAG="--device /dev/dri"
fi
docker pull lmsysorg/sglang:v0.4.5-rocm630
docker pull lmsysorg/sglang:v0.4.5.post2-rocm630
docker run -dt --user root --device=/dev/kfd $DEVICE_FLAG \
-v ${{ github.workspace }}:/sglang-checkout --ipc=host --group-add video \
--cap-add=SYS_PTRACE -e HF_TOKEN=${HF_TOKEN} --security-opt seccomp=unconfined \
-w /sglang-checkout --name ci_sglang \
lmsysorg/sglang:v0.4.5-rocm630
lmsysorg/sglang:v0.4.5.post2-rocm630
- name: Install dependencies
run: |
@@ -149,7 +149,7 @@ jobs:
finish:
if: always()
needs: [
accuracy-test-1-gpu-amd, mla-test-1-gpu-amd
accuracy-test-1-gpu-amd, mla-test-1-gpu-amd, bench-test-2-gpu-amd
]
runs-on: ubuntu-latest
steps: