Another try

Signed-off-by: Huy Do <huydhn@gmail.com>

Author: huydhn

.github/scripts/gpt-oss/run_accuracy_checks.sh

@@ -12,6 +12,8 @@ elif [[ "${DEVICE_NAME}" == *rocm* ]]; then
   export VLLM_ROCM_USE_AITER=1
   export VLLM_USE_AITER_UNIFIED_ATTENTION=1
   export VLLM_ROCM_USE_AITER_MHA=0
+else
+  export VLLM_FLASH_ATTN_VERSION=2
 fi
 
 tp=0
@@ -44,9 +46,7 @@ mkdir -p /tmp/gpqa_openai
 
 # Not sure why this is needed on ROCm image
 if [[ "${DEVICE_NAME}" == "rocm" ]]; then
-  ls -la gpt_oss
-  ls -la gpt_oss/evals
-  export PYTHONPATH=$(pwd):$PYTHONPATH
+  export PYTHONPATH=$(pwd)
 fi
 
 # Low

.github/scripts/gpt-oss/run_benchmarks.sh

@@ -12,6 +12,8 @@ elif [[ "${DEVICE_NAME}" == *rocm* ]]; then
   export VLLM_ROCM_USE_AITER=1
   export VLLM_USE_AITER_UNIFIED_ATTENTION=1
   export VLLM_ROCM_USE_AITER_MHA=0
+else
+  export VLLM_FLASH_ATTN_VERSION=2
 fi
 
 pushd vllm-benchmarks/vllm
@@ -22,10 +24,6 @@ if [[ "${DEVICE_NAME}" != "rocm" ]]; then
   pip install --pre vllm==0.10.1+gptoss \
     --extra-index-url https://wheels.vllm.ai/gpt-oss/ \
     --extra-index-url https://download.pytorch.org/whl/nightly/cu128
-
-  export TORCH_CUDA_ARCH_LIST='8.9 9.0'
-  pip install --no-build-isolation "git+https://github.com/facebookresearch/xformers@v0.0.31" \
-    --extra-index-url https://download.pytorch.org/whl/nightly/cu128
 fi
 
 pip freeze

.github/workflows/gpt-oss-benchmark.yml

@@ -16,7 +16,7 @@ jobs:
       matrix:
         include:
           # gpt-oss-120b
-          - runner: linux.aws.h100.4
+          - runner: linux.aws.h100.8
             model: openai/gpt-oss-120b
             docker-image: 'public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:6d8d0a24c02bfd84d46b3016b865a44f048ae84b'
           - runner: linux.dgx.b200.8