b200 test

Author: Oseltamivir

.github/workflows/eval-tmpl.yml

@@ -82,7 +82,6 @@ jobs:
       - name: Resource cleanup
         run: |
           # Helper to avoid indefinite hangs on flaky tools (Docker/Slurm)
-          sudo rm -rf /home/nvadmin/actions-runner/_work/InferenceMAX/InferenceMAX/gpt-oss-120b/
           safe_timeout() {
             if command -v timeout >/dev/null 2>&1; then
               timeout -k 5 30s "$@"
@@ -138,7 +137,7 @@ jobs:
         with:
           fetch-depth: 0
           # Avoid aggressive workspace deletion if stale, rely on git reset/clean later
-          clean: true
+          clean: false
 
       - name: Launch eval via runner script
         env:

benchmarks/gptoss_fp4_b200_docker.sh

@@ -14,6 +14,8 @@
 
 nvidia-smi
 
+rm -rf /home/nvadmin/actions-runner/_work/InferenceMAX/InferenceMAX/gpt-oss-120b/
+
 # To improve CI stability, we patch this helper function to prevent a race condition that
 # happens 1% of the time. ref: https://github.com/flashinfer-ai/flashinfer/pull/1779
 sed -i '102,108d' /usr/local/lib/python3.12/dist-packages/flashinfer/jit/cubin_loader.py