debug

Author: cccclai

.github/workflows/trunk.yml

@@ -787,7 +787,6 @@ jobs:
   #       done
 
   test-huggingface-transformers-xnnpack:
-    # NB: Don't run this on fork PRs because they won't have access to the secret and would fail anyway
     if: ${{ !github.event.pull_request.head.repo.fork }}
     name: test-huggingface-transformers-xnnpack
     uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
@@ -796,33 +795,17 @@ jobs:
       contents: read
     secrets: inherit
     strategy:
-      matrix:
-        config: [
-          # XNNPack.
-          llama3.2-1b|xnnpack|--quantize,
-          qwen3-0.6b|xnnpack|--quantize,
-          qwen3-1.7b|xnnpack|--quantize,
-          gemma3-1b|xnnpack|--quantize,
-          phi4-mini|xnnpack|--quantize,
-          smollm2-135m|xnnpack|--quantize,
-          smollm3-3b|xnnpack|--quantize
-        ]
       fail-fast: false
+      matrix:
+        model: ["gemma3-4b"]  # llava gives segfault so not covering.
     with:
       secrets-env: EXECUTORCH_HF_TOKEN
-      runner: linux.2xlarge.memory
+      runner: linux.24xlarge
       docker-image: ci-image:executorch-ubuntu-22.04-clang12
       submodules: 'recursive'
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
       timeout: 90
-      upload-artifact: profiling-artifacts-${{ strategy.job-index }}
       script: |
-        set -eux
-        IFS='|' read -r MODEL RECIPE QUANTIZE <<< "${{ matrix.config }}"
-        echo "Model: $MODEL"
-        echo "Recipe: $RECIPE"
-        echo "Quantize: $QUANTIZE"
-
         # The generic Linux job chooses to use base env, not the one setup by the image
         CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")
         conda activate "${CONDA_ENV}"
@@ -838,23 +821,8 @@ jobs:
         pip install git+https://github.com/huggingface/optimum-executorch.git@${OPTIMUM_ET_VERSION}
         echo "::endgroup::"
 
-        python -c "import torch, torchvision; print(torch.__version__, torchvision.__version__)"
-
-        echo "::group::Run tests"
-        export OUTPUT_DIR="$(pwd)/${MODEL}_${RECIPE}_${QUANTIZE}"
-        python .ci/scripts/test_huggingface_optimum_model.py --model ${MODEL} --recipe ${RECIPE} ${QUANTIZE} --model_dir ${OUTPUT_DIR}
-        echo "::endgroup::"
-
-        echo "::group::Generate artifacts for performance profiling"
-        ./cmake-out/executor_runner \
-          --model_path ${OUTPUT_DIR}/model.pte \
-          --etdump_path ${OUTPUT_DIR}/etdump.etdp
-
-        export TSV_PATH=artifacts-to-be-uploaded/${MODEL}_op_prof.tsv
-        mkdir -p $(dirname "$TSV_PATH")
-        python3 -m devtools.inspector.inspector_cli \
-          --etdump_path ${OUTPUT_DIR}/etdump.etdp \
-          --tsv_path ${TSV_PATH}
+        echo "::group::Test ${{ matrix.model }}"
+        python .ci/scripts/test_huggingface_optimum_model.py --model ${{ matrix.model }} --quantize --recipe xnnpack
         echo "::endgroup::"
 
   # test-huggingface-transformers-macos: