Update CI for HF Optimum models

Author: Guang Yang

.github/workflows/trunk.yml

@@ -555,11 +555,10 @@ jobs:
     strategy:
       matrix:
         hf_model_id: [
-          google/gemma-2-2b,
-          Qwen/Qwen2.5-0.5B,
+          google/gemma-3-1b-it,
+          Qwen/Qwen3-0.6B,
           HuggingFaceTB/SmolLM2-135M,
-          meta-llama/Llama-3.2-1B,
-          allenai/OLMo-1B-hf
+          allenai/OLMo-1B-hf,
         ]
       fail-fast: false
     with:
@@ -583,25 +582,37 @@ jobs:
         git clone https://github.com/huggingface/optimum-executorch
         cd optimum-executorch
         # There is no release yet, for CI stability, always test from the same commit on main
-        git checkout 577a2b19670e4c643a5c6ecb09bf47b9a699e7c6
+        git checkout da80c9e35b3db5c7eea8731b7d660482fb4870a8
         pip install .[tests]
         pip list
         echo "::endgroup::"
 
         echo "::group::Export and Run ${{ matrix.hf_model_id }}"
         # Pass matrix variable as environment variable
         export MODEL_ID="${{ matrix.hf_model_id }}"
+        export OUTPUT_DIR="${MODEL_ID}_custom_sdpa_8da4w"
+
+        optimum-cli export executorch \
+          --model ${MODEL_ID} \
+          --task text-generation \
+          --recipe xnnpack \
+          --use_custom_sdpa \
+          --output_dir ${OUTPUT_DIR} \
+          --qlinear
+
+        ls -FlAGhp ${OUTPUT_DIR}
+
         python -c "
         import os
         from optimum.executorch import ExecuTorchModelForCausalLM
         from transformers import AutoTokenizer
 
         model_id = os.getenv('MODEL_ID')
-        print(f'Loading model: {model_id}')
-        model = ExecuTorchModelForCausalLM.from_pretrained(model_id, recipe='xnnpack')
-        tokenizer = AutoTokenizer.from_pretrained(model_id)
+        pte_dir = os.getenv('OUTPUT_DIR')
+        print(f"Loading model '{model_id}' from '{pte_dir}'.")
+        model = ExecuTorchModelForCausalLM.from_pretrained(pte_dir)
         generated_text = model.text_generation(
-          tokenizer=tokenizer,
+          tokenizer=AutoTokenizer.from_pretrained(model_id),
           prompt='Simply put, the theory of relativity states that',
           max_seq_len=64
         )