Add Qwen 3.5 MoE E2E CI test with prequantized HQQ-INT4 checkpoint (#18445)

Add end-to-end CI testing for Qwen3.5-35B-A3B MoE model using a
prequantized HQQ-INT4 checkpoint from HuggingFace. The CI flow:

1. Download prequantized model
(SocialLocalMobile/Qwen3.5-35B-A3B-HQQ-INT4)
2. Run inference.py sanity check (verify correct output in eager mode)
3. Export to .pte/.ptd via export.py --prequantized
4. Run C++ runner, validate output contains "Paris"

Remove the standalone runner build from unittest-cuda (now covered by
the E2E test job).

Author: mergennachin

.ci/scripts/export_model_artifact.sh

@@ -184,9 +184,17 @@ case "$HF_MODEL" in
     PREPROCESSOR_FEATURE_SIZE=""
     PREPROCESSOR_OUTPUT=""
     ;;
+  SocialLocalMobile/Qwen3.5-35B-A3B-HQQ-INT4)
+    MODEL_NAME="qwen3_5_moe"
+    TASK=""
+    MAX_SEQ_LEN=""
+    EXTRA_PIP=""
+    PREPROCESSOR_FEATURE_SIZE=""
+    PREPROCESSOR_OUTPUT=""
+    ;;
   *)
     echo "Error: Unsupported model '$HF_MODEL'"
-    echo "Supported models: mistralai/Voxtral-Mini-3B-2507, mistralai/Voxtral-Mini-4B-Realtime-2602, openai/whisper-{small, medium, large, large-v2, large-v3, large-v3-turbo}, google/gemma-3-4b-it, Qwen/Qwen3-0.6B, nvidia/diar_streaming_sortformer_4spk-v2, nvidia/parakeet-tdt, facebook/dinov2-small-imagenet1k-1-layer"
+    echo "Supported models: mistralai/Voxtral-Mini-3B-2507, mistralai/Voxtral-Mini-4B-Realtime-2602, openai/whisper-{small, medium, large, large-v2, large-v3, large-v3-turbo}, google/gemma-3-4b-it, Qwen/Qwen3-0.6B, nvidia/diar_streaming_sortformer_4spk-v2, nvidia/parakeet-tdt, facebook/dinov2-small-imagenet1k-1-layer, SocialLocalMobile/Qwen3.5-35B-A3B-HQQ-INT4"
     exit 1
     ;;
 esac
@@ -380,6 +388,45 @@ if [ "$MODEL_NAME" = "voxtral_realtime" ]; then
   exit 0
 fi
 
+# Qwen 3.5 MoE uses a prequantized checkpoint and custom export script
+if [ "$MODEL_NAME" = "qwen3_5_moe" ]; then
+  pip install safetensors huggingface_hub
+  pip install -r examples/models/qwen3_5_moe/requirements.txt
+
+  # Download prequantized model outside OUTPUT_DIR to avoid uploading on failure
+  LOCAL_MODEL_DIR=$(mktemp -d)
+  INDUCTOR_CACHE=$(mktemp -d)
+  trap 'rm -rf "$LOCAL_MODEL_DIR" "$INDUCTOR_CACHE"' EXIT
+
+  python -c "from huggingface_hub import snapshot_download; snapshot_download('${HF_MODEL}', local_dir='${LOCAL_MODEL_DIR}')"
+
+  # Sanity check: run inference on the prequantized model
+  echo "::group::Inference sanity check"
+  python -m executorch.examples.models.qwen3_5_moe.inference \
+      --prequantized "$LOCAL_MODEL_DIR" \
+      --prompt "What is the capital of France?" \
+      --max-new-tokens 32 \
+      --temperature 0 \
+      --no-compile
+  echo "::endgroup::"
+
+  # Copy tokenizer for the runner
+  cp "$LOCAL_MODEL_DIR/tokenizer.json" "${OUTPUT_DIR}/tokenizer.json"
+
+  # Export to .pte/.ptd (short cache dir avoids objcopy symbol length issues)
+  echo "::group::Export"
+  TORCHINDUCTOR_CACHE_DIR="$INDUCTOR_CACHE" \
+  python -m executorch.examples.models.qwen3_5_moe.export \
+      --prequantized "$LOCAL_MODEL_DIR" \
+      --output-dir "${OUTPUT_DIR}"
+  echo "::endgroup::"
+
+  test -f "${OUTPUT_DIR}/model.pte"
+  test -f "${OUTPUT_DIR}/aoti_cuda_blob.ptd"
+  ls -al "${OUTPUT_DIR}"
+  exit 0
+fi
+
 MAX_SEQ_LEN_ARG=""
 if [ -n "$MAX_SEQ_LEN" ]; then
   MAX_SEQ_LEN_ARG="--max_seq_len $MAX_SEQ_LEN"

.ci/scripts/test_model_e2e.sh

@@ -216,9 +216,21 @@ case "$HF_MODEL" in
     AUDIO_FILE="test_audio.wav"
     IMAGE_PATH=""
     ;;
+  SocialLocalMobile/Qwen3.5-35B-A3B-HQQ-INT4)
+    MODEL_NAME="qwen3_5_moe"
+    RUNNER_TARGET="qwen3_5_moe_runner"
+    RUNNER_PATH="qwen3_5_moe"
+    EXPECTED_OUTPUT="Paris"
+    PREPROCESSOR=""
+    TOKENIZER_URL=""
+    TOKENIZER_FILE="tokenizer.json"
+    AUDIO_URL=""
+    AUDIO_FILE=""
+    IMAGE_PATH=""
+    ;;
   *)
     echo "Error: Unsupported model '$HF_MODEL'"
-    echo "Supported models: mistralai/Voxtral-Mini-3B-2507, mistralai/Voxtral-Mini-4B-Realtime-2602, nvidia/diar_streaming_sortformer_4spk-v2, openai/whisper series (whisper-{small, medium, large, large-v2, large-v3, large-v3-turbo}), google/gemma-3-4b-it, Qwen/Qwen3-0.6B, nvidia/parakeet-tdt, facebook/dinov2-small-imagenet1k-1-layer"
+    echo "Supported models: mistralai/Voxtral-Mini-3B-2507, mistralai/Voxtral-Mini-4B-Realtime-2602, nvidia/diar_streaming_sortformer_4spk-v2, openai/whisper series (whisper-{small, medium, large, large-v2, large-v3, large-v3-turbo}), google/gemma-3-4b-it, Qwen/Qwen3-0.6B, nvidia/parakeet-tdt, facebook/dinov2-small-imagenet1k-1-layer, SocialLocalMobile/Qwen3.5-35B-A3B-HQQ-INT4"
     exit 1
     ;;
 esac
@@ -232,7 +244,7 @@ echo "::group::Prepare $MODEL_NAME Artifacts"
 
 
 # Download tokenizer files (skip for models that bundle tokenizer in export or do not use one)
-if [ "$MODEL_NAME" != "parakeet" ] && [ "$MODEL_NAME" != "voxtral_realtime" ] && [ "$MODEL_NAME" != "sortformer" ] && [ "$MODEL_NAME" != "dinov2" ]; then
+if [ "$MODEL_NAME" != "parakeet" ] && [ "$MODEL_NAME" != "voxtral_realtime" ] && [ "$MODEL_NAME" != "sortformer" ] && [ "$MODEL_NAME" != "dinov2" ] && [ "$MODEL_NAME" != "qwen3_5_moe" ]; then
   if [ "$TOKENIZER_FILE" != "" ]; then
     curl -L $TOKENIZER_URL/$TOKENIZER_FILE -o $MODEL_DIR/$TOKENIZER_FILE
   else
@@ -341,6 +353,9 @@ EOF
       RUNNER_ARGS="$RUNNER_ARGS --data_path ${MODEL_DIR}/aoti_cuda_blob.ptd"
     fi
     ;;
+  qwen3_5_moe)
+    RUNNER_ARGS="$RUNNER_ARGS --tokenizer_path ${MODEL_DIR}/$TOKENIZER_FILE --prompt 'What is the capital of France?' --max_new_tokens 32"
+    ;;
   voxtral_realtime)
     RUNNER_ARGS="--model_path ${MODEL_DIR}/model.pte --tokenizer_path ${MODEL_DIR}/$TOKENIZER_FILE --preprocessor_path ${MODEL_DIR}/$PREPROCESSOR --audio_path ${MODEL_DIR}/$AUDIO_FILE --temperature 0"
     # Add CUDA data path if present
@@ -359,7 +374,7 @@ EOF
     ;;
 esac
 
-OUTPUT=$($RUNNER_BIN $RUNNER_ARGS 2>&1)
+OUTPUT=$(eval $RUNNER_BIN $RUNNER_ARGS 2>&1)
 EXIT_CODE=$?
 set -e
 

.github/workflows/cuda.yml

@@ -148,9 +148,6 @@ jobs:
         # Run quantize roundtrip tests (Qwen 3.5 MoE save/load prequantized)
         python -m pytest examples/models/qwen3_5_moe/test_quantize_roundtrip.py -v -o "addopts="
 
-        # Build Qwen3.5 MoE runner (ExecuTorch already built above)
-        cd examples/models/qwen3_5_moe && cmake --workflow --preset qwen3-5-moe-cuda
-
   export-model-cuda-artifact:
     name: export-model-cuda-artifact
     # Skip this job if the pull request is from a fork (HuggingFace secrets are not available)
@@ -182,6 +179,8 @@ jobs:
             name: "parakeet-tdt"
           - repo: "facebook"
             name: "dinov2-small-imagenet1k-1-layer"
+          - repo: "SocialLocalMobile"
+            name: "Qwen3.5-35B-A3B-HQQ-INT4"
         quant:
           - "non-quantized"
           - "quantized-int4-tile-packed"
@@ -192,6 +191,15 @@ jobs:
               repo: "google"
               name: "gemma-3-4b-it"
             quant: "quantized-int4-weight-only"
+          # Qwen3.5 MoE uses a prequantized checkpoint, only tile-packed
+          - model:
+              repo: "SocialLocalMobile"
+              name: "Qwen3.5-35B-A3B-HQQ-INT4"
+            quant: "non-quantized"
+          - model:
+              repo: "SocialLocalMobile"
+              name: "Qwen3.5-35B-A3B-HQQ-INT4"
+            quant: "quantized-int4-weight-only"
           # Voxtral Realtime only supports int4-tile-packed on CUDA
           - model:
               repo: "mistralai"
@@ -246,7 +254,7 @@ jobs:
     with:
       timeout: 90
       secrets-env: EXECUTORCH_HF_TOKEN
-      runner: linux.g5.4xlarge.nvidia.gpu
+      runner: ${{ matrix.model.name == 'Qwen3.5-35B-A3B-HQQ-INT4' && 'linux.aws.a100' || 'linux.g5.4xlarge.nvidia.gpu' }}
       gpu-arch-type: cuda
       gpu-arch-version: 12.6
       use-custom-docker-registry: false
@@ -301,6 +309,8 @@ jobs:
             name: "parakeet-tdt"
           - repo: "facebook"
             name: "dinov2-small-imagenet1k-1-layer"
+          - repo: "SocialLocalMobile"
+            name: "Qwen3.5-35B-A3B-HQQ-INT4"
         quant:
           - "non-quantized"
           - "quantized-int4-tile-packed"
@@ -311,6 +321,15 @@ jobs:
               repo: "google"
               name: "gemma-3-4b-it"
             quant: "quantized-int4-weight-only"
+          # Qwen3.5 MoE uses a prequantized checkpoint, only tile-packed
+          - model:
+              repo: "SocialLocalMobile"
+              name: "Qwen3.5-35B-A3B-HQQ-INT4"
+            quant: "non-quantized"
+          - model:
+              repo: "SocialLocalMobile"
+              name: "Qwen3.5-35B-A3B-HQQ-INT4"
+            quant: "quantized-int4-weight-only"
           # Voxtral Realtime only supports int4-tile-packed on CUDA
           - model:
               repo: "mistralai"
@@ -359,7 +378,7 @@ jobs:
             quant: "non-quantized"
     with:
       timeout: 90
-      runner: linux.g5.4xlarge.nvidia.gpu
+      runner: ${{ matrix.model.name == 'Qwen3.5-35B-A3B-HQQ-INT4' && 'linux.aws.a100' || 'linux.g5.4xlarge.nvidia.gpu' }}
       gpu-arch-type: cuda
       gpu-arch-version: 12.6
       use-custom-docker-registry: false