From d753f03dc0fef76d086f924909c7b6809a8c9792 Mon Sep 17 00:00:00 2001
From: Bin Bao <binbao@meta.com>
Date: Wed, 15 Oct 2025 17:40:38 -0700
Subject: [PATCH 1/7] Add int4mm test to the CUDA CI flow

---
 .github/workflows/cuda.yml | 148 +++++++++++++++++++++++++++++++++++++
 1 file changed, 148 insertions(+)

diff --git a/.github/workflows/cuda.yml b/.github/workflows/cuda.yml
index 8a9dc101ff3..15885fb155f 100644
--- a/.github/workflows/cuda.yml
+++ b/.github/workflows/cuda.yml
@@ -275,6 +275,154 @@ jobs:
           exit 1
         fi
 
+        if [ $EXIT_CODE -ne 0 ]; then
+          echo "Unexpected exit code: $EXIT_CODE"
+          exit $EXIT_CODE
+        fi
+          echo "::endgroup::"
+
+  export-voxtral-cuda-quantized:
+    name: export-voxtral-cuda-quantized
+    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
+    permissions:
+      id-token: write
+      contents: read
+    secrets: inherit
+    strategy:
+      fail-fast: false
+    with:
+      timeout: 90
+      secrets-env: EXECUTORCH_HF_TOKEN
+      runner: linux.g5.4xlarge.nvidia.gpu
+      gpu-arch-type: cuda
+      gpu-arch-version: 12.6
+      use-custom-docker-registry: false
+      submodules: recursive
+      upload-artifact: voxtral-cuda-quantized-export
+      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      script: |
+        set -eux
+
+        echo "::group::Setup ExecuTorch"
+        ./install_executorch.sh
+        echo "::endgroup::"
+
+        echo "::group::Setup Huggingface"
+        pip install -U "huggingface_hub[cli]" accelerate
+        huggingface-cli login --token $SECRET_EXECUTORCH_HF_TOKEN
+        OPTIMUM_ET_VERSION=$(cat .ci/docker/ci_commit_pins/optimum-executorch.txt)
+        pip install git+https://github.com/huggingface/optimum-executorch.git@${OPTIMUM_ET_VERSION}
+        pip install mistral-common librosa
+        pip list
+        echo "::endgroup::"
+
+        echo "::group::Export Voxtral with Quantization"
+        optimum-cli export executorch \
+            --model "mistralai/Voxtral-Mini-3B-2507" \
+            --task "multimodal-text-to-text" \
+            --recipe "cuda" \
+            --dtype bfloat16 \
+            --device cuda \
+            --max_seq_len 1024 \
+            --qlinear 4w \
+            --qlinear_encoder 4w \
+            --qlinear_packing_format tile_packed_to_4d \
+            --qlinear_encoder_packing_format tile_packed_to_4d \
+            --output_dir ./
+        python -m executorch.extension.audio.mel_spectrogram \
+            --feature_size 128 \
+            --stack_output \
+            --max_audio_len 300 \
+            --output_file voxtral_preprocessor.pte
+
+        test -f model.pte
+        test -f aoti_cuda_blob.ptd
+        test -f voxtral_preprocessor.pte
+        echo "::endgroup::"
+
+        echo "::group::Store Voxtral Quantized Artifacts"
+        mkdir -p "${RUNNER_ARTIFACT_DIR}"
+        cp model.pte "${RUNNER_ARTIFACT_DIR}/"
+        cp aoti_cuda_blob.ptd "${RUNNER_ARTIFACT_DIR}/"
+        cp voxtral_preprocessor.pte "${RUNNER_ARTIFACT_DIR}/"
+        ls -al "${RUNNER_ARTIFACT_DIR}"
+        echo "::endgroup::"
+
+  test-voxtral-cuda-quantized-e2e:
+    name: test-voxtral-cuda-quantized-e2e
+    needs: export-voxtral-cuda-quantized
+    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
+    permissions:
+      id-token: write
+      contents: read
+    strategy:
+      fail-fast: false
+    with:
+      timeout: 90
+      runner: linux.g5.4xlarge.nvidia.gpu
+      gpu-arch-type: cuda
+      gpu-arch-version: 12.6
+      use-custom-docker-registry: false
+      submodules: recursive
+      download-artifact: voxtral-cuda-quantized-export
+      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      script: |
+        set -eux
+
+        echo "::group::Setup ExecuTorch Requirements"
+        CMAKE_ARGS="-DEXECUTORCH_BUILD_CUDA=ON" ./install_requirements.sh
+        pip list
+        echo "::endgroup::"
+
+        echo "::group::Prepare Voxtral Quantized Artifacts"
+        cp "${RUNNER_ARTIFACT_DIR}/model.pte" .
+        cp "${RUNNER_ARTIFACT_DIR}/aoti_cuda_blob.ptd" .
+        cp "${RUNNER_ARTIFACT_DIR}/voxtral_preprocessor.pte" .
+        TOKENIZER_URL="https://huggingface.co/mistralai/Voxtral-Mini-3B-2507/resolve/main/tekken.json"
+        curl -L $TOKENIZER_URL -o tekken.json
+        ls -al model.pte aoti_cuda_blob.ptd voxtral_preprocessor.pte tekken.json
+        echo "::endgroup::"
+
+        echo "::group::Download Test Audio File"
+        AUDIO_URL="https://github.com/voxserv/audio_quality_testing_samples/raw/refs/heads/master/testaudio/16000/test01_20s.wav"
+        curl -L $AUDIO_URL -o poem.wav
+        echo "::endgroup::"
+
+        echo "::group::Build Voxtral Runner"
+        cmake --preset llm \
+              -DEXECUTORCH_BUILD_CUDA=ON \
+              -DCMAKE_INSTALL_PREFIX=cmake-out \
+              -DCMAKE_BUILD_TYPE=Release \
+              -Bcmake-out -S.
+        cmake --build cmake-out -j$(( $(nproc) - 1 )) --target install --config Release
+
+        cmake -DEXECUTORCH_BUILD_CUDA=ON \
+              -DCMAKE_BUILD_TYPE=Release \
+              -Sexamples/models/voxtral \
+              -Bcmake-out/examples/models/voxtral/
+        cmake --build cmake-out/examples/models/voxtral --target voxtral_runner --config Release
+        echo "::endgroup::"
+
+        echo "::group::Run Voxtral Runner with Quantized Model"
+        set +e
+        export LD_LIBRARY_PATH=/opt/conda/lib:$LD_LIBRARY_PATH
+        OUTPUT=$(cmake-out/examples/models/voxtral/voxtral_runner \
+              --model_path model.pte \
+              --data_path aoti_cuda_blob.ptd \
+              --tokenizer_path tekken.json \
+              --audio_path poem.wav \
+              --processor_path voxtral_preprocessor.pte \
+              --temperature 0 2>&1)
+        EXIT_CODE=$?
+        set -e
+
+        echo "$OUTPUT"
+
+        if ! echo "$OUTPUT" | grep -iq "poem"; then
+          echo "Expected output 'poem' not found in output"
+          exit 1
+        fi
+
         if [ $EXIT_CODE -ne 0 ]; then
           echo "Unexpected exit code: $EXIT_CODE"
           exit $EXIT_CODE

From 8456310f457d790eb0ec167f7bfca52a662bc9bc Mon Sep 17 00:00:00 2001
From: Bin Bao <binbao@meta.com>
Date: Wed, 15 Oct 2025 18:40:43 -0700
Subject: [PATCH 2/7] Update optimum-executorch pin

---
 .ci/docker/ci_commit_pins/optimum-executorch.txt | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/.ci/docker/ci_commit_pins/optimum-executorch.txt b/.ci/docker/ci_commit_pins/optimum-executorch.txt
index 49b079047a3..3c085a7ef3a 100644
--- a/.ci/docker/ci_commit_pins/optimum-executorch.txt
+++ b/.ci/docker/ci_commit_pins/optimum-executorch.txt
@@ -1 +1 @@
-44d8d54e38c0258357d4e92e1fefe21e845947a3
+09fdbd0a0639b128f712a4f5202ed42ca4c60957

From fcbefe312eb20a481e53730b83829778def3da43 Mon Sep 17 00:00:00 2001
From: Bin Bao <binbao@meta.com>
Date: Thu, 16 Oct 2025 13:45:27 -0700
Subject: [PATCH 3/7] 1. Test more quantization format 2. Update
 models/voxtral/READEME.md with CUDA quantization

---
 .github/workflows/cuda.yml        | 114 +++++++++++++++++++++++-------
 examples/models/voxtral/README.md |  18 ++++-
 2 files changed, 105 insertions(+), 27 deletions(-)

diff --git a/.github/workflows/cuda.yml b/.github/workflows/cuda.yml
index 15885fb155f..6d39c808458 100644
--- a/.github/workflows/cuda.yml
+++ b/.github/workflows/cuda.yml
@@ -281,8 +281,8 @@ jobs:
         fi
           echo "::endgroup::"
 
-  export-voxtral-cuda-quantized:
-    name: export-voxtral-cuda-quantized
+  test-voxtral-cuda-quantized-4w-tile-packed:
+    name: test-voxtral-cuda-quantized-4w-tile-packed
     uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
     permissions:
       id-token: write
@@ -298,7 +298,6 @@ jobs:
       gpu-arch-version: 12.6
       use-custom-docker-registry: false
       submodules: recursive
-      upload-artifact: voxtral-cuda-quantized-export
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
       script: |
         set -eux
@@ -316,7 +315,7 @@ jobs:
         pip list
         echo "::endgroup::"
 
-        echo "::group::Export Voxtral with Quantization"
+        echo "::group::Export Voxtral with Quantization (4w-tile-packed)"
         optimum-cli export executorch \
             --model "mistralai/Voxtral-Mini-3B-2507" \
             --task "multimodal-text-to-text" \
@@ -328,6 +327,7 @@ jobs:
             --qlinear_encoder 4w \
             --qlinear_packing_format tile_packed_to_4d \
             --qlinear_encoder_packing_format tile_packed_to_4d \
+            --qembedding 4w \
             --output_dir ./
         python -m executorch.extension.audio.mel_spectrogram \
             --feature_size 128 \
@@ -340,50 +340,114 @@ jobs:
         test -f voxtral_preprocessor.pte
         echo "::endgroup::"
 
-        echo "::group::Store Voxtral Quantized Artifacts"
-        mkdir -p "${RUNNER_ARTIFACT_DIR}"
-        cp model.pte "${RUNNER_ARTIFACT_DIR}/"
-        cp aoti_cuda_blob.ptd "${RUNNER_ARTIFACT_DIR}/"
-        cp voxtral_preprocessor.pte "${RUNNER_ARTIFACT_DIR}/"
-        ls -al "${RUNNER_ARTIFACT_DIR}"
+        echo "::group::Download Tokenizer and Test Audio"
+        TOKENIZER_URL="https://huggingface.co/mistralai/Voxtral-Mini-3B-2507/resolve/main/tekken.json"
+        curl -L $TOKENIZER_URL -o tekken.json
+        AUDIO_URL="https://github.com/voxserv/audio_quality_testing_samples/raw/refs/heads/master/testaudio/16000/test01_20s.wav"
+        curl -L $AUDIO_URL -o poem.wav
+        echo "::endgroup::"
+
+        echo "::group::Build Voxtral Runner"
+        cmake --preset llm \
+              -DEXECUTORCH_BUILD_CUDA=ON \
+              -DCMAKE_INSTALL_PREFIX=cmake-out \
+              -DCMAKE_BUILD_TYPE=Release \
+              -Bcmake-out -S.
+        cmake --build cmake-out -j$(( $(nproc) - 1 )) --target install --config Release
+
+        cmake -DEXECUTORCH_BUILD_CUDA=ON \
+              -DCMAKE_BUILD_TYPE=Release \
+              -Sexamples/models/voxtral \
+              -Bcmake-out/examples/models/voxtral/
+        cmake --build cmake-out/examples/models/voxtral --target voxtral_runner --config Release
         echo "::endgroup::"
 
-  test-voxtral-cuda-quantized-e2e:
-    name: test-voxtral-cuda-quantized-e2e
-    needs: export-voxtral-cuda-quantized
+        echo "::group::Run Voxtral Runner with Quantized Model (4w-tile-packed)"
+        set +e
+        export LD_LIBRARY_PATH=/opt/conda/lib:$LD_LIBRARY_PATH
+        OUTPUT=$(cmake-out/examples/models/voxtral/voxtral_runner \
+              --model_path model.pte \
+              --data_path aoti_cuda_blob.ptd \
+              --tokenizer_path tekken.json \
+              --audio_path poem.wav \
+              --processor_path voxtral_preprocessor.pte \
+              --temperature 0 2>&1)
+        EXIT_CODE=$?
+        set -e
+
+        echo "$OUTPUT"
+
+        if ! echo "$OUTPUT" | grep -iq "poem"; then
+          echo "Expected output 'poem' not found in output"
+          exit 1
+        fi
+
+        if [ $EXIT_CODE -ne 0 ]; then
+          echo "Unexpected exit code: $EXIT_CODE"
+          exit $EXIT_CODE
+        fi
+        echo "::endgroup::"
+
+  test-voxtral-cuda-quantized-4w-weight-only:
+    name: test-voxtral-cuda-quantized-4w-weight-only
     uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
     permissions:
       id-token: write
       contents: read
+    secrets: inherit
     strategy:
       fail-fast: false
     with:
       timeout: 90
+      secrets-env: EXECUTORCH_HF_TOKEN
       runner: linux.g5.4xlarge.nvidia.gpu
       gpu-arch-type: cuda
       gpu-arch-version: 12.6
       use-custom-docker-registry: false
       submodules: recursive
-      download-artifact: voxtral-cuda-quantized-export
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
       script: |
         set -eux
 
-        echo "::group::Setup ExecuTorch Requirements"
-        CMAKE_ARGS="-DEXECUTORCH_BUILD_CUDA=ON" ./install_requirements.sh
+        echo "::group::Setup ExecuTorch"
+        ./install_executorch.sh
+        echo "::endgroup::"
+
+        echo "::group::Setup Huggingface"
+        pip install -U "huggingface_hub[cli]" accelerate
+        huggingface-cli login --token $SECRET_EXECUTORCH_HF_TOKEN
+        OPTIMUM_ET_VERSION=$(cat .ci/docker/ci_commit_pins/optimum-executorch.txt)
+        pip install git+https://github.com/huggingface/optimum-executorch.git@${OPTIMUM_ET_VERSION}
+        pip install mistral-common librosa
         pip list
         echo "::endgroup::"
 
-        echo "::group::Prepare Voxtral Quantized Artifacts"
-        cp "${RUNNER_ARTIFACT_DIR}/model.pte" .
-        cp "${RUNNER_ARTIFACT_DIR}/aoti_cuda_blob.ptd" .
-        cp "${RUNNER_ARTIFACT_DIR}/voxtral_preprocessor.pte" .
-        TOKENIZER_URL="https://huggingface.co/mistralai/Voxtral-Mini-3B-2507/resolve/main/tekken.json"
-        curl -L $TOKENIZER_URL -o tekken.json
-        ls -al model.pte aoti_cuda_blob.ptd voxtral_preprocessor.pte tekken.json
+        echo "::group::Export Voxtral with Quantization (4w-weight-only)"
+        optimum-cli export executorch \
+            --model "mistralai/Voxtral-Mini-3B-2507" \
+            --task "multimodal-text-to-text" \
+            --recipe "cuda" \
+            --dtype bfloat16 \
+            --device cuda \
+            --max_seq_len 1024 \
+            --qlinear 4w \
+            --qlinear_encoder 4w \
+            --qembedding 4w \
+            --output_dir ./
+        python -m executorch.extension.audio.mel_spectrogram \
+            --feature_size 128 \
+            --stack_output \
+            --max_audio_len 300 \
+            --output_file voxtral_preprocessor.pte
+
+        test -f model.pte
+        test -f aoti_cuda_blob.ptd
+        test -f voxtral_preprocessor.pte
         echo "::endgroup::"
 
-        echo "::group::Download Test Audio File"
+        echo "::group::Download Tokenizer and Test Audio"
+        TOKENIZER_URL="https://huggingface.co/mistralai/Voxtral-Mini-3B-2507/resolve/main/tekken.json"
+        curl -L $TOKENIZER_URL -o tekken.json
         AUDIO_URL="https://github.com/voxserv/audio_quality_testing_samples/raw/refs/heads/master/testaudio/16000/test01_20s.wav"
         curl -L $AUDIO_URL -o poem.wav
         echo "::endgroup::"
@@ -403,7 +467,7 @@ jobs:
         cmake --build cmake-out/examples/models/voxtral --target voxtral_runner --config Release
         echo "::endgroup::"
 
-        echo "::group::Run Voxtral Runner with Quantized Model"
+        echo "::group::Run Voxtral Runner with Quantized Model (4w-weight-only)"
         set +e
         export LD_LIBRARY_PATH=/opt/conda/lib:$LD_LIBRARY_PATH
         OUTPUT=$(cmake-out/examples/models/voxtral/voxtral_runner \
diff --git a/examples/models/voxtral/README.md b/examples/models/voxtral/README.md
index 861043fe2a7..1c211f2dcb3 100644
--- a/examples/models/voxtral/README.md
+++ b/examples/models/voxtral/README.md
@@ -39,8 +39,6 @@ This exports Voxtral with XNNPack backend acceleration and 4-bit weight/8-bit ac
 ## CUDA Support
 If your environment has CUDA support, you can enable the runner to run on CUDA for improved performance. Follow the export and runtime commands below:
 
-**Note:** We are currently working on quantization support for CUDA. Currently, only bfloat16 dtype is supported for CUDA execution.
-
 ### Exporting with CUDA
 ```
 optimum-cli export executorch \
@@ -57,6 +55,22 @@ This will generate:
 - `model.pte` - The exported model
 - `aoti_cuda_blob.ptd` - The CUDA kernel blob required for runtime
 
+Furthermore, we support several quantization formats on CUDA.
+To export Voxtral with int4 weights-only quantization, use
+```
+optimum-cli export executorch \
+  --model "mistralai/Voxtral-Mini-3B-2507" \
+  --task "multimodal-text-to-text" \
+  --recipe "cuda" \
+  --dtype bfloat16 \
+  --device cuda \
+  --max_seq_len 1024 \
+  --qlinear 4w \
+  --qlinear_encoder 4w \
+  --qembedding 4w \
+  --output_dir="voxtral"
+```
+
 See the "Building the multimodal runner" section below for instructions on building with CUDA support, and the "Running the model" section for runtime instructions.
 
 # Running the model

From 95e2e5193b802d5f2acd915b1938a1e11aa8ec2e Mon Sep 17 00:00:00 2001
From: Bin Bao <binbao@meta.com>
Date: Thu, 16 Oct 2025 14:11:26 -0700
Subject: [PATCH 4/7] Consolidate e2e workflows

---
 .github/workflows/cuda.yml        | 203 ++++++++++--------------------
 examples/models/voxtral/README.md |   2 +-
 2 files changed, 69 insertions(+), 136 deletions(-)

diff --git a/.github/workflows/cuda.yml b/.github/workflows/cuda.yml
index 6d39c808458..0ca0a42c2e5 100644
--- a/.github/workflows/cuda.yml
+++ b/.github/workflows/cuda.yml
@@ -200,89 +200,8 @@ jobs:
 
         echo "::endgroup::"
 
-  test-voxtral-cuda-e2e:
-    name: test-voxtral-cuda-e2e
-    needs: export-voxtral-cuda-artifact
-    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
-    permissions:
-      id-token: write
-      contents: read
-    strategy:
-      fail-fast: false
-    with:
-      timeout: 90
-      runner: linux.g5.4xlarge.nvidia.gpu
-      gpu-arch-type: cuda
-      gpu-arch-version: 12.6
-      use-custom-docker-registry: false
-      submodules: recursive
-      download-artifact: voxtral-cuda-export
-      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
-      script: |
-        set -eux
-
-        echo "::group::Setup ExecuTorch Requirements"
-        CMAKE_ARGS="-DEXECUTORCH_BUILD_CUDA=ON" ./install_requirements.sh
-        pip list
-        echo "::endgroup::"
-
-        echo "::group::Prepare Voxtral Artifacts"
-        cp "${RUNNER_ARTIFACT_DIR}/model.pte" .
-        cp "${RUNNER_ARTIFACT_DIR}/aoti_cuda_blob.ptd" .
-        cp "${RUNNER_ARTIFACT_DIR}/voxtral_preprocessor.pte" .
-        TOKENIZER_URL="https://huggingface.co/mistralai/Voxtral-Mini-3B-2507/resolve/main/tekken.json"
-        curl -L $TOKENIZER_URL -o tekken.json
-        ls -al model.pte aoti_cuda_blob.ptd voxtral_preprocessor.pte tekken.json
-        echo "::endgroup::"
-
-        echo "::group::Download Test Audio File"
-        AUDIO_URL="https://github.com/voxserv/audio_quality_testing_samples/raw/refs/heads/master/testaudio/16000/test01_20s.wav"
-        curl -L $AUDIO_URL -o poem.wav
-        echo "::endgroup::"
-
-        echo "::group::Build Voxtral Runner"
-        cmake --preset llm \
-              -DEXECUTORCH_BUILD_CUDA=ON \
-              -DCMAKE_INSTALL_PREFIX=cmake-out \
-              -DCMAKE_BUILD_TYPE=Release \
-              -Bcmake-out -S.
-        cmake --build cmake-out -j$(( $(nproc) - 1 )) --target install --config Release
-
-        cmake -DEXECUTORCH_BUILD_CUDA=ON \
-              -DCMAKE_BUILD_TYPE=Release \
-              -Sexamples/models/voxtral \
-              -Bcmake-out/examples/models/voxtral/
-        cmake --build cmake-out/examples/models/voxtral --target voxtral_runner --config Release
-        echo "::endgroup::"
-
-        echo "::group::Run Voxtral Runner"
-        set +e
-        export LD_LIBRARY_PATH=/opt/conda/lib:$LD_LIBRARY_PATH
-        OUTPUT=$(cmake-out/examples/models/voxtral/voxtral_runner \
-              --model_path model.pte \
-              --data_path aoti_cuda_blob.ptd \
-              --tokenizer_path tekken.json \
-              --audio_path poem.wav \
-              --processor_path voxtral_preprocessor.pte \
-              --temperature 0 2>&1)
-        EXIT_CODE=$?
-        set -e
-
-        echo "$OUTPUT"
-
-        if ! echo "$OUTPUT" | grep -iq "poem"; then
-          echo "Expected output 'poem' not found in output"
-          exit 1
-        fi
-
-        if [ $EXIT_CODE -ne 0 ]; then
-          echo "Unexpected exit code: $EXIT_CODE"
-          exit $EXIT_CODE
-        fi
-          echo "::endgroup::"
-
-  test-voxtral-cuda-quantized-4w-tile-packed:
-    name: test-voxtral-cuda-quantized-4w-tile-packed
+  export-voxtral-cuda-quantized-int4-tile-packed:
+    name: export-voxtral-cuda-quantized-int4-tile-packed
     uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
     permissions:
       id-token: write
@@ -298,6 +217,7 @@ jobs:
       gpu-arch-version: 12.6
       use-custom-docker-registry: false
       submodules: recursive
+      upload-artifact: voxtral-cuda-quantized-int4-tile-packed
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
       script: |
         set -eux
@@ -315,7 +235,7 @@ jobs:
         pip list
         echo "::endgroup::"
 
-        echo "::group::Export Voxtral with Quantization (4w-tile-packed)"
+        echo "::group::Export Voxtral with Quantization (int4-tile-packed)"
         optimum-cli export executorch \
             --model "mistralai/Voxtral-Mini-3B-2507" \
             --task "multimodal-text-to-text" \
@@ -340,56 +260,16 @@ jobs:
         test -f voxtral_preprocessor.pte
         echo "::endgroup::"
 
-        echo "::group::Download Tokenizer and Test Audio"
-        TOKENIZER_URL="https://huggingface.co/mistralai/Voxtral-Mini-3B-2507/resolve/main/tekken.json"
-        curl -L $TOKENIZER_URL -o tekken.json
-        AUDIO_URL="https://github.com/voxserv/audio_quality_testing_samples/raw/refs/heads/master/testaudio/16000/test01_20s.wav"
-        curl -L $AUDIO_URL -o poem.wav
-        echo "::endgroup::"
-
-        echo "::group::Build Voxtral Runner"
-        cmake --preset llm \
-              -DEXECUTORCH_BUILD_CUDA=ON \
-              -DCMAKE_INSTALL_PREFIX=cmake-out \
-              -DCMAKE_BUILD_TYPE=Release \
-              -Bcmake-out -S.
-        cmake --build cmake-out -j$(( $(nproc) - 1 )) --target install --config Release
-
-        cmake -DEXECUTORCH_BUILD_CUDA=ON \
-              -DCMAKE_BUILD_TYPE=Release \
-              -Sexamples/models/voxtral \
-              -Bcmake-out/examples/models/voxtral/
-        cmake --build cmake-out/examples/models/voxtral --target voxtral_runner --config Release
-        echo "::endgroup::"
-
-        echo "::group::Run Voxtral Runner with Quantized Model (4w-tile-packed)"
-        set +e
-        export LD_LIBRARY_PATH=/opt/conda/lib:$LD_LIBRARY_PATH
-        OUTPUT=$(cmake-out/examples/models/voxtral/voxtral_runner \
-              --model_path model.pte \
-              --data_path aoti_cuda_blob.ptd \
-              --tokenizer_path tekken.json \
-              --audio_path poem.wav \
-              --processor_path voxtral_preprocessor.pte \
-              --temperature 0 2>&1)
-        EXIT_CODE=$?
-        set -e
-
-        echo "$OUTPUT"
-
-        if ! echo "$OUTPUT" | grep -iq "poem"; then
-          echo "Expected output 'poem' not found in output"
-          exit 1
-        fi
-
-        if [ $EXIT_CODE -ne 0 ]; then
-          echo "Unexpected exit code: $EXIT_CODE"
-          exit $EXIT_CODE
-        fi
+        echo "::group::Store Voxtral Quantized Artifacts (int4-tile-packed)"
+        mkdir -p "${RUNNER_ARTIFACT_DIR}"
+        cp model.pte "${RUNNER_ARTIFACT_DIR}/"
+        cp aoti_cuda_blob.ptd "${RUNNER_ARTIFACT_DIR}/"
+        cp voxtral_preprocessor.pte "${RUNNER_ARTIFACT_DIR}/"
+        ls -al "${RUNNER_ARTIFACT_DIR}"
         echo "::endgroup::"
 
-  test-voxtral-cuda-quantized-4w-weight-only:
-    name: test-voxtral-cuda-quantized-4w-weight-only
+  export-voxtral-cuda-quantized-int4-weight-only:
+    name: export-voxtral-cuda-quantized-int4-weight-only
     uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
     permissions:
       id-token: write
@@ -405,6 +285,7 @@ jobs:
       gpu-arch-version: 12.6
       use-custom-docker-registry: false
       submodules: recursive
+      upload-artifact: voxtral-cuda-quantized-int4-weight-only
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
       script: |
         set -eux
@@ -422,7 +303,7 @@ jobs:
         pip list
         echo "::endgroup::"
 
-        echo "::group::Export Voxtral with Quantization (4w-weight-only)"
+        echo "::group::Export Voxtral with Quantization (int4-weight-only)"
         optimum-cli export executorch \
             --model "mistralai/Voxtral-Mini-3B-2507" \
             --task "multimodal-text-to-text" \
@@ -445,9 +326,61 @@ jobs:
         test -f voxtral_preprocessor.pte
         echo "::endgroup::"
 
-        echo "::group::Download Tokenizer and Test Audio"
+        echo "::group::Store Voxtral Quantized Artifacts (int4-weight-only)"
+        mkdir -p "${RUNNER_ARTIFACT_DIR}"
+        cp model.pte "${RUNNER_ARTIFACT_DIR}/"
+        cp aoti_cuda_blob.ptd "${RUNNER_ARTIFACT_DIR}/"
+        cp voxtral_preprocessor.pte "${RUNNER_ARTIFACT_DIR}/"
+        ls -al "${RUNNER_ARTIFACT_DIR}"
+        echo "::endgroup::"
+
+  test-voxtral-cuda-e2e:
+    name: test-voxtral-cuda-e2e-${{ matrix.format.name }}
+    needs:
+      - export-voxtral-cuda-artifact
+      - export-voxtral-cuda-quantized-int4-tile-packed
+      - export-voxtral-cuda-quantized-int4-weight-only
+    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
+    permissions:
+      id-token: write
+      contents: read
+    strategy:
+      fail-fast: false
+      matrix:
+        format:
+          - name: "non-quantized"
+            artifact: "voxtral-cuda-export"
+          - name: "quantized-int4-tile-packed"
+            artifact: "voxtral-cuda-quantized-int4-tile-packed"
+          - name: "quantized-int4-weight-only"
+            artifact: "voxtral-cuda-quantized-int4-weight-only"
+    with:
+      timeout: 90
+      runner: linux.g5.4xlarge.nvidia.gpu
+      gpu-arch-type: cuda
+      gpu-arch-version: 12.6
+      use-custom-docker-registry: false
+      submodules: recursive
+      download-artifact: ${{ matrix.format.artifact }}
+      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      script: |
+        set -eux
+
+        echo "::group::Setup ExecuTorch Requirements"
+        CMAKE_ARGS="-DEXECUTORCH_BUILD_CUDA=ON" ./install_requirements.sh
+        pip list
+        echo "::endgroup::"
+
+        echo "::group::Prepare Voxtral Artifacts (${{ matrix.format.name }})"
+        cp "${RUNNER_ARTIFACT_DIR}/model.pte" .
+        cp "${RUNNER_ARTIFACT_DIR}/aoti_cuda_blob.ptd" .
+        cp "${RUNNER_ARTIFACT_DIR}/voxtral_preprocessor.pte" .
         TOKENIZER_URL="https://huggingface.co/mistralai/Voxtral-Mini-3B-2507/resolve/main/tekken.json"
         curl -L $TOKENIZER_URL -o tekken.json
+        ls -al model.pte aoti_cuda_blob.ptd voxtral_preprocessor.pte tekken.json
+        echo "::endgroup::"
+
+        echo "::group::Download Test Audio File"
         AUDIO_URL="https://github.com/voxserv/audio_quality_testing_samples/raw/refs/heads/master/testaudio/16000/test01_20s.wav"
         curl -L $AUDIO_URL -o poem.wav
         echo "::endgroup::"
@@ -467,7 +400,7 @@ jobs:
         cmake --build cmake-out/examples/models/voxtral --target voxtral_runner --config Release
         echo "::endgroup::"
 
-        echo "::group::Run Voxtral Runner with Quantized Model (4w-weight-only)"
+        echo "::group::Run Voxtral Runner (${{ matrix.format.name }})"
         set +e
         export LD_LIBRARY_PATH=/opt/conda/lib:$LD_LIBRARY_PATH
         OUTPUT=$(cmake-out/examples/models/voxtral/voxtral_runner \
diff --git a/examples/models/voxtral/README.md b/examples/models/voxtral/README.md
index 1c211f2dcb3..51a68e4f8f6 100644
--- a/examples/models/voxtral/README.md
+++ b/examples/models/voxtral/README.md
@@ -56,7 +56,7 @@ This will generate:
 - `aoti_cuda_blob.ptd` - The CUDA kernel blob required for runtime
 
 Furthermore, we support several quantization formats on CUDA.
-To export Voxtral with int4 weights-only quantization, use
+To export Voxtral with int4 weight-only quantization, use
 ```
 optimum-cli export executorch \
   --model "mistralai/Voxtral-Mini-3B-2507" \

From d27e10aa4edb8b7ab1b77f2689490b95acad2652 Mon Sep 17 00:00:00 2001
From: Bin Bao <binbao@meta.com>
Date: Thu, 16 Oct 2025 18:58:46 -0700
Subject: [PATCH 5/7] Remove --qembedding

---
 .github/workflows/cuda.yml | 2 --
 1 file changed, 2 deletions(-)

diff --git a/.github/workflows/cuda.yml b/.github/workflows/cuda.yml
index 0ca0a42c2e5..15e2e0ef960 100644
--- a/.github/workflows/cuda.yml
+++ b/.github/workflows/cuda.yml
@@ -247,7 +247,6 @@ jobs:
             --qlinear_encoder 4w \
             --qlinear_packing_format tile_packed_to_4d \
             --qlinear_encoder_packing_format tile_packed_to_4d \
-            --qembedding 4w \
             --output_dir ./
         python -m executorch.extension.audio.mel_spectrogram \
             --feature_size 128 \
@@ -313,7 +312,6 @@ jobs:
             --max_seq_len 1024 \
             --qlinear 4w \
             --qlinear_encoder 4w \
-            --qembedding 4w \
             --output_dir ./
         python -m executorch.extension.audio.mel_spectrogram \
             --feature_size 128 \

From 8cd4c6a6578fbdac9f3a228dd0e84e62284d50bb Mon Sep 17 00:00:00 2001
From: Bin Bao <binbao@meta.com>
Date: Fri, 17 Oct 2025 15:49:49 -0700
Subject: [PATCH 6/7] Remove --qlinear as the produced result is invalid

---
 .github/workflows/cuda.yml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/.github/workflows/cuda.yml b/.github/workflows/cuda.yml
index 15e2e0ef960..1e4849a8843 100644
--- a/.github/workflows/cuda.yml
+++ b/.github/workflows/cuda.yml
@@ -303,6 +303,7 @@ jobs:
         echo "::endgroup::"
 
         echo "::group::Export Voxtral with Quantization (int4-weight-only)"
+        # TODO: --qlinear 4w will produce wrong result. Need more investigation.
         optimum-cli export executorch \
             --model "mistralai/Voxtral-Mini-3B-2507" \
             --task "multimodal-text-to-text" \
@@ -310,7 +311,6 @@ jobs:
             --dtype bfloat16 \
             --device cuda \
             --max_seq_len 1024 \
-            --qlinear 4w \
             --qlinear_encoder 4w \
             --output_dir ./
         python -m executorch.extension.audio.mel_spectrogram \

From 45cbb8457519de506d3610a220466b6c7fd61078 Mon Sep 17 00:00:00 2001
From: Bin Bao <binbao@meta.com>
Date: Fri, 17 Oct 2025 16:22:42 -0700
Subject: [PATCH 7/7] Consolidate export workflow

---
 .github/workflows/cuda.yml        | 159 ++++--------------------------
 examples/models/voxtral/README.md |   5 +-
 2 files changed, 22 insertions(+), 142 deletions(-)

diff --git a/.github/workflows/cuda.yml b/.github/workflows/cuda.yml
index 1e4849a8843..9ee72a34ef0 100644
--- a/.github/workflows/cuda.yml
+++ b/.github/workflows/cuda.yml
@@ -88,7 +88,7 @@ jobs:
         PYTHON_EXECUTABLE=python source .ci/scripts/test_model.sh "${{ matrix.model }}" cmake cuda
 
   export-voxtral-cuda-artifact:
-    name: export-voxtral-cuda-artifact
+    name: export-voxtral-cuda-${{ matrix.quant.name }}
     uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
     permissions:
       id-token: write
@@ -96,6 +96,18 @@ jobs:
     secrets: inherit
     strategy:
       fail-fast: false
+      matrix:
+        quant:
+          - name: "non-quantized"
+            artifact: "voxtral-cuda-export"
+            extra_args: ""
+          - name: "quantized-int4-tile-packed"
+            artifact: "voxtral-cuda-quantized-int4-tile-packed"
+            extra_args: "--qlinear 4w --qlinear_encoder 4w --qlinear_packing_format tile_packed_to_4d --qlinear_encoder_packing_format tile_packed_to_4d"
+          - name: "quantized-int4-weight-only"
+            artifact: "voxtral-cuda-quantized-int4-weight-only"
+            # TODO: adding "--qlinear 4w" produces invalid results. Need further investigation.
+            extra_args: "--qlinear_encoder 4w"
     with:
       timeout: 90
       secrets-env: EXECUTORCH_HF_TOKEN
@@ -104,7 +116,7 @@ jobs:
       gpu-arch-version: 12.6
       use-custom-docker-registry: false
       submodules: recursive
-      upload-artifact: voxtral-cuda-export
+      upload-artifact: ${{ matrix.quant.artifact }}
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
       script: |
         set -eux
@@ -122,7 +134,8 @@ jobs:
         pip list
         echo "::endgroup::"
 
-        echo "::group::Export Voxtral"
+        echo "::group::Export Voxtral (${{ matrix.quant.name }})"
+        EXTRA_ARGS="${{ matrix.quant.extra_args }}"
         optimum-cli export executorch \
             --model "mistralai/Voxtral-Mini-3B-2507" \
             --task "multimodal-text-to-text" \
@@ -130,6 +143,7 @@ jobs:
             --dtype bfloat16 \
             --device cuda \
             --max_seq_len 1024 \
+            ${EXTRA_ARGS} \
             --output_dir ./
         python -m executorch.extension.audio.mel_spectrogram \
             --feature_size 128 \
@@ -142,7 +156,7 @@ jobs:
         test -f voxtral_preprocessor.pte
         echo "::endgroup::"
 
-        echo "::group::Store Voxtral Artifacts"
+        echo "::group::Store Voxtral Artifacts (${{ matrix.quant.name }})"
         mkdir -p "${RUNNER_ARTIFACT_DIR}"
         cp model.pte "${RUNNER_ARTIFACT_DIR}/"
         cp aoti_cuda_blob.ptd "${RUNNER_ARTIFACT_DIR}/"
@@ -200,144 +214,9 @@ jobs:
 
         echo "::endgroup::"
 
-  export-voxtral-cuda-quantized-int4-tile-packed:
-    name: export-voxtral-cuda-quantized-int4-tile-packed
-    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
-    permissions:
-      id-token: write
-      contents: read
-    secrets: inherit
-    strategy:
-      fail-fast: false
-    with:
-      timeout: 90
-      secrets-env: EXECUTORCH_HF_TOKEN
-      runner: linux.g5.4xlarge.nvidia.gpu
-      gpu-arch-type: cuda
-      gpu-arch-version: 12.6
-      use-custom-docker-registry: false
-      submodules: recursive
-      upload-artifact: voxtral-cuda-quantized-int4-tile-packed
-      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
-      script: |
-        set -eux
-
-        echo "::group::Setup ExecuTorch"
-        ./install_executorch.sh
-        echo "::endgroup::"
-
-        echo "::group::Setup Huggingface"
-        pip install -U "huggingface_hub[cli]" accelerate
-        huggingface-cli login --token $SECRET_EXECUTORCH_HF_TOKEN
-        OPTIMUM_ET_VERSION=$(cat .ci/docker/ci_commit_pins/optimum-executorch.txt)
-        pip install git+https://github.com/huggingface/optimum-executorch.git@${OPTIMUM_ET_VERSION}
-        pip install mistral-common librosa
-        pip list
-        echo "::endgroup::"
-
-        echo "::group::Export Voxtral with Quantization (int4-tile-packed)"
-        optimum-cli export executorch \
-            --model "mistralai/Voxtral-Mini-3B-2507" \
-            --task "multimodal-text-to-text" \
-            --recipe "cuda" \
-            --dtype bfloat16 \
-            --device cuda \
-            --max_seq_len 1024 \
-            --qlinear 4w \
-            --qlinear_encoder 4w \
-            --qlinear_packing_format tile_packed_to_4d \
-            --qlinear_encoder_packing_format tile_packed_to_4d \
-            --output_dir ./
-        python -m executorch.extension.audio.mel_spectrogram \
-            --feature_size 128 \
-            --stack_output \
-            --max_audio_len 300 \
-            --output_file voxtral_preprocessor.pte
-
-        test -f model.pte
-        test -f aoti_cuda_blob.ptd
-        test -f voxtral_preprocessor.pte
-        echo "::endgroup::"
-
-        echo "::group::Store Voxtral Quantized Artifacts (int4-tile-packed)"
-        mkdir -p "${RUNNER_ARTIFACT_DIR}"
-        cp model.pte "${RUNNER_ARTIFACT_DIR}/"
-        cp aoti_cuda_blob.ptd "${RUNNER_ARTIFACT_DIR}/"
-        cp voxtral_preprocessor.pte "${RUNNER_ARTIFACT_DIR}/"
-        ls -al "${RUNNER_ARTIFACT_DIR}"
-        echo "::endgroup::"
-
-  export-voxtral-cuda-quantized-int4-weight-only:
-    name: export-voxtral-cuda-quantized-int4-weight-only
-    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
-    permissions:
-      id-token: write
-      contents: read
-    secrets: inherit
-    strategy:
-      fail-fast: false
-    with:
-      timeout: 90
-      secrets-env: EXECUTORCH_HF_TOKEN
-      runner: linux.g5.4xlarge.nvidia.gpu
-      gpu-arch-type: cuda
-      gpu-arch-version: 12.6
-      use-custom-docker-registry: false
-      submodules: recursive
-      upload-artifact: voxtral-cuda-quantized-int4-weight-only
-      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
-      script: |
-        set -eux
-
-        echo "::group::Setup ExecuTorch"
-        ./install_executorch.sh
-        echo "::endgroup::"
-
-        echo "::group::Setup Huggingface"
-        pip install -U "huggingface_hub[cli]" accelerate
-        huggingface-cli login --token $SECRET_EXECUTORCH_HF_TOKEN
-        OPTIMUM_ET_VERSION=$(cat .ci/docker/ci_commit_pins/optimum-executorch.txt)
-        pip install git+https://github.com/huggingface/optimum-executorch.git@${OPTIMUM_ET_VERSION}
-        pip install mistral-common librosa
-        pip list
-        echo "::endgroup::"
-
-        echo "::group::Export Voxtral with Quantization (int4-weight-only)"
-        # TODO: --qlinear 4w will produce wrong result. Need more investigation.
-        optimum-cli export executorch \
-            --model "mistralai/Voxtral-Mini-3B-2507" \
-            --task "multimodal-text-to-text" \
-            --recipe "cuda" \
-            --dtype bfloat16 \
-            --device cuda \
-            --max_seq_len 1024 \
-            --qlinear_encoder 4w \
-            --output_dir ./
-        python -m executorch.extension.audio.mel_spectrogram \
-            --feature_size 128 \
-            --stack_output \
-            --max_audio_len 300 \
-            --output_file voxtral_preprocessor.pte
-
-        test -f model.pte
-        test -f aoti_cuda_blob.ptd
-        test -f voxtral_preprocessor.pte
-        echo "::endgroup::"
-
-        echo "::group::Store Voxtral Quantized Artifacts (int4-weight-only)"
-        mkdir -p "${RUNNER_ARTIFACT_DIR}"
-        cp model.pte "${RUNNER_ARTIFACT_DIR}/"
-        cp aoti_cuda_blob.ptd "${RUNNER_ARTIFACT_DIR}/"
-        cp voxtral_preprocessor.pte "${RUNNER_ARTIFACT_DIR}/"
-        ls -al "${RUNNER_ARTIFACT_DIR}"
-        echo "::endgroup::"
-
   test-voxtral-cuda-e2e:
     name: test-voxtral-cuda-e2e-${{ matrix.format.name }}
-    needs:
-      - export-voxtral-cuda-artifact
-      - export-voxtral-cuda-quantized-int4-tile-packed
-      - export-voxtral-cuda-quantized-int4-weight-only
+    needs: export-voxtral-cuda-artifact
     uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
     permissions:
       id-token: write
diff --git a/examples/models/voxtral/README.md b/examples/models/voxtral/README.md
index 51a68e4f8f6..65085e30c1d 100644
--- a/examples/models/voxtral/README.md
+++ b/examples/models/voxtral/README.md
@@ -56,7 +56,7 @@ This will generate:
 - `aoti_cuda_blob.ptd` - The CUDA kernel blob required for runtime
 
 Furthermore, we support several quantization formats on CUDA.
-To export Voxtral with int4 weight-only quantization, use
+For example, to export Voxtral with int4 weight and int4mm for linear layers, you can use the following command,
 ```
 optimum-cli export executorch \
   --model "mistralai/Voxtral-Mini-3B-2507" \
@@ -67,7 +67,8 @@ optimum-cli export executorch \
   --max_seq_len 1024 \
   --qlinear 4w \
   --qlinear_encoder 4w \
-  --qembedding 4w \
+  --qlinear_packing_format tile_packed_to_4d \
+  --qlinear_encoder_packing_format tile_packed_to_4d \
   --output_dir="voxtral"
 ```