Update base for Update on "Reduce allocation overhead in quantized sdpa"

For small models dequantizing portions of v cache causes extra alloc overhead.

Probably a better way to handle this is to dequantize entire v cache outside the model

There isnt significant perf advantage from this yet but subsequent diffs will use caching allocator where this refactor help.

Differential Revision: [D85532077](https://our.internmc.facebook.com/intern/diff/D85532077/)

[ghstack-poisoned]

Author: kimishpatel

.ci/scripts/export_model_cuda_artifact.sh .ci/scripts/export_model_artifact.sh.ci/scripts/export_model_cuda_artifact.sh renamed to .ci/scripts/export_model_artifact.sh

@@ -5,15 +5,17 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
-# Export model to CUDA format with optional quantization
+# Export model to CUDA/Metal format with optional quantization
 
 show_help() {
   cat << EOF
-Usage: export_model_cuda_artifact.sh <hf_model> [quant_name] [output_dir]
+Usage: export_model_artifact.sh <device> <hf_model> [quant_name] [output_dir]
 
-Export a HuggingFace model to CUDA format with optional quantization.
+Export a HuggingFace model to CUDA/Metal format with optional quantization.
 
 Arguments:
+  device       cuda or metal (required)
+
   hf_model     HuggingFace model ID (required)
                Supported models:
                  - mistralai/Voxtral-Mini-3B-2507
@@ -29,9 +31,9 @@ Arguments:
   output_dir   Output directory for artifacts (optional, default: current directory)
 
 Examples:
-  export_model_cuda_artifact.sh "openai/whisper-small"
-  export_model_cuda_artifact.sh "mistralai/Voxtral-Mini-3B-2507" "quantized-int4-tile-packed"
-  export_model_cuda_artifact.sh "google/gemma-3-4b-it" "non-quantized" "./output"
+  export_model_artifact.sh metal "openai/whisper-small"
+  export_model_artifact.sh cuda "mistralai/Voxtral-Mini-3B-2507" "quantized-int4-tile-packed"
+  export_model_artifact.sh cuda "google/gemma-3-4b-it" "non-quantized" "./output"
 EOF
 }
 
@@ -48,9 +50,22 @@ fi
 
 set -eux
 
-HF_MODEL="$1"
-QUANT_NAME="${2:-non-quantized}"
-OUTPUT_DIR="${3:-.}"
+DEVICE="$1"
+HF_MODEL="$2"
+QUANT_NAME="${3:-non-quantized}"
+OUTPUT_DIR="${4:-.}"
+
+case "$DEVICE" in
+  cuda)
+    ;;
+  metal)
+    ;;
+  *)
+    echo "Error: Unsupported device '$DEVICE'"
+    echo "Supported devices: cuda, metal"
+    exit 1
+    ;;
+esac
 
 # Determine model configuration based on HF model ID
 case "$HF_MODEL" in
@@ -75,6 +90,10 @@ case "$HF_MODEL" in
     fi
     ;;
   google/gemma-3-4b-it)
+    if [ "$DEVICE" = "metal" ]; then
+      echo "Error: Export for device 'metal' is not yet tested for model '$HF_MODEL'"
+      exit 1
+    fi
     MODEL_NAME="gemma3"
     TASK="multimodal-text-to-text"
     MAX_SEQ_LEN="64"
@@ -95,9 +114,17 @@ case "$QUANT_NAME" in
     EXTRA_ARGS=""
     ;;
   quantized-int4-tile-packed)
+    if [ "$DEVICE" = "metal" ]; then
+      echo "Error: Metal backend does not yet support quantization '$QUANT_NAME'"
+      exit 1
+    fi
     EXTRA_ARGS="--qlinear 4w --qlinear_encoder 4w --qlinear_packing_format tile_packed_to_4d --qlinear_encoder_packing_format tile_packed_to_4d"
     ;;
   quantized-int4-weight-only)
+    if [ "$DEVICE" = "metal" ]; then
+      echo "Error: Metal backend does not yet support quantization '$QUANT_NAME'"
+      exit 1
+    fi
     EXTRA_ARGS="--qlinear_encoder 4w"
     ;;
   *)
@@ -118,12 +145,18 @@ MAX_SEQ_LEN_ARG=""
 if [ -n "$MAX_SEQ_LEN" ]; then
   MAX_SEQ_LEN_ARG="--max_seq_len $MAX_SEQ_LEN"
 fi
+
+DEVICE_ARG=""
+if [ "$DEVICE" = "cuda" ]; then
+  DEVICE_ARG="--device cuda"
+fi
+
 optimum-cli export executorch \
     --model "$HF_MODEL" \
     --task "$TASK" \
-    --recipe "cuda" \
+    --recipe "$DEVICE" \
     --dtype bfloat16 \
-    --device cuda \
+    ${DEVICE_ARG} \
     ${MAX_SEQ_LEN_ARG} \
     ${EXTRA_ARGS} \
     --output_dir ./
@@ -137,18 +170,18 @@ if [ -n "$PREPROCESSOR_OUTPUT" ]; then
 fi
 
 test -f model.pte
-test -f aoti_cuda_blob.ptd
+test -f aoti_${DEVICE}_blob.ptd
 if [ -n "$PREPROCESSOR_OUTPUT" ]; then
   test -f $PREPROCESSOR_OUTPUT
 fi
 echo "::endgroup::"
 
 echo "::group::Store $MODEL_NAME Artifacts"
 mkdir -p "${OUTPUT_DIR}"
-cp model.pte "${OUTPUT_DIR}/"
-cp aoti_cuda_blob.ptd "${OUTPUT_DIR}/"
+mv model.pte "${OUTPUT_DIR}/"
+mv aoti_${DEVICE}_blob.ptd "${OUTPUT_DIR}/"
 if [ -n "$PREPROCESSOR_OUTPUT" ]; then
-  cp $PREPROCESSOR_OUTPUT "${OUTPUT_DIR}/"
+  mv $PREPROCESSOR_OUTPUT "${OUTPUT_DIR}/"
 fi
 ls -al "${OUTPUT_DIR}"
 echo "::endgroup::"

.ci/scripts/test_model_cuda_e2e.sh .ci/scripts/test_model_e2e.sh.ci/scripts/test_model_cuda_e2e.sh renamed to .ci/scripts/test_model_e2e.sh

@@ -5,15 +5,17 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
-# Test CUDA model end-to-end, need to run .ci/scripts/export_model_cuda_artifact.sh first
+# Test CUDA/Metal model end-to-end, need to run .ci/scripts/export_model_artifact.sh first
 
 show_help() {
   cat << EOF
-Usage: test_model_cuda_e2e.sh <hf_model> <quant_name> [model_dir]
+Usage: test_model_e2e.sh <device> <hf_model> <quant_name> [model_dir]
 
-Build and run end-to-end tests for CUDA models.
+Build and run end-to-end tests for CUDA/Metal models.
 
 Arguments:
+  device      cuda or metal (required)
+
   hf_model    HuggingFace model ID (required)
               Supported models:
                 - mistralai/Voxtral-Mini-3B-2507
@@ -27,12 +29,12 @@ Arguments:
                 - quantized-int4-weight-only
 
   model_dir   Directory containing model artifacts (optional, default: current directory)
-              Expected files: model.pte, aoti_cuda_blob.ptd
+              Expected files: model.pte, aoti_cuda_blob.ptd/aoti_metal_blob.ptd
               Tokenizers and test files will be downloaded to this directory
 
 Examples:
-  test_model_cuda_e2e.sh "openai/whisper-small" "non-quantized"
-  test_model_cuda_e2e.sh "mistralai/Voxtral-Mini-3B-2507" "quantized-int4-tile-packed" "./model_output"
+  test_model_e2e.sh metal "openai/whisper-small" "non-quantized"
+  test_model_e2e.sh cuda "mistralai/Voxtral-Mini-3B-2507" "quantized-int4-tile-packed" "./model_output"
 EOF
 }
 
@@ -55,20 +57,21 @@ fi
 
 set -eux
 
-HF_MODEL="$1"
-QUANT_NAME="$2"
+DEVICE="$1"
+HF_MODEL="$2"
+QUANT_NAME="$3"
 # Download tokenizers, audio, and image files to this directory
-MODEL_DIR="${3:-.}"
+MODEL_DIR="${4:-.}"
 
 echo "Testing model: $HF_MODEL (quantization: $QUANT_NAME)"
 
-# Make sure model.pte and aoti_cuda_blob.ptd exist
+# Make sure model.pte and aoti_${DEVICE}_blob.ptd exist
 if [ ! -f "$MODEL_DIR/model.pte" ]; then
   echo "Error: model.pte not found in $MODEL_DIR"
   exit 1
 fi
-if [ ! -f "$MODEL_DIR/aoti_cuda_blob.ptd" ]; then
-  echo "Error: aoti_cuda_blob.ptd not found in $MODEL_DIR"
+if [ ! -f "$MODEL_DIR/aoti_${DEVICE}_blob.ptd" ]; then
+  echo "Error: aoti_${DEVICE}_blob.ptd not found in $MODEL_DIR"
   exit 1
 fi
 # Locate EXECUTORCH_ROOT from the directory of this script
@@ -152,14 +155,24 @@ ls -al
 echo "::endgroup::"
 
 echo "::group::Build $MODEL_NAME Runner"
+
+if [ "$DEVICE" = "cuda" ]; then
+  BUILD_BACKEND="EXECUTORCH_BUILD_CUDA"
+elif [ "$DEVICE" = "metal" ]; then
+  BUILD_BACKEND="EXECUTORCH_BUILD_METAL"
+else
+  echo "Error: Unsupported device '$DEVICE'. Must be 'cuda' or 'metal'."
+  exit 1
+fi
+
 cmake --preset llm \
-      -DEXECUTORCH_BUILD_CUDA=ON \
+      -D${BUILD_BACKEND}=ON \
       -DCMAKE_INSTALL_PREFIX=cmake-out \
       -DCMAKE_BUILD_TYPE=Release \
       -Bcmake-out -S.
 cmake --build cmake-out -j$(nproc) --target install --config Release
 
-cmake -DEXECUTORCH_BUILD_CUDA=ON \
+cmake -D${BUILD_BACKEND}=ON \
       -DCMAKE_BUILD_TYPE=Release \
       -Sexamples/models/$RUNNER_PATH \
       -Bcmake-out/examples/models/$RUNNER_PATH/
@@ -168,11 +181,13 @@ echo "::endgroup::"
 
 echo "::group::Run $MODEL_NAME Runner"
 set +e
-export LD_LIBRARY_PATH=/opt/conda/lib:$LD_LIBRARY_PATH
+if [ "$DEVICE" = "cuda" ]; then
+  export LD_LIBRARY_PATH=/opt/conda/lib:$LD_LIBRARY_PATH
+fi
 
 # Build runner command with common arguments
 RUNNER_BIN="cmake-out/examples/models/$RUNNER_PATH/$RUNNER_TARGET"
-RUNNER_ARGS="--model_path ${MODEL_DIR}/model.pte --data_path ${MODEL_DIR}/aoti_cuda_blob.ptd --temperature 0"
+RUNNER_ARGS="--model_path ${MODEL_DIR}/model.pte --data_path ${MODEL_DIR}/aoti_${DEVICE}_blob.ptd --temperature 0"
 
 # Add model-specific arguments
 case "$MODEL_NAME" in

.github/workflows/cuda.yml

@@ -142,7 +142,7 @@ jobs:
         pip install git+https://github.com/huggingface/optimum-executorch.git@${OPTIMUM_ET_VERSION}
         echo "::endgroup::"
 
-        source .ci/scripts/export_model_cuda_artifact.sh "${{ matrix.model.repo }}/${{ matrix.model.name }}" "${{ matrix.quant }}" "${RUNNER_ARTIFACT_DIR}"
+        source .ci/scripts/export_model_artifact.sh cuda "${{ matrix.model.repo }}/${{ matrix.model.name }}" "${{ matrix.quant }}" "${RUNNER_ARTIFACT_DIR}"
 
   benchmark-model-cuda:
     name: benchmark-model-cuda
@@ -249,4 +249,4 @@ jobs:
       download-artifact: ${{ matrix.model.repo }}-${{ matrix.model.name }}-cuda-${{ matrix.quant }}
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
       script: |
-        source .ci/scripts/test_model_cuda_e2e.sh "${{ matrix.model.repo }}/${{ matrix.model.name }}" "${{ matrix.quant }}" "${RUNNER_ARTIFACT_DIR}"
+        source .ci/scripts/test_model_e2e.sh cuda "${{ matrix.model.repo }}/${{ matrix.model.name }}" "${{ matrix.quant }}" "${RUNNER_ARTIFACT_DIR}"