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/docker/ci_commit_pins/optimum-executorch.txt

@@ -1 +1 @@
-4361747abfc55e40e929396ed986efe775d745f9
+d03e90c2cd9048e6d9a75285c0355f033cd016fc

.ci/docker/ci_commit_pins/pytorch.txt

@@ -1 +1 @@
-556fc09a9f67f24ca5591ec049c5d0c347c5f62a
+b31bad1b8f1331bf43d47f46602cf6141db56844

.ci/docker/requirements-ci.txt

@@ -30,7 +30,6 @@ sphinx-reredirects==0.1.4
 matplotlib>=3.9.4
 sphinx-copybutton==0.5.2
 # PyTorch Theme
--e git+https://github.com/pytorch/pytorch_sphinx_theme.git@pytorch_sphinx_theme2#egg=pytorch_sphinx_theme2
-
+pytorch_sphinx_theme2==0.2.0
 # script unit test requirements
 yaspin==3.1.0

.ci/scripts/export_model_cuda_artifact.sh

@@ -17,7 +17,7 @@ Arguments:
   hf_model     HuggingFace model ID (required)
                Supported models:
                  - mistralai/Voxtral-Mini-3B-2507
-                 - openai/whisper-small
+                 - openai/whisper series (whisper-{small, medium, large, large-v2, large-v3, large-v3-turbo})
                  - google/gemma-3-4b-it
 
   quant_name   Quantization type (optional, default: non-quantized)
@@ -62,13 +62,17 @@ case "$HF_MODEL" in
     PREPROCESSOR_FEATURE_SIZE="128"
     PREPROCESSOR_OUTPUT="voxtral_preprocessor.pte"
     ;;
-  openai/whisper-small)
+  openai/whisper-*)
     MODEL_NAME="whisper"
     TASK="automatic-speech-recognition"
     MAX_SEQ_LEN=""
     EXTRA_PIP="librosa"
-    PREPROCESSOR_FEATURE_SIZE="80"
     PREPROCESSOR_OUTPUT="whisper_preprocessor.pte"
+    if [[ "$HF_MODEL" == *"large-v3"* ]]; then
+      PREPROCESSOR_FEATURE_SIZE="128"
+    else
+      PREPROCESSOR_FEATURE_SIZE="80"
+    fi
     ;;
   google/gemma-3-4b-it)
     MODEL_NAME="gemma3"
@@ -80,7 +84,7 @@ case "$HF_MODEL" in
     ;;
   *)
     echo "Error: Unsupported model '$HF_MODEL'"
-    echo "Supported models: mistralai/Voxtral-Mini-3B-2507, openai/whisper-small, google/gemma-3-4b-it"
+    echo "Supported models: mistralai/Voxtral-Mini-3B-2507, openai/whisper-{small, medium, large, large-v2, large-v3, large-v3-turbo}, google/gemma-3-4b-it"
     exit 1
     ;;
 esac

.ci/scripts/test_model_cuda_e2e.sh

@@ -17,7 +17,7 @@ Arguments:
   hf_model    HuggingFace model ID (required)
               Supported models:
                 - mistralai/Voxtral-Mini-3B-2507
-                - openai/whisper-small
+                - openai/whisper series (whisper-{small, medium, large, large-v2, large-v3, large-v3-turbo})
                 - google/gemma-3-4b-it
 
   quant_name  Quantization type (required)
@@ -91,13 +91,13 @@ case "$HF_MODEL" in
     AUDIO_FILE="poem.wav"
     IMAGE_PATH=""
     ;;
-  openai/whisper-small)
-    MODEL_NAME="whisper"
+  openai/whisper-*)
+    MODEL_NAME="${HF_MODEL#openai/}"
     RUNNER_TARGET="whisper_runner"
     RUNNER_PATH="whisper"
     EXPECTED_OUTPUT="Mr. Quilter is the apostle of the middle classes"
     PREPROCESSOR="whisper_preprocessor.pte"
-    TOKENIZER_URL="https://huggingface.co/openai/whisper-small/resolve/main" # @lint-ignore
+    TOKENIZER_URL="https://huggingface.co/${HF_MODEL}/resolve/main" # @lint-ignore
     TOKENIZER_FILE=""
     AUDIO_URL=""
     AUDIO_FILE="output.wav"
@@ -117,7 +117,7 @@ case "$HF_MODEL" in
     ;;
   *)
     echo "Error: Unsupported model '$HF_MODEL'"
-    echo "Supported models: mistralai/Voxtral-Mini-3B-2507, openai/whisper-small, google/gemma-3-4b-it"
+    echo "Supported models: mistralai/Voxtral-Mini-3B-2507, openai/whisper series (whisper-{small, medium, large, large-v2, large-v3, large-v3-turbo}), google/gemma-3-4b-it"
     exit 1
     ;;
 esac
@@ -142,7 +142,7 @@ fi
 # Download test files
 if [ "$AUDIO_URL" != "" ]; then
   curl -L $AUDIO_URL -o ${MODEL_DIR}/$AUDIO_FILE
-elif [ "$MODEL_NAME" = "whisper" ]; then
+elif [[ "$MODEL_NAME" == *whisper* ]]; then
   conda install -y -c conda-forge "ffmpeg<8"
   pip install datasets soundfile torchcodec
   python -c "from datasets import load_dataset;import soundfile as sf;sample = load_dataset('distil-whisper/librispeech_long', 'clean', split='validation')[0]['audio'];sf.write('${MODEL_DIR}/$AUDIO_FILE', sample['array'][:sample['sampling_rate']*30], sample['sampling_rate'])"
@@ -179,8 +179,8 @@ case "$MODEL_NAME" in
   voxtral)
     RUNNER_ARGS="$RUNNER_ARGS --tokenizer_path ${MODEL_DIR}/$TOKENIZER_FILE --audio_path ${MODEL_DIR}/$AUDIO_FILE --processor_path ${MODEL_DIR}/$PREPROCESSOR"
     ;;
-  whisper)
-    RUNNER_ARGS="$RUNNER_ARGS --tokenizer_path ${MODEL_DIR}/ --audio_path ${MODEL_DIR}/$AUDIO_FILE --processor_path ${MODEL_DIR}/$PREPROCESSOR"
+  whisper-*)
+    RUNNER_ARGS="$RUNNER_ARGS --tokenizer_path ${MODEL_DIR}/ --audio_path ${MODEL_DIR}/$AUDIO_FILE --processor_path ${MODEL_DIR}/$PREPROCESSOR --model_name ${MODEL_NAME}"
     ;;
   gemma3)
     RUNNER_ARGS="$RUNNER_ARGS --tokenizer_path ${MODEL_DIR}/ --image_path $IMAGE_PATH"

.ci/scripts/test_qnn_static_llm.sh

@@ -81,7 +81,7 @@ elif [[ "${TASK_NAME}" == "stories_260k_bc" ]]; then
     fi
 
 elif [[ "${TASK_NAME}" == "smollm2_135m" ]]; then
-    $PYTHON_EXECUTABLE backends/qualcomm/tests/test_qnn_delegate.py -k TestExampleLLMScript.test_static_smollm2 --model SM8650 --build_folder build-x86/ --executorch_root . --artifact_dir ./static_smollm2 --enable_x86_64
+    $PYTHON_EXECUTABLE backends/qualcomm/tests/test_qnn_delegate.py -k TestExampleLLMScript.test_static_llm_model --model_name smollm2_135m --model SM8650 --build_folder build-x86/ --executorch_root . --artifact_dir ./static_smollm2 --enable_x86_64
     exit_code1=$?
     if [ $exit_code1 -ne 0 ]; then
         exit 1

.github/workflows/cuda.yml

@@ -104,6 +104,8 @@ jobs:
             name: "Voxtral-Mini-3B-2507"
           - repo: "openai"
             name: "whisper-small"
+          - repo: "openai"
+            name: "whisper-large-v3-turbo"
           - repo: "google"
             name: "gemma-3-4b-it"
         quant:
@@ -223,6 +225,8 @@ jobs:
             name: "Voxtral-Mini-3B-2507"
           - repo: "openai"
             name: "whisper-small"
+          - repo: "openai"
+            name: "whisper-large-v3-turbo"
           - repo: "google"
             name: "gemma-3-4b-it"
         quant:

.github/workflows/doc-build.yml

@@ -46,14 +46,15 @@ jobs:
           export CHANNEL=nightly
         fi
 
-        # Get the version of ExecuTorch from REF_NAME and save as ET_VERSION_DOCS
-        # ET_VERSION_DOCS will be pulled during the doc build to add to the version dropdown
-        # on the website. See docs/source/conf.py for details
-
+        # Set RELEASE environment variable for tagged releases
         GITHUB_REF=${{ github.ref }}
-        echo "$GITHUB_REF"
-        export ET_VERSION_DOCS="${GITHUB_REF}"
-        echo "$ET_VERSION_DOCS"
+        if [[ "${GITHUB_REF}" =~ ^refs/tags/v[0-9]+\.[0-9]+ ]]; then
+          export RELEASE=true
+          echo "Building release docs (RELEASE=true)"
+        else
+          export RELEASE=false
+          echo "Building main docs (RELEASE=false)"
+        fi
 
         set -eux
 

.lintrunner.toml

@@ -449,3 +449,24 @@ command = [
   "--",
   "@{{PATHSFILE}}",
 ]
+
+[[linter]]
+code = 'ETVKNODEBUG'
+include_patterns = [
+    "backends/vulkan/**/*.glsl",
+]
+command = [
+    'python',
+    '-m',
+    'lintrunner_adapters',
+    'run',
+    'grep_linter',
+    '--pattern=((DEBUG_MODE)|(GL_EXT_debug_printf))',
+    '--linter-name=ETVKNODEBUG',
+    '--error-name=Using DEBUG_MODE or GL_EXT_debug_printf in Vulkan shader',
+    """--error-description=\
+        #define DEBUG_MODE or #extension GL_EXT_debug_printf should only be used during development!
+    """,
+    '--',
+    '@{{PATHSFILE}}',
+]

.mypy.ini

@@ -74,6 +74,12 @@ ignore_missing_imports = True
 [mypy-pytorch_sphinx_theme]
 ignore_missing_imports = True
 
+[mypy-pytorch_sphinx_theme2]
+ignore_missing_imports = True
+
+[mypy-executorch.version]
+ignore_missing_imports = True
+
 [mypy-ruamel]
 ignore_missing_imports = True