Update base for Update on "[ET][jni] Provide opt-out config for vulkan backend dependency"

## Context

IGVR can't use downloadable modules so an opt out path is required to avoid regressing APK size from including the vulkan backend.

Differential Revision: [D86547777](https://our.internmc.facebook.com/intern/diff/D86547777/)

cc manuelcandales digantdesai cbilgin

[ghstack-poisoned]

Author: ssjia

.ci/docker/ci_commit_pins/optimum-executorch.txt

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

.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"

.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:

backends/arm/CMakeLists.txt

@@ -48,44 +48,17 @@ endif()
 
 # VGF backend builds
 if(EXECUTORCH_BUILD_VGF)
+
+  # include libvgf
+  set(LIBVGF_PATH
+      "${EXECUTORCH_ROOT}/examples/arm/ethos-u-scratch/ml-sdk-for-vulkan-manifest/sw/vgf-lib/"
+  )
+
   set(VULKAN_THIRD_PARTY_PATH ${EXECUTORCH_ROOT}/backends/vulkan/third-party)
   set(VULKAN_HEADERS_PATH ${VULKAN_THIRD_PARTY_PATH}/Vulkan-Headers/include)
   set(VOLK_HEADERS_PATH ${VULKAN_THIRD_PARTY_PATH}/volk)
 
-  if(APPLE
-     OR CMAKE_SYSTEM_PROCESSOR MATCHES "^(arm64|aarch64)$"
-     OR EXISTS
-        "${EXECUTORCH_ROOT}/examples/arm/ethos-u-scratch/ml-sdk-for-vulkan-manifest/"
-  )
-    message(STATUS "libvgf sourced from local scratch tree")
-
-    # Legacy layout: libvgf sourced from local scratch tree
-    set(LIBVGF_PATH
-        "${EXECUTORCH_ROOT}/examples/arm/ethos-u-scratch/ml-sdk-for-vulkan-manifest/sw/vgf-lib/"
-    )
-    set(LIBVGF_STATIC "${LIBVGF_PATH}/build/src/libvgf.a")
-  else()
-    message(STATUS "libvgf installed from pip package")
-
-    set(Python3_FIND_VIRTUALENV FIRST)
-    if(EXECUTORCH_ROOT AND EXISTS "${EXECUTORCH_ROOT}/env")
-      set(Python3_EXECUTABLE "${EXECUTORCH_ROOT}/env/bin/python3")
-    endif()
-
-    find_package(Python3 REQUIRED COMPONENTS Interpreter)
-
-    # Prefer arch-specific site-packages if present, else pure
-    set(_vgf_site_arch "${Python3_SITEARCH}/vgf_lib/binaries")
-    set(_vgf_site_pure "${Python3_SITELIB}/vgf_lib/binaries")
-    if(EXISTS "${_vgf_site_arch}")
-      set(LIBVGF_PATH "${_vgf_site_arch}")
-    else()
-      set(LIBVGF_PATH "${_vgf_site_pure}")
-    endif()
-
-    set(LIBVGF_STATIC "${LIBVGF_PATH}/lib/libvgf.a")
-  endif()
-
+  set(LIBVGF_STATIC "${LIBVGF_PATH}/build/src/libvgf.a")
   set(LIBVGF_INCLUDE "${LIBVGF_PATH}/include/")
 
   add_library(vgf STATIC IMPORTED)

backends/arm/requirements-arm-vgf.txt

@@ -205,51 +205,7 @@ function setup_path_emulation_layer() {
     model_emulation_layer_path="$(cd "${mlsdk_manifest_dir}/sw/emulation-layer/" && pwd)"
     prepend_env_in_setup_path LD_LIBRARY_PATH "${model_emulation_layer_path}/deploy/lib"
     prepend_env_in_setup_path DYLD_LIBRARY_PATH "${model_emulation_layer_path}/deploy/lib"
-    prepend_env_in_setup_path VK_LAYER_PATH "${model_emulation_layer_path}/deploy/share/vulkan/explicit_layer.d"
     prepend_env_in_setup_path VK_INSTANCE_LAYERS VK_LAYER_ML_Tensor_Emulation
     prepend_env_in_setup_path VK_INSTANCE_LAYERS VK_LAYER_ML_Graph_Emulation
-}
-
-function setup_path_emulation_layer_from_pip() {
-    if ! command -v emulation_layer >/dev/null 2>&1; then
-        echo "[mlsdk_utils] 'emulation_layer' command not found; skipping pip emulation layer path setup"
-        return
-    fi
-
-    local output
-    if ! output=$(emulation_layer 2>/dev/null); then
-        echo "[mlsdk_utils] Failed to query emulation_layer environment; skipping"
-        return
-    fi
-
-    local exports
-    exports=$(echo "$output" | grep '^export ' || true)
-
-    local ld_line
-    ld_line=$(echo "$exports" | grep 'LD_LIBRARY_PATH=' || true)
-    if [[ -n "${ld_line}" ]]; then
-        local ld_value=${ld_line#export LD_LIBRARY_PATH=}
-        ld_value=${ld_value%%:\$LD_LIBRARY_PATH*}
-        if [[ -n "${ld_value}" ]]; then
-            prepend_env_in_setup_path LD_LIBRARY_PATH "${ld_value}"
-        fi
-    fi
-
-    local vk_add_line
-    vk_add_line=$(echo "$exports" | grep 'VK_ADD_LAYER_PATH=' || true)
-    if [[ -n "${vk_add_line}" ]]; then
-        local vk_add_value=${vk_add_line#export VK_ADD_LAYER_PATH=}
-        if [[ -n "${vk_add_value}" ]]; then
-            prepend_env_in_setup_path VK_ADD_LAYER_PATH "${vk_add_value}"
-        fi
-    fi
-
-    local vk_instance_line
-    vk_instance_line=$(echo "$exports" | grep 'VK_INSTANCE_LAYERS=' || true)
-    if [[ -n "${vk_instance_line}" ]]; then
-        local vk_instance_value=${vk_instance_line#export VK_INSTANCE_LAYERS=}
-        if [[ -n "${vk_instance_value}" ]]; then
-            prepend_env_in_setup_path VK_INSTANCE_LAYERS "${vk_instance_value}"
-        fi
-    fi
+    prepend_env_in_setup_path VK_LAYER_PATH "${model_emulation_layer_path}/deploy/share/vulkan/explicit_layer.d"
 }

backends/arm/scripts/mlsdk_utils.sh

@@ -50,14 +50,9 @@ if [[ -z ${model} ]]; then echo "Model name needs to be provided"; exit 1; fi
 
 source ${setup_path_script}
 
-if ! command -v "${converter}" >/dev/null 2>&1; then
-    if command -v model_converter >/dev/null 2>&1; then
-        converter="model_converter"
-    fi
-fi
-
-command -v "${converter}" >/dev/null 2>&1 \
-    || { echo "Could not find a model converter executable (tried model-converter, model_converter). ${_setup_msg}"; exit 1; }
+# basic checks before we get started
+hash ${converter} \
+    || { echo "Could not find ${converter} on PATH, ${_setup_msg}"; exit 1; }
 
 
 

backends/arm/scripts/run_vkml.sh

@@ -8,17 +8,23 @@
 
 from typing import Tuple
 
+import pytest
 import torch
+from executorch.backends.arm.quantizer.arm_quantizer import (
+    get_symmetric_a16w8_quantization_config,
+    TOSAQuantizer,
+)
+from executorch.backends.arm.test import common, conftest
 
-from executorch.backends.arm.test import common
 from executorch.backends.arm.test.tester.test_pipeline import (
     EthosU55PipelineINT,
     EthosU85PipelineINT,
     TosaPipelineFP,
     TosaPipelineINT,
     VgfPipeline,
 )
-
+from executorch.backends.arm.tosa import TosaSpecification
+from executorch.backends.xnnpack.test.tester import Quantize
 
 aten_op = "torch.ops.aten.rsqrt.default"
 input_t1 = Tuple[torch.Tensor]  # Input x
@@ -104,3 +110,99 @@ def test_rsqrt_vgf_INT(test_tensor: torch.Tensor):
         tosa_version="TOSA-1.0+INT",
     )
     pipeline.run()
+
+
+def get_symmetric_a16w8_rsqrt_quantizer(
+    u55_config=False, per_channel_quantization=False
+):
+    tosa_version = conftest.get_option("tosa_version")
+    tosa_profiles = {
+        "1.0": TosaSpecification.create_from_string("TOSA-1.0+INT+int16"),
+    }
+
+    quantizer = TOSAQuantizer(tosa_profiles[tosa_version])
+    quantizer.set_global(
+        get_symmetric_a16w8_quantization_config(is_per_channel=per_channel_quantization)
+    )
+
+    return Quantize(
+        quantizer,
+        get_symmetric_a16w8_quantization_config(
+            is_per_channel=per_channel_quantization
+        ),
+    )
+
+
+@common.parametrize("test_tensor", Rsqrt.test_parameters)
+@pytest.mark.xfail(
+    reason="MLETORCH-707: AssertionError: Output 0 does not match reference output."
+)
+def test_rsqrt_16a8w_tosa_INT(test_tensor: torch.Tensor):
+    """Test rsqrt operation with int16 quantization"""
+    pipeline = TosaPipelineINT[input_t1](
+        Rsqrt(),
+        test_tensor(),
+        aten_op,
+        exir_op=[],
+        per_channel_quantization=False,
+        use_to_edge_transform_and_lower=True,
+        tosa_extensions=["int16"],
+    )
+
+    pipeline.change_args(
+        "quantize",
+        get_symmetric_a16w8_rsqrt_quantizer(per_channel_quantization=False),
+    )
+    # Run the pipeline
+    pipeline.run()
+
+
+@common.parametrize("test_tensor", Rsqrt.test_parameters)
+@common.XfailIfNoCorstone300
+@pytest.mark.xfail(
+    reason="MLETORCH-707: AssertionError: Output 0 does not match reference output."
+)
+def test_rsqrt_16a8w_u55_INT16(test_tensor: torch.Tensor):
+    """Test rsqrt operation with int16 quantization on U55"""
+    pipeline = EthosU55PipelineINT[input_t1](
+        Rsqrt(),
+        test_tensor(),
+        aten_op,
+        exir_ops=[],
+        per_channel_quantization=True,
+        use_to_edge_transform_and_lower=True,
+        atol=1e-03,
+        rtol=1e-03,
+        run_on_fvp=True,
+    )
+
+    pipeline.change_args(
+        "quantize",
+        get_symmetric_a16w8_rsqrt_quantizer(per_channel_quantization=True),
+    )
+    pipeline.run()
+
+
+@common.parametrize("test_tensor", Rsqrt.test_parameters)
+@common.XfailIfNoCorstone320
+@pytest.mark.xfail(
+    reason="MLETORCH-707: AssertionError: Output 0 does not match reference output."
+)
+def test_rsqrt_16a8w_u85_INT16(test_tensor: torch.Tensor):
+    """Test rsqrt operation with int16 quantization on U85"""
+    pipeline = EthosU85PipelineINT[input_t1](
+        Rsqrt(),
+        test_tensor(),
+        aten_op,
+        exir_ops=[],
+        use_to_edge_transform_and_lower=True,
+        atol=1e-03,
+        rtol=1e-03,
+        run_on_fvp=True,
+    )
+
+    pipeline.change_args(
+        "quantize",
+        get_symmetric_a16w8_rsqrt_quantizer(per_channel_quantization=False),
+    )
+    pipeline.run()

backends/arm/test/ops/test_rsqrt.py

@@ -31,7 +31,6 @@
 from executorch.backends.arm.tosa.compile_spec import TosaCompileSpec
 from executorch.backends.arm.tosa.specification import Tosa_1_00, TosaSpecification
 from executorch.backends.arm.vgf import VgfCompileSpec
-from executorch.backends.arm.vgf.model_converter import find_model_converter_binary
 from executorch.exir import ExecutorchProgramManager, ExportedProgram
 from executorch.exir.lowered_backend_module import LoweredBackendModule
 from torch.fx.node import Node
@@ -679,15 +678,11 @@ def corstone320_installed() -> bool:
 
 
 def model_converter_installed() -> bool:
-    model_converter = find_model_converter_binary()
-    if model_converter is None:
-        return False
-
+    cmd = ["model-converter", "--version"]
     try:
-        _run_cmd([model_converter, "--version"], check=True)
-    except Exception:
+        _run_cmd(cmd, check=True)
+    except:
         return False
-
     return True