Update on "[ET Device Support] DeviceAllocator interface and DeviceAllocatorRegistry"

This diff introduces the `DeviceAllocator` abstract interface and `DeviceAllocatorRegistry` for device-specific memory allocation. This is a foundational abstraction that enables the runtime to dispatch memory operations to the appropriate device backend other than CPU (CUDA, etc.).

**DeviceAllocator interface provides:**
- `init_buffer()` - Initialize memory buffer pools for memory-planned tensors
- `get_offset_address()` - Get pointer to offset within pre-allocated buffer
- `allocate()` / `deallocate()` - Dynamic device memory allocation
- `copy_host_to_device()` / `copy_device_to_host()` - Data transfer between host and device
- `device_type()` - Returns the device type this allocator handles

**DeviceAllocatorRegistry provides:**
- Singleton registry mapping DeviceType → DeviceAllocator
- `register_allocator()` / `get_allocator()` methods
- Fixed-size array indexed by device type (no dynamic allocation, embedded-friendly)

**Design notes:**
- Registry stores raw pointers (non-owning) - allocators are expected to be singletons with static lifetime
- Follows ExecuTorch's embedded-first philosophy (no std::unique_ptr, no heap allocation in registry)
- Convenience free functions `register_device_allocator()` and `get_device_allocator()` for ease of use

Differential Revision: [D93635656](https://our.internmc.facebook.com/intern/diff/D93635656/)

[ghstack-poisoned]

Author: Gasoonjia

.ci/docker/ci_commit_pins/pytorch.txt

@@ -1 +1 @@
-659af3c353e49b35c191cdd2dba3b3c79d0e6822
+release/2.11

.ci/docker/common/install_pytorch.sh

@@ -32,9 +32,9 @@ install_pytorch_and_domains() {
   pip_install "$(echo dist/*.whl)"
 
   # Grab the pinned audio and vision commits from PyTorch
-  TORCHAUDIO_VERSION=$(cat .github/ci_commit_pins/audio.txt)
+  TORCHAUDIO_VERSION=release/2.11
   export TORCHAUDIO_VERSION
-  TORCHVISION_VERSION=$(cat .github/ci_commit_pins/vision.txt)
+  TORCHVISION_VERSION=release/0.26
   export TORCHVISION_VERSION
 
   install_domains

.ci/scripts/export_model_artifact.sh

@@ -24,6 +24,7 @@ Arguments:
                  - google/gemma-3-4b-it
                  - nvidia/diar_streaming_sortformer_4spk-v2
                  - nvidia/parakeet-tdt
+                 - facebook/dinov2-small-imagenet1k-1-layer
 
   quant_name   Quantization type (optional, default: non-quantized)
                Options:
@@ -167,6 +168,14 @@ case "$HF_MODEL" in
     PREPROCESSOR_FEATURE_SIZE=""
     PREPROCESSOR_OUTPUT=""
     ;;
+  facebook/dinov2-small-imagenet1k-1-layer)
+    MODEL_NAME="dinov2"
+    TASK=""
+    MAX_SEQ_LEN=""
+    EXTRA_PIP=""
+    PREPROCESSOR_FEATURE_SIZE=""
+    PREPROCESSOR_OUTPUT=""
+    ;;
   mistralai/Voxtral-Mini-4B-Realtime-2602)
     MODEL_NAME="voxtral_realtime"
     TASK=""
@@ -177,7 +186,7 @@ case "$HF_MODEL" in
     ;;
   *)
     echo "Error: Unsupported model '$HF_MODEL'"
-    echo "Supported models: mistralai/Voxtral-Mini-3B-2507, mistralai/Voxtral-Mini-4B-Realtime-2602, openai/whisper-{small, medium, large, large-v2, large-v3, large-v3-turbo}, google/gemma-3-4b-it, Qwen/Qwen3-0.6B, nvidia/diar_streaming_sortformer_4spk-v2, nvidia/parakeet-tdt"
+    echo "Supported models: mistralai/Voxtral-Mini-3B-2507, mistralai/Voxtral-Mini-4B-Realtime-2602, openai/whisper-{small, medium, large, large-v2, large-v3, large-v3-turbo}, google/gemma-3-4b-it, Qwen/Qwen3-0.6B, nvidia/diar_streaming_sortformer_4spk-v2, nvidia/parakeet-tdt, facebook/dinov2-small-imagenet1k-1-layer"
     exit 1
     ;;
 esac
@@ -293,6 +302,23 @@ if [ "$MODEL_NAME" = "sortformer" ]; then
   exit 0
 fi
 
+# DINOv2 uses a custom export script
+if [ "$MODEL_NAME" = "dinov2" ]; then
+  pip install -r examples/models/dinov2/install_requirements.txt
+
+  python -m executorch.examples.models.dinov2.export_dinov2 \
+      --backend "$DEVICE" \
+      --output-dir "${OUTPUT_DIR}"
+
+  test -f "${OUTPUT_DIR}/model.pte"
+  if [ "$DEVICE" = "cuda" ] || [ "$DEVICE" = "cuda-windows" ]; then
+    test -f "${OUTPUT_DIR}/aoti_cuda_blob.ptd"
+  fi
+  ls -al "${OUTPUT_DIR}"
+  echo "::endgroup::"
+  exit 0
+fi
+
 # Voxtral Realtime uses a custom export script
 if [ "$MODEL_NAME" = "voxtral_realtime" ]; then
   pip install safetensors huggingface_hub
@@ -308,6 +334,7 @@ if [ "$MODEL_NAME" = "voxtral_realtime" ]; then
     VR_QUANT_ARGS="--qlinear-encoder 8da4w --qlinear 8da4w --qlinear-group-size 32 --qembedding 8w"
   elif [ "$QUANT_NAME" = "quantized-int4-metal" ]; then
     VR_QUANT_ARGS="--qlinear-encoder fpa4w --qlinear fpa4w"
+    VR_DTYPE_ARGS="--dtype bf16"
   elif [ "$QUANT_NAME" = "quantized-int4-tile-packed" ]; then
     VR_QUANT_ARGS="--qlinear-encoder 4w --qlinear-encoder-packing-format tile_packed_to_4d --qlinear 4w --qlinear-packing-format tile_packed_to_4d --qembedding 8w"
     VR_DTYPE_ARGS="--dtype bf16"
@@ -347,6 +374,7 @@ if [ "$MODEL_NAME" = "voxtral_realtime" ]; then
   fi
   # Copy tokenizer from downloaded model weights
   cp "$LOCAL_MODEL_DIR/tekken.json" "${OUTPUT_DIR}/tekken.json"
+  rm -rf "$LOCAL_MODEL_DIR"
   ls -al "${OUTPUT_DIR}"
   echo "::endgroup::"
   exit 0

.ci/scripts/test_backend.sh

@@ -46,7 +46,7 @@ if [[ "$FLOW" == *qnn* ]]; then
     export LD_LIBRARY_PATH"=$QNN_X86_LIB_DIR:$QNN_SDK_ROOT/lib/x86_64-linux-clang/:${LD_LIBRARY_PATH:-}"
 
     # TODO Get SDK root from install scripts
-    EXTRA_BUILD_ARGS+=" -DEXECUTORCH_BUILD_QNN=ON -DQNN_SDK_ROOT=$QNN_SDK_ROOT"
+    EXTRA_BUILD_ARGS+=" -DEXECUTORCH_BUILD_QNN=ON -DQNN_SDK_ROOT=$QNN_SDK_ROOT -DEXECUTORCH_BUILD_EXTENSION_TENSOR=ON"
 fi
 
 if [[ "$FLOW" == *vulkan* ]]; then

.ci/scripts/test_lora.sh

@@ -139,7 +139,8 @@ Okay, so I need to calculate 15% of 80."
 EXPECTED_QUANT_LORA_PREFIX="
 <|im_start|>user Calculate 15% of 80?<|im_end|><|im_start|>assistant
 To calculate 15% of 80, we can multiply 80 by 15/100.
-So, 15% of 80 is equal to (80 * 15) / 100 = 1200 / 100 = 12.
+80 * 15/100 = 12.
+So, 15% of 80 is 12.
 #### 12
 The answer is: 12<|im_end|>"
 

.ci/scripts/test_model_e2e.sh

@@ -24,6 +24,7 @@ Arguments:
                 - google/gemma-3-4b-it
                 - Qwen/Qwen3-0.6B
                 - nvidia/parakeet-tdt
+                - facebook/dinov2-small-imagenet1k-1-layer
                 - mistralai/Voxtral-Mini-4B-Realtime-2602
 
   quant_name  Quantization type (required)
@@ -190,6 +191,19 @@ case "$HF_MODEL" in
     AUDIO_FILE="poem.wav"
     IMAGE_PATH=""
     ;;
+  facebook/dinov2-small-imagenet1k-1-layer)
+    MODEL_NAME="dinov2"
+    RUNNER_TARGET="dinov2_runner"
+    RUNNER_PATH="dinov2"
+    EXPECTED_OUTPUT="Samoyed"
+    PREPROCESSOR=""
+    TOKENIZER_URL=""
+    TOKENIZER_FILE=""
+    AUDIO_URL=""
+    AUDIO_FILE=""
+    IMAGE_URL="https://github.com/pytorch/hub/raw/master/images/dog.jpg"
+    IMAGE_PATH=""
+    ;;
   mistralai/Voxtral-Mini-4B-Realtime-2602)
     MODEL_NAME="voxtral_realtime"
     RUNNER_TARGET="voxtral_realtime_runner"
@@ -204,7 +218,7 @@ case "$HF_MODEL" in
     ;;
   *)
     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"
+    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"
     exit 1
     ;;
 esac
@@ -218,7 +232,7 @@ 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" ] && [ "$MODEL_NAME" != "sortformer" ]; then
+if [ "$MODEL_NAME" != "parakeet" ] && [ "$MODEL_NAME" != "voxtral_realtime" ] && [ "$MODEL_NAME" != "sortformer" ] && [ "$MODEL_NAME" != "dinov2" ]; then
   if [ "$TOKENIZER_FILE" != "" ]; then
     curl -L $TOKENIZER_URL/$TOKENIZER_FILE -o $MODEL_DIR/$TOKENIZER_FILE
   else
@@ -234,10 +248,15 @@ if [ "$AUDIO_URL" != "" ]; then
 elif [[ "$MODEL_NAME" == *whisper* ]] || [ "$MODEL_NAME" = "voxtral_realtime" ]; then
   conda install -y -c conda-forge "ffmpeg<8"
   pip install datasets soundfile
-  pip install torchcodec==0.11.0.dev20260217 --extra-index-url https://download.pytorch.org/whl/nightly/cpu
+  pip install torchcodec==0.11.0 --extra-index-url https://download.pytorch.org/whl/test/cpu
   python -c "from datasets import load_dataset;import soundfile as sf;sample = load_dataset('distil-whisper/librispeech_long', 'clean', split='validation')[0]['audio'];sf.write('${MODEL_DIR}/$AUDIO_FILE', sample['array'][:sample['sampling_rate']*30], sample['sampling_rate'])"
 fi
 
+# Download test image for vision models
+if [ -n "${IMAGE_URL:-}" ]; then
+  curl -L "$IMAGE_URL" -o "${MODEL_DIR}/test_image.jpg"
+fi
+
 ls -al
 echo "::endgroup::"
 
@@ -316,6 +335,12 @@ EOF
       RUNNER_ARGS="$RUNNER_ARGS --data_path ${MODEL_DIR}/aoti_cuda_blob.ptd"
     fi
     ;;
+  dinov2)
+    RUNNER_ARGS="--model_path ${MODEL_DIR}/model.pte --image_path ${MODEL_DIR}/test_image.jpg"
+    if [ "$DEVICE" = "cuda" ]; then
+      RUNNER_ARGS="$RUNNER_ARGS --data_path ${MODEL_DIR}/aoti_cuda_blob.ptd"
+    fi
+    ;;
   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

.ci/scripts/test_model_e2e_windows.ps1

@@ -13,7 +13,8 @@ param(
     [Parameter(Mandatory = $true)]
     [string]$QuantName,
     [string]$ModelDir = ".",
-    [string]$ExpectedCudaVersion = ""
+    [string]$ExpectedCudaVersion = "",
+    [string]$Mode = ""
 )
 
 Set-StrictMode -Version Latest
@@ -25,6 +26,15 @@ if ($Device -ne "cuda-windows") {
     throw "Unsupported device '$Device'. Expected 'cuda-windows'."
 }
 
+if ($Mode -ne "") {
+    if ($Mode -notin @("vr-streaming", "vr-offline")) {
+        throw "Unsupported mode '$Mode'. Supported modes: vr-streaming, vr-offline"
+    }
+    if ($HfModel -ne "mistralai/Voxtral-Mini-4B-Realtime-2602") {
+        throw "Mode '$Mode' can only be used with Voxtral Realtime model"
+    }
+}
+
 Write-Host "Testing model: $HfModel (quantization: $QuantName)"
 
 $resolvedModelDir = (Resolve-Path -Path $ModelDir).Path
@@ -79,15 +89,28 @@ switch ($HfModel) {
         $runnerTarget = "voxtral_realtime_runner"
         $runnerPath = "voxtral_realtime"
         $runnerPreset = "voxtral-realtime-cuda"
-        $expectedOutput = "Loading audio from"
+        $expectedOutput = "Quilter"
         $preprocessor = "preprocessor.pte"
         $tokenizerUrl = ""
         $tokenizerFile = "tekken.json"
         $audioUrl = "https://github.com/voxserv/audio_quality_testing_samples/raw/refs/heads/master/testaudio/16000/test01_20s.wav"
         $audioFile = "poem.wav"
     }
+    "facebook/dinov2-small-imagenet1k-1-layer" {
+        $runnerTarget = "dinov2_runner"
+        $runnerPath = "dinov2"
+        $runnerPreset = "dinov2-cuda"
+        $expectedOutput = "Samoyed"
+        $preprocessor = ""
+        $tokenizerUrl = ""
+        $tokenizerFile = ""
+        $audioUrl = ""
+        $audioFile = ""
+        $imageUrl = "https://github.com/pytorch/hub/raw/master/images/dog.jpg"
+        $imageFile = "test_image.jpg"
+    }
     default {
-        throw "Unsupported model '$HfModel'. Supported: mistralai/Voxtral-Mini-3B-2507, mistralai/Voxtral-Mini-4B-Realtime-2602, nvidia/diar_streaming_sortformer_4spk-v2, nvidia/parakeet-tdt"
+        throw "Unsupported model '$HfModel'. Supported: mistralai/Voxtral-Mini-3B-2507, mistralai/Voxtral-Mini-4B-Realtime-2602, nvidia/diar_streaming_sortformer_4spk-v2, nvidia/parakeet-tdt, facebook/dinov2-small-imagenet1k-1-layer"
     }
 }
 
@@ -162,6 +185,9 @@ try {
     if ($audioUrl -ne "") {
         Download-IfNeeded -Url $audioUrl -OutFile (Join-Path -Path $resolvedModelDir -ChildPath $audioFile)
     }
+    if ((Get-Variable -Name imageUrl -ErrorAction SilentlyContinue) -and $imageUrl -ne "") {
+        Download-IfNeeded -Url $imageUrl -OutFile (Join-Path -Path $resolvedModelDir -ChildPath $imageFile)
+    }
     Get-ChildItem -Path $resolvedModelDir
     Write-Host "::endgroup::"
 
@@ -207,6 +233,16 @@ try {
                 "--audio_path", (Join-Path -Path $resolvedModelDir -ChildPath $audioFile),
                 "--preprocessor_path", (Join-Path -Path $resolvedModelDir -ChildPath $preprocessor)
             )
+            if ($Mode -ne "vr-offline") {
+                $runnerArgs += "--streaming"
+            }
+        }
+        "facebook/dinov2-small-imagenet1k-1-layer" {
+            $runnerArgs = @(
+                "--model_path", $modelPte,
+                "--data_path", $cudaBlob,
+                "--image_path", (Join-Path -Path $resolvedModelDir -ChildPath $imageFile)
+            )
         }
     }
 

.ci/scripts/test_wheel_package_qnn.sh

@@ -158,17 +158,17 @@ print(module_vars["TORCH_VERSION"])
 PY
 )
 
-  NIGHTLY_VERSION=$(
-  "$PYBIN" - <<'PY'
-import runpy
-module_vars = runpy.run_path("torch_pin.py")
-print(module_vars["NIGHTLY_VERSION"])
-PY
-)
-  echo "=== [$LABEL] Install torch==${TORCH_VERSION}.${NIGHTLY_VERSION} ==="
-
-  # Install torchao based on the pinned PyTorch version
-  "$PIPBIN" install torch=="${TORCH_VERSION}.${NIGHTLY_VERSION}" --index-url "https://download.pytorch.org/whl/nightly/cpu"
+#   NIGHTLY_VERSION=$(
+#   "$PYBIN" - <<'PY'
+# import runpy
+# module_vars = runpy.run_path("torch_pin.py")
+# print(module_vars["NIGHTLY_VERSION"])
+# PY
+# )
+  echo "=== [$LABEL] Install torch==${TORCH_VERSION} ==="
+
+  # Install torch based on the pinned PyTorch version, preferring the PyTorch test index
+  "$PIPBIN" install torch=="${TORCH_VERSION}" --extra-index-url "https://download.pytorch.org/whl/test"
   "$PIPBIN" install wheel
 
   # Install torchao based on the pinned commit from third-party/ao submodule

.ci/scripts/unittest-linux-cmake.sh

@@ -19,6 +19,8 @@ if ! python -c "import tosa_serializer" >/dev/null 2>&1; then
     TOSA_SERIALIZATION_DIR="${TOSA_TOOLS_DIR}/serialization"
   fi
 
+  # NOTE: Will be removed when tosa-tools is installed via pypi
+  python -m pip install pybind11==2.10.4
   CMAKE_POLICY_VERSION_MINIMUM=3.5 BUILD_PYBIND=1 \
     python -m pip install --no-dependencies \
     "${TOSA_SERIALIZATION_DIR}"

.ci/scripts/utils.sh

@@ -53,7 +53,7 @@ dedupe_macos_loader_path_rpaths() {
   pushd ..
   torch_lib_dir=$(python -c "import importlib.util; print(importlib.util.find_spec('torch').submodule_search_locations[0])")/lib
   popd
-  
+
   if [[ -z "${torch_lib_dir}" || ! -d "${torch_lib_dir}" ]]; then
     return
   fi
@@ -141,9 +141,9 @@ install_pytorch_and_domains() {
 
   dedupe_macos_loader_path_rpaths
   # Grab the pinned audio and vision commits from PyTorch
-  TORCHAUDIO_VERSION=$(cat .github/ci_commit_pins/audio.txt)
+  TORCHAUDIO_VERSION=release/2.11
   export TORCHAUDIO_VERSION
-  TORCHVISION_VERSION=$(cat .github/ci_commit_pins/vision.txt)
+  TORCHVISION_VERSION=release/0.26
   export TORCHVISION_VERSION
 
   install_domains