Merge branch 'main' into main

Author: psiddh

.ci/docker/build.sh

@@ -40,6 +40,15 @@ case "${IMAGE_NAME}" in
     LINTRUNNER=""
     GCC_VERSION=11
     ;;
+  executorch-ubuntu-22.04-gcc11-aarch64-android)
+    LINTRUNNER=""
+    GCC_VERSION=11
+    ANDROID_NDK_VERSION=r28c
+    ;;
+  executorch-ubuntu-22.04-gcc11-aarch64-arm-sdk)
+    ARM_SDK=yes
+    GCC_VERSION=11
+    ;;
   executorch-ubuntu-22.04-linter)
     LINTRUNNER=yes
     CLANG_VERSION=12

.ci/docker/ci_commit_pins/pytorch.txt

@@ -1 +1 @@
-7a79b41e29a790ebb4b530eb98a89381e2d7de29
+659af3c353e49b35c191cdd2dba3b3c79d0e6822

.ci/docker/common/install_android.sh

@@ -40,8 +40,14 @@ install_ndk() {
   rm -rf "${NDK_INSTALLATION_DIR}" && mkdir -p "${NDK_INSTALLATION_DIR}"
 
   pushd /tmp
-  # The NDK installation is cached on ossci-android S3 bucket
-  curl -Os --retry 3 "https://ossci-android.s3.amazonaws.com/android-ndk-${ANDROID_NDK_VERSION}-linux.zip"
+  ARCH=$(uname -m)
+  if [ "${ARCH}" = "aarch64" ]; then
+    # aarch64 NDK is not cached on S3, download from Google directly
+    curl -Os --retry 3 "https://dl.google.com/android/repository/android-ndk-${ANDROID_NDK_VERSION}-linux.zip"
+  else
+    # The NDK installation is cached on ossci-android S3 bucket
+    curl -Os --retry 3 "https://ossci-android.s3.amazonaws.com/android-ndk-${ANDROID_NDK_VERSION}-linux.zip"
+  fi
   unzip -qo "android-ndk-${ANDROID_NDK_VERSION}-linux.zip"
 
   # Print the content for manual verification
@@ -73,7 +79,10 @@ install_sdk() {
   yes | /opt/cmdline-tools/bin/sdkmanager --sdk_root="${SDK_INSTALLATION_DIR}" --install "build-tools;35.0.0"
   # And some more tools for future emulator tests
   yes | /opt/cmdline-tools/bin/sdkmanager --sdk_root="${SDK_INSTALLATION_DIR}" --install "platform-tools"
-  yes | /opt/cmdline-tools/bin/sdkmanager --sdk_root="${SDK_INSTALLATION_DIR}" --install "tools"
+  # The 'tools' package (emulator) is not available on aarch64
+  if [ "$(uname -m)" != "aarch64" ]; then
+    yes | /opt/cmdline-tools/bin/sdkmanager --sdk_root="${SDK_INSTALLATION_DIR}" --install "tools"
+  fi
 }
 
 install_prerequiresites

.ci/docker/requirements-ci.txt

@@ -30,6 +30,6 @@ sphinx-reredirects==0.1.4
 matplotlib>=3.9.4
 sphinx-copybutton==0.5.2
 # PyTorch Theme
-pytorch_sphinx_theme2==0.2.0
+pytorch_sphinx_theme2==0.4.4
 # script unit test requirements
 yaspin==3.1.0

.ci/scripts/cleanup-samsung-linux-deps.sh

@@ -0,0 +1,23 @@
+#!/bin/bash
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# Copyright (c) Samsung Electronics Co. LTD
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+set +ex
+
+if [[ "${DEVICE_ACQUIRED:-0}" != "1" ]]; then
+  exit 0
+fi
+
+if ! command -v devicefarm-cli >/dev/null 2>&1; then
+  echo "[WARN] Skip device disconnect (devicefarm-cli not installed)." >&2
+  exit 0
+fi
+
+echo "[INFO] Disconnecting device (-d)..."
+devicefarm-cli -d || echo "::warning::Device disconnect failed (ignored)"
+
+set -ex

.ci/scripts/export_model_artifact.sh

@@ -9,7 +9,7 @@
 
 show_help() {
   cat << EOF
-Usage: export_model_artifact.sh <device> <hf_model> [quant_name] [output_dir]
+Usage: export_model_artifact.sh <device> <hf_model> [quant_name] [output_dir] [mode]
 
 Export a HuggingFace model to CUDA/Metal/XNNPACK format with optional quantization.
 
@@ -19,6 +19,7 @@ Arguments:
   hf_model     HuggingFace model ID (required)
                Supported models:
                  - mistralai/Voxtral-Mini-3B-2507
+                 - mistralai/Voxtral-Mini-4B-Realtime-2602
                  - openai/whisper series (whisper-{small, medium, large, large-v2, large-v3, large-v3-turbo})
                  - google/gemma-3-4b-it
                  - nvidia/parakeet-tdt
@@ -33,13 +34,22 @@ Arguments:
 
   output_dir   Output directory for artifacts (optional, default: current directory)
 
+  mode         Export mode (optional, default: auto-detect based on model and device)
+               Supported modes:
+                 - vr-streaming: Voxtral Realtime streaming mode
+                 - vr-offline: Voxtral Realtime offline mode
+
 Examples:
   export_model_artifact.sh metal "openai/whisper-small"
   export_model_artifact.sh metal "nvidia/parakeet-tdt" "quantized-int4-metal"
+  export_model_artifact.sh metal "mistralai/Voxtral-Mini-4B-Realtime-2602" "quantized-int4-metal"
+  export_model_artifact.sh metal "mistralai/Voxtral-Mini-4B-Realtime-2602" "non-quantized" "." "vr-streaming"
   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"
   export_model_artifact.sh cuda "nvidia/parakeet-tdt" "non-quantized" "./output"
   export_model_artifact.sh xnnpack "nvidia/parakeet-tdt" "quantized-8da4w" "./output"
+  export_model_artifact.sh xnnpack "mistralai/Voxtral-Mini-4B-Realtime-2602" "quantized-8da4w" "./output"
+  export_model_artifact.sh xnnpack "mistralai/Voxtral-Mini-4B-Realtime-2602" "non-quantized" "./output" "vr-offline"
 EOF
 }
 
@@ -60,6 +70,26 @@ DEVICE="$1"
 HF_MODEL="$2"
 QUANT_NAME="${3:-non-quantized}"
 OUTPUT_DIR="${4:-.}"
+MODE="${5:-}"
+
+# Validate mode if specified
+if [ -n "$MODE" ]; then
+  case "$MODE" in
+    vr-streaming|vr-offline)
+      # Voxtral Realtime modes require Voxtral Realtime model
+      if [ "$HF_MODEL" != "mistralai/Voxtral-Mini-4B-Realtime-2602" ]; then
+        echo "Error: Mode '$MODE' can only be used with Voxtral Realtime model"
+        echo "Provided model: $HF_MODEL"
+        exit 1
+      fi
+      ;;
+    *)
+      echo "Error: Unsupported mode '$MODE'"
+      echo "Supported modes: vr-streaming, vr-offline"
+      exit 1
+      ;;
+  esac
+fi
 
 case "$DEVICE" in
   cuda)
@@ -119,9 +149,17 @@ case "$HF_MODEL" in
     PREPROCESSOR_FEATURE_SIZE=""
     PREPROCESSOR_OUTPUT=""
     ;;
+  mistralai/Voxtral-Mini-4B-Realtime-2602)
+    MODEL_NAME="voxtral_realtime"
+    TASK=""
+    MAX_SEQ_LEN=""
+    EXTRA_PIP="mistral-common librosa"
+    PREPROCESSOR_FEATURE_SIZE=""
+    PREPROCESSOR_OUTPUT=""
+    ;;
   *)
     echo "Error: Unsupported model '$HF_MODEL'"
-    echo "Supported models: mistralai/Voxtral-Mini-3B-2507, openai/whisper-{small, medium, large, large-v2, large-v3, large-v3-turbo}, google/gemma-3-4b-it, nvidia/parakeet-tdt"
+    echo "Supported models: mistralai/Voxtral-Mini-3B-2507, mistralai/Voxtral-Mini-4B-Realtime-2602, openai/whisper-{small, medium, large, large-v2, large-v3, large-v3-turbo}, google/gemma-3-4b-it, nvidia/parakeet-tdt"
     exit 1
     ;;
 esac
@@ -201,6 +239,64 @@ if [ "$MODEL_NAME" = "parakeet" ]; then
   exit 0
 fi
 
+# Voxtral Realtime uses a custom export script
+if [ "$MODEL_NAME" = "voxtral_realtime" ]; then
+  pip install safetensors huggingface_hub
+
+  # Download model weights from HuggingFace (requires HF_TOKEN for gated model)
+  LOCAL_MODEL_DIR="${OUTPUT_DIR}/model_weights"
+  python -c "from huggingface_hub import snapshot_download; snapshot_download('${HF_MODEL}', local_dir='${LOCAL_MODEL_DIR}')"
+
+  # Per-component quantization flags
+  VR_QUANT_ARGS=""
+  if [ "$QUANT_NAME" = "quantized-8da4w" ]; then
+    VR_QUANT_ARGS="--qlinear-encoder 8da4w --qlinear 8da4w --qlinear-group-size 32 --qembedding 8w"
+  elif [ "$QUANT_NAME" = "quantized-int4-metal" ]; then
+    VR_QUANT_ARGS="--qlinear-encoder fpa4w --qlinear fpa4w"
+  fi
+
+  # Determine streaming mode based on MODE parameter
+  USE_STREAMING="false"
+  if [ "$MODE" = "vr-streaming" ]; then
+    USE_STREAMING="true"
+  elif [ "$MODE" = "vr-offline" ]; then
+    USE_STREAMING="false"
+  elif [ -z "$MODE" ]; then
+    # Auto-detect: XNNPACK uses streaming, others use offline
+    if [ "$DEVICE" = "xnnpack" ]; then
+      USE_STREAMING="true"
+    fi
+  fi
+
+  # Configure export and preprocessor based on streaming mode
+  STREAMING_ARG=""
+  PREPROCESSOR_ARGS="--feature_size 128 --output_file ${OUTPUT_DIR}/preprocessor.pte"
+  if [ "$USE_STREAMING" = "true" ]; then
+    STREAMING_ARG="--streaming"
+    PREPROCESSOR_ARGS="$PREPROCESSOR_ARGS --streaming"
+  else
+    PREPROCESSOR_ARGS="$PREPROCESSOR_ARGS --stack_output --max_audio_len 300"
+  fi
+
+  python -m executorch.examples.models.voxtral_realtime.export_voxtral_rt \
+      --model-path "$LOCAL_MODEL_DIR" \
+      --backend "$DEVICE" \
+      ${STREAMING_ARG} \
+      --output-dir "${OUTPUT_DIR}" \
+      ${VR_QUANT_ARGS}
+
+  # Export preprocessor
+  python -m executorch.extension.audio.mel_spectrogram ${PREPROCESSOR_ARGS}
+
+  test -f "${OUTPUT_DIR}/model.pte"
+  test -f "${OUTPUT_DIR}/preprocessor.pte"
+  # Copy tokenizer from downloaded model weights
+  cp "$LOCAL_MODEL_DIR/tekken.json" "${OUTPUT_DIR}/tekken.json"
+  ls -al "${OUTPUT_DIR}"
+  echo "::endgroup::"
+  exit 0
+fi
+
 MAX_SEQ_LEN_ARG=""
 if [ -n "$MAX_SEQ_LEN" ]; then
   MAX_SEQ_LEN_ARG="--max_seq_len $MAX_SEQ_LEN"

.ci/scripts/setup-samsung-linux-deps.sh

@@ -14,41 +14,40 @@ if [[ -z "${API_KEY}" ]]; then
   exit 1
 fi
 
-OS_NAME="Ubuntu 22.04"
-LITECORE_BASE="https://soc-developer.semiconductor.samsung.com/api/v1/resource/ai-litecore/download"
-DEVICEFARM_BASE="https://soc-developer.semiconductor.samsung.com/api/v1/resource/remotelab/download"
-
-parse_url() {
-  local json="$1"
-  if command -v jq >/dev/null 2>&1; then
-    jq -r '.data // empty' <<<"$json"
-  else
-    sed -n 's/.*"data":[[:space:]]*"\([^"]*\)".*/\1/p' <<<"$json"
-  fi
-}
+export DEVICE_CONNECT_ENABLED=1
+
+while [[ $# -gt 0 ]]; do
+  case "$1" in
+    --skip-device-connect)
+      export DEVICE_CONNECT_ENABLED=0
+      shift
+      ;;
+    *)
+      # Unknown option
+      shift
+      ;;
+  esac
+done
+
+LITECORE_VERSION="v1.0"
+LITECORE_FILE_NAME="ai-litecore-ubuntu2204-${LITECORE_VERSION}.tar.gz"
+DEVICEFARM_CLI_VERSION="beta-v1.1.0"
+DEVICEFARM_FILE_NAME="devicefarmcli-${DEVICEFARM_CLI_VERSION}.zip"
+
+LITECORE_URL="https://soc-developer.semiconductor.samsung.com/api/v1/resource/download-file/${LITECORE_FILE_NAME}"
+DEVICEFARM_URL="https://soc-developer.semiconductor.samsung.com/api/v1/resource/download-file/${DEVICEFARM_FILE_NAME}"
 
 download_and_extract() {
-  local base_url="$1"
-  local version="$2"
-  local out_dir="$3"
-  local out_file="$4"
+  local download_url="$1"
+  local out_dir="$2"
+  local out_file="$3"
 
-  local resp
-  resp=$(curl -fsSL -G \
+  echo "Downloading from ${download_url}..."
+  curl -fsSL --retry 3 \
     -H "apikey: ${API_KEY}" \
-    --data-urlencode "version=${version}" \
-    --data-urlencode "os=${OS_NAME}" \
-    "${base_url}")
-
-  local download_url
-  download_url=$(parse_url "$resp")
-  if [[ -z "${download_url}" ]]; then
-    echo "ERROR: It failed to download from ${base_url} ."
-    echo "Response: $resp" >&2
-    exit 1
-  fi
+    -o "${out_file}" \
+    "${download_url}"
 
-  curl -fsSL -L --retry 3 -o "${out_file}" "${download_url}"
   echo "Download completed: ${out_file}"
 
   mkdir -p "${out_dir}"
@@ -60,7 +59,7 @@ download_and_extract() {
 
   zip)
     echo "Extracting ZIP..."
-    unzip -q -d "${out_dir}" "${out_file}"
+    unzip -qo -d "${out_dir}" "${out_file}"
     ;;
 
   *)
@@ -71,13 +70,12 @@ download_and_extract() {
 }
 
 download_ai_lite_core() {
-  local litecore_version="${1:-1.0}"
-  local litecore_out="/tmp/exynos-ai-litecore-v${litecore_version}.tar.gz"
+  local litecore_version="${1:-${LITECORE_VERSION}}"
+  local litecore_out="/tmp/${LITECORE_FILE_NAME}"
   local litecore_dir="/tmp/exynos_ai_lite_core"
 
   download_and_extract \
-    "${LITECORE_BASE}" \
-    "${litecore_version}" \
+    "${LITECORE_URL}" \
     "${litecore_dir}" \
     "${litecore_out}"
 
@@ -86,22 +84,21 @@ download_ai_lite_core() {
 }
 
 install_devicefarm_cli() {
-  local cli_version="${1:-beta-1.0.9}"
-  local cli_out="/tmp/devicefarm-cli-v${cli_version}.zip"
+  local cli_version="${1:-${DEVICEFARM_CLI_VERSION}}"
+  local cli_out="/tmp/${DEVICEFARM_FILE_NAME}"
   local cli_dir="/tmp/devicefarm_cli"
 
   download_and_extract \
-    "${DEVICEFARM_BASE}" \
-    "${cli_version}" \
+    "${DEVICEFARM_URL}" \
     "${cli_dir}" \
     "${cli_out}"
 
   export PATH="${PATH%:}:${cli_dir}"
   chmod +x "${cli_dir}/devicefarm-cli"
 }
 
-Enqueue_device_request() {
-  export DEVICE_RESERVED=0
+acquire_device() {
+  export DEVICE_ACQUIRED=0
   if ! command -v devicefarm-cli >/dev/null 2>&1; then
     echo "[WARN] devicefarm-cli is not installed." >&2
     return 1
@@ -138,7 +135,7 @@ Enqueue_device_request() {
 	echo "$out"
 	# Execute test command
 	devicefarm-cli -E "ls /" || true
-	export DEVICE_RESERVED=1
+	export DEVICE_ACQUIRED=1
 	echo "[INFO] Device successfully assigned and connected."
 	return 0
 	;;
@@ -173,10 +170,12 @@ install_enn_backend() {
   export PYTHONPATH="${PYTHONPATH:-}:${EXECUTORCH_ROOT}/.."
 }
 
-litecore_ver="1.0"
-devicefarm_ver="beta-1.0.9"
-
-download_ai_lite_core ${litecore_ver}
-install_devicefarm_cli "${devicefarm_ver}"
+download_ai_lite_core ${LITECORE_VERSION}
 install_enn_backend
-Enqueue_device_request
+
+if [[ "${DEVICE_CONNECT_ENABLED}" == "1" ]]; then
+  install_devicefarm_cli "${DEVICEFARM_CLI_VERSION}"
+  acquire_device
+else
+  export DEVICE_ACQUIRED=0
+fi