Merge branch 'main' into main

Author: psiddh

.ci/docker/common/install_openssl.sh

@@ -5,8 +5,10 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
-# NB: This script is needed for sccache and is adopted from PyTorch core repo at
-# https://github.com/pytorch/pytorch/blob/main/.ci/docker/common/install_openssl.sh
+# NB: This script is needed for sccache and is adopted from PyTorch core repo.
+# Core deleted its copy in pytorch/pytorch#179513, so this points at the last
+# revision that still had it instead of at a branch.
+# https://github.com/pytorch/pytorch/blob/9274b93eac9f026e9f51e282449b0300b8e05482/.ci/docker/common/install_openssl.sh
 set -ex
 
 OPENSSL=openssl-1.1.1k

.ci/docker/common/install_pytorch.sh

@@ -49,8 +49,29 @@ install_pytorch_and_domains() {
     export MAX_JOBS="${PYTORCH_BUILD_MAX_JOBS}"
   fi
   configure_pytorch_compiler
-  # Then build and install PyTorch
-  conda_run python setup.py bdist_wheel
+  # PyTorch no longer supports "python setup.py bdist_wheel"; it now builds
+  # through scikit-build-core (PEP 517). Build the wheel with the standard
+  # frontend and keep build isolation off, so PyTorch builds against this
+  # environment's numpy and toolchain (avoids an ABI mismatch) and reuses
+  # sccache.
+  #
+  # With isolation off the frontend does not fetch the PEP 517 build
+  # requirements, so install them here. They go into a throwaway venv rather
+  # than into the image, because scikit-build-core registers a setuptools
+  # build_ext plugin: left in the image it hijacks every later
+  # "pip install --no-build-isolation" and turns on C++20 module scanning that
+  # the image compiler cannot satisfy. The venv inherits the image's
+  # site-packages, so PyTorch still builds against the same numpy.
+  #
+  # Keep the list in sync with pytorch/pyproject.toml [build-system].requires.
+  local build_venv=/tmp/pytorch-build-venv
+  rm -rf "${build_venv}"
+  conda_run python -m venv --system-site-packages "${build_venv}"
+  conda_run "${build_venv}/bin/pip" install build "scikit-build-core>=1.0" \
+    "setuptools>=77.0.0,<82" "cmake>=3.27,<4" ninja "packaging>=24.2" \
+    "typing-extensions>=4.10.0" pyyaml six
+  conda_run "${build_venv}/bin/python" -m build --wheel --no-isolation
+  rm -rf "${build_venv}"
   pip_install "$(echo dist/*.whl)"
 
   # Grab the pinned audio and vision commits from PyTorch

.ci/scripts/export_model_artifact.sh

@@ -25,6 +25,7 @@ Arguments:
                  - nvidia/diar_streaming_sortformer_4spk-v2
                  - nvidia/parakeet-tdt
                  - facebook/dinov2-small-imagenet1k-1-layer
+                 - meta-models/Muse-Glimmer-30B-GGUF
 
   quant_name   Quantization type (optional, default: non-quantized)
                Options:
@@ -33,13 +34,17 @@ Arguments:
                  - quantized-int4-weight-only (CUDA only)
                  - quantized-int4-metal (Metal only)
                  - quantized-8da4w (XNNPACK only)
+                 - kquant-17gb (Muse Glimmer only)
+                 - kquant-dynamic (Muse Glimmer only)
 
   output_dir   Output directory for artifacts (optional, default: current directory)
 
   mode         Export mode (optional, default: vr-streaming)
                Supported modes:
                  - vr-streaming: Voxtral Realtime streaming mode
                  - vr-offline: Voxtral Realtime offline mode
+                 - solo-text: Muse Glimmer solo text mode
+                 - dflash-image: Muse Glimmer DFlash vision mode
 
 Examples:
   export_model_artifact.sh metal "openai/whisper-small"
@@ -53,6 +58,7 @@ Examples:
   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"
+  export_model_artifact.sh cuda "meta-models/Muse-Glimmer-30B-GGUF" "kquant-17gb" "./output" "solo-text"
 EOF
 }
 
@@ -89,9 +95,16 @@ if [ -n "$MODE" ]; then
         exit 1
       fi
       ;;
+    solo-text|dflash-image)
+      if [ "$HF_MODEL" != "meta-models/Muse-Glimmer-30B-GGUF" ]; then
+        echo "Error: Mode '$MODE' can only be used with Muse Glimmer model"
+        echo "Provided model: $HF_MODEL"
+        exit 1
+      fi
+      ;;
     *)
       echo "Error: Unsupported mode '$MODE'"
-      echo "Supported modes: vr-streaming, vr-offline"
+      echo "Supported modes: vr-streaming, vr-offline, solo-text, dflash-image"
       exit 1
       ;;
   esac
@@ -203,9 +216,17 @@ case "$HF_MODEL" in
     PREPROCESSOR_FEATURE_SIZE=""
     PREPROCESSOR_OUTPUT=""
     ;;
+  meta-models/Muse-Glimmer-30B-GGUF)
+    MODEL_NAME="muse_glimmer"
+    TASK=""
+    MAX_SEQ_LEN=""
+    EXTRA_PIP=""
+    PREPROCESSOR_FEATURE_SIZE=""
+    PREPROCESSOR_OUTPUT=""
+    ;;
   *)
     echo "Error: Unsupported model '$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, facebook/dinov2-small-imagenet1k-1-layer, SocialLocalMobile/Qwen3.5-35B-A3B-HQQ-INT4, unsloth/gemma-4-31B-it-GGUF"
+    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, SocialLocalMobile/Qwen3.5-35B-A3B-HQQ-INT4, unsloth/gemma-4-31B-it-GGUF, meta-models/Muse-Glimmer-30B-GGUF"
     exit 1
     ;;
 esac
@@ -243,13 +264,36 @@ case "$QUANT_NAME" in
     fi
     EXTRA_ARGS="--qlinear 8da4w --qlinear_group_size 32 --qlinear_encoder 8da4w --qlinear_encoder_group_size 32"
     ;;
+  kquant-17gb|kquant-dynamic)
+    if [ "$HF_MODEL" != "meta-models/Muse-Glimmer-30B-GGUF" ]; then
+      echo "Error: Quantization '$QUANT_NAME' can only be used with Muse Glimmer model"
+      echo "Provided model: $HF_MODEL"
+      exit 1
+    fi
+    EXTRA_ARGS=""
+    ;;
   *)
     echo "Error: Unsupported quantization '$QUANT_NAME'"
-    echo "Supported quantizations: non-quantized, quantized-int4-tile-packed, quantized-int4-weight-only, quantized-int4-metal, quantized-8da4w"
+    echo "Supported quantizations: non-quantized, quantized-int4-tile-packed, quantized-int4-weight-only, quantized-int4-metal, quantized-8da4w, kquant-17gb, kquant-dynamic"
     exit 1
     ;;
 esac
 
+if [ "$MODEL_NAME" = "muse_glimmer" ]; then
+  if [ "$DEVICE" != "cuda" ]; then
+    echo "Error: Muse Glimmer is only supported with the cuda device"
+    exit 1
+  fi
+  if [ "$QUANT_NAME" != "kquant-17gb" ] && [ "$QUANT_NAME" != "kquant-dynamic" ]; then
+    echo "Error: Muse Glimmer requires quantization 'kquant-17gb' or 'kquant-dynamic'"
+    exit 1
+  fi
+  if [ "$MODE" != "solo-text" ] && [ "$MODE" != "dflash-image" ]; then
+    echo "Error: Muse Glimmer requires mode 'solo-text' or 'dflash-image'"
+    exit 1
+  fi
+fi
+
 echo "::group::Export $MODEL_NAME"
 
 if [ -n "$EXTRA_PIP" ]; then
@@ -470,6 +514,73 @@ if [ "$MODEL_NAME" = "qwen3_5_moe" ]; then
   exit 0
 fi
 
+# Muse Glimmer: download the selected GGUFs and export the requested CUDA configuration.
+if [ "$MODEL_NAME" = "muse_glimmer" ]; then
+  pip install safetensors huggingface_hub gguf
+
+  LOCAL_MODEL_DIR=$(mktemp -d)
+  INDUCTOR_CACHE=$(mktemp -d "${RUNNER_TEMP:-/tmp}/inductor_cache_XXXXXX")
+  INDUCTOR_TMPDIR=$(mktemp -d "${RUNNER_TEMP:-/tmp}/tmpdir_XXXXXX")
+  trap 'rm -rf "$LOCAL_MODEL_DIR" "$INDUCTOR_CACHE" "$INDUCTOR_TMPDIR"' EXIT
+
+  case "$QUANT_NAME" in
+    kquant-17gb)
+      TARGET_GGUF_FILE="Muse-Glimmer-30B-KQuant-17GB-Q4_K_M.gguf"
+      ;;
+    kquant-dynamic)
+      TARGET_GGUF_FILE="Muse-Glimmer-30B-KQuant-Dynamic-Q4_K_XL.gguf"
+      ;;
+  esac
+
+  python -c "from huggingface_hub import hf_hub_download; hf_hub_download('${HF_MODEL}', '${TARGET_GGUF_FILE}', local_dir='${LOCAL_MODEL_DIR}')"
+  TARGET_GGUF_PATH="${LOCAL_MODEL_DIR}/${TARGET_GGUF_FILE}"
+
+  echo "::group::Export"
+  case "$MODE" in
+    solo-text)
+      EXPORT_START_SECONDS=$SECONDS
+      TMPDIR="$INDUCTOR_TMPDIR" \
+      TORCHINDUCTOR_CACHE_DIR="$INDUCTOR_CACHE" \
+      python -m executorch.examples.models.muse_glimmer.export.export_solo \
+          --gguf "$TARGET_GGUF_PATH" \
+          --backend cuda \
+          --output-dir "${OUTPUT_DIR}"
+      ;;
+    dflash-image)
+      DRAFT_GGUF_FILE="dflash-Muse-Glimmer-30B-Q4_K_M.gguf"
+      MMPROJ_GGUF_FILE="mmproj-Muse-Glimmer-30B-Q4_K_M.gguf"
+      python -c "from huggingface_hub import hf_hub_download; hf_hub_download('${HF_MODEL}', '${DRAFT_GGUF_FILE}', local_dir='${LOCAL_MODEL_DIR}')"
+      python -c "from huggingface_hub import hf_hub_download; hf_hub_download('${HF_MODEL}', '${MMPROJ_GGUF_FILE}', local_dir='${LOCAL_MODEL_DIR}')"
+      EXPORT_START_SECONDS=$SECONDS
+      TMPDIR="$INDUCTOR_TMPDIR" \
+      TORCHINDUCTOR_CACHE_DIR="$INDUCTOR_CACHE" \
+      python -m executorch.examples.models.muse_glimmer.export.export_dflash \
+          --target-gguf "$TARGET_GGUF_PATH" \
+          --draft-gguf "${LOCAL_MODEL_DIR}/${DRAFT_GGUF_FILE}" \
+          --mmproj "${LOCAL_MODEL_DIR}/${MMPROJ_GGUF_FILE}" \
+          --backend cuda \
+          --output-dir "${OUTPUT_DIR}"
+      ;;
+    *)
+      echo "Error: Muse Glimmer requires mode 'solo-text' or 'dflash-image'"
+      exit 1
+      ;;
+  esac
+  EXPORT_DURATION_SECONDS=$((SECONDS - EXPORT_START_SECONDS))
+  EXPORT_DURATION_MINUTES=$(awk -v seconds="$EXPORT_DURATION_SECONDS" 'BEGIN {printf "%.2f", seconds / 60}')
+  echo "Muse Glimmer took ${EXPORT_DURATION_MINUTES} minutes to export."
+  echo "::endgroup::"
+
+  test -f "${OUTPUT_DIR}/model.pte"
+  test -f "${OUTPUT_DIR}/aoti_cuda_blob.ptd"
+  if [ "$MODE" = "dflash-image" ]; then
+    test -f "${OUTPUT_DIR}/pos_embed.bin"
+  fi
+  ls -al "${OUTPUT_DIR}"
+
+  exit 0
+fi
+
 # Gemma 4 31B: download the Q4_K_M GGUF and export via the GGUF loader
 if [ "$MODEL_NAME" = "gemma4_31b" ]; then
   pip install safetensors huggingface_hub gguf

.ci/scripts/setup-windows-arm64.ps1

@@ -0,0 +1,48 @@
+Set-PSDebug -Trace 1
+$ErrorActionPreference = 'Stop'
+$PSNativeCommandUseErrorActionPreference = $true
+
+$vsWhere = "${env:ProgramFiles(x86)}\Microsoft Visual Studio\Installer\vswhere.exe"
+$vsInstallPath = & $vsWhere -latest -products * `
+    -requires Microsoft.VisualStudio.Component.VC.Tools.ARM64 `
+    -property installationPath
+if (-not $vsInstallPath) {
+    throw "Visual Studio with the ARM64 C++ toolchain was not found."
+}
+$vsDevShell = Join-Path $vsInstallPath "Common7\Tools\Launch-VsDevShell.ps1"
+& $vsDevShell -Arch arm64 -HostArch amd64
+
+$buildDir = "cmake-out-windows-arm64"
+if (Test-Path -Path $buildDir) {
+    Remove-Item -Path $buildDir -Recurse -Force
+}
+
+# XNNPACK's optional ARM ISA and assembly microkernels do not build with MSVC.
+# Baseline ARM64 NEON kernels remain enabled.
+cmake -S . -B $buildDir `
+    -G "Visual Studio 17 2022" `
+    -A ARM64 `
+    -DCMAKE_BUILD_TYPE=Release `
+    -DCMAKE_CXX_STANDARD=20 `
+    -DEXECUTORCH_BUILD_EXECUTOR_RUNNER=ON `
+    -DEXECUTORCH_BUILD_EXTENSION_DATA_LOADER=ON `
+    -DEXECUTORCH_BUILD_EXTENSION_MODULE=ON `
+    -DEXECUTORCH_BUILD_EXTENSION_TENSOR=ON `
+    -DEXECUTORCH_BUILD_EXTENSION_FLAT_TENSOR=ON `
+    -DEXECUTORCH_BUILD_EXTENSION_NAMED_DATA_MAP=ON `
+    -DEXECUTORCH_BUILD_KERNELS_OPTIMIZED=ON `
+    -DEXECUTORCH_BUILD_KERNELS_CUSTOM=OFF `
+    -DEXECUTORCH_BUILD_KERNELS_CUSTOM_AOT=OFF `
+    -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=OFF `
+    -DEXECUTORCH_BUILD_XNNPACK=ON `
+    -DEXECUTORCH_BUILD_EXTENSION_LLM=ON `
+    -DEXECUTORCH_BUILD_EXTENSION_LLM_RUNNER=ON `
+    -DXNNPACK_ENABLE_ASSEMBLY=OFF `
+    -DXNNPACK_ENABLE_ARM_BF16=OFF `
+    -DXNNPACK_ENABLE_ARM_DOTPROD=OFF `
+    -DXNNPACK_ENABLE_ARM_FP16_SCALAR=OFF `
+    -DXNNPACK_ENABLE_ARM_FP16_VECTOR=OFF
+
+cmake --build $buildDir --config Release -j $env:NUMBER_OF_PROCESSORS
+
+Write-Host "Windows ARM64 build completed successfully!"

.ci/scripts/test_llava.sh

@@ -114,7 +114,7 @@ export_llava() {
 # Download a new image
 download_image() {
     echo "Downloading image"
-    curl -o basketball.jpg https://upload.wikimedia.org/wikipedia/commons/7/73/Chicago_Bulls_and_New_Jersey_Nets%2C_March_28%2C_1991.jpg
+    curl -fL -o basketball.jpg https://upload.wikimedia.org/wikipedia/commons/3/3e/Chicago_Bulls_-_New_Jersey_Nets_match_on_March_28%2C_1991.jpg
 }
 
 run_and_verify() {