Merge branch 'main' into change-1105490

Author: AdrianLundell

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

@@ -1 +1 @@
-44d8d54e38c0258357d4e92e1fefe21e845947a3
+467660923a5a25e4718e1d6697b93ff1bab4e807

.ci/docker/ci_commit_pins/pytorch.txt

@@ -1 +1 @@
-53a2908a10f414a2f85caa06703a26a40e873869
+e6f766c7d750d40603eee3f66c5915bac606b3ea

.ci/docker/requirements-ci.txt

@@ -1,12 +1,12 @@
 mpmath==1.3.0
 numpy>=2.0.0; python_version >= '3.10'
 PyYAML==6.0.1
-ruamel.yaml==0.17.32
+ruamel.yaml==0.18.15
 sympy==1.12
 timm==0.6.13
 tomli==2.0.1
 torchsr==1.0.4
-transformers==4.47.1
+transformers==4.56.1
 zstd==1.5.5.1
 pandas>=2.2.2; python_version >= '3.10'
 pytest==7.2.0

.ci/scripts/test-cuda-build.sh

@@ -27,9 +27,6 @@ test_executorch_cuda_build() {
     nvcc --version || echo "nvcc not found"
     nvidia-smi || echo "nvidia-smi not found"
 
-    # Set CMAKE_ARGS to enable CUDA build - ExecuTorch will handle PyTorch installation automatically
-    export CMAKE_ARGS="-DEXECUTORCH_BUILD_CUDA=ON"
-
     echo "=== Starting ExecuTorch Installation ==="
     # Install ExecuTorch with CUDA support with timeout and error handling
     timeout 5400 ./install_executorch.sh || {

.ci/scripts/test_llama_lora.sh

@@ -55,7 +55,7 @@ cmake_build_llama_runner
 # Constants.
 RUNTIME_ARGS="--tokenizer_path=${DOWNLOADED_PATH}/tokenizer.model --temperature=0 --seq_len=20 --warmup=1"
 PROMPT="What happens if you eat watermelon seeds?"
-EXPECTED_PREFIX="What happens if you eat watermelon seeds? Watermelon seeds are a good source of vitamin C,"
+EXPECTED_PREFIX="What happens if you eat watermelon seeds? Watermelon seeds are a good source of vitamin C and"
 
 # Export LoRA PTE file.
 MODEL_NAME="llama_3_2_1B_lora"
@@ -94,7 +94,7 @@ else
   exit 1
 fi
 
-# Export LoRA PTE, PTD file.
+# Export LoRA PTE, foundation PTD file.
 MODEL_SEPARATE="${MODEL_NAME}_separate"
 $PYTHON_EXECUTABLE -m extension.llm.export.export_llm \
     base.checkpoint="${DOWNLOADED_PATH}/consolidated.00.pth" \
@@ -114,20 +114,62 @@ $PYTHON_EXECUTABLE -m extension.llm.export.export_llm \
 NOW=$(date +"%H:%M:%S")
 echo "Starting to run llama runner at ${NOW}"
 # shellcheck source=/dev/null
-cmake-out/examples/models/llama/llama_main --model_path=${MODEL_SEPARATE}.pte --data_path=${MODEL_SEPARATE}.ptd --prompt="${PROMPT}" ${RUNTIME_ARGS} > result2.txt
+cmake-out/examples/models/llama/llama_main --model_path=${MODEL_SEPARATE}.pte --data_paths=${MODEL_SEPARATE}.ptd --prompt="${PROMPT}" ${RUNTIME_ARGS} > result2.txt
 NOW=$(date +"%H:%M:%S")
 echo "Finished at ${NOW}"
 
 RESULT2=$(cat result2.txt)
 if [[ "${RESULT2}" == "${EXPECTED_PREFIX}"* ]]; then
   echo "Expected result prefix: ${EXPECTED_PREFIX}"
   echo "Actual result: ${RESULT2}"
+  # Do not clean up files if test passes, as they're re-used in the next test.
   echo "Success"
-  cleanup_files
 else
   echo "Expected result prefix: ${EXPECTED_PREFIX}"
   echo "Actual result: ${RESULT2}"
   echo "Failure; results not the same"
   cleanup_files
   exit 1
 fi
+
+# Export LoRA PTE, LoRA PTD, foundation PTD file.
+MODEL_PROGRAM_ONLY="${MODEL_NAME}_program"
+MODEL_LORA_WEIGHTS="lora_weights"
+MODEL_FOUNDATION_WEIGHTS="foundation_weights"
+$PYTHON_EXECUTABLE -m extension.llm.export.export_llm \
+    base.checkpoint="${DOWNLOADED_PATH}/consolidated.00.pth" \
+    base.params="${DOWNLOADED_PATH}/params.json" \
+    base.adapter_checkpoint="${DOWNLOADED_PATH}/adapter_model.pt" \
+    base.adapter_config="${DOWNLOADED_PATH}/adapter_config.json" \
+    base.tokenizer_path="${DOWNLOADED_PATH}/tokenizer.model" \
+    model.use_kv_cache=true \
+    model.use_sdpa_with_kv_cache=true \
+    model.dtype_override="fp32" \
+    backend.xnnpack.enabled=true \
+    backend.xnnpack.extended_ops=true \
+    export.output_name="${MODEL_PROGRAM_ONLY}.pte" \
+    export.foundation_weights_file="${MODEL_FOUNDATION_WEIGHTS}.ptd" \
+    export.lora_weights_file="${MODEL_LORA_WEIGHTS}.ptd"
+
+# Run llama runner.
+NOW=$(date +"%H:%M:%S")
+echo "Starting to run llama runner at ${NOW}"
+# shellcheck source=/dev/null
+cmake-out/examples/models/llama/llama_main --model_path=${MODEL_PROGRAM_ONLY}.pte --data_paths="${MODEL_FOUNDATION_WEIGHTS}.ptd,${MODEL_LORA_WEIGHTS}.ptd" --prompt="${PROMPT}" ${RUNTIME_ARGS} > result3.txt
+NOW=$(date +"%H:%M:%S")
+echo "Finished at ${NOW}"
+
+RESULT3=$(cat result3.txt)
+if [[ "${RESULT3}" == "${EXPECTED_PREFIX}"* ]]; then
+  echo "Expected result prefix: ${EXPECTED_PREFIX}"
+  echo "Actual result: ${RESULT3}"
+  echo "Success"
+else
+  echo "Expected result prefix: ${EXPECTED_PREFIX}"
+  echo "Actual result: ${RESULT3}"
+  echo "Failure; results not the same"
+  cleanup_files
+  exit 1
+fi
+
+cleanup_files

.ci/scripts/test_phi_3_mini.sh

@@ -36,34 +36,33 @@ cmake_build_phi_3_mini() {
   cmake --build ${BUILD_DIR}/${MODEL_DIR} -j${NPROC} --config ${BUILD_TYPE}
 }
 
-# Download and convert tokenizer.model
+# Download tokenizer.model
 prepare_tokenizer() {
-  echo "Downloading and converting tokenizer.model"
-  wget -O tokenizer.model "https://huggingface.co/microsoft/Phi-3-mini-128k-instruct/resolve/main/tokenizer.model?download=true"
-  $PYTHON_EXECUTABLE -m pytorch_tokenizers.tools.llama2c.convert -t tokenizer.model -o tokenizer.bin
+  echo "Downloading tokenizer.model"
+  wget -O tokenizer.model "https://huggingface.co/microsoft/Phi-3-mini-4k-instruct/resolve/main/tokenizer.model?download=true"
 }
 
 # Export phi-3-mini model to pte
 export_phi_3_mini () {
   echo "Exporting phi-3-mini. This will take a few minutes"
-  $PYTHON_EXECUTABLE -m executorch.examples.models.phi-3-mini.export_phi-3-mini -c "4k" -s 128 -o phi-3-mini.pte
+  optimum-cli export executorch --model microsoft/Phi-3-mini-4k-instruct --task text-generation --recipe xnnpack --output_dir ./
 }
 
 run_and_verify() {
     NOW=$(date +"%H:%M:%S")
     echo "Starting to run phi-3-mini runner at ${NOW}"
-    if [[ ! -f "phi-3-mini.pte" ]]; then
-        echo "Export failed. Abort"
+    if [[ ! -f "model.pte" ]]; then
+        echo "Missing model artifact. Abort"
         exit 1
     fi
-    if [[ ! -f "tokenizer.bin" ]]; then
-        echo "tokenizer.bin is missing."
+    if [[ ! -f "tokenizer.model" ]]; then
+        echo "tokenizer.model is missing."
         exit 1
     fi
 
     ${BUILD_DIR}/${MODEL_DIR}/phi_3_mini_runner \
-    --model_path=phi-3-mini.pte \
-    --tokenizer_path=tokenizer.bin \
+    --model_path=model.pte \
+    --tokenizer_path=tokenizer.model \
     --seq_len=60 \
     --temperature=0 \
     --prompt="<|system|>
@@ -92,7 +91,7 @@ What is the capital of France?<|end|>
 cmake_install_executorch_libraries
 cmake_build_phi_3_mini
 
-# Step 2. Export the tokenizer and model
+# Step 2. Export the model
 prepare_tokenizer
 export_phi_3_mini
 

.ci/scripts/utils.sh

@@ -44,6 +44,44 @@ install_pip_dependencies() {
   popd || return
 }
 
+dedupe_macos_loader_path_rpaths() {
+  if [[ "$(uname)" != "Darwin" ]]; then
+    return
+  fi
+
+  local torch_lib_dir
+  pushd ..
+  torch_lib_dir=$(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
+
+  local torch_libs=(
+    "libtorch_cpu.dylib"
+    "libtorch.dylib"
+    "libc10.dylib"
+  )
+
+  for lib_name in "${torch_libs[@]}"; do
+    local lib_path="${torch_lib_dir}/${lib_name}"
+    if [[ ! -f "${lib_path}" ]]; then
+      continue
+    fi
+
+    local removed=0
+    # Repeatedly remove the @loader_path rpath entries until none remain.
+    while install_name_tool -delete_rpath @loader_path "${lib_path}" 2>/dev/null; do
+      removed=1
+    done
+
+    if [[ "${removed}" == "1" ]]; then
+      install_name_tool -add_rpath @loader_path "${lib_path}" || true
+    fi
+  done
+}
+
 install_domains() {
   echo "Install torchvision and torchaudio"
   pip install --no-use-pep517 --user "git+https://github.com/pytorch/audio.git@${TORCHAUDIO_VERSION}"
@@ -101,6 +139,7 @@ install_pytorch_and_domains() {
     echo "Use cached wheel at ${cached_torch_wheel}"
   fi
 
+  dedupe_macos_loader_path_rpaths
   # Grab the pinned audio and vision commits from PyTorch
   TORCHAUDIO_VERSION=$(cat .github/ci_commit_pins/audio.txt)
   export TORCHAUDIO_VERSION

.githooks/pre-commit

@@ -1,22 +1,24 @@
 #!/usr/bin/env bash
 
-# Pre-commit hook to automatically update PyTorch commit pin when torch_pin.py changes
+# Pre-commit hook to automatically update PyTorch commit pin and sync c10 directories when torch_pin.py changes
 
 # Check if torch_pin.py is being committed
 if git diff --cached --name-only | grep -q "^torch_pin.py$"; then
     echo "🔍 Detected changes to torch_pin.py"
-    echo "📝 Updating PyTorch commit pin..."
+    echo "📝 Updating PyTorch commit pin and syncing c10 directories..."
 
-    # Run the update script
+    # Run the update script (which now also syncs c10 directories)
     if python .github/scripts/update_pytorch_pin.py; then
-        # Check if pytorch.txt was modified
+        # Stage any modified files (pytorch.txt and grafted c10 files)
         if ! git diff --quiet .ci/docker/ci_commit_pins/pytorch.txt; then
-            echo "✅ PyTorch commit pin updated successfully"
-            # Stage the updated file
             git add .ci/docker/ci_commit_pins/pytorch.txt
             echo "📌 Staged .ci/docker/ci_commit_pins/pytorch.txt"
-        else
-            echo "ℹ️  PyTorch commit pin unchanged"
+        fi
+
+        # Stage any grafted c10 files
+        if ! git diff --quiet runtime/core/portable_type/c10/; then
+            git add runtime/core/portable_type/c10/
+            echo "📌 Staged grafted c10 files"
         fi
     else
         echo "❌ Failed to update PyTorch commit pin"