Update base for Update on "Export lora weights to sep file"

Differential Revision: [D83777195](https://our.internmc.facebook.com/intern/diff/D83777195/)

[ghstack-poisoned]

Author: lucylq

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

@@ -1 +1 @@
-bd06b54e627fbfd354a2cffa4c80fb21883209a9
+44d8d54e38c0258357d4e92e1fefe21e845947a3

.github/workflows/cuda.yml

@@ -86,3 +86,87 @@ jobs:
         PYTHON_EXECUTABLE=python CMAKE_ARGS="-DEXECUTORCH_BUILD_CUDA=ON" ./install_executorch.sh
         export LD_LIBRARY_PATH=/opt/conda/lib:$LD_LIBRARY_PATH
         PYTHON_EXECUTABLE=python source .ci/scripts/test_model.sh "${{ matrix.model }}" cmake cuda
+
+  test-voxtral-cuda-e2e:
+    name: test-voxtral-cuda-e2e
+    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
+    permissions:
+      id-token: write
+      contents: read
+    secrets: inherit
+    strategy:
+      fail-fast: false
+    with:
+      timeout: 90
+      secrets-env: EXECUTORCH_HF_TOKEN
+      runner: linux.g5.4xlarge.nvidia.gpu
+      gpu-arch-type: cuda
+      gpu-arch-version: 12.6
+      use-custom-docker-registry: false
+      submodules: recursive
+      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      script: |
+        set -eux
+
+        echo "::group::Setup ExecuTorch"
+        CMAKE_ARGS="-DEXECUTORCH_BUILD_CUDA=ON" ./install_executorch.sh
+        echo "::endgroup::"
+
+        echo "::group::Setup Huggingface"
+        pip install -U "huggingface_hub[cli]" accelerate
+        huggingface-cli login --token $SECRET_EXECUTORCH_HF_TOKEN
+        OPTIMUM_ET_VERSION=$(cat .ci/docker/ci_commit_pins/optimum-executorch.txt)
+        pip install git+https://github.com/huggingface/optimum-executorch.git@${OPTIMUM_ET_VERSION}
+        pip install mistral-common librosa
+        echo "::endgroup::"
+
+        echo "::group::Export Voxtral"
+        optimum-cli export executorch \
+            --model "mistralai/Voxtral-Mini-3B-2507" \
+            --task "multimodal-text-to-text" \
+            --recipe "cuda" \
+            --dtype bfloat16 \
+            --device cuda \
+            --max_seq_len 1024 \
+            --output_dir ./
+        echo "::endgroup::"
+
+        echo "::group::Build Voxtral Runner"
+        cmake -DCMAKE_BUILD_TYPE=Release \
+              -DEXECUTORCH_BUILD_CUDA=ON \
+              -DEXECUTORCH_BUILD_EXTENSION_TENSOR=ON \
+              -DEXECUTORCH_BUILD_EXTENSION_MODULE=ON \
+              -DEXECUTORCH_BUILD_EXTENSION_NAMED_DATA_MAP=ON \
+              -DEXECUTORCH_BUILD_TESTS=ON \
+              -Bcmake-out .
+        cmake --build cmake-out -j$(( $(nproc) - 1 )) --target voxtral_runner
+        echo "::endgroup::"
+
+        echo "::group::Run Voxtral Runner"
+        # Capture output and allow exit code 139 if we have the expected printout
+        set +e
+        export LD_LIBRARY_PATH=/opt/conda/lib:$LD_LIBRARY_PATH
+        OUTPUT=$(cmake-out/backends/cuda/voxtral_runner model.pte aoti_cuda_blob.ptd 2>&1)
+        EXIT_CODE=$?
+        set -e
+
+        echo "$OUTPUT"
+
+        # Check if the output contains "Run latency (ms):"
+        if echo "$OUTPUT" | grep -q "Run latency (ms):"; then
+          echo "Found expected output: 'Run latency (ms):'"
+          if [ $EXIT_CODE -eq 139 ]; then
+            echo "Exit code 139 (segfault) detected, but passing since we have the expected output"
+            exit 0
+          elif [ $EXIT_CODE -ne 0 ]; then
+            echo "Unexpected exit code: $EXIT_CODE"
+            exit $EXIT_CODE
+          else
+            echo "Command succeeded with exit code 0"
+            exit 0
+          fi
+        else
+          echo "Expected output 'Run latency (ms):' not found in output"
+          exit 1
+        fi
+        echo "::endgroup::"

CMakeLists.txt

@@ -266,6 +266,18 @@ if(EXECUTORCH_BUILD_PTHREADPOOL)
   executorch_move_interface_include_directories_to_build_time_only(
     pthreadpool_interface
   )
+
+  if(APPLE)
+    # Use hidden visibility for pthreadpool on Apple platforms to avoid issues
+    # with pthreadpool symbols from libtorch_cpu taking precedence over the ones
+    # from the pthreadpool library statically linked in _portable_lib. The
+    # pthreadpool public APIs are marked as weak by default on some Apple
+    # platforms, so setting to hidden visibility works around this by not
+    # putting the symbol in the indirection table. See
+    # https://github.com/pytorch/executorch/issues/14321 for more details.
+    target_compile_options(pthreadpool PRIVATE -fvisibility=hidden)
+  endif()
+
   install(
     TARGETS pthreadpool pthreadpool_interface fxdiv
     EXPORT ExecuTorchTargets

backends/aoti/CMakeLists.txt

@@ -40,13 +40,8 @@ target_compile_options(aoti_common PUBLIC -fexceptions -frtti -fPIC)
 # Ensure symbols are exported properly
 target_link_options(aoti_common PUBLIC -Wl,--export-dynamic)
 
-# Link against PyTorch libraries and standard libraries
-target_link_libraries(
-  aoti_common
-  PUBLIC extension_tensor ${CMAKE_DL_LIBS}
-         # Link PyTorch libraries for AOTI functions
-         ${TORCH_LIBRARIES}
-)
+# Link against ExecuTorch libraries and standard libraries
+target_link_libraries(aoti_common PUBLIC extension_tensor ${CMAKE_DL_LIBS})
 executorch_target_link_options_shared_lib(aoti_common)
 
 install(

backends/aoti/aoti_model_container.h

@@ -77,6 +77,8 @@ struct AOTIDelegateHandle {
   void* so_handle;
   std::string so_path;
   AOTInductorModelContainerHandle container_handle;
+  void* cuda_stream; // cudaStream_t stored as void* to avoid CUDA header
+                     // dependency
 };
 
 } // namespace aoti

backends/aoti/common_shims.cpp

@@ -127,11 +127,18 @@ int32_t aoti_torch_layout_strided() {
 }
 
 // Dtype constants - these return the PyTorch dtype codes
-// Currently only float32 is supported, but using robust enum-based approach
 int32_t aoti_torch_dtype_float32() {
   return 6; // PyTorch's float32 dtype code
 }
 
+int32_t aoti_torch_dtype_bfloat16() {
+  return 15; // PyTorch's bfloat16 dtype code
+}
+
+int32_t aoti_torch_dtype_int64() {
+  return 4; // PyTorch's int64 dtype code
+}
+
 // Cleanup functions
 void cleanup_tensor_metadata() {
   internal::tensor_to_sizes.clear();

backends/aoti/common_shims.h

@@ -58,6 +58,8 @@ AOTITorchError aoti_torch_get_dim(Tensor* tensor, int64_t* ret_dim);
 int32_t aoti_torch_device_type_cpu();
 int32_t aoti_torch_layout_strided();
 int32_t aoti_torch_dtype_float32();
+int32_t aoti_torch_dtype_bfloat16();
+int32_t aoti_torch_dtype_int64();
 
 // Autograd mode functions
 int32_t aoti_torch_grad_mode_is_enabled();

backends/aoti/targets.bzl

@@ -51,7 +51,7 @@ def define_common_targets():
         link_whole = True,
         supports_python_dlopen = True,
         visibility = ["@EXECUTORCH_CLIENTS"],
-        deps = [
+        exported_deps = [
             ":common_shims",
             ":model_container",
         ],

backends/aoti/utils.h

@@ -34,6 +34,8 @@ inline executorch::aten::ScalarType dtype_to_scalar_type(int32_t dtype) {
   // Convert based on known PyTorch dtype codes (without CUDA-specific
   // dependency)
   switch (dtype) {
+    case 4: // PyTorch's int64 dtype code
+      return executorch::aten::ScalarType::Long;
     case 6: // PyTorch's float32 dtype code
       return executorch::aten::ScalarType::Float;
     case 15: // PyTorch's bfloat16 dtype code

backends/arm/_passes/__init__.py

@@ -91,6 +91,7 @@
     ReplaceScalarWithTensorArgPassTOSABI,
     ReplaceScalarWithTensorArgPassTOSAMI,
 )
+from .rewrite_matmul import RewriteMatmulPass  # noqa
 from .rewrite_upsample import RewriteUpsamplePass  # noqa
 from .scalars_to_attribute_pass import ScalarsToAttributePass  # noqa
 from .size_adjust_input_pass import SizeAdjustInputPass  # noqa