Merge branch 'main' into cuda-quiet-symbol-promotion

Author: edmundshieh

.ci/scripts/setup-emscripten.sh

@@ -15,10 +15,15 @@ install_node() {
 }
 
 install_emscripten() {
+    # The WebGPU backend's instance API (WGPUInstanceFeatureName /
+    # WGPUInstanceDescriptor::requiredFeatures) needs the emdawnwebgpu port
+    # pinned by Emscripten 4.0.13 or newer; 4.0.10 ships the older
+    # capabilities-based descriptor and does not compile it.
+    local emscripten_version="${EMSCRIPTEN_VERSION:-4.0.19}"
     git clone https://github.com/emscripten-core/emsdk.git
     pushd emsdk || return
-    ./emsdk install 4.0.10
-    ./emsdk activate 4.0.10
+    ./emsdk install "${emscripten_version}"
+    ./emsdk activate "${emscripten_version}"
     source ./emsdk_env.sh
     popd || return
 }

.ci/scripts/tests/test_cu134_dependencies.py

@@ -150,7 +150,17 @@ def test_python315_cpu_domains_use_available_torchaudio(self):
                     self.assertIn(f"torchaudio=={audio}", domains)
                     self.assertIn("torchvision==0.29.0", domains)
 
-    def test_python315_preserves_cuda_and_source_pins(self):
+    def test_python315_cuda_domains_use_available_torchaudio(self):
+        for machine in ("x86_64", "aarch64"):
+            with self.subTest(machine=machine):
+                _, _, domains, _ = self.install_commands(
+                    (13, 2), machine=machine, python_version=(3, 15)
+                )
+                self.assertIn("torchaudio==2.11.1", domains)
+                self.assertIn("torchvision==0.29.0", domains)
+                self.assertIn("https://download.pytorch.org/whl/test/cu132", domains)
+
+    def test_python315_preserves_cu134_and_source_pins(self):
         for cuda in ((13, 2), (13, 4)):
             for nightly in (True, False):
                 with self.subTest(cuda=cuda, nightly=nightly):
@@ -160,7 +170,7 @@ def test_python315_preserves_cuda_and_source_pins(self):
                     audio = (
                         "torchaudio==2.11.0.dev20260922+cu134"
                         if nightly and cuda == (13, 4)
-                        else "torchaudio==2.11.0" if nightly else "torchaudio"
+                        else "torchaudio==2.11.1" if nightly else "torchaudio"
                     )
                     self.assertIn(audio, domains)
 

.ci/scripts/wheel/envvar_linux.sh

@@ -9,6 +9,3 @@
 # any variables so that subprocesses will see them.
 
 source "${GITHUB_WORKSPACE}/${REPOSITORY}/.ci/scripts/wheel/envvar_base.sh"
-
-# The reusable smoke installer selects PyTorch's index; ET also needs PyPI packages.
-export PIP_EXTRA_INDEX_URL="${PIP_EXTRA_INDEX_URL:+${PIP_EXTRA_INDEX_URL} }https://pypi.org/simple"

.ci/scripts/wheel/pre_build_script.sh

@@ -11,6 +11,10 @@ set -euxo pipefail
 
 # This script is run before building ExecuTorch binaries
 
+# The reusable smoke installer selects PyTorch's index; ET also needs PyPI packages.
+export PIP_EXTRA_INDEX_URL="${PIP_EXTRA_INDEX_URL:+${PIP_EXTRA_INDEX_URL} }https://pypi.org/simple"
+echo "PIP_EXTRA_INDEX_URL=${PIP_EXTRA_INDEX_URL}" >> "${GITHUB_ENV}"
+
 # Initialize submodules here instead of during checkout so we can use OpenSSL
 # on Windows (schannel fails with SEC_E_ILLEGAL_MESSAGE on some gitlab hosts).
 UNAME_S=$(uname -s)
@@ -160,6 +164,21 @@ fi
 
 "${GITHUB_WORKSPACE}/${REPOSITORY}/install_requirements.sh" --example
 
+if [[ $UNAME_S == *"MINGW"* || $UNAME_S == *"MSYS"* ]]; then
+  PYTHON_BIN=$(command -v python || command -v python3)
+  if "${PYTHON_BIN}" -c 'import sys; sys.exit(sys.version_info < (3, 15))'; then
+    # No cp315 tokenizer wheel is published. Build the pinned source with the
+    # existing build dependencies and Visual Studio environment before smoke install.
+    # MSBuild's file tracker needs a short path even during CMake compiler checks.
+    tokenizer_source_dir=$(mktemp -d /c/et-tok.XXXXXX)
+    # Dereference submodule symlinks because MSYS cannot recreate them on every runner.
+    cp -RL extension/llm/tokenizers/. "${tokenizer_source_dir}"
+    cmd //c '.ci\scripts\wheel\vc_env_helper.bat' python -m pip install \
+      --no-build-isolation "$(cygpath -w "${tokenizer_source_dir}")"
+    rm -rf "${tokenizer_source_dir}"
+  fi
+fi
+
 # Enable VGF in pybind wheel builds when the platform-specific build input is
 # available from pip.
 if [[ "$UNAME_S" == "Linux" || "$UNAME_S" == "Darwin" ]]; then

.github/workflows/pull.yml

@@ -1581,6 +1581,10 @@ jobs:
         # Test selective build
         PYTHON_EXECUTABLE=python bash examples/wasm/test_build_wasm.sh
 
+        # Build the WebGPU configuration of the Wasm bindings. Build only:
+        # running a delegated model needs a browser with WebGPU.
+        PYTHON_EXECUTABLE=python bash scripts/build_wasm_tests.sh --webgpu
+
   unittest-wasm-bindings:
     needs: docker-image
     name: unittest-wasm-bindings

CMakeLists.txt

@@ -1543,10 +1543,6 @@ if(EXECUTORCH_BUILD_PYBIND)
   executorch_target_copy_mlx_metallib(portable_lib)
 endif()
 
-if(EXECUTORCH_BUILD_WASM)
-  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/extension/wasm)
-endif()
-
 if(EXECUTORCH_BUILD_TOKENIZERS_WASM)
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/extension/wasm/tokenizers)
 endif()
@@ -1602,6 +1598,12 @@ if(EXECUTORCH_BUILD_VGF)
   list(APPEND _executorch_backends vgf_backend)
 endif()
 
+# Added after the backends so EXECUTORCH_BUILD_WASM_WEBGPU can link
+# webgpu_backend, which has to already be defined at that point.
+if(EXECUTORCH_BUILD_WASM)
+  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/extension/wasm)
+endif()
+
 # Top-level interface targets.
 
 # A target containing all configured backends.

backends/cortex_m/passes/cortex_m_pass_manager.py

@@ -111,13 +111,15 @@ class CortexMPassManager(ExportedProgramPassManager):
     legacy_pass_list: list[PassClass] = [
         # Run before folding so qparams attach to max_pool2d values, not tuple + getitem.
         RemoveGetItemPass,
-        FoldAndAnnotateQParamsPass,
+        # Retrace scalar ops while the graph still has floating-point Q/DQ inputs.
         ReplaceScalarWithTensorArgPass,
-        ReplaceQuantNodesPass,
+        FoldAndAnnotateQParamsPass,
         ActivationFusionPass,
         QuantizedClampActivationPass,
         DecomposeHardswishPass,
         AtenToCortexMPass,
+        # Retrace only after quantized ATen ops have become Cortex-M kernels.
+        ReplaceQuantNodesPass,
         FuseConvPaddingPass,
         InitializeScratchBuffersPass,
         LiftConstantTensorsPass,

backends/cortex_m/test/build_test_runner.sh

@@ -51,8 +51,11 @@ ops_list=(
     aten::cat.out
     aten::full.out
     aten::ge.Tensor_out
+    aten::permute_copy.out
+    aten::slice_copy.Tensor_out
     aten::unsqueeze_copy.out
     aten::squeeze_copy.dim_out
+    aten::squeeze_copy.dims_out
     aten::select_copy.int_out
     aten::amax.out
     cortex_m::quantize_per_tensor.out

backends/cortex_m/test/ops/test_attention.py

@@ -0,0 +1,58 @@
+# Copyright 2026 Arm Limited and/or its affiliates.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+from executorch.backends.arm.test.common import parametrize
+from executorch.backends.cortex_m.test.tester import CortexMTester, McuTestCase
+
+
+class CortexMAttention(torch.nn.Module):
+    ops_before_transforms = {
+        "executorch_exir_dialects_edge__ops_aten_linear_default": 1,
+        "executorch_exir_dialects_edge__ops_aten_bmm_default": 2,
+        "executorch_exir_dialects_edge__ops_aten_mul_Tensor": 1,
+        "executorch_exir_dialects_edge__ops_aten__softmax_default": 1,
+    }
+
+    ops_after_transforms = {
+        "executorch_exir_dialects_edge__ops_cortex_m_quantized_linear_default": 1,
+        "executorch_exir_dialects_edge__ops_cortex_m_quantized_batch_matmul_default": 2,
+        "executorch_exir_dialects_edge__ops_cortex_m_softmax_default": 1,
+    }
+
+    def __init__(self):
+        super().__init__()
+        self.qkv = torch.nn.Linear(12, 36)
+
+    def forward(self, x):
+        qkv = self.qkv(x).reshape(1, 5, 3, 3, 4).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv.unbind(0)
+        scores = (q * 0.5) @ k.transpose(-2, -1)
+        return (scores.softmax(dim=-1) @ v).transpose(1, 2).reshape(1, 5, 12)
+
+
+test_cases = {
+    "qkv_attention": McuTestCase(CortexMAttention(), (torch.randn(1, 5, 12),)),
+}
+
+
+@parametrize("test_case", test_cases)
+def test_dialect_attention(test_case, cortex_m_target):
+    tester = CortexMTester(
+        test_case.model, test_case.example_inputs, target_config=cortex_m_target
+    )
+    tester.test_dialect(
+        test_case.model.ops_before_transforms,
+        test_case.model.ops_after_transforms,
+        qtol=2,
+    )
+
+
+@parametrize("test_case", test_cases)
+def test_implementation_attention(test_case, cortex_m_target):
+    tester = CortexMTester(
+        test_case.model, test_case.example_inputs, target_config=cortex_m_target
+    )
+    tester.test_implementation(qtol=2)

backends/cortex_m/test/ops/test_sdpa.py

@@ -129,9 +129,6 @@ def _qkv(query_shape, key_shape, value_shape, dtype=torch.float32):
     "explicit_scale": (_SCALE_NOT_FOLDED, RuntimeError),
     "multihead_attention": (_LINEAR_RETRACED_AS_INT32, Exception),
 }
-xfail_cases_impl: dict[str, xfail_type] = {
-    "multihead_attention": (_LINEAR_RETRACED_AS_INT32, Exception),
-}
 
 
 def _transform_for_annotation(model, example_inputs, dynamic_shapes=None):
@@ -161,7 +158,7 @@ def test_dialect_sdpa(test_case, cortex_m_target):
     )
 
 
-@parametrize("test_case", test_cases, xfails=xfail_cases_impl)
+@parametrize("test_case", test_cases)
 def test_implementation_sdpa(test_case, cortex_m_target):
     tester = CortexMTester(
         test_case.model, test_case.example_inputs, target_config=cortex_m_target