Update

[ghstack-poisoned]

Author: swolchok

.ci/docker/ci_commit_pins/pytorch.txt

@@ -1 +1 @@
-295f2ed4d103017f7e19a7b8263ece606cd629db
+7ae0ce6360b6e4f944906502d20da24c04debee5

.ci/scripts/unittest-linux.sh

@@ -21,8 +21,7 @@ if [[ "$BUILD_TOOL" == "cmake" ]]; then
     source .ci/scripts/setup-vulkan-linux-deps.sh
 
     PYTHON_EXECUTABLE=python \
-    EXECUTORCH_BUILD_PYBIND=ON \
-    CMAKE_ARGS="-DEXECUTORCH_BUILD_XNNPACK=ON -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON" \
+    CMAKE_ARGS="-DEXECUTORCH_BUILD_PYBIND=ON -DEXECUTORCH_BUILD_XNNPACK=ON -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON" \
     .ci/scripts/setup-linux.sh "$@"
 
     # Install llama3_2_vision dependencies.

.ci/scripts/unittest-macos.sh

@@ -21,8 +21,7 @@ trap 'rm -rfv ${TMP_DIR}' EXIT
 
 # Setup MacOS dependencies as there is no Docker support on MacOS atm
 PYTHON_EXECUTABLE=python \
-EXECUTORCH_BUILD_PYBIND=ON \
-CMAKE_ARGS="-DEXECUTORCH_BUILD_COREML=ON -DEXECUTORCH_BUILD_MPS=ON -DEXECUTORCH_BUILD_XNNPACK=ON -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON" \
+CMAKE_ARGS="-DEXECUTORCH_BUILD_PYBIND=ON -DEXECUTORCH_BUILD_COREML=ON -DEXECUTORCH_BUILD_MPS=ON -DEXECUTORCH_BUILD_XNNPACK=ON -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON" \
 ${CONDA_RUN} --no-capture-output \
 .ci/scripts/setup-macos.sh "$@"
 

.ci/scripts/wheel/envvar_base.sh

@@ -8,13 +8,10 @@
 # should typically only contain shell variable assignments. Be sure to export
 # any variables so that subprocesses will see them.
 
-# Enable pybindings so that users can execute ExecuTorch programs from python.
-export EXECUTORCH_BUILD_PYBIND=1
-
 # Ensure that CMAKE_ARGS is defined before referencing it. Defaults to empty
 # if not defined.
 export CMAKE_ARGS="${CMAKE_ARGS:-}"
 
 # Link the XNNPACK backend into the pybindings runtime so that users can execute
 # ExecuTorch programs that delegate to it.
-CMAKE_ARGS="${CMAKE_ARGS} -DEXECUTORCH_BUILD_XNNPACK=ON"
+CMAKE_ARGS="${CMAKE_ARGS} -DEXECUTORCH_BUILD_PYBIND=ON -DEXECUTORCH_BUILD_XNNPACK=ON"

.github/workflows/pull.yml

@@ -365,8 +365,7 @@ jobs:
         # build module for executorch.extension.pybindings.portable_lib
         BUILD_TOOL="cmake"
         PYTHON_EXECUTABLE=python \
-        EXECUTORCH_BUILD_XNNPACK=ON \
-        EXECUTORCH_BUILD_PYBIND=ON \
+        CMAKE_ARGS="-DEXECUTORCH_BUILD_PYBIND=ON -DEXECUTORCH_BUILD_XNNPACK=ON" \
         bash .ci/scripts/setup-linux.sh --build-tool "${BUILD_TOOL}"
 
         # see if we can import the module successfully
@@ -504,7 +503,7 @@ jobs:
 
         # Setup MacOS dependencies as there is no Docker support on MacOS atm
         PYTHON_EXECUTABLE=python \
-        EXECUTORCH_BUILD_PYBIND=ON \
+        CMAKE_ARGS="-DEXECUTORCH_BUILD_PYBIND=ON" \
         EXECUTORCH_BUILD_ARM_BAREMETAL=ON \
         .ci/scripts/setup-linux.sh --build-tool "${BUILD_TOOL}"
 

.github/workflows/trunk.yml

@@ -261,7 +261,7 @@ jobs:
 
         # build module for executorch.extension.pybindings.portable_lib
         BUILD_TOOL=${{ matrix.build-tool }}
-        EXECUTORCH_BUILD_PYBIND=ON PYTHON_EXECUTABLE=python ${CONDA_RUN} bash .ci/scripts/setup-macos.sh --build-tool "${BUILD_TOOL}"
+        CMAKE_ARGS="-DEXECUTORCH_BUILD_PYBIND=ON" PYTHON_EXECUTABLE=python ${CONDA_RUN} bash .ci/scripts/setup-macos.sh --build-tool "${BUILD_TOOL}"
 
         # see if we can import the module successfully
         ${CONDA_RUN} python -c "from executorch.extension.pybindings import portable_lib; print('success!')"
@@ -536,8 +536,9 @@ jobs:
         git clone https://github.com/huggingface/optimum-executorch
         cd optimum-executorch
         # There is no release yet, for CI stability, always test from the same commit on main
-        git checkout 1907349524b5f2d61f9c04d2e985da826d4308ba
-        pip install .[tests]
+        git checkout 6a7e83f3eee2976fa809335bfb78a45b1ea1cb25
+        pip install .
+        pip install accelerate sentencepiece
         pip list
         echo "::endgroup::"
 

backends/arm/operator_support/convolution_support.py

@@ -55,7 +55,7 @@ def _is_node_supported_u55(self, node: fx.Node):
 
         C_in = shape_in[1]
         C_out = shape_out[1]
-        if (C_in == group) and (C_out % C_in) == 0:
+        if (C_in == group) and (C_out % C_in) == 0 and len(shape_in) <= 4:
             # Depthwise convolution
             for dim in shape_in[1:]:
                 if not 1 <= dim <= 65536:
@@ -74,13 +74,19 @@ def _is_node_supported_u55(self, node: fx.Node):
 
         kernel_w = kernel[2]
         kernel_h = kernel[3] if len(kernel) > 3 else 1
+        kernel_z = kernel[4] if len(kernel) > 4 else 1
         # Kernel condition misses constraint on sum of absolute weights
         if not 1 <= kernel_h <= 64 or not 1 <= kernel_w * kernel_h <= 4096:
             self.reporter.report_reject(
                 node,
                 f"Convolution needs to have kernel_y<=64, kernel_x*kernel_y<=4096, got kernel ({kernel_w}, {kernel_h})",
             )
             return False
+        if kernel_z != 1:
+            self.reporter.report_reject(
+                node, f"Convolution3d needs to have kernel_z==1, got {kernel_z}."
+            )
+            return False
 
         if not self._stride_condition(node):
             self.reporter.report_reject(
@@ -107,6 +113,14 @@ def _stride_condition(self, node: fx.Node) -> bool:
         if len(strides) == 1:
             strides = [strides[0]] * 2
 
+        if len(strides) > 2:
+            stride_z = strides[2]
+            if stride_z > 1:
+                self.reporter.report_reject(
+                    node, f"Convolution3d only supports stride_z<=1, got {stride_z}."
+                )
+                return False
+
         for stride, dilation in zip(strides, dilations):
             stride_condition = 1 <= stride <= 3
             dilation_condition = (not has_padding) and (dilation == 1)

backends/arm/test/models/test_llama.py

@@ -52,7 +52,7 @@ def prepare_model(self):
                 params_file, str
             ), "invalid input for --llama_inputs"
         else:
-            logging.warning(
+            logger.warning(
                 "Skipping Llama test because of lack of input. To run use --llama_inputs <.pt> <.json>"
             )
             return None, None, None
@@ -61,6 +61,8 @@ def prepare_model(self):
             params_file
         ), "Invalid file paths"
 
+        logger.info("Running test_llama.py")
+
         # TODO: Enable key value cache
         args = [
             "--disable_dynamic_shape",
@@ -112,9 +114,11 @@ def test_llama_tosa_MI(self):
                 )
                 .export()
                 .to_edge_transform_and_lower()
-                .check_count({"torch.ops.higher_order.executorch_call_delegate": 14})
+                .check_count({"torch.ops.higher_order.executorch_call_delegate": 26})
                 .to_executorch()
                 .run_method_and_compare_outputs(
-                    inputs=llama_inputs, atol=1.8, rtol=0.01  # TODO: decrease tolerance
+                    inputs=llama_inputs,
+                    atol=4.3,
+                    rtol=1.1,  # TODO: MLETORCH-825 decrease tolerance
                 )
             )

backends/arm/test/ops/test_conv2d.py

@@ -8,10 +8,10 @@
 
 import torch
 from executorch.backends.arm.test import common
-from executorch.backends.arm.test.tester.arm_tester import ArmTester
 from executorch.backends.arm.test.tester.test_pipeline import (
     EthosU55PipelineBI,
     EthosU85PipelineBI,
+    OpNotSupportedPipeline,
     TosaPipelineBI,
     TosaPipelineMI,
 )
@@ -34,9 +34,9 @@ def __init__(
         in_channels: Union[List, int, None] = None,
         out_channels: Union[List, int, None] = None,
         kernel_size: Union[List, Tuple, None] = None,
-        stride: Union[List, Tuple, None] = None,
-        padding: Union[List, Tuple, None] = None,
-        dilation: Union[List, Tuple, None] = None,
+        stride: Union[List, Tuple, int, None] = None,
+        padding: Union[List, Tuple, int, None] = None,
+        dilation: Union[List, Tuple, int, None] = None,
         groups: Union[List, int, None] = None,
         bias: Union[List, bool, None] = None,
         padding_mode: Union[List, str, None] = None,
@@ -446,17 +446,9 @@ def test_convolution_2d_u85_BI_on_fvp(test_module):
 def test_reject_convolution_2d_u55_BI(
     module: Conv2d,
 ):
-    (
-        ArmTester(
-            module,
-            example_inputs=module.get_inputs(),
-            compile_spec=common.get_u55_compile_spec(),
-        )
-        .quantize()
-        .export()
-        .check_count({"torch.ops.aten.conv2d.default": 1})
-        .check(["torch.ops.quantized_decomposed"])
-        .to_edge_transform_and_lower()
-        .check(["executorch_exir_dialects_edge__ops_aten_convolution_default"])
-        .check_count({"torch.ops.higher_order.executorch_call_delegate": 0})
-    )
+    OpNotSupportedPipeline(
+        module,
+        module.get_inputs(),
+        "TOSA-0.80+BI+u55",
+        {"executorch_exir_dialects_edge__ops_aten_convolution_default": 1},
+    ).run()