Update

[ghstack-poisoned]

Author: swolchok

.ci/scripts/test_qnn_static_llama.sh

@@ -34,11 +34,11 @@ $PYTHON_EXECUTABLE -m extension.llm.tokenizer.tokenizer -t tokenizer.model -o to
 
 set +e
 # Compile only as weight sharing is not applicable on x86
-$PYTHON_EXECUTABLE backends/qualcomm/tests/test_qnn_delegate.py -k TestExampleScript.test_stories_single_llama --model SM8650 --build_folder build-android/ --executorch_root . --artifact_dir . --compile_only
+$PYTHON_EXECUTABLE backends/qualcomm/tests/test_qnn_delegate.py -k TestExampleLLMScript.test_llama_stories_110m --model SM8650 --build_folder build-android/ --executorch_root . --artifact_dir . --llama_artifacts . --compile_only
 exit_code1=$?
 
 # Checks accuracy with weight sharing disabled since x86 does not support weight sharing.
-$PYTHON_EXECUTABLE backends/qualcomm/tests/test_qnn_delegate.py -k TestExampleScript.test_stories_single_llama --model SM8650 --build_folder build-x86/ --executorch_root . --artifact_dir . --enable_x86_64
+$PYTHON_EXECUTABLE backends/qualcomm/tests/test_qnn_delegate.py -k TestExampleLLMScript.test_llama_stories_110m --model SM8650 --build_folder build-x86/ --executorch_root . --artifact_dir . --llama_artifacts . --enable_x86_64
 exit_code2=$?
 
 # Check the exit codes and print messages

.ci/scripts/unittest-macos.sh

@@ -41,9 +41,9 @@ ${CONDA_RUN} --no-capture-output \
 PYTHON_EXECUTABLE=python ./examples/models/llama3_2_vision/install_requirements.sh
 
 if [[ "$BUILD_TOOL" == "cmake" ]]; then
-    .ci/scripts/unittest-macos-cmake.sh "$BUILD_MODE"
+    .ci/scripts/unittest-macos-cmake.sh
 elif [[ "$BUILD_TOOL" == "buck2" ]]; then
-    .ci/scripts/unittest-buck2.sh "$BUILD_MODE"
+    .ci/scripts/unittest-buck2.sh
 else
     echo "Unknown build tool $BUILD_TOOL"
     exit 1

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -21,7 +21,7 @@ body:
       A clear and concise description of what the bug is.
 
       ```python
-      # Sample code to reproduce the problem
+      # Sample code to reproduce the problem. If applicable, also include your model export command.
       ```
 
       ```

.github/workflows/trunk.yml

@@ -159,7 +159,7 @@ jobs:
         sudo sysctl fs.inotify.max_user_watches=1048576 # 1024 * 1024
 
         # Test ethos-u delegate examples with run.sh
-        backends/arm/test/test_arm_baremetal.sh test_run_ethosu_fvp
+        backends/arm/test/test_arm_baremetal.sh test_full_ethosu_fvp
 
 
   test-arm-reference-delegation:

CODEOWNERS

@@ -0,0 +1,85 @@
+# IMPORTANT:
+# This file is ONLY used to subscribe for notifications for PRs
+# related to a specific file path. Approvals from people in this
+# file are not required for merges.
+
+/backends/apple @shoumikhin @cccclai
+/backends/apple/mps @cccclai @DenisVieriu97
+/backends/arm @digantdesai
+/backends/cadence @tarun292
+/backends/example @iseeyuan @JacobSzwejbka @larryliu0820
+/backends/mediatek @cccclai @neuropilot-captain
+/backends/qualcomm @cccclai @chunit-quic @haowhsu-quic @shewu-quic @winskuo-quic
+/backends/test @cccclai
+/backends/transforms @kimishpatel
+/backends/vulkan @SS-JIA
+/backends/xnnpack @digantdesai @mcr229
+
+/build @GregoryComer @dbort @kirklandsign
+
+/codegen @larryliu0820 @lucylq
+
+/devtools @tarun292 @Gasoonjia
+
+/docs @mergennachin
+
+/examples/apple @shoumikhin
+/examples/apple/coreml @cccclai @metascroy @cymbalrush @YifanShenSZ
+/examples/arm @digantdesai
+/examples/cadence @tarun292
+/examples/demo-apps @shoumikhin @kirklandsign
+/examples/devtools @tarun292
+/examples/llm_manual @larryliu0820
+/examples/llm_pte_finetuning @JacobSzwejbka
+/examples/mediatek @cccclai
+/examples/models @lucylq
+/examples/portable @larryliu0820 @manuelcandales
+/examples/qualcomm @cccclai
+/examples/selective_build @lucylq @larryliu0820 @JacobSzwejbka
+/examples/xnnpack @digantdesai @mcr229
+
+/exir/backend @cccclai @kimishpatel @JacobSzwejbka @tarun292
+/exir @JacobSzwejbka @tarun292 @larryliu0820
+
+
+/extension/android @kirklandsign
+/extension/android_test @kirklandsign
+/extension/apple @shoumikhin
+/extension/aten_util @JacobSzwejbka
+/extension/benchmark @tarun292
+/extension/data_loader @JacobSzwejbka @lucylq @dbort
+/extension/evalue_util @GregoryComer @dbort
+/extension/export_util @kimishpatel
+/extension/flat_tensor @lucylq
+/extension/gguf_util @larryliu0820
+/extension/kernel_util @kimishpatel @manuelcandales
+/extension/llm @jackzhxng @iseeyuan @larryliu0820
+/extension/memory_allocator @JacobSzwejbka @dbort
+/extension/module @shoumikhin
+/extension/parallel @kimishpatel
+/extension/pybindings @JacobSzwejbka @larryliu0820
+/extension/pytree @JacobSzwejbka
+/extension/runner_util @dbort
+/extension/tensor @shoumikhin
+/extension/testing_util @dbort
+/extension/threadpool @kimishpatel
+/extension/training @JacobSzwejbka
+
+/kernels @manuelcandales
+
+/profiler @tarun292 @Gasoonjia
+
+/runtime @dbort @JacobSzwejbka @lucylq
+/runtime/backend @cccclai
+
+/schema @dbort @JacobSzwejbka @lucylq
+
+/scripts @GregoryComer
+
+/shim @larryliu0820 @GregoryComer
+
+/third-party @GregoryComer
+
+/test @larryliu0820 @kirklandsign
+
+/util @tarun292

backends/arm/README.md

@@ -55,10 +55,10 @@ To run the unit test suite with Corstone3x0 FVP simulator support use
 backends/arm/test/test_arm_baremetal.sh test_pytest_ethosu_fvp
 ```
 
-You can test to run some models with the run.sh flow
+You can test to run some models with the full fvp test flow
 
 ```
-backends/arm/test/test_arm_baremetal.sh test_run_ethosu_fvp
+backends/arm/test/test_arm_baremetal.sh test_full_ethosu_fvp
 ```
 
 ## Unit tests

backends/arm/_passes/arm_pass_manager.py

@@ -27,6 +27,7 @@
 from executorch.backends.arm._passes.convert_squeezes_to_view import (  # type: ignore[import-not-found]
     ConvertSqueezesToViewPass,
 )
+from executorch.backends.arm._passes.convert_to_clamp import ConvertToClampPass
 from executorch.backends.arm._passes.decompose_batchnorm_pass import (
     DecomposeBatchNormPass,
 )
@@ -104,6 +105,7 @@ def _tosa_080_BI_pipeline(self, exported_program: ExportedProgram) -> GraphModul
         self.add_pass(DecomposeLinearPass())
         self.add_pass(ConvertMeanDimToAveragePoolPass())
         self.add_pass(ConvertFullLikeToFullPass())
+        self.add_pass(ConvertToClampPass())
 
         self.add_pass(ReplaceScalarWithTensorArgPass())
         self.add_pass(AnnotateDecomposedMatmulPass())
@@ -144,6 +146,8 @@ def _tosa_080_MI_pipeline(self, exported_program: ExportedProgram) -> GraphModul
         self.add_pass(DecomposeDivPass())
         self.add_pass(DecomposeSoftmaxesPass())
         self.add_pass(ConvertFullLikeToFullPass())
+        self.add_pass(ConvertToClampPass())
+
         self.add_pass(AnnotateDecomposedMatmulPass())
         self.add_pass(QuantizeOperatorArguments())
         self.add_pass(FoldAndAnnotateQParamsPass())  # type: ignore[call-arg]

backends/arm/_passes/convert_to_clamp.py

@@ -0,0 +1,36 @@
+# Copyright 2025 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.
+
+from typing import Tuple
+
+from executorch.exir.dialects._ops import ops as exir_ops
+from executorch.exir.pass_base import ExportPass
+
+edge_operators = {
+    exir_ops.edge.aten.hardtanh.default,
+    exir_ops.edge.aten.relu.default,
+}
+
+
+def get_clamp_params(op, args) -> Tuple[float | None, float | None]:
+    if op == exir_ops.edge.aten.hardtanh.default:
+        return args[1], args[2]
+    elif op == exir_ops.edge.aten.relu.default:
+        return 0.0, None
+    else:
+        raise ValueError(f"Getting clamp parameters for op {op} is not implemented.")
+
+
+class ConvertToClampPass(ExportPass):
+    def call_operator(self, op, args, kwargs, meta):
+        if op not in edge_operators:
+            return super().call_operator(op, args, kwargs, meta)
+
+        return super().call_operator(
+            exir_ops.edge.aten.clamp.default,
+            (args[0], *get_clamp_params(op, args)),
+            {},
+            meta,
+        )

backends/arm/_passes/decompose_select.py

@@ -35,8 +35,9 @@ def call(self, graph_module: torch.fx.GraphModule):
             input_node, dim, index = node.args
 
             rank = len(input_node.meta["val"].size())
+            shape = input_node.meta["val"].shape
             dim = dim % rank if dim < 0 else dim
-            index = index % rank if index < 0 else index
+            index = index % shape[dim] if index < 0 else index
 
             with graph_module.graph.inserting_before(node):
                 slice_node = create_node(

backends/arm/_passes/insert_table_ops.py

@@ -39,6 +39,7 @@ class InsertTableOpsPass(ExportPass):
 
     table_ops: Dict[EdgeOpOverload, Callable[[torch.Tensor], torch.Tensor]] = {
         exir_ops.edge.aten.exp.default: torch.exp,
+        exir_ops.edge.aten.floor.default: torch.floor,
         exir_ops.edge.aten.log.default: torch.log,
         exir_ops.edge.aten.reciprocal.default: torch.reciprocal,
         exir_ops.edge.aten.rsqrt.default: torch.rsqrt,

backends/arm/operator_support/tosa_supported_operators.py

@@ -91,6 +91,7 @@ class BaseTOSASupportList(OperatorSupportBase):
 
     def is_node_supported(self, submodules, node: fx.Node) -> bool:
         supported = node.op == "call_function" and node.target in [
+            exir_ops.edge.aten.abs.default,
             exir_ops.edge.aten.add.Tensor,
             exir_ops.edge.aten.expand_copy.default,
             exir_ops.edge.aten.cat.default,
@@ -106,6 +107,7 @@ def is_node_supported(self, submodules, node: fx.Node) -> bool:
             exir_ops.edge.aten.log.default,
             exir_ops.edge.aten.linear.default,
             exir_ops.edge.aten.split_with_sizes_copy.default,
+            exir_ops.edge.aten.floor.default,
             exir_ops.edge.aten.full.default,
             exir_ops.edge.aten.full_like.default,
             exir_ops.edge.aten.ge.Tensor,

backends/arm/operators/__init__.py

@@ -7,6 +7,7 @@
 
 from . import (  # noqa
     node_visitor,
+    op_abs,
     op_add,
     op_avg_pool2d,
     op_bmm,
@@ -20,7 +21,6 @@
     op_ge,
     op_get_item,
     op_gt,
-    op_hardtanh,
     op_le,
     op_log,
     op_lt,
@@ -30,7 +30,6 @@
     op_mul,
     op_permute,
     op_reciprocal,
-    op_relu,
     op_repeat,
     op_rescale,
     op_rshift,
@@ -47,4 +46,5 @@
     op_upsample_nearest2d,
     op_view,
     ops_binary,
+    ops_unary,
 )