Merge branch 'release/1.0' into cherry-pick-14786-by-pytorch_bot_bot_

Author: GregoryComer

backends/arm/_passes/annotate_decomposed_matmul.py

@@ -68,7 +68,10 @@ def call(self, graph_module: GraphModule) -> PassResult:
                 node for node in partition.nodes if node.target in matmul_targets
             ][0]
 
-            if quantized_input:
+            if quantized_input and not all(
+                input_node.target in DQ_OPS
+                for input_node in matmul_node.all_input_nodes
+            ):
                 matmul_args = matmul_node.all_input_nodes
                 for node in matmul_args:
                     # Find the dq-node connected to this mm/bmm arg
@@ -94,7 +97,9 @@ def call(self, graph_module: GraphModule) -> PassResult:
 
             partition_output = list(partition.output_nodes[0].users)[0]
             quantized_output = partition_output.target in Q_OPS
-            if quantized_output:
+            if quantized_output and not all(
+                user.target in Q_OPS for user in matmul_node.users
+            ):
                 with graph_module.graph.inserting_after(matmul_node):
                     # Create q-node after matmul
                     q_node = create_node(

backends/arm/test/ops/test_matmul.py

@@ -22,6 +22,7 @@
 
 class MatMul(torch.nn.Module):
     test_data_generators = {
+        "rand_rand_2d": lambda: (torch.rand(5, 5), torch.rand(5, 2)),
         "rand_rand_3d": lambda: (torch.rand(2, 3, 5), torch.rand(2, 5, 2)),
         "rand_rand_4d": lambda: (torch.rand(1, 2, 3, 5), torch.rand(1, 2, 5, 2)),
     }
@@ -32,6 +33,7 @@ def forward(self, x: torch.Tensor, y: torch.Tensor):
 
 class MatMulSingleInput(torch.nn.Module):
     test_data_generators = {
+        "rand_2d": lambda: (torch.rand(5, 5),),
         "rand_3d": lambda: (torch.rand(2, 5, 5),),
         "rand_4d": lambda: (torch.rand(1, 2, 5, 5),),
     }
@@ -42,6 +44,11 @@ def forward(self, x: torch.Tensor):
 
 class MatMulCombo(torch.nn.Module):
     test_data_generators = {
+        "rand_rand_rand_2d": lambda: (
+            torch.rand(5, 5),
+            torch.rand(5, 2),
+            torch.rand(2, 5),
+        ),
         "rand_rand_rand_3d": lambda: (
             torch.rand(2, 5, 5),
             torch.rand(2, 5, 2),

backends/arm/test/tester/arm_tester.py

@@ -458,6 +458,10 @@ def run_method_and_compare_outputs(
         for run_iteration in range(num_runs):
             reference_input = inputs if inputs else next(self.generate_random_inputs())
 
+            # Avoid issues with inplace operators
+            test_input = copy.deepcopy(reference_input)
+            original_input = copy.deepcopy(reference_input)
+
             input_shapes = [
                 generated_input.shape if hasattr(generated_input, "shape") else (1,)
                 for generated_input in reference_input
@@ -472,16 +476,16 @@ def run_method_and_compare_outputs(
                 # Run exported module directly
                 test_outputs, _ = pytree.tree_flatten(
                     self._calculate_reference_output(
-                        exported_program.module(), reference_input
+                        exported_program.module(), test_input
                     )
                 )
             else:
                 # Run lowered model with target
                 test_outputs, _ = pytree.tree_flatten(
-                    test_stage.run_artifact(reference_input)
+                    test_stage.run_artifact(test_input)
                 )
 
-            logger.info(f"\n      Input: {reference_input}")
+            logger.info(f"\n      Input: {original_input}")
             logger.info(f"\n Ref output: {reference_outputs}")
             logger.info(f"\nTest output: {test_outputs}")
 

docs/source/advanced-topics-section.md

@@ -0,0 +1,112 @@
+(advanced-topics-section)=
+
+# Advanced
+
+Deep dive into ExecuTorch's advanced features for optimization, customization, and integration.
+
+This section covers advanced concepts for developers who need to customize ExecuTorch for specific use cases, optimize performance, or integrate with custom hardware backends.
+
+## Quantization & Optimization
+
+Techniques for model compression and performance optimization.
+
+**→ {doc}`quantization-optimization` — Quantization strategies and performance optimization**
+
+Key topics:
+
+- Quantization strategies and techniques
+- Performance profiling and optimization
+
+## Model Export
+
+Learn the core ExecuTorch workflow, exporting PyTorch models to the `.pte` format for edge deployment.
+
+**→ {doc}`using-executorch-export`** - Model Export & Lowering
+
+Key topics:
+
+- Export and Lowering Workflow
+- Hardware Backend Selection & Optimization
+- Dynamic Shapes & Advanced Model Features
+
+
+## Kernel Library
+
+Deep dive into ExecuTorch's kernel implementation and customization.
+
+**→ {doc}`kernel-library-advanced` — Kernel library deep dive and customization**
+
+Key topics:
+
+- Kernel library architecture
+- Custom kernel implementation
+- Selective build and optimization
+
+## Backend & Delegates
+
+**→ {doc}`backend-delegate-advanced` — Backend delegate integration**
+
+Key topics:
+
+- Learn how to integrate Backend Delegate into ExecuTorch and more
+- XNNPACK Delegate Internals
+- Debugging Delegation
+
+
+## Runtime & Integration
+
+Advanced runtime features and backend integration.
+
+**→ {doc}`runtime-integration-advanced` — Runtime customization and backend integration**
+
+Key topics:
+
+- Backend delegate implementation
+- Platform abstraction layer
+- Custom runtime integration
+
+## Compiler & IR
+
+Advanced compiler features and intermediate representation details.
+
+**→ {doc}`compiler-ir-advanced` — Compiler passes and IR specification**
+
+Key topics:
+
+- Custom compiler passes
+- Memory planning strategies
+- Backend dialect and EXIR
+- Ops set definition
+
+
+## File Formats
+
+ExecuTorch file format specifications and internals.
+
+**→ {doc}`file-formats-advanced` — PTE and PTD file format specifications**
+
+Key topics:
+
+- PTE file format internals
+- PTD file format specification
+- Custom file format handling
+
+## Next Steps
+
+After exploring advanced topics:
+
+- **{doc}`tools-sdk-section`** - Developer tools for debugging and profiling
+- **{doc}`api-section`** - Complete API reference documentation
+
+```{toctree}
+:hidden:
+:maxdepth: 2
+:caption: Advanced Topics
+
+quantization-optimization
+using-executorch-export
+kernel-library-advanced
+backend-delegate-advanced
+runtime-integration-advanced
+compiler-ir-advanced
+file-formats-advanced

docs/source/android-arm-vgf.md

@@ -0,0 +1 @@
+```{include} backends-arm-vgf.md

docs/source/android-backends.md

@@ -0,0 +1,28 @@
+(android-backends)=
+# Backends
+
+Available hardware acceleration backends for Android deployment.
+
+## CPU Acceleration
+
+- {doc}`android-xnnpack` — XNNPACK CPU acceleration
+
+## GPU Acceleration
+
+- {doc}`android-vulkan` — Vulkan GPU acceleration
+
+## NPU/Accelerator Backends
+
+- {doc}`android-qualcomm` — Qualcomm AI Engine (NPU)
+- {doc}`android-mediatek` — MediaTek NPU acceleration
+- {doc}`android-arm-vgf` — ARM VGF Backend
+- {doc}`android-samsung-exynos` — Samsung Exynos NPU
+
+```{toctree}
+:hidden:
+android-xnnpack
+android-vulkan
+android-qualcomm
+android-mediatek
+android-arm-vgf
+android-samsung-exynos

docs/source/android-examples.md

@@ -0,0 +1,9 @@
+# Examples & Demos
+
+- [Working with LLMs - Android Examples](https://github.com/meta-pytorch/executorch-examples/tree/main/llm/android)
+- [Demo Apps](https://github.com/meta-pytorch/executorch-examples/tree/main/dl3/android/DeepLabV3Demo#executorch-android-demo-app)
+- {doc}`tutorial-arm-vgf` — Export a simple PyTorch model for the ExecuTorch VGF backend
+
+```{toctree}
+:hidden:
+tutorial-arm-vgf

docs/source/android-mediatek.md

@@ -0,0 +1 @@
+```{include} backends-mediatek.md

docs/source/android-qualcomm.md

@@ -0,0 +1 @@
+```{include} backends-qualcomm.md

docs/source/android-samsung-exynos.md

@@ -0,0 +1 @@
+```{include} backends-samsung-exynos.md