Update on "[Executorch][llama] Enable quantized sdpa"

Enable leveraging quantized sdpa op when quantized kv cache is used. Instead of adding yet another arg, at the moment I have chosen to leverage quantize_kv_cache option.

Differential Revision: [D71833064](https://our.internmc.facebook.com/intern/diff/D71833064/)

[ghstack-poisoned]

Author: kimishpatel

.ci/scripts/build_android_instrumentation.sh

@@ -12,10 +12,10 @@ if [[ -z "${PYTHON_EXECUTABLE:-}" ]]; then
 fi
 which "${PYTHON_EXECUTABLE}"
 
-mkdir -p "${BUILD_AAR_DIR}"/executorch_android/src/androidTest/resources
-cp extension/module/test/resources/add.pte "${BUILD_AAR_DIR}"/executorch_android/src/androidTest/resources
+mkdir -p extension/android/executorch_android/src/androidTest/resources
+cp extension/module/test/resources/add.pte extension/android/executorch_android/src/androidTest/resources
 
-pushd "${BUILD_AAR_DIR}"
+pushd extension/android
 ANDROID_HOME="${ANDROID_SDK:-/opt/android/sdk}" ./gradlew :executorch_android:testDebugUnitTest
 ANDROID_HOME="${ANDROID_SDK:-/opt/android/sdk}" ./gradlew :executorch_android:assembleAndroidTest
 popd

.github/workflows/_android.yml

@@ -37,7 +37,7 @@ jobs:
 
         mkdir -p ${ARTIFACTS_DIR_NAME}/library_test_dir
         bash .ci/scripts/build_android_instrumentation.sh
-        cp ${BUILD_AAR_DIR}/executorch_android/build/outputs/apk/androidTest/debug/executorch_android-debug-androidTest.apk "${ARTIFACTS_DIR_NAME}/library_test_dir"
+        cp extension/android/executorch_android/build/outputs/apk/androidTest/debug/executorch_android-debug-androidTest.apk "${ARTIFACTS_DIR_NAME}/library_test_dir"
 
         mkdir -p ${ARTIFACTS_DIR_NAME}/fp32-xnnpack-custom
         bash examples/models/llama/install_requirements.sh

CONTRIBUTING.md

@@ -1,7 +1,6 @@
 Thank you for your interest in contributing to ExecuTorch! We want to make
 it easy to contribute to this project.
 
- 
 
 ## Dev Install
 
@@ -91,7 +90,7 @@ executorch
 │   └── <a href="runtime/platform">platform</a> - Layer between architecture specific code and portable C++.
 ├── <a href="schema">schema</a> - ExecuTorch PTE file format flatbuffer schemas.
 ├── <a href="scripts">scripts</a> - Utility scripts for building libs, size management, dependency management, etc.
-├── <a href="shim">shim</a> - Compatibility layer between OSS and Internal builds.
+├── <a href="shim_et">shim_et</a> - Compatibility layer between OSS and Internal builds.
 ├── <a href="test">test</a> - Broad scoped end-to-end tests.
 ├── <a href="third-party">third-party</a> - Third-party dependencies.
 ├── <a href="tools">tools</a> - Tools for building ExecuTorch from source, for different built tools (CMake, Buck).
@@ -192,9 +191,6 @@ in the Github repo.
 
 ## Coding Style
 
-Goal: Encourage standards that make it easier to read, edit, maintain, and debug
-the ExecuTorch code.
-
 ### lintrunner
 
 We use [`lintrunner`](https://pypi.org/project/lintrunner/) to help make sure the
@@ -259,7 +255,7 @@ toolchains, and having access to relatively modern C++ features.
 
 #### C/C++ standard library usage
 
-**Restricted usage of the C++ standard library.**
+**Restricted usage of the C++ standard library**
 
 Rationale: ExecuTorch is intended to be portable to bare-metal systems that lack
 certain features, like dynamic memory, threading, and locking, required by parts
@@ -280,7 +276,7 @@ careful to also manually destroy objects initialized in this way.
 
 #### C++ language features
 
-**Exceptions: Do not use.**
+**Exceptions: Do not use**
 - Rationale: Exceptions are not widely supported on some classes of
   microcontrollers and DSPs, and they can significantly increase binary size.
 
@@ -289,12 +285,12 @@ must work with threading**
 - Rationale: The core runtime must work on systems that do not have threading
   support.
 
-**RTTI, dynamic_cast, and `<typeid>`: Do not use.**
+**RTTI, dynamic_cast, and `<typeid>`: Do not use**
 - Rationale: RTTI adds extra data to every virtual class. ExecuTorch doesn't
   have a strong need for `dynamic_cast` and friends, so it's better to reduce
   the binary size.
 
-**Templates and template metaprogramming: Be careful and avoid if possible.**
+**Templates and template metaprogramming: Be careful and avoid if possible**
 - Rationale: Most templating results in code generation, and is one of the most
   common sources of binary bloat. Some use of templates is fine (e.g. an
   `ArrayRef<T>`, or code that handles multiple `ScalarType` types), but for the
@@ -359,7 +355,7 @@ docs](https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/
 for basics.
 
 1. Push your branch to your fork of `pytorch/executorch`. Most people do not
-  have permission to push a branch directoy to the upstream repo.
+  have permission to push a branch directory to the upstream repo.
 1. Create your PR
    - Use the `main` branch as the base.
    - Give the PR a clear and descriptive title. It will become the title of the

README.md

@@ -49,9 +49,9 @@ Key value propositions of ExecuTorch are:
 ## Getting Started
 To get started you can:
 
-- Visit the [Step by Step Tutorial](https://pytorch.org/executorch/main/index.html) on getting things running locally and deploy a model to a device
+- Visit the [Step by Step Tutorial](https://pytorch.org/executorch/main/index.html) to get things running locally and deploy a model to a device
 - Use this [Colab Notebook](https://pytorch.org/executorch/stable/getting-started-setup.html#quick-setup-colab-jupyter-notebook-prototype) to start playing around right away
-- Jump straight into LLMs use cases by following specific instructions for [Llama](./examples/models/llama/README.md) and [Llava](./examples/models/llava/README.md)
+- Jump straight into LLM use cases by following specific instructions for [Llama](./examples/models/llama/README.md) and [Llava](./examples/models/llava/README.md)
 
 ## Feedback and Engagement
 

backends/qualcomm/_passes/decompose_einsum.py

@@ -8,6 +8,8 @@
 from executorch.exir.pass_base import ExportPass, PassResult
 from torch.fx.experimental.proxy_tensor import make_fx
 
+from .utils import copy_nn_module_stack
+
 
 class DecomposeEinsum(ExportPass):
     """
@@ -36,6 +38,7 @@ def call(self, graph_module: torch.fx.GraphModule) -> PassResult:
                         remap[f"arg1_{i+1}"] = arg
 
                     for decomposed_node in decomposed_module.graph.nodes:
+                        copy_nn_module_stack(node, decomposed_node)
                         # This is the arg[0] equation string, which is not required anymore after decomposition
                         if "arg0" in decomposed_node.name:
                             continue

backends/qualcomm/_passes/decompose_linalg_vector_norm.py

@@ -8,6 +8,8 @@
 from executorch.exir import to_edge
 from executorch.exir.pass_base import ExportPass, PassResult
 
+from .utils import copy_nn_module_stack
+
 
 class LinalgVectorNorm(torch.nn.Module):
     def __init__(self, exp, dim, keepdim):
@@ -62,6 +64,7 @@ def call(self, graph_module: torch.fx.GraphModule) -> PassResult:
                     remap = {"x": node.args[0]}
 
                     for decomposed_node in decomposed_module.graph.nodes:
+                        copy_nn_module_stack(node, decomposed_node)
                         # no need to copy existent 'output'
                         if decomposed_node.op == "output":
                             for user in node.users.copy():

backends/qualcomm/_passes/utils.py

@@ -121,6 +121,14 @@ def get_passes_dependency_for_capture_program():
     }
 
 
+def copy_nn_module_stack(src, target):
+    """
+    Copy meta["nn_module_stack"] from src node to target node if existing.
+    """
+    if value := src.meta.get("nn_module_stack"):
+        target.meta["nn_module_stack"] = value
+
+
 def is_float_tensor(node: torch.fx.Node) -> bool:
     if "val" not in node.meta or not isinstance(node.meta["val"], FakeTensor):
         return False