Merge branch 'main' into android-0.6-docs

Author: kirklandsign

.ci/scripts/utils.sh

@@ -32,7 +32,7 @@ install_executorch() {
   which pip
   # Install executorch, this assumes that Executorch is checked out in the
   # current directory.
-  ./install_executorch.sh --pybind xnnpack "$@"
+  ./install_executorch.sh "$@"
   # Just print out the list of packages for debugging
   pip list
 }

.github/workflows/apple.yml

@@ -37,7 +37,7 @@ jobs:
         id: set_version
         shell: bash
         run: |
-          VERSION="0.5.0.$(TZ='PST8PDT' date +%Y%m%d)"
+          VERSION="0.7.0.$(TZ='PST8PDT' date +%Y%m%d)"
           echo "version=$VERSION" >> "$GITHUB_OUTPUT"
 
   build-demo-ios:

backends/xnnpack/operators/node_visitor.py

@@ -210,7 +210,7 @@ def get_serialized_dtype(
         self,
         quant_params: Optional[QuantParams],
         node: torch.fx.Node,
-        fp32_static_weight: bool = False,
+        force_fp32: bool = False,
     ) -> XNNDatatype:
         # Default initialization
         dtype = XNNDatatype.xnn_datatype_fp32
@@ -267,7 +267,7 @@ def get_per_channel_dtype(
             if node_dtype is not None and node_dtype == torch.float16:
                 dtype = (
                     XNNDatatype.xnn_datatype_fp32
-                    if fp32_static_weight
+                    if force_fp32
                     else XNNDatatype.xnn_datatype_fp16
                 )
 
@@ -348,7 +348,7 @@ def define_tensor(  # noqa: C901
         convert_to_nhwc: bool = False,
         swap_in_out_for_weights: bool = False,
         quant_params: Optional[QuantParams] = None,
-        fp32_static_weights: bool = False,
+        force_fp32: bool = False,
         groups: int = 1,
     ) -> None:
         """
@@ -368,7 +368,7 @@ def define_tensor(  # noqa: C901
                         constant data. If used along with convert_to_nhwc, this
                         swap will happen before converting to nhwc.
             quant_params: Quantization meta data for this tensor, None if it is not quantized
-            fp32_static_weights: XNN_FLAG_FP32_STATIC_WEIGHTS for fp16 conv
+            force_fp32: forces tensor to be serialize as fp32, used for bias of dynamically quantized ops
             groups: number of groups for swap_in_out_for_weights
         """
 
@@ -405,7 +405,7 @@ def define_tensor(  # noqa: C901
             convert_to_nhwc,
             swap_in_out_for_weights,
             quant_params,
-            fp32_static_weights,
+            force_fp32,
             groups,
         )
 
@@ -417,9 +417,7 @@ def define_tensor(  # noqa: C901
             check_or_raise(len(dims) == 4, "Converting to nhwc requires 4d tensor")
             dims = [dims[i] for i in PERM_NCHW_TO_NHWC]
 
-        dtype = self.get_serialized_dtype(
-            quant_params, tensor, fp32_static_weight=fp32_static_weights
-        )
+        dtype = self.get_serialized_dtype(quant_params, tensor, force_fp32=force_fp32)
 
         tvalue = XNNTensorValue(
             datatype=dtype,
@@ -504,7 +502,7 @@ def get_serialized_buffer_index(
         convert_to_nhwc: bool,
         swap_in_out_for_weights: bool,
         quant_params: Optional[QuantParams],
-        fp32_static_weights: bool = False,
+        force_fp32: bool = False,
         groups: int = 1,
     ) -> int:
         """
@@ -525,7 +523,7 @@ def get_serialized_buffer_index(
                         constant data. If used along with convert_to_nhwc, this
                         swap will happen before converting to nhwc.
             quant_params: Quantization meta data for this tensor, None if it is not quantize
-            fp32_static_weights: bool to indicate whether tensor is fp32 static weights
+            force_fp32: bool to indicate whether tensor is fp32 static weights
             groups: groups for swap_in_out_for_weights
 
         Returns:
@@ -554,7 +552,7 @@ def get_serialized_buffer_index(
         # Quantize buffer if static data is indeed quantized
         if quant_params is not None and not quant_params.is_dynamic:
             const_val = quant_params.quantize_tensor(const_val).contiguous()
-        elif const_val.dtype != torch.float16 or fp32_static_weights:
+        elif const_val.dtype != torch.float16 or force_fp32:
             # ensure that the const is fp32
             const_val = const_val.to(dtype=torch.float32).contiguous()
 

backends/xnnpack/operators/op_conv2d.py

@@ -82,7 +82,6 @@ def define_node(
         weight_quant_params = QuantParams.from_weights(
             kernel_node, self._exported_program
         )
-        fp32_static_weights = kernel_node.meta["val"].dtype == torch.float16
 
         if weight_quant_params is not None and weight_quant_params.per_channel:
             if is_transpose:
@@ -102,8 +101,8 @@ def define_node(
             convert_to_nhwc=True,
             swap_in_out_for_weights=is_depthwise_conv or is_transpose,
             quant_params=weight_quant_params,
-            fp32_static_weights=fp32_static_weights,
             groups=groups if is_transpose else 1,
+            force_fp32=True,
         )
         kwargs["filter_id"] = vals_to_ids[get_input_node(node, 1)]
 
@@ -127,13 +126,14 @@ def define_node(
             bias_quant_params = QuantParams.from_bias(
                 bias_node, weight_quant_params, input_quant_params
             )
+
             self.define_tensor(
                 get_input_node(node, 2),
                 xnn_graph,
                 vals_to_ids,
                 convert_to_nhwc=False,
                 quant_params=bias_quant_params,
-                fp32_static_weights=fp32_static_weights,
+                force_fp32=True,
             )
             kwargs["bias_id"] = vals_to_ids[get_input_node(node, 2)]
 

backends/xnnpack/operators/op_linear.py

@@ -59,7 +59,6 @@ def define_node(
             xnn_graph,
             vals_to_ids,
             quant_params=weight_quant_params,
-            fp32_static_weights=True,
         )
         filter_id = vals_to_ids[weight_node]
 
@@ -69,12 +68,18 @@ def define_node(
             bias_quant_params = QuantParams.from_bias(
                 bias_node, weight_quant_params, input_quant_params
             )
+            # For dynamic quantization, there are no kernels with fp16 bias
+            # So we need to force the fp16 bias to fp32
+            force_fp32 = False
+            if input_quant_params is not None and input_quant_params.is_dynamic:
+                force_fp32 = True
+
             self.define_tensor(
                 get_input_node(node, 2),
                 xnn_graph,
                 vals_to_ids,
                 quant_params=bias_quant_params,
-                fp32_static_weights=True,
+                force_fp32=force_fp32,
             )
             bias_id = vals_to_ids[bias_node]
         else:

backends/xnnpack/test/ops/test_linear.py

@@ -605,9 +605,7 @@ def _test_qd8_linear_per_tensor_unsupported(self, dtype: torch.dtype = torch.flo
 
                     if legacy_partitioner:
                         tester.to_edge()
-                        tester.partition(
-                            Partition(DynamicallyQuantizedPartitioner)
-                        ).dump_artifact()
+                        tester.partition(Partition(DynamicallyQuantizedPartitioner))
                         # should have [add]mm node
                         if uses_bias:
                             tester.check(
@@ -624,7 +622,7 @@ def _test_qd8_linear_per_tensor_unsupported(self, dtype: torch.dtype = torch.flo
                     else:
                         tester.to_edge_transform_and_lower(
                             ToEdgeTransformAndLower([DynamicallyQuantizedPartitioner])
-                        ).dump_artifact()
+                        )
                         # should not have a delegate node
                         tester.check_not(
                             [
@@ -717,7 +715,7 @@ def test_fp16_linear(self):
                     num_batch_dims=num_batch_dims,
                     uses_bias=use_bias,
                     dtype=torch.float16,
-                    atol=5e-2,  # TODO(T212995726): Investigate right atol for rand[n] inputs
+                    atol=5e-3,  # TODO(T212995726): Investigate right atol for rand[n] inputs
                 )
 
     def test_fp32_linear(self):

docs/README.md

@@ -39,33 +39,32 @@ To build the documentation locally:
 
 1. Clone the ExecuTorch repo to your machine.
 
-1. If you don't have it already, start a conda environment:
+   ```bash
+   git clone -b viable/strict https://github.com/pytorch/executorch.git && cd executorch
+   ```
 
-   ```{note}
-   The below command generates a completely new environment and resets
-   any existing dependencies. If you have an environment already, skip
-   the `conda create` command.
+1. If you don't have it already, start either a Python virtual envitonment:
+
+   ```bash
+   python3 -m venv .venv && source .venv/bin/activate && pip install --upgrade pip
    ```
 
+   Or a Conda environment:
+
    ```bash
-   conda create -yn executorch python=3.10.0
-   conda activate executorch
+   conda create -yn executorch python=3.10.0 && conda activate executorch
    ```
 
 1. Install dependencies:
 
    ```bash
    pip3 install -r ./.ci/docker/requirements-ci.txt
    ```
-1. Update submodules
 
-   ```bash
-   git submodule sync && git submodule update --init
-   ```
 1. Run:
 
    ```bash
-   bash install_executorch.sh
+   ./install_executorch.sh
    ```
 
 1. Go to the `docs/` directory.

docs/source/getting-started.md

@@ -137,7 +137,7 @@ For a full example of running a model on Android, see the [DeepLabV3AndroidDemo]
 #### Installation
 ExecuTorch supports both iOS and MacOS via C++, as well as hardware backends for CoreML, MPS, and CPU. The iOS runtime library is provided as a collection of .xcframework targets and are made available as a Swift PM package.
 
-To get started with Xcode, go to File > Add Package Dependencies. Paste the URL of the ExecuTorch repo into the search bar and select it. Make sure to change the branch name to the desired ExecuTorch version in format “swiftpm-”, (e.g. “swiftpm-0.5.0”).  The ExecuTorch dependency can also be added to the package file manually. See [Using ExecuTorch on iOS](using-executorch-ios.md) for more information.
+To get started with Xcode, go to File > Add Package Dependencies. Paste the URL of the ExecuTorch repo into the search bar and select it. Make sure to change the branch name to the desired ExecuTorch version in format “swiftpm-”, (e.g. “swiftpm-0.6.0”).  The ExecuTorch dependency can also be added to the package file manually. See [Using ExecuTorch on iOS](using-executorch-ios.md) for more information.
 
 #### Runtime APIs
 Models can be loaded and run from Objective-C using the C++ APIs.
@@ -151,7 +151,7 @@ ExecuTorch provides C++ APIs, which can be used to target embedded or mobile dev
 CMake is the preferred build system for the ExecuTorch C++ runtime. To use with CMake, clone the ExecuTorch repository as a subdirectory of your project, and use CMake's `add_subdirectory("executorch")` to include the dependency. The `executorch` target, as well as kernel and backend targets will be made available to link against. The runtime can also be built standalone to support diverse toolchains. See [Using ExecuTorch with C++](using-executorch-cpp.md) for a detailed description of build integration, targets, and cross compilation.
 
 ```
-git clone -b release/0.5 https://github.com/pytorch/executorch.git
+git clone -b viable/strict https://github.com/pytorch/executorch.git
 ```
 ```python
 # CMakeLists.txt

docs/source/llm/getting-started.md

@@ -43,15 +43,17 @@ Instructions on installing miniconda can be [found here](https://docs.anaconda.c
 mkdir et-nanogpt
 cd et-nanogpt
 
-# Clone the ExecuTorch repository and submodules.
+# Clone the ExecuTorch repository.
 mkdir third-party
-git clone -b release/0.4 https://github.com/pytorch/executorch.git third-party/executorch
-cd third-party/executorch
-git submodule update --init
+git clone -b viable/strict https://github.com/pytorch/executorch.git third-party/executorch && cd third-party/executorch
 
-# Create a conda environment and install requirements.
-conda create -yn executorch python=3.10.0
-conda activate executorch
+# Create either a Python virtual environment:
+python3 -m venv .venv && source .venv/bin/activate && pip install --upgrade pip
+
+# Or a Conda environment:
+conda create -yn executorch python=3.10.0 && conda activate executorch
+
+# Install requirements
 ./install_executorch.sh
 
 cd ../..
@@ -76,11 +78,8 @@ pyenv install -s 3.10
 pyenv virtualenv 3.10 executorch
 pyenv activate executorch
 
-# Clone the ExecuTorch repository and submodules.
-mkdir third-party
-git clone -b release/0.4 https://github.com/pytorch/executorch.git third-party/executorch
-cd third-party/executorch
-git submodule update --init
+# Clone the ExecuTorch repository.
+git clone -b viable/strict https://github.com/pytorch/executorch.git third-party/executorch && cd third-party/executorch
 
 # Install requirements.
 PYTHON_EXECUTABLE=python ./install_executorch.sh

docs/source/using-executorch-building-from-source.md

@@ -36,27 +36,23 @@ portability details.
 
 ## Environment Setup
 
-### Create a Virtual Environment
+### Clone ExecuTorch
 
-[Install conda on your machine](https://conda.io/projects/conda/en/latest/user-guide/install/index.html). Then, create a virtual environment to manage our dependencies.
    ```bash
-   # Create and activate a conda environment named "executorch"
-   conda create -yn executorch python=3.10.0
-   conda activate executorch
+   # Clone the ExecuTorch repo from GitHub
+   git clone -b viable/strict https://github.com/pytorch/executorch.git && cd executorch
    ```
 
-### Clone ExecuTorch
+### Create a Virtual Environment
 
+Create and activate a Python virtual environment:
    ```bash
-   # Clone the ExecuTorch repo from GitHub
-   # 'main' branch is the primary development branch where you see the latest changes.
-   # 'viable/strict' contains all of the commits on main that pass all of the necessary CI checks.
-   git clone --branch viable/strict https://github.com/pytorch/executorch.git
-   cd executorch
-
-   # Update and pull submodules
-   git submodule sync
-   git submodule update --init
+   python3 -m venv .venv && source .venv/bin/activate && pip install --upgrade pip
+   ```
+
+Or alternatively, [install conda on your machine](https://conda.io/projects/conda/en/latest/user-guide/install/index.html). Then, create a Conda environment named "executorch".
+   ```bash
+   conda create -yn executorch python=3.10.0 && conda activate executorch
    ```
 
 ## Install ExecuTorch pip package from Source

docs/source/using-executorch-ios.md

@@ -25,7 +25,7 @@ The prebuilt ExecuTorch runtime, backend, and kernels are available as a [Swift
 
 #### Xcode
 
-In Xcode, go to `File > Add Package Dependencies`. Paste the URL of the [ExecuTorch repo](https://github.com/pytorch/executorch) into the search bar and select it. Make sure to change the branch name to the desired ExecuTorch version in format "swiftpm-<version>", (e.g. "swiftpm-0.5.0"), or a branch name in format "swiftpm-<version>.<year_month_date>" (e.g. "swiftpm-0.5.0-20250228") for a nightly build on a specific date.
+In Xcode, go to `File > Add Package Dependencies`. Paste the URL of the [ExecuTorch repo](https://github.com/pytorch/executorch) into the search bar and select it. Make sure to change the branch name to the desired ExecuTorch version in format "swiftpm-<version>", (e.g. "swiftpm-0.6.0"), or a branch name in format "swiftpm-<version>.<year_month_date>" (e.g. "swiftpm-0.7.0-20250401") for a nightly build on a specific date.
 
 ![](_static/img/swiftpm_xcode1.png)
 
@@ -58,7 +58,7 @@ let package = Package(
   ],
   dependencies: [
     // Use "swiftpm-<version>.<year_month_day>" branch name for a nightly build.
-    .package(url: "https://github.com/pytorch/executorch.git", branch: "swiftpm-0.5.0")
+    .package(url: "https://github.com/pytorch/executorch.git", branch: "swiftpm-0.6.0")
   ],
   targets: [
     .target(
@@ -97,7 +97,7 @@ xcode-select --install
 2. Clone ExecuTorch:
 
 ```bash
-git clone https://github.com/pytorch/executorch.git --depth 1 --recurse-submodules --shallow-submodules && cd executorch
+git clone -b viable/strict https://github.com/pytorch/executorch.git && cd executorch
 ```
 
 3. Set up [Python](https://www.python.org/downloads/macos/) 3.10+ and activate a virtual environment:
@@ -106,15 +106,16 @@ git clone https://github.com/pytorch/executorch.git --depth 1 --recurse-submodul
 python3 -m venv .venv && source .venv/bin/activate && ./install_requirements.sh
 ```
 
-4. Install the required dependencies, including those needed for the backends like [Core ML](backends-coreml.md) or [MPS](backends-mps.md). Choose one:
+4. Install the required dependencies, including those needed for the backends like [Core ML](backends-coreml.md) or [MPS](backends-mps.md). Choose one, or both:
 
 ```bash
 # ExecuTorch with xnnpack and CoreML backend
-./install_executorch.sh --pybind xnnpack
+./backends/apple/coreml/scripts/install_requirements.sh
+./install_executorch.sh --pybind coreml xnnpack
 
-# Optional: ExecuTorch with xnnpack, CoreML, and MPS backend
+# ExecuTorch with xnnpack and MPS backend
 ./backends/apple/mps/install_requirements.sh
-./install_executorch.sh --pybind xnnpack mps
+./install_executorch.sh --pybind mps xnnpack
 ```
 
 5. Install [CMake](https://cmake.org):