Merge branch 'main' into op_softmax

Author: Erik-Lundell

.ci/scripts/build_llama_android.sh

@@ -19,7 +19,6 @@ install_executorch_and_backend_lib() {
   cmake -DBUCK2="${BUCK2}" \
     -DCMAKE_TOOLCHAIN_FILE="${ANDROID_NDK}/build/cmake/android.toolchain.cmake" \
     -DANDROID_ABI="${ANDROID_ABI}" \
-    -DANDROID_PLATFORM=android-23 \
     -DCMAKE_INSTALL_PREFIX=cmake-android-out \
     -DCMAKE_BUILD_TYPE=Release \
     -DEXECUTORCH_BUILD_EXTENSION_DATA_LOADER=ON \
@@ -41,7 +40,6 @@ build_llama_runner() {
     cmake -DBUCK2="${BUCK2}" \
     -DCMAKE_TOOLCHAIN_FILE="$ANDROID_NDK"/build/cmake/android.toolchain.cmake  \
     -DANDROID_ABI="${ANDROID_ABI}" \
-    -DANDROID_PLATFORM=android-23 \
     -DCMAKE_INSTALL_PREFIX=cmake-android-out \
     -DCMAKE_BUILD_TYPE=Release -DPYTHON_EXECUTABLE=python \
     -DEXECUTORCH_BUILD_XNNPACK=ON \

.ci/scripts/test_llava.sh

@@ -56,7 +56,6 @@ cmake_install_executorch_libraries_for_android() {
     cmake                                                                       \
         -DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake \
         -DANDROID_ABI=arm64-v8a                                                 \
-        -DANDROID_PLATFORM=android-23                                           \
         ${EXECUTORCH_COMMON_CMAKE_ARGS}                                         \
         -B${BUILD_DIR} .
 
@@ -93,7 +92,6 @@ cmake_build_llava_runner_for_android() {
     cmake                                                                       \
         -DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake \
         -DANDROID_ABI=arm64-v8a                                                 \
-        -DANDROID_PLATFORM=android-23                                           \
         ${LLAVA_COMMON_CMAKE_ARGS}                                              \
         -DCMAKE_PREFIX_PATH="$python_lib"                                       \
         -DLLAVA_RUNNER_NO_TORCH_DUMMY_IMAGE=ON                                  \

.github/scripts/propose_ghstack_orig_pr.py

@@ -0,0 +1,135 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import argparse
+import os
+import re
+
+from typing import List
+
+# Provided by the PyGithub pip package.
+from github import Auth, Github
+from github.Repository import Repository
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(
+        description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter
+    )
+    parser.add_argument(
+        "--repo",
+        type=str,
+        help='The github repo to modify: e.g. "pytorch/executorch".',
+        required=True,
+    )
+    parser.add_argument(
+        "--pr",
+        type=int,
+        help="Number of the PR in the stack to check and create corresponding PR",
+        required=True,
+    )
+    return parser.parse_args()
+
+
+def extract_stack_from_body(pr_body: str) -> List[int]:
+    """Extracts a list of PR numbers from a ghexport-generated PR body.
+
+    The base of the stack is in index 0.
+    """
+
+    # Expected format. The `__->__` could appear on any line. Stop parsing
+    # after the blank line. This would return [1, 2, 3].
+    """
+    Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
+    * #3
+    * __->__ #2
+    * #1
+
+    <PR description details>
+    """
+
+    prs = []
+    ghstack_begin = (
+        "Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):"
+    )
+    ghstack_begin_seen = False
+    for line in pr_body.splitlines():
+        if ghstack_begin in line:
+            ghstack_begin_seen = True
+        if not ghstack_begin_seen:
+            continue
+        match = re.match(r"\*(?:.*?)? #(\d+)", line)
+        if match:
+            # It's a bullet followed by an integer.
+            prs.append(int(match.group(1)))
+    return list(reversed(prs))
+
+
+def get_pr_stack_from_number(pr_number: int, repo: Repository) -> List[int]:
+    pr_stack = extract_stack_from_body(repo.get_pull(pr_number).body)
+
+    if not pr_stack:
+        raise Exception(
+            f"Could not find PR stack in body of #{pr_number}. "
+            + "Please make sure that the PR was created with ghstack."
+        )
+
+    return pr_stack
+
+
+def create_prs_for_orig_branch(pr_stack: List[int], repo: Repository):
+    # For the first PR, we want to merge to `main` branch, and we will update
+    # as we go through the stack
+    orig_branch_merge_base = "main"
+    for i in range(len(pr_stack)):
+        pr = repo.get_pull(pr_stack[i])
+        if not pr.is_merged():
+            print("The PR (and stack above) is not merged yet, skipping")
+            return
+        # Check for invariant: For the current PR, it must be gh/user/x/base <- gh/user/x/head
+        assert pr.base.ref.replace("base", "head") == pr.head.ref
+        # The PR we want to create is then "branch_to_merge" <- gh/user/x/orig
+        # gh/user/x/orig is the clean diff between gh/user/x/base <- gh/user/x/head
+        orig_branch_merge_head = pr.base.ref.replace("base", "orig")
+        bot_metadata = f"""This PR was created by the merge bot to help merge the original PR into the main branch.
+ghstack PR number: https://github.com/pytorch/executorch/pull/{pr.number}
+^ Please use this as the source of truth for the PR details, comments, and reviews
+ghstack PR base: https://github.com/pytorch/executorch/tree/{pr.base.ref}
+ghstack PR head: https://github.com/pytorch/executorch/tree/{pr.head.ref}
+Merge bot PR base: https://github.com/pytorch/executorch/tree/{orig_branch_merge_base}
+Merge bot PR head: https://github.com/pytorch/executorch/tree/{orig_branch_merge_head}"""
+
+        existing_orig_pr = repo.get_pulls(
+            head="pytorch:" + orig_branch_merge_head,
+            base=orig_branch_merge_base,
+            state="open",
+        )
+        if existing_orig_pr.totalCount > 0:
+            print(
+                f"PR for {orig_branch_merge_head} already exists {existing_orig_pr[0]}"
+            )
+            # We don't need to create/edit because the head PR is merged and orig is finalized.
+        else:
+            repo.create_pull(
+                base=orig_branch_merge_base,
+                head=orig_branch_merge_head,
+                title=pr.title,
+                body=bot_metadata,
+            )
+        # Advance the base for the next PR
+        orig_branch_merge_base = orig_branch_merge_head
+
+
+def main():
+    args = parse_args()
+
+    with Github(auth=Auth.Token(os.environ["GITHUB_TOKEN"])) as gh:
+        repo = gh.get_repo(args.repo)
+        create_prs_for_orig_branch(get_pr_stack_from_number(args.pr, repo), repo)
+
+
+if __name__ == "__main__":
+    main()

.github/workflows/ghstack_land.yml

@@ -0,0 +1,40 @@
+name: Propose to merge ghstack orig PRs to main
+on:
+  pull_request:
+    types: [closed]
+    branches:
+      - 'gh/cccclai/[0-9]+/base'
+      - 'gh/dbort/[0-9]+/base'
+      - 'gh/guangy10/[0-9]+/base'
+      - 'gh/helunwencser/[0-9]+/base'
+      - 'gh/jorgep31415/[0-9]+/base'
+      - 'gh/kimishpatel/[0-9]+/base'
+      - 'gh/kirklandsign/[0-9]+/base'
+      - 'gh/larryliu0820/[0-9]+/base'
+      - 'gh/manuelcandales/[0-9]+/base'
+      - 'gh/mcr229/[0-9]+/base'
+      - 'gh/swolchok/[0-9]+/base'
+      - 'gh/SS-JIA/[0-9]+/base'
+
+jobs:
+  ghstack_merge_to_main:
+    name: Try to create a PR with ghstack /orig branch
+    runs-on: ubuntu-22.04
+    environment: cherry-pick-bot
+    steps:
+    - uses: actions/checkout@v3
+      with:
+        fetch-depth: '0'
+    - uses: actions/setup-python@v4
+      with:
+        python-version: '3.10'
+    - name: Try to merge PR to main
+      run: |
+        pip install pygithub
+
+        PR_NUMBER=$(echo "$GITHUB_REF" | grep -oE '[0-9]+')
+
+        python .github/scripts/propose_ghstack_orig_pr.py --pr $PR_NUMBER --repo pytorch/executorch
+      env:
+        GITHUB_TOKEN: ${{ secrets.GH_PYTORCHBOT_CHERRY_PICK_TOKEN }}
+        GITHUB_REF: ${{ github.ref }}

CMakeLists.txt

@@ -372,6 +372,9 @@ endif()
 # Detect if an Android toolchain is set.
 if(CMAKE_TOOLCHAIN_FILE MATCHES ".*android\.toolchain\.cmake$")
   set(CMAKE_TOOLCHAIN_ANDROID ON)
+if(NOT ANDROID_PLATFORM)
+  set(ANDROID_PLATFORM android-30)
+endif()
 else()
   set(CMAKE_TOOLCHAIN_ANDROID OFF)
 endif()

backends/arm/_passes/annotate_channels_last_dim_order_pass.py

@@ -9,6 +9,7 @@
 from typing import cast
 
 import torch
+from executorch.backends.arm._passes.arm_pass_utils import get_first_fake_tensor
 from executorch.backends.arm.tosa_quant_utils import dq_op
 from executorch.backends.arm.tosa_utils import is_consumer_node_depthwise_conv2d
 from executorch.exir.pass_base import ExportPass, PassResult
@@ -52,12 +53,7 @@ def call(self, graph_module: torch.fx.GraphModule):
         NHWC_Order = (0, 2, 3, 1)
         HWCM_Order = (2, 3, 0, 1)
         for node in graph_module.graph.nodes:
-            if isinstance(
-                node.meta["val"], (tuple, torch.fx.immutable_collections.immutable_list)
-            ):
-                node_data = node.meta["val"][0].data
-            else:
-                node_data = node.meta["val"].data
+            node_data = get_first_fake_tensor(node).data
 
             if len(node_data.shape) == 4:
                 dim_order = NHWC_Order

backends/arm/_passes/arm_pass_manager.py

@@ -25,6 +25,7 @@
 from executorch.backends.arm._passes.insert_squeeze_after_sum_pass import (
     InsertSqueezeAfterSumPass,
 )
+from executorch.backends.arm._passes.match_arg_ranks_pass import MatchArgRanksPass
 from executorch.backends.arm._passes.meandim_to_averagepool_pass import (
     ConvertMeanDimToAveragePool,
 )
@@ -33,6 +34,9 @@
     ScalarsToAttributePass,
 )
 from executorch.backends.arm._passes.size_adjust_conv2d_pass import SizeAdjustConv2DPass
+from executorch.backends.arm._passes.unsqueeze_scalar_placeholders_pass import (
+    UnsqueezeScalarPlaceholdersPass,
+)
 from executorch.exir import ExportedProgram
 from executorch.exir.backend.compile_spec_schema import CompileSpec
 from executorch.exir.pass_manager import PassManager
@@ -48,10 +52,12 @@ def transform_to_backend_pipeline(
     ):
         """Apply passes before transforming program to backend"""
         self.add_pass(CastInt64ToInt32Pass(exported_program))
+        self.add_pass(UnsqueezeScalarPlaceholdersPass(exported_program))
         self.add_pass(SizeAdjustConv2DPass())
         self.add_pass(RemoveClonePass())
         self.add_pass(ConvertExpandCopyToRepeatPass())
         self.add_pass(ConvertMeanDimToAveragePool())
+        self.add_pass(MatchArgRanksPass(exported_program))
         self.add_pass(DecomposeDivPass())
         self.add_pass(InsertSqueezeAfterSumPass())
         self.add_pass(ConvertSplitToSlicePass())
@@ -65,7 +71,7 @@ def transform_to_backend_pipeline(
         return self._transform(exported_program.graph_module)
 
     def transform_for_annotation_pipeline(self, graph_module: torch.fx.GraphModule):
-        self.add_pass(DecomposeDivPass())
         self.add_pass(ScalarsToAttributePass())
         self.add_pass(DecomposeSoftmaxesPass())
+        self.add_pass(DecomposeDivPass())
         return self._transform(graph_module)

backends/arm/_passes/arm_pass_utils.py

@@ -7,9 +7,11 @@
 from typing import Optional
 
 import torch
+import torch.fx
 
 from executorch.exir.dialects._ops import ops as exir_ops
 from torch._ops import OpOverload
+from torch._subclasses.fake_tensor import FakeTensor
 
 
 def create_node(
@@ -64,3 +66,21 @@ def insert_q_dq_pair(
     # node's first use
     q.args = (anchor,) + q_params
     return dq
+
+
+def get_first_fake_tensor(node: torch.fx.Node) -> FakeTensor:
+    """
+    Returns a FakeTensor from the meta field of 'node'.
+    If the node contains many fake tensors, return the first one.
+    """
+    if isinstance(
+        node.meta["val"], (tuple, torch.fx.immutable_collections.immutable_list)
+    ):
+        fake_tensor = node.meta["val"][0]
+    else:
+        fake_tensor = node.meta["val"]
+
+    assert isinstance(
+        fake_tensor, FakeTensor
+    ), f'Found {fake_tensor} in meta["val"] of {node}, expected to find FakeTensor.'
+    return fake_tensor

backends/arm/_passes/decompose_div_pass.py

@@ -8,15 +8,18 @@
 from executorch.exir.dialects._ops import ops as exir_ops
 from executorch.exir.pass_base import ExportPass
 
+edge_div_ops = (exir_ops.edge.aten.div.Tensor,)
+aten_div_ops = (torch.ops.aten.div.Tensor, torch.ops.aten.div_.Tensor)
+
 
 def get_div_decomposition(op) -> tuple:
     """
     Returns the the (reciprocal_op, mul_op), where the ops depends on if
     the div op is in exir_ops torch.ops.aten.
     """
-    if op == exir_ops.edge.aten.div.Tensor:
+    if op in edge_div_ops:
         return (exir_ops.edge.aten.reciprocal.default, exir_ops.edge.aten.mul.Tensor)
-    if op == torch.ops.aten.div.Tensor:
+    if op in aten_div_ops:
         return (torch.ops.aten.reciprocal.default, torch.ops.aten.mul.Tensor)
     raise RuntimeError(f"Can't get div decomposition for op {op}")
 
@@ -33,7 +36,7 @@ class DecomposeDivPass(ExportPass):
     """
 
     def call_operator(self, op, args, kwargs, meta):
-        if op not in (exir_ops.edge.aten.div.Tensor, torch.ops.aten.div.Tensor):
+        if op not in (edge_div_ops + aten_div_ops):
             return super().call_operator(op, args, kwargs, meta)
 
         reciprocal_op, mul_op = get_div_decomposition(op)