Merge branch 'main' into stable_diffusion_lcm

Author: suryasidd

.ci/docker/requirements-ci.txt

@@ -1,7 +1,7 @@
 mpmath==1.3.0
 numpy>=2.0.0; python_version >= '3.10'
 PyYAML==6.0.1
-ruamel.yaml==0.17.32
+ruamel.yaml==0.18.15
 sympy==1.12
 timm==0.6.13
 tomli==2.0.1

.ci/scripts/test_llama_lora.sh

@@ -55,7 +55,7 @@ cmake_build_llama_runner
 # Constants.
 RUNTIME_ARGS="--tokenizer_path=${DOWNLOADED_PATH}/tokenizer.model --temperature=0 --seq_len=20 --warmup=1"
 PROMPT="What happens if you eat watermelon seeds?"
-EXPECTED_PREFIX="What happens if you eat watermelon seeds? Watermelon seeds are a good source of vitamin C,"
+EXPECTED_PREFIX="What happens if you eat watermelon seeds? Watermelon seeds are a good source of vitamin C and"
 
 # Export LoRA PTE file.
 MODEL_NAME="llama_3_2_1B_lora"

.githooks/pre-commit

@@ -8,7 +8,11 @@ if git diff --cached --name-only | grep -q "^torch_pin.py$"; then
     echo "📝 Updating PyTorch commit pin..."
 
     # Run the update script
-    if python .github/scripts/update_pytorch_pin.py; then
+    hook_output=$(python .github/scripts/update_pytorch_pin.py 2>&1)
+    hook_status=$?
+    echo "$hook_output"
+
+    if [ $hook_status -eq 0 ]; then
         # Check if pytorch.txt was modified
         if ! git diff --quiet .ci/docker/ci_commit_pins/pytorch.txt; then
             echo "✅ PyTorch commit pin updated successfully"
@@ -19,9 +23,14 @@ if git diff --cached --name-only | grep -q "^torch_pin.py$"; then
             echo "ℹ️  PyTorch commit pin unchanged"
         fi
     else
-        echo "❌ Failed to update PyTorch commit pin"
-        echo "Please run: python .github/scripts/update_pytorch_pin.py"
-        exit 1
+        if echo "$hook_output" | grep -qi "rate limit exceeded"; then
+            echo "⚠️  PyTorch commit pin not updated due to GitHub API rate limiting."
+            echo "   Please manually update .ci/docker/ci_commit_pins/pytorch.txt if needed."
+        else
+            echo "❌ Failed to update PyTorch commit pin"
+            echo "Please run: python .github/scripts/update_pytorch_pin.py"
+            exit 1
+        fi
     fi
 fi
 

.github/scripts/update_pytorch_pin.py

@@ -4,7 +4,6 @@
 import re
 import sys
 import urllib.request
-from datetime import datetime
 
 
 def parse_nightly_version(nightly_version):
@@ -53,7 +52,7 @@ def get_commit_hash_for_nightly(date_str):
         Commit hash string
     """
     api_url = "https://api.github.com/repos/pytorch/pytorch/commits"
-    params = f"?sha=nightly&per_page=100"
+    params = f"?sha=nightly&per_page=50"
     url = api_url + params
 
     req = urllib.request.Request(url)
@@ -74,14 +73,21 @@ def get_commit_hash_for_nightly(date_str):
         commit_msg = commit.get("commit", {}).get("message", "")
         # Check if the first line of commit message matches
         first_line = commit_msg.split("\n")[0].strip()
-        if first_line == target_title or first_line.startswith(f"{date_str} nightly"):
-            return commit["sha"]
+        if first_line.startswith(f"{date_str} nightly"):
+            return extract_hash_from_title(first_line)
 
     raise ValueError(
         f"Could not find commit with title matching '{target_title}' in nightly branch"
     )
 
 
+def extract_hash_from_title(title):
+    match = re.search(r"\(([0-9a-fA-F]{7,40})\)", title)
+    if not match:
+        raise ValueError(f"Could not extract commit hash from title '{title}'")
+    return match.group(1)
+
+
 def update_pytorch_pin(commit_hash):
     """
     Update .ci/docker/ci_commit_pins/pytorch.txt with the new commit hash.

.github/workflows/pull.yml

@@ -892,6 +892,7 @@ jobs:
         # Install test requirements
         pip install -r backends/nxp/requirements-tests-pypi.txt
         pip install -r backends/nxp/requirements-tests-eiq.txt
+        PYTHON_EXECUTABLE=python bash examples/nxp/setup.sh
 
         # Run pytest
         PYTHON_EXECUTABLE=python bash backends/nxp/run_unittests.sh

backends/aoti/CMakeLists.txt

@@ -26,7 +26,7 @@ include(${EXECUTORCH_ROOT}/tools/cmake/Utils.cmake)
 find_package_torch()
 
 # Common AOTI functionality - combines all AOTI common components
-set(_aoti_common_sources aoti_model_container.cpp common_shims.cpp)
+set(_aoti_common_sources common_shims.cpp)
 add_library(aoti_common STATIC ${_aoti_common_sources})
 target_include_directories(
   aoti_common

backends/aoti/aoti_model_container.h renamed to backends/aoti/aoti_delegate_handle.h

@@ -60,36 +60,17 @@ using AOTInductorModelContainerRunFunc = AOTIRuntimeError (*)(
     AOTInductorStreamHandle stream_handle,
     AOTIProxyExecutorHandle proxy_executor_handle);
 
-// Global function pointers (will be loaded dynamically)
-extern AOTInductorModelContainerCreateWithDeviceFunc
-    AOTInductorModelContainerCreateWithDevice;
-extern AOTInductorModelContainerDeleteFunc AOTInductorModelContainerDelete;
-extern AOTInductorModelContainerGetNumInputsFunc
-    AOTInductorModelContainerGetNumInputs;
-extern AOTInductorModelContainerGetNumOutputsFunc
-    AOTInductorModelContainerGetNumOutputs;
-extern AOTInductorModelContainerRunFunc AOTInductorModelContainerRun;
-
 // Retrieves the name of an input tensor by index from the AOTI model container.
-// Needed by Metal backend
 using AOTInductorModelContainerGetInputNameFunc = AOTIRuntimeError (*)(
     AOTInductorModelContainerHandle container_handle,
     size_t input_idx,
     const char** input_name);
 
 // Retrieves the number of constants from the AOTI model container.
-// Needed by Metal backend
 using AOTInductorModelContainerGetNumConstantsFunc = AOTIRuntimeError (*)(
     AOTInductorModelContainerHandle container_handle,
     size_t* num_constants);
 
-// Global function pointers (will be loaded dynamically).
-// Needed by Metal backend
-extern AOTInductorModelContainerGetInputNameFunc
-    AOTInductorModelContainerGetInputName;
-extern AOTInductorModelContainerGetNumConstantsFunc
-    AOTInductorModelContainerGetNumConstants;
-
 } // extern "C"
 
 // AOTI Delegate Handle structure
@@ -99,6 +80,13 @@ struct AOTIDelegateHandle {
   AOTInductorModelContainerHandle container_handle;
   void* cuda_stream; // cudaStream_t stored as void* to avoid CUDA header
                      // dependency
+
+  // Function pointers specific to this handle's shared library
+  AOTInductorModelContainerCreateWithDeviceFunc create_with_device;
+  AOTInductorModelContainerDeleteFunc delete_container;
+  AOTInductorModelContainerGetNumInputsFunc get_num_inputs;
+  AOTInductorModelContainerGetNumOutputsFunc get_num_outputs;
+  AOTInductorModelContainerRunFunc run;
 };
 
 } // namespace aoti

backends/aoti/aoti_model_container.cpp

@@ -0,0 +1,17 @@
+load("@fbsource//xplat/executorch/build:runtime_wrapper.bzl", "runtime")
+
+oncall("executorch")
+
+runtime.python_library(
+    name = "passes",
+    srcs = [
+      "replace_view_copy_with_view.py",
+    ],
+    visibility = [
+        "//executorch/...",
+    ],
+    deps = [
+        "//caffe2:torch",
+        "//executorch/exir:pass_base",
+    ],
+)

backends/aoti/passes/TARGETS

@@ -4,9 +4,12 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
-# pyre-strict
+# This pass replaces view_copy ops with view ops. This is different than
+# exir/passes/replace_view_copy_with_view.py and exir/passes/reinplace.py
+# because this should only be used in the AOTInductor backend, as it
+# has less restrictions on whether the tensor memory is densely packed,
 
-from typing import Dict, Iterable, Tuple
+from typing import Dict, Iterable
 
 import torch
 from executorch.exir.dialects._ops import ops
@@ -15,33 +18,30 @@
 from torch import fx
 
 
-_SLICE_COPY_TARGETS: Tuple[torch._ops.OpOverload | EdgeOpOverload] = (
-    torch.ops.aten.slice_copy.Tensor,
-    ops.edge.aten.slice_copy.Tensor,
-)
-
-_SLICE_TARGETS: Dict[
+_VIEW_TARGETS: Dict[
     torch._ops.OpOverload | EdgeOpOverload, torch._ops.OpOverload | EdgeOpOverload
 ] = {
     torch.ops.aten.slice_copy.Tensor: torch.ops.aten.slice.Tensor,
     ops.edge.aten.slice_copy.Tensor: ops.edge.aten.slice.Tensor,
+    torch.ops.aten.select_copy.int: torch.ops.aten.select.int,
+    ops.edge.aten.select_copy.int: ops.edge.aten.select.int,
 }
 
 
-class ReplaceSliceCopyWithSlicePass(ExportPass):
-    """Replace non-mutated ``slice_copy`` results with ``slice`` views."""
+class ReplaceViewCopyWithViewPass(ExportPass):
+    """Replace non-mutated ``view_copy`` type of ops with ``view`` ops."""
 
     def call(self, graph_module: fx.GraphModule) -> PassResult:
         graph_changed = False
 
         for node in graph_module.graph.nodes:
-            if node.op != "call_function" or node.target not in _SLICE_COPY_TARGETS:
+            if node.op != "call_function" or node.target not in _VIEW_TARGETS:
                 continue
 
             if self._has_blocking_user(node, node.users.keys()):
                 continue
 
-            node.target = _SLICE_TARGETS[node.target]
+            node.target = _VIEW_TARGETS[node.target]
             graph_changed = True
 
         if graph_changed: