Merge branch 'main' into cuda-err-msg

Author: Gasoonjia

.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.

backends/aoti/utils.h

@@ -100,6 +100,64 @@ inline bool is_tensor_contiguous(
 
 } // extern "C"
 
+// Utility function to convert sizes pointer to vector
+inline std::vector<executorch::aten::SizesType> convert_sizes_to_vector(
+    int64_t ndim,
+    const int64_t* sizes_ptr) {
+  std::vector<executorch::aten::SizesType> sizes(ndim);
+  for (int i = 0; i < ndim; i++) {
+    sizes[i] = static_cast<executorch::aten::SizesType>(sizes_ptr[i]);
+  }
+  return sizes;
+}
+
+// Utility function to convert strides pointer to vector or calculate from sizes
+inline std::vector<executorch::aten::StridesType> convert_strides_to_vector(
+    int64_t ndim,
+    const int64_t* sizes_ptr,
+    const int64_t* strides_ptr) {
+  std::vector<executorch::aten::StridesType> strides(ndim);
+
+  if (strides_ptr != nullptr) {
+    // Use provided strides.
+    for (int64_t i = 0; i < ndim; i++) {
+      strides[i] = static_cast<executorch::aten::StridesType>(strides_ptr[i]);
+    }
+  } else {
+    // Calculate strides from sizes.
+    if (ndim > 0) {
+      strides[ndim - 1] = static_cast<executorch::aten::StridesType>(
+          1); // Last dimension has stride 1
+      for (int64_t i = ndim - 2; i >= 0; i--) {
+        if (sizes_ptr[i + 1] == 0) {
+          strides[i] = strides[i + 1]; // Copy stride when size is 0
+        } else {
+          strides[i] = static_cast<executorch::aten::StridesType>(
+              static_cast<int64_t>(strides[i + 1]) * sizes_ptr[i + 1]);
+        }
+      }
+    }
+  }
+  return strides;
+}
+
+// Check if tensor is in contiguous memory format (NCHW for 4D tensors)
+// Contiguous format means strides decrease from left to right:
+// For NCHW: strides = [C*H*W, H*W, W, 1]
+inline bool is_contiguous_tensor(
+    std::vector<executorch::aten::SizesType>& sizes,
+    std::vector<executorch::aten::StridesType>& strides) {
+  int64_t ndim = static_cast<int64_t>(strides.size());
+  int64_t expected_stride = 1;
+  for (int64_t i = ndim - 1; i >= 0; i--) {
+    if (strides[i] != expected_stride) {
+      return false;
+    }
+    expected_stride *= sizes[i];
+  }
+  return true;
+}
+
 } // namespace aoti
 } // namespace backends
 } // namespace executorch

backends/apple/metal/runtime/shims/tensor_attribute.cpp

@@ -0,0 +1,37 @@
+/*
+ * 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.
+ */
+
+#include <executorch/backends/apple/metal/runtime/shims/tensor_attribute.h>
+#include <executorch/backends/apple/metal/runtime/shims/utils.h>
+#include <iostream>
+
+namespace executorch {
+namespace backends {
+namespace metal {
+
+extern "C" {
+
+// Metal-specific device type constant
+__attribute__((__visibility__("default"))) int32_t
+aoti_torch_device_type_mps() {
+  return 13; // Consistent with c10/core/DeviceType.h
+}
+
+// Override aoti_torch_get_device_type to return MPS device type
+AOTITorchError aoti_torch_get_device_type(
+    AOTITensorHandle tensor,
+    int32_t* ret_device_type) {
+  *ret_device_type = aoti_torch_device_type_mps();
+  return Error::Ok;
+}
+
+} // extern "C"
+
+} // namespace metal
+} // namespace backends
+} // namespace executorch

backends/apple/metal/runtime/shims/tensor_attribute.h

@@ -0,0 +1,32 @@
+/*
+ * 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.
+ */
+
+#pragma once
+
+#include <executorch/backends/aoti/common_shims.h>
+#include <executorch/backends/apple/metal/runtime/shims/types.h>
+
+namespace executorch {
+namespace backends {
+namespace metal {
+
+extern "C" {
+
+// Metal-specific device type function
+int32_t aoti_torch_device_type_mps();
+
+// Override aoti_torch_get_device_type to return MPS device type
+AOTITorchError aoti_torch_get_device_type(
+    AOTITensorHandle tensor,
+    int32_t* ret_device_type);
+
+} // extern "C"
+
+} // namespace metal
+} // namespace backends
+} // namespace executorch

backends/apple/metal/runtime/shims/types.h

@@ -0,0 +1,35 @@
+/*
+ * 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.
+ */
+
+#pragma once
+
+#include <executorch/extension/tensor/tensor.h>
+#include <executorch/runtime/core/error.h>
+#include <cstdint>
+
+namespace executorch {
+namespace backends {
+namespace metal {
+
+// Common using declarations for ExecutorTorch types
+using executorch::runtime::Error;
+using executorch::runtime::etensor::Tensor;
+
+extern "C" {
+
+// Common AOTI type aliases
+// Note: AOTITensorHandle is aliased to Tensor* for ExecutorTorch compatibility
+using AOTITensorHandle = Tensor*;
+using AOTIRuntimeError = Error;
+using AOTITorchError = Error;
+
+} // extern "C"
+
+} // namespace metal
+} // namespace backends
+} // namespace executorch

backends/apple/metal/runtime/shims/utils.cpp

@@ -0,0 +1,51 @@
+/*
+ * 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.
+ */
+
+#include <executorch/backends/apple/metal/runtime/shims/utils.h>
+#include <executorch/runtime/platform/log.h>
+#include <cstdint>
+
+namespace executorch {
+namespace backends {
+namespace metal {
+
+extern "C" {
+
+// Helper function to check if a dtype is supported in Metal backend
+bool is_dtype_supported_in_et_metal(int32_t dtype) {
+  switch (dtype) {
+    case static_cast<int32_t>(SupportedDTypes::INT64):
+    case static_cast<int32_t>(SupportedDTypes::FLOAT32):
+    case static_cast<int32_t>(SupportedDTypes::BFLOAT16):
+      return true;
+    default:
+      return false;
+  }
+}
+
+// Metal-specific dtype validation utility function
+AOTITorchError validate_dtype(int32_t dtype) {
+  if (is_dtype_supported_in_et_metal(dtype)) {
+    return Error::Ok;
+  }
+
+  ET_LOG(
+      Error,
+      "Unsupported dtype: %d. Supported dtypes: %d (int64), %d (float32), %d (bfloat16)",
+      dtype,
+      static_cast<int32_t>(SupportedDTypes::INT64),
+      static_cast<int32_t>(SupportedDTypes::FLOAT32),
+      static_cast<int32_t>(SupportedDTypes::BFLOAT16));
+  return Error::InvalidArgument;
+}
+
+} // extern "C"
+
+} // namespace metal
+} // namespace backends
+} // namespace executorch

backends/apple/metal/runtime/shims/utils.h

@@ -0,0 +1,46 @@
+/*
+ * 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.
+ */
+
+#pragma once
+
+#include <executorch/backends/aoti/utils.h>
+#include <executorch/backends/apple/metal/runtime/shims/types.h>
+#include <executorch/runtime/core/exec_aten/exec_aten.h>
+#include <cstdint>
+
+namespace executorch {
+namespace backends {
+namespace metal {
+
+// Enum for supported data types in et-metal backend
+enum class SupportedDTypes : int32_t {
+  // UINT8 = 0,    // PyTorch's uint8 dtype code
+  // INT8 = 1,     // PyTorch's int8 dtype code
+  // INT16 = 2,    // PyTorch's int16 dtype code
+  // INT32 = 3,    // PyTorch's int32 dtype code
+  INT64 = 4, // PyTorch's int64 dtype code
+  // FLOAT16 = 5,  // PyTorch's float16 dtype code
+  FLOAT32 = 6, // PyTorch's float32 dtype code
+  // FLOAT64 = 7,  // PyTorch's float64 dtype code
+  // BOOL = 11,    // PyTorch's bool dtype code
+  BFLOAT16 = 15 // PyTorch's bfloat16 dtype code
+};
+
+extern "C" {
+
+// Helper function to check if a dtype is supported in Metal backend
+bool is_dtype_supported_in_et_metal(int32_t dtype);
+
+// Metal-specific dtype validation utility function
+AOTITorchError validate_dtype(int32_t dtype);
+
+} // extern "C"
+
+} // namespace metal
+} // namespace backends
+} // namespace executorch

backends/cadence/aot/compiler.py

@@ -38,7 +38,7 @@
     ExecutorchProgramManager,
 )
 from executorch.exir.passes import ToOutVarPass
-from executorch.exir.passes.sym_shape_eval_pass import HintBasedSymShapeEvalPass
+from executorch.exir.passes.sym_shape_eval_pass import ConstraintBasedSymShapeEvalPass
 from executorch.exir.program._program import to_edge
 
 from torch.export.exported_program import ExportedProgram
@@ -460,7 +460,7 @@ def _lower_ep_to_cadence_gen_etrecord(
             emit_stacktrace=False,
             to_out_var_pass=ToOutVarPass(),
             extract_delegate_segments=False,
-            sym_shape_eval_pass=HintBasedSymShapeEvalPass(),
+            sym_shape_eval_pass=ConstraintBasedSymShapeEvalPass(),
         ),
     )
 

backends/cadence/aot/quantizer/quantizer.py

@@ -342,3 +342,16 @@ def __init__(self, quantizers: Optional[list[Quantizer]] = None) -> None:
             quantizers = get_cadence_default_quantizers()
         quantizers.append(CadenceAtenQuantizer(SoftmaxPattern(), qconfig_A16))
         super().__init__(quantizers)
+
+
+class CadenceWith16BitLinearActivationsQuantizer(CadenceQuantizer):
+    """
+    Quantizer including A16 fully_connected
+    """
+
+    def __init__(self, quantizers: Optional[list[Quantizer]] = None) -> None:
+        if quantizers is None:
+            quantizers = []
+        # Add 16-bit quantizers for LinearPattern
+        quantizers.append(CadenceAtenQuantizer(LinearPattern(), qconfig_A16))
+        super().__init__(quantizers)