onnxscript 0.7.1.dev20260602__tar.gz → 0.7.1.dev20260609__tar.gz

{onnxscript-0.7.1.dev20260602/onnxscript.egg-info → onnxscript-0.7.1.dev20260609}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: onnxscript
-Version: 0.7.1.dev20260602
+Version: 0.7.1.dev20260609
 Summary: Naturally author ONNX functions and models using a subset of Python
 Author-email: Microsoft Corporation <onnx@microsoft.com>
 License: MIT License
@@ -27,7 +27,7 @@ License: MIT License
         
 Project-URL: Homepage, https://microsoft.github.io/onnxscript/
 Project-URL: Repository, https://github.com/microsoft/onnxscript
-Project-URL: Commit, https://github.com/microsoft/onnxscript/tree/1fb43af420966ff10b8baabbc3c1c50514a472fc
+Project-URL: Commit, https://github.com/microsoft/onnxscript/tree/33d144535078fc06121d3cb17d32cb41f851afb0
 Classifier: Development Status :: 4 - Beta
 Classifier: Environment :: Console
 Classifier: Intended Audience :: Developers

{onnxscript-0.7.1.dev20260602 → onnxscript-0.7.1.dev20260609}/onnxscript/function_libs/torch_lib/ops/core.py

@@ -4091,7 +4091,7 @@ def aten_from_file(
 def aten_full(
     size: Union[INT64, INT32],
     fill_value: TensorType,
-    dtype: int = FLOAT.dtype,
+    dtype: int = -1,
     layout: str = "",
     device: str = "",
     pin_memory: bool = False,
@@ -4348,6 +4348,11 @@ def aten_gru(
         # Extract hidden_size from hx shape: [num_layers * num_directions, batch, hidden_size]
         hidden_size_attr = hx.shape[2]
 
+        # linear_before_reset=1 matches PyTorch's GRU formulation where the linear
+        # transformation is applied before multiplying by the reset gate:
+        #   ht = g(Xt*(Wh^T) + rt (.) (Ht-1*(Rh^T) + Rbh) + Wbh)
+        # The ONNX default (linear_before_reset=0) uses a different equation and
+        # would produce numerically incorrect results.
         if B is not None:
             Y, Y_h = op.GRU(
                 current_input,
@@ -4357,6 +4362,7 @@ def aten_gru(
                 initial_h=layer_h,
                 direction=direction,
                 hidden_size=hidden_size_attr,
+                linear_before_reset=1,
             )
         else:
             Y, Y_h = op.GRU(
@@ -4366,6 +4372,7 @@ def aten_gru(
                 initial_h=layer_h,
                 direction=direction,
                 hidden_size=hidden_size_attr,
+                linear_before_reset=1,
             )
 
         # Y shape: [seq_length, num_directions, batch_size, hidden_size]
@@ -6081,10 +6088,12 @@ def aten_masked_fill(self: TTensor, mask: BOOL, value: TTensor) -> TTensor:
 def aten_masked_scatter(self: TTensor, mask: TTensor, source: TTensor) -> TTensor:
     """masked_scatter(Tensor self, Tensor mask, Tensor source) -> Tensor"""
 
-    if len(mask.shape) < len(self.shape):
-        mask = op.Expand(mask, op.Shape(self))
-    else:
-        self = op.Expand(self, op.Shape(mask))
+    # Broadcast self and mask to their common shape so NonZero enumerates every
+    # masked element. The previous rank-only check missed same-rank broadcasting
+    # (e.g. mask (1, S, 1) vs self (1, S, D)): it left mask un-expanded, so only a
+    # subset of masked positions were scattered (pytorch/pytorch#186146).
+    self = op.Expand(self, op.Shape(mask))
+    mask = op.Expand(mask, op.Shape(self))
     index = op.Transpose(op.NonZero(mask), perm=[1, 0])
 
     # NOTE: source can have more elements than needed.

{onnxscript-0.7.1.dev20260602 → onnxscript-0.7.1.dev20260609}/onnxscript/function_libs/torch_lib/ops/vision.py

@@ -8,10 +8,12 @@
 from __future__ import annotations
 
 import warnings
-from typing import Sequence
+from typing import Optional, Sequence
 
 from onnxscript.function_libs.torch_lib.registration import torch_op
+from onnxscript.function_libs.torch_lib.tensor_typing import TFloat
 from onnxscript.onnx_opset import opset18 as op
+from onnxscript.onnx_opset import opset19
 from onnxscript.onnx_types import FLOAT, INT64
 
 _INT64_MAX = 0x7FFFFFFFFFFFFFFF
@@ -91,3 +93,36 @@ def torchvision_roi_pool(input, boxes, output_size: Sequence[int], spatial_scale
         pooled_shape=(pooled_height, pooled_width),
         spatial_scale=spatial_scale,
     )
+
+
+@torch_op("torchvision::deform_conv2d", trace_only=True)
+def torchvision_deform_conv2d(
+    input: TFloat,
+    weight: TFloat,
+    offset: TFloat,
+    mask: Optional[TFloat],
+    bias: Optional[TFloat],
+    stride_h: int,
+    stride_w: int,
+    pad_h: int,
+    pad_w: int,
+    dilation_h: int,
+    dilation_w: int,
+    groups: int,
+    offset_groups: int,
+    use_mask: bool,
+):
+    """torchvision::deform_conv2d(Tensor input, Tensor weight, Tensor offset, Tensor? mask, Tensor? bias, int stride_h, int stride_w, int pad_h, int pad_w, int dilation_h, int dilation_w, int groups, int offset_groups, bool use_mask) -> Tensor"""
+
+    return opset19.DeformConv(
+        X=input,
+        W=weight,
+        offset=offset,
+        B=bias,
+        mask=mask if use_mask else None,
+        dilations=(dilation_h, dilation_w),
+        strides=(stride_h, stride_w),
+        pads=(pad_h, pad_w, pad_h, pad_w),
+        group=groups,
+        offset_group=offset_groups,
+    )

{onnxscript-0.7.1.dev20260602 → onnxscript-0.7.1.dev20260609}/onnxscript/rewriter/rules/common/_fuse_pad_into_conv.py

@@ -146,6 +146,9 @@ class _FuseConvPadBase(orp.RewriteRuleClassBase):
         if np.any(self._pads_list[:2] + self._pads_list[x_rank : x_rank + 2]):
             self._pads_list = None
             return check_result.fail(f"{pads.name} must be zero in non-spatial dimensions.")
+        if any(p < 0 for p in self._pads_list):
+            self._pads_list = None
+            return check_result.fail(f"{pads.name} must not contain negative values.")
 
         return check_result
 

{onnxscript-0.7.1.dev20260602 → onnxscript-0.7.1.dev20260609}/onnxscript/rewriter/rules/common/_fuse_relus_clips.py

@@ -74,7 +74,9 @@ class _FuseReluClipBase(RewriteRuleClassBase, abc.ABC):
                 min_clip = min_input.const_value.numpy()
 
         if len(node.inputs) > 2:
-            max_clip = node.inputs[2].const_value.numpy()
+            max_clip = node.inputs[2]
+            if max_clip is not None:
+                max_clip = max_clip.const_value.numpy()
 
         return min_clip, max_clip, dtype
 

{onnxscript-0.7.1.dev20260602 → onnxscript-0.7.1.dev20260609/onnxscript.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: onnxscript
-Version: 0.7.1.dev20260602
+Version: 0.7.1.dev20260609
 Summary: Naturally author ONNX functions and models using a subset of Python
 Author-email: Microsoft Corporation <onnx@microsoft.com>
 License: MIT License
@@ -27,7 +27,7 @@ License: MIT License
         
 Project-URL: Homepage, https://microsoft.github.io/onnxscript/
 Project-URL: Repository, https://github.com/microsoft/onnxscript
-Project-URL: Commit, https://github.com/microsoft/onnxscript/tree/1fb43af420966ff10b8baabbc3c1c50514a472fc
+Project-URL: Commit, https://github.com/microsoft/onnxscript/tree/33d144535078fc06121d3cb17d32cb41f851afb0
 Classifier: Development Status :: 4 - Beta
 Classifier: Environment :: Console
 Classifier: Intended Audience :: Developers