onnxslim 0.1.83__py3-none-any.whl → 0.1.84__py3-none-any.whl

onnxslim/core/pattern/elimination/slice.py

@@ -39,6 +39,16 @@ class SlicePatternMatcher(PatternMatcher):
                 first_slice_node_axes = first_slice_node_inputs[3].values.tolist()
                 first_slice_node_steps = first_slice_node_inputs[4].values.tolist()
 
+                # Check all users upfront before modifying the graph.
+                # If any user has overlapping axes, skip the optimization entirely
+                # to avoid corrupting the graph (fixes GitHub issue #277).
+                for user_node in first_slice_node_users:
+                    second_slice_node_inputs = list(user_node.inputs)
+                    second_slice_node_axes = second_slice_node_inputs[3].values.tolist()
+                    new_axes = first_slice_node_axes + second_slice_node_axes
+                    if len(new_axes) != len(set(new_axes)):
+                        return match_case
+
                 for user_node in first_slice_node_users:
                     second_slice_node = user_node
                     second_slice_node_inputs = list(second_slice_node.inputs)
@@ -52,33 +62,30 @@ class SlicePatternMatcher(PatternMatcher):
                     new_axes = first_slice_node_axes + second_slice_node_axes
                     new_steps = first_slice_node_steps + second_slice_node_steps
 
-                    if len(new_axes) != len(set(new_axes)):
-                        continue
-
                     inputs = []
+                    output_name = second_slice_node.outputs[0].name
                     inputs.extend(
                         (
                             next(iter(first_slice_node.inputs)),
                             gs.Constant(
-                                second_slice_node_inputs[1].name + "_starts",
+                                output_name + "_starts",
                                 values=np.array(new_starts, dtype=np.int64),
                             ),
                             gs.Constant(
-                                second_slice_node_inputs[2].name + "_ends",
+                                output_name + "_ends",
                                 values=np.array(new_ends, dtype=np.int64),
                             ),
                             gs.Constant(
-                                second_slice_node_inputs[3].name + "_axes",
+                                output_name + "_axes",
                                 values=np.array(new_axes, dtype=np.int64),
                             ),
                             gs.Constant(
-                                second_slice_node_inputs[4].name + "_steps",
+                                output_name + "_steps",
                                 values=np.array(new_steps, dtype=np.int64),
                             ),
                         )
                     )
                     outputs = list(second_slice_node.outputs)
-
                     first_slice_node.outputs.clear()
                     second_slice_node.inputs.clear()
                     second_slice_node.outputs.clear()

onnxslim/core/pattern/fusion/concat_reshape.py

@@ -36,9 +36,11 @@ class ConcatReshapeMatcher(PatternMatcher):
     def rewrite(self, opset=11):
         match_case = {}
         concat_node = self.concat_0
+        reshape_node = self.reshape_0
         index = next(idx for idx, i in enumerate(concat_node.inputs) if isinstance(i, gs.Variable))
+        output_name = reshape_node.outputs[0].name
         constant = gs.Constant(
-            concat_node.inputs[index].name + "_fixed",
+            output_name + "_fixed",
             values=np.array([-1], dtype=np.int64),
         )
         concat_node.inputs.pop(index)

onnxslim/core/pattern/fusion/convadd.py

@@ -44,12 +44,8 @@ class ConvAddMatcher(PatternMatcher):
             inputs = []
             inputs.append(next(iter(conv_node.inputs)))
             inputs.append(conv_weight)
-            weight_name = list(conv_node.inputs)[1].name
-            if weight_name.endswith("weight"):
-                bias_name = f"{weight_name[:-6]}bias"
-            else:
-                bias_name = f"{weight_name}_bias"
-            inputs.append(gs.Constant(bias_name, values=conv_bias))
+            output_name = add_node.outputs[0].name
+            inputs.append(gs.Constant(output_name + "_bias", values=conv_bias))
             outputs = list(add_node.outputs)
 
             conv_node.outputs.clear()

onnxslim/core/pattern/fusion/convbn.py

@@ -52,15 +52,11 @@ class ConvBatchNormMatcher(PatternMatcher):
 
             inputs = []
             inputs.append(next(iter(conv_transpose_node.inputs)))
-            weight_name = list(conv_transpose_node.inputs)[1].name
-            if weight_name.endswith("weight"):
-                bias_name = f"{weight_name[:-6]}bias"
-            else:
-                bias_name = f"{weight_name}_bias"
+            output_name = bn_node.outputs[0].name
             inputs.extend(
                 (
-                    gs.Constant(weight_name + "_weight", values=conv_w),
-                    gs.Constant(bias_name, values=conv_b),
+                    gs.Constant(output_name + "_weight", values=conv_w),
+                    gs.Constant(output_name + "_bias", values=conv_b),
                 )
             )
             outputs = list(bn_node.outputs)

onnxslim/core/pattern/fusion/convmul.py

@@ -38,14 +38,13 @@ class ConvMulMatcher(PatternMatcher):
                 inputs = []
                 inputs.append(next(iter(conv_node.inputs)))
 
-                weight_name = list(conv_node.inputs)[1].name
-                inputs.append(gs.Constant(weight_name, values=new_weight))
+                output_name = mul_node.outputs[0].name
+                inputs.append(gs.Constant(output_name + "_weight", values=new_weight))
 
                 if len(conv_node.inputs) == 3:
                     conv_bias = conv_node.inputs[2].values
                     new_bias = conv_bias * mul_constant.squeeze()
-                    bias_name = list(conv_node.inputs)[2].name
-                    inputs.append(gs.Constant(bias_name, values=new_bias))
+                    inputs.append(gs.Constant(output_name + "_bias", values=new_bias))
 
                 outputs = list(mul_node.outputs)
 

onnxslim/core/pattern/fusion/gemm.py

@@ -76,7 +76,7 @@ class MatMulAddPatternMatcher(PatternMatcher):
                 output_variable.outputs.remove(add_node)
 
                 matmul_bias_transpose_constant = gs.Constant(
-                    matmul_bias_variable.name, values=matmul_bias_variable.values.T
+                    f"{matmul_node.name}_weight", values=matmul_bias_variable.values.T
                 )
 
                 inputs = []
@@ -143,7 +143,7 @@ class MatMulAddPatternMatcher(PatternMatcher):
                 output_variable.outputs.remove(add_node)
 
                 matmul_bias_transpose_constant = gs.Constant(
-                    matmul_bias_variable.name, values=matmul_bias_variable.values.T
+                    f"{matmul_node.name}_weight", values=matmul_bias_variable.values.T
                 )
 
                 inputs = []
@@ -235,14 +235,15 @@ class GemmMulPatternMatcher(PatternMatcher):
                     gemm_weight_fused = gemm_weight * mul_weight[:, None]
                 else:
                     gemm_weight_fused = gemm_weight * mul_weight
-                gemm_weight_fused_constant = gs.Constant(gemm_weight_constant.name + "_fused", values=gemm_weight_fused)
+                output_name = reshape_node.outputs[0].name
+                gemm_weight_fused_constant = gs.Constant(output_name + "_weight_fused", values=gemm_weight_fused)
                 gemm_node.inputs[1] = gemm_weight_fused_constant
 
                 if gemm_bias_constant:
                     gemm_bias = gemm_bias_constant.values
                     mul_bias = mul_bias_variable.values
                     gemm_bias_fused = gemm_bias * mul_bias
-                    gemm_bias_fused_constant = gs.Constant(gemm_bias_constant.name + "_fused", values=gemm_bias_fused)
+                    gemm_bias_fused_constant = gs.Constant(output_name + "_bias_fused", values=gemm_bias_fused)
                     gemm_node.inputs[2] = gemm_bias_fused_constant
 
                 mul_node.replace_all_uses_with(reshape_node)
@@ -312,7 +313,8 @@ class GemmAddPatternMatcher(PatternMatcher):
                     and add_bias.ndim <= 2
                 ):
                     gemm_bias_fused = gemm_bias + add_bias
-                    gemm_bias_fused_constant = gs.Constant(gemm_bias_constant.name + "_fused", values=gemm_bias_fused)
+                    output_name = reshape_node.outputs[0].name
+                    gemm_bias_fused_constant = gs.Constant(output_name + "_bias_fused", values=gemm_bias_fused)
                     gemm_node.inputs[2] = gemm_bias_fused_constant
                 else:
                     return match_case

onnxslim/third_party/onnx_graphsurgeon/ir/graph.py

@@ -794,109 +794,6 @@ class Graph:
                     tensor.to_constant(arr)
                     tensor.inputs.clear()
 
-        # Pass 2: Run shape-tensor cast elision
-        def run_cast_elision(node):
-            """Perform cast elision optimization on an ONNX node to eliminate unnecessary cast operations."""
-            import onnx
-
-            # Search for Cast(s) (from int -> float) -> intermediate operator (with float constants) -> Cast(s) (back to int)
-            # This pattern is problematic for TensorRT since these operations may be performed on Shape Tensors, which
-            # are not allowed to be floating point type. Attempt to fold the pattern here
-            VALID_CAST_ELISION_OPS = {
-                "Add",
-                "Sub",
-                "Mul",
-                "Div",
-                "Max",
-                "Min",
-                "Equal",
-                "Greater",
-                "Less",
-                "Concat",
-            }
-
-            if node.op not in VALID_CAST_ELISION_OPS:
-                return
-
-            # If the uncasted outputs of this node have any consumers other than "Cast" nodes,
-            # then we cannot elide the cast.
-            for out_tensor in node.outputs:
-                if out_tensor in self.outputs:
-                    return
-
-                if any(out_node.op != "Cast" for out_node in out_tensor.outputs):
-                    return
-
-            # Get list of input nodes that cast to float32
-            inp_casts = [
-                inp_node
-                for inp_tensor in node.inputs
-                for inp_node in inp_tensor.inputs
-                if inp_node.op == "Cast" and inp_node.attrs["to"] == onnx.TensorProto.DataType.FLOAT
-            ]
-
-            # No cast nodes found, return early
-            if not inp_casts:
-                return
-
-            # Ensure that all input cast nodes are casting from the same type
-            inp_dtypes = [dtype_to_onnx(inp_cast.inputs[0].dtype) for inp_cast in inp_casts]
-            if len(set(inp_dtypes)) != 1:
-                return
-
-            final_type = inp_dtypes[0]
-
-            # Get list of output nodes that cast to int32 or int64
-            out_casts = [
-                out_node
-                for out_tensor in node.outputs
-                for out_node in out_tensor.outputs
-                if out_node.op == "Cast"
-                and out_node.attrs["to"] in {onnx.TensorProto.DataType.INT32, onnx.TensorProto.DataType.INT64}
-            ]
-
-            # No cast node found on outputs, return early
-            if not out_casts:
-                return
-
-            # Ensure that all output cast nodes are casting to the same type and that this
-            # matches the original type before the inputs were casted.
-            out_dtypes = [out_cast.attrs["to"] for out_cast in out_casts]
-            if len(set(out_dtypes)) != 1 or out_dtypes[0] != final_type:
-                return
-
-            # If all checks passed, reconnect inputs/outputs to the consumers/producers
-            # of the Cast nodes.
-            # Note that we need to be careful in how we rebind tensors since they may
-            # be used by multiple nodes. Thus, it is not necessarily safe to assume that
-            # `cast_node.inputs[0].outputs[0] == cast_node`.
-            for index, inp in enumerate(node.inputs):
-                if isinstance(inp, Constant):
-                    inp.values = inp.values.astype(onnx.helper.tensor_dtype_to_np_dtype(final_type))
-
-                for cast in inp_casts:
-                    if cast.outputs[0] == inp:
-                        node.inputs[index] = cast.inputs[0]
-
-            for index, out in enumerate(node.outputs):
-                for cast in out_casts:
-                    if cast.inputs[0] == out:
-                        out_tensor = cast.outputs[0]
-                        out_tensor.inputs.clear()  # Disconnect from Cast
-                        node.outputs[index] = out_tensor
-
-        if fold_shapes:
-            # Perform shape tensor cast elision prior to most other folding
-            G_LOGGER.debug(f"Performing shape tensor cast elision in {self.name}")
-            try:
-                with self.node_ids():
-                    for node in self.nodes:
-                        run_cast_elision(node)
-            except Exception as err:
-                if not error_ok:
-                    raise err
-                G_LOGGER.warning("'{:}' routine failed with: {:}".format("Shape tensor cast elision", err))
-
         # Note that most of the remaining passes operate on a clone of the original graph.
         # Pass 3: Find all descendants of constant tensors
 

{onnxslim-0.1.83.dist-info → onnxslim-0.1.84.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: onnxslim
-Version: 0.1.83
+Version: 0.1.84
 Summary: OnnxSlim: A Toolkit to Help Optimize Onnx Model
 Project-URL: homepage, https://github.com/inisis/OnnxSlim
 Project-URL: issues, https://github.com/inisis/OnnxSlim/issues

{onnxslim-0.1.83.dist-info → onnxslim-0.1.84.dist-info}/RECORD

@@ -16,15 +16,15 @@ onnxslim/core/pattern/elimination/__init__.py,sha256=C9EwJj7DQmaXVvGx6wxvqvCdQGE
 onnxslim/core/pattern/elimination/concat.py,sha256=RmN3B0qtVixE_7QfgxsJHj2MUPOEdp8oxrcFN2oSR5Q,2261
 onnxslim/core/pattern/elimination/reshape.py,sha256=XwvuPAZnXCCEwJb2n1guigstnsl3wlxGygytH3GZXN8,3109
 onnxslim/core/pattern/elimination/reshape_as.py,sha256=FI3LYR0pzbp2pDmaX13duHrQ4uqwaKNu4bG78en-7wY,2034
-onnxslim/core/pattern/elimination/slice.py,sha256=moZibU-TbtdwtmGIUwyjnjf3oRCeCBcQq0M1gY5ZWDk,5033
+onnxslim/core/pattern/elimination/slice.py,sha256=aOfxc7h4mottkK78gq8qoKYtLWBwnxoa7lnY1Z15hSc,5547
 onnxslim/core/pattern/elimination/unsqueeze.py,sha256=v7Rin3qB6F49ETrxXWEQQxUgtlF18nvHb6JFarf0kwQ,3855
 onnxslim/core/pattern/fusion/__init__.py,sha256=3ajHvRurL7WHL4tfNsBoLQh6Sq2fyiqH-VsPuftYMGg,183
-onnxslim/core/pattern/fusion/concat_reshape.py,sha256=LvknixTAsSUqUkGSuoEA1QpC-TmBrsx6AHZoeT0gTbI,1615
-onnxslim/core/pattern/fusion/convadd.py,sha256=ONORwlZbQ1kYJVAnCyGY6KLIicOOELmKm7-l2vbe078,3245
-onnxslim/core/pattern/fusion/convbn.py,sha256=ZsVDuAxe41f_eN9rt2psJLKQyzGMjO2RCcX9FKRNM1Y,4118
-onnxslim/core/pattern/fusion/convmul.py,sha256=aqq2fMtnMt7cXgQxdwu2hIk2kl-SI7FwpyCxtt9lT1w,3380
+onnxslim/core/pattern/fusion/concat_reshape.py,sha256=9q1cPpOpO7s87k0r9qUFuLLMuTGJXOkOX3l7Xl1KiAQ,1685
+onnxslim/core/pattern/fusion/convadd.py,sha256=4nOB6OGbKIBaM2nlxSdnOP_Ayer-1O7hu_hdaXVzF8M,3082
+onnxslim/core/pattern/fusion/convbn.py,sha256=e8EXGSWmlBFrM1tkTTZIXaLwSXh82V3XKie4D2cm1nY,3944
+onnxslim/core/pattern/fusion/convmul.py,sha256=2QbbqxtzATXZMsCtcP4EcZQ1vj8Rb1yFFSiC72zG22Q,3335
 onnxslim/core/pattern/fusion/gelu.py,sha256=uR67AJ_tL1gboY6VsTdqajHxW3Pbu656UMhCe1mQZDY,1469
-onnxslim/core/pattern/fusion/gemm.py,sha256=Ti9yZAfEprFRvW1FiAD0zvewELOJbRjposIk3yjjXfQ,12928
+onnxslim/core/pattern/fusion/gemm.py,sha256=-Fdp3FkD54Kw1yC-2FXQ1NzaSvr4IRxmR7ObL5_cJTI,13035
 onnxslim/core/pattern/fusion/padconv.py,sha256=oF-Z4tlyu-AAWJMQDoszNITNgd2mb0vAg2gi0RwQuMo,3838
 onnxslim/core/pattern/fusion/reduce.py,sha256=dMC7CPlFglrJxugsJWjcc-jQCIa_GIbW1y9K2FRvvcE,2755
 onnxslim/core/shape_inference/__init__.py,sha256=iMAX6y6LsR8S3DOpeshPaMQLS3Plj4zYBdSaLGRYIts,16833
@@ -172,7 +172,7 @@ onnxslim/third_party/onnx_graphsurgeon/importers/base_importer.py,sha256=ESIul1p
 onnxslim/third_party/onnx_graphsurgeon/importers/onnx_importer.py,sha256=qa86Ne8yWCmpoAPBWV2lV1hlCvnQ6UPe-M1JXSfnMqM,23097
 onnxslim/third_party/onnx_graphsurgeon/ir/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 onnxslim/third_party/onnx_graphsurgeon/ir/function.py,sha256=X1Rd1ZQlHhK6crg788a-LCmQSzv446LGfw376_Cz8Co,11820
-onnxslim/third_party/onnx_graphsurgeon/ir/graph.py,sha256=RU1luTR5sGMbPbRXtKGsYtBIv5BlXZOo7gU6bv0L5FY,70494
+onnxslim/third_party/onnx_graphsurgeon/ir/graph.py,sha256=BEHXQoQMYclhEld5_o2MeA9zgpPZECfe6J9VenhiPgk,66101
 onnxslim/third_party/onnx_graphsurgeon/ir/node.py,sha256=lHrJCNRhtPRZrE7vuvQkG_wfEsJzDW7Wf-T_kr4OJHI,9996
 onnxslim/third_party/onnx_graphsurgeon/ir/tensor.py,sha256=bypjlsVp1qByPhJRbTSjSrPpoatmMykjnJ9_cnnmz9Y,19265
 onnxslim/third_party/onnx_graphsurgeon/logger/__init__.py,sha256=b6lAvvrKZKNtCZOgcvz2Aj9lUO5mw5JM8UFP5BqBOnQ,83
@@ -180,8 +180,8 @@ onnxslim/third_party/onnx_graphsurgeon/logger/logger.py,sha256=L12rrwn33RHH-2WLv
 onnxslim/third_party/onnx_graphsurgeon/util/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 onnxslim/third_party/onnx_graphsurgeon/util/exception.py,sha256=KrsHbKEQ4237UbjlODsUzvkXoAY72LZi23ApBeFANWg,786
 onnxslim/third_party/onnx_graphsurgeon/util/misc.py,sha256=kyxInD2SCRLU4wHMeiDEYEHB3871fGks6kQTuF9uATY,8960
-onnxslim-0.1.83.dist-info/METADATA,sha256=Npm1SQ2CnsjAh0NF6Z5twoqjiu9IJLfrLRh4KkvEALo,10651
-onnxslim-0.1.83.dist-info/WHEEL,sha256=WLgqFyCfm_KASv4WHyYy0P3pM_m7J5L9k2skdKLirC8,87
-onnxslim-0.1.83.dist-info/entry_points.txt,sha256=O2QgceCVeGeRhnxRSDRcGiFd0ZNfElwrTiRo1W2V7KA,47
-onnxslim-0.1.83.dist-info/licenses/LICENSE,sha256=oHZXw-yrBwdNVGu4JtlZhMgmQHKIZ7BJJlJdhu1HKvI,1062
-onnxslim-0.1.83.dist-info/RECORD,,
+onnxslim-0.1.84.dist-info/METADATA,sha256=ZoGC6wTTau3dqyYbSA8rtLL8ghV8TV5KwDXSmZw9yjo,10651
+onnxslim-0.1.84.dist-info/WHEEL,sha256=WLgqFyCfm_KASv4WHyYy0P3pM_m7J5L9k2skdKLirC8,87
+onnxslim-0.1.84.dist-info/entry_points.txt,sha256=O2QgceCVeGeRhnxRSDRcGiFd0ZNfElwrTiRo1W2V7KA,47
+onnxslim-0.1.84.dist-info/licenses/LICENSE,sha256=oHZXw-yrBwdNVGu4JtlZhMgmQHKIZ7BJJlJdhu1HKvI,1062
+onnxslim-0.1.84.dist-info/RECORD,,