onnx-diagnostic 0.8.4__py3-none-any.whl → 0.8.5__py3-none-any.whl

onnx_diagnostic/__init__.py

@@ -3,5 +3,5 @@ Patches, Investigates onnx models.
 Functions, classes to dig into a model when this one is right, slow, wrong...
 """
 
-__version__ = "0.8.4"
+__version__ = "0.8.5"
 __author__ = "Xavier Dupré"

onnx_diagnostic/_command_lines_parser.py

@@ -198,15 +198,19 @@ def get_parser_print() -> ArgumentParser:
     )
     parser.add_argument(
         "fmt",
-        choices=["pretty", "raw", "text", "printer"],
+        choices=["dot", "pretty", "printer", "raw", "shape", "text"],
         default="pretty",
         help=textwrap.dedent(
             """
             Prints out a model on the standard output.
-            raw     - just prints the model with print(...)
-            printer - onnx.printer.to_text(...)
+
+            dot     - converts the graph into dot
             pretty  - an improved rendering
+            printer - onnx.printer.to_text(...)
+            raw     - just prints the model with print(...)
+            shape   - prints every node node with input and output shapes
             text    - uses GraphRendering
+
             """.strip(
                 "\n"
             )
@@ -232,6 +236,14 @@ def _cmd_print(argv: List[Any]):
         from .helpers.graph_helper import GraphRendering
 
         print(GraphRendering(onx).text_rendering())
+    elif args.fmt == "shape":
+        from experimental_experiment.xbuilder import GraphBuilder
+
+        print(GraphBuilder(onx).pretty_text())
+    elif args.fmt == "dot":
+        from .helpers.dot_helper import to_dot
+
+        print(to_dot(onx))
     else:
         raise ValueError(f"Unexpected value fmt={args.fmt!r}")
 
@@ -517,12 +529,12 @@ def get_parser_validate(name: str = "validate") -> ArgumentParser:
         nargs="*",
         help=textwrap.dedent(
             """
-        Applies patches before exporting, it can be a boolean
-        to enable to disable the patches or be more finetuned
-        (default is True). It is possible to disable patch for torch
-        by adding:
-            --patch "patch_sympy=False" --patch "patch_torch=False"
-        """.strip(
+            Applies patches before exporting, it can be a boolean
+            to enable to disable the patches or be more finetuned
+            (default is True). It is possible to disable patch for torch
+            by adding:
+                --patch "patch_sympy=False" --patch "patch_torch=False"
+            """.strip(
                 "\n"
             )
         ),

onnx_diagnostic/export/api.py

@@ -64,6 +64,7 @@ def to_onnx(
     exporter_kwargs: Optional[Dict[str, Any]] = None,
     save_ep: Optional[str] = None,
     optimize: bool = True,
+    optimizer_for_ort: bool = True,
     use_control_flow_dispatcher: bool = False,
     onnx_plugs: Optional[List[EagerDirectReplacementWithOnnx]] = None,
     inline: bool = True,
@@ -88,6 +89,7 @@ def to_onnx(
     :param exporter_kwargs: additional parameters sent to the exporter
     :param save_ep: saves the exported program
     :param optimize: optimizes the model
+    :param optimizer_for_ort: optimizes the model for onnxruntime
     :param use_control_flow_dispatcher: use the dispatcher created to supported
         custom loops (see :func:`onnx_diagnostic.export.control_flow_onnx.loop_for_onnx`)
     :param onnx_plugs: the code was modified to replace some parts with onnx translation
@@ -126,8 +128,10 @@ def to_onnx(
         options = None
         if exporter_kwargs is not None:
             options = exporter_kwargs.pop("options", None)
-        if options is None:
-            options = OptimizationOptions(patterns="default+onnxruntime")
+        if options is None and optimize:
+            options = OptimizationOptions(
+                patterns="default+onnxruntime" if optimizer_for_ort else "default"
+            )
         main_dispatcher = (
             get_main_dispatcher(use_control_flow_dispatcher, onnx_plugs)
             if onnx_plugs or use_control_flow_dispatcher
@@ -161,6 +165,9 @@ def to_onnx(
         assert (
             not output_dynamic_shapes
         ), f"output_dynamic_shapes not supported for exporter={exporter!r}"
+        assert (
+            optimize
+        ), f"torch.onnx.export always optimizes the model but optimize={optimize}"
         custom_translation_table = {}
         if onnx_plugs:
             for plug in onnx_plugs:
@@ -180,7 +187,7 @@ def to_onnx(
             custom_translation_table=custom_translation_table,
             **(exporter_kwargs or {}),
         )
-        if not inline and optimize:
+        if not inline and optimize and optimizer_for_ort:
             ort_fusions.optimize_for_ort(epo.model)
 
         if onnx_plugs:
@@ -207,7 +214,7 @@ def to_onnx(
                 common_passes.InlinePass()(epo.model)
                 common_passes.RemoveUnusedOpsetsPass()(epo.model)
 
-        if inline and optimize:
+        if inline and optimize and optimizer_for_ort:
             ort_fusions.optimize_for_ort(epo.model)
         if filename:
             epo.save(filename, external_data=True)
@@ -232,6 +239,10 @@ def to_onnx(
             f"Only a specified set of inputs is supported for exporter={exporter!r}, "
             f"but it is {list(kwargs)}"  # type: ignore[arg-type]
         )
+        assert optimizer_for_ort and optimize, (
+            f"ModelBuilder only produces model optimized for onnxruntime but "
+            f"optimizer_for_ort={optimizer_for_ort} and optimize={optimize}"
+        )
         flat_inputs = flatten_object(kwargs, drop_keys=True)
         first = flat_inputs[0]
         first_float = [

onnx_diagnostic/export/onnx_plug.py

@@ -128,7 +128,61 @@ class EagerDirectReplacementWithOnnx:
 
         print(pretty_onnx(onx))
 
-        # And with :func:`torch.onnx.export`:
+    We do the same with :func:`torch.onnx.export`:
+
+    .. runpython::
+        :showcode:
+
+        import onnx.helper as oh
+        import torch
+        from onnx_diagnostic.helpers.onnx_helper import pretty_onnx
+        from onnx_diagnostic.export.onnx_plug import EagerDirectReplacementWithOnnx
+        from onnx_diagnostic.export.api import to_onnx
+        from onnx_diagnostic.torch_export_patches.patch_inputs import use_dyn_not_str
+
+
+        def demo_customsub(x, y):
+            return x - y
+
+
+        def demo_customsub_shape(x, y):
+            return torch.empty(torch.broadcast_shapes(x.shape, y.shape), dtype=x.dtype)
+
+
+        def make_function_proto():
+            return oh.make_function(
+                "onnx_plug",
+                "demo_customsub",
+                ["x", "y"],
+                ["z"],
+                [oh.make_node("Sub", ["x", "y"], ["z"])],
+                opset_imports=[oh.make_opsetid("", 22)],
+            )
+
+
+        class Model(torch.nn.Module):
+            def forward(self, x):
+                y = x.sum(axis=1, keepdim=True)
+                d = torch.ops.onnx_plug.demo_customsub(x, y)
+                return torch.abs(d)
+
+
+        replacements = [
+            EagerDirectReplacementWithOnnx(
+                demo_customsub, demo_customsub_shape, make_function_proto(), 2, 1
+            )
+        ]
+
+        x = torch.randn((3, 4), dtype=torch.float32)
+        model = Model()
+        ds = ({0: "d1", 1: "d2"},)
+
+        # The exported program shows a custom op.
+        ep = torch.export.export(model, (x,), dynamic_shapes=use_dyn_not_str(ds))
+        print("ep")
+
+        # As the exporter knows how the replace this custom op.
+        # Let's export.
 
         onx = to_onnx(
             model,
@@ -152,8 +206,8 @@ class EagerDirectReplacementWithOnnx:
             dtype = first_tensor.dtype
             itype = torch_dtype_to_onnx_dtype(dtype)
             if dtype == torch.float32:
-                if opset >= 24:
-                    return "LOOPA24", itype
+                if opset >= 23:
+                    return "LOOPA23", itype
                 return "LOOPMHA", itype
             if dtype == torch.float16:
                 if first_tensor.is_cuda:
@@ -175,9 +229,9 @@ class EagerDirectReplacementWithOnnx:
                 ("PACKED", onnx.TensorProto.FLOAT16): _add_com_microsoft_opset(
                     PackedAttention.to_function_proto()
                 ),
-                ("LOOPA24", onnx.TensorProto.FLOAT): LoopAttention24.to_function_proto(),
-                ("LOOPA24", onnx.TensorProto.FLOAT16): _update_sequence_type(
-                    onnx.TensorProto.FLOAT16, LoopAttention24.to_function_proto()
+                ("LOOPA23", onnx.TensorProto.FLOAT): LoopAttention23.to_function_proto(),
+                ("LOOPA23", onnx.TensorProto.FLOAT16): _update_sequence_type(
+                    onnx.TensorProto.FLOAT16, LoopAttention23.to_function_proto()
                 ),
                 ("LOOPMHA", onnx.TensorProto.FLOAT): _add_com_microsoft_opset(
                     LoopMHAAttention.to_function_proto()

onnx_diagnostic/helpers/helper.py

@@ -2,6 +2,7 @@ import ast
 import enum
 import inspect
 import itertools
+import json
 from dataclasses import is_dataclass, fields
 from typing import Any, Callable, Dict, List, Optional, Set, Tuple, Union
 import numpy as np
@@ -1373,11 +1374,7 @@ def max_diff(
         if hist:
             if isinstance(hist, bool):
                 hist = np.array([0, 0.0001, 0.001, 0.01, 0.1, 1, 10, 100], dtype=diff.dtype)
-            ind = np.digitize(diff.reshape((-1,)), hist, right=True)
-            cou = np.bincount(ind, minlength=ind.shape[0] + 1)
-            res["rep"] = dict(
-                zip([f">{x}" for x in hist], [int(i) for i in (cou.sum() - np.cumsum(cou))])
-            )
+            res["rep"] = {f">{h}": (diff > h).sum().item() for h in hist}
         return res  # type: ignore
 
     if isinstance(expected, torch.Tensor) and isinstance(got, torch.Tensor):
@@ -1493,27 +1490,11 @@ def max_diff(
             dev=dev,
         )
         if hist:
-            if isinstance(hist, list) and len(hist) == 1:
-                res["rep"] = {f">{hist[0]}": (diff > hist[0]).sum().item()}
-            elif isinstance(hist, list) and len(hist) == 2:
-                res["rep"] = {
-                    f">{hist[0]}": (diff > hist[0]).sum().item(),
-                    f">{hist[1]}": (diff > hist[1]).sum().item(),
-                }
-            else:
-                if isinstance(hist, bool):
-                    hist = torch.tensor(
-                        [0, 0.0001, 0.001, 0.01, 0.1, 1, 10, 100], dtype=diff.dtype
-                    )
-                hist = torch.tensor(hist).to(diff.device)
-                ind = torch.bucketize(diff.reshape((-1,)), hist, right=False)
-                cou = torch.bincount(ind, minlength=ind.shape[0] + 1)
-                res["rep"] = dict(
-                    zip(
-                        [f">{x}" for x in hist],
-                        [int(i) for i in (cou.sum() - torch.cumsum(cou, 0))],
-                    )
+            if isinstance(hist, bool):
+                hist = torch.tensor(
+                    [0, 0.0001, 0.001, 0.01, 0.1, 1, 10, 100], dtype=diff.dtype
                 )
+            res["rep"] = {f">{h}": (diff > h).sum().item() for h in hist}
         return res  # type: ignore
 
     if isinstance(expected, int) and isinstance(got, torch.Tensor):
@@ -1750,8 +1731,26 @@ def max_diff(
     )
 
 
-def string_diff(diff: Dict[str, Any]) -> str:
-    """Renders discrepancies return by :func:`max_diff` into one string."""
+def string_diff(diff: Dict[str, Any], js: bool = False, ratio: bool = False, **kwargs) -> str:
+    """
+    Renders discrepancies return by :func:`max_diff` into one string.
+
+    :param diff: differences
+    :param js: json format
+    :param ratio: display mismatch ratio
+    :param kwargs: addition values to add in the json format
+    """
+    if js:
+        if "rep" in diff:
+            rep = diff["rep"]
+            diff = {**{k: v for k, v in diff.items() if k != "rep"}, **rep}
+            if ratio:
+                for k, v in rep.items():
+                    diff[f"%{k}"] = v / diff["n"]
+                diff["mean"] = diff["sum"] / diff["n"]
+            diff.update(kwargs)
+        return json.dumps(diff)
+
     # dict(abs=, rel=, sum=, n=n_diff, dnan=)
     if "dev" in diff:
         ddiff = {k: v for k, v in diff.items() if k != "dev"}

onnx_diagnostic/torch_export_patches/patches/_patch_transformers_masking_utils.py

@@ -1,3 +1,4 @@
+import inspect
 from typing import Callable, List, Optional, Tuple
 import torch
 
@@ -19,6 +20,12 @@ if patch_masking_utils:
         prepare_padding_mask,
     )
 
+    _prepare_padding_mask_kwargs = (
+        dict(_slice=False)
+        if "_slice" in inspect.signature(prepare_padding_mask).parameters
+        else {}
+    )
+
     try:
         # transformers>=5.0
         from transformers.masking_utils import (
@@ -132,7 +139,9 @@ if patch_masking_utils:
     ) -> Optional[torch.Tensor]:
         """manual patch for function ``transformers.masking_utils.sdpa_mask_recent_torch``."""
         q_length = cache_position.shape[0]
-        padding_mask = prepare_padding_mask(attention_mask, kv_length, kv_offset, _slice=False)
+        padding_mask = prepare_padding_mask(
+            attention_mask, kv_length, kv_offset, **_prepare_padding_mask_kwargs
+        )
         if allow_is_causal_skip and _ignore_causal_mask_sdpa(
             padding_mask, q_length, kv_length, kv_offset, local_size
         ):

onnx_diagnostic/torch_export_patches/patches/_patch_transformers_qwen2_5.py

@@ -24,7 +24,7 @@ if patch_qwen2_5:
 
     onnx_plugs_op = onnxscript.values.Opset("onnx_plug", 1)
     op = onnxscript.opset22
-    op24 = onnxscript.onnx_opset.opset24
+    op23 = onnxscript.onnx_opset.opset23
     msft_op = onnxscript.values.Opset("com.microsoft", 1)
     STOPAT = (
         int(os.environ.get("STOPAT", None))
@@ -101,7 +101,7 @@ if patch_qwen2_5:
         return attn_output_4d
 
     @onnxscript.script(opset=onnx_plugs_op)
-    def LoopAttention24(
+    def LoopAttention23(
         query_states,
         key_states,
         value_states,
@@ -109,26 +109,26 @@ if patch_qwen2_5:
         scaling: float = 0.11180339887498948,
         num_heads: int = 16,
     ):
-        to_3d_shape = op24.Constant(value_ints=[0, 0, -1])
-        query_transposed = op24.Transpose(query_states, perm=[0, 2, 1, 3])
-        output_shape = op24.Shape(query_transposed)
-        query_3d = op24.Reshape(query_transposed, to_3d_shape)
-        value_3d = op24.Reshape(op24.Transpose(value_states, perm=[0, 2, 1, 3]), to_3d_shape)
-        key_3d = op24.Reshape(op24.Transpose(key_states, perm=[0, 2, 1, 3]), to_3d_shape)
-        cu_seqlens = op24.Cast(cu_seqlens, to=onnx.TensorProto.INT32)
-        num_patches = op24.Size(cu_seqlens) - 1
-        seq_axis = op24.Constant(value_ints=[1])
-        seq_axis_int32 = op24.Cast(seq_axis, to=onnx.TensorProto.INT32)
-        seq_attn = op24.SequenceEmpty(dtype=onnx.TensorProto.FLOAT)
+        to_3d_shape = op23.Constant(value_ints=[0, 0, -1])
+        query_transposed = op23.Transpose(query_states, perm=[0, 2, 1, 3])
+        output_shape = op23.Shape(query_transposed)
+        query_3d = op23.Reshape(query_transposed, to_3d_shape)
+        value_3d = op23.Reshape(op23.Transpose(value_states, perm=[0, 2, 1, 3]), to_3d_shape)
+        key_3d = op23.Reshape(op23.Transpose(key_states, perm=[0, 2, 1, 3]), to_3d_shape)
+        cu_seqlens = op23.Cast(cu_seqlens, to=onnx.TensorProto.INT32)
+        num_patches = op23.Size(cu_seqlens) - 1
+        seq_axis = op23.Constant(value_ints=[1])
+        seq_axis_int32 = op23.Cast(seq_axis, to=onnx.TensorProto.INT32)
+        seq_attn = op23.SequenceEmpty(dtype=onnx.TensorProto.FLOAT)
         for i_patch in range(num_patches):
-            i_1d = op24.Reshape(i_patch, [1])
+            i_1d = op23.Reshape(i_patch, [1])
             i_plus_1_1d = i_1d + 1
-            start = op24.Gather(cu_seqlens, i_1d, axis=0)
-            end = op24.Gather(cu_seqlens, i_plus_1_1d, axis=0)
-            query_i = op24.Slice(query_3d, start, end, seq_axis_int32)
-            key_i = op24.Slice(key_3d, start, end, seq_axis_int32)
-            value_i = op24.Slice(value_3d, start, end, seq_axis_int32)
-            mha_output = op24.Attention(
+            start = op23.Gather(cu_seqlens, i_1d, axis=0)
+            end = op23.Gather(cu_seqlens, i_plus_1_1d, axis=0)
+            query_i = op23.Slice(query_3d, start, end, seq_axis_int32)
+            key_i = op23.Slice(key_3d, start, end, seq_axis_int32)
+            value_i = op23.Slice(value_3d, start, end, seq_axis_int32)
+            mha_output = op23.Attention(
                 query_i,
                 key_i,
                 value_i,
@@ -137,9 +137,9 @@ if patch_qwen2_5:
                 kv_num_heads=num_heads,
                 softmax_precision=onnx.TensorProto.FLOAT,
             )
-            seq_attn = op24.SequenceInsert(seq_attn, mha_output)
-        attn_output = op24.ConcatFromSequence(seq_attn, axis=1)
-        attn_output_4d = op24.Reshape(attn_output, output_shape)
+            seq_attn = op23.SequenceInsert(seq_attn, mha_output)
+        attn_output = op23.ConcatFromSequence(seq_attn, axis=1)
+        attn_output_4d = op23.Reshape(attn_output, output_shape)
         return attn_output_4d
 
     @onnxscript.script(opset=onnx_plugs_op)
@@ -263,8 +263,8 @@ if patch_qwen2_5:
         if strategy is not None:
             return strategy, itype
         if dtype == torch.float32 or itype == onnx.TensorProto.FLOAT:
-            if opset >= 24:
-                return "LOOPA24", itype
+            if opset >= 23:
+                return "LOOPA23", itype
             return "LOOPMHA", itype
         if dtype == torch.float16 or itype == onnx.TensorProto.FLOAT16:
             # first_tensor may be a SymbolicTensor (onnx).
@@ -288,9 +288,9 @@ if patch_qwen2_5:
             ("PACKED", onnx.TensorProto.FLOAT16): _add_com_microsoft_opset(
                 PackedAttention.to_function_proto()
             ),
-            ("LOOPA24", onnx.TensorProto.FLOAT): LoopAttention24.to_function_proto(),
-            ("LOOPA24", onnx.TensorProto.FLOAT16): _update_sequence_type(
-                onnx.TensorProto.FLOAT16, LoopAttention24.to_function_proto()
+            ("LOOPA23", onnx.TensorProto.FLOAT): LoopAttention23.to_function_proto(),
+            ("LOOPA23", onnx.TensorProto.FLOAT16): _update_sequence_type(
+                onnx.TensorProto.FLOAT16, LoopAttention23.to_function_proto()
             ),
             ("LOOPMHA", onnx.TensorProto.FLOAT): _add_com_microsoft_opset(
                 LoopMHAAttention.to_function_proto()

{onnx_diagnostic-0.8.4.dist-info → onnx_diagnostic-0.8.5.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: onnx-diagnostic
-Version: 0.8.4
+Version: 0.8.5
 Summary: Tools to help converting pytorch models into ONNX.
 Home-page: https://github.com/sdpython/onnx-diagnostic
 Author: Xavier Dupré

{onnx_diagnostic-0.8.4.dist-info → onnx_diagnostic-0.8.5.dist-info}/RECORD

@@ -1,16 +1,16 @@
-onnx_diagnostic/__init__.py,sha256=BuvSD4fYmz8ZVadwiG4GCHeb46p5sNX7-_GM16OtKW0,173
+onnx_diagnostic/__init__.py,sha256=dCiBK_S7EOo_rAsmsgv-laLhtKzE2uny0XIR5aO4eDk,173
 onnx_diagnostic/__main__.py,sha256=YmyV_Aq_ianDlHyKLHMa6h8YK3ZmFPpLVHLKjM91aCk,79
-onnx_diagnostic/_command_lines_parser.py,sha256=Xkh_7fIDbT5ghTpLqlVhx4cIxAUXqv6zvPdnN3aCdOY,52254
+onnx_diagnostic/_command_lines_parser.py,sha256=ZFJdQP1Ee8D5a_xUch-0CHaYbbILztejTjVdyc9KrMw,52667
 onnx_diagnostic/api.py,sha256=BhCl_yCd78N7TlVtPOHjeYv1QBEy39TjZ647rcHqLh0,345
 onnx_diagnostic/doc.py,sha256=t3RELgfooYnVMAi0JSpggWkQEgUsREz8NmRvn0TnLI8,2829
 onnx_diagnostic/ext_test_case.py,sha256=rVZWqFEfnvwnsD3wF4jeDblh5uj5ckZ8C6DZQ0RGb_E,49599
 onnx_diagnostic/export/__init__.py,sha256=yEIoWiOeTwBsDhyYt2fTKuhtA0Ya1J9u9ZzMTOTWaWs,101
-onnx_diagnostic/export/api.py,sha256=c3ZASq2upUAoiQ4aymm8vOAEySN_Yk6l0o1hWf6Ailo,10065
+onnx_diagnostic/export/api.py,sha256=BX4c99gMlRYsBWk3P15FMRogArxjP4dXYXP5gILjgIk,10626
 onnx_diagnostic/export/control_flow.py,sha256=zU5n_QYhNcBllyMsl1_i6ohZt2CshqG2MokJghrvA60,7751
 onnx_diagnostic/export/control_flow_onnx.py,sha256=sODOD4v7EJj6LWhrfcdCW68r9nYKsRM4SRnqDw4TrSI,18049
 onnx_diagnostic/export/control_flow_research.py,sha256=RuYz9_eM42Bk6TKSiPV6dS68LIMZu-6WBCFCKoSvjrk,5422
 onnx_diagnostic/export/dynamic_shapes.py,sha256=M2hlpHSTbkzZwGKAbrpQXng5HQrwjF5Z6wGGxEgnp74,42061
-onnx_diagnostic/export/onnx_plug.py,sha256=WqqdTBk2pV26AplNQysIdhR9y3ZFdQ-5KXu5ogTNcgI,21053
+onnx_diagnostic/export/onnx_plug.py,sha256=U13fL0BjnhMzcDGxaAOqM4TQte5Z4zKDg4ESS0iktjM,22704
 onnx_diagnostic/export/shape_helper.py,sha256=m628y0oRCQbeZkeh8JDHIfWMsSjoJoeX-IPiPGDHT-w,11273
 onnx_diagnostic/export/validate.py,sha256=_PGUql2DJhIgGKo0WjTGUc5AgsZUx8fEs00MePy-w98,6043
 onnx_diagnostic/helpers/__init__.py,sha256=GJ2GT7cgnlIveVUwMZhuvUwidbTJaKv8CsSIOpZDsJg,83
@@ -23,7 +23,7 @@ onnx_diagnostic/helpers/doc_helper.py,sha256=pl5MZd3_FaE8BqQnqoBuSBxoNCFcd2OJd3e
 onnx_diagnostic/helpers/dot_helper.py,sha256=hwgTJsbsUv0qq7euyPDnc1NsBZDGOwv32JXSZxIHJkE,8118
 onnx_diagnostic/helpers/fake_tensor_helper.py,sha256=J7wnK3WTuVKnYiMzLVTAPkdJr3hQfIfMC9ZlOu7oGmI,11024
 onnx_diagnostic/helpers/graph_helper.py,sha256=hevQT5a7_QuriVPQcbT5qe18n99Doyl5h3-qshx1-uk,14093
-onnx_diagnostic/helpers/helper.py,sha256=f5w53QR0IO1-zqAMacgeGpZNA8uAo0c2k_ZYXP_BRhE,65840
+onnx_diagnostic/helpers/helper.py,sha256=x8EYQmgrz_G5QS_IsbeFIoDcN_sUs-CslJMHseBj1Fw,65482
 onnx_diagnostic/helpers/log_helper.py,sha256=0lJiTF87lliI-LmgpUH_V2N8NuzJ0LryH0mSYpkRaL8,93272
 onnx_diagnostic/helpers/memory_peak.py,sha256=M3m4_thWFIwP5HytbJYEqaijXIv5v5BW_vlcJowIYI4,6434
 onnx_diagnostic/helpers/mini_onnx_builder.py,sha256=jR2lkRZEQ0N30H0FqeBwaxJd_w_6kyxFagrnulqFjhE,23883
@@ -117,9 +117,9 @@ onnx_diagnostic/torch_export_patches/patches/_patch_transformers_dynamic_cache.p
 onnx_diagnostic/torch_export_patches/patches/_patch_transformers_gemma3.py,sha256=nVgYQk0xXpHiictN1wOHVMN2lTH9b0vfIJ4ie-uKopg,1999
 onnx_diagnostic/torch_export_patches/patches/_patch_transformers_generation_mixin.py,sha256=VIZsVHgR8NmAcBQalPl5I6ZzNgcBxjGb6ars31m9gRg,21936
 onnx_diagnostic/torch_export_patches/patches/_patch_transformers_idefics.py,sha256=kTjuTRsfkGGGhspJnMxAMQSchZgGC_IruJzpHh_FmI8,6348
-onnx_diagnostic/torch_export_patches/patches/_patch_transformers_masking_utils.py,sha256=R4YwnN9ktxjjImiJtLRxiKtKLr9LuFlwkPXkTJ6BTIo,6895
+onnx_diagnostic/torch_export_patches/patches/_patch_transformers_masking_utils.py,sha256=HE3fovyvMiYe9EPz1UjdD9AWopX3H188SMwPb8w5mzM,7111
 onnx_diagnostic/torch_export_patches/patches/_patch_transformers_qwen2.py,sha256=OxYdlLrwtd_KGHt3E17poduxvWFg-CfGS57-yN1i6gI,3827
-onnx_diagnostic/torch_export_patches/patches/_patch_transformers_qwen2_5.py,sha256=icjSrI3LFrShtV_AYQ8F2qiMFHZ74Qg5I2c-V23uEgg,31601
+onnx_diagnostic/torch_export_patches/patches/_patch_transformers_qwen2_5.py,sha256=GS7IDHyRaLAsbZE5k7KN-ZT5-ezbmEUzXPJ_xG4SulA,31601
 onnx_diagnostic/torch_export_patches/patches/_patch_transformers_qwen3.py,sha256=cND9Iqo1aKdlX-BXGr9Qlq_Y4EW1L5VWSwZfqYTVazU,4888
 onnx_diagnostic/torch_export_patches/patches/_patch_transformers_rotary_embedding.py,sha256=4bJ_z2gizZQla_fcCVt0dmuhzO9Vu-D7CCMWdxMlrKM,16893
 onnx_diagnostic/torch_export_patches/patches/_patch_transformers_sam_mask_decoder.py,sha256=-6TuBm3sLAFEGuW3vRfOTtE5uP6aINFfu7xMnl27Dws,5703
@@ -146,8 +146,8 @@ onnx_diagnostic/torch_onnx/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJ
 onnx_diagnostic/torch_onnx/runtime_info.py,sha256=u1bD6VXqzBCRmqmbzQtDswaPs1PH_ygr1r-CrcfXpNU,8562
 onnx_diagnostic/torch_onnx/sbs.py,sha256=8okBEIupMgw7TtKc80YFimMtwnY3GchdY05FsA9ooa0,40749
 onnx_diagnostic/torch_onnx/sbs_dataclasses.py,sha256=UctdBjzoPTQG1LS0tZ8A6E9hpoq5HWUYaJLPOPJc9FI,20299
-onnx_diagnostic-0.8.4.dist-info/licenses/LICENSE.txt,sha256=Vv6TXglX6Rc0d-f8aREhayhT-6PMQXEyOmI2NKlUCMc,1045
-onnx_diagnostic-0.8.4.dist-info/METADATA,sha256=8S9bFx2lTef7dTeL_dTCtGj1MalIyoUvs5dMzrMffNg,6734
-onnx_diagnostic-0.8.4.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-onnx_diagnostic-0.8.4.dist-info/top_level.txt,sha256=KwNkXewmcobM3ZT1DJLVWH6ebJzA5qKg7cWqKfpGNT4,16
-onnx_diagnostic-0.8.4.dist-info/RECORD,,
+onnx_diagnostic-0.8.5.dist-info/licenses/LICENSE.txt,sha256=Vv6TXglX6Rc0d-f8aREhayhT-6PMQXEyOmI2NKlUCMc,1045
+onnx_diagnostic-0.8.5.dist-info/METADATA,sha256=A54IonPIcnualwiRJhvjRMfhF3p3jdXhEH1vTtZBgyE,6734
+onnx_diagnostic-0.8.5.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+onnx_diagnostic-0.8.5.dist-info/top_level.txt,sha256=KwNkXewmcobM3ZT1DJLVWH6ebJzA5qKg7cWqKfpGNT4,16
+onnx_diagnostic-0.8.5.dist-info/RECORD,,