onnx-diagnostic 0.7.3__py3-none-any.whl → 0.7.4__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__ = "0.7.3"
+__version__ = "0.7.4"
 __author__ = "Xavier Dupré"

onnx_diagnostic/_command_lines_parser.py

@@ -349,6 +349,15 @@ def get_parser_validate() -> ArgumentParser:
             python -m onnx_diagnostic validate -m microsoft/Phi-4-mini-reasoning \\
                 --run -v 1 -o dump_test --no-quiet --repeat 2 --warmup 2 \\
                 --dtype float16 --device cuda --export modelbuilder
+
+            position_ids is usually not needed, they can be removed by adding:
+
+            --drop position_ids
+
+            The behaviour may be modified compare the original configuration,
+            the following argument can be rope_scaling to dynamic:
+
+            --mop \"rope_scaling={'rope_type': 'dynamic', 'factor': 10.0}\""
             """
         ),
         formatter_class=RawTextHelpFormatter,
@@ -403,10 +412,12 @@ def get_parser_validate() -> ArgumentParser:
     )
     parser.add_argument(
         "--inputs2",
-        default=True,
-        action=BooleanOptionalAction,
+        default=1,
+        type=int,
         help="Validates the model on a second set of inputs\n"
-        "to check the exported model supports dynamism.",
+        "to check the exported model supports dynamism. The values is used "
+        "as an increment to the first set of inputs. A high value may trick "
+        "a different behavior in the model and missed by the exporter.",
     )
     parser.add_argument(
         "--runtime",
@@ -422,7 +433,8 @@ def get_parser_validate() -> ArgumentParser:
     parser.add_argument(
         "--drop",
         help="Drops the following inputs names, it should be a list\n"
-        "with comma separated values.",
+        "with comma separated values, example:\n"
+        "--drop position_ids",
     )
     parser.add_argument(
         "--opset",

onnx_diagnostic/export/shape_helper.py

@@ -30,6 +30,77 @@ def all_dynamic_shape_from_inputs(inputs: Any, dim_prefix: Any = "d") -> Any:
         )
         ds = all_dynamic_shape_from_inputs(inputs)
         pprint.pprint(ds)
+
+    For this function to work, patches must be enabled if :epkg:`transformers`
+    does not implement the serialization functions.
+
+    .. runpython::
+        :showcode:
+
+        import pprint
+        import torch
+        from onnx_diagnostic.helpers.cache_helper import (
+            make_dynamic_cache,
+            make_encoder_decoder_cache,
+            make_mamba_cache,
+            make_sliding_window_cache,
+            make_static_cache,
+        )
+        from onnx_diagnostic.export.shape_helper import all_dynamic_shape_from_inputs
+        from onnx_diagnostic.torch_export_patches import torch_export_patches
+
+        caches = [
+            make_dynamic_cache(
+                [
+                    (torch.rand((4, 4, 4)), torch.rand((4, 4, 4))),
+                    (torch.rand((4, 4, 4)), torch.rand((4, 4, 4))),
+                    (torch.rand((4, 4, 4)), torch.rand((4, 4, 4))),
+                ]
+            ),
+            make_encoder_decoder_cache(
+                make_dynamic_cache(
+                    [
+                        (torch.rand((4, 4, 4)), torch.rand((4, 4, 4))),
+                        (torch.rand((4, 4, 4)), torch.rand((4, 4, 4))),
+                        (torch.rand((4, 4, 4)), torch.rand((4, 4, 4))),
+                    ]
+                ),
+                make_dynamic_cache(
+                    [
+                        (torch.rand((5, 5, 5)), torch.rand((5, 5, 5))),
+                        (torch.rand((5, 5, 5)), torch.rand((5, 5, 5))),
+                        (torch.rand((5, 5, 5)), torch.rand((5, 5, 5))),
+                    ]
+                ),
+            ),
+            make_sliding_window_cache(
+                [
+                    (torch.rand((4, 5, 6, 7)), torch.rand((4, 5, 6, 7))),
+                    (torch.rand((4, 5, 6, 7)), torch.rand((4, 5, 6, 7))),
+                    (torch.rand((4, 5, 6, 7)), torch.rand((4, 5, 6, 7))),
+                ]
+            ),
+            make_static_cache(
+                [
+                    (torch.rand((4, 5, 6, 7)), torch.rand((4, 5, 6, 7))),
+                    (torch.rand((4, 5, 6, 7)), torch.rand((4, 5, 6, 7))),
+                    (torch.rand((4, 5, 6, 7)), torch.rand((4, 5, 6, 7))),
+                ],
+                max_cache_len=15,
+            ),
+            make_mamba_cache(
+                [
+                    (torch.rand((4, 4, 4)), torch.rand((4, 4, 4))),
+                    (torch.rand((4, 4, 4)), torch.rand((4, 4, 4))),
+                    (torch.rand((4, 4, 4)), torch.rand((4, 4, 4))),
+                ]
+            ),
+        ]
+
+        with torch_export_patches(patch_transformers=True):
+            for cache in caches:
+                print(f"-- {cache.__class__.__name__}")
+                pprint.pprint(all_dynamic_shape_from_inputs(cache))
     """
     if isinstance(dim_prefix, str):
         prefixes: Set[str] = set()

onnx_diagnostic/helpers/cache_helper.py

@@ -39,11 +39,21 @@ def flatten_unflatten_for_dynamic_shapes(
         subtrees.append(value)
         start = end
     if use_dict:
-        if spec.type is dict or spec.context:
+        if spec.type is dict:
             # This a dictionary.
             return dict(zip(spec.context, subtrees))
         if spec.type is tuple:
             return tuple(subtrees)
+        if spec.type is list:
+            return list(subtrees)
+        if spec.context:
+            # This is a custom class with attributes.
+            # It is returned as a list.
+            return list(subtrees)
+        raise ValueError(
+            f"Unable to interpret spec type {spec.type} "
+            f"(type is {type(spec.type)}, context is {spec.context})."
+        )
     # This is a list.
     return subtrees
 

onnx_diagnostic/reference/ops/op_cast_like.py

@@ -1,13 +1,17 @@
 from onnx.onnx_pb import TensorProto
 from onnx.reference.op_run import OpRun
-from onnx.reference.ops.op_cast import (
-    bfloat16,
-    cast_to,
-    float8e4m3fn,
-    float8e4m3fnuz,
-    float8e5m2,
-    float8e5m2fnuz,
-)
+
+try:
+    from onnx.reference.ops.op_cast import (
+        bfloat16,
+        cast_to,
+        float8e4m3fn,
+        float8e4m3fnuz,
+        float8e5m2,
+        float8e5m2fnuz,
+    )
+except ImportError:
+    from onnx.reference.ops.op_cast import cast_to
 from ...helpers.onnx_helper import np_dtype_to_tensor_dtype
 
 

onnx_diagnostic/tasks/automatic_speech_recognition.py

@@ -33,7 +33,7 @@ def get_inputs(
     head_dim: int,
     batch_size: int = 2,
     sequence_length: int = 30,
-    add_second_input: bool = False,
+    add_second_input: int = 1,
     **kwargs,  # unused
 ):
     """
@@ -132,6 +132,9 @@ def get_inputs(
     )
     res = dict(inputs=inputs, dynamic_shapes=shapes)
     if add_second_input:
+        assert (
+            add_second_input > 0
+        ), f"Not implemented for add_second_input={add_second_input}."
         res["inputs2"] = get_inputs(
             model=model,
             config=config,
@@ -144,7 +147,8 @@ def get_inputs(
             decoder_layers=decoder_layers,
             head_dim=head_dim,
             batch_size=batch_size + 1,
-            sequence_length=sequence_length + 1,
+            sequence_length=sequence_length + add_second_input,
+            add_second_input=0,
             **kwargs,
         )["inputs"]
     return res

onnx_diagnostic/tasks/feature_extraction.py

@@ -22,7 +22,7 @@ def get_inputs(
     batch_size: int,
     sequence_length: int,
     dummy_max_token_id: int,
-    add_second_input: bool = False,
+    add_second_input: int = 1,
     **kwargs,  # unused
 ):
     """
@@ -52,12 +52,16 @@ def get_inputs(
     )
     res = dict(inputs=inputs, dynamic_shapes=shapes)
     if add_second_input:
+        assert (
+            add_second_input > 0
+        ), f"Not implemented for add_second_input={add_second_input}."
         res["inputs2"] = get_inputs(
             model=model,
             config=config,
             batch_size=batch_size + 1,
-            sequence_length=sequence_length + 1,
+            sequence_length=sequence_length + add_second_input,
             dummy_max_token_id=dummy_max_token_id,
+            add_second_input=0,
             **kwargs,
         )["inputs"]
     return res

onnx_diagnostic/tasks/fill_mask.py

@@ -22,7 +22,7 @@ def get_inputs(
     batch_size: int,
     sequence_length: int,
     dummy_max_token_id: int,
-    add_second_input: bool = False,
+    add_second_input: int = 1,
     **kwargs,  # unused
 ):
     """
@@ -54,12 +54,16 @@ def get_inputs(
     )
     res = dict(inputs=inputs, dynamic_shapes=shapes)
     if add_second_input:
+        assert (
+            add_second_input > 0
+        ), f"Not implemented for add_second_input={add_second_input}."
         res["inputs2"] = get_inputs(
             model=model,
             config=config,
             batch_size=batch_size + 1,
-            sequence_length=sequence_length + 1,
+            sequence_length=sequence_length + add_second_input,
             dummy_max_token_id=dummy_max_token_id,
+            add_second_input=0,
             **kwargs,
         )["inputs"]
     return res

onnx_diagnostic/tasks/image_classification.py

@@ -34,7 +34,7 @@ def get_inputs(
     input_channels: int,
     batch_size: int = 2,
     dynamic_rope: bool = False,
-    add_second_input: bool = False,
+    add_second_input: int = 1,
     **kwargs,  # unused
 ):
     """
@@ -75,14 +75,18 @@ def get_inputs(
         shapes["interpolate_pos_encoding"] = None  # type: ignore[assignment]
     res = dict(inputs=inputs, dynamic_shapes=shapes)
     if add_second_input:
+        assert (
+            add_second_input > 0
+        ), f"Not implemented for add_second_input={add_second_input}."
         res["inputs2"] = get_inputs(
             model=model,
             config=config,
-            input_width=input_width + 1,
-            input_height=input_height + 1,
+            input_width=input_width + add_second_input,
+            input_height=input_height + add_second_input,
             input_channels=input_channels,
             batch_size=batch_size + 1,
             dynamic_rope=dynamic_rope,
+            add_second_input=0,
             **kwargs,
         )["inputs"]
     return res

onnx_diagnostic/tasks/image_text_to_text.py

@@ -32,7 +32,7 @@ def get_inputs(
     sequence_length2: int = 3,
     n_images: int = 2,
     dynamic_rope: bool = False,
-    add_second_input: bool = False,
+    add_second_input: int = 1,
     **kwargs,  # unused
 ):
     """
@@ -105,6 +105,9 @@ def get_inputs(
     )
     res = dict(inputs=inputs, dynamic_shapes=shapes)
     if add_second_input:
+        assert (
+            add_second_input > 0
+        ), f"Not implemented for add_second_input={add_second_input}."
         res["inputs2"] = get_inputs(
             model=model,
             config=config,
@@ -116,10 +119,11 @@ def get_inputs(
             height=height,
             num_channels=num_channels,
             batch_size=batch_size + 1,
-            sequence_length=sequence_length + 1,
+            sequence_length=sequence_length + add_second_input,
             sequence_length2=sequence_length2 + 1,
             n_images=n_images + 1,
             dynamic_rope=dynamic_rope,
+            add_second_input=0,
             **kwargs,
         )["inputs"]
     return res

onnx_diagnostic/tasks/mixture_of_expert.py

@@ -41,7 +41,7 @@ def get_inputs(
     sequence_length2: int = 3,
     n_images: int = 2,
     dynamic_rope: bool = False,
-    add_second_input: bool = False,
+    add_second_input: int = 1,
     **kwargs,  # unused
 ):
     """

onnx_diagnostic/tasks/object_detection.py

@@ -27,7 +27,7 @@ def get_inputs(
     input_channels: int,
     batch_size: int = 2,
     dynamic_rope: bool = False,
-    add_second_input: bool = False,
+    add_second_input: int = 1,
     **kwargs,  # unused
 ):
     """
@@ -65,14 +65,18 @@ def get_inputs(
     )
     res = dict(inputs=inputs, dynamic_shapes=shapes)
     if add_second_input:
+        assert (
+            add_second_input > 0
+        ), f"Not implemented for add_second_input={add_second_input}."
         res["inputs2"] = get_inputs(
             model=model,
             config=config,
-            input_width=input_width + 1,
-            input_height=input_height + 1,
+            input_width=input_width + add_second_input,
+            input_height=input_height + add_second_input,
             input_channels=input_channels,
             batch_size=batch_size + 1,
             dynamic_rope=dynamic_rope,
+            add_second_input=0,
             **kwargs,
         )["inputs"]
     return res

onnx_diagnostic/tasks/sentence_similarity.py

@@ -22,7 +22,7 @@ def get_inputs(
     batch_size: int,
     sequence_length: int,
     dummy_max_token_id: int,
-    add_second_input: bool = False,
+    add_second_input: int = 1,
     **kwargs,  # unused
 ):
     """
@@ -54,12 +54,16 @@ def get_inputs(
     )
     res = dict(inputs=inputs, dynamic_shapes=shapes)
     if add_second_input:
+        assert (
+            add_second_input > 0
+        ), f"Not implemented for add_second_input={add_second_input}."
         res["inputs2"] = get_inputs(
             model=model,
             config=config,
             batch_size=batch_size + 1,
-            sequence_length=sequence_length + 1,
+            sequence_length=sequence_length + add_second_input,
             dummy_max_token_id=dummy_max_token_id,
+            add_second_input=0,
             **kwargs,
         )["inputs"]
     return res

onnx_diagnostic/tasks/summarization.py

@@ -29,7 +29,7 @@ def get_inputs(
     batch_size: int = 2,
     sequence_length: int = 30,
     sequence_length2: int = 3,
-    add_second_input: bool = False,
+    add_second_input: int = 1,
     **kwargs,  # unused
 ):
     """
@@ -144,6 +144,9 @@ def get_inputs(
     )
     res = dict(inputs=inputs, dynamic_shapes=shapes)
     if add_second_input:
+        assert (
+            add_second_input > 0
+        ), f"Not implemented for add_second_input={add_second_input}."
         res["inputs2"] = get_inputs(
             model=model,
             config=config,
@@ -154,8 +157,9 @@ def get_inputs(
             head_dim_encoder=head_dim_encoder,
             head_dim_decoder=head_dim_decoder,
             batch_size=batch_size + 1,
-            sequence_length=sequence_length + 1,
+            sequence_length=sequence_length + add_second_input,
             sequence_length2=sequence_length2 + 1,
+            add_second_input=0,
             **kwargs,
         )["inputs"]
     return res

onnx_diagnostic/tasks/text2text_generation.py

@@ -30,7 +30,7 @@ def get_inputs(
     batch_size: int = 2,
     sequence_length: int = 30,
     sequence_length2: int = 3,
-    add_second_input: bool = False,
+    add_second_input: int = 1,
     **kwargs,  # unused
 ):
     """
@@ -149,6 +149,9 @@ def get_inputs(
     )
     res = dict(inputs=inputs, dynamic_shapes=shapes)
     if add_second_input:
+        assert (
+            add_second_input > 0
+        ), f"Not implemented for add_second_input={add_second_input}."
         res["inputs2"] = get_inputs(
             model=model,
             config=config,
@@ -160,8 +163,9 @@ def get_inputs(
             head_dim_decoder=head_dim_decoder,
             encoder_dim=encoder_dim,
             batch_size=batch_size + 1,
-            sequence_length=sequence_length + 1,
+            sequence_length=sequence_length + add_second_input,
             sequence_length2=sequence_length2 + 1,
+            add_second_input=0,
             **kwargs,
         )["inputs"]
     return res

onnx_diagnostic/tasks/text_classification.py

@@ -22,7 +22,7 @@ def get_inputs(
     batch_size: int,
     sequence_length: int,
     dummy_max_token_id: int,
-    add_second_input: bool = False,
+    add_second_input: int = 1,
     **kwargs,  # unused
 ):
     """
@@ -54,12 +54,16 @@ def get_inputs(
     )
     res = dict(inputs=inputs, dynamic_shapes=shapes)
     if add_second_input:
+        assert (
+            add_second_input > 0
+        ), f"Not implemented for add_second_input={add_second_input}."
         res["inputs2"] = get_inputs(
             model=model,
             config=config,
             batch_size=batch_size + 1,
-            sequence_length=sequence_length + 1,
+            sequence_length=sequence_length + add_second_input,
             dummy_max_token_id=dummy_max_token_id,
+            add_second_input=0,
             **kwargs,
         )["inputs"]
     return res

onnx_diagnostic/tasks/text_generation.py

@@ -72,7 +72,7 @@ def get_inputs(
     num_key_value_heads: Optional[int] = None,
     head_dim: Optional[int] = None,
     cls_cache: Optional[Union[type, str]] = None,
-    add_second_input: bool = False,
+    add_second_input: int = 1,
     **kwargs,  # unused
 ):
     """
@@ -260,13 +260,15 @@ def get_inputs(
             config=config,
             dummy_max_token_id=dummy_max_token_id,
             num_hidden_layers=num_hidden_layers,
-            batch_size=batch_size + 1,
+            batch_size=(batch_size + 1) if add_second_input > 0 else 1,
             sequence_length=sequence_length + 1,
-            sequence_length2=sequence_length2 + 1,
+            sequence_length2=sequence_length2
+            + (add_second_input if add_second_input > 0 else -add_second_input),
             dynamic_rope=dynamic_rope,
             num_key_value_heads=num_key_value_heads,
             head_dim=head_dim,
             cls_cache=cls_cache,
+            add_second_input=0,
             **kwargs,
         )["inputs"]
     return res

onnx_diagnostic/tasks/text_to_image.py

@@ -25,7 +25,7 @@ def get_inputs(
     in_channels: int,
     sample_size: int,
     cross_attention_dim: int,
-    add_second_input: bool = False,
+    add_second_input: int = 1,
     **kwargs,  # unused
 ):
     """
@@ -58,15 +58,19 @@ def get_inputs(
     )
     res = dict(inputs=inputs, dynamic_shapes=shapes)
     if add_second_input:
+        assert (
+            add_second_input > 0
+        ), f"Not implemented for add_second_input={add_second_input}."
         res["inputs2"] = get_inputs(
             model=model,
             config=config,
             batch_size=batch_size + 1,
             sequence_length=sequence_length,
-            cache_length=cache_length + 1,
+            cache_length=cache_length + add_second_input,
             in_channels=in_channels,
             sample_size=sample_size,
             cross_attention_dim=cross_attention_dim,
+            add_second_input=0,
             **kwargs,
         )["inputs"]
     return res

onnx_diagnostic/tasks/zero_shot_image_classification.py

@@ -34,7 +34,7 @@ def get_inputs(
     input_height: int = 224,
     input_channels: int = 3,
     batch_size_image=3,
-    add_second_input: bool = False,
+    add_second_input: int = 1,
     **kwargs,  # unused
 ):
     """
@@ -87,16 +87,20 @@ def get_inputs(
     )
     res = dict(inputs=inputs, dynamic_shapes=shapes)
     if add_second_input:
+        assert (
+            add_second_input > 0
+        ), f"Not implemented for add_second_input={add_second_input}."
         res["inputs2"] = get_inputs(
             model=model,
             config=config,
             dummy_max_token_id=dummy_max_token_id,
             batch_size=batch_size + 1,
-            sequence_length=sequence_length + 1,
+            sequence_length=sequence_length + add_second_input,
             input_width=input_width,
             input_height=input_height,
             input_channels=input_channels,
             batch_size_image=batch_size_image + 1,
+            add_second_input=0,
             **kwargs,
         )["inputs"]
     return res

onnx_diagnostic/torch_export_patches/onnx_export_errors.py

@@ -420,7 +420,11 @@ def torch_export_patches(
                 patch_transformers_list, verbose=verbose
             )
 
-            if masking_utils and hasattr(masking_utils, "_vmap_for_bhqkv"):
+            if (
+                masking_utils
+                and patch_transformers_list.patch_masking_utils
+                and hasattr(masking_utils, "_vmap_for_bhqkv")
+            ):
                 if verbose:
                     print(
                         "[torch_export_patches] patches "
@@ -429,6 +433,27 @@ def torch_export_patches(
                 f_transformers__vmap_for_bhqkv = masking_utils._vmap_for_bhqkv
                 masking_utils._vmap_for_bhqkv = patch_transformers_list.patched__vmap_for_bhqkv
 
+            if (
+                masking_utils
+                and patch_transformers_list.patch_masking_utils
+                and hasattr(masking_utils, "eager_mask")
+            ):
+                if verbose:
+                    print(
+                        "[torch_export_patches] patches "
+                        "transformers.masking_utils.eager_mask"
+                    )
+                f_transformers_eager_mask = masking_utils.eager_mask
+                masking_utils.eager_mask = patch_transformers_list.patched_eager_mask
+                if (
+                    "eager" in masking_utils.ALL_MASK_ATTENTION_FUNCTIONS
+                    and masking_utils.ALL_MASK_ATTENTION_FUNCTIONS["eager"]
+                    == f_transformers_eager_mask
+                ):
+                    masking_utils.ALL_MASK_ATTENTION_FUNCTIONS["eager"] = (
+                        patch_transformers_list.patched_eager_mask
+                    )
+
         if custom_patches:
             if verbose:
                 print("[torch_export_patches] applies custom patches")
@@ -511,7 +536,7 @@ def torch_export_patches(
 
             if custom_patches:
                 if verbose:
-                    print("[torch_export_patches] unpatch custom patches")
+                    print("[torch_export_patches] unpatches custom patches")
                 unpatch_module_or_classes(
                     custom_patches, revert_custom_patches_info, verbose=verbose
                 )
@@ -526,18 +551,43 @@ def torch_export_patches(
                 except ImportError:
                     masking_utils = None
                 if verbose:
-                    print("[torch_export_patches] unpatch transformers")
+                    print("[torch_export_patches] unpatches transformers")
                 unpatch_module_or_classes(
                     patch_transformers_list, revert_patches_info, verbose=verbose
                 )
 
-                if masking_utils and hasattr(masking_utils, "_vmap_for_bhqkv"):
+                if (
+                    masking_utils
+                    and patch_transformers_list.patch_masking_utils
+                    and hasattr(masking_utils, "_vmap_for_bhqkv")
+                ):
+                    masking_utils._vmap_for_bhqkv = f_transformers__vmap_for_bhqkv
                     if verbose:
                         print(
-                            "[torch_export_patches] unpatch "
+                            "[torch_export_patches] restored "
                             "transformers.masking_utils._vmap_for_bhqkv"
                         )
-                    masking_utils._vmap_for_bhqkv = f_transformers__vmap_for_bhqkv
+
+                if (
+                    masking_utils
+                    and patch_transformers_list.patch_masking_utils
+                    and hasattr(masking_utils, "eager_mask")
+                ):
+                    f_transformers_eager_mask = masking_utils.eager_mask
+                    masking_utils.eager_mask = f_transformers_eager_mask
+                    if (
+                        "eager" in masking_utils.ALL_MASK_ATTENTION_FUNCTIONS
+                        and masking_utils.ALL_MASK_ATTENTION_FUNCTIONS["eager"]
+                        == patch_transformers_list.patched_eager_mask
+                    ):
+                        masking_utils.ALL_MASK_ATTENTION_FUNCTIONS["eager"] = (
+                            f_transformers_eager_mask
+                        )
+                    if verbose:
+                        print(
+                            "[torch_export_patches] restored "
+                            "transformers.masking_utils.eager_mask"
+                        )
 
             ########
             # caches

onnx_diagnostic/torch_export_patches/patches/patch_transformers.py

@@ -7,59 +7,107 @@ import torch
 import transformers
 from transformers.modeling_attn_mask_utils import AttentionMaskConverter
 from transformers.cache_utils import StaticCache, Cache, DynamicCache
+
+try:
+    import transformers.masking_utils
+
+    patch_masking_utils = True
+except ImportError:
+    patch_masking_utils = False
+
 from ...ext_test_case import has_transformers
 from ...helpers.torch_helper import is_torchdynamo_exporting
 
 
-def patched__vmap_for_bhqkv(mask_function: Callable, bh_indices: bool = True) -> Callable:
-    """manual patch for function ``transformers.masking_utils._vmap_for_bhqkv``."""
-    from ...helpers import string_type
-
-    dimensions: List[Tuple[Optional[int], ...]] = [
-        (None, None, None, 0),
-        (None, None, 0, None),
-    ]
-    if bh_indices:
-        dimensions.extend([(None, 0, None, None), (0, None, None, None)])
-    # reshape
-    dimensions = [tuple(1 if d is None else -1 for d in shape) for shape in dimensions]
-    dimensions = tuple(reversed(dimensions))
-    indices = tuple(shape.index(-1) for shape in dimensions)
-
-    # unsqueeze
-    udimensions = [tuple(di for di, d in enumerate(shape) if d == 1) for shape in dimensions]
-
-    def vector_mask_function(
-        *args, mask_function=mask_function, dimensions=dimensions, indices=indices
-    ):
-        assert len(args) == len(dimensions) == len(udimensions), (
-            f"Mismatch between args={string_type(args)} and dimensions={dimensions} "
-            f"and udimensions={udimensions}."
-        )
-        assert len(indices) == len(args), (
-            f"Mismatch between args={string_type(args)} and indices={indices}, "
-            f"they should have the same length."
+if patch_masking_utils:
+    # Introduced in 4.52
+    from transformers.masking_utils import causal_mask_function, sdpa_mask
+
+    def patched__vmap_for_bhqkv(mask_function: Callable, bh_indices: bool = True) -> Callable:
+        """manual patch for function ``transformers.masking_utils._vmap_for_bhqkv``."""
+        from ...helpers import string_type
+
+        dimensions: List[Tuple[Optional[int], ...]] = [
+            (None, None, None, 0),
+            (None, None, 0, None),
+        ]
+        if bh_indices:
+            dimensions.extend([(None, 0, None, None), (0, None, None, None)])
+        # reshape
+        dimensions = [tuple(1 if d is None else -1 for d in shape) for shape in dimensions]
+        dimensions = tuple(reversed(dimensions))
+        indices = tuple(shape.index(-1) for shape in dimensions)
+
+        # unsqueeze
+        udimensions = [
+            tuple(di for di, d in enumerate(shape) if d == 1) for shape in dimensions
+        ]
+
+        def vector_mask_function(
+            *args, mask_function=mask_function, dimensions=dimensions, indices=indices
+        ):
+            assert len(args) == len(dimensions) == len(udimensions), (
+                f"Mismatch between args={string_type(args)} and dimensions={dimensions} "
+                f"and udimensions={udimensions}."
+            )
+            assert len(indices) == len(args), (
+                f"Mismatch between args={string_type(args)} and indices={indices}, "
+                f"they should have the same length."
+            )
+            for a in args:
+                assert (
+                    a.ndim == 1
+                ), f"Expected a tensor with 1 dimension not {string_type(a, with_shape=True)}"
+                torch._check(a.shape[0] > 0)
+
+            new_args = [a.reshape(shape) for a, shape in zip(args, dimensions)]
+            # new_args = [
+            #    a.unsqueeze(dims[0]).unsqueeze(dims[1]).unsqueeze(dims[2])
+            #    for a, dims in zip(args, udimensions)
+            # ]
+            max_shape = tuple(args[i].shape[0] for i in indices)
+            # if is_torchdynamo_exporting():
+            #     for a in args:
+            #         # The exporter should export with a dimension > 1
+            #         # to make sure it is dynamic.
+            #         torch._check(a.shape[0] > 1)
+            expanded_args = [a.expand(max_shape) for a in new_args]
+            return mask_function(*expanded_args)
+
+        return vector_mask_function
+
+    def patched_eager_mask(
+        batch_size: int,
+        cache_position: torch.Tensor,
+        kv_length: int,
+        kv_offset: int = 0,
+        mask_function: Callable = causal_mask_function,
+        attention_mask: Optional[torch.Tensor] = None,
+        dtype: torch.dtype = torch.float32,
+        **kwargs,
+    ) -> torch.Tensor:
+        """manual patch for function ``transformers.masking_utils.eager_mask``."""
+        # The masks for eager attention are simply boolean mask from sdpa, casted to 0 and -inf
+        _ = kwargs.pop("allow_is_causal_skip", None)
+        mask = sdpa_mask(
+            batch_size=batch_size,
+            cache_position=cache_position,
+            kv_length=kv_length,
+            kv_offset=kv_offset,
+            mask_function=mask_function,
+            attention_mask=attention_mask,
+            allow_is_causal_skip=False,
+            allow_torch_fix=False,
+            **kwargs,
         )
-        for a in args:
-            assert (
-                a.ndim == 1
-            ), f"Expected a tensor with 1 dimension not {string_type(a, with_shape=True)}"
-            torch._check(a.shape[0] > 0)
-
-        new_args = [a.reshape(shape) for a, shape in zip(args, dimensions)]
-        # new_args = [
-        #    a.unsqueeze(dims[0]).unsqueeze(dims[1]).unsqueeze(dims[2])
-        #    for a, dims in zip(args, udimensions)
-        # ]
-        max_shape = tuple(args[i].shape[0] for i in indices)
-        # if is_torchdynamo_exporting():
-        #     for a in args:
-        #         # The exporter should export with a dimension > 1 to make sure it is dynamic.
-        #         torch._check(a.shape[0] > 1)
-        expanded_args = [a.expand(max_shape) for a in new_args]
-        return mask_function(*expanded_args)
-
-    return vector_mask_function
+        min_dtype = torch.finfo(dtype).min
+        # The patched line.
+        # we need 0s where the tokens should be taken into account,
+        # and -inf otherwise (mask is already of boolean type)
+        # mask =
+        #   torch.where(mask, torch.tensor(0.0, device=mask.device, dtype=dtype), min_dtype)
+        mask = (~mask).to(dtype) * min_dtype
+        return mask
 
 
 def _patch_make_causal_mask(

onnx_diagnostic/torch_models/hghub/model_inputs.py

@@ -26,7 +26,7 @@ def get_untrained_model_with_inputs(
     use_pretrained: bool = False,
     same_as_pretrained: bool = False,
     use_preinstalled: bool = True,
-    add_second_input: bool = False,
+    add_second_input: int = 1,
     subfolder: Optional[str] = None,
     use_only_preinstalled: bool = False,
 ) -> Dict[str, Any]:

onnx_diagnostic/torch_models/validate.py

@@ -18,7 +18,6 @@ from ..helpers.cache_helper import flatten_unflatten_for_dynamic_shapes
 from ..tasks import random_input_kwargs
 from ..torch_export_patches import torch_export_patches
 from ..torch_export_patches.patch_inputs import use_dyn_not_str
-from ..reference import TorchOnnxEvaluator
 from .hghub import get_untrained_model_with_inputs
 
 
@@ -157,6 +156,12 @@ def version_summary() -> Dict[str, Union[int, float, str]]:
         "version_torch": torch.__version__,
         "version_numpy": numpy.__version__,
     }
+    try:
+        import scipy
+
+        summary["version_scipy"] = getattr(scipy, "__version__", "?")
+    except ImportError:
+        pass
     try:
         import transformers
 
@@ -181,6 +186,12 @@ def version_summary() -> Dict[str, Union[int, float, str]]:
         summary["version_onnxruntime"] = getattr(onnxruntime, "__version__", "?")
     except ImportError:
         pass
+    try:
+        import onnx_ir
+
+        summary["version_onnx_ir"] = getattr(onnx_ir, "__version__", "?")
+    except ImportError:
+        pass
     import onnx_diagnostic
 
     summary["version_onnx_diagnostic"] = onnx_diagnostic.__version__
@@ -276,7 +287,7 @@ def validate_model(
     runtime: str = "onnxruntime",
     repeat: int = 1,
     warmup: int = 0,
-    inputs2: bool = True,
+    inputs2: int = 1,
 ) -> Tuple[Dict[str, Union[int, float, str]], Dict[str, Any]]:
     """
     Validates a model.
@@ -325,7 +336,8 @@ def validate_model(
     :param repeat: number of time to measure the model
     :param warmup: warmup the model first
     :param inputs2: checks that the second set of inputs is reunning as well,
-        this ensures that the model does support dynamism
+        this ensures that the model does support dynamism, the value is used
+        as an increment to the first set of values (added to dimensions)
     :return: two dictionaries, one with some metrics,
         another one with whatever the function produces
 
@@ -1054,7 +1066,7 @@ def validate_onnx_model(
     runtime: str = "onnxruntime",
     repeat: int = 1,
     warmup: int = 0,
-    inputs2: bool = True,
+    inputs2: int = 1,
 ) -> Tuple[Dict[str, Any], Dict[str, Any]]:
     """
     Verifies that an onnx model produces the same
@@ -1070,8 +1082,9 @@ def validate_onnx_model(
     :param runtime: onnx runtime to use, onnxruntime or torch
     :param repeat: run that number of times the model
     :param warmup: warmup the model
-    :param inputs: to validate the model on the second input set
-        to make sure the exported model supports dynamism
+    :param inputs2: to validate the model on the second input set
+        to make sure the exported model supports dynamism, the value is
+        used as an increment added to the first set of inputs (added to dimensions)
     :return: two dictionaries, one with some metrics,
         another one with whatever the function produces
     """
@@ -1113,6 +1126,9 @@ def validate_onnx_model(
             f"{providers}..., flavour={flavour!r}"
         )
 
+    if runtime != "onnxruntime":
+        from ..reference import TorchOnnxEvaluator
+
     cls_runtime = (
         (
             lambda model, providers: onnxruntime.InferenceSession(
@@ -1122,7 +1138,7 @@ def validate_onnx_model(
         )
         if runtime == "onnxruntime"
         else (
-            lambda model, providers: TorchOnnxEvaluator(
+            lambda model, providers, _cls_=TorchOnnxEvaluator: _cls_(  # type: ignore[misc]
                 model, providers=providers, verbose=max(verbose - 1, 0)
             )
         )

{onnx_diagnostic-0.7.3.dist-info → onnx_diagnostic-0.7.4.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: onnx-diagnostic
-Version: 0.7.3
+Version: 0.7.4
 Summary: Investigate ONNX models
 Home-page: https://github.com/sdpython/onnx-diagnostic
 Author: Xavier Dupré

{onnx_diagnostic-0.7.3.dist-info → onnx_diagnostic-0.7.4.dist-info}/RECORD

@@ -1,17 +1,17 @@
-onnx_diagnostic/__init__.py,sha256=N1lf8_afRytDUnulPCeDVDPA-M4k7y9x7LbWwX0USZs,173
+onnx_diagnostic/__init__.py,sha256=dmZNMpFkDRd7ZCC6bC2fEFqvAhHhsqqua8ZE5LbOC9s,173
 onnx_diagnostic/__main__.py,sha256=YmyV_Aq_ianDlHyKLHMa6h8YK3ZmFPpLVHLKjM91aCk,79
-onnx_diagnostic/_command_lines_parser.py,sha256=TliHXedXFerv-zO6cBigxKbuKHE0-6TUhhsk1pdkz9M,28072
+onnx_diagnostic/_command_lines_parser.py,sha256=65oUjJ2tgPxQgIKgOAI04jhOFRnGUSNivUNDVMZ-urU,28597
 onnx_diagnostic/api.py,sha256=BhCl_yCd78N7TlVtPOHjeYv1QBEy39TjZ647rcHqLh0,345
 onnx_diagnostic/doc.py,sha256=t3RELgfooYnVMAi0JSpggWkQEgUsREz8NmRvn0TnLI8,2829
 onnx_diagnostic/ext_test_case.py,sha256=Bq4vdlM0P72H1orlKJTeOBqm1YGHTK-ylAlNsBe4LeA,43438
 onnx_diagnostic/export/__init__.py,sha256=yEIoWiOeTwBsDhyYt2fTKuhtA0Ya1J9u9ZzMTOTWaWs,101
 onnx_diagnostic/export/dynamic_shapes.py,sha256=HYf2OEi7PmRSj8uxMD-wbdVxxejkWXTPBAkxoFeM27A,40811
-onnx_diagnostic/export/shape_helper.py,sha256=C9cEq_x8I40RKuD89qWIholN1XZoWhaKPfbZQhiPD3g,4725
+onnx_diagnostic/export/shape_helper.py,sha256=EQXHRVxwGpHRYhx8Y44Crqs640pmaIuKSwW1KJOW0IU,7501
 onnx_diagnostic/export/validate.py,sha256=_PGUql2DJhIgGKo0WjTGUc5AgsZUx8fEs00MePy-w98,6043
 onnx_diagnostic/helpers/__init__.py,sha256=GJ2GT7cgnlIveVUwMZhuvUwidbTJaKv8CsSIOpZDsJg,83
 onnx_diagnostic/helpers/args_helper.py,sha256=SRWnqC7EENg09RZlA50B_PcdiIhdbgA4C3ACfzl5nMs,4419
 onnx_diagnostic/helpers/bench_run.py,sha256=CGA6VMJZMH2gDhVueT9ypNm4PMcjGrrGFYp08nhWj9k,16539
-onnx_diagnostic/helpers/cache_helper.py,sha256=1slql7FdL89wzRWGoehLI7G1BY0cOsqGQmaZvLBvCIA,11229
+onnx_diagnostic/helpers/cache_helper.py,sha256=TeBUuGvqIMO-dsLDy7keaVt3ImZeifldwTgx6TEjBo8,11595
 onnx_diagnostic/helpers/config_helper.py,sha256=9h1NWC9RLmu43Yf5Cz9usjMdLiyLWXMhwgE4Lg-eOU8,3889
 onnx_diagnostic/helpers/doc_helper.py,sha256=pl5MZd3_FaE8BqQnqoBuSBxoNCFcd2OJd3eITUSku5c,5897
 onnx_diagnostic/helpers/graph_helper.py,sha256=hevQT5a7_QuriVPQcbT5qe18n99Doyl5h3-qshx1-uk,14093
@@ -35,7 +35,7 @@ onnx_diagnostic/reference/ops/op_add_add_mul_mul.py,sha256=CXQVtgVrT066gDJFwxL4n
 onnx_diagnostic/reference/ops/op_attention.py,sha256=ThALMDF53v3QeG1bohi0bvX2o90HZhGJbbAFOtwEHPE,2027
 onnx_diagnostic/reference/ops/op_average_pool_grad.py,sha256=zMcOtjB7hWySfIIFXogcmj3xxCWwxEX_g2VKg-SOAEs,2360
 onnx_diagnostic/reference/ops/op_bias_softmax.py,sha256=dcXsw2chxc8-puIkI0LFsBxKOJaCSovDcF1HkgboQp0,524
-onnx_diagnostic/reference/ops/op_cast_like.py,sha256=qpA_3KenP7pK9w6O-OvqNLG-rzQO3w9dsWFKqaMR8Ns,1274
+onnx_diagnostic/reference/ops/op_cast_like.py,sha256=tH2pYFPluHK9r_cdkqnk9k2q3dPzaeKgQZt_Pjw9N4o,1383
 onnx_diagnostic/reference/ops/op_complex.py,sha256=OobDrRNYcktdCdTJzOQBesrKC8vsKxuHIi7Yev1DJrs,651
 onnx_diagnostic/reference/ops/op_concat.py,sha256=seW71-QDKzv9QQhjhjThKip0Y3d9nkVd7Hs1A2nNQjk,519
 onnx_diagnostic/reference/ops/op_constant_of_shape.py,sha256=3G9TRxaUoqYudnKrVHBEblo_16qMl0c9wNZqnQeNJJ4,2009
@@ -72,22 +72,22 @@ onnx_diagnostic/reference/torch_ops/sequence_ops.py,sha256=3EiVKpGfN4d1Iry4hgnr3
 onnx_diagnostic/reference/torch_ops/shape_ops.py,sha256=pJrNR2UB4PlWl6cv4EDl1uGl8YTBUUMQkhJcsh5K4sA,4291
 onnx_diagnostic/reference/torch_ops/unary_ops.py,sha256=E8Ys1eZsOTsucBKoXb1_Kl5LbBDygniDvW2BvN4IPMo,1708
 onnx_diagnostic/tasks/__init__.py,sha256=0BYtrAnr0zKN3om71oi-OVz5wFYDp9WWIk51qWjjyCw,2450
-onnx_diagnostic/tasks/automatic_speech_recognition.py,sha256=7OspFypNHLSL6huvP9ms_euhHqYXyTZjAJ8Z7EXGimk,6967
-onnx_diagnostic/tasks/feature_extraction.py,sha256=CbxbGsv3JvEQ2J9tO2DOpMHcJj5ZlCwY81ZB3hPB4D4,2339
-onnx_diagnostic/tasks/fill_mask.py,sha256=ZWz8swzEeRbkmbY9oZ4CM1LYCWWUxnS5CqrKmUVw-u0,2457
-onnx_diagnostic/tasks/image_classification.py,sha256=UjUAFYnwXIdPMXJdHR5MDzpsfMeIvyuKR4RqJVpGV_Q,4449
-onnx_diagnostic/tasks/image_text_to_text.py,sha256=LmpMdH6oF_EN3WIACzSip4fPZOjZWFOoXg4k8qAio6Q,7639
-onnx_diagnostic/tasks/mixture_of_expert.py,sha256=C0ugEc8OWmVyEZpsh8MJq_te1zgOHhpITtnSmGC16Ls,2801
-onnx_diagnostic/tasks/object_detection.py,sha256=1lF5e2f2Coz1urSptEKgvUGCOSFBf0Anuq_QYOC00dA,4046
-onnx_diagnostic/tasks/sentence_similarity.py,sha256=3MvNxjC1iEMtQL_jH1c8bmrVc5IG1lfUygrCZ0SORJk,2472
-onnx_diagnostic/tasks/summarization.py,sha256=NLwqhpiQrU8UWd3u30VsNA3FsL315S3nlQ7ycUzJueo,8105
-onnx_diagnostic/tasks/text2text_generation.py,sha256=mYvsq-O69fr5pitX0mWugT76QuK4xUs40Vsz9ru_XK8,8522
-onnx_diagnostic/tasks/text_classification.py,sha256=NCCKobBQyCc7dSVj7_5N6S_RuvBlRMAdWkS2rVvrzck,2528
-onnx_diagnostic/tasks/text_generation.py,sha256=fraa5w2_eexXLAcDAR1e4g9zWKdvWybMYi1zu7XU7J4,12988
-onnx_diagnostic/tasks/text_to_image.py,sha256=6z-rFG6MX9aBi8YoYtYI_8OV3M3Tfoi45V8S9csQ7k4,2799
-onnx_diagnostic/tasks/zero_shot_image_classification.py,sha256=GKaXm8g7cK23h3wJEUc6Q-6mpmLAzQ4YkJbd-eGP7Y4,4496
+onnx_diagnostic/tasks/automatic_speech_recognition.py,sha256=H94rxeiAjcJdECA1g95G_U9fZfpXk6dfjNKKYuvc4Qc,7130
+onnx_diagnostic/tasks/feature_extraction.py,sha256=MptOP-1ZSIMTIJ0woSqKLR7TlB9m7kww9V8cfEgZJTY,2502
+onnx_diagnostic/tasks/fill_mask.py,sha256=Rvrz0j_DQu-vf4CSSAZMBMXb2EuHvOCzRZwj8Cy8yfg,2620
+onnx_diagnostic/tasks/image_classification.py,sha256=x1XfeWAOe0r_s9kU6WENoYxjfoRTp1pkwKgIveoLbUw,4627
+onnx_diagnostic/tasks/image_text_to_text.py,sha256=-vbZMA_ruo0WR_96YMYRvoNfq1plpElBJWXC2klAf7Q,7802
+onnx_diagnostic/tasks/mixture_of_expert.py,sha256=DgIsbwzV4smysOK83wny91k3ix1Qt2tSFXLGLoz4WOo,2796
+onnx_diagnostic/tasks/object_detection.py,sha256=xRBH9JZxBQf0SVSTJP6d-VVCKqrw7JAeif1joHfiYng,4224
+onnx_diagnostic/tasks/sentence_similarity.py,sha256=soL6QxLvyjtQ-3tQ3nCFxrcrk_4a8tuAjil8zYQ_pXk,2635
+onnx_diagnostic/tasks/summarization.py,sha256=LZ8A8wl6cd8kWSc6k5vLHa_XZkm35rYkTRv8iUYtr6I,8268
+onnx_diagnostic/tasks/text2text_generation.py,sha256=Pk-H_qX5Y-2dzk45N9jbQ73S3O_d5-D11MyUhUfUwuM,8685
+onnx_diagnostic/tasks/text_classification.py,sha256=dO_LLbwwv0OJfIa9DqxQqAGUDuz3iIF1XkafzaYJdJw,2691
+onnx_diagnostic/tasks/text_generation.py,sha256=tW9Gnum_eck3czNyctuUISA-Ek7pO37v5-11GC8QBW8,13124
+onnx_diagnostic/tasks/text_to_image.py,sha256=mOS3Ruosi3hzRMxXLDN7ZkAbi7NnQb7MWwQP_okGVHs,2962
+onnx_diagnostic/tasks/zero_shot_image_classification.py,sha256=1iqYamkq5kZNXEXsySw748ernc0O94GkwpYAIEl6Kj4,4659
 onnx_diagnostic/torch_export_patches/__init__.py,sha256=0SaZedwznm1hQUCvXZsGZORV5vby954wEExr5faepGg,720
-onnx_diagnostic/torch_export_patches/onnx_export_errors.py,sha256=h_txSp30QmF1R_Q2wL4qpPqY59Dund2P9nAAsvucS8A,21245
+onnx_diagnostic/torch_export_patches/onnx_export_errors.py,sha256=ZsUSOnKxeq4-dP86c5dTIbHMJFy_y690vvU4yfo6tNs,23438
 onnx_diagnostic/torch_export_patches/onnx_export_serialization.py,sha256=wFE2fNxihAA3iua79AEB97_RBVv4wvGxwS9g4RJaSIc,10715
 onnx_diagnostic/torch_export_patches/patch_expressions.py,sha256=vr4tt61cbDnaaaduzMj4UBZ8OUtr6GfDpIWwOYqjWzs,3213
 onnx_diagnostic/torch_export_patches/patch_inputs.py,sha256=9b4pmyT00BwLqi7WG-gliep1RUy3gXEgW6BDnlSSA-M,7689
@@ -97,26 +97,26 @@ onnx_diagnostic/torch_export_patches/eval/__init__.py,sha256=57x62uZNA80XiWgkG8F
 onnx_diagnostic/torch_export_patches/eval/model_cases.py,sha256=DTvdHPtNQh25Akv5o3D4Jxf1L1-SJ7w14tgvj8AAns8,26577
 onnx_diagnostic/torch_export_patches/patches/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 onnx_diagnostic/torch_export_patches/patches/patch_torch.py,sha256=KaZ8TjDa9ATgT4HllYzzoNf_51q_yOj_GuF5NYjPCrU,18913
-onnx_diagnostic/torch_export_patches/patches/patch_transformers.py,sha256=QGfXf1IlsOToQKF9NOHau_l62u1alQbJ6KeiSKFb980,44061
+onnx_diagnostic/torch_export_patches/patches/patch_transformers.py,sha256=nVOg69e_cXvpcP5WIW9pIHCgnF-P_Ne87mRC6ep0g-I,45847
 onnx_diagnostic/torch_export_patches/serialization/__init__.py,sha256=BHLdRPtNAtNPAS-bPKEj3-foGSPvwAbZXrHzGGPDLEw,1876
 onnx_diagnostic/torch_export_patches/serialization/diffusers_impl.py,sha256=drq3EH_yjcSuIWYsVeUWm8Cx6YCZFU6bP_1PLtPfY5I,945
 onnx_diagnostic/torch_export_patches/serialization/transformers_impl.py,sha256=9u2jkqnuyBkIF3R2sDEO0Jlkedl-cQhBNXxXXDLSEwE,8885
 onnx_diagnostic/torch_models/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 onnx_diagnostic/torch_models/llms.py,sha256=soyg4yC87ptGoeulJhKqw5opGmuLvH1pn_ZDXZ4Jr8E,90
-onnx_diagnostic/torch_models/validate.py,sha256=uUWkrCNE766c6BYKe-g5jQW3w89cvb7F-xnIkKsZKSw,63254
+onnx_diagnostic/torch_models/validate.py,sha256=dlWeRNLcQ2h3fxx07MA0NHoOPqT3-Afejgrj2Rozvck,63796
 onnx_diagnostic/torch_models/hghub/__init__.py,sha256=vi1Q7YHdddj1soiBN42MSvJdFqe2_KUoWafHISjwOu8,58
 onnx_diagnostic/torch_models/hghub/hub_api.py,sha256=Bvr-sTAhS6s6UCkt-KsY_7Mdai08-AQzvHrzbYCSuvk,13186
 onnx_diagnostic/torch_models/hghub/hub_data.py,sha256=NTTDsCtIVvYnr5J3rlcq0GSGDOzOPzq9Tsnb3oVf4q8,8309
 onnx_diagnostic/torch_models/hghub/hub_data_cached_configs.py,sha256=zZvIxTbmL55x44kCj3-T5Kg3Qzm9KB_Xj-MCcU9-LuQ,268245
-onnx_diagnostic/torch_models/hghub/model_inputs.py,sha256=tuTGfcFKK3ugA9Fh909THHWkrNyXkU5IhV6RumlNUiQ,10830
+onnx_diagnostic/torch_models/hghub/model_inputs.py,sha256=qDw03KsLd_ZAMHBso--rUriCAZIewKFZG9n4-1zvGo8,10825
 onnx_diagnostic/torch_models/untrained/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 onnx_diagnostic/torch_models/untrained/llm_phi2.py,sha256=ynBTDHJHCk44NjLT_t6OiFDBdPP0rFGPteiONDxvztw,3708
 onnx_diagnostic/torch_models/untrained/llm_tiny_llm.py,sha256=QXw_Bs2SzfeiQMf-tmtVl83SmVOL4-Um7Qy-f0E48QI,2507
 onnx_diagnostic/torch_onnx/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 onnx_diagnostic/torch_onnx/runtime_info.py,sha256=1g9F_Jf9AAgYQU4stbsrFXwQl-30mWlQrFbQ7val8Ps,9268
 onnx_diagnostic/torch_onnx/sbs.py,sha256=1EL25DeYFzlBSiFG_XjePBLvsiItRXbdDrr5-QZW2mA,16878
-onnx_diagnostic-0.7.3.dist-info/licenses/LICENSE.txt,sha256=Vv6TXglX6Rc0d-f8aREhayhT-6PMQXEyOmI2NKlUCMc,1045
-onnx_diagnostic-0.7.3.dist-info/METADATA,sha256=uMYdGJbm6K04yoi5UidH3uA0nK_PXrnVXe83s9v6yPE,7431
-onnx_diagnostic-0.7.3.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-onnx_diagnostic-0.7.3.dist-info/top_level.txt,sha256=KwNkXewmcobM3ZT1DJLVWH6ebJzA5qKg7cWqKfpGNT4,16
-onnx_diagnostic-0.7.3.dist-info/RECORD,,
+onnx_diagnostic-0.7.4.dist-info/licenses/LICENSE.txt,sha256=Vv6TXglX6Rc0d-f8aREhayhT-6PMQXEyOmI2NKlUCMc,1045
+onnx_diagnostic-0.7.4.dist-info/METADATA,sha256=6l0XAH0UYEmqkPDswFSZFGrU5hYVEDB8YByQCiBkwlI,7431
+onnx_diagnostic-0.7.4.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+onnx_diagnostic-0.7.4.dist-info/top_level.txt,sha256=KwNkXewmcobM3ZT1DJLVWH6ebJzA5qKg7cWqKfpGNT4,16
+onnx_diagnostic-0.7.4.dist-info/RECORD,,