onnx-diagnostic 0.8.11__py3-none-any.whl → 0.9.0__py3-none-any.whl

onnx_diagnostic/torch_export_patches/onnx_export_errors.py

@@ -562,6 +562,7 @@ def _patch_transformers(
                 "[torch_export_patches] patches "
                 "transformers.masking_utils.sdpa_mask_recent_torch"
             )
+
         f_transformers_sdpa_mask_recent_torch = masking_utils.sdpa_mask_recent_torch
         masking_utils.sdpa_mask_recent_torch = (
             patch_transformers_list.patched_sdpa_mask_recent_torch
@@ -574,7 +575,9 @@ def _patch_transformers(
             )
         if masking_utils.sdpa_mask == f_transformers_sdpa_mask_recent_torch:
             if verbose:
-                print("[torch_export_patches] patches transformers.masking_utils.sdpa_mask")
+                print(
+                    "[torch_export_patches] patches transformers.masking_utils.sdpa_mask (1)"
+                )
             f_transformers_sdpa_mask = masking_utils.sdpa_mask
             masking_utils.sdpa_mask = patch_transformers_list.patched_sdpa_mask_recent_torch
             if patch_details:
@@ -583,8 +586,23 @@ def _patch_transformers(
                     f_transformers_sdpa_mask,
                     patch_transformers_list.patched_sdpa_mask_recent_torch,
                 )
-        else:
-            f_transformers_sdpa_mask = None
+
+    if (  # vmap
+        masking_utils
+        and patch_transformers_list.patch_masking_utils
+        and hasattr(masking_utils, "sdpa_mask")
+        and f_transformers_sdpa_mask is None
+    ):
+        if verbose:
+            print("[torch_export_patches] patches transformers.masking_utils.sdpa_mask (3)")
+        f_transformers_sdpa_mask = masking_utils.sdpa_mask
+        masking_utils.sdpa_mask = patch_transformers_list.patched_sdpa_mask
+        if patch_details:
+            patch_details.append(
+                "transformers",
+                f_transformers_sdpa_mask,
+                patch_transformers_list.patched_sdpa_mask,
+            )
 
     if (  # eager_mask
         masking_utils
@@ -742,17 +760,17 @@ def _unpatch_transformers(
                 "transformers.masking_utils.sdpa_mask_recent_torch"
             )
 
-        if f_transformers_sdpa_mask is not None:
-            assert f_transformers_sdpa_mask.__name__ in (
-                "sdpa_mask",
-                "sdpa_mask_recent_torch",
-            ), (
-                f"corrupted function 'sdpa_mask', its name is "
-                f"{f_transformers_sdpa_mask.__name__!r}"
-            )
-            masking_utils.sdpa_mask = f_transformers_sdpa_mask
-            if verbose:
-                print("[torch_export_patches] restored transformers.masking_utils.sdpa_mask")
+    if f_transformers_sdpa_mask is not None:
+        assert f_transformers_sdpa_mask.__name__ in (
+            "sdpa_mask",
+            "sdpa_mask_recent_torch",
+        ), (
+            f"corrupted function 'sdpa_mask', its name is "
+            f"{f_transformers_sdpa_mask.__name__!r}"
+        )
+        masking_utils.sdpa_mask = f_transformers_sdpa_mask
+        if verbose:
+            print("[torch_export_patches] restored transformers.masking_utils.sdpa_mask")
 
     if (  # eager_mask
         masking_utils

onnx_diagnostic/torch_export_patches/patches/_patch_transformers_masking_utils.py

@@ -36,6 +36,26 @@ if patch_masking_utils:
         _ignore_bidirectional_mask_sdpa = None
         bidirectional_mask_function = None
 
+    try:
+        from transformers.masking_utils import _non_vmap_expansion_sdpa
+    except ImportError:
+
+        def _non_vmap_expansion_sdpa(
+            batch_indices: torch.Tensor,
+            head_indices: torch.Tensor,
+            q_indices: torch.Tensor,
+            kv_indices: torch.Tensor,
+        ):
+            """
+            https://github.com/huggingface/optimum-onnx/blob/
+            c123e8f4fab61b54a8e0e31ce74462bcacca576e/optimum/exporters/onnx/model_patcher.py#L362-L365
+            """
+            batch_indices = batch_indices[:, None, None, None]
+            head_indices = head_indices[None, :, None, None]
+            q_indices = q_indices[None, None, :, None]
+            kv_indices = kv_indices[None, None, None, :]
+            return batch_indices, head_indices, q_indices, kv_indices
+
     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
@@ -146,12 +166,13 @@ if patch_masking_utils:
             padding_mask, q_length, kv_length, kv_offset, local_size
         ):
             return None
-        if (
-            allow_is_bidirectional_skip
-            and _ignore_bidirectional_mask_sdpa
-            and _ignore_bidirectional_mask_sdpa(padding_mask)
-        ):
-            return None
+        if allow_is_bidirectional_skip and _ignore_bidirectional_mask_sdpa:
+            # transformers<=5.0: 1 parameter, 3 for transformers>5.0
+            n_parameters = len(inspect.signature(_ignore_bidirectional_mask_sdpa).parameters)
+            if _ignore_bidirectional_mask_sdpa(
+                *[padding_mask, kv_length, kv_offset][:n_parameters]
+            ):
+                return None
 
         if mask_function is bidirectional_mask_function:
             if padding_mask is not None:
@@ -180,3 +201,83 @@ if patch_masking_utils:
             batch_arange, head_arange, cache_position, kv_arange
         )
         return causal_mask
+
+    def patched_sdpa_mask(
+        batch_size: int,
+        cache_position: torch.Tensor,
+        kv_length: int,
+        kv_offset: int = 0,
+        mask_function: Callable = causal_mask_function,
+        attention_mask: torch.Tensor | None = None,
+        local_size: int | None = None,
+        allow_is_causal_skip: bool = True,
+        allow_is_bidirectional_skip: bool = False,
+        allow_torch_fix: bool = True,
+        use_vmap: bool = False,
+        **kwargs,
+    ) -> torch.Tensor | None:
+        """manual patch for function ``transformers.masking_utils.sdpa_mask``."""
+        q_length = cache_position.shape[0]
+
+        # Potentially pad the 2D mask
+        padding_mask = prepare_padding_mask(attention_mask, kv_length, kv_offset)
+
+        # Under specific conditions, we can avoid materializing the mask
+        #   1. Causal masks can rely on the `is_causal` argument
+        #   2. Bidirectional do not need any further processing (no bias)
+        if allow_is_causal_skip and _ignore_causal_mask_sdpa(
+            padding_mask, q_length, kv_length, kv_offset, local_size
+        ):
+            return None
+        if allow_is_bidirectional_skip and _ignore_bidirectional_mask_sdpa(
+            padding_mask, kv_length, local_size
+        ):
+            return None
+
+        # Potentially add the padding 2D mask
+        if padding_mask is not None:
+            mask_function = and_masks(mask_function, padding_mask_function(padding_mask))
+
+        batch_arange = torch.arange(batch_size, device=cache_position.device)
+        head_arange = torch.arange(1, device=cache_position.device)
+        # Similar to `kv_arange = torch.arange(start=kv_offset,
+        # end=kv_offset + kv_length, device=cache_position.device)`
+        # but without data-dependent slicing (i.e. torch.compile friendly)
+        kv_arange = torch.arange(kv_length, device=cache_position.device) + kv_offset
+
+        # Actual mask creation
+        # Option 1: Fast non-vmap mask creation (default)
+        # PATCHED
+        use_vmap = False
+        if not use_vmap:
+            # Apply mask function element-wise through broadcasting
+            attention_mask = mask_function(
+                *_non_vmap_expansion_sdpa(batch_arange, head_arange, cache_position, kv_arange)
+            )
+            # Expand the mask to match batch size
+            # and query length if they weren't used in the mask function
+            attention_mask = attention_mask.expand(batch_size, -1, q_length, kv_length)
+
+        # Option 2: Vmap mask creation (torch>=2.6 and custom patterns)
+        # elif _is_torch_greater_or_equal_than_2_6:
+        # This creates the 4D mask easily.
+        # Note that we need this context manager as vmap cannot handle slicing a tensor from
+        # scalar tensor (it internally calls `.item()` which vmap does not allow,
+        # but this context works around it
+        # We don't need to add an offset to the mask_function either,
+        # as we vmap directly the correct indices for k and kv indices
+        #    with TransformGetItemToIndex():
+        #        attention_mask = _vmap_expansion_sdpa(mask_function)(
+        #            batch_arange, head_arange, cache_position, kv_arange
+        #        )
+
+        # Option 3: Error out since it indicates that the user did something custom,
+        # which they shouldn't have (torch<2.6)
+        else:
+            raise ValueError(
+                "The vmap functionality for mask creation "
+                "is only supported from torch>=2.6. "
+                "Please update your torch version or use "
+                "`use_vmap=False` with index-based masks."
+            )
+        return attention_mask

onnx_diagnostic/torch_export_patches/patches/patch_torch.py

@@ -188,6 +188,11 @@ def patched__broadcast_shapes(*_shapes):
     return common_shape
 
 
+def value_ranges_is_positive(value_ranges: torch.utils._sympy.value_ranges.ValueRanges):
+    """Tells if an interval is equivalent to a positive or null integer."""
+    return value_ranges.lower == 0 and value_ranges.upper > 4623372036854775806
+
+
 class patched_ShapeEnv:
 
     def _check_frozen(
@@ -281,7 +286,10 @@ class patched_ShapeEnv:
                     )
                     self._update_var_to_range(b, b_bound, self.var_to_range_sloc[a])
                     tgt_bound = self.bound_sympy(tgt)
-                    assert tgt_bound.issubset(
+                    assert (
+                        value_ranges_is_positive(tgt_bound)
+                        and value_ranges_is_positive(src_bound)
+                    ) or tgt_bound.issubset(
                         src_bound
                     ), f"{tgt_bound=} not a subset of {src_bound=}"
 
@@ -524,8 +532,10 @@ class patched_ShapeEnv:
 
             transmute_into_runtime_assert = False
 
-            backed_var_to_val = getattr(
-                self, "backed_var_to_val", getattr(self, "var_to_val", {})
+            backed_var_to_val = (
+                self.backed_var_to_val
+                if hasattr(self, "backed_var_to_val")
+                else self.var_to_val
             )
             concrete_val = None
             if not (expr.free_symbols <= backed_var_to_val.keys()):

onnx_diagnostic/torch_export_patches/patches/patch_transformers.py

@@ -38,6 +38,7 @@ if patch_masking_utils:
     from ._patch_transformers_masking_utils import (
         patched__vmap_for_bhqkv,
         patched_eager_mask,
+        patched_sdpa_mask,
         patched_sdpa_mask_recent_torch,
     )
 

onnx_diagnostic/torch_export_patches/serialization/transformers_impl.py

@@ -24,10 +24,12 @@ from ...helpers.cache_helper import make_dynamic_cache, make_static_cache, Cache
 from . import make_serialization_function_for_dataclass
 
 SUPPORTED_DATACLASSES: Set[type] = set()
+
 WRONG_REGISTRATIONS = {
     DynamicCache: "4.50",
     BaseModelOutput: None,
 }
+
 SHORTEN_LAYER_NAMES = {
     "DynamicLayer": "D",
     "DynamicSlidingWindowLayer": "W",
@@ -39,6 +41,20 @@ SHORTEN_LAYER_NAMES = {
     "X": "StaticSlidingWindowLayer",
 }
 
+KWARGS_LAYER_NAMES = {
+    "DynamicLayer": lambda layer: "",
+    "DynamicSlidingWindowLayer": lambda layer: str(layer.sliding_window),
+    "StaticLayer": lambda layer: "",
+    "StaticSlidingWindowLayer": lambda layer: str(layer.sliding_window),
+}
+
+PARSE_LAYER_NAMES = {
+    "DynamicLayer": lambda skw: {},
+    "DynamicSlidingWindowLayer": lambda skw: dict(sliding_window=int(skw[1:])),
+    "StaticLayer": lambda skw: {},
+    "StaticSlidingWindowLayer": lambda skw: dict(sliding_window=int(skw[1:])),
+}
+
 
 def _flatten_key_value_cache(cache: Cache) -> Tuple[List[Any], torch.utils._pytree.Context]:
     ca = CacheKeyValue(cache)
@@ -59,7 +75,11 @@ def _flatten_key_value_cache(cache: Cache) -> Tuple[List[Any], torch.utils._pytr
     keys = []
     for i in range(len(ca.key_cache)):
         letter = SHORTEN_LAYER_NAMES[ca.cls_layers[i].__name__]
-        keys.extend([f"key_{letter}{i}", f"value_{letter}{i}"])
+        if hasattr(cache, "layers"):
+            kwargs = KWARGS_LAYER_NAMES[ca.cls_layers[i].__name__](cache.layers[i])
+        else:
+            kwargs = ""
+        keys.extend([f"key_{letter}{kwargs}_{i}", f"value_{letter}{kwargs}_{i}"])
     return flat, keys
 
 
@@ -86,10 +106,16 @@ def _unflatten_cache(
         res = make_cache(list(zip(values[::2], values[1::2])))
     else:
         cls_layer_names = [SHORTEN_LAYER_NAMES[name.split("_")[1][0]] for name in context][::2]
+        cls_kwargs = [
+            PARSE_LAYER_NAMES[SHORTEN_LAYER_NAMES[name.split("_")[1][0]]](name.split("_")[1])
+            for name in context
+        ][::2]
         cls_layers = [
             getattr(transformers.cache_utils, cls_name) for cls_name in cls_layer_names
         ]
-        res = make_cache(list(zip(values[::2], values[1::2])), cls_layers=cls_layers)
+        res = make_cache(
+            list(zip(values[::2], values[1::2])), cls_layers=cls_layers, cls_kwargs=cls_kwargs
+        )
 
     assert output_type is None or isinstance(
         res, output_type

{onnx_diagnostic-0.8.11.dist-info → onnx_diagnostic-0.9.0.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: onnx-diagnostic
-Version: 0.8.11
+Version: 0.9.0
 Summary: Tools to help converting pytorch models into ONNX.
 Home-page: https://github.com/sdpython/onnx-diagnostic
 Author: Xavier Dupré
@@ -91,7 +91,7 @@ Enlightening Examples
 * `Export microsoft/phi-2
   <https://sdpython.github.io/doc/onnx-diagnostic/dev/auto_examples/plot_export_tiny_phi2.html>`_
 * `Export a LLM through method generate (with Tiny-LLM)
-  <https://sdpython.github.io/doc/onnx-diagnostic/dev/auto_examples/plot_export_tiny_llm_method_generate.html>`_
+  <https://sdpython.github.io/doc/onnx-diagnostic/dev/auto_final/plot_export_tiny_llm_method_generate.html>`_
 
 **Torch Export**
 

{onnx_diagnostic-0.8.11.dist-info → onnx_diagnostic-0.9.0.dist-info}/RECORD

@@ -1,4 +1,4 @@
-onnx_diagnostic/__init__.py,sha256=Rvw2q5b1Qw5vuqQ3xWEgU1oyUhLffzhnDfncZoOx6Qc,174
+onnx_diagnostic/__init__.py,sha256=hp8Jeb4rpAUldATxz_DyWOTlzHxT2YqTum8AVG-NRGk,173
 onnx_diagnostic/__main__.py,sha256=YmyV_Aq_ianDlHyKLHMa6h8YK3ZmFPpLVHLKjM91aCk,79
 onnx_diagnostic/_command_lines_parser.py,sha256=nR-WI-15da9e0o-z5mIKezCH4dGFU66gCbyuAFu2FzY,59315
 onnx_diagnostic/doc.py,sha256=-w2qaes0G0TM0PQFTzBVcJPh0r_IR6Vd2j2fj9iV65Y,10478
@@ -6,13 +6,16 @@ onnx_diagnostic/ext_test_case.py,sha256=5WQrMdILgr-j8hrqBs7Vi6IFMD3TXNHD7Dlr9K5y
 onnx_diagnostic/typing.py,sha256=okLjQGA_ikThoMQ1ikjeu-pCi9FiaO8uhSoKz1ufCQ8,409
 onnx_diagnostic/ci_models/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 onnx_diagnostic/ci_models/ci_helpers.py,sha256=lblOF7z2kLcCRAwMOdqp-Tz1EL1oBywHfVokhqiTQRg,15592
-onnx_diagnostic/ci_models/export_phi4_mm.py,sha256=hekQikmMJi4ihUf494zXVTgMFm0l7RXyfklzroObsYs,41646
+onnx_diagnostic/ci_models/export_phi4_mm.py,sha256=anCnx5lt7aGZVrzd4XCJv28-isoCX7feEHHEZDtrW-Y,41735
 onnx_diagnostic/ci_models/export_qwen25_vl.py,sha256=_rYPr8PPraWizr2MPcGuYjrJ55ilJOyKl8kg0wq4L90,20405
+onnx_diagnostic/ci_models/data/Blanca_Lake_Hudak.jpg,sha256=O_mpR1XvIxBpQix0Tgit39Vd1PvAefbPquTJ6ChEwcA,690409
+onnx_diagnostic/ci_models/data/Ice_worm_glacier.jpg,sha256=GkteS7CrGfMQ8Hdyk-ShGtXr5HomFy1mKivJEHXdM2Q,610733
+onnx_diagnostic/ci_models/data/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 onnx_diagnostic/export/__init__.py,sha256=yEIoWiOeTwBsDhyYt2fTKuhtA0Ya1J9u9ZzMTOTWaWs,101
-onnx_diagnostic/export/api.py,sha256=CL2mwRAJU5tv4ujdlLgMLiC5GoqqeheTW4-xdVRry2M,43172
+onnx_diagnostic/export/api.py,sha256=3O21VcaJ0CyZ0X8FwXjPXTCoNfWu1q6iBHeoeDvtOmc,43583
 onnx_diagnostic/export/cf_simple_loop_for.py,sha256=OHPGQc9AC-0TBtCYpP6cm-iHP9gmNt8WYRrPlO9ewlc,21158
 onnx_diagnostic/export/control_flow_onnx.py,sha256=izGlctqQANrHzSxPMbT7hoauNbnIBdx6hb8ry7HtVmM,18263
-onnx_diagnostic/export/dynamic_shapes.py,sha256=_3xnWXa4n7fTkAT0NSLyLkcEJqm896wHKA8ilEyxvm0,44746
+onnx_diagnostic/export/dynamic_shapes.py,sha256=AOfKmGYIs_5hZPy7vi9pZDbpvu-wIlD2cD15lV8rl9E,44763
 onnx_diagnostic/export/onnx_plug.py,sha256=U13fL0BjnhMzcDGxaAOqM4TQte5Z4zKDg4ESS0iktjM,22704
 onnx_diagnostic/export/shape_helper.py,sha256=K3y8-vFXYGg5R1tgeVOm_RjCZ8-yyrvYkQ4b3ILM5H4,10990
 onnx_diagnostic/export/validate.py,sha256=t8NZjgLqLjabKpUDjO5A5nBk_-BAsN1cdq-6UJ0Jm88,6127
@@ -20,25 +23,25 @@ onnx_diagnostic/helpers/__init__.py,sha256=GJ2GT7cgnlIveVUwMZhuvUwidbTJaKv8CsSIO
 onnx_diagnostic/helpers/_log_helper.py,sha256=OTwQH0OIxs9B6nrSvR7MoxMimSw_8mU0mj133NvLk5o,16832
 onnx_diagnostic/helpers/args_helper.py,sha256=qHxXJAM4ovBSYeufVVGMRQwrOW2AuP_HFQgqgU-QBQM,4463
 onnx_diagnostic/helpers/bench_run.py,sha256=9T26icFQOSo3HB1gEw31bf-mD5BrVATrRDhuyWESxz8,16543
-onnx_diagnostic/helpers/cache_helper.py,sha256=p7BB7Ba8n2hAPtpvhEuWJ9h9nQF3TlMQZPiMjJZQe70,32029
+onnx_diagnostic/helpers/cache_helper.py,sha256=INfp9dkGTl3HDEQYbK_fC0T9dMgTuk0pNCNXjwB3Z9w,35033
 onnx_diagnostic/helpers/config_helper.py,sha256=cWRETgFhZ7tayIZPnMqF8BF5AvTU64G2BMqyzgO7lzs,5670
 onnx_diagnostic/helpers/doc_helper.py,sha256=9rigmq36D20yCKG5VkQnODiWMG2eJ0p22UdHew66UrQ,5983
 onnx_diagnostic/helpers/dot_helper.py,sha256=hwgTJsbsUv0qq7euyPDnc1NsBZDGOwv32JXSZxIHJkE,8118
 onnx_diagnostic/helpers/fake_tensor_helper.py,sha256=59046wDIw84or6PJxLaa2CFqaWT7Y3mpYr-BB2shcBE,12027
 onnx_diagnostic/helpers/graph_helper.py,sha256=ailTFdJu1Kir1VQ4GDsm9bUDIevZmgsJVJBGu7CFbCM,14100
-onnx_diagnostic/helpers/helper.py,sha256=7rBu2mQVxQxXrW1YuK433Lp68lvgNOhEqUNkuv9GxJw,66454
+onnx_diagnostic/helpers/helper.py,sha256=hqWZ-VDrnNDehdM8nJN6mblUxZeH-He4dvy378a5LMM,67773
 onnx_diagnostic/helpers/log_helper.py,sha256=njGse2PeD1muXVYvEMl6dsWTDja76wCWipxrEMAkLiw,94504
 onnx_diagnostic/helpers/memory_peak.py,sha256=kigKbD3cq6ZCkE_nhOgNIkr9zHP9zp7ezZhut8GTrcI,6596
 onnx_diagnostic/helpers/mini_onnx_builder.py,sha256=3IAJRf7bphKN0sY714me7H1UzEA1ejJQNYY7cMbGkCo,23878
 onnx_diagnostic/helpers/model_builder_helper.py,sha256=r8LG3xh_-gwp4RODVM9jhB9HNWvccuNypwygWHoU2jA,18548
-onnx_diagnostic/helpers/onnx_helper.py,sha256=1C-ZOsSS0mPHSMT02MD_0kBXpvJTkbsaom45IPiyu_4,64447
+onnx_diagnostic/helpers/onnx_helper.py,sha256=DosgZN1ZU47xA2FN49JCfcmKF9rC8jeO9Vu5icstdOY,64437
 onnx_diagnostic/helpers/optim_helper.py,sha256=0NiYRwV9GLTub4SEny0dqEhLcajRjEhcgkeBDVr9bGQ,4424
-onnx_diagnostic/helpers/ort_session.py,sha256=NNgJZ756f2zXCTkJidzoTQ5HiQLlkzQzk8hYQp9b0qM,30778
-onnx_diagnostic/helpers/rt_helper.py,sha256=YhVB7v5iR_UT9IqCeqyg58IEIx18f_TnCO6u6GhUciQ,38711
+onnx_diagnostic/helpers/ort_session.py,sha256=gYSUTuUFd8vf7Hk48HZ7B73f83ZUeCcTa6BE6ylk8as,30975
+onnx_diagnostic/helpers/rt_helper.py,sha256=ZuWYTI4Nc_z556PXnS5qtpT__YlyTOwn9ocQ1vw-hfY,40352
 onnx_diagnostic/helpers/torch_fx_graph_helper.py,sha256=7xFe4svdbr4gV3OTNcx8eJejjDyHAv4hD_RNNKSxL0c,6571
-onnx_diagnostic/helpers/torch_helper.py,sha256=zy2F3pSKnby_0gHH-1rBiE1dvu6NVp_pLNro560ATgs,39449
+onnx_diagnostic/helpers/torch_helper.py,sha256=WrC6pr7oT7d8P-34FVDrM7ZYtVEx62OaSGlwK2N_7sc,39782
 onnx_diagnostic/investigate/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-onnx_diagnostic/investigate/input_observer.py,sha256=O42MxwjlrFBgb8nmg1UGLRYUBxoVv0JsSkUARxl0HQk,13882
+onnx_diagnostic/investigate/input_observer.py,sha256=vA3I_fDizIIL4m3ETl6qb5qPmmf6QcSYox06FJXOqh0,40576
 onnx_diagnostic/reference/__init__.py,sha256=rLZsxOlnb7-81F2CzepGnZLejaROg4JvgFaGR9FwVQA,208
 onnx_diagnostic/reference/evaluator.py,sha256=5PcPL_5xuEmm4j16KK0o2N5UkdUpuTSxQYc48MPWLpw,8847
 onnx_diagnostic/reference/ort_evaluator.py,sha256=J3keuxFGnQIKP9V80U61y9rpU3zE26RR7HPozeaeiYo,33929
@@ -106,7 +109,7 @@ onnx_diagnostic/tasks/zero_shot_image_classification.py,sha256=jJCMWuOqGv5ahCfjr
 onnx_diagnostic/tasks/data/__init__.py,sha256=uJoemrWgEjI6oA-tMX7r3__x-b3siPmkgqaY7bgIles,401
 onnx_diagnostic/tasks/data/dummies_imagetext2text_generation_gemma3.onnx,sha256=UbtvmWMqcZOKJ-I-HXWI1A6YR6QDaFS5u_yXm5C3ZBw,10299
 onnx_diagnostic/torch_export_patches/__init__.py,sha256=KcBsUXOzw0n-Eo6uFQUUCnJjDsHLnR_J7u-SJu0d6bo,719
-onnx_diagnostic/torch_export_patches/onnx_export_errors.py,sha256=XHYtU7w3vsaTMCuF5X1YtOKxgwL8eEuktXzVZpRz55o,43431
+onnx_diagnostic/torch_export_patches/onnx_export_errors.py,sha256=1q615Li3qQskArEjDT31uB9emeY6ByfHtOO7XIkDWhI,44023
 onnx_diagnostic/torch_export_patches/onnx_export_serialization.py,sha256=KZgr8WbLvmZte1G1hKA0zIlio5ZHz9MKIwGPqNamB6E,12690
 onnx_diagnostic/torch_export_patches/patch_details.py,sha256=UHBo4QTLF3ZgQ4951yYHIQqxOeRYNaG7x56XFcRTtg4,11794
 onnx_diagnostic/torch_export_patches/patch_expressions.py,sha256=VOsv71FsR_UZtxz4-5_VKL2sHQhOkHy9RkPJME2h7UU,3271
@@ -124,18 +127,18 @@ onnx_diagnostic/torch_export_patches/patches/_patch_transformers_funnel.py,sha25
 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=HE3fovyvMiYe9EPz1UjdD9AWopX3H188SMwPb8w5mzM,7111
+onnx_diagnostic/torch_export_patches/patches/_patch_transformers_masking_utils.py,sha256=WD7424vssc_uEpcoG_E5V7edxugVratMQJSng5s3zW8,11808
 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=oYz0tr-6KH0DabpgaISytnXAGxQosoA8gV5LpksO4yI,34834
 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=AQTAAGhpg_Jlzc_26c9rmoZRuhmJcrFX488_W_Z0QD8,16945
 onnx_diagnostic/torch_export_patches/patches/_patch_transformers_sam_mask_decoder.py,sha256=-6TuBm3sLAFEGuW3vRfOTtE5uP6aINFfu7xMnl27Dws,5703
 onnx_diagnostic/torch_export_patches/patches/patch_helper.py,sha256=kK_CGW643iVXxa-m6pttDBS7HTyMQaPypza7iqIInn4,721
-onnx_diagnostic/torch_export_patches/patches/patch_torch.py,sha256=IpvyM5yhA3tioyDAwAEe3gZDvbRUwCI_CzKmm76zwyU,45297
-onnx_diagnostic/torch_export_patches/patches/patch_transformers.py,sha256=1W3iKVYx2QT2xJTKlz1UmtjySuwv-rfT5yVL9DjOfzI,3376
+onnx_diagnostic/torch_export_patches/patches/patch_torch.py,sha256=xoWyNn-hj9d6TW-DVYSUJfhYzJJgpLiG3HZzcQA3k7s,45743
+onnx_diagnostic/torch_export_patches/patches/patch_transformers.py,sha256=UXfHCu-v1uw0CY7XSjPz35GOQKjtdwxKZUOm_ACelNw,3403
 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=divFi7xZcSIm7pERsMg0Eb9a7gDSmXPsdebA_yjoZk4,11530
+onnx_diagnostic/torch_export_patches/serialization/transformers_impl.py,sha256=mBJ8iMngSqI9JJPceq2At9rDWAlRmmRRYFudSl7wgok,12421
 onnx_diagnostic/torch_models/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 onnx_diagnostic/torch_models/code_sample.py,sha256=D0Hv5PJQj2J8Nivoha-RzCdvVhsWYFVFEoG2D2yKm5k,12882
 onnx_diagnostic/torch_models/llms.py,sha256=soyg4yC87ptGoeulJhKqw5opGmuLvH1pn_ZDXZ4Jr8E,90
@@ -154,8 +157,8 @@ onnx_diagnostic/torch_onnx/compare.py,sha256=pXPZ168pT7J_-UOOvsbmS_dCJcwaPjA7Gxi
 onnx_diagnostic/torch_onnx/runtime_info.py,sha256=Mv6WQXNtfNF31zgLm5Hk3x2fTDS6p8UjPh51_Jn4tl0,8565
 onnx_diagnostic/torch_onnx/sbs.py,sha256=xeq4jrMnTeLvZ3UjFlMNpm-VBbP_16fPwTi1_Ujvilo,40752
 onnx_diagnostic/torch_onnx/sbs_dataclasses.py,sha256=WPu9hdz8dIXqvFYy-xtPX9liD_DUVkXOWO_vElOGksY,20277
-onnx_diagnostic-0.8.11.dist-info/licenses/LICENSE.txt,sha256=Vv6TXglX6Rc0d-f8aREhayhT-6PMQXEyOmI2NKlUCMc,1045
-onnx_diagnostic-0.8.11.dist-info/METADATA,sha256=aQ94UbORi24koH-hkkyOjrpCFqnVLYLWXSMrqC_IbOc,6904
-onnx_diagnostic-0.8.11.dist-info/WHEEL,sha256=qELbo2s1Yzl39ZmrAibXA2jjPLUYfnVhUNTlyF1rq0Y,92
-onnx_diagnostic-0.8.11.dist-info/top_level.txt,sha256=KwNkXewmcobM3ZT1DJLVWH6ebJzA5qKg7cWqKfpGNT4,16
-onnx_diagnostic-0.8.11.dist-info/RECORD,,
+onnx_diagnostic-0.9.0.dist-info/licenses/LICENSE.txt,sha256=Vv6TXglX6Rc0d-f8aREhayhT-6PMQXEyOmI2NKlUCMc,1045
+onnx_diagnostic-0.9.0.dist-info/METADATA,sha256=-gBCPiRO-6gtbL_khtaTN8GC_UyZNzlY2PmYUo3lCW4,6900
+onnx_diagnostic-0.9.0.dist-info/WHEEL,sha256=wUyA8OaulRlbfwMtmQsvNngGrxQHAvkKcvRmdizlJi0,92
+onnx_diagnostic-0.9.0.dist-info/top_level.txt,sha256=KwNkXewmcobM3ZT1DJLVWH6ebJzA5qKg7cWqKfpGNT4,16
+onnx_diagnostic-0.9.0.dist-info/RECORD,,

{onnx_diagnostic-0.8.11.dist-info → onnx_diagnostic-0.9.0.dist-info}/WHEEL

@@ -1,5 +1,5 @@
 Wheel-Version: 1.0
-Generator: setuptools (80.10.1)
+Generator: setuptools (80.10.2)
 Root-Is-Purelib: true
 Tag: py3-none-any