onnx-diagnostic 0.7.14__py3-none-any.whl → 0.7.16__py3-none-any.whl

onnx_diagnostic/torch_models/code_sample.py

@@ -0,0 +1,343 @@
+import os
+import textwrap
+import torch
+from typing import Any, Dict, List, Optional, Tuple, Union
+from ..helpers import flatten_object
+from ..helpers.torch_helper import to_any
+from .hghub.model_inputs import _preprocess_model_id
+from .hghub import get_untrained_model_with_inputs
+from .validate import filter_inputs, make_patch_kwargs
+
+
+CODE_SAMPLES = {
+    "imports": "from typing import Any\nimport torch",
+    "get_model_with_inputs": textwrap.dedent(
+        """
+    def get_model_with_inputs(
+        model_id:str,
+        subfolder: str | None = None,
+        dtype: str | torch.dtype | None = None,
+        device: str | torch.device | None = None,
+        same_as_pretrained: bool = False,
+        use_pretrained: bool = False,
+        input_options: dict[str, Any] | None = None,
+        model_options: dict[str, Any] | None = None,
+    ) -> dict[str, Any]:
+        if use_pretrained:
+            import transformers
+            assert same_as_pretrained, (
+                "same_as_pretrained must be True if use_pretrained is True"
+            )
+            # tokenizer = AutoTokenizer.from_pretrained(model_path)
+            model = transformers.AutoModel.from_pretrained(
+                model_id,
+                trust_remote_code=True,
+                subfolder=subfolder,
+                dtype=dtype,
+                device=device,
+            )
+            data = {"model": model}
+            assert not input_options, "Not implemented yet with input_options{input_options}"
+            assert not model_options, "Not implemented yet with input_options{model_options}"
+        else:
+            from onnx_diagnostic.torch_models.hghub import get_untrained_model_with_inputs
+            data = get_untrained_model_with_inputs(
+                model_id,
+                use_pretrained=use_pretrained,
+                same_as_pretrained=same_as_pretrained,
+                inputs_kwargs=input_options,
+                model_kwargs=model_options,
+                subfolder=subfolder,
+                add_second_input=False,
+            )
+            if dtype:
+                data["model"] = data["model"].to(
+                    getattr(torch, dtype) if isinstance(dtype, str) else dtype
+                )
+            if device:
+                data["model"] = data["model"].to(device)
+        return data["model"]
+    """
+    ),
+}
+
+
+def make_code_for_inputs(inputs: Dict[str, torch.Tensor]) -> str:
+    """
+    Creates a code to generate random inputs.
+
+    :param inputs: dictionary
+    :return: code
+    """
+    codes = []
+    for k, v in inputs.items():
+        if isinstance(v, (int, bool, float)):
+            code = f"{k}={v}"
+        elif isinstance(v, torch.Tensor):
+            shape = tuple(map(int, v.shape))
+            if v.dtype in (torch.int32, torch.int64):
+                code = f"{k}=torch.randint({v.max()}, size={shape}, dtype={v.dtype})"
+            elif v.dtype in (torch.float32, torch.float16, torch.bfloat16):
+                code = f"{k}=torch.rand({shape}, dtype={v.dtype})"
+            else:
+                raise ValueError(f"Unexpected dtype = {v.dtype} for k={k!r}")
+        elif v.__class__.__name__ == "DynamicCache":
+            obj = flatten_object(v)
+            cc = [f"torch.rand({tuple(map(int,_.shape))}, dtype={_.dtype})" for _ in obj]
+            va = [f"({a},{b})" for a, b in zip(cc[: len(cc) // 2], cc[len(cc) // 2 :])]
+            va2 = ", ".join(va)
+            code = f"{k}=make_dynamic_cache([{va2}])"
+        else:
+            raise ValueError(f"Unexpected type {type(v)} for k={k!r}")
+        codes.append(code)
+    st = ", ".join(codes)
+    return f"dict({st})"
+
+
+def make_export_code(
+    exporter: str,
+    optimization: Optional[str] = None,
+    patch_kwargs: Optional[Dict[str, Any]] = None,
+    stop_if_static: int = 0,
+    dump_folder: Optional[str] = None,
+    opset: Optional[int] = None,
+    dynamic_shapes: Optional[Dict[str, Any]] = None,
+    output_names: Optional[List[str]] = None,
+    verbose: int = 0,
+) -> Tuple[str, str]:
+    args = [f"dynamic_shapes={dynamic_shapes}"]
+    if output_names:
+        args.append(f"output_names={output_names}")
+    code = []
+    imports = []
+    if dump_folder:
+        code.append(f"os.makedirs({dump_folder!r})")
+        imports.append("import os")
+        filename = os.path.join(dump_folder, "model.onnx")
+    if exporter == "custom":
+        if opset:
+            args.append(f"target_opset={opset}")
+        if optimization:
+            args.append(f"options=OptimizationOptions(patterns={optimization!r})")
+            args.append(f"large_model=True, filename={filename!r}")
+        sargs = ", ".join(args)
+        imports.extend(
+            [
+                "from experimental_experiment.torch_interpreter import to_onnx",
+                "from experimental_experiment.xbuilder import OptimizationOptions",
+            ]
+        )
+        code.extend([f"onx = to_onnx(model, inputs, {sargs})"])
+    elif exporter == "onnx-dynamo":
+        if opset:
+            args.append(f"opset_version={opset}")
+        sargs = ", ".join(args)
+        code.extend([f"epo = torch.onnx.export(model, args=(), kwargs=inputs, {sargs})"])
+        if optimization:
+            imports.append("import onnxscript")
+            code.extend(["onnxscript.optimizer.optimize_ir(epo.model)"])
+            if "os_ort" in optimization:
+                imports.append("import onnxscript.rewriter.ort_fusions as ort_fusions")
+                code.extend(["ort_fusions.optimize_for_ort(epo.model)"])
+        if dump_folder:
+            code.extend([f"epo.save({filename!r})"])
+    else:
+        raise ValueError(f"Unexpected exporter {exporter!r}")
+    if not patch_kwargs:
+        return "\n".join(imports), "\n".join(code)
+
+    imports.append("from onnx_diagnostic.torch_export_patches import torch_export_patches")
+    if stop_if_static:
+        patch_kwargs["stop_if_static"] = stop_if_static
+    sargs = ", ".join(f"{k}={v}" for k, v in patch_kwargs.items())
+    code = [f"with torch_export_patches({sargs}):", *["    " + _ for _ in code]]
+    return "\n".join(imports), "\n".join(code)
+
+
+def code_sample(
+    model_id: str,
+    task: Optional[str] = None,
+    do_run: bool = False,
+    exporter: Optional[str] = None,
+    do_same: bool = False,
+    verbose: int = 0,
+    dtype: Optional[Union[str, torch.dtype]] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    same_as_pretrained: bool = False,
+    use_pretrained: bool = False,
+    optimization: Optional[str] = None,
+    quiet: bool = False,
+    patch: Union[bool, str, Dict[str, bool]] = False,
+    rewrite: bool = False,
+    stop_if_static: int = 1,
+    dump_folder: Optional[str] = None,
+    drop_inputs: Optional[List[str]] = None,
+    input_options: Optional[Dict[str, Any]] = None,
+    model_options: Optional[Dict[str, Any]] = None,
+    subfolder: Optional[str] = None,
+    opset: Optional[int] = None,
+    runtime: str = "onnxruntime",
+    output_names: Optional[List[str]] = None,
+) -> str:
+    """
+    This generates a code to export a model with the proper settings.
+
+    :param model_id: model id to validate
+    :param task: task used to generate the necessary inputs,
+        can be left empty to use the default task for this model
+        if it can be determined
+    :param do_run: checks the model works with the defined inputs
+    :param exporter: exporter the model using this exporter,
+        available list: ``export-strict``, ``export-nostrict``, ...
+        see below
+    :param do_same: checks the discrepancies of the exported model
+    :param verbose: verbosity level
+    :param dtype: uses this dtype to check the model
+    :param device: do the verification on this device
+    :param same_as_pretrained: use a model equivalent to the trained,
+        this is not always possible
+    :param use_pretrained: use the trained model, not the untrained one
+    :param optimization: optimization to apply to the exported model,
+        depend on the the exporter
+    :param quiet: if quiet, catches exception if any issue
+    :param patch: applies patches (``patch_transformers=True, path_diffusers=True``)
+        if True before exporting
+        see :func:`onnx_diagnostic.torch_export_patches.torch_export_patches`,
+        a string can be used to specify only one of them
+    :param rewrite: applies known rewriting (``patch_transformers=True``) before exporting,
+        see :func:`onnx_diagnostic.torch_export_patches.torch_export_patches`
+    :param stop_if_static: stops if a dynamic dimension becomes static,
+        see :func:`onnx_diagnostic.torch_export_patches.torch_export_patches`
+    :param dump_folder: dumps everything in a subfolder of this one
+    :param drop_inputs: drops this list of inputs (given their names)
+    :param input_options: additional options to define the dummy inputs
+        used to export
+    :param model_options: additional options when creating the model such as
+        ``num_hidden_layers`` or ``attn_implementation``
+    :param subfolder: version or subfolders to uses when retrieving a model id
+    :param opset: onnx opset to use for the conversion
+    :param runtime: onnx runtime to use to check about discrepancies,
+        possible values ``onnxruntime``, ``torch``, ``orteval``,
+        ``orteval10``, ``ref`` only if `do_run` is true
+    :param output_names: output names the onnx exporter should use
+    :return: a code
+
+    .. runpython::
+        :showcode:
+
+        from onnx_diagnostic.torch_models.code_sample import code_sample
+
+        print(
+            code_sample(
+                "arnir0/Tiny-LLM",
+                exporter="onnx-dynamo",
+                optimization="ir",
+                patch=True,
+            )
+        )
+    """
+    model_id, subfolder, same_as_pretrained, use_pretrained = _preprocess_model_id(
+        model_id,
+        subfolder,
+        same_as_pretrained=same_as_pretrained,
+        use_pretrained=use_pretrained,
+    )
+    patch_kwargs = make_patch_kwargs(patch=patch, rewrite=rewrite)
+
+    iop = input_options or {}
+    mop = model_options or {}
+    data = get_untrained_model_with_inputs(
+        model_id,
+        verbose=verbose,
+        task=task,
+        use_pretrained=use_pretrained,
+        same_as_pretrained=same_as_pretrained,
+        inputs_kwargs=iop,
+        model_kwargs=mop,
+        subfolder=subfolder,
+        add_second_input=False,
+    )
+    if drop_inputs:
+        update = {}
+        for k in data:
+            if k.startswith("inputs"):
+                update[k], ds = filter_inputs(
+                    data[k],
+                    drop_names=drop_inputs,
+                    model=data["model"],
+                    dynamic_shapes=data["dynamic_shapes"],
+                )
+        update["dynamic_shapes"] = ds
+        data.update(update)
+
+    update = {}
+    for k in data:
+        if k.startswith("inputs"):
+            v = data[k]
+            if dtype:
+                update[k] = v = to_any(
+                    v, getattr(torch, dtype) if isinstance(dtype, str) else dtype
+                )
+            if device:
+                update[k] = v = to_any(v, device)
+    if update:
+        data.update(update)
+
+    args = [f"{model_id!r}"]
+    if subfolder:
+        args.append(f"subfolder={subfolder!r}")
+    if dtype:
+        args.append(f"dtype={dtype!r}")
+    if device:
+        args.append(f"device={device!r}")
+    if same_as_pretrained:
+        args.append(f"same_as_pretrained={same_as_pretrained!r}")
+    if use_pretrained:
+        args.append(f"use_pretrained={use_pretrained!r}")
+    if input_options:
+        args.append(f"input_options={input_options!r}")
+    if model_options:
+        args.append(f"model_options={model_options!r}")
+    model_args = ", ".join(args)
+    imports, exporter_code = (
+        make_export_code(
+            exporter=exporter,
+            patch_kwargs=patch_kwargs,
+            verbose=verbose,
+            optimization=optimization,
+            stop_if_static=stop_if_static,
+            dump_folder=dump_folder,
+            opset=opset,
+            dynamic_shapes=data["dynamic_shapes"],
+        )
+        if exporter is not None
+        else ([], [])
+    )
+    input_code = make_code_for_inputs(data["inputs"])
+    cache_import = (
+        "from onnx_diagnostic.helpers.cache_helper import make_dynamic_cache"
+        if "dynamic_cache" in input_code
+        else ""
+    )
+
+    pieces = [
+        CODE_SAMPLES["imports"],
+        imports,
+        cache_import,
+        CODE_SAMPLES["get_model_with_inputs"],
+        textwrap.dedent(
+            f"""
+            model = get_model_with_inputs({model_args})
+                        """
+        ),
+        f"inputs = {input_code}",
+        exporter_code,
+    ]
+    code = "\n".join(pieces)  # type: ignore[arg-type]
+    try:
+        import black
+    except ImportError:
+        # No black formatting.
+        return code
+
+    return black.format_str(code, mode=black.Mode())

onnx_diagnostic/torch_models/hghub/hub_data_cached_configs.py

@@ -4865,3 +4865,41 @@ def _ccached_google_gemma_3_4b_it_like():
             },
         }
     )
+
+
+def _ccached_hf_internal_testing_tiny_random_gemma3_for_causal_lm():
+    "hf-internal-testing/tiny-random-Gemma3ForCausalLM"
+    return transformers.Gemma3TextConfig(
+        **{
+            "architectures": ["Gemma3ForCausalLM"],
+            "attention_bias": false,
+            "attention_dropout": 0.0,
+            "attn_logit_softcapping": null,
+            "bos_token_id": 2,
+            "cache_implementation": "hybrid",
+            "eos_token_id": [1, 106],
+            "final_logit_softcapping": null,
+            "head_dim": 8,
+            "hidden_activation": "gelu_pytorch_tanh",
+            "hidden_size": 16,
+            "initializer_range": 0.02,
+            "intermediate_size": 32,
+            "max_position_embeddings": 32768,
+            "model_type": "gemma3_text",
+            "num_attention_heads": 2,
+            "num_hidden_layers": 2,
+            "num_key_value_heads": 1,
+            "pad_token_id": 0,
+            "query_pre_attn_scalar": 256,
+            "rms_norm_eps": 1e-06,
+            "rope_local_base_freq": 10000,
+            "rope_scaling": null,
+            "rope_theta": 1000000,
+            "sliding_window": 512,
+            "sliding_window_pattern": 6,
+            "torch_dtype": "float32",
+            "transformers_version": "4.52.0.dev0",
+            "use_cache": true,
+            "vocab_size": 262144,
+        }
+    )

onnx_diagnostic/torch_models/hghub/model_inputs.py

@@ -95,6 +95,8 @@ def get_untrained_model_with_inputs(
         print("-- dynamic shapes:", pprint.pformat(data['dynamic_shapes']))
         print("-- configuration:", pprint.pformat(data['configuration']))
     """
+    if task == "":
+        task = None
     assert not use_preinstalled or not use_only_preinstalled, (
         f"model_id={model_id!r}, preinstalled model is only available "
         f"if use_only_preinstalled is False."
@@ -120,14 +122,16 @@ def get_untrained_model_with_inputs(
             **(model_kwargs or {}),
         )
 
-    model, task, mkwargs, diff_config = None, None, {}, None
+    model, task_, mkwargs, diff_config = None, None, {}, None
     if use_pretrained and same_as_pretrained:
         if model_id in HANDLED_MODELS:
-            model, task, config = load_specific_model(model_id, verbose=verbose)
+            model, task_, config = load_specific_model(model_id, verbose=verbose)
 
+    if task is None:
+        task = task_
     if model is None:
         arch = architecture_from_config(config)
-        if arch is None:
+        if task is None and arch is None:
             task = task_from_id(model_id, subfolder=subfolder)
         assert task is not None or arch is not None, (
             f"Unable to determine the architecture for model {model_id!r}, "

onnx_diagnostic/torch_models/validate.py

@@ -4,7 +4,7 @@ import inspect
 import os
 import pprint
 import sys
-from typing import Any, Callable, Dict, List, Optional, Sequence, Tuple, Union
+from typing import Any, Callable, Dict, List, Optional, Sequence, Set, Tuple, Union
 import time
 import numpy as np
 import onnx
@@ -117,11 +117,21 @@ def _make_folder_name(
     drop_inputs: Optional[List[str]] = None,
     same_as_pretrained: bool = False,
     use_pretrained: bool = False,
+    task: Optional[str] = None,
 ) -> str:
     "Creates a filename unique based on the given options."
     els = [model_id.replace("/", "_")]
     if subfolder:
         els.append(subfolder.replace("/", "_"))
+    if task:
+        els.append(task)
+    if drop_inputs:
+        ii = "-".join(f"{s[0]}{s[-1]}" for s in drop_inputs)
+        els.append(f"I-{ii.upper()}")
+    if use_pretrained:
+        els.append("TRAINED")
+    elif same_as_pretrained:
+        els.append("SAMESIZE")
     if exporter:
         els.append(exporter)
     if optimization:
@@ -142,14 +152,7 @@ def _make_folder_name(
         els.append(sdev)
     if opset is not None:
         els.append(f"op{opset}")
-    if drop_inputs:
-        ii = "-".join(f"{s[0]}{s[-1]}" for s in drop_inputs)
-        els.append(f"I-{ii.upper()}")
-    if use_pretrained:
-        els.append("TRAINED")
-    elif same_as_pretrained:
-        els.append("SAMESIZE")
-    return "-".join(els)
+    return "/".join([e for e in els if e])
 
 
 def version_summary() -> Dict[str, Union[int, float, str]]:
@@ -290,6 +293,33 @@ def shrink_config(cfg: Dict[str, Any]) -> Dict[str, Any]:
     return new_cfg
 
 
+def make_patch_kwargs(
+    patch: Union[bool, str, Dict[str, bool]] = False,
+    rewrite: bool = False,
+) -> Dict[str, Any]:
+    """Creates patch arguments."""
+    default_patch = dict(patch_transformers=True, patch_diffusers=True, patch=True)
+    if isinstance(patch, bool):
+        patch_kwargs = default_patch if patch else dict(patch=False)
+    elif isinstance(patch, str):
+        patch_kwargs = {"patch": True, **{p: True for p in patch.split(",")}}  # noqa: C420
+    else:
+        assert isinstance(patch, dict), f"Unable to interpret patch={patch!r}"
+        patch_kwargs = patch.copy()
+        if "patch" not in patch_kwargs:
+            if any(patch_kwargs.values()):
+                patch_kwargs["patch"] = True
+        elif len(patch) == 1 and patch.get("patch", False):
+            patch_kwargs.update(default_patch)
+
+    assert not rewrite or patch_kwargs.get("patch", False), (
+        f"rewrite={rewrite}, patch={patch}, patch_kwargs={patch_kwargs} "
+        f"patch must be True to enable rewriting, "
+        f"if --patch=0 was specified on the command line, rewrites are disabled."
+    )
+    return patch_kwargs
+
+
 def validate_model(
     model_id: str,
     task: Optional[str] = None,
@@ -319,6 +349,7 @@ def validate_model(
     inputs2: int = 1,
     output_names: Optional[List[str]] = None,
     ort_logs: bool = False,
+    quiet_input_sets: Optional[Set[str]] = None,
 ) -> Tuple[Dict[str, Union[int, float, str]], Dict[str, Any]]:
     """
     Validates a model.
@@ -373,6 +404,8 @@ def validate_model(
         or an empty cache for example
     :param output_names: output names the onnx exporter should use
     :param ort_logs: increases onnxruntime verbosity when creating the session
+    :param quiet_input_sets: avoid raising an exception if the inputs belongs to that set
+        even if quiet is False
     :return: two dictionaries, one with some metrics,
         another one with whatever the function produces
 
@@ -414,25 +447,8 @@ def validate_model(
         use_pretrained=use_pretrained,
     )
     time_preprocess_model_id = time.perf_counter() - main_validation_begin
-    default_patch = dict(patch_transformers=True, patch_diffusers=True, patch=True)
-    if isinstance(patch, bool):
-        patch_kwargs = default_patch if patch else dict(patch=False)
-    elif isinstance(patch, str):
-        patch_kwargs = {"patch": True, **{p: True for p in patch.split(",")}}  # noqa: C420
-    else:
-        assert isinstance(patch, dict), f"Unable to interpret patch={patch!r}"
-        patch_kwargs = patch.copy()
-        if "patch" not in patch_kwargs:
-            if any(patch_kwargs.values()):
-                patch_kwargs["patch"] = True
-        elif len(patch) == 1 and patch.get("patch", False):
-            patch_kwargs.update(default_patch)
+    patch_kwargs = make_patch_kwargs(patch=patch, rewrite=rewrite)
 
-    assert not rewrite or patch_kwargs.get("patch", False), (
-        f"rewrite={rewrite}, patch={patch}, patch_kwargs={patch_kwargs} "
-        f"patch must be True to enable rewriting, "
-        f"if --patch=0 was specified on the command line, rewrites are disabled."
-    )
     summary = version_summary()
     summary.update(
         dict(
@@ -473,6 +489,7 @@ def validate_model(
             drop_inputs=drop_inputs,
             use_pretrained=use_pretrained,
             same_as_pretrained=same_as_pretrained,
+            task=task,
         )
         dump_folder = os.path.join(dump_folder, folder_name)
         if not os.path.exists(dump_folder):
@@ -487,6 +504,8 @@ def validate_model(
             print(f"[validate_model] validate model id {model_id!r}, subfolder={subfolder!r}")
         else:
             print(f"[validate_model] validate model id {model_id!r}")
+        if task:
+            print(f"[validate_model] with task {task!r}")
         print(f"[validate_model] patch={patch!r}")
         if model_options:
             print(f"[validate_model] model_options={model_options!r}")
@@ -762,6 +781,10 @@ def validate_model(
             ep = data["exported_program"]
             if verbose:
                 print(f"[validate_model] -- dumps exported program in {dump_folder!r}...")
+            assert isinstance(
+                folder_name, str
+            ), f"folder_name={folder_name!r} should be a string"
+            folder_name = folder_name.replace("/", "-")
             with open(os.path.join(dump_folder, f"{folder_name}.ep"), "w") as f:
                 f.write(str(ep))
             torch.export.save(ep, os.path.join(dump_folder, f"{folder_name}.pt2"))
@@ -770,6 +793,10 @@ def validate_model(
             if verbose:
                 print("[validate_model] done (dump ep)")
         if "onnx_program" in data:
+            assert isinstance(
+                folder_name, str
+            ), f"folder_name={folder_name!r} should be a string"
+            folder_name = folder_name.replace("/", "-")
             epo = data["onnx_program"]
             if verbose:
                 print(f"[validate_model] dumps onnx program in {dump_folder!r}...")
@@ -842,6 +869,7 @@ def validate_model(
             warmup=warmup,
             second_input_keys=second_input_keys,
             ort_logs=ort_logs,
+            quiet_input_sets=quiet_input_sets,
         )
         summary.update(summary_valid)
         summary["time_total_validation_onnx"] = time.perf_counter() - validation_begin
@@ -904,6 +932,7 @@ def validate_model(
                     repeat=repeat,
                     warmup=warmup,
                     second_input_keys=second_input_keys,
+                    quiet_input_sets=quiet_input_sets,
                 )
                 summary.update(summary_valid)
 
@@ -1289,6 +1318,7 @@ def validate_onnx_model(
     warmup: int = 0,
     second_input_keys: Optional[List[str]] = None,
     ort_logs: bool = False,
+    quiet_input_sets: Optional[Set[str]] = None,
 ) -> Tuple[Dict[str, Any], Dict[str, Any]]:
     """
     Verifies that an onnx model produces the same
@@ -1308,6 +1338,7 @@ def validate_onnx_model(
         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)
     :param ort_logs: triggers the logs for onnxruntime
+    :param quiet_input_sets: avoid raising an exception for these sets of inputs
     :return: two dictionaries, one with some metrics,
         another one with whatever the function produces
     """
@@ -1431,6 +1462,8 @@ def validate_onnx_model(
     keys = [("inputs", "run_expected", "")]
     if second_input_keys:
         keys.extend([(k, f"run_expected2{k[6:]}", f"2{k[6:]}") for k in second_input_keys])
+    if verbose:
+        print(f"[validate_onnx_model] -- keys={keys}")
     for k_input, k_expected, suffix in keys:
         # make_feeds
         assert k_input in data, f"Unable to find {k_input!r} in {sorted(data)}"
@@ -1455,10 +1488,12 @@ def validate_onnx_model(
 
         # run ort
         if verbose:
-            print("[validate_onnx_model] run session...")
+            print(f"[validate_onnx_model] run session on inputs 'inputs{suffix}'...")
+            if quiet_input_sets and f"inputs{suffix}" in quiet_input_sets:
+                print(f"[validate_onnx_model] quiet_input_sets={quiet_input_sets}")
 
         got = _quiet_or_not_quiet(
-            quiet,
+            quiet or (quiet_input_sets is not None and f"inputs{suffix}" in quiet_input_sets),
             _mk(f"run_onnx_ort{suffix}"),
             summary,
             data,
@@ -1865,6 +1900,7 @@ def call_torch_export_custom(
         "custom-nostrict-all-noinline",
         "custom-dec",
         "custom-decall",
+        "custom-fake",
     }
     assert exporter in available, f"Unexpected value for exporter={exporter!r} in {available}"
     assert "model" in data, f"model is missing from data: {sorted(data)}"
@@ -1873,6 +1909,14 @@ def call_torch_export_custom(
     strict = "-strict" in exporter
     args, kwargs = split_args_kwargs(data["inputs_export"])
     ds = data.get("dynamic_shapes", None)
+    if "-fake" in exporter:
+        from onnx_diagnostic.export.shape_helper import make_fake_with_dynamic_dimensions
+
+        if verbose:
+            print("[call_torch_export_custom] switching to FakeTensor")
+        assert not args, f"Exporter {exporter!r} not implemented with fake tensors."
+        kwargs = torch_deepcopy(kwargs)
+        kwargs, _ = make_fake_with_dynamic_dimensions(kwargs, dynamic_shapes=ds)
     opset = data.get("model_opset", None)
     if opset:
         summary["export_opset"] = opset

{onnx_diagnostic-0.7.14.dist-info → onnx_diagnostic-0.7.16.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: onnx-diagnostic
-Version: 0.7.14
+Version: 0.7.16
 Summary: Tools to help converting pytorch models into ONNX.
 Home-page: https://github.com/sdpython/onnx-diagnostic
 Author: Xavier Dupré
@@ -60,9 +60,9 @@ You need then to remove those which are not dynamic in your model.
 
 .. code-block:: python
 
-  from onnx_diagnostic.export.shape_helper import all_dynamic_shape_from_inputs
+  from onnx_diagnostic.export.shape_helper import all_dynamic_shapes_from_inputs
 
-  dynamic_shapes = all_dynamic_shape_from_inputs(cache)
+  dynamic_shapes = all_dynamic_shapes_from_inputs(cache)
 
 It also implements tools to investigate, validate exported models (ExportedProgramm, ONNXProgram, ...).
 See `documentation of onnx-diagnostic <https://sdpython.github.io/doc/onnx-diagnostic/dev/>`_ and
@@ -126,13 +126,13 @@ Snapshot of usefuls tools
       ep = torch.export.export(model, args, kwargs=kwargs, dynamic_shapes=dynamic_shapes)
       # ...
 
-**all_dynamic_shape_from_inputs**
+**all_dynamic_shapes_from_inputs**
 
 .. code-block:: python
 
-  from onnx_diagnostic.export.shape_helper import all_dynamic_shape_from_inputs
+  from onnx_diagnostic.export.shape_helper import all_dynamic_shapes_from_inputs
 
-  dynamic_shapes = all_dynamic_shape_from_inputs(cache)
+  dynamic_shapes = all_dynamic_shapes_from_inputs(cache)
 
 **torch_export_rewrite**