Check on bigger inputs

_unittests/ut_torch_models/test_hghub_model.py

@@ -17,7 +17,7 @@ class TestHuggingFaceHubModel(ExtTestCase):
     @hide_stdout()
     def test_get_untrained_model_with_inputs_tiny_llm(self):
         mid = "arnir0/Tiny-LLM"
-        data = get_untrained_model_with_inputs(mid, verbose=1)
+        data = get_untrained_model_with_inputs(mid, verbose=1, add_second_input=0)
         self.assertEqual(
             set(data),
             {

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/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