Fix gpt bigcode ONNX export for transformers<4.39.0 (#1973)

* fix gpt bigcode ONNX export for transformers<=4.36.0

* fix gpt bigcode ONNX export for transformers<4.39.0

* bump min transformers version

* fix min transformers version

* should be done in another pr

optimum/exporters/onnx/model_configs.py

@@ -279,7 +279,7 @@ class LlamaOnnxConfig(TextDecoderWithPositionIdsOnnxConfig):
 
 
 class Qwen2OnnxConfig(LlamaOnnxConfig):
-    pass
+    MIN_TRANSFORMERS_VERSION = version.parse("4.37.0")
 
 
 class GemmaOnnxConfig(LlamaOnnxConfig):
@@ -291,6 +291,7 @@ class GemmaOnnxConfig(LlamaOnnxConfig):
 class PhiOnnxConfig(TextDecoderWithPositionIdsOnnxConfig):
     DEFAULT_ONNX_OPSET = 14  # Phi now uses F.scaled_dot_product_attention by default for torch>=2.1.1.
     NORMALIZED_CONFIG_CLASS = NormalizedTextConfig
+    MIN_TRANSFORMERS_VERSION = version.parse("4.36.0")
 
 
 class Phi3OnnxConfig(PhiOnnxConfig):
@@ -299,6 +300,7 @@ class Phi3OnnxConfig(PhiOnnxConfig):
     ) + TextDecoderOnnxConfig.DUMMY_INPUT_GENERATOR_CLASSES
     DUMMY_PKV_GENERATOR_CLASS = MistralDummyPastKeyValuesGenerator
     NORMALIZED_CONFIG_CLASS = NormalizedTextConfigWithGQA
+    MIN_TRANSFORMERS_VERSION = version.parse("4.41.0")
 
 
 class MistralOnnxConfig(TextDecoderWithPositionIdsOnnxConfig):
@@ -1173,7 +1175,7 @@ def outputs(self) -> Dict[str, Dict[int, str]]:
 
 
 class OwlV2OnnxConfig(OwlViTOnnxConfig):
-    pass
+    MIN_TRANSFORMERS_VERSION = version.parse("4.35.0")
 
 
 class LayoutLMOnnxConfig(TextAndVisionOnnxConfig):

optimum/exporters/onnx/model_patcher.py

@@ -276,13 +276,16 @@ def __init__(
             model.decoder.model.decoder.config.use_cache = True
 
 
-def _unmask_unattended_patched(
-    expanded_mask: torch.Tensor,
-    min_dtype: float,
+def _unmask_unattended_patched_legacy(
+    expanded_mask: torch.Tensor, attention_mask: torch.Tensor, unmasked_value: Union[bool, float]
 ):
     return expanded_mask
 
 
+def _unmask_unattended_patched(expanded_mask: torch.Tensor, min_dtype: float):
+    return expanded_mask
+
+
 def _make_causal_mask_patched(
     input_ids_shape: torch.Size,
     dtype: torch.dtype,
@@ -316,7 +319,11 @@ def _make_causal_mask_patched(
 
 
 _make_causal_mask_patched_staticmethod = staticmethod(_make_causal_mask_patched)
-_unmask_unattended_patched_staticmethod = staticmethod(_unmask_unattended_patched)
+
+if _transformers_version >= version.parse("4.39.0"):
+    _unmask_unattended_patched_staticmethod = staticmethod(_unmask_unattended_patched)
+else:
+    _unmask_unattended_patched_staticmethod = staticmethod(_unmask_unattended_patched_legacy)
 
 
 # Adapted from _prepare_4d_causal_attention_mask

setup.py

@@ -15,7 +15,7 @@
 REQUIRED_PKGS = [
     "coloredlogs",
     "sympy",
-    "transformers[sentencepiece]>=4.26.0,<4.43.0",
+    "transformers[sentencepiece]>=4.29.0,<4.43.0",
     "torch>=1.11",
     "packaging",
     "numpy<2.0",  # transformers requires numpy<2.0 https://github.com/huggingface/transformers/pull/31569