Fix for GenerationConfig issue for pipeline init changes by transformers update (#2266)

Co-authored-by: Zhou, Huijuan <huijuan.zhou@intel.com>

Author: 2 people

optimum/habana/transformers/modeling_utils.py

@@ -309,6 +309,7 @@
     gaudi_XLMRoberta_Sdpa_SelfAttention_forward,
 )
 from .models.deepseek_v2.modeling_deepseek_v2 import DeepseekV2ForCausalLM as GaudiDeepseekV2ForCausalLM
+from .pipelines import GaudiImageToTextPipeline
 
 
 def adapt_transformers_to_gaudi():
@@ -388,6 +389,9 @@ def adapt_transformers_to_gaudi():
     transformers.generation.MaxTimeCriteria.__call__ = gaudi_MaxTimeCriteria_call
     transformers.generation.EosTokenCriteria.__call__ = gaudi_EosTokenCriteria_call
     transformers.generation.StoppingCriteriaList.__call__ = gaudi_StoppingCriteriaList_call
+    transformers.pipelines.image_to_text.ImageToTextPipeline._default_generation_config = (
+        GaudiImageToTextPipeline._default_generation_config
+    )
 
     # Optimization for BLOOM generation on Gaudi
     transformers.models.bloom.modeling_bloom.BloomAttention.forward = gaudi_bloom_attention_forward

optimum/habana/transformers/pipelines/__init__.py

@@ -0,0 +1 @@
+from .image_to_text import GaudiImageToTextPipeline

optimum/habana/transformers/pipelines/image_to_text.py

@@ -0,0 +1,9 @@
+from transformers.pipelines.image_to_text import ImageToTextPipeline
+
+from ..generation import GaudiGenerationConfig
+
+
+class GaudiImageToTextPipeline(ImageToTextPipeline):
+    _default_generation_config = GaudiGenerationConfig(
+        max_new_tokens=256,
+    )