Merge pull request #368 from Aintor/main

Fix for Mask Patch Failure and Quantization Issues in Latest `transformers` Versions

Author: aseemw

python_coreml_stable_diffusion/torch2coreml.py

@@ -358,6 +358,8 @@ def bundle_resources_for_swift_cli(args):
 from transformers.models.clip import modeling_clip
 
 # Copied from https://github.com/huggingface/transformers/blob/v4.30.0/src/transformers/models/clip/modeling_clip.py#L677C1-L692C1
+# Starting from transformers >= 4.35.0, the _make_causal_mask function is replaced by _create_4d_causal_attention_mask in modeling_clip.
+# For backward compatibility with versions < 4.35.0, both functions are patched here.
 def patched_make_causal_mask(input_ids_shape, dtype, device, past_key_values_length: int = 0):
     """ Patch to replace torch.finfo(dtype).min with -1e4
     """
@@ -370,8 +372,9 @@ def patched_make_causal_mask(input_ids_shape, dtype, device, past_key_values_len
     if past_key_values_length > 0:
         mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype, device=device), mask], dim=-1)
     return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length)
-
-modeling_clip._make_causal_mask = patched_make_causal_mask
+    
+modeling_clip._make_causal_mask = patched_make_causal_mask # For transformers >= 4.30.0 and transformers < 4.35.0
+modeling_clip._create_4d_causal_attention_mask = patched_make_causal_mask # For transformers >= 4.35.0
 
 def convert_text_encoder(text_encoder, tokenizer, submodule_name, args):
     """ Converts the text encoder component of Stable Diffusion