byt5_patch.py

#ByT5 Compatibility Patch for imagen-pytorch

import sys

# Replace T5Tokenizer with AutoTokenizer in transformers
from transformers import AutoTokenizer

class T5TokenizerReplacement:
    #Replacement for T5Tokenizer that uses AutoTokenizer
    
    @classmethod
    def from_pretrained(cls, pretrained_model_name_or_path, *args, **kwargs):
        # Use AutoTokenizer which works with ByT5
        return AutoTokenizer.from_pretrained(pretrained_model_name_or_path, *args, **kwargs)

# Replace T5Tokenizer before transformers.models.t5 is imported
import transformers.models.t5.tokenization_t5 as t5_module
t5_module.T5Tokenizer = T5TokenizerReplacement

print("[PATCH] Replaced T5Tokenizer with AutoTokenizer (ByT5 compatible)")

# MPS compatibility for project function
def patch_imagen_project():
    
    try:
        import imagen_pytorch.imagen_pytorch as imagen_module
        
        def project_mps_compatible(x, y):
            #Use float32 instead of float64 for MPS
            import torch
            original_dtype = x.dtype
            x_compute = x.float() if x.device.type == "mps" else x.double()
            y_compute = y.float() if y.device.type == "mps" else y.double()
            
            x_compute = torch.nn.functional.normalize(x_compute, dim=-1)
            parallel = (x_compute * y_compute).sum(dim=-1, keepdim=True) * x_compute
            orthogonal = y_compute - parallel
            
            return parallel.to(original_dtype), orthogonal.to(original_dtype)
        
        imagen_module.project = project_mps_compatible
        print("[PATCH] Applied MPS compatibility patch for project function")
    except ImportError:
        pass


patch_imagen_project()