torch_patch.py

"""
Patch for torch module to make it compatible with newer diffusers versions 
while using PyTorch 2.0.1
"""
import torch
import sys
import warnings
import types
import functools

# Check if the attributes already exist
if not hasattr(torch, 'float8_e4m3fn'):
    # Add missing attributes for compatibility
    # These won't actually function, but they'll allow imports to succeed
    torch.float8_e4m3fn = torch.float16  # Use float16 as a placeholder type
    warnings.warn(
        "Added placeholder for torch.float8_e4m3fn. Actual 8-bit operations won't work, "
        "but imports should succeed. Using PyTorch 2.0.1 with newer diffusers."
    )

if not hasattr(torch, 'float8_e5m2'):
    torch.float8_e5m2 = torch.float16  # Use float16 as a placeholder type

# Add other missing torch types that might be referenced
for type_name in ['bfloat16', 'bfloat8', 'float8_e4m3fnuz']:
    if not hasattr(torch, type_name):
        setattr(torch, type_name, torch.float16)
        
# Create a placeholder for torch._dynamo if it doesn't exist
if not hasattr(torch, '_dynamo'):
    torch._dynamo = types.ModuleType('torch._dynamo')
    sys.modules['torch._dynamo'] = torch._dynamo
    
    # Add common attributes/functions used by torch._dynamo
    torch._dynamo.config = types.SimpleNamespace(suppress_errors=True)
    torch._dynamo.optimize = lambda *args, **kwargs: lambda f: f
    torch._dynamo.disable = lambda: None
    torch._dynamo.reset_repro_cache = lambda: None

# Add torch.compile if it doesn't exist
if not hasattr(torch, 'compile'):
    # Just return the function unchanged
    torch.compile = lambda fn, **kwargs: fn

# Create a placeholder for torch.cuda.amp if it doesn't exist
if not hasattr(torch.cuda, 'amp'):
    torch.cuda.amp = types.ModuleType('torch.cuda.amp')
    sys.modules['torch.cuda.amp'] = torch.cuda.amp
    
    # Mock autocast
    class MockAutocast:
        def __init__(self, *args, **kwargs):
            pass
        def __enter__(self):
            return self
        def __exit__(self, *args):
            pass
        def __call__(self, func):
            @functools.wraps(func)
            def wrapper(*args, **kwargs):
                return func(*args, **kwargs)
            return wrapper
    
    torch.cuda.amp.autocast = MockAutocast

print("PyTorch patched for compatibility with newer diffusers - using latest diffusers with PyTorch 2.0.1")