79 lines
2.8 KiB
Python
79 lines
2.8 KiB
Python
from antlr4.error.ErrorListener import ErrorListener
|
|
from antlr4 import InputStream
|
|
from .Parser.MathExprLexer import MathExprLexer
|
|
from .Parser.MathExprParser import MathExprParser
|
|
from .Parser.TensorEvalVisitor import TensorEvalVisitor
|
|
from antlr4 import CommonTokenStream
|
|
import torch
|
|
from .helper_functions import getIndexTensorAlongDim, ThrowingErrorListener
|
|
import comfy.utils
|
|
|
|
def calculate_patches(Model, a, b=None, c=None, d=None, w=0.0, x=0.0, y=0.0, z=0.0):
|
|
# Parse expression
|
|
input_stream = InputStream(Model)
|
|
lexer = MathExprLexer(input_stream)
|
|
stream = CommonTokenStream(lexer)
|
|
parser = MathExprParser(stream)
|
|
parser.addErrorListener(ThrowingErrorListener())
|
|
tree = parser.expr()
|
|
|
|
sd_a = a.model.state_dict()
|
|
sd_b = b.model.state_dict() if b is not None else {}
|
|
sd_c = c.model.state_dict() if c is not None else {}
|
|
sd_d = d.model.state_dict() if d is not None else {}
|
|
|
|
patches = {}
|
|
layer_count = len(sd_a)
|
|
|
|
pbar = comfy.utils.ProgressBar(layer_count)
|
|
|
|
# Iterate over all keys in the main model 'a'
|
|
for i, (key, tens_a) in enumerate(sd_a.items()):
|
|
# Get corresponding tensors from other models, defaulting to zeros if missing or models not provided
|
|
tens_b = sd_b.get(key, None)
|
|
if tens_b is None: tens_b = torch.zeros_like(tens_a,device=tens_a.device)
|
|
else: tens_b = tens_b.to(tens_a.device)
|
|
|
|
tens_c = sd_c.get(key, None)
|
|
if tens_c is None: tens_c = torch.zeros_like(tens_a,device=tens_a.device)
|
|
else: tens_c = tens_c.to(tens_a.device)
|
|
|
|
tens_d = sd_d.get(key, None)
|
|
if tens_d is None: tens_d = torch.zeros_like(tens_a,device=tens_a.device)
|
|
else: tens_d = tens_d.to(tens_a.device)
|
|
|
|
# Variables for the visitor
|
|
variables = {
|
|
'a': tens_a,
|
|
'b': tens_b,
|
|
'c': tens_c,
|
|
'd': tens_d,
|
|
'w': w, 'x': x, 'y': y, 'z': z,
|
|
'L': i, 'layer': i,
|
|
'LC': layer_count, 'layer_count': layer_count
|
|
}
|
|
|
|
for dim_idx in range(tens_a.ndim):
|
|
idx_tensor = getIndexTensorAlongDim(tens_a, dim_idx)
|
|
idx_tensor = idx_tensor.to(tens_a.device)
|
|
variables[f'D{dim_idx}'] = idx_tensor
|
|
variables[f'dim_{dim_idx}'] = idx_tensor
|
|
|
|
visitor = TensorEvalVisitor(variables, tens_a.shape)
|
|
result_tensor = visitor.visit(tree)
|
|
|
|
# Calculate difference for patching
|
|
# The patch should be: result - original
|
|
# Because ComfyUI applies: original + patch
|
|
diff = result_tensor - tens_a
|
|
|
|
# Allow skipping zero patches to save memory
|
|
if torch.all(diff == 0):
|
|
continue
|
|
|
|
# Store patch. ComfyUI expects { key: (tensor,) } usually
|
|
patches[key] = (diff,)
|
|
|
|
pbar.update(1)
|
|
|
|
return patches |