xinyun/ComfyUI
1
0
Fork 0
mirror of https://git.datalinker.icu/comfyanonymous/ComfyUI synced 2026-08-15 17:36:44 +08:00
Code Issues Packages Projects Releases Wiki Activity

Ensure backward compatabilty with 3rd party tools like GGUF

Browse Source
This commit is contained in:
lspindler 2025-09-08 13:01:18 +02:00
parent fdeb63df91
commit b5f30790f7
2 changed files with 11 additions and 18 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
comfy/ops.py
View File

@ -411,7 +411,6 @@ op_class_list = [
def operator_factory(**factory_kwargs): def operator_factory(**factory_kwargs):
class OpSet: class OpSet:
pass pass
op_set = OpSet()
for k, v in factory_kwargs.items(): for k, v in factory_kwargs.items():
assert hasattr(CastWeightBiasOp, k) assert hasattr(CastWeightBiasOp, k)

28
comfy/quant_tensor.py
View File

@ -1,20 +1,10 @@
import torch import torch
import logging
from abc import ABC, abstractmethod
from dataclasses import dataclass
from typing import Tuple
from torch.utils._triton import has_triton
from typing import Dict
Q_TYPES = [torch.float8_e4m3fn] Q_TYPES = [torch.float8_e4m3fn, torch.float8_e5m2]
if has_triton():
q_compile_decorator = torch.compile()
else:
q_compile_decorator = lambda func: func
def get_quantizer_with_constraints(target_dtype: torch.dtype): def get_quantizer_with_constraints(target_dtype: torch.dtype):
if target_dtype == torch.float8_e4m3fn: if target_dtype in Q_TYPES:
q_fn = dynamic_tensor_quantizer q_fn = dynamic_tensor_quantizer
else: else:
raise ValueError(f"Unsupported dtype {target_dtype}") raise ValueError(f"Unsupported dtype {target_dtype}")
@ -24,26 +14,29 @@ def get_quantizer_with_constraints(target_dtype: torch.dtype):
def fn(x, **kwargs): def fn(x, **kwargs):
if alignment_check_fn(x): if alignment_check_fn(x):
return x, None return x, None
return q_fn(x, **kwargs) if x.dtype == target_dtype:
return x, None
return q_fn(x, dtype=target_dtype, **kwargs)
return fn return fn
@q_compile_decorator
def dynamic_tensor_quantizer(x: torch.Tensor, dtype=torch.dtype, *args, **kwargs): def dynamic_tensor_quantizer(x: torch.Tensor, dtype: torch.dtype, *args, **kwargs):
input_scale = torch.abs(x).max() / torch.finfo(dtype).max input_scale = torch.abs(x).max() / torch.finfo(dtype).max
x = (x / input_scale).clamp(torch.finfo(dtype).min, torch.finfo(dtype).max).to(dtype=dtype) x = (x / input_scale).clamp(torch.finfo(dtype).min, torch.finfo(dtype).max).to(dtype=dtype)
return x, input_scale.float() return x, input_scale.float()
@q_compile_decorator
def tensor_quantizer(x: torch.Tensor, scale: torch.Tensor, dtype: torch.dtype): def tensor_quantizer(x: torch.Tensor, scale: torch.Tensor, dtype: torch.dtype):
x = (x / scale).clamp(torch.finfo(dtype).min, torch.finfo(dtype).max).to(dtype=dtype).contiguous() x = (x / scale).clamp(torch.finfo(dtype).min, torch.finfo(dtype).max).to(dtype=dtype).contiguous()
return x, scale.float() return x, scale.float()
@q_compile_decorator
def tensor_dequantizer(x: torch.Tensor, scale: torch.Tensor, dtype: torch.dtype): def tensor_dequantizer(x: torch.Tensor, scale: torch.Tensor, dtype: torch.dtype):
x = x.to(dtype=dtype) * scale.to(dtype=dtype) x = x.to(dtype=dtype) * scale.to(dtype=dtype)
return x return x
def woq_fwd(self, x): def woq_fwd(self, x):
dq_weight = self.dequantizer(self.weight, scale=self.scale_weight, dtype=x.dtype) dq_weight = self.dequantizer(self.weight, scale=self.scale_weight, dtype=x.dtype)
bias = self.bias bias = self.bias
@ -51,6 +44,7 @@ def woq_fwd(self, x):
bias = self.dequantizer(bias, torch.ones_like(self.scale_weight), x.dtype) bias = self.dequantizer(bias, torch.ones_like(self.scale_weight), x.dtype)
return torch.nn.functional.linear(x, dq_weight, bias) return torch.nn.functional.linear(x, dq_weight, bias)
def quantized_fwd(self, input): def quantized_fwd(self, input):
tensor_2d = False tensor_2d = False
if len(input.shape) == 2: if len(input.shape) == 2: