xinyun/ComfyUI
1
0
Fork 0
mirror of https://git.datalinker.icu/comfyanonymous/ComfyUI synced 2026-08-15 05:10:02 +08:00
Code Issues Packages Projects Releases Wiki Activity

Qwen VL WAR

Browse Source
This commit is contained in:
lspindler 2025-08-25 10:33:44 +02:00
parent 1404916962
commit f4bdd45496
1 changed files with 8 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
comfy/ops.py
View File

@ -43,7 +43,10 @@ def dequantizer(x: torch.Tensor, scale: torch.Tensor, dtype: torch.dtype):
def woq_fwd(self, x): def woq_fwd(self, x):
dq_weight = self.dequantizer(self.weight, self.scale_weight, x.dtype) dq_weight = self.dequantizer(self.weight, self.scale_weight, x.dtype)
return torch.nn.functional.linear(x, dq_weight, self.bias) bias = self.bias
if bias is not None and bias.dtype != x.dtype:
bias = self.dequantizer(bias, self.scale_weight, x.dtype)
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
@ -80,7 +83,6 @@ def get_dequantizer_fn(scale_weight, scale_input):
def scaled_dot_product_attention(q, k, v, *args, **kwargs): def scaled_dot_product_attention(q, k, v, *args, **kwargs):
return torch.nn.functional.scaled_dot_product_attention(q, k, v, *args, **kwargs) return torch.nn.functional.scaled_dot_product_attention(q, k, v, *args, **kwargs)
try: try:
if torch.cuda.is_available(): if torch.cuda.is_available():
from torch.nn.attention import SDPBackend, sdpa_kernel from torch.nn.attention import SDPBackend, sdpa_kernel
@ -199,8 +201,10 @@ class disable_weight_init:
scale_input = state_dict.get(f"{prefix}scale_input", None) scale_input = state_dict.get(f"{prefix}scale_input", None)
if scale_input is not None: if scale_input is not None:
self.register_buffer('scale_input', scale_input.to(device=self.device, dtype=torch.float32)) if scale_input != 1: # TODO not really nice but e.g. Qwen VL has an input scale but does not use it?
self.register_buffer('scale_input', scale_input.to(device=self.device, dtype=torch.float32))
else:
scale_input = None
self.forward = types.MethodType(get_quantized_forward(scale_weight, scale_input), self) self.forward = types.MethodType(get_quantized_forward(scale_weight, scale_input), self)
setattr(self, "quantizer", get_quantizer_fn(scale_weight, scale_input)) setattr(self, "quantizer", get_quantizer_fn(scale_weight, scale_input))
setattr(self, "dequantizer", get_dequantizer_fn(scale_weight, scale_input)) setattr(self, "dequantizer", get_dequantizer_fn(scale_weight, scale_input))