wangbo/ComfyUI
1
0
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2026-01-12 15:20:51 +08:00
Code Issues Actions 9 Packages Projects Releases Wiki Activity

Merge branch 'comfyanonymous:master' into master

Browse Source
This commit is contained in:
patientx 2024-08-22 00:07:14 +03:00 committed by GitHub
commit febf8601dc
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 22 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
comfy/ldm/flux/layers.py
View File

@ -178,7 +178,7 @@ class DoubleStreamBlock(nn.Module):
txt += txt_mod2.gate * self.txt_mlp((1 + txt_mod2.scale) * self.txt_norm2(txt) + txt_mod2.shift)
if txt.dtype == torch.float16:
txt = txt.clip(-65504, 65504)
txt = torch.nan_to_num(txt, nan=0.0, posinf=65504, neginf=-65504)
return img, txt
@ -233,7 +233,7 @@ class SingleStreamBlock(nn.Module):
output = self.linear2(torch.cat((attn, self.mlp_act(mlp)), 2))
x += mod.gate * output
if x.dtype == torch.float16:
x = x.clip(-65504, 65504)
x = torch.nan_to_num(x, nan=0.0, posinf=65504, neginf=-65504)
return x

1
comfy/model_management.py
View File

@ -682,6 +682,7 @@ def unet_manual_cast(weight_dtype, inference_device, supported_dtypes=[torch.flo
if bf16_supported and weight_dtype == torch.bfloat16:
return None
fp16_supported = should_use_fp16(inference_device, prioritize_performance=True)
for dt in supported_dtypes:
if dt == torch.float16 and fp16_supported:
return torch.float16

21
comfy/ops.py
View File

@ -254,16 +254,33 @@ def fp8_linear(self, input):
non_blocking = comfy.model_management.device_supports_non_blocking(input.device)
w = cast_to(self.weight, device=input.device, non_blocking=non_blocking).t()
scale_weight = self.scale_weight
scale_input = self.scale_input
if scale_weight is None:
scale_weight = torch.ones((1), device=input.device, dtype=torch.float32)
if scale_input is None:
scale_input = scale_weight
if scale_input is None:
scale_input = torch.ones((1), device=input.device, dtype=torch.float32)
if self.bias is not None:
o, _ = torch._scaled_mm(inn, w, out_dtype=input.dtype, bias=cast_to_input(self.bias, input, non_blocking=non_blocking))
o = torch._scaled_mm(inn, w, out_dtype=input.dtype, bias=cast_to_input(self.bias, input, non_blocking=non_blocking), scale_a=scale_input, scale_b=scale_weight)
else:
o, _ = torch._scaled_mm(inn, w, out_dtype=input.dtype)
o = torch._scaled_mm(inn, w, out_dtype=input.dtype, scale_a=scale_input, scale_b=scale_weight)
if isinstance(o, tuple):
o = o[0]
return o.reshape((-1, input.shape[1], self.weight.shape[0]))
return None
class fp8_ops(manual_cast):
class Linear(manual_cast.Linear):
def reset_parameters(self):
self.scale_weight = None
self.scale_input = None
return None
def forward_comfy_cast_weights(self, input):
out = fp8_linear(self, input)
if out is not None: