298977887/ComfyUI
1
0
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2025-03-14 21:47:07 +00:00
Code Issues Packages Projects Releases Wiki Activity

Support fp8_scaled diffusion models that don't use fp8 matrix mult.

Browse Source
This commit is contained in:
comfyanonymous 2025-03-07 04:37:58 -05:00
parent e62d72e8ca
commit e1474150de
3 changed files with 8 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
comfy/model_base.py
View File

@ -108,7 +108,7 @@ class BaseModel(torch.nn.Module):
if not unet_config.get("disable_unet_model_creation", False): if not unet_config.get("disable_unet_model_creation", False):
if model_config.custom_operations is None: if model_config.custom_operations is None:
fp8 = model_config.optimizations.get("fp8", model_config.scaled_fp8 is not None) fp8 = model_config.optimizations.get("fp8", False)
operations = comfy.ops.pick_operations(unet_config.get("dtype", None), self.manual_cast_dtype, fp8_optimizations=fp8, scaled_fp8=model_config.scaled_fp8) operations = comfy.ops.pick_operations(unet_config.get("dtype", None), self.manual_cast_dtype, fp8_optimizations=fp8, scaled_fp8=model_config.scaled_fp8)
else: else:
operations = model_config.custom_operations operations = model_config.custom_operations

4
comfy/model_detection.py
View File

@ -471,6 +471,10 @@ def model_config_from_unet(state_dict, unet_key_prefix, use_base_if_no_match=Fal
model_config.scaled_fp8 = scaled_fp8_weight.dtype model_config.scaled_fp8 = scaled_fp8_weight.dtype
if model_config.scaled_fp8 == torch.float32: if model_config.scaled_fp8 == torch.float32:
model_config.scaled_fp8 = torch.float8_e4m3fn model_config.scaled_fp8 = torch.float8_e4m3fn
if scaled_fp8_weight.nelement() == 2:
model_config.optimizations["fp8"] = False
else:
model_config.optimizations["fp8"] = True
return model_config return model_config

4
comfy/ops.py
View File

@ -17,6 +17,7 @@
""" """
import torch import torch
import logging
import comfy.model_management import comfy.model_management
from comfy.cli_args import args, PerformanceFeature from comfy.cli_args import args, PerformanceFeature
import comfy.float import comfy.float
@ -308,6 +309,7 @@ class fp8_ops(manual_cast):
return torch.nn.functional.linear(input, weight, bias) return torch.nn.functional.linear(input, weight, bias)
def scaled_fp8_ops(fp8_matrix_mult=False, scale_input=False, override_dtype=None): def scaled_fp8_ops(fp8_matrix_mult=False, scale_input=False, override_dtype=None):
logging.info("Using scaled fp8: fp8 matrix mult: {}, scale input: {}".format(fp8_matrix_mult, scale_input))
class scaled_fp8_op(manual_cast): class scaled_fp8_op(manual_cast):
class Linear(manual_cast.Linear): class Linear(manual_cast.Linear):
def __init__(self, *args, **kwargs): def __init__(self, *args, **kwargs):
@ -358,7 +360,7 @@ def scaled_fp8_ops(fp8_matrix_mult=False, scale_input=False, override_dtype=None
def pick_operations(weight_dtype, compute_dtype, load_device=None, disable_fast_fp8=False, fp8_optimizations=False, scaled_fp8=None): def pick_operations(weight_dtype, compute_dtype, load_device=None, disable_fast_fp8=False, fp8_optimizations=False, scaled_fp8=None):
fp8_compute = comfy.model_management.supports_fp8_compute(load_device) fp8_compute = comfy.model_management.supports_fp8_compute(load_device)
if scaled_fp8 is not None: if scaled_fp8 is not None:
return scaled_fp8_ops(fp8_matrix_mult=fp8_compute, scale_input=True, override_dtype=scaled_fp8) return scaled_fp8_ops(fp8_matrix_mult=fp8_compute and fp8_optimizations, scale_input=True, override_dtype=scaled_fp8)
if ( if (
fp8_compute and fp8_compute and