wangbo/ComfyUI
1
0
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2025-12-17 01:52:59 +08:00
Code Issues Actions 13 Packages Projects Releases Wiki Activity

Don't try fp8 matrix mult in quantized ops if not supported by hardware. (#10874)

Browse Source
This commit is contained in:
comfyanonymous 2025-11-24 23:55:49 -08:00 committed by GitHub
parent b6805429b9
commit acfaa5c4a1
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 3 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
comfy/ops.py
View File

@ -646,11 +646,12 @@ def mixed_precision_ops(layer_quant_config={}, compute_dtype=torch.bfloat16, ful
return MixedPrecisionOps return MixedPrecisionOps
def pick_operations(weight_dtype, compute_dtype, load_device=None, disable_fast_fp8=False, fp8_optimizations=False, scaled_fp8=None, model_config=None): def pick_operations(weight_dtype, compute_dtype, load_device=None, disable_fast_fp8=False, fp8_optimizations=False, scaled_fp8=None, model_config=None):
fp8_compute = comfy.model_management.supports_fp8_compute(load_device) # TODO: if we support more ops this needs to be more granular
if model_config and hasattr(model_config, 'layer_quant_config') and model_config.layer_quant_config: if model_config and hasattr(model_config, 'layer_quant_config') and model_config.layer_quant_config:
logging.info(f"Using mixed precision operations: {len(model_config.layer_quant_config)} quantized layers") logging.info(f"Using mixed precision operations: {len(model_config.layer_quant_config)} quantized layers")
return mixed_precision_ops(model_config.layer_quant_config, compute_dtype) return mixed_precision_ops(model_config.layer_quant_config, compute_dtype, full_precision_mm=not fp8_compute)
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 and fp8_optimizations, scale_input=fp8_optimizations, override_dtype=scaled_fp8) return scaled_fp8_ops(fp8_matrix_mult=fp8_compute and fp8_optimizations, scale_input=fp8_optimizations, override_dtype=scaled_fp8)