comfyanonymous/ComfyUI
1
0
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2026-01-26 11:09:50 +00:00
Code Issues Packages Projects Releases Wiki Activity

Make old scaled fp8 format use the new mixed quant ops system. (#11000)

Browse Source
This commit is contained in:
comfyanonymous
2025-12-05 11:35:42 -08:00
committed by GitHub
parent 0ec05b1481
commit 43071e3de3
24 changed files with 278 additions and 275 deletions
Download Patch File Download Diff File Expand all files Collapse all files

33
comfy/model_detection.py
View File

@@ -6,20 +6,6 @@ import math
import logging
import torch
def detect_layer_quantization(metadata):
quant_key = "_quantization_metadata"
if metadata is not None and quant_key in metadata:
quant_metadata = metadata.pop(quant_key)
quant_metadata = json.loads(quant_metadata)
if isinstance(quant_metadata, dict) and "layers" in quant_metadata:
logging.info(f"Found quantization metadata (version {quant_metadata.get('format_version', 'unknown')})")
return quant_metadata["layers"]
else:
raise ValueError("Invalid quantization metadata format")
return None
def count_blocks(state_dict_keys, prefix_string):
count = 0
while True:
@@ -767,22 +753,11 @@ def model_config_from_unet(state_dict, unet_key_prefix, use_base_if_no_match=Fal
if model_config is None and use_base_if_no_match:
model_config = comfy.supported_models_base.BASE(unet_config)
scaled_fp8_key = "{}scaled_fp8".format(unet_key_prefix)
if scaled_fp8_key in state_dict:
scaled_fp8_weight = state_dict.pop(scaled_fp8_key)
model_config.scaled_fp8 = scaled_fp8_weight.dtype
if model_config.scaled_fp8 == torch.float32:
model_config.scaled_fp8 = torch.float8_e4m3fn
if scaled_fp8_weight.nelement() == 2:
model_config.optimizations["fp8"] = False
else:
model_config.optimizations["fp8"] = True
# Detect per-layer quantization (mixed precision)
layer_quant_config = detect_layer_quantization(metadata)
if layer_quant_config:
model_config.layer_quant_config = layer_quant_config
logging.info(f"Detected mixed precision quantization: {len(layer_quant_config)} layers quantized")
quant_config = comfy.utils.detect_layer_quantization(state_dict, unet_key_prefix)
if quant_config:
model_config.quant_config = quant_config
logging.info("Detected mixed precision quantization")
return model_config