andreas/ComfyUI
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Cleanup and fix issues with text encoder quants. (#10872)

Browse Source
This commit is contained in:
comfyanonymous
2025-11-24 22:48:53 -08:00
committed by GitHub
parent 22a2644e57
commit 25022e0b09
7 changed files with 128 additions and 102 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
comfy/sd1_clip.py
View File

@@ -109,13 +109,23 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
operations = model_options.get("custom_operations", None)
scaled_fp8 = None
quantization_metadata = model_options.get("quantization_metadata", None)
if operations is None:
scaled_fp8 = model_options.get("scaled_fp8", None)
if scaled_fp8 is not None:
operations = comfy.ops.scaled_fp8_ops(fp8_matrix_mult=False, override_dtype=scaled_fp8)
layer_quant_config = None
if quantization_metadata is not None:
layer_quant_config = json.loads(quantization_metadata).get("layers", None)
if layer_quant_config is not None:
operations = comfy.ops.mixed_precision_ops(layer_quant_config, dtype, full_precision_mm=True)
logging.info(f"Using MixedPrecisionOps for text encoder: {len(layer_quant_config)} quantized layers")
else:
operations = comfy.ops.manual_cast
# Fallback to scaled_fp8_ops for backward compatibility
scaled_fp8 = model_options.get("scaled_fp8", None)
if scaled_fp8 is not None:
operations = comfy.ops.scaled_fp8_ops(fp8_matrix_mult=False, override_dtype=scaled_fp8)
else:
operations = comfy.ops.manual_cast
self.operations = operations
self.transformer = model_class(config, dtype, device, self.operations)