jaidaken/ComfyUI
1
0
Fork 0
Code Issues Pull Requests Actions 1 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

12
comfy/text_encoders/cosmos.py
View File

@@ -7,10 +7,10 @@ from transformers import T5TokenizerFast
class T5XXLModel(sd1_clip.SDClipModel):
def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, attention_mask=True, model_options={}):
textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_old_config_xxl.json")
t5xxl_scaled_fp8 = model_options.get("t5xxl_scaled_fp8", None)
if t5xxl_scaled_fp8 is not None:
t5xxl_quantization_metadata = model_options.get("t5xxl_quantization_metadata", None)
if t5xxl_quantization_metadata is not None:
model_options = model_options.copy()
model_options["scaled_fp8"] = t5xxl_scaled_fp8
model_options["quantization_metadata"] = t5xxl_quantization_metadata
super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, zero_out_masked=attention_mask, model_options=model_options)
@@ -30,12 +30,12 @@ class CosmosT5Tokenizer(sd1_clip.SD1Tokenizer):
super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="t5xxl", tokenizer=T5XXLTokenizer)
def te(dtype_t5=None, t5xxl_scaled_fp8=None):
def te(dtype_t5=None, t5_quantization_metadata=None):
class CosmosTEModel_(CosmosT5XXL):
def __init__(self, device="cpu", dtype=None, model_options={}):
if t5xxl_scaled_fp8 is not None and "t5xxl_scaled_fp8" not in model_options:
if t5_quantization_metadata is not None:
model_options = model_options.copy()
model_options["t5xxl_scaled_fp8"] = t5xxl_scaled_fp8
model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata
if dtype is None:
dtype = dtype_t5
super().__init__(device=device, dtype=dtype, model_options=model_options)