mirror of
https://github.com/comfyanonymous/ComfyUI.git
synced 2025-04-19 10:53:29 +00:00
3e8155f7a3
Add clip g long clip support. Text encoder refactor. Support llama models with different vocab sizes.
30 lines
1.1 KiB
Python
30 lines
1.1 KiB
Python
from comfy import sd1_clip
|
|
import os
|
|
|
|
|
|
def model_options_long_clip(sd, tokenizer_data, model_options):
|
|
w = sd.get("clip_l.text_model.embeddings.position_embedding.weight", None)
|
|
if w is None:
|
|
w = sd.get("clip_g.text_model.embeddings.position_embedding.weight", None)
|
|
else:
|
|
model_name = "clip_g"
|
|
|
|
if w is None:
|
|
w = sd.get("text_model.embeddings.position_embedding.weight", None)
|
|
if w is not None:
|
|
if "text_model.encoder.layers.30.mlp.fc1.weight" in sd:
|
|
model_name = "clip_g"
|
|
elif "text_model.encoder.layers.1.mlp.fc1.weight" in sd:
|
|
model_name = "clip_l"
|
|
else:
|
|
model_name = "clip_l"
|
|
|
|
if w is not None:
|
|
tokenizer_data = tokenizer_data.copy()
|
|
model_options = model_options.copy()
|
|
model_config = model_options.get("model_config", {})
|
|
model_config["max_position_embeddings"] = w.shape[0]
|
|
model_options["{}_model_config".format(model_name)] = model_config
|
|
tokenizer_data["{}_max_length".format(model_name)] = w.shape[0]
|
|
return tokenizer_data, model_options
|