jinx/ComfyUI
1
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2025-08-03 07:26:31 +08:00
Code Issues Activity

More flexible long clip support.

Browse Source 
Add clip g long clip support.

Text encoder refactor.

Support llama models with different vocab sizes.
This commit is contained in:
comfyanonymous
2025-04-15 10:32:21 -04:00
parent 8a438115fb
commit 3e8155f7a3
17 changed files with 95 additions and 66 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
comfy/text_encoders/hydit.py
View File

@@ -9,24 +9,26 @@ import torch
class HyditBertModel(sd1_clip.SDClipModel):
def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, model_options={}):
textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "hydit_clip.json")
model_options = {**model_options, "model_name": "hydit_clip"}
super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"start": 101, "end": 102, "pad": 0}, model_class=BertModel, enable_attention_masks=True, return_attention_masks=True, model_options=model_options)
class HyditBertTokenizer(sd1_clip.SDTokenizer):
def __init__(self, embedding_directory=None, tokenizer_data={}):
tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "hydit_clip_tokenizer")
super().__init__(tokenizer_path, pad_with_end=False, embedding_size=1024, embedding_key='chinese_roberta', tokenizer_class=BertTokenizer, pad_to_max_length=False, max_length=512, min_length=77)
super().__init__(tokenizer_path, pad_with_end=False, embedding_size=1024, embedding_key='chinese_roberta', tokenizer_class=BertTokenizer, pad_to_max_length=False, max_length=512, min_length=77, tokenizer_data=tokenizer_data)
class MT5XLModel(sd1_clip.SDClipModel):
def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, model_options={}):
textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "mt5_config_xl.json")
model_options = {**model_options, "model_name": "mt5xl"}
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=True, return_attention_masks=True, model_options=model_options)
class MT5XLTokenizer(sd1_clip.SDTokenizer):
def __init__(self, embedding_directory=None, tokenizer_data={}):
#tokenizer_path = os.path.join(os.path.join(os.path.dirname(os.path.realpath(__file__)), "mt5_tokenizer"), "spiece.model")
tokenizer = tokenizer_data.get("spiece_model", None)
super().__init__(tokenizer, pad_with_end=False, embedding_size=2048, embedding_key='mt5xl', tokenizer_class=SPieceTokenizer, has_start_token=False, pad_to_max_length=False, max_length=99999999, min_length=256)
super().__init__(tokenizer, pad_with_end=False, embedding_size=2048, embedding_key='mt5xl', tokenizer_class=SPieceTokenizer, has_start_token=False, pad_to_max_length=False, max_length=99999999, min_length=256, tokenizer_data=tokenizer_data)
def state_dict(self):
return {"spiece_model": self.tokenizer.serialize_model()}
@@ -35,7 +37,7 @@ class HyditTokenizer:
def __init__(self, embedding_directory=None, tokenizer_data={}):
mt5_tokenizer_data = tokenizer_data.get("mt5xl.spiece_model", None)
self.hydit_clip = HyditBertTokenizer(embedding_directory=embedding_directory)
self.mt5xl = MT5XLTokenizer(tokenizer_data={"spiece_model": mt5_tokenizer_data}, embedding_directory=embedding_directory)
self.mt5xl = MT5XLTokenizer(tokenizer_data={**tokenizer_data, "spiece_model": mt5_tokenizer_data}, embedding_directory=embedding_directory)
def tokenize_with_weights(self, text:str, return_word_ids=False, **kwargs):
out = {}