ryan/ComfyUI
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki 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

15
comfy/text_encoders/sd3_clip.py
View File

@@ -15,6 +15,7 @@ class T5XXLModel(sd1_clip.SDClipModel):
model_options = model_options.copy()
model_options["scaled_fp8"] = t5xxl_scaled_fp8
model_options = {**model_options, "model_name": "t5xxl"}
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, model_options=model_options)
@@ -31,17 +32,16 @@ def t5_xxl_detect(state_dict, prefix=""):
return out
class T5XXLTokenizer(sd1_clip.SDTokenizer):
def __init__(self, embedding_directory=None, tokenizer_data={}):
def __init__(self, embedding_directory=None, tokenizer_data={}, min_length=77):
tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_tokenizer")
super().__init__(tokenizer_path, embedding_directory=embedding_directory, pad_with_end=False, embedding_size=4096, embedding_key='t5xxl', tokenizer_class=T5TokenizerFast, has_start_token=False, pad_to_max_length=False, max_length=99999999, min_length=77)
super().__init__(tokenizer_path, embedding_directory=embedding_directory, pad_with_end=False, embedding_size=4096, embedding_key='t5xxl', tokenizer_class=T5TokenizerFast, has_start_token=False, pad_to_max_length=False, max_length=99999999, min_length=min_length, tokenizer_data=tokenizer_data)
class SD3Tokenizer:
def __init__(self, embedding_directory=None, tokenizer_data={}):
clip_l_tokenizer_class = tokenizer_data.get("clip_l_tokenizer_class", sd1_clip.SDTokenizer)
self.clip_l = clip_l_tokenizer_class(embedding_directory=embedding_directory)
self.clip_g = sdxl_clip.SDXLClipGTokenizer(embedding_directory=embedding_directory)
self.t5xxl = T5XXLTokenizer(embedding_directory=embedding_directory)
self.clip_l = sd1_clip.SDTokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data)
self.clip_g = sdxl_clip.SDXLClipGTokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data)
self.t5xxl = T5XXLTokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data)
def tokenize_with_weights(self, text:str, return_word_ids=False, **kwargs):
out = {}
@@ -61,8 +61,7 @@ class SD3ClipModel(torch.nn.Module):
super().__init__()
self.dtypes = set()
if clip_l:
clip_l_class = model_options.get("clip_l_class", sd1_clip.SDClipModel)
self.clip_l = clip_l_class(layer="hidden", layer_idx=-2, device=device, dtype=dtype, layer_norm_hidden_state=False, return_projected_pooled=False, model_options=model_options)
self.clip_l = sd1_clip.SDClipModel(layer="hidden", layer_idx=-2, device=device, dtype=dtype, layer_norm_hidden_state=False, return_projected_pooled=False, model_options=model_options)
self.dtypes.add(dtype)
else:
self.clip_l = None