xinyun/ComfyUI
1
0
Fork 0
mirror of https://git.datalinker.icu/comfyanonymous/ComfyUI synced 2026-09-09 07:37:03 +08:00
Code Issues Packages Projects Releases Wiki Activity

unlock max length for some text encoders (t5 and llama)

Browse Source
This commit is contained in:
power88 2025-04-20 12:21:48 +08:00
parent fd27494441
commit 1db597f39c
2 changed files with 23 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

22
comfy/sd.py
View File

@ -793,6 +793,7 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
clip_target.clip = comfy.text_encoders.sd3_clip.sd3_clip(clip_l=False, clip_g=True, t5=False) clip_target.clip = comfy.text_encoders.sd3_clip.sd3_clip(clip_l=False, clip_g=True, t5=False)
clip_target.tokenizer = comfy.text_encoders.sd3_clip.SD3Tokenizer clip_target.tokenizer = comfy.text_encoders.sd3_clip.SD3Tokenizer
elif clip_type == CLIPType.HIDREAM: elif clip_type == CLIPType.HIDREAM:
tokenizer_data['clip_g_max_length'] = 77
clip_target.clip = comfy.text_encoders.hidream.hidream_clip(clip_l=False, clip_g=True, t5=False, llama=False, dtype_t5=None, dtype_llama=None, t5xxl_scaled_fp8=None, llama_scaled_fp8=None) clip_target.clip = comfy.text_encoders.hidream.hidream_clip(clip_l=False, clip_g=True, t5=False, llama=False, dtype_t5=None, dtype_llama=None, t5xxl_scaled_fp8=None, llama_scaled_fp8=None)
clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer
else: else:
@ -816,6 +817,7 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
clip_target.tokenizer = comfy.text_encoders.wan.WanT5Tokenizer clip_target.tokenizer = comfy.text_encoders.wan.WanT5Tokenizer
tokenizer_data["spiece_model"] = clip_data[0].get("spiece_model", None) tokenizer_data["spiece_model"] = clip_data[0].get("spiece_model", None)
elif clip_type == CLIPType.HIDREAM: elif clip_type == CLIPType.HIDREAM:
tokenizer_data['t5xxl_max_length'] = 128
clip_target.clip = comfy.text_encoders.hidream.hidream_clip(**t5xxl_detect(clip_data), clip_target.clip = comfy.text_encoders.hidream.hidream_clip(**t5xxl_detect(clip_data),
clip_l=False, clip_g=False, t5=True, llama=False, dtype_llama=None, llama_scaled_fp8=None) clip_l=False, clip_g=False, t5=True, llama=False, dtype_llama=None, llama_scaled_fp8=None)
clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer
@ -836,6 +838,7 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
clip_target.tokenizer = comfy.text_encoders.lumina2.LuminaTokenizer clip_target.tokenizer = comfy.text_encoders.lumina2.LuminaTokenizer
tokenizer_data["spiece_model"] = clip_data[0].get("spiece_model", None) tokenizer_data["spiece_model"] = clip_data[0].get("spiece_model", None)
elif te_model == TEModel.LLAMA3_8: elif te_model == TEModel.LLAMA3_8:
tokenizer_data['llama_max_length'] = 4096
clip_target.clip = comfy.text_encoders.hidream.hidream_clip(**llama_detect(clip_data), clip_target.clip = comfy.text_encoders.hidream.hidream_clip(**llama_detect(clip_data),
clip_l=False, clip_g=False, t5=False, llama=True, dtype_t5=None, t5xxl_scaled_fp8=None) clip_l=False, clip_g=False, t5=False, llama=True, dtype_t5=None, t5xxl_scaled_fp8=None)
clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer
@ -845,6 +848,7 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
clip_target.clip = comfy.text_encoders.sd3_clip.sd3_clip(clip_l=True, clip_g=False, t5=False) clip_target.clip = comfy.text_encoders.sd3_clip.sd3_clip(clip_l=True, clip_g=False, t5=False)
clip_target.tokenizer = comfy.text_encoders.sd3_clip.SD3Tokenizer clip_target.tokenizer = comfy.text_encoders.sd3_clip.SD3Tokenizer
elif clip_type == CLIPType.HIDREAM: elif clip_type == CLIPType.HIDREAM:
tokenizer_data['clip_l_max_length'] = 77
clip_target.clip = comfy.text_encoders.hidream.hidream_clip(clip_l=True, clip_g=False, t5=False, llama=False, dtype_t5=None, dtype_llama=None, t5xxl_scaled_fp8=None, llama_scaled_fp8=None) clip_target.clip = comfy.text_encoders.hidream.hidream_clip(clip_l=True, clip_g=False, t5=False, llama=False, dtype_t5=None, dtype_llama=None, t5xxl_scaled_fp8=None, llama_scaled_fp8=None)
clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer
else: else:
@ -880,6 +884,18 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
t5_kwargs = t5xxl_detect(clip_data) if t5 else {} t5_kwargs = t5xxl_detect(clip_data) if t5 else {}
llama_kwargs = llama_detect(clip_data) if llama else {} llama_kwargs = llama_detect(clip_data) if llama else {}
# Unlock max_length
tokenizer_updates = {
'clip_l': ('clip_l_max_length', 77),
'clip_g': ('clip_g_max_length', 77),
't5': ('t5xxl_max_length', 128),
'llama': ('llama_max_length', 4096),
}
for model_type, (key, max_length) in tokenizer_updates.items():
if locals().get(model_type): # Checks if the variable exists and is truthy
tokenizer_data[key] = max_length
clip_target.clip = comfy.text_encoders.hidream.hidream_clip(clip_l=clip_l, clip_g=clip_g, t5=t5, llama=llama, **t5_kwargs, **llama_kwargs) clip_target.clip = comfy.text_encoders.hidream.hidream_clip(clip_l=clip_l, clip_g=clip_g, t5=t5, llama=llama, **t5_kwargs, **llama_kwargs)
clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer
else: else:
@ -889,6 +905,12 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
clip_target.clip = comfy.text_encoders.sd3_clip.sd3_clip(**t5xxl_detect(clip_data)) clip_target.clip = comfy.text_encoders.sd3_clip.sd3_clip(**t5xxl_detect(clip_data))
clip_target.tokenizer = comfy.text_encoders.sd3_clip.SD3Tokenizer clip_target.tokenizer = comfy.text_encoders.sd3_clip.SD3Tokenizer
elif len(clip_data) == 4: elif len(clip_data) == 4:
tokenizer_data.update({
'clip_l_max_length': 77,
'clip_g_max_length': 77,
't5xxl_max_length': 128,
'llama_max_length': 4096,
})
clip_target.clip = comfy.text_encoders.hidream.hidream_clip(**t5xxl_detect(clip_data), **llama_detect(clip_data)) clip_target.clip = comfy.text_encoders.hidream.hidream_clip(**t5xxl_detect(clip_data), **llama_detect(clip_data))
clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer

1
comfy/text_encoders/hidream.py
View File

@ -9,6 +9,7 @@ import logging
class HiDreamTokenizer: class HiDreamTokenizer:
def __init__(self, embedding_directory=None, tokenizer_data={}): def __init__(self, embedding_directory=None, tokenizer_data={}):
# TODO: Load tokenizers dynamically only when needed, based on the active model.
self.clip_l = sd1_clip.SDTokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data) 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.clip_g = sdxl_clip.SDXLClipGTokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data)
self.t5xxl = sd3_clip.T5XXLTokenizer(embedding_directory=embedding_directory, min_length=128, max_length=128, tokenizer_data=tokenizer_data) self.t5xxl = sd3_clip.T5XXLTokenizer(embedding_directory=embedding_directory, min_length=128, max_length=128, tokenizer_data=tokenizer_data)