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

14
comfy/sdxl_clip.py
View File

@@ -9,6 +9,7 @@ class SDXLClipG(sd1_clip.SDClipModel):
layer_idx=-2
textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_config_bigg.json")
model_options = {**model_options, "model_name": "clip_g"}
super().__init__(device=device, freeze=freeze, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype,
special_tokens={"start": 49406, "end": 49407, "pad": 0}, layer_norm_hidden_state=False, return_projected_pooled=True, model_options=model_options)
@@ -17,14 +18,13 @@ class SDXLClipG(sd1_clip.SDClipModel):
class SDXLClipGTokenizer(sd1_clip.SDTokenizer):
def __init__(self, tokenizer_path=None, embedding_directory=None, tokenizer_data={}):
super().__init__(tokenizer_path, pad_with_end=False, embedding_directory=embedding_directory, embedding_size=1280, embedding_key='clip_g')
super().__init__(tokenizer_path, pad_with_end=False, embedding_directory=embedding_directory, embedding_size=1280, embedding_key='clip_g', tokenizer_data=tokenizer_data)
class SDXLTokenizer:
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 = SDXLClipGTokenizer(embedding_directory=embedding_directory)
self.clip_l = sd1_clip.SDTokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data)
self.clip_g = SDXLClipGTokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data)
def tokenize_with_weights(self, text:str, return_word_ids=False, **kwargs):
out = {}
@@ -41,8 +41,7 @@ class SDXLTokenizer:
class SDXLClipModel(torch.nn.Module):
def __init__(self, device="cpu", dtype=None, model_options={}):
super().__init__()
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, model_options=model_options)
self.clip_l = sd1_clip.SDClipModel(layer="hidden", layer_idx=-2, device=device, dtype=dtype, layer_norm_hidden_state=False, model_options=model_options)
self.clip_g = SDXLClipG(device=device, dtype=dtype, model_options=model_options)
self.dtypes = set([dtype])
@@ -75,7 +74,7 @@ class SDXLRefinerClipModel(sd1_clip.SD1ClipModel):
class StableCascadeClipGTokenizer(sd1_clip.SDTokenizer):
def __init__(self, tokenizer_path=None, embedding_directory=None, tokenizer_data={}):
super().__init__(tokenizer_path, pad_with_end=True, embedding_directory=embedding_directory, embedding_size=1280, embedding_key='clip_g')
super().__init__(tokenizer_path, pad_with_end=True, embedding_directory=embedding_directory, embedding_size=1280, embedding_key='clip_g', tokenizer_data=tokenizer_data)
class StableCascadeTokenizer(sd1_clip.SD1Tokenizer):
def __init__(self, embedding_directory=None, tokenizer_data={}):
@@ -84,6 +83,7 @@ class StableCascadeTokenizer(sd1_clip.SD1Tokenizer):
class StableCascadeClipG(sd1_clip.SDClipModel):
def __init__(self, device="cpu", max_length=77, freeze=True, layer="hidden", layer_idx=-1, dtype=None, model_options={}):
textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_config_bigg.json")
model_options = {**model_options, "model_name": "clip_g"}
super().__init__(device=device, freeze=freeze, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype,
special_tokens={"start": 49406, "end": 49407, "pad": 49407}, layer_norm_hidden_state=False, enable_attention_masks=True, return_projected_pooled=True, model_options=model_options)