mirrors/ComfyUI
1
0
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2025-09-12 04:27:21 +00:00
Code Issues Packages Projects Releases Wiki Activity

Make it possible to load tokenizer data from checkpoints.

Browse Source
This commit is contained in:
comfyanonymous
2024-07-24 16:43:53 -04:00
parent ce80e69fb8
commit 10c919f4c7
8 changed files with 26 additions and 31 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
comfy/text_encoders/aura_t5.py
View File

@@ -9,13 +9,13 @@ class PT5XlModel(sd1_clip.SDClipModel):
super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 2, "pad": 1}, model_class=comfy.text_encoders.t5.T5, enable_attention_masks=True, zero_out_masked=True)
class PT5XlTokenizer(sd1_clip.SDTokenizer):
def __init__(self, embedding_directory=None):
def __init__(self, embedding_directory=None, tokenizer_data={}):
tokenizer_path = os.path.join(os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_pile_tokenizer"), "tokenizer.model")
super().__init__(tokenizer_path, pad_with_end=False, embedding_size=2048, embedding_key='pile_t5xl', tokenizer_class=SPieceTokenizer, has_start_token=False, pad_to_max_length=False, max_length=99999999, min_length=256, pad_token=1)
class AuraT5Tokenizer(sd1_clip.SD1Tokenizer):
def __init__(self, embedding_directory=None):
super().__init__(embedding_directory=embedding_directory, clip_name="pile_t5xl", tokenizer=PT5XlTokenizer)
def __init__(self, embedding_directory=None, tokenizer_data={}):
super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="pile_t5xl", tokenizer=PT5XlTokenizer)
class AuraT5Model(sd1_clip.SD1ClipModel):
def __init__(self, device="cpu", dtype=None, **kwargs):

6
comfy/text_encoders/sa_t5.py
View File

@@ -9,13 +9,13 @@ class T5BaseModel(sd1_clip.SDClipModel):
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, zero_out_masked=True)
class T5BaseTokenizer(sd1_clip.SDTokenizer):
def __init__(self, embedding_directory=None):
def __init__(self, embedding_directory=None, tokenizer_data={}):
tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_tokenizer")
super().__init__(tokenizer_path, pad_with_end=False, embedding_size=768, embedding_key='t5base', tokenizer_class=T5TokenizerFast, has_start_token=False, pad_to_max_length=False, max_length=99999999, min_length=128)
class SAT5Tokenizer(sd1_clip.SD1Tokenizer):
def __init__(self, embedding_directory=None):
super().__init__(embedding_directory=embedding_directory, clip_name="t5base", tokenizer=T5BaseTokenizer)
def __init__(self, embedding_directory=None, tokenizer_data={}):
super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="t5base", tokenizer=T5BaseTokenizer)
class SAT5Model(sd1_clip.SD1ClipModel):
def __init__(self, device="cpu", dtype=None, **kwargs):

13
comfy/text_encoders/sd3_clip.py
View File

@@ -13,22 +13,13 @@ class T5XXLModel(sd1_clip.SDClipModel):
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)
class T5XXLTokenizer(sd1_clip.SDTokenizer):
def __init__(self, embedding_directory=None):
def __init__(self, embedding_directory=None, tokenizer_data={}):
tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_tokenizer")
super().__init__(tokenizer_path, 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)
class SDT5XXLTokenizer(sd1_clip.SD1Tokenizer):
def __init__(self, embedding_directory=None):
super().__init__(embedding_directory=embedding_directory, clip_name="t5xxl", tokenizer=T5XXLTokenizer)
class SDT5XXLModel(sd1_clip.SD1ClipModel):
def __init__(self, device="cpu", dtype=None, **kwargs):
super().__init__(device=device, dtype=dtype, clip_name="t5xxl", clip_model=T5XXLModel, **kwargs)
class SD3Tokenizer:
def __init__(self, embedding_directory=None):
def __init__(self, embedding_directory=None, tokenizer_data={}):
self.clip_l = sd1_clip.SDTokenizer(embedding_directory=embedding_directory)
self.clip_g = sdxl_clip.SDXLClipGTokenizer(embedding_directory=embedding_directory)
self.t5xxl = T5XXLTokenizer(embedding_directory=embedding_directory)

4
comfy/text_encoders/spiece_tokenizer.py
View File

@@ -1,4 +1,5 @@
import os
import torch
class SPieceTokenizer:
add_eos = True
@@ -9,6 +10,9 @@ class SPieceTokenizer:
def __init__(self, tokenizer_path):
import sentencepiece
if torch.is_tensor(tokenizer_path):
tokenizer_path = tokenizer_path.numpy().tobytes()
if isinstance(tokenizer_path, bytes):
self.tokenizer = sentencepiece.SentencePieceProcessor(model_proto=tokenizer_path, add_eos=self.add_eos)
else: