Karl/comfy_fm24_newgens
1
0
Fork 0
mirror of https://github.com/karl0ss/comfy_fm24_newgens.git synced 2025-10-24 12:13:58 +01:00
Code Issues Packages Projects Releases Wiki Activity

text chunking

Browse Source
This commit is contained in:
Karl 2025-09-23 15:49:06 +01:00
parent fd999ec1e6
commit 650369b06f
1 changed files with 10 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
lib/text_chunker.py
View File

@ -8,12 +8,12 @@ class CLIPTextChunker:
Using a conservative limit of 70 tokens to account for special tokens.
"""
def __init__(self, max_tokens: int = 70):
def __init__(self, max_tokens: int = 40):
"""
Initialize the text chunker.
Args:
max_tokens (int): Maximum number of tokens per chunk (default: 70 for CLIP, being conservative)
max_tokens (int): Maximum number of tokens per chunk (default: 40 for CLIP, being very conservative)
"""
self.max_tokens = max_tokens
self._tokenizer = None
@ -24,8 +24,9 @@ class CLIPTextChunker:
if self._tokenizer is None:
try:
from transformers import CLIPTokenizer
self._tokenizer = CLIPTokenizer.from_pretrained("openai/clip-vit-base-patch32")
except ImportError:
# Use a simpler model that should be more reliable
self._tokenizer = CLIPTokenizer.from_pretrained("openai/clip-vit-base-patch32", local_files_only=False)
except Exception as e:
# Fallback to character-based estimation if transformers not available
self._tokenizer = None
return self._tokenizer
@ -42,7 +43,8 @@ class CLIPTextChunker:
"""
if self.tokenizer is None:
# Fallback to character count if tokenizer not available
return len(text)
# Use a very conservative estimate: ~0.6 characters per token for CLIP
return int(len(text) * 0.6)
tokens = self.tokenizer(
text,
@ -153,14 +155,14 @@ class CLIPTextChunker:
# Fallback to regular chunking
return self.chunk_text(text)
def chunk_prompt_for_clip(prompt: str, max_tokens: int = 70) -> List[str]:
def chunk_prompt_for_clip(prompt: str, max_tokens: int = 40) -> List[str]:
"""
Convenience function to chunk a prompt for CLIP processing.
Uses a conservative 70 token limit to be safe.
Uses a conservative 40 token limit to be safe.
Args:
prompt (str): The prompt to chunk
max_tokens (int): Maximum tokens per chunk (default: 70 for safety)
max_tokens (int): Maximum tokens per chunk (default: 40 for safety)
Returns:
List[str]: List of prompt chunks