ceos/unsloth
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Studio: point RAG captioner at the pre-quantized Unsloth bnb-4bit repo

Browse source
This commit is contained in:
Roland Tannous 2026-05-26 21:52:17 +04:00
commit c1bc79bc84
1 changed files with 5 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

6
studio/backend/core/rag/captioner.py
View file

@ -21,7 +21,11 @@ from typing import Any
logger = logging.getLogger(__name__)
_CAPTION_MODEL_NAME = "Qwen/Qwen3-VL-2B-Instruct"
# Pre-quantized Unsloth bnb-4bit repo (see unsloth/models/mapper.py).
# Using this name directly skips the FLOAT_TO_INT_MAPPER redirect so the
# initial download fetches the 4-bit weights (~1.5 GB) instead of bf16
# (~5 GB followed by on-the-fly quantization).
_CAPTION_MODEL_NAME = "unsloth/Qwen3-VL-2B-Instruct-unsloth-bnb-4bit"
_PROMPT = (
"Describe this figure in <=60 words. Focus on factual content "
"(axes, labels, captions, visible text, main objects). "