#[Unsloth Ecosystem] DeepSeek-OCR-2 deployment 98% complete on RTX 2080 Ti (Turing) – stuck on final

9 messages · Page 1 of 1 (latest)

alpine shore
#

Thanks to Unsloth, I've made great progress running DeepSeek-OCR-2 on my RTX 2080 Ti (Turing, 22GB):

✅ Model loads successfully with 9.5GB VRAM (lower than official 12GB!)
✅ Inference flow works – can see generated token IDs
✅ flash-attention-triton verified on Turing (might help other old GPU users)

Progress: 98% done!

Last blocker:
When using Unsloth 2026.2.1 (installed with --no-deps to protect my transformers 4.46.3), I get:
ImportError: cannot import name 'is_trackio_available' from 'transformers'

DeepSeek-OCR-2 officially requires transformers==4.46.3, so I can't upgrade.
Tried older Unsloth versions (2025.1.8, 2024.12.12) – each gave different errors.

Questions:

  1. Is there any Unsloth version compatible with transformers 4.46.3?
  2. Any config flags to bypass this import check?

Just need that final 2%! Will share my Turing debugging经验 in return. 🙏

image.png
tiny summit
#

did you use the unsloth version of deepseek?

#

that works with transformers v5

alpine shore
#

Thanks for the reply and the link!

I checked the model card on Hugging Face, and it still lists transformers==4.46.3 as a requirement — the same version I'm currently using. But you mentioned that the unsloth version is optimized for transformers v5.

This seems to create a contradiction: does the unsloth/DeepSeek-OCR-2 actually require v5, or can it work with 4.46.3? Has anyone successfully run it with 4.46.3?

The reason I'm stuck with 4.46.3 is that DeepSeek-OCR-2's official docs explicitly require it. My current setup (transformers 4.46.3 + model loading) works fine — inference runs, I can see token IDs — but I hit this import error when using Unsloth 2026.2.1.

So my core question is: Is v5 mandatory for Unsloth to work with this model? Or is there a way to make it work with 4.46.3?

Thanks for any clarification!

alpine shore
#

Update: Found the root cause! vLLM 0.8.5 requires ALL_ATTENTION_FUNCTIONS from
transformers.modeling_utils, which doesn't exist in transformers 4.46.3.
This explains the import error.

So the real question becomes: Does DeepSeek-OCR-2 work with newer transformers
versions (>=4.47.0)? Has anyone tested this?

normal field
#

try 4.57.1

#

or 4.57.3

normal field
#

/.,mnbvcxz`A~SZDXRFTGYHUIJOKPL[;]