Can you also release the 224*224 pre-trained weights?
#7
by
qx-Sean
- opened
I really appreciate the released model.
However, when I try to compose a LLAVA-Vicuna7B model with RadDino, I get an OOM issue with 2 Nvidia A6000 GPUs.
I tried DeepSpeed Stage 2 without offload. But still have this issue.
I'm wondering if you can also share the 224*224 pre-trained weights.
That would help a lot.
Thanks!
Hi, @qx-Sean . Unfortunately, we won't be able to share that checkpoint.
That being said, RAD-DINO has 87 M parameters, vs 7B for your LLM (i.e., 1.2%). I'd suggest you focus on the LLM instead to address your OOM error. You could also use the CLS token instead of the patch tokens.
fepegar
changed discussion status to
closed