PG-InstructBLIP checkpoint for blip2_t5_instruct + flant5xl (to run on 16GB GPUs)
#4
by
Baonq - opened
Hi! Thanks for releasing PG-InstructBLIP and the PhysObjects/PG-VLM work.
I’m trying to run PG-InstructBLIP on a 16GB GPU (Tesla P100/V100). The released checkpoint (pgvlm_weights.bin) targets blip2_t5_instruct with model_type="flant5xxl", which is difficult to run reliably on 16GB GPUs.
In the paper you mention using a smaller VLM (FLAN-T5 XL). Could you please share/upload a PG-InstructBLIP checkpoint fine-tuned with model_type="flant5xl" (e.g., a link to a Hugging Face repo or a direct checkpoint file)?
Thanks a lot!
If you prefer not to share it publicly, I can receive the checkpoint via email (if that’s easier): qbao1607@gmail.com.