centroid based vLLM implementation for lora instruct.

Input

5614-dim embedding

Encoder

88 x CNN with 48 heads

Output

mAP projection