inference-sh/JEV-27B-VL-merged

JEV-27B-VL, decision LoRA merged

autotrust/JEV-27B-VL at revision f34b598d4ef4bcefd337bee8d8e7ddd3b7733ccc with its decision LoRA (adapter_vllm, r=32, alpha=64, every language-model linear and lm_head) folded into the base weights: W + 2 * B @ A, computed in fp32 and rounded to bf16 once. The vision tower and MTP weights are unchanged. Modified from the original by inference.sh; Apache-2.0 as the original.

Serve it without --enable-lora. serve_decide.py is the repository's, patched to read decision_head.json from JEV_DECIDE_ADAPTER_DIR (point it at adapter_vllm/ here) and run with JEV_DECIDE_LORA set to the served model name. Only System 1 (thinking off) uses this checkpoint; System 2 needs the unmerged base model.