useful-quants/LFM2.5-VL-3B-NVFP4-FP8-Mixed

Serving fix: CUDA graph execution default (was --enforce-eager), corrected perf tables (Gate 7C), postmortem, startup warmup, MinVRAM now default recommendation

docs: correct raw tensor payload byte count in model card

Initial private release: LFM2.5-VL-3B NVFP4/FP8 Mixed (Candidate B + Vision-M)

Initial private release: LFM2.5-VL-3B NVFP4/FP8 Mixed (Candidate B + Vision-M)

initial commit