2026-11-10 · America/Los_Angeles · 社区动态 · #17

RustingStrata vs PolyStrata vs llama.cpp on one RTX 3060 12 GB: Qwen3.6-35B-A3B

RTX 3060 12 GB, Ryzen 5 7600X, 32 GB RAM, one GGUF file: Qwen3.6-35B-A3B-UD-Q3_K_M.gguf (Unsloth, MTP layer included, 15.9 GiB, does not fit in VRAM). Prompt of 5290 tokens, 128 generated tokens. RustingStrata PolyStrata llama.cpp b12119 Prefill, 5290-token prompt 2330 tok/s 1458-1549 tok/s 1692 tok/s (-ncmoe 22, -ub 2048) Decode, no MTP 90.9 tok/s 78.5 tok/s (--no-draft, see below) 53.5 tok/s Decode with MTP drafts 128.4 tok/s 97.5-100.4 tok/s 77.7 tok/s (--spec-type draft-mtp, 92 of 102 drafts accepted) How each was run: RustingStrata: rustingstrata bench --gpu --mtp 0|1 --pp 5290 --tg 128…

热度 28.9 / 100；排名与评分保留该期记录。