llama.cpp fork for significantly improved performance on Ampere (especially RTX 3090 / 3090 Ti): TurboQuant KV cache, MTP speculative decoding with a 64K draft-vocabulary shortlist, custom SM86 + Qwen kernels. 90 tok/s over a 100K-token generation at temperature 1.
Serving Qwen3.8-27B-FP8 on a single DGX Spark (GB10): 7.88 to 58.5 tok/s single-stream from decode strategy alone, weights untouched. Speculative decoding and prefix caching benchmarked, plus DFlash 2 — the only Qwen3.8-27B build that can serve it under vLLM.
