Weschera/Qwen3.8-27B-NVFP4-DFlash2-DGX-Spark

Qwen3.8-27B on one DGX Spark: SGLang + DFlash2 recipe, engine image, benchmarks, data and charts. By @redp314

Qwen3.8-Flash-Next on one DGX Spark: tuned vLLM launch, context sweep, tuning ladder. By @redp314

recipe for running Qwen3.8-Flash-Next on a single DGX Spark (fork)

Qwen3.8-27B NVFP4 on SGLang (DGX Spark / SM121): official image pin, native DSpark/EAGLE click-run, matched evals

Qwen3.8 Flash at 74 tok/s on a single DGX Spark / GB10. vLLM + MTP recipe, 212 tok/s at 8 streams.

Qwen3.8 27B on SGLang for DGX Spark

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.

Qwen3.8-Flash-Next (NVFP4) on DGX Spark in vLLM: one Spark 43.9 tok/s with our disk-backed n-gram table patch, staged gather and reduced-vocab MTP draft; TP2 SPEED 53.7; TP4 CONTEXT 9M KV pool. Launchers, patch, 40-prompt harness, KV ledger, credits. SGLang TP2 day-0 lane kept.

qwen38-flash-next recipe for solo GB10