lovedheart

Qwen3.8-Flash-Next on sglang (#37500 merged, FP4/FP8 QSA indexer, SM120 b12x)

Qwen3.8-Flash-Next on sglang (#37500 merged, FP4/FP8 QSA indexer, SM120 b12x)

One-command recipes for vLLM-Moet: auto-download + serve DeepSeek-V4-Flash, GLM-5.2, Kimi-K2.7-Code.

Frontier MoE on Blackwell: DeepSeek-V4-Flash, GLM-5.2, Kimi-K2.7 with 2-bit experts + FP4 recovery.

A vLLM extension for GPU + NUMA dual parallel inference