beamivalice/sushi v1.3.0 on GitHub

v1.3.0 — Qwen EXL3 Quantizer, Adaptive DFlash2, KV6, Dashboard

Qwen3.8-Flash-Next EXL3 quantizer: sushi convert exl3 writes a Sushi pack from the BF16 checkpoint at the rate you pick, with the trunk in the format the packs serve; runs resume where they stopped, keep finished layers and packs, and fit their Metal threadgroups to any Apple GPU. Check the result with sushi kld. Thanks to @cosmic-quokka for the M2 Pro report (#57).
Adaptive DFlash2 for Qwen3.8-Flash-Next: point --drafter at the DFlash2 drafter and Qwen verifies 2 to 8 draft rows per round as the drafter's confidence pays, up to 20% faster than MTP on Sushi-2.6bpw and level with it on Sushi-4bpw (--drafter-min-depth / --drafter-max-depth); MTP stays the default. Thanks to @1RAFAL for the request (#59).
--kv-quant 6 on all three models: a 6-bit KV cache between kv4 and kv8 for Qwen, MiMo and GLM, also as kv_quant 6 in model-settings.json and per request; GLM stores its MLA latent at 6 bits and keeps DFlash2 and prefix reuse.
Web UI Dashboard: a Chat | Dashboard switch in the top bar with live throughput, requests, GPU and memory charts plus sessions and models tables (needs --metrics); switching never interrupts a chat. Thanks to @cosmic-quokka for the dashboard (#55).
Requests that used to fall back now run as asked: MTP stays on with presence_penalty and repeat_penalty; Qwen serves high and max effort as xhigh, GLM serves medium as high and xhigh as max, each with a notice in the log and the response.