Qwen3.8-27B-W4A16-AutoRound-fast-ru

Qwen3.8-27B W4A16 AutoRound — fast variant with a Russian+English draft vocabulary

This is the syv-ai/qwen38-27b-rtx3090 fast variant (GPTQ-int4 lm_head + MTP module) built on the [dbirks W4A16 AutoRound body], with the draft vocabulary rebuilt for Russian-first workloads: 18,195 ids counted over the model's own outputs on a mixed ru/en/code corpus (97.9% held-out coverage). Drop-in for the syv serving stack.

Why

The shipped draft vocabulary is counted over an English-heavy corpus; on Russian text the drafter misses often enough to cost real speed. This rebuild restores full MTP k=3 speculation quality on Russian while keeping English and code coverage.

Measured (RTX 3090 Ti, Proxmox VM, syv stack main-7db2d1f, MTP k=3, prefix cache on)

• single-stream ~118 tok/s (RU prompts, bf16 KV), 8-stream aggregate ~351 tok/s (1.5k-token prompts) / ~430 tok/s (decode-phase micro-prompt ladder)
• works with vLLM OffloadingConnector (KV eviction to RAM) and VISION=1 simultaneously — see bench notes and field report

Use

Point the syv launcher at this folder (MODEL=/app/models/<this repo>); the draft head and vocab (mtp_draft_vocab_ids.pt, model_extra_tensors.safetensors) are picked up automatically.

Built and benched by the Digital Assistant team homelab; not affiliated with syv-ai or dbirks.