Optimizing Qwen 3.8 27B with MTP and KV Cache Quantization
gabrielesilinic · reddit · 2026-08-31
Detailed guide on running Qwen 3.8 27B-UD-IQ4XS on a 7900XTX using llama-server.
Key Optimizations:
- MTP Tuning: Used --spec-draft-p-min 0.70 to reject low-confidence drafts, fixing slowdowns in long chats.
- KV Cache Quantization: Enabled q80 for both K and V caches without stability issues.
- Performance: Achieves 22-30 t/s at 140k context, utilizing 22.1GB VRAM.
Stack: llama-swap for management, systemd for service orchestration.
More from Infra
- Same 1,000 Cloudflare Durable Objects: $4,050/month cut to $22 with one trick — threepointone · 2026-10-16
- NVIDIA ships FP4-quantized Qwen3.8-Flash for local use, trending on Hugging Face — nvidia · 2026-10-16
- Baidu's Virtual Head Attention tames the KV cache, FLOPs, quality triangle in long-context LLMs — baidu · 2026-10-16
- DuoDiscover agents auto-build fast eval proxies, scoring 2.68x TPU kernel speedup — berkeley_ai · 2026-10-16
- Google Raises 2026 Capital Spending to Up to $205 Billion — emmanuelvivier · 2026-10-16
- Google open-sources AX, a Kubernetes-style orchestrator for billions of AI agent tasks — thisdudelikesAI · 2026-10-16