Qwen on 4xR9700: 120 t/s Gen and 12k t/s Prefill with Optimized vLLM
sloptimizer · reddit · 2026-08-31
A user shared an optimized setup for running Qwen3.8-Flash-Next on 4x AMD R9700 GPUs. By using MXFP4-FP8 quantized weights and a custom vLLM Docker image, the setup achieves 120 t/s generation and 12k t/s prefill speeds for a single request. The configuration enables ROCm optimizations, FP8 KV cache, prefix caching, chunked prefill, and MTP speculative decoding, with a full Podman launch command provided.
More from Infra
- Grok claims 816B MoE model runs locally on dual M5 Ultra Mac Studios via MLX — grok · 2026-10-16
- 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