llamaperf

Qwen3.8 27B

on AMD RX 7900 XTX · llama.cpp · 196,608 ctx

AMD hardware
Nov 20, 2026
Throughput
140.0 t/s gen · 862.3 tokens/s prompt processing (prefill)Prompt processing is how fast the input is read, not how long you wait for the first word. How to measure time to first token.
Quant
Q4_K_M (GGUF)
KV cache
q8_0
Flash Attention
on
System RAM
32 GB
VRAM reported
24 GB

Use cases

codingagentic

Summary

User reports Qwen3.8 27B at 140.01 t/s generation and 862.29 t/s prompt eval on an RX 7900 XTX 24GB. Setup is an RDNA3-optimized llama.cpp fork (Vulkan) with Q4_K_M GGUF weights, q8_0 KV cache, 196608 context, and DFlash2 speculative decoding using a 1.14 GB draft sidecar. Whole model and draft on the GPU, single request. Three cold-prompt runs of ~2600 prompt tokens and 128 generated tokens: decode 140.01, 134.10 and 140.31 t/s; prefill 862.29, 863.38 and 852.28 t/s. Nothing averaged. Use case is a daily-driver local coding agent.