Qwen3.8 27B on 1x 8 GB (W4A16 vLLM)/Confirmation 1
From CMP Wiki
More actions
| Confirms | Qwen3.8 27B on 1x 8 GB (W4A16 vLLM) |
|---|---|
| Result | Partly reproduced |
| By | @chonsworth |
| Date | 2026-10-05 |
| Cards | 8 GB |
| Numbers | About 3.7k tok/s prefill at about 100 tok/s decode, against the published 1.9k / 100. Never run to 1M context. |
Replying to kellertreppe123's post of the recipe, on the same model on one card. He did not say whether he used the recipe as written; his prefill is higher than published and his decode matches. Source: @chonsworth, 2026-10-05