Qwen3.8 Flash Next on 2x 8 GB (ElectroHobo W4A16 TP2)/Confirmation 1
From CMP Wiki
More actions
| Confirms | Qwen3.8 Flash Next on 2x 8 GB (ElectroHobo W4A16 TP2) |
|---|---|
| Result | Reproduced |
| By | @darkmaniac7 |
| Date | 2026-10-28 |
| Cards | 8 GB |
| Numbers | One user code 263, structured 266, prose 215 tok/s; 8 users 916 / 961 / 729; prefill 4.6k tok/s on a 17k prompt; 53% draft acceptance. Four instances on 8 cards: 3,617 / 3,316 / 2,664 tok/s aggregate at 32 users, 17.4k tok/s prefill. |
Release 0.2.0 on Gen2 x16 cards with P2P, 1,024-token answers at temperature 0. Vision works. Four instances share one 48 GiB table through PLE_SHARED_DIR on tmpfs (77 GiB host RAM for all four); each keeps 90 to 100% of its lone speed. Custom all-reduce off cost 17 to 25% single-stream, so it stays on. Source: @darkmaniac7, 2026-10-28 (lab notes)