Qwen3.8 27B Swift-1.5
on 3× 2x Radeon VII + NVIDIA laptop 8GB · Strata
Use cases
agentic
Summary
User reports a fork of Strata serving a ~47 GiB MoE model (Swift-1.5 / Qwen3.8-Flash-Next) on 2x Radeon VII (gfx906) plus one 8 GiB NVIDIA laptop card, at a pooled median of 34.7 tok/s decode versus 23.3 tok/s for same-day stock.
Setup uses DFlash2 speculative decoding on HIP, with the NVIDIA card serving expert and prefill rows to the AMD engine over a socket.
User reports 1.4-2.4x faster short-prompt reads (200-600 token agent traffic), 6.2 GiB less host RAM, and a 12% drop on fresh ~16k prefill reads; long-context reads gain instead. Greedy 80k-token infill is byte-identical to stock.
Compare this run
Medians of plain runs: one device, one request, no speculative decoding, the whole model in the device's memory. The report above is one setup, so read it beside them and not as one of them.
Qwen3.8 27B at 4-bit by hardware
- NVIDIA RTX 309024 GB67.0 t/s · one run
- 37.5 t/s · one run
Bars are medians on one scale, 0 to 67.0 tokens/s. Press ▶ to watch text arrive at a row's speed.