Maya是个好项目。
Replying to @PeasantSmith
Getting this project out and about is extremely hard work, please consider sending some support via buymeacoffee.com/peasantsmit…
Please send this guy a Strix Halo, and preferably a couple of R9700s while you're at it. His Strata work really could help the AMD product community thrive. @AMD @ColbyCakesss @AzorFrank
Strix Halo is now officially supported 🥳
- AMD gpu improvements
- Multi GPU improvements (batching)
- Improvements on decode
- Turing improvements
- Intel Arc improvements
github.com/Niko1221/Strata/r…
Thanks @coldniko for taking my multi-GPU helper optimization codes, happy to contribute to the Strata community!
Here are a few observations from my testing.
Strata offers several multi-GPU execution paths, with two main approaches: layer splitting and expert helpers. Auto split is a strong all-round option, especially for prefill. However, when more expert weights remain in system RAM, the optimized helper path can offer better decode throughput—but faster decode doesn’t always mean a faster total request.
Enable the optimized helper path with:
--expert-cache-device1 auto --remote-expert-opt
Use this instead of --layer-split auto.
Testing setup: Strata v0.1.39, 2× RTX 4090 with 24 GB VRAM each, 128GB RAM. All tests used MTP (--spec 4), --pcie-frac 0, and the same 31,400-token input with 2,048 output tokens. Results are medians of three runs after warmup, with zero prompt reuse.
Compared with auto split, helper opt’s decode advantage grew with the model’s weight footprint: from just +9% on IQ3_XXS to +25% on Q3_K_XL, reaching around +40% on bothnQ4 models. But prefill consistently favors auto split: it was 25–51% faster across all four models tested.
Overall, auto split is probably still the reliable choice for most everyday workloads. But personally, I prefer the heterogeneous helper approach—it’s just too much fun. LLM is great! Strata is great!










