Reddit user reports 20–30% throughput gain from disabling MTP
A Reddit user says disabling MTP improved token throughput by 20–30% in tests involving GyroStrata and Strata. The tester used a GeForce RTX 3060 with 32GB of RAM and reports that the issue may arise when a model is large relative to available memory; these results remain anecdotal.
Local LLM users running large models on limited hardware may want to test whether disabling MTP improves throughput, but the reported gain still needs independent verification.
Little reminder, with all my testing I found out that mtp can decrease performance when using very big model, since many are starting to use strata or similar and I see many app activating it as default try and disable it, in my case, testing trough gyrostrata and strata with a 3060 and 32gb of ram both with gyro-m and swift 1.5 I got 20-30% improvement in tok/s just by disabling mtp.