Windows user tunes Strata parameters for local Qwen inference on RTX 4090

On October 26, 2026, a user posted in Reddit's LocalLLaMA community about tuning Strata configuration options when running Qwen models on Windows. Their test system combined an RTX 4090 with an additional 48 GB of VRAM and a Ryzen 3900X processor, running the IQ3_S quantization of Qwen3.8-Flash-Next at a 200k context.

The user essentially ran a grid search across a large set of Strata flags. Adjusting three options lifted decode throughput from 97 tok/s to 142 tok/s, a 45% gain, with the most relevant settings being --adapt-async 1, --kv-resident 20480, and STRATA_ADAPT_LAG=2. They also called out several configurations that performed poorly: --expert-cache takes MiB rather than slots, so filling the card with it cut performance by 72%; --ple-io ram dropped by 90% on the 48 GB setup; and --no-prefill-borrow fell by 46%.

The post further notes one lever still locked: on Windows, Strata does not pin memory with CUDA, so every host-to-GPU expert copy is driven by the CPU, which holds back expert transfer performance on that platform. So far the thread reports only these tuning outcomes and the platform-level observation, with no follow-up progress shared.