Hobbyist boosts Strata decoding for Qwen3.8-Flash-Next
A hobbyist running IQ3_S Qwen3.8-Flash-Next on Strata 0.1.43 with an RTX 4090, 48 GB RAM and a Ryzen 9 3900X reported 2,076-token prompt processing and an average decode rate of 95–105 tokens per second at a 200,000-token context. A separate grid-search result rose from 97 to 142 tokens per second with --adapt-async 1, --kv-resident 20480 and STRATA_ADAPT_LAG=2. The author also found that --ple-io ram, a large explicit --expert-cache and --no-prefill-borrow reduced throughput by 90%, 72% and 46%, respectively.
The report provides a detailed, reproducible starting point for local users serving Qwen3.8-Flash-Next at long context on Strata, while its single-source and non-official status limit broader confidence.
TLDR: IQ3_S Qwen3.8-Flash-Next, RTX 4090 + 48 GB + Ryzen 3900X, Windows, 200k context. Ran essentially a grid search on a ton of flags: decode 97 → 142 tok/s (+45%) for three flags. Most relevant: --adapt-async 1, --kv-resident 20480, STRATA_ADAPT_LAG=2. Worst: --expert-cache is in MiB not slots (−72% if you fill the card), --ple-io ram (−90% on 48 GB), --no-prefill-borrow (−46%). One lever still locked: Windows sucks and doesnt pin memory w cuda , so every host -> GPU expert copy is CPU-driven.
Been using strata for a few weeks since thats all the rage (waiting until llama.cpp catches up, but tbh the age of specialized inference engines might just be how it is). Version 0.43, so im probably missing some newer flags. It updates so frequently and i hear that people were pretty mad with version 0.2.0 so im just holding off. if it aint broke dont fix it
I basically told Qwen 3.8 FN to tune itself and took a very hands-off approach. I let it do web research + spawn sub-agents to search thru strata's code.
Some source content is omitted. Continue reading on the original site.
#Strata#Qwen3.8-Flash-Next#NVIDIA RTX 4090#AMD Ryzen 9 3900X#Windows#llama.cpp