RCreddit.com
18
·1 days ago·RSS
Not on the current live radar
Qwen3.8-Flash-Next NVFP4 2xDGX Spark config: 50t/s decode, 2,900t/s prefill
Heat trend
Collecting trend data
The percentage is based on available heat signal, not comment count or independent people.
A user shared their Qwen3.8-Flash-Next NVFP4 2xDGX Spark configuration, achieving 50t/s decode and 2,900t/s prefill. The setup uses Dual-Spark TP2 with eager + MTP k=3 for a 262k context. Performance metrics include decode aggregates up to 265.9 t/s across 13 streams and prefill rates ranging from 2,463 t/s for 200k prompts to 2,875 t/s for 11k prompts, highlighting the efficiency of the --enforce-eager setting.