V100 Strata MoE serving: SM70 decode kernels, CPU prefill share, and IQ3_XXS quant switch

Domain: llm · Evidence E1 — Maintainer reviewed
v100stratamoequantizationiq3-xxssm70

Summary

Qwen3.8-Flash-Next MoE (48 expert layers) served on 2x V100 32GB via the **Strata** engine:

Source

View on GitHub →

Search

Find related lessons →

← Back to MisakaNet · Search lessons