LocalLLaMA
Auto-fit vs tuned MoE offload: 564 → 1330 pp tok/s, unchanged decode (Qwen3.6-35B-A3B Q6 / RTX 3090)
TL;DR: On a Qwen3.6-35B-A3B Q6 setup sized for 64K context on a 24GB RTX 3090, spilling eight MoE expert layers to CPU freed enough VRAM to increase -b from 512 to 1024 and -ub from 128 to 512. Prompt processing improved by 2.36×, while generation speed remained unchanged within measurement noise. Benchmark Auto-fit baseline Tuned Result PP4K 564.5 tok/s 1330.0 tok/s 2.36× TG4K 97.4 tok/s 97.7 tok/s Within noise TG32K 81.6 tok/s 84.0 tok/...
Read the full discussion on Reddit
This article was aggregated from Reddit. Click to join the conversation.
View on Reddit
Related Trends in LocalLLaMA
Building an open-source control plane for self-hosted vLLM, what would you want in it?
5090: Windows or Linux for Qwen3.8.27b
RTX Pro 4500 vs. A100 (x4) - purchase being made today
Best llm at understaning middle eastern languages? Arabic/Persian etc
Qwen3.8-27B abliterated FP8: refusal 64–99% → 0–6%, and MMLU/GSM8K move less than 1.3 points
llama.cpp Windows Manager