← sources

Extending context window past 262144?

source: reddit (reddit:LocalLLaMA)published 2026-09-23 14:33ingested 2026-09-24 02:40

retained text

I currently use Qwen3.8 27B NVFP4 with 1cat-vllm and the full 262144 token context window in FP8 and I have some VRAM to spare (4x16GB Tesla V100). I'm looking to increase it as far as it'll go.

I'll probably end up trying it but I was wondering if anyone had advice on using RoPE scaling / YaRN specifically with qwen3.8 27B -- how do I get the best performance/precision out of it? Is it worth it or does the model become dumb? Have you used it successfully?

document fetched 2026-09-24 00:45 · canonical https://reddit.com/r/LocalLLaMA/comments/1wo7otf/extending_context_window_past_262144