# CUDA: enable sparse fa for qwen4 by am17an · Pull Request #28770 · ggml-org/llama.cpp — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/85916435cae691e1

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-09-20 16:03 ingested 2026-09-21 02:38

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1wlkxjw/cuda_enable_sparse_fa_for_qwen4_by_am17an_pull/)

## retained text



```
Another day, another Qwen Flash Next speedup 
```



document fetched 2026-09-21 00:36 · canonical https://reddit.com/r/LocalLLaMA/comments/1wlkxjw/cuda_enable_sparse_fa_for_qwen4_by_am17an_pull
