# Which speculative decoding are you using and why? — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/01a900295202242b

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


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-09-21 09:09 ingested 2026-09-22 02:40

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1wm7kws/which_speculative_decoding_are_you_using_and_why/)

## retained text



```
I have 2 working models at the moment `Qwen3.8-Flash-Next-UD-Q4_K_XL` and `Qwen3.8-27B-UD-Q5_K_XL`. For Flash model I'm using these params

          spec-draft-device: Vulkan0
          spec-draft-n-max: 4
          spec-draft-ngl: 99
          spec-draft-p-min: 0.5
          spec-type: draft-mtp,ngram-map-k

And for 27B one I use these

          spec-draft-device: Vulkan0
          spec-draft-n-max: 4
          spec-draft-ngl: 99
          spec-draft-p-min: 0.5
          spec-type: draft-dflash,ngram-map-k

I don't know if ngram-* like decodings make sense at all or maybe they make it worse. I didn't see much difference on my working tasks and you can't use llama-bench with speculative models to check it without bias.
My hardware: AMD Ryzen AI 9 HX 470 w/ Radeon 890M, 96Gb DDR5 RAM, llama-server Vulkan backend from [Nathan](https://github.com/Nathanw1014/strix-halo-llamacpp)
```



document fetched 2026-09-22 00:41 · canonical https://reddit.com/r/LocalLLaMA/comments/1wm7kws/which_speculative_decoding_are_you_using_and_why
