---
title: "donald991234 (@donald991234) | Hashnode"
description: "donald991234 (@donald991234) writes on Hashnode. Read their latest articles and follow their work."
url: "https://hashnode.com/@donald991234"
---

@donald991234

# donald991234

@donald991234 · Joined September 2026

## Comments

[Good question, and it sent me back to the source (master at 8019dc5). Short answer: the snapshots are budgeted, and the post should have said so. They don't live in the draft context at all. common\_context\_params\_to\_llama sets n\_rs\_seq = speculative.need\_n\_rs\_seq(), which is --spec-draft-n-max for MTP, EAGLE3, DFlash and DSpark. The target model's llama\_memory\_recurrent constructor then allocates rs\_size x (1 + n\_rs\_seq) state rows up front, with rs\_size = n\_seq\_max. fit.cpp measures the main model by creating a context and reading llama\_get\_memory\_breakdown, so those rows land in its context term and scale with n\_max and -np. The draft is measured separately with n\_rs\_seq = 0, so a failed draft measurement doesn't drop them, and they don't grow with prompt length. What fit doesn't cover is --ctx-checkpoints, which are host-side copies. I haven't verified this on hardware; running with --verbose and comparing fit's projected memory with the "RS buffer size" log line would confirm it. Thanks for pushing on this.](https://hashnode.com/posts/how-much-extra-vram-speculative-decoding-needs-in-llama-cpp-mtp-dflash-draft-models/6abb4e53f601645d75bc6149/comment/6abb8b20126a47d11ee0a643#reply-6abb9986ce57b82d0614f168)

Reply · Article · Sep 29 · [How much extra VRAM speculative decoding needs in llama.cpp (MTP, DFlash, draft models)](https://hashnode.com/posts/how-much-extra-vram-speculative-decoding-needs-in-llama-cpp-mtp-dflash-draft-models/6abb4e53f601645d75bc6149)
