Note: this in a continuation of my Laguna S2.1 118B bench and tune for AMD Strix Halo on Ubuntu 26 HERE. Hint: You need the offical Poolside.ai build llama.cpp. Here I focus on the mixed reports in the wild from “the model works” to “throw it away”

When we look online we see mix results from the Laguna S2.1 118B’s usage though only a small % of the community doing any real testing.

This is a follow up on that problem with community findings applied to my benched and tuned config to dial in the intelligence side of things.

TL:DR:

Speculative D-Flash decoding like this works best for Strix Halo (–spec-draft-n-max 4). Tune for your AI inference platform as this most likely will not match yours

llama-server \
  --host 0.0.0.0 \
  --port 8080 \
  --ctx-size 250000 \
  -ngl 999 \
  -m laguna-s-2.1-Q4_K_M.gguf \
  -md laguna-s-2.1-DFlash-BF16.gguf \
  --spec-type draft-dflash \
  --spec-draft-n-max 4 \
  --spec-draft-p-min 0.55 \
  --jinja \
  --reasoning off \
  --device Vulkan0 \
  -fa on \
  -fit off \
  -b 16384 \
  -ub 2048 \
  -ctk q8_0 \
  -ctv q8_0 \
  --temp 1.0

• Added –chat-template-kwargs ‘{“enable_thinking”:false}’ as explicitly turns thinking off (the structural off-switch). Community testing found thinking-on net-negative on held-out work producing worse scores, invented bugs, over-refusal, long-loop wedging).
• Added -fit off Disables the memory auto-fitter (known to hang on load with this model on some llama.cpp forks/builds). Also, help in preventing silent performance drops: Because the auto-fitter will quietly push layers to your CPU if it runs out of VRAM, your generation speeds (tokens per second) can plummet without you knowing exactly why. Turning it off forces the application to OOM and crash, giving you a clear, hard limit to work from.
• Added –temp 0.6 — Sets the recommended default temperature from the same testing (works well for coding/agent use; clients can still override).
• –jinja Ensures the native template so tool-calling works. The recommended default.
• Existing DFlash settings (n-max 4, p-min), Q4_K_M, flash-attn, high-precision KV (q8_0), large batch/ubatch, and ~250k context . See how I got this via bench / decode tune in article linked on line 1 on post.

Source Twitter post (the main one that consolidates the four key recommendations after independent testing + the 12-hour soak):
https://x.com/no_stp_on_snek/status/2081028096304599090

(Related follow-up with sharper thinking-control details: https://x.com/no_stp_on_snek/status/2081365791484633233.

Run a quick tool-call test after starting to confirm structured tool_calls appear (not just prose).