kyle/imladris
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

straylight: unwire Laguna DFlash until the GGUF matches llama.cpp

Browse Source 
Official laguna-s-2.1-DFlash-BF16.gguf is on disk but this llama.cpp
rejects it (expected 76 tensors, got 69). Leave the draft unwired so
Laguna loads; keep amdgpu DPM high.
This commit is contained in:
kyle
2026-09-18 00:00:37 -07:00
parent 14b3b3c6b8
commit 10dddaab7a
1 changed files with 3 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files
hw/straylight/default.nix
+3 -8
View File
@@ -39,16 +39,11 @@ let
; Poolside Laguna S 2.1 (118B total, 8B active): primary coding agent.
; ~69 GB of weights, so it only loads once the raised TTM/GTT limit is active (reboot).
; DFlash draft (~2.2 GB BF16) from poolside/Laguna-S-2.1-GGUF. Keep draft KV in
; f16; q8 draft KV collapses acceptance. spec-draft-n-max 4 is the gfx1151
; sweet spot (n=15 is the trained cap and is slower here).
; DFlash GGUF is on disk (laguna-s-2.1-DFlash-BF16.gguf). nixpkgs llama.cpp 0.4.0
; rejects it at load ("expected 76, got 69" tensors); Poolside's file needs their
; laguna-branch loader. Leave unwired so Laguna still starts.
[laguna-s-2.1]
model = ${llamaModelsDir}/laguna-s-2.1-UD-Q4_K_S/Laguna-S-2.1-UD-Q4_K_S-00001-of-00003.gguf
model-draft = ${llamaModelsDir}/laguna-s-2.1-UD-Q4_K_S/laguna-s-2.1-DFlash-BF16.gguf
spec-type = draft-dflash
spec-draft-n-max = 4
cache-type-k-draft = f16
cache-type-v-draft = f16
; Two slots sharing one unified 131072-token KV pool: a second request (subagent,
; another client) runs alongside instead of queueing, and either one may still use
; the full 131072 as long as their combined context fits.