From 4a4a6fc803a84605862b04f819fc45e8cdc1c5a5 Mon Sep 17 00:00:00 2001 From: Vuong Hoang Date: Wed, 27 May 2026 00:03:03 -0700 Subject: [PATCH] =?UTF-8?q?llama-swap:=20artemis-31b-v1i=20=E2=80=94=20add?= =?UTF-8?q?=20--reasoning=20on=20+=20deepseek=20format?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Worldtree-dev's --reasoning-format gemma suggestion isn't supported in the deployed llama.cpp build (accepts only none|deepseek|deepseek-legacy). Falling back to deepseek, which also populates reasoning_content — the field Worldtree's GemmaProvider fallback path checks. Verified via test inference: 608 reasoning_content deltas + 199 content deltas + 0 raw <|channel> marker leaks. --- stacks/llama-swap/conf/config.yaml | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/stacks/llama-swap/conf/config.yaml b/stacks/llama-swap/conf/config.yaml index 09b035d..1ac775a 100644 --- a/stacks/llama-swap/conf/config.yaml +++ b/stacks/llama-swap/conf/config.yaml @@ -320,7 +320,7 @@ models: "artemis-31b-v1i": name: "Artemis 31B v1i Q6_K" - description: "BeaverAI Artemis-31B-v1i — Gemma 4 31B dense fine-tune, Q6_K (~28.6GB). 256K context." + description: "BeaverAI Artemis-31B-v1i — Gemma 4 31B dense fine-tune, Q6_K (~28.6GB). 256K context native; served at 128K." ttl: 600 cmd: | /app/llama-server @@ -328,13 +328,15 @@ models: -hf BeaverAI/Artemis-31B-v1i-GGUF:Q6_K --port ${PORT} --n-gpu-layers 999 - --ctx-size 32768 + --ctx-size 131072 --flash-attn on --temp 1.0 --top-p 0.95 --top-k 64 --repeat-penalty 1.0 --chat-template-kwargs '{"enable_thinking":true}' + --reasoning on + --reasoning-format deepseek # ========================================================================== # GLM MODELS