Koko210/miku-discord
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Disable model warmup to improve switching speed

Browse Source 
- Added --no-warmup flag to both llama3.1 and vision models
- Reduces model switch time by 2-5 seconds per swap
- No impact on response quality, only minor first-token latency
- Better for frequent model switching use case and tight VRAM budget
This commit is contained in:
koko210Serve
2025-12-10 10:09:37 +02:00
parent a7f3a0a0ee
commit 675bb21653
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
llama-swap-config.yaml
View File

@@ -4,7 +4,7 @@
models:
# Main text generation model (Llama 3.1 8B)
llama3.1:
cmd: /app/llama-server --port ${PORT} --model /models/Llama-3.1-8B-Instruct-UD-Q4_K_XL.gguf -ngl 99 -nkvo -c 16384 --host 0.0.0.0
cmd: /app/llama-server --port ${PORT} --model /models/Llama-3.1-8B-Instruct-UD-Q4_K_XL.gguf -ngl 99 -nkvo -c 16384 --host 0.0.0.0 --no-warmup
ttl: 1800 # Unload after 30 minutes of inactivity (1800 seconds)
aliases:
- llama3.1
@@ -12,7 +12,7 @@ models:
# Vision/Multimodal model (MiniCPM-V-4.5 - supports images, video, and GIFs)
vision:
cmd: /app/llama-server --port ${PORT} --model /models/MiniCPM-V-4_5-Q3_K_S.gguf --mmproj /models/MiniCPM-V-4_5-mmproj-f16.gguf -ngl 99 -c 4096 --host 0.0.0.0
cmd: /app/llama-server --port ${PORT} --model /models/MiniCPM-V-4_5-Q3_K_S.gguf --mmproj /models/MiniCPM-V-4_5-mmproj-f16.gguf -ngl 99 -c 4096 --host 0.0.0.0 --no-warmup
ttl: 900 # Vision model used less frequently, shorter TTL (15 minutes = 900 seconds)
aliases:
- vision