Draft acceptance tweaks?
A user is seeking advice on optimizing the performance of MTP, specifically asking if their current setup is within the expected range or if further tweaks are possible. The provided configuration details include a command for running llama-server with a Qwen3.6-35B-A3B-MTP-UD-Q8_K_XL.gguf model, specifying parameters like --ctx-size 100352, --gpu-layers 999, --batch-size 1024, and --spec-type draft-mtp, among others, to achieve better draft acceptance.
- Published
- Sep 7, 2026, 02:19
- Source type
- Dev community
- Tier
- Community
- Source status
- Healthy
Times shown in UTC
More details
What can I do to get more from MTP if anything? Or am I sitting at the expected range. Thanks
/opt/llama.cpp/build/bin/llama-server \ --host 10.210.44.126 \ --port 11434 \ --model /opt/models/Qwen3.6-35B-A3B-MTP-UD-Q8_K_XL.gguf \ --mmproj /opt/models/mmproj-35B-A3B-BF16.gguf \ --ctx-size 100352 \ --cache-type-k f16 \ --cache-type-v f16 \ --gpu-layers 999 \ --main-gpu 1 \ --poll 0 \ --batch-size 1024 \ --ubatch-size 256 \ --threads 6 \ --split-mode tensor \ --tensor-split 34,40,26 \ --jinja \ --reasoning-format deepseek \ --image-min-tokens 1024 \ --alias Qwen3.6-35B-MTP-Q8-XL,frigate \ --flash-attn on \ --reasoning on \ --spec-type draft-mtp \ --spec-draft-n-max 5 \ --ctx-checkpoints 48 \ --cont-batching \ --min-p 0.05 \ --top-p 0.80 \ --parallel 2 \ --metrics \ --kv-unified \ --temp 0.6 \ --top-k 20 \ --load-mode none \ --lazy-mode off \ --fit off \ --slots