A user shares their setup using two modded RTX 2080 Ti GPUs with 22GB VRAM each to run Qwen 3.6 27B at 38 tokens/s with llama.cpp, including tips on power limiting, tensor split mode, and KV cache settings.
PLEASE KEEP IN MIND BOTH OF MY CARDS ARE POWER LIMITED TO 150W (i hate noise) \------- Just wanted to share my current setup, that might help some users out there... services: llama-server: image: ghcr.io/ggml-org/llama.cpp:full-cuda12-b9128 container_name: llama-server restart: unless-stopped ports: - "16384:8080" volumes: - ./models:/models:ro command: > --server --model /models/Qwen3.6-27B-IQ4_XS-uc.gguf --alias "Qwen3.6 27B" --temp 0.6 --top-p 0.95 --min-p 0.00 --top-k 20 --port 8080 --host 0.0.0.0 --cache-type-k f16 --cache-type-v f16 --fit on --presence-penalty 1.32 --repeat-penalty 1.0 --jinja --chat-template-file /models/Qwen3.6.jinja --mmproj /models/Qwen3.6-27B-mmproj-BF16.gguf --webui --spec-default --chat-template-kwargs '{"preserve_thinking": true}' --reasoning-budget 8192 --reasoning-budget-message "... thinking budget exceeded, let's answer now.\n" --split-mode tensor user: "1000:1000" deploy: resources: reservations: devices: - driver: nvidia count: all capabilities: [gpu] environment: - NVIDIA_VISIBLE_DEVICES=all This is my exact config, my 2 extremely old 2080Ti gpus where upgraded in china to have 22GB vram each... and on ebay i bought a NVLINK (i do not recommend bying it, as no meassurable difference appears) Quantisation i run is IQ4\_XS if i change the kv cache to q8\_0 it sometimes happens during long coding sessions that the model loops, this is why i run kv-cache@f16 and never have this problem since then. i use the hauhaucs qwen3.6 model uncensored on IQ4 matrix quants. You can also forget about MTP as you are compute bound with those cards and not bandwidth bound. The absolut biggest boost came from --split-mode tensor , this gave me a boost from 14 token/s to 38t/s i think without the power limit we should get 45 token/s what i also never did think about is the --fit on ... i always declared context length manually worked great but it looks like its not a good idea to always run at 95% vram consumption. fit on also improved token gen a little. Btw. this is a < 1k USD setup running on 400w peak on the wall, and it works great with hermes and opencode. the jinja template i use is this one: [https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates](https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates) (in this setup template 11, i did not yet test the newer templates) https://preview.redd.it/gasb8yo8ga1h1.png?width=476&format=png&auto=webp&s=0450efcae279b0bcbd33f9d6d4f7241d8e3581d4
Technical post sharing performance stats for running Qwen 27B on 2x RTX 5070 Ti GPUs with vLLM cu129-nightly, achieving up to 94-87 tps decode and 170k GPU KV cache.
A user shares optimized settings for running Qwen3.6 27B (Q8_0) on a dual GPU setup (RTX 4090 + RTX 3090) with llama.cpp, achieving 75-100 t/s and 1500 pp with 250k context.
A performance report and guide for running the Qwen3.8-27B model on dual RTX 3090 GPUs using vLLM, achieving 308 tokens per second with 32 concurrent streams, with a GitHub repository for local deployment configurations.
An adaptive KV cache streaming fork of llama.cpp enables running the Qwen 3.8 27B model with 262K context on a 16GB RTX 5070 Ti GPU, achieving ~25 tok/s by efficiently managing memory between RAM and VRAM.
A user on a dual RTX 3090 setup discovers that llama.cpp's --split-mode tensor runs prompt processing on CPU (~400 t/s), while switching to --split-mode layer unlocks GPU prompt processing over 1600 t/s, with only a modest drop in token generation speed.