It’s amazing how far open source LLMs have come.
Qwen3-32b recreated the Windows95 Starfield screensaver as a web app with the bonus feature to enable “warp drive” on click. This was generated with reasoning disabled (/no_think) using a 4-bit quant running locally on a 4090.
Here’s the result: https://codepen.io/mekelef486/pen/xbbWGpX
Model: Qwen3-32B-Q4_K_M.gguf (Unsloth quant)
Llama.cpp Server Docker Config:
docker run \
-p 8080:8080 \
-v /path/to/models:/models \
--name llama-cpp-qwen3-32b \
--gpus all \
ghcr.io/ggerganov/llama.cpp:server-cuda \
-m /models/qwen3-32b-q4_k_m.gguf \
--host 0.0.0.0 --port 8080 \
--n-gpu-layers 65 \
--ctx-size 13000 \
--temp 0.7 \
--top-p 0.8 \
--top-k 20 \
--min-p 0
System Prompt:
You are a helpful expert and aid. Communicate clearly and succinctly. Avoid emojis.
User Prompt:
Create a simple web app that uses javascript to visualize a simple starfield, where the user is racing forward through the stars from a first person point of view like in the old Microsoft screensaver. Stars must be uniformly distributed. Clicking inside the window enables “warp speed” mode, where the visualization speeds up and star trails are added. The app must be fully contained in a single HTML file. /no_think
Is 13k your max context at Q4K_M?
Exllamav3 is still in development so it’s not fully optimized and could have bugs, but I get 16k context with 4bpw (which has very similar perplexity as Q4_K_M, according to developer’s own measurements) using only 22GB VRAM, since I also run my desktop env on the same computer.
I’m close to the limit at 23886MiB / 24564MiB of VRAM used when the server is running. I like to have a bit of headroom for other tasks.
But I’m by no means a llama.cpp expert. If you have any tips for better performance I’d love to hear them!
Enable flash attention if you havent already
22466MiB / 24564MiB, awesome, thank you!
You’re welcome. Also, whats your gpu and are you using cublas (nvidia) or vulcan(universal amd+nvidia) or something else for gpu postprocessing?
It’s a 4090 using cublas. I just run the stock llama.cpp server with CUDA support. Do you know if there’d be any advantage to building it from source or using something else?
If you were running amd GPU theres some versions of llama.cpp engine you can compile with rocm compat. If your ever tempted to run a huge model with partial offloaded CPU/ram inferencing you can set the program to run with highest program niceness priority which believe it or not pushes up the token speed slightly
People implement this on their calculator during class. This is the kind of thing you would write to learn programming, the definition of entry-level. You’re using a device that can execute billions of trigonometric calculations per millisecond to produce code that calculates X and Y coordinates for few dozens of points on a radial trajectory.
What the fuck…
There is, seriously, no pleasing some people. You appear to be a vocal member of this highly undignified and odious demographic.
Fair point. My original prompt asked for more, but the model wasn’t capable enough. Not sure if the “warp drive” part would be part of any standard algo.
Any ideas on challenges that are new and more fun than the “balls rolling in a hexa-,hepta-,octagon” or “simulate a solar system” prompts everyone’s using these days?
