X / @superalesha · alesha.pro · t.me/fuckup_files
By day I run infrastructure. The rest of the time I bench LLMs on a hand-built 4x RTX 3090 rig with 96 GB of VRAM, and I publish every measurement with raw traces and per-run provenance attached. People argue about quants and reasoning effort from vibes. I run the thing and count.
| Project | Stars | What it is |
|---|---|---|
| atlas |  |
Interactive canvas that takes an LLM apart tensor by tensor. It measures INT8/INT4/FP8 error, distributions, spectra and outlier channels on every weight tensor |
| evalatro |  |
The open benchmark where language models play real Balatro |
| 4x3090-llm-benchmarks |  |
Public archive of local LLM speed measurements from the rig, with exact launch commands and per-run provenance |
| vllm-cross-model-kv |  |
Cross-model KV cache transfer in vLLM: a smaller model prefills, a bigger one answers from its cache |
| qwen38-27b-bench-4x3090 |  |
Raw traces from the 67-hour Qwen3.8-27B quant grinder, GGUF sweeps plus a 288-point vLLM vs SGLang matrix |
| tools |  |
ComfyUI workflows and benchmark configs from the rig |
The rig is 4x RTX 3090 with 96 GB of VRAM in total. P2P is verified on all 12 pairs, the cards run at 220 W each on PCIe 3.0 x16. Engines on the box: vLLM, llama.cpp, exllamav3 + tabbyAPI, plus SGLang when it behaves. Ampere sm_86 has no native FP8, and most quant advice out there assumes everyone owns a Hopper. I enjoy measuring what actually wins on these cards.
- I took Qwen3.8-27B apart to see how it works inside. I tapped every FFN neuron, all 1,114,112 of them, and pushed 3.92M tokens through in 87 minutes. Neurons that fire on more than 90% of tokens: 2.
- Settling the quant debate for Qwen3.8 27B once and for all. I ran 8 quants through 720 real agentic and reasoning tasks each. That is 5,760 runs and 67 hours on home hardware.
- 1-bit quant scoring 76.9 vs 79.1 for bf16. The table looked too good to be true, so I downloaded the model and checked it on real agent loops.
- Three local models shipped a game, a video and a soundtrack. Qwen3.8 27B wrote the shooter, Minimax H3 rebuilt my run in its own render, Minimax Music 3 scored it. Three local models, one apartment, zero cloud.
- Reasoning effort, measured. I burned 30.9M tokens across 83 arms. xhigh costs 4.1x the tokens of off for 12 extra solved tasks out of 150.
Every number on this profile comes from a real run. If it is not in a repo with raw data, I did not say it.