0xSero TurboQuant: consumer-GPU KV compression measurements on RTX 3090 and RTX 5090

Independent reimplementation of TurboQuant (arXiv 2504.19874, ICLR 2026) with a vLLM adapter and Triton kernels. README read 2026-10-18. Claims the README marks measured, not theoretical: 3-bit keys and 2-bit values give 5.3x compression (16-bit to 3-bit); 4-bit values reach cosine similarity 0.997 against 0.94 at 2-bit values; on 4x RTX 3090 running Qwen3.5-27B, TurboQuant freed 30 GB of KV memory (the README's '2x context on dense model' row). A separate row says 8x RTX 3090 with Qwen3.5-35B-A3B-FP8 at TP=8 reaches 131K context where bf16 KV OOMs at 32K. Tested environment named in the README: vLLM 0.18.0, PyTorch 2.10, CUDA 12.8, one RTX 5090 32GB with Qwen3.5-27B-AWQ, and 8x RTX 3090 24GB with Qwen3.5-35B-A3B at TP=8. The 30 GB figure is the one this collection records; the 131K-vs-32K row is the same author's second measurement and is kept in the kv record notes rather than as a second record. Author-reported, no linked raw log, so confidence is 0.7. Not a Google Research number; the paper itself is a separate source.