4 Commits
Author SHA1 Message Date
Ashwin GopinathandAnirudh B. Vangara 19ebecbc4d SpectralQuant v2: calibrated eigenbasis rotation with water-filled bit allocation
Upgrades SpectralQuant from uniform bit allocation to greedy water-filling
inside the semantic subspace, with per-dimension Lloyd-Max codebooks fit to
the empirical scalar marginal. Adds full three-way evaluation harness
(perplexity, generation, latency, LongBench) on Modal H200 GPUs, comprehensive
documentation, and the consolidated arxiv-ready manuscript.

v1 is preserved on the 'v1' branch.

Co-Authored-By: Anirudh B. Vangara <anirudh@sentra.app>
2026-05-15 12:24:05 -07:00
Ashwin Gopinath d5dd891b8b Add multiregime, optimal allocation, and push-to-0.95 experiments
- experiments/multiregime_sweep.py + results (Qwen-1.5B, Qwen-7B, Llama-8B)
- experiments/optimal_allocation_sweep.py + results (Qwen-1.5B, Qwen-7B, Llama-8B)
- experiments/push_to_095.py + results (Qwen-7B)
2026-04-07 19:28:28 +08:00
Ashwin Gopinath 1254311a9c Add shaped cache sweep experiment, results, and figures
- experiments/shaped_cache_sweep.py
- results/shaped_cache/ (Qwen-1.5B, Qwen-7B, Llama-8B, combined)
- figures/ (pareto, kv_sensitivity, heatmap)
- gen_shaped_cache_figures.py
2026-04-07 01:30:31 +05:30
Ashwin Gopinath 8d5a431276 SpectralQuant: 3% Is All You Need
Complete repository for "3% Is All You Need: Breaking TurboQuant's
Compression Limit via Spectral Structure" — submitted to arXiv.

Includes source code, experiment scripts, raw results, paper (LaTeX + PDF),
figures, tests, and full reproduction instructions.
2026-04-06 16:34:46 +05:30