Merge pull request #8 from cksac/copilot/implement-turboquant-model

Address code review: use actual prime for table_size, clarify test assertions

Add hash-based weight compression module (TurboQuant-Model)

Initial plan

feat: per-group blockwise calibration (4-bit PPL -0.51, KLD -26%)\n\nAdd per-group alpha correction to blockwise calibration. Instead of a\nsingle scalar per row (M,), each group gets its own learnable correction\n(M,G), injected into weight_norms during the forward pass for gradient\nflow through the per-group norm scaling.\n\nResults on Qwen3.5-0.8B-Base (4-bit, 4s/50i):\n- Per-row: PPL 13.6971 (-0.259), KLD 0.1170 (-10%), 12.9 min\n- Per-group: PPL 13.4427 (-0.514), KLD 0.0959 (-26%), 14.0 min\n\nPer-group is 2x better PPL and 2.6x better KLD at only 8% more time.\nRecovers 28.1% of the quantization gap (vs 14.2% for per-row).\n\nChanges:\n- CalibrationConfig.per_group: default True\n- Blockwise cal: per-group alpha via weight_norms injection\n- _fold_alpha: handle (M,G) alpha shape\n- Test script: --per-group flag\n- Updated docs and site with per-group results"

feat: block-wise norm calibration (4-bit PPL -0.26, KLD -10%)\n\nAdd block-wise end-to-end norm calibration that optimizes per-row norms\nthrough each transformer block sequentially, using MSE + angular + KLD\nloss against pre-captured FP targets.\n\nKey results on Qwen3.5-0.8B-Base (4-bit):\n- PPL: 13.9564 → 13.6971 (-0.2592)\n- KLD: 0.1301 → 0.1170 (-10.1%)\n- Calibration time: ~13 min (4 samples, 50 iters)\n\nPer-layer calibration was harmful (+0.07 PPL), confirming that\nlocally optimal norms don't compose through the network.\n4+4 residual doesn't benefit (already cos≈1.000).\n\nChanges:\n- norm_calibration.py: calibrate_norms_blockwise() with sequential\n block processing, FP target pre-capture, exp parameterization\n- cli.py: --calibrate flag and calibrate subcommand now use blockwise\n- Defaults: n_samples=4, n_iters=50 (3.9x faster, equal quality)\n- docs/techniques/blockwise-calibration.md\n- site/src/app/techniques/blockwise-calibration/page.tsx\n- Updated nav chain: Norm Compression → Block-wise Cal → QJL"

feat: CPU offload for pass 2 + embedding quantization (INT8/INT4)

feat: add comprehensive survey of recent quantization papers and their compatibility with TurboQuant

feat: add MMLU benchmark for f16 reference and 4+4bit per-layer rotation with factored_int8 norm

feat: implement per-layer rotation strategy in quantization and add corresponding tests

Untrack .venv-wsl and add to .gitignore

feat: enhance norm codec support with factored_int4 and update related tests

Add comprehensive tests for polar decomposition quantization and compression methods

Add documentation for quantization techniques and optimizations

Add norm codec documentation and implement entropy coding technique