CompressionV4

Reproduction artifact for the v4 paper: Practical Limits of Lossless Compression for bf16 Transformer LLM Weights.

Status: Draft v4.0.6 (final). The artifact set is end-to-end operational; the smoke test passes 3/3 locally; figures, fixtures, prompts, the regenerated bf16.zl (617 B), and the eight inference-sanity prompts are all vendored. Canonical archival snapshot: GitHub release v4.0.3 (immutable git tag; release ZIP is byte-identical to this mirror; current head v4.0.6 contains only post-archive housekeeping). Supplementary persistent identifier: Software Heritage SWHID swh:1:rel:55d910f5af170c22719cc9346f4d8a5029f09164 covers the v4.0.3 commit (SH save-task 2330392, archived 2026-05-14). No Zenodo DOI was minted; the Zenodo↔GitHub webhook installation did not complete on this repository.

Headline empirical findings

Repository

Source & results live at github.com/NimoRotem/llm-compression-limits (tag v4.0.3, release page: github.com/NimoRotem/llm-compression-limits/releases/tag/v4.0.3).

Browse

• README.md — repo entry point: status, layout, citation, reproduce recipe live
• DEVIATIONS.md — live catalog of drift between paper text and as-shipped artifacts live
• REMAINING_WORK.md — post-completion status (no work remaining for v4.0.6) live
• code/ — Apache-2.0 core harness sources live
• methods/bf16-split/ — bf16 byte-split predictor + mixture-CDF live
• methods/qk-mixture-cdf/ — Q4_K mixture-CDF coder + 277 B qb_k4.profile live
• reproductions/dfloat11/ — pointer to official kernels (Hopper-only) scoped out
• reproductions/unweight/ — pointer to official kernels (Hopper-only) scoped out
• reproductions/zipnn/ — pointer to official repo (default config roundtrip-fails) scoped out
• results/results.jsonl.zst — 7,960 rows / 11,942 verified method-evals live
• manifest/model-manifest.json — 14 source models with resolved HF SHAs live
• manifest/expected-checksums.json — three measured prefixes; smoke test 3/3 passing live
• manifest/dependencies.txt — pinned versions + official reproduction URLs live
• prompts/inference-sanity-prompts.json — 8 real prompts; gpt2 verified 8/8 (sanity_check_gpt2.jsonl) live
• profiles/bf16.zl — 617 B, regenerated v4.0.3 (fp16 deferred — see paper §6) live
• profiles/bench_bf16zl.jsonl — 290-tensor benchmark, 1.4986 geomean live
• fixtures/ — TinyLlama bf16 (8 MiB) + Llama-3.2-3B Q4_K (13.5 MiB) + Qwen2.5-0.5B GGUF (379 MiB) live
• figures/ — 5 figures × {pdf, png} rendered and committed live
• notebooks/figures.ipynb — populated (5 figure cells); render_figures.py CLI alongside live
• notebooks/lomo-tables.ipynb — populated (LOMO aggregation cells) live
• cross-layer-pairs.csv — 250 (model, role, layer-pair) rows from the cross-layer correlation atlas (paper §8) live
• provenance/iter4/FINAL_VERDICT.md live
• provenance/iter5/FINAL_VERDICT.md live
• provenance/iter6/FINAL_VERDICT.md live
• archive/ — links the canonical GitHub release v4.0.3 and the Software Heritage SWHID (previously a Zenodo-deposit stub; Zenodo deposit not pursued) live
• smoke-test/ — fixtures + checksum gate, ./reproducibility_smoke_test.sh exits 0 live

How to reproduce

git clone https://github.com/NimoRotem/llm-compression-limits.git
cd llm-compression-limits
git checkout v4.0.3
./reproducibility_smoke_test.sh    # 3/3 sha256[:16] checks pass

Expected output (sha256[:16] of compressed bytes):

• bf16_split(TinyLlama_layer3_q_proj.bin) → fc3544c35489eb94
• qb_mixture_k4(Llama32_3B_layer14_gate.q4k.bin) → c79972a64d11b717
• decomp_perstream_zstd19_bg(Qwen2.5-0.5B-Q4_K_M.gguf) → 1c6f077f8c228cf2

The script auto-fetches the 379 MiB GGUF from this mirror if absent. End-to-end roundtrip is byte-exact for all three methods.

License

Code: Apache-2.0 (LICENSE). Data / results / profiles / figures: CC-BY-4.0 (LICENSE-DATA).

v4.0.6 — last update 2026-05-14. Paper draft is final (v4.0.6); SWHID is the persistent identifier.