Data for "Accuracy is not compatibility: two Evo 2 checkpoints hinge on one FP8 matrix multiplication"
Authors/Creators
Description
Raw recordings, logits, plans, run outputs and analyses behind the preprint "Accuracy is not compatibility: two Evo 2 checkpoints hinge on one FP8 matrix multiplication", which the software repository summarises but does not hold. D1: the RTX 4060 FP8 tensor-core hardware ground truth (173 recordings, manifest SHA-256 be1270ed64432a1311ab5e2c98ccb69028f72e3424ad4c6148d4cd468fba6695), as one tar. D2: the raw output of the H100 PCIe probe of 15 Oct 2026, including the full output bits of the 21 FP8 projections of Evo 2 1B. D3: the live split-K calls of the H100 PCIe sweep of 16 Oct 2026. D4, D8: the H100 PCIe official Evo 2 1B and 40B bf16 logits. D5: RTX 5070 Ti Evo 2 1B per-position log-probabilities. D6: the RTX 4060 official Evo 2 1B and 40B logits. D7: the plans of the two H100 rentals (Chinese, with English summaries in the README). D9: the plan and results of runs A7, A9, A11 and A12. D10: the locally hashed (self-attested) localisation plan PLAN_C with its deviation log, code, outputs and block-0 tables. D11, D12: the matched residual interventions (35 Evo 2 1B cells; plan, code, results; captured tensors). D13: the RTX PRO 6000 follow-up on a CUDA 13 stack (1B, 7B, 40B; forward, generation and BRCA1 tests; operator comparisons). D14: the RTX 5070 Ti Evo 2 40B BRCA1 500-variant follow-up. D15: the RTX 4060 exploratory probes of 10-11 Oct 2026. D16: the paper's number ledger, post hoc analyses and figure scripts. The software, mmaemu 0.1.0, emulates a reference GPU's FP8 GEMMs bit for bit on the recorded evidence; the rest of Evo 2 runs natively, so whole-model outputs are close to the reference GPUs', not identical. Plans and logs were scrubbed of instance identifiers, host names, key names, AI-assistant session paths and local account information (exceptions kept as recorded where checksums pin the files: six members of the D1 tar and one PLAN_C script, see README.md); README.md lists every placeholder and FILES.tsv every file with its SHA-256.
Technical info (English)
Each item directory is deposited as one zip archive. Download all files into one directory, unzip every archive there (for z in *.zip; do unzip -q "$z"; done) and verify with sha256sum -c SHA256SUMS (macOS: shasum -a 256 -c SHA256SUMS); ARCHIVES.sha256 lists the SHA-256 of every other deposited file. D12 (about 13 GB of captured tensors) belongs to D11; copy its runs/ into D11's directory to check D11's MANIFEST.sha256. Arrays are numpy .npz (load with allow_pickle=False). PyTorch files (.pt: block-0 tables, captured tensors, two recordings in D1 and D2) are Python pickles: load them only with torch.load(..., weights_only=True). D1 replays with tools/replay_truth.py of the mmaemu repository (MMAEMU_TRUTH_DIR=<dir>/tcemu_ada_truth). Scripts read inputs from absolute paths on the author's workstation; README.md maps them to the deposited items. Clock times are Pacific time (PDT) unless marked Z (UTC).
Files
D10_plan_c.zip
Files
(14.8 GB)
| Name | Size | |
|---|---|---|
|
md5:97342ca631add301fd9d55f0949105be
|
1.9 kB | Download |
|
md5:b632912ed02e8a1f0261af37cf2a8b58
|
798.9 MB | Preview Download |
|
md5:4e7776dfefeee43301f8fd219ca1b2ae
|
584.5 kB | Preview Download |
|
md5:c5b183d9f19f4d0394609cb77be74398
|
9.0 GB | Preview Download |
|
md5:aa2a164fbc1151a34e57f0181ab3789f
|
1.3 GB | Preview Download |
|
md5:48f6257a34fdda488981b42d9d4e2585
|
22.4 MB | Preview Download |
|
md5:d4551f487780a8a2a880741d0b56fa00
|
32.1 MB | Preview Download |
|
md5:06f0ef8c61409467f4b2e935bf652517
|
657.4 kB | Preview Download |
|
md5:dc60723982a7834eeb274e96d465a902
|
2.1 GB | Preview Download |
|
md5:c53faf4393e0204f8998b243a7ad2733
|
1.2 GB | Preview Download |
|
md5:082187d6dacee7269c2b70141fb8a546
|
66.2 MB | Preview Download |
|
md5:9707ec2667ebd49ea65971e346e4996b
|
2.7 MB | Preview Download |
|
md5:a771f20af5989af2054b8b23249e2613
|
7.9 MB | Preview Download |
|
md5:d21dedd20da796dad497a26a09b7dae8
|
5.7 MB | Preview Download |
|
md5:a7961dff90f80fdc702a34673828aae6
|
15.7 kB | Preview Download |
|
md5:dfd249e2a333bdb49d9dc3c215da334b
|
14.7 MB | Preview Download |
|
md5:ddcf539759f313d66ffe0e4d8ea5474e
|
197.5 MB | Preview Download |
|
md5:40646c79d5ff40c708a4134053e680ae
|
233.5 kB | Download |
|
md5:2e32edf775b479e447393b668ec611c5
|
5.6 kB | Preview Download |
|
md5:5caa66883efc7b5af1851325874187fc
|
36.1 kB | Preview Download |
|
md5:7d7493b73237808fc2769f5cd2a5b007
|
15.7 kB | Download |
|
md5:17230c1b870246d1c0eff36798e2a235
|
200.7 kB | Download |
Additional details
Additional titles
- Subtitle (English)
- FP8 GEMM hardware recordings (RTX 4060, H100 PCIe), Evo 2 logits, localisation and intervention runs, follow-ups on two sm_120 GPUs, and the paper's number ledger
Related works
- Is supplement to
- Software: 10.5281/zenodo.23440103 (DOI)
- Software: https://github.com/astding/mmaemu (URL)
- References
- Journal article: 10.1145/3830409 (DOI)
- Journal article: 10.1038/s41586-026-10176-5 (DOI)
- Journal article: 10.1038/s41586-018-0461-z (DOI)
Dates
- Collected
-
2026-10-10/2026-10-21
References
- Khattak, F. and Mikaitis, M. (2026). Accurate Models of NVIDIA Tensor Cores. ACM Transactions on Architecture and Code Optimization 23(3), 1-25. https://doi.org/10.1145/3830409
- Brixi, G. et al. (2026). Genome modelling and design across all domains of life with Evo 2. Nature 652(8112), 1349-1361. https://doi.org/10.1038/s41586-026-10176-5
- Findlay, G. M. et al. (2018). Accurate classification of BRCA1 variants with saturation genome editing. Nature 562(7726), 217-222. https://doi.org/10.1038/s41586-018-0461-z