Evo 2 on the RTX 4060: the official-path reference outputs, 11-12 October 2026¶

Before the RTX 4060 was replaced by the RTX 5070 Ti, the official Evo 2 path (evo2 with Transformer Engine FP8) ran on it. Every Evo 2 comparison in this repository is against these captures:

examples/data/evo2_1b_ada_rtx4060_te.npz and .json, used by the notebook and the README's Evo 2 1B results;
rtx4060_official_te_40b.npz in the Evo 2 40B log.

Hardware and software: NVIDIA GeForce RTX 4060 (sm_89, 24 SMs, 8 GB), driver 572.16, WSL2; Python 3.12.2, torch 2.7.1+cu128, cuBLAS/cuBLASLt 12.8.3.14, Transformer Engine 2.3.0, evo2 0.6.0, vtx (vortex) 1.1.0, flash-attn 2.8.0.post2 (each capture's .json in data/ records them).

Where each shipped array comes from¶

Shipped	Produced by	Run
`evo2_1b_ada_rtx4060_te.npz`: `<w>/ids`, `<w>/acgt`, `<w>/next` (p0-p3, w8k, w2k)	golden_real.py: `evo2.Evo2("evo2_1b_base")`, the real upstream model, through run_te.sh	12 Oct, 00:21; data/evo2_1b_base_real-te-fp8.json
`brca1/rs`, `brca1/vs` (435 reference and 500 variant windows)	brca1_source_script.py, its first pass (`run("TE_FP8")`: upstream `score_sequences` with TE FP8), cached as `brca1_subset500_te_scores.npz`	11 Oct, 09:34-09:56; logs/brca1_run2.log, AUROC 0.7378
`brca1/pos`, `brca1/ref`, `brca1/alt`, `brca1/lof`	the upstream BRCA1 notebook's table (Findlay et al. 2018), `sample(n=500, random_state=0)`, as in the script
`.json`: `windows` (bp, mean log p(next), SHA-256 of the full logits)	the same capture as the arrays	data/evo2_1b_base_real-te-fp8.json
`.json`: `projection0_w2k`	the hardware recording `evo2.evo2_1b_w2k`	docs/validation.md
`.json`: `h100_official_te`	the H100 run of the official path	2026-10-14/evo2-h100-sm90
40B log: `rtx4060_official_te_40b.npz` (w2k, w4k: `acgt`, `next`)	golden_oracle.py `--model evo2_40b` (see below)	w2k: 12 Oct, 00:26-00:31, logs/40b_w2k.log; w4k: 00:51-00:54 by chain_40b.sh, logs/40b_w4k.log, logs/chain_40b.log

The two shipped .npz files were assembled from these captures; the assembly was not kept as a script. check_reference.py checks every shipped array against its capture, and the SHA-256 of every capture's full logits against its .json: 0 failures (check_reference.txt). The captures with the full bf16 logits (3.5 MB for the 1B) are not in the repository; their .json files here record the SHA-256 of each window's logits.

The windows: p0-p3 are the four prompts of evo2/test/data/prompts.csv, w8k the first 8,192 bp of their concatenation, w2k and w4k its first 2,048 and 4,096 bp. They are defined by windows() of bench.py (evo2-anywhere, commit 5d6a1d9, which common.py imports), and acgt and next are computed by its summarize(): log_softmax of the bf16 logits in fp32. The rest of bench.py runs evo2-anywhere's own models and was not used.

The 40B reference: the official blocks, one at a time¶

The official 40B path does not fit in 8 GB. golden_oracle.py runs the upstream model one block at a time: each block is built by vortex's get_block and loaded by vortex's own StripedHyena.to_bfloat16_except_pr_lc and custom_load_state_dict (the TE FP8 recipe, the Wqkv permutation), and no other code touches the weights. Run the same way on the 1B, it gives the logits of the real upstream model bit for bit on all six windows (check_reference.txt; the two captures, data/evo2_1b_base_real-te-fp8.json and data/evo2_1b_base_oracle-te-fp8_p0-p1-p2-p3-w8k-w2k.json, record the same logits SHA-256). For the 40B, only w2k and w4k fit in 8 GB; p0-p3 and w8k ran out of memory (logs/chain_40b.log). The same w2k and w4k arrays had been produced on 11 October by an earlier block-by-block runner in evo2-anywhere, bit for bit.

Files¶

scripts/: the capture scripts as run, with their paths on the capture host: common.py, golden_real.py, golden_oracle.py, run_te.sh (the TE environment, with a cuBLASLt trace of every call), chain_40b.sh, compare_goldens.py, brca1_source_script.py, bench.py. /data/quant/tcemu_audit/gpu.sh, called by chain_40b.sh, only takes a lock on the GPU (flock).
data/: the metadata of the captures used (windows, logits SHA-256, versions, block hashes).
logs/: the 40B runs and the BRCA1 run. The owner column of the ls listing in chain_40b.log is redacted.
check_reference.py, check_reference.txt: the check above, run on 15 October 2026 with --goldens <captures> --xlsx 41586_2018_461_MOESM3_ESM.xlsx.

Not used by anything in this repository: the second pass of brca1_source_script.py (a lookup-table experiment; its result is also in the log), the 4 kb prefixes of p0-p3 and the cuBLASLt heuristic table that chain_40b.sh also produced. The first attempt at the BRCA1 run (10 October) stopped with a CUDA error in that second pass, before the scores were saved, so the BRCA1 scores were recorded on 11 October, not on 10 October as the BRCA1 capture's .json and, until 15 October, examples/data/evo2_1b_ada_rtx4060_te.json said.