Skip to content

T2: full truth replay on the reviewed tree, 2026-10-15

Verdict: PASS (compare.txt), with every count equal to the run of 2026-10-14 (../../2026-10-14/t2/README.md, which explains the totals): 173 of 173 recordings in the report; 942,445,814 of 942,451,934 values match; 0 zero-sign and 0 NaN-payload mismatches; the same 6,120 value mismatches at the old emulator recording's NaN-decode outputs; hash-only 2,251 of 2,251; hardware errors 113 of 193 reproduced; 2,466 not applicable; fastacc cross-check 0 mismatches in 512,512 values.

  • Code under test: the tree of 49f1ecf (git archive, so repo_commit in replay_mmaemu.json is empty). In the reports <src> stands for that tree and <out> for the output directory.
  • Command: python tools/replay_truth.py --out <out> --compare $MMAEMU_TRUTH_DIR/replay/replay_fixed.json --threads 12, on the CPU with CUDA_VISIBLE_DEVICES="", torch 2.7.1+cpu and Python 3.12.2 (the run of 2026-10-14 used torch 2.7.1+cu128 with no visible GPU).
  • Wall time: 26 min (1547 s).
  • Truth: manifest.json sha256 be1270ed64432a1311ab5e2c98ccb69028f72e3424ad4c6148d4cd468fba6695, unchanged after the run.

Compared field by field with the reports of 2026-10-14, compare.json and replay_mmaemu.json differ only in the message of the UnsupportedFeatureError for a fused bias, which 49f1ecf rewords (it appears in the not-applicable reasons), and in the run metadata (start time, durations, paths, torch build, commit).