Skip to content

T2: full truth replay, 2026-10-14

Verdict: PASS (compare.txt). The report lists all 173 recordings. 147 have outputs that mmaemu models; they are compared with the hardware and, count by count, with the audit's FixedModel replay. The 15 tables are listed without comparison, and the 11 bias, row-wise, fp16-accumulator or HMMA recordings have no output that mmaemu models.

  • Command: python tools/replay_truth.py --out <dir> --compare $MMAEMU_TRUTH_DIR/replay/replay_fixed.json --threads 12. It ran on the CPU with CUDA_VISIBLE_DEVICES="", torch 2.7.1+cu128 and Python 3.12.2.
  • Wall time: 26 min (1568 s).
  • Truth: manifest.json sha256 be1270ed64432a1311ab5e2c98ccb69028f72e3424ad4c6148d4cd468fba6695. The truth tree was unchanged after the run.
  • Code under test: the run started at commit 6492713 (tree c4b4cb0). The repo_commit field in replay_mmaemu.json says e173ecc, because HEAD was read when the report was written and the branch moved during the run. The two commits have identical mmaemu/ Python sources, tools/truth_adapter.py and tools/replay_truth.py. The only mmaemu/ change is mmaemu/backends/cuda/bin/, which a CPU replay does not load.

Totals: mmaemu vs FixedModel

Quantity mmaemu FixedModel
Matching values 942,445,814 of 942,451,934 1,023,913,574 of 1,023,919,694
Zero-sign mismatches 0 0
NaN-payload mismatches 0 0
Value mismatches 6,120 6,120
Hash-only recordings equal 2,251 / 2,251 2,539 / 2,539
Hardware errors reproduced 113 / 193 113 / 193
Not applicable 2,466 255

All 6,120 value mismatches are in emulator.tcemu_kernels_fixed_inputs, for both models, with equal counts per output. The adapter's emulator check places every one of them at the old-kernel emulator's NaN-decode outputs (docs/testing.md), where mmaemu gives the canonical NaN.

FixedModel also produces the outputs that mmaemu reports as not applicable, so its value counts are larger. The 2,211 additional not-applicable outputs are features that mmaemu 0.1 rejects with UnsupportedFeatureError (see docs/limitations.md): fused bias (1,962) and row-wise scales (249). The fastacc cross-check found 0 mismatches in 512,512 values.