Skip to content

T3 on an RTX 5070 Ti (sm_120) on fa0fd90, 2026-10-16

The GPU test suite on the fatbins of the release-candidate freeze (../rc-fatbin/README.md), run by tools/run_gpu_tests.sh --di-full --out <this directory> from a clean worktree of fa0fd90 (63b7395, later, renamed the profiles; T3 on its code: ../rtx-5070-ti-renamed/README.md), with the ground truth and the audit data present: one pytest run per test file under the GPU lock, its output in test_<file>.log and di_full_*.log, and summary.txt, the script's one line per run. Every run exits 0 (or 5, nothing collected). <src> is the worktree, <out> the scratch output directory.

Against the run of 2026-10-14 on e173ecc (../../2026-10-14/t3-rtx-5070-ti-sm120/README.md), file by file:

  • equal: scaffold, backend reference, epilogue, fast kernels (91 passed; 91 skip, no tc3 binary for sm_120), gates (1 passed, 3 skipped), hardware mini (3 passed, 3 skipped), device independence (27 passed);
  • test_kernels_ieee 167 passed (79 then) and test_cuda_runtime 19 passed (13 then): the tests of h1 and h4, each kernel against the CPU reference of its own preset;
  • test_live_cublas 2 skipped (1 then): its second test, added since, also needs the profile device (the RTX 4060);
  • test_h100_mini (new in the script): 310 passed, its GPU cases (the 124 live H100 calls through the reference backend on the GPU and through auto, which runs h4 here, and the E4M3 x E4M3 ones through h1); -m gpu deselects its 125 CPU cases;
  • test_backend_selection exits 5, nothing collected: a CPU-only test file with no GPU-marked test (it runs in T0).

Full accumulators (--di-full, MMAEMU_DI_FULL=<backend>, 7 Evo 2 calls per locked run): the reference backend, v1 and v4 reproduce the RTX 4060's full-accumulator sha256 on all 21 calls each; cuda (an alias, not a status entry), v9 (gated, sm_120 not in its validated_on) and tc3 (no sm_120 binary) skip.

The same suite ran on 7612fd8 (the release-candidate freeze) earlier the same day with the same counts in every file but test_scaffold (102 deselected then, 103 now: a CPU test added since). A first run on 7612fd8 pointed MMAEMU_AUDIT_DIR one directory too high, so the fast-kernel cases that read the audit data were not collected (13 passed, 17 skipped); the run with the correct directory replaced it.