T3 on an H100 PCIe (sm_90), 2026-10-14

Verdict: every per-file run exits 0 (summary.txt). Tests that need the ground truth were skipped, so this run does not include the device-independence check.

• Hardware: NVIDIA H100 PCIe (sm_90, 114 SMs), rented; torch 2.7.1+cu128, CUDA 12.8 (fixed-fp8.json, fixed-fp8.log).
• Software: this run did not record the driver, Python or cuBLASLt; the Evo 2 baseline on the rented H100 the same day recorded driver 570.148.08, Python 3.12.15 and cuBLASLt 12.8.3.14 (../evo2-h100-sm90/windows.json).
• Code under test: the tree of 69cd231 (git archive), with the frozen release fatbins (rc1/v0.1.0, which include sm_90 SASS).
• Command: PYTHON=<venv>/bin/python tools/run_gpu_tests.sh --out <dir>. MMAEMU_TRUTH_DIR and MMAEMU_AUDIT_DIR were not set, because the ground truth and the audit data were not copied to the rented host.

File	Result	Skips
test_kernels_ieee	79 passed: v1 and v4 from the fatbin and from NVRTC against the reference, and NVRTC equal to the fatbin	none
test_cuda_runtime	13 passed	none
test_hw_mini	3 passed: reference, cuda:v1 and cuda:v4 on the hardware-mini cases recorded on the RTX 4060	cuda is not a status entry; cuda:v9 is gated and sm_90 is not in its validated_on; cuda:tc3 has no sm_90 binary
test_fast_kernels	12 passed (v9)	12: no tc3 binary for sm_90; 6: no ground truth or audit data
test_gates_hmma	1 passed	3: they need a device where the tc3 gates pass
test_device_independence	0 passed	all 138: MMAEMU_TRUTH_DIR is not set
test_live_cublas	skipped	T3p runs only on the profile device (RTX 4060)
test_scaffold, test_backend_reference, test_epilogue	2, 3 and 5 passed	none

Evo 2 fixed inputs (fixed-fp8.json)

The same host ran the test bundle's fixed_fp8.py before T3.

• Code: mmaemu 5c15e8a, with v1 and v4 compiled by NVRTC (the bundle had no fatbins). Their sources differ from 69cd231 only in comments. T3's test_nvrtc_equals_fatbin passed on this GPU (see test_kernels_ieee above).
• Inputs: the recorded FP8 operands of the 21 input-projection GEMMs of one Evo 2 1B 2 kb window (evo2.evo2_1b_w2k).
• v1 and v4: all 21 unit-scale fp32 accumulators and all 21 real-scale bf16 outputs are bitwise equal to the RTX 4060 recording (42 of 42 rows PASS).
• Native torch._scaled_mm on this GPU differs from the RTX 4060 output on all 21 calls (native_observations).

Why sm_90 is not in validated_on

v1 and v4 list sm_89 and sm_120. sm_120 was added after the full device-independence suite and the wheel check without nvcc (validation_logs/2026-10-13-sm120). Neither ran on sm_90. v1 and v4 are not gated, so AUTO still selects v4 on sm_90; validated_here is False there.