File size: 1,347 Bytes
e22b68b | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 | # Validation
Required on NVIDIA Thor with CUDA 13 and CUTLASS DSL 4.4.x, 4.5.x, or 4.6.x:
```bash
python tests/test_fa4_cute_runtime.py --backend source --mode full
python tests/test_fa4_cute_runtime.py --backend installed --mode full \
--artifact build/torch-cuda
```
The gate covers non-causal and causal GQA, head dimensions 48/72/128, and the
PI0.5 D256 encoder family. D256 cases include dense lengths 320/968 and padded
`seqused_k` lengths 456/712 in a 968-token buffer. Every case reports
max/p99/mean/cosine metrics and requires bitwise CUDA Graph replay after JIT
warmup.
On 2026-08-06, the current source snapshot passed `8/8` rows on NVIDIA Thor
(SM110), PyTorch `2.13.0+cu130`, CUDA 13.0, and CUTLASS DSL 4.4.2. D48/D72/D128
and all four D256 rows passed; worst p99 absolute error was `0.000244` and
worst cosine was `0.99999982`. CUDA Graph replay was bitwise deterministic.
The torch-noarch final-layout import probe also passed from the isolated Thor
environment.
The same `8/8` source matrix passed on the same Thor with CUTLASS DSL 4.6.1.
The 4.6 compatibility aliases cover public API relocations only; D48/D72/D128
and D256 dense/seqused results matched the 4.4.2 gate.
Installed-artifact execution and performance must still be rerun after every
Hub rebuild. Source qualification alone is not a release artifact claim.
|