Update VALIDATION.md
Browse files- VALIDATION.md +29 -0
VALIDATION.md
ADDED
|
@@ -0,0 +1,29 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Validation
|
| 2 |
+
|
| 3 |
+
Required on NVIDIA Thor with CUDA 13 and CUTLASS DSL 4.4.x, 4.5.x, or 4.6.x:
|
| 4 |
+
|
| 5 |
+
```bash
|
| 6 |
+
python tests/test_fa4_cute_runtime.py --backend source --mode full
|
| 7 |
+
python tests/test_fa4_cute_runtime.py --backend installed --mode full \
|
| 8 |
+
--artifact build/torch-cuda
|
| 9 |
+
```
|
| 10 |
+
|
| 11 |
+
The gate covers non-causal and causal GQA, head dimensions 48/72/128, and the
|
| 12 |
+
PI0.5 D256 encoder family. D256 cases include dense lengths 320/968 and padded
|
| 13 |
+
`seqused_k` lengths 456/712 in a 968-token buffer. Every case reports
|
| 14 |
+
max/p99/mean/cosine metrics and requires bitwise CUDA Graph replay after JIT
|
| 15 |
+
warmup.
|
| 16 |
+
|
| 17 |
+
On 2026-08-06, the current source snapshot passed `8/8` rows on NVIDIA Thor
|
| 18 |
+
(SM110), PyTorch `2.13.0+cu130`, CUDA 13.0, and CUTLASS DSL 4.4.2. D48/D72/D128
|
| 19 |
+
and all four D256 rows passed; worst p99 absolute error was `0.000244` and
|
| 20 |
+
worst cosine was `0.99999982`. CUDA Graph replay was bitwise deterministic.
|
| 21 |
+
The torch-noarch final-layout import probe also passed from the isolated Thor
|
| 22 |
+
environment.
|
| 23 |
+
|
| 24 |
+
The same `8/8` source matrix passed on the same Thor with CUTLASS DSL 4.6.1.
|
| 25 |
+
The 4.6 compatibility aliases cover public API relocations only; D48/D72/D128
|
| 26 |
+
and D256 dense/seqused results matched the 4.4.2 gate.
|
| 27 |
+
|
| 28 |
+
Installed-artifact execution and performance must still be rerun after every
|
| 29 |
+
Hub rebuild. Source qualification alone is not a release artifact claim.
|