Add Cosmos3-Edge BF16 production gates
Browse files- VALIDATION.md +34 -0
VALIDATION.md
ADDED
|
@@ -0,0 +1,34 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Validation
|
| 2 |
+
|
| 3 |
+
Source correctness covers:
|
| 4 |
+
|
| 5 |
+
- BF16 add/euler/CFG/cast for shapes:
|
| 6 |
+
- `(1024,)`
|
| 7 |
+
- `(1025,)`
|
| 8 |
+
- `(4, 4096)`
|
| 9 |
+
- `(2, 16, 32, 64)`
|
| 10 |
+
- FP16 CFG on the same shapes.
|
| 11 |
+
- Video ops for:
|
| 12 |
+
- `(1, 4, 5, 16, 16)`
|
| 13 |
+
- `(2, 8, 9, 8, 8)`
|
| 14 |
+
- `(1, 16, 17, 16, 24)`
|
| 15 |
+
- Action-tail boundaries `(32,7)`, `(257,51)` and `(4096,1024)`.
|
| 16 |
+
- Bias/tail matrices `(1,16,7)`, `(51,64,32)` and `(105,257,256)`.
|
| 17 |
+
- Cosmos3-Edge production contracts:
|
| 18 |
+
- flat velocity/action tail `(1,201,920, 3,840)`;
|
| 19 |
+
- bias and zero tail `(60,64,9)`;
|
| 20 |
+
- exact direct output and bitwise CUDA Graph replay.
|
| 21 |
+
- UniPC orders `(0,1)`, `(1,1)`, `(1,2)` and `(2,2)`.
|
| 22 |
+
|
| 23 |
+
Correctness gates:
|
| 24 |
+
|
| 25 |
+
- exact BF16/FP16 rounded match against PyTorch reference for covered ops.
|
| 26 |
+
- exact FP32 match for BF16-to-FP32 cast and decode postprocess.
|
| 27 |
+
- fullgraph compile parity for action-tail and UniPC APIs.
|
| 28 |
+
- raw Cosmos Edge native parity and direct/CUDA Graph performance comparison.
|
| 29 |
+
|
| 30 |
+
Run:
|
| 31 |
+
|
| 32 |
+
```bash
|
| 33 |
+
python diffusion-step-ops/tests/test_diffusion_step_ops.py --backend source
|
| 34 |
+
```
|