Validation
Source correctness covers:
- BF16 add/euler/CFG/cast for shapes:
(1024,)(1025,)(4, 4096)(2, 16, 32, 64)
- FP16 CFG on the same shapes.
- Video ops for:
(1, 4, 5, 16, 16)(2, 8, 9, 8, 8)(1, 16, 17, 16, 24)
- Action-tail boundaries
(32,7),(257,51)and(4096,1024). - Bias/tail matrices
(1,16,7),(51,64,32)and(105,257,256). - Cosmos3-Edge production contracts:
- flat velocity/action tail
(1,201,920, 3,840); - bias and zero tail
(60,64,9); - exact direct output and bitwise CUDA Graph replay.
- flat velocity/action tail
- UniPC orders
(0,1),(1,1),(1,2)and(2,2).
Correctness gates:
- exact BF16/FP16 rounded match against PyTorch reference for covered ops.
- exact FP32 match for BF16-to-FP32 cast and decode postprocess.
- fullgraph compile parity for action-tail and UniPC APIs.
- raw Cosmos Edge native parity and direct/CUDA Graph performance comparison.
Run:
python diffusion-step-ops/tests/test_diffusion_step_ops.py --backend source