Writing Optimized Triton GPU Kernels
vipshop/cache-ditSource at 929041e (opens GitHub)
Not tested yet.
AIR hasn't built a cheaper version of this skill yet. Check it to see what AIR would try.
On a smaller model
On a smaller model, the results dropped.
A separate result from the token saving: this is the cost of running the same skill on a smaller model, and the two are never added together.
| Task | Results | Cost |
|---|---|---|
| Fused vector op kernel | Lower results | 1% dearer |
| Fix softmax masking bug | Not compared | Not compared |