加入
KernelBench operators
Faster GPU kernels for PyTorch operators, exactly as correct as the reference (placeholder)
比较改动
选两个分支看看改了什么，再发起 PR 把工作带过去。
可以合并。 这两个分支可以自动合并。
登录后才能发起 PR。
2 个提交
Score: 2026-09-24: the check failed, softmax differs from the reference beyond atol 1e-2
ada-codex 上个月提交
28e58c6 ada-codex 上个月提交Start fp16-everywhere
ada-codex 上个月提交
dc585d8 ada-codex 上个月提交+16 −0 routes/fp16-everywhere/README.md 新增 查看文件
| @@ -0,0 +1,16 @@ | ||
| 1 | +--- | |
| 2 | +issue: 3 | |
| 3 | +status: done | |
| 4 | +tags: systems | |
| 5 | +--- | |
| 6 | +# fp16-everywhere | |
| 7 | + | |
| 8 | +Half precision throughout. | |
| 9 | + | |
| 10 | +## What happened | |
| 11 | + | |
| 12 | +- 2026-09-24: the check failed, softmax differs from the reference beyond atol 1e-2 | |
| 13 | + | |
| 14 | +## Next | |
| 15 | + | |
| 16 | +Fails the tolerance on softmax and cumsum; keep fp32 accumulators. | |