[BUG] some triton kernels are unable to handle tensors whose shapes aren't divisible by 16, 32, or 64.

[MaxwellDing]

I encountered NaNs when training wan-2.1, NaN tensor in shape (1536, 1560) . Upon investigation, I found that some triton kernels in coat don't use boundary checks for load and store operations. This leads to memory corruption, which in turn causes NaN outputs.

kernels: transpose, division, division_transpose...

[haochengxi]

Hi, if you need the kernels urgently, you can leave your email, and I can send you some kernels with boundary checks.

[MaxwellDing]

Hi, if you need the kernels urgently, you can leave your email, and I can send you some kernels with boundary checks.

I'm good—I can fix it up right away. The reason I haven't submitted a pull request is that I'm concerned I might not be familiar with all the kernels and could end up missing some.