Jason Ansel
Jason Ansel
@pytorchbot merge
How would inductor know about MLIR filenames? Seems internal to Triton unless I am missing something...
@pytorchbot merge
@albanD it would be better for decomps to not use cudnn, inductor can pattern match to cudnn if it needs to.
You want to do something like: ``` git checkout origin/main third_party/kineto git submodule update ```
@pytorchbot merge
I approved the CI for you
@pytorchbot merge
> in torch._inductor.config. Is there a way to control these for a decomposition or should I manually do the unrolling in the decomposition? We should try not to special case...
A new inductor prim could work for masked loads. In inductor IR, a masked load is: ``` ops.masked(cond, lambda: ops.load(...), 0) ``` or if you are inlining something the `lambda:`...