eqy

icon location `nvprof` is dead, long live `nsys nvprof`

Results 65 comments of eqy

[cuDNN][SDPA] Remove `TORCH_CUDNN_SDPA_ENABLED=1`, enable cuDNN SDPA by default on H100 and 2nd on other archs >= sm80

@pytorchmergebot merge

[cuDNN][SDPA] Remove `TORCH_CUDNN_SDPA_ENABLED=1`, enable cuDNN SDPA by default on H100 and 2nd on other archs >= sm80

@pytorchmergebot merge

[CUDNN] Remove defunct cuDNN V8 API build flag

TODO: investigate if maxwell failures are due to no int8.

[CUDNN] Remove defunct cuDNN V8 API build flag

@pytorchmergebot rebase

[CUDNN] Remove defunct cuDNN V8 API build flag

int8 maxpooling issue forwarded to cuDNN, as it is surfacing even in recent generations...

[CUDNN] Remove defunct cuDNN V8 API build flag

@pytorchmergebot rebase

[CUDNN] Remove defunct cuDNN V8 API build flag

Note: requires cuDNN >= 9.1 for maxpooling fix

[CUDNN] Remove defunct cuDNN V8 API build flag

@pytorchmergebot rebase

[CUDNN] Remove defunct cuDNN V8 API build flag

@pytorchmergebot merge

[CUDNN] Remove defunct cuDNN V8 API build flag

@pytorchmergebot merge