Currently there is `test_vertical_fusion1` which fuses entirely during
the lowering stage and no buffers are realized. This adds
`test_scheduler_vertical_fusion1` which is the same test but with
several intermediate calculations realized so the scheduler is left
to do the fusion.
To support the test, this PR also adds:
- `metrics.ir_nodes_pre_fusion` which when compared with
`generated_kernel_count` tells us how many nodes were fused.
- `torch._test_inductor_realize` which is an identity operator in
eager, but under inductor also forces the input to be realized.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/90014
Approved by: https://github.com/jansel
This improves hf_Bert 1.139x->1.21x, currently lowmem dropout doesn't work for nn.Dropout module, and before this change we were recomputing all the dropout masks in a very inefficient kernel. This change pushes dropout masks to be saved in the dropout kernels where they are first computed.
cc @mlazos @soumith @voznesenskym @yanboliang @penguinwu @anijain2305 @EikanWang @jgong5 @Guobing-Chen @chunyuan-w @XiaobingSuper @zhuhaozhe @blzheng @Xia-Weiwen @wenzhe-nrv @jiayisunx
Pull Request resolved: https://github.com/pytorch/pytorch/pull/88046
Approved by: https://github.com/Chillee
