grid_sampler_2d: removed lowering #1134
Conversation
There is now a decomposition in pytorch that seems to have better performance, see benchmarks at pytorch/pytorch#84350 [ghstack-poisoned]
There is now a decomposition in pytorch that seems to have better performance, see benchmarks at pytorch/pytorch#84350 ghstack-source-id: d8f7d54 Pull Request resolved: #1134
There is now a decomposition in pytorch that seems to have better performance, see benchmarks at pytorch/pytorch#84350 [ghstack-poisoned]
There is now a decomposition in pytorch that seems to have better performance, see benchmarks at pytorch/pytorch#84350 ghstack-source-id: 7ff6b8a Pull Request resolved: #1134
| make_fallback(aten._embedding_bag) | ||
| make_fallback(aten._embedding_bag_forward_only) | ||
| make_fallback(aten._fused_moving_avg_obs_fq_helper) | ||
| make_fallback(aten.grid_sampler_2d_backward) |
There was a problem hiding this comment.
I don't think this PR resolves that, no? Since we're currently applying decompositions after autograd.
There was a problem hiding this comment.
You are right! Thanks, fixed now.
There is now a decomposition in pytorch that seems to have better performance, see benchmarks at pytorch/pytorch#84350 [ghstack-poisoned]
There is now a decomposition in pytorch that seems to have better performance, see benchmarks at pytorch/pytorch#84350 ghstack-source-id: 3e6559d Pull Request resolved: #1134
|
The test failures are in |
|
@fdrocha yeah they're not caused by this PR, but by updating the pin. |
Thanks @Chillee . I was able to get the tests to go green by replacing "fake_result" with "val" in Should I make those changes? Don't understand those parts of the code base, but it seems like the right thing... There also a bunch of other instances of "fake_result" in torch repo that should probably be changed to "val" FWIW |
|
@fdrocha Yeah we just need to get a PR in updating all the uses (also see pytorch/pytorch#84432). |
I see. I guess I will wait for that PR to be merged before trying to merge this one. |
|
We could also just skip the tests until the upstream breaking is fixed. |
|
Yeah let's do that and land this. |
There is now a decomposition in pytorch that seems to have better performance, see benchmarks at pytorch/pytorch#84350 [ghstack-poisoned]
There is now a decomposition in pytorch that seems to have better performance, see benchmarks at pytorch/pytorch#84350 ghstack-source-id: c91a2c5 Pull Request resolved: #1134
There is now a decomposition in pytorch that seems to have better performance, see benchmarks at pytorch/pytorch#84350 [ghstack-poisoned]
There is now a decomposition in pytorch that seems to have better performance, see benchmarks at pytorch/pytorch#84350 ghstack-source-id: 986b35a Pull Request resolved: #1134
There is now a decomposition in pytorch that seems to have better performance, see benchmarks at pytorch/pytorch#84350 [ghstack-poisoned]
There is now a decomposition in pytorch that seems to have better performance, see benchmarks at pytorch/pytorch#84350 ghstack-source-id: adef10f Pull Request resolved: #1134
There is now a decomposition in pytorch that seems to have better performance, see benchmarks at pytorch/pytorch#84350 [ghstack-poisoned]
There is now a decomposition in pytorch that seems to have better performance, see benchmarks at pytorch/pytorch#84350 ghstack-source-id: 4f2951f Pull Request resolved: #1134
There is now a decomposition in pytorch that seems to have better performance, see benchmarks at pytorch/pytorch#84350 [ghstack-poisoned]
There is now a decomposition in pytorch that seems to have better performance, see benchmarks at pytorch/pytorch#84350 ghstack-source-id: 872e1f1 Pull Request resolved: #1134
There is now a decomposition in pytorch that seems to have better performance, see benchmarks at pytorch/pytorch#84350 [ghstack-poisoned]
There is now a decomposition in pytorch that seems to have better performance, see benchmarks at pytorch/pytorch#84350 ghstack-source-id: 93522ab Pull Request resolved: #1134
|
This looks like it was not properly merged @fdrocha -- the merge button doesn't work with ghstack |
Drat. I did use Any idea on how to fix this? |
|
@ezyang what permissions do we need to land ghstack PRs? |
|
This is landed now. Apparently you need to be repo admin lol |
|
oops actualy not yet |
There is now a decomposition in pytorch that seems to have better performance, see benchmarks at pytorch/pytorch#84350 ghstack-source-id: 93522ab Pull Request resolved: #1134
|
@fdrocha we should start sending PRs without ghstack then :( |
|
@ezyang this is not merged yet, right? |
|
I think this was merged a couple days ago: |
|
Ok, never mind then!
…On Mon, Sep 19, 2022 at 10:22 AM Mario Lezcano Casado < ***@***.***> wrote:
I think this was merged a couple days ago:
[image: image]
<https://user-images.githubusercontent.com/3291265/190987637-f4b3a3a1-ff7b-49c1-85b9-c3f15724f7d8.png>
—
Reply to this email directly, view it on GitHub
<#1134 (comment)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AX23VOJ7C2332KWTG2HXJCTV7AWGVANCNFSM6AAAAAAQGB3E4U>
.
You are receiving this because you were mentioned.Message ID:
***@***.***>
|
|
I was talking to @ezyang and his opinion was that "we should just ask him to merge ghstack PRs if we want them landed" (for now). |
Stack from ghstack (oldest at bottom):
There is now a decomposition in pytorch that seems to have
better performance, see benchmarks at
pytorch/pytorch#84350