[dynamic shapes] unbacked safe conv1d #154089
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/154089
Note: Links to docs will display an error until the docs builds have been completed. ❌ 1 New FailureAs of commit 3fe085e with merge base ab6cb85 ( NEW FAILURE - The following job has failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
| return torch.contiguous_format | ||
| elif input_tensor.is_contiguous(memory_format=torch.preserve_format): | ||
| return torch.preserve_format |
There was a problem hiding this comment.
I might be wrong, but it seems:
- the result is used in a
.to()call here:pytorch/torch/_meta_registrations.py
Line 2472 in 7053930
- and
memory_format=Noneis the same asmemory_format=torch.preserve_format - I'm not sure "contiguous according to
torch.preserve_format" makes sense? This is mostly used for copying the original tensor, to say "keep the original format" - how can you be contiguous w.r.t that? There's some indications this isn't supported:pytorch/aten/src/ATen/native/TensorProperties.cpp
Lines 115 to 124 in 7053930
There was a problem hiding this comment.
@bdhirsh is that something you are familiar with?
| @@ -2447,10 +2450,8 @@ def pick_memory_format(): | ||
| else: | ||
| if is_channels_last(input_tensor): | ||
| return torch.channels_last | ||
| if input_tensor.is_contiguous(memory_format=torch.contiguous_format): | ||
| if utils.definitely_contiguous_for_memory_format(input_tensor, memory_format=torch.contiguous_format): |
There was a problem hiding this comment.
we are just changing the meta function, is there a decomposition for conv that asserts a matching behaviour with the meta for unbacked?
or does the conv kernel generate the correct expected memory format anyway? if the later i wonder if its safe to do the change?
| @@ -2447,10 +2450,8 @@ def pick_memory_format(): | ||
| else: | ||
| if is_channels_last(input_tensor): | ||
| return torch.channels_last | ||
| if input_tensor.is_contiguous(memory_format=torch.contiguous_format): | ||
| if utils.definitely_contiguous_for_memory_format(input_tensor, memory_format=torch.contiguous_format): |
There was a problem hiding this comment.
do you want to return None explicitly ?
There was a problem hiding this comment.
We need the striding of convolution to be correct because we fallback to aten convolution for it. If don't guard, and we get the striding wrong, inductor output code will be incorrect.
We would need a mechanism to force the contiguous/channels last path for the convolution if we're not able to infer the meta striding. but thats not implemented here, we're just changing the meta.
Ran into this exporting https://github.com/SWivid/F5-TTS
cc @ezyang @SherlockNoMad @EikanWang @jgong5 @wenzhe-nrv