[STABLE ABI] Port forced_align #4079
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/audio/4079
Note: Links to docs will display an error until the docs builds have been completed. ❗ 1 Active SEVsThere are 1 currently active SEVs. If your PR is affected, please view them below: ✅ No FailuresAs of commit b308882 with merge base 3b0e7a6 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
pearu
force-pushed
the
pearu/stable-abi-forced_align
branch
from
b3a5a0e to
282bd7f
Compare
samanklesaria
left a comment
samanklesaria
left a comment
There was a problem hiding this comment.
Looks good to me!
pearu
force-pushed
the
pearu/stable-abi-forced_align
branch
from
007452e to
bae56dd
Compare
pearu
force-pushed
the
pearu/stable-abi-forced_align
branch
from
bae56dd to
b308882
Compare
NicolasHug
left a comment
NicolasHug
left a comment
There was a problem hiding this comment.
Thanks a lot for the PR @pearu .
Agreed to move forward with this, and let's try to upstream the components in libtorchaudio/stable as much as possible.
| STD_TORCH_CHECK( | ||
| targets.size(1) == torchaudio::util::max<index_t>(targetLengths), | ||
| "target length mismatch"); | ||
| }); |
There was a problem hiding this comment.
On the above, can you help me understand why we need to use STABLE_DISPATCH_INDEX_TYPES now? Basically, which part of at::max(inputLengths).item().toInt() do we need to work-around? Is this a permanent workaround, or something we'll be able to simplify in the future when more APIs are ported to the stable part?
There was a problem hiding this comment.
Currently, targetLengths can be a tensor with int32 or int64 items. If one uses, say, torchaudio::util::max<int32_t>(targetLengths) when targetLengths dtype is int64, the result of max may be wrong. The usage of STABLE_DISPATCH_INDEX_TYPES makes calling max safe.
We can simplify this in future after we are able to rewrite at::max(inputLengths).item().toInt() using stable ABI methods.
3 participants
As in the title.
This PR is marked as Draft as it requires the torch PR stack starting at pytorch/pytorch#161891 (read: CI failures are expected).
Same as #4022 but using C++ torch::stable::Tensor API instead of C shim Tensor API and this PR includes also the GPU port.
PLEASE NOTE THAT THE TORCHAUDIO REPOSITORY IS NO LONGER ACTIVELY MONITORED. You may not get a response. For open discussions, visit https://discuss.pytorch.org/.