Open JBloodless opened 11 months ago
@JBloodless check the pytorch issue I think they might have fixed it? You'll need to use a version of pytorch from after the patch in that issue
(we also need to get turbine setup with a nightly pytorch channel)
@JBloodless check the pytorch issue I think they might have fixed it? You'll need to use a version of pytorch from after the patch in that issue
Still getting the same error on torch 2.2.0.dev20231113. Also, the code runs (with same error), but pip throws
torch-mlir 20230924.971 requires torch==2.2.0.dev20230922, but you have torch 2.2.0.dev20231113 which is incompatible.
shark-turbine 0.9.1.dev3 requires iree-runtime>=20231026.688, but you have iree-runtime 20231004.665 which is incompatible.
which was mentioned by @stellaraccident in previous comment
This is a PyTorch error. Is there a linked PyTorch issue?
If working with a nightly you're going to need to use pip with --no-deps and manage manually. We don't currently test the nighties and the public API is set to change so we're keeping the pin to stable until we do.
This is a PyTorch error. Is there a linked PyTorch issue?
If working with a nightly you're going to need to use pip with --no-deps and manage manually. We don't currently test the nighties and the public API is set to change so we're keeping the pin to stable until we do.
Yeah, there is Pytorch issue (couple of comments up), and guys here said that the initial problem with pack/pad could be resolved in nightly, but it is not. torch-mlir
and turbine
seem to be working with nightlies, I just mentioned the logs from pip to clarify if it's okay (seems to be), thanks for response.
I pinged torch issue, but maybe there is some more development here?
@stellaraccident @dan-garvey maybe there is any progress here :)
I don't have any secret channel for getting these things resolved in upstream. Also, I don't have any RNNs from pytorch in my immediate priority list so am unlikely to have the time to find any creative solution.
Request description
Hi. Tried to convert the model with pack_padded_sequence/pad_packed_sequence and found out that they apparently are not supported yet. Since these functions are crucial for some high-perfomance models (which are aimed at fast training and inference) and I turned to IREE to further improve perfomance, I think it would be useful to add support for them.
Minimal repro:
Current error log:
What component(s) does this issue relate to?
Frontends, MLIR, Python
Additional context
No response