ARM-software / Tool-Solutions

Tutorials & examples for Arm software development tools.
Apache License 2.0
252 stars 138 forks source link

Build Failure in Tool-Solutions/docker/pytorch-aarch64 #254

Open vishwascm opened 1 week ago

vishwascm commented 1 week ago

I tried to build Pytorch using latest commit in Tool-Solutions/docker/pytorch-aarch64 and found the following error:

Command used Tool-Solutions/docker/pytorch-aarch64/build.sh

Error

repo_url=https://review.mlplatform.org/ml/ComputeLibrary/
++ echo https://review.mlplatform.org/c/ml/ComputeLibrary/+/12818/1
++ awk -F/ '{print "refs/changes/" substr($(NF-1),length($(NF-1))-1,2) "/" $(NF-1) "/" $(NF)}'
+ refname=refs/changes/18/12818/1
+ git fetch https://review.mlplatform.org/ml/ComputeLibrary/ refs/changes/18/12818/1
From https://review.mlplatform.org/ml/ComputeLibrary
 * branch                  refs/changes/18/12818/1 -> FETCH_HEAD
+ git cherry-pick FETCH_HEAD
Committer identity unknown

*** Please tell me who you are.

Run

  git config --global user.email "you@example.com"
  git config --global user.name "Your Name"

to set your account's default identity.
Omit --global to set the identity only in this repository.

fatal: unable to auto-detect email address (got 'root@ip-172-32-5-129.(none)')

My git credentials are set properly.

jondea commented 1 week ago

Hi @vishwascm, thanks for the bug report!

I'm not sure how your machine is set up, but it does look like you haven't set up an email address and name for git. Have you tried running those two suggested commands (even with some example email and name)?

That being said, I think we can fix this another way. Could you try adding --no-commit to the cherry-pick here please and let me know if it works for you? https://github.com/ARM-software/Tool-Solutions/blob/4d593cb2559f19da9b92929b1f5f7ea5e1f9b8f2/docker/pytorch-aarch64/get-source.sh#L59

vishwascm commented 4 days ago

Hi @vishwascm, thanks for the bug report!

I'm not sure how your machine is set up, but it does look like you haven't set up an email address and name for git. Have you tried running those two suggested commands (even with some example email and name)?

That being said, I think we can fix this another way. Could you try adding --no-commit to the cherry-pick here please and let me know if it works for you?

https://github.com/ARM-software/Tool-Solutions/blob/4d593cb2559f19da9b92929b1f5f7ea5e1f9b8f2/docker/pytorch-aarch64/get-source.sh#L59

@jondea thanks for the suggestion, build was successful by adding --no-commit to the cherry-pick. But I found a new observation. I had earlier use Tool-solutions repo to build docker image along with pytorch build: https://github.com/ARM-software/Tool-Solutions/commit/ac72dca47fcb8ba2d5861214984e343156793d63 , and now I am using the latest commit method (in which the methodology is somewhat different from first): https://github.com/ARM-software/Tool-Solutions/commit/2dd009fa95e74e341fc90320bdb3b3c465f81d61.
In the second method I am finding scaling issues (from 32T to 64T, timing of hugging face models is increasing). In the first method this issue was not there (from 32T to 64T, timing of hugging face models decreased).

jondea commented 4 days ago

Glad to hear, I'll put that change up for review so that you shouldn't see the error again.

That's interesting, we actually had a few improvements to the threading so this is unexpected. Could you share the exact hardware and scripts you ran so that we can look into it please? Could I also ask you to make a new issue for it so that we can track the regression better please?

vishwascm commented 3 days ago

@jondea, thanks for reply. I will create a new issue for scaling Issue.