Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

chore!: Update PyTorch to 2.5 #465

Merged
merged 1 commit into from
Jan 9, 2025
+1 −1
Merged

chore!: Update PyTorch to 2.5 #465

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion requirements.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,6 +17,6 @@ openai>=1.13.3,<2.0.0
sentencepiece>=0.2.0
tabulate>=0.9.0
tenacity>=8.3.0,!=8.4.0
torch>=2.3.0,<2.5.0
torch>=2.3.0,<2.6.0
Copy link
Contributor

@courtneypacheco courtneypacheco Jan 8, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Shouldn't this be

Suggested change
torch>=2.3.0,<2.6.0
torch>=2.5.0,<2.6.0

...to enforce the use of 2.5? If other required packages in this file change (or their versions change), then pip may end up pulling 2.3.z or 2.4.z when trying to automatically resolve dependencies.

I'm referencing the title: Update PyTorch to 2.5

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Having an older version is fine if that's a requirement for one component. That's unlikely to happen though, unless a library is downgrading its dependencies. And that's also what we've done for the instructlab/instructlab repo: instructlab/instructlab#2865.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Gotcha. So we're allowing 2.5 if other packages support it, but not enforcing the use of 2.5.

I was confused by the PR title because it suggested we were actually updating the version.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks for hightling this @courtneypacheco ! Looking in the CI logs, even though this raises the cap it looks like CI is actually installing torch 2.4.x and running with that. So, we may have more work to do here to figure out what dependency is forcing our torch to 2.4.x instead of 2.5.x.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Ahh, I see in the logs:

2025-01-08T12:55:07.4571834Z Collecting torch<2.5.0,>=2.3.0 (from instructlab==0.21.0a2.dev258)
2025-01-08T12:55:07.4574409Z   Obtaining dependency information for torch<2.5.0,>=2.3.0 from

Our CI tests install instructlab from latest main, so we must have picked up instructlab before the torch bump was merged there. Kicking this off again to ensure it picks up torch 2.5.x now.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I suspect that it's because we're still on vLLM 0.6.2, which depends on PyTorch 2.4.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Ahh, yes. So in other words, neither instructlab/instructlab nor instructlab/sdg can actually bump our torch version in use until we move to a newer vllm than 0.6.2 that depends on torch 2.5.x. Regardless of this version range, our vllm dependency specifies an exact dependency on torch 2.4.0.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes, that seems to be the case. I have opened a new PR in instructlab to update vLLM to 0.6.6.post1: instructlab/instructlab#2892.

transformers>=4.41.2
xdg-base-dirs>=6.0.1
Loading