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.

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

GQA into Mitchich65 #443

Merged
merged 92 commits into from
Mar 18, 2024
Merged

GQA into Mitchich65 #443

merged 92 commits into from
Mar 18, 2024

Conversation

dirkgr
Copy link
Member

@dirkgr dirkgr commented Feb 10, 2024

TODO

  • Doesn't run on AMD like this

dirkgr and others added 30 commits February 7, 2024 18:11
@dirkgr dirkgr marked this pull request as ready for review March 18, 2024 18:41
Copy link
Member Author

@dirkgr dirkgr left a comment

Choose a reason for hiding this comment

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

I cannot approve this because I created the OR. I am OK with this, with two concerns in comments.

@@ -1047,6 +1049,11 @@ class TrainConfig(BaseConfig):
The activation checkpointing strategy to use.
"""

fused_loss: Optional[bool] = None
"""
Whether to use the fused CE loss function from `flash-attn`.
Copy link
Member Author

Choose a reason for hiding this comment

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

What's the default?

Copy link
Member

Choose a reason for hiding this comment

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

Defaults to False.

Comment on lines 59 to 60
out[f"eval/ppl/{label}/CrossEntropyLoss"] = loss.item()
out[f"eval/ppl/{label}/Perplexity"] = torch.exp(loss).item()
Copy link
Member Author

Choose a reason for hiding this comment

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

This will make comparisons impossible?

Copy link
Member

Choose a reason for hiding this comment

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

Hmm yea could be difficult. It just annoys me these aren't grouped together in W&B. Reverted: 856860d

try:
mp.set_start_method("spawn", force=True)
except RuntimeError as e:
print(f"failed to set multiprocessing start method: {e}")
Copy link
Member Author

Choose a reason for hiding this comment

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

I did not know this could fail?

Copy link
Member

Choose a reason for hiding this comment

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

Yea I saw this happen a couple times, not sure why exactly.

@epwalsh epwalsh merged commit fa8ec33 into mitchish65-2 Mar 18, 2024
1 check passed
@epwalsh epwalsh deleted the mitchish65-2-gqa branch March 18, 2024 19:04
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

2 participants