OptimalScale / OptimalScale/LMFlow

Some question on code of LISA

Open
#897 1 comment 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
Python
Stars
8.5k
Forks
822
PR merge metrics
No merged PRs in 30d

Description

I have known that LISA's core code in src\lmflow\pipeline\finetuner.py, mainly in class DynamicLayerActivationCallback. I read it with Algorithm 1 Layerwise Importance Sampling AdamW (LISA) in paper aside.

So where is step2: Freeze all layers except the embedding and language modeling head layer? I can only find def freeze_all_layers(self) in class DynamicLayerActivationCallback, not excluding embedding and head layer

And i'm curious on the notation k in paper Algorithm 1:
step 4: Run AdamW for K iterations with ${η_t}_{t=ik}^{ik=k-1}$ Is the k same as K ?

My english is bad so tell me if any understanding problem,thanks for answering

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Read src\lmflow\pipeline\finetuner.py, focusing on DynamicLayerActivationCallback and freeze_all_layers, then compare its behavior with Algorithm 1 in the LISA paper. Trace how layer freezing and optimizer iterations are represented, and document whether the implementation matches the paper or needs clarification; there is no test named in the issue.

Written by the indexing model from the issue text.

Assessment

Tech stack
python, pytorch
Domain
machine-learning
Issue type
Documentation
Difficulty
4/5
Estimated time
3-5 days
Activity status
Stale
Clarity
Needs clarification
Newbie friendliness
15/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.