feat: update knowledge distillation tutorial for using vllm with Qwen model #2960
+320
−91
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
RexBearIU
force-pushed
the
jackyf/docs/distillation
branch
from
14091ae to
2fb1059
Compare
RexBearIU
requested review from
A9isha,
RissyRan,
bvandermoon,
gagika,
gobbleturk,
jacoguzo,
jiangjy1982,
richjames0,
shralex and
vipannalla
as code owners
| export USERNAME_OR_ORG = <Owner of Hugging Face repository> | ||
| export RUN_NAME = <unique name for the run> | ||
| export HF_TOKEN=<your-hf-token> # e.g., hf_BA6... | ||
| export BASE_DIRECTORY=<your-base-directory> # e.g., knowledge-distillation |
Collaborator
Author
There was a problem hiding this comment.
Thanks. We use a mounted Hyperdisk because writing large model files and many small I/O ops directly to gs:// is often much slower. The tutorial writes to /mnt/hyperdisk for performance and reproducibility, and I fixed the duplicated env export in the doc.
| uv pip install -r dependencies/requirements/requirements.txt | ||
| To install MaxText and its dependencies for post-training (including vLLM for the teacher), run the following: | ||
|
|
||
| 1. Follow the [MaxText installation instructions](https://maxtext.readthedocs.io/en/latest/install_maxtext.html#install-maxtext): |
|
|
||
| We will use vLLM to generate the dataset from the teacher model. | ||
|
|
||
| Create a python script named `generate_distillation_data_vllm.py` with the following content (this script writes a Parquet dataset compatible with MaxText SFT): |
Collaborator
There was a problem hiding this comment.
Can we make this script available in MaxText like this: https://github.com/AI-Hypercomputer/maxtext/blob/main/tools/data_generation/generate_distillation_data.py
Collaborator
Author
There was a problem hiding this comment.
Agreed. I’ve already created the script.
RexBearIU
force-pushed
the
jackyf/docs/distillation
branch
from
2fb1059 to
8005986
Compare
RexBearIU
requested review from
NicoGrande,
NuojCheng,
aireenmei,
hengtaoguo,
jesselu-google,
khatwanimohit and
suexu1025
as code owners
Codecov Report✅ All modified and coverable lines are covered by tests. 📢 Thoughts on this report? Let us know! |
RexBearIU
force-pushed
the
jackyf/docs/distillation
branch
from
8005986 to
84aa2ed
Compare
RexBearIU
force-pushed
the
jackyf/docs/distillation
branch
from
84aa2ed to
eb215d2
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Description
This pull request significantly updates and modernizes the knowledge distillation tutorial for MaxText, aligning it with current best practices and tooling. The guide now uses Qwen3-32B as the teacher model (via vLLM) and Llama-3.1-8B as the student, streamlines the setup with Hyperdisk storage, and provides new scripts and commands for dataset generation and fine-tuning. The instructions have been clarified, unnecessary conversion steps removed for the teacher, and the fine-tuning process updated for the latest MaxText and vLLM workflows.
Tests
Manually triggered the distillation pipeline and monitored the execution flow step-by-step. Confirmed that the training loop finished and resources were released.
Checklist
Before submitting this PR, please make sure (put X in square brackets):
gemini-reviewlabel.