mlcommons / mlcommons/inference
[405b-SUT] Max number of output tokens
Open
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 1.6k
- Forks
- 650
- Avg merge
- 1d 22h
- Merged PRs (30d)
- 6
Description
For 405B the sampling parameter config sets the max output tokens to be 20k.
However, given the reference output distribution with max output length of 1.7k, I don't think we should set this parameter in the sampler that high.
@nvzhihanj @arjunsuresh @mrmhodak
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start in language/llama3.1-405b/SUT_VLLM.py at the sampling parameter configuration around line 75. Compare the 20k max output-token setting with the reference output distribution, whose maximum output length is about 1.7k. Done means the sampler no longer permits an unnecessarily high maximum for the 405B workload.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- machine-learning
- Issue type
- Bug
- Difficulty
- 2/5
- Estimated time
- 1-3 hours
- Activity status
- Stale
- Clarity
- Mostly clear
- Newbie friendliness
- 42/100