Skip to content

Conversation

@22quinn
Copy link
Collaborator

@22quinn 22quinn commented May 17, 2025

In classification problems, two labels are often reduced to one label because they are equivalent, both representing binary classification. For a single label, we should apply sigmoid instead of softmax

Related to #18052

cc @maxdebayser @DarkLight1337 @WoosukKwon @houseroad

Signed-off-by: 22quinn <33176974+22quinn@users.noreply.github.com>
@github-actions
Copy link

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

@DarkLight1337
Copy link
Member

Could you explain why this is needed? I thought softmax is just a generalization of sigmoid so they should be equivalent for the binary case.

@22quinn
Copy link
Collaborator Author

22quinn commented May 18, 2025

Could you explain why this is needed? I thought softmax is just a generalization of sigmoid so they should be equivalent for the binary case.

@DarkLight1337 When the output has only a single class, applying softmax will always give you 1: e^x / e^x = 1. This can be verified by running the code below without this PR.
Softmax only works for >=2 classes. A binary classification problem can be represented with either two-class or one-class. I think one class is more common as two classes seem not necessary and involve more computation.

from vllm import LLM
MODEL = "22quinn/Llama-3.2-1B-1Label-dummy"
PROMPTS = ["Hello my name is Robert", "ok I got it"]
model = LLM(MODEL, task="classify", enforce_eager=True, enable_prefix_caching=True)
outputs = model.classify(PROMPTS)
for output in outputs:
    print(output.outputs.probs)

Copy link
Member

@DarkLight1337 DarkLight1337 left a comment

Choose a reason for hiding this comment

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

Oh, I get it now. Thanks for the explanation!

@DarkLight1337 DarkLight1337 enabled auto-merge (squash) May 18, 2025 05:09
@github-actions github-actions bot added the ready ONLY add when PR is ready to merge/full CI is needed label May 18, 2025
@vllm-bot vllm-bot merged commit 908733a into vllm-project:main May 18, 2025
74 of 76 checks passed
zzzyq pushed a commit to zzzyq/vllm that referenced this pull request May 24, 2025
Signed-off-by: 22quinn <33176974+22quinn@users.noreply.github.com>
Signed-off-by: Yuqi Zhang <yuqizhang@google.com>
@22quinn 22quinn deleted the classify-1label branch June 20, 2025 04:00
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

ready ONLY add when PR is ready to merge/full CI is needed

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants