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

[Bugfix]Disable the post_norm layer of the vision encoder for LLaVA models #9653

Merged
merged 2 commits into from
Oct 24, 2024

Conversation

litianjian
Copy link
Contributor

Disable the post_norm layer of vision encoder for LLaVA-Onevision models based on #9217.

Copy link

👋 Hi! Thank you for contributing to the vLLM project.
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 do one of these:

  • Add ready label to the PR
  • Enable auto-merge.

🚀

@litianjian litianjian changed the title [Bugfix]Disable post_norm for llava-ov models. [Bugfix]Disable the post_norm layer of the vision encoder for LLaVA-Onevision Oct 24, 2024
@DarkLight1337
Copy link
Member

Does this issue occur for the other LLaVA models as well?

@litianjian
Copy link
Contributor Author

Does this issue occur for the other LLaVA models as well?

Other LLaVA models can still use num_hidden_layers_override to determine whether enable post_layer_norm layer. However, I believe that consistently using this parameter for judgment would be better.

@DarkLight1337
Copy link
Member

Let's update the other models then, since from a quick look at the HF implementation, they don't use post_layernorm.

@litianjian
Copy link
Contributor Author

Let's update the other models then, since from a quick look at the HF implementation, they don't use post_layernorm.

OK, all LLaVA models don't use post_layernorm.

@litianjian litianjian changed the title [Bugfix]Disable the post_norm layer of the vision encoder for LLaVA-Onevision [Bugfix]Disable the post_norm layer of the vision encoder for LLaVA models Oct 24, 2024
@DarkLight1337 DarkLight1337 added the ready ONLY add when PR is ready to merge/full CI is needed label Oct 24, 2024
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.

Thanks for fixing!

@DarkLight1337 DarkLight1337 enabled auto-merge (squash) October 24, 2024 10:16
@simon-mo simon-mo merged commit f584549 into vllm-project:main Oct 24, 2024
68 of 71 checks passed
MErkinSag pushed a commit to MErkinSag/vllm that referenced this pull request Oct 26, 2024
…odels (vllm-project#9653)

Signed-off-by: Erkin Sagiroglu <erkin@infra-aipipeline-1-at1-prox-prod-a.ipa.corp.telnyx.com>
cooleel pushed a commit to cooleel/vllm that referenced this pull request Oct 28, 2024
…odels (vllm-project#9653)

Signed-off-by: Shanshan Wang <shanshan.wang@h2o.ai>
cooleel pushed a commit to cooleel/vllm that referenced this pull request Oct 28, 2024
…odels (vllm-project#9653)

Signed-off-by: Shanshan Wang <shanshan.wang@h2o.ai>
FerdinandZhong pushed a commit to FerdinandZhong/vllm that referenced this pull request Oct 29, 2024
…odels (vllm-project#9653)

Signed-off-by: qishuai <ferdinandzhong@gmail.com>
NickLucche pushed a commit to NickLucche/vllm that referenced this pull request Oct 31, 2024
NickLucche pushed a commit to NickLucche/vllm that referenced this pull request Oct 31, 2024
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