Skip to content

Conversation

sarckk
Copy link
Collaborator

@sarckk sarckk commented Aug 29, 2025

Purpose

Multi-modal processor tests for H2OVLChatModel and KimiVLForConditionalGeneration are broken on trunk after merging #22628.

I was previously unable to reproduce this issue locally as it requires that these test cases come after the Gemma3n test case. The failures are:

H20VLChatModel

TypeError: _LazyConfigMapping.__init__() missing 1 required positional argument: 'mapping'

Kimi (OOM)

(EngineCore_0 pid=1298081)   File "/data/users/yhshin/gitrepos/vllm/vllm/model_executor/models/moonvit.py", line 141, in sdpa_attention
(EngineCore_0 pid=1298081)     attn_output = F.scaled_dot_product_attention(q,
(EngineCore_0 pid=1298081)                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(EngineCore_0 pid=1298081) torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 254.47 GiB. GPU 0 has a total capacity of 94.99 GiB of which 78.18 GiB is free. Including non-PyTorch memory, this process has 16.80 GiB memory in use. Of the allocated memory 16.03 GiB is allocated by PyTorch, and 104.27 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid ...

This PR reverts gemma3n changes in #22628 (commit c498483) which fixes the tests.

Test Plan

pytest tests/models/multimodal/processing/test_tensor_schema.py::test_model_tensor_schema -s 

Test Result

Multi-Modal Processor Test should pass in CI


Essential Elements of an Effective PR Description Checklist
  • The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
  • The test plan, such as providing test command.
  • The test results, such as pasting the results comparison before and after, or e2e results
  • (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
  • (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.

Copy link
Contributor

@gemini-code-assist gemini-code-assist bot left a comment

Choose a reason for hiding this comment

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

Code Review

This PR reverts the gemma3n fast prefill feature to fix failing tests. The changes within the model files are correct. However, the revert is incomplete, leaving behind dead code, configuration options, and tests related to the removed feature. I've added a critical comment pointing out the need for a complete revert to maintain code health.

@heheda12345 heheda12345 enabled auto-merge (squash) August 29, 2025 04:49
@heheda12345 heheda12345 disabled auto-merge August 29, 2025 04:49
@heheda12345
Copy link
Collaborator

Need to ensure Multi-modal processor tests pass before merging this PR.

@heheda12345 heheda12345 added the ready ONLY add when PR is ready to merge/full CI is needed label Aug 29, 2025
@heheda12345
Copy link
Collaborator

@sarckk v1-test-e2e-plus-engine fails and is related to this PR.

@mergify mergify bot added the v1 label Aug 29, 2025
@heheda12345 heheda12345 enabled auto-merge (squash) August 29, 2025 16:39
@zou3519 zou3519 mentioned this pull request Aug 29, 2025
10 tasks
@simon-mo simon-mo disabled auto-merge August 29, 2025 19:16
@simon-mo simon-mo merged commit 8c3e199 into vllm-project:main Aug 29, 2025
40 of 42 checks passed
nopperl pushed a commit to pfnet/vllm that referenced this pull request Sep 3, 2025
MatthewBonanni pushed a commit to MatthewBonanni/vllm that referenced this pull request Sep 3, 2025
Signed-off-by: Yong Hoon Shin <[email protected]>
Signed-off-by: Matthew Bonanni <[email protected]>
MatthewBonanni pushed a commit to MatthewBonanni/vllm that referenced this pull request Sep 3, 2025
842974287 pushed a commit to 842974287/vllm that referenced this pull request Sep 3, 2025
Signed-off-by: Yong Hoon Shin <[email protected]>
Signed-off-by: Shiyan Deng <[email protected]>
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 v1
Projects
None yet
Development

Successfully merging this pull request may close these issues.

4 participants