Skip to content

[CI] switch DeepSeek-V2-Lite E2E from h100_4 to l4_4 - #250

Merged
jiangkuaixue123 merged 2 commits into
vllm-project:mainfrom
yenuo26:CI
Aug 26, 2026
Merged

[CI] switch DeepSeek-V2-Lite E2E from h100_4 to l4_4#250
jiangkuaixue123 merged 2 commits into
vllm-project:mainfrom
yenuo26:CI

Conversation

@yenuo26

@yenuo26 yenuo26 commented Aug 14, 2026

Copy link
Copy Markdown
Collaborator

Summary

  • Add the l4_4 Buildkite hardware preset (gpu_4_queue) in ci_mirror_hardwares.yml.
  • Point L2 ready and L3 merge-test DeepSeek-V2-Lite E2E steps at l4_4 instead of h100_4.
  • Keep h100_4 available as an unused/optional preset for later use.

Test plan

  • Confirm Buildkite expands mirror_hardwares: l4_4 to gpu_4_queue + CI image
  • Label PR ready → unit (l4_1) + DeepSeek-V2-Lite E2E (l4_4)
  • Label PR merge-test → same L3 path
  • Confirm E2E scenarios complete on 4×L4 without OOM (including baseline-graph)

@yenuo26 yenuo26 added the ready Used to trigger ready CI in PRs. label Aug 14, 2026
@jiangkuaixue123

Copy link
Copy Markdown
Collaborator

It seems CI failed.

@jiangkuaixue123

Copy link
Copy Markdown
Collaborator

@yujuancao07 cc

@yenuo26

yenuo26 commented Aug 15, 2026

Copy link
Copy Markdown
Collaborator Author

@yujuancao07 cc

I seems still OOM

[2026-08-14T10:15:42Z] [ffn] (EngineCore pid=5468) torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 352.00 MiB. GPU 0 has a total capacity of 22.05 GiB of which 346.12 MiB is free. Including non-PyTorch memory, this process has 21.70 GiB memory in use. Of the allocated memory 21.47 GiB is allocated by PyTorch, and 864.00 KiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation.  See documentation for Memory Management  (https://docs.pytorch.org/docs/stable/notes/cuda.html#optimizing-memory-usage-with-pytorch-cuda-alloc-conf)

@jiangkuaixue123 jiangkuaixue123 added ready Used to trigger ready CI in PRs. and removed ready Used to trigger ready CI in PRs. labels Aug 17, 2026
@yenuo26 yenuo26 added ready Used to trigger ready CI in PRs. and removed ready Used to trigger ready CI in PRs. labels Aug 17, 2026
…nces

- Introduced a new hardware configuration for `l4_4` in `ci_mirror_hardwares.yml` with appropriate agent and plugin settings.
- Updated `test-merge.yml` and `test-ready.yml` to reference the new `l4_4` hardware instead of `h100_4`.

Signed-off-by: [Your Name] <[Your Email]>
Signed-off-by: wangyu <410167048@qq.com>
@yujuancao07
yujuancao07 force-pushed the CI branch 2 times, most recently from ba76635 to dbf7e0b Compare August 25, 2026 02:03
@jiangkuaixue123 jiangkuaixue123 added ready Used to trigger ready CI in PRs. and removed ready Used to trigger ready CI in PRs. labels Aug 25, 2026
Rename the 2A1F scenarios with an explicit -2a1f suffix and make the CI
gate run the 2A2F variants (baseline-graph, afd-eager-2a2f,
afd-graph-2a2f, afd-graph-dbo-2a2f) so DeepSeek-V2-Lite fits in the
4x L4 24GB budget. Update the runner, docs, skill, and unit tests to
match, and align the Qwen3 MoE local suite scenario names.

Signed-off-by: yujuancao07 <yujuancao07@gmail.com>
@jiangkuaixue123 jiangkuaixue123 added ready Used to trigger ready CI in PRs. and removed ready Used to trigger ready CI in PRs. labels Aug 25, 2026
@jiangkuaixue123
jiangkuaixue123 merged commit af01a1a into vllm-project:main Aug 26, 2026
3 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

ready Used to trigger ready CI in PRs.

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants