Reserve more memory for DeepSeekOCR model and adjust server start timeout for DeepGEMM to reduce flakiness#15277
Conversation
|
/tag-and-rerun-ci |
Summary of ChangesHello @Kangyan-Zhou, I'm Gemini Code Assist1! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed! This pull request addresses test flakiness in the DeepSeekOCR model by adjusting its memory allocation and CUDA graph settings within the test suite. By reserving more static memory and optimizing batching, the changes aim to enhance the stability and reliability of the DeepSeekOCR tests. Highlights
Using Gemini Code AssistThe full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips. Invoking Gemini You can request assistance from Gemini at any point by creating a comment using either
Customization To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a Limitations & Feedback Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here. You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension. Footnotes
|
There was a problem hiding this comment.
Code Review
This pull request aims to reduce flakiness in the TestDeepseekOCRServer test by reserving more memory for the model. This is done by adding --mem-fraction-static=0.70 and --cuda-graph-max-bs=4 as extra arguments for the server launch. However, my review found a critical issue in the base test class TestOpenAIMLLMServerBase that prevents these new arguments from being used, making the change ineffective. I've provided a detailed explanation and a suggested fix for the base class logic.
| extra_args = [ | ||
| "--mem-fraction-static=0.70", | ||
| "--cuda-graph-max-bs=4", | ||
| ] |
There was a problem hiding this comment.
These extra_args will not be passed to the server due to an issue in the setUpClass method of the base class TestOpenAIMLLMServerBase.
In test/srt/test_vision_openai_server_common.py, the other_args are set as follows:
other_args=(
cls.extra_args + cls.fixed_args + ["--trust-remote-code"]
if cls.trust_remote_code
else []
),Since TestDeepseekOCRServer sets trust_remote_code = False, other_args becomes an empty list, and your new arguments are ignored.
To fix this, you should modify test/srt/test_vision_openai_server_common.py to correctly handle the arguments. For example, you could replace the other_args logic in TestOpenAIMLLMServerBase.setUpClass with the following:
other_args = list(cls.extra_args)
if cls.trust_remote_code:
other_args.extend(cls.fixed_args)
else:
other_args.extend(arg for arg in cls.fixed_args if arg != "--trust-remote-code")
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
api_key=cls.api_key,
other_args=other_args,
)Also, note that ["--trust-remote-code"] is appended redundantly when trust_remote_code is true, as it's already in fixed_args. The suggested fix above also resolves this.
Since this change is ineffective without the fix in the base class, I'm marking this as critical.
|
/rerun-stage list |
|
❌ Stage NVIDIA stages:
AMD stages:
Other stages will be added soon. For now, use |
|
/rerun-stage unit-test-backend-1-gpu |
|
/tag-and-rerun-ci |
…n3_pp * 'main' of https://github.com/sgl-project/sglang: (74 commits) [bug fix][pp] fix inconsistent latency between tp (sgl-project#15379) Fix warp illegal instruction in kimi k2 thinking PCG (sgl-project#15306) Fix gpt-oss yarn with `truncate` argument (sgl-project#14270) Monkey patch deepseek-ocr's `v_head_dim` (sgl-project#15384) [model-gateway] Replace PolicyRegistry RwLock with DashMap for lock-free policy lookups (sgl-project#15361) [PP] Fix dynamic chunking strategy for PP (sgl-project#15372) Fix issue: ENABLE_BELOW_SM90 cannot be enabled on aarch64 CPU (sgl-project#12967) Split test_piecewise_cuda_graph.py to optimize CI resource usage (sgl-project#15290) unified management of environment variables for vlm cuda ipc transport (sgl-project#14501) Mistral Large 3 NVFP4 TRTLLM MoE support (sgl-project#15049) fix: adjust time for test_epd_disaggregation.py (sgl-project#15354) Add doc for qwen3 next (sgl-project#15337) feat: DeepSeek-V3.2 Streaming tool call output (sgl-project#15278) Feature/trtllm mha workspace size configurable sgl-project#15089 (sgl-project#15131) [VLM] Support cos sin cache for Qwen3-VL & GLM-4.1V (sgl-project#15205) [Deepseek V3.2] Support Overlap Spec + NSA (sgl-project#15307) Add request-level timestamp for when prefill finishes (sgl-project#14860) [CI] Migrate LoRA tests to test/registered/lora/ (sgl-project#15176) Reserve more memory for DeepSeekOCR model and adjust server start timeout for DeepGEMM to reduce flakiness (sgl-project#15277) Fix condition check for require_gathered_buffer (sgl-project#15328) ...
…eout for DeepGEMM to reduce flakiness (sgl-project#15277)
…eout for DeepGEMM to reduce flakiness (sgl-project#15277)
…eout for DeepGEMM to reduce flakiness (sgl-project#15277)
Motivation
Modifications
Accuracy Tests
Benchmarking and Profiling
Checklist