[Feature]: add lineinfo or --generate-line-info for all cuda device compilations

[youkaichao]

🚀 The feature, motivation and pitch

Given the blogpost https://blog.vllm.ai/2025/08/11/cuda-debugging.html and the fact that we have 800 MiB wheel size limit now ( see pypi/support#6326 ), i think we should be able to add lineinfo flags in vLLM’s nvcc compilation to add file line number information, which helps debugging a lot.

Alternatives

No response

Additional context

No response

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

[youkaichao]

cc @mgoin if you agree

[ycycse]

Hi, I'm willing to work on this issue.

[wenchen76]

I’ve opened a draft PR for this issue. Hi @ycycse, would love to collaborate with you on this, thank you!

[mgoin]

I'm only okay with it if the binary size increase is ~10%. Just because we have more potential space in pypi doesn't mean we should balloon the binary size to fill it.

[ycycse]

I’ve opened a draft PR for this issue. Hi @ycycse, would love to collaborate with you on this, thank you!

That's very helpful!
I build the code with python setup.py bdist_wheel --dist-dir=dist got the following result (.whl):

• origin(824MB)
• after changed(915MB)

After add export CMAKE_BUILD_TYPE=Release, the results are (.whl):

• origin(249MB)
• after changed(339MB)

[ycycse]

I’ve opened a draft PR for this issue. Hi @ycycse, would love to collaborate with you on this, thank you!

By the way, I use cuobjdump --dump-ptx to analyze all the .so file in .whl. Try to find .loc col added by --addlineinfo, and I notice only part of files got the symbol:

• vllm/vllm_flash_attn/_vllm_fa3_C.abi3.so [NONE]
• vllm/vllm_flash_attn/_vllm_fa2_C.abi3.so [NONE]
• vllm/_moe_C.abi3.so [GOT LOC]
• vllm/cumem_allocator.abi3.so [NO DEVICE CODE]
• vllm/_C.abi3.so [GOT LOC]
• vllm/_flashmla_C.abi3.so [NONE]

Because VLLM_GPU_FLAGS is only used to compile _moe_C and _C. If we want add line info to other external project, we need to do more.

[youkaichao]

An update: since the binary size grows quite a lot with line information, I don't think we can enable it by default. Instead, I have updated the blog at vllm-project/vllm-project.github.io#91 to instruct people to compile with debug symbols.

thank you all for the contribution!