[Bugfix] Fix _has_module to verify native deps via trial import by esmeetu · Pull Request #39873 · vllm-project/vllm

esmeetu · 2026-04-15T06:35:20Z

Summary

_has_module previously used importlib.util.find_spec() which only checks whether a module spec exists on disk, but does not verify that the module can actually be imported
For native-extension packages whose shared libraries are missing (e.g. nixl_ep installed without libcudart.so.12), find_spec returns a valid spec but the real import crashes with ImportError
This caused has_nixl_ep() (and potentially other has_* guards) to return True when the module was not actually usable, leading to import-time crashes like:
```
ImportError: libcudart.so.12: cannot open shared object file: No such file or directory
```
The fix adds a trial importlib.import_module() after the find_spec pre-check, catching ImportError/OSError and returning False when native dependencies are unsatisfied. The @cache decorator ensures this only runs once per module name.

Not a duplicate

No existing open PRs address this issue. Searched for _has_module, find_spec, and has_module ImportError.

Test plan

AI disclosure

This PR was authored with AI assistance (Claude). The human submitter has reviewed all changed lines.

🤖 Generated with Claude Code

`_has_module` used `importlib.util.find_spec` which only checks whether a module spec exists on disk. For native-extension packages whose shared libraries are missing (e.g. `nixl_ep` without `libcudart.so.12`), `find_spec` succeeds but the real import crashes. Add a trial `importlib.import_module` after the `find_spec` pre-check so that `has_nixl_ep()` (and all other `has_*` guards) correctly return `False` when a package is installed but its native dependencies are unsatisfied. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> Signed-off-by: esmeetu <jasonailu87@gmail.com>

gemini-code-assist

Code Review

This pull request updates the _has_module utility in vllm/utils/import_utils.py to perform a trial import of the specified module instead of just checking for its existence via find_spec. This change ensures that native dependencies and shared libraries are correctly satisfied before reporting the module as available. I have no feedback to provide.

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 9a77e42a67

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector · 2026-04-15T06:38:14Z

+    try:
+        importlib.import_module(module_name)
+        return True
+    except (ImportError, OSError) as exc:


Catch non-ImportError failures in availability probe

The new trial import in _has_module only handles ImportError/OSError, so optional modules that fail import with other exception types now crash capability checks instead of returning False. This is a real path in-tree: vllm/utils/deep_gemm.py explicitly documents/catches RuntimeError from vendored deep_gemm import (_C.init() failure), but has_deep_gemm() now routes through _has_module first and would raise before that fallback logic runs. This turns a best-effort feature probe into a startup/runtime hard failure for broken optional installs.

Useful? React with 👍 / 👎.

Tests cover: - Returns True for importable stdlib modules - Returns False for nonexistent modules - Returns False when find_spec succeeds but import raises ImportError (e.g. missing libcudart.so) - Returns False when import raises OSError - Result is cached across repeated calls Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> Signed-off-by: esmeetu <jasonailu87@gmail.com>

jhsmith409 · 2026-04-15T15:54:17Z

Real-world repro that this PR would fix: #39923.

Summary: vllm/vllm-openai:cu130-nightly (digest sha256:9f815b8c…, vLLM 0.19.1rc1.dev296+gbcc2306ce) ships a nixl_ep/ package whose compiled .so needs libcudart.so.12 — not present anywhere in the cu13 image. has_nixl_ep() currently returns True via find_spec, so all2all_utils.py then eagerly imports the broken module and the server crash-loops at startup:

File ".../nixl_ep/__init__.py", line 23, in <module>
    from . import nixl_ep_cpp as _nixl_ep_cpp
ImportError: libcudart.so.12: cannot open shared object file: No such file or directory

With this PR's trial-import change, _has_module("nixl_ep") would catch the ImportError, has_nixl_ep() would return False, and the server would start normally — the nixl_ep code path isn't needed for single-GPU inference anyway.

Happy to test a nightly once this merges and report back. Thanks for the fix.

benchislett · 2026-04-15T18:56:57Z

Duplicate #39851

esmeetu · 2026-04-18T00:57:37Z

@benchislett yeah, it does solve that issue. But i think this PR introduce a robust way to avoid similar events happen. :)

jeffreywang88 · 2026-05-29T06:45:52Z

Hey @esmeetu, this PR would've fixed nixl_ep import issues on cuda13 images (thread) that #39851 didn't fully resolve. I'll open a PR with the fix you have here. Let me know if you want to re-open this PR instead.

mergify Bot added the bug Something isn't working label Apr 15, 2026

gemini-code-assist Bot reviewed Apr 15, 2026

View reviewed changes

chatgpt-codex-connector Bot reviewed Apr 15, 2026

View reviewed changes

jhsmith409 mentioned this pull request Apr 15, 2026

[Bug]: vllm/vllm-openai:cu130-nightly ships broken nixl_ep linked against libcudart.so.12 — server fails at import #39923

Closed

benchislett closed this Apr 15, 2026

esmeetu deleted the fix/has-module-trial-import branch May 25, 2026 05:47

This was referenced May 29, 2026

[llm] Upgrade vLLM to 0.21.0 ray-project/ray#63396

Merged

[BugFix] Fix _has_module to verify native deps via trial import #44035

Merged

Jwrede mentioned this pull request Jun 9, 2026

[serve.llm] Add TTFT, TPOT, and generation time metrics ray-project/ray#63856

Open

8 tasks

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[Bugfix] Fix _has_module to verify native deps via trial import#39873

[Bugfix] Fix _has_module to verify native deps via trial import#39873
esmeetu wants to merge 2 commits into
mainfrom
fix/has-module-trial-import

esmeetu commented Apr 15, 2026 •

edited

Loading

Uh oh!

gemini-code-assist Bot left a comment

Uh oh!

chatgpt-codex-connector Bot left a comment

Uh oh!

chatgpt-codex-connector Bot Apr 15, 2026

Uh oh!

jhsmith409 commented Apr 15, 2026

Uh oh!

benchislett commented Apr 15, 2026

Uh oh!

esmeetu commented Apr 18, 2026

Uh oh!

jeffreywang88 commented May 29, 2026 •

edited

Loading

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Uh oh!

Conversation

esmeetu commented Apr 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Not a duplicate

Test plan

AI disclosure

Uh oh!

gemini-code-assist Bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector Bot Apr 15, 2026

Choose a reason for hiding this comment

Uh oh!

jhsmith409 commented Apr 15, 2026

Uh oh!

benchislett commented Apr 15, 2026

Uh oh!

esmeetu commented Apr 18, 2026

Uh oh!

jeffreywang88 commented May 29, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

esmeetu commented Apr 15, 2026 •

edited

Loading

jeffreywang88 commented May 29, 2026 •

edited

Loading