vllm-project/vllm: technical due diligence, read from its repository

The median source file here is 206 lines. 50 files exceed 2,060 lines, the largest being 8,943. Size is not a defect on its own — some problems genuinely live in one place — but these are where merge conflicts, review fatigue and single-owner knowledge concentrate, and they are the first thing that slows an inheriting engineer down.

34 tests are individually marked to retry. A retry turns a flaky test green without saying why it failed, so a green build means the tests passed eventually. Ask how many retries a typical run needs, and which tests need them.

The newest release, v0.32.0, is not on the branch this survey read; the newest release on it is v0.11.1. The shipped product may differ from what this report describes. Survey the release branch too.

Dependency versions are declared as ranges and no lockfile is committed, so a build today does not necessarily produce what a build last month produced. That matters twice over in a transaction: the thing being bought cannot be reproduced exactly, and a compromised upstream release arrives without anyone choosing to take it.

1 third-party CI action is referenced by a tag or branch rather than a commit, so they run whatever that tag points to on the day; 2 steps pipe a download straight into a shell. Tags have been moved to malicious commits in real supply-chain attacks, and a piped script is whatever the server returns that minute. Pinning to a commit is the fix, and it is small.

10 HTTP calls use a client with no timeout (Python `requests` without `timeout=`, or Go's default client). When a dependency slows down, each one holds a worker until the connection gives up, and enough of them take the service down with the dependency.

2 places run a command through a shell (`shell=True`, `os.system`, `exec` with a template) rather than passing arguments. Where any part comes from outside, it is command injection. Ask where each one's input comes from.

19 places read data with a deserialiser that executes what it reads (pickle). That is safe on data the system wrote itself and remote code execution on anything else. Ask where each one's input comes from.

19 of 631 locked dependencies match 22 published advisories in OSV. A lockfile says which version is installed, not whether the flawed code is reachable from this product, so each needs a look. Most are fixed by an upgrade, and the advisory names the version. This lockfile pins development tools alongside what ships and does not say which is which, so some of these may never reach a user.

The newest was published on 2026-10-21. A project that publishes advisories is telling its users about its fixes, and the count and pace say how much security work it carries. Whether 52 of them are fixed in the code you ship could not be told from this history: its version is only known to be at least 0.11.1. Answer lives in: the release you run, set against each fixed version. GHSA-q97c-p368-75c4 (moderate), GHSA-6757-hggm-hp2w (high), GHSA-vfp2-c8pq-v6h6 (high) and 50 more are fixed in the surveyed commit; which release carries them cannot be told from this history, which does not place the release users run. Answer lives in: the version you run. The record is incomplete: NVD: 8 records name vllm with nothing tying them to this repository (CVE-2024-11041, CVE-2024-8939, CVE-2024-9053, …).

Nothing in the service initialises an error reporter (Sentry, Rollbar, Bugsnag, Honeybadger or an APM agent). An exception in production reaches a person only if someone is reading the logs, which usually means a customer reports it first.