How a Fake vLLM PR Exposed the Risks of AI‑Generated Resume Padding

vLLM recently posted a tweet that was later used by a training institute to showcase a fabricated pull request (PR) on the project, and the institute even provided an English translation of the PR on a Chinese social platform.

PR Appearance and Claimed Fix

The PR title reads fix(eagle3): read norm_before_fc from eagle_config for NVIDIA checkpoint . It claims that NVIDIA’s Eagle3 checkpoint hides the norm_before_fc field inside a nested eagle_config dictionary, which vLLM only reads at the top level, causing RMSNorm to be silently skipped and lowering acceptance rates.

Test Plan – The “Soul” of the PR

The author attached three Python scripts as a test plan:

Correctness test : compare greedy decoding output of the baseline and the Eagle3 branch.

Performance test : run a single request of 512 tokens and compare tokens‑per‑second (tok/s) between baseline and Eagle3.

Unit test : construct a mock config and verify that both code paths correctly read norm_before_fc.

The PR concludes with the note “Tested with Qwen3-32B + Qwen3-32B_eagle3 on 1x H200”.

Actual Outcome

All four correctness checks passed, and a performance figure (baseline xx tok/s → Eagle3 xx tok/s) was reported. The PR passed all 63 CI checks, received the ready label, and was merged into main.

Why the PR Was Fake

After merging, community members reported that the alleged bug does not exist: the NVIDIA Eagle3 checkpoint already reads norm_before_fc correctly, and the “RMSNorm being skipped” scenario was invented to make the PR look technical. The underlying motive was the “PR training” workflow advertised by the institute for resume building.

Mechanics of the Resume‑Padding Scheme

The author, who identifies as a ZTE engineer working on a domestic NPU inference engine, explained the high barrier to contributing genuine fixes to projects like vLLM, PyTorch, or SGLang (requiring deep knowledge of hundreds of thousands of lines of C++/CUDA/Python, reproducing real bugs, and multi‑GPU validation). To bypass this, the institute teaches participants to:

Learn vLLM’s code patterns.

Pick a harmless logic branch and masquerade it as a bug.

Use AI coding agents to generate a convincing Purpose and Test Plan.

Claim the PR was run on high‑end hardware such as H200 or 8×A100 (which reviewers cannot easily verify).

Merge the PR, screenshot the merge, and list it on their résumé as a “vLLM core contribution”.

This low‑cost, low‑risk, high‑reward approach succeeded until community members reported the fabricated nature of the contribution, leading vLLM to ban the author and publicly expose the practice.

Broader Implications

The incident illustrates how AI coding agents can reduce the cost of generating large volumes of low‑quality PRs, forcing maintainers to spend real time reviewing them. vLLM now receives over a thousand PRs per month, growing exponentially, which strains the core team. Recruiters may start verifying open‑source contributions, eroding trust in Chinese‑language contributors, and genuine contributors risk being suspected of “resume padding”.

Final Thoughts

The author acknowledges that AI tools like Claude Code, Codex, and Cursor dramatically boost personal productivity, but warns that AI also amplifies one’s shortcomings. Using AI as a genuine problem‑solver can be a productivity revolution, whereas using it as a PR generator to deceive employers burns the trust assets of the open‑source ecosystem.