In a vLLM contributor's test, a 4,175,968-byte reproducer allocated 3,072 MiB before the shape check rejected it, and a one-row view passed with 3 GiB of storage attached. The fix merged into main on November 16th, 2026 checks archive size before torch.load.
By [RuntimeWire Staff](https://runtimewire.com/author/runtimewire-staff)
· Published
Contributor Leizhen Zhang tested vLLM, an open-source inference and serving engine, with a 4,175,968-byte base64 embedding payload that allocated 3,072 MiB in 'safe_load_prompt_embeds' before the shape check rejected it. A second version using a one-row view passed the 2 GiB check while keeping the full 3 GiB storage attached. The merged pull request moves the check to the front of the path: the fix on vLLM's main branch sums the declared uncompressed size inside the archive and rejects oversized payloads before 'torch.load' runs.…