Serving
SafeTensors is a tensor checkpoint format with a metadata header and raw tensor data. It avoids arbitrary code execution during deserialization and supports memory-mapped access patterns.
A data-only format narrows the trust boundary when loading external checkpoints. Mapping and predictable offsets can simplify loading, but they do not guarantee zero-copy device placement.
Loaders validate offsets, data types, and shapes before exposing tensor views. Actual transfer behavior still depends on framework, storage, operating system, and destination device.
SafeTensors is a tensor checkpoint format with a metadata header and raw tensor data. It avoids arbitrary code execution during deserialization and supports memory-mapped access patterns. Loaders validate offsets, data types, and shapes before exposing tensor views. Actual transfer behavior still depends on framework, storage, operating system, and destination device.
A data-only format narrows the trust boundary when loading external checkpoints. Mapping and predictable offsets can simplify loading, but they do not guarantee zero-copy device placement.
RunInfraby RightNow© 2026 RunInfra. All rights reserved.
