Expand description
Typed safetensors GPTQ adapter for the Marlin physical ABI.
The adapter performs validation and repacking once while static plan resources are initialized. Providers receive only plan-owned device regions and never parse checkpoint metadata or repack on a request path.
Structs§
- Gptq
Marlin Safetensors Source - Mmap-backed safetensors archive with an explicit GPTQ-to-Marlin cold-path adapter. Dense components retain the archive’s zero-copy behavior.