Expand description
FTW multi-shard loader is file/memmap2-backed — native hosts only.
FTW (FreeToken Weight) format loader and multi-shard memory mapper (Work Package F9/F10).
Loads freetoken_weight.json and memory-maps .ftw binary shards, providing
zero-copy access to shared weights and MoE expert banks (256 experts per layer).
Structs§
- FtwExpert
Bank Slice - Slices within the memory-mapped shards for a single expert in a given layer.
- FtwExpert
Data - Zero-copy byte slices for one expert’s NVFP4 tensors.
- FtwManifest
- FtwModel
Package - Container owning memory-mapped FTW shards and tensor index.
- FtwShard
Entry - FtwTensor
Entry