Module read 

Format-agnostic checkpoint loading.

Supported inputs (detected by extension, with a magic-byte fallback):

• .safetensors — single file or several shard files
• model.safetensors.index.json — HF sharded-model index (shards resolved relative to the index file)
• .pt / .pth / .bin / .ckpt — PyTorch pickle checkpoints (zip-based torch.save format) via candle’s pickle reader
• .npy / .npz — NumPy arrays
• .gguf — existing GGUF files (tensors dequantized to F32; metadata carried through), which makes dtype re-quantization possible

Every tensor is returned as raw little-endian bytes plus a SrcDtype (F32/F16/BF16 preserved exactly; other dtypes — F64, integers — are cast to F32 with a warning).

Structs

Checkpoint

A loaded checkpoint: all tensors plus any metadata carried over from the source (only GGUF inputs have metadata).

LoadOptions

RawTensor

One tensor read from a checkpoint: name, row-major (python-order) shape, source dtype, and raw little-endian bytes.

Functions

load_checkpoint

Load one or more checkpoint files into a single Checkpoint. Multiple inputs (e.g. safetensors shards) are merged; duplicate tensor names across inputs are an error.