API Reference

save_delta_from_paths

dt.save_delta_from_paths(
    out_path,
    finetuned_dir,
    base_dir,
    strategy="sparse",
    prefetch=2,
    **kwargs,
) -> str

Streaming delta save. Peak RAM is O(prefetch tensors), not O(two full models).

Args:

Parameter	Type	Description
`out_path`	`str \\| Path`	Output `.wdelta` file path
`finetuned_dir`	`str \\| Path`	Folder containing fine-tuned safetensors shards
`base_dir`	`str \\| Path`	Folder containing base safetensors shards
`strategy`	`str`	`"sparse"`, `"quantized"`, or `"int4"`
`prefetch`	`int`	Number of tensor pairs to prefetch (default 2)
`**kwargs`		Strategy-specific options (see below)

Strategy kwargs:

Strategy	kwarg	default	description
`sparse`	`sparsity`	`0.9`	Fraction of weights to zero out
`int4`	`outlier_fraction`	`0.01`	Fraction of weights stored as float16 outliers

Returns: SHA-256 hex hash of the base model.

load_delta_from_paths

dt.load_delta_from_paths(
    path,
    base_dir,
    verify=True,
) -> Dict[str, np.ndarray]

Reconstruct a fine-tuned model from a .wdelta file and a base model directory. Loads each base shard once — O(n_shards) file opens rather than O(n_tensors × n_shards).

Args:

Parameter	Type	Description
`path`	`str \\| Path`	Path to the `.wdelta` file
`base_dir`	`str \\| Path`	Folder containing base safetensors shards
`verify`	`bool`	SHA-256 verify base before reconstructing (default `True`)

Returns: Reconstructed state dict as Dict[str, np.ndarray].

inspect

dt.inspect(path) -> dict

Return metadata from a .wdelta file without loading the base model.

Args:

Parameter	Type	Description
`path`	`str \\| Path`	Path to the `.wdelta` file

Returns:

{
    "path": "checkpoint.wdelta",
    "size_mb": 294.2,
    "parent_hash": "e1810a...",
    "strategy": "int4",
    "n_tensors": 290,
    "tensors": {
        "model.embed_tokens.weight": {"shape": [151936, 896], "dtype": "float32"},
        ...
    }
}

save_delta

dt.save_delta(
    path,
    finetuned,
    base,
    strategy="sparse",
    **kwargs,
) -> str

Compute and save the delta between finetuned and base. Loads both models fully into RAM — for models larger than ~3B use save_delta_from_paths instead.

Args:

Parameter	Type	Description
`path`	`str \\| Path`	Output `.wdelta` file path
`finetuned`	`Dict[str, np.ndarray \\| Tensor]`	Fine-tuned state dict
`base`	`Dict[str, np.ndarray \\| Tensor]`	Base state dict
`strategy`	`str`	`"sparse"`, `"quantized"`, or `"int4"`

Returns: SHA-256 hex hash of the base model.

load_delta

dt.load_delta(
    path,
    base,
    verify=True,
) -> Dict[str, np.ndarray]

Reconstruct a fine-tuned model from a .wdelta file and a base state dict. Requires the full base loaded in RAM — for large models use load_delta_from_paths instead.

Args:

Parameter	Type	Description
`path`	`str \\| Path`	Path to the `.wdelta` file
`base`	`Dict[str, np.ndarray \\| Tensor]`	Base state dict
`verify`	`bool`	SHA-256 verify base before reconstructing (default `True`)

Returns: Reconstructed state dict as Dict[str, np.ndarray].