import os import sys from collections import defaultdict from typing import Any, Dict, List, Optional, Set, Tuple, Union import torch from safetensors import deserialize, safe_open, serialize, serialize_file def storage_ptr(tensor: torch.Tensor) -> int: try: return tensor.untyped_storage().data_ptr() except Exception: # Fallback for torch==1.10 try: return tensor.storage().data_ptr() except NotImplementedError: # Fallback for meta storage return 0 def _end_ptr(tensor: torch.Tensor) -> int: if tensor.nelement(): stop = tensor.view(-1)[-1].data_ptr() + _SIZE[tensor.dtype] else: stop = tensor.data_ptr() return stop def storage_size(tensor: torch.Tensor) -> int: try: return tensor.untyped_storage().nbytes() except AttributeError: # Fallback for torch==1.10 try: return tensor.storage().size() * _SIZE[tensor.dtype] except NotImplementedError: # Fallback for meta storage # On torch >=2.0 this is the tensor size return tensor.nelement() * _SIZE[tensor.dtype] def _filter_shared_not_shared(tensors: List[Set[str]], state_dict: Dict[str, torch.Tensor]) -> List[Set[str]]: filtered_tensors = [] for shared in tensors: if len(shared) < 2: filtered_tensors.append(shared) continue areas = [] for name in shared: tensor = state_dict[name] areas.append((tensor.data_ptr(), _end_ptr(tensor), name)) areas.sort() _, last_stop, last_name = areas[0] filtered_tensors.append({last_name}) for start, stop, name in areas[1:]: if start >= last_stop: filtered_tensors.append({name}) else: filtered_tensors[-1].add(name) last_stop = stop return filtered_tensors def _find_shared_tensors(state_dict: Dict[str, torch.Tensor]) -> List[Set[str]]: tensors = defaultdict(set) for k, v in state_dict.items(): if v.device != torch.device("meta") and storage_ptr(v) != 0 and storage_size(v) != 0: # Need to add device as key because of multiple GPU. tensors[(v.device, storage_ptr(v), storage_size(v))].add(k) tensors = list(sorted(tensors.values())) tensors = _filter_shared_not_shared(tensors, state_dict) return tensors def _is_complete(tensor: torch.Tensor) -> bool: return tensor.data_ptr() == storage_ptr(tensor) and tensor.nelement() * _SIZE[tensor.dtype] == storage_size(tensor) def _remove_duplicate_names( state_dict: Dict[str, torch.Tensor], *, preferred_names: Optional[List[str]] = None, discard_names: Optional[List[str]] = None, ) -> Dict[str, List[str]]: if preferred_names is None: preferred_names = [] preferred_names = set(preferred_names) if discard_names is None: discard_names = [] discard_names = set(discard_names) shareds = _find_shared_tensors(state_dict) to_remove = defaultdict(list) for shared in shareds: complete_names = set([name for name in shared if _is_complete(state_dict[name])]) if not complete_names: raise RuntimeError( "Error while trying to find names to remove to save state dict, but found no suitable name to keep" f" for saving amongst: {shared}. None is covering the entire storage.Refusing to save/load the model" " since you could be storing much more memory than needed. Please refer to" " https://huggingface.co/docs/safetensors/torch_shared_tensors for more information. Or open an" " issue." ) keep_name = sorted(list(complete_names))[0] # Mechanism to preferentially select keys to keep # coming from the on-disk file to allow # loading models saved with a different choice # of keep_name preferred = complete_names.difference(discard_names) if preferred: keep_name = sorted(list(preferred))[0] if preferred_names: preferred = preferred_names.intersection(complete_names) if preferred: keep_name = sorted(list(preferred))[0] for name in sorted(shared): if name != keep_name: to_remove[keep_name].append(name) return to_remove def save_model( model: torch.nn.Module, filename: str, metadata: Optional[Dict[str, str]] = None, force_contiguous: bool = True ): """ Saves a given torch model to specified filename. This method exists specifically to avoid tensor sharing issues which are not allowed in `safetensors`. [More information on tensor sharing](../torch_shared_tensors) Args: model (`torch.nn.Module`): The model to save on disk. filename (`str`): The filename location to save the file metadata (`Dict[str, str]`, *optional*): Extra information to save along with the file. Some metadata will be added for each dropped tensors. This information will not be enough to recover the entire shared structure but might help understanding things force_contiguous (`boolean`, *optional*, defaults to True): Forcing the state_dict to be saved as contiguous tensors. This has no effect on the correctness of the model, but it could potentially change performance if the layout of the tensor was chosen specifically for that reason. """ state_dict = model.state_dict() to_removes = _remove_duplicate_names(state_dict) for kept_name, to_remove_group in to_removes.items(): for to_remove in to_remove_group: if metadata is None: metadata = {} if to_remove not in metadata: # Do not override user data metadata[to_remove] = kept_name del state_dict[to_remove] if force_contiguous: state_dict = {k: v.contiguous() for k, v in state_dict.items()} try: save_file(state_dict, filename, metadata=metadata) except ValueError as e: msg = str(e) msg += " Or use save_model(..., force_contiguous=True), read the docs for potential caveats." raise ValueError(msg) def load_model(model: torch.nn.Module, filename: Union[str, os.PathLike], strict: bool = True, device: Union[str, int] = "cpu") -> Tuple[List[str], List[str]]: """ Loads a given filename onto a torch model. This method exists specifically to avoid tensor sharing issues which are not allowed in `safetensors`. [More information on tensor sharing](../torch_shared_tensors) Args: model (`torch.nn.Module`): The model to load onto. filename (`str`, or `os.PathLike`): The filename location to load the file from. strict (`bool`, *optional*, defaults to True): Whether to fail if you're missing keys or having unexpected ones. When false, the function simply returns missing and unexpected names. device (`Union[str, int]`, *optional*, defaults to `cpu`): The device where the tensors need to be located after load. available options are all regular torch device locations. Returns: `(missing, unexpected): (List[str], List[str])` `missing` are names in the model which were not modified during loading `unexpected` are names that are on the file, but weren't used during the load. """ state_dict = load_file(filename, device=device) model_state_dict = model.state_dict() to_removes = _remove_duplicate_names(model_state_dict, preferred_names=state_dict.keys()) missing, unexpected = model.load_state_dict(state_dict, strict=False) missing = set(missing) for to_remove_group in to_removes.values(): for to_remove in to_remove_group: if to_remove not in missing: unexpected.append(to_remove) else: missing.remove(to_remove) if strict and (missing or unexpected): missing_keys = ", ".join([f'"{k}"' for k in sorted(missing)]) unexpected_keys = ", ".join([f'"{k}"' for k in sorted(unexpected)]) error = f"Error(s) in loading state_dict for {model.__class__.__name__}:" if missing: error += f"\n Missing key(s) in state_dict: {missing_keys}" if unexpected: error += f"\n Unexpected key(s) in state_dict: {unexpected_keys}" raise RuntimeError(error) return missing, unexpected def save(tensors: Dict[str, torch.Tensor], metadata: Optional[Dict[str, str]] = None) -> bytes: """ Saves a dictionary of tensors into raw bytes in safetensors format. Args: tensors (`Dict[str, torch.Tensor]`): The incoming tensors. Tensors need to be contiguous and dense. metadata (`Dict[str, str]`, *optional*, defaults to `None`): Optional text only metadata you might want to save in your header. For instance it can be useful to specify more about the underlying tensors. This is purely informative and does not affect tensor loading. Returns: `bytes`: The raw bytes representing the format Example: ```python from safetensors.torch import save import torch tensors = {"embedding": torch.zeros((512, 1024)), "attention": torch.zeros((256, 256))} byte_data = save(tensors) ``` """ serialized = serialize(_flatten(tensors), metadata=metadata) result = bytes(serialized) return result def save_file( tensors: Dict[str, torch.Tensor], filename: Union[str, os.PathLike], metadata: Optional[Dict[str, str]] = None, ): """ Saves a dictionary of tensors into raw bytes in safetensors format. Args: tensors (`Dict[str, torch.Tensor]`): The incoming tensors. Tensors need to be contiguous and dense. filename (`str`, or `os.PathLike`)): The filename we're saving into. metadata (`Dict[str, str]`, *optional*, defaults to `None`): Optional text only metadata you might want to save in your header. For instance it can be useful to specify more about the underlying tensors. This is purely informative and does not affect tensor loading. Returns: `None` Example: ```python from safetensors.torch import save_file import torch tensors = {"embedding": torch.zeros((512, 1024)), "attention": torch.zeros((256, 256))} save_file(tensors, "model.safetensors") ``` """ serialize_file(_flatten(tensors), filename, metadata=metadata) def load_file(filename: Union[str, os.PathLike], device: Union[str, int] = "cpu") -> Dict[str, torch.Tensor]: """ Loads a safetensors file into torch format. Args: filename (`str`, or `os.PathLike`): The name of the file which contains the tensors device (`Union[str, int]`, *optional*, defaults to `cpu`): The device where the tensors need to be located after load. available options are all regular torch device locations. Returns: `Dict[str, torch.Tensor]`: dictionary that contains name as key, value as `torch.Tensor` Example: ```python from safetensors.torch import load_file file_path = "./my_folder/bert.safetensors" loaded = load_file(file_path) ``` """ result = {} with safe_open(filename, framework="pt", device=device) as f: for k in f.keys(): result[k] = f.get_tensor(k) return result def load(data: bytes) -> Dict[str, torch.Tensor]: """ Loads a safetensors file into torch format from pure bytes. Args: data (`bytes`): The content of a safetensors file Returns: `Dict[str, torch.Tensor]`: dictionary that contains name as key, value as `torch.Tensor` on cpu Example: ```python from safetensors.torch import load file_path = "./my_folder/bert.safetensors" with open(file_path, "rb") as f: data = f.read() loaded = load(data) ``` """ flat = deserialize(data) return _view2torch(flat) # torch.float8 formats require 2.1; we do not support these dtypes on earlier versions _float8_e4m3fn = getattr(torch, "float8_e4m3fn", None) _float8_e5m2 = getattr(torch, "float8_e5m2", None) _SIZE = { torch.int64: 8, torch.float32: 4, torch.int32: 4, torch.bfloat16: 2, torch.float16: 2, torch.int16: 2, torch.uint8: 1, torch.int8: 1, torch.bool: 1, torch.float64: 8, _float8_e4m3fn: 1, _float8_e5m2: 1, } _TYPES = { "F64": torch.float64, "F32": torch.float32, "F16": torch.float16, "BF16": torch.bfloat16, "I64": torch.int64, # "U64": torch.uint64, "I32": torch.int32, # "U32": torch.uint32, "I16": torch.int16, # "U16": torch.uint16, "I8": torch.int8, "U8": torch.uint8, "BOOL": torch.bool, "F8_E4M3": _float8_e4m3fn, "F8_E5M2": _float8_e5m2, } def _getdtype(dtype_str: str) -> torch.dtype: return _TYPES[dtype_str] def _view2torch(safeview) -> Dict[str, torch.Tensor]: result = {} for k, v in safeview: dtype = _getdtype(v["dtype"]) arr = torch.frombuffer(v["data"], dtype=dtype).reshape(v["shape"]) if sys.byteorder == "big": arr = torch.from_numpy(arr.numpy().byteswap(inplace=False)) result[k] = arr return result def _tobytes(tensor: torch.Tensor, name: str) -> bytes: if tensor.layout != torch.strided: raise ValueError( f"You are trying to save a sparse tensor: `{name}` which this library does not support." " You can make it a dense tensor before saving with `.to_dense()` but be aware this might" " make a much larger file than needed." ) if not tensor.is_contiguous(): raise ValueError( f"You are trying to save a non contiguous tensor: `{name}` which is not allowed. It either means you" " are trying to save tensors which are reference of each other in which case it's recommended to save" " only the full tensors, and reslice at load time, or simply call `.contiguous()` on your tensor to" " pack it before saving." ) if tensor.device.type != "cpu": # Moving tensor to cpu before saving tensor = tensor.to("cpu") import ctypes import numpy as np # When shape is empty (scalar), np.prod returns a float # we need a int for the following calculations length = int(np.prod(tensor.shape).item()) bytes_per_item = _SIZE[tensor.dtype] total_bytes = length * bytes_per_item ptr = tensor.data_ptr() if ptr == 0: return b"" newptr = ctypes.cast(ptr, ctypes.POINTER(ctypes.c_ubyte)) data = np.ctypeslib.as_array(newptr, (total_bytes,)) # no internal copy if sys.byteorder == "big": NPDTYPES = { torch.int64: np.int64, torch.float32: np.float32, torch.int32: np.int32, # XXX: This is ok because both have the same width torch.bfloat16: np.float16, torch.float16: np.float16, torch.int16: np.int16, torch.uint8: np.uint8, torch.int8: np.int8, torch.bool: bool, torch.float64: np.float64, # XXX: This is ok because both have the same width and byteswap is a no-op anyway _float8_e4m3fn: np.uint8, _float8_e5m2: np.uint8, } npdtype = NPDTYPES[tensor.dtype] # Not in place as that would potentially modify a live running model data = data.view(npdtype).byteswap(inplace=False) return data.tobytes() def _flatten(tensors: Dict[str, torch.Tensor]) -> Dict[str, Dict[str, Any]]: if not isinstance(tensors, dict): raise ValueError(f"Expected a dict of [str, torch.Tensor] but received {type(tensors)}") invalid_tensors = [] for k, v in tensors.items(): if not isinstance(v, torch.Tensor): raise ValueError(f"Key `{k}` is invalid, expected torch.Tensor but received {type(v)}") if v.layout != torch.strided: invalid_tensors.append(k) if invalid_tensors: raise ValueError( f"You are trying to save a sparse tensors: `{invalid_tensors}` which this library does not support." " You can make it a dense tensor before saving with `.to_dense()` but be aware this might" " make a much larger file than needed." ) shared_pointers = _find_shared_tensors(tensors) failing = [] for names in shared_pointers: if len(names) > 1: failing.append(names) if failing: raise RuntimeError( f""" Some tensors share memory, this will lead to duplicate memory on disk and potential differences when loading them again: {failing}. A potential way to correctly save your model is to use `save_model`. More information at https://huggingface.co/docs/safetensors/torch_shared_tensors """ ) return { k: { "dtype": str(v.dtype).split(".")[-1], "shape": v.shape, "data": _tobytes(v, k), } for k, v in tensors.items() }