Skip to content

[utils] sync save_state_dict from tensornvme #6128

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
botbw wants to merge 2 commits into from
Closed

[utils] sync save_state_dict from tensornvme #6128

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
177e842
[utils] sync save_state_dict from tensornvme
botbw Nov 13, 2024
555f5f8
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot] Nov 13, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
26 changes: 17 additions & 9 deletions colossalai/utils/safetensors.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,7 +1,7 @@
# a python safetensors serializer modified from https://github.com/huggingface/safetensors/blob/41bd1acf38ad28ac559522d40596c6c802f79453/safetensors/src/tensor.rs#L214
import json
from dataclasses import asdict, dataclass
from typing import Dict, List, Tuple
from typing import Dict, List, Optional, Tuple

import torch
from safetensors.torch import _TYPES
Expand All @@ -27,10 +27,10 @@ class PreparedData:
offset: int


def prepare(data: Dict[str, torch.Tensor]) -> Tuple[PreparedData, List[torch.Tensor]]:
def prepare(data: Dict[str, torch.Tensor]) -> Tuple[PreparedData, List[str]]:
sorted_data = sorted(data.items(), key=lambda x: (x[1].dtype, x[0]))

tensors = []
tensor_keys = []
metadata = {}
offset = 0

Expand All @@ -41,7 +41,7 @@ def prepare(data: Dict[str, torch.Tensor]) -> Tuple[PreparedData, List[torch.Ten
)
offset += n
metadata[name] = asdict(tensor_info)
tensors.append(tensor)
tensor_keys.append(name)

metadata_buf = json.dumps(metadata).encode("utf-8")

Expand All @@ -50,15 +50,23 @@ def prepare(data: Dict[str, torch.Tensor]) -> Tuple[PreparedData, List[torch.Ten

n = len(metadata_buf)

return PreparedData(n=n, header_bytes=metadata_buf, offset=offset), tensors
return PreparedData(n=n, header_bytes=metadata_buf, offset=offset), tensor_keys


def save(f_writer: AsyncFileWriter, state_dict: Dict[str, torch.Tensor]) -> None:
prepared_data, tensors = prepare(state_dict)
def save_state_dict(
f_writer: AsyncFileWriter,
state_dict: Dict[str, torch.Tensor],
state_dict_pinned: Optional[Dict[str, torch.Tensor]] = None,
) -> None:
prepared_data, tensor_keys = prepare(state_dict)
n, header_bytes, _ = prepared_data.n, prepared_data.header_bytes, prepared_data.offset

f_writer.write(n.to_bytes(8, byteorder="little"))
f_writer.write(header_bytes)

for tensor in tensors:
f_writer.write_raw(tensor, tensor.data_ptr(), tensor.numel() * tensor.element_size(), f_writer.offset)
f_writer.register_h2d(len(tensor_keys))
for name in tensor_keys:
if state_dict_pinned:
f_writer.write_tensor(state_dict[name], state_dict_pinned[name])
else:
f_writer.write_tensor(state_dict[name])