Use streaming transfers

Added in a new streaming_save/load implementations that don't use an
extra layer of zipfile serialization

Co-authored-by: Tristan Rice <[email protected]>
Co-authored-by: Krishn Parasar <[email protected]>

torchft/checkpointing.py

@@ -20,11 +20,10 @@
 from contextlib import contextmanager
 from datetime import timedelta
 from http.server import BaseHTTPRequestHandler
-from typing import Generator, Generic, List, Optional, TypeVar
-
-import torch
+from typing import Generator, Generic, List, Optional, TypeVar, cast
 
 from torchft.http import _IPv6HTTPServer
+from torchft.serialization import streaming_load, streaming_save
 
 logger: logging.Logger = logging.getLogger(__name__)
 
@@ -161,7 +160,7 @@ def do_GET(self):
 
                         state_dict = ckpt_server._state_dict
 
-                        torch.save(state_dict, self.wfile)
+                        streaming_save(state_dict, self.wfile)
                 except Exception as e:
                     logger.exception(
                         f"Exception in checkpoint server when handling {self.path=}: {e}",
@@ -198,9 +197,8 @@ def load_from_address(cls, address: str, timeout: timedelta) -> T:
             data = f.read()
 
         reader = io.BytesIO(data)
-        # We have to set weights_only to False as there are some non-tensor
-        # states like lr_scheduler.
-        return torch.load(reader, weights_only=False)
+        state_dict = streaming_load(reader)
+        return cast(T, state_dict)
 
     def address(self) -> str:
         """

torchft/checkpointing_test.py

@@ -5,6 +5,7 @@
 # LICENSE file in the root directory of this source tree.
 
 import threading
+import unittest
 import urllib.error
 from datetime import timedelta
 from unittest import TestCase
@@ -103,3 +104,7 @@ def test_timed_acquire(self) -> None:
                 pass
 
         self.assertTrue(lock.locked())
+
+
+if __name__ == "__main__":
+    unittest.main()

torchft/serialization.py

@@ -0,0 +1,100 @@
+import pickle
+from dataclasses import dataclass
+from io import BufferedIOBase
+from typing import Dict, Tuple
+
+import torch
+
+
+@dataclass
+class _Entry:
+    key: str
+    dtype: object
+    is_storage: bool
+    length: int
+
+
+class _InMemoryStateDict:
+    def __init__(self) -> None:
+        self.records: Dict[str, Tuple[object, int]] = {}
+
+    def write_record(self, key: str, data: object, length: int) -> None:
+        self.records[key] = (data, length)
+
+    def write_to(self, f: BufferedIOBase) -> None:
+        entries = []
+        for key, (data, length) in self.records.items():
+            entries.append(
+                _Entry(
+                    key=key,
+                    is_storage=isinstance(data, torch.UntypedStorage),
+                    dtype=type(data),
+                    length=length,
+                )
+            )
+
+        pickle.dump(entries, f)
+
+        for key, (data, length) in self.records.items():
+            if isinstance(data, bytes):
+                f.write(data)
+            elif isinstance(data, str):
+                f.write(data.encode("utf-8"))
+            elif isinstance(data, torch.UntypedStorage):
+                data._write_file(f, False, False, 1)
+            else:
+                raise TypeError(f"unknown type: {type(data)}")
+
+    def read_from(self, f: BufferedIOBase) -> None:
+        entries = pickle.load(f)
+
+        for entry in entries:
+            data = f.read(entry.length)
+            if entry.is_storage:
+                storage = torch.frombuffer(
+                    data,
+                    dtype=torch.uint8,
+                ).untyped_storage()
+
+                self.records[entry.key] = (
+                    storage,
+                    entry.length,
+                )
+            else:
+                self.records[entry.key] = (data, entry.length)
+
+    def has_record(self, key: str) -> bool:
+        return key in self.records
+
+    def get_record(self, key: str) -> object:
+        return self.records[key][0]
+
+    def get_storage_from_record(
+        self, key: str, _length: int, _type: int
+    ) -> torch.Tensor:
+        return torch.tensor(self.records[key][0], dtype=torch.uint8)
+
+    def serialization_id(self) -> str:
+        return "torchft"
+
+
+def streaming_save(obj: object, f: BufferedIOBase) -> None:
+    out = _InMemoryStateDict()
+    torch.serialization._save(
+        obj,
+        zip_file=out,
+        pickle_module=pickle,
+        pickle_protocol=2,
+        _disable_byteorder_record=False,
+    )
+    out.write_to(f)
+
+
+def streaming_load(f: BufferedIOBase) -> object:
+    out = _InMemoryStateDict()
+    out.read_from(f)
+    return torch.serialization._load(
+        zip_file=out,
+        map_location=None,
+        pickle_module=pickle,
+    )

torchft/serialization_test.py

@@ -0,0 +1,106 @@
+from io import BytesIO
+from typing import cast
+from unittest import TestCase
+
+import torch
+import torch.distributed as dist
+from torch.distributed.tensor import DeviceMesh, DTensor, distribute_tensor
+
+from torchft.serialization import streaming_load, streaming_save
+
+
+class MyClass:
+    def __init__(self, a: int) -> None:
+        self.a = a
+
+    def __eq__(self, other: "MyClass") -> bool:
+        return self.a == other.a
+
+
+class TestCheckpointingSerialization(TestCase):
+    def test_scalar_tensor(self) -> None:
+        tensor = torch.tensor(42, dtype=torch.int32)
+        state_dict = {"scalar": tensor}
+        file = BytesIO()
+        streaming_save(state_dict, file)
+        file.seek(0)
+
+        result = streaming_load(file)
+        torch.testing.assert_close(result, state_dict)
+
+    def test_strided_tensor(self) -> None:
+        base_tensor = torch.arange(16, dtype=torch.float32).reshape(4, 4)
+        strided_tensor = base_tensor[::2, ::2]
+        state_dict = {"strided": strided_tensor}
+        file = BytesIO()
+        streaming_save(state_dict, file)
+        file.seek(0)
+
+        result = streaming_load(file)
+        torch.testing.assert_close(result, state_dict)
+
+    def test_tensor_with_offset(self) -> None:
+        base_tensor = torch.arange(10, dtype=torch.float64)
+        offset_tensor = base_tensor[2:]
+        state_dict = {"offset": offset_tensor}
+        file = BytesIO()
+        streaming_save(state_dict, file)
+        file.seek(0)
+
+        result = streaming_load(file)
+        torch.testing.assert_close(result, state_dict)
+
+    def test_nested_tensors(self) -> None:
+        tensor1 = torch.tensor([1, 2, 3], dtype=torch.int32)
+        tensor2 = torch.tensor([[1.5, 2.5], [3.5, 4.5]], dtype=torch.float64)
+        state_dict = {"nested": {"tensor1": tensor1, "tensor2": tensor2}}
+        file = BytesIO()
+        streaming_save(state_dict, file)
+        file.seek(0)
+
+        result = streaming_load(file)
+        torch.testing.assert_close(result, state_dict)
+
+    def test_various_data_types(self) -> None:
+        tensor_float32 = torch.tensor([1.0, 2.0, 3.0], dtype=torch.float32)
+        tensor_int16 = torch.tensor([1, 2, 3], dtype=torch.int16)
+        tensor_bool = torch.tensor([True, False, True], dtype=torch.bool)
+        state_dict = {
+            "float32": tensor_float32,
+            "int16": tensor_int16,
+            "bool": tensor_bool,
+        }
+        file = BytesIO()
+        streaming_save(state_dict, file)
+        file.seek(0)
+
+        result = streaming_load(file)
+        torch.testing.assert_close(result, state_dict)
+
+    def test_dtensor(self) -> None:
+        dist.init_process_group(
+            backend="gloo", rank=0, world_size=1, store=dist.HashStore()
+        )
+
+        device_mesh = DeviceMesh("cpu", 1)
+        tensor = torch.randn(4, 4, device="cuda")
+        dtensor = distribute_tensor(tensor, device_mesh, [])
+        state_dict = dtensor
+        file = BytesIO()
+        streaming_save(state_dict, file)
+        file.seek(0)
+
+        result = cast(DTensor, streaming_load(file))
+        torch.testing.assert_close(result.to_local(), state_dict.to_local())
+
+    def test_python_object(self) -> None:
+        state_dict = {
+            "obj": MyClass(42),
+        }
+
+        file = BytesIO()
+        streaming_save(state_dict, file)
+        file.seek(0)
+
+        result = streaming_load(file)
+        self.assertEqual(result, state_dict)