checkpointer 2.1.0__tar.gz → 2.5.0__tar.gz

{checkpointer-2.1.0 → checkpointer-2.5.0}/LICENSE

@@ -1,4 +1,4 @@
-Copyright 2024 Hampus Hallman
+Copyright 2018-2025 Hampus Hallman
 
 Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
 

{checkpointer-2.1.0 → checkpointer-2.5.0}/PKG-INFO

@@ -1,18 +1,18 @@
-Metadata-Version: 2.3
+Metadata-Version: 2.4
 Name: checkpointer
-Version: 2.1.0
+Version: 2.5.0
 Summary: A Python library for memoizing function results with support for multiple storage backends, async runtimes, and automatic cache invalidation
 Project-URL: Repository, https://github.com/Reddan/checkpointer.git
 Author: Hampus Hallman
-License: Copyright 2024 Hampus Hallman
+License: Copyright 2018-2025 Hampus Hallman
         
         Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
         
         The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
         
         THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+License-File: LICENSE
 Requires-Python: >=3.12
-Requires-Dist: relib
 Description-Content-Type: text/markdown
 
 # checkpointer · [![License](https://img.shields.io/badge/license-MIT-blue)](https://github.com/Reddan/checkpointer/blob/master/LICENSE) [![pypi](https://img.shields.io/pypi/v/checkpointer)](https://pypi.org/project/checkpointer/) [![Python 3.12](https://img.shields.io/badge/python-3.12-blue)](https://pypi.org/project/checkpointer/)
@@ -108,7 +108,7 @@ Layer caches by stacking checkpoints:
 @dev_checkpoint  # Adds caching during development
 def some_expensive_function():
     print("Performing a time-consuming operation...")
-    return sum(i * i for i in range(10**6))
+    return sum(i * i for i in range(10**8))
 ```
 
 - **In development**: Both `dev_checkpoint` and `memory` caches are active.
@@ -153,6 +153,18 @@ Access cached results without recalculating:
 stored_result = expensive_function.get(4)
 ```
 
+### Refresh Function Hash
+
+When using `capture=True`, changes to captured variables are included in the function's hash to determine cache invalidation. However, `checkpointer` does not automatically update this hash during a running Python session—it recalculates between sessions or when you explicitly refresh it.
+
+Use the `reinit` method to manually refresh the function's hash within the same session:
+
+```python
+expensive_function.reinit()
+```
+
+This forces `checkpointer` to recalculate the hash of `expensive_function`, considering any changes to captured variables. It's useful when you've modified external variables that the function depends on and want the cache to reflect these changes immediately.
+
 ---
 
 ## Storage Backends
@@ -189,9 +201,9 @@ from checkpointer import checkpoint, Storage
 from datetime import datetime
 
 class CustomStorage(Storage):
+    def store(self, path, data): ...  # Save the checkpoint data
     def exists(self, path) -> bool: ...  # Check if a checkpoint exists at the given path
     def checkpoint_date(self, path) -> datetime: ...  # Return the date the checkpoint was created
-    def store(self, path, data): ...  # Save the checkpoint data
     def load(self, path): ...  # Return the checkpoint data
     def delete(self, path): ...  # Delete the checkpoint
 

{checkpointer-2.1.0 → checkpointer-2.5.0}/README.md

@@ -91,7 +91,7 @@ Layer caches by stacking checkpoints:
 @dev_checkpoint  # Adds caching during development
 def some_expensive_function():
     print("Performing a time-consuming operation...")
-    return sum(i * i for i in range(10**6))
+    return sum(i * i for i in range(10**8))
 ```
 
 - **In development**: Both `dev_checkpoint` and `memory` caches are active.
@@ -136,6 +136,18 @@ Access cached results without recalculating:
 stored_result = expensive_function.get(4)
 ```
 
+### Refresh Function Hash
+
+When using `capture=True`, changes to captured variables are included in the function's hash to determine cache invalidation. However, `checkpointer` does not automatically update this hash during a running Python session—it recalculates between sessions or when you explicitly refresh it.
+
+Use the `reinit` method to manually refresh the function's hash within the same session:
+
+```python
+expensive_function.reinit()
+```
+
+This forces `checkpointer` to recalculate the hash of `expensive_function`, considering any changes to captured variables. It's useful when you've modified external variables that the function depends on and want the cache to reflect these changes immediately.
+
 ---
 
 ## Storage Backends
@@ -172,9 +184,9 @@ from checkpointer import checkpoint, Storage
 from datetime import datetime
 
 class CustomStorage(Storage):
+    def store(self, path, data): ...  # Save the checkpoint data
     def exists(self, path) -> bool: ...  # Check if a checkpoint exists at the given path
     def checkpoint_date(self, path) -> datetime: ...  # Return the date the checkpoint was created
-    def store(self, path, data): ...  # Save the checkpoint data
     def load(self, path): ...  # Return the checkpoint data
     def delete(self, path): ...  # Delete the checkpoint
 

checkpointer-2.5.0/checkpointer/__init__.py

@@ -0,0 +1,20 @@
+import gc
+import tempfile
+from typing import Callable
+from .checkpoint import Checkpointer, CheckpointError, CheckpointFn
+from .object_hash import ObjectHash
+from .storages import MemoryStorage, PickleStorage, Storage
+
+create_checkpointer = Checkpointer
+checkpoint = Checkpointer()
+capture_checkpoint = Checkpointer(capture=True)
+memory_checkpoint = Checkpointer(format="memory", verbosity=0)
+tmp_checkpoint = Checkpointer(root_path=tempfile.gettempdir() + "/checkpoints")
+
+def cleanup_all(invalidated=True, expired=True):
+  for obj in gc.get_objects():
+    if isinstance(obj, CheckpointFn):
+      obj.cleanup(invalidated=invalidated, expired=expired)
+
+def get_function_hash(fn: Callable, capture=False) -> str:
+  return CheckpointFn(Checkpointer(capture=capture), fn).fn_hash

{checkpointer-2.1.0 → checkpointer-2.5.0}/checkpointer/checkpoint.py

@@ -1,15 +1,15 @@
 from __future__ import annotations
 import inspect
-import relib.hashing as hashing
-from typing import Generic, TypeVar, Type, TypedDict, Callable, Unpack, Literal, Any, cast, overload
-from pathlib import Path
+import re
 from datetime import datetime
 from functools import update_wrapper
-from .types import Storage
-from .function_body import get_function_hash
-from .utils import unwrap_fn, sync_resolve_coroutine, resolved_awaitable
-from .storages import STORAGE_MAP
+from pathlib import Path
+from typing import Any, Callable, Generic, Iterable, Literal, Type, TypedDict, TypeVar, Unpack, cast, overload
+from .fn_ident import get_fn_ident
+from .object_hash import ObjectHash
 from .print_checkpoint import print_checkpoint
+from .storages import STORAGE_MAP, Storage
+from .utils import resolved_awaitable, sync_resolve_coroutine, unwrap_fn
 
 Fn = TypeVar("Fn", bound=Callable)
 
@@ -50,22 +50,47 @@ class Checkpointer:
 
 class CheckpointFn(Generic[Fn]):
   def __init__(self, checkpointer: Checkpointer, fn: Fn):
-    wrapped = unwrap_fn(fn)
-    file_name = Path(wrapped.__code__.co_filename).name
-    update_wrapper(cast(Callable, self), wrapped)
-    storage = STORAGE_MAP[checkpointer.format] if isinstance(checkpointer.format, str) else checkpointer.format
     self.checkpointer = checkpointer
     self.fn = fn
-    self.fn_hash, self.depends = get_function_hash(wrapped, self.checkpointer.capture)
-    self.fn_id = f"{file_name}/{wrapped.__name__}"
+
+  def _set_ident(self, force=False):
+    if not hasattr(self, "fn_hash_raw") or force:
+      self.fn_hash_raw, self.depends = get_fn_ident(unwrap_fn(self.fn), self.checkpointer.capture)
+    return self
+
+  def _lazyinit(self):
+    wrapped = unwrap_fn(self.fn)
+    fn_file = Path(wrapped.__code__.co_filename).name
+    fn_name = re.sub(r"[^\w.]", "", wrapped.__qualname__)
+    update_wrapper(cast(Callable, self), wrapped)
+    store_format = self.checkpointer.format
+    Storage = STORAGE_MAP[store_format] if isinstance(store_format, str) else store_format
+    deep_hashes = [child._set_ident().fn_hash_raw for child in iterate_checkpoint_fns(self)]
+    self.fn_hash = str(ObjectHash().update_hash(self.fn_hash_raw, iter=deep_hashes))
+    self.fn_subdir = f"{fn_file}/{fn_name}/{self.fn_hash[:16]}"
     self.is_async = inspect.iscoroutinefunction(wrapped)
-    self.storage = storage(checkpointer)
+    self.storage = Storage(self)
+    self.cleanup = self.storage.cleanup
+
+  def __getattribute__(self, name: str) -> Any:
+    return object.__getattribute__(self, "_getattribute")(name)
+
+  def _getattribute(self, name: str) -> Any:
+    setattr(self, "_getattribute", super().__getattribute__)
+    self._lazyinit()
+    return self._getattribute(name)
+
+  def reinit(self, recursive=False):
+    pointfns = list(iterate_checkpoint_fns(self)) if recursive else [self]
+    for pointfn in pointfns:
+      pointfn._set_ident(True)
+    for pointfn in pointfns:
+      pointfn._lazyinit()
 
   def get_checkpoint_id(self, args: tuple, kw: dict) -> str:
     if not callable(self.checkpointer.path):
-      # TODO: use digest size before digesting instead of truncating the hash
-      call_hash = hashing.hash((self.fn_hash, args, kw), "blake2b")[:32]
-      return f"{self.fn_id}/{call_hash}"
+      call_hash = ObjectHash(self.fn_hash, args, kw, digest_size=16)
+      return f"{self.fn_subdir}/{call_hash}"
     checkpoint_id = self.checkpointer.path(*args, **kw)
     if not isinstance(checkpoint_id, str):
       raise CheckpointError(f"path function must return a string, got {type(checkpoint_id)}")
@@ -74,13 +99,13 @@ class CheckpointFn(Generic[Fn]):
   async def _store_on_demand(self, args: tuple, kw: dict, rerun: bool):
     checkpoint_id = self.get_checkpoint_id(args, kw)
     checkpoint_path = self.checkpointer.root_path / checkpoint_id
-    should_log = self.checkpointer.verbosity > 0
+    verbose = self.checkpointer.verbosity > 0
     refresh = rerun \
       or not self.storage.exists(checkpoint_path) \
       or (self.checkpointer.should_expire and self.checkpointer.should_expire(self.storage.checkpoint_date(checkpoint_path)))
 
     if refresh:
-      print_checkpoint(should_log, "MEMORIZING", checkpoint_id, "blue")
+      print_checkpoint(verbose, "MEMORIZING", checkpoint_id, "blue")
       data = self.fn(*args, **kw)
       if inspect.iscoroutine(data):
         data = await data
@@ -89,12 +114,12 @@ class CheckpointFn(Generic[Fn]):
 
     try:
       data = self.storage.load(checkpoint_path)
-      print_checkpoint(should_log, "REMEMBERED", checkpoint_id, "green")
+      print_checkpoint(verbose, "REMEMBERED", checkpoint_id, "green")
       return data
     except (EOFError, FileNotFoundError):
-      print_checkpoint(should_log, "CORRUPTED", checkpoint_id, "yellow")
-      self.storage.delete(checkpoint_path)
-      return await self._store_on_demand(args, kw, rerun)
+      pass
+    print_checkpoint(verbose, "CORRUPTED", checkpoint_id, "yellow")
+    return await self._store_on_demand(args, kw, True)
 
   def _call(self, args: tuple, kw: dict, rerun=False):
     if not self.checkpointer.when:
@@ -107,8 +132,8 @@ class CheckpointFn(Generic[Fn]):
     try:
       val = self.storage.load(checkpoint_path)
       return resolved_awaitable(val) if self.is_async else val
-    except:
-      raise CheckpointError("Could not load checkpoint")
+    except Exception as ex:
+      raise CheckpointError("Could not load checkpoint") from ex
 
   def exists(self, *args: tuple, **kw: dict) -> bool:
     return self.storage.exists(self.checkpointer.root_path / self.get_checkpoint_id(args, kw))
@@ -116,3 +141,15 @@ class CheckpointFn(Generic[Fn]):
   __call__: Fn = cast(Fn, lambda self, *args, **kw: self._call(args, kw))
   rerun: Fn = cast(Fn, lambda self, *args, **kw: self._call(args, kw, True))
   get: Fn = cast(Fn, lambda self, *args, **kw: self._get(args, kw))
+
+  def __repr__(self) -> str:
+    return f"<CheckpointFn {self.fn.__name__} {self.fn_hash[:6]}>"
+
+def iterate_checkpoint_fns(pointfn: CheckpointFn, visited: set[CheckpointFn] = set()) -> Iterable[CheckpointFn]:
+  visited = visited or set()
+  if pointfn not in visited:
+    yield pointfn
+    visited.add(pointfn)
+    for depend in pointfn.depends:
+      if isinstance(depend, CheckpointFn):
+        yield from iterate_checkpoint_fns(depend, visited)

checkpointer-2.5.0/checkpointer/fn_ident.py

@@ -0,0 +1,94 @@
+import dis
+import inspect
+from collections.abc import Callable
+from itertools import takewhile
+from pathlib import Path
+from types import CodeType, FunctionType, MethodType
+from typing import Any, Generator, Type, TypeGuard
+from .object_hash import ObjectHash
+from .utils import AttrDict, distinct, get_cell_contents, iterate_and_upcoming, transpose, unwrap_fn
+
+cwd = Path.cwd()
+
+def is_class(obj) -> TypeGuard[Type]:
+  # isinstance works too, but needlessly triggers __getattribute__
+  return issubclass(type(obj), type)
+
+def extract_classvars(code: CodeType, scope_vars: AttrDict) -> dict[str, dict[str, Type]]:
+  attr_path: tuple[str, ...] = ()
+  scope_obj = None
+  classvars: dict[str, dict[str, Type]] = {}
+  for instr, upcoming_instrs in iterate_and_upcoming(dis.get_instructions(code)):
+    if instr.opname in scope_vars and not attr_path:
+      attrs = takewhile(lambda instr: instr.opname == "LOAD_ATTR", upcoming_instrs)
+      attr_path = (instr.opname, instr.argval, *(str(x.argval) for x in attrs))
+    elif instr.opname == "CALL":
+      obj = scope_vars.get_at(attr_path)
+      attr_path = ()
+      if is_class(obj):
+        scope_obj = obj
+    elif instr.opname in ("STORE_FAST", "STORE_DEREF") and scope_obj:
+      load_key = instr.opname.replace("STORE", "LOAD")
+      classvars.setdefault(load_key, {})[instr.argval] = scope_obj
+      scope_obj = None
+  return classvars
+
+def extract_scope_values(code: CodeType, scope_vars: AttrDict) -> Generator[tuple[tuple[str, ...], Any], None, None]:
+  classvars = extract_classvars(code, scope_vars)
+  scope_vars = scope_vars.set({k: scope_vars[k].set(v) for k, v in classvars.items()})
+  for instr, upcoming_instrs in iterate_and_upcoming(dis.get_instructions(code)):
+    if instr.opname in scope_vars:
+      attrs = takewhile(lambda instr: instr.opname == "LOAD_ATTR", upcoming_instrs)
+      attr_path: tuple[str, ...] = (instr.opname, instr.argval, *(str(x.argval) for x in attrs))
+      val = scope_vars.get_at(attr_path)
+      if val is not None:
+        yield attr_path, val
+  for const in code.co_consts:
+    if isinstance(const, CodeType):
+      yield from extract_scope_values(const, scope_vars)
+
+def get_self_value(fn: Callable) -> type | object | None:
+  if isinstance(fn, MethodType):
+    return fn.__self__
+  parts = tuple(fn.__qualname__.split(".")[:-1])
+  cls = parts and AttrDict(fn.__globals__).get_at(parts)
+  if is_class(cls):
+    return cls
+
+def get_fn_captured_vals(fn: Callable) -> list[Any]:
+  self_value = get_self_value(fn)
+  scope_vars = AttrDict({
+    "LOAD_FAST": AttrDict({"self": self_value} if self_value else {}),
+    "LOAD_DEREF": AttrDict(get_cell_contents(fn)),
+    "LOAD_GLOBAL": AttrDict(fn.__globals__),
+  })
+  vals = dict(extract_scope_values(fn.__code__, scope_vars))
+  return list(vals.values())
+
+def is_user_fn(candidate_fn) -> TypeGuard[Callable]:
+  if not isinstance(candidate_fn, (FunctionType, MethodType)):
+    return False
+  fn_path = Path(inspect.getfile(candidate_fn)).resolve()
+  return cwd in fn_path.parents and ".venv" not in fn_path.parts
+
+def get_depend_fns(fn: Callable, capture: bool, captured_vals_by_fn: dict[Callable, list[Any]] = {}) -> dict[Callable, list[Any]]:
+  from .checkpoint import CheckpointFn
+  captured_vals_by_fn = captured_vals_by_fn or {}
+  captured_vals = get_fn_captured_vals(fn)
+  captured_vals_by_fn[fn] = [val for val in captured_vals if not callable(val)] * capture
+  child_fns = (unwrap_fn(val, checkpoint_fn=True) for val in captured_vals if callable(val))
+  for child_fn in child_fns:
+    if isinstance(child_fn, CheckpointFn):
+      captured_vals_by_fn[child_fn] = []
+    elif child_fn not in captured_vals_by_fn and is_user_fn(child_fn):
+      get_depend_fns(child_fn, capture, captured_vals_by_fn)
+  return captured_vals_by_fn
+
+def get_fn_ident(fn: Callable, capture: bool) -> tuple[str, list[Callable]]:
+  from .checkpoint import CheckpointFn
+  captured_vals_by_fn = get_depend_fns(fn, capture)
+  depends, depend_captured_vals = transpose(captured_vals_by_fn.items(), 2)
+  depends = distinct(fn.__func__ if isinstance(fn, MethodType) else fn for fn in depends)
+  unwrapped_depends = [fn for fn in depends if not isinstance(fn, CheckpointFn)]
+  fn_hash = str(ObjectHash(fn, unwrapped_depends).update(depend_captured_vals, tolerate_errors=True))
+  return fn_hash, depends

checkpointer-2.5.0/checkpointer/object_hash.py

@@ -0,0 +1,186 @@
+import ctypes
+import hashlib
+import io
+import re
+from collections.abc import Iterable
+from contextlib import nullcontext
+from decimal import Decimal
+from itertools import chain
+from pickle import HIGHEST_PROTOCOL as PROTOCOL
+from types import BuiltinFunctionType, FunctionType, GeneratorType, MethodType, ModuleType, UnionType
+from typing import Any, TypeAliasType, TypeVar
+from .utils import ContextVar, get_fn_body
+
+try:
+  import numpy as np
+except:
+  np = None
+try:
+  import torch
+except:
+  torch = None
+
+def encode_type(t: type | FunctionType) -> str:
+  return f"{t.__module__}:{t.__qualname__}"
+
+def encode_val(v: Any) -> str:
+  return encode_type(type(v))
+
+class ObjectHashError(Exception):
+  def __init__(self, obj: Any, cause: Exception):
+    super().__init__(f"{type(cause).__name__} error when hashing {obj}")
+    self.obj = obj
+
+class ObjectHash:
+  def __init__(self, *obj: Any, iter: Iterable[Any] = [], digest_size=64, tolerate_errors=False) -> None:
+    self.hash = hashlib.blake2b(digest_size=digest_size)
+    self.current: dict[int, int] = {}
+    self.tolerate_errors = ContextVar(tolerate_errors)
+    self.update(iter=chain(obj, iter))
+
+  def copy(self) -> "ObjectHash":
+    new = ObjectHash(tolerate_errors=self.tolerate_errors.value)
+    new.hash = self.hash.copy()
+    return new
+
+  def hexdigest(self) -> str:
+    return self.hash.hexdigest()
+
+  __str__ = hexdigest
+
+  def update_hash(self, *data: bytes | str, iter: Iterable[bytes | str] = []) -> "ObjectHash":
+    for d in chain(data, iter):
+      self.hash.update(d.encode() if isinstance(d, str) else d)
+    return self
+
+  def header(self, *args: Any) -> "ObjectHash":
+    return self.update_hash(":".join(map(str, args)))
+
+  def update(self, *objs: Any, iter: Iterable[Any] = [], tolerate_errors: bool | None=None) -> "ObjectHash":
+    with nullcontext() if tolerate_errors is None else self.tolerate_errors.set(tolerate_errors):
+      for obj in chain(objs, iter):
+        try:
+          self._update_one(obj)
+        except Exception as ex:
+          if self.tolerate_errors.value:
+            self.header("error").update(type(ex))
+            continue
+          raise ObjectHashError(obj, ex) from ex
+      return self
+
+  def _update_one(self, obj: Any) -> None:
+    match obj:
+      case None:
+        self.header("null")
+
+      case bool() | int() | float() | complex() | Decimal() | ObjectHash():
+        self.header("number", encode_val(obj), obj)
+
+      case str() | bytes() | bytearray() | memoryview():
+        self.header("bytes", encode_val(obj), len(obj)).update_hash(obj)
+
+      case set() | frozenset():
+        self.header("set", encode_val(obj), len(obj))
+        try:
+          items = sorted(obj)
+        except:
+          self.header("unsortable")
+          items = sorted(str(ObjectHash(item, tolerate_errors=self.tolerate_errors.value)) for item in obj)
+        self.update(iter=items)
+
+      case TypeVar():
+        self.header("TypeVar").update(obj.__name__, obj.__bound__, obj.__constraints__, obj.__contravariant__, obj.__covariant__)
+
+      case TypeAliasType():
+        self.header("TypeAliasType").update(obj.__name__, obj.__value__)
+
+      case UnionType():
+        self.header("UnionType").update(obj.__args__)
+
+      case BuiltinFunctionType():
+        self.header("builtin", obj.__qualname__)
+
+      case FunctionType():
+        self.header("function", encode_type(obj)).update(get_fn_body(obj), obj.__defaults__, obj.__kwdefaults__, obj.__annotations__)
+
+      case MethodType():
+        self.header("method").update(obj.__func__, obj.__self__.__class__)
+
+      case ModuleType():
+        self.header("module", obj.__name__, obj.__file__)
+
+      case GeneratorType():
+        self.header("generator", obj.__qualname__)._update_iterator(obj)
+
+      case io.TextIOWrapper() | io.FileIO() | io.BufferedRandom() | io.BufferedWriter() | io.BufferedReader():
+        self.header("file", encode_val(obj)).update(obj.name, obj.mode, obj.tell())
+
+      case type():
+        self.header("type", encode_type(obj))
+
+      case _ if np and isinstance(obj, np.dtype):
+        self.header("dtype").update(obj.__class__, obj.descr)
+
+      case _ if np and isinstance(obj, np.ndarray):
+        self.header("ndarray", encode_val(obj), obj.shape, obj.strides).update(obj.dtype)
+        if obj.dtype.hasobject:
+          self.update(obj.__reduce_ex__(PROTOCOL))
+        else:
+          array = np.ascontiguousarray(obj if obj.base is None else obj.base).view(np.uint8)
+          self.update_hash(array.data)
+
+      case _ if torch and isinstance(obj, torch.Tensor):
+        self.header("tensor", encode_val(obj), str(obj.dtype), tuple(obj.shape), obj.stride(), str(obj.device))
+        if obj.device.type != "cpu":
+          obj = obj.cpu()
+        storage = obj.storage()
+        buffer = (ctypes.c_ubyte * (storage.nbytes())).from_address(storage.data_ptr())
+        self.update_hash(memoryview(buffer))
+
+      case _ if id(obj) in self.current:
+        self.header("circular", self.current[id(obj)])
+
+      case _:
+        try:
+          self.current[id(obj)] = len(self.current)
+          match obj:
+            case list() | tuple():
+              self.header("list", encode_val(obj), len(obj)).update(iter=obj)
+            case dict():
+              try:
+                items = sorted(obj.items())
+              except:
+                items = sorted((str(ObjectHash(key, tolerate_errors=self.tolerate_errors.value)), val) for key, val in obj.items())
+              self.header("dict", encode_val(obj), len(obj)).update(iter=chain.from_iterable(items))
+            case _:
+              self._update_object(obj)
+        finally:
+          del self.current[id(obj)]
+
+  def _update_iterator(self, obj: Iterable) -> None:
+    self.header("iterator", encode_val(obj)).update(iter=obj).header(b"iterator-end")
+
+  def _update_object(self, obj: object) -> "ObjectHash":
+    self.header("instance", encode_val(obj))
+    try:
+      reduced = obj.__reduce_ex__(PROTOCOL) if hasattr(obj, "__reduce_ex__") else obj.__reduce__()
+    except:
+      reduced = None
+    if isinstance(reduced, str):
+      return self.header("reduce-str").update(reduced)
+    if reduced:
+      reduced = list(reduced)
+      it = reduced.pop(3) if len(reduced) >= 4 else None
+      self.header("reduce").update(reduced)
+      if it is not None:
+        self._update_iterator(it)
+      return self
+    if state := hasattr(obj, "__getstate__") and obj.__getstate__():
+      return self.header("getstate").update(state)
+    if len(getattr(obj, "__slots__", [])):
+      slots = {slot: getattr(obj, slot, None) for slot in getattr(obj, "__slots__")}
+      return self.header("slots").update(slots)
+    if d := getattr(obj, "__dict__", {}):
+      return self.header("dict").update(d)
+    repr_str = re.sub(r"\s+(at\s+0x[0-9a-fA-F]+)(>)$", r"\2", repr(obj))
+    return self.header("repr").update(repr_str)

{checkpointer-2.1.0 → checkpointer-2.5.0}/checkpointer/storages/__init__.py

@@ -1,5 +1,5 @@
 from typing import Type
-from ..types import Storage
+from .storage import Storage
 from .pickle_storage import PickleStorage
 from .memory_storage import MemoryStorage
 from .bcolz_storage import BcolzStorage

{checkpointer-2.1.0 → checkpointer-2.5.0}/checkpointer/storages/bcolz_storage.py

@@ -1,7 +1,7 @@
 import shutil
 from pathlib import Path
 from datetime import datetime
-from ..types import Storage
+from .storage import Storage
 
 def get_data_type_str(x):
   if isinstance(x, tuple):
@@ -73,9 +73,8 @@ class BcolzStorage(Storage):
 
   def delete(self, path):
     # NOTE: Not recursive
-    metapath = get_metapath(path)
-    try:
-      shutil.rmtree(metapath)
-      shutil.rmtree(path)
-    except FileNotFoundError:
-      pass
+    shutil.rmtree(get_metapath(path), ignore_errors=True)
+    shutil.rmtree(path, ignore_errors=True)
+
+  def cleanup(self, invalidated=True, expired=True):
+    raise NotImplementedError("cleanup() not implemented for bcolz storage")

checkpointer-2.5.0/checkpointer/storages/memory_storage.py

@@ -0,0 +1,39 @@
+from typing import Any
+from pathlib import Path
+from datetime import datetime
+from .storage import Storage
+
+item_map: dict[Path, dict[str, tuple[datetime, Any]]] = {}
+
+def get_short_path(path: Path):
+  return path.parts[-1]
+
+class MemoryStorage(Storage):
+  def get_dict(self):
+    return item_map.setdefault(self.checkpointer.root_path / self.checkpoint_fn.fn_subdir, {})
+
+  def store(self, path, data):
+    self.get_dict()[get_short_path(path)] = (datetime.now(), data)
+
+  def exists(self, path):
+    return get_short_path(path) in self.get_dict()
+
+  def checkpoint_date(self, path):
+    return self.get_dict()[get_short_path(path)][0]
+
+  def load(self, path):
+    return self.get_dict()[get_short_path(path)][1]
+
+  def delete(self, path):
+    del self.get_dict()[get_short_path(path)]
+
+  def cleanup(self, invalidated=True, expired=True):
+    curr_key = self.checkpointer.root_path / self.checkpoint_fn.fn_subdir
+    for key, calldict in list(item_map.items()):
+      if key.parent == curr_key.parent:
+        if invalidated and key != curr_key:
+          del item_map[key]
+        elif expired and self.checkpointer.should_expire:
+          for callid, (date, _) in list(calldict.items()):
+            if self.checkpointer.should_expire(date):
+              del calldict[callid]

checkpointer-2.5.0/checkpointer/storages/pickle_storage.py

@@ -0,0 +1,45 @@
+import pickle
+import shutil
+from pathlib import Path
+from datetime import datetime
+from .storage import Storage
+
+def get_path(path: Path):
+  return path.with_name(f"{path.name}.pkl")
+
+class PickleStorage(Storage):
+  def store(self, path, data):
+    full_path = get_path(path)
+    full_path.parent.mkdir(parents=True, exist_ok=True)
+    with full_path.open("wb") as file:
+      pickle.dump(data, file, -1)
+
+  def exists(self, path):
+    return get_path(path).exists()
+
+  def checkpoint_date(self, path):
+    return datetime.fromtimestamp(get_path(path).stat().st_mtime)
+
+  def load(self, path):
+    with get_path(path).open("rb") as file:
+      return pickle.load(file)
+
+  def delete(self, path):
+    get_path(path).unlink(missing_ok=True)
+
+  def cleanup(self, invalidated=True, expired=True):
+    version_path = self.checkpointer.root_path.resolve() / self.checkpoint_fn.fn_subdir
+    fn_path = version_path.parent
+    if invalidated:
+      old_dirs = [path for path in fn_path.iterdir() if path.is_dir() and path != version_path]
+      for path in old_dirs:
+        shutil.rmtree(path)
+      print(f"Removed {len(old_dirs)} invalidated directories for {self.checkpoint_fn.__qualname__}")
+    if expired and self.checkpointer.should_expire:
+      count = 0
+      for pkl_path in fn_path.rglob("*.pkl"):
+        path = pkl_path.with_suffix("")
+        if self.checkpointer.should_expire(self.checkpoint_date(path)):
+          count += 1
+          self.delete(path)
+      print(f"Removed {count} expired checkpoints for {self.checkpoint_fn.__qualname__}")