Add CUDA process checkpointing helpers

Author: kkraus14

cuda_core/cuda/core/__init__.py

@@ -28,7 +28,7 @@ def _import_versioned_module():
 del _import_versioned_module
 
 
-from cuda.core import system, utils
+from cuda.core import checkpoint, system, utils
 from cuda.core._device import Device
 from cuda.core._event import Event, EventOptions
 from cuda.core._graphics import GraphicsResource

cuda_core/cuda/core/checkpoint.py

@@ -0,0 +1,192 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+#
+# SPDX-License-Identifier: Apache-2.0
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from enum import IntEnum
+from typing import Any, Iterable
+
+
+class ProcessState(IntEnum):
+    """
+    CUDA checkpoint state for a process.
+    """
+
+    RUNNING = 0
+    LOCKED = 1
+    CHECKPOINTED = 2
+    FAILED = 3
+
+
+@dataclass(frozen=True)
+class GpuPair:
+    """
+    GPU UUID remapping pair used while restoring a checkpointed CUDA process.
+
+    Attributes
+    ----------
+    old_uuid
+        UUID of the GPU that was checkpointed.
+    new_uuid
+        UUID of the GPU to restore onto.
+    """
+
+    old_uuid: Any
+    new_uuid: Any
+
+
+@dataclass(frozen=True)
+class Process:
+    """
+    CUDA process that can be locked, checkpointed, restored, and unlocked.
+
+    Parameters
+    ----------
+    pid : int
+        Process ID of the CUDA process.
+    """
+
+    pid: int
+
+    def __post_init__(self):
+        _check_pid(self.pid)
+
+    @property
+    def state(self) -> ProcessState:
+        """
+        CUDA checkpoint state for this process.
+        """
+        driver = _get_driver()
+        state = _handle_return(driver.cuCheckpointProcessGetState(self.pid))
+        return ProcessState(int(state))
+
+    @property
+    def restore_thread_id(self) -> int:
+        """
+        CUDA restore thread ID for this process.
+        """
+        driver = _get_driver()
+        return _handle_return(driver.cuCheckpointProcessGetRestoreThreadId(self.pid))
+
+    def lock(self, timeout_ms: int = 0) -> None:
+        """
+        Lock this process, blocking further CUDA API calls.
+
+        Parameters
+        ----------
+        timeout_ms : int, optional
+            Timeout in milliseconds. A value of 0 indicates no timeout.
+        """
+        driver = _get_driver()
+        args = driver.CUcheckpointLockArgs()
+        args.timeoutMs = _check_timeout_ms(timeout_ms)
+        _handle_return(driver.cuCheckpointProcessLock(self.pid, args))
+
+    def checkpoint(self) -> None:
+        """
+        Checkpoint the GPU memory contents of this locked process.
+        """
+        driver = _get_driver()
+        _handle_return(driver.cuCheckpointProcessCheckpoint(self.pid, None))
+
+    def restore(self, gpu_pairs: Iterable[GpuPair] | None = None) -> None:
+        """
+        Restore this checkpointed process.
+
+        Parameters
+        ----------
+        gpu_pairs : iterable of GpuPair, optional
+            GPU UUID remapping pairs. If provided, the array must contain every
+            checkpointed GPU.
+        """
+        driver = _get_driver()
+        args = _make_restore_args(driver, gpu_pairs)
+        _handle_return(driver.cuCheckpointProcessRestore(self.pid, args))
+
+    def unlock(self) -> None:
+        """
+        Unlock this locked process so it can resume CUDA API calls.
+        """
+        driver = _get_driver()
+        _handle_return(driver.cuCheckpointProcessUnlock(self.pid, None))
+
+
+def _get_driver():
+    try:
+        from cuda.bindings import driver
+    except ImportError:
+        from cuda import cuda as driver
+
+    required = (
+        "cuCheckpointProcessCheckpoint",
+        "cuCheckpointProcessGetRestoreThreadId",
+        "cuCheckpointProcessGetState",
+        "cuCheckpointProcessLock",
+        "cuCheckpointProcessRestore",
+        "cuCheckpointProcessUnlock",
+        "CUcheckpointGpuPair",
+        "CUcheckpointLockArgs",
+        "CUcheckpointRestoreArgs",
+    )
+    missing = [name for name in required if not hasattr(driver, name)]
+    if missing:
+        raise RuntimeError(
+            f"CUDA checkpointing requires cuda.bindings with CUDA checkpoint API support. Missing: {', '.join(missing)}"
+        )
+    return driver
+
+
+def _handle_return(result):
+    from cuda.core._utils.cuda_utils import handle_return
+
+    return handle_return(result)
+
+
+def _check_pid(pid: int) -> int:
+    if isinstance(pid, bool) or not isinstance(pid, int):
+        raise TypeError("pid must be an int")
+    if pid <= 0:
+        raise ValueError("pid must be a positive int")
+    return pid
+
+
+def _check_timeout_ms(timeout_ms: int) -> int:
+    if isinstance(timeout_ms, bool) or not isinstance(timeout_ms, int):
+        raise TypeError("timeout_ms must be an int")
+    if timeout_ms < 0:
+        raise ValueError("timeout_ms must be >= 0")
+    return timeout_ms
+
+
+def _make_restore_args(driver, gpu_pairs: Iterable[GpuPair] | None):
+    if gpu_pairs is None:
+        return None
+
+    pairs = []
+    for gpu_pair in gpu_pairs:
+        if isinstance(gpu_pair, driver.CUcheckpointGpuPair):
+            pair = gpu_pair
+        elif isinstance(gpu_pair, GpuPair):
+            pair = driver.CUcheckpointGpuPair()
+            pair.oldUuid = gpu_pair.old_uuid
+            pair.newUuid = gpu_pair.new_uuid
+        else:
+            raise TypeError("gpu_pairs must contain GpuPair or cuda.bindings.driver.CUcheckpointGpuPair objects")
+        pairs.append(pair)
+
+    if not pairs:
+        return None
+
+    args = driver.CUcheckpointRestoreArgs()
+    args.gpuPairs = pairs
+    args.gpuPairsCount = len(pairs)
+    return args
+
+
+__all__ = [
+    "GpuPair",
+    "Process",
+    "ProcessState",
+]

cuda_core/cuda/core/system/__init__.py

@@ -1,4 +1,4 @@
-# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 #
 # SPDX-License-Identifier: Apache-2.0
 

cuda_core/docs/source/api.rst

@@ -174,6 +174,29 @@ CUDA compilation toolchain
    LinkerOptions
 
 
+CUDA process checkpointing
+--------------------------
+
+.. autosummary::
+   :toctree: generated/
+
+   :template: class.rst
+
+   checkpoint.Process
+
+.. autosummary::
+   :toctree: generated/
+
+   :template: dataclass.rst
+
+   checkpoint.GpuPair
+
+.. autosummary::
+   :toctree: generated/
+
+   checkpoint.ProcessState
+
+
 CUDA system information and NVIDIA Management Library (NVML)
 ------------------------------------------------------------
 

cuda_core/docs/source/release/1.0.0-notes.rst

@@ -16,7 +16,10 @@ Highlights
 New features
 ------------
 
-- TBD
+- Added the :mod:`cuda.core.checkpoint` module for CUDA process checkpointing,
+  including process state queries, lock/checkpoint/restore/unlock operations,
+  and GPU UUID remapping support for restore.
+  (`#1343 <https://github.com/NVIDIA/cuda-python/issues/1343>`__)
 
 
 Fixes and enhancements