Post v13.5 (#470)

This commit is contained in:
Dheemanth
2026-11-04 12:48:54 -08:00
committed by GitHub
parent 5443602d89
commit 0ec046d5e9
53 changed files with 1750 additions and 183 deletions

View File

@@ -23,7 +23,7 @@ Query and display detailed properties of all CUDA-capable devices in your system
- `Device.get_all_devices()` - Get tuple of all available Device instances
- `Device(device_id)` - Get Device object for specific device ID
- `system.get_driver_version()` - Query CUDA driver version
- `system.get_user_mode_driver_version()` - Query CUDA driver version
- `Device.set_current()` - Set the current device for API calls
- `Device.properties` - Access comprehensive device properties
- `Device.name` - Get device name string
@@ -90,7 +90,7 @@ Query and display detailed properties of all CUDA-capable devices in your system
- CUDA Toolkit 13.0 or newer (recommended; matches `cuda-python` 13.x)
- Python 3.10 or newer
- `cuda-python` package (>=13.0.0)
- `cuda-core` package (>=1.0.0)
- `cuda-core` package (>=1.2.0)
## Installation
@@ -103,7 +103,7 @@ pip install -r requirements.txt
The requirements.txt installs:
- `cuda-python` (>=13.0.0)
- `cuda-core` (>=1.0.0)
- `cuda-core` (>=1.2.0)
## How to Run

View File

@@ -87,7 +87,12 @@ def print_driver_info() -> None:
"NVML bindings are not compatible with this driver; "
"device info will be limited."
)
print(f"Current process: {system.get_process_name(os.getpid())}")
try:
# On WSL, NVML can only name a PID that already holds a CUDA
# context, which this process doesn't have yet.
print(f"Current process: {system.get_process_name(os.getpid())}")
except system.NotFoundError as e:
print(f"Current process: unavailable ({e})")
def print_device_info(device: "system.Device") -> None:

View File

@@ -147,14 +147,27 @@ def run_lifecycle(proc: checkpoint.Process, lock_timeout_ms: int) -> List[StepTi
raise CheckpointUnsupportedError from exc
timings.append(StepTiming("lock", ms, proc.state))
ms = _time_call(proc.checkpoint)
timings.append(StepTiming("checkpoint", ms, proc.state))
# Always unlock, even on failure, or the process stays permanently
# "locked" and later cleanup calls deadlock against it. Keep the
# original error primary if unlock() also fails.
op_error: BaseException | None = None
try:
ms = _time_call(proc.checkpoint)
timings.append(StepTiming("checkpoint", ms, proc.state))
ms = _time_call(proc.restore)
timings.append(StepTiming("restore", ms, proc.state))
ms = _time_call(proc.restore)
timings.append(StepTiming("restore", ms, proc.state))
except BaseException as exc:
op_error = exc
ms = _time_call(proc.unlock)
timings.append(StepTiming("unlock", ms, proc.state))
try:
ms = _time_call(proc.unlock)
timings.append(StepTiming("unlock", ms, proc.state))
except Exception:
if op_error is None:
raise
if op_error is not None:
raise op_error
return timings

View File

@@ -13,6 +13,8 @@ This sample demonstrates how to add a custom GPU operation to PyTorch using the
- cuda-python >= 13.0.0
- cuda-core >=1.0.0
**PyTorch GPU compatibility:** Install a CUDA-enabled PyTorch build that supports your target GPU. GPU architecture support varies by platform and CUDA build. Check the [PyTorch CUDA Support Matrix](https://github.com/pytorch/pytorch/blob/main/RELEASE.md#pytorch-cuda-support-matrix) before installing PyTorch.
## Installation
```bash

View File

@@ -95,9 +95,9 @@ def check_ipc_support(device) -> bool:
return True
def child_worker(q_in, q_out, n_elements, parent_seed, child_seed):
def child_worker(q_in, q_out, n_elements, parent_seed, child_seed, device_id):
"""Runs in a separate process. Verifies and modifies the shared buffer."""
device = Device(0)
device = Device(device_id)
device.set_current()
pid = mp.current_process().pid
@@ -188,7 +188,7 @@ def main() -> int:
q_from_child = mp.Queue()
child = mp.Process(
target=child_worker,
args=(q_to_child, q_from_child, N, parent_seed, child_seed),
args=(q_to_child, q_from_child, N, parent_seed, child_seed, args.device),
)
child.start()
q_to_child.put(buffer)

View File

@@ -6,7 +6,7 @@ This sample demonstrates peer-to-peer (P2P) memory access between multiple GPUs
## What you will learn
- How to detect multiple CUDA-capable GPUs using `system.get_num_devices()` and `Device(id)`
- How to detect multiple CUDA-capable GPUs using `Device.get_all_devices()`
- How to check P2P capability between GPU pairs using `device.can_access_peer()`
- How to enable and disable peer access using `DeviceMemoryResource.peer_accessible_by`
- How to allocate device memory on specific GPUs using `DeviceMemoryResource`
@@ -26,9 +26,7 @@ This sample demonstrates peer-to-peer (P2P) memory access between multiple GPUs
## Key APIs
**From cuda.core:**
- `system` – Pre-instantiated singleton for system-level CUDA information
- `system.get_num_devices()` – Get number of CUDA-capable devices
- `Device(id)` – Get specific CUDA device handle
- `Device.get_all_devices()` – Get handles for all CUDA-visible devices
- `device.can_access_peer(peer)` – Check if this device can access peer device memory
- `device.set_current()` – Set active device for subsequent operations
- `device.create_stream()` – Create CUDA stream for kernel execution
@@ -94,7 +92,7 @@ pip install -r requirements.txt
Or manually:
```bash
pip install numpy>=2.3.2 cuda-core>=1.0.0 cuda-python>=13.0.0
pip install numpy>=2.3.2 cuda-core>=1.2.0 cuda-python>=13.0.0
```
## How to run

View File

@@ -39,7 +39,6 @@ try:
Program,
ProgramOptions,
launch,
system,
)
except ImportError as e:
print(f"Error: Required package not found: {e}")
@@ -91,7 +90,8 @@ def run(num_elements=1024 * 1024 * 16):
# Check for multiple GPUs
print("\nChecking for multiple GPUs...")
num_devices = system.get_num_devices()
devices = Device.get_all_devices()
num_devices = len(devices)
print(f"CUDA-capable device count: {num_devices}")
if num_devices < 2:
@@ -101,9 +101,6 @@ def run(num_elements=1024 * 1024 * 16):
)
return 2
# Get device properties
devices = [Device(i) for i in range(num_devices)]
# Check for P2P capability
print("\nChecking GPU(s) for support of peer to peer memory access...")