Release v13.3 of the CUDA samples with CUDA 13.3 Toolkit (#435)

This is the release of the CUDA 13.3 samples, which include additions for CUDA Tile C++, and updated CCCL and Python samples.

python/4_DistributedComputing/ipcMemoryPool/README.md

@@ -68,8 +68,8 @@ round-trip test:
 - CUDA Toolkit 13.0 or newer (matches `cuda-python` 13.x)
 - Python 3.10 or newer
 - `cuda-python` (>=13.0.0)
-- `cuda-core` (>=0.6.0)
-- `cupy-cuda13x` (>=13.0.0)
+- `cuda-core` (>=1.0.0)
+- `cupy-cuda13x` (>=14.0.0)
 
 ## Installation
 
@@ -83,8 +83,8 @@ pip install -r requirements.txt
 The `requirements.txt` installs:
 
 - `cuda-python` (>=13.0.0)
-- `cuda-core` (>=0.6.0)
-- `cupy-cuda13x` (>=13.0.0)
+- `cuda-core` (>=1.0.0)
+- `cupy-cuda13x` (>=14.0.0)
 
 ## How to Run
 

python/4_DistributedComputing/ipcMemoryPool/ipcMemoryPool.py

@@ -175,7 +175,7 @@ def main() -> int:
         f"(is_ipc_enabled={mr.is_ipc_enabled})"
     )
 
-    buffer = mr.allocate(nbytes)
+    buffer = mr.allocate(nbytes, stream=device.default_stream)
     try:
         # Fill the buffer with a known pattern from the parent side.
         arr = cp.from_dlpack(buffer).view(dtype=cp.float32)

python/4_DistributedComputing/ipcMemoryPool/requirements.txt

@@ -1,4 +1,4 @@
 cuda-python>=13.0.0
-cuda-core>=0.6.0
-cupy-cuda13x>=13.0.0
+cuda-core>=1.0.0
+cupy-cuda13x>=14.0.0
 numpy>=1.24.0

python/4_DistributedComputing/multiGPUGradientAverage/README.md

@@ -8,7 +8,7 @@ This sample demonstrates gradient averaging across multiple GPUs using MPI and c
 
 - How to initialize MPI for multi-process GPU communication
 - How to map MPI ranks to CUDA devices consistently
-- How to integrate cuda.core streams with CuPy using `ExternalStream`
+- How to integrate cuda.core streams with CuPy using `Stream.from_external`
 - How to compile and launch custom CUDA kernels using cuda.core
 - How to use cuda.core Event for GPU timing measurements
 - How to use MPI Allreduce with host-staging for universal compatibility
@@ -23,12 +23,17 @@ This sample demonstrates gradient averaging across multiple GPUs using MPI and c
 ## Installation
 
 ```bash
-pip install mpi4py cupy-cuda13x cuda-python cuda-core
+pip install -r requirements.txt
 ```
 
 ## Running
 
-**IMPORTANT:** This sample **MUST** be run with `mpirun` with at least 2 processes.
+**IMPORTANT:** This sample **MUST** be launched by an MPI runtime with at
+least 2 processes. On Linux/macOS this is typically `mpirun`; on Windows with
+Microsoft MPI the launcher is `mpiexec` (and the flag for process count is
+`-n`). Either form is accepted by most MPI stacks.
+
+Linux / macOS (OpenMPI, MPICH, Intel MPI):
 
 ```bash
 # Single node (2 GPUs)
@@ -41,6 +46,14 @@ mpirun -np 4 python multiGPUGradientAverage.py --size 10000
 CUDA_VISIBLE_DEVICES=0,2 mpirun -np 2 python multiGPUGradientAverage.py
 ```
 
+Windows (Microsoft MPI — `mpiexec` is installed under
+`C:\Program Files\Microsoft MPI\Bin\` and is not on PATH by default):
+
+```powershell
+& "C:\Program Files\Microsoft MPI\Bin\mpiexec.exe" -n 2 `
+    python multiGPUGradientAverage.py --size 10000
+```
+
 ## Sample Output
 
 ```
@@ -82,11 +95,11 @@ Demo complete.
 
 ## Key Technical Details
 
-The sample uses cuda.core streams and makes CuPy use them via `ExternalStream`:
+The sample uses cuda.core streams and makes CuPy use them via `Stream.from_external`:
 
 ```python
 stream = device.create_stream()
-cp.cuda.ExternalStream(int(stream.handle)).use()
+cp.cuda.Stream.from_external(stream).use()
 ```
 
 GPU timing is measured using cuda.core Event:
@@ -107,4 +120,9 @@ The host-staging pattern transfers data GPU → CPU → MPI → CPU → GPU for
 
 **Error: "This sample requires at least 2 MPI processes!"**
 
-Solution: Run with `mpirun -np 2 python multiGPUGradientAverage.py`
+Solution:
+- Linux / macOS: `mpirun -np 2 python multiGPUGradientAverage.py`
+- Windows (Microsoft MPI): `& "C:\Program Files\Microsoft MPI\Bin\mpiexec.exe" -n 2 python multiGPUGradientAverage.py`
+  (or `mpiexec -n 2 ...` after adding `C:\Program Files\Microsoft MPI\Bin\` to `PATH`).
+
+See the **Running** section above for fully-formed examples.

python/4_DistributedComputing/multiGPUGradientAverage/multiGPUGradientAverage.py

@@ -117,7 +117,7 @@ def init_device(rank: int):
 
     # Create cuda.core stream and make CuPy use it
     stream = device.create_stream()
-    cp.cuda.ExternalStream(int(stream.handle)).use()
+    cp.cuda.Stream.from_external(stream).use()
 
     return device, stream
 
@@ -317,7 +317,7 @@ def main():
 
         # Step 1: Compute local gradients on each GPU
         # Use cuda.core Event for GPU timing measurements
-        timing_options = EventOptions(enable_timing=True)
+        timing_options = EventOptions(timing_enabled=True)
         start_event = stream.record(options=timing_options)
 
         local_grad = compute_local_gradients(num_elements, device, stream, rank)

python/4_DistributedComputing/multiGPUGradientAverage/requirements.txt

@@ -5,14 +5,14 @@ mpi4py>=3.1.4
 
 # GPU array library (NumPy-compatible arrays on CUDA)
 # Use cupy-cuda11x, cupy-cuda12x, or cupy-cuda13x depending on your CUDA version
-cupy-cuda13x>=13.0.0
+cupy-cuda13x>=14.0.0
 
 # CUDA Python bindings (low-level CUDA driver API)
 cuda-python>=13.0.0
 
 # cuda.core - Modern Python interface for CUDA
 # Provides Program, LaunchConfig, Device, and launch APIs
-cuda-core>=0.6.0
+cuda-core>=1.0.0
 
 # Note: This sample uses host-staging for MPI communication
 # Standard MPI installation is sufficient (no CUDA-aware MPI required)

python/4_DistributedComputing/simpleP2P/README.md

@@ -94,7 +94,7 @@ pip install -r requirements.txt
 
 Or manually:
 ```bash
-pip install numpy>=2.3.2 cuda-core>=0.6.0 cuda-python>=13.0.0
+pip install numpy>=2.3.2 cuda-core>=1.0.0 cuda-python>=13.0.0
 ```
 
 ## How to run

python/4_DistributedComputing/simpleP2P/requirements.txt

@@ -3,4 +3,4 @@
 
 numpy>=2.3.2
 cuda-python>=13.0.0
-cuda-core>=0.6.0
+cuda-core>=1.0.0

python/4_DistributedComputing/simpleP2P/simpleP2P.py

@@ -161,13 +161,13 @@ def run(num_elements=1024 * 1024 * 16):
     dev0.set_current()
     mr0 = DeviceMemoryResource(dev0)
     mr0.peer_accessible_by = [gpuid[1]]  # Grant GPU 1 access to GPU 0's memory
-    g0 = mr0.allocate(buf_size)
+    g0 = mr0.allocate(buf_size, stream=dev0.default_stream)
 
     # Allocate on GPU 1 and grant access to GPU 0
     dev1.set_current()
     mr1 = DeviceMemoryResource(dev1)
     mr1.peer_accessible_by = [gpuid[0]]  # Grant GPU 0 access to GPU 1's memory
-    g1 = mr1.allocate(buf_size)
+    g1 = mr1.allocate(buf_size, stream=dev1.default_stream)
 
     print(f"  Peer access enabled: GPU{gpuid[0]} <-> GPU{gpuid[1]}")
     print(
@@ -177,7 +177,7 @@ def run(num_elements=1024 * 1024 * 16):
 
     # Allocate pinned host memory
     pinned_mr = PinnedMemoryResource()
-    h0 = pinned_mr.allocate(buf_size)
+    h0 = pinned_mr.allocate(buf_size, stream=dev0.default_stream)
 
     print("  Memory allocated successfully")
 
@@ -190,7 +190,7 @@ def run(num_elements=1024 * 1024 * 16):
         print("\nMeasuring P2P bandwidth...")
         print("  Performing 100 ping-pong copies between GPUs...")
 
-        event_options = EventOptions(enable_timing=True)
+        event_options = EventOptions(timing_enabled=True)
         sync_event0 = None
         sync_event1 = None
 
@@ -206,7 +206,7 @@ def run(num_elements=1024 * 1024 * 16):
                 # Copy g0 -> g1 on stream0
                 g1.copy_from(g0, stream=stream0)
                 # Record event on stream0 to signal completion of this copy
-                sync_event0 = stream0.record(options=EventOptions(enable_timing=False))
+                sync_event0 = stream0.record(options=EventOptions(timing_enabled=False))
             else:
                 # Wait for previous stream0 copy to complete
                 if sync_event0 is not None:
@@ -214,7 +214,7 @@ def run(num_elements=1024 * 1024 * 16):
                 # Copy g1 -> g0 on stream1
                 g0.copy_from(g1, stream=stream1)
                 # Record event on stream1 to signal completion of this copy
-                sync_event1 = stream1.record(options=EventOptions(enable_timing=False))
+                sync_event1 = stream1.record(options=EventOptions(timing_enabled=False))
 
         # Wait for last stream1 copy to complete
         if sync_event1 is not None: