MCPcopy Create free account
hub / github.com/NVIDIA/cuda-samples / main

Function main

python/2_CoreConcepts/processCheckpoint/processCheckpoint.py:192–296  ·  view source on GitHub ↗
()

Source from the content-addressed store, hash-verified

190
191
192def main():
193 parser = argparse.ArgumentParser(
194 description="CUDA process checkpoint sample using cuda.core",
195 formatter_class=argparse.RawDescriptionHelpFormatter,
196 )
197 parser.add_argument(
198 "--device", type=int, default=0, help="CUDA device ID (default: 0)"
199 )
200 parser.add_argument(
201 "--buffer-mib",
202 type=int,
203 default=16,
204 help="GPU buffer size in MiB (default: 16)",
205 )
206 parser.add_argument(
207 "--lock-timeout-ms",
208 type=int,
209 default=5000,
210 help="Timeout passed to Process.lock in ms (default: 5000)",
211 )
212 args = parser.parse_args()
213
214 if sys.platform != "linux":
215 print("CUDA process checkpointing is Linux-only.")
216 return 2
217
218 if args.buffer_mib <= 0:
219 print("Error: --buffer-mib must be positive")
220 return 1
221
222 print("[Process Checkpoint Sample using CUDA Core API]")
223 print(f"PID: {os.getpid()}")
224
225 device = Device(args.device)
226 device.set_current()
227 print(f"Device: {device.name}")
228 print(f"Compute Capability: sm_{device.arch}")
229 print(f"Buffer size: {args.buffer_mib} MiB")
230 print(f"Lock timeout: {args.lock_timeout_ms} ms")
231
232 # CUDA process checkpointing relies on kernel-mode driver features
233 # that aren't shipped on integrated-GPU platforms (e.g. Tegra /
234 # Jetson / Thor). There Process.lock() still succeeds and only
235 # Process.checkpoint() reports the missing support, which leaves the
236 # process locked and hanging on exit, so the check below cannot cover
237 # it. Remove this guard once integrated platforms gain checkpoint
238 # support.
239 if device.properties.integrated:
240 print(
241 f"CUDA process checkpointing is not supported on integrated "
242 f"GPUs (sm_{device.arch}), waiving this sample."
243 )
244 return 2
245
246 print()
247 print("Compiling kernel ...")
248 fill_kernel = compile_fill_kernel(device)
249

Callers 1

Calls 5

compile_fill_kernelFunction · 0.85
hash_device_bufferFunction · 0.85
run_lifecycleFunction · 0.85
print_timingsFunction · 0.85
emptyMethod · 0.45

Tested by

no test coverage detected