dusty-nv opened a new issue, #15874:
URL: https://github.com/apache/tvm/issues/15874
Hello, I am using an NVIDIA Jetson device which has unified CPU/GPU memory,
and I'm trying to eliminate unneeded CPU<->GPU memory copies. I noticed there
are `kDLCUDAManaged` and `kDLCUDAHost` device types. But when I try to
construct a device with `kDLCUDAManaged`:
```python
import tvm
device = tvm.runtime.Device(tvm.runtime.Device.kDLCUDAManaged, 0) #
tvm.runtime.cuda(0)
print(f"device={device}, name={device.device_name},
compute={device.compute_version}, max_clocks={device.max_clock_rate},
multiprocessors={device.multi_processor_count},
max_thread_dims={device.max_thread_dimensions},
api_version={device.api_version}, driver_version={device.driver_version}")
```
```
tvm.error.InternalError: Traceback (most recent call last):
[bt] (5)
/usr/local/lib/python3.8/dist-packages/tvm/libtvm.so(TVMFuncCall+0x64)
[0xffff5d979a3c]
[bt] (4) /usr/local/lib/python3.8/dist-packages/tvm/libtvm.so(+0x32efb14)
[0xffff5d97ab14]
[bt] (3)
/usr/local/lib/python3.8/dist-packages/tvm/libtvm.so(tvm::runtime::DeviceAPIManager::GetAPI(int,
bool)+0x1ec) [0xffff5d97d114]
[bt] (2)
/usr/local/lib/python3.8/dist-packages/tvm/libtvm.so(tvm::runtime::DeviceAPIManager::GetAPI(std::__cxx11::basic_string<char,
std::char_traits<char>, std::allocator<char> >, bool)+0x30c) [0xffff5d97cd44]
[bt] (1)
/usr/local/lib/python3.8/dist-packages/tvm/libtvm.so(tvm::runtime::detail::LogFatal::Entry::Finalize()+0x78)
[0xffff5b69c3a0]
[bt] (0)
/usr/local/lib/python3.8/dist-packages/tvm/libtvm.so(tvm::runtime::Backtrace[abi:cxx11]()+0x30)
[0xffff5d9c6d50]
File "/opt/mlc-llm/3rdparty/tvm/src/runtime/c_runtime_api.cc", line 133
InternalError: Check failed: (allow_missing) is false: Device API
cuda_managed is not enabled.
```
What is the proper way to use `kDLCUDAManaged` ?
### Environment
TVM Unity 0.12
Ubuntu 20.04, JetPack 5.1.2, CUDA 11.4
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
