def is_device_version_less_than(min_version=(7, 0)):
"""
Returns True if the version of CUDA being used is less than min_version
"""
major_version = getDeviceAttribute(
cudaDeviceAttr.cudaDevAttrComputeCapabilityMajor, 0)
minor_version = getDeviceAttribute(
cudaDeviceAttr.cudaDevAttrComputeCapabilityMinor, 0)
if major_version > min_version[0]:
return False
if major_version < min_version[0]:
return True
if minor_version < min_version[1]:
return True
return False
def validate_setup():
import os
# TODO: Remove the following check once we arrive at a solution for #4827
# This is a temporary workaround to unblock internal testing
# related issue: https://github.com/rapidsai/cudf/issues/4827
if (
"RAPIDS_NO_INITIALIZE" in os.environ
or "CUDF_NO_INITIALIZE" in os.environ
):
return
import warnings
from rmm._cuda.gpu import (
CUDARuntimeError,
cudaDeviceAttr,
cudaError,
deviceGetName,
driverGetVersion,
getDeviceAttribute,
getDeviceCount,
runtimeGetVersion,
)
def _try_get_old_or_new_symbols():
try:
# CUDA 10.2+ symbols
return [
cudaError.cudaErrorDeviceUninitialized,
cudaError.cudaErrorTimeout,
]
except AttributeError:
# CUDA 10.1 symbols
return [cudaError.cudaErrorDeviceUninitilialized]
notify_caller_errors = {
cudaError.cudaErrorInitializationError,
cudaError.cudaErrorInsufficientDriver,
cudaError.cudaErrorInvalidDeviceFunction,
cudaError.cudaErrorInvalidDevice,
cudaError.cudaErrorStartupFailure,
cudaError.cudaErrorInvalidKernelImage,
cudaError.cudaErrorAlreadyAcquired,
cudaError.cudaErrorOperatingSystem,
cudaError.cudaErrorNotPermitted,
cudaError.cudaErrorNotSupported,
cudaError.cudaErrorSystemNotReady,
cudaError.cudaErrorSystemDriverMismatch,
cudaError.cudaErrorCompatNotSupportedOnDevice,
*_try_get_old_or_new_symbols(),
cudaError.cudaErrorUnknown,
cudaError.cudaErrorApiFailureBase,
}
try:
gpus_count = getDeviceCount()
except CUDARuntimeError as e:
if e.status in notify_caller_errors:
raise e
# If there is no GPU detected, set `gpus_count` to -1
gpus_count = -1
if gpus_count > 0:
# Cupy throws RunTimeException to get GPU count,
# hence obtaining GPU count by in-house cpp api above
# 75 - Indicates to get "cudaDevAttrComputeCapabilityMajor" attribute
# 0 - Get GPU 0
major_version = getDeviceAttribute(
cudaDeviceAttr.cudaDevAttrComputeCapabilityMajor, 0
)
if major_version >= 6:
# You have a GPU with NVIDIA Pascal™ architecture or better
# Hardware Generation Compute Capability
# Turing 7.5
# Volta 7.x
# Pascal 6.x
# Maxwell 5.x
# Kepler 3.x
# Fermi 2.x
pass
else:
device_name = deviceGetName(0)
minor_version = getDeviceAttribute(
cudaDeviceAttr.cudaDevAttrComputeCapabilityMinor, 0
)
warnings.warn(
f"You will need a GPU with NVIDIA Pascal™ or "
f"newer architecture"
f"\nDetected GPU 0: {device_name} \n"
f"Detected Compute Capability: "
f"{major_version}.{minor_version}"
)
cuda_runtime_version = runtimeGetVersion()
if cuda_runtime_version >= 10000:
# CUDA Runtime Version Check: Runtime version is greater than 10000
pass
else:
from cudf.errors import UnSupportedCUDAError
minor_version = cuda_runtime_version % 100
major_version = (cuda_runtime_version - minor_version) // 1000
raise UnSupportedCUDAError(
f"Detected CUDA Runtime version is "
f"{major_version}.{str(minor_version)[0]}"
f"Please update your CUDA Runtime to 10.0 or above"
)
cuda_driver_supported_rt_version = driverGetVersion()
# Though Yes, Externally driver version is represented like `418.39`
# and cuda runtime version like `10.1`. It is not the similar case
# at cuda api's level. Coming down to APIs they follow a uniform
# convention of an integer which corresponds to the versioning
# like (1000 major + 10 minor) for 10.1 Driver version API doesn't
# actually indicate driver version, it indicates only the latest
# CUDA version supported by the driver.
# For reference :
# https://docs.nvidia.com/deploy/cuda-compatibility/index.html
if cuda_driver_supported_rt_version == 0:
from cudf.errors import UnSupportedCUDAError
raise UnSupportedCUDAError(
"We couldn't detect the GPU driver "
"properly. Please follow the linux installation guide to "
"ensure your driver is properly installed "
": https://docs.nvidia.com/cuda/cuda-installation-guide-linux/"
)
elif cuda_driver_supported_rt_version >= cuda_runtime_version:
# CUDA Driver Version Check:
# Driver Runtime version is >= Runtime version
pass
else:
from cudf.errors import UnSupportedCUDAError
raise UnSupportedCUDAError(
f"Please update your NVIDIA GPU Driver to support CUDA "
f"Runtime.\n"
f"Detected CUDA Runtime version : {cuda_runtime_version}"
f"\n"
f"Latest version of CUDA supported by current "
f"NVIDIA GPU Driver : {cuda_driver_supported_rt_version}"
)
else:
warnings.warn("No NVIDIA GPU detected")
def validate_setup():
import os
# TODO: Remove the following check once we arrive at a solution for #4827
# This is a temporary workaround to unblock internal testing
# related issue: https://github.com/rapidsai/cudf/issues/4827
if ("RAPIDS_NO_INITIALIZE" in os.environ
or "CUDF_NO_INITIALIZE" in os.environ):
return
import warnings
from cuda.cudart import cudaDeviceAttr, cudaError_t
from rmm._cuda.gpu import (
CUDARuntimeError,
deviceGetName,
driverGetVersion,
getDeviceAttribute,
getDeviceCount,
runtimeGetVersion,
)
from cudf.errors import UnsupportedCUDAError
notify_caller_errors = {
cudaError_t.cudaErrorInitializationError,
cudaError_t.cudaErrorInsufficientDriver,
cudaError_t.cudaErrorInvalidDeviceFunction,
cudaError_t.cudaErrorInvalidDevice,
cudaError_t.cudaErrorStartupFailure,
cudaError_t.cudaErrorInvalidKernelImage,
cudaError_t.cudaErrorAlreadyAcquired,
cudaError_t.cudaErrorOperatingSystem,
cudaError_t.cudaErrorNotPermitted,
cudaError_t.cudaErrorNotSupported,
cudaError_t.cudaErrorSystemNotReady,
cudaError_t.cudaErrorSystemDriverMismatch,
cudaError_t.cudaErrorCompatNotSupportedOnDevice,
cudaError_t.cudaErrorDeviceUninitialized,
cudaError_t.cudaErrorTimeout,
cudaError_t.cudaErrorUnknown,
cudaError_t.cudaErrorApiFailureBase,
}
try:
gpus_count = getDeviceCount()
except CUDARuntimeError as e:
if e.status in notify_caller_errors:
raise e
# If there is no GPU detected, set `gpus_count` to -1
gpus_count = -1
if gpus_count > 0:
# Cupy throws RunTimeException to get GPU count,
# hence obtaining GPU count by in-house cpp api above
major_version = getDeviceAttribute(
cudaDeviceAttr.cudaDevAttrComputeCapabilityMajor, 0)
if major_version < 6:
# A GPU with NVIDIA Pascal™ architecture or newer is required.
# Reference: https://developer.nvidia.com/cuda-gpus
# Hardware Generation Compute Capability
# Ampere 8.x
# Turing 7.5
# Volta 7.0, 7.2
# Pascal 6.x
# Maxwell 5.x
# Kepler 3.x
# Fermi 2.x
device_name = deviceGetName(0)
minor_version = getDeviceAttribute(
cudaDeviceAttr.cudaDevAttrComputeCapabilityMinor, 0)
warnings.warn(
"A GPU with NVIDIA Pascal™ (Compute Capability 6.0) "
"or newer architecture is required.\n"
f"Detected GPU 0: {device_name}\n"
f"Detected Compute Capability: {major_version}.{minor_version}"
)
cuda_runtime_version = runtimeGetVersion()
if cuda_runtime_version < 11000:
# Require CUDA Runtime version 11.0 or greater.
major_version = cuda_runtime_version // 1000
minor_version = (cuda_runtime_version % 1000) // 10
raise UnsupportedCUDAError(
"Detected CUDA Runtime version is "
f"{major_version}.{minor_version}. "
"Please update your CUDA Runtime to 11.0 or above.")
cuda_driver_supported_rt_version = driverGetVersion()
# Though Yes, Externally driver version is represented like `418.39`
# and cuda runtime version like `10.1`. It is not the similar case
# at cuda api's level. Coming down to APIs they follow a uniform
# convention of an integer which corresponds to the versioning
# like (1000 major + 10 minor) for 10.1 Driver version API doesn't
# actually indicate driver version, it indicates only the latest
# CUDA version supported by the driver.
# For reference :
# https://docs.nvidia.com/deploy/cuda-compatibility/index.html
if cuda_driver_supported_rt_version == 0:
raise UnsupportedCUDAError(
"We couldn't detect the GPU driver properly. Please follow "
"the installation guide to ensure your driver is properly "
"installed: "
"https://docs.nvidia.com/cuda/cuda-installation-guide-linux/")
elif cuda_driver_supported_rt_version >= cuda_runtime_version:
# CUDA Driver Version Check:
# Driver Runtime version is >= Runtime version
pass
elif (cuda_driver_supported_rt_version >= 11000
and cuda_runtime_version >= 11000):
# With cuda enhanced compatibility any code compiled
# with 11.x version of cuda can now run on any
# driver >= 450.80.02. 11000 is the minimum cuda
# version 450.80.02 supports.
pass
else:
raise UnsupportedCUDAError(
"Please update your NVIDIA GPU Driver to support CUDA "
"Runtime.\n"
f"Detected CUDA Runtime version : {cuda_runtime_version}\n"
"Latest version of CUDA supported by current "
f"NVIDIA GPU Driver : {cuda_driver_supported_rt_version}")
else:
warnings.warn("No NVIDIA GPU detected")