Python bindings to the NVIDIA Management Library

Provides a Python interface to GPU management and monitoring functions.

This is a wrapper around the NVML library. For information about the NVML library, see the NVML developer page http://developer.nvidia.com/nvidia-management-library-nvml

As of version 11.0.0, the NVML-wrappers used in pynvml are identical to those published through nvidia-ml-py.

Note that this file can be run with 'python -m doctest -v README.txt' although the results are system dependent

Requires

Python 3, or an earlier version with the ctypes module.

Installation

pip install .

Usage

You can use the lower level nvml bindings

>>> from pynvml import *
>>> nvmlInit()
>>> print("Driver Version:", nvmlSystemGetDriverVersion())
Driver Version: 410.00
>>> deviceCount = nvmlDeviceGetCount()
>>> for i in range(deviceCount):
...     handle = nvmlDeviceGetHandleByIndex(i)
...     print("Device", i, ":", nvmlDeviceGetName(handle))
...
Device 0 : Tesla V100

>>> nvmlShutdown()

Or the higher level nvidia_smi API

from pynvml.smi import nvidia_smi
nvsmi = nvidia_smi.getInstance()
nvsmi.DeviceQuery('memory.free, memory.total')

from pynvml.smi import nvidia_smi
nvsmi = nvidia_smi.getInstance()
print(nvsmi.DeviceQuery('--help-query-gpu'), end='\n')

Functions

Python methods wrap NVML functions, implemented in a C shared library. Each function's use is the same with the following exceptions:

• Instead of returning error codes, failing error codes are raised as Python exceptions.

  >>> try:
  ...     nvmlDeviceGetCount()
  ... except NVMLError as error:
  ...     print(error)
  ...
  Uninitialized
  

• C function output parameters are returned from the corresponding Python function left to right.

  nvmlReturn_t nvmlDeviceGetEccMode(nvmlDevice_t device,
                                    nvmlEnableState_t *current,
                                    nvmlEnableState_t *pending);
  

  >>> nvmlInit()
  >>> handle = nvmlDeviceGetHandleByIndex(0)
  >>> (current, pending) = nvmlDeviceGetEccMode(handle)
  

• C structs are converted into Python classes.

  nvmlReturn_t DECLDIR nvmlDeviceGetMemoryInfo(nvmlDevice_t device,
                                               nvmlMemory_t *memory);
  typedef struct nvmlMemory_st {
      unsigned long long total;
      unsigned long long free;
      unsigned long long used;
  } nvmlMemory_t;
  

  >>> info = nvmlDeviceGetMemoryInfo(handle)
  >>> print "Total memory:", info.total
  Total memory: 5636292608
  >>> print "Free memory:", info.free
  Free memory: 5578420224
  >>> print "Used memory:", info.used
  Used memory: 57872384
  

• Python handles string buffer creation.

  nvmlReturn_t nvmlSystemGetDriverVersion(char* version,
                                          unsigned int length);
  

  >>> version = nvmlSystemGetDriverVersion();
  >>> nvmlShutdown()
  

For usage information see the NVML documentation.

Variables

All meaningful NVML constants and enums are exposed in Python.

The NVML_VALUE_NOT_AVAILABLE constant is not used. Instead None is mapped to the field.

NVML Permissions

Many of the pynvml wrappers assume that the underlying NVIDIA Management Library (NVML) API can be used without admin/root privileges. However, it is certainly possible for the system permissions to prevent pynvml from querying GPU performance counters. For example:

$ nvidia-smi nvlink -g 0
GPU 0: Tesla V100-SXM2-32GB (UUID: GPU-96ab329d-7a1f-73a8-a9b7-18b4b2855f92)
NVML: Unable to get the NvLink link utilization counter control for link 0: Insufficient Permissions

A simple way to check the permissions status is to look for RmProfilingAdminOnly in the driver params file (Note that RmProfilingAdminOnly == 1 means that admin/sudo access is required):

$ cat /proc/driver/nvidia/params | grep RmProfilingAdminOnly
RmProfilingAdminOnly: 1

For more information on setting/unsetting the relevant admin privileges, see these notes on resolving ERR_NVGPUCTRPERM errors.

Release Notes

• Version 2.285.0
  • Added new functions for NVML 2.285. See NVML documentation for more information.
  • Ported to support Python 3.0 and Python 2.0 syntax.
  • Added nvidia_smi.py tool as a sample app.
• Version 3.295.0
  • Added new functions for NVML 3.295. See NVML documentation for more information.
  • Updated nvidia_smi.py tool
    • Includes additional error handling
• Version 4.304.0
  • Added new functions for NVML 4.304. See NVML documentation for more information.
  • Updated nvidia_smi.py tool
• Version 4.304.3
  • Fixing nvmlUnitGetDeviceCount bug
• Version 5.319.0
  • Added new functions for NVML 5.319. See NVML documentation for more information.
• Version 6.340.0
  • Added new functions for NVML 6.340. See NVML documentation for more information.
• Version 7.346.0
  • Added new functions for NVML 7.346. See NVML documentation for more information.
• Version 7.352.0
  • Added new functions for NVML 7.352. See NVML documentation for more information.
• Version 8.0.0
  • Refactor code to a nvidia_smi singleton class
  • Added DeviceQuery that returns a dictionary of (name, value).
  • Added filter parameters on DeviceQuery to match query api in nvidia-smi
  • Added filter parameters on XmlDeviceQuery to match query api in nvidia-smi
  • Added integer enumeration for filter strings to reduce overhead for performance monitoring.
  • Added loop(filter) method with async and callback support
• Version 8.0.1
  • Restructuring directories into two packages (pynvml and nvidia_smi)
  • Adding initial tests for both packages
  • Some name-convention cleanup in pynvml
• Version 8.0.2
  • Added NVLink function wrappers for pynvml module
• Version 8.0.3
  • Added versioneer
  • Fixed nvmlDeviceGetNvLinkUtilizationCounter bug
• Version 8.0.4
  • Added nvmlDeviceGetTotalEnergyConsumption
  • Added notes about NVML permissions
  • Fixed version-check testing
• Version 11.0.0
  • Updated nvml.py to CUDA 11
  • Updated smi.py DeviceQuery to R460
  • Aligned nvml.py with latest nvidia-ml-py deployment