Release Notes#
The CUPTI Python Release Notes include new features, important bug fixes, known issues, supported platforms, and supported GPUs.
Updates in CUPTI Python Releases#
Updates in CUPTI Python 13.4.0#
Refer CUPTI C 13.4.0 release notes.
New Features
Added support for function parameter structures for CUDA Runtime, CUDA Driver and NVTX APIs. These allow for convenient consumption of
CallbackData.function_paramsandNvtxData.function_paramsin CUPTI callbacks. See Function Parameter Classes for CUDA and NVTX APIs.Added support for GPU clock control. See Clock Control.
Added support for API
cupti.cupti.get_cuda_event_id().New activity classes:
cupti.cupti.ActivityGreenContext3,cupti.cupti.ActivityKernel13,cupti.cupti.ActivityMemcpy7,cupti.cupti.ActivityMemset5.
API Changes
Renamed classes and enums to follow PEP 8 naming conventions. The old names remain available as aliases for backwards compatibility.
Old Name
New Name
driver_api_trace_cbidruntime_api_trace_cbidDriver_api_trace_cbidRuntime_api_trace_cbidNvtx_api_trace_cbidPmSampling_TriggerModePmSampling_DecodeStopReasonPmSampling_HardwareBuffer_AppendModeDevice_GetChipName_ParamsPmSampling_CounterDataImage_Initialize_ParamsPmSampling_CounterData_GetSampleInfo_ParamsPmSampling_DecodeData_ParamsPmSampling_Disable_ParamsPmSampling_Enable_ParamsPmSampling_GetCounterAvailability_ParamsPmSampling_GetCounterDataInfo_ParamsPmSampling_GetCounterDataSize_ParamsPmSampling_SetConfig_ParamsPmSampling_Start_ParamsPmSampling_Stop_ParamsProfiler_DeInitialize_ParamsProfiler_Host_ConfigAddMetrics_ParamsProfiler_Host_Deinitialize_ParamsProfiler_Host_EvaluateToGpuValues_ParamsProfiler_Host_GetBaseMetrics_ParamsProfiler_Host_GetConfigImageSize_ParamsProfiler_Host_GetConfigImage_ParamsProfiler_Host_GetMaxNumHardwareMetricsPerPass_ParamsProfiler_Host_GetMetricProperties_ParamsProfiler_Host_GetMetricsInSinglePassSet_ParamsProfiler_Host_GetNumOfPasses_ParamsProfiler_Host_GetRangeName_ParamsProfiler_Host_GetSinglePassSets_ParamsProfiler_Host_GetSubMetrics_ParamsProfiler_Host_GetSupportedChips_ParamsProfiler_Host_Initialize_ParamsProfiler_Initialize_ParamsIn
cupti.pm_sampling,HardwareBuffer_AppendModehas been renamed toHardwareBufferAppendMode. The old name remains available as an alias.
Deprecated features
Deprecated activity classes:
cupti.cupti.ActivityGreenContext2,cupti.cupti.ActivityKernel12,cupti.cupti.ActivityMemcpy6,cupti.cupti.ActivityMemset4.
Updates in CUPTI Python 13.3.1#
Refer CUPTI C 13.3.1 release notes.
Backwards Compatibility Fixes
Restored activity classes that were previously dropped, now marked as deprecated for backwards compatibility.
cupti.cupti.ActivitySynchronization,cupti.cupti.ActivityCudaEvent,cupti.cupti.ActivityUnifiedMemoryCounter2,cupti.cupti.ActivityMemcpy5(dropped in 12.8.0)cupti.cupti.ActivityKernel9,cupti.cupti.ActivityMemoryPool2(dropped in 13.0.0)cupti.cupti.ActivityKernel10,cupti.cupti.ActivityContext3,cupti.cupti.ActivityMarkerData,cupti.cupti.ActivityDevice5(dropped in 13.2.0)cupti.cupti.ActivityKernel11,cupti.cupti.ActivityGreenContext(dropped in 13.3.0)
Moving forward, CUPTI Python will be consistent with CUPTI C and also include types moved to
cupti_activity_deprecated.h.
Updates in CUPTI Python 13.3.0#
Refer CUPTI C 13.3.0 release notes.
New Features
Added support for CUPTI’s PM Sampling API: 1:1 bindings and a higher-level pythonic API in
cupti.pm_sampling.Added support for CUPTI’s Profiler Host API: 1:1 bindings and a higher-level pythonic API in
cupti.profiler_host.CUPTI Python now uses cuda-pathfinder to locate and load
libcupti.so. For more details, see Locating the CUPTI Library.New activity classes:
cupti.cupti.ActivityKernel12,cupti.cupti.ActivityGreenContext2.Added support for API
cupti.cupti.is_tracing_session_running().
Deprecated and dropped features
Dropped activity classes:
cupti.cupti.ActivityKernel11,cupti.cupti.ActivityGreenContext.
Updates in CUPTI Python 13.2.0#
Refer CUPTI C 13.2.0 release notes.
New Features
The
cupti.cupti.CallbackDomain.NVTXcallback domain is now supported.The
cupti.cupti.CallbackData.function_paramsfield ofcupti.cupti.CallbackDatais now supported.New Activity kinds and classes:
cupti.cupti.ActivityKind.GRAPH_HOST_NODEwithcupti.cupti.ActivityGraphHostNodecupti.cupti.ActivityKind.COMPUTE_ENGINE_CTX_SWITCHwithcupti.cupti.ActivityComputeEngineCtxSwitchcupti.cupti.ActivityKind.GREEN_CONTEXTwithcupti.cupti.ActivityGreenContextcupti.cupti.ActivityKind.HOST_LAUNCHwithcupti.cupti.ActivityHostLaunch
New activity classes:
cupti.cupti.ActivityKernel11,cupti.cupti.ActivityContext4,cupti.cupti.ActivityMarkerData2,cupti.cupti.ActivityDevice6Support for Python 3.14 has been added.
Deprecated and dropped features
Dropped activity classes:
cupti.cupti.ActivityKernel10,cupti.cupti.ActivityContext3,cupti.cupti.ActivityMarkerData,cupti.cupti.ActivityDevice5Support for Python 3.9 has been dropped.
Updates in CUPTI Python 13.0.1#
Refer CUPTI C 13.0.1 release notes.
New Features
Relaxed nvidia-cuda-cupti dependency:
cupti-pythonnow supportsnvidia-cuda-cuptiversions>=13.0.0and<14.Relaxed cuda-bindings dependency:
cupti-pythonnow supportscuda-bindingsversions>=13.0.0and<14.
Updates in CUPTI Python 13.0.0#
Refer CUPTI C 13.0.0 release notes.
New Features
New Activity Kind
cupti.cupti.ActivityKind.ROTATIONand its corresponding activity classcupti.cupti.ActivityConfidentialComputeRotationhas been added.Support for Activity Kind
cupti.cupti.ActivityKind.RUNTIMEand Callback Domaincupti.cupti.CallbackDomain.RUNTIME_APIhave been added.New activity classes
cupti.cupti.ActivityKernel10andcupti.cupti.ActivityMemoryPool3have been introduced.Support for APIs
cupti.cupti.subscribe_v2(),cupti.cupti.activity_enable_cuda_event_device_timestamps()andcupti.cupti.activity_enable_all_sync_records()has been added.Added
cupti.pyistub file to improve IDE support, including type hints and auto-completion.Support for Python 3.13 has been added.
Support for Linux (aarch64 sbsa) architecture has been added.
Deprecated and dropped features
Support for activity classes
cupti.cupti.ActivityKernel9andcupti.cupti.ActivityMemoryPool2has been removed.
Updates in CUPTI Python 12.8.0#
Refer CUPTI C 12.8.0 release notes.
New Features
New Activity Kind
cupti.cupti.ActivityKind.MEM_DECOMPRESSand its corresponding activity classcupti.cupti.ActivityMemDecompresshas been added.New Activity Kind
cupti.cupti.ActivityKind.DEVICE_GRAPH_TRACEand its corresponding activity classcupti.cupti.ActivityDeviceGraphTracehas been added.New activity classes
cupti.cupti.ActivityMemcpy6,cupti.cupti.ActivityUnifiedMemoryCounter3,cupti.cupti.ActivityCudaEvent2andcupti.cupti.ActivitySynchronization2have been introduced.Support for API
cupti.cupti.activity_enable_hw_trace()has been added.
Deprecated and dropped features
Support for activity classes
cupti.cupti.ActivityMemcpy5,cupti.cupti.ActivityUnifiedMemoryCounter2,cupti.cupti.ActivityCudaEventandcupti.cupti.ActivitySynchronizationhas been removed.
Updates in CUPTI Python 12.6.0#
This was the first CUPTI Python release. Refer CUPTI C 12.6.0 release notes.
Limitations#
Known Issues#
Libraries like NCCL that emit NVTX annotations with extended payloads can now be decoded using CUPTI; however, this capability is not yet supported in CUPTI Python.
Known Limitations#
CUPTI Python cannot collect tracing data when the Python interpreter is tearing down or is not in a proper state. The callbacks registered by functions
cupti.cupti.activity_register_callbacks(),cupti.cupti.activity_register_timestamp_callback(),cupti.cupti.subscribe(), andcupti.cupti.subscribe_v2()will not be dispatched in such cases. For example, CUDA API calls that occur in a class’s__del__method will not be traced.The following CUDA Runtime API function parameter structures are not supported by CUPTI Python:
cudaMalloc3D_v3020_paramscudaMalloc3DArray_v3020_paramscudaMemset3D_v3020_paramscudaMemset3DAsync_v3020_paramscudaMallocMipmappedArray_v5000_paramscudaMemset3D_ptds_v7000_paramscudaMemset3DAsync_ptsz_v7000_paramscudaMemAdvise_v12020_paramscudaMemPrefetchAsync_v12020_paramscudaMemPrefetchAsync_ptsz_v12020_params
CUPTI Python does not support the following CUDA Driver API function parameter class members:
Python Class
Member
handlehandlehandlelocationlocationlocation
Support#
Information on supported python versions, platforms, and GPUs.
Python Support#
Python 3.10 - 3.14
Platform Support#
Linux (x86_64)
Linux (aarch64 sbsa)
GPU Support#
Turing and later GPU architectures, i.e. devices with compute capability 7.5 and higher.