Advanced Configuration
The default settings for Intel® Extension for PyTorch* are sufficient for most use cases. However, if users want to customize Intel® Extension for PyTorch*, advanced configuration is available at build time and runtime.
Build Time Configuration
The following build options are supported by Intel® Extension for PyTorch*. Users who install Intel® Extension for PyTorch* via source compilation could override the default configuration by explicitly setting a build option ON or OFF, and then build.
| Build Option | Default Value |
Description |
|---|---|---|
| USE_ONEMKL | ON | Use oneMKL BLAS |
| USE_CHANNELS_LAST_1D | ON | Use channels last 1d |
| USE_PERSIST_STREAM | ON | Use persistent oneDNN stream |
| USE_SCRATCHPAD_MODE | ON | Use oneDNN scratchpad mode |
| USE_PRIMITIVE_CACHE | ON | Cache oneDNN primitives by FRAMEWORK for specific operators |
| USE_QUEUE_BARRIER | ON | Use queue submit_barrier, otherwise use dummy kernel |
| USE_PTI | ON | Build XPU Profiler with PTI support. |
| USE_DS_KERNELS | ON | Build deepspeed kernels |
| USE_SYCL_ASSERT | OFF | Enables assert in sycl kernel |
| USE_ITT_ANNOTATION | OFF | Enables ITT annotation in sycl kernel |
| USE_SPLIT_FP64_LOOPS | ON | Split FP64 loops into separate kernel for element-wise kernels |
| BUILD_BY_PER_KERNEL | OFF | Build by DPC++ per_kernel option (exclusive with USE_AOT_DEVLIST) |
| BUILD_INTERNAL_DEBUG | OFF | Use internal debug code path |
| BUILD_SEPARATE_OPS | OFF | Build each operator in separate library |
| BUILD_SIMPLE_TRACE | ON | Build simple trace for each registered operator |
| USE_AOT_DEVLIST | "" | Set device list for AOT build |
| USE_XETLA | "ON" | Use XeTLA based customer kernels; Specify a comma-sep list of gpu architectures (e.g. xe_lpg,xe_hpg) to only enable kernels for specific platforms |
| USE_ONEDNN_DIR | "" | Specify oneDNN source path which contains its include directory and lib directory |
| USE_XETLA_SRC | "${IPEX_GPU_ROOT_DIR}/aten/operators/xetla/kernels/" | Specify XETLA source path which contains its include dir |
| BUILD_OPT_LEVEL | "" | Add build option -Ox, accept values: 0/1 |
| BUILD_WITH_SANITIZER | "" | Build with sanitizer check. Support one of address, thread, and leak options at a time. The default option is address. |
For above build options which can be configured to ON or OFF, users can configure them to 1 or 0 also, while ON equals to 1 and OFF equals to 0.
Runtime Configuration
The following launch options are supported in Intel® Extension for PyTorch*. Users who execute AI models on XPU could override the default configuration by explicitly setting the option value at runtime using environment variables, and then launch the execution.
| Launch Option CPU, GPU |
Default Value |
Description |
|---|---|---|
| IPEX_FP32_MATH_MODE | FP32 | Set values for FP32 math mode (valid values: FP32, TF32, BF32). Refer to API Documentation for details. |
| Launch Option GPU ONLY |
Default Value |
Description |
|---|---|---|
| IPEX_VERBOSE | 0 | Set verbose level with synchronization execution mode, will be deprecated very soon. Please use IPEX_LOG_LEVEL instead. |
| IPEX_XPU_SYNC_MODE | 0 | Set 1 to enforce synchronization execution mode, will be deprecated very soon. |
| IPEX_LOG_LEVEL | -1 | Set log level to trace the execution and get log information, pls refer to 'ipex_log.md' for different log level. |
| IPEX_LOG_COMPONENT | "ALL" | Set IPEX_LOG_COMPONENT = ALL to log all component message. Use ';' as separator to log more than one components, such as "OPS;RUNTIME". Use '/' as separator to log subcomponents. |
| IPEX_LOG_ROTATE_SIZE | -1 | Set Rotate file size in MB for IPEX_LOG, less than 0 means unuse this setting. |
| IPEX_LOG_SPLIT_SIZE | -1 | Set split file size in MB for IPEX_LOG, less than 0 means unuse this setting. |
| IPEX_LOG_OUTPUT | "" | Set output file path for IPEX_LOG, default is null |
| Launch Option Experimental |
Default Value |
Description |
|---|---|---|
| IPEX_SIMPLE_TRACE | 0 | Set 1 to enable simple trace for all operators*, will be deprecated very soon. Please use IPEX_LOG_LEVEL instead. |
| Distributed Option GPU ONLY |
Default Value |
Description |
|---|---|---|
| TORCH_LLM_ALLREDUCE | 0 | This is a prototype feature to provide better scale-up performance by enabling optimized collective algorithms in oneCCL and asynchronous execution in torch-ccl. This feature requires XeLink enabled for cross-cards communication. By default, this feature is not enabled with setting 0. |
| CCL_BLOCKING_WAIT | 0 | This is a prototype feature to control over whether collectives execution on XPU is host blocking or non-blocking. By default, setting 0 enables blocking behavior. |
| CCL_SAME_STREAM | 0 | This is a prototype feature to allow using a computation stream as communication stream to minimize overhead for streams synchronization. By default, setting 0 uses separate streams for communication. |
For above launch options which can be configured to 1 or 0, users can configure them to ON or OFF also, while ON equals to 1 and OFF equals to 0.
Examples to configure the launch options:
Set one or more options before running the model
export IPEX_LOG_LEVEL=1
export IPEX_FP32_MATH_MODE=TF32
...
python ResNet50.py
Set one option when running the model
IPEX_LOG_LEVEL=1 python ResNet50.py
Set more than one options when running the model
IPEX_LOG_LEVEL=1 IPEX_FP32_MATH_MODE=TF32 python ResNet50.py