Appendix

Public Network Addresses Included in the Software

The websites in the Agent SDK installation package will be cleared and will not be accessed after installation, posing no security risks.

The SDK itself does not access the public URLs and email addresses in this manual. Therefore, they do not pose any risk.

For more public network addresses, see AgentSDK Public Network Addresses.xlsx.

Environment Variable Usage

The Agent SDK may use the following environment variables during runtime.

Environment Variable	Description
LOCAL_RANK	A torch distributed training setting that describes the rank of the current thread on the current node. The value range is [0, 8).
RANK	A torch distributed training setting that describes the rank of the current thread across all nodes. The value range is [0, 8).

When Agent SDK starts, it validates environment variables against a whitelist. Only the following environment variables are retained.

Environment Variable	Description
ASCEND_WORK_PATH	Location for files generated during unified CANN runtime.
ASCEND_AICPU_PATH	Installation path of AICPU in ascend-toolkit.
ASCEND_HOME_PATH	Same as ASCEND_TOOLKIT_HOME. This is the file storage path after you install the CANN Toolkit.
ASCEND_OPP_PATH	Root directory of the operator library.
ASCEND_TOOLKIT_HOME	Path for storing files after the CANN Toolkit software package is installed.
ASDOPS_LOG_LEVEL	Operator library log level.
ASDOPS_LOG_PATH	Path for storing operator library logs.
ASDOPS_LOG_TO_BOOST_TYPE	Name of the acceleration library log directory.
ASDOPS_LOG_TO_FILE	Specifies whether to output operator library logs to a file.
ASDOPS_LOG_TO_FILE_FLUSH	Specifies whether to flush the log file after writes.
ASDOPS_LOG_TO_STDOUT	Specifies whether to output operator library logs to the console.
ATB_COMPARE_TILING_EVERY_KERNEL	After each kernel runs, the system compares the tiling content on the NPU before and after the running to check whether tiling memory corruption occurs.
ATB_DEVICE_TILING_BUFFER_BLOCK_NUM	Number of DeviceTilingBuffer blocks in the context. The number is related to the maximum number of parallel operators. Generally, the default value is used should not be changed unless necessary.
ATB_HOME_PATH	Path for storing files after the nnal package is installed.
ATB_HOST_TILING_BUFFER_BLOCK_NUM	Number of HostTilingBuffer blocks in the context. The number is related to the maximum number of parallel operators. Generally, the default value is used should not be changed unless necessary.
ATB_MATMUL_SHUFFLE_K_ENABLE	Specifies whether to enable Shuffle-K to control the consistency of the accumulation order of the matrix multiplication result when different positions of the matrix are computed. This variable affects the internal accumulation order of the matmul operator.
ATB_OPSRUNNER_KERNEL_CACHE_GLOABL_COUNT	Number of slots for the global kernel cache. When the number of slots increases, the cache hit rate is improved but the retrieval efficiency is reduced. When the number of slots decreases, the retrieval efficiency is improved but the cache hit rate is reduced.
ATB_OPSRUNNER_KERNEL_CACHE_LOCAL_COUNT	Number of slots for the local kernel cache. When the number of slots increases, the cache hit rate is improved but the retrieval efficiency is reduced. When the number of slots decreases, the retrieval efficiency is improved but the cache hit rate is reduced.
ATB_OPSRUNNER_SETUP_CACHE_ENABLE	Specifies whether to enable the SetupCache function of the ATB. When detecting that the input and output tensors of the operation do not change, this function skips most of the processes of the setup to improve the performance on the scheduling side. This function is enabled by default for performance acceleration.
ATB_STREAM_SYNC_EVERY_KERNEL_ENABLE	Used to locate the kernel where the error is reported. When the variable is set to 1, stream synchronization is performed when the execution of each kernel ends.
ATB_STREAM_SYNC_EVERY_OPERATION_ENABLE	Used to locate the operation where the error is reported. When the variable is set to 1, synchronization is performed during the execution of each operation.
ATB_STREAM_SYNC_EVERY_RUNNER_ENABLE	Used to locate the runner where the error is reported. When the variable is set to 1, stream synchronization is performed during the execution of each runner.
ATB_SHARE_MEMORY_NAME_SUFFIX	Name suffix of the shared memory. When the communication operator is used by multiple users, this variable needs to be set to distinguish the shared memory.
ATB_WORKSPACE_MEM_ALLOC_ALG_TYPE	Workspace memory allocation algorithms. The ATB selects different algorithms to compute the workspace size and allocation based on the environment variable configuration. You can select different algorithms to test the workspace allocation.
ATB_WORKSPACE_MEM_ALLOC_GLOBAL	Specifies whether to use the global intermediate tensor memory allocation algorithm. After this algorithm is used, the size of the intermediate tensor memory is computed and allocated.
HOME	Home directory path of the current user
LCCL_DETERMINISTIC	Specifies whether to enable deterministic LCCL AllReduce (with order-preserving addition). Note that this function takes effect when rankSize ≤ 8. When enabled, it affects Performance of some communication operators Accumulation order of the lccl communication operators
LD_LIBRARY_PATH	Dynamic library search path (for Linux only).
PATH	Executable file search path.
PYTHONPATH	Python module search path.
TOOLCHAIN_HOME	Installation path of the Toolkit toolchain.