Concept算子库
Kernel Library
The collection of low-level computational implementations available to an accelerator, such as cuBLAS and cuDNN on GPUs or custom operators for a specific NPU. The degree of uniformity in their interfaces and semantics directly affects the framework-level invocation paths and model execution efficiency.
Timeline (last 90 days)
Oct 12
Most of PyTorch's operators are lowered to operator libraries for execution, and whether an accelerator can run models efficiently depends to a large extent on its operator library's coverage and performance for commonly used PyTorch operators.
Unverified50%