NEML2 1.4.0
|
Ctrl
+F
or Cmd
+F
to search the entire page.Refer to System Documentation for detailed explanation about this system.
Global settings for tensors, models, etc.
buffer_name_separator
Nested buffer name separator. The default is '.'. For example, a sub-model 'foo' which declares a buffer 'bar' will have a buffer named 'foo.bar'.default_device
Default device on which tensors are created and models are evaluated. The string supplied must follow the following schema: (cpu|cuda)[:<device-index>] where cpu or cuda specifies the device type, and :<device-index> optionally specifies a device index. For example, device='cpu' sets the target compute device to be CPU, and device='cuda:1' sets the target compute device to be CUDA with device ID 1.default_floating_point_type
Default floating point type for tensors. Options are Float64 Float32 Float16 default_integer_type
Default integer type for tensors. Options are Int64 Int32 Int16 Int8 interop_threads
Number threads used for inter-ops parallelism. If set to 0, defaults to number of CPU cores.intraop_threads
Number threads used for intra-ops parallelism. If set to 0, defaults to number of CPU cores.machine_precision
Machine precision used at various places to workaround singularities like division-by-zero.parameter_name_separator
Parameter name separator. The default is '.'. For example, a sub-model 'foo' which declares a parameter 'bar' will have a parameter named 'foo.bar'.tighter_tolerance
A tighter tolerance used in various algorithms.tolerance
Tolerance used in various algorithms.Detailed documentation link