export Namespace Reference

Classes
class	OnnxWrapper
class	OnnxBackboneWrapper
class	OnnxExpertWrapper

Functions
Dict[str, torch.Tensor]	_extract_state_dict (object ckpt_obj)
	_write_u32s (f, values)
	_write_f32_array (f, np.ndarray arr)
None	export_native_weights (Path out_path, ChessNetFactorizedMoE model)
None	export_inputs_and_refs (Path inputs_path, Path refs_path, ChessNetFactorizedMoE model, int n_samples, int seed)
None	main ()

Variables
int	MAGIC_WEIGHTS = 0x32454F4D
int	MAGIC_INPUTS = 0x32504E49
int	MAGIC_REFS = 0x32464552
int	VERSION = 1
int	MAX_BRANCH_PLANES = 10
dict	POOL_TO_CODE

Detailed Description

@file export.py
@brief Export parity artifacts for training/model.py

1) Native float32 weights for a C++ loader
2) ONNX models (monolithic + split backbone/experts)
3) Deterministic input samples and PyTorch reference outputs

Function Documentation

_extract_state_dict()

Dict[str, torch.Tensor] _extract_state_dict ( object ckpt_obj )

protected

Here is the caller graph for this function:

_write_f32_array()

_write_f32_array ( f, np.ndarray arr )

protected

Here is the caller graph for this function:

_write_u32s()

_write_u32s ( f, values )

protected

Here is the caller graph for this function:

export_inputs_and_refs()

None export_inputs_and_refs	(	Path	inputs_path,
		Path	refs_path,
		ChessNetFactorizedMoE	model,
		int	n_samples,
		int	seed )

Here is the call graph for this function:

Here is the caller graph for this function:

export_native_weights()

None export_native_weights	(	Path	out_path,
		ChessNetFactorizedMoE	model )

Here is the call graph for this function:

Here is the caller graph for this function:

main()

None main

(

)

Here is the call graph for this function:

Here is the caller graph for this function:

Variable Documentation

MAGIC_INPUTS

int export.MAGIC_INPUTS = 0x32504E49

MAGIC_REFS

int export.MAGIC_REFS = 0x32464552

MAGIC_WEIGHTS

int export.MAGIC_WEIGHTS = 0x32454F4D

MAX_BRANCH_PLANES

int export.MAX_BRANCH_PLANES = 10

POOL_TO_CODE

dict export.POOL_TO_CODE

Initial value:

=  {
    "flat": 0,
    "gap": 1,
    "pool2avg": 2,
    "pool2max": 3,
}

VERSION

int export.VERSION = 1