model Namespace Reference

Classes
class	PieceBranch
class	LightExpert
class	ChessNetFactorizedMoE

Variables
	model = ChessNetFactorizedMoE(expert_bottleneck=16, mixer_out=512)
list	planes_list = [torch.randn(1, in_ch, 8, 8) for in_ch in ChessNetFactorizedMoE.PLANES_PER_TYPE]
	bypass = torch.randn(1, 12, 8, 8)
	global_v = torch.randn(1, 21)
	wdl = model(planes_list, bypass, global_v)
	total_params = sum(p.numel() for p in model.parameters())
	branches_params = sum(p.numel() for n, p in model.named_parameters() if "branches" in n)
	stem_global_params = sum(p.numel() for n, p in model.named_parameters() if "stem_global" in n)
	pointwise_mixer_params = sum(p.numel() for n, p in model.named_parameters() if "pointwise_mixer" in n)
	backbone_params = sum(p.numel() for n, p in model.named_parameters() if "branches" in n or "pointwise_mixer" in n or "stem_global" in n)
	expert_params = sum(p.numel() for n, p in model.named_parameters() if "experts" in n)

Detailed Description

@file model.py
@brief Lightweight MoE Model for Chess Evaluation (Factorized Cache Architecture).

RULES FOR C++ PERFORMANCE:
1. NO 3x3 CONVOLUTIONS AFTER THE MIXER. (Preserves spatial caching).
2. Dense layers must be protected by a 1x1 bottleneck to minimize Flat size.

Variable Documentation

backbone_params

model.backbone_params = sum(p.numel() for n, p in model.named_parameters() if "branches" in n or "pointwise_mixer" in n or "stem_global" in n)

branches_params

model.branches_params = sum(p.numel() for n, p in model.named_parameters() if "branches" in n)

bypass

model.bypass = torch.randn(1, 12, 8, 8)

expert_params

model.expert_params = sum(p.numel() for n, p in model.named_parameters() if "experts" in n)

global_v

model.global_v = torch.randn(1, 21)

model

model.model = ChessNetFactorizedMoE(expert_bottleneck=16, mixer_out=512)

planes_list

list model.planes_list = [torch.randn(1, in_ch, 8, 8) for in_ch in ChessNetFactorizedMoE.PLANES_PER_TYPE]

pointwise_mixer_params

model.pointwise_mixer_params = sum(p.numel() for n, p in model.named_parameters() if "pointwise_mixer" in n)

stem_global_params

model.stem_global_params = sum(p.numel() for n, p in model.named_parameters() if "stem_global" in n)

total_params

model.total_params = sum(p.numel() for p in model.parameters())

wdl

model.wdl = model(planes_list, bypass, global_v)