fz::FusedQuantAdaptiveLorenzoConfig Struct Reference

#include <quant_adaptive_lorenzo_stage.h>

Public Attributes
uint8_t	coder_block_size
	Fixed at 32.
uint8_t	eb_mode
	ErrorBoundMode cast to uint8_t.
double	value_base_f64
	NOA/PREL scan result; 0 for ABS.
float	user_error_bound
	Original config_.error_bound, narrow copy (debug only).

Detailed Description

Serialized config for FusedQuantAdaptiveLorenzoStage, stored in FZMBufferEntry.stage_config. abs_error_bound_f64 is the resolved absolute bound the forward pass actually quantized against (ABS: config_.error_bound unchanged; NOA/PREL: eb * value_base) — the inverse reads it back to dequant with ebx2 = 2 * abs_error_bound_f64, exactly mirroring LorenzoQuantConfig's own error_bound_f64 field and the reasons given there for keeping it double-precision (a narrow float copy of a NOA-resolved bound already carries ~6e-08 relative error before the quantizer sees it).