LanguageModelGSPOLossConfig

Module: fast_llm.layers.language_model.loss.config

Variant of: LanguageModelLossConfig — select with type: gspo

Inherits from: LanguageModelPolicyGradientLossConfig, CombinableLossConfig, LanguageModelLossConfig

Fields

weight — core

Type: float    Default: 1.0

Weight for this loss in the total loss computation.

logits_scale_factor — feature

Type: float    Default: 1.0

Extra logits scale factor applied for this loss only, stacked on top of the model's logits_scale_factor.

metrics — feature

Type: PolicyMetricsLevel    Default: "auto"

Diagnostic metrics to log. basic: importance-ratio, KL, advantage statistics and the policy entropy. auto: basic when pipeline_parallel == 1, else none. none: disable, adding no cost. basic is not supported with pipeline_parallel > 1.

epsilon_high

Type: float    Default: 0.2

Upper clip parameter for ratio of log probs

epsilon_low

Type: float    Default: 0.2

Lower clip parameter for ratio of log probs

use_triton — expert

Type: bool or None    Default: None

Enable triton implementation. Default: use if available.