diff --git a/src/maxtext/configs/models/deepseek4-284b.yml b/src/maxtext/configs/models/deepseek4-284b.yml index 5623878b6c..5bef652d42 100644 --- a/src/maxtext/configs/models/deepseek4-284b.yml +++ b/src/maxtext/configs/models/deepseek4-284b.yml @@ -56,6 +56,8 @@ routed_bias_update_rate: 0.001 load_balance_loss_weight: 0.0001 log_moe_bias_norms: false routed_scaling_factor: 1.5 +# Reference Gate computes router logits in fp32 (linear(x.float(), weight.float())) regardless of model dtype. +float32_gate_logits: true # --- Attention configuration ---