Update default fp32 residual

Signed-off-by: John St. John <jstjohn@nvidia.com>

Author: jstjohn

bionemo-recipes/recipes/evo2_megatron/src/bionemo/evo2/models/evo2_provider.py

@@ -306,7 +306,7 @@ class HyenaModelProvider(TransformerConfig, ModelProviderMixin[MCoreHyenaModel])
     apply_rope_fusion: bool = True
     make_vocab_size_divisible_by: int = 128
     gated_linear_unit: bool = True
-    fp32_residual_connection: bool = True
+    fp32_residual_connection: bool = False
     normalization: str = "RMSNorm"
     add_bias_linear: bool = False
     hidden_dropout: float = 0.0

bionemo-recipes/recipes/evo2_megatron/src/bionemo/evo2/run/predict.py

@@ -1046,9 +1046,6 @@ def predict(
         # Set the model to use fewer layers and skip post-processing (output heads).
         model_provider.num_layers = target_num_layers
         model_provider.post_process = False
-        if hasattr(model_provider, "fp32_residual_connection"):
-            model_provider.fp32_residual_connection = False
-
         # Also truncate the hybrid_override_pattern if it exists, since it must match num_layers
         if hasattr(model_provider, "hybrid_override_pattern") and model_provider.hybrid_override_pattern is not None:
             original_pattern = model_provider.hybrid_override_pattern