fix: Huber energy loss uses raw model energy instead of energy_pred that includes atom_ener_coeff weighting, so coefficients are silently dropped when use_huber=True

Author: Han Wang

deepmd/pt/loss/ener.py

@@ -224,8 +224,8 @@ def forward(
                     loss += atom_norm * (pref_e * l2_ener_loss)
                 else:
                     l_huber_loss = custom_huber_loss(
-                        atom_norm * model_pred["energy"],
-                        atom_norm * label["energy"],
+                        atom_norm * energy_pred,
+                        atom_norm * energy_label,
                         delta=self.huber_delta,
                     )
                     loss += pref_e * l_huber_loss