Update configuration_neollm.py

configuration_neollm.py

@@ -572,7 +572,7 @@ class NeoLLMConfig(PretrainedConfig):
         mea_groupnorm_eps=1e-6,
         use_lucid_attention=False,
         lucid_attention_eps=1e-6,
-        use_affine_scaled_attention=True,
+        use_affine_scaled_attention=False,
         affine_momentum=0.9,
         use_xsa=True,
         xsa_eps=1e-6,
@@ -620,7 +620,7 @@ class NeoLLMConfig(PretrainedConfig):
         repo_start_layer=None,
         repo_d_p=None,
         use_repo_grape=True,
-        use_repo_goat_prior=True,
+        use_repo_goat_prior=False,
         repo_goat_num_frequencies=3,
         repo_goat_sink_decay=4.0,
         # ── VersatileFFN (Nie et al., 2026) ───────────────────────────────
@@ -638,7 +638,7 @@ class NeoLLMConfig(PretrainedConfig):
         use_laurel_lr=False,
         laurel_lr_rank=32,
         # ── Interleaved Head Attention (Duvvuri et al., 2026) ─────────────
-        use_iha=False,
+        use_iha=True,
         iha_num_pseudo_heads=2,          # P=2 → 2×2=4 patrones por head
         iha_local_global_pattern="LLLLG", # 4 locales + 1 global (paper §5.1)
         iha_sliding_window=None,          # auto = N // (2*P^2) usando la longitud real del batch