[config] Recognize custom hybrid SWA models via hf_text_config.is_hybrid_swa (#23988)

python/sglang/srt/configs/model_config.py

@@ -504,7 +504,7 @@ class ModelConfig:
     def _derive_hybrid_model(self):
         # Use self.context_len after it has been initialized to prevent using context_len which may be None.
         self.is_hybrid_swa = (
-            is_hybrid_swa_model(self.hf_config.architectures)
+            is_hybrid_swa_model(self.hf_config.architectures, self.hf_text_config)
             and not self.disable_hybrid_swa_memory
         )
 
@@ -1671,7 +1671,10 @@ def compute_mla_mscale_scaling(rope_scaling: dict, base_scaling: float) -> float
     return base_scaling * mscale * mscale
 
 
-def is_hybrid_swa_model(model_architectures: List[str]):
+def is_hybrid_swa_model(
+    model_architectures: List[str],
+    hf_text_config: Optional[PretrainedConfig] = None,
+):
 
     hybrid_swa_archs = {
         "Llama4ForConditionalGeneration",
@@ -1687,7 +1690,13 @@ def is_hybrid_swa_model(model_architectures: List[str]):
         "Gemma4ForConditionalGeneration",
         "LagunaForCausalLM",
     }
-    return any(arch in hybrid_swa_archs for arch in model_architectures)
+    if any(arch in hybrid_swa_archs for arch in model_architectures):
+        return True
+    # Also recognize models that explicitly opt-in via their HF text config,
+    # so custom hybrid-SWA architectures don't need to be added to the allowlist.
+    if hf_text_config is not None and getattr(hf_text_config, "is_hybrid_swa", False):
+        return True
+    return False
 
 
 def get_hybrid_layer_ids(
@@ -1758,6 +1767,17 @@ def get_hybrid_layer_ids(
         full_attention_layer_ids = [
             i for i, x in enumerate(layer_types) if x == "full_attention"
         ]
+    elif getattr(hf_text_config, "hybrid_layer_pattern", None) is not None:
+        # Generic fallback for custom hybrid SWA models that opt in via
+        # hf_text_config.is_hybrid_swa and expose a hybrid_layer_pattern
+        # (1 = SWA, 0 = full) without needing to be added to the allowlist.
+        hybrid_layer_pattern = hf_text_config.hybrid_layer_pattern
+        swa_attention_layer_ids = [
+            i for i in range(num_hidden_layers) if hybrid_layer_pattern[i] == 1
+        ]
+        full_attention_layer_ids = [
+            i for i in range(num_hidden_layers) if hybrid_layer_pattern[i] == 0
+        ]
     else:
         swa_attention_layer_ids = None
         full_attention_layer_ids = None