fix: fallback to triton for attention-sink models (flashinfer unsupported) (#23139)

Co-authored-by: Liangsheng Yin <hnyls2002@gmail.com>
Co-authored-by: hnyls2002 <lsyincs@gmail.com>

python/sglang/srt/configs/model_config.py

@@ -381,6 +381,8 @@ class ModelConfig:
                 )
             )
 
+        self.has_attention_sinks = self._detect_attention_sinks()
+
         self.is_hybrid_swa_compress = self.hf_config.architectures[0] in [
             "MiMoV2FlashForCausalLM",
             "MiMoV2MTP",
@@ -388,6 +390,25 @@ class ModelConfig:
             "Gemma4ForConditionalGeneration",
         ]
 
+    def _detect_attention_sinks(self) -> bool:
+        """Check whether the model uses learned attention sinks.
+
+        Attention sinks are per-head scalars added to the softmax denominator
+        to compensate for evicted KV-cache entries under sliding-window
+        attention.  Not every hybrid-SWA model uses them.
+        """
+        archs = self.hf_config.architectures or []
+        # GptOss always creates sinks unconditionally.
+        if "GptOssForCausalLM" in archs:
+            return True
+
+        # MiMoV2 creates sinks only when the config flags are set.
+        if any(a in archs for a in ("MiMoV2FlashForCausalLM", "MiMoV2MTP")):
+            return getattr(
+                self.hf_text_config, "add_swa_attention_sink_bias", False
+            ) or getattr(self.hf_text_config, "add_full_attention_sink_bias", False)
+        return False
+
     def _derive_context_length(self, context_length: int):
         is_draft_model = self.is_draft_model
         derived_context_len = get_context_length(self.hf_text_config)

python/sglang/srt/server_args.py

@@ -2422,7 +2422,10 @@ class ServerArgs:
             elif is_mps():
                 return "torch_native"
             else:
-                return "flashinfer" if is_flashinfer_available() else "triton"
+                # FlashInfer does not support attention sinks.
+                if is_flashinfer_available() and not model_config.has_attention_sinks:
+                    return "flashinfer"
+                return "triton"
         else:
             # MLA architecture
             if is_hopper_with_cuda_12_3():
@@ -3161,9 +3164,14 @@ class ServerArgs:
 
         # If decode backend is implicit, pick a safe backend without changing io backend.
         if not self.use_mla_backend():
-            self.decode_attention_backend = (
-                "flashinfer" if is_flashinfer_available() else "triton"
-            )
+            # FlashInfer does not support attention sinks.
+            if (
+                is_flashinfer_available()
+                and not self.get_model_config().has_attention_sinks
+            ):
+                self.decode_attention_backend = "flashinfer"
+            else:
+                self.decode_attention_backend = "triton"
         else:
             self.decode_attention_backend = (
                 "flashinfer" if is_sm100_supported() else "triton"