[Linear Attn] Add CUSTOM enum and plugin extensibility for kernel backends (#24937)

python/sglang/srt/layers/attention/linear/utils.py

@@ -16,6 +16,11 @@ class LinearAttnKernelBackend(Enum):
     TRITON = "triton"
     CUTEDSL = "cutedsl"
     FLASHINFER = "flashinfer"
+    CUSTOM = "custom"
+
+    @classmethod
+    def _missing_(cls, value):
+        return cls.CUSTOM
 
     def is_triton(self):
         return self == LinearAttnKernelBackend.TRITON
@@ -26,6 +31,9 @@ class LinearAttnKernelBackend(Enum):
     def is_flashinfer(self):
         return self == LinearAttnKernelBackend.FLASHINFER
 
+    def is_custom(self):
+        return self == LinearAttnKernelBackend.CUSTOM
+
 
 LINEAR_ATTN_DECODE_BACKEND: Optional[LinearAttnKernelBackend] = None
 LINEAR_ATTN_PREFILL_BACKEND: Optional[LinearAttnKernelBackend] = None
@@ -41,11 +49,8 @@ def initialize_linear_attn_config(server_args: ServerArgs):
 
     LINEAR_ATTN_DECODE_BACKEND = LinearAttnKernelBackend(decode)
     LINEAR_ATTN_PREFILL_BACKEND = LinearAttnKernelBackend(prefill)
-    rank0_log(
+
-        f"Linear attention kernel backend: "
+    rank0_log(f"Linear attention kernel backend: decode={decode}, prefill={prefill}")
-        f"decode={LINEAR_ATTN_DECODE_BACKEND.value}, "
-        f"prefill={LINEAR_ATTN_PREFILL_BACKEND.value}"
-    )
 
 
 def get_linear_attn_decode_backend() -> LinearAttnKernelBackend:

python/sglang/srt/server_args.py

@@ -305,6 +305,10 @@ def add_rl_on_policy_target_choices(choices):
     RL_ON_POLICY_TARGET_CHOICES.extend(choices)
 
 
+def add_linear_attn_kernel_backend_choices(choices):
+    LINEAR_ATTN_KERNEL_BACKEND_CHOICES.extend(choices)
+
+
 def _resolve_speculative_algorithm_alias(
     speculative_algorithm: Optional[str],
     speculative_draft_model_path: Optional[str],