[Kernel] Enable Helion backend for Kimi Delta-Attention (#32593)

Co-authored-by: Ethan Che <eche@meta.com>

test/registered/unit/server_args/test_page_major_backend_allowlist.py

@@ -37,7 +37,14 @@ from sglang.test.ci.ci_register import register_cpu_ci
 register_cpu_ci(est_time=5, suite="base-a-test-cpu")
 
 
-def _accepts(backend: str, *, use_mla: bool, unified: bool = True) -> bool:
+def _accepts(
+    backend: str,
+    *,
+    use_mla: bool,
+    unified: bool = True,
+    linear_decode: str | None = None,
+    linear_prefill: str | None = None,
+) -> bool:
     """Run just `_handle_page_major_kv_layout` against a minimal stand-in.
 
     ServerArgs' real constructor pulls in a model config; this exercises the
@@ -53,8 +60,8 @@ def _accepts(backend: str, *, use_mla: bool, unified: bool = True) -> bool:
         "prefill_attention_backend": None,
         "decode_attention_backend": None,
         "linear_attn_backend": "triton",
-        "linear_attn_decode_backend": None,
-        "linear_attn_prefill_backend": None,
+        "linear_attn_decode_backend": linear_decode,
+        "linear_attn_prefill_backend": linear_prefill,
         "mamba_backend": "triton",
     }.items():
         object.__setattr__(sa, name, value)
@@ -115,6 +122,17 @@ class TestPageMajorBackendAllowlist(unittest.TestCase):
                     f"{backend} has no dense-id remapping and must be rejected",
                 )
 
+    def test_helion_linear_attention_is_kda_only(self):
+        for unified in (True, False):
+            for phase in ("decode", "prefill"):
+                kwargs = {f"linear_{phase}": "helion"}
+                self.assertTrue(
+                    _accepts("triton", use_mla=True, unified=unified, **kwargs)
+                )
+                self.assertFalse(
+                    _accepts("triton", use_mla=False, unified=unified, **kwargs)
+                )
+
 
 if __name__ == "__main__":
     unittest.main()