[MiniMax-M3] Fix UnboundLocalError in the sparse backend on the MSA path

python/sglang/srt/layers/attention/minimax_sparse_backend.py

@@ -309,8 +309,6 @@ def __init__(self, runner: ModelRunner):
         # that wide. Head split mirrors MiniMaxM3 sparse attention's.
         self._idx_group_size = 1
         if self.index_cache_enabled:
-            from sglang.srt.runtime_context import get_parallel
-
             _num_idx_heads = max(
                 sparse_cfg["sparse_num_index_heads"] // get_parallel().attn_tp_size, 1
             )

test/registered/unit/layers/attention/test_minimax_sparse_backend_scoping.py

@@ -0,0 +1,27 @@
+import unittest
+
+from sglang.test.ci.ci_register import register_cpu_ci
+from sglang.test.test_utils import CustomTestCase
+
+register_cpu_ci(est_time=5, suite="base-a-test-cpu")
+
+
+class TestMiniMaxSparseBackendScoping(CustomTestCase):
+    """MiniMaxSparseAttnBackend.__init__ calls the module-level get_parallel() on
+    the MSA path. A function-level import of get_parallel further down the same
+    method made the name local to all of __init__, so a GPU start that took the
+    MSA path raised UnboundLocalError. No CPU runner builds this backend, so the
+    scoping is checked directly."""
+
+    def test_get_parallel_is_not_a_local_of_init(self):
+        from sglang.srt.layers.attention.minimax_sparse_backend import (
+            MiniMaxSparseAttnBackend,
+        )
+
+        self.assertNotIn(
+            "get_parallel", MiniMaxSparseAttnBackend.__init__.__code__.co_varnames
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()