refactor: wrap split backends once on full-attention backends (#31439)

python/sglang/srt/model_executor/model_runner_components/attention_backend_setup.py

@@ -184,17 +184,26 @@ def _build_resolved_backend(
             HybridAttnBackend,
         )
 
-        attn_backend = HybridAttnBackend(
+        # Compose the two full-attention backends first, then apply model-level
+        # wrappers once.  Wrapping each child independently duplicates the
+        # linear/sparse side backend for hybrid models (for example, two GDN
+        # dispatchers for Qwen3.5 when prefill and decode use different MHA
+        # backends), duplicating initialization and associated state while only
+        # one side backend can be active in a forward pass.
+        attn_backend = attn_backend_wrapper(
             model_runner,
-            decode_backend=_build_backend_from_str(
+            HybridAttnBackend(
                 model_runner=model_runner,
-                backend_str=resolved.decode,
+                decode_backend=_build_full_attention_backend_from_str(
-                init_new_workspace=init_new_workspace,
+                    model_runner=model_runner,
-            ),
+                    backend_str=resolved.decode,
-            prefill_backend=_build_backend_from_str(
+                    init_new_workspace=init_new_workspace,
-                model_runner=model_runner,
+                ),
-                backend_str=resolved.prefill,
+                prefill_backend=_build_full_attention_backend_from_str(
-                init_new_workspace=init_new_workspace,
+                    model_runner=model_runner,
+                    backend_str=resolved.prefill,
+                    init_new_workspace=init_new_workspace,
+                ),
             ),
         )
         logger.info(
@@ -217,9 +226,21 @@ def _build_resolved_backend(
 
 def _build_backend_from_str(
     *, model_runner: ModelRunner, backend_str: str, init_new_workspace: bool
+) -> AttentionBackend:
+    return attn_backend_wrapper(
+        model_runner,
+        _build_full_attention_backend_from_str(
+            model_runner=model_runner,
+            backend_str=backend_str,
+            init_new_workspace=init_new_workspace,
+        ),
+    )
+
+
+def _build_full_attention_backend_from_str(
+    *, model_runner: ModelRunner, backend_str: str, init_new_workspace: bool
 ) -> AttentionBackend:
     if backend_str not in ATTENTION_BACKENDS:
         raise ValueError(f"Invalid attention backend: {backend_str}")
     model_runner.init_new_workspace = init_new_workspace
-    full_attention_backend = ATTENTION_BACKENDS[backend_str](model_runner)
+    return ATTENTION_BACKENDS[backend_str](model_runner)
-    return attn_backend_wrapper(model_runner, full_attention_backend)

test/registered/unit/model_executor/model_runner_components/test_attention_backend_setup.py

@@ -0,0 +1,70 @@
+import sys
+from types import SimpleNamespace
+from unittest.mock import patch
+
+import pytest
+
+from sglang.srt.layers.attention.hybrid_attn_backend import HybridAttnBackend
+from sglang.srt.model_executor.model_runner_components import (
+    attention_backend_setup,
+)
+from sglang.srt.model_executor.model_runner_components.attention_backend_setup import (
+    ResolvedAttentionBackendStr,
+)
+from sglang.test.ci.ci_register import register_cpu_ci
+
+register_cpu_ci(est_time=5, suite="base-a-test-cpu")
+
+
+class _FakeBackend:
+    def __init__(self, name):
+        self.name = name
+
+
+def test_split_full_attention_applies_model_wrapper_once():
+    runner = SimpleNamespace(
+        server_args=SimpleNamespace(speculative_attention_mode="prefill"),
+        kv_cache_dtype=None,
+        token_to_kv_pool=object(),
+        req_to_token_pool=object(),
+        init_new_workspace=None,
+    )
+    wrapper_inputs = []
+    wrapped_backend = object()
+
+    def wrap_once(model_runner, backend):
+        assert model_runner is runner
+        wrapper_inputs.append(backend)
+        return wrapped_backend
+
+    constructors = {
+        "decode-test": lambda model_runner: _FakeBackend("decode"),
+        "prefill-test": lambda model_runner: _FakeBackend("prefill"),
+    }
+    resolved = ResolvedAttentionBackendStr(decode="decode-test", prefill="prefill-test")
+
+    with (
+        patch.dict(attention_backend_setup.ATTENTION_BACKENDS, constructors),
+        patch.object(
+            attention_backend_setup,
+            "attn_backend_wrapper",
+            side_effect=wrap_once,
+        ),
+    ):
+        result = attention_backend_setup._build_resolved_backend(
+            model_runner=runner,
+            resolved=resolved,
+            init_new_workspace=True,
+        )
+
+    assert result is wrapped_backend
+    assert len(wrapper_inputs) == 1
+    split_backend = wrapper_inputs[0]
+    assert isinstance(split_backend, HybridAttnBackend)
+    assert split_backend.decode_backend.name == "decode"
+    assert split_backend.prefill_backend.name == "prefill"
+    assert runner.init_new_workspace is True
+
+
+if __name__ == "__main__":
+    sys.exit(pytest.main([__file__, "-v"]))