[Qwen3-next] Add PD disaggregation support for mamba with extra_buffer (#15180)

Signed-off-by: Shangming Cai <csmthu@gmail.com>
Co-authored-by: ybyang <10629930+whybeyoung@users.noreply.github.com>

python/sglang/srt/disaggregation/decode.py

@@ -144,6 +144,7 @@ class HybridMambaDecodeReqToTokenPool(HybridReqToTokenPool):
         enable_memory_saver: bool,
         cache_params: "Mamba2CacheParams",
         speculative_num_draft_tokens: int,
+        enable_mamba_extra_buffer: bool,
         pre_alloc_size: int,
     ):
         DecodeReqToTokenPool.__init__(
@@ -154,10 +155,11 @@ class HybridMambaDecodeReqToTokenPool(HybridReqToTokenPool):
             enable_memory_saver=enable_memory_saver,
             pre_alloc_size=pre_alloc_size,
         )
-        self.enable_memory_saver = enable_memory_saver
+        self.mamba_ping_pong_track_buffer_size = (
-        self.enable_mamba_extra_buffer = (
+            2 if speculative_num_draft_tokens is None else 1
-            False  # TODO: add PD support for mamba cache extra_buffer
         )
+        self.enable_mamba_extra_buffer = enable_mamba_extra_buffer
+        self.enable_memory_saver = enable_memory_saver
         self._init_mamba_pool(
             size=size + pre_alloc_size,
             mamba_spec_state_size=size + pre_alloc_size,

python/sglang/srt/model_executor/model_runner.py

@@ -1797,6 +1797,7 @@ class ModelRunner:
                         enable_memory_saver=self.server_args.enable_memory_saver,
                         cache_params=config.mamba2_cache_params,
                         speculative_num_draft_tokens=self.server_args.speculative_num_draft_tokens,
+                        enable_mamba_extra_buffer=self.server_args.enable_mamba_extra_buffer(),
                         pre_alloc_size=pre_alloc_size,
                     )
                 else:

python/sglang/srt/server_args.py

@@ -1382,9 +1382,6 @@ class ServerArgs:
                 assert (
                     is_cuda()
                 ), "Mamba extra_buffer is only supported on CUDA devices with FLA backend"
-                assert (
-                    self.disaggregation_mode == "null"
-                ), "Mamba extra_buffer is not compatible with disaggregation mode yet."
                 if self.speculative_num_draft_tokens is not None:
                     assert (
                         self.mamba_track_interval >= self.speculative_num_draft_tokens

test/srt/run_suite.py

@@ -157,7 +157,7 @@ suites = {
     "per-commit-8-gpu-h200": [
         TestFile("test_deepseek_v3_basic.py", 275),
         TestFile("test_deepseek_v3_mtp.py", 275),
-        TestFile("test_disaggregation_hybrid_attention.py", 200),
+        TestFile("test_disaggregation_hybrid_attention.py", 600),
         TestFile("models/test_kimi_k2_models.py", 200),
         TestFile("test_deepseek_v32_basic.py", 275),
         TestFile("test_deepseek_v32_mtp.py", 275),

test/srt/test_disaggregation_hybrid_attention.py

@@ -79,5 +79,77 @@ class TestDisaggregationHybridAttentionMamba(PDDisaggregationServerBase):
         self.assertGreater(metrics["accuracy"], 0.93)
 
 
+class TestDisaggregationHybridAttentionMambaExtraBuffer(PDDisaggregationServerBase):
+    @classmethod
+    def setUpClass(cls):
+        super().setUpClass()
+        cls.model = "Qwen/Qwen3-Next-80B-A3B-Instruct"
+
+        # Non blocking start servers
+        cls.start_prefill()
+        cls.start_decode()
+
+        # Block until both
+        cls.wait_server_ready(cls.prefill_url + "/health")
+        cls.wait_server_ready(cls.decode_url + "/health")
+
+        cls.launch_lb()
+
+    @classmethod
+    def start_prefill(cls):
+        prefill_args = [
+            "--trust-remote-code",
+            "--disaggregation-mode",
+            "prefill",
+            "--tp",
+            "4",
+            "--mamba-scheduler-strategy",
+            "extra_buffer",
+        ]
+        prefill_args += cls.transfer_backend + cls.rdma_devices
+        cls.process_prefill = popen_launch_pd_server(
+            cls.model,
+            cls.prefill_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=prefill_args,
+        )
+
+    @classmethod
+    def start_decode(cls):
+        decode_args = [
+            "--trust-remote-code",
+            "--disaggregation-mode",
+            "decode",
+            "--tp",
+            "4",
+            "--base-gpu-id",
+            "4",
+            "--mamba-scheduler-strategy",
+            "extra_buffer",
+        ]
+        decode_args += cls.transfer_backend + cls.rdma_devices
+        cls.process_decode = popen_launch_pd_server(
+            cls.model,
+            cls.decode_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=decode_args,
+        )
+
+    def test_gsm8k(self):
+        args = SimpleNamespace(
+            num_shots=5,
+            data_path=None,
+            num_questions=200,
+            max_new_tokens=512,
+            parallel=128,
+            host=f"http://{self.base_host}",
+            port=int(self.lb_port),
+        )
+        metrics = run_eval_few_shot_gsm8k(args)
+        print(f"Evaluation metrics: {metrics}")
+
+        self.assertGreater(metrics["accuracy"], 0.93)
+
+
 if __name__ == "__main__":
     unittest.main()