refactor(attn): init hisparse_coordinator before attn_backend; replace lazy property with init-time capture (#26012)

Co-authored-by: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>

python/sglang/srt/layers/attention/deepseek_v4_backend.py

@@ -354,14 +354,9 @@ class DeepseekV4AttnBackend(
         self.page_size = model_runner.page_size
         assert self.page_size == 256, "the system hardcodes page_size=256"
 
-        # Pool refs — captured at construction so they survive deletion of the
-        # corresponding ForwardBatch fields.
         self.req_to_token_pool = model_runner.req_to_token_pool
         self.token_to_kv_pool: DeepSeekV4TokenToKVPool = model_runner.token_to_kv_pool
-        # Keep a runner ref to read live state set after backend construction
-        # (e.g. hisparse_coordinator is built in model_runner *after*
-        # init_attention_backend()).
-        self.model_runner = model_runner
+        self.hisparse_coordinator = model_runner.hisparse_coordinator
         self.req_to_token = model_runner.req_to_token_pool.req_to_token
         self.MAX_SEQ_LEN_FOR_CAPTURE = self.req_to_token.shape[1]
 
@@ -385,12 +380,6 @@ class DeepseekV4AttnBackend(
         ] = None
         self._replay_forward_batch: Optional[ForwardBatch] = None  # FIXME: out-of-band
 
-    @property
-    def hisparse_coordinator(self):
-        # Live read: model_runner builds the coordinator *after*
-        # init_attention_backend(), so we cannot capture at __init__ time.
-        return self.model_runner.hisparse_coordinator
-
     def _move_to_device(self, x: List[int]) -> torch.Tensor:
         pin_tensor = torch.tensor(x, dtype=torch.int32, pin_memory=True)
         return pin_tensor.to(self.device, non_blocking=True)
@@ -1196,7 +1185,6 @@ class DeepseekV4MultiStepBackend(DeepseekV4AttnBackend):
         self, model_runner: ModelRunner, topk: int, speculative_num_steps: int
     ):
         super().__init__(model_runner)
-        self.model_runner = model_runner
         self.topk = topk
         self.speculative_num_steps = speculative_num_steps
         self.attn_backends: List[DeepseekV4AttnBackend] = []

python/sglang/srt/layers/attention/deepseek_v4_backend_hip_radix.py

@@ -348,14 +348,9 @@ class DeepseekV4HipRadixBackend(
         self.page_size = model_runner.page_size
         assert self.page_size == 256, "the system hardcodes page_size=256"
 
-        # Pool refs — captured at construction so they survive deletion of the
-        # corresponding ForwardBatch fields.
         self.req_to_token_pool = model_runner.req_to_token_pool
         self.token_to_kv_pool: DeepSeekV4TokenToKVPool = model_runner.token_to_kv_pool
-        # Keep a runner ref to read live state set after backend construction
-        # (e.g. hisparse_coordinator is built in model_runner *after*
-        # init_attention_backend()).
-        self.model_runner = model_runner
+        self.hisparse_coordinator = model_runner.hisparse_coordinator
         self.req_to_token = model_runner.req_to_token_pool.req_to_token
         self.MAX_SEQ_LEN_FOR_CAPTURE = self.req_to_token.shape[1]
 
@@ -379,12 +374,6 @@ class DeepseekV4HipRadixBackend(
         ] = None
         self._replay_forward_batch: Optional[ForwardBatch] = None  # FIXME: out-of-band
 
-    @property
-    def hisparse_coordinator(self):
-        # Live read: model_runner builds the coordinator *after*
-        # init_attention_backend(), so we cannot capture at __init__ time.
-        return self.model_runner.hisparse_coordinator
-
     def _move_to_device(self, x: List[int]) -> torch.Tensor:
         pin_tensor = torch.tensor(x, dtype=torch.int32, pin_memory=True)
         return pin_tensor.to(self.device, non_blocking=True)
@@ -1196,7 +1185,6 @@ class DeepseekV4MultiStepBackend(DeepseekV4HipRadixBackend):
         self, model_runner: ModelRunner, topk: int, speculative_num_steps: int
     ):
         super().__init__(model_runner)
-        self.model_runner = model_runner
         self.topk = topk
         self.speculative_num_steps = speculative_num_steps
         self.attn_backends: List[DeepseekV4HipRadixBackend] = []

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

@@ -330,14 +330,9 @@ class DeepseekSparseAttnBackend(
         self.qk_rope_head_dim = model_runner.model_config.qk_rope_head_dim
 
         assert model_runner.req_to_token_pool is not None
-        # Pool refs — captured at construction so they survive deletion of the
-        # corresponding ForwardBatch fields.
         self.req_to_token_pool = model_runner.req_to_token_pool
         self.token_to_kv_pool = model_runner.token_to_kv_pool
-        # Keep a runner ref to read live state set after backend construction
-        # (e.g. hisparse_coordinator is built in model_runner *after*
-        # init_attention_backend()).
-        self.model_runner = model_runner
+        self.hisparse_coordinator = model_runner.hisparse_coordinator
         self.req_to_token = model_runner.req_to_token_pool.req_to_token
 
         self.use_mha: bool = False
@@ -400,12 +395,6 @@ class DeepseekSparseAttnBackend(
         else:
             self.workspace_buffer = None
 
-    @property
-    def hisparse_coordinator(self):
-        # Live read: model_runner builds the coordinator *after*
-        # init_attention_backend(), so we cannot capture at __init__ time.
-        return self.model_runner.hisparse_coordinator
-
     def get_device_int32_arange(self, l: int) -> torch.Tensor:
         if l > len(self._arange_buf):
             next_pow_of_2 = 1 << (l - 1).bit_length()
@@ -2318,7 +2307,6 @@ class DeepseekSparseAttnMultiStepBackend:
     def __init__(
         self, model_runner: ModelRunner, topk: int, speculative_num_steps: int
     ):
-        self.model_runner = model_runner
         self.topk = topk
         self.speculative_num_steps = speculative_num_steps
         self.attn_backends = []

python/sglang/srt/model_executor/model_runner.py

@@ -761,9 +761,6 @@ class ModelRunner(ModelRunnerKVCacheMixin):
 
         if self.device == "cuda" or self.device == "musa":
             self.init_cublas()
-            self.init_attention_backend()
-            self.kernel_warmup()
-            # Init hisparse coordinator (must happen before CUDA graph capture)
             if self.enable_hisparse:
                 from sglang.srt.managers.hisparse_coordinator import HiSparseCoordinator
                 from sglang.srt.mem_cache.sparsity import parse_hisparse_config
@@ -785,6 +782,8 @@ class ModelRunner(ModelRunnerKVCacheMixin):
                     ),
                     host_to_device_ratio=hisparse_cfg.host_to_device_ratio,
                 )
+            self.init_attention_backend()
+            self.kernel_warmup()
             self._pre_initialize_flashinfer_allreduce_workspace()
             self.init_device_graphs()
         elif self.device == "cpu":