fix ds3.2 nsa backend prefill TBO (#14901)

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

@@ -185,6 +185,9 @@ class TboAttnBackend(AttentionBackend):
     def forward_decode(self, *args, **kwargs):
         return self.primary.forward_decode(*args, **kwargs)
 
+    def get_indexer_metadata(self, layer_id: int, forward_batch: "ForwardBatch"):
+        return self.primary.get_indexer_metadata(layer_id, forward_batch)
+
 
 def _init_forward_metadata_cuda_graph_split(
     fn_name: str,

python/sglang/srt/models/deepseek_v2.py

@@ -67,6 +67,7 @@ from sglang.srt.layers.attention.nsa.utils import (
     is_nsa_enable_prefill_cp,
     prepare_input_dp_with_cp_dsa,
 )
+from sglang.srt.layers.attention.tbo_backend import TboAttnBackend
 from sglang.srt.layers.attention.utils import concat_and_cast_mha_k_triton
 from sglang.srt.layers.communicator import (
     LayerCommunicator,
@@ -425,7 +426,10 @@ def handle_attention_nsa(attn, forward_batch):
     Dispatch logic is centralized in NativeSparseAttnBackend.set_nsa_prefill_impl and executed
     in init_forward_metadata. Read the decision from backend.use_mha.
     """
+
     backend = forward_batch.attn_backend
+    if isinstance(backend, TboAttnBackend):  # if enable tbo, get primary backend
+        backend = backend.primary
     if hasattr(backend, "use_mha") and backend.use_mha:
         return AttnForwardMethod.MHA_ONE_SHOT
     return AttnForwardMethod.MLA
@@ -2670,7 +2674,10 @@ class DeepseekV2AttentionMLA(nn.Module):
 
         Returns: (kv_a, k_pe) both in BF16
         """
-        kv_indices = forward_batch.attn_backend.forward_metadata.page_table_1_flattened
+        backend = forward_batch.attn_backend
+        if isinstance(backend, TboAttnBackend):  # if enable tbo, get primary backend
+            backend = backend.primary
+        kv_indices = backend.forward_metadata.page_table_1_flattened
         assert (
             kv_indices is not None
         ), "page_table_1_flattened should have been generated for FP8 MHA path"

python/sglang/srt/server_args.py

@@ -727,6 +727,9 @@ class ServerArgs:
         # Handle any other necessary validations.
         self._handle_other_validations()
 
+        # Handle two-batch overlap settings.
+        self._handle_two_batch_overlap()
+
     def _handle_deprecated_args(self):
         # Handle deprecated tool call parsers
         deprecated_tool_call_parsers = {"qwen25": "qwen", "glm45": "glm"}
@@ -2392,6 +2395,12 @@ class ServerArgs:
                     self.preferred_sampling_params
                 )
 
+    def _handle_two_batch_overlap(self):
+        if self.enable_two_batch_overlap and self.moe_a2a_backend == "none":
+            raise ValueError(
+                "When enabling two batch overlap, moe_a2a_backend cannot be 'none'."
+            )
+
     @staticmethod
     def add_cli_args(parser: argparse.ArgumentParser):