[Bugfix] fix kv buffer register & dp attention & deepepmoe (#9327)

python/sglang/srt/disaggregation/ascend/conn.py

@@ -23,9 +23,7 @@ class AscendKVManager(MooncakeKVManager):
         )
 
     def register_buffer_to_engine(self):
-        self.engine.register(
-            self.kv_args.kv_data_ptrs[0], sum(self.kv_args.kv_data_lens)
-        )
+        self.engine.batch_register(self.kv_args.kv_data_ptrs, self.kv_args.kv_data_lens)
         # The Ascend backend optimize batch registration for small memory blocks.
         self.engine.batch_register(
             self.kv_args.aux_data_ptrs, self.kv_args.aux_data_lens

python/sglang/srt/layers/dp_attention.py

@@ -234,7 +234,7 @@ def initialize_dp_attention(
     _DpGatheredBufferWrapper.set_metadata(
         hidden_size=model_config.hidden_size,
         dtype=model_config.dtype,
-        device=torch.device("cuda"),
+        device=torch.device(server_args.device),
     )
 
 

python/sglang/srt/layers/moe/ep_moe/layer.py

@@ -736,7 +736,7 @@ class DeepEPMoE(EPMoE):
             assert isinstance(dispatch_output, AscendDeepEPLLOutput)
         hidden_states, topk_idx, topk_weights, _, seg_indptr, _ = dispatch_output
         assert self.quant_method is not None
-        assert self.activation == "silu"
+        assert self.moe_runner_config.activation == "silu"
 
         # NOTE: Ascend's Dispatch & Combine does not support FP16
         output_dtype = torch.bfloat16