[Fix] Fix several bugs on DSA models (#22430)

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

@@ -2129,11 +2129,9 @@ class NativeSparseAttnBackend(
             # disable for MTP
             self.nsa_kv_cache_store_fp8
             and self.nsa_prefill_impl == "flashmla_sparse"
+            and forward_mode == ForwardMode.EXTEND
         ):
             topk_transform_method = TopkTransformMethod.RAGGED
-
-            if forward_mode is not None and (forward_mode.is_decode_or_idle()):
-                topk_transform_method = TopkTransformMethod.PAGED
         else:
             topk_transform_method = TopkTransformMethod.PAGED
         return topk_transform_method

python/sglang/srt/server_args.py

@@ -1491,8 +1491,10 @@ class ServerArgs:
             self.nsa_decode_backend = "tilelang"
         elif kv_cache_dtype == "fp8_e4m3":
             if major >= 10:
-                self.nsa_prefill_backend = "trtllm"
-                self.nsa_decode_backend = "trtllm"
+                if not user_set_prefill:
+                    self.nsa_prefill_backend = "trtllm"
+                if not user_set_decode:
+                    self.nsa_decode_backend = "trtllm"
             else:
                 # flashmla_auto dispatches to flashmla_sparse/flashmla_kv based on hardware and heuristics
                 if not user_set_prefill: