refine aiter_backend for mtp (#7279)

Co-authored-by: HAI <hixiao@gmail.com>

python/sglang/srt/managers/schedule_batch.py

@@ -1722,6 +1722,7 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
             or attention_backend_str == "cutlass_mla"
             or attention_backend_str == "ascend"
             or attention_backend_str == "trtllm_mha"
+            or attention_backend_str == "aiter"
             or global_server_args_dict["enable_two_batch_overlap"]
         ):
             seq_lens_cpu = (

python/sglang/srt/speculative/eagle_worker.py

@@ -226,6 +226,22 @@ class EAGLEWorker(TpModelWorker):
                 self.draft_model_runner,
                 skip_prefill=False,
             )
+        elif self.server_args.attention_backend == "aiter":
+            from sglang.srt.layers.attention.aiter_backend import (
+                AiterAttnBackend,
+                AiterMultiStepDraftBackend,
+            )
+
+            self.draft_attn_backend = AiterMultiStepDraftBackend(
+                self.draft_model_runner,
+                self.topk,
+                self.speculative_num_steps,
+            )
+            self.draft_extend_attn_backend = AiterAttnBackend(
+                self.draft_model_runner,
+                skip_prefill=False,
+            )
+            self.has_prefill_wrapper_verify = False
         elif self.server_args.attention_backend == "fa3":
             from sglang.srt.layers.attention.flashattention_backend import (
                 FlashAttentionBackend,