Reduce one step decode for draft model. (#11561)

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

@@ -916,7 +916,7 @@ class FlashInferMLAMultiStepDraftBackend:
         )
 
         self.attn_backends = []
-        for i in range(self.speculative_num_steps):
+        for i in range(self.speculative_num_steps - 1):
             self.attn_backends.append(
                 FlashInferMLAAttnBackend(
                     model_runner,
@@ -998,7 +998,7 @@ class FlashInferMLAMultiStepDraftBackend:
             device="cuda",
         )
 
-        for i in range(self.speculative_num_steps):
+        for i in range(self.speculative_num_steps - 1):
             self.attn_backends[i].init_cuda_graph_state(
                 max_bs, max_num_tokens, kv_indices_buf=self.cuda_graph_kv_indices[i]
             )