修复新request kv cache移动错误

vllm/spec_decode/spec_decode_worker.py

@@ -531,6 +531,9 @@ class SpecDecodeWorker(LoraNotSupportedWorkerBase):
         not called, meaning that the kv-cache in proposer for requests is not
         updated, so they cannot enable spec decode in the rest decoding.
         """
+        if self.tree_style_spec_decoding and self.kvcache_slot_to_be_moved is not None:
+            execute_model_req.kvcache_slot_to_be_moved = self.kvcache_slot_to_be_moved
+            self.kvcache_slot_to_be_moved = None
 
         sampler_output = self.scorer_worker.execute_model(execute_model_req)
         assert len(sampler_output) == 1