diff --git a/fastdeploy/engine/sched/resource_manager_v1.py b/fastdeploy/engine/sched/resource_manager_v1.py
index 9d8925ceaf..c1e92b69ec 100644
--- a/fastdeploy/engine/sched/resource_manager_v1.py
+++ b/fastdeploy/engine/sched/resource_manager_v1.py
@@ -333,7 +333,7 @@ def schedule(self):
                             # Prepare decoding task
                             scheduled_reqs.append(self._prepare_decode_task(request))
                         num_decoding_req_nums += 1
-                        token_budget -= 1
+                    token_budget -= 1
                 else:  # need to prefill
                     llm_logger.debug(
                         f"scheduler prefill task: {request} request.need_prefill_tokens {request.need_prefill_tokens} request.num_computed_tokens {request.num_computed_tokens}"
diff --git a/fastdeploy/worker/gpu_model_runner.py b/fastdeploy/worker/gpu_model_runner.py
index 3817b49326..a51b5c2a2e 100644
--- a/fastdeploy/worker/gpu_model_runner.py
+++ b/fastdeploy/worker/gpu_model_runner.py
@@ -1167,11 +1167,12 @@ def initialize_kv_cache(self, profile: bool = False) -> None:
                     fill_value=0,
                     dtype=cache_type,
                 )
-                cache_kvs[f"value_caches_{i}"] = paddle.full(
-                    shape=kv_cache_shape,
-                    fill_value=0,
-                    dtype=cache_type,
-                )
+                if "deepseek" not in self.fd_config.model_config.model_type:
+                    cache_kvs[f"value_caches_{i}"] = paddle.full(
+                        shape=kv_cache_shape,
+                        fill_value=0,
+                        dtype=cache_type,
+                    )
                 if kv_cache_quant_type == "block_wise_fp8":
                     cache_kvs[f"key_cache_scales_{i}"] = paddle.full(
                         shape=kv_cache_scale_shape,
@@ -1898,7 +1899,15 @@ def cal_theortical_kvcache(self):
             if self.speculative_method in ["mtp"]
             else self.model_config.num_hidden_layers
         )
-        required_memory = byte_of_dtype * 2 * (self.cache_config.block_size * hidden_dim) * num_layers  # k + v
+        if "deepseek" in self.fd_config.model_config.model_type:
+            required_memory = (
+                byte_of_dtype
+                * (self.fd_config.model_config.kv_lora_rank + self.fd_config.model_config.qk_rope_head_dim)
+                * (self.cache_config.block_size)
+                * num_layers
+            )  # compress_kv + k_pe
+        else:
+            required_memory = byte_of_dtype * 2 * (self.cache_config.block_size * hidden_dim) * num_layers  # k + v
         return required_memory
 
     def not_need_stop(self) -> bool: