[FDConfig]Remove splitwise_role and engine_worker_queue_port in FDConfig (#4147)

* remove splitwise_role and engine_worker_queue_port

* fix xpu

* fix xpu

* fix xpu

* fix unittest

* resolve conflct

Author: YuanRisheng

fastdeploy/config.py

@@ -296,8 +296,6 @@ def __init__(
         # Do profile or not
         self.do_profile: bool = False
 
-        # splitwise role
-        self.splitwise_role: str = "mixed"
         # guided decoding backend
         self.guided_decoding_backend: str = None
         # disable any whitespace for guided decoding
@@ -319,14 +317,6 @@ def __init__(
         else:
             self.expert_parallel_size = 1
         self.use_ep = self.expert_parallel_size > 1
-        if self.splitwise_role == "mixed":
-            self.moe_phase = MoEPhase(phase="prefill")
-        elif self.splitwise_role == "prefill":
-            self.moe_phase = MoEPhase(phase="prefill")
-        elif self.splitwise_role == "decode":
-            self.moe_phase = MoEPhase(phase="decode")
-        else:
-            raise NotImplementedError
 
         # pd_disaggregation
         use_pd_disaggregation: int = int(os.getenv("FLAGS_use_pd_disaggregation", 0))
@@ -1116,10 +1106,8 @@ def __init__(
         max_model_len: int = 8192,
         ips: str = None,
         use_warmup: bool = False,
-        engine_worker_queue_port: str = "8002",
         limit_mm_per_prompt: Optional[Dict[str, Any]] = None,
         mm_processor_kwargs: Optional[Dict[str, Any]] = None,
-        splitwise_role: str = "mixed",
         innode_prefill_ports: Optional[List[int]] = None,
         max_num_partial_prefills: int = 1,
         max_long_partial_prefills: int = 1,
@@ -1182,19 +1170,14 @@ def __init__(
         self.limit_mm_per_prompt = limit_mm_per_prompt
         self.mm_processor_kwargs = mm_processor_kwargs
         self.use_warmup = use_warmup
-        self.splitwise_role = splitwise_role
         self.innode_prefill_ports = innode_prefill_ports
         self.max_num_partial_prefills = max_num_partial_prefills
         self.max_long_partial_prefills = max_long_partial_prefills
         self.long_prefill_token_threshold = long_prefill_token_threshold
         self.reasoning_parser = reasoning_parser
         self.guided_decoding_backend = guided_decoding_backend
         self.disable_any_whitespace = disable_any_whitespace
-        self.engine_worker_queue_port = engine_worker_queue_port
         self._str_to_list("innode_prefill_ports", int)
-        if isinstance(engine_worker_queue_port, int):
-            self.engine_worker_queue_port = str(engine_worker_queue_port)
-        self._str_to_list("engine_worker_queue_port", str)
 
         if envs.FD_FOR_TORCH_MODEL_FORMAT:
             self.model_config.model_format = "torch"
@@ -1267,6 +1250,15 @@ def postprocess(self):
             else:
                 self.guided_decoding_backend = "xgrammar"
 
+        if self.scheduler_config.splitwise_role == "mixed":
+            self.model_config.moe_phase = MoEPhase(phase="prefill")
+        elif self.scheduler_config.splitwise_role == "prefill":
+            self.model_config.moe_phase = MoEPhase(phase="prefill")
+        elif self.scheduler_config.splitwise_role == "decode":
+            self.model_config.moe_phase = MoEPhase(phase="decode")
+        else:
+            raise NotImplementedError
+
     def check(self):
         """
         check the legality of config
@@ -1301,7 +1293,7 @@ def check(self):
             f"max_long_partial_prefills: {self.max_long_partial_prefills} should "
             f"be less than or equal to max_num_partial_prefills: {self.max_num_partial_prefills}"
         )
-        assert self.splitwise_role in ["mixed", "prefill", "decode"]
+        assert self.scheduler_config.splitwise_role in ["mixed", "prefill", "decode"]
         # TODO(@wufeisheng): TP and EP need to be supported simultaneously.
         assert (self.parallel_config.tensor_parallel_size == 1 and self.parallel_config.expert_parallel_size >= 1) or (
             self.parallel_config.tensor_parallel_size >= 1 and self.parallel_config.expert_parallel_size == 1
@@ -1387,16 +1379,18 @@ def init_cache_info(self):
         initialize cache info
         """
         disaggregate_info = {}
-        if self.splitwise_role != "mixed":
-            disaggregate_info["role"] = self.splitwise_role
+        if self.scheduler_config.splitwise_role != "mixed":
+            disaggregate_info["role"] = self.scheduler_config.splitwise_role
             disaggregate_info["cache_info"] = dict()
             current_protocol = self.cache_config.cache_transfer_protocol.split(",")
             disaggregate_info["transfer_protocol"] = current_protocol
             for protocol in current_protocol:
                 if protocol == "ipc":
                     disaggregate_info["cache_info"][protocol] = {
                         "ip": self.host_ip,
-                        "port": self.engine_worker_queue_port[self.parallel_config.local_data_parallel_id],
+                        "port": self.parallel_config.engine_worker_queue_port[
+                            self.parallel_config.local_data_parallel_id
+                        ],
                         "device_ids": self.local_device_ids,
                     }
                 elif protocol == "rdma":

fastdeploy/engine/args_utils.py

@@ -1019,6 +1019,11 @@ def create_engine_config(self) -> FDConfig:
                 else:
                     self.max_num_batched_tokens = self.max_model_len
 
+        if isinstance(self.engine_worker_queue_port, int):
+            self.engine_worker_queue_port = str(self.engine_worker_queue_port)
+        if isinstance(self.engine_worker_queue_port, str):
+            self.engine_worker_queue_port = self.engine_worker_queue_port.split(",")
+
         all_dict = asdict(self)
         all_dict["model_cfg"] = model_cfg
         cache_cfg = CacheConfig(all_dict)
@@ -1032,11 +1037,6 @@ def create_engine_config(self) -> FDConfig:
         early_stop_cfg = self.create_early_stop_config()
         early_stop_cfg.update_enable_early_stop(self.enable_early_stop)
 
-        if isinstance(self.engine_worker_queue_port, int):
-            self.engine_worker_queue_port = str(self.engine_worker_queue_port)
-        if isinstance(self.engine_worker_queue_port, str):
-            self.engine_worker_queue_port = self.engine_worker_queue_port.split(",")
-
         assert is_port_available(
             "0.0.0.0", int(self.engine_worker_queue_port[parallel_cfg.local_data_parallel_id])
         ), f"The parameter `engine_worker_queue_port`:{self.engine_worker_queue_port} is already in use."
@@ -1052,12 +1052,10 @@ def create_engine_config(self) -> FDConfig:
             speculative_config=speculative_cfg,
             ips=self.ips,
             use_warmup=self.use_warmup,
-            engine_worker_queue_port=self.engine_worker_queue_port,
             limit_mm_per_prompt=self.limit_mm_per_prompt,
             mm_processor_kwargs=self.mm_processor_kwargs,
             reasoning_parser=self.reasoning_parser,
             tool_parser=self.tool_call_parser,
-            splitwise_role=self.splitwise_role,
             innode_prefill_ports=self.innode_prefill_ports,
             max_num_partial_prefills=self.max_num_partial_prefills,
             max_long_partial_prefills=self.max_long_partial_prefills,

fastdeploy/engine/common_engine.py

@@ -76,10 +76,10 @@ def __init__(self, cfg, start_queue=True):
                 cfg.scheduler_config.max_num_seqs,
                 cfg,
                 cfg.parallel_config.tensor_parallel_size,
-                cfg.splitwise_role,
+                cfg.scheduler_config.splitwise_role,
                 cfg.parallel_config.local_data_parallel_id,
             )
-            if cfg.splitwise_role != "mixed":
+            if cfg.scheduler_config.splitwise_role != "mixed":
                 raise NotImplementedError(
                     "Currently ENABLE_V1_KVCACHE_SCHEDULER=1 only supported in mixed sampling now."
                 )
@@ -88,13 +88,13 @@ def __init__(self, cfg, start_queue=True):
                 cfg.scheduler_config.max_num_seqs,
                 cfg,
                 cfg.parallel_config.tensor_parallel_size,
-                cfg.splitwise_role,
+                cfg.scheduler_config.splitwise_role,
                 cfg.parallel_config.local_data_parallel_id,
             )
 
         self.start_worker_queue_service(start_queue)
 
-        os.environ["INFERENCE_MSG_QUEUE_ID"] = self.cfg.engine_worker_queue_port[
+        os.environ["INFERENCE_MSG_QUEUE_ID"] = self.cfg.parallel_config.engine_worker_queue_port[
             self.cfg.parallel_config.local_data_parallel_id
         ]
 
@@ -137,7 +137,9 @@ def start(self):
         self.token_processor.run()
 
     def _init_worker_monitor_signals(self):  # exist_task_signal 用于各worker进程感知是否有新Task需要处理
-        current_suffix = int(self.cfg.engine_worker_queue_port[self.cfg.parallel_config.local_data_parallel_id])
+        current_suffix = int(
+            self.cfg.parallel_config.engine_worker_queue_port[self.cfg.parallel_config.local_data_parallel_id]
+        )
         llm_logger.info(f"current_suffix: {current_suffix}")
         exist_task_signal_data = np.zeros([1], dtype=np.int32)
         self.exist_task_signal = IPCSignal(
@@ -195,7 +197,7 @@ def start_worker_queue_service(self, start_queue):
         """
         address = (
             self.cfg.master_ip,
-            int(self.cfg.engine_worker_queue_port[self.cfg.parallel_config.local_data_parallel_id]),
+            int(self.cfg.parallel_config.engine_worker_queue_port[self.cfg.parallel_config.local_data_parallel_id]),
         )
 
         if start_queue and (self.cfg.host_ip == self.cfg.master_ip or self.cfg.master_ip == "0.0.0.0"):
@@ -209,7 +211,7 @@ def start_worker_queue_service(self, start_queue):
 
             if (
                 self.cfg.cache_config.enable_prefix_caching
-                or self.cfg.splitwise_role != "mixed"
+                or self.cfg.scheduler_config.splitwise_role != "mixed"
                 and self.cfg.parallel_config.local_data_parallel_id == 0
             ):
                 self.cache_task_queue = EngineCacheQueue(
@@ -253,7 +255,10 @@ def insert_tasks(self, tasks, current_id=-1, allocated=False):
                 del self.resource_manager.req_dict[task.request_id]
                 cur_task = self.resource_manager.tasks_list[cur_task_idx]
                 cur_task.prompt_token_ids[0] = task.outputs.token_ids[0]
-                if self.cfg.speculative_config.method in ["mtp"] and self.cfg.splitwise_role == "decode":
+                if (
+                    self.cfg.speculative_config.method in ["mtp"]
+                    and self.cfg.scheduler_config.splitwise_role == "decode"
+                ):
                     cur_task.draft_token_ids = copy.deepcopy(task.outputs.draft_token_ids)
                 if task.error_code != 200:
                     self.resource_manager.stop_flags[cur_task_idx] = True
@@ -478,7 +483,10 @@ def _insert_task_to_worker(self):
                     time.sleep(0.001)
                     continue
                 if hasattr(self, "exist_prefill_task_signal") and self.exist_prefill_task_signal.value[0] > 0:
-                    if self.cfg.splitwise_role == "mixed" or self.split_connector.has_splitwise_tasks():
+                    if (
+                        self.cfg.scheduler_config.splitwise_role == "mixed"
+                        or self.split_connector.has_splitwise_tasks()
+                    ):
                         time.sleep(0.005)
                         continue
                 if self.engine_worker_queue.num_cache_infos() > 0:
@@ -507,7 +515,7 @@ def _insert_task_to_worker(self):
                     continue
 
                 current_id = (current_id + 1) % 100003
-                if self.cfg.splitwise_role != "mixed":
+                if self.cfg.scheduler_config.splitwise_role != "mixed":
                     llm_logger.info("Inserting splitwise tasks")
                     self.split_connector.send_splitwise_tasks(tasks, current_id)
 
@@ -759,7 +767,7 @@ def start_cache_service(self, device_ids, ipc_signal_suffix):
             device_ids=device_ids,
             pod_ip=self.cfg.master_ip,
             engine_worker_queue_port=int(
-                self.cfg.engine_worker_queue_port[self.cfg.parallel_config.local_data_parallel_id]
+                self.cfg.parallel_config.engine_worker_queue_port[self.cfg.parallel_config.local_data_parallel_id]
             ),
             pid_suffix=ipc_signal_suffix,
         )

fastdeploy/engine/engine.py

@@ -115,7 +115,7 @@ def start(self, api_server_pid=None):
         start_time = time.time()
 
         self.api_server_pid = api_server_pid
-        self.ipc_signal_suffix = self.cfg.engine_worker_queue_port[0]
+        self.ipc_signal_suffix = self.cfg.parallel_config.engine_worker_queue_port[0]
         self._init_worker_signals()
 
         self.data_processor = self.input_processor.create_processor()
@@ -127,7 +127,7 @@ def start(self, api_server_pid=None):
             self.engine.start_zmq_service(api_server_pid)
 
         if self.do_profile == 0 and (
-            self.cfg.cache_config.enable_prefix_caching or self.cfg.splitwise_role != "mixed"
+            self.cfg.cache_config.enable_prefix_caching or self.cfg.scheduler_config.splitwise_role != "mixed"
         ):
             device_ids = self.cfg.device_ids.split(",")
             self.cache_manager_processes = self.engine.start_cache_service(device_ids, self.ipc_signal_suffix)
@@ -161,7 +161,7 @@ def check_worker_initialize_status_func(res: dict):
             self._stop_profile()
         # Launch components: scheduler, cache_manager, expert_service et.al.
         self.launch_components()
-        if self.cfg.cache_config.enable_prefix_caching or self.cfg.splitwise_role != "mixed":
+        if self.cfg.cache_config.enable_prefix_caching or self.cfg.scheduler_config.splitwise_role != "mixed":
             self.launched_cache_manager_signal.value[0] = 1
 
         # Worker launched
@@ -311,7 +311,7 @@ def _init_worker_signals(self):
         )
 
         # launched_cache_manager_signal 用于感知engine是否启动了cache_manager
-        if self.cfg.cache_config.enable_prefix_caching or self.cfg.splitwise_role != "mixed":
+        if self.cfg.cache_config.enable_prefix_caching or self.cfg.scheduler_config.splitwise_role != "mixed":
             launched_cache_manager_signal_data = np.zeros([1], dtype=np.int32)
             self.launched_cache_manager_signal = IPCSignal(
                 name="launched_cache_manager_signal",
@@ -426,10 +426,10 @@ def _setting_environ_variables(self):
             }
         )
 
-        if self.cfg.splitwise_role != "mixed":
+        if self.cfg.scheduler_config.splitwise_role != "mixed":
             variables["FLAGS_use_pd_disaggregation"] = 1
             # TODO dynamic load environment variable
-            if self.cfg.splitwise_role == "prefill":
+            if self.cfg.scheduler_config.splitwise_role == "prefill":
                 variables["FLAGS_fmt_write_cache_completed_signal"] = 1
 
         if self.cfg.model_config.enable_mm:
@@ -463,7 +463,7 @@ def _start_worker_service(self):
             else len(self.data_processor.tokenizer.vocab)
         )
 
-        ports = ",".join(self.cfg.engine_worker_queue_port)
+        ports = ",".join(self.cfg.parallel_config.engine_worker_queue_port)
         ips = None
         if self.cfg.ips is not None:
             ips = ",".join(self.cfg.ips)
@@ -481,9 +481,9 @@ def _start_worker_service(self):
             f" --enc_dec_block_num {self.cfg.cache_config.enc_dec_block_num}"
             f" --eos_tokens_lens {self.data_processor.eos_token_id_len}"
             f" --pad_token_id {self.data_processor.pad_token_id}"
-            f" --engine_pid {self.cfg.engine_worker_queue_port[0]}"
+            f" --engine_pid {self.cfg.parallel_config.engine_worker_queue_port[0]}"
             f" --max_num_batched_tokens {self.cfg.scheduler_config.max_num_batched_tokens}"
-            f" --splitwise_role {self.cfg.splitwise_role}"
+            f" --splitwise_role {self.cfg.scheduler_config.splitwise_role}"
             f" --kv_cache_ratio {self.cfg.cache_config.kv_cache_ratio}"
             f" --expert_parallel_size {self.cfg.parallel_config.expert_parallel_size}"
             f" --data_parallel_size {self.cfg.parallel_config.data_parallel_size}"
@@ -602,7 +602,7 @@ def _stop_profile(self):
         num_gpu_blocks = self.get_profile_block_num_signal.value[0]
         self.cfg.cache_config.reset(num_gpu_blocks)
         self.engine.resource_manager.reset_cache_config(self.cfg.cache_config)
-        if self.cfg.cache_config.enable_prefix_caching or self.cfg.splitwise_role != "mixed":
+        if self.cfg.cache_config.enable_prefix_caching or self.cfg.scheduler_config.splitwise_role != "mixed":
             device_ids = self.cfg.device_ids.split(",")
             self.cache_manager_processes = self.engine.start_cache_service(device_ids, self.ipc_signal_suffix)
 
@@ -619,7 +619,7 @@ def check_health(self, time_interval_threashold=30):
         return True, ""
 
     def launch_components(self):
-        if self.cfg.splitwise_role != "mixed":
+        if self.cfg.scheduler_config.splitwise_role != "mixed":
             # 单机逻辑
             self.engine.engine_worker_queue.available_prefill_instances.put(1)
             self.engine.split_mode_get_tasks()
@@ -632,7 +632,7 @@ def launch_components(self):
 
         self.cfg.init_cache_info()
 
-        role = self.cfg.splitwise_role
+        role = self.cfg.scheduler_config.splitwise_role
         host_ip = self.cfg.host_ip
         disaggregate = self.cfg.disaggregate_info
         if self.cfg.scheduler_config.name == "splitwise":
@@ -649,7 +649,7 @@ def launch_components(self):
                 ):
                     address = (
                         self.cfg.master_ip,
-                        int(self.cfg.engine_worker_queue_port[i]),
+                        int(self.cfg.parallel_config.engine_worker_queue_port[i]),
                     )
                     llm_logger.info(f"dp start queue service {address}")
                     self.dp_engine_worker_queue_server.append(

fastdeploy/engine/expert_service.py

@@ -50,13 +50,13 @@ def __init__(self, cfg, local_data_parallel_id, start_queue=True):
         self.cfg = cfg
         start_pos = (local_data_parallel_id * self.cfg.parallel_config.tensor_parallel_size) % cfg.worker_num_per_node
         end_pos = start_pos + self.cfg.parallel_config.tensor_parallel_size
-        if cfg.splitwise_role != "mixed":
+        if cfg.scheduler_config.splitwise_role != "mixed":
             self.cfg.cache_config.rdma_comm_ports = self.cfg.cache_config.rdma_comm_ports[start_pos:end_pos]
         self.cfg.local_device_ids = self.cfg.device_ids.split(",")[start_pos:end_pos]
         llm_logger.info(f"local_data_parallel_id: {local_data_parallel_id}")
         self.cfg.disaggregate_info = None
 
-        if cfg.splitwise_role != "mixed":
+        if cfg.scheduler_config.splitwise_role != "mixed":
             if len(self.cfg.cache_config.pd_comm_port) == 1:
                 self.cfg.cache_config.pd_comm_port[0] = (
                     int(self.cfg.cache_config.pd_comm_port[0]) + local_data_parallel_id
@@ -84,21 +84,21 @@ def start(self, ipc_signal_suffix, local_data_parallel_id):
             self.api_server_pid = ipc_signal_suffix
             self.engine.start_zmq_service(ipc_signal_suffix)
         else:
-            ipc_signal_suffix = self.cfg.engine_worker_queue_port[0]
+            ipc_signal_suffix = self.cfg.parallel_config.engine_worker_queue_port[0]
 
         llm_logger.info(f"start expert service {local_data_parallel_id}")
-        if self.cfg.splitwise_role != "mixed":
+        if self.cfg.scheduler_config.splitwise_role != "mixed":
             self.engine.start_cache_service(self.cfg.local_device_ids, ipc_signal_suffix)
             self.engine.split_mode_get_tasks()
 
         if self.cfg.scheduler_config.name == "splitwise":
             self.cfg.init_cache_info()
-            role = self.cfg.splitwise_role
+            role = self.cfg.scheduler_config.splitwise_role
             host_ip = self.cfg.host_ip
             disaggregate = self.cfg.disaggregate_info
             self.engine.scheduler.start(role, host_ip, disaggregate)
 
-        if self.cfg.splitwise_role != "mixed":
+        if self.cfg.scheduler_config.splitwise_role != "mixed":
             self.splitwise_receive_thread = threading.Thread(
                 target=self.engine.split_connector.start_receiver, args=()
             )