update

Author: yuanlehome

fastdeploy/model_executor/model_loader/default_loader.py

@@ -95,31 +95,3 @@ def load_model(self, fd_config: FDConfig) -> nn.Layer:
         # TODO(gongshaotian): Now, only support safetensor
         self.load_weights(model, fd_config, architectures)
         return model
-
-    def load_rl_mock_model(self, fd_config: FDConfig) -> nn.Layer:
-        """use for rl model load"""
-        # (TODO:gaoziyuan) optimze
-        assert fd_config.load_config.load_strategy == "normal", fd_config.load_config.load_strategy
-        original_architectures = fd_config.model_config.architectures[0]
-        logger.info(f"Starting to load model {original_architectures}.")
-
-        import fastdeploy.rl  # noqa
-
-        if fd_config.speculative_config.model_type != "mtp":
-            model_architectures = original_architectures.replace("Ernie5ForCausalLM", "Ernie5MoeForCausalLM")
-        else:
-            model_architectures = original_architectures.replace("Ernie5ForCausalLM", "Ernie5MTPForCausalLM")
-
-        model_architectures += "RL"
-        context = contextlib.nullcontext()
-
-        with context:
-            model_cls = ModelRegistry.get_class(model_architectures)
-            model = model_cls(fd_config)
-
-        model.eval()
-
-        # normal strategy need load weight and architectures need without "RL"
-        self.load_weights(model, fd_config, original_architectures)
-        # RL model not need set_state_dict
-        return model

fastdeploy/model_executor/model_loader/default_loader_v1.py

@@ -98,30 +98,3 @@ def load_model(self, fd_config: FDConfig) -> nn.Layer:
             return model
         self.load_weights(model, fd_config, enable_cache)
         return model
-
-    def load_rl_mock_model(self, fd_config: FDConfig) -> nn.Layer:
-        """use for rl model load"""
-        # (TODO:gaoziyuan) optimze
-        assert fd_config.load_config.load_strategy == "normal", fd_config.load_config.load_strategy
-        original_architectures = fd_config.model_config.architectures[0]
-
-        import fastdeploy.rl  # noqa
-
-        if fd_config.speculative_config.model_type != "mtp":
-            model_architectures = original_architectures.replace("Ernie5ForCausalLM", "Ernie5MoeForCausalLM")
-        else:
-            model_architectures = original_architectures.replace("Ernie5ForCausalLM", "Ernie5MTPForCausalLM")
-
-        model_architectures += "RL"
-        context = paddle.LazyGuard()
-
-        with context:
-            model_cls = ModelRegistry.get_class(model_architectures)
-            model = model_cls(fd_config)
-
-        model.eval()
-
-        # normal strategy need load weight and architectures need without "RL"
-        self.load_weights(model, fd_config, original_architectures)
-        # RL model not need set_state_dict
-        return model

fastdeploy/rl/dynamic_weight_manager.py

@@ -14,7 +14,6 @@
 # limitations under the License.
 """
 
-import gc
 import os
 import time
 from multiprocessing.shared_memory import SharedMemory
@@ -88,7 +87,6 @@ def update_parameters(self, pid: int = 0, restart_process_group=False) -> None:
         strategy_handlers = {
             "ipc_snapshot": self._update_ipc_snapshot,
             "ipc": self._update_ipc,
-            "normal": self._normal_load_weight,
         }
 
         if handler := strategy_handlers.get(self.load_config.load_strategy):
@@ -103,18 +101,6 @@ def update_parameters(self, pid: int = 0, restart_process_group=False) -> None:
         # step5: recapture cuda_graph
         # step6: update weight status signal
 
-    def _normal_load_weight(self):
-        """use for RL mock."""
-        from fastdeploy.model_executor.model_loader import get_model_loader
-
-        model_loader = get_model_loader(load_config=self.fd_config.load_config)
-        model = model_loader.load_rl_mock_model(fd_config=self.fd_config)
-        state_dict = model.state_dict()
-        self._update_model_from_state(state_dict, "raw")
-        del model
-        del state_dict
-        gc.collect()
-
     def _update_ipc_snapshot(self):
         """Update using IPC snapshot strategy for elastic recovery."""
         model_path = os.path.join(