add fleet fallback

Author: xiaoguoguo626807

fastdeploy/config.py

@@ -60,7 +60,7 @@
 _ResolvedTask = Literal["generate", "encode", "embed"]
 
 # Model implementation backend options
-ModelImpl = Literal["auto", "fastdeploy", "paddleformers"]
+ModelImpl = Literal["auto", "fastdeploy", "paddleformers", "paddlefleet"]
 
 _RUNNER_CONVERTS: dict[RunnerType, list[ConvertType]] = {
     "generate": [],

fastdeploy/engine/args_utils.py

@@ -139,6 +139,7 @@ class EngineArgs:
     'auto': Use native FastDeploy implementation when available, fallback to PaddleFormers.
     'fastdeploy': Use only native FastDeploy implementations.
     'paddleformers': Use PaddleFormers backend with FastDeploy optimizations.
+    'paddlefleet': Use PaddleFleet backend.
     """
     override_pooler_config: Optional[Union[dict, PoolerConfig]] = None
     """
@@ -636,7 +637,7 @@ def __post_init__(self):
                     "kvcache_storage_backend is only supported when ENABLE_V1_KVCACHE_SCHEDULER=1"
                 )
 
-        valid_model_impls = ["auto", "fastdeploy", "paddleformers"]
+        valid_model_impls = ["auto", "fastdeploy", "paddleformers", "paddlefleet"]
         if self.model_impl not in valid_model_impls:
             raise NotImplementedError(
                 f"not support model_impl: '{self.model_impl}'. " f"Must be one of: {', '.join(valid_model_impls)}"
@@ -974,13 +975,14 @@ def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
         model_group.add_argument(
             "--model-impl",
             type=str,
-            choices=["auto", "fastdeploy", "paddleformers"],
+            choices=["auto", "fastdeploy", "paddleformers", "paddlefleet"],
             default=EngineArgs.model_impl,
             help=(
                 "Model implementation backend. "
                 "'auto': Use native FastDeploy when available, fallback to PaddleFormers. "
                 "'fastdeploy': Use only native FastDeploy implementations. "
                 "'paddleformers': Use PaddleFormers backend with FastDeploy optimizations."
+                "'paddlefleet': Use PaddleFleet backend."
             ),
         )
 

fastdeploy/model_executor/models/model_base.py

@@ -194,6 +194,8 @@ def _try_resolve_paddleformers(
         elif model_impl == "auto" and is_fallback:
             # Auto mode fallback when no native implementation exists
             backend_arch = "PaddleFormersForCausalLM"
+        elif model_impl == "paddlefleet":
+            backend_arch = "PaddleFleetForCausalLM"
         elif model_impl == "fastdeploy":
             return None
         else:

fastdeploy/model_executor/models/paddleformers/__init__.py

@@ -21,10 +21,12 @@
 )
 
 from .base import PaddleFormersModelBase
+from .base_fleet import PaddleFleetModelBase
 from .causallm import CausalLMMixin
 
 __all__ = [
     "PaddleFormersForCausalLM",
+    "PaddleFleetForCausalLM",
 ]
 
 
@@ -38,3 +40,14 @@ class PaddleFormersForCausalLM(CausalLMMixin, PaddleFormersModelBase, ModelForCa
     @classmethod
     def name(cls):
         return "PaddleFormersForCausalLM"
+
+
+@ModelRegistry.register_model_class(
+    architecture="PaddleFleetForCausalLM",
+    module_name="paddleformers",
+    category=ModelCategory.TEXT_GENERATION,
+)
+class PaddleFleetForCausalLM(PaddleFleetModelBase, ModelForCasualLM):
+    @classmethod
+    def name(cls):
+        return "PaddleFleetForCausalLM"