fix ce bug (#7874)

Author: liuruyan

fastdeploy/model_executor/layers/moe/ep.py

@@ -27,11 +27,8 @@
 import fastdeploy
 from fastdeploy import envs
 from fastdeploy.config import MoEPhase
-from fastdeploy.model_executor.utils import try_import
 from fastdeploy.utils import singleton
 
-paddlefleet_ops = try_import(["paddlefleet.ops"])
-
 
 def load_deep_ep() -> ModuleType:
     """
@@ -46,7 +43,10 @@ def load_deep_ep() -> ModuleType:
             # Enable paddle.enable_compat before importing deep_ep (required by PFCC/PaddleFleet variants)
             paddle.enable_compat(scope={"deep_ep"})
             try:
-                import paddlefleet_ops.deep_ep as deep_ep  # type: ignore
+                try:
+                    import paddlefleet.ops.deep_ep as deep_ep  # type: ignore
+                except:
+                    import paddlefleet_ops.deep_ep as deep_ep  # type: ignore
 
                 logger.info("FD use PaddleFleet/DeepEP now.")
                 return deep_ep

fastdeploy/model_executor/layers/quantization/fp8_utils.py

@@ -27,7 +27,7 @@
     from fastdeploy.model_executor.ops.gpu import per_token_group_fp8_quant
 
 
-paddlefleet_ops = try_import(["paddlefleet.ops"])
+paddlefleet_ops = try_import(["paddlefleet.ops", "paddlefleet_ops"])
 
 
 def load_deep_gemm():
@@ -45,7 +45,10 @@ def load_deep_gemm():
             try:
                 import logging
 
-                import paddlefleet_ops.deep_gemm as deep_gemm
+                try:
+                    import paddlefleet.ops.deep_gemm as deep_gemm
+                except:
+                    import paddlefleet_ops.deep_gemm as deep_gemm
 
                 logging.getLogger().handlers.clear()
                 logger.info("Detected sm100, use PaddleFleet DeepGEMM")