update fleet_ops (#7859)

* update fleet_ops

* use try_import in ep

Author: liuruyan

fastdeploy/model_executor/layers/moe/ep.py

@@ -27,8 +27,11 @@
 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:
     """
@@ -43,7 +46,7 @@ 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
+                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

@@ -14,46 +14,19 @@
 # limitations under the License.
 """
 
-import importlib
-
 import paddle
 import triton
 from paddleformers.utils.log import logger
 
+from fastdeploy.model_executor.layers.utils import get_sm_version
 from fastdeploy.model_executor.ops.triton_ops import _per_token_group_quant_fp8
+from fastdeploy.model_executor.utils import try_import
 from fastdeploy.platforms import current_platform
 
 if current_platform.is_cuda():
     from fastdeploy.model_executor.ops.gpu import per_token_group_fp8_quant
 
 
-from ..utils import get_sm_version
-
-
-def try_import(modules, name=None, fail_msg=None):
-    """
-    try_import
-    """
-    if not isinstance(modules, (list, tuple)):
-        modules = [modules]
-
-    for m in modules:
-        assert isinstance(m, str), m
-        try:
-            m = importlib.import_module(m)
-        except ImportError:
-            m = None
-
-        if m is not None:
-            if name is None:
-                return m
-            elif hasattr(m, name):
-                return getattr(m, name)
-
-    if fail_msg is not None:
-        logger.warning(fail_msg)
-
-
 paddlefleet_ops = try_import(["paddlefleet.ops"])
 
 
@@ -72,7 +45,7 @@ def load_deep_gemm():
             try:
                 import logging
 
-                import paddlefleet.ops.deep_gemm as deep_gemm
+                import paddlefleet_ops.deep_gemm as deep_gemm
 
                 logging.getLogger().handlers.clear()
                 logger.info("Detected sm100, use PaddleFleet DeepGEMM")

fastdeploy/model_executor/utils.py

@@ -555,6 +555,30 @@ def fn(loaded_weight_name, is_moe):
     return fn
 
 
+def try_import(modules, name=None, fail_msg=None):
+    """
+    try_import
+    """
+    if not isinstance(modules, (list, tuple)):
+        modules = [modules]
+
+    for m in modules:
+        assert isinstance(m, str), m
+        try:
+            m = importlib.import_module(m)
+        except ImportError:
+            m = None
+
+        if m is not None:
+            if name is None:
+                return m
+            elif hasattr(m, name):
+                return getattr(m, name)
+
+    if fail_msg is not None:
+        logger.warning(fail_msg)
+
+
 def has_flashinfer():
     return importlib.util.find_spec("flashinfer") is not None