[XPU] refine fused moe (#4219)

custom_ops/xpu_ops/src/ops/moe_layer.cc

@@ -228,8 +228,9 @@ MoeLayer(const paddle::Tensor &x, const paddle::Tensor &gate_weight,
                quant_method == "weight_only_int4") {
         APPLY_MOE_LAYER_KERNEL(paddle::bfloat16, int4_t);
     } else {
-        PD_THROW("MoeLayer not support x_type==%d, w_type==%d",
-                 static_cast<int>(x_type), static_cast<int>(w_type));
+        PD_THROW("MoeLayer not support x_type=", static_cast<int>(x_type),
+                 ", w_type=", static_cast<int>(w_type),
+                 ", quant_method=", quant_method);
         return {};
     }
 #undef APPLY_MOE_LAYER_KERNEL

fastdeploy/model_executor/layers/backends/xpu/__init__.py

@@ -16,16 +16,11 @@
 xpu backend methods
 """
 
-from .moe.fused_moe import (
-    XPUMoEMethod,
-    XPUWeightOnlyMoeEpMethod,
-    XPUWeightOnlyMoEMethod,
-)
+from .moe.fused_moe import XPUMoEMethod, XPUWeightOnlyMoEMethod
 from .quantization.weight_only import XPUWeightOnlyLinearMethod
 
 __all__ = [
     "XPUWeightOnlyLinearMethod",
     "XPUMoEMethod",
     "XPUWeightOnlyMoEMethod",
-    "XPUWeightOnlyMoeEpMethod",
 ]

fastdeploy/model_executor/layers/moe/fused_moe_backend_base.py

@@ -30,15 +30,22 @@ class MoEMethodBase(QuantMethodBase):
 
     def __init__(self, quant_config):
         super().__init__()
-        if quant_config is None:
+        self.quant_config = quant_config
+        if self.quant_config is None:
             self.moe_quant_type = "w16a16"
+        elif hasattr(quant_config, "algo"):
+            self.moe_quant_type = quant_config.algo
         else:
-            self.quant_config = quant_config
+            self.moe_quant_type = quant_config.name()
         self.added_weight_attrs = ["up_gate_proj_weight", "down_proj_weight"]
         self.added_scale_attrs = [
             "up_gate_proj_weight_scale",
             "down_proj_weight_scale",
         ]
+        self.added_in_scale_attrs = [
+            "up_gate_proj_in_scale",
+            "down_proj_in_scale",
+        ]
         self.pack_num = 1
         self.ep_prefill_runner = None
         self.ep_decoder_runner = None

fastdeploy/model_executor/layers/quantization/weight_only.py

@@ -79,18 +79,11 @@ class WeightOnlyConfig(QuantConfigBase):
     def get_quant_method(self, layer) -> Optional[QuantMethodBase]:
         if current_platform.is_xpu():
             if isinstance(layer, FusedMoE):
-                if layer.ep_size > 1:
-                    from fastdeploy.model_executor.layers.backends import (
-                        XPUWeightOnlyMoeEpMethod,
-                    )
+                from fastdeploy.model_executor.layers.backends import (
+                    XPUWeightOnlyMoEMethod,
+                )
 
-                    return XPUWeightOnlyMoeEpMethod(self)
-                else:
-                    from fastdeploy.model_executor.layers.backends import (
-                        XPUWeightOnlyMoEMethod,
-                    )
-
-                    return XPUWeightOnlyMoEMethod(self)
+                return XPUWeightOnlyMoEMethod(self)
             else:
                 from fastdeploy.model_executor.layers.backends import (
                     XPUWeightOnlyLinearMethod,

fastdeploy/worker/worker_process.py

@@ -404,9 +404,9 @@ class PaddleDisWorkerProc:
 
             if num_blocks_local <= 0:
                 raise ValueError(
-                    "The total number of blocks cannot be less than zero."
-                    "Please increase gpu_memory_utilization"
-                    "Or decrease max_num_batched_tokens(max model length) "
+                    "The total number of blocks cannot be less than zero. "
+                    "Please increase gpu_memory_utilization "
+                    "Or decrease max_num_batched_tokens(max model length)."
                 )
 
             if self.ranks > 1:

fastdeploy/worker/xpu_model_runner.py

@@ -1227,7 +1227,8 @@ class XPUModelRunner(ModelRunnerBase):
         """
         Clear the block tables and kv cache after profiling.
         """
-        del self.share_inputs["caches"]
+        if hasattr(self.share_inputs, "caches"):
+            del self.share_inputs["caches"]
         if self.forward_meta is not None:
             del self.forward_meta.caches
         paddle.device.xpu.empty_cache()