for ci

Author: yangjianfengo1

fastdeploy/model_executor/layers/moe/fused_moe_cutlass_backend.py

@@ -736,6 +736,8 @@ def process_prequanted_weights(self, layer: nn.Layer, state_dict):
         """
         Paddle cutlass process prequanted weights.
         """
+        if state_dict is None:
+            return
         up_gate_proj_expert_weight_key = layer.weight_key_map.get("up_gate_proj_expert_weight_key", None)
         down_proj_expert_weight_key = layer.weight_key_map.get("down_proj_expert_weight_key", None)
         up_gate_proj_expert_weight_scale_key = layer.weight_key_map.get("up_gate_proj_expert_weight_scale_key", None)

test/layers/test_w4afp8.py

@@ -0,0 +1,26 @@
+import paddle
+from paddle import nn
+
+from fastdeploy.model_executor.layers.moe.fused_moe_cutlass_backend import (
+    CutlassW4AFP8MoEMethod,
+)
+
+
+class W4AFP8Layer(nn.Layer):
+    def __init__(self):
+        super().__init__()
+        self.weight_key_map = {}
+        self.ep_size = 1
+
+    def load_experts_weight(self, state_dict, up_gate_proj_expert_weight_key, down_proj_expert_weight_key):
+        return (
+            paddle.zeros([64, 128, 128]),
+            paddle.zeros([64, 128, 128]),
+            paddle.zeros([64, 128, 128]),
+            paddle.zeros([64, 128, 128]),
+        )
+
+
+w4afp8 = CutlassW4AFP8MoEMethod({})
+layer = W4AFP8Layer()
+w4afp8.process_prequanted_weights(layer, None)