remove load_up_proj_weight_first (#6932)

fastdeploy/model_executor/layers/moe/moe.py

@@ -322,8 +322,8 @@ class FusedMoE(nn.Layer):
                 shard_dim=SHARD_ID_TO_SHARDED_DIM[shard_id],
             )
 
-    def _load_gate_up_weight(self, param, expert_id, loaded_weight, shard_id, shard_dim=None, is_sharded=False):
-        if self.tp_size > 1 and not is_sharded and not self.fd_config.load_config.is_pre_sharded:
+    def _load_gate_up_weight(self, param, expert_id, loaded_weight, shard_id, shard_dim=None):
+        if self.tp_size > 1 and not self.fd_config.load_config.is_pre_sharded:
             tp_shard_dim = shard_dim
             weight_dim = -1 if tp_shard_dim else 0
             size = loaded_weight.shape[weight_dim]
@@ -334,8 +334,7 @@ class FusedMoE(nn.Layer):
         expert_param = param[expert_id - self.expert_id_offset]
         dim = -1 if shard_dim else 0
         param_shard_size = expert_param.shape[dim] // 2
-        switch_w13 = getattr(self.quant_method, "load_up_proj_weight_first", False)
-        if (shard_id == "gate" and not switch_w13) or (shard_id == "up" and switch_w13):
+        if shard_id == "gate":
             param_shard_offset = 0
         else:
             param_shard_offset = param_shard_size

fastdeploy/model_executor/layers/quantization/nvfp4.py

@@ -503,14 +503,16 @@ class ModelOptNvFp4FusedMoE(MoEMethodBase):
         set_weight_attrs(layer.up_gate_proj_input_scale, {**extra_weight_attrs, "weight_type": "input_scale"})
         set_weight_attrs(layer.down_proj_input_scale, {**extra_weight_attrs, "weight_type": "input_scale"})
 
-    @property
-    def load_up_proj_weight_first(self) -> bool:
-        # FlashInfer CUTLASS kernel assumes [Up, Gate] Proj as W13
-        # 目前默认给True
-        return True
-
     def process_weights_after_loading(self, layer):
         """ """
+
+        # FlashInfer CUTLASS kernel assumes [Up, Gate] Proj as W13
+
+        [a, b] = layer.up_gate_proj_weight.split(2, axis=1)
+        layer.up_gate_proj_weight.set_value(paddle.concat([b, a], axis=1))
+        [a, b] = layer.up_gate_proj_weight_scale.split(2, axis=1)
+        layer.up_gate_proj_weight_scale.set_value(paddle.concat([b, a], axis=1))
+
         up_gate_proj_weight_scale_2 = layer.up_gate_proj_weight_scale_2[:, 0]
         free_tensor(layer.up_gate_proj_weight_scale_2)
         create_parameter_and_copy(layer, name="up_gate_proj_weight_scale_2", weight=up_gate_proj_weight_scale_2)

tests/quantization/test_modelopt_nvfp4.py

@@ -432,7 +432,6 @@ class TestModelOptNvFp4FusedMoE(unittest.TestCase):
             scale = paddle.ones([1, 64, 16], dtype=paddle.float16)
             swizzled = _process_scale_interleaved(scale)
         self.assertEqual(list(swizzled.shape), [1, 128, 16])
-        self.assertTrue(method.load_up_proj_weight_first)
 
     def test_process_weights_after_loading(self):
         """Test post-loading weight processing logic for FusedMoE layers."""