[BugFix]fix v1 loader lm head fp32 (#5270) (#5287)

fastdeploy/model_executor/layers/mtp_linear.py

@@ -120,10 +120,10 @@ class ParallelEHProjection(nn.Layer):
         weight_tensor = get_tensor(state_dict.pop(self.weight_key)).astype(paddle.get_default_dtype())
         if self.linear.weight.shape != weight_tensor.shape:
             weight_tensor = weight_tensor.transpose([1, 0])
-        self.linear.weight.set_value(weight_tensor)
+        self.linear.weight.set_value(weight_tensor.astype(self.linear.weight.dtype))
 
         if self.bias_key is not None:
-            bias = get_tensor(state_dict.pop(self.bias_key)).astype(paddle.get_default_dtype())
+            bias = get_tensor(state_dict.pop(self.bias_key)).astype(self.linear.bias.dtype)
             self.linear.bias.set_value(bias)
 
     def forward(self, input):

fastdeploy/model_executor/models/qwen2_5_vl/qwen2_5_vl.py

@@ -251,7 +251,9 @@ class Qwen2_5_VLForConditionalGeneration(ModelForCasualLM):
         self.model.load_state_dict(state_dict)
         self.visual.load_state_dict(state_dict)
         if self.tie_word_embeddings:
-            self.lm_head.linear.weight.set_value(self.model.embed_tokens.embeddings.weight.transpose([1, 0]))
+            self.lm_head.linear.weight.set_value(
+                self.model.embed_tokens.embeddings.weight.transpose([1, 0]).astype(self.lm_head.linear.weight.dtype)
+            )
         else:
             self.lm_head.load_state_dict(state_dict)
 

tests/entrypoints/test_generation.py

@@ -50,6 +50,7 @@ class TestGeneration(unittest.TestCase):
                 model=MODEL_NAME,
                 max_num_batched_tokens=4096,
                 tensor_parallel_size=1,
+                lm_head_fp32=True,
                 engine_worker_queue_port=int(os.getenv("FD_ENGINE_QUEUE_PORT")),
                 cache_queue_port=int(os.getenv("FD_CACHE_QUEUE_PORT")),
             )