[bugfix]: qwen2_vl rope_type default (#1129)

Author: shihaobai

lightllm/models/llama/model.py

@@ -118,7 +118,9 @@ def _init_custom(self):
             scaling_type = rope_scaling["type"]
         else:
             raise ValueError(f"Unknown RoPE scaling format {rope_scaling}")
-        if scaling_type == "yarn":
+        if scaling_type == "default":
+            self._init_to_get_rotary()
+        elif scaling_type == "yarn":
             self._init_to_get_yarn_rotary()
         elif scaling_type == "dynamic":
             self._init_to_get_dynamic_ntk_rotary()

lightllm/models/qwen2_vl/infer_struct.py

@@ -11,6 +11,11 @@ def __init__(self):
         self.position_sin = None
 
     def init_some_extra_state(self, model, input_ids: torch.Tensor):
+        rope_scaling = model.config.get("rope_scaling", {})
+        self.rope_type = rope_scaling.get("rope_type", rope_scaling.get("type", None))
+        if self.rope_type != "mrope":
+            super().init_some_extra_state(model, input_ids)
+            return
         InferStateInfo.init_some_extra_state(self, model, input_ids)
         if self.is_prefill:
             position_ids = self.position_ids

lightllm/models/qwen2_vl/layer_infer/transformer_layer_infer.py

@@ -19,6 +19,8 @@ def __init__(self, layer_num, network_config, mode=[]):
         self.axis_map = torch.tensor(axis_map, dtype=torch.int32, device="cuda")
 
     def _get_qkv(self, input, infer_state, layer_weight):
+        if infer_state.rope_type != "mrope":
+            return super()._get_qkv(input, infer_state, layer_weight)
         q = layer_weight.q_proj.mm(input)
         cache_kv = layer_weight.kv_proj.mm(input).view(-1, (self.tp_k_head_num_ + self.tp_v_head_num_), self.head_dim_)
         seq_len, _ = q.shape