Add rope_scaling to Qwen (#1210)

vllm/model_executor/models/qwen.py

@@ -8,7 +8,7 @@
 The input of the model is flattened to a 1D tensor of tokens. The model uses
 InputMetadata to extract the original 2D shape of the input.
 """
-from typing import List, Optional, Tuple
+from typing import Any, Dict, List, Optional, Tuple
 
 import torch
 from torch import nn
@@ -76,13 +76,12 @@ class QWenMLP(nn.Module):
 
 class QWenAttention(nn.Module):
 
-    def __init__(
-        self,
-        hidden_size: int,
-        num_heads: int,
-        max_position_embeddings: int,
-        rope_theta: float = 10000,
-    ):
+    def __init__(self,
+                 hidden_size: int,
+                 num_heads: int,
+                 max_position_embeddings: int,
+                 rope_theta: float = 10000,
+                 rope_scaling: Optional[Dict[str, Any]] = None):
         super().__init__()
         self.hidden_size = hidden_size
         tensor_model_parallel_world_size = get_tensor_model_parallel_world_size(
@@ -116,7 +115,7 @@ class QWenAttention(nn.Module):
             rotary_dim=self.head_dim,
             base=rope_theta,
             max_position=max_position_embeddings,
-        )
+            rope_scaling=rope_scaling)
 
     def forward(
         self,
@@ -144,10 +143,12 @@ class QWenBlock(nn.Module):
         self.ln_1 = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
 
         rope_theta = getattr(config, "rope_theta", 10000)
+        rope_scaling = getattr(config, "rope_scaling", None)
         self.attn = QWenAttention(config.hidden_size,
                                   config.num_attention_heads,
                                   config.max_position_embeddings,
-                                  rope_theta=rope_theta)
+                                  rope_theta=rope_theta,
+                                  rope_scaling=rope_scaling)
 
         self.ln_2 = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)