Reduce some magical numbers

Author: phoenixdong

tools/checkpoint/llava_onevision/convert_to_fs_qwen2.5_1.5b.py

@@ -224,9 +224,9 @@ def convert(input_path, output_path, tensor_parallel_size, use_te):
 
     # Process vision tower
     # Indices from mapping pytorch multihead attention to megatron.
-    kv_channels = 72
     hidden_dim = 1152
     num_heads = 16
+    kv_channels = hidden_dim // num_heads
     indices = []
     for i in range(num_heads):
         lb = i * kv_channels