fix float16 support for kimi-vl (#17156)

Co-authored-by: zhouzaida <[email protected]>

Author: zhouzaida

vllm/model_executor/models/kimi_vl.py

@@ -340,8 +340,7 @@ def _parse_and_validate_image_input(
         else:
             pixel_values = pixel_values.reshape(-1, num_channels, patch_size,
                                                 patch_size)
-        # fp32 -> bf16
-        pixel_values = pixel_values.to(torch.bfloat16)
+        pixel_values = pixel_values.to(self.vision_tower.dtype)
         # image_grid_hws.shape = (N, 2)
         assert image_grid_hws.ndim == 2, f"unexpected shape for image_grid_hws: {image_grid_hws.shape}"