Fix Qwen3-VL M-RoPE on the Transformers modeling backend (grids + compile) (#49292)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com> Co-authored-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Aritra Roy Gosthipaty committed
Jul 21, 2026 at 23:58 UTC
33178f90063e5eec7d3e2a6288f3ca43d39ae8d3
1 file changed
+3
-7
vllm/model_executor/models/transformers/multimodal.py
+3
-7
@@ -337,7 +337,7 @@ class MultiModalMixin(SupportsMultiModal, SupportsMRoPE):
337
# Positions shape handling for MRoPE models
338
if self.model_config.uses_mrope:
339
# [3, seq_len] -> [3, 1, seq_len]
340
- positions = positions[:, None]
340
+ positions = positions[:, None].contiguous()
341
model_output = super().forward(
342
input_ids, positions, intermediate_tensors, inputs_embeds
343
)
@@ -481,12 +481,8 @@ class MultiModalMixin(SupportsMultiModal, SupportsMRoPE):
481
image_grid_thw = kwargs.get("image_grid_thw", [])
482
video_grid_thw = kwargs.get("video_grid_thw", [])
483
484
- image_grid_thw = (torch.stack if image_grid_thw else torch.tensor)(
485
- image_grid_thw
486
- )
487
- video_grid_thw = (torch.stack if video_grid_thw else torch.tensor)(
488
- video_grid_thw
489
- )
484
+ image_grid_thw = torch.stack(image_grid_thw) if image_grid_thw else None
485
+ video_grid_thw = torch.stack(video_grid_thw) if video_grid_thw else None
486
487
# `get_rope_index` doesn't always accept arbitrary `kwargs`
488
kwargs = {}