A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[Models][Qwen3VL] Optimise `_validate_and_reshape_mm_tensor` (#26426)
Signed-off-by: Lukas Geiger <lukas.geiger94@gmail.com>
Lukas Geiger committed
Oct 9, 2025 at 11:25 UTC
0426e3c5e149837c7ce74c35faaf3aea3542f327
1 file changed
+1
-1
vllm/model_executor/models/qwen3_vl.py
+1
-1
@@ -1304,7 +1304,7 @@ class Qwen3VLForConditionalGeneration(
1304
f"Got ndim: {mm_input.ndim} "
1305
f"(shape={mm_input.shape})"
1306
)
1307
- return torch.concat(list(mm_input))
1307
+ return mm_input.reshape(-1, mm_input.shape[-1])
1308
else:
1309
return torch.concat(mm_input)
1310