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
[Model][Bugfix] Fix MiDashengLM audio encoder mask by removing incorrect `logical_not` (#25925)
Signed-off-by: zhoukz <me@zhoukz.com>
Zhou Jiahao committed
Sep 30, 2025 at 16:15 UTC
2e1b8bc2b6d1796d65cea6ed514759d12a72cf17
1 file changed
+1
-2
vllm/model_executor/models/midashenglm.py
+1
-2
@@ -426,8 +426,7 @@ class DashengAudioTransformer(nn.Module):
426
assert x_length.ndim == 1, "Lengths are of size (B,)"
427
scaled_lengths = (x_length / (self.hop_length * 4)).long()
428
mask = self._to_mask(max_length=t, lengths=scaled_lengths)
429
- split_masks = mask.logical_not().split(target_length_in_patches,
430
- dim=-1)
429
+ split_masks = mask.split(target_length_in_patches, dim=-1)
430
else:
431
mask = None
432
split_masks = [None] * len(input_splits)