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
Invert pattern order to make sure that out_proj layers are identified (#24781)
Signed-off-by: Alexandre Marques <almarque@redhat.com>
Alexandre Marques committed
Sep 12, 2025 at 17:45 UTC
c89ed8de43e9ce513cdbb683862b07984cf51aae
1 file changed
+3
-3
vllm/model_executor/models/voxtral.py
+3
-3
@@ -585,12 +585,12 @@ class VoxtralForConditionalGeneration(nn.Module, SupportsMultiModal,
585
r"language_model.model.layers.\1.mlp.down_proj"),
586
(r"layers\.(\d+)\.feed_forward\.w3",
587
r"language_model.model.layers.\1.mlp.up_proj"),
588
- (r"mm_whisper_embeddings\.whisper_encoder\.transformer\.layers\.(\d+)\.attention.wo",
589
- r"whisper_encoder.whisper_encoder.layers.\1.layers.self_attn.out_proj"
590
- ),
588
(r"mm_whisper_embeddings\.whisper_encoder\.transformer\.layers\.(\d+)\.attention.w(.*)",
589
r"whisper_encoder.whisper_encoder.layers.\1.layers.self_attn.\2_proj"
590
),
591
+ (r"mm_whisper_embeddings\.whisper_encoder\.transformer\.layers\.(\d+)\.attention.wo",
592
+ r"whisper_encoder.whisper_encoder.layers.\1.layers.self_attn.out_proj"
593
+ ),
594
(r"mm_whisper_embeddings\.whisper_encoder\.transformer\.layers\.(\d+)\.feed_forward.w(\d+)",
595
r"whisper_encoder.whisper_encoder.layers.\1.layers.mlp.fc\2"),
596
(r"mm_whisper_embeddings\.whisper_encoder\.conv_layers\.0",