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 7, 05:15 UTC
ParakeetProjection.norm = RMSNorm instead of nn.LayerNorm (#36133)
Signed-off-by: Netanel Haber <58652339+netanel-haber@users.noreply.github.com>
Netanel Haber committed
Mar 5, 2026 at 19:29 UTC
b93a9e6f6d91baf59e39089ce8dbf2f2a3f0f6c9
1 file changed
+2
-1
vllm/model_executor/models/parakeet.py
+2
-1
@@ -14,6 +14,7 @@ from transformers import ParakeetEncoder as HFParakeetEncoder
14
from transformers import ParakeetFeatureExtractor, PretrainedConfig
15
16
from vllm.model_executor.layers.activation import ReLUSquaredActivation
17
+from vllm.model_executor.layers.layernorm import RMSNorm
18
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
19
from vllm.transformers_utils.configs.parakeet import ExtractorConfig, ParakeetConfig
20
@@ -26,7 +27,7 @@ class ParakeetProjection(nn.Module):
27
llm_hidden_size = config.llm_hidden_size
28
bias = config.projection_bias
29
29
- self.norm = nn.LayerNorm(sound_hidden_size, eps=config.projection_eps)
30
+ self.norm = RMSNorm(sound_hidden_size, eps=config.projection_eps)
31
self.linear1 = nn.Linear(sound_hidden_size, proj_hidden_size, bias=bias)
32
self.activation = ReLUSquaredActivation()
33
self.linear2 = nn.Linear(proj_hidden_size, llm_hidden_size, bias=bias)