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 9, 02:30 UTC
Fix accessing hidden_act from model config (#32686)
Signed-off-by: Grzegorz Karch <gkarch@nvidia.com>
Grzegorz K. Karch committed
Feb 2, 2026 at 12:11 UTC
78061ef584ad1bc124b10f7515edaf879464f6ff
1 file changed
+6
-1
vllm/model_executor/models/nemotron_nas.py
+6
-1
@@ -177,10 +177,15 @@ class DeciLMDecoderLayer(nn.Module):
177
else:
178
intermediate_size = block_config.ffn.intermediate_size
179
180
+ if hasattr(block_config.ffn, "hidden_act"):
181
+ hidden_act = block_config.ffn.hidden_act
182
+ else:
183
+ hidden_act = config.hidden_act
184
+
185
self.mlp = LlamaMLP(
186
hidden_size=self.hidden_size,
187
intermediate_size=intermediate_size,
183
- hidden_act=config.hidden_act,
188
+ hidden_act=hidden_act,
189
quant_config=quant_config,
190
bias=getattr(config, "mlp_bias", False),
191
prefix=f"{prefix}.mlp",