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 11, 08:45 UTC
Fix nemotron_nas intermediate_size computation (#30795)
Signed-off-by: Grzegorz Karch <gkarch@nvidia.com>
Grzegorz K. Karch committed
Dec 17, 2025 at 02:06 UTC
f5db6385a19b04e76b5834618305485753e75544
1 file changed
+7
-4
vllm/model_executor/models/nemotron_nas.py
+7
-4
@@ -169,10 +169,13 @@ class DeciLMDecoderLayer(nn.Module):
169
self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
170
171
if not self._is_no_op_ffn:
172
- ffn_mult = block_config.ffn.ffn_mult
173
- intermediate_size = _ffn_mult_to_intermediate_size(
174
- ffn_mult, config.hidden_size
175
- )
172
+ if hasattr(block_config.ffn, "ffn_mult"):
173
+ ffn_mult = block_config.ffn.ffn_mult
174
+ intermediate_size = _ffn_mult_to_intermediate_size(
175
+ ffn_mult, config.hidden_size
176
+ )
177
+ else:
178
+ intermediate_size = block_config.ffn.intermediate_size
179
180
self.mlp = LlamaMLP(
181
hidden_size=self.hidden_size,