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
[Bugfix] Correct LayerNorm epsilon parameter in modernbert.py (#27008)
Signed-off-by: bogdanm <152898065+bogdan01m@users.noreply.github.com> Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
bogdanm committed
Oct 16, 2025 at 17:27 UTC
dcbb3f18715b8c1e23d63bc6daa80482552ebcfa
1 file changed
+5
-2
vllm/model_executor/models/modernbert.py
+5
-2
@@ -39,9 +39,12 @@ class ModernBertEmbeddings(nn.Module):
39
self.tok_embeddings = VocabParallelEmbedding(
40
config.vocab_size, config.hidden_size
41
)
42
- self.norm = nn.LayerNorm(
43
- config.hidden_size, eps=config.layer_norm_eps, bias=config.norm_bias
42
+ eps = (
43
+ getattr(config, "norm_eps", None)
44
+ or getattr(config, "layer_norm_eps", None)
45
+ or 1e-5
46
)
47
+ self.norm = nn.LayerNorm(config.hidden_size, eps=eps, bias=config.norm_bias)
48
49
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
50
return self.tok_embeddings(input_ids)