[NemotronH] Use ReplicatedLinear for fc1_latent_proj (#16569)

Signed-off-by: Roi Koren <roik@nvidia.com>
2026-01-14 11:08:27 +02:00
parent ba625c2d90
commit 72bacc88c8
1 changed files with 1 additions and 2 deletions
--- a/python/sglang/srt/models/nemotron_h.py
+++ b/python/sglang/srt/models/nemotron_h.py
@@ -191,12 +191,11 @@ class NemotronHMoE(nn.Module):
            self.shared_experts = None

        if self.use_latent_moe:
-            self.fc1_latent_proj = ColumnParallelLinear(
+            self.fc1_latent_proj = ReplicatedLinear(
                input_size=config.hidden_size,
                output_size=self.moe_hidden_size,
                bias=config.mlp_bias,
                quant_config=quant_config,
-                gather_output=True,
                prefix=f"{prefix}.fc1_latent_proj",
            )
            self.fc2_latent_proj = ReplicatedLinear(