From 72bacc88c8a08c4a028ef452add7ccf550ae052e Mon Sep 17 00:00:00 2001 From: roikoren755 <26850796+roikoren755@users.noreply.github.com> Date: Wed, 14 Jan 2026 11:08:27 +0200 Subject: [PATCH] [NemotronH] Use ReplicatedLinear for fc1_latent_proj (#16569) Signed-off-by: Roi Koren --- python/sglang/srt/models/nemotron_h.py | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-) diff --git a/python/sglang/srt/models/nemotron_h.py b/python/sglang/srt/models/nemotron_h.py index a42b1f2c0..0ff9be822 100644 --- 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(