From 1d5b79111ad48d1b7840cad839540c440b489317 Mon Sep 17 00:00:00 2001 From: Foke Singh Date: Mon, 9 Dec 2024 16:43:29 +0000 Subject: [PATCH] modernbert: set default epsilon value for embeddings layernorm. --- examples/modernbert/modernbert.zig | 1 + 1 file changed, 1 insertion(+) diff --git a/examples/modernbert/modernbert.zig b/examples/modernbert/modernbert.zig index 831610c..42eb939 100644 --- a/examples/modernbert/modernbert.zig +++ b/examples/modernbert/modernbert.zig @@ -56,6 +56,7 @@ pub const ModernBertModel = struct { pub fn init(self: *ModernBertModel, options: ModernBertOptions) void { self.options = options; self.final_norm.eps = 1e-5; + self.embeddings.norm.eps = 1e-5; for (self.layers, 0..) |*encoder_layer, layer_idx| { encoder_layer.attn.Wqkv.weight = encoder_layer.attn.Wqkv.weight.withSharding(.{0}); encoder_layer.attn.Wo.weight = encoder_layer.attn.Wo.weight.withSharding(.{1});