Do not apply experts weight scale if it is 1

This commit is contained in:
Iwan Kawrakow
2025-10-20 08:21:21 +03:00
parent 599c812f12
commit acb0bc63fc

View File

@@ -863,7 +863,7 @@ llm_expert_gating_func_type gating_op,
weights = ggml_reshape_3d(ctx, weights, 1, n_expert_used, n_tokens);
}
if (scale_w) {
if (scale_w && std::abs(w_scale-1) > 1e-5f) {
weights = ggml_scale(ctx, weights, w_scale);
cb(weights, "ffn_moe_weights_scaled", il);
}