From 5b4673b3dd8e65f74b81538f992395a89180e1f9 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Stanis=C5=82aw=20Szymczyk?= Date: Fri, 3 Jan 2025 14:57:56 +0100 Subject: [PATCH] llama : rename expert_weights_b to exp_probs_b --- src/llama.cpp | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/src/llama.cpp b/src/llama.cpp index 1ab930e3e..9e1094f8d 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -10264,7 +10264,7 @@ static struct ggml_tensor * llm_build_moe_ffn( struct ggml_tensor * up_exps, struct ggml_tensor * gate_exps, struct ggml_tensor * down_exps, - struct ggml_tensor * expert_weights_b, + struct ggml_tensor * exp_probs_b, int64_t n_expert, int64_t n_expert_used, llm_ffn_op_type type_op, @@ -10298,8 +10298,8 @@ llm_expert_gating_func_type gating_op, // add experts selection bias - introduced in DeepSeek V3 // leave probs unbiased as it's later used to get expert weights ggml_tensor * selection_probs = probs; - if (expert_weights_b != nullptr) { - selection_probs = ggml_add(ctx, probs, expert_weights_b); + if (exp_probs_b != nullptr) { + selection_probs = ggml_add(ctx, probs, exp_probs_b); cb(selection_probs, "ffn_moe_probs_biased", il); }