vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

llama : rename expert_weights_b to exp_probs_b

Browse source
This commit is contained in:
Stanisław Szymczyk 2025-01-03 14:57:56 +01:00
parent 140eb29264
commit 5b4673b3dd
1 changed files with 3 additions and 3 deletions
Download patch file Download diff file Expand all files Collapse all files

6
src/llama.cpp
View file

@ -10264,7 +10264,7 @@ static struct ggml_tensor * llm_build_moe_ffn(
struct ggml_tensor * up_exps, struct ggml_tensor * up_exps,
struct ggml_tensor * gate_exps, struct ggml_tensor * gate_exps,
struct ggml_tensor * down_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,
int64_t n_expert_used, int64_t n_expert_used,
llm_ffn_op_type type_op, 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 // add experts selection bias - introduced in DeepSeek V3
// leave probs unbiased as it's later used to get expert weights // leave probs unbiased as it's later used to get expert weights
ggml_tensor * selection_probs = probs; ggml_tensor * selection_probs = probs;
if (expert_weights_b != nullptr) { if (exp_probs_b != nullptr) {
selection_probs = ggml_add(ctx, probs, expert_weights_b); selection_probs = ggml_add(ctx, probs, exp_probs_b);
cb(selection_probs, "ffn_moe_probs_biased", il); cb(selection_probs, "ffn_moe_probs_biased", il);
} }