CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-07 09:57:00 +00:00
Code Issues Packages Projects Releases Wiki Activity

model : add BailingMoeV2 support (#16063)

Browse Source 
* add BailingMoeV2 support

* update llm types

* undo

* undo

* update llm types

* add model collection link

* update

* almost working

* correct group selection and rename n_group_exp

* avoid large top_k and use argmax instead for now

if we had something like argmax2 that would be equivalent, but this works fine until then

* poke

* skip group selection when there are no tokens

* fix 1T conversion

* hopefully fixed expert group selection

third time's the charm?

* make expert group selection generally available

The new LLaDA2Moe model uses this method too, make it generally available regardless of architecture.

* allow n_expert_groups to be 1 (Kimi K2)

* address review suggestions
This commit is contained in:
Sigbjørn Skjæret
2025-10-20 21:38:20 +02:00
committed by GitHub
parent c9c1972e2c
commit 84bf3c6778
15 changed files with 521 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

30
src/llama-graph.cpp
View File

@@ -950,6 +950,31 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
cb(selection_probs, "ffn_moe_probs_biased", il);
}
// select top n_group_used expert groups
// https://huggingface.co/deepseek-ai/DeepSeek-V3/blob/e815299b0bcbac849fa540c768ef21845365c9eb/modeling_deepseek.py#L440-L457
if (hparams.n_expert_groups > 1 && n_tokens > 0) {
const int64_t n_exp_per_group = n_expert / hparams.n_expert_groups;
// organize experts into n_expert_groups
ggml_tensor * selection_groups = ggml_reshape_3d(ctx0, selection_probs, n_exp_per_group, hparams.n_expert_groups, n_tokens); // [n_exp_per_group, n_expert_groups, n_tokens]
ggml_tensor * group_scores = ggml_top_k(ctx0, selection_groups, 2); // [2, n_expert_groups, n_tokens]
group_scores = ggml_get_rows(ctx0, ggml_reshape_4d(ctx0, selection_groups, 1, selection_groups->ne[0], selection_groups->ne[1], selection_groups->ne[2]), group_scores); // [1, 2, n_expert_groups, n_tokens]
// get top n_group_used expert groups
group_scores = ggml_sum_rows(ctx0, ggml_reshape_3d(ctx0, group_scores, group_scores->ne[1], group_scores->ne[2], group_scores->ne[3])); // [1, n_expert_groups, n_tokens]
group_scores = ggml_reshape_2d(ctx0, group_scores, group_scores->ne[1], group_scores->ne[2]); // [n_expert_groups, n_tokens]
ggml_tensor * expert_groups = ggml_top_k(ctx0, group_scores, hparams.n_group_used); // [n_group_used, n_tokens]
cb(expert_groups, "ffn_moe_group_topk", il);
// mask out the other groups
selection_probs = ggml_get_rows(ctx0, selection_groups, expert_groups); // [n_exp_per_group, n_group_used, n_tokens]
selection_probs = ggml_set_rows(ctx0, ggml_scale_bias(ctx0, selection_groups, 0.0f, -INFINITY), selection_probs, expert_groups); // [n_exp_per_group, n_expert_groups, n_tokens]
selection_probs = ggml_reshape_2d(ctx0, selection_probs, n_expert, n_tokens); // [n_expert, n_tokens]
cb(selection_probs, "ffn_moe_probs_masked", il);
}
// select experts
ggml_tensor * selected_experts = ggml_top_k(ctx0, selection_probs, n_expert_used); // [n_expert_used, n_tokens]
cb(selected_experts->src[0], "ffn_moe_argsort", il);
@@ -981,6 +1006,11 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
ggml_tensor * weights_sum = ggml_sum_rows(ctx0, weights); // [1, n_tokens]
cb(weights_sum, "ffn_moe_weights_sum", il);
if (arch == LLM_ARCH_BAILINGMOE2) {
weights_sum = ggml_scale_bias(ctx0, weights_sum, 1.0, 1e-20);
cb(weights_sum, "ffn_moe_weights_sum_biased", il);
}
weights = ggml_div(ctx0, weights, weights_sum); // [n_expert_used, n_tokens]
cb(weights, "ffn_moe_weights_norm", il);