sleepy/llama.cpp
1
0
Fork 0
Code Issues 16 Pull Requests Actions 70 Packages Projects Releases Wiki Activity

ggml : add ggml_top_k (#17365)

Browse Source 
* ggml : add ggml_top_k

* cont : add ggml_argsort_top_k

* metal : add top_k support

* ggml : cleanup

* tests : add virtual err() function for test_case

* ggml : add comments
This commit is contained in:
Georgi Gerganov
2025-11-25 15:31:43 +02:00
committed by GitHub
parent 05872ac885
commit 583cb83416
15 changed files with 516 additions and 85 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llama-graph.cpp
+3 -3
View File
@@ -961,14 +961,14 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
// 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]
ggml_tensor * group_scores = ggml_argsort_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]
ggml_tensor * expert_groups = ggml_argsort_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
@@ -979,7 +979,7 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
}
// select experts
ggml_tensor * selected_experts = ggml_top_k(ctx0, selection_probs, n_expert_used); // [n_expert_used, n_tokens]
ggml_tensor * selected_experts = ggml_argsort_top_k(ctx0, selection_probs, n_expert_used); // [n_expert_used, n_tokens]
cb(selected_experts->src[0], "ffn_moe_argsort", il);
cb(selected_experts, "ffn_moe_topk", il);