ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-19 12:50:11 +00:00
Code Issues Packages Projects Releases Wiki Activity

Slightly better

Browse Source
This commit is contained in:
Kawrakow
2025-11-30 07:51:34 +00:00
parent c37c1bdc33
commit bbb1b1da6c
1 changed files with 5 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
src/llama-build-context.cpp
View File

@@ -1130,7 +1130,6 @@ llm_expert_gating_func_type gating_op,
if (shared_out->ne[1] > 32) {
shared_out = ggml_cast(ctx, shared_out, GGML_TYPE_F16);
}
ggml_build_forward_expand(graph, shared_out);
results[id] = shared_out;
}
cur = ggml_add(ctx, results[0], results[1]);
@@ -1140,10 +1139,12 @@ llm_expert_gating_func_type gating_op,
cur = ggml_add(ctx, cur, results[id]);
cb(cur, "ffn_shared_combined", il);
}
if (cur->type == GGML_TYPE_F16) {
cur = ggml_cast(ctx, cur, GGML_TYPE_F32);
if (routed_out->ne[1] > 32) {
auto routed_out_f16 = ggml_cast(ctx, routed_out, GGML_TYPE_F16);
cur = ggml_add(ctx, routed_out_f16, cur);
} else {
cur = ggml_add(ctx, routed_out, cur);
}
cur = ggml_add(ctx, routed_out, cur);
cb(cur, "ffn_out", il);
} else {
//printf("Using non-split ffn for shared experts in layer %d\n", il);