ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-01-26 17:20:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

Leave FFN partial results as f16

Browse Source
This commit is contained in:
Kawrakow
2025-11-28 07:25:20 +00:00
parent 259718f8cb
commit ec45020e37
1 changed files with 6 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

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

@@ -691,9 +691,9 @@ ggml_tensor * llm_build_context::llm_build_ffn(
if (ffn.size() > 2) {
cur->op_params[0] = 0xff;
}
if (cur->type != GGML_TYPE_F32) {
cur = ggml_cast(ctx, cur, GGML_TYPE_F32);
}
//if (cur->type != GGML_TYPE_F32) {
// cur = ggml_cast(ctx, cur, GGML_TYPE_F32);
//}
return cur;
}
@@ -9002,6 +9002,9 @@ ggml_tensor * llm_build_context::build_std_attention(ggml_cgraph * gf, ggml_tens
cur = llm_build_norm(ctx0, cur, hparams, split_norm, NULL, LLM_NORM_RMS, cb, il);
cb(cur, "attn_norm", il_cb);
}
else if (cur->type != GGML_TYPE_F32) {
cur = ggml_cast(ctx0, cur, GGML_TYPE_F32);
}
auto [Qcur, Kcur, Vcur] = llm_build_mul_mat_qkv(gf, cur, nullptr, nullptr, nullptr, nullptr,
split_wq, nullptr, split_wk, nullptr, split_wv, nullptr,
model.layers[il].attn_q_norm, model.layers[il].attn_k_norm, f_attn_scale, il_cb);