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

Do not allocate KV cache for unused layers (#843)

Browse Source 
* Do not allocate KV cache for unused layers

* Do not apply experts weight scale if it is 1

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow
2025-10-20 10:09:39 +03:00
committed by GitHub
parent 36f9601e8d
commit 1f072ab135
2 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
src/llama-build-context.cpp
View File

@@ -863,7 +863,7 @@ llm_expert_gating_func_type gating_op,
weights = ggml_reshape_3d(ctx, weights, 1, n_expert_used, n_tokens);
}
if (scale_w) {
if (scale_w && std::abs(w_scale-1) > 1e-5f) {
weights = ggml_scale(ctx, weights, w_scale);
cb(weights, "ffn_moe_weights_scaled", il);
}

2
src/llama.cpp
View File

@@ -532,7 +532,7 @@ static bool llama_kv_cache_init(
const struct llama_hparams & hparams = model.hparams;
const int64_t n_layer = hparams.n_layer;
const int64_t n_layer = hparams.n_layer - hparams.nextn_predict_layers;
cache.has_shift = false;