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

Guard against attempts to use MLA for non-MLA models (#320)

Browse Source 
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow
2025-04-08 08:47:24 +02:00
committed by GitHub
parent d223b26daf
commit c50d00d0dc
1 changed files with 17 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

25
src/llama.cpp
View File

@@ -3245,13 +3245,15 @@ static bool llama_kv_cache_init(
cache.ctxs.push_back(ctx);
}
cache.k_l.reserve(n_layer);
cache.v_l.reserve(n_layer);
// DeepSeek MLA
cache.kv_l.reserve(n_layer);
if (cparams.mla_attn == 1 && !cparams.flash_attn) {
cache.kvt_l.reserve(n_layer);
if (model.arch == LLM_ARCH_DEEPSEEK2 && cparams.mla_attn) {
// DeepSeek MLA
cache.kv_l.reserve(n_layer);
if (cparams.mla_attn == 1 && !cparams.flash_attn) {
cache.kvt_l.reserve(n_layer);
}
} else {
cache.k_l.reserve(n_layer);
cache.v_l.reserve(n_layer);
}
bool warn = true;
@@ -3299,7 +3301,7 @@ static bool llama_kv_cache_init(
cache.v_l.push_back(v);
}
}
if (cparams.mla_attn && n_mla < n_layer && n_mla > 0) {
if (model.arch == LLM_ARCH_DEEPSEEK2 && cparams.mla_attn && n_mla < n_layer && n_mla > 0) {
LLAMA_LOG_ERROR("%s: unexpected situation with %d out of %d layers having MLA enabled\n", __func__, n_mla, int(n_layer));
LLAMA_LOG_ERROR("%s: bailing out\n", __func__);
GGML_ABORT("fatal error");
@@ -18568,6 +18570,13 @@ struct llama_context * llama_new_context_with_model(
params.seed = time(NULL);
}
if (model->arch != LLM_ARCH_DEEPSEEK2 && cparams.mla_attn > 0) {
LLAMA_LOG_WARN("=====================================================================\n");
LLAMA_LOG_WARN(" MLA is only available for LLM_ARCH_DEEPSEEK2 -> turning off MLA\n");
LLAMA_LOG_WARN("=====================================================================\n");
cparams.mla_attn = 0;
}
LLAMA_LOG_INFO("%s: n_ctx = %u\n", __func__, cparams.n_ctx);
LLAMA_LOG_INFO("%s: n_batch = %u\n", __func__, cparams.n_batch);
LLAMA_LOG_INFO("%s: n_ubatch = %u\n", __func__, cparams.n_ubatch);