mikhail/llama.cpp-mtp-turboquant
1
0
Fork 0
Code Issues Pull Requests Actions 32 Packages Projects Releases Wiki Activity

model: using single llm_build per arch (#21970)

Browse Source 
* model: using single llm_build per arch

* fix merge

* nits
This commit is contained in:
Xuan-Son Nguyen
2026-04-16 21:10:22 +02:00
committed by GitHub
parent e45dbdece8
commit 4fbdabdc61
7 changed files with 139 additions and 128 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llama-model.cpp
+12 -19
View File
@@ -1271,8 +1271,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
}
// Set non-causal attention for diffusion models
hparams.causal_attn = false;
}
break;
} break;
case LLM_ARCH_LLADA:
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@@ -1286,8 +1285,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
}
// Set non-causal attention for diffusion models
hparams.causal_attn = false;
}
break;
} break;
case LLM_ARCH_LLADA_MOE:
{
ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp, false);
@@ -8553,9 +8551,9 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
case LLM_ARCH_LLAMA4:
{
if (hparams.swa_type == LLAMA_SWA_TYPE_NONE) {
llm = std::make_unique<llm_build_llama<false>>(*this, params);
llm = std::make_unique<llm_build_llama4<false>>(*this, params);
} else {
llm = std::make_unique<llm_build_llama_iswa>(*this, params);
llm = std::make_unique<llm_build_llama4<true>>(*this, params);
}
} break;
case LLM_ARCH_LLAMA_EMBED:
@@ -8633,23 +8631,19 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
case LLM_ARCH_DREAM:
{
llm = std::make_unique<llm_build_dream>(*this, params);
}
break;
} break;
case LLM_ARCH_LLADA:
{
llm = std::make_unique<llm_build_llada>(*this, params);
}
break;
} break;
case LLM_ARCH_LLADA_MOE:
{
llm = std::make_unique<llm_build_llada_moe>(*this, params);
}
break;
} break;
case LLM_ARCH_RND1:
{
llm = std::make_unique<llm_build_rnd1>(*this, params);
}
break;
} break;
case LLM_ARCH_QWEN2VL:
{
llm = std::make_unique<llm_build_qwen2vl>(*this, params);
@@ -8839,11 +8833,11 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
{
switch (params.gtype) {
case LLM_GRAPH_TYPE_ENCODER:
llm = std::make_unique<llm_build_t5_enc>(*this, params);
llm = std::make_unique<llm_build_t5<true>>(*this, params);
break;
case LLM_GRAPH_TYPE_DEFAULT:
case LLM_GRAPH_TYPE_DECODER:
llm = std::make_unique<llm_build_t5_dec>(*this, params);
llm = std::make_unique<llm_build_t5<false>>(*this, params);
break;
default:
GGML_ABORT("invalid graph type");
@@ -8851,9 +8845,8 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
} break;
case LLM_ARCH_T5ENCODER:
{
llm = std::make_unique<llm_build_t5_enc>(*this, params);
}
break;
llm = std::make_unique<llm_build_t5encoder>(*this, params);
} break;
case LLM_ARCH_JAIS:
{
llm = std::make_unique<llm_build_jais>(*this, params);