mikhail/llama.cpp-mtp-turboquant
1
0
Fork 0
Code Issues Pull Requests Actions 31 Packages Projects Releases Wiki Activity

model : support for LlamaBidirectionalModel architecture (#18220)

Browse Source 
* model: llama-embed-nemotron

* minor: python lint

* changed arch-name

* templated llm_build_llama to be used for both llama and llama-embed arch
This commit is contained in:
Saba Fallah
2025-12-24 14:02:36 +01:00
committed by GitHub
parent 2a9ea2020c
commit 54132f1b1f
7 changed files with 60 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llama-model.cpp
+10 -3
View File
@@ -606,7 +606,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
ml.get_key(LLM_KV_ROPE_DIMENSION_COUNT, hparams.n_rot, false);
if (arch == LLM_ARCH_LLAMA || arch == LLM_ARCH_DECI || arch == LLM_ARCH_FALCON) {
if (arch == LLM_ARCH_LLAMA || arch == LLM_ARCH_DECI || arch == LLM_ARCH_FALCON || arch == LLM_ARCH_LLAMA_EMBED) {
if (hparams.n_rot != hparams.n_embd_head_k) {
throw std::runtime_error(format("invalid n_rot: %u, expected %u", hparams.n_rot, hparams.n_embd_head_k));
}
@@ -630,6 +630,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
// arch-specific KVs
switch (arch) {
case LLM_ARCH_LLAMA:
case LLM_ARCH_LLAMA_EMBED:
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@@ -2652,6 +2653,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
case LLM_ARCH_GRANITE:
case LLM_ARCH_GRANITE_MOE:
case LLM_ARCH_MISTRAL3:
case LLM_ARCH_LLAMA_EMBED:
{
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
@@ -7269,16 +7271,20 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
switch (arch) {
case LLM_ARCH_LLAMA:
{
llm = std::make_unique<llm_build_llama>(*this, params);
llm = std::make_unique<llm_build_llama<false>>(*this, params);
} break;
case LLM_ARCH_LLAMA4:
{
if (hparams.swa_type == LLAMA_SWA_TYPE_NONE) {
llm = std::make_unique<llm_build_llama>(*this, params);
llm = std::make_unique<llm_build_llama<false>>(*this, params);
} else {
llm = std::make_unique<llm_build_llama_iswa>(*this, params);
}
} break;
case LLM_ARCH_LLAMA_EMBED:
{
llm = std::make_unique<llm_build_llama<true>>(*this, params);
} break;
case LLM_ARCH_DECI:
{
llm = std::make_unique<llm_build_deci>(*this, params);
@@ -7874,6 +7880,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
case LLM_ARCH_ERNIE4_5:
case LLM_ARCH_ERNIE4_5_MOE:
case LLM_ARCH_MISTRAL3:
case LLM_ARCH_LLAMA_EMBED:
return LLAMA_ROPE_TYPE_NORM;
// the pairs of head values are offset by n_rot/2