vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

lora : improve compat with mergekit-extract-lora (#11131)

Browse source 
* (wip) support mergekit-extracted lora

* support mergekit-extract-lora

* use lora->get_scale

* correct comment

* correct norm name & condition

* add some hints
This commit is contained in:
Xuan Son Nguyen 2025-01-08 15:59:53 +01:00 • committed by GitHub
parent c07d437bbd
commit 4d2b3d8804
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
4 changed files with 74 additions and 12 deletions
Download patch file Download diff file Expand all files Collapse all files

24
src/llama-adapter.cpp
View file

@ -242,6 +242,10 @@ static void llama_lora_adapter_init_impl(struct llama_model & model, const char
} else {
ab_map[name].b = cur;
}
} else if (str_endswith(name, "_norm.weight")) {
// TODO: add support for norm vector
// for now, we don't really care because most adapters still work fine without it
continue;
} else {
throw std::runtime_error("LoRA tensor '" + name + "' has unexpected suffix");
}
@ -251,6 +255,7 @@ static void llama_lora_adapter_init_impl(struct llama_model & model, const char
for (auto & it : ab_map) {
const std::string & name = it.first;
llama_lora_weight & w = it.second;
bool is_token_embd = str_endswith(name, "token_embd.weight");
if (!w.a || !w.b) {
throw std::runtime_error("LoRA tensor pair for '" + name + "' is missing one component");
@ -259,16 +264,23 @@ static void llama_lora_adapter_init_impl(struct llama_model & model, const char
// device buft and device ctx
auto * model_tensor = llama_model_get_tensor(model, name.c_str());
if (!model_tensor) {
throw std::runtime_error("LoRA tensor '" + name + "' does not exist in base model");
throw std::runtime_error("LoRA tensor '" + name + "' does not exist in base model (hint: maybe wrong base model?)");
}
struct ggml_context * dev_ctx = ctx_for_buft(ggml_backend_buffer_get_type(model_tensor->buffer));
// validate tensor shape
if (model_tensor->ne[0] != w.a->ne[0] || model_tensor->ne[1] != w.b->ne[1]) {
throw std::runtime_error("tensor '" + name + "' has incorrect shape");
}
if (w.a->ne[1] != w.b->ne[0]) {
throw std::runtime_error("lora_a tensor is not transposed (hint: adapter from \"finetune\" example is no longer supported)");
if (is_token_embd) {
// expect B to be non-transposed, A and B are flipped; see llm_build_inp_embd()
if (model_tensor->ne[0] != w.b->ne[1] || model_tensor->ne[1] != w.a->ne[1]) {
throw std::runtime_error("tensor '" + name + "' has incorrect shape (hint: maybe wrong base model?)");
}
} else {
if (model_tensor->ne[0] != w.a->ne[0] || model_tensor->ne[1] != w.b->ne[1]) {
throw std::runtime_error("tensor '" + name + "' has incorrect shape (hint: maybe wrong base model?)");
}
if (w.a->ne[1] != w.b->ne[0]) {
throw std::runtime_error("lora_a tensor is not transposed (hint: adapter from \"finetune\" example is no longer supported)");
}
}
// save tensor to adapter