sleepy/llama.cpp
1
0
Fork 0
Code Issues 8 Pull Requests 1 Actions 176 Packages Projects Releases Wiki Activity

llama : update LoRA API. + fix excessive graph reserves (#19280)

Browse Source 
* Refactoring to use new llama_put_adapter_loras

* cont : alternative lora API

---------

Co-authored-by: Jake Chavis <jakechavis6@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
agent-enemy-2
2026-02-14 03:06:27 -05:00
committed by GitHub
parent eb145c0753
commit 2d8015e8a4
4 changed files with 62 additions and 88 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llama-context.cpp
+44 -59
View File
@@ -1057,51 +1057,43 @@ bool llama_context::set_sampler(llama_seq_id seq_id, llama_sampler * sampler) {
return true;
}
void llama_context::set_adapter_lora(
llama_adapter_lora * adapter,
float scale) {
LLAMA_LOG_DEBUG("%s: adapter = %p, scale = %f\n", __func__, (void *) adapter, scale);
void llama_context::set_adapters_lora(llama_adapter_lora ** adapters, size_t n_adapters, float * scales) {
LLAMA_LOG_DEBUG("%s: adapters = %p\n", __func__, (void *) adapters);
if (auto it = loras.find(adapter); it != loras.end()) {
if (it->second == scale) {
return;
}
}
loras[adapter] = scale;
sched_need_reserve = true;
}
bool llama_context::rm_adapter_lora(
llama_adapter_lora * adapter) {
LLAMA_LOG_DEBUG("%s: adapter = %p\n", __func__, (void *) adapter);
auto it = loras.find(adapter);
if (it != loras.end()) {
loras.erase(it);
sched_need_reserve = true;
return true;
}
return false;
}
void llama_context::clear_adapter_lora() {
LLAMA_LOG_DEBUG("%s: call\n", __func__);
if (loras.empty()) {
if (adapters_lora_are_same(adapters, n_adapters, scales)) {
return;
}
loras.clear();
for (size_t i = 0; i < n_adapters; i ++) {
if (scales[i] != 0.0f) {
loras[adapters[i]] = scales[i];
}
}
sched_need_reserve = true;
}
bool llama_context::apply_adapter_cvec(
bool llama_context::adapters_lora_are_same(llama_adapter_lora ** adapters, size_t n_adapters, float * scales) {
LLAMA_LOG_DEBUG("%s: adapters = %p\n", __func__, (void *) adapters);
if (n_adapters != loras.size()) {
return false;
}
for (size_t i = 0; i < n_adapters; i ++) {
auto it = loras.find(adapters[i]);
if (it == loras.end() || it->second != scales[i]) {
return false;
}
}
return true;
}
bool llama_context::set_adapter_cvec(
const float * data,
size_t len,
int32_t n_embd,
@@ -3209,35 +3201,28 @@ uint32_t llama_get_sampled_probs_count_ith(llama_context * ctx, int32_t i) {
// llama adapter API
int32_t llama_set_adapter_lora(
int32_t llama_set_adapters_lora(
llama_context * ctx,
llama_adapter_lora * adapter,
float scale) {
ctx->set_adapter_lora(adapter, scale);
llama_adapter_lora ** adapters,
size_t n_adapters,
float * scales) {
if (adapters == nullptr || scales == nullptr) {
GGML_ASSERT(n_adapters == 0 && "invalid llama_set_adapters_lora call");
}
ctx->set_adapters_lora(adapters, n_adapters, scales);
return 0;
}
int32_t llama_rm_adapter_lora(
llama_context * ctx,
llama_adapter_lora * adapter) {
bool res = ctx->rm_adapter_lora(adapter);
return res ? 0 : -1;
}
void llama_clear_adapter_lora(llama_context * ctx) {
ctx->clear_adapter_lora();
}
int32_t llama_apply_adapter_cvec(
int32_t llama_set_adapter_cvec(
llama_context * ctx,
const float * data,
size_t len,
int32_t n_embd,
int32_t il_start,
int32_t il_end) {
bool res = ctx->apply_adapter_cvec(data, len, n_embd, il_start, il_end);
const float * data,
size_t len,
int32_t n_embd,
int32_t il_start,
int32_t il_end) {
bool res = ctx->set_adapter_cvec(data, len, n_embd, il_start, il_end);
return res ? 0 : -1;
}