sdgoij/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-03-17 16:44:07 +00:00
Code Issues 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

15
common/common.cpp
View File

@@ -1223,7 +1223,7 @@ common_init_result_ptr common_init_from_params(common_params & params) {
return res; return res;
} }
int err = llama_apply_adapter_cvec( int err = llama_set_adapter_cvec(
lctx, lctx,
cvec.data.data(), cvec.data.data(),
cvec.data.size(), cvec.data.size(),
@@ -1325,12 +1325,15 @@ std::string get_model_endpoint() {
} }
void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora) { void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora) {
llama_clear_adapter_lora(ctx); std::vector<llama_adapter_lora *> loras;
for (auto & la : lora) { std::vector<float> scales;
if (la.scale != 0.0f) {
llama_set_adapter_lora(ctx, la.ptr, la.scale); for (auto & la: lora) {
} loras.push_back(la.ptr);
scales.push_back(la.scale);
} }
llama_set_adapters_lora(ctx, loras.data(), loras.size(), scales.data());
} }
struct llama_model_params common_model_params_to_llama(common_params & params) { struct llama_model_params common_model_params_to_llama(common_params & params) {

21
include/llama.h
View File

@@ -656,21 +656,12 @@ extern "C" {
// The following functions operate on a llama_context, hence the naming: llama_verb_... // The following functions operate on a llama_context, hence the naming: llama_verb_...
// Add a loaded LoRA adapter to given context // Set LoRa adapters on the context. Will only modify if the adapters currently in context are different.
// This will not modify model's weight LLAMA_API int32_t llama_set_adapters_lora(
LLAMA_API int32_t llama_set_adapter_lora(
struct llama_context * ctx, struct llama_context * ctx,
struct llama_adapter_lora * adapter, struct llama_adapter_lora ** adapters,
float scale); size_t n_adapters,
float * scales);
// Remove a specific LoRA adapter from given context
// Return -1 if the adapter is not present in the context
LLAMA_API int32_t llama_rm_adapter_lora(
struct llama_context * ctx,
struct llama_adapter_lora * adapter);
// Remove all LoRA adapters from given context
LLAMA_API void llama_clear_adapter_lora(struct llama_context * ctx);
// Apply a loaded control vector to a llama_context, or if data is NULL, clear // Apply a loaded control vector to a llama_context, or if data is NULL, clear
// the currently loaded vector. // the currently loaded vector.
@@ -678,7 +669,7 @@ extern "C" {
// to an n_embd x n_layers buffer starting from layer 1. // to an n_embd x n_layers buffer starting from layer 1.
// il_start and il_end are the layer range the vector should apply to (both inclusive) // il_start and il_end are the layer range the vector should apply to (both inclusive)
// See llama_control_vector_load in common to load a control vector. // See llama_control_vector_load in common to load a control vector.
LLAMA_API int32_t llama_apply_adapter_cvec( LLAMA_API int32_t llama_set_adapter_cvec(
struct llama_context * ctx, struct llama_context * ctx,
const float * data, const float * data,
size_t len, size_t len,

103
src/llama-context.cpp
View File

@@ -1057,51 +1057,43 @@ bool llama_context::set_sampler(llama_seq_id seq_id, llama_sampler * sampler) {
return true; return true;
} }
void llama_context::set_adapter_lora( void llama_context::set_adapters_lora(llama_adapter_lora ** adapters, size_t n_adapters, float * scales) {
llama_adapter_lora * adapter, LLAMA_LOG_DEBUG("%s: adapters = %p\n", __func__, (void *) adapters);
float scale) {
LLAMA_LOG_DEBUG("%s: adapter = %p, scale = %f\n", __func__, (void *) adapter, scale);
if (auto it = loras.find(adapter); it != loras.end()) { if (adapters_lora_are_same(adapters, n_adapters, scales)) {
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()) {
return; return;
} }
loras.clear(); 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; 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, const float * data,
size_t len, size_t len,
int32_t n_embd, 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 // llama adapter API
int32_t llama_set_adapter_lora( int32_t llama_set_adapters_lora(
llama_context * ctx, llama_context * ctx,
llama_adapter_lora * adapter, llama_adapter_lora ** adapters,
float scale) { size_t n_adapters,
ctx->set_adapter_lora(adapter, scale); 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; return 0;
} }
int32_t llama_rm_adapter_lora( int32_t llama_set_adapter_cvec(
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(
llama_context * ctx, llama_context * ctx,
const float * data, const float * data,
size_t len, size_t len,
int32_t n_embd, int32_t n_embd,
int32_t il_start, int32_t il_start,
int32_t il_end) { int32_t il_end) {
bool res = ctx->apply_adapter_cvec(data, len, n_embd, il_start, il_end); bool res = ctx->set_adapter_cvec(data, len, n_embd, il_start, il_end);
return res ? 0 : -1; return res ? 0 : -1;
} }

11
src/llama-context.h
View File

@@ -105,16 +105,11 @@ struct llama_context {
void set_causal_attn(bool value); void set_causal_attn(bool value);
void set_warmup(bool value); void set_warmup(bool value);
void set_adapter_lora( void set_adapters_lora(llama_adapter_lora ** adapters, size_t n_adapters, float * scales);
llama_adapter_lora * adapter,
float scale);
bool rm_adapter_lora( bool adapters_lora_are_same(llama_adapter_lora ** adapters, size_t n_adapters, float * scales);
llama_adapter_lora * adapter);
void clear_adapter_lora(); bool set_adapter_cvec(
bool apply_adapter_cvec(
const float * data, const float * data,
size_t len, size_t len,
int32_t n_embd, int32_t n_embd,