ver4a/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

llama : refactor llama_context, llama_kv_cache, llm_build_context (#12181)

Browse source 
* llama : refactor llama_context, llama_kv_cache, llm_build_context

ggml-ci

* graph : don't mutate the KV cache during defrag

ggml-ci

* context : reduce virtuals + remove test function

ggml-ci

* context : move interface implementation to source file + factory

ggml-ci

* graph : move KV cache build functions to llama_context impl

ggml-ci

* graph : remove model reference from build_pooling

ggml-ci

* graph : remove llama_model reference

ggml-ci

* kv_cache : provide rope factors

ggml-ci

* graph : rework inputs to use only unique_ptr, remove attn input abstraction

ggml-ci

* context : remove llama_context_i abstraction

ggml-ci

* context : clean-up

ggml-ci

* graph : clean-up

ggml-ci

* llama : remove redundant keywords (struct, enum)

ggml-ci

* model : adapt gemma3

ggml-ci

* graph : restore same attention ops as on master

ggml-ci

* llama : remove TODO + fix indent

ggml-ci
This commit is contained in:
Georgi Gerganov 2025-03-13 12:35:44 +02:00 • committed by GitHub
parent 2048b5913d
commit e0dbec0bc6
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
46 changed files with 13903 additions and 12190 deletions
Download patch file Download diff file Expand all files Collapse all files

20
src/llama-adapter.h
View file

@ -15,11 +15,11 @@
//
struct llama_adapter_cvec {
struct ggml_tensor * tensor_for(int il) const;
ggml_tensor * tensor_for(int il) const;
struct ggml_tensor * apply_to(struct ggml_context * ctx, struct ggml_tensor * cur, int il) const;
ggml_tensor * apply_to(ggml_context * ctx, ggml_tensor * cur, int il) const;
int32_t apply(
bool apply(
const llama_model & model,
const float * data,
size_t len,
@ -36,7 +36,7 @@ private:
std::vector<ggml_context_ptr> ctxs;
std::vector<ggml_backend_buffer_ptr> bufs;
std::vector<struct ggml_tensor *> tensors; // per layer
std::vector<ggml_tensor *> tensors; // per layer
};
//
@ -44,8 +44,8 @@ private:
//
struct llama_adapter_lora_weight {
struct ggml_tensor * a = nullptr;
struct ggml_tensor * b = nullptr;
ggml_tensor * a = nullptr;
ggml_tensor * b = nullptr;
// get actual scale based on rank and alpha
float get_scale(float alpha, float adapter_scale) const {
@ -55,12 +55,12 @@ struct llama_adapter_lora_weight {
}
llama_adapter_lora_weight() = default;
llama_adapter_lora_weight(struct ggml_tensor * a, struct ggml_tensor * b) : a(a), b(b) {}
llama_adapter_lora_weight(ggml_tensor * a, ggml_tensor * b) : a(a), b(b) {}
};
struct llama_adapter_lora {
// map tensor name to lora_a_b
std::unordered_map<std::string, struct llama_adapter_lora_weight> ab_map;
std::unordered_map<std::string, llama_adapter_lora_weight> ab_map;
std::vector<ggml_context_ptr> ctxs;
std::vector<ggml_backend_buffer_ptr> bufs;
@ -70,5 +70,7 @@ struct llama_adapter_lora {
llama_adapter_lora() = default;
~llama_adapter_lora() = default;
llama_adapter_lora_weight * get_weight(struct ggml_tensor * w);
llama_adapter_lora_weight * get_weight(ggml_tensor * w);
};
using llama_adapter_loras = std::unordered_map<llama_adapter_lora *, float>;