tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-06-27 20:05:20 +00:00
Code Issues Packages Projects Releases Wiki Activity

kv-cache : separate recurrent vs non-recurrent impl (#12799)

Browse Source 
* kv-cache : serparate recurrent vs non-recurrent impl (wip)

ggml-ci

* kv-cache : init -> contructor + add llama_memory_params

ggml-ci

* kv-cache : fix callback reference

ggml-ci

* context : llama_kv_cache -> llama_memory_i

ggml-ci

* context : move memory creation logic to model

ggml-ci

* llama : remove reference of memory during encode

ggml-ci

* kv-cache : hide padding details in the implementation

ggml-ci

* kv-cache : add ubatch_next()

ggml-ci

* context : simplify sbatch logic

ggml-ci

* kv-cache : hide defrag logic in the implementation

ggml-ci

* context : hide kv cache details in implementation

ggml-ci

* build : fix

ggml-ci

* cont : another fix

ggml-ci

* kv-cache : simplify interface (wip)

ggml-ci

* kv-cache : use separate KV cell structs for unified/recurrent

ggml-ci

* kv-cache : clean-up

ggml-ci

* model : better llama_model::create_model() signature

ggml-ci

* kv-cache : fix recurrent seq_rm()

ggml-ci

* kv-cache : replace `struct callbacks` with `llama_model &`

ggml-ci

* kv-cache : replace `struct graph_params` with `llama_context &`

ggml-ci

* kv-cache : fix offload check

ggml-ci

* context : avoid passing unique_ptr

ggml-ci

* kv-cache : avoid using the backends from the llama_context

ref #13113

ggml-ci

* kv-cache : more consistent debug logs [no ci]

* kv-cache : do not pass the full llama_context for kv graphs

ggml-ci

* kv-cache : remove comment

* kv-cache : ggml_rope_ext_inplace -> ggml_rope_ext

ggml-ci

* kv-cache : fix recurrent multi-user case

ggml-ci

* memory : remove comments [no ci]
This commit is contained in:
Georgi Gerganov
2025-05-02 17:48:36 +03:00
committed by GitHub
parent cb06a3c363
commit c642bc014c
11 changed files with 1960 additions and 1048 deletions
Download Patch File Download Diff File Expand all files Collapse all files

70
src/llama-model.cpp
View File

@ -4445,6 +4445,19 @@ const ggml_tensor * llama_model::get_tensor(const char * name) const {
return it->second;
}
ggml_tensor * llama_model::get_rope_factors(uint32_t n_ctx_per_seq, int il) const {
// choose long/short freq factors based on the context size
if (layers[il].rope_freqs != nullptr) {
return layers[il].rope_freqs;
}
if (n_ctx_per_seq > hparams.n_ctx_orig_yarn) {
return layers[il].rope_long;
}
return layers[il].rope_short;
}
struct llm_build_llama : public llm_graph_context {
llm_build_llama(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
const int64_t n_embd_head = hparams.n_embd_head_v;
@ -4485,7 +4498,7 @@ struct llm_build_llama : public llm_graph_context {
// self-attention
{
// rope freq factors for llama3; may return nullptr for llama2 and other models
ggml_tensor * rope_factors = static_cast<const llama_kv_cache_unified *>(memory)->cbs.get_rope_factors(n_ctx_per_seq, il);
ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
// compute Q and K and RoPE them
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
@ -4710,7 +4723,7 @@ struct llm_build_deci : public llm_graph_context {
} else if (n_head > 0) {
// self-attention
// rope freq factors for llama3; may return nullptr for llama2 and other models
ggml_tensor * rope_factors = static_cast<const llama_kv_cache_unified *>(memory)->cbs.get_rope_factors(n_ctx_per_seq, il);
ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
// compute Q and K and RoPE them
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
@ -7192,7 +7205,7 @@ struct llm_build_phi3 : public llm_graph_context {
// self-attention
{
// rope freq factors for 128k context
ggml_tensor * rope_factors = static_cast<const llama_kv_cache_unified *>(memory)->cbs.get_rope_factors(n_ctx_per_seq, il);
ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
ggml_tensor* attn_norm_output = build_norm(inpL,
model.layers[il].attn_norm,
@ -7944,7 +7957,7 @@ struct llm_build_minicpm3 : public llm_graph_context {
for (int il = 0; il < n_layer; ++il) {
ggml_tensor * inpSA = inpL;
ggml_tensor * rope_factors = static_cast<const llama_kv_cache_unified *>(memory)->cbs.get_rope_factors(n_ctx_per_seq, il);
ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
// norm
cur = build_norm(inpL,
@ -8711,7 +8724,7 @@ struct llm_build_mamba : public llm_graph_context {
ggml_tensor * state_mask,
const llama_ubatch & ubatch,
int il) const {
const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
const llama_kv_cache_recurrent * kv_self = static_cast<const llama_kv_cache_recurrent *>(memory);
const auto kv_head = kv_self->head;
@ -9012,7 +9025,7 @@ struct llm_build_cohere2 : public llm_graph_context {
// self-attention
{
// rope freq factors for 128k context
ggml_tensor * rope_factors = static_cast<const llama_kv_cache_unified *>(memory)->cbs.get_rope_factors(n_ctx_per_seq, il);
ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
// compute Q and K and RoPE them
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
@ -9950,7 +9963,7 @@ struct llm_build_deepseek : public llm_graph_context {
// self-attention
{
// rope freq factors for llama3; may return nullptr for llama2 and other models
ggml_tensor * rope_factors = static_cast<const llama_kv_cache_unified *>(memory)->cbs.get_rope_factors(n_ctx_per_seq, il);
ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
// compute Q and K and RoPE them
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
@ -11314,7 +11327,7 @@ struct llm_build_exaone : public llm_graph_context {
// self-attention
{
// rope freq factors for llama3; may return nullptr for llama2 and other models
ggml_tensor * rope_factors = static_cast<const llama_kv_cache_unified *>(memory)->cbs.get_rope_factors(n_ctx_per_seq, il);
ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
// compute Q and K and RoPE them
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
@ -11459,7 +11472,7 @@ struct llm_build_rwkv6_base : public llm_graph_context {
ggml_tensor * state_mask,
const llama_ubatch & ubatch,
int il) const {
const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
const llama_kv_cache_recurrent * kv_self = static_cast<const llama_kv_cache_recurrent *>(memory);
const auto n_tokens = ubatch.n_tokens;
const auto n_seqs = ubatch.n_seqs;
@ -11855,7 +11868,7 @@ struct llm_build_rwkv7_base : public llm_graph_context {
ggml_tensor *& first_layer_value,
const llama_ubatch & ubatch,
int il) const {
const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
const llama_kv_cache_recurrent * kv_self = static_cast<const llama_kv_cache_recurrent *>(memory);
const auto n_tokens = ubatch.n_tokens;
const auto n_seqs = ubatch.n_seqs;
@ -12695,7 +12708,7 @@ struct llm_build_bailingmoe : public llm_graph_context {
// self-attention
{
// rope freq factors for llama3; may return nullptr for llama2 and other models
ggml_tensor * rope_factors = static_cast<const llama_kv_cache_unified *>(memory)->cbs.get_rope_factors(n_ctx_per_seq, il);
ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
// compute Q and K and RoPE them
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
@ -12815,7 +12828,7 @@ struct llm_build_bailingmoe : public llm_graph_context {
}
};
llama_memory_i * llama_model::create_memory() const {
llama_memory_i * llama_model::create_memory(const llama_memory_params & params, llama_cparams & cparams) const {
llama_memory_i * res;
switch (arch) {
@ -12825,26 +12838,29 @@ llama_memory_i * llama_model::create_memory() const {
case LLM_ARCH_RWKV7:
case LLM_ARCH_ARWKV7:
{
res = new llama_kv_cache_unified(hparams, {
/*.get_rope_factors =*/ nullptr
});
res = new llama_kv_cache_recurrent(
*this,
GGML_TYPE_F32,
GGML_TYPE_F32,
cparams.offload_kqv,
std::max((uint32_t) 1, cparams.n_seq_max));
} break;
default:
{
res = new llama_kv_cache_unified(hparams, {
/*.get_rope_factors =*/ [this](uint32_t n_ctx_per_seq, int il) {
// choose long/short freq factors based on the context size
if (layers[il].rope_freqs != nullptr) {
return layers[il].rope_freqs;
}
const auto padding = llama_kv_cache_unified::get_padding(cparams);
if (n_ctx_per_seq > hparams.n_ctx_orig_yarn) {
return layers[il].rope_long;
}
cparams.n_ctx = GGML_PAD(cparams.n_ctx, padding);
return layers[il].rope_short;
}
});
LLAMA_LOG_DEBUG("%s: n_ctx = %u (padded)\n", __func__, cparams.n_ctx);
res = new llama_kv_cache_unified(
*this,
params.type_k,
params.type_v,
!cparams.flash_attn,
cparams.offload_kqv,
cparams.n_ctx,
padding);
}
}