181 lines
8.0 KiB
C++
181 lines
8.0 KiB
C++
#include "llama-kv-cache.h"
|
|
#include "llama-model.h"
|
|
|
|
#include <numeric>
|
|
#include <string>
|
|
#include <vector>
|
|
|
|
static void init_model(llama_model & model, uint32_t n_layer) {
|
|
model.hparams = {};
|
|
model.hparams.n_layer_all = n_layer;
|
|
model.hparams.n_embd_head_k_full = 4;
|
|
model.hparams.n_embd_head_v_full = 4;
|
|
model.hparams.n_head_arr.fill(1);
|
|
model.hparams.n_head_kv_arr.fill(1);
|
|
std::iota(model.hparams.kv_source_impl.begin(), model.hparams.kv_source_impl.end(), 0);
|
|
model.layers.resize(n_layer);
|
|
}
|
|
|
|
int main() {
|
|
for (int bit = 0; bit < 64; ++bit) {
|
|
GGML_ASSERT(ctz64(uint64_t{1} << bit) == bit);
|
|
GGML_ASSERT(ctz64(UINT64_MAX << bit) == bit);
|
|
}
|
|
|
|
llama_backend_init();
|
|
|
|
{
|
|
std::string errors;
|
|
llama_log_set([](ggml_log_level level, const char * text, void * user_data) {
|
|
if (level == GGML_LOG_LEVEL_ERROR) {
|
|
*static_cast<std::string *>(user_data) += text;
|
|
}
|
|
}, &errors);
|
|
|
|
for (llm_arch arch : { LLM_ARCH_DEEPSEEK2, LLM_ARCH_DEEPSEEK32, LLM_ARCH_DEEPSEEK4,
|
|
LLM_ARCH_DEEPSEEK41, LLM_ARCH_GLM_DSA, LLM_ARCH_GLM5_NEXT, LLM_ARCH_DOTS3NOTE,
|
|
LLM_ARCH_MISTRAL_LARGE4, LLM_ARCH_HY_V4, LLM_ARCH_MINIMAX_M3, LLM_ARCH_QWEN4EXP,
|
|
LLM_ARCH_DFLASH }) {
|
|
std::unique_ptr<llama_model> model(llama_model_create(arch, llama_model_default_params()));
|
|
init_model(*model, 1);
|
|
if (arch == LLM_ARCH_DEEPSEEK2) {
|
|
model->hparams.n_embd_head_k_mla_impl = 4;
|
|
model->hparams.n_embd_head_v_mla_impl = 4;
|
|
}
|
|
|
|
for (ggml_type type_k : { GGML_TYPE_F16, GGML_TYPE_F8_E4M3 }) {
|
|
for (ggml_type type_v : { GGML_TYPE_F16, GGML_TYPE_F8_E4M3 }) {
|
|
if (type_k == GGML_TYPE_F16 && type_v == GGML_TYPE_F16) {
|
|
continue;
|
|
}
|
|
auto params = llama_context_default_params();
|
|
params.type_k = type_k;
|
|
params.type_v = type_v;
|
|
errors.clear();
|
|
llama_context * ctx = llama_init_from_model(model.get(), params);
|
|
GGML_ASSERT(ctx == nullptr);
|
|
if (type_k == type_v || (!model->hparams.is_mla() && arch != LLM_ARCH_DEEPSEEK4 && arch != LLM_ARCH_DEEPSEEK41)) {
|
|
GGML_ASSERT(errors.find("FP8 cache is not supported for architecture") != std::string::npos);
|
|
} else {
|
|
GGML_ASSERT(errors.find("model does not support different K") != std::string::npos);
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
llama_log_set(nullptr, nullptr);
|
|
}
|
|
|
|
{
|
|
std::unique_ptr<llama_model> target(llama_model_create(LLM_ARCH_GEMMA4, llama_model_default_params()));
|
|
std::unique_ptr<llama_model> assistant(llama_model_create(LLM_ARCH_GEMMA4_ASSISTANT, llama_model_default_params()));
|
|
init_model(*target, 3);
|
|
init_model(*assistant, 2);
|
|
|
|
ggml_context_ptr ctx(ggml_init({ 4096, nullptr, false }));
|
|
target->layers[2].k_cache_s = ggml_new_tensor_1d(ctx.get(), GGML_TYPE_F32, 1);
|
|
target->layers[2].v_cache_s = ggml_new_tensor_1d(ctx.get(), GGML_TYPE_F32, 1);
|
|
assistant->layers[0].k_cache_s = ggml_new_tensor_1d(ctx.get(), GGML_TYPE_F32, 1);
|
|
assistant->layers[0].v_cache_s = ggml_new_tensor_1d(ctx.get(), GGML_TYPE_F32, 1);
|
|
assistant->layers[1].k_cache_s = assistant->layers[0].k_cache_s;
|
|
assistant->layers[1].v_cache_s = assistant->layers[0].v_cache_s;
|
|
*static_cast<float *>(target->layers[2].k_cache_s->data) = 0.5f;
|
|
*static_cast<float *>(target->layers[2].v_cache_s->data) = 0.25f;
|
|
*static_cast<float *>(assistant->layers[0].k_cache_s->data) = 2.0f;
|
|
*static_cast<float *>(assistant->layers[0].v_cache_s->data) = 4.0f;
|
|
|
|
for (ggml_type type_k : { GGML_TYPE_F16, GGML_TYPE_F8_E4M3 }) {
|
|
for (ggml_type type_v : { GGML_TYPE_F16, GGML_TYPE_F8_E4M3 }) {
|
|
llama_kv_cache source(*target, false, target->hparams, type_k, type_v,
|
|
false, false, true, 32, 1, 0, LLAMA_SWA_TYPE_NONE, nullptr, nullptr,
|
|
[](uint32_t il) { return il == 0 ? 2 : -1; }, nullptr);
|
|
llama_kv_cache shared(*assistant, false, assistant->hparams, type_k, type_v,
|
|
false, false, true, 32, 1, 0, LLAMA_SWA_TYPE_NONE, &source, nullptr, nullptr,
|
|
[](int32_t il) { return il == 0 ? 2 : 1; });
|
|
|
|
GGML_ASSERT(source.get_k_scale(2) == (type_k == GGML_TYPE_F8_E4M3 ? target->layers[2].k_cache_s : nullptr));
|
|
GGML_ASSERT(source.get_v_scale(2) == (type_v == GGML_TYPE_F8_E4M3 ? target->layers[2].v_cache_s : nullptr));
|
|
GGML_ASSERT(source.get_k_scale(0) == source.get_k_scale(2));
|
|
GGML_ASSERT(source.get_v_scale(0) == source.get_v_scale(2));
|
|
GGML_ASSERT(shared.get_k_storage(0) == source.get_k_storage(2));
|
|
GGML_ASSERT(shared.get_k_scale(0) == source.get_k_scale(2));
|
|
GGML_ASSERT(shared.get_v_scale(0) == source.get_v_scale(2));
|
|
GGML_ASSERT(shared.get_k_scale(1) == nullptr);
|
|
GGML_ASSERT(shared.get_v_scale(1) == nullptr);
|
|
}
|
|
}
|
|
}
|
|
|
|
// seq_rm removes [p0, p1) of one sequence, a negative bound is open
|
|
{
|
|
std::unique_ptr<llama_model> model(llama_model_create(LLM_ARCH_LLAMA, llama_model_default_params()));
|
|
init_model(*model, 1);
|
|
|
|
llama_kv_cache kv(*model, false, model->hparams, GGML_TYPE_F16, GGML_TYPE_F16,
|
|
false, false, true, 32, 2, 0, LLAMA_SWA_TYPE_NONE, nullptr, nullptr, nullptr, nullptr);
|
|
|
|
// returns the first cell that takes the tokens
|
|
auto add = [&](llama_seq_id seq_id, llama_pos p0, uint32_t n) {
|
|
std::vector<llama_token> token(n, 1);
|
|
std::vector<llama_pos> pos(n);
|
|
std::vector<int32_t> n_seq_id(n, 1);
|
|
std::vector<llama_seq_id *> seq_ids(n, &seq_id);
|
|
std::iota(pos.begin(), pos.end(), p0);
|
|
|
|
llama_ubatch ubatch = {};
|
|
ubatch.n_tokens = n;
|
|
ubatch.n_seq_tokens = n;
|
|
ubatch.n_seqs = 1;
|
|
ubatch.n_seqs_unq = 1;
|
|
ubatch.n_pos = 1;
|
|
ubatch.token = token.data();
|
|
ubatch.pos = pos.data();
|
|
ubatch.n_seq_id = n_seq_id.data();
|
|
ubatch.seq_id = seq_ids.data();
|
|
ubatch.seq_id_unq = &seq_id;
|
|
|
|
const auto sinfo = kv.find_slot(ubatch, false);
|
|
GGML_ASSERT(!sinfo.empty());
|
|
kv.apply_ubatch(sinfo, ubatch);
|
|
|
|
return sinfo.head();
|
|
};
|
|
|
|
GGML_ASSERT(add(0, 0, 8) == 0);
|
|
GGML_ASSERT(add(1, 0, 4) == 8);
|
|
|
|
// the tail of sequence 0, the other sequence keeps its cells
|
|
GGML_ASSERT(kv.seq_rm(0, 5, -1));
|
|
GGML_ASSERT(kv.seq_pos_min(0) == 0 && kv.seq_pos_max(0) == 4);
|
|
GGML_ASSERT(kv.seq_pos_min(1) == 0 && kv.seq_pos_max(1) == 3);
|
|
|
|
// the next tokens go to the lowest freed cell
|
|
GGML_ASSERT(add(0, 5, 2) == 5);
|
|
|
|
// p1 is not part of the range
|
|
GGML_ASSERT(kv.seq_rm(0, -1, 2));
|
|
GGML_ASSERT(kv.seq_pos_min(0) == 2 && kv.seq_pos_max(0) == 6);
|
|
GGML_ASSERT(kv.seq_rm(0, 6, 6));
|
|
GGML_ASSERT(kv.seq_pos_max(0) == 6);
|
|
|
|
// freed cells are taken in cell order, so a higher position can sit in a lower cell
|
|
GGML_ASSERT(add(0, 7, 3) == 0);
|
|
GGML_ASSERT(kv.seq_rm(0, 6, -1));
|
|
GGML_ASSERT(add(0, 6, 1) == 0);
|
|
|
|
// a cell shared with another sequence stays for that sequence
|
|
kv.seq_cp(0, 1, 5, -1);
|
|
GGML_ASSERT(kv.seq_rm(0, 5, -1));
|
|
GGML_ASSERT(kv.seq_pos_max(0) == 4 && kv.seq_pos_max(1) == 6);
|
|
|
|
// every cell of a repeated position is removed
|
|
add(0, 5, 1);
|
|
add(0, 5, 1);
|
|
GGML_ASSERT(kv.seq_rm(0, 5, 6));
|
|
GGML_ASSERT(kv.seq_pos_max(0) == 4);
|
|
}
|
|
|
|
llama_backend_free();
|
|
return 0;
|
|
}
|