happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

llama : add "virtual sequences" 

ggml-ci
This commit is contained in:
Georgi Gerganov 2025-06-23 16:29:02 +03:00
parent 30b4d4e1b3
commit dfceb012ee
No known key found for this signature in database
GPG Key ID: 449E073F9DC10735
16 changed files with 648 additions and 256 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
examples/parallel/parallel.cpp
View File

@ -235,7 +235,7 @@ int main(int argc, char ** argv) {
// the max batch size is as large as the context to handle cases where we get very long input prompt from multiple
// users. regardless of the size, the main loop will chunk the batch into a maximum of params.n_batch tokens at a time
llama_batch batch = llama_batch_init(n_ctx, 0, 1);
llama_batch batch = llama_batch_init(n_ctx*n_clients, 0, 1);
int32_t n_total_prompt = 0;
int32_t n_total_gen = 0;
@ -289,6 +289,7 @@ int main(int argc, char ** argv) {
// all sequences have ended - clear the entire KV cache
for (int i = 1; i <= n_clients; ++i) {
llama_memory_seq_rm(mem, i, -1, -1);
// but keep the system prompt
llama_memory_seq_cp(mem, 0, i, -1, -1);
}

1
ggml/src/ggml.c
View File

@ -4696,7 +4696,6 @@ struct ggml_tensor * ggml_flash_attn_ext(
if (mask) {
GGML_ASSERT(ggml_is_contiguous(mask));
GGML_ASSERT(mask->ne[2] == q->ne[3]);
GGML_ASSERT(mask->ne[1] >= GGML_PAD(q->ne[1], GGML_KQ_MASK_PAD) &&
"the Flash-Attention kernel requires the mask to be padded to GGML_KQ_MASK_PAD and at least n_queries big");
//GGML_ASSERT(ggml_can_repeat_rows(mask, qk));

28
src/llama-batch.cpp
View File

@ -166,6 +166,8 @@ bool llama_batch_allocr::init(
// note: tracking the other way around is not necessary for now
//seq_cpl[s0][s1] = true;
has_cpl = true;
}
}
}
@ -405,6 +407,10 @@ uint32_t llama_batch_allocr::get_n_outputs() const {
return n_outputs;
}
uint32_t llama_batch_allocr::get_n_used() const {
return n_used;
}
std::vector<int32_t> & llama_batch_allocr::get_out_ids() {
return out_ids;
}
@ -420,6 +426,8 @@ llama_pos llama_batch_allocr::seq_pos_max(llama_seq_id seq_id) const {
void llama_batch_allocr::split_reset() {
out_ids.clear();
n_used = 0;
used.clear();
used.resize(get_n_tokens(), false);
@ -444,6 +452,7 @@ llama_ubatch llama_batch_allocr::split_simple(uint32_t n_ubatch) {
idxs.push_back(cur_idx);
used[cur_idx] = true;
++n_used;
++cur_idx;
@ -459,9 +468,17 @@ llama_ubatch llama_batch_allocr::split_simple(uint32_t n_ubatch) {
return ubatch_add(idxs, idxs.size(), false);
}
llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch) {
llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch, bool sequential) {
if (sequential && has_cpl) {
LLAMA_LOG_ERROR("%s: sequential split is not supported when there are coupled sequences in the input batch\n", __func__);
return {};
}
std::vector<seq_set_t> cur_seq_set;
llama_seq_id last_seq_id = -1;
// determine the non-overlapping sequence sets participating in this ubatch
for (int32_t i = 0; i < batch.n_tokens; ++i) {
if (used[i]) {
@ -478,9 +495,16 @@ llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch) {
}
}
// accept only increasing sequence ids
if (sequential) {
add = add && (cur_seq_set.empty() || batch.seq_id[i][0] == last_seq_id + 1);
}
if (add) {
cur_seq_set.push_back(seq_set[i]);
last_seq_id = batch.seq_id[i][0];
if (cur_seq_set.size() > n_ubatch) {
break;
}
@ -529,6 +553,7 @@ llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch) {
idxs_per_seq[s].push_back(idx);
used[idx] = true;
++n_used;
++cur_idx[s];
}
@ -570,6 +595,7 @@ llama_ubatch llama_batch_allocr::split_seq(uint32_t n_ubatch) {
idxs.push_back(cur_idx);
used[cur_idx] = true;
++n_used;
if (idxs.size() >= n_ubatch) {
break;

9
src/llama-batch.h
View File

@ -54,6 +54,7 @@ public:
uint32_t get_n_tokens() const;
uint32_t get_n_outputs() const;
uint32_t get_n_used() const;
// the array of output indices in the order they were encountered during the ubatch splitting
std::vector<int32_t> & get_out_ids();
@ -69,7 +70,8 @@ public:
llama_ubatch split_simple(uint32_t n_ubatch);
// make ubatches of equal-length sequences sets
llama_ubatch split_equal(uint32_t n_ubatch);
// if sequential == true, the tokens in the ubatch will have increasing sequential sequence ids
llama_ubatch split_equal(uint32_t n_ubatch, bool sequential);
// sequence-set-wise split - each ubatch contains a single sequence-set
llama_ubatch split_seq(uint32_t n_ubatch);
@ -112,6 +114,9 @@ private:
using pos_set_t = std::set<llama_pos>;
using seq_cpl_t = std::vector<bool>;
// helper flag to quickly determine if there are any coupled sequences in the batch
bool has_cpl;
std::vector<pos_set_t> seq_pos; // seq_pos[s]: the set of positions in sequence s
std::vector<seq_cpl_t> seq_cpl; // seq_cpl[s0][s1]: if sequence s0 is coupled to sequence s1
@ -125,6 +130,8 @@ private:
// batch indices of the output
std::vector<int32_t> out_ids;
uint32_t n_used;
// used[i] indicates if token i has already been used in a previous ubatch
std::vector<bool> used;

6
src/llama-context.cpp
View File

@ -33,6 +33,9 @@ llama_context::llama_context(
throw std::runtime_error("n_seq_max must be <= " + std::to_string(LLAMA_MAX_SEQ));
}
const char * LLAMA_HT = getenv("LLAMA_HT");
cparams.n_seq_virt = LLAMA_HT ? cparams.n_seq_max : 1;
cparams.n_threads = params.n_threads;
cparams.n_threads_batch = params.n_threads_batch;
cparams.yarn_ext_factor = params.yarn_ext_factor;
@ -1308,7 +1311,8 @@ ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, u
this->n_outputs = n_outputs;
llama_batch_allocr balloc(model.hparams.n_pos_per_embd());
llama_ubatch ubatch = balloc.ubatch_reserve(n_tokens/n_seqs, n_seqs);
//llama_ubatch ubatch = balloc.ubatch_reserve(n_tokens/n_seqs, n_seqs);
llama_ubatch ubatch = balloc.ubatch_reserve(n_tokens, 1);
auto * gf = graph_init();
auto res = graph_build(ctx_compute.get(), gf, ubatch, LLM_GRAPH_TYPE_DEFAULT, mctx);

5
src/llama-cparams.h
View File

@ -11,8 +11,9 @@ struct llama_cparams {
uint32_t n_batch;
uint32_t n_ubatch;
uint32_t n_seq_max;
int n_threads; // number of threads to use for generation
int n_threads_batch; // number of threads to use for batch processing
uint32_t n_seq_virt;
int32_t n_threads; // number of threads to use for generation
int32_t n_threads_batch; // number of threads to use for batch processing
float rope_freq_base;
float rope_freq_scale;

25
src/llama-graph.cpp
View File

@ -1000,13 +1000,13 @@ llm_graph_input_mem_hybrid * llm_graph_context::build_inp_mem_hybrid() const {
{
GGML_ASSERT(hparams.swa_type == LLAMA_SWA_TYPE_NONE && "Hybrid recurrent is not supported with SWA attention layers");
const auto n_kv = inp->mctx->get_attn()->get_n_kv();
const auto n_kv = inp->mctx->get_attn()->get_n_kv();
const auto n_seqs = cparams.n_seq_virt > 1 ? ubatch.n_seqs_unq : 1;
inp->self_k_idxs = mctx_cur->get_attn()->build_input_k_idxs(ctx0, ubatch);
inp->self_v_idxs = mctx_cur->get_attn()->build_input_v_idxs(ctx0, ubatch);
inp->self_kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
//cb(inp->self_kq_mask, "KQ_mask", -1);
inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens/n_seqs, GGML_KQ_MASK_PAD), 1, n_seqs);
ggml_set_input(inp->self_kq_mask);
inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
@ -1033,6 +1033,10 @@ ggml_tensor * llm_graph_context::build_attn_mha(
float kq_scale) const {
const bool v_trans = v->nb[1] > v->nb[2];
const auto n_seqs = cparams.n_seq_virt > 1 ? ubatch.n_seqs_unq : 1;
q = ggml_reshape_4d(ctx0, q, q->ne[0], q->ne[1], q->ne[2]/n_seqs, n_seqs);
q = ggml_permute(ctx0, q, 0, 2, 1, 3);
k = ggml_permute(ctx0, k, 0, 2, 1, 3);
v = ggml_permute(ctx0, v, 0, 2, 1, 3);
@ -1081,7 +1085,7 @@ ggml_tensor * llm_graph_context::build_attn_mha(
#endif
}
cur = ggml_reshape_2d(ctx0, cur, cur->ne[0]*n_head, n_tokens);
cur = ggml_reshape_2d(ctx0, cur, cur->ne[0]*n_head, n_tokens*n_seqs);
} else {
ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
@ -1126,7 +1130,7 @@ ggml_tensor * llm_graph_context::build_attn_mha(
cur = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
cur = ggml_cont_2d(ctx0, cur, cur->ne[0]*n_head, n_tokens);
cur = ggml_cont_2d(ctx0, cur, cur->ne[0]*n_head, n_tokens*n_seqs);
if (!cparams.offload_kqv) {
// all nodes between the KV store and the attention output are run on the CPU
@ -1204,12 +1208,13 @@ llm_graph_input_attn_kv_unified * llm_graph_context::build_attn_inp_kv_unified()
{
GGML_ASSERT(hparams.swa_type == LLAMA_SWA_TYPE_NONE && "Use llama_kv_cache_unified_iswa for SWA");
const auto n_kv = mctx_cur->get_n_kv();
const auto n_kv = mctx_cur->get_n_kv();
const auto n_seqs = cparams.n_seq_virt > 1 ? ubatch.n_seqs_unq : 1;
inp->self_k_idxs = mctx_cur->build_input_k_idxs(ctx0, ubatch);
inp->self_v_idxs = mctx_cur->build_input_v_idxs(ctx0, ubatch);
inp->self_kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens/n_seqs, GGML_KQ_MASK_PAD), 1, n_seqs);
ggml_set_input(inp->self_kq_mask);
inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
@ -1451,13 +1456,15 @@ llm_graph_input_attn_kv_unified_iswa * llm_graph_context::build_attn_inp_kv_unif
auto inp = std::make_unique<llm_graph_input_attn_kv_unified_iswa>(hparams, cparams, mctx_cur);
const auto n_seqs = cparams.n_seq_virt > 1 ? ubatch.n_seqs_unq : 1;
{
const auto n_kv = mctx_cur->get_base()->get_n_kv();
inp->self_k_idxs = mctx_cur->get_base()->build_input_k_idxs(ctx0, ubatch);
inp->self_v_idxs = mctx_cur->get_base()->build_input_v_idxs(ctx0, ubatch);
inp->self_kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens/n_seqs, GGML_KQ_MASK_PAD), 1, n_seqs);
ggml_set_input(inp->self_kq_mask);
inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
@ -1471,7 +1478,7 @@ llm_graph_input_attn_kv_unified_iswa * llm_graph_context::build_attn_inp_kv_unif
inp->self_k_idxs_swa = mctx_cur->get_swa()->build_input_k_idxs(ctx0, ubatch);
inp->self_v_idxs_swa = mctx_cur->get_swa()->build_input_v_idxs(ctx0, ubatch);
inp->self_kq_mask_swa = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
inp->self_kq_mask_swa = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens/n_seqs, GGML_KQ_MASK_PAD), 1, n_seqs);
ggml_set_input(inp->self_kq_mask_swa);
inp->self_kq_mask_swa_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask_swa, GGML_TYPE_F16) : inp->self_kq_mask_swa;

18
src/llama-graph.h
View File

@ -255,10 +255,10 @@ public:
ggml_tensor * get_kq_mask() const { return self_kq_mask_cnv; }
ggml_tensor * self_k_idxs = nullptr; // I64 [n_batch]
ggml_tensor * self_v_idxs = nullptr; // I64 [n_batch]
ggml_tensor * self_v_idxs = nullptr; // I64 [n_batch] or [n_batch*n_embd_v_gqa]
ggml_tensor * self_kq_mask = nullptr; // F32 [n_kv, n_batch]
ggml_tensor * self_kq_mask_cnv = nullptr; // [n_kv, n_batch]
ggml_tensor * self_kq_mask = nullptr; // F32 [n_kv, n_batch/n_seqs, n_seqs]
ggml_tensor * self_kq_mask_cnv = nullptr; // [n_kv, n_batch/n_seqs, n_seqs]
const llama_hparams & hparams;
const llama_cparams & cparams;
@ -289,14 +289,14 @@ public:
ggml_tensor * get_kq_mask_swa() const { return self_kq_mask_swa_cnv; }
ggml_tensor * self_k_idxs = nullptr; // I64 [n_batch]
ggml_tensor * self_v_idxs = nullptr; // I64 [n_batch]
ggml_tensor * self_v_idxs = nullptr; // I64 [n_batch] or [n_batch*n_embd_v_gqa]
ggml_tensor * self_k_idxs_swa = nullptr; // I64 [n_batch]
ggml_tensor * self_v_idxs_swa = nullptr; // I64 [n_batch]
ggml_tensor * self_v_idxs_swa = nullptr; // I64 [n_batch] or [n_batch*n_embd_v_gqa]
ggml_tensor * self_kq_mask = nullptr; // F32 [n_kv, n_batch]
ggml_tensor * self_kq_mask_cnv = nullptr; // [n_kv, n_batch]
ggml_tensor * self_kq_mask_swa = nullptr; // F32 [n_kv, n_batch]
ggml_tensor * self_kq_mask_swa_cnv = nullptr; // [n_kv, n_batch]
ggml_tensor * self_kq_mask = nullptr; // F32 [n_kv, n_batch/n_seqs, n_seqs]
ggml_tensor * self_kq_mask_cnv = nullptr; // [n_kv, n_batch/n_seqs, n_seqs]
ggml_tensor * self_kq_mask_swa = nullptr; // F32 [n_kv, n_batch/n_seqs, n_seqs]
ggml_tensor * self_kq_mask_swa_cnv = nullptr; // [n_kv, n_batch/n_seqs, n_seqs]
const llama_hparams & hparams;
const llama_cparams & cparams;

26
src/llama-kv-cache-unified-iswa.cpp
View File

@ -20,14 +20,15 @@ llama_kv_cache_unified_iswa::llama_kv_cache_unified_iswa(
bool swa_full,
uint32_t kv_size,
uint32_t n_seq_max,
uint32_t n_seq_virt,
uint32_t n_ubatch,
uint32_t n_pad) : hparams(model.hparams) {
uint32_t n_pad) : hparams(model.hparams), n_seq_virt(n_seq_virt) {
llama_kv_cache_unified::layer_filter_cb filter_base = [&](int32_t il) { return !model.hparams.is_swa(il); };
llama_kv_cache_unified::layer_filter_cb filter_swa = [&](int32_t il) { return model.hparams.is_swa(il); };
const uint32_t size_base = kv_size;
uint32_t size_swa = std::min(size_base, GGML_PAD(hparams.n_swa*n_seq_max + n_ubatch, n_pad));
uint32_t size_swa = std::min(size_base, GGML_PAD(hparams.n_swa*(n_seq_max/n_seq_virt) + n_ubatch, n_pad));
// when using full-size SWA cache, we set the SWA cache size to be equal to the base cache size
if (swa_full) {
@ -41,14 +42,14 @@ llama_kv_cache_unified_iswa::llama_kv_cache_unified_iswa(
kv_base = std::make_unique<llama_kv_cache_unified>(
model, std::move(filter_base), type_k, type_v,
v_trans, offload, size_base, n_seq_max, n_pad,
v_trans, offload, size_base, n_seq_max, n_seq_virt, n_pad,
0, LLAMA_SWA_TYPE_NONE);
LLAMA_LOG_INFO("%s: creating SWA KV cache, size = %u cells\n", __func__, size_swa);
kv_swa = std::make_unique<llama_kv_cache_unified>(
model, std::move(filter_swa), type_k, type_v,
v_trans, offload, size_swa, n_seq_max, n_pad,
v_trans, offload, size_swa, n_seq_max, n_seq_virt, n_pad,
hparams.n_swa, hparams.swa_type);
}
@ -100,6 +101,11 @@ llama_memory_context_ptr llama_kv_cache_unified_iswa::init_batch(llama_batch_all
// first try simple split
do {
if (n_seq_virt > 1) {
// requires equal splits, so we skip the simple split
break;
}
balloc.split_reset();
std::vector<llama_ubatch> ubatches;
@ -113,6 +119,11 @@ llama_memory_context_ptr llama_kv_cache_unified_iswa::init_batch(llama_batch_all
ubatches.push_back(std::move(ubatch)); // NOLINT
}
if (balloc.get_n_used() < balloc.get_n_tokens()) {
// failed to find a suitable split
break;
}
auto sinfos_base = kv_base->prepare(ubatches);
if (sinfos_base.empty()) {
break;
@ -135,7 +146,7 @@ llama_memory_context_ptr llama_kv_cache_unified_iswa::init_batch(llama_batch_all
std::vector<llama_ubatch> ubatches;
while (true) {
auto ubatch = balloc.split_equal(n_ubatch);
auto ubatch = balloc.split_equal(n_ubatch, n_seq_virt > 1);
if (ubatch.n_tokens == 0) {
break;
@ -144,6 +155,11 @@ llama_memory_context_ptr llama_kv_cache_unified_iswa::init_batch(llama_batch_all
ubatches.push_back(std::move(ubatch)); // NOLINT
}
if (balloc.get_n_used() < balloc.get_n_tokens()) {
// failed to find a suitable split
break;
}
auto sinfos_base = kv_base->prepare(ubatches);
if (sinfos_base.empty()) {
break;

3
src/llama-kv-cache-unified-iswa.h
View File

@ -22,6 +22,7 @@ public:
bool swa_full,
uint32_t kv_size,
uint32_t n_seq_max,
uint32_t n_seq_virt,
uint32_t n_ubatch,
uint32_t n_pad);
@ -68,6 +69,8 @@ public:
private:
const llama_hparams & hparams;
const uint32_t n_seq_virt = 1;
std::unique_ptr<llama_kv_cache_unified> kv_base;
std::unique_ptr<llama_kv_cache_unified> kv_swa;
};

702
src/llama-kv-cache-unified.cpp
View File

File diff suppressed because it is too large Load Diff

55
src/llama-kv-cache-unified.h
View File

@ -41,10 +41,31 @@ public:
// data for ggml_set_rows
using idx_vec_t = std::vector<uint32_t>;
idx_vec_t idxs;
llama_seq_id s0;
llama_seq_id s1;
std::vector<llama_seq_id> seq_id_virt;
std::vector<idx_vec_t> idxs;
uint32_t head() const {
return idxs.at(0);
GGML_ASSERT(idxs.size() == 1);
return idxs.at(0).at(0);
}
void resize(size_t n) {
seq_id_virt.resize(n);
idxs.resize(n);
}
size_t size() const {
GGML_ASSERT(idxs.size() == seq_id_virt.size());
return idxs.at(0).size();
}
size_t n_seq_virt() const {
return seq_id_virt.size();
}
bool empty() const {
@ -54,9 +75,6 @@ public:
void clear() {
idxs.clear();
}
// TODO: implement
//std::vector<idx_vec_t> seq_idxs;
};
using slot_info_vec_t = std::vector<slot_info>;
@ -70,6 +88,7 @@ public:
bool offload,
uint32_t kv_size,
uint32_t n_seq_max,
uint32_t n_seq_virt,
uint32_t n_pad,
uint32_t n_swa,
llama_swa_type swa_type);
@ -122,8 +141,8 @@ public:
uint32_t get_n_kv() const;
// get views of the current state of the cache
ggml_tensor * get_k(ggml_context * ctx, int32_t il, uint32_t n_kv) const;
ggml_tensor * get_v(ggml_context * ctx, int32_t il, uint32_t n_kv) const;
ggml_tensor * get_k(ggml_context * ctx, int32_t il, uint32_t n_kv, const slot_info & sinfo) const;
ggml_tensor * get_v(ggml_context * ctx, int32_t il, uint32_t n_kv, const slot_info & sinfo) const;
// store k_cur and v_cur in the cache based on the provided head location
ggml_tensor * cpy_k(ggml_context * ctx, ggml_tensor * k_cur, ggml_tensor * k_idxs, int32_t il, const slot_info & sinfo) const;
@ -157,8 +176,9 @@ public:
void set_input_k_idxs(ggml_tensor * dst, const llama_ubatch * ubatch, const slot_info & sinfo) const;
void set_input_v_idxs(ggml_tensor * dst, const llama_ubatch * ubatch, const slot_info & sinfo) const;
void set_input_k_shift(ggml_tensor * dst) const;
void set_input_kq_mask (ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn) const;
void set_input_k_shift (ggml_tensor * dst) const;
void set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch * ubatch) const;
private:
@ -172,15 +192,15 @@ private:
ggml_tensor * k;
ggml_tensor * v;
std::vector<ggml_tensor *> k_seq;
std::vector<ggml_tensor *> v_seq;
};
bool v_trans = true; // the value tensor is transposed
// the current index from where we start searching for a free slot in the ring buffer of KV cells (see find_slot())
// note: this is not part of the KV state and it's only used to speed-up the find_slot() method
uint32_t head = 0;
const uint32_t n_seq_max = 1;
const uint32_t n_seq_max = 1;
const uint32_t n_seq_virt = 1;
// required padding
const uint32_t n_pad = 1;
@ -200,7 +220,14 @@ private:
std::vector<ggml_context_ptr> ctxs;
std::vector<ggml_backend_buffer_ptr> bufs;
llama_kv_cells_unified cells;
// the current index from where we start searching for a free slot in the ring buffer of KV cells (see find_slot())
// note: this is not part of the KV state and it's only used to speed-up the find_slot() method
std::vector<uint32_t> v_heads;
std::vector<llama_kv_cells_unified> v_cells;
// maps from a sequence id to a virtual sequence id
std::vector<uint32_t> seq_virt_idx;
std::vector<kv_layer> layers;

8
src/llama-memory-hybrid.cpp
View File

@ -40,6 +40,7 @@ llama_memory_hybrid::llama_memory_hybrid(
offload,
kv_size,
n_seq_max,
1,
n_pad,
n_swa,
swa_type
@ -70,7 +71,7 @@ llama_memory_context_ptr llama_memory_hybrid::init_batch(llama_batch_allocr & ba
// if all tokens are output, split by sequence
ubatch = balloc.split_seq(n_ubatch);
} else {
ubatch = balloc.split_equal(n_ubatch);
ubatch = balloc.split_equal(n_ubatch, false);
}
if (ubatch.n_tokens == 0) {
@ -80,6 +81,11 @@ llama_memory_context_ptr llama_memory_hybrid::init_batch(llama_batch_allocr & ba
ubatches.push_back(std::move(ubatch)); // NOLINT
}
if (balloc.get_n_used() < balloc.get_n_tokens()) {
// failed to find a suitable split
break;
}
// prepare the recurrent batches first
if (!mem_recr->prepare(ubatches)) {
// TODO: will the recurrent cache be in an undefined context at this point?

5
src/llama-memory-recurrent.cpp
View File

@ -374,10 +374,11 @@ llama_memory_context_ptr llama_memory_recurrent::init_batch(llama_batch_allocr &
// if all tokens are output, split by sequence
ubatch = balloc.split_seq(n_ubatch);
} else {
ubatch = balloc.split_equal(n_ubatch);
ubatch = balloc.split_equal(n_ubatch, false);
}
if (ubatch.n_tokens == 0) {
if (balloc.get_n_used() < balloc.get_n_tokens()) {
// failed to find a suitable split
break;
}

2
src/llama-model.cpp
View File

@ -14499,6 +14499,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
params.swa_full,
cparams.n_ctx,
cparams.n_seq_max,
cparams.n_seq_virt,
cparams.n_ubatch,
padding);
} else {
@ -14513,6 +14514,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
cparams.offload_kqv,
cparams.n_ctx,
cparams.n_seq_max,
cparams.n_seq_virt,
padding,
hparams.n_swa,
hparams.swa_type);

8
tools/batched-bench/batched-bench.cpp
View File

@ -61,7 +61,7 @@ int main(int argc, char ** argv) {
const int32_t n_kv_max = llama_n_ctx(ctx);
llama_batch batch = llama_batch_init(n_kv_max, 0, 1);
llama_batch batch = llama_batch_init(n_kv_max*8, 0, 1); // TODO: tmp!!!
// decode in batches of ctx_params.n_batch tokens
auto decode_helper = [](llama_context * ctx, llama_batch & batch, int32_t n_batch) {
@ -119,9 +119,9 @@ int main(int argc, char ** argv) {
const int n_ctx_req = is_pp_shared ? pp + pl*tg : pl*(pp + tg);
if (n_ctx_req > n_kv_max) {
continue;
}
//if (n_ctx_req > n_kv_max) {
// continue;
//}
common_batch_clear(batch);