vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-11 09:34:37 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge branch 'master' into concedo_experimental

Browse source 
# Conflicts:
#	Makefile
#	README.md
This commit is contained in:
Concedo 2023-11-27 14:06:14 +08:00
commit 8acd7be734
20 changed files with 1274 additions and 136 deletions
Download patch file Download diff file Expand all files Collapse all files

243
llama.cpp
View file

@ -1123,6 +1123,12 @@ static std::string llama_token_to_str(const struct llama_context * ctx, llama_to
//
struct llama_state {
llama_state() {
#ifdef GGML_USE_METAL
ggml_metal_log_set_callback(log_callback, log_callback_user_data);
#endif
}
// We save the log callback globally
ggml_log_callback log_callback = llama_log_callback_default;
void * log_callback_user_data = nullptr;
@ -1285,6 +1291,7 @@ struct llama_kv_cache {
// cannot be freely changed after a slot has been allocated.
uint32_t head = 0;
uint32_t size = 0;
uint32_t used = 0; // used cells (i.e. at least one seq_id)
// computed before each graph build
uint32_t n = 0;
@ -1514,6 +1521,7 @@ static bool llama_kv_cache_init(
cache.head = 0;
cache.size = n_ctx;
cache.used = 0;
cache.cells.clear();
cache.cells.resize(n_ctx);
@ -1615,6 +1623,8 @@ static bool llama_kv_cache_find_slot(
}
}
cache.used += n_tokens;
return true;
}
@ -1635,6 +1645,7 @@ static void llama_kv_cache_clear(struct llama_kv_cache & cache) {
cache.cells[i].seq_id.clear();
}
cache.head = 0;
cache.used = 0;
}
static void llama_kv_cache_seq_rm(
@ -1657,6 +1668,9 @@ static void llama_kv_cache_seq_rm(
continue;
}
if (cache.cells[i].seq_id.empty()) {
// keep count of the number of used cells
if (cache.cells[i].pos >= 0) cache.used--;
cache.cells[i].pos = -1;
if (new_head == cache.size) new_head = i;
}
@ -1664,7 +1678,7 @@ static void llama_kv_cache_seq_rm(
}
// If we freed up a slot, set head to it so searching can start there.
if (new_head != cache.size) cache.head = new_head;
if (new_head != cache.size && new_head < cache.head) cache.head = new_head;
}
static void llama_kv_cache_seq_cp(
@ -1690,6 +1704,7 @@ static void llama_kv_cache_seq_keep(struct llama_kv_cache & cache, llama_seq_id
for (uint32_t i = 0; i < cache.size; ++i) {
if (!cache.cells[i].has_seq_id(seq_id)) {
if (cache.cells[i].pos >= 0) cache.used--;
cache.cells[i].pos = -1;
cache.cells[i].seq_id.clear();
if (new_head == cache.size) new_head = i;
@ -1700,7 +1715,7 @@ static void llama_kv_cache_seq_keep(struct llama_kv_cache & cache, llama_seq_id
}
// If we freed up a slot, set head to it so searching can start there.
if (new_head != cache.size) cache.head = new_head;
if (new_head != cache.size && new_head < cache.head) cache.head = new_head;
}
static void llama_kv_cache_seq_shift(
@ -1721,6 +1736,7 @@ static void llama_kv_cache_seq_shift(
cache.cells[i].delta += delta;
if (cache.cells[i].pos < 0) {
if (!cache.cells[i].seq_id.empty()) cache.used--;
cache.cells[i].pos = -1;
cache.cells[i].seq_id.clear();
if (new_head == cache.size) new_head = i;
@ -3489,7 +3505,7 @@ static void llm_build_k_shift(
struct ggml_cgraph * graph,
llm_rope_type type,
int64_t n_ctx,
int64_t n_rot,
int n_rot,
float freq_base,
float freq_scale,
const llm_build_cb & cb) {
@ -3521,7 +3537,7 @@ static void llm_build_k_shift(
// we rotate only the first n_rot dimensions
ggml_rope_custom_inplace(ctx,
ggml_view_3d(ctx, kv.k,
n_rot, n_head_kv, n_ctx,
n_embd_head, n_head_kv, n_ctx,
ggml_element_size(kv.k)*n_embd_head,
ggml_element_size(kv.k)*n_embd_gqa,
ggml_element_size(kv.k)*n_embd_gqa*n_ctx*il),
@ -4844,92 +4860,34 @@ struct llm_build_context {
// self-attention
{
// compute Q and K and RoPE them
struct ggml_tensor * tmpq = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
cb(tmpq, "tmpq", il);
struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
cb(Qcur, "Qcur", il);
struct ggml_tensor * tmpk = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
cb(tmpk, "tmpk", il);
struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
cb(Kcur, "Kcur", il);
struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
cb(Vcur, "Vcur", il);
// RoPE the first n_rot of q/k, pass the other half, and concat.
struct ggml_tensor * qrot = ggml_cont(ctx0, ggml_view_3d(
ctx0, tmpq, hparams.n_rot, n_head, n_tokens,
ggml_element_size(tmpq) * n_embd_head,
ggml_element_size(tmpq) * n_embd_head * n_head,
0
));
cb(qrot, "qrot", il);
struct ggml_tensor * krot = ggml_cont(ctx0, ggml_view_3d(
ctx0, tmpk, hparams.n_rot, n_head, n_tokens,
ggml_element_size(tmpk) * n_embd_head,
ggml_element_size(tmpk) * n_embd_head * n_head_kv,
0
));
cb(krot, "krot", il);
// get the second half of tmpq, e.g tmpq[n_rot:, :, :]
struct ggml_tensor * qpass = ggml_view_3d(
ctx0, tmpq, (n_embd_head - hparams.n_rot), n_head, n_tokens,
ggml_element_size(tmpq) * n_embd_head,
ggml_element_size(tmpq) * n_embd_head * n_head,
ggml_element_size(tmpq) * hparams.n_rot
);
cb(qpass, "qpass", il);
struct ggml_tensor * kpass = ggml_view_3d(
ctx0, tmpk, (n_embd_head - hparams.n_rot), n_head_kv, n_tokens,
ggml_element_size(tmpk) * (n_embd_head),
ggml_element_size(tmpk) * (n_embd_head) * n_head_kv,
ggml_element_size(tmpk) * hparams.n_rot
);
cb(kpass, "kpass", il);
struct ggml_tensor * qrotated = ggml_rope_custom(
ctx0, qrot, inp_pos, hparams.n_rot, 2, 0, n_orig_ctx,
freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
Qcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos,
hparams.n_rot, 2, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(qrotated, "qrotated", il);
struct ggml_tensor * krotated = ggml_rope_custom(
ctx0, krot, inp_pos, hparams.n_rot, 2, 0, n_orig_ctx,
freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
);
cb(krotated, "krotated", il);
// ggml currently only supports concatenation on dim=2
// so we need to permute qrot, qpass, concat, then permute back.
qrotated = ggml_cont(ctx0, ggml_permute(ctx0, qrotated, 2, 1, 0, 3));
cb(qrotated, "qrotated", il);
krotated = ggml_cont(ctx0, ggml_permute(ctx0, krotated, 2, 1, 0, 3));
cb(krotated, "krotated", il);
qpass = ggml_cont(ctx0, ggml_permute(ctx0, qpass, 2, 1, 0, 3));
cb(qpass, "qpass", il);
kpass = ggml_cont(ctx0, ggml_permute(ctx0, kpass, 2, 1, 0, 3));
cb(kpass, "kpass", il);
struct ggml_tensor * Qcur = ggml_concat(ctx0, qrotated, qpass);
cb(Qcur, "Qcur", il);
struct ggml_tensor * Kcur = ggml_concat(ctx0, krotated, kpass);
cb(Kcur, "Kcur", il);
struct ggml_tensor * Q = ggml_cont(ctx0, ggml_permute(ctx0, Qcur, 2, 1, 0, 3));
cb(Q, "Q", il);
Kcur = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 2, 1, 0, 3));
Kcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
hparams.n_rot, 2, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Kcur, "Kcur", il);
llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);
cur = llm_build_kqv(ctx0, hparams, kv_self,
model.layers[il].wo, NULL,
Q, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il);
Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il);
cb(cur, "kqv_out", il);
}
@ -5557,6 +5515,12 @@ static int llama_decode_internal(
batch.seq_id = seq_id_arr.data();
}
// if we have enough unused cells before the current head ->
// better to start searching from the beginning of the cache, hoping to fill it
if (kv_self.head > kv_self.used + 2*n_tokens) {
kv_self.head = 0;
}
if (!llama_kv_cache_find_slot(kv_self, batch)) {
return 1;
}
@ -5567,7 +5531,7 @@ static int llama_decode_internal(
//kv_self.n = std::max(32, GGML_PAD(llama_kv_cache_cell_max(kv_self), 32)); // TODO: this might be better for CUDA?
kv_self.n = std::min((int32_t) cparams.n_ctx, std::max(32, llama_kv_cache_cell_max(kv_self)));
//printf("kv_self.n = %d\n", kv_self.n);
//printf("kv_self.n = %5d, kv_self.used = %5d, kv_self.head = %5d\n", kv_self.n, kv_self.used, kv_self.head);
ggml_allocr_reset(lctx.alloc);
@ -6716,10 +6680,13 @@ struct llama_grammar_candidate {
// pointer. If an invalid sequence is encountered, returns `llama_partial_utf8.n_remain == -1`.
static std::pair<std::vector<uint32_t>, llama_partial_utf8> decode_utf8(
const char * src,
size_t n_src,
llama_partial_utf8 partial_start) {
static const int lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 2, 2, 3, 4 };
const char * pos = src;
std::vector<uint32_t> code_points;
// common english strings have the same number of codepoints and bytes. `+ 1` for the terminating 0.
code_points.reserve(n_src + 1);
uint32_t value = partial_start.value;
int n_remain = partial_start.n_remain;
@ -6770,6 +6737,13 @@ static std::pair<std::vector<uint32_t>, llama_partial_utf8> decode_utf8(
return std::make_pair(std::move(code_points), llama_partial_utf8{ value, n_remain });
}
static std::pair<std::vector<uint32_t>, llama_partial_utf8> decode_utf8(
std::string src,
llama_partial_utf8 partial_start
) {
return decode_utf8(src.c_str(), src.size(), partial_start);
}
// returns true iff pos points to the end of one of the definitions of a rule
static bool llama_grammar_is_end_of_sequence(const llama_grammar_element * pos) {
switch (pos->type) {
@ -7422,7 +7396,7 @@ void llama_sample_grammar(struct llama_context * ctx, llama_token_data_array * c
} else if (piece.empty() || piece[0] == 0) {
candidates->data[i].logit = -INFINITY;
} else {
candidates_decoded.push_back(decode_utf8(piece.c_str(), grammar->partial_utf8));
candidates_decoded.push_back(decode_utf8(piece, grammar->partial_utf8));
candidates_grammar.push_back({ i, candidates_decoded.back().first.data(), candidates_decoded.back().second });
}
}
@ -7629,7 +7603,7 @@ void llama_grammar_accept_token(struct llama_context * ctx, struct llama_grammar
const std::string piece = llama_token_to_str(ctx, token);
// Note terminating 0 in decoded string
const auto decoded = decode_utf8(piece.c_str(), grammar->partial_utf8);
const auto decoded = decode_utf8(piece, grammar->partial_utf8);
const auto & code_points = decoded.first;
for (auto it = code_points.begin(), end = code_points.end() - 1; it != end; ++it) {
grammar->stacks = llama_grammar_accept(grammar->rules, grammar->stacks, *it);
@ -8877,8 +8851,6 @@ struct llama_context * llama_new_context_with_model(
#ifdef GGML_USE_METAL
if (model->n_gpu_layers > 0) {
ggml_metal_log_set_callback(llama_log_callback_default, NULL);
ctx->ctx_metal = ggml_metal_init(1);
if (!ctx->ctx_metal) {
LLAMA_LOG_ERROR("%s: ggml_metal_init() failed\n", __func__);
@ -9113,8 +9085,107 @@ int llama_model_apply_lora_from_file(const struct llama_model * model, const cha
}
}
struct llama_kv_cache_view llama_kv_cache_view_init(const struct llama_context * ctx, int32_t n_max_seq) {
struct llama_kv_cache_view result = {
/*.n_cells = */ 0,
/*.n_max_seq = */ n_max_seq,
/*.token_count = */ 0,
/*.used_cells = */ llama_get_kv_cache_used_cells(ctx),
/*.max_contiguous = */ 0,
/*.max_contiguous_idx = */ -1,
/*.cells = */ nullptr,
/*.cells_sequences = */ nullptr,
};
return result;
}
void llama_kv_cache_view_free(struct llama_kv_cache_view * view) {
if (view->cells != nullptr) {
free(view->cells);
view->cells = nullptr;
}
if (view->cells_sequences != nullptr) {
free(view->cells_sequences);
view->cells_sequences = nullptr;
}
}
void llama_kv_cache_view_update(const struct llama_context * ctx, struct llama_kv_cache_view * view) {
if (uint32_t(view->n_cells) < ctx->kv_self.size || view->cells == nullptr) {
view->n_cells = int32_t(ctx->kv_self.size);
void * p = realloc(view->cells, sizeof(struct llama_kv_cache_view_cell) * view->n_cells);
GGML_ASSERT(p != nullptr && "Failed to alloc kv_cache_view cells");
view->cells = (struct llama_kv_cache_view_cell *)p;
p = realloc(view->cells_sequences, sizeof(llama_seq_id) * view->n_max_seq * view->n_cells);
GGML_ASSERT(p != nullptr && "Failed to alloc kv_cache_view cells sequences");
view->cells_sequences = (llama_seq_id *)p;
}
const std::vector<llama_kv_cell> & kv_cells = ctx->kv_self.cells;
llama_kv_cache_view_cell * c_curr = view->cells;
llama_seq_id * cs_curr = view->cells_sequences;
int32_t used_cells = 0;
int32_t token_count = 0;
int32_t curr_contig_idx = -1;
uint32_t max_contig = 0;
int32_t max_contig_idx = -1;
for (int32_t i = 0; i < int32_t(ctx->kv_self.size); i++, c_curr++, cs_curr += view->n_max_seq) {
const size_t curr_size = kv_cells[i].seq_id.size();
token_count += curr_size;
c_curr->pos = kv_cells[i].pos + kv_cells[i].delta;
if (curr_size > 0) {
if (curr_contig_idx >= 0 && uint32_t(i - curr_contig_idx) > max_contig) {
max_contig = i - curr_contig_idx;
max_contig_idx = curr_contig_idx;
}
curr_contig_idx = -1;
} else if (curr_contig_idx < 0) {
curr_contig_idx = i;
}
int seq_idx = 0;
for (const llama_seq_id it : kv_cells[i].seq_id) {
if (seq_idx >= view->n_max_seq) {
break;
}
cs_curr[seq_idx] = it;
seq_idx++;
}
if (seq_idx != 0) {
used_cells++;
}
for (; seq_idx < view->n_max_seq; seq_idx++) {
cs_curr[seq_idx] = -1;
}
}
if (curr_contig_idx >= 0 && kv_cells.size() - curr_contig_idx > max_contig) {
max_contig_idx = curr_contig_idx;
max_contig = kv_cells.size() - curr_contig_idx;
}
view->max_contiguous = max_contig;
view->max_contiguous_idx = max_contig_idx;
view->token_count = token_count;
view->used_cells = used_cells;
if (uint32_t(used_cells) != ctx->kv_self.used) {
LLAMA_LOG_ERROR("%s: used cells mismatch. kv_cache says %d but we calculated %d\n",
__func__, ctx->kv_self.used, used_cells);
}
}
int llama_get_kv_cache_token_count(const struct llama_context * ctx) {
return ctx->kv_self.head;
int result = 0;
for (uint32_t i = 0; i < ctx->kv_self.size; i++) {
result += ctx->kv_self.cells[i].seq_id.size();
}
return result;
}
int llama_get_kv_cache_used_cells(const struct llama_context * ctx) {
return ctx->kv_self.used;
}
void llama_kv_cache_clear(struct llama_context * ctx) {
@ -9284,10 +9355,12 @@ static void llama_copy_state_data_internal(struct llama_context * ctx, llama_dat
const size_t kv_buf_size = kv_self.buf.size;
const uint32_t kv_head = kv_self.head;
const uint32_t kv_size = kv_self.size;
const uint32_t kv_used = kv_self.used;
data_ctx->write(&kv_buf_size, sizeof(kv_buf_size));
data_ctx->write(&kv_head, sizeof(kv_head));
data_ctx->write(&kv_size, sizeof(kv_size));
data_ctx->write(&kv_used, sizeof(kv_used));
if (kv_buf_size) {
const size_t elt_size = ggml_element_size(kv_self.k);
@ -9410,10 +9483,12 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) {
size_t kv_buf_size;
uint32_t kv_head;
uint32_t kv_size;
uint32_t kv_used;
memcpy(&kv_buf_size, inp, sizeof(kv_buf_size)); inp += sizeof(kv_buf_size);
memcpy(&kv_head, inp, sizeof(kv_head)); inp += sizeof(kv_head);
memcpy(&kv_size, inp, sizeof(kv_size)); inp += sizeof(kv_size);
memcpy(&kv_used, inp, sizeof(kv_used)); inp += sizeof(kv_used);
if (kv_buf_size) {
GGML_ASSERT(kv_self.buf.size == kv_buf_size);
@ -9448,6 +9523,7 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) {
ctx->kv_self.head = kv_head;
ctx->kv_self.size = kv_size;
ctx->kv_self.used = kv_used;
ctx->kv_self.cells.resize(kv_size);
@ -9923,6 +9999,9 @@ const std::vector<std::pair<std::string, struct ggml_tensor *>> & llama_internal
void llama_log_set(ggml_log_callback log_callback, void * user_data) {
g_state.log_callback = log_callback ? log_callback : llama_log_callback_default;
g_state.log_callback_user_data = user_data;
#ifdef GGML_USE_METAL
ggml_metal_log_set_callback(g_state.log_callback, g_state.log_callback_user_data);
#endif
}
static void llama_log_internal_v(ggml_log_level level, const char * format, va_list args) {