vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-11 01:24:36 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge branch 'upstream' into concedo_experimental

Browse source 
# Conflicts:
#	.github/workflows/bench.yml
#	.github/workflows/build.yml
#	.github/workflows/python-check-requirements.yml
#	README.md
#	docs/backend/SYCL.md
#	flake.lock
#	ggml/CMakeLists.txt
#	ggml/src/kompute-shaders/op_rope_f16.comp
#	ggml/src/kompute-shaders/op_rope_f32.comp
#	ggml/src/kompute-shaders/rope_common.comp
This commit is contained in:
Concedo 2024-08-14 22:25:43 +08:00
commit e8de0af3ec
18 changed files with 1326 additions and 101 deletions
Download patch file Download diff file Expand all files Collapse all files

38
examples/server/server.cpp
View file

@ -632,6 +632,7 @@ struct server_context {
bool clean_kv_cache = true;
bool add_bos_token = true;
bool has_eos_token = false;
int32_t n_ctx; // total context for all clients / slots
@ -694,7 +695,7 @@ struct server_context {
n_ctx = llama_n_ctx(ctx);
add_bos_token = llama_should_add_bos_token(model);
GGML_ASSERT(llama_add_eos_token(model) != 1);
has_eos_token = llama_add_eos_token(model) != 1;
return true;
}
@ -754,13 +755,13 @@ struct server_context {
default_generation_settings_for_props = get_formated_generation(slots.front());
default_generation_settings_for_props["seed"] = -1;
// the update_slots() logic will always submit a maximum of n_batch tokens
// the update_slots() logic will always submit a maximum of n_batch or n_parallel tokens
// note that n_batch can be > n_ctx (e.g. for non-causal attention models such as BERT where the KV cache is not used)
{
const int32_t n_batch = llama_n_batch(ctx);
// only a single seq_id per token is needed
batch = llama_batch_init(n_batch, 0, 1);
batch = llama_batch_init(std::max(n_batch, params.n_parallel), 0, 1);
}
metrics.init();
@ -1032,7 +1033,7 @@ struct server_context {
{
slot.sparams.logit_bias.clear();
if (json_value(data, "ignore_eos", false)) {
if (json_value(data, "ignore_eos", false) && has_eos_token) {
slot.sparams.logit_bias[llama_token_eos(model)] = -INFINITY;
}
@ -1137,28 +1138,19 @@ struct server_context {
if (!system_prompt.empty()) {
system_tokens = ::llama_tokenize(ctx, system_prompt, true);
llama_batch_clear(batch);
for (int i = 0; i < (int)system_tokens.size(); ++i) {
llama_batch_add(batch, system_tokens[i], i, { 0 }, false);
}
const int32_t n_batch = llama_n_batch(ctx);
const int32_t n_tokens_prompt = system_tokens.size();
for (int32_t i = 0; i < batch.n_tokens; i += n_batch) {
const int32_t n_tokens = std::min(params.n_batch, batch.n_tokens - i);
llama_batch batch_view = {
n_tokens,
batch.token + i,
nullptr,
batch.pos + i,
batch.n_seq_id + i,
batch.seq_id + i,
batch.logits + i,
0, 0, 0, // unused
};
for (int32_t i = 0; i < n_tokens_prompt; i += n_batch) {
const int32_t n_tokens = std::min(n_batch, n_tokens_prompt - i);
if (llama_decode(ctx, batch_view) != 0) {
llama_batch_clear(batch);
for (int32_t j = 0; j < n_tokens; ++j) {
llama_batch_add(batch, system_tokens[i + j], i + j, { 0 }, false);
}
if (llama_decode(ctx, batch) != 0) {
LOG_ERROR("llama_decode() failed", {});
return;
}