Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-19 01:05:09 +02:00
Code Issues Packages Projects Releases Wiki Activity

traded my ink for a pen

Browse Source
This commit is contained in:
Concedo
2025-03-14 11:58:15 +08:00
parent 52cf1ded0c e0dbec0bc6
commit 0db4ae6237
25 changed files with 13853 additions and 12134 deletions
Download Patch File Download Diff File Expand all files Collapse all files
examples/server/server.cpp
+28 -11
View File
@@ -2040,6 +2040,18 @@ struct server_context {
return ret;
}
bool can_be_detokenized(const struct llama_context * ctx, const std::vector<llama_token> & tokens) {
const llama_model * model = llama_get_model(ctx);
const llama_vocab * vocab = llama_model_get_vocab(model);
const int32_t n_vocab = llama_vocab_n_tokens(vocab);
for (const auto & token : tokens) {
if (token < 0 || token >= n_vocab) {
return false;
}
}
return true;
}
bool launch_slot_with_task(server_slot & slot, const server_task & task) {
slot.reset();
slot.id_task = task.id;
@@ -2054,6 +2066,11 @@ struct server_context {
slot.lora = task.params.lora;
}
bool can_detokenize = can_be_detokenized(ctx, slot.prompt_tokens);
if (!can_detokenize) {
send_error(task, "Prompt contains invalid tokens", ERROR_TYPE_INVALID_REQUEST);
return false;
}
SLT_DBG(slot, "launching slot : %s\n", safe_json_to_str(slot.to_json()).c_str());
if (slot.n_predict > 0 && slot.params.n_predict > slot.n_predict) {
@@ -2096,7 +2113,7 @@ struct server_context {
SRV_DBG("%s", "clearing KV cache\n");
// clear the entire KV cache
llama_kv_cache_clear(ctx);
llama_kv_self_clear(ctx);
clean_kv_cache = false;
}
@@ -2638,8 +2655,8 @@ struct server_context {
res->n_tasks_deferred = queue_tasks.queue_tasks_deferred.size();
res->t_start = metrics.t_start;
res->kv_cache_tokens_count = llama_get_kv_cache_token_count(ctx);
res->kv_cache_used_cells = llama_get_kv_cache_used_cells(ctx);
res->kv_cache_tokens_count = llama_kv_self_n_tokens(ctx);
res->kv_cache_used_cells = llama_kv_self_used_cells(ctx);
res->n_prompt_tokens_processed_total = metrics.n_prompt_tokens_processed_total;
res->t_prompt_processing_total = metrics.t_prompt_processing_total;
@@ -2755,7 +2772,7 @@ struct server_context {
// Erase token cache
const size_t n_erased = slot->cache_tokens.size();
llama_kv_cache_seq_rm(ctx, slot->id, -1, -1);
llama_kv_self_seq_rm(ctx, slot->id, -1, -1);
slot->cache_tokens.clear();
auto res = std::make_unique<server_task_result_slot_erase>();
@@ -2823,8 +2840,8 @@ struct server_context {
SLT_WRN(slot, "slot context shift, n_keep = %d, n_left = %d, n_discard = %d\n", n_keep, n_left, n_discard);
llama_kv_cache_seq_rm (ctx, slot.id, n_keep , n_keep + n_discard);
llama_kv_cache_seq_add(ctx, slot.id, n_keep + n_discard, slot.n_past, -n_discard);
llama_kv_self_seq_rm (ctx, slot.id, n_keep , n_keep + n_discard);
llama_kv_self_seq_add(ctx, slot.id, n_keep + n_discard, slot.n_past, -n_discard);
if (slot.params.cache_prompt) {
for (size_t i = n_keep + n_discard; i < slot.cache_tokens.size(); i++) {
@@ -3015,8 +3032,8 @@ struct server_context {
const int64_t kv_shift = (int64_t) head_p - (int64_t) head_c;
llama_kv_cache_seq_rm (ctx, slot.id, head_p, head_c);
llama_kv_cache_seq_add(ctx, slot.id, head_c, head_c + n_match, kv_shift);
llama_kv_self_seq_rm (ctx, slot.id, head_p, head_c);
llama_kv_self_seq_add(ctx, slot.id, head_c, head_c + n_match, kv_shift);
for (size_t i = 0; i < n_match; i++) {
slot.cache_tokens[head_p + i] = slot.cache_tokens[head_c + i];
@@ -3054,9 +3071,9 @@ struct server_context {
}
// keep only the common part
if (!llama_kv_cache_seq_rm(ctx, slot.id, slot.n_past, -1)) {
if (!llama_kv_self_seq_rm(ctx, slot.id, slot.n_past, -1)) {
// could not partially delete (likely using a non-Transformer model)
llama_kv_cache_seq_rm(ctx, slot.id, -1, -1);
llama_kv_self_seq_rm(ctx, slot.id, -1, -1);
// there is no common part left
slot.n_past = 0;
@@ -3296,7 +3313,7 @@ struct server_context {
slot.cache_tokens.push_back(id);
slot.cache_tokens.insert(slot.cache_tokens.end(), ids.begin(), ids.end() - 1);
llama_kv_cache_seq_rm(ctx, slot.id, slot.n_past, -1);
llama_kv_self_seq_rm(ctx, slot.id, slot.n_past, -1);
for (size_t i = 0; i < ids.size(); ++i) {
completion_token_output result;
examples/server/tests/utils.py
+1 -1
View File
@@ -302,7 +302,7 @@ class ServerPreset:
server.model_hf_repo = "ggml-org/models"
server.model_hf_file = "tinyllamas/stories260K.gguf"
server.model_alias = "tinyllama-2"
server.n_ctx = 256
server.n_ctx = 512
server.n_batch = 32
server.n_slots = 2
server.n_predict = 64