tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-06-28 12:25:03 +00:00
Code Issues Packages Projects Releases Wiki Activity

llama : deprecate llama_kv_self_ API (#14030)

Browse Source 
* llama : deprecate llama_kv_self_ API

ggml-ci

* llama : allow llama_memory_(nullptr)

ggml-ci

* memory : add flag for optional data clear in llama_memory_clear

ggml-ci
This commit is contained in:
Georgi Gerganov
2025-06-06 14:11:15 +03:00
committed by GitHub
parent 487a5e0401
commit 745aa5319b
34 changed files with 206 additions and 127 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
tools/server/server.cpp
View File

@ -2006,7 +2006,7 @@ struct server_context {
}
}
if (!llama_kv_self_can_shift(ctx)) {
if (!llama_memory_can_shift(llama_get_memory(ctx))) {
if (params_base.ctx_shift) {
params_base.ctx_shift = false;
SRV_WRN("%s\n", "ctx_shift is not supported by this context, it will be disabled");
@ -2224,7 +2224,7 @@ struct server_context {
SRV_DBG("%s", "clearing KV cache\n");
// clear the entire KV cache
llama_kv_self_clear(ctx);
llama_memory_clear(llama_get_memory(ctx), true);
clean_kv_cache = false;
}
@ -2910,7 +2910,7 @@ struct server_context {
// Erase token cache
const size_t n_erased = slot->cache_tokens.size();
llama_kv_self_seq_rm(ctx, slot->id, -1, -1);
llama_memory_seq_rm(llama_get_memory(ctx), slot->id, -1, -1);
slot->cache_tokens.clear();
auto res = std::make_unique<server_task_result_slot_erase>();
@ -2985,8 +2985,8 @@ 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_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);
llama_memory_seq_rm (llama_get_memory(ctx), slot.id, n_keep , n_keep + n_discard);
llama_memory_seq_add(llama_get_memory(ctx), slot.id, n_keep + n_discard, slot.n_past, -n_discard);
// add generated tokens to cache
{
@ -3189,8 +3189,8 @@ struct server_context {
const int64_t kv_shift = (int64_t) head_p - (int64_t) head_c;
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);
llama_memory_seq_rm (llama_get_memory(ctx), slot.id, head_p, head_c);
llama_memory_seq_add(llama_get_memory(ctx), slot.id, head_c, head_c + n_match, kv_shift);
for (size_t i = 0; i < n_match; i++) {
slot.cache_tokens.set_token(head_p + i, slot.cache_tokens[head_c + i]);
@ -3212,7 +3212,7 @@ struct server_context {
}
if (slot.n_past > 0 && slot.n_past < (int) slot.cache_tokens.size()) {
const auto pos_min = llama_kv_self_seq_pos_min(ctx, slot.id);
const auto pos_min = llama_memory_seq_pos_min(llama_get_memory(ctx), slot.id);
if (pos_min == -1) {
SLT_ERR(slot, "n_past = %d, cache_tokens.size() = %d, seq_id = %d, pos_min = %d\n", slot.n_past, (int) slot.cache_tokens.size(), slot.id, pos_min);
GGML_ABORT("pos_min == -1, but n_past > 0 - should not happen: https://github.com/ggml-org/llama.cpp/pull/13833#discussion_r2116181237");
@ -3247,9 +3247,9 @@ struct server_context {
}
// keep only the common part
if (!llama_kv_self_seq_rm(ctx, slot.id, slot.n_past, -1)) {
if (!llama_memory_seq_rm(llama_get_memory(ctx), slot.id, slot.n_past, -1)) {
// could not partially delete (likely using a non-Transformer model)
llama_kv_self_seq_rm(ctx, slot.id, -1, -1);
llama_memory_seq_rm(llama_get_memory(ctx), slot.id, -1, -1);
// there is no common part left
slot.n_past = 0;
@ -3589,7 +3589,7 @@ struct server_context {
slot.cache_tokens.push_back(id);
slot.cache_tokens.insert({ids.begin(), ids.end() - 1});
llama_kv_self_seq_rm(ctx, slot.id, slot.n_past, -1);
llama_memory_seq_rm(llama_get_memory(ctx), slot.id, slot.n_past, -1);
for (size_t i = 0; i < ids.size(); ++i) {
completion_token_output result;