CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-01 09:01:57 +00:00
Code Issues Packages Projects Releases Wiki Activity

memory : remove KV cache size padding (#16812)

Browse Source 
* memory : remove KV cache size padding

* cont : restore padding for n_kv tensor shape

* server : use slot context size instead of training context size

* server : simplify context limit logic
This commit is contained in:
Georgi Gerganov
2025-10-28 20:19:44 +02:00
committed by GitHub
parent a8ca18b4b8
commit 85a7d8677b
6 changed files with 14 additions and 54 deletions
Download Patch File Download Diff File Expand all files Collapse all files

27
tools/server/server.cpp
View File

@@ -2866,10 +2866,12 @@ struct server_context {
// if context shifting is disabled, make sure that we don't run out of context
if (!params_base.ctx_shift && slot.n_past + 1 >= slot.n_ctx) {
slot.truncated = true;
slot.stop = STOP_TYPE_LIMIT;
slot.has_next_token = false;
SLT_DBG(slot, "stopped due to running out of context, n_past = %d, n_ctx = %d\n", slot.n_past, slot.n_ctx);
SLT_DBG(slot, "stopped due to running out of context capacity, n_past = %d, n_prompt_tokens = %d, n_decoded = %d, n_ctx = %d\n",
slot.n_decoded, slot.n_prompt_tokens(), slot.n_past, slot.n_ctx);
}
// check the limits
@@ -2929,16 +2931,6 @@ struct server_context {
}
}
// if context shift is disabled, we stop when it reaches the context limit
if (slot.n_past >= slot.n_ctx) {
slot.truncated = true;
slot.stop = STOP_TYPE_LIMIT;
slot.has_next_token = false;
SLT_DBG(slot, "stopped due to running out of context capacity, n_past = %d, n_prompt_tokens = %d, n_decoded = %d, n_ctx = %d\n",
slot.n_decoded, slot.n_prompt_tokens(), slot.n_past, slot.n_ctx);
}
if (llama_vocab_is_eog(vocab, result.tok)) {
slot.stop = STOP_TYPE_EOS;
slot.has_next_token = false;
@@ -2946,19 +2938,6 @@ struct server_context {
SLT_DBG(slot, "%s", "stopped by EOS\n");
}
const auto n_ctx_train = llama_model_n_ctx_train(model);
if (slot.task->params.n_predict < 1 && slot.n_prompt_tokens() + slot.n_decoded >= n_ctx_train) {
slot.truncated = true;
slot.stop = STOP_TYPE_LIMIT;
slot.has_next_token = false; // stop prediction
SLT_WRN(slot,
"n_predict (%d) is set for infinite generation. "
"Limiting generated tokens to n_ctx_train (%d) to avoid EOS-less generation infinite loop\n",
slot.task->params.n_predict, n_ctx_train);
}
SLT_DBG(slot, "n_decoded = %d, n_remaining = %d, next token: %5d '%s'\n", slot.n_decoded, slot.n_remaining, result.tok, token_str.c_str());
return slot.has_next_token; // continue