CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-04 09:32:00 +00:00
Code Issues Packages Projects Releases Wiki Activity

server : use slot context size instead of training context size

Browse Source
This commit is contained in:
Georgi Gerganov
2025-10-28 11:39:07 +02:00
parent e776168267
commit 7ebe7f77a1
2 changed files with 4 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
tools/server/server.cpp
View File

@@ -2946,17 +2946,15 @@ 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) {
if (slot.task->params.n_predict < 1 && slot.n_prompt_tokens() + slot.n_decoded >= slot.n_ctx) {
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);
"Limiting generated tokens to slot.n_ctx (%d) to avoid EOS-less generation infinite loop\n",
slot.task->params.n_predict, slot.n_ctx);
}
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());