|
|
@@ -3796,7 +3796,7 @@ struct server_context {
|
|
|
|
|
|
// when the prompt prefix does not match, print the tokens around the mismatch
|
|
|
// this is useful for debugging prompt caching
|
|
|
- {
|
|
|
+ if (slots_debug) {
|
|
|
const int np0 = std::max<int>(n_past - 4, 0);
|
|
|
const int np1 = std::min<int>(n_past + 6, std::min(slot.prompt.tokens.size(), slot.task->tokens.size()));
|
|
|
|