Bug fix: repeated prompt must not be ignored.

1 files changed, 8 insertions, 0 deletions

diff --git a/llm.c b/llm.c
index e979929..784df27 100644
--- a/llm.c
+++ b/llm.c
@@ -257,6 +257,14 @@ void llm_run(struct llm_ctx *llm, const char *user_prompt, FILE *out)
 		syslog(LOG_WARNING, "cached prefix is shorter than the system prompt");
 	}
 
+	/* Force at least one real decode every request. If n_common hit
+	 * n_new_tokens, decode gets skipped entirely, so ctx keeps the
+	 * stale logits from the previous request's EOG -- sampling them
+	 * again just reproduces EOG immediately (empty response on an
+	 * exact repeat). */
+	if (n_common == n_new_tokens)
+		n_common--;
+
 	/* Drop everything in the cache past the common prefix. */
 	llama_memory_seq_rm(llama_get_memory(llm->ctx), 
 		0,          /* sequence id: 0 - we only use one sequence */