kv-cache : fix handling when find_slot fails

ggerganov · ggerganov · commit c96c48c6c8a9 · 2025-07-04T17:59:08.000+03:00
ggml-ci
diff --git a/src/llama-kv-cache-unified.cpp b/src/llama-kv-cache-unified.cpp
@@ -573,7 +573,7 @@ llama_kv_cache_unified::slot_info_vec_t llama_kv_cache_unified::prepare(const st
         apply_ubatch(sinfo_new, ubatch);
     }
 
-    GGML_ASSERT(!states.empty());
+    GGML_ASSERT(!states.empty() || !success);
 
     // iterate backwards and restore the cells to their original state
     for (auto it = states.rbegin(); it != states.rend(); ++it) {

Original file line number	Diff line number	Diff line change
`@@ -573,7 +573,7 @@ llama_kv_cache_unified::slot_info_vec_t llama_kv_cache_unified::prepare(const st`
`573`	`573`	`apply_ubatch(sinfo_new, ubatch);`
`574`	`574`	`}`
`575`	`575`
`576`		`- GGML_ASSERT(!states.empty());`
	`576`	`+ GGML_ASSERT(!states.empty() \|\| !success);`
`577`	`577`
`578`	`578`	`// iterate backwards and restore the cells to their original state`
`579`	`579`	`for (auto it = states.rbegin(); it != states.rend(); ++it) {`