server : fix kv cache management (#3588)

diff --git a/examples/server/server.cpp b/examples/server/server.cpp
index d992feeef7026ece01e2484db463c8586ce38dc2..ee0ababb1d5cee93bb2ce40b709f40da70cbec6d 100644 (file)
--- a/examples/server/server.cpp
+++ b/examples/server/server.cpp
@@ -405,6 +405,7 @@ struct llama_server_context
         // compare the evaluated prompt with the new prompt
         n_past = common_part(embd, prompt_tokens);
         embd = prompt_tokens;
+
         if (n_past == num_prompt_tokens)
         {
             // we have to evaluate at least 1 token to generate logits.
@@ -412,6 +413,9 @@ struct llama_server_context
             n_past--;
         }
 
+        // since #3228 we now have to manually manage the KV cache
+        llama_kv_cache_seq_rm(ctx, 0, n_past, -1);
+
         LOG_VERBOSE("prompt ingested", {
                                            {"n_past", n_past},
                                            {"cached", tokens_to_str(ctx, embd.cbegin(), embd.cbegin() + n_past)},
@@ -461,9 +465,6 @@ struct llama_server_context
         // compare the evaluated prompt with the new prompt
         n_past = common_part(embd, prompt_tokens);
 
-        // since #3228 we now have to manually manage the KV cache
-        llama_kv_cache_seq_rm(ctx, 0, n_past, -1);
-
         embd = prompt_tokens;
         if (n_past == num_prompt_tokens)
         {
@@ -471,6 +472,9 @@ struct llama_server_context
             n_past--;
         }
 
+        // since #3228 we now have to manually manage the KV cache
+        llama_kv_cache_seq_rm(ctx, 0, n_past, -1);
+
         LOG_VERBOSE("prompt ingested", {
                                            {"n_past", n_past},
                                            {"cached", tokens_to_str(ctx, embd.cbegin(), embd.cbegin() + n_past)},