token healing : change argument order

Author: mare5x

common/sampling.cpp

@@ -128,8 +128,8 @@ static llama_token_healing_output llama_token_healing_get_prefix(
 
 llama_token_healing_output llama_token_healing_rollback(
                            const llama_context * ctx_main,
-                           llama_token_healing_type th_type,
                            std::vector<llama_token> & tokens,
+                           llama_token_healing_type th_type,
                            int max_to_remove) {
     // NB. To avoid returning empty `tokens`, at least 1 token will remain in `tokens` after rolling back.
     //     It is the caller's responsibility to add BOS to the start of the prompt if they want to roll back the whole prompt.

common/sampling.h

@@ -189,8 +189,8 @@ struct llama_token_healing_output {
 // Call `llama_token_healing_set_prefix` with the returned prefix before the first sampling.
 llama_token_healing_output llama_token_healing_rollback(
                            const llama_context * ctx_main,
-                           llama_token_healing_type th_type,
                            std::vector<llama_token> & tokens,
+                           llama_token_healing_type th_type,
                            int max_to_remove = -1);
 
 void llama_token_healing_set_prefix(llama_sampling_context * ctx_sampling, const std::string & prefix);

examples/main/main.cpp

@@ -282,8 +282,8 @@ int main(int argc, char ** argv) {
     }
     llama_token_healing_output token_healing_out{};
     if (!params.interactive_first && sparams.token_healing.enabled) {
-        token_healing_out = llama_token_healing_rollback(ctx, sparams.token_healing.type, embd_inp,
-                                                              sparams.token_healing.n_rollback);
+        token_healing_out = llama_token_healing_rollback(ctx, embd_inp,
+            sparams.token_healing.type, sparams.token_healing.n_rollback);
     }
 
     // Should not run without any tokens
@@ -917,7 +917,7 @@ int main(int argc, char ** argv) {
                         const int max_to_remove = sparams.token_healing.n_rollback < 0
                                                    ? n_new_tokens
                                                    : std::min(sparams.token_healing.n_rollback, n_new_tokens);
-                        token_healing_out = llama_token_healing_rollback(ctx, sparams.token_healing.type, embd_inp, max_to_remove);
+                        token_healing_out = llama_token_healing_rollback(ctx, embd_inp, sparams.token_healing.type, max_to_remove);
                         n_bytes_to_skip = token_healing_out.prefix.size();
                     }
 

examples/server/server.cpp

@@ -2076,8 +2076,8 @@ struct server_context {
 
                             if (slot.sparams.token_healing.enabled) {
                                 // For FIM roll back only the prefix part (i.e. cursor location)
-                                token_healing_out = llama_token_healing_rollback(ctx, slot.sparams.token_healing.type,
-                                    prefix_tokens, slot.sparams.token_healing.n_rollback);
+                                token_healing_out = llama_token_healing_rollback(ctx, prefix_tokens,
+                                    slot.sparams.token_healing.type, slot.sparams.token_healing.n_rollback);
                             }
 
                             auto embd_inp = params.spm_infill ? suffix_tokens : prefix_tokens;
@@ -2097,8 +2097,8 @@ struct server_context {
                             prompt_tokens = tokenize(slot.prompt, system_prompt.empty()); // add BOS if there isn't system prompt
 
                             if (slot.sparams.token_healing.enabled) {
-                                token_healing_out = llama_token_healing_rollback(ctx, slot.sparams.token_healing.type,
-                                    prompt_tokens, slot.sparams.token_healing.n_rollback);
+                                token_healing_out = llama_token_healing_rollback(ctx, prompt_tokens,
+                                    slot.sparams.token_healing.type, slot.sparams.token_healing.n_rollback);
                             }
                         }