Use syslog, move the sock to dexd subdirectory.

1 files changed, 20 insertions, 12 deletions

diff --git a/llm.c b/llm.c
index 4d3dc58..e979929 100644
--- a/llm.c
+++ b/llm.c
@@ -1,6 +1,7 @@
 #include <stdio.h>
 #include <err.h>
 #include <string.h>
+#include <syslog.h>
 
 #include "llm.h"
 #include "mem.h"
@@ -47,9 +48,17 @@ struct llm_ctx {
 static void llm_log_cb(enum ggml_log_level level, const char *s, void *ud)
 {
 	(void)ud;
-	if (level < GGML_LOG_LEVEL_WARN)
+
+	switch (level) {
+	case GGML_LOG_LEVEL_WARN:
+		syslog(LOG_WARNING, "%s", s);
+		return;
+	case GGML_LOG_LEVEL_ERROR:
+		syslog(LOG_ERR, "%s", s);
 		return;
-	fputs(s, stderr);
+	default:
+		return;
+	}
 }
 
 /* Tokenize `text` (already a fully rendered chat-template prompt) into 
@@ -105,7 +114,7 @@ struct llm_ctx *llm_init(const char *model_path)
 	}
 
 	if (!llama_model_chat_template(llm->model, NULL))
-		fprintf(stderr, "WARN: model has no embedded chat template\n");
+		syslog(LOG_WARNING, "model has no embedded chat template");
 
 	/* Creation of context allocates KV cache and compute buffers.
 	 * Do this once in init, instead of per request. */
@@ -206,14 +215,14 @@ void llm_run(struct llm_ctx *llm, const char *user_prompt, FILE *out)
 	 * assumptions about the model's internal chat structure. */
 	prompt_len = llama_chat_apply_template(NULL, messages, 2, true, NULL, 0);
 	if (prompt_len <= 0) {
-		fprintf(stderr, "ERROR: Chat template size calculation failed\n");
+		syslog(LOG_ERR, "chat template size calculation failed");
 		return;
 	}
 
 	prompt = MALLOC((size_t)prompt_len + 1);
 	if (llama_chat_apply_template(NULL, messages, 2, true, prompt, 
 		prompt_len + 1) < 0) {
-		fprintf(stderr, "ERROR: failed to apply chat template\n");
+		syslog(LOG_ERR, "failed to apply chat template");
 		free(prompt);
 		return;
 	}
@@ -221,12 +230,12 @@ void llm_run(struct llm_ctx *llm, const char *user_prompt, FILE *out)
 	n_new_tokens = tokenize_prompt(llm->vocab, prompt, prompt_len, &new_tokens);
 	free(prompt);
 	if (n_new_tokens <= 0) {
-		fprintf(stderr, "ERROR: tokenization failed\n");
+		syslog(LOG_ERR, "tokenization failed");
 		return;
 	}
 
 	if (n_new_tokens + MAX_TOKENS > N_CTX) {
-		fprintf(stderr, "ERROR: token count exceeds context size\n");
+		syslog(LOG_ERR, "token count exceeds context window");
 		free(new_tokens);
 		return;
 	}
@@ -244,9 +253,8 @@ void llm_run(struct llm_ctx *llm, const char *user_prompt, FILE *out)
 
 	/* Sanity check */
 	if (llm->n_cached_tokens > 0 && n_common < llm->n_system_tokens) {
-		fprintf(stderr, "WARN: Cached prefix is shorter than the "
-			"system prompt. Chat template may be rendering "
-			"inconsistently between requests.");
+		/* Template may be rendering inconsistently between requests */
+		syslog(LOG_WARNING, "cached prefix is shorter than the system prompt");
 	}
 
 	/* Drop everything in the cache past the common prefix. */
@@ -261,7 +269,7 @@ void llm_run(struct llm_ctx *llm, const char *user_prompt, FILE *out)
 		batch = llama_batch_get_one(new_tokens + n_common, 
 			n_new_tokens - n_common);
 		if (llama_decode(llm->ctx, batch) != 0) {
-			fprintf(stderr, "ERROR: prompt evaluation failed\n");
+			syslog(LOG_ERR, "prompt evaluation failed");
 			free(new_tokens);
 			llm->n_cached_tokens = n_common; 
 			return;
@@ -299,7 +307,7 @@ void llm_run(struct llm_ctx *llm, const char *user_prompt, FILE *out)
 		
 		batch = llama_batch_get_one(&new_token_id, 1);
 		if (llama_decode(llm->ctx, batch) != 0) {
-			fprintf(stderr, "ERROR: llama_decode failed!\n");
+			syslog(LOG_ERR, "decode() error in generation loop");
 			break;
 		}