Use syslog, move the sock to dexd subdirectory.

4 files changed, 76 insertions, 40 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;
 		}
 
diff --git a/main.c b/main.c
index e60c00d..b9c42b0 100644
--- a/main.c
+++ b/main.c
@@ -1,14 +1,16 @@
 #include <stdio.h>
+#include <stdlib.h>
 #include <err.h>
 #include <errno.h>
 #include <string.h>
+#include <syslog.h>
 #include <unistd.h>
 #include <sys/un.h>
 #include <sys/socket.h>
 
 #include "llm.h"
 
-#define SOCK_PATH  "/var/run/lex.sock"
+#define SOCK_PATH  "/var/run/lexd/sock"
 #define MODEL_PATH "qwen2.5-3b-instruct-q4_k_m.gguf"
 
 static void usage(const char *prog)
@@ -27,12 +29,12 @@ static void loop(struct llm_ctx *llm, int lsock)
 		csock = accept(lsock, NULL, NULL);
 		if (csock == -1) {
 			if (errno != EINTR)
-				warn("accept()");
+				syslog(LOG_WARNING, "accept() failed: %m");
 			continue;
 		}
 
 		if (!(cfp = fdopen(csock, "r+"))) {
-			warn("cfp fdopen()");
+			syslog(LOG_WARNING, "fdopen() failed for cfp: %m");
 			close(csock);
 			continue;
 		}
@@ -40,7 +42,7 @@ static void loop(struct llm_ctx *llm, int lsock)
 		if (fgets(line, sizeof(line), cfp) != NULL) {
 			len = strcspn(line, "\n");
 			if (line[len] == '\0' && len == sizeof(line) - 1) {
-				warnx("request too large");
+				syslog(LOG_WARNING, "request too large");
 				// drain and skip
 				while ((c = fgetc(cfp)) != '\n' && c != EOF)
 					;
@@ -63,8 +65,10 @@ int sock_init(void)
 	struct sockaddr_un addr;
 
 	sock = socket(AF_UNIX, SOCK_STREAM, 0);
-	if (sock == -1)
-		err(1, "socket()");
+	if (sock == -1) {
+		syslog(LOG_ERR, "socket error: %m");
+		exit(1);
+	}
 
 	memset(&addr, 0, sizeof(addr));
 	addr.sun_family = AF_UNIX;
@@ -72,11 +76,15 @@ int sock_init(void)
 
 	unlink(SOCK_PATH); /* remove stale socket from a previous run */
 	
-	if (bind(sock, (struct sockaddr *)&addr, sizeof(addr)) == -1)
-		err(1, "bind()");
+	if (bind(sock, (struct sockaddr *)&addr, sizeof(addr)) == -1) {
+		syslog(LOG_ERR, "bind failed: %m");
+		exit(1);
+	}
 
-	if (listen(sock, 1) == -1) /* backlog = 1 */
-		err(1, "listen()");
+	if (listen(sock, 1) == -1) { /* backlog = 1 */
+		syslog(LOG_ERR, "listen(): %m");
+		exit(1);
+	}
 
 	return sock;
 }
@@ -91,6 +99,8 @@ int main(int argc , char *argv[])
 	prog = argv[0];
 	model_path = MODEL_PATH;
 
+	openlog("lexd", LOG_PID | LOG_NDELAY, LOG_DAEMON);
+
 	while ((opt = getopt(argc, argv, "m:")) != -1) {
 		switch (opt) {
 		case 'm':
@@ -106,23 +116,34 @@ int main(int argc , char *argv[])
 	if (argc != 0) 
 		usage(prog);
 
-	if (unveil(model_path, "r") == -1)
-		err(1, "unveil %s failed", model_path);
-	if (unveil(SOCK_PATH, "rwc") == -1)
-		err(1, "unveil %s failed", SOCK_PATH);
-	if (unveil(NULL, NULL) == -1)
-		err(1, "unveil lock failed");
+	if (unveil(model_path, "r") == -1) {
+		syslog(LOG_ERR, "unveil %s: %m", model_path);
+		exit(1);
+	}
+
+	if (unveil(SOCK_PATH, "rwc") == -1) {
+		syslog(LOG_ERR, "unveil %s: %m", SOCK_PATH);
+		exit(1);
+	}
+
+	if (unveil(NULL, NULL) == -1) {
+		syslog(LOG_ERR, "unveil lock failed: %m");
+		exit(1);
+	}
 
 	sock = sock_init();
 
-	if (!(llm = llm_init(model_path)))
-		errx(1, "Failed to load model");
+	if (!(llm = llm_init(model_path))) {
+		syslog(LOG_ERR, "failed to load model");
+		exit(1);
+	}
 
 	/* llm_init() calls mlock, hence pledge() after that. */
-	if (pledge("stdio unix", NULL) == -1)
-		err(1, "secondary pledge failed");
+	if (pledge("stdio unix", NULL) == -1) {
+		syslog(LOG_ERR, "secondary pledge: %m");
+		exit(1);
+	}
 
 	loop(llm, sock); /* doesn't return */
-
 	return 0;
 }
diff --git a/mem.h b/mem.h
index 2cdb36e..852de0a 100644
--- a/mem.h
+++ b/mem.h
@@ -3,6 +3,7 @@
 
 #include <err.h>
 #include <stdlib.h>
+#include <syslog.h>
 
 #define MALLOC(s)      xmalloc((s), __FILE__, __LINE__)
 #define CALLOC(n, s)   xcalloc((n), (s), __FILE__, __LINE__)
@@ -12,8 +13,10 @@ static inline void *xmalloc(size_t s, const char *file, int line)
 {
 	void *p;
 
-	if (!(p = malloc(s)))
-		err(1, "%s:%d: malloc", file, line);
+	if (!(p = malloc(s))) {
+		syslog(LOG_ERR, "malloc: %s (line %d)", file, line);
+		exit(1);
+	}
 	return p;
 }
 
@@ -21,8 +24,10 @@ static inline void *xcalloc(size_t n, size_t s, const char *file, int line)
 {
 	void *p;
 
-	if (!(p = calloc(n, s)))
-		err(1, "%s:%d: calloc", file, line);
+	if (!(p = calloc(n, s))) {
+		syslog(LOG_ERR, "calloc: %s (line %d)", file, line);
+		exit(1);
+	}
 	return p;
 }
 
@@ -30,8 +35,10 @@ static inline void *xrealloc(void *ptr, size_t s, const char *file, int line)
 {
 	void *p;
 
-	if (!(p = realloc(ptr, s)))
-		err(1, "%s:%d: realloc", file, line);
+	if (!(p = realloc(ptr, s))) {
+		syslog(LOG_ERR, "realloc: %s (line %d)", file, line);
+		exit(1);
+	}
 	return p;
 }
 
diff --git a/test_prompts.sh b/test_prompts.sh
index 0424e62..1de3822 100755
--- a/test_prompts.sh
+++ b/test_prompts.sh
@@ -2,7 +2,7 @@
 
 # Usage: ./test_prompts.sh <model_path> [prompt_file]
 
-SOCK_PATH="/var/run/lex.sock"
+SOCK_PATH="/var/run/lexd/sock"
 EXECUTABLE="./lex"
 
 # Mandatory Model Argument