aboutsummaryrefslogtreecommitdiff
path: root/tools/server/server_internal.h
diff options
context:
space:
mode:
authorOnur Gündüz <ogunduz@gmail.com>2026-06-24 02:29:52 +0200
committerOnur Gündüz <ogunduz@gmail.com>2026-06-24 02:29:52 +0200
commitdc29ac64a203a954161026ad5f6288df201e11ad (patch)
tree019f84f6a15b53b6134d75ef7f539f986a918fda /tools/server/server_internal.h
parentcb85a499f3d5178a752c85063bf9ff277ca3a71e (diff)
downloadffwd-dc29ac64a203a954161026ad5f6288df201e11ad.tar.gz
ffwd-dc29ac64a203a954161026ad5f6288df201e11ad.zip
server: fix tokenizer data race under concurrency
Diffstat (limited to 'tools/server/server_internal.h')
-rw-r--r--tools/server/server_internal.h8
1 files changed, 4 insertions, 4 deletions
diff --git a/tools/server/server_internal.h b/tools/server/server_internal.h
index b332404..df6b9ea 100644
--- a/tools/server/server_internal.h
+++ b/tools/server/server_internal.h
@@ -105,9 +105,10 @@ typedef struct {
int worker_init_rc;
job *job_head;
job *job_tail;
- /* Tokenizer stage: dispatch enqueues a raw job here when the pipeline is
- * busy, and a dedicated thread parses + tokenizes it off the worker so
- * tokenization overlaps the GPU instead of serializing in front of it. */
+ /* Tokenizer stage: dispatch enqueues every raw job here, and a single
+ * dedicated thread parses + tokenizes it off the worker so tokenization
+ * overlaps the GPU. One thread, so the shared model tokenizer is never
+ * touched concurrently. */
pthread_t tokenizer;
pthread_cond_t raw_cv;
int raw_stopping;
@@ -330,7 +331,6 @@ void execute_rerank_request(rerank_request *r);
/* ---- server_schedule.c: job queue, micro-batching, completion ---- */
void enqueue_job(job *j);
int worker_has_pending_jobs(http_server *s);
-int server_has_backlog(http_server *s);
void enqueue_raw_job(job *j);
job *dequeue_raw_job(http_server *s);
void tokenize_job(http_server *s, job *j);