A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 13, 15:45 UTC
[Log] update shm wait time msg (#28255)
Boyuan Feng committed
Nov 7, 2025 at 06:43 UTC
0f872b7977e50ab65b920d58898b37016cf49a6f
1 file changed
+12
-10
vllm/distributed/device_communicators/shm_broadcast.py
+12
-10
@@ -49,6 +49,16 @@ def to_bytes_big(value: int, size: int) -> bytes:
49
logger = init_logger(__name__)
50
51
52
+def long_wait_time_msg(threshold: int) -> str:
53
+ return (
54
+ "No available shared memory broadcast block found "
55
+ f"in {threshold} seconds. This typically happens "
56
+ "when some processes are hanging or doing some "
57
+ "time-consuming work (e.g. compilation, "
58
+ "weight/kv cache quantization)."
59
+ )
60
+
61
+
62
class SpinTimer:
63
def record_activity(self):
64
pass
@@ -422,11 +432,7 @@ class MessageQueue:
432
# if we wait for a long time, log a message
433
if elapsed > VLLM_RINGBUFFER_WARNING_INTERVAL * n_warning:
434
logger.info(
425
- "No available shared memory broadcast block found"
426
- " in %s seconds. This typically happens when some"
427
- " processes are hanging or doing some"
428
- " time-consuming work (e.g. compilation)",
429
- VLLM_RINGBUFFER_WARNING_INTERVAL,
435
+ long_wait_time_msg(VLLM_RINGBUFFER_WARNING_INTERVAL)
436
)
437
n_warning += 1
438
@@ -493,11 +499,7 @@ class MessageQueue:
499
elapsed > VLLM_RINGBUFFER_WARNING_INTERVAL * n_warning
500
):
501
logger.info(
496
- "No available shared memory broadcast block found"
497
- " in %s seconds. This typically happens when some"
498
- " processes are hanging or doing some"
499
- " time-consuming work (e.g. compilation).",
500
- VLLM_RINGBUFFER_WARNING_INTERVAL,
502
+ long_wait_time_msg(VLLM_RINGBUFFER_WARNING_INTERVAL)
503
)
504
n_warning += 1
505