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 15, 11:00 UTC
[BugFix] Fix `--disable-log-stats` in V1 server mode (#17600)
Signed-off-by: Nick Hill <nhill@redhat.com>
Nick Hill committed
May 7, 2025 at 21:08 UTC
3d13ca0e242a99ef1ca53de1828689130924b3f5
2 files changed
+7
-5
vllm/v1/engine/async_llm.py
+3
-2
@@ -120,8 +120,9 @@ class AsyncLLM(EngineClient):
120
executor_class=executor_class,
121
log_stats=self.log_stats,
122
)
123
- for stat_logger in self.stat_loggers[0]:
124
- stat_logger.log_engine_initialized()
123
+ if self.stat_loggers:
124
+ for stat_logger in self.stat_loggers[0]:
125
+ stat_logger.log_engine_initialized()
126
self.output_handler: Optional[asyncio.Task] = None
127
try:
128
# Start output handler eagerly if we are in the asyncio eventloop.
vllm/v1/engine/core_client.py
+4
-3
@@ -442,9 +442,10 @@ class MPClient(EngineCoreClient):
442
logger.info("Core engine process %d ready.", eng_id)
443
identities.discard(eng_id)
444
# Setup KV cache config with initialization state from
445
- # engine core process.
446
- self.vllm_config.cache_config.num_gpu_blocks = message_dict[
447
- 'num_gpu_blocks']
445
+ # engine core process. Sum values from all engines in DP case.
446
+ num_gpu_blocks = self.vllm_config.cache_config.num_gpu_blocks or 0
447
+ num_gpu_blocks += message_dict['num_gpu_blocks']
448
+ self.vllm_config.cache_config.num_gpu_blocks = num_gpu_blocks
449
450
def _init_core_engines(
451
self,