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 16, 18:15 UTC
[multi-step] Raise error if not using async engine (#7703)
William Lin committed
Aug 21, 2024 at 11:49 UTC
91f4522cbf85df0a65f619b25f6751edf2d5f0d6
1 file changed
+5
vllm/engine/llm_engine.py
+5
@@ -1302,6 +1302,11 @@ class LLMEngine:
1302
raise NotImplementedError(
1303
"Pipeline parallelism is only supported through AsyncLLMEngine "
1304
"as performance will be severely degraded otherwise.")
1305
+
1306
+ if self.scheduler_config.num_scheduler_steps > 1:
1307
+ raise NotImplementedError(
1308
+ "Multiple scheduler steps (multi-step) are only supported "
1309
+ "through AsyncLLMEngine. ")
1310
seq_group_metadata_list, scheduler_outputs = self.scheduler[
1311
0].schedule()
1312