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, 00:45 UTC
[V1] Improve Error Message for Unsupported Config (#12535)
Co-authored-by: Michael Goin <michael@neuralmagic.com>
Robert Shaw committed
Jan 28, 2025 at 23:56 UTC
5f671cb4c3145194e94ffb393ee459432f7fa2b8
1 file changed
+7
-2
vllm/platforms/cuda.py
+7
-2
@@ -120,13 +120,18 @@ class CudaPlatformBase(Platform):
120
if parallel_config.worker_cls == "auto":
121
if scheduler_config.is_multi_step:
122
if envs.VLLM_USE_V1:
123
- raise NotImplementedError
123
+ raise NotImplementedError(
124
+ "Multi-step scheduling is not supported (and not "
125
+ "needed) on VLLM V1. Please launch without "
126
+ "--num-scheduler-steps.")
127
else:
128
parallel_config.worker_cls = \
129
"vllm.worker.multi_step_worker.MultiStepWorker"
130
elif vllm_config.speculative_config:
131
if envs.VLLM_USE_V1:
129
- raise NotImplementedError
132
+ raise NotImplementedError(
133
+ "Speculative decoding is not yet supported on VLLM V1."
134
+ )
135
else:
136
parallel_config.worker_cls = \
137
"vllm.spec_decode.spec_decode_worker.create_spec_worker"