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, 18:30 UTC
[Doc] Fix misleading log during multi-modal profiling (#14955)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Cyrus Leung committed
Mar 17, 2025 at 21:14 UTC
166a168b0fa606608bbd3b4be1ab5d904a4e3927
1 file changed
+3
-1
vllm/multimodal/profiling.py
+3
-1
@@ -218,8 +218,10 @@ class MultiModalProfiler(Generic[_I]):
218
219
# V0 does not support chunked prefill.
220
if total_len > seq_len and not envs.VLLM_USE_V1:
221
+ # `max_num_batched_tokens` is defined by `SchedulerConfig`
222
logger.warning(
222
- "The context length (%d) of the model is too short "
223
+ "The sequence length used for profiling ("
224
+ "max_num_batched_tokens / max_num_seqs = %d) is too short "
225
"to hold the multi-modal embeddings in the worst case "
226
"(%d tokens in total, out of which %s are reserved for "
227
"multi-modal embeddings). This may cause certain "