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 14, 05:30 UTC
[Misc] Clean up MM profiling warnings (#25222)
Signed-off-by: Roger Wang <hey@rogerw.io>
Roger Wang committed
Sep 18, 2025 at 21:46 UTC
31a8a2a7bccb29612bc58c9a69252bfb78f5abe4
1 file changed
-29
vllm/multimodal/profiling.py
-29
@@ -234,19 +234,6 @@ class MultiModalProfiler(Generic[_I]):
234
prompt_token_ids = mm_inputs["prompt_token_ids"]
235
total_len = len(prompt_token_ids)
236
237
- # V0 does not support chunked prefill.
238
- if total_len > seq_len and not envs.VLLM_USE_V1:
239
- # `max_num_batched_tokens` is defined by `SchedulerConfig`
240
- logger.warning_once(
241
- "The sequence length used for profiling (max_num_batched_tokens / max_num_seqs = %d) " # noqa: E501
242
- "is too short to hold the multi-modal embeddings in the worst case (%d tokens in total, out of which %s are reserved for multi-modal embeddings). " # noqa: E501
243
- "This may cause certain multi-modal inputs to fail during inference, even when the input text is short. " # noqa: E501
244
- "To avoid this, you should increase `max_model_len`, reduce `max_num_seqs`, and/or reduce `mm_counts`.", # noqa: E501
245
- seq_len,
246
- total_len,
247
- str(self._get_mm_num_tokens(mm_inputs)),
248
- )
249
-
237
if total_len < seq_len:
238
prompt_token_ids.extend([0] * (seq_len - total_len))
239
@@ -270,22 +257,6 @@ class MultiModalProfiler(Generic[_I]):
257
mm_counts=mm_counts,
258
)
259
if max_tokens_per_item is not None:
273
- if mm_counts is None:
274
- total_mm_tokens = sum(max_tokens_per_item.values())
275
- else:
276
- total_mm_tokens = sum(max_tokens_per_item[k] * mm_counts[k]
277
- for k in max_tokens_per_item.keys()
278
- & mm_counts.keys())
279
- if total_mm_tokens > seq_len:
280
- logger.warning_once(
281
- "The sequence length (%d) is smaller than the pre-defined"
282
- " worst-case total number of multimodal tokens (%d). "
283
- "This may cause certain multi-modal inputs to fail during "
284
- "inference. To avoid this, you should increase "
285
- "`max_model_len` or reduce `mm_counts`.",
286
- seq_len,
287
- total_mm_tokens,
288
- )
260
return max_tokens_per_item
261
262
mm_inputs = self._get_dummy_mm_inputs(seq_len, mm_counts)