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, 16:00 UTC
Print the warning only once (#16193)
Signed-off-by: Gregory Shtrasberg <Gregory.Shtrasberg@amd.com>
Gregory Shtrasberg committed
Apr 7, 2025 at 14:30 UTC
ad434d4cfe1349059868c8b4132dd5237a3e147f
1 file changed
+12
-10
vllm/multimodal/profiling.py
+12
-10
@@ -216,17 +216,18 @@ class MultiModalProfiler(Generic[_I]):
216
# Encoder-decoder multimodal models only support v0
217
if total_len > seq_len:
218
# `max_num_batched_tokens` is defined by `SchedulerConfig`
219
- logger.warning(
219
+ logger.warning_once(
220
"The encoder sequence length used for profiling ("
221
- "max_num_batched_tokens / max_num_seqs = %d) is too short "
221
+ f"max_num_batched_tokens / max_num_seqs = {seq_len}) "
222
+ " is too short "
223
"to hold the multi-modal embeddings in the worst case "
223
- "(%d tokens in total, out of which %s are reserved for "
224
+ f"({total_len} tokens in total, out of which "
225
+ f"{total_placeholders_by_modality} are reserved for "
226
"multi-modal embeddings). This may cause certain "
227
"multi-modal inputs to fail during inference, even when "
228
"the input text is short. To avoid this, you should "
229
"increase `max_model_len`, reduce `max_num_seqs`, "
228
- "and/or reduce `mm_counts`.", seq_len, total_len,
229
- total_placeholders_by_modality)
230
+ "and/or reduce `mm_counts`.")
231
232
processor = cast(EncDecMultiModalProcessor, self.processor)
233
if processor.pad_dummy_encoder_prompt:
@@ -251,17 +252,18 @@ class MultiModalProfiler(Generic[_I]):
252
# V0 does not support chunked prefill.
253
if total_len > seq_len and not envs.VLLM_USE_V1:
254
# `max_num_batched_tokens` is defined by `SchedulerConfig`
254
- logger.warning(
255
+ logger.warning_once(
256
"The sequence length used for profiling ("
256
- "max_num_batched_tokens / max_num_seqs = %d) is too short "
257
+ f"max_num_batched_tokens / max_num_seqs = {seq_len}) "
258
+ "is too short "
259
"to hold the multi-modal embeddings in the worst case "
258
- "(%d tokens in total, out of which %s are reserved for "
260
+ f"({total_len} tokens in total, out of which "
261
+ f"{total_placeholders_by_modality} are reserved for "
262
"multi-modal embeddings). This may cause certain "
263
"multi-modal inputs to fail during inference, even when "
264
"the input text is short. To avoid this, you should "
265
"increase `max_model_len`, reduce `max_num_seqs`, "
263
- "and/or reduce `mm_counts`.", seq_len, total_len,
264
- total_placeholders_by_modality)
266
+ "and/or reduce `mm_counts`.")
267
268
if total_len < seq_len:
269
prompt_token_ids.extend([0] * (seq_len - total_len))