[Bugfix][Multimodal] Bound renderer warmup to the prefill token budget (#55448)

Signed-off-by: Luca Motz <luca.motz@icloud.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
This commit is contained in:
lucamotz
2026-09-05 16:43:28 +00:00
committed by GitHub
co-authored by OpenAI Codex
parent 7985444339
commit f4eccdadef
4 changed files with 59 additions and 8 deletions
+9 -1
View File
@@ -29,7 +29,12 @@ from .processing import (
)
if TYPE_CHECKING:
from vllm.config import ModelConfig, ObservabilityConfig, VllmConfig
from vllm.config import (
ModelConfig,
ObservabilityConfig,
SchedulerConfig,
VllmConfig,
)
from vllm.model_executor.models.interfaces import SupportsMultiModal
logger = init_logger(__name__)
@@ -243,6 +248,7 @@ class MultiModalRegistry:
*,
cache: BaseMultiModalProcessorCache | None = None,
processor: BaseMultiModalProcessor | None = None,
scheduler_config: "SchedulerConfig | None" = None,
) -> MultiModalInput:
"""
Create dummy data for profiling the memory usage of a model.
@@ -250,6 +256,8 @@ class MultiModalRegistry:
The model is identified by `model_config`.
"""
seq_len = model_config.max_model_len
if scheduler_config is not None and scheduler_config.enable_chunked_prefill:
seq_len = min(seq_len, scheduler_config.max_num_batched_tokens)
if processor is None:
processor = self.create_processor(model_config, cache=cache)