From 4d022cbc757e89b2296daf3b33303eeea4ba4a62 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Nicol=C3=B2=20Lucchesi?= Date: Fri, 11 Apr 2025 19:06:14 +0200 Subject: [PATCH] [TPU][V1] Make `--disable_chunked_mm_input` mandatory for serving MM models (#16483) Signed-off-by: NickLucche --- vllm/platforms/tpu.py | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/vllm/platforms/tpu.py b/vllm/platforms/tpu.py index ada599c27b446..d8807a72ba2f3 100644 --- a/vllm/platforms/tpu.py +++ b/vllm/platforms/tpu.py @@ -120,6 +120,13 @@ class TpuPlatform(Platform): assert not vllm_config.speculative_config, ( "Speculative decoding is not yet supported for TPU backend") + if scheduler_config.is_multimodal_model and not \ + scheduler_config.disable_chunked_mm_input: + logger.warning("TPU does not support running Multimodal models"\ + " without setting `--disable_chunked_mm_input`. " \ + "Forcing --disable_chunked_mm_input.") + scheduler_config.disable_chunked_mm_input = True + @classmethod def is_pin_memory_available(cls): logger.warning("Pin memory is not supported on TPU.")