From eeee25791327d94bd5ea5460f5df7c1f7198d4e8 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E6=B1=AA=E5=BF=97=E9=B9=8F?= Date: Mon, 7 Sep 2026 07:21:21 +0800 Subject: [PATCH] [Bugfix][Diffusion] Restore the TRTLLM attention default for the MiniMax-H3 modular alias (#7162) Signed-off-by: princepride --- vllm_omni/diffusion/model_metadata.py | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) diff --git a/vllm_omni/diffusion/model_metadata.py b/vllm_omni/diffusion/model_metadata.py index 67c58fbec..38df83846 100644 --- a/vllm_omni/diffusion/model_metadata.py +++ b/vllm_omni/diffusion/model_metadata.py @@ -48,12 +48,16 @@ _DIFFUSION_MODEL_METADATA: dict[str, DiffusionModelMetadata] = { attention_mask_free=True, ), # The modular alias is served by MiniMaxH3Pipeline and has the same - # Ref2VA request contract. Keep admission limits in sync with it. + # Ref2VA request contract and packed-sequence layout. Every field must stay + # in sync with it: the repository root ``model_index.json`` declares this + # alias, so serving the checkpoint by repo id resolves here rather than to + # ``MiniMaxH3Pipeline``. "MiniMaxH3ModularPipeline": DiffusionModelMetadata( supports_multimodal_inputs=True, max_multimodal_image_inputs=9, supports_mixed_reference_inputs=True, final_output_type="video", + attention_mask_free=True, ), "Magi2Pipeline": DiffusionModelMetadata( supports_multimodal_inputs=True,