From b864374089749ad3b4e6d01cb5d4f627beeb6174 Mon Sep 17 00:00:00 2001 From: Rakesh Kariya <83279947+rk9595@users.noreply.github.com> Date: Mon, 31 Aug 2026 11:26:35 +0400 Subject: [PATCH] [Bugfix][MiMo-Audio] Restore make_empty_intermediate_tensors on the talker for vLLM 0.28 (#6803) Signed-off-by: rk9595 --- .../model_executor/models/mimo_audio/mimo_audio_llm.py | 8 ++++++++ 1 file changed, 8 insertions(+) diff --git a/vllm_omni/model_executor/models/mimo_audio/mimo_audio_llm.py b/vllm_omni/model_executor/models/mimo_audio/mimo_audio_llm.py index 19a339056..c90adb467 100644 --- a/vllm_omni/model_executor/models/mimo_audio/mimo_audio_llm.py +++ b/vllm_omni/model_executor/models/mimo_audio/mimo_audio_llm.py @@ -1,3 +1,6 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project + # Copyright 2025 Xiaomi Corporation. import logging import threading @@ -541,6 +544,11 @@ class MiMoAudioLLMForConditionalGeneration(nn.Module, SupportsMultiModal, Suppor architectures=["Qwen2ForCausalLM"], ) + # vLLM 0.28 turned SupportsPP.make_empty_intermediate_tensors from a + # method into a bare annotation, so declaring SupportsPP no longer + # supplies one. Delegate to the inner model, as the other omni talkers do. + self.make_empty_intermediate_tensors = self.model.make_empty_intermediate_tensors + self.device = current_omni_platform.get_torch_device() # global_sampler MUST stay greedy (do_sample=False) so its token decision # matches vLLM's external sampler (SamplingParams temperature=0.0). Both