From c3745387af15521930c0433dc34a353f360ca5a6 Mon Sep 17 00:00:00 2001 From: tzhouam Date: Fri, 23 Oct 2026 02:59:46 +0000 Subject: [PATCH] knowledge(vllm-omni): 1 change(s) from refine --- .../components/model-executor/_index.md | 1 + .../model-executor/rules-mimo-audio-device.md | 19 +++++++++++++++++++ 2 files changed, 20 insertions(+) create mode 100644 knowledge/repos/vllm-omni/components/model-executor/rules-mimo-audio-device.md diff --git a/knowledge/repos/vllm-omni/components/model-executor/_index.md b/knowledge/repos/vllm-omni/components/model-executor/_index.md index e83daf55..144b16e8 100644 --- a/knowledge/repos/vllm-omni/components/model-executor/_index.md +++ b/knowledge/repos/vllm-omni/components/model-executor/_index.md @@ -67,3 +67,4 @@ sources: [] - [MiniCPM duplex vision-only unit rules](rules-minicpm-vision-units.md) - [Model Executor — NPU 基础镜像与 Ascend 升级规则](rules-npu-base-image.md) - [MiniCPM-o duplex turn-end drain 规则](rules-minicpmo-turn-end-drain.md) +- [MiMo-Audio 设备与平台支持](rules-mimo-audio-device.md) diff --git a/knowledge/repos/vllm-omni/components/model-executor/rules-mimo-audio-device.md b/knowledge/repos/vllm-omni/components/model-executor/rules-mimo-audio-device.md new file mode 100644 index 00000000..725a1cbe --- /dev/null +++ b/knowledge/repos/vllm-omni/components/model-executor/rules-mimo-audio-device.md @@ -0,0 +1,19 @@ +--- +title: "MiMo-Audio 设备与平台支持" +created: 2026-10-23 +updated: 2026-10-23 +type: rule +tags: [vllm-omni, components, model-executor] +sources: ["PR #8654"] +--- + +# MiMo-Audio 设备与平台支持 + +## VLLM-OMNI-PR8654-MIMO-DEVICE — MiMo-Audio 本地 token 默认设备必须加速器无关 + +- 触发：修改 `vllm_omni/model_executor/models/mimo_audio/mimo_audio_llm.py` 中 `base_local_forward` 或 `local_forward` 的 `tokens_device` 默认参数，或修改 `docs/models/supported_models.md` 中 `MiMoAudioModel` 行的硬件平台勾选。 +- 强制：`tokens_device` 默认值必须以 `torch.accelerator.is_available()` 做可用性门控，并用 `torch.accelerator.current_accelerator().type` 与 `torch.accelerator.current_device_index()` 拼接设备字符串，仅在无加速器时回退 `cpu`；`base_local_forward` 与 `local_forward` 两处默认表达式必须保持一致。为 `MiMoAudioModel` 勾选新硬件列（本 PR 为 Intel GPU）时，必须确认设备推导已按上述方式做到加速器无关。 +- 禁止：在这两个默认参数里硬编码 `"cuda:"` 前缀并用 `torch.cuda.is_available()` 门控——在 XPU 等非 CUDA 加速器上该默认值会静默解析为 `cpu`，使 embedding 查找偏离模块所在设备；不得把通用 accelerator API 与硬编码后端前缀混用。 +- 验收：在 diff 中逐行核对两个签名的默认设备表达式一致且不含硬编码后端名；supported_models.md 中 MiMo-Audio 的平台勾选与设备推导实际支持的加速器后端一一对应。^[PR #8654] + +