| [feature] add MiMoV2.5 model
Co-authored-by: ghoshaw<chenzhiguo6@huawei.com>
# message auto-generated for no-merge-commit merge:
!2448 merge master into master
[feature] add MiMoV2.5 model
Created-by: ghoshaw
Commit-by: ghoshaw
Merged-by: ascend-robot
Description: ## What this PR does / why we need it?
增加 MiMoV2.5 模型支持。
这个是原仓合入的代码,所以代码量比较大。
## Does this PR introduce any user-facing change?
无
## How was this patch tested?
在NPU 上减层(2 层)加载qwen2vl 数据集跑通功能,loss收敛. 未测试语音输入 。测试时视觉部分和语音部分均冻结。 开源权重为量化版,功能跑通使用随机初始化权重。
See merge request: Ascend/MindSpeed-MM!2448 | 26 天前 |