[NPU] Add MiniMax M2.5 support

[Fulin-Gao]

Summary

This issue documents the end-to-end deployment and training workflow for running MiniMax M2.5 GRPO training with Vime on Ascend NPU (Atlas A3), intended as a reference for the community.

Item	Version	Commit
Ascend HDK	25.5.1.B030	N/A
CANN	9.0.0	N/A
python	3.12.13	N/A
torch_npu	2.10.0.post2	N/A
triton-ascend	3.2.1	N/A
vllm	0.22.1+empty	0decac0
vllm-ascend	0.22.1rc1	5f6faa0c
Megatron-Bridge	N/A	7f0fb345
Megatron-LM	0.16.0rc0	1dcf0dafa
MegatronAdaptor	0.1.0	15582ad
TransformerEngineNPU	2.13.0	d743c83
flash-linear-attention-npu	1.0.0	14c2c92

Usage

1. Setup

git clone -b ascend https://github.com/vllm-project/vime.git
cd vime
docker build -f docker/Dockerfile.npu -t vime-ascend:latest .

export IMAGE=vime-ascend:latest

docker run -d --name vime-npu -it --net=host --shm-size=1024g \
--privileged=true \
--cap-add=SYS_PTRACE \
--device=/dev/davinci_manager \
--device=/dev/hisi_hdc \
--device=/dev/devmm_svm \
-v /usr/local/Ascend/driver:/usr/local/Ascend/driver \
-v /usr/local/dcmi:/usr/local/dcmi \
-v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \
-v /usr/local/sbin:/usr/local/sbin \
-v /home:/home \
-v /mnt:/mnt \
-v /tmp:/tmp \
-v /data:/data \
-v /usr/share/zoneinfo/Asia/Shanghai:/etc/localtime \
$IMAGE

docker exec -it vime-npu bash

2. Download

# (1) Model
hf download PrimeIntellect/MiniMax-M2.5-bf16 --local-dir /path/to/MiniMax-M2.5-bf16

# (2) Data
hf download --repo-type dataset zhuzilin/dapo-math-17k --local-dir /path/to/dapo-math-17k
hf download --repo-type dataset zhuzilin/aime-2024 --local-dir /path/to/aime-2024

3. Trim

trim_safetensor_layers.py

python /path/to/trim_safetensor_layers.py \
  --model_dir /path/to/MiniMax-M2.5-bf16 \
  --out_dir /path/to/MiniMax-M2.5-bf16-2layer \
  --num_layers 2 \
  --write_skipped \
  --asset-include "*.md" \
  --asset-include "*.jinja" \
  --asset-include "*.txt" \
  --asset-include "*.py"

4. Training

cd /root/vime
bash scripts/run-minimax-m2-npu.sh

[aoshen02]

One context is that slime remove some models and after we sync, there maybe some models that people need to bring back, and we can refer to this pr #393 to add it back.