如何才能快速训练GLM 5.2

Your Question

采用slime框架sft GLM 5.2，有点崩溃。256卡 h200，试了各种训练配置，太慢了，一个训练 step 4～5min。。。。如何才能加快训练速度？不知道是不是我设置的原因，真的有点崩溃

What I've Tried

我试了各种TP/PP/CP的组合，都没有什么效果，一个训练step就是大概4-5min....

Environment (if relevant)

slime version:
v0.3.1-30-g41014d1f

Git commit: 41014d1f29e201137fdffce737bb8bac65bc5219
Working tree: dirty（包含当前本地修改）

Python version:
3.12.3
GCC 13.3.0

Python version:
3.12.3
GCC 13.3.0

PyTorch version:
2.11.0+cu129

CUDA/ROCm version:
CUDA 12.9
CUDA compiler build: 12.9.r12.9
NVIDIA driver CUDA API: 13.2
NCCL: 2.28.9+cuda12.9
ROCm: N/A

GPU type and count:
256 × NVIDIA H200（约 141 GB/GPU）
32 nodes × 8 GPUs/node

OS:
Ubuntu 24.04.2 LTS (Noble Numbat)
Linux kernel 5.15.0-174-generic
x86_64, glibc 2.39

训练镜像：

registry.h.pjlab.org.cn/ailab-puyullmgpu-puyullm_gpu/lvhaijun:
slime_nightly-dev-20260810a-cu129

Additional Context

No response

Pre-submission Checklist

• I have read the CONTRIBUTING.md and understand the collaboration scope.
• I have read the documentation and FAQ and my question is not answered there.
• I have searched for existing issues and my question has not been asked before.

No file, test, or entry point is named. Start by reproducing the SFT run with the reported environment and profiling a single training step across the tried TP/PP/CP configurations. Done means identifying the bottleneck or missing configuration detail and documenting a validated faster setup.