|
| 1 | +# 1. 安装 |
| 2 | + |
| 3 | +**环境依赖** |
| 4 | + |
| 5 | +|Chip type|Driver version| |
| 6 | +|-|-| |
| 7 | +|MUSA S5000|5.1.0-server| |
| 8 | + |
| 9 | +* **机器:** MUSA S5000 80GB 8-card machine |
| 10 | +* **GCC path:** /usr/bin/gcc (12.3.0) |
| 11 | +* **python version:** 3.10 |
| 12 | + |
| 13 | +要验证 MUSA GPU 是否正常,可以使用`mthreads-gmi`命令 |
| 14 | + |
| 15 | +```shell |
| 16 | +mthreads-gmi |
| 17 | +#example:$ mthreads-gmi |
| 18 | + |
| 19 | +Tue Jun 23 17:48:04 2026 |
| 20 | +--------------------------------------------------------------------- |
| 21 | + mthreads-gmi:2.4.1 Driver Version:5.1.0-server |
| 22 | +--------------------------------------------------------------------- |
| 23 | +ID Name |PCIe |%GPU Mem |
| 24 | + Device Type Perf |Pcie Lane Width |Temp MPC Capable |
| 25 | + | ECC Mode |
| 26 | ++-------------------------------------------------------------------+ |
| 27 | +0 MTT S5000 |00000000:18:00.0 |0% 0MiB(81920MiB) |
| 28 | + Physical P0 |16x(16x) |48C YES |
| 29 | + | On-die, EDC |
| 30 | ++-------------------------------------------------------------------+ |
| 31 | +1 MTT S5000 |00000000:3a:00.0 |0% 0MiB(81920MiB) |
| 32 | + Physical P0 |16x(16x) |48C YES |
| 33 | + | On-die, EDC |
| 34 | ++-------------------------------------------------------------------+ |
| 35 | +2 MTT S5000 |00000000:4b:00.0 |0% 0MiB(81920MiB) |
| 36 | + Physical P0 |16x(16x) |49C YES |
| 37 | + | On-die, EDC |
| 38 | ++-------------------------------------------------------------------+ |
| 39 | +3 MTT S5000 |00000000:5c:00.0 |0% 0MiB(81920MiB) |
| 40 | + Physical P0 |16x(16x) |48C YES |
| 41 | + | On-die, EDC |
| 42 | ++-------------------------------------------------------------------+ |
| 43 | +4 MTT S5000 |00000000:9a:00.0 |0% 0MiB(81920MiB) |
| 44 | + Physical P0 |16x(16x) |48C YES |
| 45 | + | On-die, EDC |
| 46 | ++-------------------------------------------------------------------+ |
| 47 | +5 MTT S5000 |00000000:ba:00.0 |0% 0MiB(81920MiB) |
| 48 | + Physical P0 |16x(16x) |50C YES |
| 49 | + | On-die, EDC |
| 50 | ++-------------------------------------------------------------------+ |
| 51 | +6 MTT S5000 |00000000:ca:00.0 |0% 0MiB(81920MiB) |
| 52 | + Physical P0 |16x(16x) |48C YES |
| 53 | + | On-die, EDC |
| 54 | ++-------------------------------------------------------------------+ |
| 55 | +7 MTT S5000 |00000000:da:00.0 |0% 0MiB(81920MiB) |
| 56 | + Physical P0 |16x(16x) |49C YES |
| 57 | + | On-die, EDC |
| 58 | +--------------------------------------------------------------------- |
| 59 | + |
| 60 | +--------------------------------------------------------------------- |
| 61 | +Processes: |
| 62 | +ID PID Process name GPU Memory |
| 63 | + Usage |
| 64 | ++-------------------------------------------------------------------+ |
| 65 | + No running processes found |
| 66 | +--------------------------------------------------------------------- |
| 67 | +``` |
| 68 | +**安装依赖** |
| 69 | + |
| 70 | +1. 拉取镜像 |
| 71 | + |
| 72 | +```shell |
| 73 | +docker pull registry.mthreads.com/public/paddle_musa:musa-5.1.0-py310-ubuntu22.04 |
| 74 | +``` |
| 75 | + |
| 76 | +2. 启动 docker |
| 77 | + |
| 78 | +```shell |
| 79 | +docker run \ |
| 80 | + -it \ |
| 81 | + --network=host \ |
| 82 | + --privileged \ |
| 83 | + --env MTHREADS_VISIBLE_DEVICES=all \ |
| 84 | + --shm-size=80g \ |
| 85 | + --name <container_name> \ |
| 86 | + -v <mnt>:<mnt> \ |
| 87 | + registry.mthreads.com/public/paddle_musa:musa-5.1.0-py310-ubuntu22.04 \ |
| 88 | + /bin/bash |
| 89 | +``` |
| 90 | + |
| 91 | +3. 进入容器运行环境 |
| 92 | + |
| 93 | +```shell |
| 94 | +docker exec -it <container_name> |
| 95 | +``` |
| 96 | + |
| 97 | +4. 拉取代码分支 |
| 98 | + |
| 99 | +```shell |
| 100 | +git clone https://github.com/PaddlePaddle/PaddleFormers.git |
| 101 | +cd PaddleFormers |
| 102 | +python -m pip install -e . |
| 103 | +``` |
| 104 | + |
| 105 | + |
| 106 | +# 2. 开始训练 |
| 107 | + |
| 108 | +所有示例脚本都位于 examples/config/musa 下。 |
| 109 | +下面以 ERNIE-4.5-21B-A3B-PT 的 SFT 为例,可以参考 examples/best_practices/ERNIE-4.5/README.md 进行数据和模型准备。 |
| 110 | + |
| 111 | +如果需要对 ERNIE-4.5-21B-A3B-PT 进行全参数 SFT: |
| 112 | + |
| 113 | +```shell |
| 114 | +bash ./examples/config/musa/ERNIE-4.5-21B-A3B/sft/run_sft.sh |
| 115 | +``` |
0 commit comments