【AMD ROCm 实战】云端 AI 开发系列（一）：在 ModelScope 上部署 MI300X 并运行第一个 PyTorch 程序

【AMD ROCm 实战】云端 AI 开发系列（一）：在 ModelScope 上部署 MI300X 并运行第一个 PyTorch 程序

摘要: 本文详细记录在魔搭创空间（ModelScope）申请 AMD Instinct MI300X 云实例、配置 ROCm 6.2 环境、安装 PyTorch for ROCm 并运行第一个 GPU 加速程序的完整流程。包含详细的截图、命令输出和性能基准测试数据。

🎯 1. 为什么选择 AMD ROCm 云端开发？

作为一名后端架构师，我在企业级 AI 部署中一直面临三大痛点：

#mermaid-svg-RRUA563aDpamN6pP{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-RRUA563aDpamN6pP .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-RRUA563aDpamN6pP .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-RRUA563aDpamN6pP .error-icon{fill:#552222;}#mermaid-svg-RRUA563aDpamN6pP .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-RRUA563aDpamN6pP .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-RRUA563aDpamN6pP .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-RRUA563aDpamN6pP .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-RRUA563aDpamN6pP .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-RRUA563aDpamN6pP .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-RRUA563aDpamN6pP .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-RRUA563aDpamN6pP .marker{fill:#333333;stroke:#333333;}#mermaid-svg-RRUA563aDpamN6pP .marker.cross{stroke:#333333;}#mermaid-svg-RRUA563aDpamN6pP svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-RRUA563aDpamN6pP p{margin:0;}#mermaid-svg-RRUA563aDpamN6pP .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-RRUA563aDpamN6pP .cluster-label text{fill:#333;}#mermaid-svg-RRUA563aDpamN6pP .cluster-label span{color:#333;}#mermaid-svg-RRUA563aDpamN6pP .cluster-label span p{background-color:transparent;}#mermaid-svg-RRUA563aDpamN6pP .label text,#mermaid-svg-RRUA563aDpamN6pP span{fill:#333;color:#333;}#mermaid-svg-RRUA563aDpamN6pP .node rect,#mermaid-svg-RRUA563aDpamN6pP .node circle,#mermaid-svg-RRUA563aDpamN6pP .node ellipse,#mermaid-svg-RRUA563aDpamN6pP .node polygon,#mermaid-svg-RRUA563aDpamN6pP .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-RRUA563aDpamN6pP .rough-node .label text,#mermaid-svg-RRUA563aDpamN6pP .node .label text,#mermaid-svg-RRUA563aDpamN6pP .image-shape .label,#mermaid-svg-RRUA563aDpamN6pP .icon-shape .label{text-anchor:middle;}#mermaid-svg-RRUA563aDpamN6pP .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-RRUA563aDpamN6pP .rough-node .label,#mermaid-svg-RRUA563aDpamN6pP .node .label,#mermaid-svg-RRUA563aDpamN6pP .image-shape .label,#mermaid-svg-RRUA563aDpamN6pP .icon-shape .label{text-align:center;}#mermaid-svg-RRUA563aDpamN6pP .node.clickable{cursor:pointer;}#mermaid-svg-RRUA563aDpamN6pP .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-RRUA563aDpamN6pP .arrowheadPath{fill:#333333;}#mermaid-svg-RRUA563aDpamN6pP .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-RRUA563aDpamN6pP .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-RRUA563aDpamN6pP .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-RRUA563aDpamN6pP .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-RRUA563aDpamN6pP .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-RRUA563aDpamN6pP .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-RRUA563aDpamN6pP .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-RRUA563aDpamN6pP .cluster text{fill:#333;}#mermaid-svg-RRUA563aDpamN6pP .cluster span{color:#333;}#mermaid-svg-RRUA563aDpamN6pP div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-RRUA563aDpamN6pP .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-RRUA563aDpamN6pP rect.text{fill:none;stroke-width:0;}#mermaid-svg-RRUA563aDpamN6pP .icon-shape,#mermaid-svg-RRUA563aDpamN6pP .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-RRUA563aDpamN6pP .icon-shape p,#mermaid-svg-RRUA563aDpamN6pP .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-RRUA563aDpamN6pP .icon-shape .label rect,#mermaid-svg-RRUA563aDpamN6pP .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-RRUA563aDpamN6pP .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-RRUA563aDpamN6pP .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-RRUA563aDpamN6pP :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}

企业 AI 部署痛点

NVIDIA 算力成本高

显存瓶颈限制大模型

CUDA 生态锁定

AMD MI300X成本降低 50-60%

192GB HBM3 显存轻松部署 70B+ 模型

ROCm 开源生态避免厂商锁定

AMD ROCm 的核心优势:

但很多开发者担心：ROCm 真的成熟了吗？迁移成本高吗？

带着这些问题，我决定通过真实的云端实测来验证。

🛠️ 2. 环境准备：在 ModelScope 申请 AMD 实例

2.1 注册与登录

访问 魔搭创空间，使用阿里云账号或手机号注册。

2.2 创建 Notebook 实例

点击"创建实例"，选择以下配置：

💡 提示: 首次用户可申请 AMD 提供的免费算力券（通常 50-100 小时），足够完成本系列所有实验。

2.3 启动实例并连接

点击"启动"后，等待 2-3 分钟，实例状态变为"运行中"。点击"打开 JupyterLab"进入开发环境。

🔍 3. 验证 ROCm 环境

3.1 检查 ROCm 版本

打开 JupyterLab 的终端（Terminal），执行：

# 查看 ROCm 版本信息
rocminfo | grep "Version"

# 预期输出:
# Version: 6.2.0

3.2 查看 GPU 状态

# 显示 GPU 详细信息
rocm-smi –showproductname
rocm-smi –showmeminfo vram
rocm-smi –showuse

# 预期输出:
# ======================== ROCm System Management Interface ========================
# ================================= Product Info ==================================
# GPU[0] : Card series: AMD Instinct MI300X
# GPU[0] : Card model: 0x74a1
# GPU[0] : Card vendor: Advanced Micro Devices, Inc. [AMD/ATI]
#
# ================================== Memory Info ==================================
# GPU[0] VRAM Total Memory: 196608 MB
# GPU[0] VRAM Used Memory: 512 MB
#
# =================================== Utilization =================================
# GPU[0] GPU Use (%): 0

关键指标解读:

• ✅ GPU 型号: AMD Instinct MI300X
• ✅ 总显存: 196608 MB (192 GB)
• ✅ 已用显存: 512 MB (系统占用)
• ✅ GPU 利用率: 0% (空闲状态)

🚀 4. Hello World：运行第一个 PyTorch ROCm 程序

4.1 验证 PyTorch ROCm 支持

在 JupyterLab 中创建一个新的 Python Notebook，执行：

import torch

print("=" * 60)
print("PyTorch ROCm Environment Check")
print("=" * 60)

# 1. 检查 CUDA (ROCm) 是否可用
print(f"\\n✅ CUDA (ROCm) available: {torch.cuda.is_available()}")

# 2. 查看 GPU 数量
print(f"📊 Device count: {torch.cuda.device_count()}")

# 3. 查看 GPU 名称
if torch.cuda.is_available():
print(f"🎯 Current device: {torch.cuda.get_device_name(0)}")
print(f"💾 Total memory: {torch.cuda.get_device_properties(0).total_memory / 1e9:.2f} GB")
print(f"💾 Used memory: {torch.cuda.memory_allocated(0) / 1e9:.2f} GB")
else:
print("❌ ROCm not available! Please check your installation.")

print("\\n" + "=" * 60)

预期输出:

============================================================
PyTorch ROCm Environment Check
============================================================

✅ CUDA (ROCm) available: True
📊 Device count: 1
🎯 Current device: AMD Instinct MI300X
💾 Total memory: 196.61 GB
💾 Used memory: 0.00 GB

============================================================

💡 关键发现: PyTorch 中依然使用 torch.cuda 接口调用 AMD GPU，这是 ROCm 的兼容层设计，代码几乎无需修改！

4.2 矩阵乘法性能测试

接下来，我们运行一个标准的矩阵乘法 benchmark，对比 CPU 与 GPU 的性能差异：

import torch
import time

print("=" * 60)
print("Matrix Multiplication Benchmark")
print("=" * 60)

# 定义矩阵大小
matrix_size = 10000
num_iterations = 10

# CPU 测试
print(f"\\n🖥️ Testing on CPU ({matrix_size}x{matrix_size})…")
cpu_a = torch.randn(matrix_size, matrix_size)
cpu_b = torch.randn(matrix_size, matrix_size)

start_time = time.time()
for _ in range(num_iterations):
cpu_c = torch.matmul(cpu_a, cpu_b)
cpu_time = (time.time() – start_time) / num_iterations * 1000

print(f" Average time: {cpu_time:.2f} ms")

# GPU (ROCm) 测试
if torch.cuda.is_available():
device = torch.device("cuda:0")
print(f"\\n🎮 Testing on GPU ({matrix_size}x{matrix_size})…")

gpu_a = torch.randn(matrix_size, matrix_size, device=device)
gpu_b = torch.randn(matrix_size, matrix_size, device=device)

# 预热
for _ in range(5):
torch.matmul(gpu_a, gpu_b)
torch.cuda.synchronize()

# 正式测试
start_event = torch.cuda.Event(enable_timing=True)
end_event = torch.cuda.Event(enable_timing=True)

start_event.record()
for _ in range(num_iterations):
gpu_c = torch.matmul(gpu_a, gpu_b)
end_event.record()

torch.cuda.synchronize()
gpu_time = start_event.elapsed_time(end_event) / num_iterations

print(f" Average time: {gpu_time:.2f} ms")
print(f" Speedup: {cpu_time / gpu_time:.2f}x")

# 计算 TFLOPS
operations = 2 * matrix_size ** 3
tflops = (operations / gpu_time / 1e9) * 1000
print(f" Performance: {tflops:.2f} TFLOPS")

print("\\n" + "=" * 60)

实测结果（MI300X）:

============================================================
Matrix Multiplication Benchmark
============================================================

🖥️ Testing on CPU (10000×10000)…
Average time: 8523.45 ms

🎮 Testing on GPU (10000×10000)…
Average time: 125.67 ms
Speedup: 67.82x
Performance: 15.92 TFLOPS

============================================================

性能对比分析:

💡 说明: 简单矩阵乘法无法充分利用 MI300X 的计算单元，实际深度学习任务中效率可达 30-50%。

⚠️ 5. 常见问题与解决方案

Q1: torch.cuda.is_available() 返回 False

可能原因:

解决方案:

# 1. 确认安装了 ROCm 版 PyTorch
pip list | grep torch

# 应该看到:
# torch 2.3.0+rocm6.2

# 2. 如果安装的是 CPU 版本，重新安装:
pip uninstall torch torchvision torchaudio
pip install torch torchvision torchaudio –index-url https://download.pytorch.org/whl/rocm6.2

# 3. 检查用户组权限
groups $USER

# 确保包含 video 和 render 组
sudo usermod -aG video,render $USER

# 4. 重启 JupyterLab

Q2: rocm-smi 命令找不到

可能原因: ROCm 工具未加入 PATH

解决方案:

# 临时添加
export PATH=/opt/rocm/bin:$PATH

# 永久添加（写入 ~/.bashrc）
echo 'export PATH=/opt/rocm/bin:$PATH' >> ~/.bashrc
source ~/.bashrc

Q3: 显存占用异常高

可能原因: 之前的进程未释放显存

解决方案:

# Python 中清理显存
import torch
torch.cuda.empty_cache()

# 或者重启 Jupyter Kernel

📊 6. 阶段性总结

通过本次环境搭建，我确认了：

✅ ModelScope 的 AMD 实例开箱即用，ROCm 6.2 预装完善 ✅ PyTorch 对 ROCm 的支持非常成熟，torch.cuda 接口完全兼容 ✅ MI300X 性能强劲，矩阵乘法加速比达 67x ✅ 192GB 大显存优势明显，可轻松加载 70B+ 参数大模型

🔜 7. 下一篇预告

在《第二部分：CUDA 到 ROCm 迁移实战》中，我将：

👍 如果本文对你有帮助，欢迎点赞、收藏、转发！ 💬 如果你在 ROCm 环境搭建中遇到问题，请在评论区留言，我会逐一回复！ ✍️ 行文仓促，定有不足之处，欢迎各位朋友在评论区批评指正，不胜感激!

专栏导航:

• 📖 上一篇: –
• 📖 下一篇: CUDA 到 ROCm 迁移实战：YOLOv8(内容准备中)

参考资料:

• AMD ROCm 官方文档
• PyTorch for ROCm
• 魔搭创空间 AMD 实例