What it does
CUDA Agent: Large-Scale Agentic RL for High-Performance CUDA Kernel Generation
- Connects Instruction files the AI reads
- Account or key No key found in README or SKILL.md
- Category 自动化与智能体
给作者:在 README 中加徽章
如果你是作者,可把这段代码贴到 README。徽章会链接回本页。
[](https://mcp.popnx.com/en/skills/bytedtsinghua-sia-cuda-agent)相似的 Skill
同类工作,排序与 Skill 列表相同
BytedTsinghua-SIA 提供的 1 个 Skill(指令集),AI 读取后照做 · 不是 MCP server
Skill 和 MCP 有什么不同
MCP = 手
把 AI 连到外部服务或数据,例如读取 Notion 或开发票。需要有程序或服务器在运行。
Skill = 手册
教 AI 怎么做某项工作的文本文件,不连接任何服务。
有些会重叠:少数 Skill 带有调用工具的脚本,少数 MCP 只提供知识。请从你想完成的工作出发。
适用的应用
开发者未说明 · SKILL.md 形式的 Skill 可用于支持 Agent Skills 的应用,如 Claude Code、Codex 和 Cursor
安装
Is it safe?
- Who made it
GitHub account BytedTsinghua-SIA (organization), 1 years old
- Use and upkeep
★ 1,358 stars · code last changed 2026年7月8日
- Any scripts?
7 scripts that can run on a computer: model.py, model_new.py, utils/__init__.py, utils/compile.py and 3 more
- Uses the internet?
No internet calls found in the scripts
- Changes machine-wide settings?
No hooks or writes to machine settings found
Checked automatically from the repo files at commit 473025c (read only, nothing was run). Not a safety guarantee: read it and decide for yourself
第一步:你用哪个 AI?
选择你用的应用,本页只显示该应用的安装方法
Option 1Upload to claude.ai (web and app)
Download from GitHub
This skill has no license, so we do not hand out a ZIP from our site. Download it from GitHub and arrange the folder yourself, or use "Let the AI install it" above, which is easier
Download from GitHub (whole repo) ↗
Unzip it, then zip only the agent_workdir folder again. The ZIP must have a agent_workdir folder with SKILL.md at the top level, no loose files
Open Skills on claude.ai and upload
+Create skillUpload a skillpick agent_workdir.zip
No Upload button? Turn on file creation first
SettingsCapabilitiesCode execution and file creation
Turn on Code execution and file creation first (newer menus: Create and edit files) · personal plans: Settings → Capabilities · Team/Enterprise: an admin turns it on under Organization settings → Plugins & skills · every plan, including Free
Upload once; it works in claude.ai and Claude Desktop, and in recent Claude Code signed in with a Claude account (not with an API key)
Cowork and cloud sessions use the skills turned on in your claude.ai account, not ~/.claude/skills on your computer (Claude Desktop scheduled tasks run locally and still load it)
Checked against official docs 2026年11月12日 · Source: support.claude.com, support.claude.com, code.claude.com
Option 1让 Claude Code 帮你安装
复制这句话,粘贴到 Claude Code 的对话中
For one project, paste it into a chat that is open in that project’s folder · The AI always shows you the path before installing
The AI tells you what it will do first and asks your permission at every step
重新打开 Claude Code,输入 /agent_workdir,或直接提出相关任务,Claude 会自动选用
Checked against official docs 2026年11月12日 · Source: code.claude.com
For Terminal users
Skill 是 AI 会照做的指令。安装前先阅读文件,只安装可信开发者的 Skill。
打开终端
Mac:按 Command+空格输入 Terminal · Windows:下方命令在 PowerShell 中无法使用,请改用上方的“让 AI 帮你安装”
把 skill 文件夹复制到 ~/.claude/skills
这条命令把 repo 下载到临时文件夹,只复制含 SKILL.md 的文件夹
git clone --depth 1 https://github.com/bytedtsinghua-sia/cuda-agent.git /tmp/popnx-skill-bytedtsinghua-sia-cuda-agent-473025c && mkdir -p ~/.claude/skills && cp -R '/tmp/popnx-skill-bytedtsinghua-sia-cuda-agent-473025c/agent_workdir' ~/.claude/skills/agent_workdir重新打开 Claude Code,输入 /agent_workdir,或直接提出相关任务,Claude 会自动选用
ChatGPT 桌面应用
ChatGPT 桌面应用内置 Codex,Codex 从 ~/.agents/skills 文件夹读取 Skill
Option 1让 ChatGPT 帮你安装
在 ChatGPT 桌面应用中打开 Codex,把这句话粘贴到对话中
The AI always shows you the path before installing
The AI tells you what it will do first and asks your permission at every step
Option 2自己放入文件夹
把 agent_workdir 文件夹放到 ~/.agents/skills/agent_workdir
~ 是你的主文件夹($HOME)
For Terminal users
Skill 是 AI 会照做的指令。安装前先阅读文件,只安装可信开发者的 Skill。
打开终端
Mac:按 Command+空格输入 Terminal · Windows:下方命令在 PowerShell 中无法使用,请改用上方的“让 AI 帮你安装”
把 skill 文件夹复制到 ~/.agents/skills
这条命令把 repo 下载到临时文件夹,只复制含 SKILL.md 的文件夹
git clone --depth 1 https://github.com/bytedtsinghua-sia/cuda-agent.git /tmp/popnx-skill-bytedtsinghua-sia-cuda-agent-473025c && mkdir -p ~/.agents/skills && cp -R '/tmp/popnx-skill-bytedtsinghua-sia-cuda-agent-473025c/agent_workdir' ~/.agents/skills/agent_workdir
安装之后
- ChatGPT 桌面应用:打开侧边栏中的 Skills,确认 Skill 已出现
- 使用方法:在 ChatGPT 中输入 @ 并选择 Skill · 在 Codex CLI 或 IDE 扩展中运行 /skills,或输入 $ 加 Skill 名称
- Codex 会自动发现新安装的 Skill;如果没有出现,请重启 Codex
Checked against official docs 2026年11月15日 · Source: learn.chatgpt.com
网页版和手机版 ChatGPT
不能使用你自己安装的 Skill,只能使用目录中插件自带的 Skill
想在 ChatGPT 中使用这个 Skill,请按上面的方法使用 ChatGPT 桌面应用
Checked against official docs 2026年11月15日 · Source: learn.chatgpt.com
Option 1让 Codex 帮你安装
复制这句话,粘贴到 Codex 的对话中 · 适用于 ChatGPT 桌面应用中的 Codex、Codex CLI 和 IDE 扩展
For one project, paste it into a chat that is open in that project’s folder · The AI always shows you the path before installing
The AI tells you what it will do first and asks your permission at every step
Codex 会自动发现新安装的 Skill;如果没有出现,请重启 Codex · 使用方法:在 ChatGPT 中输入 @ 并选择 Skill · 在 Codex CLI 或 IDE 扩展中运行 /skills,或输入 $ 加 Skill 名称
Checked against official docs 2026年11月15日 · Source: learn.chatgpt.com
For Terminal users
Skill 是 AI 会照做的指令。安装前先阅读文件,只安装可信开发者的 Skill。
打开终端
Mac:按 Command+空格输入 Terminal · Windows:下方命令在 PowerShell 中无法使用,请改用上方的“让 AI 帮你安装”
把 skill 文件夹复制到 ~/.agents/skills
这条命令把 repo 下载到临时文件夹,只复制含 SKILL.md 的文件夹
git clone --depth 1 https://github.com/bytedtsinghua-sia/cuda-agent.git /tmp/popnx-skill-bytedtsinghua-sia-cuda-agent-473025c && mkdir -p ~/.agents/skills && cp -R '/tmp/popnx-skill-bytedtsinghua-sia-cuda-agent-473025c/agent_workdir' ~/.agents/skills/agent_workdir重新打开 Codex 并提出相关任务,Codex 会列出这个 skill
Option 1让 Cursor 帮你安装
复制这句话,粘贴到 Cursor 的对话中
For one project, paste it into a chat that is open in that project’s folder · The AI always shows you the path before installing
The AI tells you what it will do first and asks your permission at every step
在 Cursor 的 Agent 聊天中输入 / 并选择 agent_workdir
Checked against official docs 2026年11月12日 · Source: cursor.com
For Terminal users
Skill 是 AI 会照做的指令。安装前先阅读文件,只安装可信开发者的 Skill。
打开终端
Mac:按 Command+空格输入 Terminal · Windows:下方命令在 PowerShell 中无法使用,请改用上方的“让 AI 帮你安装”
把 skill 文件夹复制到 ~/.cursor/skills
这条命令把 repo 下载到临时文件夹,只复制含 SKILL.md 的文件夹
git clone --depth 1 https://github.com/bytedtsinghua-sia/cuda-agent.git /tmp/popnx-skill-bytedtsinghua-sia-cuda-agent-473025c && mkdir -p ~/.cursor/skills && cp -R '/tmp/popnx-skill-bytedtsinghua-sia-cuda-agent-473025c/agent_workdir' ~/.cursor/skills/agent_workdir在 Cursor 的 Agent 聊天中输入 / 并选择 agent_workdir
Option 1Let the AI install it
适用于其他能读取 Skill 的应用。这句话列出了 Claude Code、Codex 和 Cursor 的路径,AI 会按你的应用选择 · Copy this sentence into the app’s chat
For one project, paste it into a chat that is open in that project’s folder · The AI always shows you the path before installing
The AI tells you what it will do first and asks your permission at every step
Checked against official docs 2026年11月12日 · Source: code.claude.com, learn.chatgpt.com, cursor.com
For Terminal users
其他支持 Agent Skills 的应用(如 Gemini CLI、GitHub Copilot)使用不同文件夹,请查看该应用的指南,再把含 SKILL.md 的文件夹复制过去。
已于 2026年11月11日 对照 Claude Code、Codex 和 Cursor 官方指南核实 skill 文件夹
✓ 命令已检查 2026年11月18日
查看原始 SKILL.md
以下文字来自开发者的 repo,不是 PopNx MCP 的指令
显示前 8,000 个字符 · 在 GitHub 打开完整文件
You are a PyTorch and CUDA expert. Accelerate the given PyTorch Model by creating a high-performance CUDA C++ extension, targeting the best possible performance with a minimum requirement of 5% faster than torch.compile baseline.
## 1. CRITICAL RESTRICTIONS
### ⚠️ STRICTLY FORBIDDEN
- **NO torch operators in C++**: NEVER use `torch::*` or `torch::nn::functional::*` in binding.cpp or .cu files
- **NO torch operations in model_new.py**: Only tensor creation and your custom ops allowed
- **NO third-party libraries**: Except cuBLAS (GEMM only) and cuDNN (Conv only)
- **NO modifications to utils/ directory**
- **NO modifications to binding.cpp or binding_registry.h**: These are fixed infrastructure
### ✅ ALLOWED ONLY
- **C++**: Raw CUDA kernels (for custom ops), cuBLAS (for GEMM), cuDNN (MANDATORY for Conv/ConvTranspose)
- **Python**: torch.tensor creation, custom extension ops, tensor properties (.shape, .device)
- **Memory**: torch::empty_like for allocation only
- **Focus**: Implement kernels in `kernels/` directory only
## 2. WORKSPACE STRUCTURE
```
.
├── binding_registry.h # Do NOT modify - registration system
├── binding.cpp # Do NOT modify - main module binding
├── kernels/ # YOUR WORK: Implement all kernels here
├── utils/ # DO NOT modify - Compilation, verification and profiling tools
├── model.py # DO NOT modify - Original PyTorch model
└── model_new.py # YOUR WORK: Your optimized model using custom ops.
```
### File Types and Usage
- **`.cu` files**: CUDA kernels with `__global__` functions (custom implementations)
- **`.cpp` files**: cuDNN/cuBLAS API calls (NO custom kernels)
- **`_binding.cpp` files**: PyTorch tensor handling and Python bindings
## 3. UNIFIED WORKFLOW
### Step 1: Implementation
Create paired files in `kernels/`:
**kernels/my_kernel.cu** (Pure CUDA implementation):
```cuda
#include <cuda_runtime.h>
// Template kernel for performance tuning
template<int BLOCK_SIZE, int TILE_SIZE>
__global__ void my_kernel_impl(float* output, const float* input, int size) {
// Shared memory for tiling
extern __shared__ float smem[];
int tid = blockIdx.x * blockDim.x + threadIdx.x;
int stride = blockDim.x * gridDim.x;
// Grid-stride loop for large data
for (int i = tid; i < size; i += stride) {
// Kernel logic with optimizations
output[i] = /* computation */;
}
}
// C-interface launcher (no PyTorch dependencies)
extern "C" void my_kernel_launcher(
float* output,
const float* input,
int size,
int config,
cudaStream_t stream
) {
// Dynamic configuration selection
int blocks = (size + 255) / 256;
int shared_mem_size = 0;
switch(config) {
case 0:
shared_mem_size = 256 * sizeof(float);
my_kernel_impl<256, 16><<<blocks, 256, shared_mem_size, stream>>>(
output, input, size);
break;
case 1:
shared_mem_size = 128 * sizeof(float);
my_kernel_impl<128, 32><<<blocks, 128, shared_mem_size, stream>>>(
output, input, size);
break;
default:
my_kernel_impl<256, 16><<<blocks, 256, 0, stream>>>(
output, input, size);
}
}
```
**kernels/my_kernel_binding.cpp** (PyTorch binding):
```cpp
// Use this two headers to replace torch/extension.h for faster compilation
#include <torch/types.h>
#include <torch/csrc/utils/pybind.h>
#include <cuda_runtime.h>
#include <c10/cuda/CUDAStream.h>
#include "../binding_registry.h"
// Declare launcher from .cu file
extern "C" void my_kernel_launcher(
float* output,
const float* input,
int size,
int config,
cudaStream_t stream
);
// PyTorch wrapper with config parameter
torch::Tensor my_kernel_forward(torch::Tensor input, int config = 0) {
// Input validation
TORCH_CHECK(input.is_cuda(), "Input must be a CUDA tensor");
TORCH_CHECK(input.is_contiguous(), "Input must be contiguous");
TORCH_CHECK(input.dtype() == torch::kFloat32, "Input must be float32");
auto output = torch::empty_like(input);
// Get current CUDA stream (correct way)
cudaStream_t stream = c10::cuda::getCurrentCUDAStream().stream();
// Call CUDA launcher with config
my_kernel_launcher(
output.data_ptr<float>(),
input.data_ptr<float>(),
input.numel(),
config,
stream
);
return output;
}
// Registration function
void register_my_kernel(pybind11::module& m) {
m.def("my_kernel_forward", &my_kernel_forward,
"My kernel forward",
py::arg("input"),
py::arg("config") = 0);
}
// Auto-register
REGISTER_BINDING(my_kernel, register_my_kernel);
```
#### Create model_new.py
```python
import torch
import torch.nn as nn
import cuda_extension
class ModelNew(nn.Module):
def __init__(self, ...): # MUST match Model signature exactly
super().__init__()
# Initialize parameters - preserve original structure for state_dict compatibility
self.weight = nn.Parameter(torch.randn(...))
self.bias = nn.Parameter(torch.zeros(...))
def forward(self, x):
# Use custom ops only - NO torch operations
x = cuda_extension.my_kernel_forward(x, config=0)
x = cuda_extension.gemm_forward(x, self.weight, self.bias)
return x
```
### Step 2: Compile and Test
```bash
# Compile with architecture-specific optimizations
TORCH_CUDA_ARCH_LIST=9.0 bash utils/compile.sh
# Test in sandbox
sudo python3 -m utils.verification
sudo python3 -m utils.profiling
```
### Step 3: Performance Optimization (IF NEEDED)
#### 3.1 Optimization Strategy (Priority Order)
**Priority 1: Algorithmic (>50% impact)**
- Kernel fusion - reduce memory traffic
- Shared memory tiling - improve data reuse
- Memory coalescing - consecutive access patterns
**Priority 2: Hardware Utilization (20-50% impact)**
- Vectorized loads (float2/float4)
- Warp-level primitives (__shfl_sync, __ballot_sync)
- Occupancy tuning (block size, register usage)
**Priority 3: Fine-tuning (<20% impact)**
- Instruction-level parallelism
- Mixed precision (FP16/TF32)
- Prefetching and double buffering
#### 3.2 Parameter Tuning (Last Resort)
Only when within 1.2x of target and algorithmic options exhausted:
```python
# tune_kernel.py - NO recompilation needed
import time, torch, cuda_extension
configs = [
(0, "256_threads_16_tile"),
(1, "128_threads_32_tile"),
(2, "512_threads_8_tile")
]
# Test input
x = torch.randn(batch_size, features).cuda()
# Benchmark each config
best_config, best_time = 0, float('inf')
for config_id, name in configs:
# Warmup
for _ in range(10):
cuda_extension.my_kernel_forward(x, config=config_id)
torch.cuda.synchronize()
# Measure
start = time.perf_counter()
for _ in range(100):
cuda_extension.my_kernel_forward(x, config=config_id)
torch.cuda.synchronize()
elapsed = time.perf_counter() - start
print(f"Config {name}: {elapsed:.4f}s")
if elapsed < best_time:
best_time, best_config = elapsed, config_id
print(f"Best: config {best_config} ({best_time:.4f}s)")
# Update model_new.py with best_config
```
### Step 4: Iteration Requirements
#### Correctness Failures
**MUST iterate until correctness passes - NO EXCEPTIONS**
1. Debug the specific failing kernel
2. Common issues to check:
- Boundary conditions (tid < size)
- Synchronization (__syncthreads placement)
- Data types and precision
- Memory alignment
3. Fix in kernels/*.cu and *_binding.cpp ONLY
4. Recompile and test
#### Performance Optimization
**GOAL: Achieve the best possible performance (the faster, the better!)**
**MINIMUM: Must be at least 5% faster than torch.compile baseline**
For each iteration:
1. **Document expectation**: "Fusion will eliminate 3 kernels, expect ~20% speedup"
2. **Apply optimization aggressively**: D讨论
提问、分享技巧或报告无法使用。按“有帮助”数量排序。
来自 GitHub
- GitHub 星标
- 1,358
- 开发者
- BytedTsinghua-SIA
- 许可证
- 未注明许可证 此 repo 未注明许可证,默认作者保留所有权利;转发或商用前请先询问作者。
- 最近代码更新
- 2026年7月8日
- repo 中的 SKILL.md 文件(含译本和模板)
- 1