cqy commited on
Commit
5e302d2
·
verified ·
1 Parent(s): 5f840c4
Files changed (1) hide show
  1. README.md +108 -25
README.md CHANGED
@@ -1,46 +1,129 @@
1
-
2
- # 🚀 RTX 50 系列显卡 AI 加速环境
3
 
4
  # 下载地址:
5
  ```
6
  https://github.com/chenxqiyu/ai_cuda_wheel
7
  ```
8
 
9
- > 适用于 Windows 平台 + ComfyUI + Nunchaku + KJNodes
10
- > 支持 CUDA 13.0、PyTorch 2.9.1xFormers、SageAttention、Triton Nunchaku INT4 推理引擎
11
- ```
12
- pip install --pre torch==2.9.1+cu130 torchvision==0.24.1+cu130 torchaudio==2.9.1+cu130 --index-url https://download.pytorch.org/whl/cu130
13
- ```
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
14
  ```
 
 
 
 
 
 
 
15
  pip install xformers-0.0.33+5d4b92a.d20260121-cp39-abi3-win_amd64.whl
16
  ```
17
- ```
 
 
18
  pip install triton_windows-3.6.0-cp311-cp311-win_amd64.whl
19
  ```
20
- ```
 
 
21
  pip install sageattention-2.2.0-cp311-cp311-win_amd64.whl
22
  ```
23
- ```
 
 
24
  pip install sageattn3-1.0.0-cp311-cp311-win_amd64.whl
25
  ```
26
- ```
 
 
27
  pip install nunchaku-1.2.0+torch2.9-cp311-cp311-win_amd64.whl
28
  ```
29
- ```
30
- https://github.com/nunchaku-ai/ComfyUI-nunchaku
31
- https://github.com/kijai/ComfyUI-KJNodes
32
- https://github.com/wallen0322/ComfyUI-SageAttention3
33
- ```
34
 
35
- ```
36
- 编译记录
 
 
 
 
 
 
 
37
 
38
- x64 Native Tools Command Prompt for VS 2022
 
39
 
40
- git修复和长路径支持
 
 
 
 
 
 
 
 
41
  git config --system core.longpaths true
42
- https://github.com/woct0rdho/triton-windows
43
- https://github.com/mengqin/SageAttention
44
- https://github.com/LagPixelLOL/xformers/tree/blackwell
45
- https://github.com/nunchaku-ai/nunchaku
46
- ```
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # 🚀 RTX 50 系列显卡 AI 加速环境(Windows)
 
2
 
3
  # 下载地址:
4
  ```
5
  https://github.com/chenxqiyu/ai_cuda_wheel
6
  ```
7
 
8
+ > **适用场景**:ComfyUI + Nunchaku + KJNodes
9
+ > **目标**:在 RTX 50 系列显卡上,构建稳定高性能的 AI 推理与加速环境 ,其他显卡需自行测试
10
+
11
+ ---
12
+
13
+ ## ✨ 环境特性
14
+
15
+ - ✅ **CUDA 13.0**
16
+ - ✅ **PyTorch 2.9.1(cu130)**
17
+ - ✅ **xFormers(Blackwell 分支)**
18
+ - ✅ **SageAttention2 / SageAttention3**
19
+ - ✅ **Triton(Windows 版)**
20
+ - ✅ **Nunchaku INT4 推理引擎**
21
+ - ✅ **ComfyUI 工作流友好**
22
+
23
+ ---
24
+
25
+ ## 🧩 基础环境
26
+
27
+ - **操作系统**:Windows 10 / 11 x64
28
+ - **Python**:3.11(推荐)
29
+ - **Visual Studio**:VS 2022(含 C++ Build Tools)
30
+ - **显卡**:RTX 50 系列
31
+
32
+ ---
33
+
34
+ ## 📦 PyTorch(CUDA 13.0)
35
+
36
+ ```bash
37
+ pip install --pre torch==2.9.1+cu130 torchvision==0.24.1+cu130 torchaudio==2.9.1+cu130 --index-url https://download.pytorch.org/whl/cu130
38
  ```
39
+
40
+ ---
41
+
42
+ ## ⚡ 加速组件安装
43
+
44
+ ### xFormers
45
+ ```bash
46
  pip install xformers-0.0.33+5d4b92a.d20260121-cp39-abi3-win_amd64.whl
47
  ```
48
+
49
+ ### Triton(Windows)
50
+ ```bash
51
  pip install triton_windows-3.6.0-cp311-cp311-win_amd64.whl
52
  ```
53
+
54
+ ### SageAttention
55
+ ```bash
56
  pip install sageattention-2.2.0-cp311-cp311-win_amd64.whl
57
  ```
58
+
59
+ ### SageAttention3
60
+ ```bash
61
  pip install sageattn3-1.0.0-cp311-cp311-win_amd64.whl
62
  ```
63
+
64
+ ### Nunchaku
65
+ ```bash
66
  pip install nunchaku-1.2.0+torch2.9-cp311-cp311-win_amd64.whl
67
  ```
 
 
 
 
 
68
 
69
+ ---
70
+
71
+ ## 🧠 ComfyUI 扩展节点
72
+
73
+ - **Nunchaku**
74
+ - https://github.com/nunchaku-ai/ComfyUI-nunchaku
75
+
76
+ - **KJNodes**
77
+ - https://github.com/kijai/ComfyUI-KJNodes
78
 
79
+ - **SageAttention3**
80
+ - https://github.com/wallen0322/ComfyUI-SageAttention3
81
 
82
+ ---
83
+
84
+ ## 🛠️ 编译与构建记录
85
+
86
+ ### 编译环境
87
+ - **工具**:x64 Native Tools Command Prompt for VS 2022
88
+
89
+ ### Git 长路径支持(必做)
90
+ ```bash
91
  git config --system core.longpaths true
92
+ ```
93
+
94
+ ### 相关源码仓库
95
+
96
+ - Triton Windows 适配
97
+ https://github.com/woct0rdho/triton-windows
98
+
99
+ - SageAttention
100
+ https://github.com/mengqin/SageAttention
101
+
102
+ - xFormers(Blackwell)
103
+ https://github.com/LagPixelLOL/xformers/tree/blackwell
104
+
105
+ - Nunchaku Core
106
+ https://github.com/nunchaku-ai/nunchaku
107
+
108
+ ---
109
+
110
+ ## 🧪 实践建议
111
+
112
+ - 🔹 优先验证 `torch.cuda.is_available()` 与 CUDA 版本一致性
113
+ - 🔹 遇到 `misaligned address`,优先检查 attention kernel / dtype / head_dim
114
+ - 🔹 SageAttention3 与 xFormers 不建议同时启用同一路径
115
+ - 🔹 INT4 推理建议搭配 **Nunchaku + FP16 输入**
116
+
117
+ ---
118
+
119
+ ## 📌 备注
120
+
121
+ 本环境主要面向 **RTX 50(Blackwell)** 架构实验与高性能推理,
122
+ 部分组件为 **非官方 / 实验性构建**,请自行评估稳定性。
123
+
124
+ ---
125
+ ## 使用sageattention3加速选其中一个就行
126
+ <img width="684" height="402" alt="image" src="https://github.com/user-attachments/assets/a6909182-577e-4dfe-b3ae-be55e607185f" />
127
+ <img width="482" height="247" alt="image" src="https://github.com/user-attachments/assets/a2348277-2062-4732-af8c-9ba25a906ee2" />
128
+
129
+ **Enjoy Blackwell 🚀**