Instructions to use csssss/com2ai-klein-4b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Diffusers
How to use csssss/com2ai-klein-4b with Diffusers:
pip install -U diffusers transformers accelerate
import torch from diffusers import DiffusionPipeline from diffusers.utils import load_image # switch to "mps" for apple devices pipe = DiffusionPipeline.from_pretrained("csssss/com2ai-klein-4b", dtype=torch.bfloat16, device_map="cuda") prompt = "Turn this cat into a dog" input_image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/cat.png") image = pipe(image=input_image, prompt=prompt).images[0] - Notebooks
- Google Colab
- Kaggle
Update README.md
Browse files
README.md
CHANGED
|
@@ -1,41 +1,54 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: apache-2.0
|
| 3 |
---
|
| 4 |
-
|
| 5 |
|
| 6 |
-
模型说明
|
| 7 |
-
本仓库为整合版权重仓库,仅做文件归集、整合打包,无任何权重训练、微调、参数修改操作:
|
| 8 |
-
- GGUF量化Transformer主干:来源于 unsloth/FLUX.2-klein-4B-GGUF
|
| 9 |
-
采用 Unsloth Dynamic 2.0 量化方案,关键层保留高精度,极致平衡推理速度与生成画质。
|
| 10 |
-
- VAE、模型配置、原生配套资源:来源于官方基座 black-forest-labs/FLUX.2-klein-4B
|
| 11 |
-
-
|
| 12 |
-
原始来源链接
|
| 13 |
-
- GGUF量化主干模型:https://huggingface.co/unsloth/FLUX.2-klein-4B-GGUF
|
| 14 |
-
- FLUX.2-klein-4B 官方原始基座:https://huggingface.co/black-forest-labs/FLUX.2-klein-4B
|
| 15 |
-
-
|
| 16 |
-
修改说明
|
| 17 |
-
本项目仅将两个上游仓库资源整合归集:将 Unsloth 量化的 GGUF Transformer 权重,与官方完整 VAE、模型配置文件合并为一体,方便本地 GGUF 推理、ComfyUI 及 diffusers 一键加载。未修改权重参数、未微调、未重量化、未蒸馏。
|
| 18 |
|
|
|
|
| 19 |
|
|
|
|
|
|
|
| 20 |
|
|
|
|
| 21 |
|
| 22 |
-
|
| 23 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 24 |
|
| 25 |
-
安装依赖
|
| 26 |
-
pip install torch sdnq diffusers huggingface-hub
|
| 27 |
-
完整可运行代码
|
| 28 |
|
| 29 |
|
|
|
|
| 30 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 31 |
import torch
|
| 32 |
from sdnq import SDNQConfig
|
| 33 |
from diffusers import Flux2KleinPipeline, Flux2Transformer2DModel, GGUFQuantizationConfig
|
| 34 |
from huggingface_hub import hf_hub_download
|
| 35 |
|
|
|
|
| 36 |
device = "cuda"
|
| 37 |
dtype = torch.bfloat16
|
| 38 |
|
|
|
|
| 39 |
gguf_path = hf_hub_download(
|
| 40 |
repo_id="csssss/com2ai-klein-4b",
|
| 41 |
filename="transformer/flux-2-klein-4b-Q8_0.gguf",
|
|
@@ -46,6 +59,7 @@ config_path = hf_hub_download(
|
|
| 46 |
filename="transformer/config.json",
|
| 47 |
)
|
| 48 |
|
|
|
|
| 49 |
transformer = Flux2Transformer2DModel.from_single_file(
|
| 50 |
gguf_path,
|
| 51 |
quantization_config=GGUFQuantizationConfig(compute_dtype=dtype),
|
|
@@ -53,14 +67,17 @@ transformer = Flux2Transformer2DModel.from_single_file(
|
|
| 53 |
config=config_path
|
| 54 |
)
|
| 55 |
|
|
|
|
| 56 |
pipe = Flux2KleinPipeline.from_pretrained(
|
| 57 |
"csssss/com2ai-klein-4b",
|
| 58 |
transformer=transformer,
|
| 59 |
torch_dtype=dtype
|
| 60 |
)
|
| 61 |
|
|
|
|
| 62 |
pipe.enable_model_cpu_offload()
|
| 63 |
|
|
|
|
| 64 |
prompt = "A cat holding a sign that says hello world"
|
| 65 |
image = pipe(
|
| 66 |
prompt=prompt,
|
|
@@ -71,5 +88,8 @@ image = pipe(
|
|
| 71 |
generator=torch.Generator(device=device).manual_seed(0)
|
| 72 |
).images[0]
|
| 73 |
|
|
|
|
| 74 |
image.save("flux-klein.png")
|
| 75 |
print("✅ 图像生成成功,已保存为 flux-klein.png")
|
|
|
|
|
|
|
|
|
| 1 |
+
# FLUX\.2\-klein\-4B 整合GGUF权重包
|
| 2 |
+
|
| 3 |
+
|
| 4 |
+
```Plain Text
|
| 5 |
---
|
| 6 |
license: apache-2.0
|
| 7 |
---
|
| 8 |
+
```
|
| 9 |
|
| 10 |
+
## 模型说明
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 11 |
|
| 12 |
+
本仓库为**整合版权重仓库**,仅做文件归集、整合打包,无任何权重训练、微调、参数修改操作:
|
| 13 |
|
| 14 |
+
- **GGUF量化Transformer主干**:来源于 `unsloth/FLUX.2-klein-4B-GGUF`
|
| 15 |
+
采用 Unsloth Dynamic 2\.0 量化方案,关键层保留高精度,极致平衡推理速度与生成画质。
|
| 16 |
|
| 17 |
+
- **VAE、模型配置、原生配套资源**:来源于官方基座 `black-forest-labs/FLUX.2-klein-4B`
|
| 18 |
|
| 19 |
+
## 原始来源链接
|
| 20 |
|
| 21 |
+
- GGUF量化主干模型:[https://huggingface\.co/unsloth/FLUX\.2\-klein\-4B\-GGUF](https://huggingface.co/unsloth/FLUX.2-klein-4B-GGUF)
|
| 22 |
+
|
| 23 |
+
- FLUX\.2\-klein\-4B 官方原始基座:[https://huggingface\.co/black\-forest\-labs/FLUX\.2\-klein\-4B](https://huggingface.co/black-forest-labs/FLUX.2-klein-4B)
|
| 24 |
+
|
| 25 |
+
## 修改说明
|
| 26 |
+
|
| 27 |
+
本项目仅将两个上游仓库资源整合归集:将 Unsloth 量化的 GGUF Transformer 权重,与官方完整 VAE、模型配置文件合并为一体,方便本地diffusers GGUF 推理、。**未修改权重参数、未微调、未重量化、未蒸馏**。
|
| 28 |
|
|
|
|
|
|
|
|
|
|
| 29 |
|
| 30 |
|
| 31 |
+
## 快速使用 / 推理示例(官方标准写法)
|
| 32 |
|
| 33 |
+
### 安装依赖
|
| 34 |
+
|
| 35 |
+
```Plain Text
|
| 36 |
+
pip install torch sdnq diffusers huggingface-hub
|
| 37 |
+
```
|
| 38 |
+
|
| 39 |
+
### 完整可运行代码
|
| 40 |
+
|
| 41 |
+
```Plain Text
|
| 42 |
import torch
|
| 43 |
from sdnq import SDNQConfig
|
| 44 |
from diffusers import Flux2KleinPipeline, Flux2Transformer2DModel, GGUFQuantizationConfig
|
| 45 |
from huggingface_hub import hf_hub_download
|
| 46 |
|
| 47 |
+
# 设备与精度配置
|
| 48 |
device = "cuda"
|
| 49 |
dtype = torch.bfloat16
|
| 50 |
|
| 51 |
+
# 精准下载本仓库 GGUF 权重与配置文件
|
| 52 |
gguf_path = hf_hub_download(
|
| 53 |
repo_id="csssss/com2ai-klein-4b",
|
| 54 |
filename="transformer/flux-2-klein-4b-Q8_0.gguf",
|
|
|
|
| 59 |
filename="transformer/config.json",
|
| 60 |
)
|
| 61 |
|
| 62 |
+
# 加载量化 Transformer
|
| 63 |
transformer = Flux2Transformer2DModel.from_single_file(
|
| 64 |
gguf_path,
|
| 65 |
quantization_config=GGUFQuantizationConfig(compute_dtype=dtype),
|
|
|
|
| 67 |
config=config_path
|
| 68 |
)
|
| 69 |
|
| 70 |
+
# 加载完整 FLUX 推理管道(含官方VAE)
|
| 71 |
pipe = Flux2KleinPipeline.from_pretrained(
|
| 72 |
"csssss/com2ai-klein-4b",
|
| 73 |
transformer=transformer,
|
| 74 |
torch_dtype=dtype
|
| 75 |
)
|
| 76 |
|
| 77 |
+
# 显存优化
|
| 78 |
pipe.enable_model_cpu_offload()
|
| 79 |
|
| 80 |
+
# 文生图推理示例
|
| 81 |
prompt = "A cat holding a sign that says hello world"
|
| 82 |
image = pipe(
|
| 83 |
prompt=prompt,
|
|
|
|
| 88 |
generator=torch.Generator(device=device).manual_seed(0)
|
| 89 |
).images[0]
|
| 90 |
|
| 91 |
+
# 保存结果
|
| 92 |
image.save("flux-klein.png")
|
| 93 |
print("✅ 图像生成成功,已保存为 flux-klein.png")
|
| 94 |
+
```
|
| 95 |
+
|