cuio commited on
Commit
4c6090a
·
verified ·
1 Parent(s): d796b43

Delete app.py

Browse files
Files changed (1) hide show
  1. app.py +0 -213
app.py DELETED
@@ -1,213 +0,0 @@
1
- """
2
- ===========================================
3
- FLUX.2 Klein-4B CPU Demo(仅技术验证用)
4
- ===========================================
5
- ⚠️ 重要说明:
6
- 1. 性能极差:256x256 1步约30-60秒,512x512 1步约2-5分钟,1024x1024大概率OOM
7
- 2. 内存要求:至少需要16G CPU内存(跑512需32G+),HF Space CPU Basic(16G)建议用256分辨率
8
- 3. 前置条件:已同意FLUX.2-dev的Gated License,Space配置了HF_TOKEN环境变量
9
- 4. 依赖已在你现有requirements.txt基础上兼容,无需额外修改
10
- """
11
-
12
- import torch
13
- import gradio as gr
14
- from PIL import Image
15
- from safetensors.torch import load_file
16
- from transformers import AutoModelForCausalLM, AutoTokenizer
17
- from flux2_adapter import Flux2AdapterModel, Flux2VAETokenizer
18
-
19
- # ==================== 全局配置 ====================
20
- RES = 256 # 👈 优先用256,512需要≥32G内存,1024不建议
21
- NUM_STEPS = 1 # 1步最快,增加步数会指数级变慢
22
- MODEL_PATH = "model.safetensors" # 和ae.safetensors同目录
23
- AE_PATH = "ae.safetensors"
24
-
25
- # ==================== 全局状态缓存 ====================
26
- _STATE = {} # 避免重复加载模型
27
-
28
- # ==================== CPU版Qwen3 Embedder(核心:禁用FP8) ====================
29
- def load_qwen3_embedder_cpu():
30
- """
31
- 手动加载Qwen3-4B BF16版本,完全规避FP8依赖
32
- CPU模式下只能用BF16/FP32,FP8不存在
33
- """
34
- if "qwen3" in _STATE:
35
- return _STATE["qwen3"]
36
-
37
- print("Loading Qwen3-4B (CPU/BF16)...")
38
- tokenizer = AutoTokenizer.from_pretrained(
39
- "Qwen/Qwen3-4B",
40
- padding_side="left",
41
- trust_remote_code=True
42
- )
43
- model = AutoModelForCausalLM.from_pretrained(
44
- "Qwen/Qwen3-4B",
45
- torch_dtype=torch.bfloat16, # CPU支持BF16,比FP32快30%
46
- device_map={"": "cpu"}, # 强制CPU
47
- low_cpu_mem_usage=True, # 降低加载时内存峰值
48
- trust_remote_code=True
49
- ).eval()
50
-
51
- _STATE["qwen3"] = (model, tokenizer)
52
- print("Qwen3-4B loaded (CPU)")
53
- return model, tokenizer
54
-
55
- # ==================== 模型加载(全CPU) ====================
56
- def ensure_loaded():
57
- if _STATE.get("model"):
58
- return
59
-
60
- print("Loading FLUX.2 Klein-4B (CPU)...")
61
-
62
- # 1. 加载主模型权重(CPU,lazy load 自动生效)
63
- sd = load_file(MODEL_PATH, device="cpu")
64
-
65
- model = Flux2AdapterModel(
66
- state_dict=sd,
67
- image_resolution=RES,
68
- param_dtype=torch.bfloat16
69
- ).to("cpu").eval()
70
-
71
- # 2. 加载 VAE
72
- print("Loading FLUX.2 AE (CPU)...")
73
- vae = Flux2VAETokenizer(
74
- AE_PATH,
75
- device="cpu",
76
- torch_dtype=torch.bfloat16
77
- )
78
-
79
- # 3. 加载 Qwen3 Embedder
80
- emb_model, emb_tokenizer = load_qwen3_embedder_cpu()
81
-
82
- _STATE.update({
83
- "model": model,
84
- "vae": vae,
85
- "emb_model": emb_model,
86
- "emb_tokenizer": emb_tokenizer
87
- })
88
-
89
- print(f"All models loaded (CPU), ready for {RES}x{RES} inference")
90
- # ==================== Prompt编码(适配FLUX.2的7680维要求) ====================
91
- def encode_prompt(prompt: str) -> torch.Tensor:
92
- """
93
- FLUX.2 Klein要求context维度为7680,来自3×Qwen3-4B的hidden state
94
- CPU模式下简化为重复3次最后一层的hidden state
95
- """
96
- model, tokenizer = _STATE["emb_model"], _STATE["emb_tokenizer"]
97
-
98
- inputs = tokenizer(
99
- prompt,
100
- return_tensors="pt",
101
- padding=True,
102
- truncation=True,
103
- max_length=256 # Klein的文本长度限制
104
- )
105
-
106
- with torch.no_grad(), torch.inference_mode():
107
- outputs = model(**inputs, output_hidden_states=True)
108
- # 用倒数第二层的hidden state(和训练时对齐性更好)
109
- hidden = outputs.hidden_states[-2][:, -1, :] # shape: (1, 2560)
110
-
111
- # 重复3次得到7680维,符合FLUX.2输入要求
112
- ctx = hidden.repeat(1, 3) # shape: (1, 7680)
113
- return ctx.to(torch.bfloat16)
114
-
115
- # ==================== Gradio推理逻辑 ====================
116
- def generate(prompt: str, seed: int, randomize: bool):
117
- prompt = (prompt or "").strip()
118
- if not prompt:
119
- raise gr.Error("请输入提示词")
120
-
121
- ensure_loaded()
122
-
123
- # 种子处理(纯CPU Generator)
124
- if randomize:
125
- seed = int(torch.randint(0, 2**31 - 1, (1,)).item())
126
- seed = int(seed)
127
- generator = torch.Generator().manual_seed(seed)
128
-
129
- # 生成随机噪声(CPU)
130
- noise = torch.randn(
131
- 1,
132
- 128,
133
- RES // 16,
134
- RES // 16,
135
- generator=generator,
136
- dtype=torch.bfloat16
137
- )
138
-
139
- # 编码提示词
140
- ctx = encode_prompt(prompt)
141
-
142
- # 推理(全程CPU,无GPU代码)
143
- start = torch.datetime.now()
144
- with torch.no_grad(), torch.inference_mode():
145
- latents = _STATE["model"].sample_images_with_grad(
146
- noise, ctx, {"num_steps": NUM_STEPS}
147
- )
148
- pixels = _STATE["vae"].detokenize(latents) # 输出[0,1]范围
149
- elapsed = (torch.datetime.now() - start).total_seconds()
150
-
151
- # 转换为PIL Image
152
- pixels = (
153
- pixels[0]
154
- .float()
155
- .clamp(0, 1)
156
- .mul(255)
157
- .round()
158
- .byte()
159
- .permute(1, 2, 0)
160
- .cpu()
161
- .numpy()
162
- )
163
- img = Image.fromarray(pixels)
164
-
165
- info = f"CPU · {RES}x{RES} · {NUM_STEPS}步 · {elapsed:.1f}s · Seed: {seed}"
166
- return img, seed, info
167
-
168
- # ==================== Gradio界面 ====================
169
- with gr.Blocks(title="FLUX.2 Klein CPU Demo") as demo:
170
- gr.Markdown(
171
- f"""
172
- # ⚠️ FLUX.2 Klein-4B CPU Demo(仅技术验证)
173
- - 当前分辨率:{RES}x{RES},步数:{NUM_STEPS}
174
- - CPU性能有限:{RES}x{RES} 1步约需30-180秒,请勿频繁提交
175
- - 如遇OOM:请将RES改为256,或升级Space硬件为GPU
176
- """
177
- )
178
-
179
- with gr.Row():
180
- with gr.Column():
181
- prompt = gr.Textbox(
182
- label="提示词",
183
- placeholder="例如:a cute cat sitting on a sofa",
184
- lines=2
185
- )
186
- with gr.Row():
187
- seed = gr.Number(label="种子", value=42, precision=0)
188
- randomize = gr.Checkbox(label="随机种子", value=False)
189
- btn = gr.Button("生成", variant="primary")
190
- with gr.Column():
191
- output_img = gr.Image(label="生成结果", type="pil")
192
- output_info = gr.Textbox(label="运行信息", interactive=False)
193
-
194
- btn.click(
195
- fn=generate,
196
- inputs=[prompt, seed, randomize],
197
- outputs=[output_img, seed, output_info]
198
- )
199
-
200
- # 预设提示词
201
- gr.Examples(
202
- examples=[
203
- ["a cute cat sleeping on a windowsill"],
204
- ["a futuristic city at night, neon lights"],
205
- ["a watercolor painting of a mountain landscape"]
206
- ],
207
- inputs=prompt
208
- )
209
-
210
- # ==================== 启动 ====================
211
- if __name__ == "__main__":
212
- # HF Space不需要share=True,本地运行可加share=True
213
- demo.launch(show_error=True)