Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -7,7 +7,7 @@ import time
|
|
7 |
from diffusers import DiffusionPipeline, AutoencoderTiny
|
8 |
from diffusers.models.attention_processor import AttnProcessor2_0
|
9 |
from custom_pipeline import FluxWithCFGPipeline
|
10 |
-
from huggingface_hub import
|
11 |
|
12 |
torch.backends.cuda.matmul.allow_tf32 = True
|
13 |
|
@@ -20,123 +20,51 @@ DEFAULT_INFERENCE_STEPS = 1
|
|
20 |
|
21 |
# Device and model setup
|
22 |
dtype = torch.float16
|
23 |
-
device = "cuda" # Explicitly set device to CUDA
|
24 |
-
|
25 |
-
# Download the LoRA weights
|
26 |
-
lora_weights_path = hf_hub_download(
|
27 |
-
repo_id="hugovntr/flux-schnell-realism",
|
28 |
-
filename="schnell-realism_v2.3.safetensors",
|
29 |
-
)
|
30 |
-
|
31 |
pipe = FluxWithCFGPipeline.from_pretrained(
|
32 |
-
"black-forest-labs/FLUX.1-schnell", torch_dtype=dtype
|
33 |
)
|
34 |
-
pipe.vae = AutoencoderTiny.from_pretrained("madebyollin/taef1", torch_dtype=dtype)
|
35 |
-
pipe.to(
|
36 |
-
|
37 |
-
# Load the LoRA weights
|
38 |
-
pipe.load_lora_weights(lora_weights_path, adapter_name="better")
|
39 |
pipe.set_adapters(["better"], adapter_weights=[1.0])
|
40 |
pipe.fuse_lora(adapter_name=["better"], lora_scale=1.0)
|
41 |
pipe.unload_lora_weights()
|
42 |
|
43 |
-
#
|
44 |
-
pipe.transformer.to(memory_format=torch.channels_last)
|
45 |
pipe.enable_xformers_memory_efficient_attention()
|
46 |
|
47 |
-
#
|
48 |
-
|
49 |
-
static_model = None
|
50 |
-
graph = None
|
51 |
-
|
52 |
-
def setup_cuda_graph(prompt, height, width, num_inference_steps):
|
53 |
-
global static_inputs, static_model, graph
|
54 |
-
|
55 |
-
batch_size = 1 if isinstance(prompt, str) else len(prompt)
|
56 |
-
num_images_per_prompt = 1
|
57 |
-
|
58 |
-
prompt_embeds, pooled_prompt_embeds, text_ids = pipe.encode_prompt(
|
59 |
-
prompt=prompt,
|
60 |
-
prompt_2=None,
|
61 |
-
prompt_embeds=None,
|
62 |
-
pooled_prompt_embeds=None,
|
63 |
-
device=device,
|
64 |
-
num_images_per_prompt=num_images_per_prompt,
|
65 |
-
max_sequence_length=300,
|
66 |
-
lora_scale=None,
|
67 |
-
)
|
68 |
-
|
69 |
-
latents, latent_image_ids = pipe.prepare_latents(
|
70 |
-
batch_size * num_images_per_prompt,
|
71 |
-
pipe.transformer.config.in_channels // 4,
|
72 |
-
height,
|
73 |
-
width,
|
74 |
-
prompt_embeds.dtype,
|
75 |
-
device,
|
76 |
-
None,
|
77 |
-
None,
|
78 |
-
)
|
79 |
-
sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps)
|
80 |
-
image_seq_len = latents.shape[1]
|
81 |
-
mu = calculate_timestep_shift(image_seq_len)
|
82 |
-
|
83 |
-
timesteps, num_inference_steps = prepare_timesteps(
|
84 |
-
pipe.scheduler,
|
85 |
-
num_inference_steps,
|
86 |
-
device,
|
87 |
-
None,
|
88 |
-
sigmas,
|
89 |
-
mu=mu,
|
90 |
-
)
|
91 |
-
|
92 |
-
guidance = torch.full([1], 3.5, device=device, dtype=torch.float16).expand(latents.shape[0]) if pipe.transformer.config.guidance_embeds else None
|
93 |
-
|
94 |
-
static_inputs = {
|
95 |
-
"hidden_states": latents.to(device),
|
96 |
-
"timestep": timesteps.to(device),
|
97 |
-
"guidance": guidance.to(device) if guidance is not None else None,
|
98 |
-
"pooled_projections": pooled_prompt_embeds.to(device),
|
99 |
-
"encoder_hidden_states": prompt_embeds.to(device),
|
100 |
-
"txt_ids": text_ids,
|
101 |
-
"img_ids": latent_image_ids,
|
102 |
-
"joint_attention_kwargs": None,
|
103 |
-
}
|
104 |
|
105 |
-
|
106 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
107 |
|
108 |
-
|
109 |
-
|
110 |
|
111 |
# Inference function
|
112 |
-
|
113 |
def generate_image(prompt, seed=24, width=DEFAULT_WIDTH, height=DEFAULT_HEIGHT, randomize_seed=False, num_inference_steps=2, progress=gr.Progress(track_tqdm=True)):
|
114 |
-
global static_inputs, graph
|
115 |
-
|
116 |
if randomize_seed:
|
117 |
seed = random.randint(0, MAX_SEED)
|
118 |
generator = torch.Generator().manual_seed(int(float(seed)))
|
119 |
|
120 |
start_time = time.time()
|
121 |
-
|
122 |
-
if static_inputs is None:
|
123 |
-
setup_cuda_graph(prompt, height, width, num_inference_steps)
|
124 |
-
|
125 |
-
static_inputs["hidden_states"].copy_(pipe.prepare_latents(
|
126 |
-
1,
|
127 |
-
pipe.transformer.config.in_channels // 4,
|
128 |
-
height,
|
129 |
-
width,
|
130 |
-
static_inputs["encoder_hidden_states"].dtype,
|
131 |
-
"cuda",
|
132 |
-
generator,
|
133 |
-
None,
|
134 |
-
)[0])
|
135 |
|
136 |
-
|
137 |
-
|
138 |
-
|
139 |
-
|
|
|
|
|
|
|
|
|
140 |
latency = f"Latency: {(time.time()-start_time):.2f} seconds"
|
141 |
return img, seed, latency
|
142 |
|
@@ -159,7 +87,7 @@ with gr.Blocks() as demo:
|
|
159 |
gr.Markdown("<span style='color: red;'>Note: Sometimes it stucks or stops generating images (I don't know why). In that situation just refresh the site.</span>")
|
160 |
|
161 |
with gr.Row():
|
162 |
-
with gr.Column(scale=2):
|
163 |
result = gr.Image(label="Generated Image", show_label=False, interactive=False)
|
164 |
with gr.Column(scale=1):
|
165 |
prompt = gr.Text(
|
@@ -192,8 +120,7 @@ with gr.Blocks() as demo:
|
|
192 |
fn=generate_image,
|
193 |
inputs=[prompt],
|
194 |
outputs=[result, seed, latency],
|
195 |
-
cache_examples=
|
196 |
-
cache_mode="lazy" # Added cache_mode
|
197 |
)
|
198 |
|
199 |
enhanceBtn.click(
|
@@ -253,4 +180,4 @@ with gr.Blocks() as demo:
|
|
253 |
)
|
254 |
|
255 |
# Launch the app
|
256 |
-
demo.
|
|
|
7 |
from diffusers import DiffusionPipeline, AutoencoderTiny
|
8 |
from diffusers.models.attention_processor import AttnProcessor2_0
|
9 |
from custom_pipeline import FluxWithCFGPipeline
|
10 |
+
from huggingface_hub import login
|
11 |
|
12 |
torch.backends.cuda.matmul.allow_tf32 = True
|
13 |
|
|
|
20 |
|
21 |
# Device and model setup
|
22 |
dtype = torch.float16
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
23 |
pipe = FluxWithCFGPipeline.from_pretrained(
|
24 |
+
"black-forest-labs/FLUX.1-schnell", torch_dtype=dtype, use_safetensors=True, variant="fp16"
|
25 |
)
|
26 |
+
pipe.vae = AutoencoderTiny.from_pretrained("madebyollin/taef1", torch_dtype=dtype, use_safetensors=True, variant="fp16")
|
27 |
+
pipe.to("cuda")
|
28 |
+
pipe.load_lora_weights('hugovntr/flux-schnell-realism', weight_name='schnell-realism_v2.3.safetensors', adapter_name="better")
|
|
|
|
|
29 |
pipe.set_adapters(["better"], adapter_weights=[1.0])
|
30 |
pipe.fuse_lora(adapter_name=["better"], lora_scale=1.0)
|
31 |
pipe.unload_lora_weights()
|
32 |
|
33 |
+
# Enable xformers
|
|
|
34 |
pipe.enable_xformers_memory_efficient_attention()
|
35 |
|
36 |
+
# Compile the model (Optional, needs further testing for stability)
|
37 |
+
# pipe.transformer = torch.compile(pipe.transformer, mode="reduce-overhead", fullgraph=True)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
38 |
|
39 |
+
# Capture CUDA Graph (Warm-up)
|
40 |
+
static_inputs = {
|
41 |
+
"prompt": "warmup",
|
42 |
+
"width": DEFAULT_WIDTH,
|
43 |
+
"height": DEFAULT_HEIGHT,
|
44 |
+
"num_inference_steps": DEFAULT_INFERENCE_STEPS,
|
45 |
+
"generator": torch.Generator().manual_seed(0),
|
46 |
+
}
|
47 |
|
48 |
+
pipe.capture_cuda_graph(**static_inputs)
|
49 |
+
torch.cuda.empty_cache()
|
50 |
|
51 |
# Inference function
|
52 |
+
@spaces.GPU(duration=25)
|
53 |
def generate_image(prompt, seed=24, width=DEFAULT_WIDTH, height=DEFAULT_HEIGHT, randomize_seed=False, num_inference_steps=2, progress=gr.Progress(track_tqdm=True)):
|
|
|
|
|
54 |
if randomize_seed:
|
55 |
seed = random.randint(0, MAX_SEED)
|
56 |
generator = torch.Generator().manual_seed(int(float(seed)))
|
57 |
|
58 |
start_time = time.time()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
59 |
|
60 |
+
# Only generate the last image in the sequence
|
61 |
+
img = pipe.generate_images(
|
62 |
+
prompt=prompt,
|
63 |
+
width=width,
|
64 |
+
height=height,
|
65 |
+
num_inference_steps=num_inference_steps,
|
66 |
+
generator=generator
|
67 |
+
)
|
68 |
latency = f"Latency: {(time.time()-start_time):.2f} seconds"
|
69 |
return img, seed, latency
|
70 |
|
|
|
87 |
gr.Markdown("<span style='color: red;'>Note: Sometimes it stucks or stops generating images (I don't know why). In that situation just refresh the site.</span>")
|
88 |
|
89 |
with gr.Row():
|
90 |
+
with gr.Column(scale=2.5):
|
91 |
result = gr.Image(label="Generated Image", show_label=False, interactive=False)
|
92 |
with gr.Column(scale=1):
|
93 |
prompt = gr.Text(
|
|
|
120 |
fn=generate_image,
|
121 |
inputs=[prompt],
|
122 |
outputs=[result, seed, latency],
|
123 |
+
cache_examples="lazy"
|
|
|
124 |
)
|
125 |
|
126 |
enhanceBtn.click(
|
|
|
180 |
)
|
181 |
|
182 |
# Launch the app
|
183 |
+
demo.launch()
|