Spaces:
Running
on
Zero
Running
on
Zero
Upload 2 files
Browse files- app.py +9 -17
- custom_pipeline.py +22 -50
app.py
CHANGED
@@ -8,10 +8,7 @@ from diffusers import DiffusionPipeline, AutoencoderTiny
|
|
8 |
from diffusers.models.attention_processor import AttnProcessor2_0
|
9 |
from custom_pipeline import FluxWithCFGPipeline
|
10 |
|
11 |
-
# Enable TF32 and set Tensor Core precision
|
12 |
torch.backends.cuda.matmul.allow_tf32 = True
|
13 |
-
torch.backends.cudnn.allow_tf32 = True
|
14 |
-
torch.set_float32_matmul_precision('high')
|
15 |
|
16 |
# Constants
|
17 |
MAX_SEED = np.iinfo(np.int32).max
|
@@ -32,10 +29,6 @@ pipe.set_adapters(["better"], adapter_weights=[1.0])
|
|
32 |
pipe.fuse_lora(adapter_name=["better"], lora_scale=1.0)
|
33 |
pipe.unload_lora_weights()
|
34 |
|
35 |
-
# Memory optimizations (optional, uncomment if needed)
|
36 |
-
# pipe.enable_model_cpu_offload()
|
37 |
-
# pipe.enable_sequential_cpu_offload()
|
38 |
-
|
39 |
torch.cuda.empty_cache()
|
40 |
|
41 |
# Inference function
|
@@ -47,15 +40,14 @@ def generate_image(prompt, seed=24, width=DEFAULT_WIDTH, height=DEFAULT_HEIGHT,
|
|
47 |
|
48 |
start_time = time.time()
|
49 |
|
50 |
-
|
51 |
-
|
52 |
-
|
53 |
-
|
54 |
-
|
55 |
-
|
56 |
-
|
57 |
-
|
58 |
-
)
|
59 |
latency = f"Latency: {(time.time()-start_time):.2f} seconds"
|
60 |
return img, seed, latency
|
61 |
|
@@ -171,4 +163,4 @@ with gr.Blocks() as demo:
|
|
171 |
)
|
172 |
|
173 |
# Launch the app
|
174 |
-
demo.launch()
|
|
|
8 |
from diffusers.models.attention_processor import AttnProcessor2_0
|
9 |
from custom_pipeline import FluxWithCFGPipeline
|
10 |
|
|
|
11 |
torch.backends.cuda.matmul.allow_tf32 = True
|
|
|
|
|
12 |
|
13 |
# Constants
|
14 |
MAX_SEED = np.iinfo(np.int32).max
|
|
|
29 |
pipe.fuse_lora(adapter_name=["better"], lora_scale=1.0)
|
30 |
pipe.unload_lora_weights()
|
31 |
|
|
|
|
|
|
|
|
|
32 |
torch.cuda.empty_cache()
|
33 |
|
34 |
# Inference function
|
|
|
40 |
|
41 |
start_time = time.time()
|
42 |
|
43 |
+
# Only generate the last image in the sequence
|
44 |
+
img = pipe.generate_images(
|
45 |
+
prompt=prompt,
|
46 |
+
width=width,
|
47 |
+
height=height,
|
48 |
+
num_inference_steps=num_inference_steps,
|
49 |
+
generator=generator
|
50 |
+
)
|
|
|
51 |
latency = f"Latency: {(time.time()-start_time):.2f} seconds"
|
52 |
return img, seed, latency
|
53 |
|
|
|
163 |
)
|
164 |
|
165 |
# Launch the app
|
166 |
+
demo.launch()
|
custom_pipeline.py
CHANGED
@@ -130,57 +130,29 @@ class FluxWithCFGPipeline(FluxPipeline):
|
|
130 |
|
131 |
# Handle guidance
|
132 |
guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float16).expand(latents.shape[0]) if self.transformer.config.guidance_embeds else None
|
133 |
-
|
134 |
-
# static method that can be jitted
|
135 |
-
@staticmethod
|
136 |
-
@torch.jit.script
|
137 |
-
def _denoising_loop_static(latents, timesteps, pooled_prompt_embeds, prompt_embeds, text_ids, latent_image_ids, guidance, transformer, scheduler):
|
138 |
-
for i, t in enumerate(timesteps):
|
139 |
-
timestep = t.expand(latents.shape[0]).to(latents.dtype)
|
140 |
-
|
141 |
-
noise_pred = transformer(
|
142 |
-
hidden_states=latents,
|
143 |
-
timestep=timestep / 1000,
|
144 |
-
guidance=guidance,
|
145 |
-
pooled_projections=pooled_prompt_embeds,
|
146 |
-
encoder_hidden_states=prompt_embeds,
|
147 |
-
txt_ids=text_ids,
|
148 |
-
img_ids=latent_image_ids,
|
149 |
-
return_dict=False,
|
150 |
-
)[0]
|
151 |
-
|
152 |
-
latents = scheduler.step(noise_pred, t, latents, return_dict=False)[0]
|
153 |
-
torch.cuda.empty_cache()
|
154 |
-
return latents
|
155 |
-
|
156 |
-
# Make the core denoising loop a static method
|
157 |
-
self._denoising_loop = torch.cuda.make_graphed_callables(
|
158 |
-
_denoising_loop_static,
|
159 |
-
(
|
160 |
-
latents.clone(), # Example inputs for warmup
|
161 |
-
timesteps.clone(),
|
162 |
-
pooled_prompt_embeds.clone(),
|
163 |
-
prompt_embeds.clone(),
|
164 |
-
text_ids.clone(),
|
165 |
-
latent_image_ids.clone(),
|
166 |
-
guidance.clone(),
|
167 |
-
self.transformer,
|
168 |
-
self.scheduler
|
169 |
-
)
|
170 |
-
)
|
171 |
|
172 |
-
#
|
173 |
-
|
174 |
-
|
175 |
-
|
176 |
-
|
177 |
-
|
178 |
-
|
179 |
-
|
180 |
-
|
181 |
-
|
182 |
-
|
183 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
184 |
|
185 |
# Final image
|
186 |
return self._decode_latents_to_image(latents, height, width, output_type)
|
|
|
130 |
|
131 |
# Handle guidance
|
132 |
guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float16).expand(latents.shape[0]) if self.transformer.config.guidance_embeds else None
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
133 |
|
134 |
+
# 6. Denoising loop
|
135 |
+
for i, t in enumerate(timesteps):
|
136 |
+
if self.interrupt:
|
137 |
+
continue
|
138 |
+
|
139 |
+
timestep = t.expand(latents.shape[0]).to(latents.dtype)
|
140 |
+
|
141 |
+
noise_pred = self.transformer(
|
142 |
+
hidden_states=latents,
|
143 |
+
timestep=timestep / 1000,
|
144 |
+
guidance=guidance,
|
145 |
+
pooled_projections=pooled_prompt_embeds,
|
146 |
+
encoder_hidden_states=prompt_embeds,
|
147 |
+
txt_ids=text_ids,
|
148 |
+
img_ids=latent_image_ids,
|
149 |
+
joint_attention_kwargs=self.joint_attention_kwargs,
|
150 |
+
return_dict=False,
|
151 |
+
)[0]
|
152 |
+
|
153 |
+
# Yield intermediate result
|
154 |
+
latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
|
155 |
+
torch.cuda.empty_cache()
|
156 |
|
157 |
# Final image
|
158 |
return self._decode_latents_to_image(latents, height, width, output_type)
|