Ryukijano commited on
Commit
b5a0af4
·
verified ·
1 Parent(s): 8b1e42e

Update custom_pipeline.py

Browse files
Files changed (1) hide show
  1. custom_pipeline.py +192 -0
custom_pipeline.py CHANGED
@@ -3,6 +3,7 @@ import numpy as np
3
  from diffusers import FluxPipeline, FlowMatchEulerDiscreteScheduler
4
  from typing import Any, Dict, List, Optional, Union
5
  from PIL import Image
 
6
 
7
  # Constants for shift calculation
8
  BASE_SEQ_LEN = 256
@@ -47,6 +48,169 @@ class FluxWithCFGPipeline(FluxPipeline):
47
  Extends the FluxPipeline to yield intermediate images during the denoising process
48
  with progressively increasing resolution for faster generation.
49
  """
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
50
  @torch.inference_mode()
51
  def generate_images(
52
  self,
@@ -71,6 +235,34 @@ class FluxWithCFGPipeline(FluxPipeline):
71
  height = height or self.default_sample_size * self.vae_scale_factor
72
  width = width or self.default_sample_size * self.vae_scale_factor
73
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
74
  # 1. Check inputs
75
  self.check_inputs(
76
  prompt,
 
3
  from diffusers import FluxPipeline, FlowMatchEulerDiscreteScheduler
4
  from typing import Any, Dict, List, Optional, Union
5
  from PIL import Image
6
+ from collections import OrderedDict
7
 
8
  # Constants for shift calculation
9
  BASE_SEQ_LEN = 256
 
48
  Extends the FluxPipeline to yield intermediate images during the denoising process
49
  with progressively increasing resolution for faster generation.
50
  """
51
+ def __init__(
52
+ self,
53
+ vae,
54
+ text_encoder,
55
+ text_encoder_2,
56
+ tokenizer,
57
+ tokenizer_2,
58
+ transformer,
59
+ scheduler: FlowMatchEulerDiscreteScheduler,
60
+ ):
61
+ super().__init__(vae, text_encoder, text_encoder_2, tokenizer, tokenizer_2, transformer, scheduler)
62
+ self.cuda_graphs = {}
63
+
64
+ def capture_cuda_graph(
65
+ self,
66
+ prompt: Union[str, List[str]] = None,
67
+ prompt_2: Optional[Union[str, List[str]]] = None,
68
+ height: Optional[int] = None,
69
+ width: Optional[int] = None,
70
+ num_inference_steps: int = 4,
71
+ guidance_scale: float = 3.5,
72
+ num_images_per_prompt: Optional[int] = 1,
73
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
74
+ latents: Optional[torch.FloatTensor] = None,
75
+ prompt_embeds: Optional[torch.FloatTensor] = None,
76
+ pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
77
+ output_type: Optional[str] = "pil",
78
+ return_dict: bool = True,
79
+ joint_attention_kwargs: Optional[Dict[str, Any]] = None,
80
+ max_sequence_length: int = 300,
81
+ **kwargs,
82
+ ):
83
+ """
84
+ Captures a static CUDA Graph for the generation process given static inputs.
85
+ """
86
+ # Use a static size for all inputs
87
+ static_height = height
88
+ static_width = width
89
+
90
+ # 1. Check inputs
91
+ self.check_inputs(
92
+ prompt,
93
+ prompt_2,
94
+ static_height,
95
+ static_width,
96
+ prompt_embeds=prompt_embeds,
97
+ pooled_prompt_embeds=pooled_prompt_embeds,
98
+ max_sequence_length=max_sequence_length,
99
+ )
100
+
101
+ self._guidance_scale = guidance_scale
102
+ self._joint_attention_kwargs = joint_attention_kwargs
103
+ self._interrupt = False
104
+
105
+ # 2. Define call parameters
106
+ batch_size = 1
107
+ device = self._execution_device
108
+
109
+ # 3. Encode prompt (with static inputs)
110
+ lora_scale = joint_attention_kwargs.get("scale", None) if joint_attention_kwargs is not None else None
111
+
112
+ # Use a static prompt for capture
113
+ static_prompt = "static prompt" if isinstance(prompt, str) else ["static prompt"]
114
+ prompt_embeds, pooled_prompt_embeds, text_ids = self.encode_prompt(
115
+ prompt=static_prompt,
116
+ prompt_2=prompt_2,
117
+ prompt_embeds=None,
118
+ pooled_prompt_embeds=None,
119
+ device=device,
120
+ num_images_per_prompt=num_images_per_prompt,
121
+ max_sequence_length=max_sequence_length,
122
+ lora_scale=lora_scale,
123
+ )
124
+
125
+ # 4. Prepare latent variables (with static inputs)
126
+ num_channels_latents = self.transformer.config.in_channels // 4
127
+ latents, latent_image_ids = self.prepare_latents(
128
+ batch_size * num_images_per_prompt,
129
+ num_channels_latents,
130
+ static_height,
131
+ static_width,
132
+ prompt_embeds.dtype,
133
+ device,
134
+ generator,
135
+ None,
136
+ )
137
+
138
+ # 5. Prepare timesteps (with static inputs)
139
+ sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps)
140
+ image_seq_len = latents.shape[1]
141
+ mu = calculate_timestep_shift(image_seq_len)
142
+ timesteps, num_inference_steps = prepare_timesteps(
143
+ self.scheduler,
144
+ num_inference_steps,
145
+ device,
146
+ None,
147
+ sigmas,
148
+ mu=mu,
149
+ )
150
+ self._num_timesteps = len(timesteps)
151
+
152
+ guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float16).expand(latents.shape[0]) if self.transformer.config.guidance_embeds else None
153
+
154
+ # Capture the graph
155
+ torch.cuda.synchronize()
156
+ stream = torch.cuda.Stream()
157
+ stream.wait_stream(torch.cuda.current_stream())
158
+ with torch.cuda.stream(stream):
159
+ for i, t in enumerate(timesteps):
160
+ timestep = t.expand(latents.shape[0]).to(latents.dtype)
161
+ noise_pred = self.transformer(
162
+ hidden_states=latents,
163
+ timestep=timestep / 1000,
164
+ guidance=guidance,
165
+ pooled_projections=pooled_prompt_embeds,
166
+ encoder_hidden_states=prompt_embeds,
167
+ txt_ids=text_ids,
168
+ img_ids=latent_image_ids,
169
+ joint_attention_kwargs=self.joint_attention_kwargs,
170
+ return_dict=False,
171
+ )[0]
172
+ latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
173
+
174
+ torch.cuda.current_stream().wait_stream(stream)
175
+ torch.cuda.synchronize()
176
+
177
+ # Capture the CUDA graph
178
+ graph = torch.cuda.CUDAGraph()
179
+ with torch.cuda.graph(graph, stream=stream):
180
+ # Create static inputs
181
+ static_inputs = OrderedDict()
182
+ static_inputs["hidden_states"] = latents.clone()
183
+ static_inputs["timestep"] = timesteps[0].expand(latents.shape[0]).to(latents.dtype)
184
+ static_inputs["guidance"] = guidance.clone() if guidance is not None else None
185
+ static_inputs["pooled_projections"] = pooled_prompt_embeds.clone()
186
+ static_inputs["encoder_hidden_states"] = prompt_embeds.clone()
187
+ static_inputs["txt_ids"] = text_ids
188
+ static_inputs["img_ids"] = latent_image_ids.clone()
189
+ static_inputs["joint_attention_kwargs"] = self.joint_attention_kwargs
190
+
191
+ # Run the static graph
192
+ for i, t in enumerate(timesteps):
193
+ timestep = static_inputs["timestep"].clone()
194
+ noise_pred = self.transformer(
195
+ hidden_states=static_inputs["hidden_states"],
196
+ timestep=timestep / 1000,
197
+ guidance=static_inputs["guidance"],
198
+ pooled_projections=static_inputs["pooled_projections"],
199
+ encoder_hidden_states=static_inputs["encoder_hidden_states"],
200
+ txt_ids=static_inputs["txt_ids"],
201
+ img_ids=static_inputs["img_ids"],
202
+ joint_attention_kwargs=static_inputs["joint_attention_kwargs"],
203
+ return_dict=False,
204
+ )[0]
205
+ static_inputs["hidden_states"] = self.scheduler.step(noise_pred, t, static_inputs["hidden_states"], return_dict=False)[0]
206
+
207
+ # Decode the latents after the loop
208
+ final_latents = static_inputs["hidden_states"]
209
+ final_image = self._decode_latents_to_image(final_latents, static_height, static_width, output_type)
210
+
211
+ # Store the graph and static inputs in the dictionary
212
+ self.cuda_graphs[(static_height, static_width, num_inference_steps)] = (graph, static_inputs, final_image)
213
+
214
  @torch.inference_mode()
215
  def generate_images(
216
  self,
 
235
  height = height or self.default_sample_size * self.vae_scale_factor
236
  width = width or self.default_sample_size * self.vae_scale_factor
237
 
238
+ # 0. Check if a CUDA graph can be used
239
+ if (height, width, num_inference_steps) in self.cuda_graphs:
240
+ graph, static_inputs, final_image = self.cuda_graphs[(height, width, num_inference_steps)]
241
+
242
+ # Update dynamic inputs (like prompt) in static_inputs
243
+ lora_scale = joint_attention_kwargs.get("scale", None) if joint_attention_kwargs is not None else None
244
+ prompt_embeds, pooled_prompt_embeds, text_ids = self.encode_prompt(
245
+ prompt=prompt,
246
+ prompt_2=prompt_2,
247
+ prompt_embeds=prompt_embeds,
248
+ pooled_prompt_embeds=pooled_prompt_embeds,
249
+ device=self._execution_device,
250
+ num_images_per_prompt=num_images_per_prompt,
251
+ max_sequence_length=max_sequence_length,
252
+ lora_scale=lora_scale,
253
+ )
254
+
255
+ # Update only the dynamic parts of static_inputs
256
+ static_inputs["pooled_projections"].copy_(pooled_prompt_embeds)
257
+ static_inputs["encoder_hidden_states"].copy_(prompt_embeds)
258
+ static_inputs["txt_ids"] = text_ids
259
+
260
+ # Replay the graph
261
+ graph.replay()
262
+ torch.cuda.empty_cache()
263
+
264
+ return final_image
265
+
266
  # 1. Check inputs
267
  self.check_inputs(
268
  prompt,