Update custom_pipeline.py

custom_pipeline.py

@@ -1,8 +1,16 @@
-import torch
 import numpy as np
-from diffusers import FluxPipeline, FlowMatchEulerDiscreteScheduler
-from typing import Any, Dict, List, Optional, Union
-from PIL import Image
+import torch
+from diffusers.pipelines.flux.pipeline_output import FluxPipeline, FluxPipelineOutput
+from typing import List, Union, Optional, Dict, Any, Callable
+from diffusers.pipelines.flux.pipeline_flux import calculate_shift, retrieve_timesteps
+from diffusers.utils import is_torch_xla_available
+
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
 
 # Constants for shift calculation
 BASE_SEQ_LEN = 256
@@ -19,7 +27,7 @@ def calculate_timestep_shift(image_seq_len: int) -> float:
     return mu
 
 def prepare_timesteps(
-    scheduler: FlowMatchEulerDiscreteScheduler,
+    scheduler,
     num_inference_steps: Optional[int] = None,
     device: Optional[Union[str, torch.device]] = None,
     timesteps: Optional[List[int]] = None,
@@ -41,20 +49,23 @@ def prepare_timesteps(
     num_inference_steps = len(timesteps)
     return timesteps, num_inference_steps
 
-# FLUX pipeline function
-class FLUXPipelineWithIntermediateOutputs(FluxPipeline):
+# FLUX pipeline with CFG and intermediate outputs
+class FluxWithCFGPipeline(FluxPipeline):
     """
-    Extends the FluxPipeline to yield intermediate images during the denoising process 
-    with progressively increasing resolution for faster generation.
+    Flux pipeline with Classifier-Free Guidance and the ability to yield 
+    intermediate images during the denoising process with progressively 
+    increasing resolution for faster generation.
     """
     @torch.inference_mode()
-    def generate_images(
+    def __call__(
         self,
         prompt: Union[str, List[str]] = None,
         prompt_2: Optional[Union[str, List[str]]] = None,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        negative_prompt_2: Optional[Union[str, List[str]]] = None,
         height: Optional[int] = None,
         width: Optional[int] = None,
-        num_inference_steps: int = 4,
+        num_inference_steps: int = 28,
         timesteps: List[int] = None,
         guidance_scale: float = 3.5,
         num_images_per_prompt: Optional[int] = 1,
@@ -62,16 +73,21 @@ class FLUXPipelineWithIntermediateOutputs(FluxPipeline):
         latents: Optional[torch.FloatTensor] = None,
         prompt_embeds: Optional[torch.FloatTensor] = None,
         pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
         output_type: Optional[str] = "pil",
         return_dict: bool = True,
         joint_attention_kwargs: Optional[Dict[str, Any]] = None,
-        max_sequence_length: int = 300,
+        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 512,
+        yield_intermediates: bool = False,  # New parameter for yielding intermediates
     ):
-        """Generates images and yields intermediate results during the denoising process."""
+
         height = height or self.default_sample_size * self.vae_scale_factor
         width = width or self.default_sample_size * self.vae_scale_factor
 
-        # 1. Check inputs
+        # 1. Check inputs. Raise error if not correct
         self.check_inputs(
             prompt,
             prompt_2,
@@ -79,6 +95,7 @@ class FLUXPipelineWithIntermediateOutputs(FluxPipeline):
             width,
             prompt_embeds=prompt_embeds,
             pooled_prompt_embeds=pooled_prompt_embeds,
+            callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
             max_sequence_length=max_sequence_length,
         )
 
@@ -87,12 +104,23 @@ class FLUXPipelineWithIntermediateOutputs(FluxPipeline):
         self._interrupt = False
 
         # 2. Define call parameters
-        batch_size = 1 if isinstance(prompt, str) else len(prompt)
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
         device = self._execution_device
 
-        # 3. Encode prompt
-        lora_scale = joint_attention_kwargs.get("scale", None) if joint_attention_kwargs is not None else None
-        prompt_embeds, pooled_prompt_embeds, text_ids = self.encode_prompt(
+        lora_scale = (
+            self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None
+        )
+        (
+            prompt_embeds,
+            pooled_prompt_embeds,
+            text_ids,
+        ) = self.encode_prompt(
             prompt=prompt,
             prompt_2=prompt_2,
             prompt_embeds=prompt_embeds,
@@ -102,6 +130,21 @@ class FLUXPipelineWithIntermediateOutputs(FluxPipeline):
             max_sequence_length=max_sequence_length,
             lora_scale=lora_scale,
         )
+        (
+            negative_prompt_embeds,
+            negative_pooled_prompt_embeds,
+            negative_text_ids,
+        ) = self.encode_prompt(
+            prompt=negative_prompt,
+            prompt_2=negative_prompt_2,
+            prompt_embeds=negative_prompt_embeds,
+            pooled_prompt_embeds=negative_pooled_prompt_embeds,
+            device=device,
+            num_images_per_prompt=num_images_per_prompt,
+            max_sequence_length=max_sequence_length,
+            lora_scale=lora_scale,
+        )
+
         # 4. Prepare latent variables
         num_channels_latents = self.transformer.config.in_channels // 4
         latents, latent_image_ids = self.prepare_latents(
@@ -114,6 +157,7 @@ class FLUXPipelineWithIntermediateOutputs(FluxPipeline):
             generator,
             latents,
         )
+
         # 5. Prepare timesteps
         sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps)
         image_seq_len = latents.shape[1]
@@ -126,43 +170,97 @@ class FLUXPipelineWithIntermediateOutputs(FluxPipeline):
             sigmas,
             mu=mu,
         )
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
         self._num_timesteps = len(timesteps)
 
-        # Handle guidance
-        guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float16).expand(latents.shape[0]) if self.transformer.config.guidance_embeds else None
-
         # 6. Denoising loop
-        for i, t in enumerate(timesteps):
-            if self.interrupt:
-                continue
-
-            timestep = t.expand(latents.shape[0]).to(latents.dtype)
-
-            noise_pred = self.transformer(
-                hidden_states=latents,
-                timestep=timestep / 1000,
-                guidance=guidance,
-                pooled_projections=pooled_prompt_embeds,
-                encoder_hidden_states=prompt_embeds,
-                txt_ids=text_ids,
-                img_ids=latent_image_ids,
-                joint_attention_kwargs=self.joint_attention_kwargs,
-                return_dict=False,
-            )[0]
-
-             # Yield intermediate result
-            latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
-            torch.cuda.empty_cache()
-
-        # Final image
-        return self._decode_latents_to_image(latents, height, width, output_type)
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                timestep = t.expand(latents.shape[0]).to(latents.dtype)
+
+                # handle guidance
+                if self.transformer.config.guidance_embeds:
+                    guidance = torch.tensor([guidance_scale], device=device)
+                    guidance = guidance.expand(latents.shape[0])
+                else:
+                    guidance = None
+
+                noise_pred_text = self.transformer(
+                    hidden_states=latents,
+                    timestep=timestep / 1000,
+                    guidance=guidance,
+                    pooled_projections=pooled_prompt_embeds,
+                    encoder_hidden_states=prompt_embeds,
+                    txt_ids=text_ids,
+                    img_ids=latent_image_ids,
+                    joint_attention_kwargs=self.joint_attention_kwargs,
+                    return_dict=False,
+                )[0]
+
+                noise_pred_uncond = self.transformer(
+                    hidden_states=latents,
+                    timestep=timestep / 1000,
+                    guidance=guidance,
+                    pooled_projections=negative_pooled_prompt_embeds,
+                    encoder_hidden_states=negative_prompt_embeds,
+                    txt_ids=negative_text_ids,
+                    img_ids=latent_image_ids,
+                    joint_attention_kwargs=self.joint_attention_kwargs,
+                    return_dict=False,
+                )[0]
+
+                noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents_dtype = latents.dtype
+                latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
+
+                if latents.dtype != latents_dtype:
+                    if torch.backends.mps.is_available():
+                        # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
+                        latents = latents.to(latents_dtype)
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+                    # Yield intermediate images if requested
+                    if yield_intermediates:
+                        yield self._decode_latents_to_image(latents, height, width, output_type)
+
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
+        # Final image decoding 
+        if output_type == "latent":
+            image = latents
+        else:
+            image = self._decode_latents_to_image(latents, height, width, output_type)
+
+        # Offload all models
         self.maybe_free_model_hooks()
-        torch.cuda.empty_cache()
 
-    def _decode_latents_to_image(self, latents, height, width, output_type, vae=None):
+        if not return_dict:
+            return (image,)
+
+        return FluxPipelineOutput(images=image)
+
+    def _decode_latents_to_image(self, latents, height, width, output_type):
         """Decodes the given latents into an image."""
-        vae = vae or self.vae
         latents = self._unpack_latents(latents, height, width, self.vae_scale_factor)
-        latents = (latents / vae.config.scaling_factor) + vae.config.shift_factor
-        image = vae.decode(latents, return_dict=False)[0]
+        latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor
+        image = self.vae.decode(latents, return_dict=False)[0]
         return self.image_processor.postprocess(image, output_type=output_type)[0]