Create inference.py

inference.py

@@ -0,0 +1,322 @@
+import inspect
+import torch
+
+from typing import Any, Callable, Dict, List, Optional, Union
+from diffusers.image_processor import PipelineImageInput, VaeImageProcessor
+from diffusers.utils import (
+    USE_PEFT_BACKEND,
+    is_torch_xla_available,
+    logging,
+    replace_example_docstring,
+    scale_lora_layers,
+    unscale_lora_layers,
+)
+from diffusers.pipelines.stable_diffusion_3.pipeline_output import StableDiffusion3PipelineOutput
+
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
+
+def retrieve_timesteps(
+        scheduler,
+        num_inference_steps: Optional[int] = None,
+        device: Optional[Union[str, torch.device]] = None,
+        timesteps: Optional[List[int]] = None,
+        sigmas: Optional[List[float]] = None,
+        **kwargs,
+):
+    if timesteps is not None and sigmas is not None:
+        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    elif sigmas is not None:
+        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accept_sigmas:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" sigmas schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+
+    return timesteps, num_inference_steps
+
+
+@torch.no_grad()
+def run(
+        self,
+        prompt: Union[str, List[str]] = None,
+        prompt_2: Optional[Union[str, List[str]]] = None,
+        prompt_3: Optional[Union[str, List[str]]] = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_inference_steps: int = 28,
+        sigmas: Optional[List[float]] = None,
+        scales: List[float] = None,
+        guidance_scale: float = 7.0,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        negative_prompt_2: Optional[Union[str, List[str]]] = None,
+        negative_prompt_3: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        ip_adapter_image: Optional[PipelineImageInput] = None,
+        ip_adapter_image_embeds: Optional[torch.Tensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+        clip_skip: Optional[int] = None,
+        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 256,
+        skip_guidance_layers: List[int] = None,
+        skip_layer_guidance_scale: float = 2.8,
+        skip_layer_guidance_stop: float = 0.2,
+        skip_layer_guidance_start: float = 0.01,
+        mu: Optional[float] = None,
+):
+    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. Raise error if not correct
+    self.check_inputs(
+        prompt,
+        prompt_2,
+        prompt_3,
+        height,
+        width,
+        negative_prompt=negative_prompt,
+        negative_prompt_2=negative_prompt_2,
+        negative_prompt_3=negative_prompt_3,
+        prompt_embeds=prompt_embeds,
+        negative_prompt_embeds=negative_prompt_embeds,
+        pooled_prompt_embeds=pooled_prompt_embeds,
+        negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+        callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
+        max_sequence_length=max_sequence_length,
+    )
+
+    self._guidance_scale = guidance_scale
+    self._skip_layer_guidance_scale = skip_layer_guidance_scale
+    self._clip_skip = clip_skip
+    self._joint_attention_kwargs = joint_attention_kwargs
+    self._interrupt = False
+
+    # 2. Define call parameters
+    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
+
+    lora_scale = (
+        self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None
+    )
+    (
+        prompt_embeds,
+        negative_prompt_embeds,
+        pooled_prompt_embeds,
+        negative_pooled_prompt_embeds,
+    ) = self.encode_prompt(
+        prompt=prompt,
+        prompt_2=prompt_2,
+        prompt_3=prompt_3,
+        negative_prompt=negative_prompt,
+        negative_prompt_2=negative_prompt_2,
+        negative_prompt_3=negative_prompt_3,
+        do_classifier_free_guidance=self.do_classifier_free_guidance,
+        prompt_embeds=prompt_embeds,
+        negative_prompt_embeds=negative_prompt_embeds,
+        pooled_prompt_embeds=pooled_prompt_embeds,
+        negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+        device=device,
+        clip_skip=self.clip_skip,
+        num_images_per_prompt=num_images_per_prompt,
+        max_sequence_length=max_sequence_length,
+        lora_scale=lora_scale,
+    )
+
+    if self.do_classifier_free_guidance:
+        if skip_guidance_layers is not None:
+            original_prompt_embeds = prompt_embeds
+            original_pooled_prompt_embeds = pooled_prompt_embeds
+        prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+        pooled_prompt_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds], dim=0)
+
+    # 4. Prepare latent variables
+    num_channels_latents = self.transformer.config.in_channels
+    latents = self.prepare_latents(
+        batch_size * num_images_per_prompt,
+        num_channels_latents,
+        height,
+        width,
+        prompt_embeds.dtype,
+        device,
+        generator,
+        latents,
+    )
+
+    # 5. Prepare timesteps
+    scheduler_kwargs = {}
+    if self.scheduler.config.get("use_dynamic_shifting", None) and mu is None:
+        _, _, height, width = latents.shape
+        image_seq_len = (height // self.transformer.config.patch_size) * (
+                width // self.transformer.config.patch_size
+        )
+        mu = calculate_shift(
+            image_seq_len,
+            self.scheduler.config.get("base_image_seq_len", 256),
+            self.scheduler.config.get("max_image_seq_len", 4096),
+            self.scheduler.config.get("base_shift", 0.5),
+            self.scheduler.config.get("max_shift", 1.16),
+        )
+        scheduler_kwargs["mu"] = mu
+    elif mu is not None:
+        scheduler_kwargs["mu"] = mu
+    timesteps, num_inference_steps = retrieve_timesteps(
+        self.scheduler,
+        num_inference_steps,
+        device,
+        sigmas=sigmas,
+        **scheduler_kwargs,
+    )
+    num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+    self._num_timesteps = len(timesteps)
+
+    # 6. Prepare image embeddings
+    if (ip_adapter_image is not None and self.is_ip_adapter_active) or ip_adapter_image_embeds is not None:
+        ip_adapter_image_embeds = self.prepare_ip_adapter_image_embeds(
+            ip_adapter_image,
+            ip_adapter_image_embeds,
+            device,
+            batch_size * num_images_per_prompt,
+            self.do_classifier_free_guidance,
+        )
+
+        if self.joint_attention_kwargs is None:
+            self._joint_attention_kwargs = {"ip_adapter_image_embeds": ip_adapter_image_embeds}
+        else:
+            self._joint_attention_kwargs.update(ip_adapter_image_embeds=ip_adapter_image_embeds)
+
+    # 7. Denoising loop
+    with self.progress_bar(total=num_inference_steps) as progress_bar:
+        for i, t in enumerate(timesteps):
+            if self.interrupt:
+                continue
+
+            # expand the latents if we are doing classifier free guidance
+            latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+            # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+            timestep = t.expand(latent_model_input.shape[0])
+
+            noise_pred = self.transformer(
+                hidden_states=latent_model_input,
+                timestep=timestep,
+                encoder_hidden_states=prompt_embeds,
+                pooled_projections=pooled_prompt_embeds,
+                joint_attention_kwargs=self.joint_attention_kwargs,
+                return_dict=False,
+            )[0]
+
+            # perform guidance
+            if self.do_classifier_free_guidance:
+                noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+                should_skip_layers = (
+                    True
+                    if i > num_inference_steps * skip_layer_guidance_start
+                       and i < num_inference_steps * skip_layer_guidance_stop
+                    else False
+                )
+                if skip_guidance_layers is not None and should_skip_layers:
+                    timestep = t.expand(latents.shape[0])
+                    latent_model_input = latents
+                    noise_pred_skip_layers = self.transformer(
+                        hidden_states=latent_model_input,
+                        timestep=timestep,
+                        encoder_hidden_states=original_prompt_embeds,
+                        pooled_projections=original_pooled_prompt_embeds,
+                        joint_attention_kwargs=self.joint_attention_kwargs,
+                        return_dict=False,
+                        skip_layers=skip_guidance_layers,
+                    )[0]
+                    noise_pred = (
+                            noise_pred + (noise_pred_text - noise_pred_skip_layers) * self._skip_layer_guidance_scale
+                    )
+
+            # compute the previous noisy sample x_t -> x_t-1
+            latents_dtype = latents.dtype
+            sigma = self.scheduler.sigmas[i]
+            sigma_next = self.scheduler.sigmas[i + 1]
+            x0_pred = (latents - sigma * noise_pred)
+            try:
+                x0_pred = torch.nn.functional.interpolate(x0_pred, size=scales[i + 1])
+            except IndexError:
+                x0_pred = x0_pred
+            noise = torch.randn(x0_pred.shape, generator=generator).to('cuda').half()
+            latents = (1 - sigma_next) * x0_pred + sigma_next * noise
+
+            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)
+                negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+                negative_pooled_prompt_embeds = callback_outputs.pop(
+                    "negative_pooled_prompt_embeds", negative_pooled_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()
+
+            if XLA_AVAILABLE:
+                xm.mark_step()
+
+    if output_type == "latent":
+        image = latents
+
+    else:
+        latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor
+
+        image = self.vae.decode(latents, return_dict=False)[0]
+        image = self.image_processor.postprocess(image, output_type=output_type)
+
+    # Offload all models
+    self.maybe_free_model_hooks()
+
+    if not return_dict:
+        return (image,)
+
+    return StableDiffusion3PipelineOutput(images=image)