Got wan 14b training to work on 24GB card.

toolkit/models/wan21.py

@@ -1,4 +1,5 @@
 # WIP, coming soon ish
+from functools import partial
 import torch
 import yaml
 from toolkit.accelerator import unwrap_model
@@ -34,6 +35,13 @@ from typing import TYPE_CHECKING, List
 from toolkit.accelerator import unwrap_model
 from toolkit.samplers.custom_flowmatch_sampler import CustomFlowMatchEulerDiscreteScheduler
 from torchvision.transforms import Resize, ToPILImage
+from tqdm import tqdm
+
+from diffusers.pipelines.wan.pipeline_output import WanPipelineOutput
+from diffusers.pipelines.wan.pipeline_wan import XLA_AVAILABLE
+# from ...callbacks import MultiPipelineCallbacks, PipelineCallback
+from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
+from typing import Any, Callable, Dict, List, Optional, Union
 
 # for generation only?
 scheduler_configUniPC = {
@@ -73,6 +81,199 @@ scheduler_config = {
 }
 
 
+class AggressiveWanUnloadPipeline(WanPipeline):
+    def __call__(
+        self: WanPipeline,
+        prompt: Union[str, List[str]] = None,
+        negative_prompt: Union[str, List[str]] = None,
+        height: int = 480,
+        width: int = 832,
+        num_frames: int = 81,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 5.0,
+        num_videos_per_prompt: Optional[int] = 1,
+        generator: Optional[Union[torch.Generator,
+                                  List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        output_type: Optional[str] = "np",
+        return_dict: bool = True,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+        callback_on_step_end: Optional[
+            Union[Callable[[int, int, Dict], None],
+                  PipelineCallback, MultiPipelineCallbacks]
+        ] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 512,
+    ):
+
+        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
+            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
+
+        # unload vae and transformer
+        vae_device = self.vae.device
+        transformer_device = self.transformer.device
+        text_encoder_device = self.text_encoder.device
+        print("Unloading vae")
+        self.vae.to("cpu")
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            negative_prompt,
+            height,
+            width,
+            prompt_embeds,
+            negative_prompt_embeds,
+            callback_on_step_end_tensor_inputs,
+        )
+
+        self._guidance_scale = guidance_scale
+        self._attention_kwargs = attention_kwargs
+        self._current_timestep = None
+        self._interrupt = False
+
+        device = self._execution_device
+
+        # 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]
+
+        # 3. Encode input prompt
+        prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+            prompt=prompt,
+            negative_prompt=negative_prompt,
+            do_classifier_free_guidance=self.do_classifier_free_guidance,
+            num_videos_per_prompt=num_videos_per_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            max_sequence_length=max_sequence_length,
+            device=device,
+        )
+
+        # unload text encoder
+        print("Unloading text encoder")
+        self.text_encoder.to("cpu")
+
+        transformer_dtype = self.transformer.dtype
+        prompt_embeds = prompt_embeds.to(transformer_dtype)
+        if negative_prompt_embeds is not None:
+            negative_prompt_embeds = negative_prompt_embeds.to(
+                transformer_dtype)
+
+        # 4. Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        timesteps = self.scheduler.timesteps
+
+        # 5. Prepare latent variables
+        num_channels_latents = self.transformer.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_videos_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            num_frames,
+            torch.float32,
+            device,
+            generator,
+            latents,
+        )
+
+        # 6. Denoising loop
+        num_warmup_steps = len(timesteps) - \
+            num_inference_steps * self.scheduler.order
+        self._num_timesteps = len(timesteps)
+
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                self._current_timestep = t
+                latent_model_input = latents.to(transformer_dtype)
+                timestep = t.expand(latents.shape[0])
+
+                noise_pred = self.transformer(
+                    hidden_states=latent_model_input,
+                    timestep=timestep,
+                    encoder_hidden_states=prompt_embeds,
+                    attention_kwargs=attention_kwargs,
+                    return_dict=False,
+                )[0]
+
+                if self.do_classifier_free_guidance:
+                    noise_uncond = self.transformer(
+                        hidden_states=latent_model_input,
+                        timestep=timestep,
+                        encoder_hidden_states=negative_prompt_embeds,
+                        attention_kwargs=attention_kwargs,
+                        return_dict=False,
+                    )[0]
+                    noise_pred = noise_uncond + guidance_scale * \
+                        (noise_pred - noise_uncond)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(
+                    noise_pred, t, latents, return_dict=False)[0]
+
+                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)
+
+                # 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()
+
+        self._current_timestep = None
+
+        # unload transformer
+        # load vae
+        print("Loading Vae")
+        self.vae.to(vae_device)
+
+        if not output_type == "latent":
+            latents = latents.to(self.vae.dtype)
+            latents_mean = (
+                torch.tensor(self.vae.config.latents_mean)
+                .view(1, self.vae.config.z_dim, 1, 1, 1)
+                .to(latents.device, latents.dtype)
+            )
+            latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to(
+                latents.device, latents.dtype
+            )
+            latents = latents / latents_std + latents_mean
+            video = self.vae.decode(latents, return_dict=False)[0]
+            video = self.video_processor.postprocess_video(
+                video, output_type=output_type)
+        else:
+            video = latents
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (video,)
+
+        return WanPipelineOutput(frames=video)
+
+
 class Wan21(BaseModel):
     def __init__(
             self,
@@ -118,6 +319,76 @@ class Wan21(BaseModel):
             if os.path.exists(te_folder_path):
                 base_model_path = model_path
 
+        self.print_and_status_update("Loading transformer")
+        transformer = WanTransformer3DModel.from_pretrained(
+            transformer_path,
+            subfolder=subfolder,
+            torch_dtype=dtype,
+        )
+
+        if self.model_config.split_model_over_gpus:
+            raise ValueError(
+                "Splitting model over gpus is not supported for Wan2.1 models")
+
+        if not self.model_config.low_vram:
+            # quantize on the device
+            transformer.to(self.quantize_device, dtype=dtype)
+            flush()
+
+        if self.model_config.assistant_lora_path is not None or self.model_config.inference_lora_path is not None:
+            raise ValueError(
+                "Assistant LoRA is not supported for Wan2.1 models currently")
+
+        if self.model_config.lora_path is not None:
+            raise ValueError(
+                "Loading LoRA is not supported for Wan2.1 models currently")
+
+        flush()
+
+        if self.model_config.quantize:
+            print("Quantizing Transformer")
+            quantization_args = self.model_config.quantize_kwargs
+            if 'exclude' not in quantization_args:
+                quantization_args['exclude'] = []
+            # patch the state dict method
+            patch_dequantization_on_save(transformer)
+            quantization_type = qfloat8
+            self.print_and_status_update("Quantizing transformer")
+            if self.model_config.low_vram:
+                print("Quantizing blocks")
+                orig_exclude = copy.deepcopy(quantization_args['exclude'])
+                # quantize each block
+                idx = 0
+                for block in tqdm(transformer.blocks):
+                    block.to(self.device_torch)
+                    quantize(block, weights=quantization_type,
+                             **quantization_args)
+                    freeze(block)
+                    idx += 1
+                    flush()
+
+                print("Quantizing the rest")
+                low_vram_exclude = copy.deepcopy(quantization_args['exclude'])
+                low_vram_exclude.append('blocks.*')
+                quantization_args['exclude'] = low_vram_exclude
+                # quantize the rest
+                transformer.to(self.device_torch)
+                quantize(transformer, weights=quantization_type,
+                         **quantization_args)
+
+                quantization_args['exclude'] = orig_exclude
+            else:
+                # do it in one go
+                quantize(transformer, weights=quantization_type,
+                         **quantization_args)
+            freeze(transformer)
+            # move it to the cpu for now
+            transformer.to("cpu")
+        else:
+            transformer.to(self.device_torch, dtype=dtype)
+
+        flush()
+
         self.print_and_status_update("Loading UMT5EncoderModel")
         tokenizer = AutoTokenizer.from_pretrained(
             base_model_path, subfolder="tokenizer", torch_dtype=dtype)
@@ -133,46 +404,10 @@ class Wan21(BaseModel):
             freeze(text_encoder)
             flush()
 
-        self.print_and_status_update("Loading transformer")
-        transformer = WanTransformer3DModel.from_pretrained(
-            transformer_path,
-            subfolder=subfolder,
-            torch_dtype=dtype,
-        )
-
-        if self.model_config.split_model_over_gpus:
-            raise ValueError(
-                "Splitting model over gpus is not supported for Wan2.1 models")
-
-        transformer.to(self.quantize_device, dtype=dtype)
-        flush()
-
-        if self.model_config.assistant_lora_path is not None or self.model_config.inference_lora_path is not None:
-            raise ValueError(
-                "Assistant LoRA is not supported for Wan2.1 models currently")
-
-        if self.model_config.lora_path is not None:
-            raise ValueError(
-                "Loading LoRA is not supported for Wan2.1 models currently")
-
-        flush()
-
-        if self.model_config.quantize:
-            quantization_args = self.model_config.quantize_kwargs
-            if 'exclude' not in quantization_args:
-                quantization_args['exclude'] = []
-            # patch the state dict method
-            patch_dequantization_on_save(transformer)
-            quantization_type = qfloat8
-            self.print_and_status_update("Quantizing transformer")
-            quantize(transformer, weights=quantization_type,
-                     **quantization_args)
-            freeze(transformer)
+        if self.model_config.low_vram:
+            print("Moving transformer back to GPU")
+            # we can move it back to the gpu now
             transformer.to(self.device_torch)
-        else:
-            transformer.to(self.device_torch, dtype=dtype)
-
-        flush()
 
         scheduler = Wan21.get_train_scheduler()
         self.print_and_status_update("Loading VAE")
@@ -213,13 +448,23 @@ class Wan21(BaseModel):
 
     def get_generation_pipeline(self):
         scheduler = UniPCMultistepScheduler(**scheduler_configUniPC)
-        pipeline = WanPipeline(
-            vae=self.vae,
-            transformer=self.unet,
-            text_encoder=self.text_encoder,
-            tokenizer=self.tokenizer,
-            scheduler=scheduler,
-        )
+        if self.model_config.low_vram:
+            pipeline = AggressiveWanUnloadPipeline(
+                vae=self.vae,
+                transformer=self.model,
+                text_encoder=self.text_encoder,
+                tokenizer=self.tokenizer,
+                scheduler=scheduler,
+            )
+        else:
+            pipeline = WanPipeline(
+                vae=self.vae,
+                transformer=self.unet,
+                text_encoder=self.text_encoder,
+                tokenizer=self.tokenizer,
+                scheduler=scheduler,
+            )
+
         return pipeline
 
     def generate_single_image(
@@ -231,6 +476,8 @@ class Wan21(BaseModel):
         generator: torch.Generator,
         extra: dict,
     ):
+        # reactivate progress bar since this is slooooow
+        pipeline.set_progress_bar_config(disable=False)
         # todo, figure out how to do video
         output = pipeline(
             prompt_embeds=conditional_embeds.text_embeds.to(
@@ -252,7 +499,7 @@ class Wan21(BaseModel):
         # shape = [1, frames, channels, height, width]
         batch_item = output[0]  # list of pil images
         if gen_config.num_frames > 1:
-            return batch_item # return the frames.
+            return batch_item  # return the frames.
         else:
             # get just the first image
             img = batch_item[0]
@@ -328,7 +575,7 @@ class Wan21(BaseModel):
         images = torch.stack(image_list)
         images = images.unsqueeze(2)
         latents = self.vae.encode(images).latent_dist.sample()
-        
+
         latents_mean = (
             torch.tensor(self.vae.config.latents_mean)
             .view(1, self.vae.config.z_dim, 1, 1, 1)
@@ -338,7 +585,7 @@ class Wan21(BaseModel):
             latents.device, latents.dtype
         )
         latents = (latents - latents_mean) * latents_std
-        
+
         latents = latents.to(device, dtype=dtype)
 
         return latents

toolkit/util/quantize.py

@@ -7,6 +7,7 @@ from optimum.quanto.tensor import Optimizer, qtype
 
 # the quantize function in quanto had a bug where it was using exclude instead of include
 
+Q_MODULES = ['QLinear', 'QConv2d', 'QEmbedding', 'QBatchNorm2d', 'QLayerNorm', 'QConvTranspose2d', 'QEmbeddingBag']
 
 def quantize(
     model: torch.nn.Module,
@@ -51,5 +52,13 @@ def quantize(
             continue
         if exclude is not None and any(fnmatch(name, pattern) for pattern in exclude):
             continue
-        _quantize_submodule(model, name, m, weights=weights,
-                            activations=activations, optimizer=optimizer)
+        try:
+            # check if m is QLinear or QConv2d
+            if m.__class__.__name__ in Q_MODULES:
+                continue
+            else:
+                _quantize_submodule(model, name, m, weights=weights,
+                                    activations=activations, optimizer=optimizer)
+        except Exception as e:
+            print(f"Failed to quantize {name}: {e}")
+            raise e