Add TeaCache4CogVideoX1.5

TeaCache4CogVideoX1.5/README.md

@@ -0,0 +1,49 @@
+<!-- ## **TeaCache4CogVideoX1.5** -->
+# TeaCache4CogVideoX1.5
+
+[TeaCache](https://github.com/LiewFeng/TeaCache) can speedup [CogVideoX1.5](https://github.com/THUDM/CogVideo) 1.8x without much visual quality degradation, in a training-free manner. The following video shows the results generated by TeaCache-ConsisID with various `rel_l1_thresh` values: 0 (original), 0.1 (1.3x speedup), 0.2 (1.8x speedup), and 0.3(2.1x speedup).
+
+https://github.com/user-attachments/assets/c444b850-3252-4b37-ad4a-122d389218d9
+
+## 📈 Inference Latency Comparisons on a Single H100 GPU
+
+| CogVideoX1.5 | TeaCache (0.1) | TeaCache (0.2) | TeaCache (0.3) |
+| :----------: | :------------: | :------------: | :------------: |
+|    ~465 s    |     ~372 s     |     ~261 s     |     ~223 s     |
+
+
+## Usage
+
+Follow [CogVideoX](https://github.com/THUDM/CogVideo) to clone the repo and finish the installation, then you can modify the `rel_l1_thresh` to obtain your desired trade-off between latency and visul quality, and change the `ckpts_path`, `prompt`, `image` to customize your identity-preserving video.
+
+For single-gpu inference, you can use the following command:
+
+```bash
+cd TeaCache4CogVideoX1.5
+
+python3 teacache_sample_video.py \
+    --rel_l1_thresh 0.2 \
+    --ckpts_path THUDM/CogVideoX1.5-5B \
+    --prompt "A clear, turquoise river flows through a rocky canyon, cascading over a small waterfall and forming a pool of water at the bottom.The river is the main focus of the scene, with its clear water reflecting the surrounding trees and rocks. The canyon walls are steep and rocky, with some vegetation growing on them. The trees are mostly pine trees, with their green needles contrasting with the brown and gray rocks. The overall tone of the scene is one of peace and tranquility.", help='Description of the video for the model to generate." \
+    --seed 42 \
+    --num_inference_steps 50 \
+    --output_path ./teacache_results
+```
+
+## Citation
+
+If you find TeaCache is useful in your research or applications, please consider giving us a star 🌟 and citing it by the following BibTeX entry.
+
+```
+@article{liu2024timestep,
+  title={Timestep Embedding Tells: It's Time to Cache for Video Diffusion Model},
+  author={Liu, Feng and Zhang, Shiwei and Wang, Xiaofeng and Wei, Yujie and Qiu, Haonan and Zhao, Yuzhong and Zhang, Yingya and Ye, Qixiang and Wan, Fang},
+  journal={arXiv preprint arXiv:2411.19108},
+  year={2024}
+}
+```
+
+
+## Acknowledgements
+
+We would like to thank the contributors to the [CogVideoX](https://github.com/THUDM/CogVideo) and [Diffusers](https://github.com/huggingface/diffusers).

TeaCache4CogVideoX1.5/teacache_smaple_video.py

@@ -0,0 +1,237 @@
+import argparse
+import torch
+import numpy as np
+from typing import Any, Dict, Optional, Tuple,  Union
+from videosys.models.transformers.cogvideox_transformer_3d import Transformer2DModelOutput
+from diffusers.utils import USE_PEFT_BACKEND, is_torch_version, scale_lora_layers, unscale_lora_layers, export_to_video
+from diffusers import CogVideoXPipeline
+
+
+def teacache_forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        timestep: Union[int, float, torch.LongTensor],
+        timestep_cond: Optional[torch.Tensor] = None,
+        ofs: Optional[Union[int, float, torch.LongTensor]] = None,
+        image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+        return_dict: bool = True,
+    ):
+        if attention_kwargs is not None:
+            attention_kwargs = attention_kwargs.copy()
+            lora_scale = attention_kwargs.pop("scale", 1.0)
+        else:
+            lora_scale = 1.0
+
+        if USE_PEFT_BACKEND:
+            # weight the lora layers by setting `lora_scale` for each PEFT layer
+            scale_lora_layers(self, lora_scale)
+        else:
+            if attention_kwargs is not None and attention_kwargs.get("scale", None) is not None:
+                logger.warning(
+                    "Passing `scale` via `attention_kwargs` when not using the PEFT backend is ineffective."
+                )
+
+        batch_size, num_frames, channels, height, width = hidden_states.shape
+
+        # 1. Time embedding
+        timesteps = timestep
+        t_emb = self.time_proj(timesteps)
+
+        # timesteps does not contain any weights and will always return f32 tensors
+        # but time_embedding might actually be running in fp16. so we need to cast here.
+        # there might be better ways to encapsulate this.
+        t_emb = t_emb.to(dtype=hidden_states.dtype)
+        emb = self.time_embedding(t_emb, timestep_cond)
+
+        if self.ofs_embedding is not None:
+            ofs_emb = self.ofs_proj(ofs)
+            ofs_emb = ofs_emb.to(dtype=hidden_states.dtype)
+            ofs_emb = self.ofs_embedding(ofs_emb)
+            emb = emb + ofs_emb
+
+        # 2. Patch embedding
+        hidden_states = self.patch_embed(encoder_hidden_states, hidden_states)
+        hidden_states = self.embedding_dropout(hidden_states)
+
+        text_seq_length = encoder_hidden_states.shape[1]
+        encoder_hidden_states = hidden_states[:, :text_seq_length]
+        hidden_states = hidden_states[:, text_seq_length:]
+
+        if self.enable_teacache:
+            if self.cnt == 0 or self.cnt == self.num_steps-1:
+                should_calc = True
+                self.accumulated_rel_l1_distance = 0
+            else: 
+                if not self.config.use_rotary_positional_embeddings:
+                    # CogVideoX-2B
+                    coefficients = [-3.10658903e+01,  2.54732368e+01, -5.92380459e+00,  1.75769064e+00, -3.61568434e-03]   
+                else:
+                    # CogVideoX-5B and CogvideoX1.5-5B
+                    coefficients = [-1.53880483e+03,  8.43202495e+02, -1.34363087e+02,  7.97131516e+00, -5.23162339e-02]
+                rescale_func = np.poly1d(coefficients)
+                self.accumulated_rel_l1_distance += rescale_func(((emb-self.previous_modulated_input).abs().mean() / self.previous_modulated_input.abs().mean()).cpu().item())
+                if self.accumulated_rel_l1_distance < self.rel_l1_thresh:
+                    should_calc = False
+                else:
+                    should_calc = True
+                    self.accumulated_rel_l1_distance = 0
+            self.previous_modulated_input = emb
+            self.cnt += 1
+            if self.cnt == self.num_steps:
+                self.cnt = 0            
+        
+        if self.enable_teacache:
+            if not should_calc:
+                hidden_states += self.previous_residual
+                encoder_hidden_states += self.previous_residual_encoder
+            else:
+                ori_hidden_states = hidden_states.clone()
+                ori_encoder_hidden_states = encoder_hidden_states.clone()
+                # 4. Transformer blocks
+                for i, block in enumerate(self.transformer_blocks):
+                    if torch.is_grad_enabled() and self.gradient_checkpointing:
+
+                        def create_custom_forward(module):
+                            def custom_forward(*inputs):
+                                return module(*inputs)
+
+                            return custom_forward
+
+                        ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                        hidden_states, encoder_hidden_states = torch.utils.checkpoint.checkpoint(
+                            create_custom_forward(block),
+                            hidden_states,
+                            encoder_hidden_states,
+                            emb,
+                            image_rotary_emb,
+                            **ckpt_kwargs,
+                        )
+                    else:
+                        hidden_states, encoder_hidden_states = block(
+                            hidden_states=hidden_states,
+                            encoder_hidden_states=encoder_hidden_states,
+                            temb=emb,
+                            image_rotary_emb=image_rotary_emb,
+                        )
+
+                self.previous_residual = hidden_states - ori_hidden_states
+                self.previous_residual_encoder = encoder_hidden_states - ori_encoder_hidden_states
+        else:
+            # 4. Transformer blocks
+            for i, block in enumerate(self.transformer_blocks):
+                if torch.is_grad_enabled() and self.gradient_checkpointing:
+
+                    def create_custom_forward(module):
+                        def custom_forward(*inputs):
+                            return module(*inputs)
+
+                        return custom_forward
+
+                    ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                    hidden_states, encoder_hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(block),
+                        hidden_states,
+                        encoder_hidden_states,
+                        emb,
+                        image_rotary_emb,
+                        **ckpt_kwargs,
+                    )
+                else:
+                    hidden_states, encoder_hidden_states = block(
+                        hidden_states=hidden_states,
+                        encoder_hidden_states=encoder_hidden_states,
+                        temb=emb,
+                        image_rotary_emb=image_rotary_emb,
+                    )
+
+        if not self.config.use_rotary_positional_embeddings:
+            # CogVideoX-2B
+            hidden_states = self.norm_final(hidden_states)
+        else:
+            # CogVideoX-5B and CogvideoX1.5-5B
+            hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
+            hidden_states = self.norm_final(hidden_states)
+            hidden_states = hidden_states[:, text_seq_length:]
+
+        # 5. Final block
+        hidden_states = self.norm_out(hidden_states, temb=emb)
+        hidden_states = self.proj_out(hidden_states)
+
+        # 6. Unpatchify
+        p = self.config.patch_size
+        p_t = self.config.patch_size_t
+
+        if p_t is None:
+            output = hidden_states.reshape(batch_size, num_frames, height // p, width // p, -1, p, p)
+            output = output.permute(0, 1, 4, 2, 5, 3, 6).flatten(5, 6).flatten(3, 4)
+        else:
+            output = hidden_states.reshape(
+                batch_size, (num_frames + p_t - 1) // p_t, height // p, width // p, -1, p_t, p, p
+            )
+            output = output.permute(0, 1, 5, 4, 2, 6, 3, 7).flatten(6, 7).flatten(4, 5).flatten(1, 2)
+
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
+        if not return_dict:
+            return (output,)
+        return Transformer2DModelOutput(sample=output)
+
+
+def main(args):
+    prompt = args.prompt
+    negative_prompt = args.negative_prompt
+    seed = args.seed
+    ckpts_path = args.ckpts_path
+    output_path = args.output_path
+    num_inference_steps = args.num_inference_steps
+    rel_l1_thresh = args.rel_l1_thresh
+    pipe = CogVideoXPipeline.from_pretrained(ckpts_path, torch_dtype=torch.bfloat16)
+
+    # TeaCache
+    pipe.transformer.__class__.enable_teacache = True
+    pipe.transformer.__class__.rel_l1_thresh = rel_l1_thresh
+    pipe.transformer.__class__.accumulated_rel_l1_distance = 0
+    pipe.transformer.__class__.previous_modulated_input = None
+    pipe.transformer.__class__.previous_residual = None
+    pipe.transformer.__class__.previous_residual_encoder = None
+    pipe.transformer.__class__.num_steps = num_inference_steps
+    pipe.transformer.__class__.cnt = 0
+    pipe.transformer.__class__.forward = teacache_forward
+
+    pipe.to("cuda")
+    pipe.vae.enable_slicing()
+    pipe.vae.enable_tiling()
+
+    video = pipe(
+        prompt=prompt,
+        negative_prompt=negative_prompt,
+        width=1360,
+        height=768,
+        num_frames=81,
+        use_dynamic_cfg=True,
+        guidance_scale=6,
+        num_inference_steps=num_inference_steps,
+        generator=torch.Generator("cuda").manual_seed(seed)
+    ).frames[0]
+    words = prompt.split()[:5]
+    video_path = f"{output_path}/teacache_cogvideox1.5-5B_{words}.mp4"
+    export_to_video(video, video_path, fps=16)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Run CogvideoX1.5-5B with given parameters")
+    
+    parser.add_argument('--seed', type=int, default=42, help='Random seed')
+    parser.add_argument('--num_inference_steps', type=int, default=50, help='Number of inference steps')
+    parser.add_argument("--output_path", type=str, default="./teacache_results", help="The path where the generated video will be saved")
+    parser.add_argument('--ckpts_path', type=str, default="/data-123/zishen/cvproject/CogVideo/THUDM/CogVideoX1.5-5B", help='Path to checkpoint')
+    parser.add_argument('--rel_l1_thresh', type=float, default=0.2, help='Higher speedup will cause to worse quality -- 0.1 for 1.3x speedup -- 0.2 for 1.8x speedup -- 0.3 for 2.1x speedup')
+    parser.add_argument('--prompt', type=str, default="A clear, turquoise river flows through a rocky canyon, cascading over a small waterfall and forming a pool of water at the bottom.The river is the main focus of the scene, with its clear water reflecting the surrounding trees and rocks. The canyon walls are steep and rocky, with some vegetation growing on them. The trees are mostly pine trees, with their green needles contrasting with the brown and gray rocks. The overall tone of the scene is one of peace and tranquility.", help='Description of the video for the model to generate')
+    parser.add_argument('--negative_prompt', type=str, default=None, help='Description of unwanted situations in model generated videos')
+    args = parser.parse_args()
+
+    main(args)