support stepvideo quantized

README.md

@@ -17,6 +17,7 @@ DiffSynth Studio is a Diffusion engine. We have restructured architectures inclu
 
 Until now, DiffSynth Studio has supported the following models:
 
+* [StepVideo](https://github.com/stepfun-ai/Step-Video-T2V)
 * [HunyuanVideo](https://github.com/Tencent/HunyuanVideo)
 * [CogVideoX](https://huggingface.co/THUDM/CogVideoX-5b)
 * [FLUX](https://huggingface.co/black-forest-labs/FLUX.1-dev)
@@ -34,6 +35,9 @@ Until now, DiffSynth Studio has supported the following models:
 * [Stable Diffusion](https://huggingface.co/runwayml/stable-diffusion-v1-5)
 
 ## News
+
+- **February 17, 2024** We support [StepVideo](https://modelscope.cn/models/stepfun-ai/stepvideo-t2v/summary)! State-of-the-art video synthesis model! See [./examples/stepvideo](./examples/stepvideo/).
+
 - **December 31, 2024** We propose EliGen, a novel framework for precise entity-level controlled text-to-image generation, complemented by an inpainting fusion pipeline to extend its capabilities to image inpainting tasks. EliGen seamlessly integrates with existing community models, such as IP-Adapter and In-Context LoRA, enhancing its versatility. For more details, see [./examples/EntityControl](./examples/EntityControl/).
   - Paper: [EliGen: Entity-Level Controlled Image Generation with Regional Attention](https://arxiv.org/abs/2501.01097)
   - Model: [ModelScope](https://www.modelscope.cn/models/DiffSynth-Studio/Eligen)

diffsynth/models/stepvideo_dit.py

@@ -238,7 +238,7 @@ class PixArtAlphaCombinedTimestepSizeEmbeddings(nn.Module):
             self.fps_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim)
 
     def forward(self, timestep, resolution=None, nframe=None, fps=None):
-        hidden_dtype = next(self.timestep_embedder.parameters()).dtype
+        hidden_dtype = timestep.dtype
 
         timesteps_proj = self.time_proj(timestep)
         timesteps_emb = self.timestep_embedder(timesteps_proj.to(dtype=hidden_dtype))  # (N, D)

diffsynth/pipelines/step_video.py

@@ -181,7 +181,7 @@ class StepVideoPipeline(BasePipeline):
         # Denoise
         self.load_models_to_device(["dit"])
         for progress_id, timestep in enumerate(progress_bar_cmd(self.scheduler.timesteps)):
-            timestep = timestep.unsqueeze(0).to(self.device)
+            timestep = timestep.unsqueeze(0).to(dtype=self.torch_dtype, device=self.device)
             print(f"Step {progress_id + 1} / {len(self.scheduler.timesteps)}")
 
             # Inference

examples/stepvideo/README.md

@@ -8,6 +8,10 @@ StepVideo is a state-of-the-art (SoTA) text-to-video pre-trained model with 30 b
 
 ## Examples
 
-See [`./stepvideo_text_to_video.py`](./stepvideo_text_to_video.py).
+For original BF16 version, please see [`./stepvideo_text_to_video.py`](./stepvideo_text_to_video.py). 80G VRAM required.
 
 https://github.com/user-attachments/assets/5954fdaa-a3cf-45a3-bd35-886e3cc4581b
+
+For FP8 quantized version, please see [`./stepvideo_text_to_video_quantized.py`](./stepvideo_text_to_video_quantized.py). 40G VRAM required.
+
+https://github.com/user-attachments/assets/f3697f4e-bc08-47d2-b00a-32d7dfa272ad

examples/stepvideo/stepvideo_text_to_video.py

@@ -13,7 +13,7 @@ torch.ops.load_library("models/stepfun-ai/stepvideo-t2v/lib/liboptimus_ths-torch
 # Load models
 model_manager = ModelManager()
 model_manager.load_models(
-    ["models/stepvideo-t2v/hunyuan_clip/clip_text_encoder/pytorch_model.bin"],
+    ["models/stepfun-ai/stepvideo-t2v/hunyuan_clip/clip_text_encoder/pytorch_model.bin"],
     torch_dtype=torch.float32, device="cpu"
 )
 model_manager.load_models(
@@ -42,6 +42,6 @@ pipe.enable_vram_management(num_persistent_param_in_dit=None)
 video = pipe(
     prompt="一名宇航员在月球上发现一块石碑，上面印有“stepfun”字样，闪闪发光。超高清、HDR 视频、环境光、杜比全景声、画面稳定、流畅动作、逼真的细节、专业级构图、超现实主义、自然、生动、超细节、清晰。",
     negative_prompt="画面暗、低分辨率、不良手、文本、缺少手指、多余的手指、裁剪、低质量、颗粒状、签名、水印、用户名、模糊。",
-    num_inference_steps=30, cfg_scale=9, num_frames=204, seed=1
+    num_inference_steps=30, cfg_scale=9, num_frames=51, seed=1
 )
 save_video(video, "video.mp4", fps=25, quality=5)

examples/stepvideo/stepvideo_text_to_video_quantized.py

@@ -0,0 +1,50 @@
+from modelscope import snapshot_download
+from diffsynth import ModelManager, StepVideoPipeline, save_video
+import torch
+
+
+# Download models
+snapshot_download(model_id="stepfun-ai/stepvideo-t2v", cache_dir="models")
+
+# Load the compiled attention for the LLM text encoder.
+# If you encounter errors here. Please select other compiled file that matches your environment or delete this line.
+torch.ops.load_library("models/stepfun-ai/stepvideo-t2v/lib/liboptimus_ths-torch2.5-cu124.cpython-310-x86_64-linux-gnu.so")
+
+# Load models
+model_manager = ModelManager()
+model_manager.load_models(
+    ["models/stepfun-ai/stepvideo-t2v/hunyuan_clip/clip_text_encoder/pytorch_model.bin"],
+    torch_dtype=torch.float32, device="cpu"
+)
+model_manager.load_models(
+    [
+        "models/stepfun-ai/stepvideo-t2v/step_llm",
+        [
+            "models/stepfun-ai/stepvideo-t2v/transformer/diffusion_pytorch_model-00001-of-00006.safetensors",
+            "models/stepfun-ai/stepvideo-t2v/transformer/diffusion_pytorch_model-00002-of-00006.safetensors",
+            "models/stepfun-ai/stepvideo-t2v/transformer/diffusion_pytorch_model-00003-of-00006.safetensors",
+            "models/stepfun-ai/stepvideo-t2v/transformer/diffusion_pytorch_model-00004-of-00006.safetensors",
+            "models/stepfun-ai/stepvideo-t2v/transformer/diffusion_pytorch_model-00005-of-00006.safetensors",
+            "models/stepfun-ai/stepvideo-t2v/transformer/diffusion_pytorch_model-00006-of-00006.safetensors",
+        ]
+    ],
+    torch_dtype=torch.float8_e4m3fn, device="cpu"
+)
+model_manager.load_models(
+    ["models/stepfun-ai/stepvideo-t2v/vae/vae_v2.safetensors"],
+    torch_dtype=torch.bfloat16, device="cpu"
+)
+pipe = StepVideoPipeline.from_model_manager(model_manager, torch_dtype=torch.bfloat16, device="cuda")
+
+# Enable VRAM management
+# This model requires 80G VRAM.
+# In order to reduce VRAM required, please set `num_persistent_param_in_dit` to a small number.
+pipe.enable_vram_management(num_persistent_param_in_dit=None)
+
+# Run!
+video = pipe(
+    prompt="一名宇航员在月球上发现一块石碑，上面印有“stepfun”字样，闪闪发光。超高清、HDR 视频、环境光、杜比全景声、画面稳定、流畅动作、逼真的细节、专业级构图、超现实主义、自然、生动、超细节、清晰。",
+    negative_prompt="画面暗、低分辨率、不良手、文本、缺少手指、多余的手指、裁剪、低质量、颗粒状、签名、水印、用户名、模糊。",
+    num_inference_steps=30, cfg_scale=9, num_frames=51, seed=1
+)
+save_video(video, "video.mp4", fps=25, quality=5)