Add support for fp8_e4m3fn model

README.md

@@ -135,6 +135,14 @@ If you encounter OOM (Out-of-Memory) issues, you can use the `--offload_model Tr
 python generate.py  --task t2v-1.3B --size 832*480 --ckpt_dir ./Wan2.1-T2V-1.3B --offload_model True --t5_cpu --sample_shift 8 --sample_guide_scale 6 --prompt "Two anthropomorphic cats in comfy boxing gear and bright gloves fight intensely on a spotlighted stage."
 ```
 
+You can also use the `--fp8` option to enable FP8 precision for reduced memory usage. Make sure to download the [FP8 model weight](https://huggingface.co/Kijai/WanVideo_comfy/resolve/main/Wan2_1-T2V-1_3B_fp8_e4m3fn.safetensors) and place it in the `Wan2.1-T2V-1.3B` folder.
+
+Additionally, an [FP8 version of the T5 model](https://huggingface.co/Kijai/WanVideo_comfy/resolve/main/umt5-xxl-enc-fp8_e4m3fn.safetensors) is available. To use the FP8 T5 model, update the configuration file:
+
+```
+t2v_1_3B.t5_checkpoint = 'umt5-xxl-enc-fp8_e4m3fn.safetensors'
+```
+
 > 💡Note: If you are using the `T2V-1.3B` model, we recommend setting the parameter `--sample_guide_scale 6`. The `--sample_shift parameter` can be adjusted within the range of 8 to 12 based on the performance.
 
 
@@ -222,6 +230,17 @@ Similar to Text-to-Video, Image-to-Video is also divided into processes with and
 python generate.py --task i2v-14B --size 1280*720 --ckpt_dir ./Wan2.1-I2V-14B-720P --image examples/i2v_input.JPG --prompt "Summer beach vacation style, a white cat wearing sunglasses sits on a surfboard. The fluffy-furred feline gazes directly at the camera with a relaxed expression. Blurred beach scenery forms the background featuring crystal-clear waters, distant green hills, and a blue sky dotted with white clouds. The cat assumes a naturally relaxed posture, as if savoring the sea breeze and warm sunlight. A close-up shot highlights the feline's intricate details and the refreshing atmosphere of the seaside."
 ```
 
+To minimize GPU memory usage, you can enable model offloading with `--offload_model True` and use FP8 precision with `--fp8`.
+
+For example, to run **Wan2.1-I2V-14B-480P** on an RTX 4090 GPU:
+
+1. First, download the [FP8 model weights](https://huggingface.co/Kijai/WanVideo_comfy/resolve/main/Wan2_1-I2V-14B-480P_fp8_e4m3fn.safetensors) and place them in the `Wan2.1-I2V-14B-480P` folder.
+2. Then, execute the following command:
+
+```
+python generate.py --task i2v-14B --size 832*480 --ckpt_dir ./Wan2.1-I2V-14B-480P --offload_model True --fp8 --image examples/i2v_input.JPG --prompt "Summer beach vacation style, a white cat wearing sunglasses sits on a surfboard. The fluffy-furred feline gazes directly at the camera with a relaxed expression. Blurred beach scenery forms the background featuring crystal-clear waters, distant green hills, and a blue sky dotted with white clouds. The cat assumes a naturally relaxed posture, as if savoring the sea breeze and warm sunlight. A close-up shot highlights the feline's intricate details and the refreshing atmosphere of the seaside." 
+```
+
 > 💡For the Image-to-Video task, the `size` parameter represents the area of the generated video, with the aspect ratio following that of the original input image.
 
 

generate.py

@@ -126,6 +126,11 @@ def _parse_args():
         action="store_true",
         default=False,
         help="Whether to use FSDP for DiT.")
+    parser.add_argument(
+        "--fp8",
+        action="store_true",
+        default=False,
+        help="Whether to use fp8.")
     parser.add_argument(
         "--save_file",
         type=str,
@@ -306,6 +311,7 @@ def generate(args):
             dit_fsdp=args.dit_fsdp,
             use_usp=(args.ulysses_size > 1 or args.ring_size > 1),
             t5_cpu=args.t5_cpu,
+            fp8=args.fp8,
         )
 
         logging.info(
@@ -363,6 +369,7 @@ def generate(args):
             dit_fsdp=args.dit_fsdp,
             use_usp=(args.ulysses_size > 1 or args.ring_size > 1),
             t5_cpu=args.t5_cpu,
+            fp8=args.fp8,
         )
 
         logging.info("Generating video ...")

wan/configs/wan_i2v_14B.py

@@ -10,12 +10,14 @@
 i2v_14B.update(wan_shared_cfg)
 
 i2v_14B.t5_checkpoint = 'models_t5_umt5-xxl-enc-bf16.pth'
+# i2v_14B.t5_checkpoint = 'umt5-xxl-enc-fp8_e4m3fn.safetensors' # fp8 model
 i2v_14B.t5_tokenizer = 'google/umt5-xxl'
 
 # clip
 i2v_14B.clip_model = 'clip_xlm_roberta_vit_h_14'
 i2v_14B.clip_dtype = torch.float16
 i2v_14B.clip_checkpoint = 'models_clip_open-clip-xlm-roberta-large-vit-huge-14.pth'
+# i2v_14B.clip_checkpoint = 'open-clip-xlm-roberta-large-vit-huge-14_fp16.safetensors' # Kijai's fp16 model
 i2v_14B.clip_tokenizer = 'xlm-roberta-large'
 
 # vae

wan/configs/wan_t2v_14B.py

@@ -10,6 +10,7 @@
 
 # t5
 t2v_14B.t5_checkpoint = 'models_t5_umt5-xxl-enc-bf16.pth'
+# t2v_14B.t5_checkpoint = 'umt5-xxl-enc-fp8_e4m3fn.safetensors' # fp8 model
 t2v_14B.t5_tokenizer = 'google/umt5-xxl'
 
 # vae

wan/configs/wan_t2v_1_3B.py

@@ -10,6 +10,7 @@
 
 # t5
 t2v_1_3B.t5_checkpoint = 'models_t5_umt5-xxl-enc-bf16.pth'
+# t2v_1_3B.t5_checkpoint = 'umt5-xxl-enc-fp8_e4m3fn.safetensors' # fp8 model
 t2v_1_3B.t5_tokenizer = 'google/umt5-xxl'
 
 # vae

wan/image2video.py

@@ -16,6 +16,10 @@
 import torchvision.transforms.functional as TF
 from tqdm import tqdm
 
+from accelerate import init_empty_weights
+from accelerate.utils import set_module_tensor_to_device
+from safetensors.torch import load_file
+
 from .distributed.fsdp import shard_model
 from .modules.clip import CLIPModel
 from .modules.model import WanModel
@@ -39,6 +43,7 @@ def __init__(
         use_usp=False,
         t5_cpu=False,
         init_on_cpu=True,
+        fp8=False,
     ):
         r"""
         Initializes the image-to-video generation model components.
@@ -62,6 +67,8 @@ def __init__(
                 Whether to place T5 model on CPU. Only works without t5_fsdp.
             init_on_cpu (`bool`, *optional*, defaults to True):
                 Enable initializing Transformer Model on CPU. Only works without FSDP or USP.
+            fp8 (`bool`, *optional*, defaults to False):
+                Enable 8-bit floating point precision for model parameters.
         """
         self.device = torch.device(f"cuda:{device_id}")
         self.config = config
@@ -73,17 +80,23 @@ def __init__(
         self.param_dtype = config.param_dtype
 
         shard_fn = partial(shard_model, device_id=device_id)
+        if config.t5_checkpoint == 'umt5-xxl-enc-fp8_e4m3fn.safetensors':
+            quantization = "fp8_e4m3fn" 
+        else:
+            quantization = "disabled"
         self.text_encoder = T5EncoderModel(
             text_len=config.text_len,
             dtype=config.t5_dtype,
             device=torch.device('cpu'),
             checkpoint_path=os.path.join(checkpoint_dir, config.t5_checkpoint),
             tokenizer_path=os.path.join(checkpoint_dir, config.t5_tokenizer),
             shard_fn=shard_fn if t5_fsdp else None,
+            quantization=quantization,
         )
 
         self.vae_stride = config.vae_stride
         self.patch_size = config.patch_size
+
         self.vae = WanVAE(
             vae_pth=os.path.join(checkpoint_dir, config.vae_checkpoint),
             device=self.device)
@@ -96,7 +109,46 @@ def __init__(
             tokenizer_path=os.path.join(checkpoint_dir, config.clip_tokenizer))
 
         logging.info(f"Creating WanModel from {checkpoint_dir}")
-        self.model = WanModel.from_pretrained(checkpoint_dir)
+        if not fp8:
+            self.model = WanModel.from_pretrained(checkpoint_dir)
+        else:
+            if '480P' in checkpoint_dir:
+                state_dict = load_file(checkpoint_dir+'/Wan2_1-I2V-14B-480P_fp8_e4m3fn.safetensors', device="cpu")
+            elif '720P' in checkpoint_dir:
+                state_dict = load_file(checkpoint_dir+'/Wan2_1-I2V-14B-720P_fp8_e4m3fn.safetensors', device="cpu")
+            dim = state_dict["patch_embedding.weight"].shape[0]
+            in_channels = state_dict["patch_embedding.weight"].shape[1]
+            ffn_dim = state_dict["blocks.0.ffn.0.bias"].shape[0]
+            model_type = "i2v" if in_channels == 36 else "t2v"
+            num_heads = 40 if dim == 5120 else 12
+            num_layers = 40 if dim == 5120 else 30
+            TRANSFORMER_CONFIG= {
+                "dim": dim,
+                "ffn_dim": ffn_dim,
+                "eps": 1e-06,
+                "freq_dim": 256,
+                "in_dim": in_channels,
+                "model_type": model_type,
+                "out_dim": 16,
+                "text_len": 512,
+                "num_heads": num_heads,
+                "num_layers": num_layers,
+            }
+
+            with init_empty_weights():
+                self.model = WanModel(**TRANSFORMER_CONFIG)
+
+            base_dtype=torch.bfloat16
+            dtype=torch.float8_e4m3fn
+            params_to_keep = {"norm", "head", "bias", "time_in", "vector_in", "patch_embedding", "time_", "img_emb", "modulation"}
+            for name, param in self.model.named_parameters():
+                dtype_to_use = base_dtype if any(keyword in name for keyword in params_to_keep) else dtype
+                # dtype_to_use = torch.bfloat16
+                # print("Assigning Parameter name: ", name, " with dtype: ", dtype_to_use)
+                set_module_tensor_to_device(self.model, name, device='cpu', dtype=dtype_to_use, value=state_dict[name])
+
+            del state_dict
+
         self.model.eval().requires_grad_(False)
 
         if t5_fsdp or dit_fsdp or use_usp:
@@ -219,13 +271,15 @@ def generate(self,
         # preprocess
         if not self.t5_cpu:
             self.text_encoder.model.to(self.device)
-            context = self.text_encoder([input_prompt], self.device)
-            context_null = self.text_encoder([n_prompt], self.device)
+            with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True):
+                context = self.text_encoder([input_prompt], self.device)
+                context_null = self.text_encoder([n_prompt], self.device)
             if offload_model:
                 self.text_encoder.model.cpu()
         else:
-            context = self.text_encoder([input_prompt], torch.device('cpu'))
-            context_null = self.text_encoder([n_prompt], torch.device('cpu'))
+            with torch.autocast(device_type="cpu", dtype=torch.bfloat16, enabled=True):
+                context = self.text_encoder([input_prompt], torch.device('cpu'))
+                context_null = self.text_encoder([n_prompt], torch.device('cpu'))
             context = [t.to(self.device) for t in context]
             context_null = [t.to(self.device) for t in context_null]
 
@@ -242,9 +296,12 @@ def generate(self,
                 torch.zeros(3, 80, h, w)
             ],
                          dim=1).to(self.device)
-        ])[0]
+        ],device=self.device)[0]
         y = torch.concat([msk, y])
 
+        if offload_model:
+            self.vae.model.cpu()
+
         @contextmanager
         def noop_no_sync():
             yield
@@ -332,9 +389,11 @@ def noop_no_sync():
             if offload_model:
                 self.model.cpu()
                 torch.cuda.empty_cache()
+                # load vae model back to device
+                self.vae.model.to(self.device)
 
             if self.rank == 0:
-                videos = self.vae.decode(x0)
+                videos = self.vae.decode(x0, device=self.device)
 
         del noise, latent
         del sample_scheduler

wan/modules/clip.py

@@ -7,6 +7,7 @@
 import torch.nn as nn
 import torch.nn.functional as F
 import torchvision.transforms as T
+from safetensors.torch import load_file
 
 from .attention import flash_attention
 from .tokenizers import HuggingfaceTokenizer
@@ -515,8 +516,13 @@ def __init__(self, dtype, device, checkpoint_path, tokenizer_path):
             device=device)
         self.model = self.model.eval().requires_grad_(False)
         logging.info(f'loading {checkpoint_path}')
-        self.model.load_state_dict(
-            torch.load(checkpoint_path, map_location='cpu'))
+        if checkpoint_path.endswith('.safetensors'):
+            state_dict = load_file(checkpoint_path, device='cpu')
+            self.model.load_state_dict(state_dict)
+        elif checkpoint_path.endswith('.pth'):
+            self.model.load_state_dict(torch.load(checkpoint_path, map_location='cpu'))
+        else:
+            raise ValueError(f'Unsupported checkpoint file format: {checkpoint_path}')
 
         # init tokenizer
         self.tokenizer = HuggingfaceTokenizer(

wan/modules/t5.py

@@ -9,6 +9,10 @@
 
 from .tokenizers import HuggingfaceTokenizer
 
+from accelerate import init_empty_weights
+from accelerate.utils import set_module_tensor_to_device
+from safetensors.torch import load_file
+
 __all__ = [
     'T5Model',
     'T5Encoder',
@@ -442,7 +446,7 @@ def _t5(name,
         model = model_cls(**kwargs)
 
     # set device
-    model = model.to(dtype=dtype, device=device)
+    # model = model.to(dtype=dtype, device=device)
 
     # init tokenizer
     if return_tokenizer:
@@ -479,21 +483,39 @@ def __init__(
         checkpoint_path=None,
         tokenizer_path=None,
         shard_fn=None,
+        quantization="disabled",
     ):
         self.text_len = text_len
         self.dtype = dtype
         self.device = device
         self.checkpoint_path = checkpoint_path
         self.tokenizer_path = tokenizer_path
 
-        # init model
-        model = umt5_xxl(
-            encoder_only=True,
-            return_tokenizer=False,
-            dtype=dtype,
-            device=device).eval().requires_grad_(False)
+
         logging.info(f'loading {checkpoint_path}')
-        model.load_state_dict(torch.load(checkpoint_path, map_location='cpu'))
+        if quantization == "disabled":
+            # init model
+            model = umt5_xxl(
+                encoder_only=True,
+                return_tokenizer=False,
+                dtype=dtype,
+                device=device).eval().requires_grad_(False)
+            model.load_state_dict(torch.load(checkpoint_path, map_location='cpu'))
+        elif quantization == "fp8_e4m3fn":
+            with init_empty_weights():
+                model = umt5_xxl(
+                    encoder_only=True,
+                    return_tokenizer=False,
+                    dtype=dtype,
+                    device=device).eval().requires_grad_(False)
+            cast_dtype = torch.float8_e4m3fn
+            state_dict = load_file(checkpoint_path, device="cpu")
+            params_to_keep = {'norm', 'pos_embedding', 'token_embedding'}
+            for name, param in model.named_parameters():
+                dtype_to_use = dtype if any(keyword in name for keyword in params_to_keep) else cast_dtype
+                set_module_tensor_to_device(model, name, device=device, dtype=dtype_to_use, value=state_dict[name])
+            del state_dict
+
         self.model = model
         if shard_fn is not None:
             self.model = shard_fn(self.model, sync_module_states=False)

wan/modules/vae.py

@@ -644,7 +644,7 @@ def __init__(self,
             z_dim=z_dim,
         ).eval().requires_grad_(False).to(device)
 
-    def encode(self, videos):
+    def encode(self, videos, device=None):
         """
         videos: A list of videos each with shape [C, T, H, W].
         """
@@ -654,7 +654,7 @@ def encode(self, videos):
                 for u in videos
             ]
 
-    def decode(self, zs):
+    def decode(self, zs, device=None):
         with amp.autocast(dtype=self.dtype):
             return [
                 self.model.decode(u.unsqueeze(0),