| 103 | |
| 104 | |
| 105 | class MultiTalkPipeline: |
| 106 | |
| 107 | def __init__( |
| 108 | self, |
| 109 | config, |
| 110 | checkpoint_dir, |
| 111 | quant_dir=None, |
| 112 | device_id=0, |
| 113 | rank=0, |
| 114 | t5_fsdp=False, |
| 115 | dit_fsdp=False, |
| 116 | use_usp=False, |
| 117 | t5_cpu=False, |
| 118 | init_on_cpu=True, |
| 119 | num_timesteps=1000, |
| 120 | use_timestep_transform=True, |
| 121 | lora_dir=None, |
| 122 | lora_scales=None, |
| 123 | quant = None |
| 124 | ): |
| 125 | r""" |
| 126 | Initializes the image-to-video generation model components. |
| 127 | |
| 128 | Args: |
| 129 | config (EasyDict): |
| 130 | Object containing model parameters initialized from config.py |
| 131 | checkpoint_dir (`str`): |
| 132 | Path to directory containing model checkpoints |
| 133 | device_id (`int`, *optional*, defaults to 0): |
| 134 | Id of target GPU device |
| 135 | rank (`int`, *optional*, defaults to 0): |
| 136 | Process rank for distributed training |
| 137 | t5_fsdp (`bool`, *optional*, defaults to False): |
| 138 | Enable FSDP sharding for T5 model |
| 139 | dit_fsdp (`bool`, *optional*, defaults to False): |
| 140 | Enable FSDP sharding for DiT model |
| 141 | use_usp (`bool`, *optional*, defaults to False): |
| 142 | Enable distribution strategy of USP. |
| 143 | t5_cpu (`bool`, *optional*, defaults to False): |
| 144 | Whether to place T5 model on CPU. Only works without t5_fsdp. |
| 145 | init_on_cpu (`bool`, *optional*, defaults to True): |
| 146 | Enable initializing Transformer Model on CPU. Only works without FSDP or USP. |
| 147 | quant (`str`, *optional*, defaults to None): |
| 148 | Quantization type, must be 'int8' or 'fp8'. |
| 149 | """ |
| 150 | if quant is not None and quant not in ("int8", "fp8"): |
| 151 | raise ValueError("quant must be 'int8', 'fp8', or None(default fp32 model)") |
| 152 | self.device = torch.device(f"cuda:{device_id}") |
| 153 | self.config = config |
| 154 | self.rank = rank |
| 155 | self.use_usp = use_usp |
| 156 | self.t5_cpu = t5_cpu |
| 157 | |
| 158 | self.num_train_timesteps = config.num_train_timesteps |
| 159 | self.param_dtype = config.param_dtype |
| 160 | |
| 161 | shard_fn = partial(shard_model, device_id=device_id) |
| 162 | self.text_encoder = T5EncoderModel( |
nothing calls this directly
no outgoing calls
no test coverage detected