| import torch |
| import torch.nn as nn |
| import torch.distributed as dist |
| import os |
| from distvae.utils import DistributedEnv |
|
|
| class Patchify(nn.Module): |
| def __init__(self): |
| super().__init__() |
| self.group_world_size = DistributedEnv.get_group_world_size() |
| self.rank_in_vae_group = DistributedEnv.get_rank_in_vae_group() |
| def forward(self, hidden_state): |
| height = hidden_state.shape[2] |
| start_idx = (height + self.group_world_size - 1) // self.group_world_size * self.rank_in_vae_group |
| end_idx = min((height + self.group_world_size - 1) // self.group_world_size * (self.rank_in_vae_group + 1), height) |
|
|
| return hidden_state[:, :, start_idx: end_idx, :].clone() |
|
|
|
|
| class DePatchify(nn.Module): |
| def __init__(self): |
| super().__init__() |
| self.group_world_size = DistributedEnv.get_group_world_size() |
| self.rank_in_vae_group = DistributedEnv.get_rank_in_vae_group() |
| self.local_rank = DistributedEnv.get_local_rank() |
| |
| def forward(self, patch_hidden_state): |
| patch_height_list = [torch.empty([1], dtype=torch.int64, device=f"cuda:{self.local_rank}") for _ in range(self.group_world_size)] |
| dist.all_gather( |
| patch_height_list, |
| torch.tensor( |
| [patch_hidden_state.shape[2]], |
| dtype=torch.int64, |
| device=f"cuda:{self.local_rank}" |
| ), |
| group=DistributedEnv.get_vae_group() |
| ) |
| patch_hidden_state_list = [ |
| torch.empty( |
| [patch_hidden_state.shape[0], patch_hidden_state.shape[1], patch_height_list[i].item(), patch_hidden_state.shape[-1]], |
| dtype=patch_hidden_state.dtype, |
| device=f"cuda:{self.local_rank}" |
| ) for i in range(self.group_world_size) |
| ] |
| dist.all_gather( |
| patch_hidden_state_list, |
| patch_hidden_state.contiguous(), |
| group=DistributedEnv.get_vae_group() |
| ) |
| return torch.cat(patch_hidden_state_list, dim=2) |
|
|
| |
