| |
| |
| import os |
| import math |
| import torch |
| import torch.nn.functional as F |
| import torch.distributed as dist |
| from torch.autograd import Function |
| from torch.distributed import group, ReduceOp |
|
|
|
|
| def get_group_rank(group_size): |
| global_rank = int(os.environ["RANK"]) |
| group_rank = global_rank % group_size |
| return group_rank |
|
|
|
|
| _parallel_state = {} |
|
|
| |
| |
| |
|
|
|
|
| def get_parallel_state(): |
| return _parallel_state |
|
|
|
|
| class _AllGather(Function): |
| @staticmethod |
| def forward(ctx, group, tensor): |
| tensor = tensor.contiguous() |
| ctx.group = group |
| group_size = dist.get_world_size(group=group) |
| out_tensor_list = [torch.empty_like(tensor) for _ in range(group_size)] |
| dist.all_gather(out_tensor_list, tensor, group=group) |
| return tuple(out_tensor_list) |
|
|
| @staticmethod |
| def backward(ctx, *grad_outputs): |
| rank = dist.get_rank(group=ctx.group) |
| gx = torch.empty_like(grad_outputs[rank]) |
| gx = gx.contiguous() |
| grad_outputs = tuple(t.contiguous() for t in grad_outputs) |
| dist.reduce_scatter(gx, list(grad_outputs), op=ReduceOp.SUM, group=ctx.group) |
| return (None, gx) |
|
|
|
|
| @torch.compiler.disable |
| def all_gather(tensor, group=group.WORLD): |
| return _AllGather.apply(group, tensor) |
|
|
|
|
| class _AllGatherVarShape(Function): |
| @staticmethod |
| def forward(ctx, group, tensor): |
| tensor = tensor.contiguous() |
| ctx.group = group |
| ctx.original_shape = tensor.shape |
|
|
| shape_info = torch.tensor( |
| list(tensor.shape), dtype=torch.long, device=tensor.device |
| ) |
|
|
| shape_list = [ |
| torch.empty_like(shape_info) |
| for _ in range(dist.get_world_size(group=group)) |
| ] |
| dist.all_gather(shape_list, shape_info, group=group) |
|
|
| all_shapes = [tuple(shape_tensor.tolist()) for shape_tensor in shape_list] |
| ctx.all_shapes = all_shapes |
|
|
| flat_tensor = tensor.flatten() |
| max_size = max(math.prod(s) for s in all_shapes) |
|
|
| if flat_tensor.numel() < max_size: |
| padded = torch.zeros(max_size, dtype=tensor.dtype, device=tensor.device) |
| padded[: flat_tensor.numel()] = flat_tensor |
| flat_tensor = padded |
|
|
| gathered_flat = [torch.empty_like(flat_tensor) for _ in range(len(all_shapes))] |
| dist.all_gather(gathered_flat, flat_tensor, group=group) |
|
|
| return tuple( |
| t[: math.prod(shape)].reshape(shape) |
| for t, shape in zip(gathered_flat, all_shapes) |
| ) |
|
|
| @staticmethod |
| def backward(ctx, *grad_outputs): |
| rank = dist.get_rank(group=ctx.group) |
|
|
| grad_input = grad_outputs[rank] |
| if grad_input is None: |
| return None, torch.zeros( |
| ctx.original_shape, device=next(iter(grad_outputs)).device |
| ) |
|
|
| max_size = max(math.prod(shape) for shape in ctx.all_shapes) |
| padded_grads = [] |
|
|
| for grad, shape in zip(grad_outputs, ctx.all_shapes): |
| if grad is not None: |
| flat_grad = grad.flatten() |
| else: |
| flat_grad = torch.zeros( |
| math.prod(shape), |
| dtype=grad_input.dtype, |
| device=grad_input.device, |
| ) |
|
|
| if flat_grad.numel() < max_size: |
| padded = torch.zeros( |
| max_size, dtype=flat_grad.dtype, device=flat_grad.device |
| ) |
| padded[: flat_grad.numel()] = flat_grad |
| padded_grads.append(padded) |
| else: |
| padded_grads.append(flat_grad) |
|
|
| result_grad = torch.empty_like(padded_grads[0]) |
| dist.reduce_scatter(result_grad, padded_grads, op=ReduceOp.SUM, group=ctx.group) |
|
|
| original_size = math.prod(ctx.original_shape) |
| return None, result_grad[:original_size].reshape(ctx.original_shape) |
|
|
|
|
| @torch.compiler.disable |
| def all_gather_var_shape(tensor, group=group.WORLD): |
| return _AllGatherVarShape.apply(group, tensor) |
|
|
|
|
| class _AllReduce(Function): |
| @staticmethod |
| def forward(ctx, _input, op, group): |
| ctx.group = group |
| ctx.op = op |
| _input = _input.clone() |
| dist.all_reduce(_input, op=op, group=group) |
| return _input |
|
|
| @staticmethod |
| def backward(ctx, grad_output): |
| grad_output = grad_output.clone() |
| dist.all_reduce(grad_output, op=ctx.op, group=ctx.group) |
| return grad_output, None, None |
|
|
|
|
| @torch.compiler.disable |
| def all_reduce(input_, op, group): |
| return _AllReduce.apply(input_, op, group) |
|
|
|
|
| class _AlltoAllSingle(Function): |
| @staticmethod |
| def forward(ctx, group, input): |
| ctx.group = group |
|
|
| world_size = dist.get_world_size(group=group) |
| if world_size == 1: |
| return input |
|
|
| input = input.contiguous() |
| output = torch.empty_like(input) |
| dist.all_to_all_single( |
| output, |
| input, |
| group=group, |
| ) |
| return output |
|
|
| @staticmethod |
| def backward(ctx, grad_output): |
| return (None, _AlltoAllSingle.apply(ctx.group, grad_output)) |
|
|
|
|
| @torch.compiler.disable |
| def all_to_all_single(input, group=group.WORLD): |
| return _AlltoAllSingle.apply(group, input) |
|
|
|
|
|
|
| @torch.compiler.disable |
| def get_subseq(input, sp_size=None): |
| if sp_size is None: |
| state = get_parallel_state() |
| if not state.get("sp_enabled", False): |
| return input |
| sp_size = state["sp_size"] |
| sp_rank = state["sp_rank"] |
| else: |
| sp_rank = get_group_rank(sp_size) |
|
|
| if sp_size == 1: |
| return input |
|
|
| if input.shape[1] % sp_size != 0: |
| raise ValueError( |
| f"Input shape {input.shape} is not divisible by sp_size {sp_size}" |
| ) |
|
|
| return torch.chunk(input, sp_size, dim=1)[sp_rank] |
|
|
|
|
| @torch.compiler.disable |
| def gather_subseq(input, sp_size=None, local_process_group=None): |
| if sp_size is None: |
| state = get_parallel_state() |
| if not state.get("sp_enabled", False): |
| return input |
| sp_size = state["sp_size"] |
| local_process_group = state["sp_process_group"] |
|
|
| if sp_size == 1: |
| return input |
|
|
| output = all_gather(input, group=local_process_group) |
| output = torch.cat(output, dim=1) |
| return output |
|
|
|
|
| @torch.compiler.disable |
| def all_to_all_4D( |
| input: torch.tensor, |
| scatter_idx: int = 2, |
| gather_idx: int = 1, |
| group=None, |
| ): |
| assert ( |
| input.dim() == 4 |
| ), f"input must be 4D tensor, got {input.dim()} and shape {input.shape}" |
|
|
| if group is None: |
| seq_world_size = 1 |
| else: |
| seq_world_size = dist.get_world_size(group) |
|
|
| if seq_world_size == 1: |
| return input |
|
|
| if scatter_idx == 2 and gather_idx == 1: |
| bs, shard_seqlen, hc, hs = input.shape |
| seqlen = shard_seqlen * seq_world_size |
| shard_hc = hc // seq_world_size |
|
|
| input_t = ( |
| input.reshape(bs, shard_seqlen, seq_world_size, shard_hc, hs) |
| .transpose(0, 2) |
| .contiguous() |
| ) |
|
|
| output = all_to_all_single(input_t, group=group) |
| output = output.reshape(seqlen, bs, shard_hc, hs) |
| output = output.transpose(0, 1).contiguous().reshape(bs, seqlen, shard_hc, hs) |
| return output |
|
|
| elif scatter_idx == 1 and gather_idx == 2: |
| bs, seqlen, shard_hc, hs = input.shape |
| hc = shard_hc * seq_world_size |
| shard_seqlen = seqlen // seq_world_size |
|
|
| input_t = ( |
| input.reshape(bs, seq_world_size, shard_seqlen, shard_hc, hs) |
| .transpose(0, 3) |
| .transpose(0, 1) |
| .contiguous() |
| .reshape(seq_world_size, shard_hc, shard_seqlen, bs, hs) |
| ) |
|
|
| output = all_to_all_single(input_t, group=group) |
| output = output.reshape(hc, shard_seqlen, bs, hs) |
| output = output.transpose(0, 2).contiguous().reshape(bs, shard_seqlen, hc, hs) |
| return output |
| else: |
| raise RuntimeError("scatter_idx must be 1 or 2 and gather_idx must be 1 or 2") |
|
|
|
|
|
|
| @torch.compiler.disable |
| def exchange_borders( |
| input_, padding, pad_mode, sp_rank, sp_size, group, dim=-1, async_op=False |
| ): |
| if async_op and input_.requires_grad: |
| raise ValueError("async_op is not supported backward, check previous commits") |
|
|
| slice_indices = [slice(None)] * input_.ndim |
| slice_indices[dim] = slice(None, padding) |
| first_tensor = input_[tuple(slice_indices)].contiguous() |
|
|
| slice_indices[dim] = slice(-padding, None) |
| last_tensor = input_[tuple(slice_indices)].contiguous() |
|
|
| if async_op: |
| first_borders = [torch.empty_like(first_tensor) for _ in range(sp_size)] |
| last_borders = [torch.empty_like(last_tensor) for _ in range(sp_size)] |
|
|
| handle_first = dist.all_gather( |
| first_borders, first_tensor, group=group, async_op=True |
| ) |
| handle_last = dist.all_gather( |
| last_borders, last_tensor, group=group, async_op=True |
| ) |
| else: |
| first_borders = all_gather(first_tensor, group=group) |
| last_borders = all_gather(last_tensor, group=group) |
|
|
| if dim < 0: |
| pad_dim = -1 - dim |
| else: |
| pad_dim = input_.ndim - 1 - dim |
|
|
| pad_size = [0] * ((input_.ndim - 2) * 2) |
| pad_size[pad_dim * 2] = padding |
| pad_size[pad_dim * 2 + 1] = padding |
| output = F.pad(input_, pad_size, mode=pad_mode) |
|
|
| slice_indices = [slice(None)] * input_.ndim |
| slice_indices[dim] = slice(-padding, None) |
|
|
| if async_op: |
| handle_first.wait() |
|
|
| if sp_rank < sp_size - 1: |
| output[tuple(slice_indices)] = first_borders[sp_rank + 1] |
| else: |
| output[tuple(slice_indices)] += first_borders[0] * 0.0 |
|
|
| slice_indices = [slice(None)] * input_.ndim |
| slice_indices[dim] = slice(None, padding) |
|
|
| if async_op: |
| handle_last.wait() |
|
|
| if sp_rank > 0: |
| output[tuple(slice_indices)] = last_borders[sp_rank - 1] |
| else: |
| output[tuple(slice_indices)] += last_borders[sp_size - 1] * 0.0 |
|
|
| return output |
|
|
|
|
| @torch.compiler.disable |
| def exchange_strides( |
| input_, pad_mode, sp_rank, sp_size, group, dim=-1, async_op=False |
| ): |
| if async_op and input_.requires_grad: |
| raise ValueError("async_op is not supported backward, check previous commits") |
|
|
| if dim not in [-1, -2]: |
| raise ValueError("dim must be -1 (W) or -2 (H) for exchange_strides") |
|
|
| if dim == -1: |
| if input_.ndim == 5: |
| input_ = F.pad(input_, (0, 0, 0, 1, 0, 0), mode=pad_mode) |
| elif input_.ndim == 4: |
| input_ = F.pad(input_, (0, 0, 0, 1), mode=pad_mode) |
| else: |
| raise ValueError(f"Input must have 4 or 5 dimensions, got {input_.ndim}") |
|
|
| left_border = input_[..., :1].contiguous() |
|
|
| if async_op: |
| left_borders = [torch.empty_like(left_border) for _ in range(sp_size)] |
| handle = dist.all_gather( |
| left_borders, left_border, group=group, async_op=True |
| ) |
| else: |
| left_borders = all_gather(left_border, group=group) |
|
|
| if input_.ndim == 5: |
| output = F.pad(input_, (0, 1, 0, 0, 0, 0), mode=pad_mode) |
| elif input_.ndim == 4: |
| output = F.pad(input_, (0, 1, 0, 0), mode=pad_mode) |
| else: |
| raise ValueError(f"Input must have 4 or 5 dimensions, got {input_.ndim}") |
|
|
| if async_op: |
| handle.wait() |
|
|
| if sp_rank != sp_size - 1: |
| output[..., -1:] = left_borders[sp_rank + 1] |
| else: |
| output[..., -1:] += left_borders[0] * 0.0 |
| else: |
| if input_.ndim == 5: |
| input_ = F.pad(input_, (0, 1, 0, 0, 0, 0), mode=pad_mode) |
| elif input_.ndim == 4: |
| input_ = F.pad(input_, (0, 1, 0, 0), mode=pad_mode) |
| else: |
| raise ValueError(f"Input must have 4 or 5 dimensions, got {input_.ndim}") |
|
|
| top_border = input_[..., :1, :].contiguous() |
|
|
| if async_op: |
| top_borders = [torch.empty_like(top_border) for _ in range(sp_size)] |
| handle = dist.all_gather( |
| top_borders, top_border, group=group, async_op=True |
| ) |
| else: |
| top_borders = all_gather(top_border, group=group) |
|
|
| if input_.ndim == 5: |
| output = F.pad(input_, (0, 0, 0, 1, 0, 0), mode=pad_mode) |
| elif input_.ndim == 4: |
| output = F.pad(input_, (0, 0, 0, 1), mode=pad_mode) |
| else: |
| raise ValueError(f"Input must have 4 or 5 dimensions, got {input_.ndim}") |
|
|
| if async_op: |
| handle.wait() |
|
|
| if sp_rank != sp_size - 1: |
| output[..., -1:, :] = top_borders[sp_rank + 1] |
| else: |
| output[..., -1:, :] += top_borders[0] * 0.0 |
|
|
| return output |
|
|