File size: 11,895 Bytes
74da989 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 | # SPDX-License-Identifier: Apache-2.0
# Adapted from https://github.com/vllm-project/vllm/blob/v0.7.3/vllm/distributed/device_communicators/base_device_communicator.py
from typing import Any
import torch
import torch.distributed as dist
from torch import Tensor
from torch.distributed import ProcessGroup, ReduceOp
class DistributedAutograd:
"""Collection of autograd functions for distributed operations.
This class provides custom autograd functions for distributed operations like all_reduce,
all_gather, and all_to_all. Each operation is implemented as a static inner class with
proper forward and backward implementations.
"""
class AllReduce(torch.autograd.Function):
"""Differentiable all_reduce operation.
The gradient of all_reduce is another all_reduce operation since the operation
combines values from all ranks equally.
"""
@staticmethod
def forward(ctx: Any,
group: ProcessGroup,
input_: Tensor,
op: dist.ReduceOp | None = None) -> Tensor:
ctx.group = group
ctx.op = op
output = input_.clone()
dist.all_reduce(output, group=group, op=op)
return output
@staticmethod
def backward(ctx: Any,
grad_output: Tensor) -> tuple[None, Tensor, None]:
grad_output = grad_output.clone()
dist.all_reduce(grad_output, group=ctx.group, op=ctx.op)
return None, grad_output, None
class AllGather(torch.autograd.Function):
"""Differentiable all_gather operation.
The operation gathers tensors from all ranks and concatenates them along a specified dimension.
The backward pass uses reduce_scatter to efficiently distribute gradients back to source ranks.
"""
@staticmethod
def forward(ctx: Any, group: ProcessGroup, input_: Tensor,
world_size: int, dim: int) -> Tensor:
ctx.group = group
ctx.world_size = world_size
ctx.dim = dim
ctx.input_shape = input_.shape
input_size = input_.size()
output_size = (input_size[0] * world_size, ) + input_size[1:]
output_tensor = torch.empty(output_size,
dtype=input_.dtype,
device=input_.device)
dist.all_gather_into_tensor(output_tensor, input_, group=group)
output_tensor = output_tensor.reshape((world_size, ) + input_size)
output_tensor = output_tensor.movedim(0, dim)
output_tensor = output_tensor.reshape(input_size[:dim] +
(world_size *
input_size[dim], ) +
input_size[dim + 1:])
return output_tensor
@staticmethod
def backward(ctx: Any,
grad_output: Tensor) -> tuple[None, Tensor, None, None]:
# Split the gradient tensor along the gathered dimension
dim_size = grad_output.size(ctx.dim) // ctx.world_size
grad_chunks = grad_output.reshape(grad_output.shape[:ctx.dim] +
(ctx.world_size, dim_size) +
grad_output.shape[ctx.dim + 1:])
grad_chunks = grad_chunks.movedim(ctx.dim, 0)
# Each rank only needs its corresponding gradient
grad_input = torch.empty(ctx.input_shape,
dtype=grad_output.dtype,
device=grad_output.device)
dist.reduce_scatter_tensor(grad_input,
grad_chunks.contiguous(),
group=ctx.group)
return None, grad_input, None, None
class AllToAll4D(torch.autograd.Function):
"""Differentiable all_to_all operation specialized for 4D tensors.
This operation is particularly useful for attention operations where we need to
redistribute data across ranks for efficient parallel processing.
The operation supports two modes:
1. scatter_dim=2, gather_dim=1: Used for redistributing attention heads
2. scatter_dim=1, gather_dim=2: Used for redistributing sequence dimensions
"""
@staticmethod
def forward(ctx: Any, group: ProcessGroup, input_: Tensor,
world_size: int, scatter_dim: int,
gather_dim: int) -> Tensor:
ctx.group = group
ctx.world_size = world_size
ctx.scatter_dim = scatter_dim
ctx.gather_dim = gather_dim
if world_size == 1:
return input_
assert input_.dim(
) == 4, f"input must be 4D tensor, got {input_.dim()} and shape {input_.shape}"
if scatter_dim == 2 and gather_dim == 1:
bs, shard_seqlen, hn, hd = input_.shape
seqlen = shard_seqlen * world_size
shard_hn = hn // world_size
input_ = input_.transpose(
0, 2).contiguous() # hn, shard_seqlen, bs, hd
output = torch.empty_like(input_)
dist.all_to_all_single(output, input_,
group=group) # hn, shard_seqlen, bs, hd
output = torch.cat(output.split(shard_hn),
dim=1) # sharded hn, seqlen, bs, hd
output = output.transpose(
0, 2).contiguous() # bs, seqlen, sharded_hn, hd
return output
elif scatter_dim == 1 and gather_dim == 2:
bs, seqlen, shard_hn, hd = input_.shape
hn = shard_hn * world_size
shard_seqlen = seqlen // world_size
input_ = input_.transpose(
0, 2).contiguous() # shard_hn, seqlen, bs, hd
input_ = input_.reshape(shard_hn, world_size, shard_seqlen, bs,
hd).transpose(0, 1).reshape(
shard_hn * world_size, shard_seqlen,
bs, hd).contiguous()
output = torch.empty_like(input_)
dist.all_to_all_single(output, input_, group=group)
output = output.transpose(
0, 2).contiguous() # bs, seqlen, sharded_hn, hd
return output
else:
raise RuntimeError(
f"Invalid scatter_dim={scatter_dim}, gather_dim={gather_dim}. "
f"Only (scatter_dim=2, gather_dim=1) and (scatter_dim=1, gather_dim=2) are supported."
)
@staticmethod
def backward(
ctx: Any,
grad_output: Tensor) -> tuple[None, Tensor, None, None, None]:
if ctx.world_size == 1:
return None, grad_output, None, None, None
# For backward pass, we swap scatter_dim and gather_dim
output = DistributedAutograd.AllToAll4D.apply(
ctx.group, grad_output, ctx.world_size, ctx.gather_dim,
ctx.scatter_dim)
return None, output, None, None, None
class DeviceCommunicatorBase:
"""
Base class for device-specific communicator with autograd support.
It can use the `cpu_group` to initialize the communicator.
If the device has PyTorch integration (PyTorch can recognize its
communication backend), the `device_group` will also be given.
"""
def __init__(self,
cpu_group: ProcessGroup,
device: torch.device | None = None,
device_group: ProcessGroup | None = None,
unique_name: str = ""):
self.device = device or torch.device("cpu")
self.cpu_group = cpu_group
self.device_group = device_group
self.unique_name = unique_name
self.rank = dist.get_rank(cpu_group)
self.world_size = dist.get_world_size(cpu_group)
self.ranks = dist.get_process_group_ranks(cpu_group)
self.global_rank = dist.get_rank()
self.global_world_size = dist.get_world_size()
self.rank_in_group = dist.get_group_rank(self.cpu_group,
self.global_rank)
def all_reduce(self,
input_: torch.Tensor,
op: dist.ReduceOp | None = ReduceOp.SUM) -> torch.Tensor:
"""Performs an all_reduce operation with gradient support."""
return DistributedAutograd.AllReduce.apply(self.device_group, input_,
op)
def all_gather(self, input_: torch.Tensor, dim: int = -1) -> torch.Tensor:
"""Performs an all_gather operation with gradient support."""
if dim < 0:
dim += input_.dim()
return DistributedAutograd.AllGather.apply(self.device_group, input_,
self.world_size, dim)
def all_to_all_4D(self,
input_: torch.Tensor,
scatter_dim: int = 2,
gather_dim: int = 1) -> torch.Tensor:
"""Performs a 4D all-to-all operation with gradient support."""
return DistributedAutograd.AllToAll4D.apply(self.device_group, input_,
self.world_size,
scatter_dim, gather_dim)
def gather(self,
input_: torch.Tensor,
dst: int = 0,
dim: int = -1) -> torch.Tensor | None:
"""
NOTE: We assume that the input tensor is on the same device across
all the ranks.
NOTE: `dst` is the local rank of the destination rank.
"""
world_size = self.world_size
assert -input_.dim() <= dim < input_.dim(), (
f"Invalid dim ({dim}) for input tensor with shape {input_.size()}")
if dim < 0:
# Convert negative dim to positive.
dim += input_.dim()
# Allocate output tensor.
if self.rank_in_group == dst:
gather_list = [torch.empty_like(input_) for _ in range(world_size)]
else:
gather_list = None
# Gather.
torch.distributed.gather(input_,
gather_list,
dst=self.ranks[dst],
group=self.device_group)
if self.rank_in_group == dst:
output_tensor = torch.cat(gather_list, dim=dim)
else:
output_tensor = None
return output_tensor
def send(self, tensor: torch.Tensor, dst: int | None = None) -> None:
"""Sends a tensor to the destination rank in a non-blocking way"""
"""NOTE: `dst` is the local rank of the destination rank."""
if dst is None:
dst = (self.rank_in_group + 1) % self.world_size
torch.distributed.send(tensor, self.ranks[dst], self.device_group)
def recv(self,
size: torch.Size,
dtype: torch.dtype,
src: int | None = None) -> torch.Tensor:
"""Receives a tensor from the source rank."""
"""NOTE: `src` is the local rank of the source rank."""
if src is None:
src = (self.rank_in_group - 1) % self.world_size
tensor = torch.empty(size, dtype=dtype, device=self.device)
torch.distributed.recv(tensor, self.ranks[src], self.device_group)
return tensor
def destroy(self) -> None:
pass
|