File size: 5,071 Bytes
b66f552
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang

from __future__ import annotations

from functools import partial
from typing import TYPE_CHECKING, Any

import torch
import torch.nn as nn
try:
    from torch.distributed import DeviceMesh
except ImportError:
    DeviceMesh = None
try:
    from torch.distributed.tensor import Placement, Replicate, Shard, distribute_module
except ImportError:
    Placement = None
    Replicate = None
    Shard = None
    distribute_module = None
try:
    from torch.distributed.tensor.parallel import ParallelStyle
except ImportError:
    class ParallelStyle:
        pass

from fla.modules.activations import swiglu, swiglu_linear

try:
    from torch.distributed.tensor import DTensor
except (ImportError, AttributeError):
    DTensor = None

if TYPE_CHECKING:
    from transformers.processing_utils import Unpack


class GatedMLP(nn.Module):

    def __init__(
        self,
        hidden_size: int,
        hidden_ratio: int | None = None,
        intermediate_size: int | None = None,
        hidden_act: str = 'swish',
        fuse_swiglu: bool = True,
    ) -> GatedMLP:
        super().__init__()

        self.hidden_size = hidden_size
        # the final number of params is `hidden_ratio * hidden_size^2`
        # `intermediate_size` is chosen to be a multiple of 256 closest to `2/3 * hidden_size * hidden_ratio`
        if hidden_ratio is None:
            hidden_ratio = 4
        if intermediate_size is None:
            intermediate_size = int(hidden_size * hidden_ratio * 2 / 3)
            intermediate_size = 256 * ((intermediate_size + 256 - 1) // 256)
        self.hidden_ratio = hidden_ratio
        self.intermediate_size = intermediate_size
        self.hidden_act = hidden_act
        self.fuse_swiglu = fuse_swiglu

        if hidden_act != 'swish':
            raise ValueError(f'Unsupported hidden_act: {hidden_act}')

        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
        if self.fuse_swiglu:
            self.swiglu_linear = SwiGLULinear()

    def forward(
        self,
        x: torch.Tensor,
        **kwargs: Unpack[Any],
    ) -> torch.Tensor:
        gate, y = self.gate_proj(x), self.up_proj(x)
        if self.fuse_swiglu:
            return self.swiglu_linear(gate, y, self.down_proj.weight, self.down_proj.bias)
        else:
            return self.down_proj(swiglu(gate, y))


class SwiGLULinear(nn.Module):

    def forward(self, x, y, weight, bias):
        return swiglu_linear(x, y, weight, bias)


class SwiGLULinearParallel(ParallelStyle):
    def __init__(
        self,
        *,
        input_layouts: Placement | None = None,
        output_layouts: Placement | None = None,
        use_local_output: bool = True,
    ):
        super().__init__()
        self.input_layouts = (input_layouts or Shard(-1),)
        self.output_layouts = (output_layouts or Replicate(),)
        self.desired_input_layouts = (Shard(-1),)
        self.use_local_output = use_local_output

    @staticmethod
    def _prepare_input_fn(
        input_layouts, desired_input_layouts, mod, inputs, device_mesh,
    ):
        x, y, weight, bias = inputs
        if not isinstance(x, DTensor):
            x = DTensor.from_local(x, device_mesh, input_layouts, run_check=False)
        if x.placements != desired_input_layouts:
            x = x.redistribute(placements=desired_input_layouts, async_op=True)

        if not isinstance(y, DTensor):
            y = DTensor.from_local(y, device_mesh, input_layouts, run_check=False)
        if y.placements != desired_input_layouts:
            y = y.redistribute(placements=desired_input_layouts, async_op=True)

        if not isinstance(weight, DTensor):
            weight = DTensor.from_local(weight, device_mesh, (Shard(1),))

        if bias is not None and not isinstance(bias, DTensor):
            bias = DTensor.from_local(bias, device_mesh, (Replicate(),))

        return x, y, weight, bias

    @staticmethod
    def _prepare_output_fn(output_layouts, use_local_output, mod, outputs, device_mesh):
        # Rowwise sharding produces partial output, depending on output layouts:
        # 1. to replicate -> allreduce
        # 2. to shard -> reduce_scatter
        if outputs.placements != output_layouts:
            outputs = outputs.redistribute(placements=output_layouts, async_op=True)
        # back to local tensor if use_local_output is True
        return outputs.to_local() if use_local_output else outputs

    def _apply(self, module: nn.Module, device_mesh: DeviceMesh) -> nn.Module:
        return distribute_module(
            module,
            device_mesh,
            partition_fn=None,
            input_fn=partial(self._prepare_input_fn, self.input_layouts, self.desired_input_layouts),
            output_fn=partial(self._prepare_output_fn, self.output_layouts, self.use_local_output),
        )