File size: 10,902 Bytes
4135502 |
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 |
# -*- coding: utf-8 -*-
# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
from __future__ import annotations
from typing import TYPE_CHECKING, Optional, Tuple
import torch
import torch.nn as nn
import torch.nn.functional as F
from einops import rearrange, repeat
from fla.modules import FusedRMSNormGated, RMSNorm, ShortConvolution
from fla.modules.activations import ACT2FN
from fla.ops.simple_gla import chunk_simple_gla, fused_recurrent_simple_gla
if TYPE_CHECKING:
from fla.models.utils import Cache
class SimpleGatedLinearAttention(nn.Module):
r"""
The layer implementaion for [Gated Linear Attention Transformers with Hardware-Efficient Training](https://arxiv.org/abs/2312.06635). # noqa
This layer calls the simplified GLA kernel in which the gating is head-wise instead of elementwise.
Args:
mode (str, Optional):
Which GLA kernel to use.
Currently available: `chunk`.
Default: `chunk`.
hidden_size (int, Optional):
The hidden size of the input. Default: 1024.
expand_k (float, Optional):
The expansion ratio for the key dim. Default: 1.0.
expand_v (float, Optional):
The expansion ratio for the value dim. Default: 1.0.
num_heads (int, Optional):
The number of heads. Default: 4.
num_kv_heads (int, Optional):
The number of key/value heads, used for MQA. Default: None.
feature_map (str, Optional):
Feature map function applied to queries/keys. Default: None.
use_short_conv (bool, Optional):
Whether to use short convolutions. Default: `False`.
conv_size (int, Optional):
The kernel size of the short convolution, only used when `use_short_conv` is `True`. Default: 4.
conv_bias (bool, Optional):
Whether to use bias in the short convolution, only used when `use_short_conv` is `True`. Default: `False`.
gate_fn (str, Optional):
The activation function for the output gate. Default: `swish`.
elementwise_affine (bool, Optional):
If `True`, applies elementwise affine to LayerNorm with learnable parameters. Default: `True`.
norm_eps (float, Optional):
The epsilon value for the layernorm/rmsnorm layer. Default: 1e-5.
gate_logit_normalizer (int, Optional):
The normalizer for the gate logits, appied after `logsigmoid`. Default: 16.
fuse_norm (bool, Optional):
Whether to fuse the norm and the output gate for better memory footprint. Default: `True`.
layer_idx (int, Optional):
The index of the layer. Default: None.
"""
def __init__(
self,
mode: str = 'chunk',
hidden_size: int = 1024,
expand_k: float = 1.,
expand_v: float = 1.,
num_heads: int = 4,
num_kv_heads: Optional[int] = None,
feature_map: Optional[str] = None,
use_short_conv: bool = True,
conv_size: int = 4,
conv_bias: bool = False,
gate_fn: str = 'swish',
elementwise_affine: Optional[bool] = True,
norm_eps: float = 1e-5,
gate_logit_normalizer: int = 16,
fuse_norm: bool = True,
layer_idx: int = None,
) -> SimpleGatedLinearAttention:
super().__init__()
self.mode = mode
self.hidden_size = hidden_size
self.expand_k = expand_k
self.expand_v = expand_v
self.num_heads = num_heads
self.num_kv_heads = num_kv_heads if num_kv_heads is not None else num_heads
self.num_kv_groups = self.num_heads // self.num_kv_heads
self.feature_map_fn = ACT2FN[feature_map] if feature_map is not None else None
self.use_short_conv = use_short_conv
self.conv_size = conv_size
self.conv_bias = conv_bias
self.key_dim = int(hidden_size * expand_k)
self.value_dim = int(hidden_size * expand_v)
self.key_dim_per_group = self.key_dim // self.num_kv_groups
self.value_dim_per_group = self.value_dim // self.num_kv_groups
self.layer_idx = layer_idx
assert mode in ['chunk', "fused_recurrent"], f"Not suppoerted mode `{mode}`."
assert self.key_dim % num_heads == 0, f"key dim must be divisible by num_heads of {num_heads}"
assert self.value_dim % num_heads == 0, f"value dim must be divisible by num_heads of {num_heads}"
self.head_k_dim = self.key_dim // num_heads
self.head_v_dim = self.value_dim // num_heads
self.q_proj = nn.Linear(hidden_size, self.key_dim, bias=False)
self.k_proj = nn.Linear(hidden_size, self.key_dim_per_group, bias=False)
self.v_proj = nn.Linear(hidden_size, self.value_dim_per_group, bias=False)
self.g_proj = nn.Linear(hidden_size, self.value_dim, bias=False)
if use_short_conv:
self.conv_size = conv_size
self.q_conv1d = ShortConvolution(self.key_dim, conv_size, activation='silu')
self.k_conv1d = ShortConvolution(self.key_dim_per_group, conv_size, activation='silu')
self.v_conv1d = ShortConvolution(self.value_dim_per_group, conv_size, activation='silu')
self.gk_proj = nn.Linear(hidden_size, self.num_heads)
if gate_fn == 'swish' and fuse_norm:
self.g_norm_swish_gate = FusedRMSNormGated(
hidden_size=self.head_v_dim,
elementwise_affine=elementwise_affine,
eps=norm_eps
)
self.fuse_norm_and_gate = True
else:
self.fuse_norm_and_gate = False
self.g_norm = RMSNorm(
hidden_size=self.head_v_dim,
elementwise_affine=elementwise_affine,
eps=norm_eps
)
self.gate_fn = ACT2FN[gate_fn]
self.o_proj = nn.Linear(self.value_dim, hidden_size, bias=False)
self.gate_logit_normalizer = gate_logit_normalizer
def forward(
self,
hidden_states: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
past_key_values: Optional[Cache] = None,
use_cache: Optional[bool] = False,
output_attentions: Optional[bool] = False,
**kwargs
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Cache]]:
if attention_mask is not None:
assert len(attention_mask.shape) == 2, (
"Expected attention_mask as a 0-1 matrix with shape [batch_size, seq_len] "
"for padding purposes (0 indicating padding). "
"Arbitrary attention masks of shape [batch_size, seq_len, seq_len] are not allowed."
)
# launching the triton kernel for just one token will actually be slower
mode = 'fused_recurrent' if hidden_states.shape[1] <= 64 else self.mode
last_state = None
if past_key_values is not None and len(past_key_values) > self.layer_idx:
last_state = past_key_values[self.layer_idx]
cu_seqlens = kwargs.get('cu_seqlens', None)
if self.use_short_conv:
conv_state_q, conv_state_k, conv_state_v = None, None, None
if last_state is not None:
conv_state_q, conv_state_k, conv_state_v = last_state['conv_state']
conv_mask = attention_mask[:, -hidden_states.shape[1]:] if attention_mask is not None else None
q, conv_state_q = self.q_conv1d(
x=self.q_proj(hidden_states),
mask=conv_mask,
cache=conv_state_q,
output_final_state=use_cache,
cu_seqlens=cu_seqlens
)
k, conv_state_k = self.k_conv1d(
x=self.k_proj(hidden_states),
mask=conv_mask,
cache=conv_state_k,
output_final_state=use_cache,
cu_seqlens=cu_seqlens
)
v, conv_state_v = self.v_conv1d(
x=self.v_proj(hidden_states),
mask=conv_mask,
cache=conv_state_v,
output_final_state=use_cache,
cu_seqlens=cu_seqlens
)
else:
q = self.q_proj(hidden_states)
k = self.k_proj(hidden_states)
v = self.v_proj(hidden_states)
gk = self.gk_proj(hidden_states)
if self.feature_map_fn is not None:
q, k = map(self.feature_map_fn, (q, k))
# dealing with left-padding
if attention_mask is not None:
v = v.mul_(attention_mask[:, -v.shape[-2]:, None])
q = rearrange(q, '... (h d) -> ... h d', h=self.num_heads)
if self.num_kv_groups > 1:
k, v = (repeat(x, '... (h d) -> ... (h g) d', h=self.num_kv_heads, g=self.num_kv_groups) for x in (k, v))
else:
k, v = (rearrange(x, '... (h d) -> ... h d', h=self.num_kv_heads) for x in (k, v))
gk = F.logsigmoid(gk) / self.gate_logit_normalizer
recurrent_state = last_state['recurrent_state'] if last_state is not None else None
if mode == 'chunk':
o, recurrent_state = chunk_simple_gla(
q=q,
k=k,
v=v,
gk=gk,
initial_state=recurrent_state,
output_final_state=use_cache,
cu_seqlens=cu_seqlens,
head_first=False
)
elif mode == 'fused_recurrent':
o, recurrent_state = fused_recurrent_simple_gla(
q=q,
k=k,
v=v,
gk=gk,
initial_state=recurrent_state,
output_final_state=use_cache,
cu_seqlens=cu_seqlens,
head_first=False
)
else:
raise NotImplementedError(f"Not supported mode `{mode}`.")
if past_key_values is not None:
past_key_values.update(
recurrent_state=recurrent_state,
conv_state=(conv_state_q, conv_state_k, conv_state_v) if self.use_short_conv else None,
layer_idx=self.layer_idx,
offset=q.shape[1]
)
g = self.g_proj(hidden_states)
if self.fuse_norm_and_gate:
g = rearrange(g, 'b t (h d) -> b t h d', h=self.num_heads)
o = self.g_norm_swish_gate(o, g)
o = rearrange(o, 'b t h d -> b t (h d)')
else:
o = rearrange(self.g_norm(o), 'b t h d -> b t (h d)')
o = o * self.gate_fn(g)
o = self.o_proj(o)
return o, None, past_key_values
def state_size(self, **kwargs) -> int:
state_size = self.key_dim * self.head_v_dim
for module in self.children():
if isinstance(module, ShortConvolution):
state_size += module.state_size
return state_size
|