jupyterjazz
commited on
Commit
•
e3681c2
1
Parent(s):
ab85772
rope-embeddings (#20)
Browse files- feat: support rope (f2e0e6205e15f5e1e23354800fe39b3ae25d9bca)
- chore: remove parallelmha (8b64fa835d32d29efa8b3a34f43e4c5011ec6f13)
- feat: default dim (77a17f7cb5d03e4bbfd8674fd49956fdcebb9cdc)
- refactor: revert alibi stuff (11ba2000440ef53e3b8ad551ababcdf2259643ed)
- chore: source (c232c27ed971d786c3e9cad242d415b2cd1fa655)
- refactor: raise error if flash attention is not installed (e8e1e150c525847c940918fa1f849997e14fcaa9)
- mha.py +14 -12
- modeling_xlm_roberta.py +2 -2
- rotary.py +575 -0
mha.py
CHANGED
@@ -1,7 +1,5 @@
|
|
1 |
-
# This implementation was adapted from https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/modules/mha.py
|
2 |
-
# Commit id: 6bbc532388e61185a92e2a563126739967b4c8c5
|
3 |
-
|
4 |
# Copyright (c) 2023, Tri Dao.
|
|
|
5 |
|
6 |
import math
|
7 |
from functools import partial
|
@@ -28,10 +26,7 @@ try:
|
|
28 |
except ImportError:
|
29 |
FusedDense, ColumnParallelLinear, RowParallelLinear = None, None, None
|
30 |
|
31 |
-
|
32 |
-
from flash_attn.layers.rotary import RotaryEmbedding
|
33 |
-
except ImportError:
|
34 |
-
RotaryEmbedding = None
|
35 |
|
36 |
|
37 |
# From https://github.com/ofirpress/attention_with_linear_biases/blob/4b92f28a005ead2567abe2359f633e73e08f3833/fairseq/models/transformer.py#L742
|
@@ -619,7 +614,6 @@ class MHA(nn.Module):
|
|
619 |
assert key_padding_mask is None
|
620 |
assert self.use_flash_attn
|
621 |
assert not self.dwconv
|
622 |
-
assert self.rotary_emb_dim == 0
|
623 |
if key_padding_mask is not None:
|
624 |
assert cu_seqlens is None
|
625 |
assert max_seqlen is None
|
@@ -643,7 +637,9 @@ class MHA(nn.Module):
|
|
643 |
else inference_params.seqlen_offset
|
644 |
)
|
645 |
)
|
646 |
-
rotary_max_seqlen =
|
|
|
|
|
647 |
batch, seqlen = x.shape[:2]
|
648 |
if not self.cross_attn and self.num_heads_kv == self.num_heads:
|
649 |
assert x_kv is None and mixer_subset is None
|
@@ -664,7 +660,10 @@ class MHA(nn.Module):
|
|
664 |
):
|
665 |
if self.rotary_emb_dim > 0:
|
666 |
qkv = self.rotary_emb(
|
667 |
-
qkv,
|
|
|
|
|
|
|
668 |
)
|
669 |
if inference_params is None:
|
670 |
if not self.checkpointing:
|
@@ -715,7 +714,11 @@ class MHA(nn.Module):
|
|
715 |
):
|
716 |
if self.rotary_emb_dim > 0:
|
717 |
q, kv = self.rotary_emb(
|
718 |
-
q,
|
|
|
|
|
|
|
|
|
719 |
)
|
720 |
if inference_params is None:
|
721 |
if not self.checkpointing:
|
@@ -730,4 +733,3 @@ class MHA(nn.Module):
|
|
730 |
context = self._apply_rotary_update_kvcache_attention(q, kv, inference_params)
|
731 |
out = self.out_proj(rearrange(context, "... h d -> ... (h d)"))
|
732 |
return out if not self.return_residual else (out, x)
|
733 |
-
|
|
|
|
|
|
|
|
|
1 |
# Copyright (c) 2023, Tri Dao.
|
2 |
+
# Adapted from https://github.com/Dao-AILab/flash-attention/pull/556
|
3 |
|
4 |
import math
|
5 |
from functools import partial
|
|
|
26 |
except ImportError:
|
27 |
FusedDense, ColumnParallelLinear, RowParallelLinear = None, None, None
|
28 |
|
29 |
+
from .rotary import RotaryEmbedding
|
|
|
|
|
|
|
30 |
|
31 |
|
32 |
# From https://github.com/ofirpress/attention_with_linear_biases/blob/4b92f28a005ead2567abe2359f633e73e08f3833/fairseq/models/transformer.py#L742
|
|
|
614 |
assert key_padding_mask is None
|
615 |
assert self.use_flash_attn
|
616 |
assert not self.dwconv
|
|
|
617 |
if key_padding_mask is not None:
|
618 |
assert cu_seqlens is None
|
619 |
assert max_seqlen is None
|
|
|
637 |
else inference_params.seqlen_offset
|
638 |
)
|
639 |
)
|
640 |
+
rotary_max_seqlen = (
|
641 |
+
inference_params.max_sequence_len if inference_params is not None else max_seqlen
|
642 |
+
)
|
643 |
batch, seqlen = x.shape[:2]
|
644 |
if not self.cross_attn and self.num_heads_kv == self.num_heads:
|
645 |
assert x_kv is None and mixer_subset is None
|
|
|
660 |
):
|
661 |
if self.rotary_emb_dim > 0:
|
662 |
qkv = self.rotary_emb(
|
663 |
+
qkv,
|
664 |
+
seqlen_offset=seqlen_offset,
|
665 |
+
cu_seqlens=cu_seqlens,
|
666 |
+
max_seqlen=rotary_max_seqlen,
|
667 |
)
|
668 |
if inference_params is None:
|
669 |
if not self.checkpointing:
|
|
|
714 |
):
|
715 |
if self.rotary_emb_dim > 0:
|
716 |
q, kv = self.rotary_emb(
|
717 |
+
q,
|
718 |
+
kv,
|
719 |
+
seqlen_offset=seqlen_offset,
|
720 |
+
cu_seqlens=cu_seqlens,
|
721 |
+
max_seqlen=rotary_max_seqlen,
|
722 |
)
|
723 |
if inference_params is None:
|
724 |
if not self.checkpointing:
|
|
|
733 |
context = self._apply_rotary_update_kvcache_attention(q, kv, inference_params)
|
734 |
out = self.out_proj(rearrange(context, "... h d -> ... (h d)"))
|
735 |
return out if not self.return_residual else (out, x)
|
|
modeling_xlm_roberta.py
CHANGED
@@ -45,7 +45,7 @@ from .embedding import XLMRobertaEmbeddings
|
|
45 |
from .mha import MHA
|
46 |
from .mlp import FusedMLP, Mlp
|
47 |
from .stochastic_depth import StochasticDepth
|
48 |
-
|
49 |
|
50 |
try:
|
51 |
from flash_attn.ops.fused_dense import FusedDense
|
@@ -91,7 +91,7 @@ def create_mixer_cls(config, cross_attn=False, return_residual=False):
|
|
91 |
rotary_kwargs = {}
|
92 |
if config.position_embedding_type == "rotary":
|
93 |
rotary_kwargs["rotary_emb_dim"] = getattr(
|
94 |
-
config, "rotary_emb_dim", config.hidden_size
|
95 |
)
|
96 |
rotary_kwargs["rotary_emb_base"] = getattr(config, "rotary_emb_base", 10000.0)
|
97 |
rotary_kwargs["rotary_emb_scale_base"] = getattr(
|
|
|
45 |
from .mha import MHA
|
46 |
from .mlp import FusedMLP, Mlp
|
47 |
from .stochastic_depth import StochasticDepth
|
48 |
+
from .rotary import RotaryEmbedding
|
49 |
|
50 |
try:
|
51 |
from flash_attn.ops.fused_dense import FusedDense
|
|
|
91 |
rotary_kwargs = {}
|
92 |
if config.position_embedding_type == "rotary":
|
93 |
rotary_kwargs["rotary_emb_dim"] = getattr(
|
94 |
+
config, "rotary_emb_dim", config.hidden_size / config.num_attention_heads
|
95 |
)
|
96 |
rotary_kwargs["rotary_emb_base"] = getattr(config, "rotary_emb_base", 10000.0)
|
97 |
rotary_kwargs["rotary_emb_scale_base"] = getattr(
|
rotary.py
ADDED
@@ -0,0 +1,575 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Adapted from https://github.com/Dao-AILab/flash-attention/pull/556
|
2 |
+
# Copyright (c) 2023, Tri Dao.
|
3 |
+
|
4 |
+
import math
|
5 |
+
from typing import Optional, Tuple, Union
|
6 |
+
|
7 |
+
import torch
|
8 |
+
from einops import rearrange, repeat
|
9 |
+
try:
|
10 |
+
from flash_attn.ops.triton.rotary import apply_rotary
|
11 |
+
except ImportError:
|
12 |
+
def apply_rotary(*args, **kwargs):
|
13 |
+
raise RuntimeError('RoPE requires flash-attention to be installed')
|
14 |
+
|
15 |
+
|
16 |
+
def rotate_half(x, interleaved=False):
|
17 |
+
if not interleaved:
|
18 |
+
x1, x2 = x.chunk(2, dim=-1)
|
19 |
+
return torch.cat((-x2, x1), dim=-1)
|
20 |
+
else:
|
21 |
+
x1, x2 = x[..., ::2], x[..., 1::2]
|
22 |
+
return rearrange(torch.stack((-x2, x1), dim=-1), "... d two -> ... (d two)", two=2)
|
23 |
+
|
24 |
+
|
25 |
+
def apply_rotary_emb_torch(x, cos, sin, interleaved=False):
|
26 |
+
"""
|
27 |
+
x: (batch_size, seqlen, nheads, headdim)
|
28 |
+
cos, sin: (seqlen, rotary_dim / 2) or (batch_size, seqlen, rotary_dim / 2)
|
29 |
+
"""
|
30 |
+
ro_dim = cos.shape[-1] * 2
|
31 |
+
assert ro_dim <= x.shape[-1]
|
32 |
+
cos = repeat(cos, "... d -> ... 1 (2 d)" if not interleaved else "... d -> ... 1 (d 2)")
|
33 |
+
sin = repeat(sin, "... d -> ... 1 (2 d)" if not interleaved else "... d -> ... 1 (d 2)")
|
34 |
+
return torch.cat(
|
35 |
+
[x[..., :ro_dim] * cos + rotate_half(x[..., :ro_dim], interleaved) * sin, x[..., ro_dim:]],
|
36 |
+
dim=-1,
|
37 |
+
)
|
38 |
+
|
39 |
+
|
40 |
+
class ApplyRotaryEmb(torch.autograd.Function):
|
41 |
+
@staticmethod
|
42 |
+
def forward(
|
43 |
+
ctx,
|
44 |
+
x,
|
45 |
+
cos,
|
46 |
+
sin,
|
47 |
+
interleaved=False,
|
48 |
+
inplace=False,
|
49 |
+
seqlen_offsets: Union[int, torch.Tensor] = 0,
|
50 |
+
cu_seqlens: Optional[torch.Tensor] = None,
|
51 |
+
max_seqlen: Optional[int] = None,
|
52 |
+
):
|
53 |
+
out = apply_rotary(
|
54 |
+
x,
|
55 |
+
cos,
|
56 |
+
sin,
|
57 |
+
seqlen_offsets=seqlen_offsets,
|
58 |
+
cu_seqlens=cu_seqlens,
|
59 |
+
max_seqlen=max_seqlen,
|
60 |
+
interleaved=interleaved,
|
61 |
+
inplace=inplace,
|
62 |
+
)
|
63 |
+
if isinstance(seqlen_offsets, int):
|
64 |
+
ctx.save_for_backward(cos, sin, cu_seqlens) # Can't save int with save_for_backward
|
65 |
+
ctx.seqlen_offsets = seqlen_offsets
|
66 |
+
else:
|
67 |
+
ctx.save_for_backward(cos, sin, cu_seqlens, seqlen_offsets)
|
68 |
+
ctx.seqlen_offsets = None
|
69 |
+
ctx.interleaved = interleaved
|
70 |
+
ctx.inplace = inplace
|
71 |
+
ctx.max_seqlen = max_seqlen
|
72 |
+
return out if not inplace else x
|
73 |
+
|
74 |
+
@staticmethod
|
75 |
+
def backward(ctx, do):
|
76 |
+
seqlen_offsets = ctx.seqlen_offsets
|
77 |
+
if seqlen_offsets is None:
|
78 |
+
cos, sin, cu_seqlens, seqlen_offsets = ctx.saved_tensors
|
79 |
+
else:
|
80 |
+
cos, sin, cu_seqlens = ctx.saved_tensors
|
81 |
+
# TD [2023-09-02]: For some reason Triton (2.0.0.post1) errors with
|
82 |
+
# "[CUDA]: invalid device context", and cloning makes it work. Idk why. Triton 2.1.0 works.
|
83 |
+
if not ctx.interleaved and not ctx.inplace:
|
84 |
+
do = do.clone()
|
85 |
+
dx = apply_rotary(
|
86 |
+
do,
|
87 |
+
cos,
|
88 |
+
sin,
|
89 |
+
seqlen_offsets=seqlen_offsets,
|
90 |
+
cu_seqlens=cu_seqlens,
|
91 |
+
max_seqlen=ctx.max_seqlen,
|
92 |
+
interleaved=ctx.interleaved,
|
93 |
+
inplace=ctx.inplace,
|
94 |
+
conjugate=True,
|
95 |
+
)
|
96 |
+
return dx, None, None, None, None, None, None, None
|
97 |
+
|
98 |
+
|
99 |
+
def apply_rotary_emb(
|
100 |
+
x,
|
101 |
+
cos,
|
102 |
+
sin,
|
103 |
+
interleaved=False,
|
104 |
+
inplace=False,
|
105 |
+
seqlen_offsets: Union[int, torch.Tensor] = 0,
|
106 |
+
cu_seqlens: Optional[torch.Tensor] = None,
|
107 |
+
max_seqlen: Optional[int] = None,
|
108 |
+
):
|
109 |
+
"""
|
110 |
+
Arguments:
|
111 |
+
x: (batch_size, seqlen, nheads, headdim) if cu_seqlens is None
|
112 |
+
else (total_seqlen, nheads, headdim)
|
113 |
+
cos, sin: (seqlen_rotary, rotary_dim / 2)
|
114 |
+
interleaved: if True, rotate pairs of even and odd dimensions (GPT-J style) instead
|
115 |
+
of 1st half and 2nd half (GPT-NeoX style).
|
116 |
+
inplace: if True, apply rotary embedding in-place.
|
117 |
+
seqlen_offsets: (batch_size,) or int. Each sequence in x is shifted by this amount.
|
118 |
+
Most commonly used in inference when we have KV cache.
|
119 |
+
cu_seqlens: (batch + 1,) or None
|
120 |
+
max_seqlen: int
|
121 |
+
Return:
|
122 |
+
out: (batch_size, seqlen, nheads, headdim) if cu_seqlens is None
|
123 |
+
else (total_seqlen, nheads, headdim)
|
124 |
+
rotary_dim must be <= headdim
|
125 |
+
Apply rotary embedding to the first rotary_dim of x.
|
126 |
+
"""
|
127 |
+
return ApplyRotaryEmb.apply(
|
128 |
+
x, cos, sin, interleaved, inplace, seqlen_offsets, cu_seqlens, max_seqlen
|
129 |
+
)
|
130 |
+
|
131 |
+
|
132 |
+
# For backward compatibility
|
133 |
+
apply_rotary_emb_func = apply_rotary_emb
|
134 |
+
|
135 |
+
|
136 |
+
class ApplyRotaryEmbQKV_(torch.autograd.Function):
|
137 |
+
@staticmethod
|
138 |
+
def forward(
|
139 |
+
ctx,
|
140 |
+
qkv,
|
141 |
+
cos,
|
142 |
+
sin,
|
143 |
+
cos_k=None,
|
144 |
+
sin_k=None,
|
145 |
+
interleaved=False,
|
146 |
+
seqlen_offsets: Union[int, torch.Tensor] = 0,
|
147 |
+
cu_seqlens: Optional[torch.Tensor] = None,
|
148 |
+
max_seqlen: Optional[int] = None,
|
149 |
+
):
|
150 |
+
# batch, seqlen, three, nheads, headdim = qkv.shape
|
151 |
+
assert qkv.shape[-3] == 3
|
152 |
+
if cos_k is None and sin_k is None and qkv.is_contiguous():
|
153 |
+
# Call 1 kernel instead of 2 kernels
|
154 |
+
# We need qkv to be contiguous so that when we reshape to combine (3, nheads)
|
155 |
+
# dimensions, we get the same tensor
|
156 |
+
qk = rearrange(qkv[..., :2, :, :], "... t h d -> ... (t h) d")
|
157 |
+
# qk = qkv[:, :, :2].reshape(batch, seqlen, -1, headdim)
|
158 |
+
apply_rotary(
|
159 |
+
qk,
|
160 |
+
cos,
|
161 |
+
sin,
|
162 |
+
seqlen_offsets=seqlen_offsets,
|
163 |
+
interleaved=interleaved,
|
164 |
+
inplace=True,
|
165 |
+
cu_seqlens=cu_seqlens,
|
166 |
+
max_seqlen=max_seqlen,
|
167 |
+
)
|
168 |
+
else:
|
169 |
+
cos_k = cos if cos_k is None else cos_k
|
170 |
+
sin_k = sin if sin_k is None else sin_k
|
171 |
+
q, k = qkv[..., 0, :, :], qkv[..., 1, :, :]
|
172 |
+
apply_rotary(
|
173 |
+
q,
|
174 |
+
cos,
|
175 |
+
sin,
|
176 |
+
seqlen_offsets,
|
177 |
+
interleaved=interleaved,
|
178 |
+
inplace=True,
|
179 |
+
cu_seqlens=cu_seqlens,
|
180 |
+
max_seqlen=max_seqlen,
|
181 |
+
)
|
182 |
+
apply_rotary(
|
183 |
+
k,
|
184 |
+
cos_k,
|
185 |
+
sin_k,
|
186 |
+
seqlen_offsets,
|
187 |
+
interleaved=interleaved,
|
188 |
+
inplace=True,
|
189 |
+
cu_seqlens=cu_seqlens,
|
190 |
+
max_seqlen=max_seqlen,
|
191 |
+
)
|
192 |
+
ctx.save_for_backward(cos, sin, cos_k, sin_k)
|
193 |
+
if isinstance(seqlen_offsets, int):
|
194 |
+
ctx.save_for_backward(cos, sin, cos_k, sin_k, cu_seqlens)
|
195 |
+
ctx.seqlen_offsets = seqlen_offsets
|
196 |
+
else:
|
197 |
+
ctx.save_for_backward(cos, sin, cos_k, sin_k, cu_seqlens, seqlen_offsets)
|
198 |
+
ctx.seqlen_offsets = None
|
199 |
+
ctx.max_seqlen = max_seqlen
|
200 |
+
ctx.interleaved = interleaved
|
201 |
+
return qkv
|
202 |
+
|
203 |
+
@staticmethod
|
204 |
+
def backward(ctx, dqkv):
|
205 |
+
seqlen_offsets = ctx.seqlen_offsets
|
206 |
+
if seqlen_offsets is None:
|
207 |
+
cos, sin, cos_k, sin_k, cu_seqlens, seqlen_offsets = ctx.saved_tensors
|
208 |
+
else:
|
209 |
+
cos, sin, cos_k, sin_k, cu_seqlens = ctx.saved_tensors
|
210 |
+
if cos_k is None and sin_k is None and dqkv.is_contiguous():
|
211 |
+
# Call 1 kernel instead of 2 kernels
|
212 |
+
# We need dqkv to be contiguous so that when we reshape to combine (3, nheads)
|
213 |
+
# dimensions, we get the same tensor
|
214 |
+
dqk = rearrange(dqkv[..., :2, :, :], "... t h d -> ... (t h) d")
|
215 |
+
apply_rotary(
|
216 |
+
dqk,
|
217 |
+
cos,
|
218 |
+
sin,
|
219 |
+
seqlen_offsets=seqlen_offsets,
|
220 |
+
interleaved=ctx.interleaved,
|
221 |
+
inplace=True,
|
222 |
+
conjugate=True,
|
223 |
+
cu_seqlens=cu_seqlens,
|
224 |
+
max_seqlen=ctx.max_seqlen,
|
225 |
+
)
|
226 |
+
else:
|
227 |
+
cos_k = cos if cos_k is None else cos_k
|
228 |
+
sin_k = sin if sin_k is None else sin_k
|
229 |
+
dq, dk = dqkv[..., 0, :, :], dqkv[..., 1, :, :]
|
230 |
+
apply_rotary(
|
231 |
+
|
232 |
+
dq,
|
233 |
+
cos,
|
234 |
+
sin,
|
235 |
+
seqlen_offsets,
|
236 |
+
interleaved=ctx.interleaved,
|
237 |
+
inplace=True,
|
238 |
+
conjugate=True,
|
239 |
+
cu_seqlens=cu_seqlens,
|
240 |
+
max_seqlen=ctx.max_seqlen,
|
241 |
+
)
|
242 |
+
apply_rotary(
|
243 |
+
dk,
|
244 |
+
cos_k,
|
245 |
+
sin_k,
|
246 |
+
seqlen_offsets,
|
247 |
+
interleaved=ctx.interleaved,
|
248 |
+
inplace=True,
|
249 |
+
conjugate=True,
|
250 |
+
cu_seqlens=cu_seqlens,
|
251 |
+
max_seqlen=ctx.max_seqlen,
|
252 |
+
)
|
253 |
+
return dqkv, None, None, None, None, None, None, None, None
|
254 |
+
|
255 |
+
|
256 |
+
def apply_rotary_emb_qkv_(
|
257 |
+
qkv,
|
258 |
+
cos,
|
259 |
+
sin,
|
260 |
+
cos_k=None,
|
261 |
+
sin_k=None,
|
262 |
+
interleaved=False,
|
263 |
+
seqlen_offsets: Union[int, torch.Tensor] = 0,
|
264 |
+
cu_seqlens: Optional[torch.Tensor] = None,
|
265 |
+
max_seqlen: Optional[int] = None,
|
266 |
+
):
|
267 |
+
"""
|
268 |
+
Arguments:
|
269 |
+
qkv: (batch_size, seqlen, 3, nheads, headdim) if cu_seqlens is None
|
270 |
+
else (total_seqlen, 3, nheads, headdim)
|
271 |
+
cos, sin: (seqlen, rotary_dim / 2)
|
272 |
+
cos_k, sin_k: (seqlen, rotary_dim / 2), optional
|
273 |
+
interleaved: if True, rotate pairs of even and odd dimensions (GPT-J style) instead of
|
274 |
+
1st half and 2nd half (GPT-NeoX style).
|
275 |
+
seqlen_offsets: (batch_size,) or int. Each sequence in Q and K is shifted by this amount.
|
276 |
+
Most commonly used in inference when we have KV cache.
|
277 |
+
cu_seqlens: (batch + 1,) or None
|
278 |
+
max_seqlen: int
|
279 |
+
Return:
|
280 |
+
qkv: (batch_size, seqlen, 3, nheads, headdim) if cu_seqlens is None
|
281 |
+
else (total_seqlen, 3, nheads, headdim)
|
282 |
+
rotary_dim must be <= headdim
|
283 |
+
Apply rotary embedding *inplace* to the first rotary_dim of Q and K.
|
284 |
+
"""
|
285 |
+
return ApplyRotaryEmbQKV_.apply(
|
286 |
+
qkv, cos, sin, cos_k, sin_k, interleaved, seqlen_offsets, cu_seqlens, max_seqlen
|
287 |
+
)
|
288 |
+
|
289 |
+
|
290 |
+
class ApplyRotaryEmbKV_(torch.autograd.Function):
|
291 |
+
@staticmethod
|
292 |
+
def forward(
|
293 |
+
ctx,
|
294 |
+
kv,
|
295 |
+
cos,
|
296 |
+
sin,
|
297 |
+
interleaved=False,
|
298 |
+
seqlen_offsets: Union[int, torch.Tensor] = 0,
|
299 |
+
cu_seqlens: Optional[torch.Tensor] = None,
|
300 |
+
max_seqlen: Optional[int] = None,
|
301 |
+
):
|
302 |
+
# batch, seqlen, two, nheads, headdim = kv.shape
|
303 |
+
assert kv.shape[-3] == 2
|
304 |
+
k = kv[..., 0, :, :]
|
305 |
+
apply_rotary(
|
306 |
+
k,
|
307 |
+
cos,
|
308 |
+
sin,
|
309 |
+
seqlen_offsets=seqlen_offsets,
|
310 |
+
interleaved=interleaved,
|
311 |
+
inplace=True,
|
312 |
+
cu_seqlens=cu_seqlens,
|
313 |
+
max_seqlen=max_seqlen,
|
314 |
+
)
|
315 |
+
if isinstance(seqlen_offsets, int):
|
316 |
+
ctx.save_for_backward(cos, sin, cu_seqlens) # Can't save int with save_for_backward
|
317 |
+
ctx.seqlen_offsets = seqlen_offsets
|
318 |
+
else:
|
319 |
+
ctx.save_for_backward(cos, sin, cu_seqlens, seqlen_offsets)
|
320 |
+
ctx.seqlen_offsets = None
|
321 |
+
ctx.max_seqlen = max_seqlen
|
322 |
+
ctx.interleaved = interleaved
|
323 |
+
return kv
|
324 |
+
|
325 |
+
@staticmethod
|
326 |
+
def backward(ctx, dkv):
|
327 |
+
seqlen_offsets = ctx.seqlen_offsets
|
328 |
+
if seqlen_offsets is None:
|
329 |
+
cos, sin, cu_seqlens, seqlen_offsets = ctx.saved_tensors
|
330 |
+
else:
|
331 |
+
cos, sin, cu_seqlens = ctx.saved_tensors
|
332 |
+
apply_rotary(
|
333 |
+
dkv[..., 0, :, :],
|
334 |
+
cos,
|
335 |
+
sin,
|
336 |
+
seqlen_offsets=seqlen_offsets,
|
337 |
+
interleaved=ctx.interleaved,
|
338 |
+
inplace=True,
|
339 |
+
conjugate=True,
|
340 |
+
cu_seqlens=cu_seqlens,
|
341 |
+
max_seqlen=ctx.max_seqlen,
|
342 |
+
)
|
343 |
+
return dkv, None, None, None, None, None, None
|
344 |
+
|
345 |
+
|
346 |
+
apply_rotary_emb_kv_ = ApplyRotaryEmbKV_.apply
|
347 |
+
|
348 |
+
|
349 |
+
def apply_rotary_emb_kv_(
|
350 |
+
kv,
|
351 |
+
cos,
|
352 |
+
sin,
|
353 |
+
interleaved=False,
|
354 |
+
seqlen_offsets: Union[int, torch.Tensor] = 0,
|
355 |
+
cu_seqlens: Optional[torch.Tensor] = None,
|
356 |
+
max_seqlen: Optional[int] = None,
|
357 |
+
):
|
358 |
+
"""
|
359 |
+
Arguments:
|
360 |
+
kv: (batch_size, seqlen, 2, nheads, headdim) if cu_seqlens is None
|
361 |
+
else (total_seqlen, 2, nheads, headdim)
|
362 |
+
cos, sin: (seqlen, rotary_dim / 2)
|
363 |
+
interleaved: if True, rotate pairs of even and odd dimensions (GPT-J style) instead of
|
364 |
+
1st half and 2nd half (GPT-NeoX style).
|
365 |
+
seqlen_offsets: (batch_size,) or int. Each sequence in Q and K is shifted by this amount.
|
366 |
+
Most commonly used in inference when we have KV cache.
|
367 |
+
cu_seqlens: (batch + 1,) or None
|
368 |
+
max_seqlen: int
|
369 |
+
Return:
|
370 |
+
kv: (batch_size, seqlen, 2, nheads, headdim) if cu_seqlens is None
|
371 |
+
else (total_seqlen, 2, nheads, headdim)
|
372 |
+
rotary_dim must be <= headdim
|
373 |
+
Apply rotary embedding *inplace* to the first rotary_dim of K.
|
374 |
+
"""
|
375 |
+
return ApplyRotaryEmbKV_.apply(
|
376 |
+
kv, cos, sin, interleaved, seqlen_offsets, cu_seqlens, max_seqlen
|
377 |
+
)
|
378 |
+
|
379 |
+
|
380 |
+
class RotaryEmbedding(torch.nn.Module):
|
381 |
+
"""
|
382 |
+
The rotary position embeddings from RoFormer_ (Su et. al).
|
383 |
+
A crucial insight from the method is that the query and keys are
|
384 |
+
transformed by rotation matrices which depend on the relative positions.
|
385 |
+
|
386 |
+
Other implementations are available in the Rotary Transformer repo_ and in
|
387 |
+
GPT-NeoX_, GPT-NeoX was an inspiration
|
388 |
+
|
389 |
+
.. _RoFormer: https://arxiv.org/abs/2104.09864
|
390 |
+
.. _repo: https://github.com/ZhuiyiTechnology/roformer
|
391 |
+
.. _GPT-NeoX: https://github.com/EleutherAI/gpt-neox
|
392 |
+
|
393 |
+
If scale_base is not None, this implements XPos (Sun et al., https://arxiv.org/abs/2212.10554).
|
394 |
+
A recommended value for scale_base is 512: https://github.com/HazyResearch/flash-attention/issues/96
|
395 |
+
Reference: https://github.com/sunyt32/torchscale/blob/main/torchscale/component/xpos_relative_position.py
|
396 |
+
"""
|
397 |
+
|
398 |
+
def __init__(
|
399 |
+
self,
|
400 |
+
dim: int,
|
401 |
+
base=10000.0,
|
402 |
+
interleaved=False,
|
403 |
+
scale_base=None,
|
404 |
+
pos_idx_in_fp32=True,
|
405 |
+
device=None,
|
406 |
+
):
|
407 |
+
"""
|
408 |
+
interleaved: if True, rotate pairs of even and odd dimensions (GPT-J style) instead
|
409 |
+
of 1st half and 2nd half (GPT-NeoX style).
|
410 |
+
pos_idx_in_fp32: if True, the position indices [0.0, ..., seqlen - 1] are in fp32,
|
411 |
+
otherwise they might be in lower precision.
|
412 |
+
This option was added because previously (before 2023-07-02), when we construct
|
413 |
+
the position indices, we use the dtype of self.inv_freq. In most cases this would
|
414 |
+
be fp32, but if the model is trained in pure bf16 (not mixed precision), then
|
415 |
+
self.inv_freq would be bf16, and the position indices are also in bf16.
|
416 |
+
Because of the limited precision of bf16 (e.g. 1995.0 is rounded to 2000.0), the
|
417 |
+
embeddings for some positions will coincide.
|
418 |
+
To maintain compatibility with models previously trained in pure bf16,
|
419 |
+
we add this option.
|
420 |
+
"""
|
421 |
+
super().__init__()
|
422 |
+
self.dim = dim
|
423 |
+
self.base = float(base)
|
424 |
+
self.pos_idx_in_fp32 = pos_idx_in_fp32
|
425 |
+
# Generate and save the inverse frequency buffer (non trainable)
|
426 |
+
inv_freq = self._compute_inv_freq(device)
|
427 |
+
self.register_buffer("inv_freq", inv_freq, persistent=False)
|
428 |
+
self.interleaved = interleaved
|
429 |
+
self.scale_base = scale_base
|
430 |
+
scale = (
|
431 |
+
(torch.arange(0, dim, 2, device=device, dtype=torch.float32) + 0.4 * dim) / (1.4 * dim)
|
432 |
+
if scale_base is not None
|
433 |
+
else None
|
434 |
+
)
|
435 |
+
self.register_buffer("scale", scale, persistent=False)
|
436 |
+
|
437 |
+
self._seq_len_cached = 0
|
438 |
+
self._cos_cached = None
|
439 |
+
self._sin_cached = None
|
440 |
+
self._cos_k_cached = None
|
441 |
+
self._sin_k_cached = None
|
442 |
+
|
443 |
+
def _compute_inv_freq(self, device=None):
|
444 |
+
return 1.0 / (
|
445 |
+
self.base
|
446 |
+
** (torch.arange(0, self.dim, 2, device=device, dtype=torch.float32) / self.dim)
|
447 |
+
)
|
448 |
+
|
449 |
+
def _update_cos_sin_cache(self, seqlen, device=None, dtype=None):
|
450 |
+
# Reset the tables if the sequence length has changed,
|
451 |
+
# if we're on a new device (possibly due to tracing for instance),
|
452 |
+
# or if we're switching from inference mode to training
|
453 |
+
if (
|
454 |
+
seqlen > self._seq_len_cached
|
455 |
+
or self._cos_cached is None
|
456 |
+
or self._cos_cached.device != device
|
457 |
+
or self._cos_cached.dtype != dtype
|
458 |
+
or (self.training and self._cos_cached.is_inference())
|
459 |
+
):
|
460 |
+
self._seq_len_cached = seqlen
|
461 |
+
# We want fp32 here, not self.inv_freq.dtype, since the model could be loaded in bf16
|
462 |
+
# And the output of arange can be quite large, so bf16 would lose a lot of precision.
|
463 |
+
# However, for compatibility reason, we add an option to use the dtype of self.inv_freq.
|
464 |
+
if self.pos_idx_in_fp32:
|
465 |
+
t = torch.arange(seqlen, device=device, dtype=torch.float32)
|
466 |
+
# We want fp32 here as well since inv_freq will be multiplied with t, and the output
|
467 |
+
# will be large. Having it in bf16 will lose a lot of precision and cause the
|
468 |
+
# cos & sin output to change significantly.
|
469 |
+
# We want to recompute self.inv_freq if it was not loaded in fp32
|
470 |
+
if self.inv_freq.dtype != torch.float32:
|
471 |
+
inv_freq = self._compute_inv_freq(device=device)
|
472 |
+
else:
|
473 |
+
inv_freq = self.inv_freq
|
474 |
+
else:
|
475 |
+
t = torch.arange(seqlen, device=device, dtype=self.inv_freq.dtype)
|
476 |
+
inv_freq = self.inv_freq
|
477 |
+
# Don't do einsum, it converts fp32 to fp16 under AMP
|
478 |
+
# freqs = torch.einsum("i,j->ij", t, self.inv_freq)
|
479 |
+
freqs = torch.outer(t, inv_freq)
|
480 |
+
if self.scale is None:
|
481 |
+
self._cos_cached = torch.cos(freqs).to(dtype)
|
482 |
+
self._sin_cached = torch.sin(freqs).to(dtype)
|
483 |
+
else:
|
484 |
+
power = (
|
485 |
+
torch.arange(seqlen, dtype=self.scale.dtype, device=self.scale.device)
|
486 |
+
- seqlen // 2
|
487 |
+
) / self.scale_base
|
488 |
+
scale = self.scale.to(device=power.device) ** rearrange(power, "s -> s 1")
|
489 |
+
# We want the multiplication by scale to happen in fp32
|
490 |
+
self._cos_cached = (torch.cos(freqs) * scale).to(dtype)
|
491 |
+
self._sin_cached = (torch.sin(freqs) * scale).to(dtype)
|
492 |
+
self._cos_k_cached = (torch.cos(freqs) / scale).to(dtype)
|
493 |
+
self._sin_k_cached = (torch.sin(freqs) / scale).to(dtype)
|
494 |
+
|
495 |
+
def forward(
|
496 |
+
self,
|
497 |
+
qkv: torch.Tensor,
|
498 |
+
kv: Optional[torch.Tensor] = None,
|
499 |
+
seqlen_offset: Union[int, torch.Tensor] = 0,
|
500 |
+
cu_seqlens: Optional[torch.Tensor] = None,
|
501 |
+
max_seqlen: Optional[int] = None,
|
502 |
+
) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
|
503 |
+
"""
|
504 |
+
qkv: (batch, seqlen, 3, nheads, headdim) if kv is none,
|
505 |
+
else it's just q of shape (batch, seqlen, nheads, headdim)
|
506 |
+
kv: (batch, seqlen, 2, nheads, headdim)
|
507 |
+
seqlen_offset: (batch_size,) or int. Each sequence in x is shifted by this amount.
|
508 |
+
Most commonly used in inference when we have KV cache.
|
509 |
+
If it's a tensor of shape (batch_size,), then to update the cos / sin cache, one
|
510 |
+
should pass in max_seqlen, which will update the cos / sin cache up to that length.
|
511 |
+
Apply rotary embedding *inplace* to qkv and / or kv.
|
512 |
+
"""
|
513 |
+
if cu_seqlens is not None:
|
514 |
+
assert max_seqlen is not None
|
515 |
+
seqlen = qkv.shape[1] if max_seqlen is None else max_seqlen
|
516 |
+
if max_seqlen is not None:
|
517 |
+
self._update_cos_sin_cache(max_seqlen, device=qkv.device, dtype=qkv.dtype)
|
518 |
+
elif isinstance(seqlen_offset, int):
|
519 |
+
self._update_cos_sin_cache(seqlen + seqlen_offset, device=qkv.device, dtype=qkv.dtype)
|
520 |
+
if kv is None:
|
521 |
+
if self.scale is None:
|
522 |
+
return apply_rotary_emb_qkv_(
|
523 |
+
qkv,
|
524 |
+
self._cos_cached,
|
525 |
+
self._sin_cached,
|
526 |
+
interleaved=self.interleaved,
|
527 |
+
seqlen_offsets=seqlen_offset,
|
528 |
+
cu_seqlens=cu_seqlens,
|
529 |
+
max_seqlen=max_seqlen,
|
530 |
+
)
|
531 |
+
else:
|
532 |
+
return apply_rotary_emb_qkv_(
|
533 |
+
qkv,
|
534 |
+
self._cos_cached,
|
535 |
+
self._sin_cached,
|
536 |
+
self._cos_k_cached,
|
537 |
+
self._sin_k_cached,
|
538 |
+
interleaved=self.interleaved,
|
539 |
+
seqlen_offsets=seqlen_offset,
|
540 |
+
cu_seqlens=cu_seqlens,
|
541 |
+
max_seqlen=max_seqlen,
|
542 |
+
)
|
543 |
+
else:
|
544 |
+
q = qkv
|
545 |
+
q = apply_rotary_emb_func(
|
546 |
+
q,
|
547 |
+
self._cos_cached,
|
548 |
+
self._sin_cached,
|
549 |
+
interleaved=self.interleaved,
|
550 |
+
inplace=True,
|
551 |
+
seqlen_offsets=seqlen_offset,
|
552 |
+
cu_seqlens=cu_seqlens,
|
553 |
+
max_seqlen=max_seqlen,
|
554 |
+
)
|
555 |
+
if self.scale is None:
|
556 |
+
kv = apply_rotary_emb_kv_(
|
557 |
+
kv,
|
558 |
+
self._cos_cached,
|
559 |
+
self._sin_cached,
|
560 |
+
interleaved=self.interleaved,
|
561 |
+
seqlen_offsets=seqlen_offset,
|
562 |
+
cu_seqlens=cu_seqlens,
|
563 |
+
max_seqlen=max_seqlen,
|
564 |
+
)
|
565 |
+
else:
|
566 |
+
kv = apply_rotary_emb_kv_(
|
567 |
+
kv,
|
568 |
+
self._cos_k_cached,
|
569 |
+
self._sin_k_cached,
|
570 |
+
interleaved=self.interleaved,
|
571 |
+
seqlen_offsets=seqlen_offset,
|
572 |
+
cu_seqlens=cu_seqlens,
|
573 |
+
max_seqlen=max_seqlen,
|
574 |
+
)
|
575 |
+
return q, kv
|