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
| import torch
import triton
import triton.language as tl
from einops import rearrange, repeat
def rotate_half(x, interleaved=False):
if not interleaved:
x1, x2 = x.chunk(2, dim=-1)
return torch.cat((-x2, x1), dim=-1)
else:
x1, x2 = x[..., ::2], x[..., 1::2]
return rearrange(torch.stack((-x2, x1), dim=-1), "... d two -> ... (d two)", two=2)
def apply_rotary_emb_torch(x, cos, sin, interleaved=False):
"""
x: (batch_size, seqlen, nheads, headdim)
cos, sin: (seqlen, rotary_dim / 2) or (batch_size, seqlen, rotary_dim / 2)
"""
ro_dim = cos.shape[-1] * 2
assert ro_dim <= x.shape[-1]
cos = repeat(cos, "... d -> ... 1 (2 d)" if not interleaved else "... d -> ... 1 (d 2)")
sin = repeat(sin, "... d -> ... 1 (2 d)" if not interleaved else "... d -> ... 1 (d 2)")
return torch.cat(
[x[..., :ro_dim] * cos + rotate_half(x[..., :ro_dim], interleaved) * sin, x[..., ro_dim:]],
dim=-1,
)
apply_rotary_emb_compile = torch.compile(apply_rotary_emb_torch)
@triton.jit
def rotary_kernel(
x_ptr, cos_ptr, sin_ptr, out_ptr,
stride_x_bs, stride_x_s, stride_x_h, stride_x_d,
stride_c_bs, stride_c_s, stride_c_d,
stride_o_bs, stride_o_s, stride_o_h, stride_o_d,
B, S, H, D, RO_DIM,
COS_HAS_BS: tl.constexpr,
INTERLEAVED: tl.constexpr,
BLOCK_D: tl.constexpr,
):
pid = tl.program_id(0)
h = pid % H
tmp = pid // H
s = tmp % S
b = tmp // S
x_base = x_ptr + b * stride_x_bs + s * stride_x_s + h * stride_x_h
o_base = out_ptr + b * stride_o_bs + s * stride_o_s + h * stride_o_h
if COS_HAS_BS:
cos_base = cos_ptr + b * stride_c_bs + s * stride_c_s
sin_base = sin_ptr + b * stride_c_bs + s * stride_c_s
else:
cos_base = cos_ptr + s * stride_c_s
sin_base = sin_ptr + s * stride_c_s
offs = tl.arange(0, BLOCK_D)
for d0 in range(0, D, BLOCK_D):
d = d0 + offs
mask = d < D
x = tl.load(
x_base + d * stride_x_d,
mask=mask,
other=0.0,
)
rope_mask = d < RO_DIM
if INTERLEAVED:
even = (d & 1) == 0
pair_idx = d >> 1
cos = tl.load(
cos_base + pair_idx * stride_c_d,
mask=rope_mask,
other=0.0,
)
sin = tl.load(
sin_base + pair_idx * stride_c_d,
mask=rope_mask,
other=0.0,
)
other_d = tl.where(even, d + 1, d - 1)
x_other = tl.load(
x_base + other_d * stride_x_d,
mask=rope_mask,
other=0.0,
)
out_rope = tl.where(
even,
x * cos - x_other * sin,
x_other * sin + x * cos,
)
else:
half = RO_DIM // 2
first_half = d < half
pair_idx = tl.where(first_half, d, d - half)
cos = tl.load(
cos_base + pair_idx * stride_c_d,
mask=rope_mask,
other=0.0,
)
sin = tl.load(
sin_base + pair_idx * stride_c_d,
mask=rope_mask,
other=0.0,
)
other_d = tl.where(first_half, d + half, d - half)
x_other = tl.load(
x_base + other_d * stride_x_d,
mask=rope_mask,
other=0.0,
)
out_rope = tl.where(
first_half,
x * cos - x_other * sin,
x_other * sin + x * cos,
)
out = tl.where(rope_mask, out_rope, x)
tl.store(
o_base + d * stride_o_d,
out,
mask=mask,
)
def apply_rotary_emb_triton(x, cos, sin, interleaved=False):
"""
x: (B, S, H, D)
cos/sin: (S, D/2) or (B, S, D/2)
"""
B, S, H, D = x.shape
ro_dim = cos.shape[-1] * 2
out = torch.empty_like(x)
grid = (B * S * H,)
COS_HAS_BS = (cos.dim() == 3)
rotary_kernel[grid](
x, cos, sin, out,
x.stride(0), x.stride(1), x.stride(2), x.stride(3),
cos.stride(0) if COS_HAS_BS else 0,
cos.stride(-2),
cos.stride(-1),
out.stride(0), out.stride(1), out.stride(2), out.stride(3),
B, S, H, D, ro_dim,
COS_HAS_BS=COS_HAS_BS,
INTERLEAVED=interleaved,
BLOCK_D=128,
)
return out
if __name__ == "__main__":
B, S, H, D = 40, 4000, 64, 128
x = torch.randn(B, S, H, D, device="cuda")
cos = torch.randn(S, D // 2, device="cuda")
sin = torch.randn(S, D // 2, device="cuda")
out_torch = apply_rotary_emb_torch(x, cos, sin, interleaved=False)
out_triton = apply_rotary_emb_triton(x, cos, sin, interleaved=False)
print("Max diff:", (out_torch - out_triton).abs().max().item())
import torch.utils.benchmark as benchmark
t_torch = benchmark.Timer(
stmt="apply_rotary_emb_torch(x, cos, sin, interleaved=False)",
globals=globals(),
).blocked_autorange(min_run_time=1.0)
t_triton = benchmark.Timer(
stmt="apply_rotary_emb_triton(x, cos, sin, interleaved=False)",
globals=globals(),
).blocked_autorange(min_run_time=1.0)
t_compiled = benchmark.Timer(
stmt="apply_rotary_emb_compile(x, cos, sin, interleaved=False)",
globals=globals(),
).blocked_autorange(min_run_time=1.0)
print(t_torch)
print(t_triton)
print(t_compiled)
|