File size: 6,075 Bytes
c335050 | 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 |
import os
import pytest
import torch
import torch.nn.functional as F
from fla.ops.rwkv6 import chunk_rwkv6
from fla.ops.rwkv6.fused_recurrent import fused_recurrent_rwkv6
from fla.utils import assert_close, device, device_platform
@pytest.mark.skipif(
device_platform == 'intel',
reason="Intel Triton Failure",
)
@pytest.mark.parametrize(
('B', 'T', 'H', 'D', 'gate_logit_normalizer', 'dtype'),
[
pytest.param(*test, id="B{}-T{}-H{}-D{}-gate_logit_normalizer{}-{}".format(*test))
for test in [
(1, 15, 2, 60, 1.0, torch.float16),
(3, 60, 3, 64, 0.1, torch.float16),
(3, 64, 2, 64, 1, torch.float16),
(4, 500, 3, 256, 1, torch.float16),
(4, 1000, 4, 64, 10, torch.float16),
(4, 2048, 4, 64, 1, torch.float16),
(4, 2048, 4, 256, 1, torch.float16),
]
],
)
def test_chunk(
B: int,
T: int,
H: int,
D: int,
gate_logit_normalizer: float,
dtype: torch.dtype,
):
torch.manual_seed(42)
os.environ['TRITON_F32_DEFAULT'] = 'ieee'
q = torch.randn((B, T, H, D), dtype=dtype, device=device).requires_grad_()
k = torch.randn((B, T, H, D), dtype=dtype, device=device).requires_grad_()
v = torch.randn((B, T, H, D), dtype=dtype, device=device).requires_grad_()
w = F.logsigmoid(torch.randn((B, T, H, D), dtype=dtype, device=device)) / gate_logit_normalizer
u = torch.randn(H, D, dtype=dtype, device=device).requires_grad_(True)
h0 = torch.randn(B, H, D, D, dtype=dtype, device=device).requires_grad_()
w = w.requires_grad_()
do = torch.randn_like(v)
ref, ref_ht = fused_recurrent_rwkv6(
q.clone(),
k.clone(),
v.clone(),
w.clone(),
u.clone(),
initial_state=h0.clone(),
output_final_state=True,
)
ref, _ = fused_recurrent_rwkv6(
q.clone(),
k.clone(),
v.clone(),
w.clone(),
u.clone(),
initial_state=h0.clone(),
output_final_state=False,
)
((ref * do).sum()).backward()
ref_dq, q.grad = q.grad.clone(), None
ref_dk, k.grad = k.grad.clone(), None
ref_dv, v.grad = v.grad.clone(), None
ref_dw, w.grad = w.grad.clone(), None
ref_du, u.grad = u.grad.clone(), None
ref_dh0, h0.grad = h0.grad.clone(), None
# triton implementation
tri, tri_ht = chunk_rwkv6(
q.clone(),
k.clone(),
v.clone(),
w.clone(),
u.clone(),
initial_state=h0.clone(),
output_final_state=True,
)
((tri * do).sum()).backward()
tri_dq, q.grad = q.grad.clone(), None
tri_dk, k.grad = k.grad.clone(), None
tri_dv, v.grad = v.grad.clone(), None
tri_dw, w.grad = w.grad.clone(), None
tri_du, u.grad = u.grad.clone(), None
tri_dh0, h0.grad = h0.grad.clone(), None
assert_close('o', ref, tri, 0.004)
assert_close('ht', ref_ht, tri_ht, 0.005)
assert_close('dq', ref_dq, tri_dq, 0.005)
assert_close('dk', ref_dk, tri_dk, 0.005)
assert_close('dv', ref_dv, tri_dv, 0.005)
assert_close('dw', ref_dw, tri_dw, 0.005)
assert_close('du', ref_du, tri_du, 0.005)
assert_close('dh0', ref_dh0, tri_dh0, 0.005)
@pytest.mark.parametrize(
('H', 'D', 'cu_seqlens', 'dtype'),
[
pytest.param(*test, id="H{}-D{}-cu_seqlens{}-{}".format(*test))
for test in [
(4, 64, [0, 15], torch.float16),
(4, 64, [0, 256, 500, 1000], torch.float16),
(4, 100, [0, 15, 100, 300, 1200, 2000], torch.float16),
]
],
)
def test_chunk_varlen(
H: int,
D: int,
cu_seqlens: list[int],
dtype: torch.dtype,
):
torch.manual_seed(42)
os.environ['TRITON_F32_DEFAULT'] = 'ieee'
N = len(cu_seqlens) - 1
T = cu_seqlens[-1]
cu_seqlens = torch.tensor(cu_seqlens, dtype=torch.int32, device=device)
# seq-first required for inputs with variable lengths
q = torch.randn((1, T, H, D), dtype=dtype, device=device).requires_grad_()
k = torch.randn((1, T, H, D), dtype=dtype, device=device).requires_grad_()
v = torch.randn((1, T, H, D), dtype=dtype, device=device).requires_grad_()
w = F.logsigmoid(torch.randn((1, T, H, D), dtype=dtype, device=device)).requires_grad_(True)
u = torch.randn(H, D, dtype=dtype, device=device).requires_grad_(True)
h0 = torch.randn((N, H, D, D), dtype=dtype, device=device).requires_grad_()
do = torch.randn_like(v)
ref, ref_ht = fused_recurrent_rwkv6(
q.clone(),
k.clone(),
v.clone(),
w.clone(),
u.clone(),
initial_state=h0.clone(),
output_final_state=True,
cu_seqlens=cu_seqlens,
)
ref, _ = fused_recurrent_rwkv6(
q.clone(),
k.clone(),
v.clone(),
w.clone(),
u.clone(),
initial_state=h0.clone(),
output_final_state=False,
cu_seqlens=cu_seqlens,
)
ref.backward(do)
ref_dq, q.grad = q.grad.clone(), None
ref_dk, k.grad = k.grad.clone(), None
ref_dv, v.grad = v.grad.clone(), None
ref_dw, w.grad = w.grad.clone(), None
ref_du, u.grad = u.grad.clone(), None
ref_dh0, h0.grad = h0.grad.clone(), None
tri, tri_ht = chunk_rwkv6(
q.clone(),
k.clone(),
v.clone(),
w.clone(),
u.clone(),
initial_state=h0.clone(),
output_final_state=True,
cu_seqlens=cu_seqlens,
)
tri.backward(do)
tri_dq, q.grad = q.grad.clone(), None
tri_dk, k.grad = k.grad.clone(), None
tri_dv, v.grad = v.grad.clone(), None
tri_dw, w.grad = w.grad.clone(), None
tri_du, u.grad = u.grad.clone(), None
tri_dh0, h0.grad = h0.grad.clone(), None
assert_close('o', ref, tri, 0.004)
assert_close('ht', ref_ht, tri_ht, 0.005)
assert_close('dq', ref_dq, tri_dq, 0.005)
assert_close('dk', ref_dk, tri_dk, 0.005)
assert_close('dv', ref_dv, tri_dv, 0.005)
assert_close('dw', ref_dw, tri_dw, 0.005)
assert_close('du', ref_du, tri_du, 0.005)
assert_close('dh0', ref_dh0, tri_dh0, 0.005)
|