attention.py 7.79 KB
Newer Older
1
from ctypes import c_uint64
PanZezhongQY's avatar
PanZezhongQY committed
2
3
4
5
6
import ctypes
import sys
import os

sys.path.insert(0, os.path.abspath(os.path.join(os.path.dirname(__file__), "..", "..")))
PanZezhong's avatar
PanZezhong committed
7
from libinfiniop import (
8
9
    LIBINFINIOP,
    TestTensor,
PanZezhong's avatar
PanZezhong committed
10
    get_test_devices,
11
    check_error,
PanZezhong's avatar
PanZezhong committed
12
    test_operator,
13
    get_args,
PanZezhong's avatar
PanZezhong committed
14
15
16
    debug,
    get_tolerance,
    profile_operation,
17
18
19
20
21
    TestWorkspace,
    InfiniDtype,
    InfiniDtypeNames,
    InfiniDeviceNames,
    infiniopOperatorDescriptor_t,
PanZezhongQY's avatar
PanZezhongQY committed
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
)

import torch


def causal_softmax(x):
    type = x.dtype
    mask = torch.tril(torch.ones_like(x), diagonal=-1).flip(dims=[-2, -1])
    y = x.clone()
    masked = torch.where(mask == 1, -torch.inf, y.to(torch.float32))
    return torch.nn.functional.softmax(masked, dim=-1).to(type)


def attention(q, k, v, k_cache, v_cache, pos):
    type = q.dtype

    n_q_head = q.shape[0]
    n_kv_head = k.shape[0]

    # Concatenate key and value caches
    k_cache = k_cache[:, :pos, :]  # (n_kv_head, pos, head_dim)
    v_cache = v_cache[:, :pos, :]  # (n_kv_head, pos, head_dim)
    k = torch.cat([k_cache, k], dim=1)  # (n_kv_head, total_seq_len, head_dim)
    v = torch.cat([v_cache, v], dim=1)  # (n_kv_head, total_seq_len, head_dim)

    total_seq_len = k.shape[1]

    head_dim = v.shape[-1]

    if n_q_head != n_kv_head:
        q = q.reshape(
            n_kv_head, -1, head_dim
        )  # (n_kv_head, n_group * seq_len, head_dim)

    # Scaled dot-product attention
    attn_scores = (
        torch.einsum("hqd,hkd->hqk", q.to(torch.float32), k.to(torch.float32))
        .to(type)
        .reshape(n_q_head, -1, total_seq_len)
    )  # (n_q_head, seq_len, total_seq_len)
    attn_scores = attn_scores / (head_dim**0.5)

    attn_weights = causal_softmax(attn_scores).reshape(
        n_kv_head, -1, total_seq_len
    )  # (n_kv_head, seq_len, total_seq_len)

    # Weighted sum of values
    attn_output = (
        torch.einsum(
            "hqk,hkd->hqd", attn_weights.to(torch.float32), v.to(torch.float32)
        )
        .to(type)
        .reshape(n_q_head, -1, head_dim)
        .permute(1, 0, 2)
    )  # ([seq_len, n_q_head, head_dim])

    return attn_output


def test(
    handle,
83
    device,
PanZezhongQY's avatar
PanZezhongQY committed
84
85
86
87
88
89
90
91
92
93
94
95
    n_q_head,
    n_kv_head,
    seq_len,
    head_dim,
    pos,
    k_cache_buf_len,
    v_cache_buf_len,
    q_stride=None,
    k_stride=None,
    v_stride=None,
    k_cache_stride=None,
    v_cache_stride=None,
96
    dtype=InfiniDtype.F16,
PanZezhong's avatar
PanZezhong committed
97
    sync=None,
PanZezhongQY's avatar
PanZezhongQY committed
98
99
):
    print(
100
101
        f"Testing Attention on {InfiniDeviceNames[device]} with n_q_head:{n_q_head} n_kv_head:{n_kv_head} seq_len:{seq_len} head_dim:{head_dim} pos:{pos} "
        f"dtype:{InfiniDtypeNames[dtype]} q_stride:{q_stride} k_stride:{k_stride} v_stride:{v_stride} k_cache_stride:{k_cache_stride} v_cache_stride:{v_cache_stride}"
PanZezhongQY's avatar
PanZezhongQY committed
102
103
    )

104
105
106
107
108
109
    out = TestTensor([seq_len, n_q_head, head_dim], None, dtype, device, mode="zeros")
    q = TestTensor([n_q_head, seq_len, head_dim], q_stride, dtype, device, scale=0.1)
    k = TestTensor([n_kv_head, seq_len, head_dim], k_stride, dtype, device, scale=0.1)
    v = TestTensor([n_kv_head, seq_len, head_dim], v_stride, dtype, device, scale=0.1)
    k_cache = TestTensor(
        [n_kv_head, k_cache_buf_len, head_dim], k_cache_stride, dtype, device, scale=0.1
PanZezhongQY's avatar
PanZezhongQY committed
110
    )
111
112
    v_cache = TestTensor(
        [n_kv_head, v_cache_buf_len, head_dim], v_cache_stride, dtype, device, scale=0.1
PanZezhongQY's avatar
PanZezhongQY committed
113
114
    )

115
116
117
118
119
120
121
122
123
124
125
    def torch_attention():
        return attention(
            q.torch_tensor(),
            k.torch_tensor(),
            v.torch_tensor(),
            k_cache.torch_tensor(),
            v_cache.torch_tensor(),
            pos,
        )

    ans = torch_attention()
PanZezhong's avatar
PanZezhong committed
126

127
128
    if sync is not None:
        sync()
PanZezhongQY's avatar
PanZezhongQY committed
129

130
    descriptor = infiniopOperatorDescriptor_t()
PanZezhongQY's avatar
PanZezhongQY committed
131
    check_error(
132
        LIBINFINIOP.infiniopCreateAttentionDescriptor(
PanZezhongQY's avatar
PanZezhongQY committed
133
134
            handle,
            ctypes.byref(descriptor),
135
136
137
138
139
140
            out.descriptor,
            q.descriptor,
            k.descriptor,
            v.descriptor,
            k_cache.descriptor,
            v_cache.descriptor,
PanZezhongQY's avatar
PanZezhongQY committed
141
142
143
144
145
            pos,
        )
    )

    # Invalidate the shape and strides in the descriptor to prevent them from being directly used by the kernel
146
147
    for tensor in [out, q, k, v, k_cache, v_cache]:
        tensor.destroy_desc()
PanZezhongQY's avatar
PanZezhongQY committed
148
149
150

    workspace_size = c_uint64(0)
    check_error(
151
152
153
        LIBINFINIOP.infiniopGetAttentionWorkspaceSize(
            descriptor, ctypes.byref(workspace_size)
        )
PanZezhongQY's avatar
PanZezhongQY committed
154
    )
155
    workspace = TestWorkspace(workspace_size.value, out.device)
PanZezhongQY's avatar
PanZezhongQY committed
156

PanZezhong's avatar
PanZezhong committed
157
158
    def lib_attention():
        check_error(
159
            LIBINFINIOP.infiniopAttention(
PanZezhong's avatar
PanZezhong committed
160
                descriptor,
161
                workspace.data(),
PanZezhong's avatar
PanZezhong committed
162
                workspace_size.value,
163
164
165
166
167
168
                out.data(),
                q.data(),
                k.data(),
                v.data(),
                k_cache.data(),
                v_cache.data(),
PanZezhong's avatar
PanZezhong committed
169
170
                None,
            )
PanZezhongQY's avatar
PanZezhongQY committed
171
172
        )

PanZezhong's avatar
PanZezhong committed
173
    lib_attention()
PanZezhongQY's avatar
PanZezhongQY committed
174

PanZezhong's avatar
PanZezhong committed
175
176
177
    # Validate results
    atol, rtol = get_tolerance(_TOLERANCE_MAP, dtype)
    if DEBUG:
178
179
        debug(out.actual_tensor(), ans, atol=atol, rtol=rtol)
    assert torch.allclose(out.actual_tensor(), ans, atol=atol, rtol=rtol)
PanZezhongQY's avatar
PanZezhongQY committed
180

PanZezhong's avatar
PanZezhong committed
181
182
183
    # Profiling workflow
    if PROFILE:
        # fmt: off
184
185
        profile_operation("PyTorch", lambda: torch_attention(), device, NUM_PRERUN, NUM_ITERATIONS)
        profile_operation("    lib", lambda: lib_attention(), device, NUM_PRERUN, NUM_ITERATIONS)
PanZezhong's avatar
PanZezhong committed
186
        # fmt: on
187
    check_error(LIBINFINIOP.infiniopDestroyAttentionDescriptor(descriptor))
PanZezhongQY's avatar
PanZezhongQY committed
188
189
190


if __name__ == "__main__":
191
    _TENSOR_DTYPES = [InfiniDtype.F16, InfiniDtype.F32]
PanZezhong's avatar
PanZezhong committed
192
193
194

    # Tolerance map for different data types
    _TOLERANCE_MAP = {
195
196
        InfiniDtype.F16: {"atol": 1e-4, "rtol": 1e-2},
        InfiniDtype.F32: {"atol": 1e-5, "rtol": 1e-3},
PanZezhong's avatar
PanZezhong committed
197
198
199
200
201
202
    }

    DEBUG = False
    PROFILE = False
    NUM_PRERUN = 10
    NUM_ITERATIONS = 1000
PanZezhongQY's avatar
PanZezhongQY committed
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
    test_cases = [
        # prefill
        (
            32,  # n_q_head
            4,  # n_kv_head
            5,  # seq_len
            64,  # head_dim
            0,  # pos
            2048,  # k_cache_buf_len
            2048,  # v_cache_buf_len
            [64, 2560, 1],  # q_stride
            [64, 2560, 1],  # k_stride
            [64, 2560, 1],  # v_stride
            [64, 11264, 1],  # k_cache_stride
            [64, 11264, 1],  # v_cache_stride
        ),
        # decode
        (
            32,  # n_q_head
            4,  # n_kv_head
            1,  # seq_len
            64,  # head_dim
            3,  # pos
            2048,  # k_cache_buf_len
            2048,  # v_cache_buf_len
            [64, 2560, 1],  # q_stride
            [64, 2560, 1],  # k_stride
            [64, 2560, 1],  # v_stride
            [64, 11264, 1],  # k_cache_stride
            [64, 11264, 1],  # v_cache_stride
        ),
        # for test
        (
            8,  # n_q_head
            4,  # n_kv_head
            2,  # seq_len
            16,  # head_dim
            1,  # pos
            8,  # k_cache_buf_len
            8,  # v_cache_buf_len
            None,  # q_stride
            None,  # k_stride
            None,  # v_stride
            None,  # k_cache_stride
            None,  # v_cache_stride
        ),
249
250
251
252
253
254
255
256
257
258
259
260
261
262
        (
            28,  # n_q_head
            28,  # n_kv_head
            15,  # seq_len
            128,  # head_dim
            0,  # pos
            2048,  # k_cache_buf_len
            2048,  # v_cache_buf_len
            [128, 10752, 1],  # q_stride
            [128, 10752, 1],  # k_stride
            [128, 10752, 1],  # v_stride
            [128, 3584, 1],  # k_cache_stride
            [128, 3584, 1],  # v_cache_stride
        ),
PanZezhongQY's avatar
PanZezhongQY committed
263
264
265
    ]
    args = get_args()

PanZezhong's avatar
PanZezhong committed
266
267
268
269
270
271
272
273
    # Configure testing options
    DEBUG = args.debug
    PROFILE = args.profile
    NUM_PRERUN = args.num_prerun
    NUM_ITERATIONS = args.num_iterations

    # Execute tests
    for device in get_test_devices(args):
274
        test_operator(device, test, test_cases, _TENSOR_DTYPES)
PanZezhongQY's avatar
PanZezhongQY committed
275
    print("\033[92mTest passed!\033[0m")