File size: 14,389 Bytes
e202b16
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
# Copyright (c) Facebook, Inc. and its affiliates. All rights reserved.
#
# This source code is licensed under the BSD license found in the
# LICENSE file in the root directory of this source tree.

import pytest
import torch

# needed to register custom ops
import xformers  # noqa: F401
import xformers.components.attention.core
from xformers.components.attention._sputnik_sparse import _csr_to_coo
from xformers.components.attention.core import (
    _broadcast_batch,
    _create_random_sparsity,
    _sparse_bmm,
)

cuda_only = pytest.mark.skipif(
    not torch.cuda.is_available() or not torch.version.cuda, reason="requires CUDA"
)

_devices = (
    ["cpu", "cuda"] if torch.cuda.is_available() and torch.version.cuda else ["cpu"]
)


def _baseline_matmul_with_sparse_mask(
    a: torch.Tensor, b: torch.Tensor, mask: torch.Tensor
) -> torch.Tensor:
    assert a.ndim == b.ndim
    assert mask.ndim == a.ndim
    assert a.shape[-1] == b.shape[-2]
    assert a.shape[-2] == mask.shape[-2], f"{a.shape}, {mask.shape}"
    assert b.shape[-1] == mask.shape[-1], f"{b.shape}, {mask.shape}"
    assert a.shape[:-2] == b.shape[:-2], f"{a.shape}, {b.shape}"
    assert a.shape[:-2] == mask.shape[:-2], f"{a.shape}, {mask.shape}"
    idxs = mask.indices().unbind()
    b = b.transpose(-2, -1)

    # compute matmul for elements within the mask
    val = (a[idxs[:-2] + (idxs[-2], slice(None))] * b[idxs[:-2] + (idxs[-1], slice(None))]).sum(-1)  # type: ignore

    out_shape = a.shape[:-1] + (b.shape[-2],)
    res = torch.sparse_coo_tensor(torch.stack(idxs), val, out_shape)
    return res


def _baseline_matmul_with_dense_mask(
    a: torch.Tensor, b: torch.Tensor, mask: torch.Tensor
) -> torch.Tensor:
    res = a @ b
    res[~mask] = float("-inf")
    return res


def _baseline_sparse_bmm(a: torch.Tensor, b: torch.Tensor) -> torch.Tensor:
    # need to use torch.sparse.mm to get gradients wrt sparse matrix a
    # TODO implement this in C++ / CUDA as this is slow!
    out = []
    for ai, bi in zip(a, b):
        out.append(torch.sparse.mm(ai, bi))
    return torch.stack(out, dim=0)


@pytest.mark.parametrize("is_sparse", [True, False])
@pytest.mark.parametrize("contiguous", [True, False])
@pytest.mark.parametrize("device", _devices)
def test_matmul_with_mask(device, contiguous, is_sparse):
    B, L, K = 8, 30, 32
    prob = 0.5
    a = torch.rand(B, L, K, device=device)
    b = torch.rand(B, K, L, device=device)
    if not contiguous:
        a = a.transpose(-2, -1).contiguous().transpose(-2, -1)
        b = b.transpose(-2, -1).contiguous().transpose(-2, -1)
    mask = torch.rand(B, L, L, device=device) > prob

    fn = torch.ops.xformers.matmul_with_mask
    fn_gt = _baseline_matmul_with_dense_mask

    if is_sparse:
        mask = mask.to_sparse()
        fn_gt = _baseline_matmul_with_sparse_mask

    res = fn(a, b, mask)
    res_gt = fn_gt(a, b, mask)

    if is_sparse:
        res = res.to_dense()
        res_gt = res_gt.to_dense()

    assert res.dtype == res_gt.dtype
    assert torch.allclose(res, res_gt)


@pytest.mark.parametrize("is_sparse", [True, False])
@pytest.mark.parametrize("contiguous", [True, False])
@pytest.mark.parametrize("device", _devices)
def test_matmul_with_mask_backward(device, contiguous, is_sparse):
    if device == "cuda" and is_sparse is False:
        # Skip test for now due to bug in torch 1.8
        # See https://github.com/pytorch/pytorch/issues/54975
        # Broken CUDA / torch 1.8 combination, awaiting an update
        return

    B, L, K = 8, 10, 16
    prob = 0.5
    a = torch.rand(B, L, K, device=device, requires_grad=True)
    b = torch.rand(B, K, L, device=device, requires_grad=True)
    if not contiguous:
        a = a.detach().transpose(-2, -1).contiguous().transpose(-2, -1).requires_grad_()
        b = b.detach().transpose(-2, -1).contiguous().transpose(-2, -1).requires_grad_()
    mask = torch.rand(B, L, L, device=device) > prob

    fn = torch.ops.xformers.matmul_with_mask
    fn_gt = _baseline_matmul_with_dense_mask

    if is_sparse:
        mask = mask.to_sparse()
        fn_gt = _baseline_matmul_with_sparse_mask

    def compute_grads(f):
        out = f(a, b, mask)
        if is_sparse:
            out = out.to_dense()
        out.sum().backward()

    compute_grads(fn)
    grad_a = a.grad.clone()
    grad_b = b.grad.clone()
    a.grad = None
    b.grad = None
    compute_grads(fn_gt)
    assert torch.allclose(grad_a, a.grad)
    assert torch.allclose(grad_b, b.grad)


@pytest.mark.parametrize("device", _devices)
def test_sddmm_sputnik(device):
    B, L, M, K = 8, 30, 16, 32
    prob = 0.5
    a = torch.rand(B, L, K, device=device)
    b = torch.rand(B, M, K, device=device).transpose(-2, -1)
    mask = _create_random_sparsity(
        torch.ones(B, L, M, dtype=torch.bool, device=device), prob
    )

    mask_csr = xformers.components.attention.core.SparseCS(mask, device)

    fn = xformers.components.attention.core._matmul_with_mask

    mask = mask.to_sparse()

    res = fn(a, b, mask_csr)
    res_gt = fn(a, b, mask)

    res = res.to_dense()
    res_gt = res_gt.to_dense()

    assert res.dtype == res_gt.dtype
    assert torch.allclose(res, res_gt)


@cuda_only
@pytest.mark.parametrize("prob", [0.5, 1])
@pytest.mark.parametrize("K", [32, 17])
@pytest.mark.parametrize("M", [30, 17])
@pytest.mark.parametrize("L", [30, 17])
def test_sddmm_csr(L, M, K, prob):
    device = torch.device("cuda")
    # TODO add more checks for different nnz
    B = 8
    a = torch.rand(B, L, K, device=device)
    b = torch.rand(B, M, K, device=device)
    mask = _create_random_sparsity(
        torch.ones(B, L, M, dtype=torch.bool, device=device), prob
    )

    mask_csr = xformers.components.attention.core.SparseCS(mask, device)
    row_indices = mask_csr.row_indices
    row_offsets = mask_csr.row_offsets
    column_indices = mask_csr.column_indices

    fn = torch.ops.xformers.csr_sddmm
    fn_gt = torch.ops.xformers.sddmm_sputnik

    res = fn(a, b, row_indices, row_offsets, column_indices)
    res_gt = fn_gt(a, b, row_indices, row_offsets, column_indices)

    assert res.dtype == res_gt.dtype
    assert torch.allclose(res, res_gt, atol=1e-6)


@cuda_only
@pytest.mark.parametrize("nnz", [0, 4, 16, 20, 36])
def test_sddmm_csr_per_nnz(nnz):
    device = torch.device("cuda")
    B = 8
    L, M, K = 1024, 1024, 32
    a = torch.rand(B, L, K, device=device)
    b = torch.rand(B, M, K, device=device)
    mask = torch.zeros(L, M, dtype=torch.bool, device=device)
    mask.view(-1)[: nnz - 1] = True
    mask[-1, -1] = True

    mask_csr = xformers.components.attention.core.SparseCS(mask, device)
    row_indices = mask_csr.row_indices
    row_offsets = mask_csr.row_offsets
    column_indices = mask_csr.column_indices

    fn = torch.ops.xformers.csr_sddmm
    fn_gt = torch.ops.xformers.sddmm_sputnik

    res = fn(a, b, row_indices, row_offsets, column_indices)
    res_gt = fn_gt(a, b, row_indices, row_offsets, column_indices)

    assert res.dtype == res_gt.dtype
    assert torch.allclose(res, res_gt, atol=1e-6)


@cuda_only
@pytest.mark.parametrize("prob", [0.5, 1])
@pytest.mark.parametrize("K", [32, 17])
@pytest.mark.parametrize("M", [30, 17])
@pytest.mark.parametrize("L", [30, 17])
def test_sddmm_coo(L, M, K, prob):
    device = torch.device("cuda")
    # TODO add more checks for different nnz
    B = 8
    a = torch.rand(B, L, K, device=device)
    b = torch.rand(B, M, K, device=device)
    mask = _create_random_sparsity(
        torch.ones(B, L, M, dtype=torch.bool, device=device), prob
    )

    mask_csr = xformers.components.attention.core.SparseCS(mask, device)
    row_indices = mask_csr.row_indices
    row_offsets = mask_csr.row_offsets
    column_indices = mask_csr.column_indices

    fn = torch.ops.xformers.coo_sddmm
    fn_gt = torch.ops.xformers.sddmm_sputnik

    # convert from csr to coo
    row_coo, _ = _csr_to_coo(L, M, row_offsets, column_indices)

    res = fn(a, b, row_indices, row_coo, column_indices)
    res_gt = fn_gt(a, b, row_indices, row_offsets, column_indices)

    assert res.dtype == res_gt.dtype
    assert torch.allclose(res, res_gt, atol=1e-6)


@pytest.mark.parametrize("device", _devices)
def test_sddmm_sputnik_backward(device):
    contiguous = True

    B, L, M, K = 8, 10, 16, 32
    prob = 0.5
    a = torch.rand(B, L, K, device=device, requires_grad=True)
    b = torch.rand(B, M, K, device=device).transpose(-2, -1).requires_grad_(True)
    if not contiguous:
        a = a.detach().transpose(-2, -1).contiguous().transpose(-2, -1).requires_grad_()
        b = b.detach().transpose(-2, -1).contiguous().transpose(-2, -1).requires_grad_()
    mask = _create_random_sparsity(
        torch.ones(B, L, M, dtype=torch.bool, device=device), prob
    )

    mask_csr = xformers.components.attention.core.SparseCS(mask, device)

    fn = xformers.components.attention.core._matmul_with_mask

    mask = mask.to_sparse()

    out_csr = fn(a, b, mask_csr)
    out_csr.values.sum().backward()
    grad_a = a.grad.clone()
    grad_b = b.grad.clone()
    a.grad = None
    b.grad = None
    # fn(a[None], b[None], mask).coalesce().values().sum().backward()  # TODO check why this fails
    fn(a, b, mask).to_dense().sum().backward()
    assert torch.allclose(grad_a, a.grad, atol=1e-7)
    assert torch.allclose(grad_b, b.grad, atol=1e-7)


@pytest.mark.parametrize("device", _devices)
def test_sparse_softmax_sputnik(device):
    B, L = 8, 30
    prob = 0.5
    a = _create_random_sparsity(torch.rand(B, L, L, device=device), prob)

    a_csr = xformers.components.attention.core.SparseCS(a, device)

    fn = xformers.components.attention.core._softmax

    a = a.to_sparse()

    res = fn(a_csr)
    res_gt = fn(a)

    res = res.to_dense()
    res_gt = res_gt.to_dense()

    assert res.dtype == res_gt.dtype
    assert torch.allclose(res, res_gt)


@pytest.mark.parametrize("device", _devices)
def test_sparse_softmax_sputnik_backward(device):
    B, L = 8, 30
    prob = 0.5
    a = _create_random_sparsity(torch.rand(B, L, L, device=device), prob)

    a_csr = xformers.components.attention.core.SparseCS(a, device)

    fn = xformers.components.attention.core._softmax

    a = a.to_sparse()

    a_csr.values.requires_grad_(True)
    fn(a_csr).values.sum().backward()
    grad_a = a_csr.values.grad.clone()
    a.requires_grad_(True)
    fn(a).coalesce().values().sum().backward()
    assert torch.allclose(
        grad_a, a.grad.coalesce().values().reshape_as(grad_a), atol=1e-7
    )


@pytest.mark.parametrize("device", _devices)
def test_spmm_sputnik(device):
    B, L, K = 8, 30, 32
    prob = 0.5

    a = _create_random_sparsity(torch.rand(B, L, L, device=device), prob)

    b = torch.rand(B, L, K, device=device)

    a_csr = xformers.components.attention.core.SparseCS(a, device)

    fn = xformers.components.attention.core.bmm

    a = a.to_sparse()

    res = fn(a_csr, b)
    res_gt = fn(a, b)

    res = res
    res_gt = res_gt

    assert res.dtype == res_gt.dtype
    assert torch.allclose(res, res_gt)


@pytest.mark.parametrize("device", _devices)
def test_spmm_sputnik_backward(device):
    B, M, L, K = 8, 16, 30, 32
    prob = 0.5

    a = _create_random_sparsity(torch.rand(B, M, L, device=device), prob)

    b = torch.rand(B, L, K, device=device)
    b.requires_grad_(True)

    a_csr = xformers.components.attention.core.SparseCS(a, device)

    fn = xformers.components.attention.core.bmm

    a = a.to_sparse()
    a.requires_grad_(True)
    a_csr.values.requires_grad_(True)

    fn(a_csr, b).sum().backward()
    grad_a = a_csr.values.grad.clone()
    grad_b = b.grad.clone()

    b.grad = None
    fn(a, b).sum().backward()

    assert torch.allclose(
        grad_a, a.grad.coalesce().values().reshape_as(grad_a), atol=1e-7
    )
    assert torch.allclose(grad_b, b.grad, atol=1e-7)


@cuda_only
def test_csr_transpose():
    B, L, K = 8, 30, 40
    prob = 0.5
    device = torch.device("cuda")

    a = _create_random_sparsity(torch.rand(B, L, K, device=device), prob)

    a_csr = xformers.components.attention.core.SparseCS(a, device)

    res = a_csr.transpose()
    res2 = res.transpose()

    assert torch.allclose(res.to_dense(), a.transpose(-2, -1))
    assert torch.allclose(res2.to_dense(), a)


@pytest.mark.parametrize("contiguous", [True, False])
@pytest.mark.parametrize("device", _devices)
@pytest.mark.parametrize("prob", [0.95, 0.996])  # cover > 0.995
@pytest.mark.parametrize("N", [32, 64, 96])  # cover > 64
def test_sparse_bmm(device, contiguous, prob, N):
    B, M = 8, 64
    a = torch.rand(B, M, N, device=device)
    a[a < prob] = 0
    a = a.to_sparse()
    b = torch.rand(B, N, M, device=device)
    if not contiguous:
        a = a + a
        b = b.transpose(-2, -1).contiguous().transpose(-2, -1)

    res = _sparse_bmm(a, b)
    res_gt = _baseline_sparse_bmm(a, b)

    assert torch.allclose(res, res_gt)


@pytest.mark.parametrize("contiguous", [True, False])
@pytest.mark.parametrize("device", _devices)
def test_sparse_bmm_backward(device, contiguous):
    if device == "cuda":
        # Skip test for now due to bug in torch 1.8
        # See https://github.com/pytorch/pytorch/issues/54975
        # Broken CUDA / torch 1.8 combination, awaiting an update
        return

    B, L, K = 8, 10, 16
    prob = 0.5
    a = torch.rand(B, L, K, device=device)
    a[a < prob] = 0
    a = a.to_sparse()
    b = torch.rand(B, K, L, device=device, requires_grad=True)
    if not contiguous:
        a = a + a
        b = b.detach().transpose(-2, -1).contiguous().transpose(-2, -1).requires_grad_()
    a.requires_grad_(True)

    def compute_grads(f):
        out = f(a, b)
        out.sum().backward()

    compute_grads(_sparse_bmm)
    grad_a = a.grad.clone().coalesce()
    grad_b = b.grad.clone()
    a.grad = None
    b.grad = None
    compute_grads(_baseline_sparse_bmm)
    new_grad_a = a.grad.coalesce()
    assert torch.allclose(grad_a.indices(), new_grad_a.indices())
    assert torch.allclose(grad_a.values(), new_grad_a.values())
    assert torch.allclose(grad_b, b.grad)


@pytest.mark.parametrize("device", _devices)
def test_sparse_coo_broadcast(device):
    B, L, K = 8, 10, 16
    prob = 0.5
    a = torch.rand(L, K, device=device)
    a[a < prob] = 0

    a_sparse = a.to_sparse()

    res = _broadcast_batch(a_sparse, B)

    res_gt = a[None, :, :].expand(B, L, K)

    assert torch.allclose(res.to_dense(), res_gt)