test_functional.py 55.3 KB
Newer Older
Tim Dettmers's avatar
Tim Dettmers committed
1
2
3
import math
import random
import time
Tim Dettmers's avatar
Tim Dettmers committed
4

5
import einops
Aarni Koskela's avatar
Aarni Koskela committed
6
import numpy as np
7
8
9
10
import pytest
import torch

import bitsandbytes as bnb
Tim Dettmers's avatar
Tim Dettmers committed
11
from bitsandbytes import functional as F
Aarni Koskela's avatar
Aarni Koskela committed
12
13
14
15
from tests.helpers import (
    BOOLEAN_TUPLES,
    TRUE_FALSE,
    describe_dtype,
16
    get_available_devices,
Aarni Koskela's avatar
Aarni Koskela committed
17
18
19
    get_test_dims,
    id_formatter,
)
Tim Dettmers's avatar
Tim Dettmers committed
20

Ruff's avatar
Ruff committed
21
torch.set_printoptions(precision=5, sci_mode=False, linewidth=120, edgeitems=20, threshold=10000)
Tim Dettmers's avatar
Tim Dettmers committed
22
23
k = 20

24

Tim Dettmers's avatar
Tim Dettmers committed
25
def assert_all_approx_close(a, b, rtol=1e-3, atol=1e-3, count=0, throw=True):
26
    idx = torch.isclose(a, b, rtol=rtol, atol=atol)
27
    sumval = (idx == 0).sum().item()
Tim Dettmers's avatar
Tim Dettmers committed
28
    if sumval > count:
Tim Dettmers's avatar
Tim Dettmers committed
29
30
        if throw:
            print(f"Too many values not close: assert {sumval} < {count}")
31
            torch.testing.assert_close(a, b, rtol=rtol, atol=atol)
Tim Dettmers's avatar
Tim Dettmers committed
32
33

    return sumval
Tim Dettmers's avatar
Tim Dettmers committed
34

35

Tim Dettmers's avatar
Tim Dettmers committed
36
37
class FFN(torch.nn.Module):
    def __init__(self, input_features, hidden_size, bias=True):
38
        super().__init__()
Tim Dettmers's avatar
Tim Dettmers committed
39
40
41
42
43
44
45
46
47
48
49
50
        self.fc1 = torch.nn.Linear(input_features, hidden_size, bias=bias)
        self.fc2 = torch.nn.Linear(hidden_size, input_features, bias=bias)

        with torch.no_grad():
            torch.nn.init.xavier_uniform_(self.fc1.weight)
            torch.nn.init.xavier_uniform_(self.fc2.weight)

    def forward(self, x):
        x = torch.relu(self.fc1(x))
        x = self.fc2(x)
        return x

51

52
class Timer:
Tim Dettmers's avatar
Tim Dettmers committed
53
54
55
56
57
    def __init__(self):
        self.starts = {}
        self.ends = {}
        self.agg = {}

58
    def tick(self, name="default"):
Tim Dettmers's avatar
Tim Dettmers committed
59
60
61
62
63
64
65
        if name not in self.starts:
            self.starts[name] = torch.cuda.Event(enable_timing=True)
            self.ends[name] = torch.cuda.Event(enable_timing=True)
            self.starts[name].record()
        else:
            ms = self.tock(name, evict=True, print_ms=False)

66
    def tock(self, name="default", evict=True, print_ms=True):
Tim Dettmers's avatar
Tim Dettmers committed
67
68
69
70
        if name in self.ends:
            self.ends[name].record()
            torch.cuda.synchronize()
            ms = self.starts[name].elapsed_time(self.ends[name])
71
72
            if name not in self.agg:
                self.agg[name] = 0.0
Tim Dettmers's avatar
Tim Dettmers committed
73
74
75
76
77
78
            self.agg[name] += ms
            if evict:
                self.starts.pop(name)
                self.ends.pop(name)

        if print_ms and name in self.agg:
79
            print(f"{name} took: {self.agg[name] / 1000.0:.5f}s")
Tim Dettmers's avatar
Tim Dettmers committed
80
81
82
83

        return self.agg[name]

    def reset(self):
84
        self.starts = {}
Tim Dettmers's avatar
Tim Dettmers committed
85
86
        self.ends = {}
        self.agg = {}
87
88
        print("Resetting benchmark data")

Tim Dettmers's avatar
Tim Dettmers committed
89

90
class Test8BitBlockwiseQuantizeFunctional:
91
    @pytest.mark.parametrize("device", get_available_devices())
92
93
94
95
    @pytest.mark.parametrize("dtype", [torch.float32, torch.float16, torch.bfloat16], ids=describe_dtype)
    @pytest.mark.parametrize("nested", TRUE_FALSE, ids=id_formatter("nested"))
    @pytest.mark.parametrize("blocksize", [4096, 2048, 1024, 512, 256, 128, 64])
    @pytest.mark.parametrize("signed", TRUE_FALSE, ids=id_formatter("signed"))
96
    def test_dynamic_blockwise_quantization(self, device, dtype, nested, blocksize, signed):
97
98
        iters = 100

99
        if device == "cpu":
100
101
            iters = 10

102
103
104
105
            # This test is slow on CPU, so avoid atypical use cases.
            if nested:
                pytest.skip("Not a typical use case.")
            if blocksize != 256:
106
                pytest.skip("Only blocksize 256 is used in CPU/XPU")
107
            if dtype != torch.float32:
108
                pytest.skip("Only float32 is used in CPU/XPU")
109

110
111
        diffs = []
        reldiffs = []
112
        for i in range(iters):
113
            A1 = torch.randn(1024, 1024, device=device, dtype=dtype)
114
115
116
117
118
119
120
121
122
123
124
125
126
127
            C, S = F.quantize_blockwise(A1, blocksize=blocksize, nested=nested)
            A2 = F.dequantize_blockwise(C, S)
            diff = torch.abs(A1 - A2).float()
            reldiff = diff / torch.abs(A1.float() + 1e-8)
            diffs.append(diff.mean().item())
            reldiffs.append(reldiff.mean().item())
        abserr = sum(diffs) / len(diffs)
        relerr = sum(reldiffs) / len(reldiffs)
        assert abserr < 0.011
        assert relerr < 0.018
        assert A2.dtype == dtype

        diffs = []
        code = F.create_dynamic_map(signed=signed)
128
        for i in range(iters):
129
            A1 = torch.rand(1024, 1024, device=device, dtype=dtype)
130
131
132
133
134
135
136
137
138
139
            C, S = F.quantize_blockwise(A1, blocksize=blocksize, nested=nested, code=code)
            A2 = F.dequantize_blockwise(C, S)
            diff = torch.abs(A1 - A2).float()
            reldiff = diff / torch.abs(A1.float() + 1e-8)
            diffs.append(diff.mean().item())
            reldiffs.append(reldiff.mean().item())
            # torch.testing.assert_close(A1, A2, atol=1e-2, rtol=0)
        abserr = sum(diffs) / len(diffs)
        relerr = sum(reldiffs) / len(reldiffs)
        if signed:
140
141
            threshold_abserr = 0.0036 if device in ("cpu", "xpu") else 0.0035
            assert abserr < 0.0036
142
143
            assert relerr < 0.015
        else:
144
            assert abserr < 0.00175 if device in ("cpu", "xpu") else 0.0023
145
146
147
            assert relerr < 0.012
        assert A2.dtype == dtype

148
149
150
151
    @pytest.mark.skipif("cpu" not in get_available_devices(), reason="CPU is required")
    @pytest.mark.parametrize("hidden", [128])
    @pytest.mark.parametrize("blocksize", [4096, 16384])
    def test_blockwise_cpu_large(self, hidden, blocksize):
152
153
154
155
        diffs = []
        reldiffs = []
        batch = 128
        seq = 128
156
157
158
159
160
161
162
163
164
165
166
167
168
169

        for i in range(2):
            A1 = torch.randn(batch, seq, hidden, device="cpu")
            t0 = time.time()
            C, S = F.quantize_blockwise(A1, blocksize=blocksize)
            A2 = F.dequantize_blockwise(C, S, blocksize=blocksize)
            print(time.time() - t0)
            diff = torch.abs(A1 - A2)
            reldiff = diff / torch.abs(A1 + 1e-8)
            diffs.append(diff.mean().item())
            reldiffs.append(reldiff.mean().item())
            assert diffs[-1] < 0.011
        # print(sum(diffs)/len(diffs))
        # print(sum(reldiffs)/len(reldiffs))
170

171
    @pytest.mark.parametrize("device", get_available_devices())
172
173
    @pytest.mark.parametrize("bits", range(2, 9), ids=id_formatter("bits"))
    @pytest.mark.parametrize("method", ["linear", "fp8", "dynamic", "quantile"])
174
    def test_few_bit_quant(self, device, bits, method):
175
176
        if device in ("cpu", "xpu") and bits != 8:
            pytest.skip("CPU/XPU implementation only supports 8 bits")
177

178
179
180
181
        abserrs = []
        relerrs = []
        code = None
        if method == "linear":
182
            code = F.create_linear_map(True, total_bits=bits).to(device)
183
184
185
        elif method == "fp8":
            ebits = math.ceil(bits / 2)
            pbits = bits - ebits - 1
186
            code = F.create_fp8_map(True, ebits, pbits, bits).to(device)
187
        elif method == "dynamic":
188
            code = F.create_dynamic_map(True, bits - 0, bits).to(device)
189
        elif method == "quantile":
190
            if device != "cuda":
191
                pytest.skip("Quantile map only works on CUDA")
192
193
194
195
196
197
198
199
200
            values = torch.randn(2048, 2048, device="cuda")
            code = F.create_quantile_map(values, bits).cuda()
        # for some data types we have no zero
        # for some data types we have one zero
        # for some data types we have two zeros
        assert torch.unique(code).numel() in [2**bits, 2**bits - 1], f"bits: {bits}, method: {method}"
        # print(method, (code==0).sum())
        assert code.numel() == 256
        for i in range(10):
201
            values = torch.randn(1, 32, device=device)
202
203
204
205
206
207
208
209
210
211
            values /= values.abs().max()
            # values[values.abs() < 1e-6] += 1e-5

            q1 = []
            v1 = []
            for v in values[0]:
                idx = torch.abs(v - code).argmin()
                q1.append(idx.item())
                v1.append(code[idx].item())

212
213
            q1 = torch.tensor(q1, device=device)
            v1 = torch.tensor(v1, device=device)
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228

            q2, S2 = F.quantize_blockwise(values, code=code)
            v2 = F.dequantize_blockwise(q2, S2)

            idx = torch.isclose(q1.int(), q2.int())
            err2 = torch.abs(v2 - values)
            abserrs.append(err2.mean().item())
            relerrs.append((err2 / (1e-10 + values).abs()).mean().item())
            if idx.sum():
                # some weird cases
                err1 = torch.abs(v1 - values).mean()
                # assert err2.mean() <= err1
            else:
                torch.testing.assert_close(q1, q2)

229
230
231
232
233
234
    @pytest.mark.parametrize("device", get_available_devices())
    def test_fp8_quant(self, device):
        # TODO
        if device == "cpu":
            pytest.skip("CPU implementation segfaults")

235
236
        for e_bits in range(1, 7):
            p_bits = 7 - e_bits
237
            code = F.create_fp8_map(True, e_bits, p_bits).to(device)
238
239
240
241

            abserr = []
            relerr = []
            for i in range(100):
242
                A1 = torch.randn(1024, 1024, device=device)
243
244
245
246
247
248
249
250
251
252
253
254
255
                C, SC = F.quantize_blockwise(A1, code=code)
                A2 = F.dequantize_blockwise(C, SC)
                diff = torch.abs(A1 - A2)
                reldiff = diff / torch.abs(A1 + 1e-8)
                abserr.append(diff.mean().item())
                relerr.append(reldiff.mean().item())
                # assert diff < 0.0075
            # print(sum(abserr)/len(abserr))
            # print(sum(relerr)/len(relerr))

            abserr = []
            relerr = []
            for i in range(100):
256
                A1 = torch.rand(1024, 1024, device=device)
257
258
259
260
261
262
263
264
265
266
267
268
269
                C, SC = F.quantize_blockwise(A1, code=code)
                A2 = F.dequantize_blockwise(C, SC)
                diff = torch.abs(A1 - A2)
                reldiff = diff / torch.abs(A1 + 1e-8)
                abserr.append(diff.mean().item())
                relerr.append(reldiff.mean().item())
                # assert diff < 0.0075
            # print(sum(abserr)/len(abserr))
            # print(sum(relerr)/len(relerr))

            abserr = []
            relerr = []
            for i in range(100):
270
                A1 = torch.randn(1024, 1024, device=device)
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
                C, SC = F.quantize_blockwise(A1)
                A2 = F.dequantize_blockwise(C, SC)
                diff = torch.abs(A1 - A2)
                reldiff = diff / torch.abs(A1 + 1e-8)
                abserr.append(diff.mean().item())
                relerr.append(reldiff.mean().item())
                # assert diff < 0.0075
            # print(3, sum(abserr)/len(abserr))
            # print(3, sum(relerr)/len(relerr))

    @pytest.mark.benchmark
    def test_bench_dequantization(self):
        a = torch.rand(1024, 1024, device="cuda").half()
        code = F.create_fp8_map(True, 3, 0, 4).cuda()
        qa, SA = F.quantize_blockwise(a, code=code)
        print(qa.max())

        max_theoretical_mu = 1024 * 1024 * 2 / 1024**3 / 672 * 1000 * 1000
        # print(max_theoretical_mu)

        torch.cuda.synchronize()
        t0 = time.time()
        for i in range(100):
            qa, SA = F.quantize_blockwise(a)
        torch.cuda.synchronize()
        # print((time.time()-t0)/1e6)
Tim Dettmers's avatar
Tim Dettmers committed
297
298


299
300
301
def test_stable_embedding():
    layer = bnb.nn.StableEmbedding(1024, 1024)
    layer.reset_parameters()
Tim Dettmers's avatar
Tim Dettmers committed
302
303


Tim Dettmers's avatar
Tim Dettmers committed
304
305
def quant(x):
    max1 = torch.abs(x).max()
306
    x = torch.round(x / max1 * 127)
Tim Dettmers's avatar
Tim Dettmers committed
307
308
    return max1, x.to(torch.int8)

309

Tim Dettmers's avatar
Tim Dettmers committed
310
def dequant(c, maxC):
311
312
    return c.float() * (maxC / 127)

Tim Dettmers's avatar
Tim Dettmers committed
313
314

def mm_dequant(maxA, maxB, C):
315
316
    return C.float() * (maxA / 127) * (maxB / 127)

Tim Dettmers's avatar
Tim Dettmers committed
317
318
319

def quant_multi(x, dim):
    max1 = torch.amax(torch.abs(x), dim=dim, keepdim=True)
320
321
    max1[max1 == 0] = 1.0
    x = torch.round(x / max1 * 127)
Tim Dettmers's avatar
Tim Dettmers committed
322
323
    return max1, x.to(torch.int8)

324

Tim Dettmers's avatar
Tim Dettmers committed
325
def quant_multi_chunk(x, dim, chunk_size=32):
326
327
328
    if dim == 1:
        x_chunked = einops.rearrange(x, "(c a) b -> c a b", c=chunk_size)
        max1 = torch.amax(torch.abs(x_chunked), dim=dim + 1, keepdim=True)
Tim Dettmers's avatar
Tim Dettmers committed
329
330
        max1 = torch.tile(max1, (1, 1, x.shape[1]))
        max1 = max1.view(x.shape)
331
332
    elif dim == 0:
        x_chunked = einops.rearrange(x, "a (b c) -> a b c", c=chunk_size)
Tim Dettmers's avatar
Tim Dettmers committed
333
334
335
        max1 = torch.amax(torch.abs(x_chunked), dim=dim, keepdim=True)
        max1 = torch.tile(max1, (x.shape[0], 1, 1))
        max1 = max1.view(x.shape)
336
337
    max1[max1 == 0] = 1.0
    x = torch.round(x / max1 * 127)
Tim Dettmers's avatar
Tim Dettmers committed
338
339
    return max1, x.to(torch.int8)

340

Tim Dettmers's avatar
Tim Dettmers committed
341
def mean(xx):
342
343
    return sum(xx) / float(len(xx))

Tim Dettmers's avatar
Tim Dettmers committed
344

Aarni Koskela's avatar
Aarni Koskela committed
345
346
methods = {
    "linear": (
347
348
349
350
351
        lambda x, dim: quant(x),
        lambda x, dim: quant(x),
        dequant,
        dequant,
        mm_dequant,
Aarni Koskela's avatar
Aarni Koskela committed
352
353
354
    ),
    "vectorwise": (quant_multi, quant_multi, dequant, dequant, mm_dequant),
}
355
356


357
@pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA is required")
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
class TestIGEMMFunctional:
    @pytest.mark.parametrize("dim1", [1024 * 2], ids=id_formatter("dim1"))
    @pytest.mark.parametrize("dim2", [1024 * 16], ids=id_formatter("dim2"))
    @pytest.mark.parametrize("quant_methods", methods.values(), ids=methods.keys())
    @pytest.mark.parametrize("batched", TRUE_FALSE, ids=id_formatter("batched"))
    def test_approx_igemm(self, dim1, dim2, quant_methods, batched):
        dim1 = dim1 - (dim1 % 32)
        dim2 = dim2 - (dim2 % 32)
        errors = []
        relerrors = []
        # print("")
        for i in range(5):
            if batched:
                A = torch.normal(0, 0.5, size=(32, dim1, dim2 // 32), device="cuda")
                B = torch.normal(0, 0.5, size=(32, dim2 // 32, dim1), device="cuda")
                maxA, Ac = quant_methods[0](A, 2)
                maxB, Bc = quant_methods[1](B, 1)
            else:
                A = torch.normal(0, 0.5, size=(dim1, dim2), device="cuda")
                B = torch.normal(0, 0.5, size=(dim2, dim1), device="cuda")
                maxA, Ac = quant_methods[0](A, 1)
                maxB, Bc = quant_methods[1](B, 0)
            torch.testing.assert_close(quant_methods[2](maxA, Ac), A, atol=0.025, rtol=0.05)
            if batched:
                out2 = torch.bmm(A, B)
                C = torch.bmm(Ac.float(), Bc.float())
            else:
                out2 = torch.mm(A, B)
                C = F.igemm(Ac, Bc)
            out = quant_methods[4](maxA, maxB, C)
            std = out2.std()
            out /= std
            out2 /= std
            err = torch.abs(out - out2)
            relerr = err / torch.abs(out2)
            errors.append(err.mean().item())
            relerrors.append(relerr.mean().item())
        # print(mean(errors))
        # print(mean(relerrors))

Matthew Douglas's avatar
Matthew Douglas committed
398
399
400
    @pytest.mark.parametrize("hidden_dim", [32, 256], ids=id_formatter("hidden_dim"))
    @pytest.mark.parametrize("batch_dim", [16, 256], ids=id_formatter("batch_dim"))
    @pytest.mark.parametrize("seq_dim", [16, 256], ids=id_formatter("seq_dim"))
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
    @pytest.mark.parametrize("transpose", BOOLEAN_TUPLES, ids=id_formatter("transpose"))
    def test_igemm(self, hidden_dim, batch_dim, transpose, seq_dim):
        hidden_dim = hidden_dim - (hidden_dim % 32)
        batch_dim = batch_dim - (batch_dim % 16)
        seq_dim = seq_dim - (seq_dim % 16)
        for i in range(k):
            shapeA = (batch_dim, hidden_dim) if not transpose[0] else (hidden_dim, batch_dim)
            shapeB = (
                (32 * random.randint(1, 4), hidden_dim) if transpose[1] else (hidden_dim, 32 * random.randint(1, 4))
            )
            A = torch.randint(-128, 127, size=shapeA, device="cuda").to(torch.int8)
            B = torch.randint(-128, 127, size=shapeB, device="cuda").to(torch.int8)
            if not transpose[0] and not transpose[1]:
                out2 = torch.matmul(A.float(), B.float())
                out = F.igemm(A, B)
            elif not transpose[0] and transpose[1]:
                out2 = torch.matmul(A.float(), B.t().float())
                out = F.igemm(A, B.t())
            elif transpose[0] and not transpose[1]:
                out2 = torch.matmul(A.t().float(), B.float())
                out = F.igemm(A.t(), B)
            elif transpose[0] and transpose[1]:
                out2 = torch.matmul(A.t().float(), B.t().float())
                out = F.igemm(A.t(), B.t())

            torch.testing.assert_close(out.float(), out2)

        for i in range(k):
            shapeA = (batch_dim, seq_dim, hidden_dim)
            shapeB = (
                (32 * random.randint(1, 4), hidden_dim) if transpose[1] else (hidden_dim, 32 * random.randint(1, 4))
            )
            A = torch.randint(-128, 127, size=shapeA, device="cuda").to(torch.int8)
            B = torch.randint(-128, 127, size=shapeB, device="cuda").to(torch.int8)
            if not transpose[0] and not transpose[1]:
                out2 = torch.matmul(A.float(), B.float())
                out = F.igemm(A, B)
            elif not transpose[0] and transpose[1]:
                out2 = torch.matmul(A.float(), B.t().float())
                out = F.igemm(A, B.t())

            torch.testing.assert_close(out.float(), out2)

Matthew Douglas's avatar
Matthew Douglas committed
444
445
446
    @pytest.mark.parametrize("seq_dim", [32, 256, 512], ids=id_formatter("seq_dim"))
    @pytest.mark.parametrize("hidden_dim", [64, 1024, 4096], ids=id_formatter("hidden_dim"))
    @pytest.mark.parametrize("batch_dim", [2, 8, 16], ids=id_formatter("batch_dim"))
447
448
449
450
451
452
453
454
455
456
457
458
459
    def test_dim3_igemm(self, seq_dim, hidden_dim, batch_dim):
        seq_dim = seq_dim - (seq_dim % 32)
        hidden_dim = hidden_dim - (hidden_dim % 32)
        batch_dim = batch_dim - (batch_dim % 2)
        for i in range(25):
            A = torch.randint(-128, 127, size=(batch_dim, seq_dim, hidden_dim), device="cuda").to(torch.int8)
            B = torch.randint(-128, 127, size=(batch_dim, seq_dim, 1024), device="cuda").to(torch.int8)
            out2 = torch.einsum("bsi, bso->io", A.float(), B.float())
            iout = torch.empty(A.shape[2], B.shape[2], dtype=torch.int32, device=A.device)
            out = F.igemm(A, B, out=iout)

            torch.testing.assert_close(out.float(), out2)

Matthew Douglas's avatar
Matthew Douglas committed
460
461
462
    @pytest.mark.parametrize("seq_dim", [32, 512], ids=id_formatter("seq_dim"))
    @pytest.mark.parametrize("hidden_dim", [32, 1024 * 4], ids=id_formatter("hidden_dim"))
    @pytest.mark.parametrize("batch_dim", [2, 16], ids=id_formatter("batch_dim"))
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
    @pytest.mark.parametrize("transpose", TRUE_FALSE, ids=id_formatter("transpose"))
    def test_minmax_igemm(self, seq_dim, hidden_dim, batch_dim, transpose):
        def min_max(x):
            maxA = torch.amax(x, dim=2, keepdim=True)
            minA = torch.amin(x, dim=2, keepdim=True)
            scale = (maxA - minA) / 2.0
            return (127 * (x - minA - scale) / scale).to(torch.int8), minA, scale

        seq_dim = seq_dim - (seq_dim % 16)
        hidden_dim = hidden_dim - (hidden_dim % 16)
        batch_dim = batch_dim - (batch_dim % 2)
        errs = []
        relerrs = []
        errs2 = []
        relerrs2 = []
        for i in range(k):
            A = torch.normal(0.0, 0.5, size=(batch_dim, seq_dim, hidden_dim), device="cuda")
            if transpose:
                B = torch.normal(0, 0.5, size=(256, hidden_dim), device="cuda")
            else:
                B = torch.normal(0, 0.5, size=(hidden_dim, 256), device="cuda")
            Ac, minA, scale = min_max(A)
            if transpose:
                maxB, Bc = quant_multi(B, dim=(1 if transpose else 0))
                out = F.igemm(Ac, Bc.t())
                out2 = torch.matmul(A, B.t())
                offset = B.t().sum(0) * (minA + scale)
                out = out.float()
                out = (out * maxB.t() * scale / (127 * 127)) + offset

                maxA, Ac = quant_multi(A, dim=2)
                out3 = F.igemm(Ac, Bc.t())
                out3 = mm_dequant(maxA, maxB.t(), out3)
            else:
                maxB, Bc = quant_multi(B, dim=0)
                offset = B.sum(0) * (minA + scale)
                out = F.igemm(Ac, Bc)
                out2 = torch.matmul(A, B)
                out = out.float()
                out = (out * maxB * scale / (127 * 127)) + offset

                maxA, Ac = quant_multi(A, dim=2)
                out3 = F.igemm(Ac, Bc)
                out3 = mm_dequant(maxA, maxB, out3)

            std = out2.std()
            out2 /= std
            out /= std
            out3 /= std

            err = torch.abs(out - out2)
            relerr = err / (torch.abs(out2) + 1e-7)

            err2 = torch.abs(out3 - out2)
            relerr2 = err2 / (torch.abs(out2) + 1e-7)

            errs.append(err.mean().item())
            relerrs.append(relerr.mean().item())
            errs2.append(err2.mean().item())
            relerrs2.append(relerr2.mean().item())
        # print(mean(errs))
        # print(mean(relerrs))
        # print(mean(errs2))
        # print(mean(relerrs2))
        assert mean(errs) < 0.015
Matthew Douglas's avatar
Matthew Douglas committed
528
529
530
531
532
533
534

        # There's a higher relerr on L40S with torch 2.4+cu118.
        is_sm89 = torch.cuda.get_device_capability() == (8, 9)
        if torch.version.cuda == "11.8" and is_sm89 and torch.__version__ < (2, 5):
            assert mean(relerrs) < 0.41
        else:
            assert mean(relerrs) < 0.3
535

Matthew Douglas's avatar
Matthew Douglas committed
536
537
538
539
    @pytest.mark.parametrize("dim1", [1, 64], ids=id_formatter("dim1"))
    @pytest.mark.parametrize("dim2", [32, 128], ids=id_formatter("dim2"))
    @pytest.mark.parametrize("dim3", [32, 256], ids=id_formatter("dim3"))
    @pytest.mark.parametrize("dim4", [32, 256], ids=id_formatter("dim4"))
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
    @pytest.mark.parametrize("transpose", BOOLEAN_TUPLES, ids=id_formatter("transpose"))
    def test_ibmm(self, dim1, dim2, dim3, dim4, transpose):
        dim2 = dim2 - (dim2 % 16)
        dim3 = dim3 - (dim3 % 16)
        dim4 = dim4 - (dim4 % 16)
        for i in range(k):
            shapeA = (dim1, dim3, dim2) if transpose[0] else (dim1, dim2, dim3)
            shapeB = (dim1, dim4, dim3) if transpose[1] else (dim1, dim3, dim4)
            A = torch.randint(-128, 127, size=shapeA, device="cuda").to(torch.int8)
            B = torch.randint(-128, 127, size=shapeB, device="cuda").to(torch.int8)

            if not transpose[0] and not transpose[1]:
                out2 = torch.bmm(A.float(), B.float())
                out = F.igemm(A, B)
            elif not transpose[0] and transpose[1]:
                out2 = torch.bmm(A.float(), B.permute([0, 2, 1]).float())
                out = F.igemm(A, B.permute([0, 2, 1]))
            elif transpose[0] and not transpose[1]:
                out2 = torch.bmm(A.permute([0, 2, 1]).float(), B.float())
                out = F.igemm(A.permute([0, 2, 1]), B)
            elif transpose[0] and transpose[1]:
                out2 = torch.bmm(A.permute([0, 2, 1]).float(), B.permute([0, 2, 1]).float())
                out = F.igemm(A.permute([0, 2, 1]), B.permute([0, 2, 1]))
            torch.testing.assert_close(out.float(), out2.float())


class TestLLMInt8Functional:
567
    @pytest.mark.parametrize("device", get_available_devices())
568
569
570
571
572
573
    @pytest.mark.parametrize("dim1", [128], ids=id_formatter("dim1"))
    @pytest.mark.parametrize("dim2", [256], ids=id_formatter("dim2"))
    @pytest.mark.parametrize("dim3", [499, 512], ids=id_formatter("dim3"))
    @pytest.mark.parametrize("dim4", [512], ids=id_formatter("dim4"))
    @pytest.mark.parametrize("dims", (2, 3), ids=id_formatter("dims"))
    @pytest.mark.parametrize("ldb", (0,), ids=id_formatter("ldb"))
574
    def test_int8_linear_matmul(self, device, dim1, dim2, dim3, dim4, dims, ldb):
575
576
        for i in range(k):
            if dims == 2:
577
                A = torch.randint(-128, 127, size=(dim1, dim3), dtype=torch.int8, device=device)
578
            elif dims == 3:
579
580
                A = torch.randint(-128, 127, size=(dim1, dim2, dim3), dtype=torch.int8, device=device)
            B = torch.randint(-128, 127, size=(dim4, dim3), dtype=torch.int8, device=device)
581
582
583
584
585
            C1 = torch.matmul(A.float(), B.t().float())

            C2 = F.int8_linear_matmul(A, B)
            torch.testing.assert_close(C1, C2.float())

586
    @pytest.mark.parametrize("device", get_available_devices())
587
588
589
590
591
    @pytest.mark.parametrize("dim1", [32], ids=id_formatter("dim1"))
    @pytest.mark.parametrize("dim2", [32], ids=id_formatter("dim2"))
    @pytest.mark.parametrize("dim3", [32], ids=id_formatter("dim3"))
    @pytest.mark.parametrize("dim4", [32], ids=id_formatter("dim4"))
    @pytest.mark.parametrize("dims", (2,), ids=id_formatter("dims"))
592
    def test_int8_linear_matmul_half(self, device, dim1, dim2, dim3, dim4, dims):
593
594
        for i in range(k):
            if dims == 2:
595
                A = torch.normal(0, 0.5, size=(dim1, dim3), device=device).half()
596
            elif dims == 3:
597
598
                A = torch.normal(0, 0.5, size=(dim1, dim2, dim3), device=device).half()
            B = torch.randn((dim4, dim3), device=device).half()
599
600
601
602
603
            torch.nn.init.xavier_uniform_(B)
            C1 = torch.matmul(A, B.t())

            A = A.view(-1, A.shape[-1])

604
            CA, statsA, _ = F.int8_vectorwise_quant(A)
605
606
607
608
609
            CB, statsB, _ = F.int8_vectorwise_quant(B)
            output = F.int8_mm_dequant(F.int8_linear_matmul(CA, CB), statsA, statsB)

            torch.testing.assert_close(C1.view(-1, C1.shape[-1]), output, atol=0.025, rtol=0.05)

610
    @pytest.mark.parametrize("device", get_available_devices())
611
612
613
614
    @pytest.mark.parametrize("dim1", (64, 256), ids=id_formatter("dim1"))
    @pytest.mark.parametrize("dim4", (64, 1024), ids=id_formatter("dim4"))
    @pytest.mark.parametrize("dims", (2,), ids=id_formatter("dims"))
    @pytest.mark.parametrize("has_bias", TRUE_FALSE, ids=id_formatter("has_bias"))
615
    def test_dequant_mm(self, device, dim1, dim4, dims, has_bias):
616
617
        inner = 128
        bias = None
Ruff's avatar
Ruff committed
618
        if has_bias:
619
            bias = torch.randn(dim4, device=device, dtype=torch.float16)
620
621

        for i in range(1):
622
623
            A = torch.randn(dim1, inner, device=device)
            B = torch.randn(dim4, inner, device=device)
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
            C1 = torch.matmul(A.half(), B.t().half())
            if has_bias:
                C1 += bias

            A1, maxA = F.vectorwise_quant(A, dim=1)
            B1, maxB = F.vectorwise_quant(B, dim=1)

            C2 = F.int8_linear_matmul(A1, B1)

            C4 = F.vectorwise_mm_dequant(C2.float(), maxA, maxB.t())
            if has_bias:
                C4 += bias

            # TODO: is something wrong here? If so, the problem goes deeper
            # n = C1.numel()
            # p = 0.06
            std = C1.std(0).view(1, -1)
            C1 /= std
            C4 /= std
            # assert_all_approx_close(C1, C4, atol=0.02, rtol=0.1, count=int(n*0.06))
            # assert (count / n < p), f"error in more than {p} of elements: {count}/{n}={count/n}"

            C5 = F.int8_mm_dequant(C2, maxA, maxB, bias=bias)
            C5 /= std
            torch.testing.assert_close(C5, C4, atol=0.015, rtol=0.1)
            n = C5.numel()
            assert_all_approx_close(C1, C4, atol=0.015, rtol=0.1, count=int(0.01 * n))

    @pytest.mark.parametrize("dim1", [1 * 1024], ids=id_formatter("dim1"))
    @pytest.mark.parametrize("dim2", [1 * 1024], ids=id_formatter("dim2"))
    @pytest.mark.parametrize("dims", (2,), ids=id_formatter("dims"))
    @pytest.mark.parametrize("threshold", [0.0, 3.0], ids=id_formatter("decomp"))
656
    @pytest.mark.deprecated
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
    def test_colrow_absmax(self, dim1, dim2, dims, threshold):
        for i in range(k):
            A = torch.randn(dim1, dim2, device="cuda").half()

            assert dims == 2

            row_stats1, _ = torch.abs(A.float()).max(1)
            col_stats1, _ = torch.abs(A.float()).max(0)

            if threshold > 0.0:
                A_truncated = A.clone()
                A_truncated[torch.abs(A_truncated) >= threshold] = 0.0
                row_stats1_trunc, _ = torch.abs(A_truncated.float()).max(1)
                col_stats1_trunc, _ = torch.abs(A_truncated.float()).max(0)

                row_stats2, col_stats2, nnz_block_ptr2 = F.get_colrow_absmax(A, threshold=threshold)

                nnz_rows1_counts = (torch.abs(A) >= threshold).sum(1).flatten()
                nnz_block_ptr1 = torch.zeros(
                    nnz_rows1_counts.shape[0] + 1,
                    dtype=nnz_rows1_counts.dtype,
                    device=nnz_rows1_counts.device,
                )
                nnz_block_ptr1[1:] = nnz_rows1_counts.cumsum(0)

                torch.testing.assert_close(col_stats1_trunc, col_stats2)
                torch.testing.assert_close(row_stats1_trunc, row_stats2)
                # torch.testing.assert_close(nnz_block_ptr1, nnz_block_ptr2)
            else:
                row_stats2, col_stats2, nnz_block_ptr2 = F.get_colrow_absmax(A, threshold=0.0)
                assert nnz_block_ptr2 is None
                torch.testing.assert_close(col_stats1, col_stats2)
                torch.testing.assert_close(row_stats1, row_stats2)

    @pytest.mark.parametrize("dim1", [2048, 4096], ids=id_formatter("dim1"))
    @pytest.mark.parametrize("dim2", [512, 1024], ids=id_formatter("dim2"))
693
    @pytest.mark.deprecated
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
    def test_int8_double_quant(self, dim1, dim2):
        for i in range(k):
            A = torch.randn(dim1, dim2, device="cuda").half()
            out_col1, Scol = F.vectorwise_quant(A, dim=0)
            out_row1, Srow = F.vectorwise_quant(A, dim=1)

            CA, CAt, statsA, statsAt, _ = F.int8_double_quant(A)

            # max difference is 1 due to rounding differences
            torch.testing.assert_close(CA, out_row1, atol=1, rtol=0)
            torch.testing.assert_close(CAt, out_col1, atol=1, rtol=0)

            n = CAt.numel()
            num_not_close_rows = (torch.isclose(CA, out_row1, atol=1) == 0).sum().item()
            num_not_close_cols = (torch.isclose(CAt, out_col1, atol=1) == 0).sum().item()

            # allow for 1:500 error due to rounding differences
            min_error = 1 / 500
            if num_not_close_cols > (min_error * n):
                print(
714
                    f"Min error exceeded {num_not_close_cols} elements are different. Error: {num_not_close_cols / n:.4f}"
715
716
717
718
                )
                assert False
            if num_not_close_rows > (min_error * n):
                print(
719
                    f"Min error exceeded {num_not_close_rows} elements are different. Error: {num_not_close_rows / n:.4f}"
720
721
722
723
724
725
                )
                assert False

            torch.testing.assert_close(Srow.flatten().float(), statsA)
            torch.testing.assert_close(Scol.flatten().float(), statsAt)

726
    @pytest.mark.parametrize("device", get_available_devices())
727
728
729
730
731
732
733
734
    @pytest.mark.parametrize(
        ("dim1", "dim4", "inner"),
        (
            pytest.param(dim1, dim4, inner, id=f"{dim1=},{dim4=},{inner=}")
            for (dim1, dim4, inner) in zip(
                (1, 8, 2048, 4096),
                (2, 128, 2048, 4096),
                (4, 256, 512, 4096),
735
            )
736
737
        ),
    )
738
    def test_integrated_int8_linear_matmul(self, device, dim1, dim4, inner):
739
740
741
        if device == "cpu" and inner > 2048:
            pytest.skip("Slow on CPU")

742
        for i in range(k):
743
744
            A = torch.randn(dim1, inner, device=device).half()
            B = torch.randn(dim4, inner, device=device).half()
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767

            out1 = torch.matmul(A.half(), B.t().half())

            C1a, stats1a, _ = F.int8_vectorwise_quant(A)
            C2a, stats2a, _ = F.int8_vectorwise_quant(B)
            A1, maxA = F.vectorwise_quant(A, dim=1)
            B1, maxB = F.vectorwise_quant(B, dim=1)

            torch.testing.assert_close(maxA.flatten().float(), stats1a)
            torch.testing.assert_close(maxB.flatten().float(), stats2a)
            torch.testing.assert_close(C1a, A1, rtol=0, atol=1)
            torch.testing.assert_close(C2a, B1, rtol=0, atol=1)

            out2 = F.int8_linear_matmul(A1, B1)

            C2 = F.int8_linear_matmul(A1, B1)

            out3 = F.vectorwise_mm_dequant(C2.float(), maxA, maxB.t())

            err1 = torch.abs(out1 - out2).mean().item()
            err2 = torch.abs(out1 - out3).mean().item()
            assert err2 <= err1 * 1.025

768
    @pytest.mark.parametrize("device", get_available_devices())
769
770
    @pytest.mark.parametrize("dim1", [512, 2048], ids=id_formatter("dim1"))
    @pytest.mark.parametrize("dim2", [1024, 4096], ids=id_formatter("dim2"))
771
    def test_coo_double_quant(self, device, dim1, dim2):
772
773
        threshold = 2.00
        for i in range(k):
774
            A = torch.randn(dim1, dim2, device=device).half()
775
776
777
778
779
780
781
782
783
784
785
786
787

            idx = torch.abs(A) >= threshold
            CA, statsA, outlier_cols = F.int8_vectorwise_quant(A, threshold=threshold)

            if outlier_cols is not None:
                A1 = A * idx
                A2 = torch.zeros_like(A) + A1
                torch.testing.assert_close(A1, A2)

                A[:, outlier_cols] = 0
                A2 = (CA.float() * statsA.unsqueeze(1) / 127).half()
                torch.testing.assert_close(A, A2, rtol=0.05, atol=1.5e-2)

788
    @pytest.mark.parametrize("device", get_available_devices())
789
790
    @pytest.mark.parametrize("dim1", [512, 2048], ids=id_formatter("dim1"))
    @pytest.mark.parametrize("dim2", [1024, 4096], ids=id_formatter("dim2"))
791
    def test_coo_int8_vectorwise_quant(self, device, dim1, dim2):
792
793
        threshold = 3.00
        for i in range(k):
794
            A = torch.randn(dim1, dim2, device=device).half()
795
796
797
798
799
800
801
802
803
804

            idx = torch.abs(A) >= threshold
            CA, statsA, outlier_cols = F.int8_vectorwise_quant(A, threshold=threshold)

            if outlier_cols is not None:
                A2 = (CA.float() * statsA.unsqueeze(1) / 127).half()
                A[:, outlier_cols] = 0
                torch.testing.assert_close(A * (idx == 0), A2, rtol=0.05, atol=1.5e-2)


805
@pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA is required")
806
class TestSpMMFunctional:
Matthew Douglas's avatar
Matthew Douglas committed
807
808
    @pytest.mark.parametrize("dim1", [256, 1024], ids=id_formatter("dim1"))
    @pytest.mark.parametrize("dim2", [128, 512], ids=id_formatter("dim2"))
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
    @pytest.mark.parametrize("transposed_B", TRUE_FALSE, ids=id_formatter("transposed_B"))
    def test_spmm_coo(self, dim1, dim2, transposed_B):
        threshold = 1.5
        dim3 = torch.randint(32, 128, size=(1,)).item()
        # dim3 = 17
        for i in range(k):
            A = torch.randn(dim1, dim2).cuda().half()
            if transposed_B:
                B = torch.randn(dim3, dim2).cuda().half()
            else:
                B = torch.randn(dim2, dim3).cuda().half()

            idx = torch.abs(A) >= threshold
            nnz = (idx == 1).sum().item()
            rows, cols = torch.where(idx)
            values = A[idx]
            cooA = F.COOSparseTensor(A.shape[0], A.shape[1], nnz, rows.int(), cols.int(), values)
            A2 = A * idx

            if transposed_B:
                out2 = F.spmm_coo(cooA, B.t())
                out1 = torch.matmul(A2, B.t())
            else:
                out2 = F.spmm_coo(cooA, B)
                out1 = torch.matmul(A2, B)

            assert_all_approx_close(out1, out2, rtol=0.01, atol=3.0e-2, count=30)

    @pytest.mark.benchmark
    def test_spmm_bench(self):
        batch = 2
        model = 1024 * 1
        hidden = model * 4
        seq = 1024
        dim1 = batch * seq
        dim2 = model
        dim3 = hidden
        threshold = 4
847
        A = torch.randn(dim1, dim2, device="cuda").half()
848
849
850
        B = torch.randn(dim2, dim3, device="cuda").half()
        for i in range(10):
            C1 = bnb.matmul(A, B.t())
Tim Dettmers's avatar
Tim Dettmers committed
851

852
853
854
855
856
857
        torch.cuda.synchronize()
        t0 = time.time()
        for i in range(k):
            C1 = bnb.matmul(A, B.t())
        torch.cuda.synchronize()
        t8 = time.time() - t0
Tim Dettmers's avatar
Tim Dettmers committed
858

859
860
861
862
863
864
        idx = torch.abs(A) >= threshold
        nnz = (idx == 1).sum().item()
        print(nnz / idx.numel())
        rows, cols = torch.where(idx)
        values = A[idx]
        cooA = F.COOSparseTensor(A.shape[0], A.shape[1], nnz, rows.int(), cols.int(), values)
Tim Dettmers's avatar
Tim Dettmers committed
865

866
867
        for i in range(10):
            out2 = F.spmm_coo(cooA, B)
Tim Dettmers's avatar
Tim Dettmers committed
868

869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
        torch.cuda.synchronize()
        t0 = time.time()
        for i in range(k):
            out2 = F.spmm_coo(cooA, B)
        torch.cuda.synchronize()
        tsp = time.time() - t0
        print(tsp, t8)
        print(tsp / t8)

    @pytest.mark.parametrize("dim1", [1 * 2048], ids=id_formatter("dim1"))
    @pytest.mark.parametrize("dim2", [12288], ids=id_formatter("dim2"))
    @pytest.mark.parametrize("dtype", [torch.float16], ids=describe_dtype)
    @pytest.mark.parametrize("out_func", ["zeros", "ones"], ids=id_formatter("out_func"))
    def test_spmm_coo_very_sparse(self, dim1, dim2, dtype, out_func):
        out_func = getattr(torch, out_func)

        threshold = 3.3
        # threshold = 2.8
        # threshold = 0.0
        A = torch.randn(dim1, dim2, device="cuda").half()
        if dtype == torch.float16:
            B = torch.randn(dim2, dim2 * 4, device="cuda").half()
            torch.nn.init.xavier_uniform_(B)
Tim Dettmers's avatar
Tim Dettmers committed
892
        else:
893
894
895
896
            B = torch.randn(dim2, dim2 * 4, device="cuda").half()
            torch.nn.init.xavier_uniform_(B)
            B, SB = F.vectorwise_quant(B, quant_type="linear")
            # B = torch.randint(-127, 127, size=(dim2, dim2*4), device='cuda').to(torch.int8)
Tim Dettmers's avatar
Tim Dettmers committed
897

898
899
900
901
902
903
904
905
906
907
908
909
        print("")
        idx = torch.abs(A) >= threshold
        nnz = (idx == 1).sum().item()
        rows, cols = torch.where(idx)
        values = A[idx]
        cooA = F.COOSparseTensor(A.shape[0], A.shape[1], nnz, rows.int(), cols.int(), values)
        A2 = A * idx
        out1 = torch.matmul(A2.half(), B.half())
        out = out_func(out1.shape, dtype=torch.float16, device=out1.device)
        out1 += out.clone()
        out2 = F.spmm_coo_very_sparse(cooA, B, out=out)
        # print(B)
910
911
        # print(out1)
        # print(out2)
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
        p = 200 / (2048 * 12288 * 4)
        n = out1.numel()
        count = math.ceil(p * n)
        std = out1.std()
        out1 /= std
        out2 /= std
        assert_all_approx_close(out1, out2.half(), rtol=0.01, atol=3.0e-2, count=count)
        # assert_all_approx_close(out1, out2.half(), rtol=0.05, atol=0.01, count=count)

        idx_col = torch.randint(0, A2.shape[-1], size=(15,))

        # torch.testing.assert_close(out1, out2.half(), rtol=0.05, atol=0.001)

        # Bt = torch.randn(dim2*4, dim2, device='cuda').half()
        # torch.cuda.synchronize()
        # t0 = time.time()
        # print(A2.shape, B.shape)
        # for i in range(100):
        #   #out3 = F.spmm_coo(cooA, Bt.t())
        #   #out2 = F.spmm_coo(cooA, B)
        #   #out2 = F.spmm_coo_very_sparse(cooA, B)
        #   #out1 = torch.matmul(A, Bt.t())

        # torch.cuda.synchronize()
        # print(time.time() - t0)

    @pytest.mark.parametrize("dim1", [1 * 2048])
    @pytest.mark.parametrize("dim2", [2048])
    @pytest.mark.parametrize("dtype", [torch.int8])
    def test_spmm_coo_dequant(self, dim1, dim2, dtype):
        threshold = 6.0
        # threshold = 2.8
        # threshold = 0.0
945
        A = torch.randn(dim1, dim2, device="cuda").half()
946
947
948
        B = torch.empty(dim2, dim2 * 4, device="cuda", dtype=torch.float16)
        torch.nn.init.xavier_uniform_(B)
        Bt = B.t().contiguous()
Tim Dettmers's avatar
Tim Dettmers committed
949

950
        CB, CBt, statsB, statsBt, coo_tensor = F.int8_double_quant(B)
951

952
        rowidx = torch.randint(0, A.shape[-1], size=(15,))
Tim Dettmers's avatar
Tim Dettmers committed
953

954
        A[:, rowidx] = 8.0
Tim Dettmers's avatar
Tim Dettmers committed
955
956
957
958
959

        idx = torch.abs(A) >= threshold
        nnz = (idx == 1).sum().item()
        rows, cols = torch.where(idx)
        values = A[idx]
Ruff's avatar
Ruff committed
960
        cooA = F.COOSparseTensor(A.shape[0], A.shape[1], nnz, rows.int(), cols.int(), values)
961
962
        A2 = A * idx
        out2 = F.spmm_coo_very_sparse(cooA, CBt, dequant_stats=statsBt)
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
        out1 = torch.matmul(A2, B.half())
        out3 = F.spmm_coo_very_sparse(cooA, CBt.half())
        out3 = out3 * statsBt.half() / 127

        values, counts = torch.unique(cooA.rowidx, return_counts=True)
        offset = counts.cumsum(0).int()
        max_count, max_idx = torch.sort(counts, descending=True)
        print(torch.median(max_count.float()))

        torch.testing.assert_close(out2, out3, rtol=0.05, atol=0.001)

        p = 200 / (2048 * 12288 * 4)
        n = out1.numel()
        count = math.ceil(p * n)
        assert_all_approx_close(out1, out2, rtol=0.01, atol=3.0e-2, count=count)

        # torch.cuda.synchronize()
        # t0 = time.time()
        # for i in range(100):
        #   out2 = F.spmm_coo_very_sparse(cooA, B)
        # torch.cuda.synchronize()
        # print('fp16', time.time() - t0)

        torch.cuda.synchronize()
        t0 = time.time()
        for i in range(100):
            out2 = F.spmm_coo(cooA, B)
        torch.cuda.synchronize()
        print("cusparse fp16", time.time() - t0)
992

993
994
995
996
997
998
        torch.cuda.synchronize()
        t0 = time.time()
        for i in range(100):
            out2 = F.spmm_coo_very_sparse(cooA, CBt)
        torch.cuda.synchronize()
        print("int8", time.time() - t0)
999

1000
1001
1002
1003
1004
1005
        torch.cuda.synchronize()
        t0 = time.time()
        for i in range(100):
            out2 = F.spmm_coo_very_sparse(cooA, CBt, dequant_stats=statsBt)
        torch.cuda.synchronize()
        print("int8+dequant", time.time() - t0)
1006

1007
1008
1009
1010
1011
1012
        torch.cuda.synchronize()
        t0 = time.time()
        for i in range(100):
            out2 = torch.matmul(A, B)
        torch.cuda.synchronize()
        print("matmul", time.time() - t0)
1013

1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
        torch.cuda.synchronize()
        t0 = time.time()
        for i in range(100):
            out1 = bnb.matmul(A, Bt)
            out2 = F.spmm_coo_very_sparse(cooA, CBt, dequant_stats=statsBt)
            out = out1 + out2
        torch.cuda.synchronize()
        print("sparse+ matmul", time.time() - t0)

        torch.cuda.synchronize()
        t0 = time.time()
        for i in range(100):
            out1 = bnb.matmul(A, Bt)
            torch.matmul(A[:, rowidx], Bt.t()[rowidx], out=out1)
        torch.cuda.synchronize()
        print("partial matmul", time.time() - t0)
1030

1031
1032
1033
1034
1035
1036
        torch.cuda.synchronize()
        t0 = time.time()
        for i in range(100):
            out1 = bnb.matmul(A, Bt)
        torch.cuda.synchronize()
        print("partial matmul", time.time() - t0)
1037

1038

1039
@pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA is required")
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
class TestSparseTensorFunctional:
    def test_coo2csr(self):
        threshold = 1
        A = torch.randn(128, 128).half().cuda()
        idx = torch.abs(A) >= threshold
        nnz = (idx == 1).sum().item()
        rows, cols = torch.where(idx)
        values = A[idx]
        cooA = F.COOSparseTensor(A.shape[0], A.shape[1], nnz, rows.int(), cols.int(), values)
        A2 = A * idx
        csrA = F.coo2csr(cooA)
        counts = csrA.rowptr[1:] - csrA.rowptr[:-1]
        assert counts.numel() == A.shape[0]
1053

1054
1055
1056
        torch.testing.assert_close(counts.long(), (A2 != 0).sum(1))
        idx = A2 != 0
        torch.testing.assert_close(A2[idx], csrA.values)
Tim Dettmers's avatar
Tim Dettmers committed
1057

1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
    def test_coo2csc(self):
        threshold = 1
        A = torch.randn(128, 128).half().cuda()
        idx = torch.abs(A) >= threshold
        nnz = (idx == 1).sum().item()
        rows, cols = torch.where(idx)
        values = A[idx]
        cooA = F.COOSparseTensor(A.shape[0], A.shape[1], nnz, rows.int(), cols.int(), values)
        A2 = A * idx
        cscA = F.coo2csc(cooA)
        counts = cscA.colptr[1:] - cscA.colptr[:-1]
        assert counts.numel() == A.shape[1]
Tim Dettmers's avatar
Tim Dettmers committed
1070

1071
1072
1073
1074
        torch.testing.assert_close(counts.long(), (A2 != 0).sum(0))
        # torch uses row-major -> use transpose to transfer to col-major
        idx = A2.t() != 0
        torch.testing.assert_close(A2.t()[idx], cscA.values)
Tim Dettmers's avatar
Tim Dettmers committed
1075

Tim Dettmers's avatar
Tim Dettmers committed
1076

1077
class TestQuantize4BitFunctional:
1078
    @pytest.mark.parametrize("device", get_available_devices())
1079
1080
1081
    @pytest.mark.parametrize("dtype", [torch.float32, torch.float16, torch.bfloat16], ids=describe_dtype)
    @pytest.mark.parametrize("quant_type", ["fp4", "nf4"])
    @pytest.mark.parametrize("blocksize", [64, 128, 256, 512, 1024, 2048, 4096])
1082
1083
    def test_4bit_quant(self, device, dtype, quant_type, blocksize):
        A1 = torch.randn(1024, 1024, device=device, dtype=dtype)
1084
1085
        qa, SA = F.quantize_4bit(A1, blocksize=blocksize, quant_type=quant_type)
        A2 = F.dequantize_4bit(qa, SA, blocksize=blocksize, quant_type=quant_type)
1086

1087
1088
1089
        err = (A1 - A2).abs().float()
        relerr = (err / (A1.abs().float() + 1e-8)).mean()
        err = err.mean()
1090

1091
        assert A2.dtype == dtype
1092

1093
1094
1095
1096
        # With larger block sizes, we can expect this to blow up.
        # At blocksize>=1024, don't even bother looking at relerr.
        if blocksize <= 64:
            assert err.item() < 0.1
1097
            assert relerr.item() < 0.28
1098
        elif blocksize <= 256:
1099
            assert err.item() < 0.11
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
            assert relerr.item() < 0.30
        elif blocksize <= 512:
            assert err.item() < 0.12
            assert relerr.item() < 0.31
        elif quant_type == "fp4":
            # 1024 => 0.48, 2048 => 0.52, 4096 => 0.56
            assert err.item() < 0.08 + math.log2(blocksize) * 4e-2
        else:
            # 1024 => 0.8, 2048 => 0.88, 4096 => 0.96
            assert err.item() < math.log2(blocksize) * 8e-2

1111
    @pytest.mark.parametrize("device", get_available_devices())
1112
    @pytest.mark.parametrize("quant_type", ["fp4", "nf4"])
Matthew Douglas's avatar
Matthew Douglas committed
1113
    @pytest.mark.parametrize("blocksize", [64, 128], ids=id_formatter("blocksize"))
1114
    def test_4bit_compressed_stats(self, device, quant_type, blocksize):
Matthew Douglas's avatar
Matthew Douglas committed
1115
1116
1117
        errs1 = []
        errs2 = []
        for i in range(10):
1118
            A1 = torch.randn(1024, 1024, device=device).half()
Matthew Douglas's avatar
Matthew Douglas committed
1119
1120
1121
1122
            q2, SA2 = F.quantize_4bit(A1, blocksize=blocksize, quant_type=quant_type)
            q3, SA3 = F.quantize_4bit(A1, blocksize=blocksize, compress_statistics=True, quant_type=quant_type)
            A2 = F.dequantize_4bit(q2, SA2, quant_type=quant_type)
            A3 = F.dequantize_4bit(q3, SA3, quant_type=quant_type)
1123

Matthew Douglas's avatar
Matthew Douglas committed
1124
1125
1126
            err = (A1 - A2).abs().float()
            relerr = (err / (A1.abs().float() + 1e-15)).mean()
            err = err.mean()
1127

Matthew Douglas's avatar
Matthew Douglas committed
1128
            errs1.append(err.item())
1129

Matthew Douglas's avatar
Matthew Douglas committed
1130
1131
            assert err.item() < 0.11
            assert relerr.item() < 0.28
1132

Matthew Douglas's avatar
Matthew Douglas committed
1133
1134
1135
            err = (A1 - A3).abs().float()
            relerr = (err / (A1.abs().float() + 1e-15)).mean()
            err = err.mean()
1136

Matthew Douglas's avatar
Matthew Douglas committed
1137
            errs2.append(err.item())
1138

Matthew Douglas's avatar
Matthew Douglas committed
1139
1140
            assert err.item() < 0.11
            assert relerr.item() < 0.28
1141
1142
1143

    # @pytest.mark.parametrize("quant_type", ['fp4', 'nf4'])
    @pytest.mark.parametrize("quant_type", ["nf4"])
1144
    @pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA is required")
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
    @pytest.mark.benchmark
    def test_bench_4bit_dequant(self, quant_type):
        blocksize = 256
        a = torch.rand(1024 * 12 * 4, 1024 * 12, device="cuda").half()
        qa, SA = F.quantize_4bit(a, blocksize=blocksize, quant_type=quant_type)

        input_size = a.numel() / 2
        output_size = a.numel() * 2
        num_bytes = input_size + output_size
        GB = num_bytes / 1e9
        max_theoretical_s = GB / 768
        # print(max_theoretical_s*1e6)
        b = torch.randn(128, 1024 * 12, device="cuda").half()

        iters = 100
        torch.cuda.synchronize()
        t0 = time.time()
        for i in range(iters):
            F.dequantize_4bit(qa, SA, blocksize=blocksize, quant_type=quant_type)
            # b.copy_(a)
        torch.cuda.synchronize()
        # print((time.time()-t0)/iters*1e6)

        # torch.cuda.synchronize()
        # t0 = time.time()
        # for i in range(iters):
        #    torch.matmul(b, a.t())
        # torch.cuda.synchronize()
        # print((time.time()-t0)/iters*1e6)

1175
    @pytest.mark.parametrize("device", get_available_devices())
1176
1177
1178
1179
1180
1181
1182
1183
1184
    @pytest.mark.parametrize("double_quant", TRUE_FALSE, ids=lambda double_quant: f"DQ_{double_quant}")
    @pytest.mark.parametrize("storage_type", ["nf4", "fp4"])
    @pytest.mark.parametrize("kind", ["fc1", "fc2", "attn", "attn_packed"])
    @pytest.mark.parametrize("dtype", [torch.float16, torch.bfloat16, torch.float32], ids=describe_dtype)
    @pytest.mark.parametrize(
        "quant_storage",
        [torch.uint8, torch.float16, torch.bfloat16, torch.float32],
        ids=describe_dtype,
    )
Matthew Douglas's avatar
Matthew Douglas committed
1185
    @pytest.mark.parametrize("dim", [128, 256, 512, 1024], ids=id_formatter("dim"))
1186
    def test_gemv_4bit(self, device, dim, dtype, storage_type, quant_storage, double_quant, kind):
Matthew Douglas's avatar
Matthew Douglas committed
1187
1188
1189
1190
1191
1192
1193
1194
1195
        errs1 = []
        errs2 = []
        errs3 = []
        relerrs1 = []
        relerrs2 = []
        relerrs3 = []
        max_errs1 = []
        max_errs2 = []
        max_errs3 = []
1196

1197
1198
1199
1200
1201
        # Large number of iterations is excessive and slow on CPU.
        # Keep for CUDA for now.
        iters = 100 if device == "cuda" else 10

        for i in range(iters):
Matthew Douglas's avatar
Matthew Douglas committed
1202
            if kind == "fc1":
1203
1204
                A = torch.randn(1, dim, dtype=dtype, device=device)
                B = torch.randn(dim * 4, dim, dtype=dtype, device=device) / math.sqrt(dim)
Matthew Douglas's avatar
Matthew Douglas committed
1205
            elif kind == "fc2":
1206
1207
                A = torch.randn(1, 4 * dim, dtype=dtype, device=device)
                B = torch.randn(dim, 4 * dim, dtype=dtype, device=device) / math.sqrt(dim)
Matthew Douglas's avatar
Matthew Douglas committed
1208
            elif kind == "attn":
1209
1210
                A = torch.randn(1, dim, dtype=dtype, device=device)
                B = torch.randn(dim, dim, dtype=dtype, device=device) / math.sqrt(dim)
Matthew Douglas's avatar
Matthew Douglas committed
1211
            elif kind == "attn_packed":
1212
1213
                A = torch.randn(1, dim, dtype=dtype, device=device)
                B = torch.randn(dim * 3, dim, dtype=dtype, device=device) / math.sqrt(dim)
Matthew Douglas's avatar
Matthew Douglas committed
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284

            qB, state = F.quantize_4bit(
                B,
                quant_type=storage_type,
                compress_statistics=double_quant,
                quant_storage=quant_storage,
            )
            C3 = torch.matmul(A, B.t())
            C2 = F.gemv_4bit(A, qB.t(), state=state)
            A.requires_grad = True
            C1 = bnb.matmul_4bit(A, qB.t(), state)

            err1 = (C1 - C2).abs().float()
            err2 = (C3 - C2).abs().float()
            err3 = (C3 - C1).abs().float()

            mag1 = torch.abs(C1).float() + 1e-5
            mag2 = torch.abs(C3).float() + 1e-5
            mag3 = torch.abs(C3).float() + 1e-5

            relerr1 = err1 / mag1
            relerr2 = err2 / mag2
            relerr3 = err3 / mag3

            max_err1 = err1.max()
            max_err2 = err2.max()
            max_err3 = err3.max()

            errs1.append(err1.mean().item())
            errs2.append(err2.mean().item())
            errs3.append(err3.mean().item())

            relerrs1.append(relerr1.mean().item())
            relerrs2.append(relerr2.mean().item())
            relerrs3.append(relerr3.mean().item())

            max_errs1.append(max_err1.item())
            max_errs2.append(max_err2.item())
            max_errs3.append(max_err3.item())

            c = int(C1.numel() * 0.0014 * (dim / 256)) + 1

            c = assert_all_approx_close(C1, C2, 1e-5, 0.01, count=0, throw=False)
        err1 = sum(errs1) / len(errs1) / math.sqrt(dim)
        err2 = sum(errs2) / len(errs2) / math.sqrt(dim)
        err3 = sum(errs3) / len(errs3) / math.sqrt(dim)
        relerr1 = sum(relerrs1) / len(relerrs1) / math.sqrt(dim)
        relerr2 = sum(relerrs2) / len(relerrs2) / math.sqrt(dim)
        relerr3 = sum(relerrs3) / len(relerrs3) / math.sqrt(dim)
        maxerr1 = sum(max_errs1) / len(max_errs1) / math.sqrt(dim)
        maxerr2 = sum(max_errs2) / len(max_errs2) / math.sqrt(dim)
        maxerr3 = sum(max_errs3) / len(max_errs3) / math.sqrt(dim)
        absratio = err2 / err3
        relratio = relerr2 / relerr3
        maxratio = relerr2 / relerr3

        # for debugging if the tests fails
        #
        # print('='*80)
        # print(f'For matmul: {A.shape}, {B.shape}, {kind}, {dtype}, {storage_type}, double_quant={double_quant}:')
        # print(C1.flatten()[-20:])
        # print(C2.flatten()[-20:])
        # print(f'inference vs training abs: {err1}')
        # print(f'inference vs training rel: {relerr1}')
        # print(f'inference vs training max: {maxerr1}')
        # print(f'inference vs training vs torch err ratio abs: {absratio}')
        # print(f'inference vs training vs torch err ratio rel: {relratio}')
        # print(f'inference vs training vs torch err ratio max: {maxratio}')
        if dtype == torch.float16:
            if dim <= 512:
                assert err1 < 7e-5
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296

                # TODO(matthewdouglas): On T4, dim=128-fp16-fc2-fp4-DQ will have relerror ~ 0.00092727
                if (
                    device == "cuda"
                    and double_quant
                    and storage_type == "fp4"
                    and kind == "fc2"
                    and torch.cuda.get_device_capability() == (7, 5)
                ):
                    assert relerr1 < 0.00093
                else:
                    assert relerr1 < 0.0008
Matthew Douglas's avatar
Matthew Douglas committed
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
            else:
                assert err1 < 6e-5
                assert relerr1 < 2e-4
            assert absratio < 1.005 and absratio > 0.995
            assert relratio < 1.005 and relratio > 0.995
            assert maxratio < 1.005 and maxratio > 0.995
        elif dtype == torch.float32:
            if dim <= 512:
                assert err1 < 5e-8
                assert relerr1 < 1e-6
                assert maxerr1 < 1e-7
            else:
                assert err1 < 5e-8
                assert relerr1 < 8e-6
                assert maxerr1 < 1e-7
            assert absratio < 1.005 and absratio > 0.995
            assert relratio < 1.005 and relratio > 0.995
            assert maxratio < 1.005 and maxratio > 0.995
        elif dtype == torch.bfloat16:
            if dim <= 512:
                assert err1 < 6e-4
                assert relerr1 < 0.007
                assert maxerr1 < 0.015
            else:
                assert err1 < 2e-4
                assert relerr1 < 0.002
                assert maxerr1 < 0.0012
            assert absratio < 1.005 and absratio > 0.995
            assert relratio < 1.04 and relratio > 0.96
            assert maxratio < 1.02 and maxratio > 0.98
1327

1328
    @pytest.mark.parametrize("device", get_available_devices())
1329
1330
1331
    @pytest.mark.parametrize("storage_type", ["nf4", "fp4"], ids=["nf4", "fp4"])
    @pytest.mark.parametrize("dtype", [torch.float16, torch.bfloat16, torch.float32], ids=describe_dtype)
    @pytest.mark.parametrize("double_quant", [False], ids=["DQ_True"])
1332
    def test_gemv_eye_4bit(self, device, storage_type, dtype, double_quant):
Matthew Douglas's avatar
Matthew Douglas committed
1333
1334
1335
        if device == "cpu" and dtype == torch.bfloat16 and torch.__version__ < (2, 3):
            pytest.skip("eye doe not support bfloat16 on CPU in torch < 2.3")

1336
1337
1338
1339
1340
1341
        dims = 10
        torch.random.manual_seed(np.random.randint(0, 412424242))
        dims = get_test_dims(0, 8192, n=dims)
        dims = [dim + (64 - (dim % 64)) for dim in dims]
        # for dim in [576, 5120, 3520, 5184, 1280, 4992, 5312, 2048]:
        for dim in dims:
1342
1343
            A = torch.normal(0, 0.1, size=(1, 1, dim), dtype=dtype, device=device)
            B = torch.eye(dim, dtype=dtype, device=device)
1344
1345
1346
1347
1348
1349

            qB, state = F.quantize_4bit(B, quant_type=storage_type, compress_statistics=double_quant)
            C3 = torch.matmul(A, B.t())
            C2 = bnb.matmul_4bit(A, qB.t(), state)
            A.requires_grad = True
            C1 = bnb.matmul_4bit(A, qB.t(), state)
1350

1351
1352
1353
1354
1355
            torch.testing.assert_close(A, C3)
            torch.testing.assert_close(A, C1)
            torch.testing.assert_close(A, C2)
        # torch.testing.assert_close(A, C1, rtol=1e-5, atol=0.00001)
        # torch.testing.assert_close(A, C2, rtol=1e-5, atol=0.080)
1356
1357
1358
1359


def test_normal_map_tree():
    code = F.create_normal_map()
Ruff's avatar
Ruff committed
1360
    values = code[:8].tolist() + code[-8:].tolist()
1361
    num_pivots = 1
Ruff's avatar
Ruff committed
1362
1363
1364
1365
    # print(values)
    while num_pivots < 16:
        idx = list(range(16 // num_pivots // 2, 16, 16 // num_pivots))
        # print(idx)
1366
1367
1368
        num_pivots *= 2
        pivots = []
        for i in idx:
Ruff's avatar
Ruff committed
1369
1370
            pivots.append((values[i - 1] + values[i]) / 2)
        # print(pivots)
1371

Tim Dettmers's avatar
Tim Dettmers committed
1372

1373
@pytest.mark.skip("Row scale has some bugs for ampere")
Tim Dettmers's avatar
Tim Dettmers committed
1374
def test_managed():
Ruff's avatar
Ruff committed
1375
    n = 32 * 10
Tim Dettmers's avatar
Tim Dettmers committed
1376
1377
1378
1379
1380
    A = F.get_paged(n, n, dtype=torch.float32)
    B = F.get_paged(n, n, dtype=torch.uint8)
    B2 = F.get_paged(n, n, dtype=torch.float32)
    assert A.is_paged
    assert B.is_paged
Ruff's avatar
Ruff committed
1381
1382
    assert A.page_deviceid == 0
    assert B.page_deviceid == 0
Tim Dettmers's avatar
Tim Dettmers committed
1383
1384
1385
    F.fill(A, 17.0)
    F.fill(B, 17)
    F.fill(B2, 2)
Ruff's avatar
Ruff committed
1386
1387
1388
1389
    assert (A == 17).sum().item() == n * n
    assert (B == 17).sum().item() == n * n
    C = A * B.float()
    assert (C == 289).sum().item() == n * n
Tim Dettmers's avatar
Tim Dettmers committed
1390
1391
1392
    F._mul(A, B2)
    F._mul(A, B2)
    F._mul(A, B2)
Ruff's avatar
Ruff committed
1393
    assert (A == 17 * (2**3)).sum().item() == n * n