test_cutlass.py 16.8 KB
Newer Older
1
2
3
4
"""Tests for cutlass kernels

Run `pytest tests/kernels/test_cutlass.py`.
"""
5
from typing import Optional, Type
6
7
8
9

import pytest
import torch

10
from tests.kernels.utils import opcheck
11
from vllm import _custom_ops as ops
12
from vllm.platforms import current_platform
13

14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
MNK_FACTORS = [
    (1, 256, 128),
    (1, 16384, 1024),
    (1, 24576, 496),
    (16, 256, 496),
    (16, 16384, 128),
    (16, 24576, 4096),
    (32, 8192, 4096),
    (32, 16384, 4096),
    (33, 1024, 1024),
    (33, 8192, 128),
    (64, 2048, 496),
    (64, 16384, 1024),
    (100, 8192, 496),
    (128, 32768, 4096),
    (256, 4096, 4096),
    (512, 256, 1024),
    (512, 8192, 4096),
    (512, 16384, 128),
    (512, 24576, 128),
]
35
36

CUDA_DEVICES = [
37
    f"cuda:{i}" for i in range(1 if torch.cuda.device_count() == 1 else 2)
38
39
]

40
41
capability = current_platform.get_device_capability()
capability = capability[0] * 10 + capability[1]
42

43

44
def to_fp8(tensor: torch.Tensor):
45
46
47
48
49
    finfo = torch.finfo(torch.float8_e4m3fn)
    return torch.round(tensor.clamp(
        min=finfo.min, max=finfo.max)).to(dtype=torch.float8_e4m3fn)


50
def to_int8(tensor: torch.Tensor):
51
52
53
    return torch.round(tensor.clamp(min=-128, max=127)).to(dtype=torch.int8)


54
55
56
57
def rand_int8(shape: tuple, device: str = "cuda"):
    return to_int8(torch.rand(shape, device=device) * 255 - 128)


58
59
60
61
62
63
def baseline_scaled_mm(a: torch.Tensor,
                       b: torch.Tensor,
                       scale_a: torch.Tensor,
                       scale_b: torch.Tensor,
                       out_dtype: Type[torch.dtype],
                       bias: Optional[torch.Tensor] = None) -> torch.Tensor:
gaoqiong's avatar
gaoqiong committed
64
    output = (scale_a * (scale_b.T * (torch.mm(
65
66
67
68
69
70
71
        a.to(dtype=torch.float32), b.to(dtype=torch.float32))))).to(out_dtype)
    if bias is not None:
        output = output + bias

    return output


72
73
74
75
76
def cutlass_fp8_gemm_helper(m: int,
                            n: int,
                            k: int,
                            per_token_act_quant: bool,
                            per_out_channel_weight_quant: bool,
77
                            use_bias: bool,
78
79
80
81
82
83
84
85
86
87
                            out_dtype: Type[torch.dtype] = torch.bfloat16,
                            device: str = "cuda"):
    # Test for a cutlass kernel with per-token activation quantization
    # and per-output channel weight quantization.
    a = to_fp8(torch.randn((m, k), device=device))
    b = to_fp8(torch.randn((n, k), device=device).t())

    m_a_scales = m if per_token_act_quant else 1
    n_b_scales = n if per_out_channel_weight_quant else 1

88
89
90
91
92
93
    scale_a = (torch.randn((m_a_scales, 1), device=device,
                           dtype=torch.float32))
    scale_b = (torch.randn((1, n_b_scales), device=device,
                           dtype=torch.float32))
    if use_bias:
        bias = torch.rand((n, ), device=device, dtype=out_dtype) * 10
94
    else:
95
        bias = None
96

97
98
    out = ops.cutlass_scaled_mm(a, b, scale_a, scale_b, out_dtype, bias)
    baseline = baseline_scaled_mm(a, b, scale_a, scale_b, out_dtype, bias)
99

100
    torch.testing.assert_close(out, baseline, rtol=1e-2, atol=5e-2)
101

102
103
104
    opcheck(torch.ops._C.cutlass_scaled_mm,
            (out, a, b, scale_a, scale_b, bias))

105
106
107
108
109
110

def cutlass_int8_gemm_helper(m: int,
                             n: int,
                             k: int,
                             per_token_act_quant: bool,
                             per_out_channel_weight_quant: bool,
111
                             use_bias: bool,
112
113
114
115
116
117
118
119
120
121
                             out_dtype: Type[torch.dtype] = torch.bfloat16,
                             device: str = "cuda"):
    # Test for a cutlass kernel with per-token activation quantization
    # and per-output channel weight quantization.
    a = to_int8(torch.randn((m, k), device=device) * 5)
    b = to_int8(torch.randn((n, k), device=device).t() * 5)

    m_a_scales = m if per_token_act_quant else 1
    n_b_scales = n if per_out_channel_weight_quant else 1

122
123
124
125
    scale_a = (torch.randn((m_a_scales, 1), device=device,
                           dtype=torch.float32))
    scale_b = (torch.randn((1, n_b_scales), device=device,
                           dtype=torch.float32))
126

127
128
    if use_bias:
        bias = torch.rand((n, ), device=device, dtype=out_dtype) * 10
129
    else:
130
131
132
133
        bias = None

    out = ops.cutlass_scaled_mm(a, b, scale_a, scale_b, out_dtype, bias)
    baseline = baseline_scaled_mm(a, b, scale_a, scale_b, out_dtype, bias)
134

135
    torch.testing.assert_close(out, baseline, rtol=1e-1, atol=1e0)
136

137
138
139
    opcheck(torch.ops._C.cutlass_scaled_mm,
            (out, a, b, scale_a, scale_b, bias))

140

zhuwenwen's avatar
zhuwenwen committed
141
# @pytest.mark.parametrize("m,n,k", MNK_FACTORS)
gaoqiong's avatar
gaoqiong committed
142
143
144
145
146
147
148
149
# @pytest.mark.parametrize("per_act_token", [True, False])
# @pytest.mark.parametrize("per_out_ch", [True, False])
# @pytest.mark.parametrize("use_bias", [True, False])
# @pytest.mark.skipif(not current_platform.has_device_capability(89),
#                     reason="FP8 is not supported on this GPU type.")
# def test_cutlass_fp8_gemm(m: int, n: int, k: int, per_act_token: bool,
#                           per_out_ch: bool, use_bias: bool):
#     cutlass_fp8_gemm_helper(m, n, k, per_act_token, per_out_ch, use_bias)
150
151


152
@pytest.mark.parametrize("m,n,k", MNK_FACTORS)
153
154
@pytest.mark.parametrize("per_act_token", [True, False])
@pytest.mark.parametrize("per_out_ch", [True, False])
155
@pytest.mark.parametrize("use_bias", [True, False])
156
def test_cutlass_int8_gemm(m: int, n: int, k: int, per_act_token: bool,
157
158
                           per_out_ch: bool, use_bias: bool):
    cutlass_int8_gemm_helper(m, n, k, per_act_token, per_out_ch, use_bias)
159
160


gaoqiong's avatar
gaoqiong committed
161
162
163
@pytest.mark.parametrize("per_act_token", [True])
@pytest.mark.parametrize("per_out_ch", [True])
@pytest.mark.parametrize("out_dtype", [ torch.float16]) #torch.bfloat16,
164
@pytest.mark.parametrize("use_bias", [True, False])
165
def test_cutlass_int8_gemm_output_dtype(per_act_token: bool, per_out_ch: bool,
166
                                        out_dtype: Type[torch.dtype],
167
                                        use_bias: bool):
168
169
170
171
172
    cutlass_int8_gemm_helper(512,
                             512,
                             512,
                             per_act_token,
                             per_out_ch,
173
                             use_bias,
174
                             out_dtype=out_dtype)
175
176


gaoqiong's avatar
gaoqiong committed
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
# @pytest.mark.parametrize("per_act_token", [True, False])
# @pytest.mark.parametrize("per_out_ch", [True, False])
# @pytest.mark.parametrize("out_dtype", [torch.bfloat16, torch.float16])
# @pytest.mark.parametrize("use_bias", [True, False])
# @pytest.mark.skipif(not current_platform.has_device_capability(89),
#                     reason="FP8 is not supported on this GPU type.")
# def test_cutlass_fp8_gemm_output_dtype(per_act_token: bool, per_out_ch: bool,
#                                        out_dtype: Type[torch.dtype],
#                                        use_bias: bool):
#     cutlass_fp8_gemm_helper(512,
#                             512,
#                             512,
#                             per_act_token,
#                             per_out_ch,
#                             use_bias,
#                             out_dtype=out_dtype)


# @pytest.mark.parametrize("per_act_token", [True, False])
# @pytest.mark.parametrize("per_out_ch", [True, False])
# @pytest.mark.parametrize("use_bias", [True, False])
# @pytest.mark.parametrize("device", CUDA_DEVICES)
# @pytest.mark.skipif(not current_platform.has_device_capability(89),
#                     reason="FP8 is not supported on this GPU type.")
# def test_cutlass_fp8_gemm_devices(per_act_token: bool, per_out_ch: bool,
#                                   use_bias: bool, device: str):
#     cutlass_fp8_gemm_helper(512, 512, 512, per_act_token, per_out_ch, use_bias,
#                             torch.bfloat16, device)


# @pytest.mark.parametrize("per_act_token", [True, False])
# @pytest.mark.parametrize("per_out_ch", [True, False])
# @pytest.mark.parametrize("use_bias", [True, False])
# @pytest.mark.parametrize("device", CUDA_DEVICES)
# def test_cutlass_int8_gemm_devices(per_act_token: bool, per_out_ch: bool,
#                                    use_bias: bool, device: str):
#     cutlass_int8_gemm_helper(512,
#                              512,
#                              512,
#                              per_act_token,
#                              per_out_ch,
#                              use_bias,
#                              out_dtype=torch.bfloat16,
#                              device=device)
221
222
223
224
225
226
227


# For the following two tests:
# N and K correspond to the size of the weight matrix and likely to be multiples
# of a large power of two. In any case, the kernel will have a naive fallback
# when N and K are not divisible by 16. But M is the number of tokens and the
# kernel must handle any M thrown at it.
gaoqiong's avatar
gaoqiong committed
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
# @pytest.mark.parametrize("per_act_token", [True, False])
# @pytest.mark.parametrize("per_out_ch", [True, False])
# @pytest.mark.parametrize("use_bias", [True, False])
# @pytest.mark.skipif(not current_platform.has_device_capability(89),
#                     reason="FP8 is not supported on this GPU type.")
# def test_cutlass_fp8_gemm_m_sweep(per_act_token: bool, per_out_ch: bool,
#                                   use_bias: bool):
#     for nk in range(32, 128, 32):
#         for m in range(1, 128):
#             cutlass_fp8_gemm_helper(m, nk, nk, per_act_token, per_out_ch,
#                                     use_bias)


# @pytest.mark.parametrize("per_act_token", [True, False])
# @pytest.mark.parametrize("per_out_ch", [True, False])
# @pytest.mark.parametrize("use_bias", [True, False])
# def test_cutlass_int8_gemm_m_sweep(per_act_token: bool, per_out_ch: bool,
#                                    use_bias: bool):
#     for nk in range(32, 128, 32):
#         for m in range(1, 128):
#             cutlass_int8_gemm_helper(m, nk, nk, per_act_token, per_out_ch,
#                                      use_bias)


# @pytest.mark.parametrize("m", [32, 64, 128])
# @pytest.mark.parametrize("n", [16, 32, 64])
# @pytest.mark.parametrize("k", [64, 128, 256])
# @pytest.mark.parametrize("out_dtype", [torch.bfloat16, torch.float16])
# @pytest.mark.skip
# def test_cutlass_int8_azp_bias_fold(m: int, n: int, k: int,
#                                     out_dtype: torch.dtype):
#     # Currently, the test is failing because folding azp into
#     # 16-bit bias loses too much precision
#     scale_a = torch.randn((1, 1), device="cuda", dtype=torch.float32) / 10
#     scale_b = torch.randn((1, n), device="cuda", dtype=torch.float32) / 10

#     aq_i8 = rand_int8((m, k))
#     bq_i8 = rand_int8((n, k)).t()

#     aq_i32 = aq_i8.to(dtype=torch.int32)
#     bq_i32 = bq_i8.to(dtype=torch.int32)

#     aq_f32 = aq_i8.to(dtype=torch.float32)
#     bq_f32 = bq_i8.to(dtype=torch.float32)

#     b_dq = scale_b * bq_f32

#     azp_a = torch.rand((1, ), device="cuda", dtype=torch.float32) * 10 + 1.5
#     azp_aq_i8 = (azp_a / scale_a).to(dtype=torch.int8)
#     azp_a = azp_aq_i8.to(dtype=torch.float32) * scale_a  # correct for rounding

#     a_dq = scale_a * (aq_i32 + azp_aq_i8).to(dtype=torch.float32)
#     torch.testing.assert_close(a_dq, scale_a * aq_f32 + azp_a)

#     baseline_dq = torch.mm(a_dq, b_dq).to(out_dtype)

#     J = torch.ones((1, k), device="cuda", dtype=torch.float32)
#     azp_bias = (azp_a * scale_b * (J @ bq_f32)).to(out_dtype)
#     assert azp_bias.shape == (1, n)
#     assert azp_bias[0, :].shape == (n, )

#     baseline_q = (scale_a.to(device='cpu') * scale_b.to(device='cpu') * (
#         (aq_i32 + azp_aq_i8).to(device='cpu') @ bq_i32.to(device='cpu'))).to(
#             dtype=out_dtype, device='cuda')

#     out = ops.cutlass_scaled_mm(aq_i8,
#                                 bq_i8,
#                                 scale_a,
#                                 scale_b,
#                                 out_dtype=out_dtype,
#                                 bias=azp_bias[0, :])
#     torch.testing.assert_close(out, baseline_dq, rtol=1e-2, atol=1e0)
#     torch.testing.assert_close(out, baseline_q, rtol=1e-2, atol=1e0)


# @pytest.mark.parametrize("m", [32, 64, 128])
# @pytest.mark.parametrize("n", [16, 32, 64])
# @pytest.mark.parametrize("k", [64, 128, 256])
# @pytest.mark.parametrize("out_dtype", [torch.bfloat16, torch.float16])
zhuwenwen's avatar
zhuwenwen committed
307
308
309
310
311
312
# @pytest.mark.skip
# def test_cutlass_int8_azp_bias_fold(m: int, n: int, k: int,
#                                     out_dtype: torch.dtype):
#     # Currently, the test is failing because folding azp into
#     # 16-bit bias loses too much precision
#     scale_a = torch.randn((1, 1), device="cuda", dtype=torch.float32) / 10
gaoqiong's avatar
gaoqiong committed
313
314
315
316
#     scale_b = torch.randn((1, n), device="cuda", dtype=torch.float32) / 10

#     aq_i8 = rand_int8((m, k))
#     bq_i8 = rand_int8((n, k)).t()
zhuwenwen's avatar
zhuwenwen committed
317
318

#     aq_i32 = aq_i8.to(dtype=torch.int32)
gaoqiong's avatar
gaoqiong committed
319
#     bq_i32 = bq_i8.to(dtype=torch.int32)
zhuwenwen's avatar
zhuwenwen committed
320
321

#     aq_f32 = aq_i8.to(dtype=torch.float32)
gaoqiong's avatar
gaoqiong committed
322
#     bq_f32 = bq_i8.to(dtype=torch.float32)
zhuwenwen's avatar
zhuwenwen committed
323

gaoqiong's avatar
gaoqiong committed
324
325
#     b_dq = scale_b * bq_f32

zhuwenwen's avatar
zhuwenwen committed
326
#     azp_a = torch.rand((1, ), device="cuda", dtype=torch.float32) * 10 + 1.5
gaoqiong's avatar
gaoqiong committed
327
328
329
#     azp_aq_i8 = (azp_a / scale_a).to(dtype=torch.int8)
#     azp_a = azp_aq_i8.to(dtype=torch.float32) * scale_a  # correct for rounding

zhuwenwen's avatar
zhuwenwen committed
330
331
332
333
334
335
336
337
338
#     a_dq = scale_a * (aq_i32 + azp_aq_i8).to(dtype=torch.float32)
#     torch.testing.assert_close(a_dq, scale_a * aq_f32 + azp_a)

#     baseline_dq = torch.mm(a_dq, b_dq).to(out_dtype)

#     J = torch.ones((1, k), device="cuda", dtype=torch.float32)
#     azp_bias = (azp_a * scale_b * (J @ bq_f32)).to(out_dtype)
#     assert azp_bias.shape == (1, n)
#     assert azp_bias[0, :].shape == (n, )
339

zhuwenwen's avatar
zhuwenwen committed
340
341
342
343
344
345
346
347
348
349
350
351
#     baseline_q = (scale_a.to(device='cpu') * scale_b.to(device='cpu') * (
#         (aq_i32 + azp_aq_i8).to(device='cpu') @ bq_i32.to(device='cpu'))).to(
#             dtype=out_dtype, device='cuda')

#     out = ops.cutlass_scaled_mm(aq_i8,
#                                 bq_i8,
#                                 scale_a,
#                                 scale_b,
#                                 out_dtype=out_dtype,
#                                 bias=azp_bias[0, :])
#     torch.testing.assert_close(out, baseline_dq, rtol=1e-2, atol=1e0)
#     torch.testing.assert_close(out, baseline_q, rtol=1e-2, atol=1e0)
352

353

354
355
356
357
358
359
360
361
362
# Test working with a subset of A and B
def test_cutlass_subset():
    big_m, big_n, big_k = 1024, 1024, 1024
    m, n, k = 512, 512, 512

    whole_a = to_int8(torch.randn((big_m, big_k), device="cuda") * 5)
    whole_b = to_int8(torch.randn((big_n, big_k), device="cuda").t() * 5)
    a = whole_a[0:m, 0:k]
    b = whole_b[0:k, 0:n]
gaoqiong's avatar
gaoqiong committed
363
364
365
366
367
    
    #变成连续内存,矩阵子模块目前不支持计算,需要重新计算lda
    a=a.contiguous().reshape(m,-1)
    b=b.contiguous().reshape(k,-1)
    
368
369
370
    scale_a = torch.randn((1, 1), device="cuda", dtype=torch.float32) / 10
    scale_b = torch.randn((1, 1), device="cuda", dtype=torch.float32) / 10

371
372
373
374
375
    out = ops.cutlass_scaled_mm(a,
                                b,
                                scale_a,
                                scale_b,
                                out_dtype=torch.bfloat16)
376
377
378
379
380
    baseline = baseline_scaled_mm(a,
                                  b,
                                  scale_a,
                                  scale_b,
                                  out_dtype=torch.bfloat16)
381

382
    torch.testing.assert_close(out, baseline, rtol=1e-1, atol=1e0)
383
384
385
386
387
388
389
390
391
392
393
394
395


# Test to make sure cuda graphs work
class CutlassLayer(torch.nn.Module):

    def __init__(self, b, scale_a, scale_b, out_dtype):
        super().__init__()
        self.b = b
        self.scale_a = scale_a
        self.scale_b = scale_b
        self.out_dtype = out_dtype

    def forward(self, a):
396
397
        return ops.cutlass_scaled_mm(a, self.b, self.scale_a, self.scale_b,
                                     self.out_dtype)
398

gaoqiong's avatar
gaoqiong committed
399
400
401
#目前只支持per-act-token+per-out-ch(fp16)
@pytest.mark.parametrize("per_act_token", [True])
@pytest.mark.parametrize("per_out_ch", [True])
402
def test_cutlass_cuda_graph(per_act_token: bool, per_out_ch: bool):
403
404
405
406
    m, n, k = 512, 512, 512

    a = to_int8(torch.randn((m, k), device="cuda"))
    b = to_int8(torch.randn((n, k), device="cuda").t())
gaoqiong's avatar
gaoqiong committed
407
408
    b=b.contiguous().reshape(k,-1)
    
409
410
411
412
413
414
    m_a_scales = m if per_act_token else 1
    n_b_scales = n if per_out_ch else 1

    scale_a = (torch.randn(
        (m_a_scales, 1), device="cuda", dtype=torch.float32) / 10)
    scale_b = (torch.randn(
gaoqiong's avatar
gaoqiong committed
415
        (n_b_scales,1), device="cuda", dtype=torch.float32) / 10)
416
417

    # Construct a trivial model with a single layer that calls a CUTLASS kernel
gaoqiong's avatar
gaoqiong committed
418
    model = CutlassLayer(b, scale_a, scale_b, torch.float16)
419
420
421
422
423
424
425
426
427
428
429

    # Run the model with a cuda graph
    stream = torch.cuda.Stream()
    with torch.cuda.stream(stream):
        g = torch.cuda.CUDAGraph()
        with torch.cuda.graph(g):
            out = model(a)
    out.zero_()
    g.replay()

    baseline = torch.mm(scale_a * a.to(dtype=torch.float32),
gaoqiong's avatar
gaoqiong committed
430
431
432
433
434
                        scale_b.T * b.to(dtype=torch.float32)).to(torch.float16)
    #print("baseline:",baseline)
    out=ops.cutlass_scaled_mm(a, b, scale_a, scale_b,
                                     torch.float16)
    #print("out:",out)
435
    torch.testing.assert_close(out, baseline, rtol=1e-1, atol=1e0)