File size: 10,164 Bytes
7e93a0e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import einops
import torch
import torch.nn.functional as F
import torch.utils.benchmark as benchmark
from torch.backends.cuda import SDPBackend

from sgm.modules.attention import BasicTransformerBlock, SpatialTransformer


def benchmark_attn():
    # Lets define a helpful benchmarking function:
    # https://pytorch.org/tutorials/intermediate/scaled_dot_product_attention_tutorial.html
    device = "cuda" if torch.cuda.is_available() else "cpu"

    def benchmark_torch_function_in_microseconds(f, *args, **kwargs):
        t0 = benchmark.Timer(
            stmt="f(*args, **kwargs)", globals={"args": args, "kwargs": kwargs, "f": f}
        )
        return t0.blocked_autorange().mean * 1e6

    # Lets define the hyper-parameters of our input
    batch_size = 32
    max_sequence_len = 1024
    num_heads = 32
    embed_dimension = 32

    dtype = torch.float16

    query = torch.rand(
        batch_size,
        num_heads,
        max_sequence_len,
        embed_dimension,
        device=device,
        dtype=dtype,
    )
    key = torch.rand(
        batch_size,
        num_heads,
        max_sequence_len,
        embed_dimension,
        device=device,
        dtype=dtype,
    )
    value = torch.rand(
        batch_size,
        num_heads,
        max_sequence_len,
        embed_dimension,
        device=device,
        dtype=dtype,
    )

    print(f"q/k/v shape:", query.shape, key.shape, value.shape)

    # Lets explore the speed of each of the 3 implementations
    from torch.backends.cuda import SDPBackend, sdp_kernel

    # Helpful arguments mapper
    backend_map = {
        SDPBackend.MATH: {
            "enable_math": True,
            "enable_flash": False,
            "enable_mem_efficient": False,
        },
        SDPBackend.FLASH_ATTENTION: {
            "enable_math": False,
            "enable_flash": True,
            "enable_mem_efficient": False,
        },
        SDPBackend.EFFICIENT_ATTENTION: {
            "enable_math": False,
            "enable_flash": False,
            "enable_mem_efficient": True,
        },
    }

    from torch.profiler import ProfilerActivity, profile, record_function

    activities = [ProfilerActivity.CPU, ProfilerActivity.CUDA]

    print(
        f"The default implementation runs in {benchmark_torch_function_in_microseconds(F.scaled_dot_product_attention, query, key, value):.3f} microseconds"
    )
    with profile(
        activities=activities, record_shapes=False, profile_memory=True
    ) as prof:
        with record_function("Default detailed stats"):
            for _ in range(25):
                o = F.scaled_dot_product_attention(query, key, value)
    print(prof.key_averages().table(sort_by="cuda_time_total", row_limit=10))

    print(
        f"The math implementation runs in {benchmark_torch_function_in_microseconds(F.scaled_dot_product_attention, query, key, value):.3f} microseconds"
    )
    with sdp_kernel(**backend_map[SDPBackend.MATH]):
        with profile(
            activities=activities, record_shapes=False, profile_memory=True
        ) as prof:
            with record_function("Math implmentation stats"):
                for _ in range(25):
                    o = F.scaled_dot_product_attention(query, key, value)
        print(prof.key_averages().table(sort_by="cuda_time_total", row_limit=10))

    with sdp_kernel(**backend_map[SDPBackend.FLASH_ATTENTION]):
        try:
            print(
                f"The flash attention implementation runs in {benchmark_torch_function_in_microseconds(F.scaled_dot_product_attention, query, key, value):.3f} microseconds"
            )
        except RuntimeError:
            print("FlashAttention is not supported. See warnings for reasons.")
        with profile(
            activities=activities, record_shapes=False, profile_memory=True
        ) as prof:
            with record_function("FlashAttention stats"):
                for _ in range(25):
                    o = F.scaled_dot_product_attention(query, key, value)
        print(prof.key_averages().table(sort_by="cuda_time_total", row_limit=10))

    with sdp_kernel(**backend_map[SDPBackend.EFFICIENT_ATTENTION]):
        try:
            print(
                f"The memory efficient implementation runs in {benchmark_torch_function_in_microseconds(F.scaled_dot_product_attention, query, key, value):.3f} microseconds"
            )
        except RuntimeError:
            print("EfficientAttention is not supported. See warnings for reasons.")
        with profile(
            activities=activities, record_shapes=False, profile_memory=True
        ) as prof:
            with record_function("EfficientAttention stats"):
                for _ in range(25):
                    o = F.scaled_dot_product_attention(query, key, value)
        print(prof.key_averages().table(sort_by="cuda_time_total", row_limit=10))


def run_model(model, x, context):
    return model(x, context)


def benchmark_transformer_blocks():
    device = "cuda" if torch.cuda.is_available() else "cpu"
    import torch.utils.benchmark as benchmark

    def benchmark_torch_function_in_microseconds(f, *args, **kwargs):
        t0 = benchmark.Timer(
            stmt="f(*args, **kwargs)", globals={"args": args, "kwargs": kwargs, "f": f}
        )
        return t0.blocked_autorange().mean * 1e6

    checkpoint = True
    compile = False

    batch_size = 32
    h, w = 64, 64
    context_len = 77
    embed_dimension = 1024
    context_dim = 1024
    d_head = 64

    transformer_depth = 4

    n_heads = embed_dimension // d_head

    dtype = torch.float16

    model_native = SpatialTransformer(
        embed_dimension,
        n_heads,
        d_head,
        context_dim=context_dim,
        use_linear=True,
        use_checkpoint=checkpoint,
        attn_type="softmax",
        depth=transformer_depth,
        sdp_backend=SDPBackend.FLASH_ATTENTION,
    ).to(device)
    model_efficient_attn = SpatialTransformer(
        embed_dimension,
        n_heads,
        d_head,
        context_dim=context_dim,
        use_linear=True,
        depth=transformer_depth,
        use_checkpoint=checkpoint,
        attn_type="softmax-xformers",
    ).to(device)
    if not checkpoint and compile:
        print("compiling models")
        model_native = torch.compile(model_native)
        model_efficient_attn = torch.compile(model_efficient_attn)

    x = torch.rand(batch_size, embed_dimension, h, w, device=device, dtype=dtype)
    c = torch.rand(batch_size, context_len, context_dim, device=device, dtype=dtype)

    from torch.profiler import ProfilerActivity, profile, record_function

    activities = [ProfilerActivity.CPU, ProfilerActivity.CUDA]

    with torch.autocast("cuda"):
        print(
            f"The native model runs in {benchmark_torch_function_in_microseconds(model_native.forward, x, c):.3f} microseconds"
        )
        print(
            f"The efficientattn model runs in {benchmark_torch_function_in_microseconds(model_efficient_attn.forward, x, c):.3f} microseconds"
        )

        print(75 * "+")
        print("NATIVE")
        print(75 * "+")
        torch.cuda.reset_peak_memory_stats()
        with profile(
            activities=activities, record_shapes=False, profile_memory=True
        ) as prof:
            with record_function("NativeAttention stats"):
                for _ in range(25):
                    model_native(x, c)
        print(prof.key_averages().table(sort_by="cuda_time_total", row_limit=10))
        print(torch.cuda.max_memory_allocated() * 1e-9, "GB used by native block")

        print(75 * "+")
        print("Xformers")
        print(75 * "+")
        torch.cuda.reset_peak_memory_stats()
        with profile(
            activities=activities, record_shapes=False, profile_memory=True
        ) as prof:
            with record_function("xformers stats"):
                for _ in range(25):
                    model_efficient_attn(x, c)
        print(prof.key_averages().table(sort_by="cuda_time_total", row_limit=10))
        print(torch.cuda.max_memory_allocated() * 1e-9, "GB used by xformers block")


def test01():
    # conv1x1 vs linear
    from sgm.util import count_params

    conv = torch.nn.Conv2d(3, 32, kernel_size=1).cuda()
    print(count_params(conv))
    linear = torch.nn.Linear(3, 32).cuda()
    print(count_params(linear))

    print(conv.weight.shape)

    # use same initialization
    linear.weight = torch.nn.Parameter(conv.weight.squeeze(-1).squeeze(-1))
    linear.bias = torch.nn.Parameter(conv.bias)

    print(linear.weight.shape)

    x = torch.randn(11, 3, 64, 64).cuda()

    xr = einops.rearrange(x, "b c h w -> b (h w) c").contiguous()
    print(xr.shape)
    out_linear = linear(xr)
    print(out_linear.mean(), out_linear.shape)

    out_conv = conv(x)
    print(out_conv.mean(), out_conv.shape)
    print("done with test01.\n")


def test02():
    # try cosine flash attention
    import time

    torch.backends.cuda.matmul.allow_tf32 = True
    torch.backends.cudnn.allow_tf32 = True
    torch.backends.cudnn.benchmark = True
    print("testing cosine flash attention...")
    DIM = 1024
    SEQLEN = 4096
    BS = 16

    print(" softmax (vanilla) first...")
    model = BasicTransformerBlock(
        dim=DIM,
        n_heads=16,
        d_head=64,
        dropout=0.0,
        context_dim=None,
        attn_mode="softmax",
    ).cuda()
    try:
        x = torch.randn(BS, SEQLEN, DIM).cuda()
        tic = time.time()
        y = model(x)
        toc = time.time()
        print(y.shape, toc - tic)
    except RuntimeError as e:
        # likely oom
        print(str(e))

    print("\n now flash-cosine...")
    model = BasicTransformerBlock(
        dim=DIM,
        n_heads=16,
        d_head=64,
        dropout=0.0,
        context_dim=None,
        attn_mode="flash-cosine",
    ).cuda()
    x = torch.randn(BS, SEQLEN, DIM).cuda()
    tic = time.time()
    y = model(x)
    toc = time.time()
    print(y.shape, toc - tic)
    print("done with test02.\n")


if __name__ == "__main__":
    # test01()
    # test02()
    # test03()

    # benchmark_attn()
    benchmark_transformer_blocks()

    print("done.")