"vscode:/vscode.git/clone" did not exist on "e38ee081a0495769e25766b894abe19bc8a6209e"
cuda.py 10.6 KB
Newer Older
1
2
import torch
from text_generation_server.utils.import_utils import SYSTEM
3
from text_generation_server.models.globals import (
4
    ATTENTION,
5
6
    BLOCK_SIZE,
)
7
from text_generation_server.layers.attention import Seqlen
8
from typing import Optional
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28

major, minor = torch.cuda.get_device_capability()
is_sm75 = major == 7 and minor == 5
_PARTITION_SIZE = 512

try:
    from vllm._C import cache_ops
except Exception as e:
    raise ImportError(
        f"Could not import vllm paged attention. Make sure your installation is correct. Complete error: {e}"
    )


def reshape_and_cache(
    key: torch.Tensor,
    value: torch.Tensor,
    key_cache: torch.Tensor,
    value_cache: torch.Tensor,
    slots: torch.Tensor,
):
29
    if ATTENTION in {"flashdecoding", "flashinfer"}:
30
31
32
33
34
35
36
        shape = key_cache.shape
        key_cache.view(-1, shape[-2], shape[-1])[slots] = key
        value_cache.view(-1, shape[-2], shape[-1])[slots] = value
    else:
        cache_ops.reshape_and_cache(
            key, value, key_cache, value_cache, slots, "auto", 1.0
        )
37
38
39
40
41
42
43
44
45


def paged_attention(
    query: torch.Tensor,
    key_cache: torch.Tensor,
    value_cache: torch.Tensor,
    kv_head_mapping: torch.Tensor,
    softmax_scale: float,
    block_tables: torch.Tensor,
46
    seqlen: Seqlen,
47
    max_s: int,
48
    softcap: Optional[float] = None,
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
):
    # Adapted from: https://github.com/vllm-project/vllm/blob/f8a1e39fae05ca610be8d5a78be9d40f5274e5fc/vllm/model_executor/layers/attention.py
    # Copyright 2023 The vLLM team. All rights
    # reserved.
    #
    # Licensed under the Apache License, Version 2.0 (the "License");
    # you may not use this file except in compliance with the License.
    # You may obtain a copy of the License at
    #
    #     http://www.apache.org/licenses/LICENSE-2.0
    #
    # Unless required by applicable law or agreed to in writing, software
    # distributed under the License is distributed on an "AS IS" BASIS,
    # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
    # See the License for the specific language governing permissions and
    # limitations under the License.
    #

    # value_cache => [num_blocks, num_heads, head_size, block_size]
68
69
    # block_size = value_cache.shape[3]
    block_size = BLOCK_SIZE
70
71
72
73
74
75
76
77
    num_seqs, num_heads, head_size = query.shape
    max_num_partitions = (max_s + _PARTITION_SIZE - 1) // _PARTITION_SIZE

    # NOTE(woosuk): We use a simple heuristic to decide whether to use
    # PagedAttention V1 or V2. If the number of partitions is 1, we use
    # V1 to avoid the overhead of reduction. Also, if the number of
    # sequences or heads is large, we use V1 since there is enough work
    # to parallelize.
78
    if ATTENTION == "flashinfer":
Nicolas Patry's avatar
Nicolas Patry committed
79
        from text_generation_server.layers.attention.flashinfer import decode_state
80
81
82
83
84
85
86

        return decode_state.get().forward(
            query.contiguous(),
            paged_kv_cache=(key_cache, value_cache),
            logits_soft_cap=softcap,
            sm_scale=softmax_scale,
        )
87
    elif ATTENTION == "flashdecoding":
88
89
90
        max_q = 1
        max_k = max_s
        import flash_attn_2_cuda
91

92
93
94
95
96
        # TODO fixme when flash contains the fix.
        # Number of splits is not correctly handled
        # by the current path
        # https://github.com/Dao-AILab/flash-attention/blob/320fb59487658f033f56711efd3d61b7c7a6f8f3/csrc/flash_attn/flash_api.cpp#L577
        # This fails becuase we're using causal, therefore window_right is set to 0 and the split logic is never applied.
97
98
        if softcap is None:
            softcap = 0.0
99
        out = flash_attn_2_cuda.varlen_fwd(
100
101
102
            query,
            key_cache,
            value_cache,
103
104
105
            None,
            seqlen.cu_seqlen_q,
            seqlen.cu_seqlen_k,
106
            None,  # pad_k
107
            None,
108
109
            block_tables,
            None,
110
111
112
113
114
115
116
117
            max_q,
            max_k,
            0.0,  # dropout
            softmax_scale,
            False,  # zero_tensors
            True,  # causal
            -1,  # Window_left
            -1,  # Window right
118
            softcap,
119
120
            False,  # return softmax
            None,  # generator
121
        )
122
        return out[0]
123
    else:
124
125
        if softcap is not None:
            raise RuntimeError("Paged attention doesn't support softcapping")
126
127
        input_lengths = seqlen.input_lengths
        from vllm._C import ops
128

129
130
        out = torch.empty_like(query)

131
132
        use_v1 = max_s <= 8192 and (
            max_num_partitions == 1 or num_seqs * num_heads > 512
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
        if use_v1:
            ops.paged_attention_v1(
                out,
                query,
                key_cache,
                value_cache,
                kv_head_mapping,
                softmax_scale,
                block_tables,
                input_lengths,
                block_size,
                max_s,
                None,
                "auto",
                1.0,
            )
        else:
            # Run PagedAttention V2.
            assert _PARTITION_SIZE % block_size == 0
            tmp_output = torch.empty(
                size=(num_seqs, num_heads, max_num_partitions, head_size),
                dtype=out.dtype,
                device=out.device,
            )
            exp_sums = torch.empty(
                size=(num_seqs, num_heads, max_num_partitions),
                dtype=torch.float32,
                device=out.device,
            )
            max_logits = torch.empty_like(exp_sums)

            ops.paged_attention_v2(
                out,
                exp_sums,
                max_logits,
                tmp_output,
                query,
                key_cache,
                value_cache,
                kv_head_mapping,
                softmax_scale,
                block_tables,
                input_lengths,
                block_size,
                max_s,
                None,
                "auto",
                1.0,
            )
    return out
184
185
186


try:
187
188
189
190
    is_ampere_or_newer = major >= 8 and minor >= 0
    if not is_ampere_or_newer:
        raise ImportError("FlashAttention only supports Ampere GPUs or newer.")

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
    import flash_attn_2_cuda

    V2 = True
except ImportError:
    try:
        import flash_attn_cuda

        V2 = False
    except ImportError as e:
        if major >= 8:
            architecture_suffix = f"-{SYSTEM}"
            raise ImportError(
                "Flash Attention V2 is not installed.\n"
                "Use the official Docker image (ghcr.io/huggingface/text-generation-inference:latest) "
                f"or install flash attention v2 with `cd server && make install install-flash-attention-v2{architecture_suffix}`"
            )
        elif is_sm75:
            raise ImportError(
                "Flash Attention is not installed.\n"
                "Use the official Docker image (ghcr.io/huggingface/text-generation-inference:latest) "
                "or install flash attention with `cd server && make install install-flash-attention`"
            ) from e
        else:
            raise ImportError(
                f"GPU with CUDA capability {major} {minor} is not supported"
            ) from e


SUPPORTS_WINDOWING = V2
220

221
if ATTENTION == "flashinfer":
222
223

    def attention(
Nicolas Patry's avatar
Nicolas Patry committed
224
225
226
        q: torch.Tensor,
        key_cache: torch.Tensor,
        value_cache: torch.Tensor,
227
228
        seqlen: Seqlen,
        block_tables: torch.Tensor,
229
230
231
232
233
        softmax_scale,
        window_size_left=-1,
        causal=True,
        softcap=0.0,
    ):
Nicolas Patry's avatar
Nicolas Patry committed
234
235
236
        from text_generation_server.layers.attention.flashinfer import (
            prefill_with_paged_kv_state,
        )
237

Nicolas Patry's avatar
Nicolas Patry committed
238
239
        return prefill_with_paged_kv_state.get().forward(
            q.contiguous(),
240
            causal=causal,
Nicolas Patry's avatar
Nicolas Patry committed
241
            paged_kv_cache=(key_cache, value_cache),
242
243
            logits_soft_cap=softcap,
            sm_scale=softmax_scale,
244
            window_left=window_size_left,
245
246
247
        )

elif V2:
248
249
250

    def attention(
        q,
Nicolas Patry's avatar
Nicolas Patry committed
251
252
        key_cache: torch.Tensor,
        value_cache: torch.Tensor,
253
254
        seqlen: Seqlen,
        block_tables: torch.Tensor,
255
256
257
        softmax_scale,
        window_size_left=-1,
        causal=True,
258
        softcap=0.0,
259
    ):
260
        out = torch.empty_like(q)
261
262
263
264
        if window_size_left <= 0 and window_size_left != -1:
            raise ValueError("`window_size_left` must be > 0 or -1")
        return flash_attn_2_cuda.varlen_fwd(
            q,
265
266
            key_cache,
            value_cache,
267
            out,
268
269
            seqlen.cu_seqlen_q,
            seqlen.cu_seqlen_k,
270
271
            None,
            None,
272
            block_tables,
273
            None,
274
275
            seqlen.max_q,
            seqlen.max_k,
276
277
278
279
280
281
            0.0,
            softmax_scale,
            False,
            causal,
            window_size_left,
            0,
282
            softcap,
283
284
            False,
            None,
285
        )[0]
286
287
288
289
290
291
292

else:

    def attention(
        q,
        k,
        v,
Nicolas Patry's avatar
Nicolas Patry committed
293
294
        key_cache: torch.Tensor,
        value_cache: torch.Tensor,
295
296
297
298
        cu_seqlens,
        max_s,
        softmax_scale,
        window_size_left=-1,
299
        causal=None,
300
        softcap=None,
301
302
303
304
305
    ):
        if window_size_left != -1:
            raise NotImplementedError(
                "window_size_left is only available with flash attn v2"
            )
306
307
        if softcap is not None:
            raise NotImplementedError("softcap is only available with flash attn v2")
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334

        # Flash attention v1 requires q, k and v to have the same number of heads
        if k.shape[1] != q.shape[1]:
            # MQA expand
            if k.shape[1] == 1:
                k = k.expand(-1, q.shape[1], -1)
            # Grouped attention reshape
            else:
                original_shape = k.shape
                k = (
                    k.unsqueeze(2)
                    .expand(-1, -1, q.shape[1] // k.shape[1], -1)
                    .reshape(original_shape[0], -1, original_shape[2])
                )
        if v.shape[1] != q.shape[1]:
            # MQA expand
            if v.shape[1] == 1:
                v = v.expand(-1, q.shape[1], -1)
            # Grouped attention reshape
            else:
                original_shape = v.shape
                v = (
                    v.unsqueeze(2)
                    .expand(-1, -1, q.shape[1] // v.shape[1], -1)
                    .reshape(original_shape[0], -1, original_shape[2])
                )

335
        out = torch.empty_like(q)
336
        flash_attn_cuda.fwd(
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
            q,
            k,
            v,
            out,
            cu_seqlens,
            cu_seqlens,
            max_s,
            max_s,
            0.0,
            softmax_scale,
            False,
            True,
            False,
            0,
            None,
352
353
        )
        return out