layer.py 20.2 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
"""Attention layer."""
4
from typing import List, Optional
5
6
7

import torch
import torch.nn as nn
8
import torch.nn.functional as F
9

10
import vllm.envs as envs
11
from vllm.attention import AttentionType
12
from vllm.attention.selector import backend_name_to_enum, get_attn_backend
13
from vllm.attention.utils.kv_sharing_utils import validate_kv_sharing_target
14
from vllm.config import CacheConfig, get_current_vllm_config
15
16
17
from vllm.distributed.kv_transfer import (get_kv_transfer_group,
                                          has_kv_transfer_group,
                                          is_v1_kv_transfer_group)
18
from vllm.forward_context import ForwardContext, get_forward_context
19
from vllm.logger import init_logger
20
from vllm.model_executor.layers.linear import UnquantizedLinearMethod
21
22
from vllm.model_executor.layers.quantization.base_config import (
    QuantizationConfig)
23
from vllm.model_executor.layers.quantization.kv_cache import BaseKVCacheMethod
24
from vllm.platforms import _Backend, current_platform
25
from vllm.utils import direct_register_custom_op
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
logger = init_logger(__name__)
USE_XFORMERS_OPS = None


def check_xformers_availability():
    global USE_XFORMERS_OPS
    if USE_XFORMERS_OPS is not None:
        return USE_XFORMERS_OPS

    if current_platform.is_cuda() and current_platform.has_device_capability(
            100):
        # Xformers FA is not compatible with B200
        USE_XFORMERS_OPS = False
    else:
        try:
            from importlib.util import find_spec

            find_spec("xformers.ops")
            USE_XFORMERS_OPS = True
        except ImportError:
            USE_XFORMERS_OPS = False

    # the warning only needs to be shown once
    if not USE_XFORMERS_OPS:
        logger.warning("Xformers is not available, falling back.")

    return USE_XFORMERS_OPS

55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74

class Attention(nn.Module):
    """Attention layer.

    This class takes query, key, and value tensors as input. The input tensors
    can either contain prompt tokens or generation tokens.
    The class does the following:

    1. Store the input key and value tensors in the KV cache.
    2. Perform (multi-head/multi-query/grouped-query) attention.
    3. Return the output tensor.
    """

    def __init__(
        self,
        num_heads: int,
        head_size: int,
        scale: float,
        num_kv_heads: Optional[int] = None,
        alibi_slopes: Optional[List[float]] = None,
75
        cache_config: Optional[CacheConfig] = None,
76
        quant_config: Optional[QuantizationConfig] = None,
77
        logits_soft_cap: Optional[float] = None,
78
        per_layer_sliding_window: Optional[int] = None,
79
        use_mla: bool = False,
80
        prefix: str = "",
81
        attn_type: str = AttentionType.DECODER,
82
        kv_sharing_target_layer_name: Optional[str] = None,
83
        **extra_impl_args,
84
    ) -> None:
85
86
87
88
        """
        The KV cache is stored inside this class and is accessed via
        `self.kv_cache`.
        """
89
        super().__init__()
90
91
92
93
94
95
96
97
98
        if per_layer_sliding_window is not None:
            # per-layer sliding window
            sliding_window = per_layer_sliding_window
        elif cache_config is not None:
            # model-level sliding window
            sliding_window = cache_config.sliding_window
        else:
            sliding_window = None

99
100
101
        if cache_config is not None:
            kv_cache_dtype = cache_config.cache_dtype
            block_size = cache_config.block_size
102
            is_attention_free = cache_config.is_attention_free
103
            calculate_kv_scales = cache_config.calculate_kv_scales
104
105
106
        else:
            kv_cache_dtype = "auto"
            block_size = 16
107
            is_attention_free = False
108
            calculate_kv_scales = False
109
110
        if num_kv_heads is None:
            num_kv_heads = num_heads
111
112
113
        assert num_heads % num_kv_heads == 0, \
            f"num_heads ({num_heads}) is not " \
            f"divisible by num_kv_heads ({num_kv_heads})"
114

115
        # The default k/v_scale is set to 1.0. This is ignored
116
117
        # when kv-cache is not fp8, and should be used with
        # kv-cache in fp8_e5m2. For kv-cache in fp8_e4m3, we
118
        # expect the pre-quantized k/v_scale to be loaded along
119
120
        # with the model weights.
        self.kv_cache_dtype = kv_cache_dtype
121
122
123
        self.calculate_kv_scales = calculate_kv_scales
        self._k_scale = torch.tensor(1.0, dtype=torch.float32)
        self._v_scale = torch.tensor(1.0, dtype=torch.float32)
124
125
126
        # FlashAttn doesn't support quantizing the kv-cache only
        # but requires q to be quantized as well.
        self._q_scale = torch.tensor(1.0, dtype=torch.float32)
127
        self._prob_scale = torch.tensor(1.0, dtype=torch.float32)
128
129
130
131
132
133

        # We also keep the float32 versions of k/v_scale for attention
        # backends that don't support tensors (Flashinfer)
        self._k_scale_float = 1.0
        self._v_scale_float = 1.0

134
        self.use_mla = use_mla
135
136
137
138
139
        self.num_heads = num_heads
        self.head_size = head_size
        self.num_kv_heads = num_kv_heads
        self.sliding_window = sliding_window

140
        quant_method = quant_config.get_quant_method(
141
            self, prefix=prefix) if quant_config else None
142
143
        if quant_method is not None and not isinstance(
                quant_method, UnquantizedLinearMethod):
144
            assert isinstance(quant_method, BaseKVCacheMethod)
145
146
            # TODO (mgoin): kv cache dtype should be specified in the FP8
            # checkpoint config and become the "auto" behavior
147
148
149
150
151
152
153
154
155
            if self.kv_cache_dtype == "fp8_e5m2":
                raise ValueError("fp8_e5m2 kv-cache is not supported with "
                                 "fp8 checkpoints.")
            # If quantization is enabled, we make "k_scale" and "v_scale"
            # parameters so that it can be loaded from the model checkpoint.
            # The k/v_scale will then be converted back to native float32
            # values after weight loading.
            self.quant_method = quant_method
            self.quant_method.create_weights(self)
156

157
158
159
        # During model initialization, the default dtype is set as the model
        # weight and activation dtype.
        dtype = torch.get_default_dtype()
160
161
162
163
164
165
        attn_backend = get_attn_backend(head_size,
                                        dtype,
                                        kv_cache_dtype,
                                        block_size,
                                        is_attention_free,
                                        use_mla=use_mla)
166
        impl_cls = attn_backend.get_impl_cls()
167
        self.impl = impl_cls(num_heads, head_size, scale, num_kv_heads,
168
                             alibi_slopes, sliding_window, kv_cache_dtype,
169
                             logits_soft_cap, attn_type,
170
                             kv_sharing_target_layer_name, **extra_impl_args)
171
        self.backend = backend_name_to_enum(attn_backend.get_name())
172
        self.dtype = dtype
173
        self.use_irope = extra_impl_args.get("use_irope", False)
174

175
176
177
178
        # For cuda-alike (CUDA and ROCM) and cpu platforms, we control how
        # torch.compile works by registering the attention as one giant
        # opaque custom op. For other platforms, we directly call them
        # and let torch.compile handle them.
179
180
181
        self.use_direct_call = not current_platform.is_cuda_alike(
        ) and not current_platform.is_cpu()

182
        self.use_output = attn_backend.accept_output_buffer
183
184
185
186
187
        compilation_config = get_current_vllm_config().compilation_config
        if prefix in compilation_config.static_forward_context:
            raise ValueError(f"Duplicate layer name: {prefix}")
        compilation_config.static_forward_context[prefix] = self
        self.layer_name = prefix
188
        self.attn_type = attn_type
189
190
191
192
193
194
195
196
197

        if kv_sharing_target_layer_name is not None:
            validate_kv_sharing_target(
                prefix,
                kv_sharing_target_layer_name,
                compilation_config.static_forward_context,
            )
        self.kv_sharing_target_layer_name = kv_sharing_target_layer_name

198
199
200
201
202
203
204
        # use a placeholder kv cache tensor during init, which will be replaced
        # by bind_kv_cache
        # this variable will not be accessed if use_direct_call is True
        self.kv_cache = [
            torch.tensor([]) for _ in range(get_current_vllm_config(
            ).parallel_config.pipeline_parallel_size)
        ]
205

206
        self.q_range = torch.tensor(envs.Q_SCALE_CONSTANT, dtype=torch.float32)
207
208
209
        self.k_range = torch.tensor(envs.K_SCALE_CONSTANT, dtype=torch.float32)
        self.v_range = torch.tensor(envs.V_SCALE_CONSTANT, dtype=torch.float32)

210
211
212
213
214
    def forward(
        self,
        query: torch.Tensor,
        key: torch.Tensor,
        value: torch.Tensor,
215
216
217
218
        # For some alternate attention backends like MLA the attention output
        # shape does not match the query shape, so we optionally let the model
        # definition specify the output tensor shape.
        output_shape: Optional[torch.Size] = None,
219
    ) -> torch.Tensor:
220
221
222
223
224
225
226
227
228
        """
        The KV cache is stored inside this class and is accessed via
        `self.kv_cache`.

        Attention metadata (`attn_metadata`) is set using a context manager in
        the model runner's `execute_model` method. It is accessed via forward
        context using
        `vllm.forward_context.get_forward_context().attn_metadata`.
        """
Chen Zhang's avatar
Chen Zhang committed
229
        if self.calculate_kv_scales:
230
231
            attn_metadata = get_forward_context().attn_metadata
            if attn_metadata.enable_kv_scales_calculation:
232
                self.calc_kv_scales(query, key, value)
233
        if self.use_output:
234
235
            output_shape = (output_shape
                            if output_shape is not None else query.shape)
236
            output = torch.zeros(output_shape,
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
                                 dtype=query.dtype,
                                 device=query.device)
            hidden_size = output_shape[-1]
            # We skip reshaping query, key and value tensors for the MLA
            # backend since these tensors have different semantics and are
            # processed differently.
            if not self.use_mla:
                # Reshape the query, key, and value tensors.
                # NOTE(woosuk): We do this outside the custom op to minimize the
                # CPU overheads from the non-CUDA-graph regions.
                query = query.view(-1, self.num_heads, self.head_size)
                output = output.view(-1, self.num_heads, self.head_size)
                if key is not None:
                    key = key.view(-1, self.num_kv_heads, self.head_size)
                if value is not None:
                    value = value.view(-1, self.num_kv_heads, self.head_size)
253
            if self.use_direct_call:
Chen Zhang's avatar
Chen Zhang committed
254
                forward_context: ForwardContext = get_forward_context()
255
                attn_metadata = forward_context.attn_metadata
256
257
                if isinstance(attn_metadata, dict):
                    attn_metadata = attn_metadata[self.layer_name]
Chen Zhang's avatar
Chen Zhang committed
258
259
260
261
262
263
                self_kv_cache = self.kv_cache[forward_context.virtual_engine]
                self.impl.forward(self,
                                  query,
                                  key,
                                  value,
                                  self_kv_cache,
264
                                  attn_metadata,
Chen Zhang's avatar
Chen Zhang committed
265
                                  output=output)
266
267
268
            else:
                torch.ops.vllm.unified_attention_with_output(
                    query, key, value, output, self.layer_name)
269
            return output.view(-1, hidden_size)
270
        else:
271
            if self.use_direct_call:
Chen Zhang's avatar
Chen Zhang committed
272
                forward_context = get_forward_context()
273
                attn_metadata = forward_context.attn_metadata
274
275
                if isinstance(attn_metadata, dict):
                    attn_metadata = attn_metadata[self.layer_name]
Chen Zhang's avatar
Chen Zhang committed
276
277
                self_kv_cache = self.kv_cache[forward_context.virtual_engine]
                return self.impl.forward(self, query, key, value,
278
                                         self_kv_cache, attn_metadata)
279
280
281
            else:
                return torch.ops.vllm.unified_attention(
                    query, key, value, self.layer_name)
282

283
284
    def calc_kv_scales(self, query, key, value):
        self._q_scale.copy_(torch.abs(query).max() / self.q_range)
285
286
287
288
289
290
291
        self._k_scale.copy_(torch.abs(key).max() / self.k_range)
        self._v_scale.copy_(torch.abs(value).max() / self.v_range)
        self._k_scale_float = self._k_scale.item()
        self._v_scale_float = self._v_scale.item()
        # We only calculate the scales once
        self.calculate_kv_scales = False

292
293
294
295
296
    def extra_repr(self) -> str:
        s = f"head_size={self.impl.head_size}"  # type: ignore
        s += f", num_heads={self.impl.num_heads}"  # type: ignore
        s += f", num_kv_heads={self.impl.num_kv_heads}"  # type: ignore
        s += f", scale={self.impl.scale}"  # type: ignore
297
        s += f", backend={self.impl.__class__.__name__}"
298
        return s
299

300
    def process_weights_after_loading(self, act_dtype: torch.dtype):
301
        if hasattr(self.impl, "process_weights_after_loading"):
302
            self.impl.process_weights_after_loading(act_dtype)
303

304

305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
class MultiHeadAttention(nn.Module):
    """Multi-headed attention without any cache, used for ViT."""

    def __init__(
        self,
        num_heads: int,
        head_size: int,
        scale: float,
        num_kv_heads: Optional[int] = None,
    ):
        super().__init__()
        self.num_heads = num_heads
        self.head_size = head_size
        self.scale = scale
        self.num_kv_heads = num_heads if num_kv_heads is None else num_kv_heads

321
322
323
        assert self.num_heads % self.num_kv_heads == 0, \
            f"num_heads ({self.num_heads}) is not " \
            f"divisible by num_kv_heads ({self.num_kv_heads})"
324
325
        self.num_queries_per_kv = self.num_heads // self.num_kv_heads

326
327
328
329
330
331
        dtype = torch.get_default_dtype()
        attn_backend = get_attn_backend(head_size,
                                        dtype,
                                        kv_cache_dtype=None,
                                        block_size=16,
                                        is_attention_free=False)
332
        backend = backend_name_to_enum(attn_backend.get_name())
333
334
335
336
        if current_platform.is_rocm():
            # currently, only torch_sdpa is supported on rocm
            self.attn_backend = _Backend.TORCH_SDPA
        else:
337
338
            if backend in (_Backend.FLASH_ATTN, _Backend.FLASH_ATTN_VLLM_V1,
                           _Backend.FLEX_ATTENTION):
339
                backend = _Backend.XFORMERS
340

341
342
343
            self.attn_backend = backend if backend in {
                _Backend.TORCH_SDPA, _Backend.XFORMERS, _Backend.PALLAS_VLLM_V1
            } else _Backend.TORCH_SDPA
344

345
346
347
348
        if (self.attn_backend == _Backend.XFORMERS
                and not check_xformers_availability()):
            self.attn_backend = _Backend.TORCH_SDPA

349
350
351
352
353
354
355
    def forward(
        self,
        query: torch.Tensor,
        key: torch.Tensor,
        value: torch.Tensor,
    ) -> torch.Tensor:
        """Input shape: batch_size x seq_len x hidden_size"""
356
        # TODO(Isotr0py): Use existing backend implementations and support FA3
357
358
359
360
361
362
363
        bsz, q_len, _ = query.size()
        kv_len = key.size(1)

        query = query.view(bsz, q_len, self.num_heads, self.head_size)
        key = key.view(bsz, kv_len, self.num_kv_heads, self.head_size)
        value = value.view(bsz, kv_len, self.num_kv_heads, self.head_size)

364
365
366
367
368
        if (num_repeat := self.num_queries_per_kv) > 1:
            # Handle MQA and GQA
            key = torch.repeat_interleave(key, num_repeat, dim=2)
            value = torch.repeat_interleave(value, num_repeat, dim=2)

369
        if self.attn_backend == _Backend.XFORMERS:
370
371
372
373
374
375
376
377
378
379
380
381
382
383
            from xformers import ops as xops

            out = xops.memory_efficient_attention_forward(query,
                                                          key,
                                                          value,
                                                          scale=self.scale)
        elif self.attn_backend == _Backend.TORCH_SDPA:
            query, key, value = (x.transpose(1, 2)
                                 for x in (query, key, value))
            out = F.scaled_dot_product_attention(query,
                                                 key,
                                                 value,
                                                 scale=self.scale)
            out = out.transpose(1, 2)
384
385
386
387
388
389
390
        elif self.attn_backend == _Backend.PALLAS_VLLM_V1:
            query, key, value = (x.transpose(1, 2)
                                 for x in (query, key, value))
            from torch_xla.experimental.custom_kernel import flash_attention
            out = flash_attention(query, key, value, sm_scale=self.scale)
            out = out.transpose(1, 2)

391
        return out.reshape(bsz, q_len, -1)
392
393


394
395
396
397
398
399
400
401
402
403
def wait_for_kv_layer_from_connector(layer_name: str):
    if not has_kv_transfer_group() or not is_v1_kv_transfer_group():
        return

    connector = get_kv_transfer_group()

    forward_context: ForwardContext = get_forward_context()
    attn_metadata = forward_context.attn_metadata
    if attn_metadata is None:
        return
404
    assert isinstance(attn_metadata, dict)
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
    connector.wait_for_layer_load(layer_name)


def maybe_save_kv_layer_to_connector(
    layer_name: str,
    kv_cache_layer: List[torch.Tensor],
):
    if not has_kv_transfer_group() or not is_v1_kv_transfer_group():
        return

    connector = get_kv_transfer_group()

    forward_context: ForwardContext = get_forward_context()
    attn_metadata = forward_context.attn_metadata
    if attn_metadata is None:
        return
421
422
423
    assert isinstance(attn_metadata, dict)
    connector.save_kv_layer(layer_name, kv_cache_layer,
                            attn_metadata[layer_name])
424
425


426
427
428
429
430
431
def unified_attention(
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    layer_name: str,
) -> torch.Tensor:
432
433
    wait_for_kv_layer_from_connector(layer_name)

434
    forward_context: ForwardContext = get_forward_context()
435
    attn_metadata = forward_context.attn_metadata
436
437
    if isinstance(attn_metadata, dict):
        attn_metadata = attn_metadata[layer_name]
438
    self = forward_context.no_compile_layers[layer_name]
439
    kv_cache = self.kv_cache[forward_context.virtual_engine]
440
441
442
443
444
    output = self.impl.forward(self, query, key, value, kv_cache,
                               attn_metadata)

    maybe_save_kv_layer_to_connector(layer_name, kv_cache)
    return output
445
446
447
448
449
450
451
452
453
454
455
456
457
458


def unified_attention_fake(
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    layer_name: str,
) -> torch.Tensor:
    return torch.empty_like(query).contiguous()


direct_register_custom_op(
    op_name="unified_attention",
    op_func=unified_attention,
459
    mutates_args=[],
460
461
462
    fake_impl=unified_attention_fake,
    dispatch_key=current_platform.dispatch_key,
)
463
464
465
466
467
468
469
470


def unified_attention_with_output(
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    output: torch.Tensor,
    layer_name: str,
471
    output_scale: Optional[torch.Tensor] = None,
472
) -> None:
473
    wait_for_kv_layer_from_connector(layer_name)
474
    forward_context: ForwardContext = get_forward_context()
475
    attn_metadata = forward_context.attn_metadata
476
477
    if isinstance(attn_metadata, dict):
        attn_metadata = attn_metadata[layer_name]
478
    self = forward_context.no_compile_layers[layer_name]
479
    kv_cache = self.kv_cache[forward_context.virtual_engine]
480
481
    self.impl.forward(self,
                      query,
482
483
484
485
                      key,
                      value,
                      kv_cache,
                      attn_metadata,
486
487
                      output=output,
                      output_scale=output_scale)
488

489
490
    maybe_save_kv_layer_to_connector(layer_name, kv_cache)

491
492
493
494
495
496
497

def unified_attention_with_output_fake(
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    output: torch.Tensor,
    layer_name: str,
498
    output_scale: Optional[torch.Tensor] = None,
499
500
501
502
503
504
505
) -> None:
    return


direct_register_custom_op(
    op_name="unified_attention_with_output",
    op_func=unified_attention_with_output,
506
    mutates_args=["output"],
507
508
509
    fake_impl=unified_attention_with_output_fake,
    dispatch_key=current_platform.dispatch_key,
)