layer.py 20.3 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 Any, Dict, 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
        blocksparse_params: Optional[Dict[str, Any]] = None,
78
        logits_soft_cap: Optional[float] = None,
79
        per_layer_sliding_window: Optional[int] = None,
80
        use_mla: bool = False,
81
        prefix: str = "",
82
        attn_type: str = AttentionType.DECODER,
83
        kv_sharing_target_layer_name: Optional[str] = None,
84
        **extra_impl_args,
85
    ) -> None:
86
87
88
89
        """
        The KV cache is stored inside this class and is accessed via
        `self.kv_cache`.
        """
90
        super().__init__()
91
92
93
94
95
96
97
98
99
        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

100
101
102
        if cache_config is not None:
            kv_cache_dtype = cache_config.cache_dtype
            block_size = cache_config.block_size
103
            is_attention_free = cache_config.is_attention_free
104
            calculate_kv_scales = cache_config.calculate_kv_scales
105
106
107
        else:
            kv_cache_dtype = "auto"
            block_size = 16
108
            is_attention_free = False
109
            calculate_kv_scales = False
110
111
        if num_kv_heads is None:
            num_kv_heads = num_heads
112
113
114
        assert num_heads % num_kv_heads == 0, \
            f"num_heads ({num_heads}) is not " \
            f"divisible by num_kv_heads ({num_kv_heads})"
115

116
        # The default k/v_scale is set to 1.0. This is ignored
117
118
        # when kv-cache is not fp8, and should be used with
        # kv-cache in fp8_e5m2. For kv-cache in fp8_e4m3, we
119
        # expect the pre-quantized k/v_scale to be loaded along
120
121
        # with the model weights.
        self.kv_cache_dtype = kv_cache_dtype
122
123
124
        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)
125
126
127
        # 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)
128
        self._prob_scale = torch.tensor(1.0, dtype=torch.float32)
129
130
131
132
133
134

        # 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

135
        self.use_mla = use_mla
136
137
138
139
140
        self.num_heads = num_heads
        self.head_size = head_size
        self.num_kv_heads = num_kv_heads
        self.sliding_window = sliding_window

141
        quant_method = quant_config.get_quant_method(
142
            self, prefix=prefix) if quant_config else None
143
144
        if quant_method is not None and not isinstance(
                quant_method, UnquantizedLinearMethod):
145
            assert isinstance(quant_method, BaseKVCacheMethod)
146
147
            # TODO (mgoin): kv cache dtype should be specified in the FP8
            # checkpoint config and become the "auto" behavior
148
149
150
151
152
153
154
155
156
            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)
157

158
159
160
        # During model initialization, the default dtype is set as the model
        # weight and activation dtype.
        dtype = torch.get_default_dtype()
161
162
163
164
165
166
167
        attn_backend = get_attn_backend(head_size,
                                        dtype,
                                        kv_cache_dtype,
                                        block_size,
                                        is_attention_free,
                                        blocksparse_params is not None,
                                        use_mla=use_mla)
168
        impl_cls = attn_backend.get_impl_cls()
169
        self.impl = impl_cls(num_heads, head_size, scale, num_kv_heads,
170
                             alibi_slopes, sliding_window, kv_cache_dtype,
171
                             blocksparse_params, logits_soft_cap, attn_type,
172
                             kv_sharing_target_layer_name, **extra_impl_args)
173
        self.backend = backend_name_to_enum(attn_backend.get_name())
174
        self.dtype = dtype
175
        self.use_irope = extra_impl_args.get("use_irope", False)
176

177
178
179
180
        # 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.
181
182
183
        self.use_direct_call = not current_platform.is_cuda_alike(
        ) and not current_platform.is_cpu()

184
        self.use_output = attn_backend.accept_output_buffer
185
186
187
188
189
        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
190
        self.attn_type = attn_type
191
192
193
194
195
196
197
198
199

        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

200
201
202
203
204
205
206
        # 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)
        ]
207

208
        self.q_range = torch.tensor(envs.Q_SCALE_CONSTANT, dtype=torch.float32)
209
210
211
        self.k_range = torch.tensor(envs.K_SCALE_CONSTANT, dtype=torch.float32)
        self.v_range = torch.tensor(envs.V_SCALE_CONSTANT, dtype=torch.float32)

212
213
214
215
216
    def forward(
        self,
        query: torch.Tensor,
        key: torch.Tensor,
        value: torch.Tensor,
217
218
219
220
        # 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,
221
    ) -> torch.Tensor:
222
223
224
225
226
227
228
229
230
        """
        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
231
        if self.calculate_kv_scales:
232
233
            attn_metadata = get_forward_context().attn_metadata
            if attn_metadata.enable_kv_scales_calculation:
234
                self.calc_kv_scales(query, key, value)
235
        if self.use_output:
236
237
            output_shape = (output_shape
                            if output_shape is not None else query.shape)
238
            output = torch.zeros(output_shape,
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
                                 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)
255
            if self.use_direct_call:
Chen Zhang's avatar
Chen Zhang committed
256
                forward_context: ForwardContext = get_forward_context()
257
                attn_metadata = forward_context.attn_metadata
258
259
                if isinstance(attn_metadata, dict):
                    attn_metadata = attn_metadata[self.layer_name]
Chen Zhang's avatar
Chen Zhang committed
260
261
262
263
264
265
                self_kv_cache = self.kv_cache[forward_context.virtual_engine]
                self.impl.forward(self,
                                  query,
                                  key,
                                  value,
                                  self_kv_cache,
266
                                  attn_metadata,
Chen Zhang's avatar
Chen Zhang committed
267
                                  output=output)
268
269
270
            else:
                torch.ops.vllm.unified_attention_with_output(
                    query, key, value, output, self.layer_name)
271
            return output.view(-1, hidden_size)
272
        else:
273
            if self.use_direct_call:
Chen Zhang's avatar
Chen Zhang committed
274
                forward_context = get_forward_context()
275
                attn_metadata = forward_context.attn_metadata
276
277
                if isinstance(attn_metadata, dict):
                    attn_metadata = attn_metadata[self.layer_name]
Chen Zhang's avatar
Chen Zhang committed
278
279
                self_kv_cache = self.kv_cache[forward_context.virtual_engine]
                return self.impl.forward(self, query, key, value,
280
                                         self_kv_cache, attn_metadata)
281
282
283
            else:
                return torch.ops.vllm.unified_attention(
                    query, key, value, self.layer_name)
284

285
286
    def calc_kv_scales(self, query, key, value):
        self._q_scale.copy_(torch.abs(query).max() / self.q_range)
287
288
289
290
291
292
293
        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

294
295
296
297
298
    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
299
        s += f", backend={self.impl.__class__.__name__}"
300
        return s
301

302
    def process_weights_after_loading(self, act_dtype: torch.dtype):
303
        if hasattr(self.impl, "process_weights_after_loading"):
304
            self.impl.process_weights_after_loading(act_dtype)
305

306

307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
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

323
324
325
        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})"
326
327
        self.num_queries_per_kv = self.num_heads // self.num_kv_heads

328
329
330
331
332
333
        dtype = torch.get_default_dtype()
        attn_backend = get_attn_backend(head_size,
                                        dtype,
                                        kv_cache_dtype=None,
                                        block_size=16,
                                        is_attention_free=False)
334
        backend = backend_name_to_enum(attn_backend.get_name())
335
336
337
338
        if current_platform.is_rocm():
            # currently, only torch_sdpa is supported on rocm
            self.attn_backend = _Backend.TORCH_SDPA
        else:
339
340
            if backend in (_Backend.FLASH_ATTN, _Backend.FLASH_ATTN_VLLM_V1,
                           _Backend.FLEX_ATTENTION):
341
                backend = _Backend.XFORMERS
342

343
344
345
            self.attn_backend = backend if backend in {
                _Backend.TORCH_SDPA, _Backend.XFORMERS, _Backend.PALLAS_VLLM_V1
            } else _Backend.TORCH_SDPA
346

347
348
349
350
        if (self.attn_backend == _Backend.XFORMERS
                and not check_xformers_availability()):
            self.attn_backend = _Backend.TORCH_SDPA

351
352
353
354
355
356
357
    def forward(
        self,
        query: torch.Tensor,
        key: torch.Tensor,
        value: torch.Tensor,
    ) -> torch.Tensor:
        """Input shape: batch_size x seq_len x hidden_size"""
358
        # TODO(Isotr0py): Use existing backend implementations and support FA3
359
360
361
362
363
364
365
        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)

366
367
368
369
370
        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)

371
        if self.attn_backend == _Backend.XFORMERS:
372
373
374
375
376
377
378
379
380
381
382
383
384
385
            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)
386
387
388
389
390
391
392
        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)

393
        return out.reshape(bsz, q_len, -1)
394
395


396
397
398
399
400
401
402
403
404
405
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
406
    assert isinstance(attn_metadata, dict)
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
    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
423
424
425
    assert isinstance(attn_metadata, dict)
    connector.save_kv_layer(layer_name, kv_cache_layer,
                            attn_metadata[layer_name])
426
427


428
429
430
431
432
433
def unified_attention(
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    layer_name: str,
) -> torch.Tensor:
434
435
    wait_for_kv_layer_from_connector(layer_name)

436
    forward_context: ForwardContext = get_forward_context()
437
    attn_metadata = forward_context.attn_metadata
438
439
    if isinstance(attn_metadata, dict):
        attn_metadata = attn_metadata[layer_name]
440
    self = forward_context.no_compile_layers[layer_name]
441
    kv_cache = self.kv_cache[forward_context.virtual_engine]
442
443
444
445
446
    output = self.impl.forward(self, query, key, value, kv_cache,
                               attn_metadata)

    maybe_save_kv_layer_to_connector(layer_name, kv_cache)
    return output
447
448
449
450
451
452
453
454
455
456
457
458
459
460


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,
461
    mutates_args=[],
462
463
464
    fake_impl=unified_attention_fake,
    dispatch_key=current_platform.dispatch_key,
)
465
466
467
468
469
470
471
472


def unified_attention_with_output(
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    output: torch.Tensor,
    layer_name: str,
473
    output_scale: Optional[torch.Tensor] = None,
474
) -> None:
475
    wait_for_kv_layer_from_connector(layer_name)
476
    forward_context: ForwardContext = get_forward_context()
477
    attn_metadata = forward_context.attn_metadata
478
479
    if isinstance(attn_metadata, dict):
        attn_metadata = attn_metadata[layer_name]
480
    self = forward_context.no_compile_layers[layer_name]
481
    kv_cache = self.kv_cache[forward_context.virtual_engine]
482
483
    self.impl.forward(self,
                      query,
484
485
486
487
                      key,
                      value,
                      kv_cache,
                      attn_metadata,
488
489
                      output=output,
                      output_scale=output_scale)
490

491
492
    maybe_save_kv_layer_to_connector(layer_name, kv_cache)

493
494
495
496
497
498
499

def unified_attention_with_output_fake(
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    output: torch.Tensor,
    layer_name: str,
500
    output_scale: Optional[torch.Tensor] = None,
501
502
503
504
505
506
507
) -> None:
    return


direct_register_custom_op(
    op_name="unified_attention_with_output",
    op_func=unified_attention_with_output,
508
    mutates_args=["output"],
509
510
511
    fake_impl=unified_attention_with_output_fake,
    dispatch_key=current_platform.dispatch_key,
)