layer.py 34.5 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
"""Attention layer."""
4

5
6
from collections.abc import Callable
from typing import cast
7
8
9

import torch
import torch.nn as nn
10
import torch.nn.functional as F
11

12
import vllm.envs as envs
13
14
15
16
17
from vllm.attention.backends.abstract import (
    AttentionBackend,
    AttentionType,
    MLAAttentionImpl,
)
18
from vllm.attention.backends.registry import AttentionBackendEnum
19
from vllm.attention.selector import get_attn_backend
20
from vllm.attention.utils.kv_sharing_utils import validate_kv_sharing_target
21
from vllm.attention.utils.kv_transfer_utils import maybe_transfer_kv_layer
22
from vllm.config import CacheConfig, get_current_vllm_config
23
from vllm.config.multimodal import MultiModalConfig
24
from vllm.config.vllm import VllmConfig
25
from vllm.forward_context import ForwardContext, get_forward_context
26
from vllm.logger import init_logger
27
from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
28
29
30
31
from vllm.model_executor.layers.linear import (
    ColumnParallelLinear,
    UnquantizedLinearMethod,
)
32
from vllm.model_executor.layers.quantization import QuantizationConfig
33
from vllm.model_executor.layers.quantization.input_quant_fp8 import QuantFP8
34
from vllm.model_executor.layers.quantization.kv_cache import BaseKVCacheMethod
35
from vllm.model_executor.layers.quantization.utils.quant_utils import GroupShape
36
from vllm.model_executor.models.vision import get_vit_attn_backend
37
from vllm.platforms import current_platform
38
from vllm.utils.torch_utils import (
39
40
41
42
43
44
45
46
47
    direct_register_custom_op,
    kv_cache_dtype_str_to_dtype,
)
from vllm.v1.kv_cache_interface import (
    FullAttentionSpec,
    KVCacheSpec,
    MLAAttentionSpec,
    SlidingWindowSpec,
)
48

49
50
51
52
53
54
if current_platform.is_rocm():
    from vllm.platforms.rocm import on_gfx9
else:
    on_gfx9 = lambda *args, **kwargs: False


55
FP8_DTYPE = current_platform.fp8_dtype()
56
57
logger = init_logger(__name__)

58

59
def maybe_get_vit_flash_attn_backend(
60
61
62
    attn_backend: AttentionBackendEnum,
    attn_backend_override: AttentionBackendEnum | None = None,
) -> tuple[AttentionBackendEnum, Callable | None]:
63
64
    if current_platform.is_rocm():
        if envs.VLLM_ROCM_USE_AITER and envs.VLLM_ROCM_USE_AITER_MHA and on_gfx9():
65
            attn_backend = AttentionBackendEnum.ROCM_AITER_FA
66
        elif (
67
            attn_backend_override is None
68
            and on_gfx9()
69
            and attn_backend == AttentionBackendEnum.FLASH_ATTN
70
        ):
71
            pass
72
        else:
73
            return AttentionBackendEnum.TORCH_SDPA, None
74
    elif current_platform.is_cuda():
75
        pass
76
    elif current_platform.is_xpu():
77
        assert attn_backend == AttentionBackendEnum.FLASH_ATTN, (
78
79
            "XPU platform only supports FLASH_ATTN as vision attention backend."
        )
80
        pass
81
    else:
82
        return AttentionBackendEnum.TORCH_SDPA, None
83

84
85
86
87
88
    if attn_backend in {
        AttentionBackendEnum.FLASH_ATTN,
        AttentionBackendEnum.ROCM_AITER_FA,
    }:
        if attn_backend == AttentionBackendEnum.ROCM_AITER_FA:
89
90
            from aiter import flash_attn_varlen_func
        else:
91
            from vllm.attention.utils.fa_utils import flash_attn_varlen_func
92
93
94
95
96
97
    else:
        flash_attn_varlen_func = None

    return attn_backend, flash_attn_varlen_func


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
def _init_kv_cache_quant(
    layer: nn.Module,
    quant_config: QuantizationConfig | None,
    prefix: str,
    kv_cache_dtype: str,
    calculate_kv_scales: bool,
) -> None:
    """Initializes KV cache scaling factors and quantization method.

    This helper function sets up the KV cache quantization attributes that are
    shared between Attention and MLAAttention layers. It initializes scale
    tensors for query, key, value, and probability, and configures the
    quantization method if applicable.

    Args:
        layer: The attention layer instance to initialize.
        quant_config: Optional quantization configuration.
        prefix: Layer name prefix for quantization method lookup.
        kv_cache_dtype: The KV cache data type string.
        calculate_kv_scales: Whether to calculate KV scales dynamically.
    """
    # The default k/v_scale is set to 1.0. This is ignored
    # when kv-cache is not fp8, and should be used with
    # kv-cache in fp8_e5m2. For kv-cache in fp8_e4m3, we
    # expect the pre-quantized k/v_scale to be loaded along
    # with the model weights.
    layer.kv_cache_dtype = kv_cache_dtype
    layer.calculate_kv_scales = calculate_kv_scales
    layer._k_scale = torch.tensor(1.0, dtype=torch.float32)
    layer._v_scale = torch.tensor(1.0, dtype=torch.float32)
    layer._q_scale = torch.tensor(1.0, dtype=torch.float32)
    layer._prob_scale = torch.tensor(1.0, dtype=torch.float32)

    # We also keep q/k/v_scale on host (cpu) memory for attention
    # backends that require the scales to be on host instead of on device.
    # e.g. Flashinfer
    layer._q_scale_float = 1.0
    layer._k_scale_float = 1.0
    layer._v_scale_float = 1.0

    # The output scale on host memory. This should be the input scale of
    # the quant op after this attention layer.
    layer._o_scale_float = None

    quant_method = (
        quant_config.get_quant_method(layer, prefix=prefix) if quant_config else None
    )
    if quant_method is not None and not isinstance(
        quant_method, UnquantizedLinearMethod
    ):
        assert isinstance(quant_method, BaseKVCacheMethod)
        # TODO (mgoin): kv cache dtype should be specified in the FP8
        # checkpoint config and become the "auto" behavior
        if 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.
        layer.quant_method = quant_method
        layer.quant_method.create_weights(layer)


161
class Attention(nn.Module, AttentionLayerBase):
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
    """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,
178
179
180
181
182
183
        num_kv_heads: int | None = None,
        alibi_slopes: list[float] | None = None,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
        logits_soft_cap: float | None = None,
        per_layer_sliding_window: int | None = None,
184
        prefix: str = "",
185
        attn_type: str = AttentionType.DECODER,
186
187
        kv_sharing_target_layer_name: str | None = None,
        attn_backend: type[AttentionBackend] | None = None,
188
        **extra_impl_args,
189
    ) -> None:
190
191
192
193
        """
        The KV cache is stored inside this class and is accessed via
        `self.kv_cache`.
        """
194
        super().__init__()
195
196
197
198
199
200
201
202
203
        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

204
        vllm_config = get_current_vllm_config()
205
206
207
        if cache_config is not None:
            kv_cache_dtype = cache_config.cache_dtype
            block_size = cache_config.block_size
208
            calculate_kv_scales = cache_config.calculate_kv_scales
209
210
211
        else:
            kv_cache_dtype = "auto"
            block_size = 16
212
            calculate_kv_scales = False
213
214
215
        self.kv_cache_torch_dtype = kv_cache_dtype_str_to_dtype(
            kv_cache_dtype, vllm_config.model_config
        )
216
217
        if num_kv_heads is None:
            num_kv_heads = num_heads
218
219
220
        assert num_heads % num_kv_heads == 0, (
            f"num_heads ({num_heads}) is not divisible by num_kv_heads ({num_kv_heads})"
        )
221

222
223
224
225
        # Initialize KV cache quantization attributes
        _init_kv_cache_quant(
            self, quant_config, prefix, kv_cache_dtype, calculate_kv_scales
        )
226

227
228
229
230
        self.num_heads = num_heads
        self.head_size = head_size
        self.num_kv_heads = num_kv_heads
        self.sliding_window = sliding_window
231
        self.has_sink = extra_impl_args.get("sinks") is not None
232

233
234
235
        # During model initialization, the default dtype is set as the model
        # weight and activation dtype.
        dtype = torch.get_default_dtype()
236
        if attn_backend is None:
237
238
239
240
241
            self.attn_backend = get_attn_backend(
                head_size,
                dtype,
                kv_cache_dtype,
                block_size,
242
                use_mla=False,
243
                has_sink=self.has_sink,
244
                attn_type=attn_type,
245
            )
246
247
248
249
        else:
            self.attn_backend = attn_backend

        impl_cls = self.attn_backend.get_impl_cls()
250
251
252
253
254
255
256
257
258
259
260
261
262
        self.impl = impl_cls(
            num_heads,
            head_size,
            scale,
            num_kv_heads,
            alibi_slopes,
            sliding_window,
            kv_cache_dtype,
            logits_soft_cap,
            attn_type,
            kv_sharing_target_layer_name,
            **extra_impl_args,
        )
263
264
        backend_name = self.attn_backend.get_name()
        self.backend = AttentionBackendEnum.__members__.get(backend_name)
265
        self.dtype = dtype
266

267
268
269
270
        # 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.
271
        self.use_direct_call = not current_platform.opaque_attention_op()
272

273
        self.use_output = self.attn_backend.accept_output_buffer
274
        compilation_config = vllm_config.compilation_config
275
276
277
278
        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
279
        self.attn_type = attn_type
280
281
282
283
284
285
286
287
288

        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

289
290
291
292
        # 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 = [
293
            torch.tensor([])
294
            for _ in range(vllm_config.parallel_config.pipeline_parallel_size)
295
        ]
296

297
298
299
300
        # Initialize q/k/v range constants.
        self.q_range = torch.tensor(envs.Q_SCALE_CONSTANT, dtype=torch.float32)
        self.k_range = torch.tensor(envs.K_SCALE_CONSTANT, dtype=torch.float32)
        self.v_range = torch.tensor(envs.V_SCALE_CONSTANT, dtype=torch.float32)
301

302
303
        # for attn backends supporting query quantization
        self.query_quant = None
304
305
        if (
            self.kv_cache_dtype.startswith("fp8")
306
            and self.impl.supports_quant_query_input
307
308
        ):
            self.query_quant = QuantFP8(static=True, group_shape=GroupShape.PER_TENSOR)
309

310
311
312
313
314
    def forward(
        self,
        query: torch.Tensor,
        key: torch.Tensor,
        value: torch.Tensor,
315
316
317
        # 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.
318
        output_shape: torch.Size | None = None,
319
    ) -> torch.Tensor:
320
321
322
323
324
325
326
327
328
        """
        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
329
        if self.calculate_kv_scales:
330
            torch.ops.vllm.maybe_calc_kv_scales(query, key, value, self.layer_name)
331
332
333
334
335
336
337
338
        output_dtype = query.dtype
        if self.query_quant is not None:
            # quantizing with a simple torch operation enables
            # torch.compile to fuse this into previous ops
            # which reduces overheads during decoding.
            # Otherwise queries are quantized using custom ops
            # which causes decoding overheads
            assert self.kv_cache_dtype in {"fp8", "fp8_e4m3"}
339
340

            # check if query quantization is supported
341
            if self.impl.supports_quant_query_input:
342
                query, _ = self.query_quant(query, self._q_scale)
343

344
        if self.use_output:
345
            output_shape = output_shape if output_shape is not None else query.shape
346
            output = torch.empty(output_shape, dtype=output_dtype, device=query.device)
347
            hidden_size = output_shape[-1]
348
349
350
351
352
353
354
355
356
            # 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)
357
            if self.use_direct_call:
Chen Zhang's avatar
Chen Zhang committed
358
                forward_context: ForwardContext = get_forward_context()
359
                attn_metadata = forward_context.attn_metadata
360
361
                if isinstance(attn_metadata, dict):
                    attn_metadata = attn_metadata[self.layer_name]
Chen Zhang's avatar
Chen Zhang committed
362
                self_kv_cache = self.kv_cache[forward_context.virtual_engine]
363
364
365
                self.impl.forward(
                    self, query, key, value, self_kv_cache, attn_metadata, output=output
                )
366
367
            else:
                torch.ops.vllm.unified_attention_with_output(
368
369
                    query, key, value, output, self.layer_name
                )
370
            return output.view(-1, hidden_size)
371
        else:
372
            if self.use_direct_call:
Chen Zhang's avatar
Chen Zhang committed
373
                forward_context = get_forward_context()
374
                attn_metadata = forward_context.attn_metadata
375
376
                if isinstance(attn_metadata, dict):
                    attn_metadata = attn_metadata[self.layer_name]
Chen Zhang's avatar
Chen Zhang committed
377
                self_kv_cache = self.kv_cache[forward_context.virtual_engine]
378
379
380
                return self.impl.forward(
                    self, query, key, value, self_kv_cache, attn_metadata
                )
381
382
            else:
                return torch.ops.vllm.unified_attention(
383
384
                    query, key, value, self.layer_name
                )
385

386
387
    def calc_kv_scales(self, query, key, value):
        self._q_scale.copy_(torch.abs(query).max() / self.q_range)
388
389
        self._k_scale.copy_(torch.abs(key).max() / self.k_range)
        self._v_scale.copy_(torch.abs(value).max() / self.v_range)
390
        self._q_scale_float = self._q_scale.item()
391
392
393
394
395
        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

396
397
398
399
400
    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
401
        s += f", backend={self.impl.__class__.__name__}"
402
        return s
403

404
    def process_weights_after_loading(self, act_dtype: torch.dtype):
405
        self.impl.process_weights_after_loading(act_dtype)
406

407
408
409
    def get_attn_backend(self) -> type[AttentionBackend]:
        return self.attn_backend

410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
    def get_kv_cache_spec(self, vllm_config: VllmConfig) -> KVCacheSpec:
        # Block size may get updated after model loading, refresh it
        block_size = vllm_config.cache_config.block_size
        # Should not be called for enc-dec or encoder-only attention.
        assert self.attn_type == AttentionType.DECODER
        if self.sliding_window is not None:
            assert not vllm_config.model_config.use_mla, (
                "MLA is not supported for slidingwindow"
            )
            return SlidingWindowSpec(
                block_size=block_size,
                num_kv_heads=self.num_kv_heads,
                head_size=self.head_size,
                dtype=self.kv_cache_torch_dtype,
                sliding_window=self.sliding_window,
            )
        else:
            return FullAttentionSpec(
                block_size=block_size,
                num_kv_heads=self.num_kv_heads,
                head_size=self.head_size,
                dtype=self.kv_cache_torch_dtype,
            )

434

435
436
437
438
439
440
441
442
class MultiHeadAttention(nn.Module):
    """Multi-headed attention without any cache, used for ViT."""

    def __init__(
        self,
        num_heads: int,
        head_size: int,
        scale: float,
443
        num_kv_heads: int | None = None,
444
445
446
        # This has no effect, it is only here to make it easier to swap
        # between Attention and MultiHeadAttention
        prefix: str = "",
447
        multimodal_config: MultiModalConfig | None = None,
448
    ) -> None:
449
450
451
452
453
        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
454
        self.layer_name = prefix
455

456
457
        assert self.num_heads % self.num_kv_heads == 0, (
            f"num_heads ({self.num_heads}) is not "
458
            f"divisible by num_kv_heads ({self.num_kv_heads})"
459
        )
460
461
        self.num_queries_per_kv = self.num_heads // self.num_kv_heads

462
463
        # During model initialization, the default dtype is set as the model
        # weight and activation dtype.
464
        dtype = torch.get_default_dtype()
465
466

        # Determine the attention backend
467
468
469
470
471
472
473
474
        attn_backend_override = None
        if multimodal_config is not None:
            attn_backend_override = multimodal_config.mm_encoder_attn_backend
        backend = get_vit_attn_backend(
            head_size=head_size,
            dtype=dtype,
            attn_backend_override=attn_backend_override,
        )
475

476
477
478
479
        self.attn_backend = (
            backend
            if backend
            in {
480
481
482
483
                AttentionBackendEnum.TORCH_SDPA,
                AttentionBackendEnum.PALLAS,
                AttentionBackendEnum.ROCM_AITER_FA,
                AttentionBackendEnum.FLASH_ATTN,
484
            }
485
            else AttentionBackendEnum.TORCH_SDPA
486
        )
487

488
489
        self.attn_backend, self._flash_attn_varlen_func = (
            maybe_get_vit_flash_attn_backend(
490
                self.attn_backend,
491
                attn_backend_override=attn_backend_override,
492
            )
493
        )
494
495

        self.is_flash_attn_backend = self.attn_backend in {
496
497
            AttentionBackendEnum.FLASH_ATTN,
            AttentionBackendEnum.ROCM_AITER_FA,
498
499
        }

500
        logger.info_once(
501
            f"Using {self.attn_backend} for MultiHeadAttention in multimodal encoder."
502
        )
503

504
505
506
507
508
509
    def forward(
        self,
        query: torch.Tensor,
        key: torch.Tensor,
        value: torch.Tensor,
    ) -> torch.Tensor:
510
        """Input shape:
511
512
513
514
        (batch_size x seq_len x hidden_size) or
        (batch_size x seq_len x num_heads x head_size)
        """
        bsz, q_len = query.size()[:2]
515
516
517
518
519
520
        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)

521
522
523
524
525
        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)

526
        if self.is_flash_attn_backend:
527
            assert self._flash_attn_varlen_func is not None
528
529
530
531
532
533
            cu_seqlens_q = torch.arange(
                0, (bsz + 1) * q_len, step=q_len, dtype=torch.int32, device=query.device
            )
            cu_seqlens_k = torch.arange(
                0, (bsz + 1) * kv_len, step=kv_len, dtype=torch.int32, device=key.device
            )
534
535
536
537
538
539
540
541
542
543
544

            out = self._flash_attn_varlen_func(
                query.flatten(0, 1),
                key.flatten(0, 1),
                value.flatten(0, 1),
                cu_seqlens_q=cu_seqlens_q,
                cu_seqlens_k=cu_seqlens_k,
                max_seqlen_q=q_len,
                max_seqlen_k=kv_len,
                softmax_scale=self.scale,
            )
545
        elif self.attn_backend == AttentionBackendEnum.TORCH_SDPA:
546
547
            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)
548
            out = out.transpose(1, 2)
549
        elif self.attn_backend == AttentionBackendEnum.PALLAS:
550
            query, key, value = (x.transpose(1, 2) for x in (query, key, value))
551
            from torch_xla.experimental.custom_kernel import flash_attention
552

553
554
            out = flash_attention(query, key, value, sm_scale=self.scale)
            out = out.transpose(1, 2)
555
556
557
        else:
            # ViT attention hasn't supported this backend yet
            raise NotImplementedError(
558
559
                f"ViT attention hasn't supported {self.attn_backend} backend yet."
            )
560

561
        return out.reshape(bsz, q_len, -1)
562
563


564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
class MLAAttention(nn.Module, AttentionLayerBase):
    """Multi-Head Latent Attention layer.

    This class takes query, and compressed key/value tensors as input.
    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,
        scale: float,
        qk_nope_head_dim: int,
        qk_rope_head_dim: int,
        v_head_dim: int,
582
        q_lora_rank: int | None,
583
584
        kv_lora_rank: int,
        kv_b_proj: ColumnParallelLinear,
585
586
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
587
588
        prefix: str = "",
        use_sparse: bool = False,
589
        indexer: object | None = None,
590
        **extra_impl_args,
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
    ):
        super().__init__()
        self.num_heads = num_heads
        self.scale = scale
        self.qk_nope_head_dim = qk_nope_head_dim
        self.qk_rope_head_dim = qk_rope_head_dim
        self.v_head_dim = v_head_dim
        self.q_lora_rank = q_lora_rank
        self.kv_lora_rank = kv_lora_rank
        self.head_size = kv_lora_rank + qk_rope_head_dim
        self.layer_name = prefix

        if cache_config is not None:
            kv_cache_dtype = cache_config.cache_dtype
            block_size = cache_config.block_size
            calculate_kv_scales = cache_config.calculate_kv_scales
        else:
            kv_cache_dtype = "auto"
            block_size = 16
            calculate_kv_scales = False
611
612
613
614
615

        # Initialize KV cache quantization attributes
        _init_kv_cache_quant(
            self, quant_config, prefix, kv_cache_dtype, calculate_kv_scales
        )
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646

        dtype = torch.get_default_dtype()
        self.attn_backend = get_attn_backend(
            self.head_size,
            dtype,
            kv_cache_dtype,
            block_size,
            use_mla=True,
            use_sparse=use_sparse,
        )
        impl_cls = cast(type[MLAAttentionImpl], self.attn_backend.get_impl_cls())
        self.impl = impl_cls(
            num_heads=self.num_heads,
            head_size=self.head_size,
            scale=self.scale,
            num_kv_heads=1,
            alibi_slopes=None,
            sliding_window=None,
            kv_cache_dtype=self.kv_cache_dtype,
            logits_soft_cap=None,
            attn_type=AttentionType.DECODER,
            kv_sharing_target_layer_name=None,
            # MLA Args
            q_lora_rank=self.q_lora_rank,
            kv_lora_rank=self.kv_lora_rank,
            qk_nope_head_dim=self.qk_nope_head_dim,
            qk_rope_head_dim=self.qk_rope_head_dim,
            qk_head_dim=self.qk_nope_head_dim + self.qk_rope_head_dim,
            v_head_dim=self.v_head_dim,
            kv_b_proj=kv_b_proj,
            indexer=indexer,
647
            **extra_impl_args,
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
        )

        self.use_direct_call = not current_platform.opaque_attention_op()

        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.kv_cache = [
            torch.tensor([])
            for _ in range(
                get_current_vllm_config().parallel_config.pipeline_parallel_size
            )
        ]

        self.use_sparse = use_sparse

        # Initialize q/k/v range constants.
667
668
669
        self.q_range = torch.tensor(envs.Q_SCALE_CONSTANT, dtype=torch.float32)
        self.k_range = torch.tensor(envs.K_SCALE_CONSTANT, dtype=torch.float32)
        self.v_range = torch.tensor(envs.V_SCALE_CONSTANT, dtype=torch.float32)
670
671
672
673
674
675

    def forward(
        self,
        q: torch.Tensor,
        kv_c_normed: torch.Tensor,
        k_pe: torch.Tensor,
676
        output_shape: torch.Size | None = None,
677
    ) -> torch.Tensor:
678
679
680
        if self.calculate_kv_scales:
            torch.ops.vllm.maybe_calc_kv_scales(q, kv_c_normed, k_pe, self.layer_name)

681
682
683
684
685
686
687
688
        if self.use_direct_call:
            forward_context: ForwardContext = get_forward_context()
            attn_metadata = forward_context.attn_metadata
            if isinstance(attn_metadata, dict):
                attn_metadata = attn_metadata[self.layer_name]
            self_kv_cache = self.kv_cache[forward_context.virtual_engine]

            if self.attn_backend.accept_output_buffer:
689
                output = torch.empty(output_shape, dtype=q.dtype, device=q.device)
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
                self.impl.forward(
                    self,
                    q,
                    kv_c_normed,
                    k_pe,
                    self_kv_cache,
                    attn_metadata,
                    output=output,
                )
                return output
            else:
                return self.impl.forward(
                    self, q, kv_c_normed, k_pe, self_kv_cache, attn_metadata
                )
        else:
            if self.attn_backend.accept_output_buffer:
706
                output = torch.empty(output_shape, dtype=q.dtype, device=q.device)
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
                torch.ops.vllm.unified_mla_attention_with_output(
                    q,
                    kv_c_normed,
                    k_pe,
                    output,
                    self.layer_name,
                )
                return output
            else:
                return torch.ops.vllm.unified_mla_attention(
                    q,
                    kv_c_normed,
                    k_pe,
                    self.layer_name,
                )

    def process_weights_after_loading(self, act_dtype: torch.dtype):
        if hasattr(self.impl, "process_weights_after_loading"):
            self.impl.process_weights_after_loading(act_dtype)

    def calc_kv_scales(
        self, q: torch.Tensor, kv_c_normed: torch.Tensor, k_pe: torch.Tensor
    ) -> None:
        """Optional scale calculation for MLA inputs.

        Mirrors Attention.calc_kv_scales. Not all MLA backends require this
        """
        # Use safe defaults if ranges are not present
        q_range = getattr(self, "q_range", torch.tensor(1.0))
        k_range = getattr(self, "k_range", torch.tensor(1.0))
        v_range = getattr(self, "v_range", torch.tensor(1.0))

        self._q_scale.copy_(torch.abs(q).max() / q_range)
        # kv_c_normed is the compressed KV representation; use it for k/v
        kv_abs_max = torch.abs(kv_c_normed).max()
        self._k_scale.copy_(kv_abs_max / k_range)
        self._v_scale.copy_(kv_abs_max / v_range)
        self._q_scale_float = self._q_scale.item()
        self._k_scale_float = self._k_scale.item()
        self._v_scale_float = self._v_scale.item()
        self.calculate_kv_scales = False

    def get_attn_backend(self) -> type[AttentionBackend]:
        return self.attn_backend

752
753
754
755
756
757
758
759
760
761
762
763
    def get_kv_cache_spec(self, vllm_config: VllmConfig) -> KVCacheSpec:
        kv_cache_dtype = kv_cache_dtype_str_to_dtype(
            self.kv_cache_dtype, vllm_config.model_config
        )
        return MLAAttentionSpec(
            block_size=vllm_config.cache_config.block_size,
            num_kv_heads=1,
            head_size=self.head_size,
            dtype=kv_cache_dtype,
            cache_dtype_str=vllm_config.cache_config.cache_dtype,
        )

764

765
766
767
768
769
770
771
def maybe_calc_kv_scales(
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    layer_name: str,
) -> None:
    forward_context: ForwardContext = get_forward_context()
772
    self = forward_context.no_compile_layers[layer_name]
773

774
775
776
    # Only calculate if the layer's calculate_kv_scales flag is True
    # This flag gets set to False after the first forward pass
    if not self.calculate_kv_scales:
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
        return

    self.calc_kv_scales(query, key, value)


def maybe_calc_kv_scales_fake(
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    layer_name: str,
) -> None:
    return


direct_register_custom_op(
    op_name="maybe_calc_kv_scales",
    op_func=maybe_calc_kv_scales,
    mutates_args=["query", "key", "value"],
    fake_impl=maybe_calc_kv_scales_fake,
)


799
def get_attention_context(
800
    layer_name: str,
801
802
803
804
805
806
807
808
) -> tuple[dict | object | None, Attention | MLAAttention, torch.Tensor]:
    """Extract attention context for a given layer.

    This helper function extracts the attention metadata, attention layer
    instance, and KV cache tensor for a specific layer.

    Args:
        layer_name: The name/identifier of the attention layer.
809

810
811
812
813
814
815
816
817
818
819
    Returns:
        A tuple containing:
        - attn_metadata: Attention metadata for this specific layer, or None if
            no metadata available
        - attn_layer: The attention layer instance (Attention or MLAAttention)
        - kv_cache: The KV cache tensor for current virtual engine

        Note: attn_metadata may be None, but attn_layer and kv_cache are always
        extracted from the forward context.
    """
820
    forward_context: ForwardContext = get_forward_context()
821
    attn_metadata = forward_context.attn_metadata
822
823
    if isinstance(attn_metadata, dict):
        attn_metadata = attn_metadata[layer_name]
824
825
826
827
828
829
830
831
832
833
834
835
836
    attn_layer: Attention | MLAAttention = forward_context.no_compile_layers[layer_name]
    kv_cache = attn_layer.kv_cache[forward_context.virtual_engine]
    return attn_metadata, attn_layer, kv_cache


@maybe_transfer_kv_layer
def unified_attention(
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    layer_name: str,
) -> torch.Tensor:
    attn_metadata, self, kv_cache = get_attention_context(layer_name)
837
    output = self.impl.forward(self, query, key, value, kv_cache, attn_metadata)
838
839

    return output
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855


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,
    fake_impl=unified_attention_fake,
)
856
857


858
@maybe_transfer_kv_layer
859
860
861
862
863
864
def unified_attention_with_output(
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    output: torch.Tensor,
    layer_name: str,
865
866
    output_scale: torch.Tensor | None = None,
    output_block_scale: torch.Tensor | None = None,
867
) -> None:
868
    attn_metadata, self, kv_cache = get_attention_context(layer_name)
869
870
871
872
873
874
875
876
877
878
879
    self.impl.forward(
        self,
        query,
        key,
        value,
        kv_cache,
        attn_metadata,
        output=output,
        output_scale=output_scale,
        output_block_scale=output_block_scale,
    )
880
881
882
883
884
885
886
887


def unified_attention_with_output_fake(
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    output: torch.Tensor,
    layer_name: str,
888
889
    output_scale: torch.Tensor | None = None,
    output_block_scale: torch.Tensor | None = None,
890
891
892
893
894
895
896
) -> None:
    return


direct_register_custom_op(
    op_name="unified_attention_with_output",
    op_func=unified_attention_with_output,
897
    mutates_args=["output", "output_block_scale"],
898
899
    fake_impl=unified_attention_with_output_fake,
)
900
901


902
@maybe_transfer_kv_layer
903
904
905
906
907
908
def unified_mla_attention(
    q: torch.Tensor,
    kv_c_normed: torch.Tensor,
    k_pe: torch.Tensor,
    layer_name: str,
) -> torch.Tensor:
909
    attn_metadata, self, kv_cache = get_attention_context(layer_name)
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
    output = self.impl.forward(self, q, kv_c_normed, k_pe, kv_cache, attn_metadata)

    return output


def unified_mla_attention_fake(
    q: torch.Tensor,
    kv_c_normed: torch.Tensor,
    k_pe: torch.Tensor,
    layer_name: str,
) -> torch.Tensor:
    return torch.empty_like(q).contiguous()


direct_register_custom_op(
    op_name="unified_mla_attention",
    op_func=unified_mla_attention,
    mutates_args=[],
    fake_impl=unified_mla_attention_fake,
    dispatch_key=current_platform.dispatch_key,
)


933
@maybe_transfer_kv_layer
934
935
936
937
938
939
def unified_mla_attention_with_output(
    q: torch.Tensor,
    kv_c_normed: torch.Tensor,
    k_pe: torch.Tensor,
    output: torch.Tensor,
    layer_name: str,
940
941
    output_scale: torch.Tensor | None = None,
    output_block_scale: torch.Tensor | None = None,
942
) -> None:
943
    attn_metadata, self, kv_cache = get_attention_context(layer_name)
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
    self.impl.forward(
        self,
        q,
        kv_c_normed,
        k_pe,
        kv_cache,
        attn_metadata,
        output=output,
        output_scale=output_scale,
        output_block_scale=output_block_scale,
    )


def unified_mla_attention_with_output_fake(
    q: torch.Tensor,
    kv_c_normed: torch.Tensor,
    k_pe: torch.Tensor,
    output: torch.Tensor,
    layer_name: str,
963
964
    output_scale: torch.Tensor | None = None,
    output_block_scale: torch.Tensor | None = None,
965
966
967
968
969
970
971
972
973
974
975
) -> None:
    return


direct_register_custom_op(
    op_name="unified_mla_attention_with_output",
    op_func=unified_mla_attention_with_output,
    mutates_args=["output", "output_block_scale"],
    fake_impl=unified_mla_attention_with_output_fake,
    dispatch_key=current_platform.dispatch_key,
)