common.py 74.8 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
"""
4
5
# MLA Common Components

6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
This file implements common components for MLA implementations.

First we define:

Sq      as Q sequence length
Skv     as KV sequence length

MLA has two possible ways of computing, a data-movement friendly approach and a
compute friendly approach, we generally want to use the compute friendly
approach for "prefill" (i.e. the ratio Sq / Skv is "small", is near 1)
and the data-movement friendly approach for "decode" (i.e. the ratio
Sq / Skv is "large").

NOTE what we deem small and large is currently determined by if its labelled
prefill or decode by the scheduler, but this is something we should probably
tune.

Main reference: DeepseekV2 paper, and FlashInfer Implementation
(https://arxiv.org/abs/2405.04434 and https://github.com/flashinfer-ai/flashinfer/pull/551).

Deepseek's MLA attention works the following way:
Matthew Bonanni's avatar
Matthew Bonanni committed
27
* Use a single latent vector to represent the per-token entry of the KV cache.
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
* For decode (i.e. the memory friendly approach) the attention "simulates" a
multi-head attention, while the compute is similar to multi-query attention.

Below is example of both paths assuming batchsize = 1

## More Extent Definitions:

C           Context length, `Skv - Sq`
H           hidden size
N           number of attention heads
Lq          latent dimension for Q              1536 in DSV3
Lkv         latent dimension for K/V            512 in DSV3
P           nope dimension, no rope.            128 in DSV3
R           rope dimension, goes through rope.  64 in DSV3
V           V head dim.                         128 in DSV3

## Vector/Matrix Definitions

h_t         hidden states (input to attention)  shape [Sq, H]
q_c         latent/compressed Q                 shape [Sq, Lq]
q_nope      uncompressed Q (no-rope)            shape [Sq, N, P]
q_pe        uncompressed Q (rope)               shape [Sq, N, R]
kv_c        latent/compressed KV                shape [Skv, Lkv]
k_pe        decoupled k position embeddings     shape [Skv, R]
new_kv_c    new kv_c from current iter          shape [Sq, Lkv]
new_k_pe    new k_pe from current iter          shape [Sq, R]
cache_kv_c  cached k_c from previous iters      shape [C, Lkv]
cache_k_pe  cached k_pe from previous iters     shape [C, R]
W_DQ        project h_t to q_c                  shape [H, Lq]
W_UQ        project q_c to q_nope               shape [Lq, N * P]
W_QR        project q_c to q_pe                 shape [Lq, N * R]
W_DKV       project h_t to kv_c                 shape [H, Lkv]
60
61
62
W_UK        project kv_c to k_nope              shape [Lkv, N, P]
W_KR        project h_t to k_pe                 shape [H, R]
W_UV        project kv_c to v                   shape [Lkv, N, V]
63
64
65
66
67
68
69
70
71
72
73
74
W_O         project v to h_t                    shape [N * V, H]


## Compute Friendly Approach (i.e. "_forward_prefill"):

q_c      = h_t @ W_DQ
q_nope   = (q_c @ W_UQ).view(Sq, N, P)
q_pe     = RoPE(q_c @ W_QR).view(Sq, N, R)
new_kv_c = h_t @ W_DKV
new_k_pe = RoPE(h_t @ W_KR)
kv_c     = torch.cat([new_kv_c, cache_kv_c], dim=0)
k_pe     = torch.cat([new_k_pe, cache_k_pe], dim=0)
75
76
k_nope   = (kv_c @ W_UK.view(Lkv, N * P)).view(Skv, N, P)
v        = (kv_c @ W_UV.view(Lkv, N * V)).view(Skv, N, V)
77
78
79
80
81
82
83
84

// MHA with QK headdim = P + R
//           V headdim = V
//      spda_o shape [Sq, N, V]
spda_o = scaled_dot_product_attention(
    torch.cat([q_nope, q_pe], dim=-1),
    torch.cat([k_nope, k_pe.unsqueeze(1).expand(-1, N, -1)], dim=-1),
    v
Matthew Bonanni's avatar
Matthew Bonanni committed
85
)
86
87
88
return spda_o @ W_O

NOTE: in the actual code,
89
90
    `kv_b_proj` is [W_UK; W_UV] concatenated per head
    `q_b_proj` is [W_UQ; W_QR] concatenated per head
91
92
93
94
95
96
97
    `out_proj` is W_O


## Data-Movement Friendly Approach (i.e. "_forward_decode"):

Runtime
q_c      = h_t @ W_DQ
98
99
q_nope   = (q_c @ W_UQ).view(-1, N, P)
ql_nope  = einsum("snh,lnh->snl", q, W_UK)
100
101
102
103
104
105
106
107
108
109
110
111
q_pe     = RoPE(q_c @ W_QR).view(Sq, N, R)
new_kv_c = h_t @ W_DKV
new_k_pe = RoPE(h_t @ W_KR)
kv_c     = torch.cat([new_kv_c, cache_kv_c], dim=0)
k_pe     = torch.cat([new_k_pe, cache_k_pe], dim=0)

// MQA with QK headdim = Lkv + R
//           V headdim = Lkv
//      spda_o shape [Sq, N, Lkv]
// NOTE: this is less compute-friendly since Lkv > P
//       but is more data-movement friendly since its MQA vs MHA
spda_o = scaled_dot_product_attention(
112
    torch.cat([ql_nope, q_pe], dim=-1),
113
114
115
    torch.cat([kv_c, k_pe], dim=-1),
    kv_c
)
116
117
118

o = einsum("snl,lnv->snv", spda_o.reshape(-1, N, Lkv), W_UV)
return o.view(-1, N * V) @ self.num_heads @ W_O
119
120
121
122


## Chunked Prefill

Matthew Bonanni's avatar
Matthew Bonanni committed
123
124
For chunked prefill we want to use the compute friendly algorithm. We are
assuming sufficiently large Sq / Skv ratio, in the future may want to switch to
125
126
127
128
129
the data-movement friendly approach if the chunk (i.e. `Sq`) is small.

However, the compute-friendly approach can potentially run out of memory if Skv
is large due to: `k_nope = (kv_c @ W_UK).view(Skv, N, P)`

Matthew Bonanni's avatar
Matthew Bonanni committed
130
131
To mitigate this, we chunk the computation of attention with respect to the
current context (i.e. `cache_kv_c` and `cache_k_pe`) so that we can used a
132
133
134
135
fixed workspace size.

The chunked prefill approach is as follows:

Matthew Bonanni's avatar
Matthew Bonanni committed
136
MCC        Max chunk of context to process per iter, computed dynamically,
137
138
139
140
141
142
143
           used to bound the memory usage

q_c        = h_t @ W_DQ
q_nope     = (q_c @ W_UQ).view(Sq, N, P)
q_pe       = RoPE(q_c @ W_QR).view(Sq, N, R)
new_kv_c   = h_t @ W_DKV
new_k_pe   = RoPE(h_t @ W_KR)
144
145
new_k_nope = (new_kv_c @ W_UK.view(Lkv, N * P)).view(Sq, N, P)
new_v      = (new_kv_c @ W_UV.view(Lkv, N * V)).view(Sq, N, V)
146
147
148
149
150
151
152
153
154
155
156
157

// MHA between queries and new KV
//     with QK headdim = P + R
//           V headdim = V
//    curr_o   shape [Sq, N, V]
//    curr_lse shape [N, Sq], this is just order FA returns
curr_o, curr_lse = scaled_dot_product_attention(
    torch.cat([q_nope, q_pe], dim=-1),
    torch.cat([new_k_nope, new_k_pe.unsqueeze(1).expand(-1, N, -1)], dim=-1),
    new_v,
    casual=True,
    return_softmax_lse=True
Matthew Bonanni's avatar
Matthew Bonanni committed
158
)
159
160
161
162
163

// Compute attention with the already existing context
for chunk_idx in range(cdiv(C, MCC)):
    chunk_start  = chunk_idx * MCC
    chunk_end    = min(chunk_start + MCC, C)
zhuwenwen's avatar
zhuwenwen committed
164
    Sc           = chunk_end - chunk_start
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
    cache_kv_c_chunk   = cache_kv_c[chunk_start:chunk_end]
    cache_k_pe_chunk   = cache_k_pe[chunk_start:chunk_end]
    cache_k_nope_chunk = (cache_kv_c_chunk @ W_UK).view(-1, N, P)
    cache_v_chunk      = (cache_kv_c_chunk @ W_UV).view(-1, N, V)

    chunk_o, chunk_lse = scaled_dot_product_attention(
        torch.cat([q_nope, q_pe], dim=-1),
        torch.cat([cache_k_nope_chunk,
                   cache_k_pe_chunk.unsqueeze(1).expand(-1, N, -1)],
                   dim=-1),
        cache_v_chunk,
        casual=False,
        return_softmax_lse=True
    )

    curr_o, curr_lse = merge_attn_states(
        suffix_output=curr_o,
        suffix_lse=curr_lse,
        prefix_output=chunk_o,
        prefix_lse=chunk_lse,
    )

return curr_o @ W_O
"""
王敏's avatar
王敏 committed
189
import os
190
191
import functools
from abc import abstractmethod
192
from dataclasses import dataclass, field
193
from typing import Generic, Optional, TypeVar, Union
194
195

import torch
196
from tqdm import tqdm
197

198
import vllm.envs as envs
199
from vllm import _custom_ops as ops
200
from vllm import envs
201
202
203
from vllm.attention.backends.abstract import (AttentionBackend, AttentionLayer,
                                              AttentionMetadata,
                                              MLAAttentionImpl)
204
from vllm.attention.backends.utils import get_mla_dims
205
from vllm.attention.ops.common import cp_lse_ag_out_rs
206
from vllm.attention.ops.merge_attn_states import merge_attn_states
207
from vllm.attention.utils.fa_utils import get_flash_attn_version
208
from vllm.config import VllmConfig, get_current_vllm_config
209
from vllm.distributed.parallel_state import get_dcp_group, is_global_first_rank
210
211
from vllm.logger import init_logger
from vllm.model_executor.layers.linear import (ColumnParallelLinear,
212
                                               LinearBase,
213
                                               UnquantizedLinearMethod)
Simon Mo's avatar
Simon Mo committed
214
from vllm.platforms import current_platform
215
from vllm.utils import cdiv, round_down
216
from vllm.utils.flashinfer import has_nvidia_artifactory
217
218
219
220
221
from vllm.v1.attention.backends.utils import (AttentionMetadataBuilder,
                                              CommonAttentionMetadata,
                                              get_per_layer_parameters,
                                              infer_global_hyperparameters,
                                              split_decodes_and_prefills)
222
from vllm.v1.kv_cache_interface import AttentionSpec
223
from vllm.v1.worker.block_table import BlockTable
224

225
226
try:
    from vllm.vllm_flash_attn import flash_attn_varlen_func
227
    is_vllm_fa = True
228
229
except ImportError:
    # For rocm use upstream flash attention
230
231
    if current_platform.is_rocm():
        from flash_attn import flash_attn_varlen_func
232
    is_vllm_fa = False
233

234
235
try:
    from flashinfer import BatchPrefillWithRaggedKVCacheWrapper
236
237
    from flashinfer.prefill import (  # noqa: F401
        cudnn_batch_prefill_with_kv_cache)
238
239
240
241
    flashinfer_available = True
except ImportError:
    flashinfer_available = False

242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263

def is_rocm_aiter_fp8bmm_enabled() -> bool:
    return current_platform.is_rocm() \
        and envs.VLLM_ROCM_USE_AITER_FP8BMM \
        and envs.VLLM_ROCM_USE_AITER


if is_rocm_aiter_fp8bmm_enabled():
    from aiter.ops.triton.batched_gemm_a8w8_a_per_token_group_prequant_w_per_batched_tensor_quant import (  # noqa: E501 # isort: skip
        batched_gemm_a8w8_a_per_token_group_prequant_w_per_batched_tensor_quant
        as aiter_triton_fp8_bmm)

    def dynamic_per_batched_tensor_quant(
            x: torch.Tensor, dtype: torch.dtype = torch.float8_e4m3fn):
        DTYPE_MAX = torch.finfo(dtype).max
        min_val, max_val = x.aminmax()
        amax = torch.maximum(min_val.abs(), max_val.abs()).clamp(min=1e-10)
        scale = DTYPE_MAX / amax
        x_scl_sat = (x * scale).clamp(min=-DTYPE_MAX, max=DTYPE_MAX)
        return x_scl_sat.to(dtype).contiguous(), scale.float().reciprocal()


264
265
logger = init_logger(__name__)

266
267
CUDNN_WORKSPACE_SIZE = 12800

268
269
270
271
272
273
274

class MLACommonBackend(AttentionBackend):

    accept_output_buffer: bool = True

    @staticmethod
    def get_name() -> str:
275
        return "TRITON_MLA"
276
277

    @staticmethod
278
    def get_metadata_cls() -> type["AttentionMetadata"]:
279
280
281
        return MLACommonMetadata

    @staticmethod
282
    def get_builder_cls() -> type["MLACommonMetadataBuilder"]:
283
284
285
286
287
288
289
290
        return MLACommonMetadataBuilder

    @staticmethod
    def get_kv_cache_shape(
        num_blocks: int,
        block_size: int,
        num_kv_heads: int,  # assumed to be 1 for MLA
        head_size: int,
291
        cache_dtype_str: str = "auto",
292
    ) -> tuple[int, ...]:
293
294
        return (num_blocks, block_size, head_size)

295
296
297
298
    @classmethod
    def get_supported_dtypes(cls) -> list[torch.dtype]:
        return [torch.float16, torch.bfloat16]

299
300
    @classmethod
    def get_supported_head_sizes(cls) -> list[int]:
301
302
        return [576]

303
304
305
306
307
308
309
310
311
312
313
    @classmethod
    def validate_head_size(cls, head_size: int) -> None:
        supported_head_sizes = cls.get_supported_head_sizes()
        if head_size not in supported_head_sizes:
            attn_type = cls.__name__.removesuffix("Backend")
            raise ValueError(
                f"Head size {head_size} is not supported by {attn_type}. "
                f"Supported head sizes are: {supported_head_sizes}. "
                "Set VLLM_ATTENTION_BACKEND=FLEX_ATTENTION to use "
                "FlexAttention backend which supports all head sizes.")

314
315

@dataclass
316
317
318
319
320
321
322
323
324
325
326
class MLACommonPrefillMetadata:
    """ Prefill Specific Metadata """

    @dataclass
    class ChunkedContextMetadata:
        # New for MLA (compared to FlashAttention)
        # For handling chunked prefill
        cu_seq_lens: torch.Tensor
        starts: torch.Tensor
        seq_tot: list[int]
        max_seq_lens: list[int]
327
        seq_lens: torch.Tensor
328
        workspace: torch.Tensor
329

330
331
332
333
334
335
336
        # for mla DCP
        cp_chunk_seq_lens: Optional[list[list[int]]] = None
        origin_context_lens: Optional[list[int]] = None
        cp_cu_seq_lens: Optional[torch.Tensor] = None
        chunk_size: Optional[int] = None
        cu_seq_lens_lst: Optional[list[list[int]]] = None

337
338
339
340
341
    block_table: torch.Tensor
    query_start_loc: torch.Tensor
    max_query_len: int
    chunked_context: Optional[ChunkedContextMetadata] = None

342

343
344
345
346
347
348
349
@dataclass
class FlashInferPrefillMetadata(MLACommonPrefillMetadata):
    prefill_main: Optional['BatchPrefillWithRaggedKVCacheWrapper'] = None
    prefill_chunks: list['BatchPrefillWithRaggedKVCacheWrapper'] = field(
        default_factory=list)


350
351
352
353
354
355
356
357
358
359
360
@dataclass
class CudnnPrefillMetadata(MLACommonPrefillMetadata):

    class ChunkedContextMetadata(
            MLACommonPrefillMetadata.ChunkedContextMetadata):
        seq_lens: torch.Tensor

    query_seq_lens: Optional[torch.Tensor] = None
    cudnn_workspace: Optional[torch.Tensor] = None


361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
@dataclass
class MLACommonDecodeMetadata:
    block_table: torch.Tensor
    seq_lens: torch.Tensor


D = TypeVar("D", bound=MLACommonDecodeMetadata)


@dataclass
class MLACommonMetadata(Generic[D]):
    """Metadata for MLACommon.

    NOTE: Please read the comment at the top of the file before trying to
    understand this class
    """
377
378
379
380
381
382
383
384
    # NOTE(sang): Definition of context_len, query_len, and seq_len.
    # |---------- N-1 iteration --------|
    # |---------------- N iteration ---------------------|
    # |- tokenA -|......................|-- newTokens ---|
    # |---------- context_len ----------|
    # |-------------------- seq_len ---------------------|
    #                                   |-- query_len ---|

385
386
    num_reqs: int
    max_query_len: int
387
    max_seq_len: int
388

389
390
391
392
    num_actual_tokens: int  # Number of tokens excluding padding.
    query_start_loc: torch.Tensor
    slot_mapping: torch.Tensor

393
394
395
396
397
398
    # New for MLA (compared to FlashAttention)
    # For handling prefill decode split
    num_decodes: int
    num_decode_tokens: int
    num_prefills: int

399
400
401
    # The dimension of the attention heads
    head_dim: Optional[int] = None

402
    decode: Optional[D] = None
403
    prefill: Optional[Union[MLACommonPrefillMetadata,
404
405
                            FlashInferPrefillMetadata,
                            CudnnPrefillMetadata]] = None
406
407

    def __post_init__(self):
408
409
        if self.head_dim is not None:
            MLACommonBackend.validate_head_size(self.head_dim)
410
411


412
M = TypeVar("M", bound=MLACommonMetadata)
413
A = TypeVar("A")
414
415


416
def use_flashinfer_prefill() -> bool:
417
    # For blackwell default to flashinfer prefill if it's available since
418
    # it is faster than FA2.
419
420
    return (not envs.VLLM_DISABLE_FLASHINFER_PREFILL and flashinfer_available
            and not envs.VLLM_USE_CUDNN_PREFILL
421
            and current_platform.is_device_capability(100))
422
423


424
def use_cudnn_prefill() -> bool:
425
426
427
    return (flashinfer_available and envs.VLLM_USE_CUDNN_PREFILL
            and current_platform.is_device_capability(100)
            and has_nvidia_artifactory())
428
429


430
# Currently 394MB, this can be tuned based on GEMM sizes used.
431
# Chosen to be the same as sglang:
432
433
434
435
#  https://github.com/sgl-project/sglang/blob/766392c6bda2558b61ce6d1c1bfd8081a549e1f1/python/sglang/global_config.py#L37
FLASHINFER_WORKSPACE_BUFFER_SIZE = 394 * 1024 * 1024


436
class MLACommonMetadataBuilder(AttentionMetadataBuilder[M]):
437
438
439
440
    """
    NOTE: Please read the comment at the top of the file before trying to
    understand this class
    """
441
    reorder_batch_threshold: int = 1
442

443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
    @staticmethod
    def determine_chunked_prefill_workspace_size(
            vllm_config: VllmConfig) -> int:
        scheduler_config = vllm_config.scheduler_config
        cache_config = vllm_config.cache_config
        model_config = vllm_config.model_config

        chunked_prefill_workspace_size = min(
            # Try for 8 full length request or at least 4 pages per-request
            max(8 * model_config.max_model_len,
                4 * scheduler_config.max_num_seqs * cache_config.block_size),
            # For long-context models try not to over-allocate limiting
            # kv-cache space, limiting it to 64k tokens,
            # which would result in the workspace being:
            #   2*(576)*(64*1024) = 144mb
            # (assuming 576 MLA head dim, and fp16)
            # which would result in up-projected context being
            #   2*(192*128)*(64*1024) = 3gb
            # (assuming 192 QK head dim, 128 heads, and fp16)
            64 * 1024)

        # Enforce that we enough for at least 1 page per request
        chunked_prefill_workspace_size = max(
            chunked_prefill_workspace_size,
            scheduler_config.max_num_seqs * cache_config.block_size)

        return chunked_prefill_workspace_size
470

471
    def __init__(self,
472
                 kv_cache_spec: AttentionSpec,
473
                 layer_names: list[str],
474
475
                 vllm_config: VllmConfig,
                 device: torch.device,
476
                 block_table: BlockTable,
477
478
479
                 metadata_cls: Optional[type[M]] = None):
        self.metadata_cls = metadata_cls \
            if metadata_cls is not None else MLACommonMetadata
480
481
482
483
        self.kv_cache_spec = kv_cache_spec
        scheduler_config = vllm_config.scheduler_config
        self.model_config = vllm_config.model_config
        parallel_config = vllm_config.parallel_config
484
485
486
        self.compilation_config = vllm_config.compilation_config
        self.device = device

487
488
489
        self.num_heads = self.model_config.get_num_attention_heads(
            parallel_config)
        self.mla_dims = get_mla_dims(self.model_config)
490
        self.aot_schedule = current_platform.is_cuda()
491
492
493
494
495
496
497
        try:
            self.dcp_world_size = get_dcp_group().world_size
            self.dcp_rank = get_dcp_group().rank_in_group
        except AssertionError:
            # DCP might not be initialized in testing
            self.dcp_world_size = 1
            self.dcp_rank = 0
498

499
        # Don't try to access the runner on AMD
500
        if self.aot_schedule:
501
            self.page_size = self.kv_cache_spec.block_size
502

503
504
505
        self.chunked_prefill_workspace_size = \
            self.determine_chunked_prefill_workspace_size(vllm_config)

506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
        if self.dcp_world_size > 1:
            # Note(hc): The local kvcache is incomplete when DCP is triggered,
            # an additional kvcache allgather across the DCP group is therefore
            # required, so the workspace has to be enlarged by 1/DCP relative
            # to the original TP allocation.
            assert self.chunked_prefill_workspace_size % \
                self.dcp_world_size == 0
            self.chunked_prefill_workspace = torch.empty(
                (self.chunked_prefill_workspace_size +
                 self.chunked_prefill_workspace_size // self.dcp_world_size,
                 self.model_config.get_head_size()),
                dtype=self.model_config.dtype,
                device=device,
            )
        else:
            self.chunked_prefill_workspace = torch.empty(
                (self.chunked_prefill_workspace_size,
                 self.model_config.get_head_size()),
                dtype=self.model_config.dtype,
                device=device,
            )
527
528

        self._use_cudnn_prefill = use_cudnn_prefill()
529
        self._use_fi_prefill = use_flashinfer_prefill()
530
531
532
533
        self.prefill_metadata_cls = (
            FlashInferPrefillMetadata
            if self._use_fi_prefill else CudnnPrefillMetadata
            if self._use_cudnn_prefill else MLACommonPrefillMetadata)
534
535
536
537
538

        if self._use_fi_prefill:
            self._workspace_buffer = torch.empty(
                FLASHINFER_WORKSPACE_BUFFER_SIZE,
                dtype=torch.uint8,
539
                device=device)
540
541
542
543
544
545
546

            self._fi_prefill_main: Optional[
                BatchPrefillWithRaggedKVCacheWrapper] = None
            self._fi_prefill_chunks: list[
                BatchPrefillWithRaggedKVCacheWrapper] = []

            self._global_hyperparameters = infer_global_hyperparameters(
547
548
                get_per_layer_parameters(vllm_config, layer_names,
                                         MLACommonImpl))
549

550
551
552
553
        if self._use_cudnn_prefill:
            self.cudnn_workspace = torch.empty(
                CUDNN_WORKSPACE_SIZE * scheduler_config.max_num_seqs,
                dtype=torch.int8,
554
                device=device,
555
556
            )

557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
    def _build_fi_prefill_wrappers(self, prefill: FlashInferPrefillMetadata):
        qo_indptr = prefill.query_start_loc

        has_context = False
        if prefill.chunked_context is not None:
            chunked_context = prefill.chunked_context
            has_context = True

        if self._fi_prefill_main is None:
            self._fi_prefill_main = BatchPrefillWithRaggedKVCacheWrapper(
                self._workspace_buffer, "NHD", backend="cutlass")

        if has_context:
            num_chunks = chunked_context.cu_seq_lens.shape[0]
            # Allocate more prefill chunk wrappers if needed
            if len(self._fi_prefill_chunks) < num_chunks:
                for _ in range(len(self._fi_prefill_chunks), num_chunks):
                    self._fi_prefill_chunks.append(
                        BatchPrefillWithRaggedKVCacheWrapper(
                            self._workspace_buffer, "NHD", backend="cutlass"))
            assert num_chunks <= len(self._fi_prefill_chunks)

        # In MLA, the non-latent num_qo_heads == num_kv_heads
580
        num_qo_heads = self.num_heads
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
        num_kv_heads = num_qo_heads

        # Sanity: Verify that num_kv_heads == 1 since it is latent space
        assert self.kv_cache_spec.num_kv_heads == 1

        # Get non-latent head_dim_qk and head_dim_vo
        head_dim_qk = (self.mla_dims.qk_nope_head_dim +
                       self.mla_dims.qk_rope_head_dim)
        head_dim_vo = self.mla_dims.v_head_dim

        # For main run, qo_indptr == kv_indptr
        kv_indptr = qo_indptr.clone()

        # Prepare main prefill
        self._fi_prefill_main.plan(
            qo_indptr=qo_indptr,
            kv_indptr=kv_indptr,
            num_qo_heads=num_qo_heads,
            num_kv_heads=num_kv_heads,
            head_dim_qk=head_dim_qk,
            head_dim_vo=head_dim_vo,
            causal=True,  # This is main run
            sm_scale=self._global_hyperparameters.sm_scale,
            window_left=self._global_hyperparameters.window_left,
            logits_soft_cap=self._global_hyperparameters.logits_soft_cap,
606
            q_data_type=self.model_config.dtype,
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
        )

        # Prepare context prefills
        if has_context:
            for i in range(num_chunks):
                kv_indptr_chunk = chunked_context.cu_seq_lens[i]

                self._fi_prefill_chunks[i].plan(
                    qo_indptr=qo_indptr,
                    kv_indptr=kv_indptr_chunk,
                    num_qo_heads=num_qo_heads,
                    num_kv_heads=num_kv_heads,
                    head_dim_qk=head_dim_qk,
                    head_dim_vo=head_dim_vo,
                    causal=False,  # This is context run
                    sm_scale=self._global_hyperparameters.sm_scale,
                    window_left=self._global_hyperparameters.window_left,
                    logits_soft_cap=self._global_hyperparameters.
                    logits_soft_cap,
626
                    q_data_type=self.model_config.dtype,
627
628
629
630
                )

        prefill.prefill_main = self._fi_prefill_main
        prefill.prefill_chunks = self._fi_prefill_chunks
631

632
    def _build_decode(self, block_table_tensor: torch.Tensor,
633
634
635
636
637
                      seq_lens_cpu: torch.Tensor,
                      seq_lens_device: torch.Tensor,
                      query_start_loc_cpu: torch.Tensor,
                      query_start_loc_device: torch.Tensor,
                      num_decode_tokens: int) -> MLACommonDecodeMetadata:
638
        return MLACommonDecodeMetadata(
639
            block_table=block_table_tensor,
640
            seq_lens=seq_lens_device,
641
642
        )

643
644
645
646
647
648
649
    def build_for_cudagraph_capture(
            self, common_attn_metadata: CommonAttentionMetadata) -> M:
        """
        This method builds the metadata for full cudagraph capture.
        Currently, only decode is supported for full cudagraphs with MLA.
        """
        m = common_attn_metadata
王敏's avatar
王敏 committed
650
651
652
653
654
655
656
        assert m.num_reqs <= (m.num_actual_tokens *
                              self.reorder_batch_threshold), \
            "MLA only supports decode-only full CUDAGraph capture. " \
            "Make sure all cudagraph capture sizes <= max_num_seq."

        assert m.max_query_len <= self.reorder_batch_threshold  # decode only

657
658
        return self.build(0, m)

659
660
661
662
    def build(self,
              common_prefix_len: int,
              common_attn_metadata: CommonAttentionMetadata,
              fast_build: bool = False) -> M:
663
        num_reqs = common_attn_metadata.num_reqs
664
        num_tokens = common_attn_metadata.num_actual_tokens
665
        max_query_len = common_attn_metadata.max_query_len
666
        max_seq_len = common_attn_metadata.max_seq_len
667

Simon Mo's avatar
Simon Mo committed
668
669
670
        # Note(simon): be careful about the CPU <> GPU memory movement in this
        # function. We should avoid GPU -> CPU sync as much as possible because
        # it blocks on all previous kernels.
671
672
673
        device = self.device
        block_table_tensor = common_attn_metadata.block_table_tensor
        slot_mapping = common_attn_metadata.slot_mapping
674

675
        query_start_loc = common_attn_metadata.query_start_loc
676
        query_start_loc_cpu = common_attn_metadata.query_start_loc_cpu
677
        seq_lens = common_attn_metadata.seq_lens
678
        seq_lens_cpu = common_attn_metadata.seq_lens_cpu
Simon Mo's avatar
Simon Mo committed
679

680
681
682
683
684
685
        query_seq_lens_cpu = query_start_loc_cpu[1:] - query_start_loc_cpu[:-1]

        num_computed_tokens_cpu = (common_attn_metadata.seq_lens_cpu -
                                   query_seq_lens_cpu)

        num_decodes, num_prefills, num_decode_tokens, num_prefill_tokens = \
686
687
            split_decodes_and_prefills(common_attn_metadata,
                                       decode_threshold=self.reorder_batch_threshold)
688

689
690
691
692
693
        # Note(hc): update seq_lens of decode reqs under DCP.
        if self.dcp_world_size > 1:
            seq_lens[:num_decodes] = seq_lens[:num_decodes] \
                // self.dcp_world_size + (self.dcp_rank <= \
                (seq_lens[:num_decodes] - 1) % self.dcp_world_size)
694
695
696
697

        assert num_decodes + num_prefills == num_reqs
        assert num_decode_tokens + num_prefill_tokens == num_tokens

698
        prefill_metadata = None
699
700
        if num_prefills > 0:
            reqs_start = num_decodes  # prefill_start
701

702
            context_lens_cpu = num_computed_tokens_cpu[reqs_start:num_reqs]
703
704
705
706
            # Note(hc): The context lengths in the perspective of dcp rank0.
            cp_context_lens_cpu = torch.ceil(context_lens_cpu.float() /
                                             self.dcp_world_size).int()
            origin_context_lens = context_lens_cpu.tolist()
Simon Mo's avatar
Simon Mo committed
707
708
            max_context_len_cpu = context_lens_cpu.max().item()
            num_prefills_with_context_cpu = (context_lens_cpu > 0).sum().item()
709
710
            prefill_query_start_loc = query_start_loc[
                reqs_start:] - query_start_loc[reqs_start]
711
712

            chunked_context_metadata = None
713
            if max_context_len_cpu > 0:
714
715
716
717
718
719
720
721
                # NOTE: it is recommend you read the `Chunked Prefill` section
                # in the comment at the top of the file before trying to
                # understand the following code

                # currently we allocate an equal amount of workspace for each
                # prefill in the batch, we could probably use a more advanced
                # algorithm here and allocate more workspace to prefills with
                # longer context lengths
Simon Mo's avatar
Simon Mo committed
722
723
                max_context_chunk = (self.chunked_prefill_workspace_size //
                                     num_prefills_with_context_cpu)
724

725
726
                if self.aot_schedule:
                    # align max_context_chunk to page_size by rounding down,
727
728
729
                    # currently the `gather_and_maybe_dequant_cache` kernel
                    # cannot handle `context_chunk_starts` that are not aligned
                    # to page_size
730
731
                    max_context_chunk = round_down(max_context_chunk,
                                                   self.page_size)
732
733

                assert max_context_chunk > 0
Simon Mo's avatar
Simon Mo committed
734
                num_chunks = cdiv(max_context_len_cpu, max_context_chunk)
735
736
737
738
739

                # if `max_context_chunk = 256`, `num_chunks = 3`, and
                #   `num_prefills_with_context = 4`, create a tensor that looks
                # like
                #  [[0, 0, 0, 0], [256, 256, 256, 256], [512, 512, 512, 512]]
Simon Mo's avatar
Simon Mo committed
740
741
                # Note(simon): this is done in CPU because of downstream's
                # of `to_list`.
742
                chunk_starts = \
Simon Mo's avatar
Simon Mo committed
743
                    torch.arange(num_chunks, dtype=torch.int32) \
744
                    .unsqueeze(1).expand(-1, num_prefills) \
745
                    * max_context_chunk
Simon Mo's avatar
Simon Mo committed
746
                chunk_ends = torch.min(context_lens_cpu.unsqueeze(0),
747
748
                                       chunk_starts + max_context_chunk)
                chunk_seq_lens = (chunk_ends - chunk_starts).clamp(min=0)
Simon Mo's avatar
Simon Mo committed
749
750

                cu_seq_lens_cpu = torch.zeros(num_chunks,
751
                                              num_prefills + 1,
Simon Mo's avatar
Simon Mo committed
752
753
754
755
756
757
                                              dtype=torch.int32,
                                              pin_memory=True)
                torch.cumsum(chunk_seq_lens,
                             dim=1,
                             out=cu_seq_lens_cpu[:, 1:],
                             dtype=torch.int32)
758

759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
                if self.dcp_world_size > 1:
                    # Note(hc): The above max_context_chunk already enforces
                    # block_size alignment, DCP just need the block_size can
                    # be divisible by dcp_world_size, because DCP use
                    # cp_gather_cache which not require `cp_chunk_starts`
                    # aligned to page_size.
                    assert max_context_chunk % self.dcp_world_size == 0
                    cp_max_context_chunk = max_context_chunk // \
                        self.dcp_world_size
                    cp_chunk_starts = \
                        torch.arange(num_chunks, dtype=torch.int32) \
                        .unsqueeze(1).expand(-1, num_prefills) \
                        * cp_max_context_chunk
                    cp_chunk_ends = torch.min(
                        cp_context_lens_cpu.unsqueeze(0),
                        cp_chunk_starts + cp_max_context_chunk)
                    cp_chunk_seq_lens = (cp_chunk_ends -
                                         cp_chunk_starts).clamp(min=0)

                    cp_cu_seq_lens_cpu = torch.zeros(num_chunks,
                                                     num_prefills + 1,
                                                     dtype=torch.int32,
                                                     pin_memory=True)
                    torch.cumsum(cp_chunk_seq_lens,
                                 dim=1,
                                 out=cp_cu_seq_lens_cpu[:, 1:],
                                 dtype=torch.int32)

787
788
789
790
                chunked_context_metadata_cls = \
                    CudnnPrefillMetadata.ChunkedContextMetadata \
                    if self._use_cudnn_prefill else \
                        MLACommonPrefillMetadata.ChunkedContextMetadata
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
                if self.dcp_world_size > 1:
                    chunked_context_metadata = \
                        chunked_context_metadata_cls(
                        cu_seq_lens=cu_seq_lens_cpu \
                            .to(device, non_blocking=True),
                        starts=cp_chunk_starts.to(device, non_blocking=True),
                        seq_tot=cp_chunk_seq_lens.sum(dim=1).tolist(),
                        max_seq_lens=chunk_seq_lens.max(dim=1).values.tolist(),
                        seq_lens=chunk_seq_lens,
                        workspace=self.chunked_prefill_workspace,
                        cp_chunk_seq_lens=cp_chunk_seq_lens.tolist(),
                        origin_context_lens=origin_context_lens,
                        cp_cu_seq_lens=cp_cu_seq_lens_cpu \
                            .to(device, non_blocking=True),
                        chunk_size=max_context_chunk,
                        cu_seq_lens_lst=cu_seq_lens_cpu.tolist(),
                    )
                else:
                    chunked_context_metadata = \
                        chunked_context_metadata_cls(
                        cu_seq_lens=cu_seq_lens_cpu \
                            .to(device, non_blocking=True),
                        starts=chunk_starts.to(device, non_blocking=True),
                        seq_tot=chunk_seq_lens.sum(dim=1).tolist(),
                        max_seq_lens=chunk_seq_lens.max(dim=1).values.tolist(),
                        seq_lens=chunk_seq_lens,
                        workspace=self.chunked_prefill_workspace,
                    )
819

820
821
822
                if self._use_cudnn_prefill:
                    chunked_context_metadata.seq_lens = chunk_seq_lens

823
824
825
                assert max(chunked_context_metadata.max_seq_lens) <= \
                    self.chunked_prefill_workspace_size

826
            prefill_metadata = self.prefill_metadata_cls(
827
                block_table=block_table_tensor[reqs_start:, ...],
828
                query_start_loc=prefill_query_start_loc,
Simon Mo's avatar
Simon Mo committed
829
                max_query_len=max_query_len,
830
831
832
                chunked_context=chunked_context_metadata,
            )

833
834
835
836
837
838
            if self._use_cudnn_prefill:
                assert isinstance(prefill_metadata, CudnnPrefillMetadata)
                prefill_metadata.query_seq_lens = prefill_query_start_loc[1:] \
                    - prefill_query_start_loc[:-1]
                prefill_metadata.cudnn_workspace = self.cudnn_workspace

zhuwenwen's avatar
zhuwenwen committed
839
        decode_metadata = None
王敏's avatar
王敏 committed
840
841
842
843
844
845
846
847
848
        if num_decodes > 0:
            decode_metadata = self._build_decode(
                block_table_tensor=block_table_tensor[:num_decodes, ...],
                seq_lens_cpu=seq_lens_cpu[:num_decodes],
                seq_lens_device=seq_lens[:num_decodes],
                query_start_loc_cpu=query_start_loc_cpu[:num_decodes + 1],
                query_start_loc_device=query_start_loc[:num_decodes + 1],
                num_decode_tokens=num_decode_tokens,
            )
849

850
        attn_metadata = self.metadata_cls(
851
852
            num_reqs=common_attn_metadata.num_reqs,
            max_query_len=common_attn_metadata.max_query_len,
853
            max_seq_len=max_seq_len,
854
            num_actual_tokens=num_tokens,
855
856
            query_start_loc=query_start_loc,
            slot_mapping=slot_mapping,
857
            head_dim=self.model_config.get_head_size(),
858
            # MLACommonMetadata Chunk prefill specific
859
860
861
            num_decodes=num_decodes,
            num_decode_tokens=num_decode_tokens,
            num_prefills=num_prefills,
862
863
            prefill=prefill_metadata,
            decode=decode_metadata,
864
865
        )

866
        if self._use_fi_prefill and num_prefills > 0:
867
868
869
870
871
            assert isinstance(attn_metadata.prefill, FlashInferPrefillMetadata)
            self._build_fi_prefill_wrappers(attn_metadata.prefill)

        return attn_metadata

872

873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
def reorg_kvcache(
    allgatered_kv_c_normed: torch.Tensor,
    allgatered_k_pe: torch.Tensor,
    cp_chunk_seq_lens_lst: list[int],
    origin_context_lens: list[int],
    cp_world_size: int,
    sum_seq_len: int,
    max_seq_len: int,
    chunk_size: int,
    chunk_idx: int,
    toks: int,
) -> tuple[torch.Tensor, torch.Tensor]:
    """
    reorg kvcache after cp local gather to tp layout for attn kernel.

    Args:
        cp_chunk_seq_lens_lst: chunk context lengths under CP.
        origin_context_lens: origin full context lengths under CP.
        cp_world_size: CP size.
        sum_seq_len: the sum of cp_chunk_seq_lens_lst.
        max_seq_len: the max value of cp_chunk_seq_lens_lst.
        chunk_size: equals to max_context_chunk from
            chunked_context_metadata building.
        chunk_idx: chunk idx of chunked_prefill.
        toks: the number of tokens for local gather cache.
    """
    kv_c_segments = []
    k_pe_segments = []
    src_token_idx = 0
    max_seq_len_check = 0
    for cp_chunk_seq_len, origin_context_len in zip(cp_chunk_seq_lens_lst,
                                                    origin_context_lens):
        chunk_context_len = chunk_size
        if cp_chunk_seq_len != 0:
            chunk_context_len = min(
                chunk_context_len, origin_context_len - chunk_size * chunk_idx)
        cp_target_rank = (chunk_context_len - 1) % cp_world_size
        cur_seq_len = 0
        for rank in range(cp_world_size):
            if rank > cp_target_rank and cp_chunk_seq_len:
                real_cp_chunk_seq_len = cp_chunk_seq_len - 1
            else:
                real_cp_chunk_seq_len = cp_chunk_seq_len
            if real_cp_chunk_seq_len:
                kv_c_segment = allgatered_kv_c_normed[rank * toks +
                                                      src_token_idx:rank *
                                                      toks + src_token_idx +
                                                      real_cp_chunk_seq_len]
                k_pe_segment = allgatered_k_pe[rank * toks +
                                               src_token_idx:rank * toks +
                                               src_token_idx +
                                               real_cp_chunk_seq_len]
                kv_c_segments.append(kv_c_segment)
                k_pe_segments.append(k_pe_segment)
                cur_seq_len += real_cp_chunk_seq_len
        max_seq_len_check = max(max_seq_len_check, cur_seq_len)
        src_token_idx += cp_chunk_seq_len
    reorganized_kv_c_normed = torch.cat(kv_c_segments, dim=0)
    reorganized_k_pe = torch.cat(k_pe_segments, dim=0)
    assert reorganized_kv_c_normed.shape[0] == sum_seq_len
    assert reorganized_k_pe.shape[0] == sum_seq_len
    assert max_seq_len_check == max_seq_len
    return reorganized_kv_c_normed, reorganized_k_pe


938
939
940
# TODO(Lucas): rename MLACommonBaseImpl -> MLACommonImpl,
# and MLACommonImpl -> MLACommonDenseImpl or somthing like that
class MLACommonBaseImpl(MLAAttentionImpl[A], Generic[A]):
941
942
943
944
945
946
947
948
949
950
951
    """
    NOTE: Please read the comment at the top of the file before trying to
    understand this class
    """

    def __init__(
        self,
        num_heads: int,
        head_size: int,
        scale: float,
        num_kv_heads: int,
952
        alibi_slopes: Optional[list[float]],
953
954
955
956
        sliding_window: Optional[int],
        kv_cache_dtype: str,
        logits_soft_cap: Optional[float],
        attn_type: str,
957
        kv_sharing_target_layer_name: Optional[str],
958
959
960
961
962
963
964
965
        # MLA Specific Arguments
        q_lora_rank: Optional[int],
        kv_lora_rank: int,
        qk_nope_head_dim: int,
        qk_rope_head_dim: int,
        qk_head_dim: int,
        v_head_dim: int,
        kv_b_proj: ColumnParallelLinear,
966
        indexer=None,
967
        q_pad_num_heads: Optional[int] = None,
968
    ) -> None:
969
970
971
        if kv_sharing_target_layer_name is not None:
            raise NotImplementedError("KV sharing is not supported for MLA")

972
973
974
975
976
977
978
979
980
981
982
983
984
        self.num_heads = num_heads
        self.head_size = head_size
        self.scale = float(scale)
        self.num_kv_heads = num_kv_heads
        self.kv_cache_dtype = kv_cache_dtype

        self.q_lora_rank = q_lora_rank
        self.kv_lora_rank = kv_lora_rank
        self.qk_nope_head_dim = qk_nope_head_dim
        self.qk_rope_head_dim = qk_rope_head_dim
        self.qk_head_dim = qk_head_dim
        self.v_head_dim = v_head_dim
        self.kv_b_proj = kv_b_proj
985

986
        self.indexer = indexer
987
        self.q_pad_num_heads = q_pad_num_heads
988
        
zhuwenwen's avatar
zhuwenwen committed
989
        self.use_llama_nn = os.environ.get('LLAMA_NN') == '1'
990

991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
    def process_weights_after_loading(self, act_dtype: torch.dtype):

        def get_layer_weight(layer):
            WEIGHT_NAMES = ("weight", "qweight", "weight_packed")
            for attr in WEIGHT_NAMES:
                if hasattr(layer, attr):
                    return getattr(layer, attr)
            raise AttributeError(
                f"Layer '{layer}' has no recognized weight attribute:"
                f" {WEIGHT_NAMES}.")

        def get_and_maybe_dequant_weights(layer: LinearBase):
            if not isinstance(layer.quant_method, UnquantizedLinearMethod):
                # NOTE: This should only be used offline, since it's O(N^3)
                eye = torch.eye(layer.input_size_per_partition,
                                dtype=act_dtype,
                                device=get_layer_weight(layer).device)
                dequant_weights = layer.quant_method.apply(layer,
                                                           eye,
                                                           bias=None)
                del eye
                # standardize to (output, input)
                return dequant_weights.T
1014
            return layer.weight if not envs.VLLM_USE_NN else layer.weight.T
1015
1016
1017
1018

        # we currently do not have quantized bmm's which are needed for
        # `W_UV` and `W_UK_T`, we just store fp16/bf16 copies and perform
        # the bmm's in 16-bit, the extra memory overhead of this is fairly low
1019
1020
1021
1022
        if self.use_llama_nn and isinstance(self.kv_b_proj.quant_method, UnquantizedLinearMethod):
            kv_b_proj_weight = get_and_maybe_dequant_weights(self.kv_b_proj)
        else:
            kv_b_proj_weight = get_and_maybe_dequant_weights(self.kv_b_proj).T
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
        assert kv_b_proj_weight.shape == (
            self.kv_lora_rank,
            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim)), (
                f"{kv_b_proj_weight.shape=}, "
                f"{self.kv_lora_rank=}, "
                f"{self.num_heads=}, "
                f"{self.qk_nope_head_dim=}, "
                f"{self.v_head_dim=}")
        kv_b_proj_weight = kv_b_proj_weight.view(
            self.kv_lora_rank,
            self.num_heads,
            self.qk_nope_head_dim + self.v_head_dim,
        )

        W_UK, W_UV = kv_b_proj_weight.split(
            [self.qk_nope_head_dim, self.v_head_dim], dim=-1)

        if is_rocm_aiter_fp8bmm_enabled():
            W_K = W_UK.transpose(0, 1)  # 16 512 128
            W_V = W_UV.permute(1, 2, 0)  # 16 128 512
            self.W_K, self.W_K_scale = dynamic_per_batched_tensor_quant(
                W_K, dtype=current_platform.fp8_dtype())
            self.W_V, self.W_V_scale = dynamic_per_batched_tensor_quant(
                W_V, dtype=current_platform.fp8_dtype())

            # The kernel operates on non-padded inputs. Hence, pre-compiling
            # triton kernel to avoid runtime compilation for unseen batch sizes
            # Pre-compile for batch sizes 1 to 1024 to cover most use-cases.
            # On DS-R1, this step adds roughly 50s to the model loading time.
            max_batch_size = 1024  # [ToDo] Find the optimal upper limit
            pre_compilation_list = list(range(1, max_batch_size + 1))
            if is_global_first_rank():
                pre_compilation_list = tqdm(
                    pre_compilation_list,
                    desc="[Aiter Triton] Pre-compiling fp8 BMM kernel",
                    total=max_batch_size,
                )

            for m in pre_compilation_list:
                x = torch.empty((self.W_K.shape[0], m, self.W_K.shape[2]),
                                dtype=torch.bfloat16,
                                device=self.W_K.device)
                aiter_triton_fp8_bmm(x,
                                     self.W_K,
                                     self.W_K_scale,
                                     group_size=128,
                                     transpose_bm=True)

                x = torch.empty((self.W_V.shape[0], m, self.W_V.shape[2]),
                                dtype=torch.bfloat16,
                                device=self.W_V.device)
                aiter_triton_fp8_bmm(x,
                                     self.W_V,
                                     self.W_V_scale,
                                     group_size=128,
                                     transpose_bm=True)
        else:
            # Convert from (L, N, V) to (N, L, V)
            self.W_UV = W_UV.transpose(0, 1)
            # Convert from (L, N, P) to (N, P, L)
            self.W_UK_T = W_UK.permute(1, 2, 0)

    def _v_up_proj(self, x: torch.Tensor, out: torch.Tensor):
        # Convert from (B, N, L) to (N, B, L)
        x = x.view(-1, self.num_heads, self.kv_lora_rank).transpose(0, 1)
        if is_rocm_aiter_fp8bmm_enabled():
            # Multiply + Transpose (N, B, L) x (N, L, V)->(N, B, V)->(B, N, V)
            x = aiter_triton_fp8_bmm(x,
                                     self.W_V,
                                     self.W_V_scale,
                                     group_size=128,
                                     transpose_bm=True)
            # Convert from (B, N, V) to (B, N * V)
            x = x.reshape(-1, self.num_heads * self.v_head_dim)
            # Copy result
            out.copy_(x)
        else:
            # Convert from (B, N * V) to (N, B, V)
            out = out.view(-1, self.num_heads, self.v_head_dim).transpose(0, 1)

            # Multiply (N, B, L) x (N, L, V) -> (N, B, V)
            torch.bmm(x, self.W_UV, out=out)  # Reuse "out" to make it "hot"

            # Convert from (N, B, V) to (B, N * V)
            out_new = out.transpose(0, 1).reshape(
                -1, self.num_heads * self.v_head_dim)

            # Adjust output buffer shape back to the original (B, N * V)
            N, B, V = out.shape
            out.resize_((B, N * V))
            out.copy_(out_new)  # Copy result


class MLACommonImpl(MLACommonBaseImpl[M], Generic[M]):
    """
    NOTE: Please read the comment at the top of the file before trying to
    understand this class
    """

    def __init__(self, *args, **kwargs) -> None:
        super().__init__(*args, **kwargs)

1125
1126
1127
1128
1129
        if use_flashinfer_prefill():
            logger.debug_once("Using FlashInfer prefill for MLA")
            self._run_prefill_context_chunk = self._run_prefill_context_chunk_fi
            self._run_prefill_new_tokens = self._run_prefill_new_tokens_fi
            self._pad_v = False
1130
1131
1132
1133
1134
1135
        elif use_cudnn_prefill():
            logger.debug_once("Using CUDNN prefill for MLA")
            self._run_prefill_context_chunk = \
                self._run_prefill_context_chunk_cudnn
            self._run_prefill_new_tokens = self._run_prefill_new_tokens_cudnn
            self._pad_v = False
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
        else:  # Use FlashAttention
            logger.debug_once("Using FlashAttention prefill for MLA")
            self._run_prefill_context_chunk = self._run_prefill_context_chunk_fa
            self._run_prefill_new_tokens = self._run_prefill_new_tokens_fa

            # Handle the differences between the flash_attn_varlen from
            # flash_attn and the one from vllm_flash_attn. The former is used on
            # RoCM and the latter has an additional parameter to control
            # FA2 vs FA3
            self.flash_attn_varlen_func = flash_attn_varlen_func
            self.vllm_flash_attn_version = get_flash_attn_version()
            if self.vllm_flash_attn_version is not None:
                self.flash_attn_varlen_func = \
                    functools.partial(flash_attn_varlen_func,
                                    fa_version=self.vllm_flash_attn_version)

            # For MLA the v head dim is smaller than qk head dim so we pad out
            # v with 0s to match the qk head dim for attention backends that do
            # not support different headdims
            # We don't need to pad V if we are on a hopper system with FA3
zhuwenwen's avatar
zhuwenwen committed
1156
1157
1158
1159
1160
1161
            if not current_platform.is_rocm():
                self._pad_v = self.vllm_flash_attn_version is None or not (
                    self.vllm_flash_attn_version == 3
                    and current_platform.get_device_capability()[0] == 9)
            else:
                self._pad_v = torch.cuda.get_device_properties(torch.cuda.current_device()).multi_processor_count == 120
1162
        
1163

1164
1165
        self.dcp_world_size: Optional[int] = None

1166
1167
1168
1169
        self.chunked_prefill_workspace_size = \
            MLACommonMetadataBuilder.determine_chunked_prefill_workspace_size(
            get_current_vllm_config())

1170
1171
1172
1173
1174
1175
1176
1177
1178
    def _flash_attn_varlen_diff_headdims(self,
                                         q,
                                         k,
                                         v,
                                         return_softmax_lse=False,
                                         softmax_scale=None,
                                         **kwargs):
        maybe_padded_v = v
        if self._pad_v:
zhuwenwen's avatar
zhuwenwen committed
1179
1180
            # maybe_padded_v = torch.nn.functional.pad(
            #     v, [0, q.shape[-1] - v.shape[-1]], value=0)
1181
            maybe_padded_v = torch.nn.functional.pad(
zhuwenwen's avatar
zhuwenwen committed
1182
1183
                    v, [0, q.shape[-1] - v.shape[-1]- 32], value=0)
            maybe_padded_v = maybe_padded_v[..., :-32].reshape(v.shape[0], v.shape[1],v.shape[2])
1184

1185
1186
1187
1188
1189
1190
1191
        if is_vllm_fa:
            kwargs["return_softmax_lse"] = return_softmax_lse
        else:
            # ROCm leverages the upstream flash_attn, which takes a parameter
            # called "return_attn_probs" instead of return_softmax_lse
            kwargs["return_attn_probs"] = return_softmax_lse

1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
        attn_out = self.flash_attn_varlen_func(
            q=q,
            k=k,
            v=maybe_padded_v,
            softmax_scale=softmax_scale,
            **kwargs,
        )

        # Unpack the output if there is multiple results
        lse = None
        if isinstance(attn_out, tuple):
            attn_out, lse = attn_out[0], attn_out[1]

        # Remain consistent with old `flash_attn_varlen_func` where there
        # is only one output tensor if `return_softmax_lse` is False.
        if return_softmax_lse:
            return attn_out, lse
        return attn_out

1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
    def _run_prefill_new_tokens_fa(self, prefill: MLACommonPrefillMetadata, q,
                                   k, v, return_softmax_lse):
        return self._flash_attn_varlen_diff_headdims(
            q=q,
            k=k,
            v=v,
            cu_seqlens_q=prefill.query_start_loc,
            cu_seqlens_k=prefill.query_start_loc,
            max_seqlen_q=prefill.max_query_len,
            max_seqlen_k=prefill.max_query_len,
            softmax_scale=self.scale,
            causal=True,
            return_softmax_lse=return_softmax_lse,
        )

    def _run_prefill_new_tokens_fi(self, prefill: MLACommonPrefillMetadata, q,
                                   k, v, return_softmax_lse):
        assert isinstance(prefill, FlashInferPrefillMetadata)
        assert prefill.prefill_main is not None
1230
        ret = prefill.prefill_main.run(
1231
1232
1233
1234
1235
1236
            q=q,
            k=k,
            v=v,
            return_lse=return_softmax_lse,
        )

1237
1238
1239
1240
1241
        if isinstance(ret, tuple):
            # Convert from (q_len, num_heads) to (num_heads, q_len)
            return ret[0], ret[1].transpose(0, 1).contiguous()
        return ret

1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
    def _run_prefill_new_tokens_cudnn(self, prefill: MLACommonPrefillMetadata,
                                      q, k, v, return_softmax_lse):
        assert isinstance(prefill, CudnnPrefillMetadata)
        assert prefill.query_seq_lens is not None
        output, lse = cudnn_batch_prefill_with_kv_cache(
            q=q,
            k_cache=k,
            v_cache=v,
            scale=self.scale,
            workspace_buffer=prefill.cudnn_workspace,
            max_token_per_sequence=prefill.max_query_len,
            max_sequence_kv=prefill.max_query_len,
            actual_seq_lens_q=prefill.query_seq_lens.view(-1, 1, 1, 1),
            actual_seq_lens_kv=prefill.query_seq_lens.view(-1, 1, 1, 1),
            causal=True,
            return_lse=True,  # do not support False for now
            is_cuda_graph_compatible=
            True,  #Indicates actual_seq_lens are on GPU or CPU.
        )
        if return_softmax_lse:
            return output, lse
        return output

1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
    def _run_prefill_context_chunk_fa(self, prefill: MLACommonPrefillMetadata,
                                      chunk_idx: int, q, k, v):
        assert prefill.chunked_context is not None
        return self._flash_attn_varlen_diff_headdims(
            q=q,
            k=k,
            v=v,
            cu_seqlens_q=prefill.query_start_loc,
            cu_seqlens_k=prefill.chunked_context.cu_seq_lens[chunk_idx],
            max_seqlen_q=prefill.max_query_len,
            max_seqlen_k=prefill.chunked_context.max_seq_lens[chunk_idx],
            softmax_scale=self.scale,
            causal=False,  # Context is unmasked
            return_softmax_lse=True,
        )

    def _run_prefill_context_chunk_fi(self, prefill: MLACommonPrefillMetadata,
                                      chunk_idx: int, q, k, v):
        assert isinstance(prefill, FlashInferPrefillMetadata)
1284
        attn_out, lse = prefill.prefill_chunks[chunk_idx].run(
1285
1286
1287
1288
1289
            q=q,
            k=k,
            v=v,
            return_lse=True,
        )
1290
1291
        # Convert from (q_len, num_heads) to (num_heads, q_len)
        return attn_out, lse.transpose(0, 1).contiguous()
1292

1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
    def _run_prefill_context_chunk_cudnn(self,
                                         prefill: MLACommonPrefillMetadata,
                                         chunk_idx: int, q, k, v):
        assert isinstance(prefill, CudnnPrefillMetadata)
        assert prefill.chunked_context is not None
        assert prefill.chunked_context.seq_lens[chunk_idx] is not None
        assert prefill.query_seq_lens is not None
        return cudnn_batch_prefill_with_kv_cache(
            q=q,
            k_cache=k,
            v_cache=v,
            scale=self.scale,
            workspace_buffer=prefill.cudnn_workspace,
            max_token_per_sequence=prefill.max_query_len,
            max_sequence_kv=prefill.chunked_context.max_seq_lens[chunk_idx],
            actual_seq_lens_q=prefill.query_seq_lens.view(-1, 1, 1, 1),
            actual_seq_lens_kv=prefill.chunked_context.seq_lens[chunk_idx].
            view(-1, 1, 1, 1),
            causal=False,
            return_lse=True,
            is_cuda_graph_compatible=
            True,  #Indicates actual_seq_lens are on GPU or CPU.
        )

1317
    def process_weights_after_loading(self, act_dtype: torch.dtype):
1318
1319

        def get_layer_weight(layer):
1320
1321
1322
1323
1324
1325
1326
            WEIGHT_NAMES = ("weight", "qweight", "weight_packed")
            for attr in WEIGHT_NAMES:
                if hasattr(layer, attr):
                    return getattr(layer, attr)
            raise AttributeError(
                f"Layer '{layer}' has no recognized weight attribute:"
                f" {WEIGHT_NAMES}.")
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339

        def get_and_maybe_dequant_weights(layer: LinearBase):
            if not isinstance(layer.quant_method, UnquantizedLinearMethod):
                # NOTE: This should only be used offline, since it's O(N^3)
                eye = torch.eye(layer.input_size_per_partition,
                                dtype=act_dtype,
                                device=get_layer_weight(layer).device)
                dequant_weights = layer.quant_method.apply(layer,
                                                           eye,
                                                           bias=None)
                del eye
                # standardize to (output, input)
                return dequant_weights.T
1340
            return layer.weight if not envs.VLLM_USE_NN else layer.weight.T
1341

1342
        # we currently do not have quantized bmm's which are needed for
1343
        # `W_UV` and `W_UK_T`, we just store fp16/bf16 copies and perform
1344
        # the bmm's in 16-bit, the extra memory overhead of this is fairly low
zhuwenwen's avatar
zhuwenwen committed
1345
1346
1347
1348
        if self.use_llama_nn and isinstance(self.kv_b_proj.quant_method, UnquantizedLinearMethod):
            kv_b_proj_weight = get_and_maybe_dequant_weights(self.kv_b_proj)
        else:
            kv_b_proj_weight = get_and_maybe_dequant_weights(self.kv_b_proj).T
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
        assert kv_b_proj_weight.shape == (
            self.kv_lora_rank,
            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim)), (
                f"{kv_b_proj_weight.shape=}, "
                f"{self.kv_lora_rank=}, "
                f"{self.num_heads=}, "
                f"{self.qk_nope_head_dim=}, "
                f"{self.v_head_dim=}")
        kv_b_proj_weight = kv_b_proj_weight.view(
            self.kv_lora_rank,
            self.num_heads,
            self.qk_nope_head_dim + self.v_head_dim,
        )

        W_UK, W_UV = kv_b_proj_weight.split(
            [self.qk_nope_head_dim, self.v_head_dim], dim=-1)

1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
        if is_rocm_aiter_fp8bmm_enabled():
            W_K = W_UK.transpose(0, 1)  # 16 512 128
            W_V = W_UV.permute(1, 2, 0)  # 16 128 512
            self.W_K, self.W_K_scale = dynamic_per_batched_tensor_quant(
                W_K, dtype=current_platform.fp8_dtype())
            self.W_V, self.W_V_scale = dynamic_per_batched_tensor_quant(
                W_V, dtype=current_platform.fp8_dtype())

            # The kernel operates on non-padded inputs. Hence, pre-compiling
            # triton kernel to avoid runtime compilation for unseen batch sizes
            # Pre-compile for batch sizes 1 to 1024 to cover most use-cases.
            # On DS-R1, this step adds roughly 50s to the model loading time.
            max_batch_size = 1024  # [ToDo] Find the optimal upper limit
            pre_compilation_list = list(range(1, max_batch_size + 1))
            if is_global_first_rank():
                pre_compilation_list = tqdm(
                    pre_compilation_list,
                    desc="[Aiter Triton] Pre-compiling fp8 BMM kernel",
                    total=max_batch_size,
                )

            for m in pre_compilation_list:
                x = torch.empty((self.W_K.shape[0], m, self.W_K.shape[2]),
                                dtype=torch.bfloat16,
                                device=self.W_K.device)
                aiter_triton_fp8_bmm(x,
                                     self.W_K,
                                     self.W_K_scale,
                                     group_size=128,
                                     transpose_bm=True)

                x = torch.empty((self.W_V.shape[0], m, self.W_V.shape[2]),
                                dtype=torch.bfloat16,
                                device=self.W_V.device)
                aiter_triton_fp8_bmm(x,
                                     self.W_V,
                                     self.W_V_scale,
                                     group_size=128,
                                     transpose_bm=True)
        else:
            # Convert from (L, N, V) to (N, L, V)
            self.W_UV = W_UV.transpose(0, 1)
            # Convert from (L, N, P) to (N, P, L)
            self.W_UK_T = W_UK.permute(1, 2, 0)
1410
1411
1412
1413
1414
1415

    def _compute_prefill_context(
        self,
        q: torch.Tensor,
        kv_c_and_k_pe_cache: torch.Tensor,
        attn_metadata: MLACommonMetadata,
1416
        k_scale: torch.Tensor,
1417
    ):
1418
1419
1420
        assert attn_metadata.prefill is not None
        prefill_metadata = attn_metadata.prefill
        assert prefill_metadata.chunked_context is not None
1421
1422

        output = None
1423
1424
        iters = len(prefill_metadata.chunked_context.seq_tot)
        workspace = prefill_metadata.chunked_context.workspace
1425
1426

        for i in range(iters):
1427
            toks = prefill_metadata.chunked_context.seq_tot[i]
1428

1429
            ops.gather_and_maybe_dequant_cache(
1430
1431
                src_cache=kv_c_and_k_pe_cache,
                dst=workspace,
1432
1433
                block_table=prefill_metadata.block_table,
                cu_seq_lens=prefill_metadata.chunked_context.cu_seq_lens[i],
1434
                batch_size=attn_metadata.num_prefills,
1435
1436
                kv_cache_dtype=self.kv_cache_dtype,
                scale=k_scale,
1437
                seq_starts=prefill_metadata.chunked_context.starts[i],
1438
1439
1440
            )

            kv_c_normed = workspace[:toks]\
1441
                [..., :self.kv_lora_rank]
1442
1443
1444
1445
1446
1447
1448
1449
            k_pe = workspace[:toks]\
                [..., self.kv_lora_rank:].unsqueeze(1)

            kv_nope = self.kv_b_proj(kv_c_normed)[0].view( \
                -1, self.num_heads, self.qk_nope_head_dim + self.v_head_dim)
            k_nope, v = kv_nope\
                .split([self.qk_nope_head_dim, self.v_head_dim], dim=-1)

1450
            if envs.VLLM_USE_OPT_CAT:
1451
                if k_nope.shape[0] > 1024:
zhuwenwen's avatar
zhuwenwen committed
1452
                    from vllm.v1.attention.backends.mla.test_concat import lightop_concat_prefill_helper
1453
1454
1455
1456
1457
                    k = lightop_concat_prefill_helper(k_nope, k_pe.expand((*k_nope.shape[:-1], -1)),
                                                      dim=2)
                else:
                    k = torch.cat((k_nope, k_pe.expand((*k_nope.shape[:-1], -1))),
                                  dim=-1)
1458
1459
1460
            else:
                k = torch.cat((k_nope, k_pe.expand((*k_nope.shape[:-1], -1))),
                            dim=-1)
1461

1462
1463
1464
            attn_output, attn_softmax_lse = self._run_prefill_context_chunk(
                prefill=prefill_metadata,
                chunk_idx=i,
1465
1466
                q=q,
                k=k,
1467
                v=v,
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
            )

            if output is None:
                output = attn_output
                output_lse = attn_softmax_lse
            else:
                output_tmp = torch.empty_like(output)
                output_lse_tmp = torch.empty_like(output_lse)
                merge_attn_states(
                    output=output_tmp,
                    output_lse=output_lse_tmp,
                    prefix_output=output,
                    prefix_lse=output_lse,
                    suffix_output=attn_output,
                    suffix_lse=attn_softmax_lse,
                )
                output = output_tmp
                output_lse = output_lse_tmp

        return output, output_lse

1489
1490
1491
1492
1493
1494
1495
1496
    def _context_parallel_compute_prefill_context(
        self,
        q: torch.Tensor,
        kv_c_and_k_pe_cache: torch.Tensor,
        attn_metadata: MLACommonMetadata,
        k_scale: torch.Tensor,
        dcp_world_size: int,
    ):
co63oc's avatar
co63oc committed
1497
        assert k_scale is None, "DCP not support scaled kvcache now."
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
        assert attn_metadata.prefill is not None
        prefill_metadata = attn_metadata.prefill
        assert prefill_metadata.chunked_context is not None
        assert prefill_metadata.chunked_context.cp_chunk_seq_lens is not None
        assert prefill_metadata.chunked_context.origin_context_lens is not None
        assert prefill_metadata.chunked_context.cp_cu_seq_lens is not None
        assert prefill_metadata.chunked_context.chunk_size is not None
        assert prefill_metadata.chunked_context.cu_seq_lens_lst is not None

        output = None
        iters = len(prefill_metadata.chunked_context.seq_tot)
        workspace = prefill_metadata.chunked_context.workspace

        for i in range(iters):
            toks = prefill_metadata.chunked_context.seq_tot[i]
            ops.cp_gather_cache(
                src_cache=kv_c_and_k_pe_cache,
                dst=workspace,
                block_table=prefill_metadata.block_table,
                cu_seq_lens=prefill_metadata.chunked_context.cp_cu_seq_lens[i],
                batch_size=attn_metadata.num_prefills,
                seq_starts=prefill_metadata.chunked_context.starts[i],
            )
            # workspace
            # |------- N tokens --------|--------- N*dcp_size tokens ----------|
            # |<- use for loca_gather ->|<--------- use for allgather -------->|
            allgather_offset = workspace.shape[0] // (dcp_world_size + 1)
            assert allgather_offset * (dcp_world_size +
                                       1) == workspace.shape[0]
            assert toks <= allgather_offset
            local_gathered_kvcache = workspace[:toks]
            cur_allgather_workspace = workspace[
                allgather_offset:allgather_offset * (1 + dcp_world_size)]
            assert toks * dcp_world_size <= cur_allgather_workspace.shape[0]
            cur_allgather_kvcache = cur_allgather_workspace[:toks *
                                                            dcp_world_size]
            cur_allgather_kvcache.copy_(get_dcp_group().all_gather(
                local_gathered_kvcache, dim=0))
            assert cur_allgather_kvcache.shape[
                -1] == self.kv_lora_rank + self.qk_rope_head_dim
            allgatered_kv_c_normed, allgatered_k_pe = \
                cur_allgather_kvcache.unsqueeze(
                1).split([self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)

            kv_c_normed, k_pe = reorg_kvcache(
                allgatered_kv_c_normed,
                allgatered_k_pe,
                cp_chunk_seq_lens_lst=prefill_metadata.chunked_context.
                cp_chunk_seq_lens[i],
                origin_context_lens=prefill_metadata.chunked_context.
                origin_context_lens,
                cp_world_size=dcp_world_size,
                sum_seq_len=prefill_metadata.chunked_context.cu_seq_lens_lst[i]
                [-1],
                max_seq_len=prefill_metadata.chunked_context.max_seq_lens[i],
                chunk_size=prefill_metadata.chunked_context.chunk_size,
                chunk_idx=i,
                toks=toks)

            kv_nope = self.kv_b_proj(kv_c_normed)[0].view( \
                -1, self.num_heads, self.qk_nope_head_dim + self.v_head_dim)
            k_nope, v = kv_nope\
                .split([self.qk_nope_head_dim, self.v_head_dim], dim=-1)
            k = torch.cat((k_nope, k_pe.expand((*k_nope.shape[:-1], -1))),
                          dim=-1)

            attn_output, attn_softmax_lse = self._run_prefill_context_chunk(
                prefill=prefill_metadata,
                chunk_idx=i,
                q=q,
                k=k,
                v=v,
            )

            if output is None:
                output = attn_output
                output_lse = attn_softmax_lse
            else:
                output_tmp = torch.empty_like(output)
                output_lse_tmp = torch.empty_like(output_lse)
                merge_attn_states(
                    output=output_tmp,
                    output_lse=output_lse_tmp,
                    prefix_output=output,
                    prefix_lse=output_lse,
                    suffix_output=attn_output,
                    suffix_lse=attn_softmax_lse,
                )
                output = output_tmp
                output_lse = output_lse_tmp

        return output, output_lse

1591
1592
1593
1594
1595
1596
1597
    def _forward_prefill(
        self,
        q: torch.Tensor,
        kv_c_normed: torch.Tensor,
        k_pe: torch.Tensor,
        kv_c_and_k_pe_cache: torch.Tensor,
        attn_metadata: MLACommonMetadata,
1598
        k_scale: torch.Tensor,
1599
    ) -> torch.Tensor:
1600
        # TODO (zyongye): Prefill function here
1601
        assert attn_metadata.prefill is not None
1602
        assert self.dcp_world_size is not None
1603

zhuwenwen's avatar
zhuwenwen committed
1604
1605
1606
1607
        if envs.VLLM_HAS_CONTEXT_DEFAULT:
            has_context = attn_metadata.prefill.chunked_context is not None
        else:
            has_context = False
1608
1609
1610
1611
1612
        kv_nope = self.kv_b_proj(kv_c_normed)[0].view(\
            -1, self.num_heads, self.qk_nope_head_dim + self.v_head_dim)
        k_nope, v = kv_nope\
            .split([self.qk_nope_head_dim, self.v_head_dim], dim=-1)

1613
        if envs.VLLM_USE_OPT_CAT:
1614
            if k_nope.shape[0] > 1024:
zhuwenwen's avatar
zhuwenwen committed
1615
                from vllm.v1.attention.backends.mla.test_concat import lightop_concat_prefill_helper
1616
1617
1618
1619
1620
                k = lightop_concat_prefill_helper(k_nope, k_pe.expand((*k_nope.shape[:-1], -1)),
                                                  dim=2)
            else:
                k = torch.cat((k_nope, k_pe.expand((*k_nope.shape[:-1], -1))),
                              dim=-1)
1621
1622
        else:
            k = torch.cat((k_nope, k_pe.expand((*k_nope.shape[:-1], -1))), dim=-1)
1623

1624
1625
        output = self._run_prefill_new_tokens(
            prefill=attn_metadata.prefill,
1626
1627
            q=q,
            k=k,
1628
            v=v,
1629
1630
1631
1632
1633
            return_softmax_lse=has_context,
        )

        if has_context:
            suffix_output, suffix_lse = output
1634
1635
1636
1637
1638
1639
1640
1641
1642
            if self.dcp_world_size > 1:
                context_output, context_lse = \
                    self._context_parallel_compute_prefill_context(
                    q, kv_c_and_k_pe_cache, attn_metadata,
                    k_scale=None, dcp_world_size=self.dcp_world_size)
            else:
                context_output, context_lse = \
                    self._compute_prefill_context(
                    q, kv_c_and_k_pe_cache, attn_metadata, k_scale)
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652

            output = torch.empty_like(suffix_output)
            merge_attn_states(
                output=output,
                prefix_output=context_output,
                prefix_lse=context_lse,
                suffix_output=suffix_output,
                suffix_lse=suffix_lse,
            )

1653
1654
1655
1656
        # unpad if necessary
        if self._pad_v:
            output = output[..., :v.shape[-1]]

1657
        return output.flatten(start_dim=-2)
1658
1659
1660
1661

    @abstractmethod
    def _forward_decode(
        self,
1662
        q: Union[torch.Tensor, tuple[torch.Tensor, torch.Tensor]],
1663
        kv_c_and_k_pe_cache: torch.Tensor,
1664
        attn_metadata: M,
1665
        layer: AttentionLayer,
1666
    ) -> tuple[torch.Tensor, Optional[torch.Tensor]]:
1667
1668
1669
1670
1671
        raise NotImplementedError

    def forward(
        self,
        layer: AttentionLayer,
1672
        q: torch.Tensor,
1673
1674
1675
        k_c_normed: torch.Tensor,  # key in unified attn
        k_pe: torch.Tensor,  # value in unified attn
        kv_cache: torch.Tensor,
1676
        attn_metadata: M,
1677
        output: Optional[torch.Tensor] = None,
1678
        output_scale: Optional[torch.Tensor] = None,
1679
        output_block_scale: Optional[torch.Tensor] = None,
1680
1681
1682
    ) -> torch.Tensor:
        assert output is not None, "Output tensor must be provided."

1683
        if output_scale is not None or output_block_scale is not None:
1684
1685
1686
1687
            raise NotImplementedError(
                "fused output quantization is not yet supported"
                " for MLACommonImpl")

1688
        if attn_metadata is None:
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
            # During the profile run try to simulate to worse case output size
            # for `self.kv_b_proj(kv_c_normed)` in `_compute_prefill_context`
            # since this can be large
            _ = torch.empty(
                (self.chunked_prefill_workspace_size, self.num_heads,
                 self.qk_nope_head_dim + self.v_head_dim),
                device=k_c_normed.device,
                dtype=k_c_normed.dtype,
            )

1699
1700
1701
1702
            # The zero fill is required when used with DP + EP
            # to ensure all ranks within a DP group compute the
            # same expert outputs.
            return output.fill_(0)
1703

1704
1705
1706
        if self.dcp_world_size is None:
            self.dcp_world_size = get_dcp_group().world_size

1707
1708
        fp8_attention = self.kv_cache_dtype.startswith("fp8")

1709
1710
1711
1712
1713
        num_actual_toks = attn_metadata.num_actual_tokens

        # Inputs and outputs may be padded for CUDA graphs
        output_padded = output
        output = output[:num_actual_toks, ...]
1714
        q = q[:num_actual_toks, ...]
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
        k_c_normed = k_c_normed[:num_actual_toks, ...]
        k_pe = k_pe[:num_actual_toks, ...]

        assert attn_metadata.num_decodes is not None and \
            attn_metadata.num_prefills is not None and \
            attn_metadata.num_decode_tokens is not None

        has_decode = attn_metadata.num_decodes > 0
        has_prefill = attn_metadata.num_prefills > 0
        num_decode_tokens = attn_metadata.num_decode_tokens

1726
        decode_q = q[:num_decode_tokens]
1727

1728
        prefill_q = q[num_decode_tokens:]
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
        prefill_k_pe = k_pe[num_decode_tokens:]
        prefill_k_c_normed = k_c_normed[num_decode_tokens:]

        # write the latent and rope to kv cache
        if kv_cache.numel() > 0:
            ops.concat_and_cache_mla(
                k_c_normed,
                k_pe.squeeze(1),
                kv_cache,
                attn_metadata.slot_mapping.flatten(),
                kv_cache_dtype=self.kv_cache_dtype,
                scale=layer._k_scale,
            )

1743
1744
1745
        if fp8_attention:
            kv_cache = kv_cache.view(current_platform.fp8_dtype())

1746
1747
1748
        if has_prefill:
            output[num_decode_tokens:] = self._forward_prefill(
                prefill_q, prefill_k_c_normed, prefill_k_pe, kv_cache,
1749
                attn_metadata, layer._k_scale)
1750
1751

        if has_decode:
1752
1753
1754
1755
1756
            assert attn_metadata.decode is not None
            decode_q_nope, decode_q_pe = decode_q.split(
                [self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)
            # Convert from (B, N, P) to (N, B, P)
            decode_q_nope = decode_q_nope.transpose(0, 1)
1757

1758
1759
1760
1761
1762
1763
1764
1765
1766
            # Pads the head_dim if necessary (for the underlying kernel)
            if self.q_pad_num_heads is not None:
                B, N, L = decode_q_pe.shape
                decode_pe_padded = decode_q_pe.new_empty(
                    (B, self.q_pad_num_heads, L))
                decode_pe_padded.resize_((B, N, L))
                decode_pe_padded.copy_(decode_q_pe)
                decode_q_pe = decode_pe_padded

1767
1768
1769
1770
1771
1772
1773
1774
            if is_rocm_aiter_fp8bmm_enabled():
                # Multiply+Transpose (N, B, P)x(N, P, L)->(N, B, L)->(B, N, L)
                decode_ql_nope = aiter_triton_fp8_bmm(decode_q_nope,
                                                      self.W_K,
                                                      self.W_K_scale,
                                                      group_size=128,
                                                      transpose_bm=True)
            else:
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
                # Pads the head_dim if necessary (for the underlying kernel)
                N, B, P = decode_q_nope.shape
                _, _, L = self.W_UK_T.shape
                if self.q_pad_num_heads is not None:
                    decode_ql_nope = decode_q_nope.new_empty(
                        (self.q_pad_num_heads, B, L))
                    decode_ql_nope.resize_((N, B, L))

                else:
                    decode_ql_nope = decode_q_nope.new_empty((N, B, L))

1786
                # Multiply (N, B, P) x (N, P, L) -> (N, B, L)
1787
                torch.bmm(decode_q_nope, self.W_UK_T, out=decode_ql_nope)
1788
1789
                # Convert from (N, B, L) to (B, N, L)
                decode_ql_nope = decode_ql_nope.transpose(0, 1)
1790

1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
            if fp8_attention:
                ql_nope_shape = decode_ql_nope.shape
                decode_ql_nope, _ = ops.scaled_fp8_quant(
                    decode_ql_nope.reshape([
                        ql_nope_shape[0], ql_nope_shape[1] * ql_nope_shape[2]
                    ]), layer._q_scale)
                decode_ql_nope = decode_ql_nope.reshape(ql_nope_shape)
                q_pe_shape = decode_q_pe.shape
                decode_q_pe, _ = ops.scaled_fp8_quant(
                    decode_q_pe.reshape(
                        [q_pe_shape[0], q_pe_shape[1] * q_pe_shape[2]]),
                    layer._q_scale)
                decode_q_pe = decode_q_pe.reshape(q_pe_shape)
1804

1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
            decode_q = (decode_ql_nope, decode_q_pe)
            if self.dcp_world_size > 1:
                assert not fp8_attention, "DCP not support fp8 kvcache now."
                # concatenate decode_ql_nope and decode_q_pe -> (B, N, L + P)
                decode_q = torch.cat(decode_q, dim=-1)
                # decode_q do allgather in head dim.
                decode_q = get_dcp_group().all_gather(decode_q, dim=1)

            # call decode attn
            attn_out, lse = self._forward_decode(decode_q, kv_cache,
                                                 attn_metadata, layer)

            # recorect dcp attn_out with lse.
            if self.dcp_world_size > 1:
                attn_out = cp_lse_ag_out_rs(attn_out, lse, get_dcp_group())

            # v_up projection
1822
            self._v_up_proj(attn_out, out=output[:num_decode_tokens])
1823
        return output_padded