common.py 51.6 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:
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
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

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)`

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:

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
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
189
190
191
    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
"""

import functools
from abc import abstractmethod
192
193

import numpy as np
194
from dataclasses import dataclass, field
195
from typing import TYPE_CHECKING, Generic, Optional, TypeVar, Union
196
197
198


import torch
zhuwenwen's avatar
zhuwenwen committed
199
import os
200

201
import vllm.envs as envs
202
from vllm import _custom_ops as ops
203
from vllm import envs
204
205
206
from vllm.attention.backends.abstract import (AttentionBackend, AttentionLayer,
                                              AttentionMetadata,
                                              MLAAttentionImpl)
207
from vllm.attention.backends.utils import get_mla_dims
208
from vllm.attention.ops.merge_attn_states import merge_attn_states
209
from vllm.attention.utils.fa_utils import get_flash_attn_version
210
from vllm.config import VllmConfig
211
212
from vllm.logger import init_logger
from vllm.model_executor.layers.linear import (ColumnParallelLinear,
213
                                               LinearBase,
214
                                               UnquantizedLinearMethod)
Simon Mo's avatar
Simon Mo committed
215
from vllm.platforms import current_platform
216
from vllm.utils import cdiv, round_down
217
218
219
220
from vllm.v1.attention.backends.utils import (
    AttentionMetadataBuilder, CommonAttentionMetadata,
    get_per_layer_parameters, infer_global_hyperparameters,
    reorder_batch_to_split_decodes_and_prefills, split_decodes_and_prefills)
221
from vllm.v1.kv_cache_interface import AttentionSpec
222
from vllm.v1.worker.block_table import BlockTable
223
224
225

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

233
234
try:
    from flashinfer import BatchPrefillWithRaggedKVCacheWrapper
235
236
    from flashinfer.prefill import (  # noqa: F401
        cudnn_batch_prefill_with_kv_cache)
237
238
239
240
    flashinfer_available = True
except ImportError:
    flashinfer_available = False

241
if TYPE_CHECKING:
242
    from vllm.v1.core.sched.output import SchedulerOutput
243
244
245
246
    from vllm.v1.worker.gpu_input_batch import InputBatch

logger = init_logger(__name__)

247
248
CUDNN_WORKSPACE_SIZE = 12800

249
250
251
252
253
254
255
256
257
258

class MLACommonBackend(AttentionBackend):

    accept_output_buffer: bool = True

    @staticmethod
    def get_name() -> str:
        return "TRITON_MLA_VLLM_V1"

    @staticmethod
259
    def get_metadata_cls() -> type["AttentionMetadata"]:
260
261
262
        return MLACommonMetadata

    @staticmethod
263
    def get_builder_cls() -> type["MLACommonMetadataBuilder"]:
264
265
266
267
268
269
270
271
        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,
272
    ) -> tuple[int, ...]:
273
274
        return (num_blocks, block_size, head_size)

275
276
277
278
    @classmethod
    def get_supported_dtypes(cls) -> list[torch.dtype]:
        return [torch.float16, torch.bfloat16]

279
280
    @classmethod
    def get_supported_head_sizes(cls) -> list[int]:
281
282
        return [576]

283
284
285
286
287
288
289
290
291
292
293
    @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.")

294
295

@dataclass
296
297
298
299
300
301
302
303
304
305
306
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]
307
        seq_lens: torch.Tensor
308
        workspace: torch.Tensor
309

310
311
312
313
314
    block_table: torch.Tensor
    query_start_loc: torch.Tensor
    max_query_len: int
    chunked_context: Optional[ChunkedContextMetadata] = None

315

316
317
318
319
320
321
322
@dataclass
class FlashInferPrefillMetadata(MLACommonPrefillMetadata):
    prefill_main: Optional['BatchPrefillWithRaggedKVCacheWrapper'] = None
    prefill_chunks: list['BatchPrefillWithRaggedKVCacheWrapper'] = field(
        default_factory=list)


323
324
325
326
327
328
329
330
331
332
333
@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


334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
@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
    """
350
351
352
353
354
355
356
357
    # NOTE(sang): Definition of context_len, query_len, and seq_len.
    # |---------- N-1 iteration --------|
    # |---------------- N iteration ---------------------|
    # |- tokenA -|......................|-- newTokens ---|
    # |---------- context_len ----------|
    # |-------------------- seq_len ---------------------|
    #                                   |-- query_len ---|

358
359
360
    num_reqs: int
    max_query_len: int

361
362
363
364
    num_actual_tokens: int  # Number of tokens excluding padding.
    query_start_loc: torch.Tensor
    slot_mapping: torch.Tensor

365
366
367
368
369
370
    # New for MLA (compared to FlashAttention)
    # For handling prefill decode split
    num_decodes: int
    num_decode_tokens: int
    num_prefills: int

371
372
373
    # The dimension of the attention heads
    head_dim: Optional[int] = None

374
    decode: Optional[D] = None
375
    prefill: Optional[Union[MLACommonPrefillMetadata,
376
377
                            FlashInferPrefillMetadata,
                            CudnnPrefillMetadata]] = None
378
379

    def __post_init__(self):
380
381
        if self.head_dim is not None:
            MLACommonBackend.validate_head_size(self.head_dim)
382
383


384
M = TypeVar("M", bound=MLACommonMetadata)
385
386


387
def use_flashinfer_prefill() -> bool:
388
    if flashinfer_available and not envs.VLLM_USE_CUDNN_PREFILL:
389
390
391
392
393
394
        # For blackwell default to flashinfer prefill if its available since
        #  its faster than FA2.
        return current_platform.has_device_capability(100)
    return False


395
396
397
398
399
400
def use_cudnn_prefill() -> bool:
    if flashinfer_available and envs.VLLM_USE_CUDNN_PREFILL:
        return current_platform.has_device_capability(100)
    return False


401
# Currently 394MB, this can be tuned based on GEMM sizes used.
402
# Chosen to be the same as sglang:
403
404
405
406
#  https://github.com/sgl-project/sglang/blob/766392c6bda2558b61ce6d1c1bfd8081a549e1f1/python/sglang/global_config.py#L37
FLASHINFER_WORKSPACE_BUFFER_SIZE = 394 * 1024 * 1024


407
class MLACommonMetadataBuilder(AttentionMetadataBuilder[M]):
408
409
410
411
412
    """
    NOTE: Please read the comment at the top of the file before trying to
    understand this class
    """

413
    def __init__(self,
414
                 kv_cache_spec: AttentionSpec,
415
416
                 vllm_config: VllmConfig,
                 device: torch.device,
417
                 block_table: BlockTable,
418
419
420
                 metadata_cls: Optional[type[M]] = None):
        self.metadata_cls = metadata_cls \
            if metadata_cls is not None else MLACommonMetadata
421
422
423
424
425
426
        self.kv_cache_spec = kv_cache_spec
        self.device = device
        scheduler_config = vllm_config.scheduler_config
        self.model_config = vllm_config.model_config
        cache_config = vllm_config.cache_config
        parallel_config = vllm_config.parallel_config
427
        self.chunked_prefill_enabled = scheduler_config.chunked_prefill_enabled
428
429
430
        self.num_heads = self.model_config.get_num_attention_heads(
            parallel_config)
        self.mla_dims = get_mla_dims(self.model_config)
431
        self.aot_schedule = current_platform.is_cuda()
432
433
434

        # Dont try to access the runner on AMD
        if self.aot_schedule:
435
            self.page_size = self.kv_cache_spec.block_size
436
437
438
439
440
441

        if self.chunked_prefill_enabled:
            self.chunked_prefill_workspace_size = min(
                # Max sure there is enough for 8 full length request or at least
                # 4 pages of cache per request
                max(
442
                    8 * self.model_config.max_model_len, 4 *
443
444
445
446
447
448
449
450
451
452
453
454
455
456
                    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)
                128 * 1024)
            assert self.chunked_prefill_workspace_size >= \
                scheduler_config.max_num_seqs * cache_config.block_size
            self.chunked_prefill_workspace = torch.empty(
                (self.chunked_prefill_workspace_size,
457
458
459
                 self.model_config.get_head_size()),
                dtype=self.model_config.dtype,
                device=device,
460
            )
461
462

        self._use_cudnn_prefill = use_cudnn_prefill()
463
        self._use_fi_prefill = use_flashinfer_prefill()
464
465
466
467
        self.prefill_metadata_cls = (
            FlashInferPrefillMetadata
            if self._use_fi_prefill else CudnnPrefillMetadata
            if self._use_cudnn_prefill else MLACommonPrefillMetadata)
468
469
470
471
472

        if self._use_fi_prefill:
            self._workspace_buffer = torch.empty(
                FLASHINFER_WORKSPACE_BUFFER_SIZE,
                dtype=torch.uint8,
473
                device=device)
474
475
476
477
478
479
480

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

            self._global_hyperparameters = infer_global_hyperparameters(
481
                get_per_layer_parameters(vllm_config, MLACommonImpl))
482

483
484
485
486
        if self._use_cudnn_prefill:
            self.cudnn_workspace = torch.empty(
                CUDNN_WORKSPACE_SIZE * scheduler_config.max_num_seqs,
                dtype=torch.int8,
487
                device=device,
488
489
            )

490
491
        self.block_table = block_table
        
492
493
        self.use_spec_decode = False
        self.num_scheduled_tokens_np = np.zeros(scheduler_config.max_num_seqs, dtype=np.int32)
494
        
495
496
497
498
        # support for cudagraph spec docoding
        self.spec_decode_block_table_tensor = None
        self.spec_decode_seq_lens = None
        
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
    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
522
        num_qo_heads = self.num_heads
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
        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,
548
            q_data_type=self.model_config.dtype,
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
            kv_data_type=self.kv_cache_spec.dtype,
        )

        # 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,
569
                    q_data_type=self.model_config.dtype,
570
571
572
573
574
                    kv_data_type=self.kv_cache_spec.dtype,
                )

        prefill.prefill_main = self._fi_prefill_main
        prefill.prefill_chunks = self._fi_prefill_chunks
575
576


577
    def reorder_batch(self, input_batch: "InputBatch",
578
                      scheduler_output: "SchedulerOutput") -> bool:
579
        self.use_spec_decode = len(
580
            scheduler_output.scheduled_spec_decode_tokens) > 0
581
582
        return reorder_batch_to_split_decodes_and_prefills(input_batch,
                                                           scheduler_output,
583
584
                                                           decode_threshold=1,
                                                           num_scheduled_tokens_np=self.num_scheduled_tokens_np)
585

586
587
    def _build_decode(self, block_table_tensor: torch.Tensor,
                      seq_lens: torch.Tensor):
588
        return MLACommonDecodeMetadata(
589
            block_table=block_table_tensor,
590
591
592
            seq_lens=seq_lens,
        )

593
594
595
596
597
598
599
    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
600
601
602
        # assert m.num_reqs == m.num_actual_tokens, \
        #     "MLA only supports decode-only full CUDAGraph capture. " \
        #     "Make sure all cudagraph capture sizes <= max_num_seq."
603

604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
        # m.max_query_len = 1  # decode-only
        
        self.use_spec_decode = m.num_speculative_tokens > 0
        # support for cudagraph spec docoding
        if self.use_spec_decode:
            for i in range(m.num_reqs):
                self.num_scheduled_tokens_np[i] = m.num_actual_tokens // m.num_reqs

            if self.spec_decode_block_table_tensor is None:
                max_num_reqs = m.seq_lens.shape[0]
                block_table_tensor = self.block_table.get_device_tensor()
                tokens_per_seq = 1+m.num_speculative_tokens

                self.spec_decode_block_table_tensor = torch.zeros((block_table_tensor.shape[0]*tokens_per_seq,
                                                                   block_table_tensor.shape[1]),
                                     dtype=block_table_tensor.dtype,
                                     device=m.seq_lens.device)
                self.spec_decode_seq_lens = torch.zeros(max_num_reqs * tokens_per_seq,
                                     dtype=m.seq_lens.dtype,
                                     device=m.seq_lens.device)
624
625
626

        return self.build(0, m)

627
628
629
630
    def build(self,
              common_prefix_len: int,
              common_attn_metadata: CommonAttentionMetadata,
              fast_build: bool = False) -> M:
631
        num_reqs = common_attn_metadata.num_reqs
632
        num_tokens = common_attn_metadata.num_actual_tokens
633
634
        max_query_len = common_attn_metadata.max_query_len

Simon Mo's avatar
Simon Mo committed
635
636
637
        # 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.
638
        device = self.device
639
        block_table = self.block_table
640
641
        block_table_tensor = common_attn_metadata.block_table_tensor
        slot_mapping = common_attn_metadata.slot_mapping
642
643
644
645
646
647
        if slot_mapping is None:
            block_table.slot_mapping[:num_tokens].copy_(
                block_table.slot_mapping_cpu[:num_tokens],
                non_blocking=True)
            block_table.slot_mapping[num_tokens:].fill_(-1)
            slot_mapping = block_table.slot_mapping[:num_tokens]
648

649
        query_start_loc = common_attn_metadata.query_start_loc
650
        query_start_loc_cpu = common_attn_metadata.query_start_loc_cpu
651
        seq_lens = common_attn_metadata.seq_lens
Simon Mo's avatar
Simon Mo committed
652

653
654
655
656
657
658
659
660
661
662
663
        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 = \
            split_decodes_and_prefills(common_attn_metadata)

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

664
        prefill_metadata = None
665
666
        if num_prefills > 0:
            reqs_start = num_decodes  # prefill_start
667

668
            context_lens_cpu = num_computed_tokens_cpu[reqs_start:num_reqs]
Simon Mo's avatar
Simon Mo committed
669
670
            max_context_len_cpu = context_lens_cpu.max().item()
            num_prefills_with_context_cpu = (context_lens_cpu > 0).sum().item()
671
672
            prefill_query_start_loc = query_start_loc[
                reqs_start:] - query_start_loc[reqs_start]
673
674

            chunked_context_metadata = None
675
            if self.chunked_prefill_enabled and num_prefills > 0 \
Simon Mo's avatar
Simon Mo committed
676
                and max_context_len_cpu > 0:
677
678
679
680
681
682
683
684
                # 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
685
686
                max_context_chunk = (self.chunked_prefill_workspace_size //
                                     num_prefills_with_context_cpu)
687

688
689
690
691
692
693
                if self.aot_schedule:
                    # align max_context_chunk to page_size by rounding down,
                    # currently the `gather_cache` kernel cannot handle
                    # `context_chunk_starts` that are not aligned to page_size
                    max_context_chunk = round_down(max_context_chunk,
                                                   self.page_size)
694
695

                assert max_context_chunk > 0
Simon Mo's avatar
Simon Mo committed
696
                num_chunks = cdiv(max_context_len_cpu, max_context_chunk)
697
698
699
700
701

                # 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
702
703
                # Note(simon): this is done in CPU because of downstream's
                # of `to_list`.
704
                chunk_starts = \
Simon Mo's avatar
Simon Mo committed
705
                    torch.arange(num_chunks, dtype=torch.int32) \
706
                    .unsqueeze(1).expand(-1, num_prefills) \
707
                    * max_context_chunk
Simon Mo's avatar
Simon Mo committed
708
                chunk_ends = torch.min(context_lens_cpu.unsqueeze(0),
709
710
                                       chunk_starts + max_context_chunk)
                chunk_seq_lens = (chunk_ends - chunk_starts).clamp(min=0)
Simon Mo's avatar
Simon Mo committed
711
712

                cu_seq_lens_cpu = torch.zeros(num_chunks,
713
                                              num_prefills + 1,
Simon Mo's avatar
Simon Mo committed
714
715
716
717
718
719
                                              dtype=torch.int32,
                                              pin_memory=True)
                torch.cumsum(chunk_seq_lens,
                             dim=1,
                             out=cu_seq_lens_cpu[:, 1:],
                             dtype=torch.int32)
720

721
722
723
724
725
                chunked_context_metadata_cls = \
                    CudnnPrefillMetadata.ChunkedContextMetadata \
                    if self._use_cudnn_prefill else \
                        MLACommonPrefillMetadata.ChunkedContextMetadata

726
                chunked_context_metadata = \
727
                    chunked_context_metadata_cls(
Simon Mo's avatar
Simon Mo committed
728
729
                    cu_seq_lens=cu_seq_lens_cpu.to(device, non_blocking=True),
                    starts=chunk_starts.to(device, non_blocking=True),
730
731
                    seq_tot=chunk_seq_lens.sum(dim=1).tolist(),
                    max_seq_lens=chunk_seq_lens.max(dim=1).values.tolist(),
732
                    seq_lens=chunk_seq_lens,
733
734
735
                    workspace=self.chunked_prefill_workspace,
                )

736
737
738
                if self._use_cudnn_prefill:
                    chunked_context_metadata.seq_lens = chunk_seq_lens

739
740
741
                assert max(chunked_context_metadata.max_seq_lens) <= \
                    self.chunked_prefill_workspace_size

742
            prefill_metadata = self.prefill_metadata_cls(
743
                block_table=block_table_tensor[reqs_start:, ...],
744
                query_start_loc=prefill_query_start_loc,
Simon Mo's avatar
Simon Mo committed
745
                max_query_len=max_query_len,
746
747
748
                chunked_context=chunked_context_metadata,
            )

749
750
751
752
753
754
            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

755
        decode_metadata = None
756
        if num_decodes > 0:
757
            if self.use_spec_decode:
758
                query_lens = self.num_scheduled_tokens_np[:num_decodes]
759
                if common_attn_metadata.num_rejected_tokens is not None:
760
                    num_rejected_tokens = common_attn_metadata.num_rejected_tokens[:num_decodes]
761
                    query_lens = query_lens - np.array(num_rejected_tokens, dtype=np.int32)
762
                    self._num_decode_tokens -= sum(num_rejected_tokens)
763
764
765
766
767
768
769
                cu_num_blocks = np.cumsum(query_lens)
                virtual_batches = cu_num_blocks[-1]
                block_offsets = np.repeat(cu_num_blocks - query_lens, query_lens)
                arange = np.arange(virtual_batches, dtype=np.int32) - block_offsets
                rarange = np.repeat(query_lens, query_lens) - arange - 1

                repeats = torch.from_numpy(query_lens).pin_memory().to(
770
                    block_table_tensor.device, non_blocking=True).contiguous()
771
                decode_block_table_tensor = torch.repeat_interleave(
772
                    block_table_tensor[:num_decodes, ...],
773
774
                    repeats, dim=0).contiguous()
                decode_seq_lens = torch.repeat_interleave(seq_lens[:num_decodes], repeats, dim=0).contiguous()
775
                seq_lens_minus = torch.from_numpy(rarange).to(torch.int32).pin_memory().to(
776
                    seq_lens.device, non_blocking=True).contiguous()
777
                decode_seq_lens = decode_seq_lens - seq_lens_minus
778

779
780
781
782
783
784
785
786
787
788
789
790
791
                if self.spec_decode_block_table_tensor is not None:
                    self.spec_decode_block_table_tensor[:self._num_decode_tokens].copy_(decode_block_table_tensor)
                    self.spec_decode_seq_lens[:self._num_decode_tokens].copy_(decode_seq_lens)

                    decode_metadata = self._build_decode(
                        block_table_tensor=self.spec_decode_block_table_tensor[:self._num_decode_tokens, ...],
                        seq_lens=self.spec_decode_seq_lens[:self._num_decode_tokens],
                    )
                else:
                    decode_metadata = self._build_decode(
                        block_table_tensor=decode_block_table_tensor,
                        seq_lens=decode_seq_lens,
                    )
792
793
            else:
                decode_metadata = self._build_decode(
794
795
                    block_table_tensor=block_table_tensor[:num_decodes, ...],
                    seq_lens=seq_lens[:num_decodes],
796
                )
797

798
        attn_metadata = self.metadata_cls(
799
800
            num_reqs=common_attn_metadata.num_reqs,
            max_query_len=common_attn_metadata.max_query_len,
801
            num_actual_tokens=num_tokens,
802
803
            query_start_loc=query_start_loc,
            slot_mapping=slot_mapping,
804
            head_dim=self.model_config.get_head_size(),
805
            # MLACommonMetadata Chunk prefill specific
806
807
808
            num_decodes=num_decodes,
            num_decode_tokens=num_decode_tokens,
            num_prefills=num_prefills,
809
810
            prefill=prefill_metadata,
            decode=decode_metadata,
811
812
        )

813
        if self._use_fi_prefill and num_prefills > 0:
814
815
816
817
818
            assert isinstance(attn_metadata.prefill, FlashInferPrefillMetadata)
            self._build_fi_prefill_wrappers(attn_metadata.prefill)

        return attn_metadata

819
820
    def can_run_in_cudagraph(
            self, common_attn_metadata: CommonAttentionMetadata) -> bool:
821
822
823
824
        # return common_attn_metadata.max_query_len == 1
        if not self.use_spec_decode:
            return common_attn_metadata.max_query_len == 1
        return self._num_prefills == 0
825

826

827
class MLACommonImpl(MLAAttentionImpl[M], Generic[M]):
828
829
830
831
832
833
834
835
836
837
838
    """
    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,
839
        alibi_slopes: Optional[list[float]],
840
841
842
843
        sliding_window: Optional[int],
        kv_cache_dtype: str,
        logits_soft_cap: Optional[float],
        attn_type: str,
844
        kv_sharing_target_layer_name: Optional[str],
845
846
847
848
849
850
851
852
853
        # 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,
    ) -> None:
854
855
856
        if kv_sharing_target_layer_name is not None:
            raise NotImplementedError("KV sharing is not supported for MLA")

857
858
859
860
861
862
863
864
865
866
867
868
869
        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
zhuwenwen's avatar
zhuwenwen committed
870
        self.use_llama_nn = os.environ.get('LLAMA_NN') == '1'
871

872
873
874
875
876
        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
877
878
879
880
881
882
        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
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
        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
            self._pad_v = self.vllm_flash_attn_version is None or not (
                self.vllm_flash_attn_version == 3
905
906
907
                and current_platform.get_device_capability()[0] == 9
                and torch.cuda.get_device_properties(torch.cuda.current_device()).multi_processor_count == 120)
        
908
909
910
911
912
913
914
915
916
917

    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
918
919
            # maybe_padded_v = torch.nn.functional.pad(
            #     v, [0, q.shape[-1] - v.shape[-1]], value=0)
920
            maybe_padded_v = torch.nn.functional.pad(
zhuwenwen's avatar
zhuwenwen committed
921
922
                    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])
923

924
925
926
927
928
929
930
        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

931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
        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

950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
    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
        return prefill.prefill_main.run(
            q=q,
            k=k,
            v=v,
            return_lse=return_softmax_lse,
        )

976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
    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

999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
    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)
        return prefill.prefill_chunks[chunk_idx].run(
            q=q,
            k=k,
            v=v,
            return_lse=True,
        )

1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
    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.
        )

1049
    def _v_up_proj(self, x):
1050
1051
1052
1053
1054
        # Convert from (B, N, L) to (N, B, L)
        x = x.view(-1, self.num_heads, self.kv_lora_rank).transpose(0, 1)
        # Multiply (N, B, L) x (N, L, V) -> (N, B, V)
        x = torch.bmm(x, self.W_UV)
        # Convert from (N, B, V) to (B, N * V)
1055
        return x.transpose(0, 1).reshape(-1, self.num_heads * self.v_head_dim)
1056

1057
    def process_weights_after_loading(self, act_dtype: torch.dtype):
1058
1059

        def get_layer_weight(layer):
1060
1061
1062
1063
1064
1065
1066
            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}.")
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079

        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
1080
            return layer.weight if not envs.VLLM_USE_NN else layer.weight.T
1081

1082
1083
1084
        # we currently do not have quantized bmm's which are needed for
        # `W_UV` and `W_UK_T`, we we just store fp16/bf16 copies and perform
        # the bmm's in 16-bit, the extra memory overhead of this is fairly low
zhuwenwen's avatar
zhuwenwen committed
1085
1086
1087
1088
        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
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
        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)

1106
1107
1108
1109
        # 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)
1110
1111
1112
1113
1114
1115
1116

    def _compute_prefill_context(
        self,
        q: torch.Tensor,
        kv_c_and_k_pe_cache: torch.Tensor,
        attn_metadata: MLACommonMetadata,
    ):
1117
1118
1119
        assert attn_metadata.prefill is not None
        prefill_metadata = attn_metadata.prefill
        assert prefill_metadata.chunked_context is not None
1120
1121

        output = None
1122
1123
        iters = len(prefill_metadata.chunked_context.seq_tot)
        workspace = prefill_metadata.chunked_context.workspace
1124
1125

        for i in range(iters):
1126
            toks = prefill_metadata.chunked_context.seq_tot[i]
1127
1128
1129
1130

            ops.gather_cache(
                src_cache=kv_c_and_k_pe_cache,
                dst=workspace,
1131
1132
                block_table=prefill_metadata.block_table,
                cu_seq_lens=prefill_metadata.chunked_context.cu_seq_lens[i],
1133
                batch_size=attn_metadata.num_prefills,
1134
                seq_starts=prefill_metadata.chunked_context.starts[i],
1135
1136
1137
            )

            kv_c_normed = workspace[:toks]\
1138
                [..., :self.kv_lora_rank]
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
            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)

            k = torch.cat((k_nope, k_pe.expand((*k_nope.shape[:-1], -1))),
                          dim=-1)

1150
1151
1152
            attn_output, attn_softmax_lse = self._run_prefill_context_chunk(
                prefill=prefill_metadata,
                chunk_idx=i,
1153
1154
                q=q,
                k=k,
1155
                v=v,
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
            )

            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

    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,
    ) -> torch.Tensor:
1185
1186
        assert attn_metadata.prefill is not None

zhuwenwen's avatar
zhuwenwen committed
1187
1188
1189
1190
        if envs.VLLM_HAS_CONTEXT_DEFAULT:
            has_context = attn_metadata.prefill.chunked_context is not None
        else:
            has_context = False
1191
1192
1193
1194
1195
1196
1197
        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)

1198
1199
        output = self._run_prefill_new_tokens(
            prefill=attn_metadata.prefill,
1200
1201
            q=q,
            k=k,
1202
            v=v,
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
            return_softmax_lse=has_context,
        )

        if has_context:
            suffix_output, suffix_lse = output
            context_output, context_lse = self._compute_prefill_context( \
                q, kv_c_and_k_pe_cache, attn_metadata)

            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,
            )

1220
1221
1222
1223
        # unpad if necessary
        if self._pad_v:
            output = output[..., :v.shape[-1]]

1224
        return output.flatten(start_dim=-2)
1225
1226
1227
1228

    @abstractmethod
    def _forward_decode(
        self,
1229
        ql_nope: torch.Tensor,
1230
1231
        q_pe: torch.Tensor,
        kv_c_and_k_pe_cache: torch.Tensor,
1232
        attn_metadata: M,
1233
1234
1235
1236
1237
1238
    ) -> torch.Tensor:
        raise NotImplementedError

    def forward(
        self,
        layer: AttentionLayer,
1239
        q: torch.Tensor,
1240
1241
1242
        k_c_normed: torch.Tensor,  # key in unified attn
        k_pe: torch.Tensor,  # value in unified attn
        kv_cache: torch.Tensor,
1243
        attn_metadata: M,
1244
        output: Optional[torch.Tensor] = None,
1245
        output_scale: Optional[torch.Tensor] = None,
1246
1247
1248
    ) -> torch.Tensor:
        assert output is not None, "Output tensor must be provided."

1249
1250
1251
1252
1253
        if output_scale is not None:
            raise NotImplementedError(
                "fused output quantization is not yet supported"
                " for MLACommonImpl")

1254
        if attn_metadata is None:
1255
1256
1257
1258
            # 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)
1259
1260
1261
1262
1263
1264

        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, ...]
1265
        q = q[:num_actual_toks, ...]
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
        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

1277
        decode_q = q[:num_decode_tokens]
1278

1279
        prefill_q = q[num_decode_tokens:]
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
        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,
            )

        if has_prefill:
            output[num_decode_tokens:] = self._forward_prefill(
                prefill_q, prefill_k_c_normed, prefill_k_pe, kv_cache,
                attn_metadata)

        if has_decode:
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
            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)
            # Multiply (N, B, P) x (N, P, L) -> (N, B, L)
            decode_ql_nope = torch.bmm(decode_q_nope, self.W_UK_T)
            # Convert from (N, B, L) to (B, N, L)
            decode_ql_nope = decode_ql_nope.transpose(0, 1)

1310
            output[:num_decode_tokens] = self._forward_decode(
1311
                decode_ql_nope, decode_q_pe, kv_cache, attn_metadata)
1312

zhuwenwen's avatar
zhuwenwen committed
1313
        return output_padded