common.py 46 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)
164
    Sc           = chunk_end - chunk_start_table
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
192
    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
from dataclasses import dataclass
193
from typing import TYPE_CHECKING, Any, Generic, Optional, TypeVar
194
195
196
from itertools import chain

import numpy as np
197
198

import torch
zhuwenwen's avatar
zhuwenwen committed
199
import os
200
201

from vllm import _custom_ops as ops
202
from vllm import envs
203
204
205
from vllm.attention.backends.abstract import (AttentionBackend, AttentionLayer,
                                              AttentionMetadata,
                                              MLAAttentionImpl)
206
from vllm.attention.backends.utils import get_mla_dims
207
from vllm.attention.ops.merge_attn_states import merge_attn_states
208
from vllm.attention.utils.fa_utils import get_flash_attn_version
209
from vllm.logger import init_logger
210
from vllm.config import get_current_vllm_config
211
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
217
from vllm.v1.attention.backends.utils import (AttentionMetadataBuilder,
                                              CommonAttentionMetadata)
218
from vllm.v1.kv_cache_interface import AttentionSpec
王敏's avatar
王敏 committed
219
from vllm.v1.worker.block_table import BlockTable
220

221
222
try:
    from vllm.vllm_flash_attn import flash_attn_varlen_func
223
    is_vllm_fa = True
224
225
except ImportError:
    # For rocm use upstream flash attention
226
227
    if current_platform.is_rocm():
        from flash_attn import flash_attn_varlen_func
228
    is_vllm_fa = False
229
230

if TYPE_CHECKING:
231
    from vllm.v1.core.sched.output import SchedulerOutput
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
    from vllm.v1.worker.gpu_input_batch import InputBatch
    from vllm.v1.worker.gpu_model_runner import GPUModelRunner

logger = init_logger(__name__)


class MLACommonBackend(AttentionBackend):

    accept_output_buffer: bool = True

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

    @staticmethod
247
    def get_metadata_cls() -> type["AttentionMetadata"]:
248
249
250
        return MLACommonMetadata

    @staticmethod
251
    def get_builder_cls() -> type["MLACommonMetadataBuilder"]:
252
253
254
255
256
257
258
259
        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,
260
    ) -> tuple[int, ...]:
261
262
        return (num_blocks, block_size, head_size)

263
264
    @classmethod
    def get_supported_head_sizes(cls) -> list[int]:
265
266
        return [576]

267
268
269
270
271
272
273
274
275
276
277
    @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.")

278
279

@dataclass
280
281
282
283
284
285
286
287
288
289
290
291
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]
        workspace: torch.Tensor
292

293
294
295
296
297
    block_table: torch.Tensor
    query_start_loc: torch.Tensor
    max_query_len: int
    chunked_context: Optional[ChunkedContextMetadata] = None

298

299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
@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
    """
315
316
317
318
319
320
321
322
323
324
325
326
    # NOTE(sang): Definition of context_len, query_len, and seq_len.
    # |---------- N-1 iteration --------|
    # |---------------- N iteration ---------------------|
    # |- tokenA -|......................|-- newTokens ---|
    # |---------- context_len ----------|
    # |-------------------- seq_len ---------------------|
    #                                   |-- query_len ---|

    num_actual_tokens: int  # Number of tokens excluding padding.
    query_start_loc: torch.Tensor
    slot_mapping: torch.Tensor

327
328
329
330
331
332
    # New for MLA (compared to FlashAttention)
    # For handling prefill decode split
    num_decodes: int
    num_decode_tokens: int
    num_prefills: int

333
334
335
    # The dimension of the attention heads
    head_dim: Optional[int] = None

336
337
    decode: Optional[D] = None
    prefill: Optional[MLACommonPrefillMetadata] = None
338
339

    def __post_init__(self):
340
341
        if self.head_dim is not None:
            MLACommonBackend.validate_head_size(self.head_dim)
342
343


344
M = TypeVar("M", bound=MLACommonMetadata)
345
346


347
class MLACommonMetadataBuilder(AttentionMetadataBuilder[M]):
348
349
350
351
352
    """
    NOTE: Please read the comment at the top of the file before trying to
    understand this class
    """

353
354
    def __init__(self,
                 runner: "GPUModelRunner",
355
356
                 kv_cache_spec: AttentionSpec,
                 block_table: BlockTable,
357
358
359
                 metadata_cls: Optional[type[M]] = None):
        self.metadata_cls = metadata_cls \
            if metadata_cls is not None else MLACommonMetadata
360
361
362
363
364
        self.runner = runner
        scheduler_config = runner.scheduler_config
        model_config = runner.model_config
        cache_config = runner.cache_config
        self.chunked_prefill_enabled = scheduler_config.chunked_prefill_enabled
365
366
367
        self.num_heads = model_config.get_num_attention_heads(
            runner.parallel_config)
        self.mla_dims = get_mla_dims(model_config)
368
        self.aot_schedule = current_platform.is_cuda()
369
        self.kv_cache_spec = kv_cache_spec
370
371
372

        # Dont try to access the runner on AMD
        if self.aot_schedule:
373
            self.page_size = self.kv_cache_spec.block_size
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398

        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(
                    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)
                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,
                 model_config.get_head_size()),
                dtype=model_config.dtype,
                device=runner.device,
            )
399
        self.block_table = block_table
400

401
402
        self.use_spec_decode = False
        self.num_scheduled_tokens_np = np.zeros(scheduler_config.max_num_seqs, dtype=np.int32)
403

王敏's avatar
王敏 committed
404
405
406
        # support for cudagraph spec docoding
        self.spec_decode_block_table_tensor = None
        self.spec_decode_seq_lens = None
407
408
409
410
411
412
        self.decode_token_num_threshold = 1
        vllm_config = get_current_vllm_config()
        speculative_config = vllm_config.speculative_config
        if speculative_config and speculative_config.num_speculative_tokens > 1:
            self.use_spec_decode = True
            self.decode_token_num_threshold = 1 + speculative_config.num_speculative_tokens
王敏's avatar
王敏 committed
413

414

415
    def reorder_batch(self, input_batch: "InputBatch",
416
                      scheduler_output: "SchedulerOutput") -> bool:
417
418
419
420
421
422
423
424
425
426
        # We now want to reorder the batch so that the "decode" requests are and
        # the front and the "prefill" requests are at the using the least amount
        # swaps possible. (NOTE for now we loosely use "decode" to mean requests
        # where attention is likely memory-bound and "prefill" to mean requests
        # where attention is likely compute-bound, TODO(lucas): figure out a
        # better naming here)
        decodes = []
        prefills = []
        num_decode_tokens = 0
        num_prefill_tokens = 0
427
428
        use_spec_decode = len(
            scheduler_output.scheduled_spec_decode_tokens) > 0
429
430
431
432
433
434
435

        for i, req_id in enumerate(input_batch.req_ids):
            num_tokens = scheduler_output.num_scheduled_tokens[req_id]
            # for now treat 1 scheduled token as "decode" even if its not,
            # we should update this to something like < 8 in the future but
            # currently the TritonMLA._forward_decode only supports
            # num_tokens = 1
436
437
438
439
440
441
442
443
444
445

            # if num_tokens == 2 or num_tokens == 1:
            #     decodes.append(i)
            #     num_decode_tokens += num_tokens
            # else:
            #     prefills.append(i)
            #     num_prefill_tokens += num_tokens
            req_idx = input_batch.req_id_to_index[req_id]
            num_computed_tokens = input_batch.num_computed_tokens_cpu[req_idx]
            num_prompt_tokens = input_batch.num_prompt_tokens[req_idx]
446
            self.num_scheduled_tokens_np[i] = num_tokens
447
            if num_computed_tokens < num_prompt_tokens or (num_tokens > self.decode_token_num_threshold):
448
449
                prefills.append(i)
                num_prefill_tokens += num_tokens
450
451
452
            else:
                decodes.append(i)
                num_decode_tokens += num_tokens
453
454
455
456
457
458
459
460
461
462
463
464
465

        # We hope that this is fairly minimal since decodes
        # should be around for a number of iterations so hopefully they are
        # relatively stationary (and new request are generally appended to the
        # persistent batch so already should be at the back)
        # To achieve this we loop over the decodes in descending order and
        # the prefills in ascending order. We swap decodes from the  "back"
        # i.e. past where the last decode should be in the reodorered with
        # prefills from the front of the batch.
        # `decodes` and `prefills` are already in ascending order just based on
        # the above loop
        num_decodes = len(decodes)
        num_prefills = len(prefills)
466
        modified_batch = False
467
468
469
470

        for i in range(1, min(num_decodes, num_prefills) + 1):
            # If the decode is at the "back" of the batch, i, we can swap it
            # with the prefill closest to the front of the batch
471
472
            decode_idx = decodes[num_decodes - i]
            if decode_idx < num_decodes:
473
474
                break

475
476
477
            input_batch.swap_states(prefills[i - 1], decode_idx)
            modified_batch = True

478
            # num_scheduled_tokens also need to be swapped
479
480
481
482
            tmp = self.num_scheduled_tokens_np[decode_idx]
            self.num_scheduled_tokens_np[decode_idx] = self.num_scheduled_tokens_np[prefills[i - 1]]
            self.num_scheduled_tokens_np[prefills[i - 1]] = tmp

483

484
485
486
487
488
489
490
        # Save for next `build` call
        # TODO(lucas): this is a bit of a hack, we should probably have a
        # better way of doing this
        self._num_decodes = num_decodes
        self._num_prefills = num_prefills
        self._num_decode_tokens = num_decode_tokens
        self._num_prefill_tokens = num_prefill_tokens
491
        
492
        self.use_spec_decode = use_spec_decode
493

494
495
        return modified_batch

496
497
    def _build_decode(self, block_table_tensor: torch.Tensor,
                      seq_lens: torch.Tensor):
498
        return MLACommonDecodeMetadata(
499
            block_table=block_table_tensor,
500
501
502
            seq_lens=seq_lens,
        )

503
504
505
506
507
508
509
    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
510
511
512
        # 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."
513

王敏's avatar
王敏 committed
514
        #m.max_query_len = 1  # decode-only
515
516
517
518
519
520

        # Update state usually set in reorder_batch.
        self._num_decodes = m.num_reqs
        self._num_decode_tokens = m.num_actual_tokens
        self._num_prefills = 0
        self._num_prefill_tokens = 0
王敏's avatar
王敏 committed
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
        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)
540
541
542
        return self.build(0, m)

    def build(self, common_prefix_len: int,
543
              common_attn_metadata: CommonAttentionMetadata) -> M:
544
545
546
547
        num_reqs = common_attn_metadata.num_reqs
        num_actual_tokens = common_attn_metadata.num_actual_tokens
        max_query_len = common_attn_metadata.max_query_len

548
549
        assert self._num_decodes + self._num_prefills == num_reqs

Simon Mo's avatar
Simon Mo committed
550
551
552
        # 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.
553
        device = self.runner.device
554
555
        block_table = self.block_table
        block_table_tensor = block_table.get_device_tensor()[:num_reqs]
556
557
558
559
560
561
562
563

        slot_mapping = common_attn_metadata.slot_mapping
        if slot_mapping is None:
            block_table.slot_mapping[:num_actual_tokens].copy_(
                block_table.slot_mapping_cpu[:num_actual_tokens],
                non_blocking=True)
            block_table.slot_mapping[num_actual_tokens:].fill_(-1)
            slot_mapping = block_table.slot_mapping[:num_actual_tokens]
564

565
566
        query_start_loc = common_attn_metadata.query_start_loc
        seq_lens = common_attn_metadata.seq_lens
Simon Mo's avatar
Simon Mo committed
567

568
569
570
571
572
573
        prefill_metadata = None
        if self._num_prefills > 0:
            reqs_start = self._num_decodes  # prefill_start

            context_lens_cpu = self.runner.input_batch.\
                num_computed_tokens_cpu_tensor[reqs_start:num_reqs]
Simon Mo's avatar
Simon Mo committed
574
575
            max_context_len_cpu = context_lens_cpu.max().item()
            num_prefills_with_context_cpu = (context_lens_cpu > 0).sum().item()
576
577
            prefill_query_start_loc = query_start_loc[
                reqs_start:] - query_start_loc[reqs_start]
578
579
580

            chunked_context_metadata = None
            if self.chunked_prefill_enabled and self._num_prefills > 0 \
Simon Mo's avatar
Simon Mo committed
581
                and max_context_len_cpu > 0:
582
583
584
585
586
587
588
589
                # 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
590
591
                max_context_chunk = (self.chunked_prefill_workspace_size //
                                     num_prefills_with_context_cpu)
592

593
594
595
596
597
598
                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)
599
600

                assert max_context_chunk > 0
Simon Mo's avatar
Simon Mo committed
601
                num_chunks = cdiv(max_context_len_cpu, max_context_chunk)
602
603
604
605
606

                # 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
607
608
                # Note(simon): this is done in CPU because of downstream's
                # of `to_list`.
609
                chunk_starts = \
Simon Mo's avatar
Simon Mo committed
610
                    torch.arange(num_chunks, dtype=torch.int32) \
611
612
                    .unsqueeze(1).expand(-1, self._num_prefills) \
                    * max_context_chunk
Simon Mo's avatar
Simon Mo committed
613
                chunk_ends = torch.min(context_lens_cpu.unsqueeze(0),
614
615
                                       chunk_starts + max_context_chunk)
                chunk_seq_lens = (chunk_ends - chunk_starts).clamp(min=0)
Simon Mo's avatar
Simon Mo committed
616
617
618
619
620
621
622
623
624

                cu_seq_lens_cpu = torch.zeros(num_chunks,
                                              self._num_prefills + 1,
                                              dtype=torch.int32,
                                              pin_memory=True)
                torch.cumsum(chunk_seq_lens,
                             dim=1,
                             out=cu_seq_lens_cpu[:, 1:],
                             dtype=torch.int32)
625
626
627

                chunked_context_metadata = \
                    MLACommonPrefillMetadata.ChunkedContextMetadata(
Simon Mo's avatar
Simon Mo committed
628
629
                    cu_seq_lens=cu_seq_lens_cpu.to(device, non_blocking=True),
                    starts=chunk_starts.to(device, non_blocking=True),
630
631
632
633
634
635
636
637
638
                    seq_tot=chunk_seq_lens.sum(dim=1).tolist(),
                    max_seq_lens=chunk_seq_lens.max(dim=1).values.tolist(),
                    workspace=self.chunked_prefill_workspace,
                )

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

            prefill_metadata = MLACommonPrefillMetadata(
639
                block_table=block_table_tensor[reqs_start:, ...],
640
                query_start_loc=prefill_query_start_loc,
Simon Mo's avatar
Simon Mo committed
641
                max_query_len=max_query_len,
642
643
644
645
646
                chunked_context=chunked_context_metadata,
            )

        decode_metadata = None
        if self._num_decodes > 0:
647
            if self.use_spec_decode and not common_attn_metadata.spec_layer_decoding:
648
649
650
651
652
653
654
655
                query_lens = self.num_scheduled_tokens_np[:self._num_decodes]
                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(
王敏's avatar
王敏 committed
656
                    block_table_tensor.device, non_blocking=True).contiguous()
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
                
                if envs.VLLM_ZERO_OVERHEAD:
                    decode_block_table_tensor = torch.empty((self._num_decode_tokens, block_table_tensor.shape[1]), 
                                                            device=block_table_tensor.device)
                    arange_np = np.arange(self._num_decodes)
                    indices_np = np.repeat(arange_np, query_lens)
                    
                    indices = torch.from_numpy(indices_np).pin_memory().to(
                        block_table_tensor.device, non_blocking=True)
                    decode_block_table_tensor = block_table_tensor[indices].contiguous()
                    decode_seq_lens = seq_lens[indices].contiguous()
                else:
                    decode_block_table_tensor = torch.repeat_interleave(
                        block_table_tensor[:self._num_decodes, ...],
                        repeats, dim=0).contiguous()
                    decode_seq_lens = torch.repeat_interleave(seq_lens[:self._num_decodes], repeats, dim=0).contiguous()
673
                seq_lens_minus = torch.from_numpy(rarange).to(torch.int32).pin_memory().to(
王敏's avatar
王敏 committed
674
                    seq_lens.device, non_blocking=True).contiguous()
675
                decode_seq_lens = decode_seq_lens - seq_lens_minus
676

王敏's avatar
王敏 committed
677
678
679
680
681
682
683
684
685
686
687
688
689
                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,
                    )
690
            else:
691
692
693
694
695
696
697
698
699
700
701
702
703
704
                self._num_decode_tokens = self._num_decodes
                if self.use_spec_decode and self.spec_decode_block_table_tensor is not None:
                    self.spec_decode_block_table_tensor[:self._num_decode_tokens].copy_(block_table_tensor[:self._num_decode_tokens, ...])
                    self.spec_decode_seq_lens[:self._num_decode_tokens].copy_(seq_lens[:self._num_decode_tokens])

                    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=block_table_tensor[:self._num_decode_tokens, ...],
                        seq_lens=seq_lens[:self._num_decode_tokens],
                    )
705
706

        return self.metadata_cls(
707
708
709
710
711
712
713
714
            num_actual_tokens=num_actual_tokens,
            query_start_loc=query_start_loc,
            slot_mapping=slot_mapping,
            head_dim=self.runner.model_config.get_head_size(),
            # MLACommonMetadata Chunk prefill specific
            num_decodes=self._num_decodes,
            num_decode_tokens=self._num_decode_tokens,
            num_prefills=self._num_prefills,
715
716
            prefill=prefill_metadata,
            decode=decode_metadata,
717
718
        )

719
720
    def can_run_in_cudagraph(
            self, common_attn_metadata: CommonAttentionMetadata) -> bool:
721
722
        if not self.use_spec_decode:
            return common_attn_metadata.max_query_len == 1
王敏's avatar
王敏 committed
723
        return self._num_prefills == 0
724

725

726
class MLACommonImpl(MLAAttentionImpl[M], Generic[M]):
727
728
729
730
731
732
733
734
735
736
737
    """
    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,
738
        alibi_slopes: Optional[list[float]],
739
740
        sliding_window: Optional[int],
        kv_cache_dtype: str,
741
        blocksparse_params: Optional[dict[str, Any]],
742
743
        logits_soft_cap: Optional[float],
        attn_type: str,
744
        kv_sharing_target_layer_name: Optional[str],
745
746
747
748
749
750
751
752
753
        # 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:
754
755
756
        if kv_sharing_target_layer_name is not None:
            raise NotImplementedError("KV sharing is not supported for MLA")

757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
        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

        # 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
775
        self.vllm_flash_attn_version = get_flash_attn_version()
776
777
778
779
        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)
zhuwenwen's avatar
zhuwenwen committed
780
781
        
        self.use_llama_nn = os.environ.get('LLAMA_NN') == '1'
782

783
784
785
786
        # 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
787
788
789
790
791
792
793
        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
            
794
795
796
797
798
799
800
801
802
803

    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
804
805
            # maybe_padded_v = torch.nn.functional.pad(
            #     v, [0, q.shape[-1] - v.shape[-1]], value=0)
806
            maybe_padded_v = torch.nn.functional.pad(
zhuwenwen's avatar
zhuwenwen committed
807
808
                    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])
809

810
811
812
813
814
815
816
        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

817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
        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

836
    def _v_up_proj(self, x):
837
838
839
840
841
        # 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)
842
        return x.transpose(0, 1).reshape(-1, self.num_heads * self.v_head_dim)
843

844
    def process_weights_after_loading(self, act_dtype: torch.dtype):
845
846

        def get_layer_weight(layer):
847
848
849
850
851
852
853
            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}.")
854
855
856
857
858
859
860
861
862
863
864
865
866

        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
867
            return layer.weight if not envs.VLLM_USE_NN else layer.weight.T
868

869
870
871
        # 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
872
873
874
875
        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
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
        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)

893
894
895
896
        # 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)
897
898
899
900
901
902

    def _compute_prefill_context(
        self,
        q: torch.Tensor,
        kv_c_and_k_pe_cache: torch.Tensor,
        attn_metadata: MLACommonMetadata,
yangql's avatar
yangql committed
903
        kv_scale=torch.tensor(1.0, dtype=torch.float32),     
904
    ):
905
906
907
        assert attn_metadata.prefill is not None
        prefill_metadata = attn_metadata.prefill
        assert prefill_metadata.chunked_context is not None
908
909

        output = None
910
911
        iters = len(prefill_metadata.chunked_context.seq_tot)
        workspace = prefill_metadata.chunked_context.workspace
912
913

        for i in range(iters):
914
            toks = prefill_metadata.chunked_context.seq_tot[i]
915
916
917
918

            ops.gather_cache(
                src_cache=kv_c_and_k_pe_cache,
                dst=workspace,
919
920
                block_table=prefill_metadata.block_table,
                cu_seq_lens=prefill_metadata.chunked_context.cu_seq_lens[i],
921
                batch_size=attn_metadata.num_prefills,
922
                seq_starts=prefill_metadata.chunked_context.starts[i],
yangql's avatar
yangql committed
923
924
                kv_dtype=self.kv_cache_dtype,
                scale=kv_scale,
925
926
927
            )

            kv_c_normed = workspace[:toks]\
928
                [..., :self.kv_lora_rank]
929
930
931
932
933
934
935
936
            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)

937
            if envs.VLLM_USE_OPT_CAT:
938
                if k_nope.shape[0] > 1024:
zhuwenwen's avatar
zhuwenwen committed
939
                    from vllm.v1.attention.backends.mla.test_concat import lightop_concat_prefill_helper
940
941
942
943
944
                    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)
945
946
947
            else:
                k = torch.cat((k_nope, k_pe.expand((*k_nope.shape[:-1], -1))),
                            dim=-1)
948

949
950
            attn_output, attn_softmax_lse = \
                self._flash_attn_varlen_diff_headdims(
951
952
                q=q,
                k=k,
953
                v=v,
954
955
956
957
                cu_seqlens_q=prefill_metadata.query_start_loc,
                cu_seqlens_k=prefill_metadata.chunked_context.cu_seq_lens[i],
                max_seqlen_q=prefill_metadata.max_query_len,
                max_seqlen_k=prefill_metadata.chunked_context.max_seq_lens[i],
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
                softmax_scale=self.scale,
                causal=False,  # Context is unmasked
                return_softmax_lse=True,
            )

            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,
yangql's avatar
yangql committed
989
        kv_scale=torch.tensor(1.0, dtype=torch.float32),
990
    ) -> torch.Tensor:
991
992
        assert attn_metadata.prefill is not None

zhuwenwen's avatar
zhuwenwen committed
993
994
995
996
        if envs.VLLM_HAS_CONTEXT_DEFAULT:
            has_context = attn_metadata.prefill.chunked_context is not None
        else:
            has_context = False
gaoqiong's avatar
gaoqiong committed
997

998
999
1000
1001
1002
        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)

1003
        if envs.VLLM_USE_OPT_CAT:
1004
            if k_nope.shape[0] > 1024:
zhuwenwen's avatar
zhuwenwen committed
1005
                from vllm.v1.attention.backends.mla.test_concat import lightop_concat_prefill_helper
1006
1007
1008
1009
1010
                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)
1011
1012
        else:
            k = torch.cat((k_nope, k_pe.expand((*k_nope.shape[:-1], -1))), dim=-1)
1013

1014
        output = self._flash_attn_varlen_diff_headdims(
1015
1016
            q=q,
            k=k,
1017
            v=v,
1018
1019
1020
1021
            cu_seqlens_q=attn_metadata.prefill.query_start_loc,
            cu_seqlens_k=attn_metadata.prefill.query_start_loc,
            max_seqlen_q=attn_metadata.prefill.max_query_len,
            max_seqlen_k=attn_metadata.prefill.max_query_len,
1022
1023
1024
1025
1026
1027
1028
1029
            softmax_scale=self.scale,
            causal=True,
            return_softmax_lse=has_context,
        )

        if has_context:
            suffix_output, suffix_lse = output
            context_output, context_lse = self._compute_prefill_context( \
yangql's avatar
yangql committed
1030
                q, kv_c_and_k_pe_cache, attn_metadata, kv_scale)
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040

            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,
            )

1041
1042
1043
1044
        # unpad if necessary
        if self._pad_v:
            output = output[..., :v.shape[-1]]

1045
        return output.flatten(start_dim=-2)
1046
1047
1048
1049

    @abstractmethod
    def _forward_decode(
        self,
1050
        ql_nope: torch.Tensor,
1051
1052
        q_pe: torch.Tensor,
        kv_c_and_k_pe_cache: torch.Tensor,
1053
        attn_metadata: M,
1054
1055
1056
1057
1058
1059
    ) -> torch.Tensor:
        raise NotImplementedError

    def forward(
        self,
        layer: AttentionLayer,
1060
        q: torch.Tensor,
1061
1062
1063
        k_c_normed: torch.Tensor,  # key in unified attn
        k_pe: torch.Tensor,  # value in unified attn
        kv_cache: torch.Tensor,
1064
        attn_metadata: M,
1065
        output: Optional[torch.Tensor] = None,
1066
        output_scale: Optional[torch.Tensor] = None,
1067
1068
1069
1070
    ) -> torch.Tensor:

        assert output is not None, "Output tensor must be provided."

1071
1072
1073
1074
1075
        if output_scale is not None:
            raise NotImplementedError(
                "fused output quantization is not yet supported"
                " for MLACommonImpl")

1076
        if attn_metadata is None:
1077
1078
1079
1080
            # 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)
1081
1082
1083
1084
1085
1086

        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, ...]
1087
        q = q[:num_actual_toks, ...]
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
        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

1099
        decode_q = q[:num_decode_tokens]
1100

1101
        prefill_q = q[num_decode_tokens:]
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
        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,
yangql's avatar
yangql committed
1119
                attn_metadata, kv_scale=layer._k_scale)
1120
1121

        if has_decode:
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
            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)

1132
            output[:num_decode_tokens] = self._forward_decode(
1133
                decode_ql_nope, decode_q_pe, kv_cache, attn_metadata, layer._k_scale, self.kv_cache_dtype)
1134

zhuwenwen's avatar
zhuwenwen committed
1135
        return output_padded