attention.py 327 KB
Newer Older
1
# Copyright (c) 2022-2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
2
3
4
5
#
# See LICENSE for license information.

"""Attention."""
6
import collections
7
from contextlib import nullcontext
8
from importlib.metadata import version as get_pkg_version
9
from importlib.metadata import PackageNotFoundError
10
import math
11
import os
12
from typing import Any, Callable, Dict, List, Optional, Tuple, Union
13
import warnings
14
import logging
15

cyanguwa's avatar
cyanguwa committed
16
import numpy as np
17
from packaging.version import Version as PkgVersion
18
19

import torch
yuguo's avatar
yuguo committed
20
from torch.utils.cpp_extension import IS_HIP_EXTENSION
21

22
import transformer_engine_torch as tex
23
24
25
26
27
from transformer_engine.pytorch.utils import (
    get_cudnn_version,
    nvtx_range_pop,
    nvtx_range_push,
)
28
from transformer_engine.pytorch.cpp_extensions.fused_attn import (
29
30
    fused_attn_fwd,
    fused_attn_bwd,
31
    FusedAttnBackend,
32
33
34
35
36
37
38
    META_QKV,
    META_O,
)
from transformer_engine.pytorch.fp8 import (
    FP8GlobalStateManager,
    get_fp8_te_dtype,
    get_fp8_torch_dtype,
39
)
40
from transformer_engine.pytorch.float8_tensor import Float8Tensor
41
from transformer_engine.pytorch.tensor._internal.float8_tensor_base import Float8TensorBase
42
from transformer_engine.pytorch.module import LayerNormLinear, Linear
43
from transformer_engine.pytorch.module.base import TransformerEngineBaseModule
44
45
46
47
48
from transformer_engine.pytorch.utils import (
    divide,
    attention_mask_func,
    split_tensor_along_dim,
    get_device_compute_capability,
49
    get_default_init_method,
50
51
52
53
)
from transformer_engine.pytorch.constants import (
    AttnMaskTypes,
    AttnTypes,
54
    AttnBiasTypes,
55
    QKVLayouts,
56
    dist_group_type,
57
    TE_DType,
58
59
60
61
)
from transformer_engine.pytorch.softmax import FusedScaleMaskSoftmax
from transformer_engine.pytorch.distributed import (
    get_distributed_world_size,
62
    get_distributed_rank,
63
    checkpoint,
64
65
66
    set_all_rng_states,
    CudaRNGStatesTracker,
    graph_safe_rng_available,
67
68
    gather_along_first_dim,
    reduce_scatter_along_first_dim,
69
)
70
from transformer_engine.pytorch.jit import jit_fuser, no_torch_dynamo
71
from transformer_engine.pytorch.graph import is_graph_capturing
72
from transformer_engine.pytorch.dot_product_attention.inference import InferenceParams
73
74
75
76
77
from transformer_engine.pytorch.tensor.quantized_tensor import (
    QuantizedTensor,
    prepare_for_saving,
    restore_from_saved,
)
78

79
80
81
82
83
# Import attention utils
import transformer_engine.pytorch.dot_product_attention.utils as dpa_utils
from transformer_engine.pytorch.dot_product_attention.utils import FlashAttentionUtils as fa_utils
from transformer_engine.pytorch.dot_product_attention.utils import AttentionLogging as attn_log
from transformer_engine.pytorch.dot_product_attention.rope import apply_rotary_pos_emb
84
85


86
87
88
# Setup Attention Logging
attn_log.setup_logging()

89
# Global vars for flash attn v2 and v3 imports
90
flash_attn_cuda_bwd = None
91
92
flash_attn_func = None
flash_attn_varlen_func = None
93
94
95
96
_flash_attn_fwd = None
_flash_attn_bwd = None
_flash_attn_varlen_fwd = None
_flash_attn_varlen_bwd = None
97
try:
98
    fa_utils.version = PkgVersion(get_pkg_version("flash-attn"))
99
except PackageNotFoundError:
100
    pass  # only print warning if use_flash_attention_2 = True in get_attention_backend
101
else:
102
    if torch.cuda.is_available() and get_device_compute_capability() >= (10, 0):
103
104
105
106
        if fa_utils.version_required_blackwell <= fa_utils.version <= fa_utils.max_version:
            fa_utils.is_installed = True
    elif fa_utils.version_required <= fa_utils.version <= fa_utils.max_version:
        fa_utils.is_installed = True
107

108
    if fa_utils.is_installed:
109
        from flash_attn_2_cuda import varlen_bwd as flash_attn_cuda_bwd
110
        from flash_attn.flash_attn_interface import flash_attn_func, flash_attn_varlen_func
111
112
        from flash_attn.flash_attn_interface import _flash_attn_forward as _flash_attn_fwd
        from flash_attn.flash_attn_interface import _flash_attn_backward as _flash_attn_bwd
113
        from flash_attn.flash_attn_interface import (
114
            _flash_attn_varlen_forward as _flash_attn_varlen_fwd,
115
116
        )
        from flash_attn.flash_attn_interface import (
117
            _flash_attn_varlen_backward as _flash_attn_varlen_bwd,
118
119
        )

120
121
        # Setup Flash attention utils
        fa_utils.set_flash_attention_version()
122
    elif (
123
        torch.cuda.is_available()
yuguo's avatar
yuguo committed
124
        and (IS_HIP_EXTENSION or get_device_compute_capability() >= (8, 0))
125
        and dpa_utils._NVTE_FLASH_ATTN
126
    ):
127
        attn_log.fa_logger.warning(
128
            "Supported flash-attn versions are %s. Found flash-attn %s.",
129
            dpa_utils._get_supported_versions(
130
                (
131
                    fa_utils.version_required
132
                    if get_device_compute_capability() < (10, 0)
133
                    else fa_utils.version_required_blackwell
134
                ),
135
                fa_utils.max_version,
136
            ),
137
            fa_utils.version,
138
        )
yuguo's avatar
yuguo committed
139
140
if not IS_HIP_EXTENSION:
    try:
141
        fa_utils.fa3_version = PkgVersion(get_pkg_version("flash-attn-3"))
yuguo's avatar
yuguo committed
142
    except PackageNotFoundError:
143
        pass  # only print warning if use_flash_attention_3 = True in get_attention_backend
yuguo's avatar
yuguo committed
144
    else:
145
146
        from flash_attn_3.flash_attn_interface import flash_attn_func as flash_attn_func_v3
        from flash_attn_3.flash_attn_interface import (
yuguo's avatar
yuguo committed
147
            flash_attn_varlen_func as flash_attn_varlen_func_v3,
148
        )
149
150
        from flash_attn_3.flash_attn_interface import (
            flash_attn_with_kvcache as flash_attn_with_kvcache_v3,
yuguo's avatar
yuguo committed
151
        )
152
153
        from flash_attn_3.flash_attn_interface import _flash_attn_forward as _flash_attn_fwd_v3
        from flash_attn_3.flash_attn_interface import _flash_attn_backward as _flash_attn_bwd_v3
yuguo's avatar
yuguo committed
154
155
    
        fa_utils.set_flash_attention_3_params()
156

157
# Global vars for available attention backends and ALiBi cache
158
159
160
_attention_backends = {
    "attention_params": None,
    "use_flash_attention": None,
161
    "flash_attention_backend": None,
162
163
164
165
    "use_fused_attention": None,
    "fused_attention_backend": None,
    "use_unfused_attention": None,
    "backend_selection_requires_update": False,
166
}
167

168
169
170
171
172
173
174
175
176
177
178
_alibi_cache = {
    "_num_heads": None,
    "_alibi_slopes": None,
    "_max_seqlen_q": None,
    "_max_seqlen_kv": None,
    "_bottom_right_alignment": True,
    "_alibi_bias": None,
    "_alibi_slopes_require_update": False,
    "_alibi_bias_require_update": False,
}

179
__all__ = ["DotProductAttention", "MultiheadAttention"]
180
181


182
183
184
185
186
def maybe_contiguous(tensor: torch.Tensor) -> torch.Tensor:
    """Make tensor contiguous if final stride is not 1."""
    return tensor.contiguous() if tensor.stride(-1) != 1 else tensor


187
188
189
def flash_attn_p2p_communicate(
    rank, send_tensor, send_dst, recv_tensor, recv_src, cp_group, batch_p2p_comm
):
190
    """Point-to-point communications of KV and dKV in Attention with context parallelism"""
191
192
193
194
    send_recv_ops = []

    if batch_p2p_comm:
        if rank % 2 == 0:
195
196
197
198
199
200
            send_op = torch.distributed.P2POp(
                torch.distributed.isend, send_tensor, send_dst, cp_group
            )
            recv_op = torch.distributed.P2POp(
                torch.distributed.irecv, recv_tensor, recv_src, cp_group
            )
201
202
203
            send_recv_ops.append(send_op)
            send_recv_ops.append(recv_op)
        else:
204
205
206
207
208
209
            recv_op = torch.distributed.P2POp(
                torch.distributed.irecv, recv_tensor, recv_src, cp_group
            )
            send_op = torch.distributed.P2POp(
                torch.distributed.isend, send_tensor, send_dst, cp_group
            )
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
            send_recv_ops.append(recv_op)
            send_recv_ops.append(send_op)
        send_recv_reqs = torch.distributed.batch_isend_irecv(send_recv_ops)
    else:
        if rank % 2 == 0:
            send_op = torch.distributed.isend(send_tensor, send_dst, cp_group)
            recv_op = torch.distributed.irecv(recv_tensor, recv_src, cp_group)
            send_recv_ops.append(send_op)
            send_recv_ops.append(recv_op)
        else:
            recv_op = torch.distributed.irecv(recv_tensor, recv_src, cp_group)
            send_op = torch.distributed.isend(send_tensor, send_dst, cp_group)
            send_recv_ops.append(recv_op)
            send_recv_ops.append(send_op)
        send_recv_reqs = send_recv_ops

    return send_recv_reqs


229
230
231
232
233
234
235
236
237
238
239
240
241
242
@jit_fuser
def flash_attn_fwd_out_correction_init(
    out_init_step: torch.Tensor,
    softmax_lse: torch.Tensor,
    softmax_lse_init_step: torch.Tensor,
    seq_dim: int,
):
    """Merge partial outputs of the first step in Attention with context parallelism"""
    softmax_lse_corrected_exp = torch.exp(softmax_lse_init_step - softmax_lse).movedim(2, seq_dim)
    softmax_lse_corrected_exp = softmax_lse_corrected_exp.unsqueeze(-1)
    out_corrected = out_init_step * softmax_lse_corrected_exp
    return out_corrected.to(out_init_step.dtype)


243
@jit_fuser
244
245
246
247
248
def flash_attn_fwd_out_correction(
    out: torch.Tensor,
    out_per_step: torch.Tensor,
    softmax_lse: torch.Tensor,
    softmax_lse_per_step: torch.Tensor,
249
    seq_dim: int,
250
):
251
    """Merge partial outputs of each step in Attention with context parallelism"""
252
    softmax_lse_corrected_exp = torch.exp(softmax_lse_per_step - softmax_lse).movedim(2, seq_dim)
253
    softmax_lse_corrected_exp = softmax_lse_corrected_exp.unsqueeze(-1)
254
    out_corrected = out_per_step * softmax_lse_corrected_exp
255
256
257
    out.add_(out_corrected)


258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
@jit_fuser
def flash_attn_fwd_second_half_out_correction(
    out: torch.Tensor,
    out_per_step: torch.Tensor,
    softmax_lse: torch.Tensor,
    softmax_lse_per_step: torch.Tensor,
    seq_dim: int,
):
    """Merge second half of partial outputs of each step in Attention with context parallelism"""
    out_ = out.select(seq_dim, 1)
    softmax_lse_ = softmax_lse.view(*softmax_lse.shape[:-1], 2, -1)[..., 1, :]
    softmax_lse_corrected_exp = torch.exp(softmax_lse_per_step - softmax_lse_).movedim(2, seq_dim)
    softmax_lse_corrected_exp = softmax_lse_corrected_exp.unsqueeze(-1)
    out_corrected = out_per_step * softmax_lse_corrected_exp
    out_.add_(out_corrected)


275
@jit_fuser
276
277
278
279
def flash_attn_fwd_softmax_lse_correction(
    softmax_lse: torch.Tensor,
    softmax_lse_per_step: torch.Tensor,
):
280
    """Merge softmax stats of each step in Attention with context parallelism"""
281
282
    max_scale = torch.max(softmax_lse, softmax_lse_per_step)
    min_scale = torch.min(softmax_lse, softmax_lse_per_step)
283
    new_scale = max_scale + torch.log1p(torch.exp(min_scale - max_scale))
284
    softmax_lse.copy_(new_scale)
285
286


287
288
289
290
291
292
293
294
295
296
297
298
299
@jit_fuser
def flash_attn_fwd_second_half_softmax_lse_correction(
    softmax_lse: torch.Tensor,
    softmax_lse_per_step: torch.Tensor,
):
    """Merge second half of softmax stats of each step in Attention with context parallelism"""
    softmax_lse_ = softmax_lse[..., 1, :]
    max_scale = torch.max(softmax_lse_, softmax_lse_per_step)
    min_scale = torch.min(softmax_lse_, softmax_lse_per_step)
    new_scale = max_scale + torch.log1p(torch.exp(min_scale - max_scale))
    softmax_lse_.copy_(new_scale)


300
301
@jit_fuser
def get_cu_seqlens_on_cp_rank(
302
303
304
305
306
307
    cu_seqlens: torch.Tensor,
    cu_seqlens_padded_on_cp_rank: torch.Tensor,
    cp_size: int,
    cp_rank: int,
    first_half: bool,
    second_half: bool,
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
):
    """Compute cu_seqlens of a context parallelism rank"""
    seqlens = cu_seqlens[1:] - cu_seqlens[:-1]
    seqlens_padded = (cu_seqlens_padded_on_cp_rank[1:] - cu_seqlens_padded_on_cp_rank[:-1]) // 2
    zeros = torch.zeros_like(seqlens)
    cu_seqlens_on_cp_rank = torch.zeros_like(cu_seqlens)
    if first_half:
        seqlens_1 = seqlens - cp_rank * seqlens_padded
        seqlens_1 = seqlens_1.clamp(zeros, seqlens_padded)
        cu_seqlens_on_cp_rank[1:].add_(seqlens_1)
    if second_half:
        seqlens_2 = seqlens - (2 * cp_size - cp_rank - 1) * seqlens_padded
        seqlens_2 = seqlens_2.clamp(zeros, seqlens_padded)
        cu_seqlens_on_cp_rank[1:].add_(seqlens_2)
    cu_seqlens_on_cp_rank.cumsum_(dim=0)
    return cu_seqlens_on_cp_rank


326
@jit_fuser
327
def get_seq_chunk_ids_for_reordering_before_attn(cp_size, device):
328
329
    """
    Context parallelism assigns two discontiguous sequence chunks to each GPU for load balancing.
330
331
332
    To make sure tokens are ordered correctly for compute, we need to reorder sequence chunks to
    be contigupus before attention compute. This function is to compute sequence chunk ids for
    reordering.
333
334
    """
    chunk_ids = torch.empty(2 * cp_size, dtype=torch.int32, device=device)
335
336
337
    for rank in range(cp_size):
        chunk_ids[rank] = 2 * rank
        chunk_ids[rank + cp_size] = 2 * cp_size - 2 * rank - 1
338
339
340
    return chunk_ids


341
@jit_fuser
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
def get_seq_chunk_ids_for_reordering_after_attn(cp_size, device):
    """
    Context parallelism assigns two discontiguous sequence chunks to each GPU for load balancing.
    We need to reorder sequence chunks back to discontiguous after attention compute. This function
    is to compute sequence chunk ids for reordering.
    """
    chunk_ids = torch.empty(2 * cp_size, dtype=torch.int32, device=device)
    for rank in range(cp_size):
        chunk_ids[2 * rank] = rank
        chunk_ids[2 * rank + 1] = 2 * cp_size - rank - 1
    return chunk_ids


@jit_fuser
def reorder_seq_chunks_for_a2a_before_attn(x, chunk_ids_for_a2a, seq_dim, cp_size):
    """Reorder sequence chunk for A2A communication before attention compute."""
    # [cp, b, s, np//cp, hn] -> [b, cp, s, np//cp, hn]
    # or [cp, s, b, np//cp, hn] -> [cp, s, b, np//cp, hn]
    x = x.movedim(0, seq_dim).contiguous()
    # [b, cp, s, np//cp, hn] -> [b, cp*2, s//2, np//cp, hn]
    # or [cp, s, b, np//cp, hn] -> [cp*2, s//2, b, np//cp, hn]
    x = x.view(*x.shape[:seq_dim], cp_size * 2, -1, *x.shape[(seq_dim + 2) :])
    # reorder the sequence chunks
    x = torch.index_select(x, dim=seq_dim, index=chunk_ids_for_a2a)
    return x


@jit_fuser
def reorder_seq_chunks_for_a2a_after_attn(x, chunk_ids_for_a2a, seq_dim, cp_size):
    """Reorder sequence chunk for A2A communication after attention compute."""
    # [b, cp*2, s//2, np//cp, hn] -> [cp*2, b, s//2, np//cp, hn]
    # or [cp*2, s//2, b, np//cp, hn] -> [cp*2, s//2, b, np//cp, hn]
    x = x.movedim(seq_dim, 0).contiguous()
    # reorder the sequence chunks
    x = torch.index_select(x, dim=0, index=chunk_ids_for_a2a)
    # [cp*2, b, s//2, np//cp, hn] -> [cp, 2, b, s//2, np//cp, hn]
    # or [cp*2, s//2, b, np//cp, hn] -> [cp, 2, s//2, b, np//cp, hn]
    x = x.view(cp_size, 2, *x.shape[1:])
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
    return x


def flash_attn_a2a_communicate(
    a2a_inputs: Union[torch.Tensor, List[torch.Tensor]],
    chunk_ids_for_a2a: torch.Tensor,
    seq_dim: int,
    cp_size: int,
    cp_group: dist_group_type,
    cp_stream: torch.cuda.Stream,
    before_attn: bool,
) -> Union[torch.Tensor, List[torch.Tensor]]:
    """A2A communication for context parallelism."""
    a2a_inputs = [a2a_inputs] if not isinstance(a2a_inputs, list) else a2a_inputs
    a2a_outputs, a2a_reqs = [None] * len(a2a_inputs), [None] * len(a2a_inputs)
    if before_attn:
        for i in range(len(a2a_inputs) + 2):
            if 0 < i < len(a2a_inputs) + 1:
                a2a_outputs[i - 1] = torch.empty_like(a2a_inputs[i - 1])
                a2a_reqs[i - 1] = torch.distributed.all_to_all_single(
                    a2a_outputs[i - 1], a2a_inputs[i - 1], group=cp_group, async_op=True
                )
            if i > 1:
                with torch.cuda.stream(cp_stream):
                    a2a_reqs[i - 2].wait()
                    x = a2a_outputs[i - 2]
                    # reorder the sequence chunks
407
408
                    x = reorder_seq_chunks_for_a2a_before_attn(
                        x, chunk_ids_for_a2a, seq_dim, cp_size
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
                    )
                    # [b, cp*2, s//2, np//cp, hn] -> [b, cp*s, np//cp, hn]
                    # or [cp*2, s//2, b, np//cp, hn] -> [cp*s, b, np//cp, hn]
                    a2a_outputs[i - 2] = x.view(*x.shape[:seq_dim], -1, *x.shape[(seq_dim + 2) :])
            if i < len(a2a_inputs):
                x = a2a_inputs[i]
                # [b, s, np, hn] -> [b, s, cp, np//cp, hn]
                # or [s, b, np, hn] -> [s, b, cp, np//cp, hn]
                x = x.view(*x.shape[:-2], cp_size, x.shape[-2] // cp_size, x.shape[-1])
                # [b, s, cp, np//cp, hn] -> [cp, b, s, np//cp, hn]
                # or [s, b, cp, np//cp, hn] -> [cp, s, b, np//cp, hn]
                a2a_inputs[i] = x.movedim(-3, 0).contiguous()
    else:
        for i in range(len(a2a_inputs) + 2):
            if 0 < i < len(a2a_inputs) + 1:
                a2a_outputs[i - 1] = torch.empty_like(a2a_inputs[i - 1])
                a2a_reqs[i - 1] = torch.distributed.all_to_all_single(
                    a2a_outputs[i - 1], a2a_inputs[i - 1], group=cp_group, async_op=True
                )
            if i < len(a2a_inputs):
                x = a2a_inputs[i]
                # [b, cp*s, np//cp, hn] -> [b, cp*2, s//2, np//cp, hn]
                # or [cp*s, b, np//cp, hn] -> [cp*2, s//2, b, np//cp, hn]
                x = x.view(*x.shape[:seq_dim], cp_size * 2, -1, *x.shape[(seq_dim + 1) :])
                # reorder the sequence chunks
434
435
                a2a_inputs[i] = reorder_seq_chunks_for_a2a_after_attn(
                    x, chunk_ids_for_a2a, seq_dim, cp_size
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
                )
            if i > 1:
                with torch.cuda.stream(cp_stream):
                    a2a_reqs[i - 2].wait()
                    x = a2a_outputs[i - 2]
                    # [cp, 2, b, s//2, np//cp, hn] -> [b, 2, s//2, cp, np//cp, hn]
                    # or [cp, 2, s//2, b, np//cp, hn] -> [2, s//2, b, cp, np//cp, hn]
                    x = x.movedim(0, -3).movedim(0, seq_dim).contiguous()
                    # [b, 2, s//2, cp, np//cp, hn] -> [b*s, np, hn]
                    # or [2, s//2, b, cp, np//cp, hn] -> [s*b, np, hn]
                    a2a_outputs[i - 2] = x.view(-1, x.shape[-3] * x.shape[-2], x.shape[-1])
    torch.cuda.current_stream().wait_stream(cp_stream)
    return a2a_outputs[0] if len(a2a_inputs) == 1 else a2a_outputs


451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
_cu_seqlens_info_with_cp_cache = {}


def _get_cu_seqlens_info_with_cp(
    batch_size: int,
    max_seqlen: int,
    cp_size: int,
    cu_seqlens: torch.Tensor,
):
    """Cumulative sequence lengths with CP being considered."""
    global _cu_seqlens_info_with_cp_cache
    if (batch_size, max_seqlen, cp_size) not in _cu_seqlens_info_with_cp_cache:
        _cu_seqlens_info_with_cp_cache[(batch_size, max_seqlen, cp_size)] = (
            cu_seqlens // cp_size,
            cu_seqlens // (cp_size * 2),
        )
    return _cu_seqlens_info_with_cp_cache[(batch_size, max_seqlen, cp_size)]


470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
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
548
549
550
551
552
def get_fa_args(
    forward: bool,
    use_flash_attn_3: bool,
    qkv_format: str,
    cu_seqlens_q=None,
    cu_seqlens_kv=None,
    max_seqlen_q=None,
    max_seqlen_kv=None,
    dq=None,
    dk=None,
    dv=None,
):
    """Get forward/backward arguments for flash-attn v2 and v3."""
    if use_flash_attn_3:
        if forward:
            if qkv_format == "thd":
                return [
                    *[None] * 4,  # k_new, v_new, qv, out
                    cu_seqlens_q,
                    cu_seqlens_kv,
                    *[None] * 3,  # cu_seqlens_k_new, seqused_q, seqused_k
                    max_seqlen_q,
                    max_seqlen_kv,
                    *[None]
                    * 8,  # page_table, kv_batch_idx, leftpad_k, rotary_cos, rotary_sin, q_descale, k_descale, v_descale
                ]
            return [
                *[None]
                * 9,  # k_new, v_new, qv, out, cu_seqlens_q, cu_seqlens_kv, cu_seqlens_k_new, seqused_q, seqused_k
                max_seqlen_q,
                max_seqlen_kv,
                *[None]
                * 8,  # page_table, kv_batch_idx, leftpad_k, rotary_cos, rotary_sin, q_descale, k_descale, v_descale
            ]
        if qkv_format == "thd":
            return [
                cu_seqlens_q,
                cu_seqlens_kv,
                None,  # sequed_q
                None,  # sequed_k
                max_seqlen_q,
                max_seqlen_kv,
                dq,
                dk,
                dv,
            ]
        return [
            None,  # cu_seqlens_q
            None,  # cu_seqlens_kv
            None,  # sequed_q
            None,  # sequed_k
            max_seqlen_q,
            max_seqlen_kv,
            dq,
            dk,
            dv,
        ]
    if forward:
        if qkv_format == "thd":
            return [
                cu_seqlens_q,
                cu_seqlens_kv,
                max_seqlen_q,
                max_seqlen_kv,
            ]
        return []
    if qkv_format == "thd":
        return [
            dq,
            dk,
            dv,
            cu_seqlens_q,
            cu_seqlens_kv,
            max_seqlen_q,
            max_seqlen_kv,
        ]
    return [
        dq,
        dk,
        dv,
    ]


553
class AttnFuncWithCPAndKVP2P(torch.autograd.Function):
554
    """
555
556
557
    Attention implementation with context parallelism. Exchange KV between CP ranks
    with P2P in ring topology. Split attention compute into multiple steps, and overlap
    current-step compute with next-step communication.
558
559
560
561
562

    This implementation also supports hierarchical CP, which parallelizes attention
    heads in low-level CP groups and parallelizes sequence dimension in high-level CP
    groups. For more details, please refer to `LongVILA <https://arxiv.org/abs/2408.10188>`_
    and `USP <https://arxiv.org/abs/2405.07719>`_.
563
564
565
    """

    @staticmethod
566
567
568
569
570
571
572
    def forward(
        ctx,
        is_training,
        q,
        k,
        v,
        cu_seqlens_q,
573
        cu_seqlens_kv,
574
        max_seqlen_q,
575
        max_seqlen_kv,
576
577
        cu_seqlens_q_padded,
        cu_seqlens_kv_padded,
578
579
580
581
582
583
584
585
        dropout_p,
        softmax_scale,
        qkv_format,
        attn_mask_type,
        attn_bias_type,
        attn_bias,
        deterministic,
        use_fused_attention,
586
587
        fp8,
        fp8_meta,
588
589
590
        cp_group,
        cp_global_ranks,
        cp_stream,
591
        quantizers,
592
        pad_between_seqs,
593
        use_flash_attn_3,
594
    ):
595
        # pylint: disable=missing-function-docstring
596
        nvtx_range_push("transformer_engine.AttnFuncWithCPAndKVP2P.forward")
597
598
599
        if softmax_scale is None:
            softmax_scale = q.shape[-1] ** (-0.5)

600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
        if isinstance(cp_group, list):
            assert (
                qkv_format != "thd"
            ), f"{qkv_format} format is not supported with hierarchical CP implementation yet!"
            assert attn_bias_type == "no_bias", (
                f"{attn_bias_type} bias type is not supported with hierarchical CP implementation"
                " yet!"
            )
            cp_group_a2a = cp_group[0]
            cp_size_a2a = get_distributed_world_size(cp_group_a2a)
            rank_a2a = get_distributed_rank(cp_group_a2a)
            cp_group = cp_group[1]
        else:
            cp_group_a2a = None
            cp_size_a2a = 1
            rank_a2a = 0

617
618
        cp_size = get_distributed_world_size(cp_group)
        rank = get_distributed_rank(cp_group)
619
620
        send_dst = cp_global_ranks[(rank + 1) % cp_size * cp_size_a2a + rank_a2a]
        recv_src = cp_global_ranks[(rank - 1) % cp_size * cp_size_a2a + rank_a2a]
621
622
        batch_p2p_comm = int(os.getenv("NVTE_BATCH_MHA_P2P_COMM", "0")) or (cp_size == 2)

623
624
        causal = "causal" in attn_mask_type
        padding = "padding" in attn_mask_type
625

626
        batch_dim = None
627
        seq_dim = None
628
        cu_seqlens_q_half, cu_seqlens_kv_half = None, None
629
        if qkv_format in ["bshd", "sbhd"]:
630
            seq_dim = qkv_format.index("s")
631
            qkv_layout = qkv_format + "_" + qkv_format[:-2] + "2" + qkv_format[-2:]
632
633
634
635
636
637
638
639
640
            cu_seqlens_q_padded, cu_seqlens_kv_padded = None, None
            if use_fused_attention:
                batch_dim = qkv_format.index("b")
                cu_seqlens_q, cu_seqlens_q_half = _get_cu_seqlens_info_with_cp(
                    q.shape[batch_dim], max_seqlen_q, cp_size, cu_seqlens_q
                )
                cu_seqlens_kv, cu_seqlens_kv_half = _get_cu_seqlens_info_with_cp(
                    q.shape[batch_dim], max_seqlen_kv, cp_size, cu_seqlens_kv
                )
641
642
        else:
            qkv_layout = qkv_format + "_" + qkv_format + "_" + qkv_format
643
644
            cu_seqlens_q_padded = cu_seqlens_q_padded // cp_size
            cu_seqlens_kv_padded = cu_seqlens_kv_padded // cp_size
645
646
647
648
649

        max_seqlen_q = max_seqlen_q // cp_size
        max_seqlen_kv = max_seqlen_kv // cp_size
        cu_seqlens_q_per_step = [None for _ in range(cp_size)]
        cu_seqlens_kv_per_step = [None for _ in range(cp_size)]
650

651
        fused_attn_backend = None
652
        qkv_dtype = q.dtype
653
654
655
        amax_per_step = None
        S_quantizer_per_step = [None for _ in range(cp_size)]
        O_CP_quantizer_per_step = [None for _ in range(cp_size)]
656
657
        # "fp8_mha" decides outputs in fp8, while inputs are inferred from the real dtype
        is_input_fp8 = False
658
659
660
661
662
663
664
665
666
667
668
        is_output_fp8 = False

        (
            QKV_quantizer,
            O_quantizer,
            O_CP_quantizer,
            S_quantizer,
            dQKV_quantizer,
            dQKV_CP_quantizer,
            dO_quantizer,
            dP_quantizer,
669
        ) = dpa_utils.get_attention_quantizers(fp8, quantizers, cp_specific_quantizers=True)
670

671
672
673
        if fp8:
            if use_fused_attention:
                fused_attn_backend = FusedAttnBackend["FP8"]
674

675
676
677
678
                assert isinstance(k, q.__class__) and isinstance(
                    v, q.__class__
                ), "q, k, and v must have the same type."
                is_input_fp8 = isinstance(q, Float8Tensor)
679
680
681
682
683
                is_output_fp8 = fp8_meta is not None and fp8_meta["recipe"].fp8_mha
                if is_input_fp8:
                    QKV_quantizer = q._quantizer
                    q, k, v = q._data, k._data, v._data
                else:
684
685
                    q_f16, k_f16, v_f16 = q, k, v
                    if cp_size_a2a == 1 or int(os.getenv("NVTE_FP8_DPA_BWD", "1")):
686
                        q = QKV_quantizer(q_f16)._data
687
                    if int(os.getenv("NVTE_FP8_DPA_BWD", "1")):
688
689
690
691
692
                        k, v = [QKV_quantizer(x)._data for x in [k_f16, v_f16]]
                amax_per_step = torch.zeros((2, cp_size), dtype=torch.float32, device=q.device)
                # partial result quantizer
                for i in range(cp_size):
                    S_quantizer_per_step[i] = S_quantizer.copy()
693
                    S_quantizer_per_step[i].amax = amax_per_step[0][i].reshape((1,))
694
                    O_CP_quantizer_per_step[i] = O_CP_quantizer.copy()
695
                    O_CP_quantizer_per_step[i].amax = amax_per_step[1][i].reshape((1,))
696
697
698
699
700
701
702
703
            else:
                assert False, "FP8 is only supported with Fused Attention!"
        else:
            q_f16 = q
            if use_fused_attention:
                fused_attn_backend = FusedAttnBackend["F16_arbitrary_seqlen"]

        if cp_size_a2a > 1:
704
            chunk_ids_for_a2a = get_seq_chunk_ids_for_reordering_before_attn(cp_size_a2a, q.device)
705

706
707
708
709
710
            q, k, v = flash_attn_a2a_communicate(
                [q, k, v], chunk_ids_for_a2a, seq_dim, cp_size_a2a, cp_group_a2a, cp_stream, True
            )
            if not fp8:
                q_f16 = q
711
            elif not is_input_fp8 and not int(os.getenv("NVTE_FP8_DPA_BWD", "1")):
712
                q_f16 = q
713
                q = QKV_quantizer(q_f16)._data
714

715
716
717
        assert qkv_format == "thd" or (
            q.shape[seq_dim] % 2 == 0 and k.shape[seq_dim] % 2 == 0
        ), "Sequence length per GPU needs to be divisible by 2!"
718
        if causal:
719
720
            if qkv_format == "bshd":
                # [b, s, np, hn] -> [b, 2, s//2, np, hn]
721
                q, k, v = [x.view(x.shape[0], 2, x.shape[1] // 2, *x.shape[2:]) for x in [q, k, v]]
722
723
            elif qkv_format == "sbhd":
                # [s, b, np, hn] -> [2, s//2, b, np, hn]
724
                q, k, v = [x.view(2, x.shape[0] // 2, *x.shape[1:]) for x in [q, k, v]]
725
        if attn_bias is not None:
726
            assert len(attn_bias.shape) == 4, (
727
728
729
                "Only support bias shape of [b, h, sq, sk] for forward, "
                "and [1, h, sq, sk] for backward!"
            )
730
731
732
            assert (
                attn_bias.shape[-2] % 2 == 0 and attn_bias.shape[-1] % (2 * cp_size) == 0
            ), "Sequence length does not meet divisible requirements!"
733
            # [b, np, sq, sk] -> [b, np, 2, sq//2, 2*cp, sk//(2*cp)]
734
735
736
737
738
739
            attn_bias_ = attn_bias.view(
                *attn_bias.shape[:-2],
                2,
                attn_bias.shape[-2] // 2,
                2 * cp_size,
                attn_bias.shape[-1] // (2 * cp_size),
740
741
            )
            # [b, np, sq, sk] -> [b, np, sq, 2*cp, sk//(2*cp)]
742
743
            attn_bias = attn_bias.view(
                *attn_bias.shape[:-1], 2 * cp_size, attn_bias.shape[-1] // (2 * cp_size)
744
            )
745
        assert q.shape[-1] % 8 == 0, "hidden size per attention head should be multiple of 8"
746

747
748
749
750
751
        softmax_lse_in_packed_format = False
        if qkv_format == "thd":
            if use_fused_attention:
                softmax_lse_in_packed_format = get_cudnn_version() >= (9, 6, 0)
            else:
752
                softmax_lse_in_packed_format = fa_utils.v2_6_0_plus or use_flash_attn_3
753

754
        flash_attn_fwd = None
755
756
        if not use_fused_attention:
            fa_forward_kwargs = {"softmax_scale": softmax_scale}
757
758
759
760
            if use_flash_attn_3:
                flash_attn_fwd = (
                    _flash_attn_fwd_v3  # pylint: disable=possibly-used-before-assignment
                )
761
762
                fa_forward_kwargs["window_size"] = (-1, 0) if causal else (-1, -1)
            else:
763
764
765
766
                if qkv_format == "thd":
                    flash_attn_fwd = _flash_attn_varlen_fwd
                else:
                    flash_attn_fwd = _flash_attn_fwd
767
768
                fa_forward_kwargs["dropout_p"] = dropout_p
                fa_forward_kwargs["return_softmax"] = False
769
                if fa_utils.v2_3_plus and not fa_utils.v2_7_0_plus:
770
                    fa_forward_kwargs["window_size"] = (-1, 0) if causal else (-1, -1)
771
                elif fa_utils.v2_7_0_plus:
772
773
                    fa_forward_kwargs["window_size_left"] = -1
                    fa_forward_kwargs["window_size_right"] = 0 if causal else -1
774
                if fa_utils.v2_4_plus:
775
                    fa_forward_kwargs["alibi_slopes"] = None
776
                if fa_utils.v2_5_7_plus and qkv_format == "thd":
777
                    fa_forward_kwargs["block_table"] = None
778
                if fa_utils.v2_6_0_plus:
779
                    fa_forward_kwargs["softcap"] = 0.0
780

781
782
783
        # Flash Attn inputs
        q_inputs = [None, None]
        kv_inputs = [None, None]
784
        attn_bias_inputs = [None, None]
785
786
787
788
        # Flash Attn outputs
        out_per_step = [None for _ in range(cp_size)]
        softmax_lse_per_step = [None for _ in range(cp_size)]
        rng_states = [None for _ in range(cp_size)]
789
        attn_biases = [None for _ in range(cp_size)]
790
791
792
793
794
795
796

        # create two streams to resolve wave quantization issue of Flash Attn in each step
        flash_attn_streams = [torch.cuda.current_stream(), cp_stream]
        # synchronize fwd results correction across steps
        fwd_results_correction_done = torch.cuda.Event()

        p2p_comm_buffers = [None for _ in range(cp_size)]
797
        if qkv_format in ["bshd", "sbhd"]:
798
799
800
            p2p_comm_buffers[0] = torch.cat((k.unsqueeze(-3), v.unsqueeze(-3)), dim=-3)
        else:
            p2p_comm_buffers[0] = torch.cat((k.unsqueeze(0), v.unsqueeze(0)), dim=0)
801
802
        send_recv_reqs = [[], []]

803
        out = None
804
        for i in range(cp_size + 1):
805
            if i < cp_size:
806
                with torch.cuda.stream(flash_attn_streams[i % 2]):
807
                    # wait until KV is received
808
                    for req in send_recv_reqs[(i + 1) % 2]:
809
810
                        req.wait()

811
812
813
814
815
816
817
818
819
820
821
822
                    if i < (cp_size - 1):
                        p2p_comm_buffers[i + 1] = torch.empty_like(p2p_comm_buffers[i])
                        send_recv_reqs[i % 2] = flash_attn_p2p_communicate(
                            rank,
                            p2p_comm_buffers[i],
                            send_dst,
                            p2p_comm_buffers[i + 1],
                            recv_src,
                            cp_group,
                            batch_p2p_comm,
                        )

823
                    if not fp8 or is_input_fp8 or int(os.getenv("NVTE_FP8_DPA_BWD", "1")):
824
825
826
                        kv_inputs[i % 2] = p2p_comm_buffers[i]
                    else:
                        # KV exchange is in BF16/FP16, cast received KV in each step
827
                        kv_inputs[i % 2] = QKV_quantizer(p2p_comm_buffers[i])._data
828
829
                    if causal:
                        if i == 0:
830
                            if pad_between_seqs:
831
832
833
834
835
836
                                cu_seqlens_q_per_step[i] = get_cu_seqlens_on_cp_rank(
                                    cu_seqlens_q, cu_seqlens_q_padded, cp_size, rank, True, True
                                )
                                cu_seqlens_kv_per_step[i] = get_cu_seqlens_on_cp_rank(
                                    cu_seqlens_kv, cu_seqlens_kv_padded, cp_size, rank, True, True
                                )
837
838
                            elif qkv_format == "thd":
                                cu_seqlens_q_per_step[i] = cu_seqlens_q // cp_size
839
                                cu_seqlens_kv_per_step[i] = cu_seqlens_kv // cp_size
840
841
842
                            else:
                                cu_seqlens_q_per_step[i] = cu_seqlens_q
                                cu_seqlens_kv_per_step[i] = cu_seqlens_kv
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
                            if qkv_format == "bshd":
                                # [b, 2, sq//2, np, hn] -> [b, sq, np, hn]
                                q_inputs[i % 2] = q.view(q.shape[0], -1, *q.shape[-2:])
                                # [b, 2, sk//2, 2, np, hn] -> [b, sk, 2, np, hn]
                                kv_inputs[i % 2] = kv_inputs[i % 2].view(
                                    k.shape[0], -1, 2, *k.shape[-2:]
                                )
                            elif qkv_format == "sbhd":
                                # [2, sq//2, b, np, hn] -> [sq, b, np, hn]
                                q_inputs[i % 2] = q.view(-1, *q.shape[-3:])
                                # [2, sk//2, b, 2, np, hn] -> [sk, b, 2, np, hn]
                                kv_inputs[i % 2] = kv_inputs[i % 2].view(
                                    -1, k.shape[2], 2, *k.shape[-2:]
                                )
                            elif qkv_format == "thd":
                                q_inputs[i % 2] = q
859
                            if use_fused_attention:
860
861
                                if attn_bias is not None:
                                    idx = (rank - i) % cp_size
862
863
864
865
866
867
                                    attn_bias_inputs[i % 2] = torch.cat(
                                        (
                                            attn_bias[..., idx, :],
                                            attn_bias[..., (2 * cp_size - idx - 1), :],
                                        ),
                                        dim=-1,
868
                                    ).contiguous()
869
870
871
872
873
874
875
876
877
878
879
880

                                q_part = q_inputs[i % 2]
                                k_part = (
                                    kv_inputs[i % 2][..., 0, :, :]
                                    if qkv_format in ["bshd", "sbhd"]
                                    else kv_inputs[i % 2][0]
                                )
                                v_part = (
                                    kv_inputs[i % 2][..., 1, :, :]
                                    if qkv_format in ["bshd", "sbhd"]
                                    else kv_inputs[i % 2][1]
                                )
881
                                fp8_meta_kwargs = {}
882
883
884
885
886
887
888
889
890
891
                                if fp8:
                                    q_part = QKV_quantizer.create_tensor_from_data(
                                        q_part, fake_dtype=qkv_dtype, internal=True
                                    )
                                    k_part = QKV_quantizer.create_tensor_from_data(
                                        k_part, fake_dtype=qkv_dtype, internal=True
                                    )
                                    v_part = QKV_quantizer.create_tensor_from_data(
                                        v_part, fake_dtype=qkv_dtype, internal=True
                                    )
892
893
                                    fp8_meta_kwargs["s_quantizer"] = S_quantizer_per_step[i]
                                    fp8_meta_kwargs["o_quantizer"] = O_CP_quantizer_per_step[i]
894

895
896
897
898
899
900
                                out_per_step[i], aux_ctx_tensors = fused_attn_fwd(
                                    is_training,
                                    max_seqlen_q,
                                    max_seqlen_kv,
                                    cu_seqlens_q_per_step[i],
                                    cu_seqlens_kv_per_step[i],
901
902
903
904
905
                                    q_part,
                                    k_part,
                                    v_part,
                                    fake_dtype=qkv_dtype,
                                    fused_attention_backend=fused_attn_backend,
906
907
908
909
910
911
912
913
914
                                    attn_scale=softmax_scale,
                                    dropout=dropout_p,
                                    qkv_layout=qkv_layout,
                                    attn_mask_type=attn_mask_type,
                                    attn_bias_type=attn_bias_type,
                                    attn_bias=attn_bias_inputs[i % 2],
                                    cu_seqlens_q_padded=cu_seqlens_q_padded,
                                    cu_seqlens_kv_padded=cu_seqlens_kv_padded,
                                    **fp8_meta_kwargs,
915
                                )
916
917
918
919
920
                                if fp8:
                                    softmax_lse_per_step[i], _, rng_states[i] = aux_ctx_tensors
                                else:
                                    softmax_lse_per_step[i], rng_states[i], *rest = aux_ctx_tensors
                                    attn_biases[i] = rest[0] if len(rest) > 0 else None
921
                            else:
922
923
924
925
926
927
928
929
930
                                fa_forward_args_thd = get_fa_args(
                                    True,
                                    use_flash_attn_3,
                                    qkv_format,
                                    cu_seqlens_q=cu_seqlens_q_per_step[i],
                                    cu_seqlens_kv=cu_seqlens_kv_per_step[i],
                                    max_seqlen_q=max_seqlen_q,
                                    max_seqlen_kv=max_seqlen_kv,
                                )
931
                                fa_outputs = flash_attn_fwd(
932
                                    q_inputs[i % 2],
933
934
935
936
937
938
939
940
941
942
943
                                    (
                                        kv_inputs[i % 2][..., 0, :, :]
                                        if qkv_format in ["bshd", "sbhd"]
                                        else kv_inputs[i % 2][0]
                                    ),
                                    (
                                        kv_inputs[i % 2][..., 1, :, :]
                                        if qkv_format in ["bshd", "sbhd"]
                                        else kv_inputs[i % 2][1]
                                    ),
                                    *fa_forward_args_thd,
944
                                    causal=True,
945
                                    **fa_forward_kwargs,
946
                                )
947
                                if not fa_utils.v2_7_0_plus:
948
949
                                    out_per_step[i] = fa_outputs[4]
                                    softmax_lse_per_step[i] = fa_outputs[5]
950
                                    if not use_flash_attn_3:
951
952
953
954
                                        rng_states[i] = fa_outputs[7]
                                else:
                                    out_per_step[i] = fa_outputs[0]
                                    softmax_lse_per_step[i] = fa_outputs[1]
955
                                    if not use_flash_attn_3:
956
                                        rng_states[i] = fa_outputs[3]
957
                        elif i <= rank:
958
                            if pad_between_seqs:
959
960
961
962
963
964
965
966
967
968
969
                                cu_seqlens_q_per_step[i] = get_cu_seqlens_on_cp_rank(
                                    cu_seqlens_q, cu_seqlens_q_padded, cp_size, rank, True, True
                                )
                                cu_seqlens_kv_per_step[i] = get_cu_seqlens_on_cp_rank(
                                    cu_seqlens_kv,
                                    cu_seqlens_kv_padded,
                                    cp_size,
                                    (rank - i) % cp_size,
                                    True,
                                    False,
                                )
970
971
                            elif qkv_format == "thd":
                                cu_seqlens_q_per_step[i] = cu_seqlens_q // cp_size
972
                                cu_seqlens_kv_per_step[i] = cu_seqlens_kv // (cp_size * 2)
973
974
975
                            else:
                                cu_seqlens_q_per_step[i] = cu_seqlens_q
                                cu_seqlens_kv_per_step[i] = cu_seqlens_kv_half
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
                            if qkv_format == "bshd":
                                # [b, 2, sq//2, np, hn] -> [b, sq, np, hn]
                                q_inputs[i % 2] = q.view(q.shape[0], -1, *q.shape[-2:])
                                # [b, 2, sk//2, 2, np, hn] -> [b, sk//2, 2, np, hn]
                                kv_inputs[i % 2] = kv_inputs[i % 2][:, 0, ...]
                            elif qkv_format == "sbhd":
                                # [2, sq//2, b, np, hn] -> [sq, b, np, hn]
                                q_inputs[i % 2] = q.view(-1, *q.shape[-3:])
                                # [2, sk//2, b, 2, np, hn] -> [sk//2, b, 2, np, hn]
                                kv_inputs[i % 2] = kv_inputs[i % 2][0]
                            elif qkv_format == "thd":
                                q_inputs[i % 2] = q
                                # [2, t, np, hn] -> [2, t/2, np, hn]
                                kv_inputs[i % 2] = tex.thd_read_half_tensor(
                                    kv_inputs[i % 2], cu_seqlens_kv_padded, 0
                                )
992
                            if use_fused_attention:
993
                                kv_inputs[i % 2] = kv_inputs[i % 2].contiguous()
994
995
                                if attn_bias is not None:
                                    idx = (rank - i) % cp_size
996
                                    attn_bias_inputs[i % 2] = attn_bias[..., idx, :].contiguous()
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008

                                q_part = q_inputs[i % 2]
                                k_part = (
                                    kv_inputs[i % 2][..., 0, :, :]
                                    if qkv_format in ["bshd", "sbhd"]
                                    else kv_inputs[i % 2][0]
                                )
                                v_part = (
                                    kv_inputs[i % 2][..., 1, :, :]
                                    if qkv_format in ["bshd", "sbhd"]
                                    else kv_inputs[i % 2][1]
                                )
1009
                                fp8_meta_kwargs = {}
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
                                if fp8:
                                    q_part = QKV_quantizer.create_tensor_from_data(
                                        q_part, fake_dtype=qkv_dtype, internal=True
                                    )
                                    k_part = QKV_quantizer.create_tensor_from_data(
                                        k_part, fake_dtype=qkv_dtype, internal=True
                                    )
                                    v_part = QKV_quantizer.create_tensor_from_data(
                                        v_part, fake_dtype=qkv_dtype, internal=True
                                    )
1020
1021
                                    fp8_meta_kwargs["s_quantizer"] = S_quantizer_per_step[i]
                                    fp8_meta_kwargs["o_quantizer"] = O_CP_quantizer_per_step[i]
1022
1023
1024
1025
1026
1027
                                out_per_step[i], aux_ctx_tensors = fused_attn_fwd(
                                    is_training,
                                    max_seqlen_q,
                                    max_seqlen_kv // 2,
                                    cu_seqlens_q_per_step[i],
                                    cu_seqlens_kv_per_step[i],
1028
1029
1030
1031
                                    q_part,
                                    k_part,
                                    v_part,
                                    qkv_dtype,
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
                                    fused_attn_backend,
                                    attn_scale=softmax_scale,
                                    dropout=dropout_p,
                                    qkv_layout=qkv_layout,
                                    attn_mask_type="padding" if padding else "no_mask",
                                    attn_bias_type=attn_bias_type,
                                    attn_bias=attn_bias_inputs[i % 2],
                                    cu_seqlens_q_padded=cu_seqlens_q_padded,
                                    cu_seqlens_kv_padded=(
                                        None
                                        if cu_seqlens_kv_padded is None
                                        else cu_seqlens_kv_padded // 2
                                    ),
                                    **fp8_meta_kwargs,
1046
                                )
1047
1048
1049
1050
1051
                                if fp8:
                                    softmax_lse_per_step[i], _, rng_states[i] = aux_ctx_tensors
                                else:
                                    softmax_lse_per_step[i], rng_states[i], *rest = aux_ctx_tensors
                                    attn_biases[i] = rest[0] if len(rest) > 0 else None
1052
                            else:
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
                                fa_forward_args_thd = get_fa_args(
                                    True,
                                    use_flash_attn_3,
                                    qkv_format,
                                    cu_seqlens_q=cu_seqlens_q_per_step[i],
                                    cu_seqlens_kv=cu_seqlens_kv_per_step[i],
                                    max_seqlen_q=max_seqlen_q,
                                    max_seqlen_kv=max_seqlen_kv // 2,
                                )
                                if use_flash_attn_3 or (
1063
                                    fa_utils.v2_3_plus and not fa_utils.v2_7_0_plus
1064
                                ):
1065
                                    fa_forward_kwargs["window_size"] = (-1, -1)
1066
                                elif fa_utils.v2_7_0_plus:
1067
1068
                                    fa_forward_kwargs["window_size_left"] = -1
                                    fa_forward_kwargs["window_size_right"] = -1
1069
                                fa_outputs = flash_attn_fwd(
1070
                                    q_inputs[i % 2],
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
                                    (
                                        kv_inputs[i % 2][..., 0, :, :]
                                        if qkv_format in ["bshd", "sbhd"]
                                        else kv_inputs[i % 2][0]
                                    ),
                                    (
                                        kv_inputs[i % 2][..., 1, :, :]
                                        if qkv_format in ["bshd", "sbhd"]
                                        else kv_inputs[i % 2][1]
                                    ),
                                    *fa_forward_args_thd,
1082
                                    causal=False,
1083
                                    **fa_forward_kwargs,
1084
                                )
1085
                                if not fa_utils.v2_7_0_plus:
1086
1087
                                    out_per_step[i] = fa_outputs[4]
                                    softmax_lse_per_step[i] = fa_outputs[5]
1088
                                    if not use_flash_attn_3:
1089
1090
1091
1092
                                        rng_states[i] = fa_outputs[7]
                                else:
                                    out_per_step[i] = fa_outputs[0]
                                    softmax_lse_per_step[i] = fa_outputs[1]
1093
                                    if not use_flash_attn_3:
1094
                                        rng_states[i] = fa_outputs[3]
1095
                        else:
1096
                            if pad_between_seqs:
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
                                cu_seqlens_q_per_step[i] = get_cu_seqlens_on_cp_rank(
                                    cu_seqlens_q, cu_seqlens_q_padded, cp_size, rank, False, True
                                )
                                cu_seqlens_kv_per_step[i] = get_cu_seqlens_on_cp_rank(
                                    cu_seqlens_kv,
                                    cu_seqlens_kv_padded,
                                    cp_size,
                                    (rank - i) % cp_size,
                                    True,
                                    True,
                                )
1108
1109
                            elif qkv_format == "thd":
                                cu_seqlens_q_per_step[i] = cu_seqlens_q // (cp_size * 2)
1110
                                cu_seqlens_kv_per_step[i] = cu_seqlens_kv // cp_size
1111
1112
1113
                            else:
                                cu_seqlens_q_per_step[i] = cu_seqlens_q_half
                                cu_seqlens_kv_per_step[i] = cu_seqlens_kv
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
                            if qkv_format == "bshd":
                                # [b, 2, sq//2, np, hn] -> [b, sq//2, np, hn]
                                q_inputs[i % 2] = q[:, 1, ...]
                                # [b, 2, sk//2, 2, np, hn] -> [b, sk, 2, np, hn]
                                kv_inputs[i % 2] = kv_inputs[i % 2].view(
                                    k.shape[0], -1, 2, *k.shape[-2:]
                                )
                            elif qkv_format == "sbhd":
                                # [2, sq//2, b, np, hn] -> [sq//2, b, np, hn]
                                q_inputs[i % 2] = q[1]
                                # [2, sk//2, b, 2, np, hn] -> [sk, b, 2, np, hn]
                                kv_inputs[i % 2] = kv_inputs[i % 2].view(
                                    -1, k.shape[2], 2, *k.shape[-2:]
                                )
                            elif qkv_format == "thd":
                                # [t, np, hn] -> [t/2, np, hn]
                                q_inputs[i % 2] = tex.thd_read_half_tensor(
                                    q, cu_seqlens_q_padded, 1
                                )
1133
                            if use_fused_attention:
1134
                                q_inputs[i % 2] = q_inputs[i % 2].contiguous()
1135
1136
                                if attn_bias is not None:
                                    idx = (rank - i) % cp_size
1137
1138
1139
1140
1141
1142
                                    attn_bias_inputs[i % 2] = torch.cat(
                                        (
                                            attn_bias_[..., 1, :, idx, :],
                                            attn_bias_[..., 1, :, (2 * cp_size - idx - 1), :],
                                        ),
                                        dim=-1,
1143
                                    ).contiguous()
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155

                                q_part = q_inputs[i % 2]
                                k_part = (
                                    kv_inputs[i % 2][..., 0, :, :]
                                    if qkv_format in ["bshd", "sbhd"]
                                    else kv_inputs[i % 2][0]
                                )
                                v_part = (
                                    kv_inputs[i % 2][..., 1, :, :]
                                    if qkv_format in ["bshd", "sbhd"]
                                    else kv_inputs[i % 2][1]
                                )
1156
                                fp8_meta_kwargs = {}
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
                                if fp8:
                                    q_part = QKV_quantizer.create_tensor_from_data(
                                        q_part, fake_dtype=qkv_dtype, internal=True
                                    )
                                    k_part = QKV_quantizer.create_tensor_from_data(
                                        k_part, fake_dtype=qkv_dtype, internal=True
                                    )
                                    v_part = QKV_quantizer.create_tensor_from_data(
                                        v_part, fake_dtype=qkv_dtype, internal=True
                                    )
1167
1168
                                    fp8_meta_kwargs["s_quantizer"] = S_quantizer_per_step[i]
                                    fp8_meta_kwargs["o_quantizer"] = O_CP_quantizer_per_step[i]
1169
1170
1171
1172
1173
1174
                                out_per_step[i], aux_ctx_tensors = fused_attn_fwd(
                                    is_training,
                                    max_seqlen_q // 2,
                                    max_seqlen_kv,
                                    cu_seqlens_q_per_step[i],
                                    cu_seqlens_kv_per_step[i],
1175
1176
1177
1178
                                    q_part,
                                    k_part,
                                    v_part,
                                    qkv_dtype,
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
                                    fused_attn_backend,
                                    attn_scale=softmax_scale,
                                    dropout=dropout_p,
                                    qkv_layout=qkv_layout,
                                    attn_mask_type="padding" if padding else "no_mask",
                                    attn_bias_type=attn_bias_type,
                                    attn_bias=attn_bias_inputs[i % 2],
                                    cu_seqlens_q_padded=(
                                        None
                                        if cu_seqlens_q_padded is None
                                        else cu_seqlens_q_padded // 2
                                    ),
                                    cu_seqlens_kv_padded=cu_seqlens_kv_padded,
                                    **fp8_meta_kwargs,
1193
                                )
1194
1195
1196
1197
1198
                                if fp8:
                                    softmax_lse_per_step[i], _, rng_states[i] = aux_ctx_tensors
                                else:
                                    softmax_lse_per_step[i], rng_states[i], *rest = aux_ctx_tensors
                                    attn_biases[i] = rest[0] if len(rest) > 0 else None
1199
                            else:
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
                                fa_forward_args_thd = get_fa_args(
                                    True,
                                    use_flash_attn_3,
                                    qkv_format,
                                    cu_seqlens_q=cu_seqlens_q_per_step[i],
                                    cu_seqlens_kv=cu_seqlens_kv_per_step[i],
                                    max_seqlen_q=max_seqlen_q // 2,
                                    max_seqlen_kv=max_seqlen_kv,
                                )
                                if use_flash_attn_3 or (
1210
                                    fa_utils.v2_3_plus and not fa_utils.v2_7_0_plus
1211
                                ):
1212
                                    fa_forward_kwargs["window_size"] = (-1, -1)
1213
                                elif fa_utils.v2_7_0_plus:
1214
1215
                                    fa_forward_kwargs["window_size_left"] = -1
                                    fa_forward_kwargs["window_size_right"] = -1
1216
                                fa_outputs = flash_attn_fwd(
1217
                                    q_inputs[i % 2],
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
                                    (
                                        kv_inputs[i % 2][..., 0, :, :]
                                        if qkv_format in ["bshd", "sbhd"]
                                        else kv_inputs[i % 2][0]
                                    ),
                                    (
                                        kv_inputs[i % 2][..., 1, :, :]
                                        if qkv_format in ["bshd", "sbhd"]
                                        else kv_inputs[i % 2][1]
                                    ),
                                    *fa_forward_args_thd,
1229
                                    causal=False,
1230
                                    **fa_forward_kwargs,
1231
                                )
1232
                                if not fa_utils.v2_7_0_plus:
1233
1234
                                    out_per_step[i] = fa_outputs[4]
                                    softmax_lse_per_step[i] = fa_outputs[5]
1235
                                    if not use_flash_attn_3:
1236
1237
1238
1239
                                        rng_states[i] = fa_outputs[7]
                                else:
                                    out_per_step[i] = fa_outputs[0]
                                    softmax_lse_per_step[i] = fa_outputs[1]
1240
                                    if not use_flash_attn_3:
1241
                                        rng_states[i] = fa_outputs[3]
1242
                    else:
1243
                        if pad_between_seqs:
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
                            cu_seqlens_q_per_step[i] = get_cu_seqlens_on_cp_rank(
                                cu_seqlens_q, cu_seqlens_q_padded, cp_size, rank, True, True
                            )
                            cu_seqlens_kv_per_step[i] = get_cu_seqlens_on_cp_rank(
                                cu_seqlens_kv,
                                cu_seqlens_kv_padded,
                                cp_size,
                                (rank - i) % cp_size,
                                True,
                                True,
                            )
1255
1256
                        elif qkv_format == "thd":
                            cu_seqlens_q_per_step[i] = cu_seqlens_q // cp_size
1257
                            cu_seqlens_kv_per_step[i] = cu_seqlens_kv // cp_size
1258
1259
1260
                        else:
                            cu_seqlens_q_per_step[i] = cu_seqlens_q
                            cu_seqlens_kv_per_step[i] = cu_seqlens_kv
1261
                        if use_fused_attention:
1262
1263
                            if attn_bias is not None:
                                idx = (rank - i) % cp_size
1264
1265
1266
1267
1268
1269
                                attn_bias_inputs[i % 2] = torch.cat(
                                    (
                                        attn_bias[..., idx, :],
                                        attn_bias[..., (2 * cp_size - idx - 1), :],
                                    ),
                                    dim=-1,
1270
                                ).contiguous()
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282

                            q_part = q
                            k_part = (
                                kv_inputs[i % 2][..., 0, :, :]
                                if qkv_format in ["bshd", "sbhd"]
                                else kv_inputs[i % 2][0]
                            )
                            v_part = (
                                kv_inputs[i % 2][..., 1, :, :]
                                if qkv_format in ["bshd", "sbhd"]
                                else kv_inputs[i % 2][1]
                            )
1283
                            fp8_meta_kwargs = {}
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
                            if fp8:
                                q_part = QKV_quantizer.create_tensor_from_data(
                                    q_part, fake_dtype=qkv_dtype, internal=True
                                )
                                k_part = QKV_quantizer.create_tensor_from_data(
                                    k_part, fake_dtype=qkv_dtype, internal=True
                                )
                                v_part = QKV_quantizer.create_tensor_from_data(
                                    v_part, fake_dtype=qkv_dtype, internal=True
                                )
1294
1295
                                fp8_meta_kwargs["s_quantizer"] = S_quantizer_per_step[i]
                                fp8_meta_kwargs["o_quantizer"] = O_CP_quantizer_per_step[i]
1296
1297
1298
1299
1300
1301
                            out_per_step[i], aux_ctx_tensors = fused_attn_fwd(
                                is_training,
                                max_seqlen_q,
                                max_seqlen_kv,
                                cu_seqlens_q_per_step[i],
                                cu_seqlens_kv_per_step[i],
1302
1303
1304
1305
                                q_part,
                                k_part,
                                v_part,
                                qkv_dtype,
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
                                fused_attn_backend,
                                attn_scale=softmax_scale,
                                dropout=dropout_p,
                                qkv_layout=qkv_layout,
                                attn_mask_type=attn_mask_type,
                                attn_bias_type=attn_bias_type,
                                attn_bias=attn_bias_inputs[i % 2],
                                cu_seqlens_q_padded=cu_seqlens_q_padded,
                                cu_seqlens_kv_padded=cu_seqlens_kv_padded,
                                **fp8_meta_kwargs,
1316
                            )
1317
1318
1319
1320
1321
                            if fp8:
                                softmax_lse_per_step[i], _, rng_states[i] = aux_ctx_tensors
                            else:
                                softmax_lse_per_step[i], rng_states[i], *rest = aux_ctx_tensors
                                attn_biases[i] = rest[0] if len(rest) > 0 else None
1322
                        else:
1323
1324
1325
1326
1327
1328
1329
1330
1331
                            fa_forward_args_thd = get_fa_args(
                                True,
                                use_flash_attn_3,
                                qkv_format,
                                cu_seqlens_q=cu_seqlens_q_per_step[i],
                                cu_seqlens_kv=cu_seqlens_kv_per_step[i],
                                max_seqlen_q=max_seqlen_q,
                                max_seqlen_kv=max_seqlen_kv,
                            )
1332
                            fa_outputs = flash_attn_fwd(
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
                                q,
                                (
                                    kv_inputs[i % 2][..., 0, :, :]
                                    if qkv_format in ["bshd", "sbhd"]
                                    else kv_inputs[i % 2][0]
                                ),
                                (
                                    kv_inputs[i % 2][..., 1, :, :]
                                    if qkv_format in ["bshd", "sbhd"]
                                    else kv_inputs[i % 2][1]
                                ),
                                *fa_forward_args_thd,
1345
                                causal=False,
1346
                                **fa_forward_kwargs,
1347
                            )
1348
                            if not fa_utils.v2_7_0_plus:
1349
1350
                                out_per_step[i] = fa_outputs[4]
                                softmax_lse_per_step[i] = fa_outputs[5]
1351
                                if not use_flash_attn_3:
1352
1353
1354
1355
                                    rng_states[i] = fa_outputs[7]
                            else:
                                out_per_step[i] = fa_outputs[0]
                                softmax_lse_per_step[i] = fa_outputs[1]
1356
                                if not use_flash_attn_3:
1357
                                    rng_states[i] = fa_outputs[3]
1358
1359
1360
1361

            if i > 0:
                # wait until fwd restuls correction of last step is done
                if i > 1:
1362
                    flash_attn_streams[(i - 1) % 2].wait_event(fwd_results_correction_done)
1363

1364
                with torch.cuda.stream(flash_attn_streams[(i - 1) % 2]):
1365
1366
1367
1368
1369
1370
1371
1372
                    if use_fused_attention:
                        # [b, np, sq, 1] -> [b, np, sq] or
                        # [t, np, 1] -> [t, np]
                        softmax_lse_per_step[i - 1].squeeze_(-1)
                        if softmax_lse_in_packed_format:
                            softmax_lse_per_step[i - 1] = (
                                softmax_lse_per_step[i - 1].transpose(0, 1).contiguous()
                            )
1373
                    if fp8:
1374
                        out_per_step[i - 1] = out_per_step[i - 1].dequantize(dtype=torch.float32)
1375
1376
                    if i == 1:
                        softmax_lse = torch.clone(softmax_lse_per_step[0]).to(torch.double)
1377
1378
                        if qkv_format == "thd":
                            out = torch.zeros_like(q if not fp8 else out_per_step[0]).view(q.shape)
1379
1380
1381
1382
                    elif (i - 1) <= rank or not causal:
                        flash_attn_fwd_softmax_lse_correction(
                            softmax_lse, softmax_lse_per_step[i - 1]
                        )
1383
                    else:
1384
                        if qkv_format == "thd":
1385
                            tex.thd_second_half_lse_correction(
1386
1387
1388
                                softmax_lse,
                                softmax_lse_per_step[i - 1],
                                cu_seqlens_q_padded,
1389
                                softmax_lse_in_packed_format,
1390
                            )
1391
                        else:
1392
1393
1394
                            flash_attn_fwd_second_half_softmax_lse_correction(
                                softmax_lse.view(*softmax_lse.shape[:-1], 2, -1),
                                softmax_lse_per_step[i - 1],
1395
                            )
1396
1397

                if i < cp_size:
1398
                    flash_attn_streams[(i - 1) % 2].record_event(fwd_results_correction_done)
1399
1400
1401

        torch.cuda.current_stream().wait_stream(flash_attn_streams[1])

1402
1403
1404
1405
        second_half_lse_seqlen = None
        if causal and rank < (cp_size - 1):
            second_half_lse_seqlen = softmax_lse_per_step[-1].shape[-1]

1406
1407
        softmax_lse = softmax_lse.to(torch.float)
        for i in range(cp_size):
1408
            if i <= rank or not causal:
1409
                if qkv_format in ["bshd", "sbhd"]:
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
                    if i == 0:
                        out = flash_attn_fwd_out_correction_init(
                            out_per_step[0],
                            softmax_lse,
                            softmax_lse_per_step[0],
                            seq_dim,
                        )
                        out = out.view(q.shape)
                    else:
                        flash_attn_fwd_out_correction(
                            out.view(*out_per_step[i].shape),
                            out_per_step[i],
                            softmax_lse,
                            softmax_lse_per_step[i],
                            seq_dim,
                        )
1426
                elif qkv_format == "thd":
1427
1428
1429
1430
1431
                    tex.thd_out_correction(
                        out,
                        out_per_step[i],
                        softmax_lse,
                        softmax_lse_per_step[i],
1432
                        cu_seqlens_q_padded,
1433
                        False,
1434
                        softmax_lse_in_packed_format,
1435
                    )
1436
            else:
1437
                if qkv_format in ["bshd", "sbhd"]:
1438
1439
                    flash_attn_fwd_second_half_out_correction(
                        out,
1440
                        out_per_step[i],
1441
                        softmax_lse,
1442
                        softmax_lse_per_step[i],
1443
                        seq_dim,
1444
                    )
1445
                elif qkv_format == "thd":
1446
1447
1448
1449
1450
                    tex.thd_out_correction(
                        out,
                        out_per_step[i],
                        softmax_lse,
                        softmax_lse_per_step[i],
1451
                        cu_seqlens_q_padded,
1452
                        True,
1453
                        softmax_lse_in_packed_format,
1454
                    )
1455
1456

        kv = p2p_comm_buffers[-1]
1457
1458
1459
1460
1461
1462
1463
1464
        if qkv_format == "bshd":
            out = out.view(out.shape[0], -1, *out.shape[-2:])
            ctx.batch_size = out.shape[0]
        elif qkv_format == "sbhd":
            out = out.view(-1, *out.shape[-3:])
            ctx.batch_size = out.shape[1]

        if cp_size_a2a > 1:
1465
            chunk_ids_for_a2a = get_seq_chunk_ids_for_reordering_after_attn(cp_size_a2a, out.device)
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
            out = flash_attn_a2a_communicate(
                out, chunk_ids_for_a2a, seq_dim, cp_size_a2a, cp_group_a2a, cp_stream, False
            )
            if use_fused_attention:
                if qkv_format == "bshd":
                    # [b*s, np, hn] -> [b, s, np, hn]
                    out = out.view(ctx.batch_size, -1, *out.shape[-2:])
                elif qkv_format == "sbhd":
                    # [s*b, np, hn] -> [s, b, np, hn]
                    out = out.view(-1, ctx.batch_size, *out.shape[-2:])
        elif not use_fused_attention:
1477
            out = out.view(-1, *out.shape[-2:])
1478

1479
1480
        if fp8 and use_fused_attention:
            amax_cp_fwd = amax_per_step.amax(dim=1)
1481
1482
            S_quantizer.amax.copy_(amax_cp_fwd[0])
            O_CP_quantizer.amax.copy_(amax_cp_fwd[1])
1483

1484
        out_fp8 = None
1485
        out_f16 = out.to(qkv_dtype)
1486

1487
        if fp8 and (is_output_fp8 or int(os.getenv("NVTE_FP8_DPA_BWD", "1"))):
1488
1489
1490
            out_fp8 = O_quantizer(out_f16)  # final result

        out_ret = out_fp8 if (fp8 and is_output_fp8) else out_f16
1491
1492

        if fp8 and int(os.getenv("NVTE_FP8_DPA_BWD", "1")):
1493
            q_save, kv_save, out_save = q, kv, out_fp8._data
1494
        elif fp8 and is_input_fp8:
1495
            q_save, kv_save, out_save = q, kv, out_f16
1496
        else:
1497
            q_f16 = q_f16.view(q.shape)
1498
1499
            q_save, kv_save, out_save = q_f16, kv, out_f16

1500
        tensors_to_save, tensor_objects = prepare_for_saving(
1501
1502
1503
            q_save,
            kv_save,
            out_save,
1504
            softmax_lse,
1505
1506
            cu_seqlens_q_padded,
            cu_seqlens_kv_padded,
1507
1508
            *cu_seqlens_q_per_step,
            *cu_seqlens_kv_per_step,
1509
1510
            *rng_states,
            *attn_biases,
1511
        )
1512
1513
1514
        ctx.save_for_backward(*tensors_to_save)
        ctx.tensor_objects = tensor_objects

1515
1516
1517
        ctx.cp_group_a2a = cp_group_a2a
        ctx.cp_size_a2a = cp_size_a2a
        ctx.rank_a2a = rank_a2a
1518
1519
        ctx.cp_group = cp_group
        ctx.cp_global_ranks = cp_global_ranks
1520
        ctx.cp_stream = cp_stream
1521
1522
        ctx.dropout_p = dropout_p
        ctx.max_seqlen_q = max_seqlen_q
1523
        ctx.max_seqlen_kv = max_seqlen_kv
1524
        ctx.softmax_scale = softmax_scale
1525
        ctx.qkv_format = qkv_format
1526
        ctx.attn_mask_type = attn_mask_type
1527
1528
        ctx.attn_bias_type = attn_bias_type
        ctx.attn_bias_shape = None if attn_bias is None else attn_bias.shape
1529
        ctx.deterministic = deterministic
1530
        ctx.use_fused_attention = use_fused_attention
1531
        ctx.softmax_lse_in_packed_format = softmax_lse_in_packed_format
1532
        ctx.second_half_lse_seqlen = second_half_lse_seqlen
1533
1534
        ctx.fp8 = fp8 and int(os.getenv("NVTE_FP8_DPA_BWD", "1"))
        ctx.fp8_meta = fp8_meta
1535
1536
        ctx.is_input_fp8 = is_input_fp8
        ctx.is_output_fp8 = is_output_fp8
1537
        ctx.use_flash_attn_3 = use_flash_attn_3
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553

        ctx.qkv_dtype = qkv_dtype
        ctx.dQKV_quantizer = dQKV_quantizer
        ctx.dQKV_CP_quantizer = dQKV_CP_quantizer
        ctx.dO_quantizer = dO_quantizer
        ctx.dP_quantizer = dP_quantizer
        ctx.QKV_quantizer = QKV_quantizer
        ctx.O_quantizer = O_quantizer
        ctx.S_quantizer = S_quantizer
        if ctx.fp8:
            ctx.QKV_quantizer = QKV_quantizer.copy()
            ctx.QKV_quantizer.scale = QKV_quantizer.scale.clone()
            ctx.O_quantizer = O_quantizer.copy()
            ctx.O_quantizer.scale = O_quantizer.scale.clone()
            ctx.S_quantizer = S_quantizer.copy()
            ctx.S_quantizer.scale = S_quantizer.scale.clone()
1554
        nvtx_range_pop("transformer_engine.AttnFuncWithCPAndKVP2P.forward")
1555

1556
        return out_ret
1557
1558
1559

    @staticmethod
    def backward(ctx, dout):
1560
        # pylint: disable=missing-function-docstring
1561
        nvtx_range_push("transformer_engine.AttnFuncWithCPAndKVP2P.backward")
1562
1563
1564
        cp_size_a2a = ctx.cp_size_a2a
        rank_a2a = ctx.rank_a2a

1565
1566
        cp_size = get_distributed_world_size(ctx.cp_group)
        rank = get_distributed_rank(ctx.cp_group)
1567
1568
        send_dst = ctx.cp_global_ranks[(rank - 1) % cp_size * cp_size_a2a + rank_a2a]
        recv_src = ctx.cp_global_ranks[(rank + 1) % cp_size * cp_size_a2a + rank_a2a]
1569
1570
        batch_p2p_comm = int(os.getenv("NVTE_BATCH_MHA_P2P_COMM", "0")) or (cp_size == 2)

1571
        q, kv, out, softmax_lse, cu_seqlens_q_padded, cu_seqlens_kv_padded, *other_tensors = (
1572
            restore_from_saved(ctx.tensor_objects, ctx.saved_tensors)
1573
1574
1575
1576
1577
        )
        cu_seqlens_q_per_step = other_tensors[:cp_size]
        cu_seqlens_kv_per_step = other_tensors[cp_size : cp_size * 2]
        rng_states = other_tensors[cp_size * 2 : cp_size * 3]
        attn_biases = other_tensors[cp_size * 3 : cp_size * 4]
1578

1579
1580
        causal = "causal" in ctx.attn_mask_type
        padding = "padding" in ctx.attn_mask_type
1581
1582

        seq_dim = None
1583
        if ctx.qkv_format in ["bshd", "sbhd"]:
1584
            seq_dim = ctx.qkv_format.index("s")
1585
1586
1587
            qkv_layout = ctx.qkv_format + "_" + ctx.qkv_format[:-2] + "2" + ctx.qkv_format[-2:]
        else:
            qkv_layout = ctx.qkv_format + "_" + ctx.qkv_format + "_" + ctx.qkv_format
1588

1589
        if attn_biases[0] is not None:
1590
1591
            # [b, np, sq, 2*cp, sk//(2*cp)]
            attn_dbias = torch.zeros(
1592
                *ctx.attn_bias_shape, dtype=attn_biases[0].dtype, device=attn_biases[0].device
1593
1594
1595
            )
            # [b, np, sq, 2*cp, sk//(2*cp)] -> [b, np, 2, sq//2, 2*cp, sk//(2*cp)]
            attn_dbias_ = attn_dbias.view(
1596
                *attn_dbias.shape[:-3], 2, attn_dbias.shape[-3] // 2, *attn_dbias.shape[-2:]
1597
1598
1599
            )
        else:
            attn_dbias = None
1600
            attn_dbias_ = None
1601

1602
1603
        softmax_lse_ = None
        if causal and ctx.second_half_lse_seqlen is not None:
1604
            if ctx.qkv_format == "thd":
1605
                softmax_lse_ = tex.thd_read_second_half_lse(
1606
1607
1608
1609
                    softmax_lse,
                    cu_seqlens_q_padded,
                    ctx.softmax_lse_in_packed_format,
                    ctx.second_half_lse_seqlen,
1610
                )
1611
1612
            else:
                # [b, np, sq] -> [b, np, 2, sq//2]
1613
                softmax_lse_ = softmax_lse.view(*softmax_lse.shape[:-1], 2, -1)
1614
                softmax_lse_ = softmax_lse_[..., 1, :].contiguous()
1615
1616
1617
1618
1619
1620
            if ctx.use_fused_attention:
                if ctx.softmax_lse_in_packed_format:
                    softmax_lse_ = softmax_lse_.transpose(0, 1).contiguous()
                # [b, np, sq//2] -> [b, np, sq//2, 1] or
                # [t//2, np] -> [t//2, np, 1]
                softmax_lse_.unsqueeze_(-1)
1621
        if ctx.use_fused_attention:
1622
1623
1624
1625
            if ctx.softmax_lse_in_packed_format:
                softmax_lse = softmax_lse.transpose(0, 1).contiguous()
            # [b, np, sq] -> [b, np, sq, 1] or
            # [t, np] -> [t, np, 1]
1626
            softmax_lse.unsqueeze_(-1)
1627
            dout = dout.contiguous()
1628

1629
        dq = None
1630
        dout_dtype = dout.dtype
1631
1632
        fused_attn_backend = None
        fused_attn_dqkv_dtype = None
1633
1634
1635
        amax_per_step = None
        dP_quantizer_per_step = [None for _ in range(cp_size)]
        dQKV_CP_quantizer_per_step = [None for _ in range(cp_size)]
1636
1637
1638
        if ctx.fp8:
            if ctx.use_fused_attention:
                fused_attn_backend = FusedAttnBackend["FP8"]
1639

1640
                if ctx.is_output_fp8:
1641
                    assert isinstance(dout, Float8Tensor), "dout must be Float8Tensors for FP8 MHA!"
1642
                    ctx.dO_quantizer = dout._quantizer
1643
                else:
1644
                    dout = ctx.dO_quantizer(dout)
1645
1646
1647
1648
1649
1650
                fused_attn_dqkv_dtype = TE_DType[dout._data.dtype]
                dq_fp8 = torch.empty((cp_size, *q.shape), dtype=dout._data.dtype, device=q.device)
                dkv_fp8 = torch.empty(
                    (cp_size, *kv.shape), dtype=dout._data.dtype, device=kv.device
                )
                dkv_fp8_ = torch.empty_like(dkv_fp8)
1651
                p2p_comm_buffers = [[kv, dkv_fp8], [torch.empty_like(kv), dkv_fp8_]]
1652
                dout = dout._data
1653
                fp8_meta_kwargs = {}
1654
                fp8_meta_kwargs["s_quantizer"] = ctx.S_quantizer
1655
1656
1657
                amax_per_step = torch.zeros((2, cp_size), dtype=torch.float32, device=q.device)
                for i in range(cp_size):
                    dP_quantizer_per_step[i] = ctx.dP_quantizer.copy()
1658
                    dP_quantizer_per_step[i].amax = amax_per_step[0][i].reshape((1,))
1659
                    dQKV_CP_quantizer_per_step[i] = ctx.dQKV_CP_quantizer.copy()
1660
                    dQKV_CP_quantizer_per_step[i].amax = amax_per_step[1][i].reshape((1,))
1661
1662
1663
            else:
                assert False, "FP8 is only supported with Fused Attention!"
        else:
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
            if ctx.fp8_meta is not None:
                if ctx.is_input_fp8:
                    q = ctx.QKV_quantizer.create_tensor_from_data(
                        q, fake_dtype=ctx.qkv_dtype, internal=True
                    )
                    kv = ctx.QKV_quantizer.create_tensor_from_data(
                        kv, fake_dtype=ctx.qkv_dtype, internal=True
                    )
                    q = q.dequantize(dtype=ctx.qkv_dtype)
                    kv = kv.dequantize(dtype=ctx.qkv_dtype)
                if ctx.is_output_fp8:
                    assert isinstance(dout, Float8Tensor), "dout must be Float8Tensors for FP8 MHA!"
                    if cp_size_a2a == 1:
                        dout = dout.dequantize(dtype=dout_dtype)
                    else:
                        ctx.dO_quantizer = dout._quantizer
                        dout = dout._data
1681
1682
1683
1684
1685
1686
1687
1688
            dq = torch.empty_like(q)
            p2p_comm_buffers = [
                torch.empty((2, *kv.shape), dtype=kv.dtype, device=kv.device),
                torch.empty((2, *kv.shape), dtype=kv.dtype, device=kv.device),
            ]
            p2p_comm_buffers[0][0].copy_(kv)
            if ctx.use_fused_attention:
                fp8_meta_kwargs = {}
1689
                fused_attn_dqkv_dtype = TE_DType[dout_dtype]
1690
1691
                fused_attn_backend = FusedAttnBackend["F16_arbitrary_seqlen"]

1692
1693
1694
1695
        if cp_size_a2a > 1:
            if not ctx.use_fused_attention:
                out = out.view(ctx.batch_size, -1, *out.shape[-2:])
                dout = dout.view(*out.shape)
1696
1697
1698
            chunk_ids_for_a2a = get_seq_chunk_ids_for_reordering_before_attn(
                cp_size_a2a, out.device
            )
1699
1700
1701
1702
1703
1704
1705
1706
1707
            out, dout = flash_attn_a2a_communicate(
                [out, dout],
                chunk_ids_for_a2a,
                seq_dim,
                cp_size_a2a,
                ctx.cp_group_a2a,
                ctx.cp_stream,
                True,
            )
1708
            if not ctx.fp8 and ctx.fp8_meta is not None and ctx.is_output_fp8:
1709
1710
1711
1712
                dout = ctx.dO_quantizer.create_tensor_from_data(
                    dout, fake_dtype=dout_dtype, internal=True
                )
                dout = dout.dequantize(dtype=dout_dtype)
1713

1714
1715
1716
1717
        out = out.view(*q.shape)
        dout = dout.view(*q.shape)
        send_recv_reqs = []

1718
        flash_attn_bwd = None
1719
1720
        if not ctx.use_fused_attention:
            fa_backward_kwargs = {"softmax_scale": ctx.softmax_scale}
1721
1722
1723
1724
            if ctx.use_flash_attn_3:
                flash_attn_bwd = (
                    _flash_attn_bwd_v3  # pylint: disable=possibly-used-before-assignment
                )
1725
1726
                fa_backward_kwargs["deterministic"] = ctx.deterministic
            else:
1727
1728
1729
1730
                if ctx.qkv_format == "thd":
                    flash_attn_bwd = _flash_attn_varlen_bwd
                else:
                    flash_attn_bwd = _flash_attn_bwd
1731
                fa_backward_kwargs["dropout_p"] = ctx.dropout_p
1732
                if fa_utils.v2_4_plus:
1733
                    fa_backward_kwargs["alibi_slopes"] = None
1734
                if fa_utils.v2_4_1_plus:
1735
                    fa_backward_kwargs["deterministic"] = ctx.deterministic
1736
                if fa_utils.v2_6_0_plus:
1737
                    fa_backward_kwargs["softcap"] = 0.0
1738

1739
1740
1741
1742
1743
        for i in range(cp_size):
            # wait until KV is received
            for req in send_recv_reqs:
                req.wait()

1744
1745
            send_tensor = p2p_comm_buffers[i % 2]
            recv_tensor = p2p_comm_buffers[(i + 1) % 2]
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
            if ctx.fp8:
                if i < cp_size - 1:
                    send_recv_reqs = flash_attn_p2p_communicate(
                        rank,
                        send_tensor[0],
                        send_dst,
                        recv_tensor[0],
                        recv_src,
                        ctx.cp_group,
                        batch_p2p_comm,
                    )
                else:
                    dkv_a2a_req = torch.distributed.all_to_all_single(
                        dkv_fp8,
                        dkv_fp8_,
                        group=ctx.cp_group,
                        async_op=True,
                    )
                    send_recv_reqs = [dkv_a2a_req]
            else:
                if i == 0:
                    send_tensor = send_tensor[0]
                    recv_tensor = recv_tensor[0]
                if i == (cp_size - 1):
                    send_tensor = send_tensor[1]
                    recv_tensor = recv_tensor[1]
                send_recv_reqs = flash_attn_p2p_communicate(
                    rank, send_tensor, send_dst, recv_tensor, recv_src, ctx.cp_group, batch_p2p_comm
                )
1775

1776
            kv = p2p_comm_buffers[i % 2][0]
1777
1778
            q_, kv_, out_, dout_ = None, None, None, None
            dq_, dk_, dv_ = None, None, None
1779
            # In reversed order of fwd
1780
            if causal:
1781
                if i == (cp_size - 1):
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
                    if ctx.qkv_format == "bshd":
                        # [b, 2, sq//2, np, hn] -> [b, sq, np, hn]
                        q_, out_, dout_ = [
                            x.view(x.shape[0], -1, *x.shape[-2:]) for x in [q, out, dout]
                        ]
                        # [b, 2, sk//2, 2, np, hn] -> [b, sk, 2, np, hn]
                        kv_ = kv.view(kv.shape[0], -1, *kv.shape[-3:])
                    elif ctx.qkv_format == "sbhd":
                        # [2, sq//2, b, np, hn] -> [sq, b, np, hn]
                        q_, out_, dout_ = [x.view(-1, *x.shape[-3:]) for x in [q, out, dout]]
                        # [2, sk//2, b, 2, np, hn] -> [sk, b, 2, np, hn]
                        kv_ = kv.view(-1, *kv.shape[-4:])
                    elif ctx.qkv_format == "thd":
                        q_, kv_, out_, dout_ = q, kv, out, dout
1796
                    if ctx.use_fused_attention:
1797
1798
1799
1800
1801
1802
1803
1804
                        if ctx.fp8:
                            aux_ctx_tensors = [
                                softmax_lse,
                                softmax_lse,
                                rng_states[cp_size - i - 1],
                            ]
                        else:
                            aux_ctx_tensors = [softmax_lse, rng_states[cp_size - i - 1]]
1805
                        if attn_dbias is not None:
1806
                            aux_ctx_tensors += [attn_biases[cp_size - i - 1]]
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
                        q_part = q_
                        k_part = kv_[..., 0, :, :] if ctx.qkv_format in ["bshd", "sbhd"] else kv_[0]
                        v_part = kv_[..., 1, :, :] if ctx.qkv_format in ["bshd", "sbhd"] else kv_[1]
                        out_part = out_
                        dout_part = dout_

                        if ctx.fp8:
                            q_part = ctx.QKV_quantizer.create_tensor_from_data(
                                q_part, fake_dtype=ctx.qkv_dtype, internal=True
                            )
                            k_part = ctx.QKV_quantizer.create_tensor_from_data(
                                k_part, fake_dtype=ctx.qkv_dtype, internal=True
                            )
                            v_part = ctx.QKV_quantizer.create_tensor_from_data(
                                v_part, fake_dtype=ctx.qkv_dtype, internal=True
                            )
                            out_part = ctx.O_quantizer.create_tensor_from_data(
                                out_part, fake_dtype=ctx.qkv_dtype, internal=True
                            )
                            dout_part = ctx.dO_quantizer.create_tensor_from_data(
1827
                                dout_part, fake_dtype=dout_dtype, internal=True
1828
                            )
1829
1830
                            fp8_meta_kwargs["dp_quantizer"] = dP_quantizer_per_step[i]
                            fp8_meta_kwargs["dqkv_quantizer"] = dQKV_CP_quantizer_per_step[i]
1831
                        dq_, dk_, dv_, dbias_ = fused_attn_bwd(
1832
                            ctx.max_seqlen_q,
1833
1834
1835
                            ctx.max_seqlen_kv,
                            cu_seqlens_q_per_step[cp_size - i - 1],
                            cu_seqlens_kv_per_step[cp_size - i - 1],
1836
1837
1838
1839
1840
                            q_part,
                            k_part,
                            v_part,
                            out_part,
                            dout_part,
1841
                            dout_dtype,
1842
                            fused_attn_dqkv_dtype,
1843
                            aux_ctx_tensors,
1844
                            fused_attn_backend,
1845
1846
                            cu_seqlens_q_padded=cu_seqlens_q_padded,
                            cu_seqlens_kv_padded=cu_seqlens_kv_padded,
1847
1848
                            attn_scale=ctx.softmax_scale,
                            dropout=ctx.dropout_p,
1849
                            qkv_layout=qkv_layout,
1850
                            attn_mask_type=ctx.attn_mask_type,
1851
                            attn_bias_type=ctx.attn_bias_type,
1852
1853
                            deterministic=ctx.deterministic,
                            **fp8_meta_kwargs,
1854
                        )
1855
1856
1857
1858
                        if ctx.fp8:
                            dq_ = dq_._data
                            dk_ = dk_._data
                            dv_ = dv_._data
1859
                    else:
1860
                        dq_ = torch.empty_like(q_)
1861
                        dkv_ = torch.empty_like(kv_)
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
                        fa_backward_args_thd = get_fa_args(
                            False,
                            ctx.use_flash_attn_3,
                            ctx.qkv_format,
                            cu_seqlens_q=cu_seqlens_q_per_step[cp_size - i - 1],
                            cu_seqlens_kv=cu_seqlens_kv_per_step[cp_size - i - 1],
                            max_seqlen_q=ctx.max_seqlen_q,
                            max_seqlen_kv=ctx.max_seqlen_kv,
                            dq=dq_,
                            dk=(
                                dkv_[..., 0, :, :]
                                if ctx.qkv_format in ["bshd", "sbhd"]
                                else dkv_[0]
                            ),
                            dv=(
                                dkv_[..., 1, :, :]
                                if ctx.qkv_format in ["bshd", "sbhd"]
                                else dkv_[1]
                            ),
                        )
                        if ctx.use_flash_attn_3 or (
                            fa_utils.v2_3_plus and not fa_utils.v2_7_0_plus
                        ):
1885
                            fa_backward_kwargs["window_size"] = (-1, 0)
1886
                        elif fa_utils.v2_7_0_plus:
1887
1888
                            fa_backward_kwargs["window_size_left"] = -1
                            fa_backward_kwargs["window_size_right"] = 0
1889
                        if not ctx.use_flash_attn_3:
1890
1891
                            fa_backward_kwargs["rng_state"] = rng_states[cp_size - i - 1]
                        flash_attn_bwd(
1892
1893
                            dout_,
                            q_,
1894
1895
                            kv_[..., 0, :, :] if ctx.qkv_format in ["bshd", "sbhd"] else kv_[0],
                            kv_[..., 1, :, :] if ctx.qkv_format in ["bshd", "sbhd"] else kv_[1],
1896
1897
                            out_,
                            softmax_lse,
1898
                            *fa_backward_args_thd,
1899
1900
                            causal=True,
                            **fa_backward_kwargs,
1901
                        )
1902
                elif i >= (cp_size - rank - 1):
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
                    if ctx.qkv_format == "bshd":
                        # [b, 2, sq//2, np, hn] -> [b, sq, np, hn]
                        q_, out_, dout_ = [
                            x.view(x.shape[0], -1, *x.shape[-2:]) for x in [q, out, dout]
                        ]
                        # [b, 2, sk//2, 2, np, hn] -> [b, sk//2, 2, np, hn]
                        kv_ = kv[:, 0]
                    elif ctx.qkv_format == "sbhd":
                        # [2, sq//2, b, np, hn] -> [sq, b, np, hn]
                        q_, out_, dout_ = [x.view(-1, *x.shape[-3:]) for x in [q, out, dout]]
                        # [2, sk//2, b, 2, np, hn] -> [sk//2, b, 2, np, hn]
                        kv_ = kv[0]
                    elif ctx.qkv_format == "thd":
                        q_, out_, dout_ = q, out, dout
                        # [2, t, np, hn] -> [2, t/2, np, hn]
                        kv_ = tex.thd_read_half_tensor(kv, cu_seqlens_kv_padded, 0)
1919
                    if ctx.use_fused_attention:
1920
                        kv_ = kv_.contiguous()
1921
1922
1923
1924
1925
1926
1927
1928
                        if ctx.fp8:
                            aux_ctx_tensors = [
                                softmax_lse,
                                softmax_lse,
                                rng_states[cp_size - i - 1],
                            ]
                        else:
                            aux_ctx_tensors = [softmax_lse, rng_states[cp_size - i - 1]]
1929
                        if attn_dbias is not None:
1930
                            aux_ctx_tensors += [attn_biases[cp_size - i - 1]]
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
                        q_part = q_
                        k_part = kv_[..., 0, :, :] if ctx.qkv_format in ["bshd", "sbhd"] else kv_[0]
                        v_part = kv_[..., 1, :, :] if ctx.qkv_format in ["bshd", "sbhd"] else kv_[1]
                        out_part = out_
                        dout_part = dout_

                        if ctx.fp8:
                            q_part = ctx.QKV_quantizer.create_tensor_from_data(
                                q_part, fake_dtype=ctx.qkv_dtype, internal=True
                            )
                            k_part = ctx.QKV_quantizer.create_tensor_from_data(
                                k_part, fake_dtype=ctx.qkv_dtype, internal=True
                            )
                            v_part = ctx.QKV_quantizer.create_tensor_from_data(
                                v_part, fake_dtype=ctx.qkv_dtype, internal=True
                            )
                            out_part = ctx.O_quantizer.create_tensor_from_data(
                                out_part, fake_dtype=ctx.qkv_dtype, internal=True
                            )
                            dout_part = ctx.dO_quantizer.create_tensor_from_data(
1951
                                dout_part, fake_dtype=dout_dtype, internal=True
1952
                            )
1953
1954
                            fp8_meta_kwargs["dp_quantizer"] = dP_quantizer_per_step[i]
                            fp8_meta_kwargs["dqkv_quantizer"] = dQKV_CP_quantizer_per_step[i]
1955
                        dq_, dk_, dv_, dbias_ = fused_attn_bwd(
1956
                            ctx.max_seqlen_q,
1957
1958
1959
                            ctx.max_seqlen_kv // 2,
                            cu_seqlens_q_per_step[cp_size - i - 1],
                            cu_seqlens_kv_per_step[cp_size - i - 1],
1960
1961
1962
1963
1964
                            q_part,
                            k_part,
                            v_part,
                            out_part,
                            dout_part,
1965
                            dout_dtype,
1966
                            fused_attn_dqkv_dtype,
1967
                            aux_ctx_tensors,
1968
                            fused_attn_backend,
1969
1970
1971
1972
                            cu_seqlens_q_padded=cu_seqlens_q_padded,
                            cu_seqlens_kv_padded=(
                                None if cu_seqlens_kv_padded is None else cu_seqlens_kv_padded // 2
                            ),
1973
1974
                            attn_scale=ctx.softmax_scale,
                            dropout=ctx.dropout_p,
1975
                            qkv_layout=qkv_layout,
1976
                            attn_mask_type="padding" if padding else "no_mask",
1977
                            attn_bias_type=ctx.attn_bias_type,
1978
1979
                            deterministic=ctx.deterministic,
                            **fp8_meta_kwargs,
1980
                        )
1981
1982
1983
1984
                        if ctx.fp8:
                            dq_ = dq_._data
                            dk_ = dk_._data
                            dv_ = dv_._data
1985
                    else:
1986
                        dq_ = torch.empty_like(q_)
1987
                        dkv_ = torch.empty_like(kv_)
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
                        fa_backward_args_thd = get_fa_args(
                            False,
                            ctx.use_flash_attn_3,
                            ctx.qkv_format,
                            cu_seqlens_q=cu_seqlens_q_per_step[cp_size - i - 1],
                            cu_seqlens_kv=cu_seqlens_kv_per_step[cp_size - i - 1],
                            max_seqlen_q=ctx.max_seqlen_q,
                            max_seqlen_kv=ctx.max_seqlen_kv // 2,
                            dq=dq_,
                            dk=(
                                dkv_[..., 0, :, :]
                                if ctx.qkv_format in ["bshd", "sbhd"]
                                else dkv_[0]
                            ),
                            dv=(
                                dkv_[..., 1, :, :]
                                if ctx.qkv_format in ["bshd", "sbhd"]
                                else dkv_[1]
                            ),
                        )
                        if ctx.use_flash_attn_3 or (
                            fa_utils.v2_3_plus and not fa_utils.v2_7_0_plus
                        ):
2011
                            fa_backward_kwargs["window_size"] = (-1, -1)
2012
                        elif fa_utils.v2_7_0_plus:
2013
2014
                            fa_backward_kwargs["window_size_left"] = -1
                            fa_backward_kwargs["window_size_right"] = -1
2015
                        if not ctx.use_flash_attn_3:
2016
2017
                            fa_backward_kwargs["rng_state"] = rng_states[cp_size - i - 1]
                        flash_attn_bwd(
2018
2019
                            dout_,
                            q_,
2020
2021
                            kv_[..., 0, :, :] if ctx.qkv_format in ["bshd", "sbhd"] else kv_[0],
                            kv_[..., 1, :, :] if ctx.qkv_format in ["bshd", "sbhd"] else kv_[1],
2022
2023
                            out_,
                            softmax_lse,
2024
                            *fa_backward_args_thd,
2025
2026
                            causal=False,
                            **fa_backward_kwargs,
2027
2028
                        )
                else:
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
                    if ctx.qkv_format == "bshd":
                        # [b, 2, sq//2, np, hn] -> [b, sq//2, np, hn]
                        q_, out_, dout_ = q[:, 1], out[:, 1], dout[:, 1]
                        # [b, 2, sk//2, 2, np, hn] -> [b, sk, 2, np, hn]
                        kv_ = kv.view(kv.shape[0], -1, *kv.shape[-3:])
                    elif ctx.qkv_format == "sbhd":
                        # [2, sq//2, b, np, hn] -> [sq//2, b, np, hn]
                        q_, out_, dout_ = q[1], out[1], dout[1]
                        # [2, sk//2, b, 2, np, hn] -> [sk, b, 2, np, hn]
                        kv_ = kv.view(-1, *kv.shape[-4:])
                    elif ctx.qkv_format == "thd":
                        # [t, np, hn] -> [t/2, np, hn]
                        q_, out_, dout_ = [
                            tex.thd_read_half_tensor(x, cu_seqlens_q_padded, 1)
                            for x in [q, out, dout]
                        ]
                        kv_ = kv
2046
                    if ctx.use_fused_attention:
2047
                        q_, out_, dout_ = [x.contiguous() for x in [q_, out_, dout_]]
2048
2049
2050
2051
2052
2053
2054
2055
                        if ctx.fp8:
                            aux_ctx_tensors = [
                                softmax_lse_,
                                softmax_lse_,
                                rng_states[cp_size - i - 1],
                            ]
                        else:
                            aux_ctx_tensors = [softmax_lse_, rng_states[cp_size - i - 1]]
2056
                        if attn_dbias is not None:
2057
                            aux_ctx_tensors += [attn_biases[cp_size - i - 1]]
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078

                        q_part = q_
                        k_part = kv_[..., 0, :, :] if ctx.qkv_format in ["bshd", "sbhd"] else kv_[0]
                        v_part = kv_[..., 1, :, :] if ctx.qkv_format in ["bshd", "sbhd"] else kv_[1]
                        out_part = out_
                        dout_part = dout_

                        if ctx.fp8:
                            q_part = ctx.QKV_quantizer.create_tensor_from_data(
                                q_part, fake_dtype=ctx.qkv_dtype, internal=True
                            )
                            k_part = ctx.QKV_quantizer.create_tensor_from_data(
                                k_part, fake_dtype=ctx.qkv_dtype, internal=True
                            )
                            v_part = ctx.QKV_quantizer.create_tensor_from_data(
                                v_part, fake_dtype=ctx.qkv_dtype, internal=True
                            )
                            out_part = ctx.O_quantizer.create_tensor_from_data(
                                out_part, fake_dtype=ctx.qkv_dtype, internal=True
                            )
                            dout_part = ctx.dO_quantizer.create_tensor_from_data(
2079
                                dout_part, fake_dtype=dout_dtype, internal=True
2080
                            )
2081
2082
                            fp8_meta_kwargs["dp_quantizer"] = dP_quantizer_per_step[i]
                            fp8_meta_kwargs["dqkv_quantizer"] = dQKV_CP_quantizer_per_step[i]
2083
                        dq_, dk_, dv_, dbias_ = fused_attn_bwd(
2084
                            ctx.max_seqlen_q // 2,
2085
2086
2087
                            ctx.max_seqlen_kv,
                            cu_seqlens_q_per_step[cp_size - i - 1],
                            cu_seqlens_kv_per_step[cp_size - i - 1],
2088
2089
2090
2091
2092
                            q_part,
                            k_part,
                            v_part,
                            out_part,
                            dout_part,
2093
                            dout_dtype,
2094
                            fused_attn_dqkv_dtype,
2095
                            aux_ctx_tensors,
2096
                            fused_attn_backend,
2097
2098
2099
2100
                            cu_seqlens_q_padded=(
                                None if cu_seqlens_q_padded is None else cu_seqlens_q_padded // 2
                            ),
                            cu_seqlens_kv_padded=cu_seqlens_kv_padded,
2101
2102
                            attn_scale=ctx.softmax_scale,
                            dropout=ctx.dropout_p,
2103
                            qkv_layout=qkv_layout,
2104
                            attn_mask_type="padding" if padding else "no_mask",
2105
                            attn_bias_type=ctx.attn_bias_type,
2106
2107
                            deterministic=ctx.deterministic,
                            **fp8_meta_kwargs,
2108
                        )
2109
2110
2111
2112
                        if ctx.fp8:
                            dq_ = dq_._data
                            dk_ = dk_._data
                            dv_ = dv_._data
2113
                    else:
2114
                        dq_ = torch.empty_like(q_)
2115
                        dkv_ = torch.empty_like(kv_)
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
                        fa_backward_args_thd = get_fa_args(
                            False,
                            ctx.use_flash_attn_3,
                            ctx.qkv_format,
                            cu_seqlens_q=cu_seqlens_q_per_step[cp_size - i - 1],
                            cu_seqlens_kv=cu_seqlens_kv_per_step[cp_size - i - 1],
                            max_seqlen_q=ctx.max_seqlen_q // 2,
                            max_seqlen_kv=ctx.max_seqlen_kv,
                            dq=dq_,
                            dk=(
                                dkv_[..., 0, :, :]
                                if ctx.qkv_format in ["bshd", "sbhd"]
                                else dkv_[0]
                            ),
                            dv=(
                                dkv_[..., 1, :, :]
                                if ctx.qkv_format in ["bshd", "sbhd"]
                                else dkv_[1]
                            ),
                        )
                        if ctx.use_flash_attn_3 or (
                            fa_utils.v2_3_plus and not fa_utils.v2_7_0_plus
                        ):
2139
                            fa_backward_kwargs["window_size"] = (-1, -1)
2140
                        elif fa_utils.v2_7_0_plus:
2141
2142
                            fa_backward_kwargs["window_size_left"] = -1
                            fa_backward_kwargs["window_size_right"] = -1
2143
                        if not ctx.use_flash_attn_3:
2144
2145
                            fa_backward_kwargs["rng_state"] = rng_states[cp_size - i - 1]
                        flash_attn_bwd(
2146
2147
                            dout_,
                            q_,
2148
2149
                            kv_[..., 0, :, :] if ctx.qkv_format in ["bshd", "sbhd"] else kv_[0],
                            kv_[..., 1, :, :] if ctx.qkv_format in ["bshd", "sbhd"] else kv_[1],
2150
2151
                            out_,
                            softmax_lse_,
2152
                            *fa_backward_args_thd,
2153
2154
                            causal=False,
                            **fa_backward_kwargs,
2155
2156
2157
                        )
            else:
                if ctx.use_fused_attention:
2158
2159
2160
2161
                    if ctx.fp8:
                        aux_ctx_tensors = [softmax_lse, softmax_lse, rng_states[cp_size - i - 1]]
                    else:
                        aux_ctx_tensors = [softmax_lse, rng_states[cp_size - i - 1]]
2162
                    if attn_dbias is not None:
2163
                        aux_ctx_tensors += [attn_biases[cp_size - i - 1]]
2164
2165
2166
2167
2168
2169
2170
2171
                    q_part = q
                    k_part = kv[..., 0, :, :] if ctx.qkv_format in ["bshd", "sbhd"] else kv[0]
                    v_part = kv[..., 1, :, :] if ctx.qkv_format in ["bshd", "sbhd"] else kv[1]
                    out_part = out
                    dout_part = dout

                    if ctx.fp8:
                        q_part = ctx.QKV_quantizer.create_tensor_from_data(
2172
                            q_part, fake_dtype=ctx.qkv_dtype, internal=True
2173
2174
                        )
                        k_part = ctx.QKV_quantizer.create_tensor_from_data(
2175
                            k_part, fake_dtype=ctx.qkv_dtype, internal=True
2176
2177
                        )
                        v_part = ctx.QKV_quantizer.create_tensor_from_data(
2178
                            v_part, fake_dtype=ctx.qkv_dtype, internal=True
2179
2180
                        )
                        out_part = ctx.O_quantizer.create_tensor_from_data(
2181
                            out_part, fake_dtype=ctx.qkv_dtype, internal=True
2182
2183
                        )
                        dout_part = ctx.dO_quantizer.create_tensor_from_data(
2184
                            dout_part, fake_dtype=dout_dtype, internal=True
2185
                        )
2186
2187
                        fp8_meta_kwargs["dp_quantizer"] = dP_quantizer_per_step[i]
                        fp8_meta_kwargs["dqkv_quantizer"] = dQKV_CP_quantizer_per_step[i]
2188
                    dq_, dk_, dv_, dbias_ = fused_attn_bwd(
2189
                        ctx.max_seqlen_q,
2190
2191
2192
                        ctx.max_seqlen_kv,
                        cu_seqlens_q_per_step[cp_size - i - 1],
                        cu_seqlens_kv_per_step[cp_size - i - 1],
2193
2194
2195
2196
2197
                        q_part,
                        k_part,
                        v_part,
                        out_part,
                        dout_part,
2198
                        dout_dtype,
2199
                        fused_attn_dqkv_dtype,
2200
                        aux_ctx_tensors,
2201
                        fused_attn_backend,
2202
2203
                        cu_seqlens_q_padded=cu_seqlens_q_padded,
                        cu_seqlens_kv_padded=cu_seqlens_kv_padded,
2204
2205
                        attn_scale=ctx.softmax_scale,
                        dropout=ctx.dropout_p,
2206
                        qkv_layout=qkv_layout,
2207
                        attn_mask_type=ctx.attn_mask_type,
2208
                        attn_bias_type=ctx.attn_bias_type,
2209
2210
                        deterministic=ctx.deterministic,
                        **fp8_meta_kwargs,
2211
                    )
2212
2213
2214
2215
2216
2217

                    if ctx.fp8:
                        dq_ = dq_._data
                        dk_ = dk_._data
                        dv_ = dv_._data

2218
                else:
2219
2220
                    dq_ = torch.empty_like(q)
                    dkv_ = torch.empty_like(kv)
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
2233
                    fa_backward_args_thd = get_fa_args(
                        False,
                        ctx.use_flash_attn_3,
                        ctx.qkv_format,
                        cu_seqlens_q=cu_seqlens_q_per_step[cp_size - i - 1],
                        cu_seqlens_kv=cu_seqlens_kv_per_step[cp_size - i - 1],
                        max_seqlen_q=ctx.max_seqlen_q,
                        max_seqlen_kv=ctx.max_seqlen_kv,
                        dq=dq_,
                        dk=dkv_[..., 0, :, :] if ctx.qkv_format in ["bshd", "sbhd"] else dkv_[0],
                        dv=dkv_[..., 1, :, :] if ctx.qkv_format in ["bshd", "sbhd"] else dkv_[1],
                    )
                    if ctx.use_flash_attn_3 or (fa_utils.v2_3_plus and not fa_utils.v2_7_0_plus):
2234
                        fa_backward_kwargs["window_size"] = (-1, -1)
2235
                    elif fa_utils.v2_7_0_plus:
2236
2237
                        fa_backward_kwargs["window_size_left"] = -1
                        fa_backward_kwargs["window_size_right"] = -1
2238
                    if not ctx.use_flash_attn_3:
2239
2240
                        fa_backward_kwargs["rng_state"] = rng_states[cp_size - i - 1]
                    flash_attn_bwd(
2241
2242
2243
2244
2245
                        dout,
                        q,
                        kv[..., 0, :, :] if ctx.qkv_format in ["bshd", "sbhd"] else kv[0],
                        kv[..., 1, :, :] if ctx.qkv_format in ["bshd", "sbhd"] else kv[1],
                        out,
2246
                        softmax_lse,
2247
                        *fa_backward_args_thd,
2248
2249
                        causal=False,
                        **fa_backward_kwargs,
2250
2251
                    )

2252
2253
            if ctx.fp8:
                dq = dq_fp8[(rank + i + 1) % cp_size]
2254
2255
2256
            if causal and ctx.qkv_format in ["bshd", "sbhd"] and i >= (cp_size - rank - 1):
                # [b, sq, np, hn] -> [b, 2, sq//2, np, hn] or
                # [sq, b, np, hn] -> [2, sq//2, b, np, hn]
2257
                dq_ = dq_.view(*dq.shape)
2258

2259
2260
2261
2262
2263
2264
2265
2266
2267
2268
2269
            if ctx.fp8:
                if i >= (cp_size - rank - 1) or not causal:
                    dq.copy_(dq_)
                else:
                    if ctx.qkv_format == "bshd":
                        dq[:, 0, ...].fill_(0)
                        dq[:, 1, ...].copy_(dq_)
                    elif ctx.qkv_format == "sbhd":
                        dq[0].fill_(0)
                        dq[1].copy_(dq_)
            elif causal:
2270
                if i > (cp_size - rank - 1):
2271
                    dq.add_(dq_)
2272
2273
                elif i == (cp_size - rank - 1):
                    if rank == (cp_size - 1):
2274
2275
                        dq.copy_(dq_)
                    else:
2276
2277
2278
2279
2280
2281
                        if ctx.qkv_format == "bshd":
                            dq[:, 0, ...].copy_(dq_[:, 0, ...])
                            dq[:, 1, ...].add_(dq_[:, 1, ...])
                        elif ctx.qkv_format == "sbhd":
                            dq[0].copy_(dq_[0])
                            dq[1].add_(dq_[1])
2282
                        elif ctx.qkv_format == "thd":
2283
                            tex.thd_grad_correction(dq, dq_, cu_seqlens_q_padded, "copy", "add")
2284
                elif i > 0:
2285
2286
2287
2288
                    if ctx.qkv_format == "bshd":
                        dq[:, 1, ...].add_(dq_)
                    elif ctx.qkv_format == "sbhd":
                        dq[1].add_(dq_)
2289
                    elif ctx.qkv_format == "thd":
2290
                        tex.thd_grad_correction(dq, dq_, cu_seqlens_q_padded, "none", "add")
2291
                else:
2292
2293
2294
2295
                    if ctx.qkv_format == "bshd":
                        dq[:, 1, ...].copy_(dq_)
                    elif ctx.qkv_format == "sbhd":
                        dq[1].copy_(dq_)
2296
                    elif ctx.qkv_format == "thd":
2297
                        tex.thd_grad_correction(dq, dq_, cu_seqlens_q_padded, "none", "copy")
2298
2299
2300
2301
2302
            else:
                if i == 0:
                    dq.copy_(dq_)
                else:
                    dq.add_(dq_)
2303

2304
            if attn_dbias is not None:
2305
                idx = (rank + i + 1) % cp_size
2306
                if i == (cp_size - 1) or not causal:
2307
                    # [b, np, sq, sk//cp] -> [b, np, sq, 2, sk//(2*cp)]
2308
                    dbias_ = dbias_.view(*dbias_.shape[:-1], 2, dbias_.shape[-1] // 2)
2309
                    attn_dbias[..., idx, :].copy_(dbias_[..., 0, :])
2310
2311
                    attn_dbias[..., (2 * cp_size - idx - 1), :].copy_(dbias_[..., 1, :])
                elif i >= (cp_size - rank - 1):
2312
2313
2314
2315
                    # [b, np, sq, sk//(2*cp)]
                    attn_dbias[..., idx, :].copy_(dbias_)
                else:
                    # [b, np, sq//2, sk//cp] -> [b, np, sq//2, 2, sk//(2*cp)]
2316
                    dbias_ = dbias_.view(*dbias_.shape[:-1], 2, dbias_.shape[-1] // 2)
2317
                    attn_dbias_[..., 1, :, idx, :].copy_(dbias_[..., 0, :])
2318
                    attn_dbias_[..., 1, :, (2 * cp_size - idx - 1), :].copy_(dbias_[..., 1, :])
2319

2320
2321
2322
            # wait until dKV is received
            for req in send_recv_reqs:
                req.wait()
2323

2324
2325
2326
2327
2328
2329
2330
            if ctx.fp8:
                if i < cp_size - 1:
                    dkv = dkv_fp8_[(rank + i + 1) % cp_size]
                else:
                    dkv = dkv_fp8[(rank + i + 1) % cp_size]
            else:
                dkv = p2p_comm_buffers[(i + 1) % 2][1]
2331
            if ctx.use_fused_attention:
2332
                if ctx.qkv_format in ["bshd", "sbhd"]:
2333
2334
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
                    dkv_ = _combine_tensors([dk_, dv_], -2)
                elif ctx.qkv_format == "thd":
                    dkv_ = torch.cat(
                        (dk_.unsqueeze(0), dv_.unsqueeze(0)), dim=0
                    )  # pylint: disable=used-before-assignment
            if ctx.qkv_format in ["bshd", "sbhd"]:
                # [b, 2, sk//2, 2, np, hn] -> [2, b, 2, sk//2, np, hn] or
                # [2, sk//2, b, 2, np, hn] -> [2, 2, sk//2, b, np, hn]
                dkv = dkv.view(2, *dkv.shape[0:-3], *dkv.shape[-2:])
                dkv_ = dkv_.movedim(-3, 0)
                if causal and (i < (cp_size - rank - 1) or i == (cp_size - 1)):
                    # [2, b, sk, np, hn] -> [2, b, 2, sk//2, np, hn] or
                    # [2, sk, b, np, hn] -> [2, 2, sk//2, b, np, hn]
                    dkv_ = dkv_.view(*dkv.shape)
2347

2348
2349
2350
2351
2352
2353
2354
2355
2356
2357
2358
            if ctx.fp8:
                if causal and i >= (cp_size - rank - 1) and i != (cp_size - 1):
                    if ctx.qkv_format == "bshd":
                        dkv[:, :, 0, ...].copy_(dkv_)
                        dkv[:, :, 1, ...].fill_(0)
                    elif ctx.qkv_format == "sbhd":
                        dkv[:, 0, ...].copy_(dkv_)
                        dkv[:, 1, ...].fill_(0)
                else:
                    dkv.copy_(dkv_)
            elif causal:
2359
                if i == (cp_size - 1):
2360
                    if rank == 0:
2361
2362
2363
2364
2365
2366
                        if ctx.qkv_format == "bshd":
                            dkv[:, :, 0, ...].add_(dkv_[:, :, 0, ...])
                            dkv[:, :, 1, ...].copy_(dkv_[:, :, 1, ...])
                        elif ctx.qkv_format == "sbhd":
                            dkv[:, 0, ...].add_(dkv_[:, 0, ...])
                            dkv[:, 1, ...].copy_(dkv_[:, 1, ...])
2367
                        elif ctx.qkv_format == "thd":
2368
                            tex.thd_grad_correction(dkv, dkv_, cu_seqlens_kv_padded, "add", "copy")
2369
2370
                    else:
                        dkv.add_(dkv_)
2371
2372
                elif i >= (cp_size - rank - 1):
                    if i == 0 and rank == (cp_size - 1):
2373
2374
2375
2376
                        if ctx.qkv_format == "bshd":
                            dkv[:, :, 0, ...].copy_(dkv_)
                        elif ctx.qkv_format == "sbhd":
                            dkv[:, 0, ...].copy_(dkv_)
2377
                        elif ctx.qkv_format == "thd":
2378
                            tex.thd_grad_correction(dkv, dkv_, cu_seqlens_kv_padded, "copy", "none")
2379
                    else:
2380
2381
2382
2383
                        if ctx.qkv_format == "bshd":
                            dkv[:, :, 0, ...].add_(dkv_)
                        elif ctx.qkv_format == "sbhd":
                            dkv[:, 0, ...].add_(dkv_)
2384
                        elif ctx.qkv_format == "thd":
2385
                            tex.thd_grad_correction(dkv, dkv_, cu_seqlens_kv_padded, "add", "none")
2386
2387
2388
2389
2390
                elif i > 0:
                    dkv.add_(dkv_)
                else:
                    dkv.copy_(dkv_)
            else:
2391
2392
2393
2394
2395
                if i == 0:
                    dkv.copy_(dkv_)
                else:
                    dkv.add_(dkv_)

2396
        if ctx.fp8 and ctx.use_fused_attention:
2397
            amax_cp_bwd = amax_per_step.amax(dim=1)
2398
2399
            ctx.dP_quantizer.amax.copy_(amax_cp_bwd[0])
            ctx.dQKV_CP_quantizer.amax.copy_(amax_cp_bwd[1])
2400
2401
2402
2403
            if ctx.qkv_format in ["bshd", "sbhd"]:
                # [cp, b, 2, sk//2, 2, np, hn] -> [cp, 2, b, 2, sk//2, np, hn] or
                # [cp, 2, sk//2, b, 2, np, hn] -> [cp, 2, 2, sk//2, b, np, hn]
                dkv_fp8 = dkv_fp8.view(cp_size, 2, *dkv_fp8.shape[1:-3], *dkv_fp8.shape[-2:])
2404
2405
2406
2407
2408
2409
2410
            dq = ctx.dQKV_CP_quantizer.create_tensor_from_data(
                dq_fp8, fake_dtype=torch.float32, internal=True
            )
            dkv = ctx.dQKV_CP_quantizer.create_tensor_from_data(
                dkv_fp8, fake_dtype=torch.float32, internal=True
            )
            dq, dkv = [x.dequantize(dtype=torch.float32) for x in [dq, dkv]]
2411
2412
            dq, dkv = [x.sum(dim=0).to(dout_dtype) for x in [dq, dkv]]

2413
        if causal:
2414
2415
            if ctx.qkv_format == "bshd":
                # [b, 2, sq//2, np, hn] -> [b, sq, np, hn]
2416
                dq = dq.view(dq.shape[0], -1, *dq.shape[-2:])
2417
                # [2, b, 2, sk//2, np, hn] -> [2, b, sk, np, hn]
2418
                dkv = dkv.view(*dkv.shape[0:2], -1, *dkv.shape[-2:])
2419
2420
            elif ctx.qkv_format == "sbhd":
                # [2, sq//2, b, np, hn] -> [sq, b, np, hn]
2421
                dq = dq.view(-1, *dq.shape[-3:])
2422
                # [2, 2, sk//2, b, np, hn] -> [2, sk, b, np, hn]
2423
2424
                dkv = dkv.view(dkv.shape[0], -1, *dkv.shape[-3:])

2425
2426
2427
        if ctx.qkv_format == "thd" and not ctx.use_fused_attention:
            dq[cu_seqlens_q_padded[-1] :].fill_(0)
            dkv[:, cu_seqlens_kv_padded[-1] :].fill_(0)
2428

2429
        if ctx.fp8 and ctx.is_input_fp8:
2430
2431
            assert torch.uint8 not in [dq.dtype, dkv.dtype]
            dq, dkv = [ctx.dQKV_quantizer(x)._data for x in [dq, dkv]]
2432
2433
2434
        dk, dv = dkv[0], dkv[1]

        if cp_size_a2a > 1:
2435
            chunk_ids_for_a2a = get_seq_chunk_ids_for_reordering_after_attn(cp_size_a2a, q.device)
2436
2437
2438
2439
2440
2441
2442
2443
2444
2445
2446
2447
2448
2449
            dq, dk, dv = flash_attn_a2a_communicate(
                [dq, dk, dv],
                chunk_ids_for_a2a,
                seq_dim,
                cp_size_a2a,
                ctx.cp_group_a2a,
                ctx.cp_stream,
                False,
            )
            if ctx.qkv_format == "bshd":
                dq, dk, dv = [x.view(ctx.batch_size, -1, *x.shape[-2:]) for x in [dq, dk, dv]]
            elif ctx.qkv_format == "sbhd":
                dq, dk, dv = [x.view(-1, ctx.batch_size, *x.shape[-2:]) for x in [dq, dk, dv]]

2450
2451
2452
        if attn_dbias is not None:
            # [b, np, sq, 2*cp, sk//(2*cp)] -> [b, np, sq, sk]
            attn_dbias = attn_dbias.view(*attn_dbias.shape[:-2], -1)
2453
2454
        # converting torch.uint8 to float8tensor
        if ctx.fp8 and ctx.is_input_fp8:
2455
2456
2457
            dq = ctx.dQKV_quantizer.create_tensor_from_data(dq, fake_dtype=dout_dtype)
            dk = ctx.dQKV_quantizer.create_tensor_from_data(dk, fake_dtype=dout_dtype)
            dv = ctx.dQKV_quantizer.create_tensor_from_data(dv, fake_dtype=dout_dtype)
2458
        nvtx_range_pop("transformer_engine.AttnFuncWithCPAndKVP2P.backward")
2459

2460
2461
2462
        return (
            None,
            dq,
2463
2464
            dk,
            dv,
2465
2466
2467
2468
2469
2470
2471
2472
2473
2474
2475
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
2476
            attn_dbias,
2477
2478
2479
2480
2481
            None,
            None,
            None,
            None,
            None,
2482
2483
            None,
            None,
2484
            None,
2485
            None,
2486
            None,
2487
        )
2488
2489


2490
2491
def get_kv_seq_info_after_all_gather(
    local_chunk_id, cp_size, max_seqlen_q, max_seqlen_kv, window_size, causal
2492
):
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503
2504
2505
2506
2507
2508
2509
2510
2511
2512
2513
2514
    """Compute KV sequence index range and update window size after all-gather."""
    local_chunk_end_idx = (local_chunk_id + 1) * max_seqlen_kv
    full_seq_end_idx = max_seqlen_kv * cp_size * 2

    if window_size is None:
        window_size = (-1, 0) if causal else (-1, -1)

    if window_size[1] == -1:
        seq_end_idx = full_seq_end_idx
        window_size_right = -1
    else:
        seq_end_idx = min(full_seq_end_idx, local_chunk_end_idx + window_size[1])
        window_size_right = local_chunk_end_idx + window_size[1] - seq_end_idx

    if window_size[0] == -1:
        seq_start_idx = 0
        window_size_left = -1
    else:
        seq_start_idx = max(0, local_chunk_end_idx - max_seqlen_q - window_size[0])
        window_size_left = window_size[0] + seq_end_idx - local_chunk_end_idx

    return (seq_start_idx, seq_end_idx), (window_size_left, window_size_right)
2515
2516
2517
2518


class AttnFuncWithCPAndKVAllGather(torch.autograd.Function):
    """
2519
2520
    Attention implementation with context parallelism. KV all-gather between CP ranks is exposed.
    Refer section 3.3.2 of `The Llama 3 Herd of Models <https://arxiv.org/abs/2407.21783>`_.
2521
2522
2523
2524
2525
2526
2527
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
    """

    @staticmethod
    def forward(
        ctx,
        is_training,
        q,
        k,
        v,
        cu_seqlens_q,
        max_seqlen_q,
        max_seqlen_kv,
        cu_seqlens_q_padded,
        dropout_p,
        softmax_scale,
        qkv_format,
        attn_mask_type,
        attn_bias_type,
        attn_bias,
        deterministic,
        use_fused_attention,
        window_size,
2543
2544
        cp_group,
        cp_stream,
2545
        use_flash_attn_3,
2546
    ):
2547
        # pylint: disable=missing-function-docstring
2548
        nvtx_range_push("transformer_engine.AttnFuncWithCPAndKVAllGather.forward")
2549
2550
2551
2552
2553
2554
        if softmax_scale is None:
            softmax_scale = q.shape[-1] ** (-0.5)

        cp_size = get_distributed_world_size(cp_group)
        rank = get_distributed_rank(cp_group)

2555
2556
        qkv_dtype = q.dtype

2557
2558
        causal = "causal" in attn_mask_type
        padding = "padding" in attn_mask_type
2559
        assert not padding, f"{attn_mask_type} mask type is not supported!"
2560
2561
2562
2563
2564
        if use_fused_attention and causal and "bottom_right" not in attn_mask_type:
            attn_mask_type = attn_mask_type + "_bottom_right"
        assert attn_bias_type == "no_bias", f"{attn_bias_type} bias type is not supported!"
        assert q.shape[-1] % 8 == 0, "Hidden size per attention head should be multiple of 8!"
        assert (
2565
            use_fused_attention or fa_utils.v2_3_plus
2566
        ), "Sliding window attention only can work with FusedAttention or FlashAttention >= 2.3!"
2567

2568
        flash_attn_fwd = None
2569
2570
        if not use_fused_attention:
            fa_forward_kwargs = {"softmax_scale": softmax_scale}
2571
2572
            if use_flash_attn_3:
                flash_attn_fwd = _flash_attn_fwd_v3
2573
            else:
2574
2575
2576
2577
                if qkv_format == "thd":
                    flash_attn_fwd = _flash_attn_varlen_fwd
                else:
                    flash_attn_fwd = _flash_attn_fwd
2578
2579
                fa_forward_kwargs["dropout_p"] = dropout_p
                fa_forward_kwargs["return_softmax"] = False
2580
                if fa_utils.v2_4_plus:
2581
                    fa_forward_kwargs["alibi_slopes"] = None
2582
                if fa_utils.v2_5_7_plus and qkv_format == "thd":
2583
                    fa_forward_kwargs["block_table"] = None
2584
                if fa_utils.v2_6_0_plus:
2585
                    fa_forward_kwargs["softcap"] = 0.0
2586
2587
2588
2589
2590
2591
2592
2593
2594
2595
2596

        assert qkv_format != "thd", f"{qkv_format} format is not supported!"
        qkv_layout = qkv_format + "_" + qkv_format + "_" + qkv_format

        seq_dim = qkv_format.index("s")
        assert (
            q.shape[seq_dim] % 2 == 0 and k.shape[seq_dim] % 2 == 0
        ), "Sequence length per GPU needs to be divisible by 2!"

        max_seqlen_q = max_seqlen_q // (2 * cp_size)
        max_seqlen_kv = max_seqlen_kv // (2 * cp_size)
2597
2598
        if use_fused_attention or qkv_format == "thd":
            cu_seqlens_q = cu_seqlens_q // (2 * cp_size)
2599
2600
2601
2602
        if cu_seqlens_q_padded is not None and qkv_format == "thd":
            cu_seqlens_q_padded = cu_seqlens_q_padded // (2 * cp_size)
        else:
            cu_seqlens_q_padded = None
2603

2604
2605
2606
2607
        # [b, s, np, hn] -> [b, 2, s//2, np, hn] or [s, b, np, hn] -> [2, s//2, b, np, hn]
        q = q.view(*q.shape[:seq_dim], 2, q.shape[seq_dim] // 2, *q.shape[(seq_dim + 1) :])
        # [b, s, np, hn] or [s, b, np, hn] -> [s, b, np, hn]
        k, v = [x.movedim(seq_dim, 0).contiguous() for x in [k, v]]
2608

2609
        # [s, b, np, hn] -> [cp, s, b, np, hn]
2610
2611
        k_ag, _ = gather_along_first_dim(k, cp_group)
        v_ag, _ = gather_along_first_dim(v, cp_group)
2612
2613

        # [cp, s, b, np, hn] -> [cp*2, s//2, b, np, hn]
2614
2615
        k_ag = k_ag.view(2 * cp_size, k.shape[0] // 2, *k.shape[1:])
        v_ag = v_ag.view(2 * cp_size, v.shape[0] // 2, *v.shape[1:])
2616
        chunk_ids_for_kv_ag = get_seq_chunk_ids_for_reordering_before_attn(cp_size, k.device)
2617
2618
2619
2620
2621
2622
2623
2624
2625
        k_ag = torch.index_select(k_ag, dim=0, index=chunk_ids_for_kv_ag)
        v_ag = torch.index_select(v_ag, dim=0, index=chunk_ids_for_kv_ag)
        # [cp*2, s//2, b, np, hn] -> [cp*s, b, np, hn]
        k_ag = k_ag.view(-1, *k.shape[1:])
        v_ag = v_ag.view(-1, *v.shape[1:])
        cp_stream.wait_stream(torch.cuda.current_stream())

        # create two streams to resolve wave quantization issue of Flash Attn in each step
        flash_attn_streams = [torch.cuda.current_stream(), cp_stream]
2626
2627

        local_seq_chunk_ids = [rank, 2 * cp_size - rank - 1]
2628
2629
2630
        kv_seq_range_per_step = [None, None]
        window_size_per_step = [None, None]
        cu_seqlens_kv_per_step = [None, None]
2631
2632
2633
2634
2635
2636
2637
2638
        out_per_step = [None, None]
        softmax_lse_per_step = [None, None]
        rng_states = [None, None]
        out = torch.empty_like(q)

        for i in range(len(local_seq_chunk_ids) + 1):
            if i < len(local_seq_chunk_ids):
                with torch.cuda.stream(flash_attn_streams[i]):
2639
2640
                    # [b, 2, sq//2, np, hn] -> [b, sq//2, np, hn]
                    # or [2, sq//2, b, np, hn] -> [sq//2, b, np, hn]
2641
2642
2643
2644
2645
2646
2647
2648
2649
                    q_ = q.select(seq_dim, i).contiguous()
                    kv_seq_range_per_step[i], window_size_per_step[i] = (
                        get_kv_seq_info_after_all_gather(
                            local_seq_chunk_ids[i],
                            cp_size,
                            max_seqlen_q,
                            max_seqlen_kv,
                            window_size,
                            causal,
2650
                        )
2651
2652
2653
2654
2655
2656
                    )
                    seq_start_idx, seq_end_idx = (
                        kv_seq_range_per_step[i][0],
                        kv_seq_range_per_step[i][1],
                    )
                    max_seqlen_kv_ = seq_end_idx - seq_start_idx
2657
                    if use_fused_attention or qkv_format == "thd":
2658
                        cu_seqlens_kv_per_step[i] = dpa_utils.get_full_cu_seqlens(
2659
2660
                            k.shape[1], max_seqlen_kv_, k.device
                        )
2661
2662
2663
                    k_, v_ = [x[seq_start_idx:seq_end_idx] for x in [k_ag, v_ag]]
                    # [s_range, b, np, hn] -> [b, s_range, np, hn] or [s_range, b, np, hn]
                    k_, v_ = [x.movedim(0, seq_dim).contiguous() for x in [k_, v_]]
2664
2665
2666
2667
                    if use_fused_attention:
                        out_per_step[i], [softmax_lse_per_step[i], rng_states[i]] = fused_attn_fwd(
                            is_training,
                            max_seqlen_q,
2668
                            max_seqlen_kv_,
2669
                            cu_seqlens_q,
2670
                            cu_seqlens_kv_per_step[i],
2671
2672
2673
                            q_,
                            k_,
                            v_,
2674
                            qkv_dtype,
2675
2676
2677
2678
2679
2680
2681
2682
                            tex.NVTE_Fused_Attn_Backend.NVTE_F16_arbitrary_seqlen,
                            attn_scale=softmax_scale,
                            dropout=dropout_p,
                            qkv_layout=qkv_layout,
                            attn_mask_type=attn_mask_type,
                            attn_bias_type=attn_bias_type,
                            attn_bias=attn_bias,
                            cu_seqlens_q_padded=cu_seqlens_q_padded,
2683
2684
                            cu_seqlens_kv_padded=cu_seqlens_kv_per_step[i],
                            window_size=window_size_per_step[i],
2685
2686
                        )
                    else:
2687
2688
2689
2690
2691
2692
2693
2694
2695
2696
                        fa_forward_args_thd = get_fa_args(
                            True,
                            use_flash_attn_3,
                            qkv_format,
                            cu_seqlens_q=cu_seqlens_q,
                            cu_seqlens_kv=cu_seqlens_kv_per_step[i],
                            max_seqlen_q=max_seqlen_q,
                            max_seqlen_kv=max_seqlen_kv_,
                        )
                        if use_flash_attn_3 or (fa_utils.v2_3_plus and not fa_utils.v2_7_0_plus):
2697
                            fa_forward_kwargs["window_size"] = window_size_per_step[i]
2698
                        elif fa_utils.v2_7_0_plus:
2699
2700
                            fa_forward_kwargs["window_size_left"] = window_size_per_step[i][0]
                            fa_forward_kwargs["window_size_right"] = window_size_per_step[i][1]
2701
2702
2703
2704
                        fa_outputs = flash_attn_fwd(
                            q_,
                            k_,
                            v_,
2705
                            *fa_forward_args_thd,
2706
2707
                            causal=causal,
                            **fa_forward_kwargs,
2708
                        )
2709
                        if not fa_utils.v2_7_0_plus:
2710
2711
                            out_per_step[i] = fa_outputs[4]
                            softmax_lse_per_step[i] = fa_outputs[5]
2712
                            if not use_flash_attn_3:
2713
2714
2715
2716
                                rng_states[i] = fa_outputs[7]
                        else:
                            out_per_step[i] = fa_outputs[0]
                            softmax_lse_per_step[i] = fa_outputs[1]
2717
                            if not use_flash_attn_3:
2718
                                rng_states[i] = fa_outputs[3]
2719
2720
2721
2722

            if i > 0:
                with torch.cuda.stream(flash_attn_streams[i - 1]):
                    if qkv_format == "bshd":
2723
                        out[:, i - 1].copy_(out_per_step[i - 1])
2724
                    elif qkv_format == "sbhd":
2725
                        out[i - 1].copy_(out_per_step[i - 1])
2726
2727
2728
2729
2730
2731
2732
2733
2734
2735
2736
2737
2738
2739
2740
2741
2742

        torch.cuda.current_stream().wait_stream(cp_stream)

        if use_fused_attention:
            if qkv_format == "bshd":
                out = out.view(out.shape[0], -1, *out.shape[-2:])
            elif qkv_format == "sbhd":
                out = out.view(-1, *out.shape[-3:])
        else:
            out = out.view(-1, *out.shape[-2:])

        ctx.save_for_backward(
            q,
            k,
            v,
            cu_seqlens_q,
            cu_seqlens_q_padded,
2743
            *cu_seqlens_kv_per_step,
2744
2745
2746
2747
            *out_per_step,
            *softmax_lse_per_step,
            *rng_states,
        )
2748
2749

        ctx.qkv_dtype = qkv_dtype
2750
2751
        ctx.kv_seq_range_per_step = kv_seq_range_per_step
        ctx.window_size_per_step = window_size_per_step
2752
2753
2754
2755
2756
2757
2758
        ctx.cp_group = cp_group
        ctx.cp_stream = cp_stream
        ctx.dropout_p = dropout_p
        ctx.max_seqlen_q = max_seqlen_q
        ctx.softmax_scale = softmax_scale
        ctx.qkv_format = qkv_format
        ctx.attn_bias_type = attn_bias_type
2759
        ctx.attn_mask_type = attn_mask_type
2760
2761
        ctx.deterministic = deterministic
        ctx.use_fused_attention = use_fused_attention
2762
        ctx.use_flash_attn_3 = use_flash_attn_3
2763
        nvtx_range_pop("transformer_engine.AttnFuncWithCPAndKVAllGather.forward")
2764
2765
2766
2767
        return out

    @staticmethod
    def backward(ctx, dout):
2768
        # pylint: disable=missing-function-docstring
2769
        nvtx_range_push("transformer_engine.AttnFuncWithCPAndKVAllGather.backward")
2770
2771
2772
        cp_size = get_distributed_world_size(ctx.cp_group)
        rank = get_distributed_rank(ctx.cp_group)

2773
2774
2775
2776
2777
2778
        (*saved_tensors,) = ctx.saved_tensors
        (q, k, v, cu_seqlens_q, cu_seqlens_q_padded) = saved_tensors[:5]
        cu_seqlens_kv_per_step = saved_tensors[5:7]
        out_per_step = saved_tensors[7:9]
        softmax_lse_per_step = saved_tensors[9:11]
        rng_states = saved_tensors[11:13]
2779
2780
        kv_seq_range_per_step = ctx.kv_seq_range_per_step
        window_size_per_step = ctx.window_size_per_step
2781

2782
        seq_dim = ctx.qkv_format.index("s")
2783
2784
        qkv_layout = ctx.qkv_format + "_" + ctx.qkv_format + "_" + ctx.qkv_format

2785
        dout = dout.view(q.shape)
2786
        dq = torch.empty_like(q)
2787
        dk = torch.zeros((k.shape[0] * cp_size, *k.shape[1:]), dtype=k.dtype, device=k.device)
2788
2789
2790
2791
2792
2793
2794
2795
2796
2797
        dv = torch.zeros_like(dk)
        dq_per_step = [None, None]
        dk_per_step = [None, None]
        dv_per_step = [None, None]

        # create two streams to resolve wave quantization issue of Flash Attn in each step
        flash_attn_streams = [torch.cuda.current_stream(), ctx.cp_stream]
        # synchronize dkv update across steps
        dkv_update_done = torch.cuda.Event()

2798
        # [s, b, np, hn] -> [cp, s, b, np, hn]
2799
2800
        k_ag, _ = gather_along_first_dim(k, ctx.cp_group)
        v_ag, _ = gather_along_first_dim(v, ctx.cp_group)
2801
2802

        # [cp, s, b, np, hn] -> [cp*2, s//2, b, np, hn]
2803
2804
        k_ag = k_ag.view(2 * cp_size, k.shape[0] // 2, *k.shape[1:])
        v_ag = v_ag.view(2 * cp_size, v.shape[0] // 2, *v.shape[1:])
2805
        chunk_ids_for_kv_ag = get_seq_chunk_ids_for_reordering_before_attn(cp_size, k.device)
2806
2807
2808
2809
2810
2811
        k_ag = torch.index_select(k_ag, dim=0, index=chunk_ids_for_kv_ag)
        v_ag = torch.index_select(v_ag, dim=0, index=chunk_ids_for_kv_ag)
        # [cp*2, s//2, b, np, hn] -> [cp*s, b, np, hn]
        k_ag = k_ag.view(-1, *k.shape[1:])
        v_ag = v_ag.view(-1, *v.shape[1:])
        ctx.cp_stream.wait_stream(torch.cuda.current_stream())
2812
2813
2814

        local_seq_chunk_ids = [rank, 2 * cp_size - rank - 1]

2815
        flash_attn_bwd = None
2816
2817
        if not ctx.use_fused_attention:
            fa_backward_kwargs = {"softmax_scale": ctx.softmax_scale}
2818
2819
            if ctx.use_flash_attn_3:
                flash_attn_bwd = _flash_attn_bwd_v3
2820
2821
                fa_backward_kwargs["deterministic"] = ctx.deterministic
            else:
2822
2823
2824
2825
                if ctx.qkv_format == "thd":
                    flash_attn_bwd = _flash_attn_varlen_bwd
                else:
                    flash_attn_bwd = _flash_attn_bwd
2826
                fa_backward_kwargs["dropout_p"] = ctx.dropout_p
2827
                if fa_utils.v2_4_plus:
2828
                    fa_backward_kwargs["alibi_slopes"] = None
2829
                if fa_utils.v2_4_1_plus:
2830
                    fa_backward_kwargs["deterministic"] = ctx.deterministic
2831
                if fa_utils.v2_6_0_plus:
2832
                    fa_backward_kwargs["softcap"] = 0.0
2833
2834
2835
2836

        for i in range(len(local_seq_chunk_ids) + 1):
            if i < len(local_seq_chunk_ids):
                with torch.cuda.stream(flash_attn_streams[i]):
2837
2838
                    # [b, 2, sq//2, np, hn] -> [b, sq//2, np, hn]
                    # or [2, sq//2, b, np, hn] -> [sq//2, b, np, hn]
2839
2840
2841
2842
2843
2844
2845
2846
2847
                    q_ = q.select(seq_dim, i).contiguous()
                    seq_start_idx, seq_end_idx = (
                        kv_seq_range_per_step[i][0],
                        kv_seq_range_per_step[i][1],
                    )
                    max_seqlen_kv = seq_end_idx - seq_start_idx
                    k_, v_ = [x[seq_start_idx:seq_end_idx] for x in [k_ag, v_ag]]
                    # [cp*s, b, np, hn] -> [b, s_range, np, hn] or [s_range, b, np, hn]
                    k_, v_ = [x.movedim(0, seq_dim).contiguous() for x in [k_, v_]]
2848
                    out_ = out_per_step[i]
2849
                    dout_ = dout.select(seq_dim, i).contiguous().view(out_.shape)
2850
2851
2852
2853
                    if ctx.use_fused_attention:
                        aux_ctx_tensors = [softmax_lse_per_step[i], rng_states[i]]
                        dq_per_step[i], dk_per_step[i], dv_per_step[i], _ = fused_attn_bwd(
                            ctx.max_seqlen_q,
2854
                            max_seqlen_kv,
2855
                            cu_seqlens_q,
2856
                            cu_seqlens_kv_per_step[i],
2857
2858
2859
2860
2861
                            q_,
                            k_,
                            v_,
                            out_,
                            dout_,
2862
                            ctx.qkv_dtype,
2863
                            TE_DType[dout.dtype],
2864
2865
2866
                            aux_ctx_tensors,
                            tex.NVTE_Fused_Attn_Backend.NVTE_F16_arbitrary_seqlen,
                            cu_seqlens_q_padded=cu_seqlens_q_padded,
2867
                            cu_seqlens_kv_padded=cu_seqlens_kv_per_step[i],
2868
2869
2870
2871
2872
                            attn_scale=ctx.softmax_scale,
                            dropout=ctx.dropout_p,
                            qkv_layout=qkv_layout,
                            attn_mask_type=ctx.attn_mask_type,
                            attn_bias_type=ctx.attn_bias_type,
2873
2874
                            window_size=window_size_per_step[i],
                            deterministic=ctx.deterministic,
2875
2876
2877
2878
2879
                        )
                    else:
                        dq_per_step[i], dk_per_step[i], dv_per_step[i] = [
                            torch.empty_like(x) for x in [q_, k_, v_]
                        ]
2880
2881
2882
2883
2884
2885
2886
2887
2888
2889
2890
2891
2892
                        fa_backward_args_thd = get_fa_args(
                            False,
                            ctx.use_flash_attn_3,
                            ctx.qkv_format,
                            cu_seqlens_q=cu_seqlens_q,
                            cu_seqlens_kv=cu_seqlens_kv_per_step[i],
                            max_seqlen_q=ctx.max_seqlen_q,
                            max_seqlen_kv=max_seqlen_kv,
                            dq=dq_per_step[i],
                            dk=dk_per_step[i],
                            dv=dv_per_step[i],
                        )
                        if not ctx.use_flash_attn_3:
2893
                            fa_backward_kwargs["rng_state"] = rng_states[i]
2894
2895
2896
                        if ctx.use_flash_attn_3 or (
                            fa_utils.v2_3_plus and not fa_utils.v2_7_0_plus
                        ):
2897
                            fa_backward_kwargs["window_size"] = window_size_per_step[i]
2898
                        elif fa_utils.v2_7_0_plus:
2899
2900
                            fa_backward_kwargs["window_size_left"] = window_size_per_step[i][0]
                            fa_backward_kwargs["window_size_right"] = window_size_per_step[i][1]
2901
                        flash_attn_bwd(
2902
2903
2904
2905
2906
2907
                            dout_,
                            q_,
                            k_,
                            v_,
                            out_,
                            softmax_lse_per_step[i],
2908
                            *fa_backward_args_thd,
2909
2910
                            causal="causal" in ctx.attn_mask_type,
                            **fa_backward_kwargs,
2911
2912
2913
2914
2915
                        )

            if i > 0:
                with torch.cuda.stream(flash_attn_streams[i - 1]):
                    if ctx.qkv_format == "bshd":
2916
                        dq[:, i - 1].copy_(dq_per_step[i - 1])
2917
                    elif ctx.qkv_format == "sbhd":
2918
2919
2920
2921
2922
2923
                        dq[i - 1].copy_(dq_per_step[i - 1])
                    # [b, s_range, np, hn] or [s_range, b, np, hn] -> [s_range, b, np, hn]
                    dk_per_step[i - 1], dv_per_step[i - 1] = [
                        x.movedim(seq_dim, 0).contiguous()
                        for x in [dk_per_step[i - 1], dv_per_step[i - 1]]
                    ]
2924
2925
2926
                    # wait until dkv update of last step is done
                    if i > 1:
                        flash_attn_streams[i - 1].wait_event(dkv_update_done)
2927
2928
2929
2930
2931
2932
                    seq_start_idx, seq_end_idx = (
                        kv_seq_range_per_step[i - 1][0],
                        kv_seq_range_per_step[i - 1][1],
                    )
                    dk[seq_start_idx:seq_end_idx].add_(dk_per_step[i - 1])
                    dv[seq_start_idx:seq_end_idx].add_(dv_per_step[i - 1])
2933
2934
2935
2936
2937
                    if i < len(local_seq_chunk_ids):
                        flash_attn_streams[i - 1].record_event(dkv_update_done)

        torch.cuda.current_stream().wait_stream(ctx.cp_stream)

2938
2939
2940
        # [cp*s, b, np, hn] -> [cp*2, s//2, b, np, hn]
        dk = dk.view(2 * cp_size, -1, *dk.shape[-3:])
        dv = dv.view(2 * cp_size, -1, *dv.shape[-3:])
2941
        chunk_ids_for_kv_ag = get_seq_chunk_ids_for_reordering_after_attn(cp_size, dk.device)
2942
2943
2944
        dk = torch.index_select(dk, dim=0, index=chunk_ids_for_kv_ag)
        dv = torch.index_select(dv, dim=0, index=chunk_ids_for_kv_ag)
        # [cp*2, s//2, b, np, hn] -> [cp*s, b, np, hn]
2945
2946
2947
2948
2949
        dk = dk.view(-1, *dk.shape[-3:])
        dv = dv.view(-1, *dv.shape[-3:])
        dk, _ = reduce_scatter_along_first_dim(dk, ctx.cp_group)
        dv, _ = reduce_scatter_along_first_dim(dv, ctx.cp_group)

2950
2951
2952
        dq = dq.view(*dq.shape[:seq_dim], -1, *dq.shape[(seq_dim + 2) :])
        dk = dk.movedim(0, seq_dim).contiguous()
        dv = dv.movedim(0, seq_dim).contiguous()
2953
        nvtx_range_pop("transformer_engine.AttnFuncWithCPAndKVAllGather.backward")
2954
2955
2956
2957
2958
2959
2960
2961
2962
2963
2964
2965
2966
2967
2968
2969
2970
2971
2972
2973
2974

        return (
            None,
            dq,
            dk,
            dv,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
2975
            None,
2976
2977
2978
2979
2980
2981
2982
2983
2984
2985
2986
2987
2988
2989
2990
2991
2992
2993
2994
2995
2996
2997
2998
2999
3000
3001
3002
3003
3004
3005
3006
3007
3008
3009
3010
        )


class AttnFuncWithCPAndQKVOA2A(torch.autograd.Function):
    """
    Attention implementation with context parallelism. Like Ulysses, applying A2A to QKVO.
    Refer the paper `DeepSpeed Ulysses <https://arxiv.org/abs/2309.14509>`_.
    """

    @staticmethod
    def forward(
        ctx,
        is_training,
        q,
        k,
        v,
        cu_seqlens_q,
        cu_seqlens_kv,
        max_seqlen_q,
        max_seqlen_kv,
        cu_seqlens_q_padded,
        cu_seqlens_kv_padded,
        dropout_p,
        softmax_scale,
        qkv_format,
        attn_mask_type,
        attn_bias_type,
        attn_bias,
        deterministic,
        use_fused_attention,
        window_size,
        fp8,
        fp8_meta,
        cp_group,
        cp_stream,
3011
        quantizers,
3012
        use_flash_attn_3,
3013
    ):
3014
        # pylint: disable=missing-function-docstring
3015
        nvtx_range_push("transformer_engine.AttnFuncWithCPAndQKVOA2A.forward")
3016
3017
3018
3019
        if softmax_scale is None:
            softmax_scale = q.shape[-1] ** (-0.5)

        cp_size = get_distributed_world_size(cp_group)
3020
        qkv_dtype = q.dtype
3021
3022
3023
3024
3025
3026
3027
3028
3029
3030

        causal = "causal" in attn_mask_type
        padding = "padding" in attn_mask_type
        assert not padding, f"{attn_mask_type} mask type is not supported!"
        assert attn_bias_type == "no_bias", f"{attn_bias_type} bias type is not supported!"
        assert q.shape[-1] % 8 == 0, "Hidden size per attention head should be multiple of 8!"
        assert (
            window_size == (-1, 0)
            or window_size == (-1, -1)
            or use_fused_attention
3031
            or fa_utils.v2_3_plus
3032
        ), "Sliding window attention only can work with FusedAttention or FlashAttention >= 2.3!"
3033

3034
        flash_attn_fwd = None
3035
3036
        if not use_fused_attention:
            fa_forward_kwargs = {"softmax_scale": softmax_scale}
3037
3038
            if use_flash_attn_3:
                flash_attn_fwd = _flash_attn_fwd_v3
3039
3040
                fa_forward_kwargs["window_size"] = window_size
            else:
3041
3042
3043
3044
                if qkv_format == "thd":
                    flash_attn_fwd = _flash_attn_varlen_fwd
                else:
                    flash_attn_fwd = _flash_attn_fwd
3045
3046
                fa_forward_kwargs["dropout_p"] = dropout_p
                fa_forward_kwargs["return_softmax"] = False
3047
                if fa_utils.v2_3_plus and not fa_utils.v2_7_0_plus:
3048
                    fa_forward_kwargs["window_size"] = window_size
3049
                elif fa_utils.v2_7_0_plus:
3050
3051
                    fa_forward_kwargs["window_size_left"] = window_size[0]
                    fa_forward_kwargs["window_size_right"] = window_size[1]
3052
                if fa_utils.v2_4_plus:
3053
                    fa_forward_kwargs["alibi_slopes"] = None
3054
                if fa_utils.v2_5_7_plus and qkv_format == "thd":
3055
                    fa_forward_kwargs["block_table"] = None
3056
                if fa_utils.v2_6_0_plus:
3057
                    fa_forward_kwargs["softcap"] = 0.0
3058
3059
3060
3061
3062
3063
3064
3065
3066
3067
3068
3069
3070
3071

        assert (
            q.shape[-2] % cp_size == 0 and k.shape[-2] % cp_size == 0
        ), "The number of attention heads needs to be divisible by CP size!"

        assert qkv_format != "thd", f"{qkv_format} format is not supported!"
        qkv_layout = qkv_format + "_" + qkv_format + "_" + qkv_format

        batch_dim = qkv_format.index("b")
        seq_dim = qkv_format.index("s")
        assert (
            q.shape[seq_dim] % 2 == 0 and k.shape[seq_dim] % 2 == 0
        ), "Sequence length per GPU needs to be divisible by 2!"

3072
        fused_attn_backend = None
3073
3074
        # "fp8_mha" decides outputs in fp8, while inputs are inferred from the real dtype
        is_input_fp8 = False
3075
3076
3077
        is_output_fp8 = False

        QKV_quantizer, O_quantizer, S_quantizer, dQKV_quantizer, dO_quantizer, dP_quantizer = (
3078
            dpa_utils.get_attention_quantizers(fp8, quantizers, cp_specific_quantizers=False)
3079
3080
3081
        )
        if fp8:
            if use_fused_attention:
3082
                fused_attn_backend = FusedAttnBackend["FP8"]
3083
3084
3085
3086
                assert isinstance(k, q.__class__) and isinstance(
                    v, q.__class__
                ), "q, k, and v must have the same type."
                is_input_fp8 = isinstance(q, Float8Tensor)
3087
                is_output_fp8 = fp8_meta is not None and fp8_meta["recipe"].fp8_mha
3088
                if is_input_fp8:
3089
                    QKV_quantizer = q._quantizer
3090
3091
3092
3093
                    q_fp8, k_fp8, v_fp8 = q, k, v
                    q, k, v = q_fp8._data, k_fp8._data, v_fp8._data
                elif int(os.getenv("NVTE_FP8_DPA_BWD", "1")):
                    q_f16, k_f16, v_f16 = q, k, v
3094
                    q, k, v = [QKV_quantizer(x)._data for x in [q_f16, k_f16, v_f16]]
3095
                fp8_meta_kwargs = {}
3096
3097
                fp8_meta_kwargs["s_quantizer"] = S_quantizer
                fp8_meta_kwargs["o_quantizer"] = O_quantizer  # partial result quantizer
3098
3099
3100
3101
3102
3103
3104
            else:
                assert False, "FP8 is only supported with Fused Attention!"
        else:
            if use_fused_attention:
                fp8_meta_kwargs = {}
                fused_attn_backend = FusedAttnBackend["F16_arbitrary_seqlen"]

3105
        chunk_ids_for_a2a = get_seq_chunk_ids_for_reordering_before_attn(cp_size, q.device)
3106
3107
3108
3109
        q, k, v = flash_attn_a2a_communicate(
            [q, k, v], chunk_ids_for_a2a, seq_dim, cp_size, cp_group, cp_stream, True
        )

3110
        if fp8 and not is_input_fp8 and not int(os.getenv("NVTE_FP8_DPA_BWD", "1")):
3111
            q_f16, k_f16, v_f16 = q, k, v
3112
            q, k, v = [QKV_quantizer(x)._data for x in [q_f16, k_f16, v_f16]]
3113
3114
3115

        batch_size = q.shape[batch_dim]
        if use_fused_attention:
3116
3117
3118
3119
3120
3121
3122
3123
3124
3125
3126
            q_part, k_part, v_part = q, k, v
            if fp8:
                q_part = QKV_quantizer.create_tensor_from_data(
                    q, fake_dtype=qkv_dtype, internal=True
                )
                k_part = QKV_quantizer.create_tensor_from_data(
                    k, fake_dtype=qkv_dtype, internal=True
                )
                v_part = QKV_quantizer.create_tensor_from_data(
                    v, fake_dtype=qkv_dtype, internal=True
                )
3127
3128
3129
3130
3131
3132
            out, aux_ctx_tensors = fused_attn_fwd(
                is_training,
                max_seqlen_q,
                max_seqlen_kv,
                cu_seqlens_q,
                cu_seqlens_kv,
3133
3134
3135
3136
                q_part,
                k_part,
                v_part,
                qkv_dtype,
3137
3138
3139
3140
3141
3142
3143
3144
3145
3146
3147
3148
                fused_attn_backend,
                attn_scale=softmax_scale,
                dropout=dropout_p,
                qkv_layout=qkv_layout,
                attn_mask_type=attn_mask_type,
                attn_bias_type=attn_bias_type,
                attn_bias=attn_bias,
                cu_seqlens_q_padded=cu_seqlens_q_padded,
                cu_seqlens_kv_padded=cu_seqlens_kv_padded,
                window_size=window_size,
                **fp8_meta_kwargs,
            )
3149
3150
            if fp8:
                out = out._data
3151
        else:
3152
3153
3154
3155
3156
3157
3158
3159
3160
            fa_forward_args_thd = get_fa_args(
                True,
                use_flash_attn_3,
                qkv_format,
                cu_seqlens_q=cu_seqlens_q,
                cu_seqlens_kv=cu_seqlens_kv,
                max_seqlen_q=max_seqlen_q,
                max_seqlen_kv=max_seqlen_kv,
            )
3161
            fa_outputs = flash_attn_fwd(
3162
3163
3164
                q,
                k,
                v,
3165
                *fa_forward_args_thd,
3166
                causal=causal,
3167
                **fa_forward_kwargs,
3168
            )
3169
            if not fa_utils.v2_7_0_plus:
3170
                out, softmax_lse = fa_outputs[4], fa_outputs[5]
3171
                rng_state = fa_outputs[7] if not use_flash_attn_3 else None
3172
3173
            else:
                out, softmax_lse = fa_outputs[0], fa_outputs[1]
3174
                rng_state = fa_outputs[3] if not use_flash_attn_3 else None
3175
3176
            aux_ctx_tensors = [softmax_lse, rng_state]

3177
        chunk_ids_for_a2a = get_seq_chunk_ids_for_reordering_after_attn(cp_size, out.device)
3178
3179
3180
3181
3182
3183
3184
3185
3186
3187
3188
3189
3190
        out = flash_attn_a2a_communicate(
            out, chunk_ids_for_a2a, seq_dim, cp_size, cp_group, cp_stream, False
        )

        if use_fused_attention:
            if qkv_format == "bshd":
                # [b*s, np, hn] -> [b, s, np, hn]
                out = out.view(batch_size, -1, *out.shape[-2:])
            elif qkv_format == "sbhd":
                # [s*b, np, hn] -> [s, b, np, hn]
                out = out.view(-1, batch_size, *out.shape[-2:])

        if fp8:
3191
            if is_output_fp8:
3192
3193
                out_fp8 = O_quantizer.create_tensor_from_data(
                    out, fake_dtype=qkv_dtype, internal=False
3194
3195
                )
                out_ret = out_fp8
3196
                out = out_fp8._data
3197
            else:
3198
                out_fp8 = O_quantizer.create_tensor_from_data(
3199
                    out, fake_dtype=qkv_dtype, internal=True
3200
                )
3201
                out_f16 = out_fp8.dequantize(dtype=qkv_dtype)
3202
3203
3204
3205
                out_ret = out_f16
        else:
            out_ret = out

3206
        if not fp8 or int(os.getenv("NVTE_FP8_DPA_BWD", "1")):
3207
            q_save, k_save, v_save, out_save = q, k, v, out
3208
3209
3210
3211
3212
3213
3214
3215
3216
        else:
            if is_input_fp8:
                q_save, k_save, v_save = q, k, v
            else:
                q_save, k_save, v_save = q_f16, k_f16, v_f16
            if is_output_fp8:
                out_save = out
            else:
                out_save = out_f16
3217

3218
        tensors_to_save, tensor_objects = prepare_for_saving(
3219
3220
3221
3222
3223
3224
3225
3226
3227
3228
            q_save,
            k_save,
            v_save,
            out_save,
            cu_seqlens_q,
            cu_seqlens_kv,
            cu_seqlens_q_padded,
            cu_seqlens_kv_padded,
            *aux_ctx_tensors,
        )
3229
3230
3231
        ctx.save_for_backward(*tensors_to_save)
        ctx.tensor_objects = tensor_objects

3232
3233
3234
3235
3236
3237
3238
3239
3240
3241
3242
3243
3244
3245
3246
        ctx.batch_size = batch_size
        ctx.cp_group = cp_group
        ctx.cp_stream = cp_stream
        ctx.dropout_p = dropout_p
        ctx.max_seqlen_q = max_seqlen_q
        ctx.max_seqlen_kv = max_seqlen_kv
        ctx.softmax_scale = softmax_scale
        ctx.qkv_format = qkv_format
        ctx.attn_mask_type = attn_mask_type
        ctx.attn_bias_type = attn_bias_type
        ctx.deterministic = deterministic
        ctx.window_size = window_size
        ctx.use_fused_attention = use_fused_attention
        ctx.fp8 = fp8 and int(os.getenv("NVTE_FP8_DPA_BWD", "1"))
        ctx.fp8_meta = fp8_meta
3247
3248
        ctx.is_input_fp8 = is_input_fp8
        ctx.is_output_fp8 = is_output_fp8
3249
        ctx.use_flash_attn_3 = use_flash_attn_3
3250
3251
3252
3253
3254
3255
3256
3257
3258
3259
3260
3261
3262
3263
3264

        ctx.qkv_dtype = qkv_dtype
        ctx.dQKV_quantizer = dQKV_quantizer
        ctx.dO_quantizer = dO_quantizer
        ctx.dP_quantizer = dP_quantizer
        ctx.QKV_quantizer = QKV_quantizer
        ctx.O_quantizer = O_quantizer
        ctx.S_quantizer = S_quantizer
        if ctx.fp8:
            ctx.QKV_quantizer = QKV_quantizer.copy()
            ctx.QKV_quantizer.scale = QKV_quantizer.scale.clone()
            ctx.O_quantizer = O_quantizer.copy()
            ctx.O_quantizer.scale = O_quantizer.scale.clone()
            ctx.S_quantizer = S_quantizer.copy()
            ctx.S_quantizer.scale = S_quantizer.scale.clone()
3265
        nvtx_range_pop("transformer_engine.AttnFuncWithCPAndQKVOA2A.forward")
3266
3267
3268
3269
        return out_ret

    @staticmethod
    def backward(ctx, dout):
3270
        # pylint: disable=missing-function-docstring
3271
        nvtx_range_push("transformer_engine.AttnFuncWithCPAndQKVOA2A.backward")
3272
3273
        cp_size = get_distributed_world_size(ctx.cp_group)

3274
3275
3276
3277
3278
3279
3280
3281
3282
3283
3284
        (
            q,
            k,
            v,
            out,
            cu_seqlens_q,
            cu_seqlens_kv,
            cu_seqlens_q_padded,
            cu_seqlens_kv_padded,
            *aux_ctx_tensors,
        ) = restore_from_saved(ctx.tensor_objects, ctx.saved_tensors)
3285
3286
3287
3288
3289

        qkv_layout = ctx.qkv_format + "_" + ctx.qkv_format + "_" + ctx.qkv_format
        causal = "causal" in ctx.attn_mask_type
        seq_dim = ctx.qkv_format.index("s")

3290
        dout_dtype = dout.dtype
3291
3292
        fused_attn_backend = None
        fused_attn_dqkv_dtype = None
3293
3294
3295
        if ctx.fp8:
            if ctx.use_fused_attention:
                fused_attn_backend = FusedAttnBackend["FP8"]
3296
                if ctx.is_output_fp8:
3297
                    assert isinstance(dout, Float8Tensor), "dout must be Float8Tensors for FP8 MHA!"
3298
                    ctx.dO_quantizer = dout._quantizer
3299
                else:
3300
                    dout = ctx.dO_quantizer(dout)
3301
                fused_attn_dqkv_dtype = TE_DType[dout._data.dtype]
3302
                dout = dout._data
3303
                fp8_meta_kwargs = {}
3304
3305
3306
3307
                fp8_meta_kwargs["s_quantizer"] = ctx.S_quantizer
                fp8_meta_kwargs["dp_quantizer"] = ctx.dP_quantizer
                fp8_meta_kwargs["dqkv_quantizer"] = ctx.dQKV_quantizer

3308
3309
3310
            else:
                assert False, "FP8 is only supported with Fused Attention!"
        else:
3311
3312
3313
3314
3315
3316
3317
3318
3319
3320
3321
3322
3323
3324
3325
3326
            if ctx.fp8_meta is not None:
                if ctx.is_output_fp8:
                    assert isinstance(dout, Float8Tensor), "dout must be Float8Tensors for FP8 MHA!"
                    ctx.dO_quantizer = dout._quantizer
                    dout = dout._data
                if ctx.is_input_fp8:
                    q = ctx.QKV_quantizer.create_tensor_from_data(
                        q, fake_dtype=ctx.qkv_dtype, internal=True
                    )
                    k = ctx.QKV_quantizer.create_tensor_from_data(
                        k, fake_dtype=ctx.qkv_dtype, internal=True
                    )
                    v = ctx.QKV_quantizer.create_tensor_from_data(
                        v, fake_dtype=ctx.qkv_dtype, internal=True
                    )
                    q, k, v = [x.dequantize(dtype=ctx.qkv_dtype) for x in [q, k, v]]
3327
3328
            if ctx.use_fused_attention:
                fp8_meta_kwargs = {}
3329
                fused_attn_dqkv_dtype = TE_DType[dout_dtype]
3330
3331
3332
3333
3334
3335
                fused_attn_backend = FusedAttnBackend["F16_arbitrary_seqlen"]

        if not ctx.use_fused_attention:
            out = out.view(ctx.batch_size, -1, *out.shape[-2:])
        dout = dout.view(*out.shape)

3336
        chunk_ids_for_a2a = get_seq_chunk_ids_for_reordering_before_attn(cp_size, out.device)
3337
3338
3339
        out, dout = flash_attn_a2a_communicate(
            [out, dout], chunk_ids_for_a2a, seq_dim, cp_size, ctx.cp_group, ctx.cp_stream, True
        )
3340
3341
3342
3343
3344
3345
3346
3347
3348
        if not ctx.fp8 and ctx.fp8_meta is not None and ctx.is_output_fp8:
            out = ctx.O_quantizer.create_tensor_from_data(
                out, fake_dtype=ctx.qkv_dtype, internal=True
            )
            dout = ctx.dO_quantizer.create_tensor_from_data(
                dout, fake_dtype=dout_dtype, internal=True
            )
            out = out.dequantize(dtype=ctx.qkv_dtype)
            dout = dout.dequantize(dtype=dout_dtype)
3349

3350
        flash_attn_bwd = None
3351
3352
        if not ctx.use_fused_attention:
            fa_backward_kwargs = {"softmax_scale": ctx.softmax_scale}
3353
3354
3355
3356
            if ctx.use_flash_attn_3:
                flash_attn_bwd = (
                    _flash_attn_bwd_v3  # pylint: disable=possibly-used-before-assignment
                )
3357
3358
3359
                fa_backward_kwargs["window_size"] = ctx.window_size
                fa_backward_kwargs["deterministic"] = ctx.deterministic
            else:
3360
3361
3362
3363
                if ctx.qkv_format == "thd":
                    flash_attn_bwd = _flash_attn_varlen_bwd
                else:
                    flash_attn_bwd = _flash_attn_bwd
3364
                fa_backward_kwargs["dropout_p"] = ctx.dropout_p
3365
                if fa_utils.v2_3_plus and not fa_utils.v2_7_0_plus:
3366
                    fa_backward_kwargs["window_size"] = ctx.window_size
3367
                elif fa_utils.v2_7_0_plus:
3368
3369
                    fa_backward_kwargs["window_size_left"] = ctx.window_size[0]
                    fa_backward_kwargs["window_size_right"] = ctx.window_size[1]
3370
                if fa_utils.v2_4_plus:
3371
                    fa_backward_kwargs["alibi_slopes"] = None
3372
                if fa_utils.v2_4_1_plus:
3373
                    fa_backward_kwargs["deterministic"] = ctx.deterministic
3374
                if fa_utils.v2_6_0_plus:
3375
                    fa_backward_kwargs["softcap"] = 0.0
3376
3377

        if ctx.use_fused_attention:
3378
3379
3380
3381
3382
3383
3384
3385
3386
3387
3388
3389
3390
3391
3392
3393
3394
3395
3396
3397
            q_part = q
            k_part = k
            v_part = v
            out_part = out
            dout_part = dout

            if ctx.fp8:
                q_part = ctx.QKV_quantizer.create_tensor_from_data(
                    q_part, fake_dtype=ctx.qkv_dtype, internal=True
                )
                k_part = ctx.QKV_quantizer.create_tensor_from_data(
                    k_part, fake_dtype=ctx.qkv_dtype, internal=True
                )
                v_part = ctx.QKV_quantizer.create_tensor_from_data(
                    v_part, fake_dtype=ctx.qkv_dtype, internal=True
                )
                out_part = ctx.O_quantizer.create_tensor_from_data(
                    out_part, fake_dtype=ctx.qkv_dtype, internal=True
                )
                dout_part = ctx.dO_quantizer.create_tensor_from_data(
3398
                    dout_part, fake_dtype=dout_dtype, internal=True
3399
3400
                )

3401
3402
3403
3404
3405
            dq, dk, dv, _ = fused_attn_bwd(
                ctx.max_seqlen_q,
                ctx.max_seqlen_kv,
                cu_seqlens_q,
                cu_seqlens_kv,
3406
3407
3408
3409
3410
                q_part,
                k_part,
                v_part,
                out_part,
                dout_part,
3411
                dout_dtype,
3412
3413
3414
3415
3416
3417
3418
3419
3420
3421
3422
3423
3424
3425
                fused_attn_dqkv_dtype,
                aux_ctx_tensors,
                fused_attn_backend,
                cu_seqlens_q_padded=cu_seqlens_q_padded,
                cu_seqlens_kv_padded=cu_seqlens_kv_padded,
                attn_scale=ctx.softmax_scale,
                dropout=ctx.dropout_p,
                qkv_layout=qkv_layout,
                attn_mask_type=ctx.attn_mask_type,
                attn_bias_type=ctx.attn_bias_type,
                window_size=ctx.window_size,
                deterministic=ctx.deterministic,
                **fp8_meta_kwargs,
            )
3426
3427
3428
3429
            if ctx.fp8:
                dq = dq._data
                dk = dk._data
                dv = dv._data
3430
3431
3432
        else:
            softmax_lse, rng_state = aux_ctx_tensors
            dq, dk, dv = [torch.empty_like(x) for x in [q, k, v]]
3433
3434
3435
3436
3437
3438
3439
3440
3441
3442
3443
3444
3445
            fa_backward_args_thd = get_fa_args(
                False,
                ctx.use_flash_attn_3,
                ctx.qkv_format,
                cu_seqlens_q=cu_seqlens_q,
                cu_seqlens_kv=cu_seqlens_kv,
                max_seqlen_q=ctx.max_seqlen_q,
                max_seqlen_kv=ctx.max_seqlen_kv,
                dq=dq,
                dk=dk,
                dv=dv,
            )
            if not ctx.use_flash_attn_3:
3446
3447
                fa_backward_kwargs["rng_state"] = rng_state
            flash_attn_bwd(
3448
3449
3450
3451
3452
3453
                dout,
                q,
                k,
                v,
                out,
                softmax_lse,
3454
                *fa_backward_args_thd,
3455
3456
                causal=causal,
                **fa_backward_kwargs,
3457
3458
            )

3459
        chunk_ids_for_a2a = get_seq_chunk_ids_for_reordering_after_attn(cp_size, q.device)
3460
3461
3462
3463
        dq, dk, dv = flash_attn_a2a_communicate(
            [dq, dk, dv], chunk_ids_for_a2a, seq_dim, cp_size, ctx.cp_group, ctx.cp_stream, False
        )

3464
        if ctx.qkv_format == "bshd":
3465
            dq, dk, dv = [x.view(ctx.batch_size, -1, *x.shape[-2:]) for x in [dq, dk, dv]]
3466
        elif ctx.qkv_format == "sbhd":
3467
3468
3469
            dq, dk, dv = [x.view(-1, ctx.batch_size, *x.shape[-2:]) for x in [dq, dk, dv]]

        if ctx.fp8:
3470
3471
3472
3473
3474
3475
3476
3477
3478
            dq = ctx.dQKV_quantizer.create_tensor_from_data(
                dq, fake_dtype=dout_dtype, internal=not ctx.is_input_fp8
            )
            dk = ctx.dQKV_quantizer.create_tensor_from_data(
                dk, fake_dtype=dout_dtype, internal=not ctx.is_input_fp8
            )
            dv = ctx.dQKV_quantizer.create_tensor_from_data(
                dv, fake_dtype=dout_dtype, internal=not ctx.is_input_fp8
            )
3479
            if not ctx.is_input_fp8:
3480
                dq, dk, dv = [x.dequantize(dtype=dout_dtype) for x in [dq, dk, dv]]
3481
        nvtx_range_pop("transformer_engine.AttnFuncWithCPAndQKVOA2A.backward")
3482
3483
3484
3485
3486
3487
3488
3489
3490
3491
3492
3493
3494
3495
3496
3497
3498
3499
3500
3501
3502
3503
3504

        return (
            None,
            dq,
            dk,
            dv,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
3505
3506
3507
            None,
            None,
            None,
3508
            None,
3509
3510
3511
        )


3512
def attn_forward_func_with_cp(
3513
3514
3515
3516
3517
    is_training,
    q,
    k,
    v,
    cu_seqlens_q,
3518
    cu_seqlens_kv,
3519
    max_seqlen_q,
3520
    max_seqlen_kv,
3521
3522
    cu_seqlens_q_padded,
    cu_seqlens_kv_padded,
3523
3524
3525
3526
    dropout_p,
    cp_group,
    cp_global_ranks,
    cp_stream,
3527
    cp_comm_type,
3528
3529
3530
3531
3532
3533
3534
    softmax_scale=None,
    qkv_format="bshd",
    attn_mask_type="causal",
    attn_bias_type="no_bias",
    attn_bias=None,
    deterministic=False,
    use_fused_attention=False,
3535
    window_size=None,
3536
3537
    fp8=False,
    fp8_meta=None,
3538
    quantizers=None,
3539
    pad_between_seqs=False,
3540
    use_flash_attn_3=False,
3541
) -> torch.Tensor:
3542
3543
3544
3545
    """
    Attention implementation with context parallelism.
    """

3546
3547
3548
3549
3550
3551
3552
3553
3554
3555
3556
3557
3558
3559
3560
3561
    if cp_comm_type == "a2a+p2p":
        assert isinstance(
            cp_group, list
        ), "Hierarchical CP implementation needs multi-level CP groups!"
        assert len(cp_group) == 2, "Current implementation only supports two-level CP groups!"
        if get_distributed_world_size(cp_group[0]) == 1:
            cp_group = cp_group[1]
            cp_comm_type = "p2p"
        elif get_distributed_world_size(cp_group[1]) == 1:
            cp_group = cp_group[0]
            cp_comm_type = "a2a"
    else:
        assert isinstance(
            cp_group, dist_group_type
        ), f"Unsupported process group for CP communication type {cp_comm_type}!"

3562
3563
3564
3565
3566
3567
3568
3569
3570
3571
3572
3573
    assert qkv_format in [
        "bshd",
        "sbhd",
        "thd",
    ], f"QKV format of {qkv_format} is not supported with context parallelism!"
    assert (
        qkv_format != "sbhd" or use_fused_attention
    ), "FlashAttention does not support sbhd format!"
    assert attn_bias is None or (use_fused_attention and "padding" not in attn_mask_type), (
        """Attention bias is only supported with FusedAttention and "causal" """
        """or "no_mask" mask types!"""
    )
3574
    assert qkv_format != "thd" or (
3575
        cu_seqlens_q_padded is not None and cu_seqlens_kv_padded is not None
3576
    ), "cu_seqlens_padded cannot be None with context parallelism + THD format!"
3577
3578
3579

    sliding_window_attn = (
        window_size is not None and window_size != (-1, 0) and window_size != (-1, -1)
3580
    )
3581
3582
3583
3584
    assert not sliding_window_attn or cp_comm_type in [
        "a2a",
        "all_gather",
    ], "The context parallel running configs cannot support sliding window attetnion!"
3585

3586
3587
3588
3589
3590
3591
3592
3593
3594
3595
3596
3597
3598
3599
3600
3601
3602
3603
3604
3605
3606
    args = [
        is_training,
        q,
        k,
        v,
        cu_seqlens_q,
        cu_seqlens_kv,
        max_seqlen_q,
        max_seqlen_kv,
        cu_seqlens_q_padded,
        cu_seqlens_kv_padded,
        dropout_p,
        softmax_scale,
        qkv_format,
        attn_mask_type,
        attn_bias_type,
        attn_bias,
        deterministic,
        use_fused_attention,
    ]

3607
    if cp_comm_type in ["p2p", "a2a+p2p"]:
3608
3609
3610
3611
3612
3613
3614
3615
3616
3617
        args += [
            fp8,
            fp8_meta,
            cp_group,
            cp_global_ranks,
            cp_stream,
            quantizers,
            pad_between_seqs,
            use_flash_attn_3,
        ]
3618
3619
3620
3621
        out = AttnFuncWithCPAndKVP2P.apply(*args)
    elif cp_comm_type == "all_gather":
        args.pop(5)
        args.pop(8)
3622
        args += [window_size, cp_group, cp_stream, use_flash_attn_3]
3623
3624
        out = AttnFuncWithCPAndKVAllGather.apply(*args)
    elif cp_comm_type == "a2a":
3625
        args += [window_size, fp8, fp8_meta, cp_group, cp_stream, quantizers, use_flash_attn_3]
3626
        out = AttnFuncWithCPAndQKVOA2A.apply(*args)
3627
3628
3629
    else:
        raise ValueError(f"Unsupported communication type: {cp_comm_type}!")

3630
3631
3632
    return out


cyanguwa's avatar
cyanguwa committed
3633
class _SplitAlongDim(torch.autograd.Function):
3634
3635
3636
    """"""

    @staticmethod
3637
3638
3639
3640
3641
    def forward(
        ctx,
        mixed_x_layer: torch.Tensor,
        split_dim: int,
        split_size_or_sections: Union[int, List[int], Tuple[int]],
3642
        squeeze=False,
3643
    ) -> Tuple[torch.Tensor, ...]:
3644
        # pylint: disable=missing-function-docstring
cyanguwa's avatar
cyanguwa committed
3645
3646
        ctx.split_dim = split_dim
        ctx.split_size_or_sections = split_size_or_sections
3647
3648
3649
3650
3651
3652
3653
3654
3655
3656
3657
3658
3659
3660
3661
3662
3663
        if isinstance(mixed_x_layer, Float8TensorBase) and not isinstance(
            mixed_x_layer, Float8Tensor
        ):
            return tuple(
                Float8TensorBase(
                    fp8_scale_inv=mixed_x_layer._scale_inv,
                    fp8_dtype=mixed_x_layer._fp8_dtype,
                    data=x.squeeze(split_dim) if squeeze else x,
                    shape=x.squeeze(split_dim).shape if squeeze else x.shape,
                    quantizer=mixed_x_layer._quantizer,
                )
                for x in torch.split(
                    mixed_x_layer._data,
                    split_size_or_sections=split_size_or_sections,
                    dim=split_dim,
                )
            )
3664
        if isinstance(mixed_x_layer, Float8Tensor):
3665
3666
3667
            return tuple(
                Float8Tensor.make_like(
                    mixed_x_layer,
3668
3669
                    data=x.squeeze(split_dim) if squeeze else x,
                    shape=x.squeeze(split_dim).shape if squeeze else x.shape,
3670
3671
                )
                for x in torch.split(
3672
3673
                    mixed_x_layer._data,
                    split_size_or_sections=split_size_or_sections,
3674
3675
3676
                    dim=split_dim,
                )
            )
3677
3678
3679
3680
        out_list = torch.split(mixed_x_layer, split_size_or_sections, dim=split_dim)
        if squeeze:
            out_list = [x.squeeze(split_dim) for x in out_list]
        return out_list
3681
3682

    @staticmethod
3683
    def backward(ctx, *grad_outputs):
3684
        # pylint: disable=missing-function-docstring
3685
3686
        assert len(grad_outputs) > 0, "No gradients received for backprop!"

cyanguwa's avatar
cyanguwa committed
3687
3688
        if isinstance(ctx.split_size_or_sections, (list, tuple)):
            split_sizes = ctx.split_size_or_sections
3689
3690
3691
            assert len(grad_outputs) == len(
                split_sizes
            ), "Unequal number of gradients vs split sections for backprop!"
cyanguwa's avatar
cyanguwa committed
3692
3693
3694
3695
3696
        if isinstance(ctx.split_size_or_sections, int):
            split_sizes = [ctx.split_size_or_sections] * len(grad_outputs)
        dims = len(grad_outputs[0].shape)
        split_dim = (ctx.split_dim + dims) % dims

3697
3698
3699
3700
3701
3702
3703
3704
        if isinstance(grad_outputs[0], Float8Tensor):
            noop_ok = True
            strides = grad_outputs[0].stride()
            data_ptr = grad_outputs[0]._data.untyped_storage().data_ptr()
            shape = list(grad_outputs[0].shape)
            for i, tensor in enumerate(grad_outputs):
                shape_i = shape
                shape_i[split_dim] = split_sizes[i]
3705
3706
3707
3708
3709
3710
3711
                offset_size = sum(split_sizes[:i]) * np.prod(shape[split_dim + 1 :])
                if (
                    tensor.stride() != strides
                    or list(tensor.shape) != shape_i
                    or tensor._data.untyped_storage().data_ptr() != data_ptr
                    or tensor.storage_offset() != offset_size
                ):
3712
3713
3714
                    noop_ok = False
                    break
            if noop_ok:
3715
3716
3717
                ret = torch.Tensor().to(
                    device=grad_outputs[0].device, dtype=grad_outputs[0]._data.dtype
                )
3718
3719
                new_shape = list(shape)
                new_shape[split_dim] = sum(split_sizes)
3720
3721
3722
3723
3724
                ret.set_(
                    grad_outputs[0]._data.untyped_storage(),
                    grad_outputs[0]._data.storage_offset(),
                    new_shape,
                    strides,
3725
                )
3726
3727
3728
3729
3730
                return (
                    Float8Tensor.make_like(grad_outputs[0], data=ret, shape=ret.shape),
                    None,
                    None,
                )
3731
3732

            grad_outputs_data = [x._data for x in grad_outputs]
3733
            data = torch.cat(grad_outputs_data, dim=split_dim)
3734
            return (
3735
3736
                Float8Tensor.make_like(grad_outputs[0], data=data, shape=data.shape),
                None,
3737
3738
3739
                None,
                None,
            )
3740
3741
        noop_ok = True
        strides = grad_outputs[0].stride()
3742
        data_ptr = grad_outputs[0].untyped_storage().data_ptr()
cyanguwa's avatar
cyanguwa committed
3743
        shape = list(grad_outputs[0].shape)
3744
        for i, tensor in enumerate(grad_outputs):
cyanguwa's avatar
cyanguwa committed
3745
3746
            shape_i = shape
            shape_i[split_dim] = split_sizes[i]
3747
3748
3749
3750
3751
3752
3753
            offset_size = sum(split_sizes[:i]) * np.prod(shape[split_dim + 1 :])
            if (
                tensor.stride() != strides
                or list(tensor.shape) != shape_i
                or tensor.untyped_storage().data_ptr() != data_ptr
                or tensor.storage_offset() != offset_size
            ):
3754
3755
3756
                noop_ok = False
                break
        if noop_ok:
3757
            ret = torch.Tensor().to(device=grad_outputs[0].device, dtype=grad_outputs[0].dtype)
3758
            new_shape = list(shape)
cyanguwa's avatar
cyanguwa committed
3759
            new_shape[split_dim] = sum(split_sizes)
3760
3761
3762
3763
3764
            ret.set_(
                grad_outputs[0].untyped_storage(),
                grad_outputs[0].storage_offset(),
                new_shape,
                strides,
3765
            )
cyanguwa's avatar
cyanguwa committed
3766
            return ret, None, None
3767

3768
        return torch.cat(grad_outputs, dim=split_dim), None, None
3769
3770
3771
3772
3773
3774
3775
3776
3777


class UnfusedDotProductAttention(torch.nn.Module):
    """Parallel attention w/o QKV and Proj Gemms
    BMM1 -> softmax + dropout -> BMM2
    """

    def __init__(
        self,
3778
        softmax_scale: float,
3779
        attention_type: str = "self",
3780
3781
3782
3783
3784
3785
        attention_dropout: float = 0.0,
        attention_dropout_ctx: Optional[Callable] = nullcontext,
        layer_number: Optional[int] = None,
    ) -> None:
        super().__init__()

3786
        self.softmax_scale = softmax_scale
3787
        self.attention_type = attention_type
3788
3789
3790
        self.attention_dropout_ctx = attention_dropout_ctx
        self.layer_number = layer_number

3791
        self.scale_mask_softmax = FusedScaleMaskSoftmax(attention_mask_func)
3792
3793
3794
3795
3796
3797

        # Dropout. Note that for a single iteration, this layer will generate
        # different outputs on different number of parallel partitions but
        # on average it should not be partition dependent.
        self.attention_dropout = torch.nn.Dropout(attention_dropout)

3798
3799
        # An FP16 training trick required for certain GPT-like models.
        self.apply_qk_layer_scaling = (
3800
3801
            bool(int(os.getenv("NVTE_APPLY_QK_LAYER_SCALING", "0"))) and layer_number is not None
        )
3802

3803
3804
3805
3806
3807
    def forward(
        self,
        query_layer: torch.Tensor,
        key_layer: torch.Tensor,
        value_layer: torch.Tensor,
3808
        qkv_layout: str = "sbh3d",
3809
3810
        cu_seqlens_q: Optional[torch.Tensor] = None,  # pylint: disable=unused-argument
        cu_seqlens_kv: Optional[torch.Tensor] = None,  # pylint: disable=unused-argument
3811
        attn_mask_type: str = "causal",
3812
        attention_mask: Optional[Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]] = None,
3813
        window_size: Optional[Tuple[int, int]] = None,
3814
3815
        core_attention_bias_type: str = "no_bias",
        core_attention_bias: Optional[torch.Tensor] = None,
3816
        alibi_slopes: Optional[torch.Tensor] = None,
3817
        inference_params: Optional[InferenceParams] = None,
3818
    ) -> torch.Tensor:
3819
        """Unfused attention fprop"""
3820
3821
3822
        assert (
            qkv_layout in QKVLayouts
        ), f"UnfusedDotProductAttention does not support qkv_layout = {qkv_layout}!"
3823
3824
3825
3826
3827
3828

        # get q_format and kv_format for training and inference
        qkv_format, q_format, _ = dpa_utils.get_qkv_format(qkv_layout, inference_params)
        if inference_params is not None and inference_params.is_paged:
            key_layer, value_layer = inference_params.convert_paged_to_nonpaged(self.layer_number)

3829
        if qkv_format == "bshd":
3830
            # convert to sbhd and use sbhd implementation for now
3831
3832
3833
            query_layer, key_layer, value_layer = [
                x.transpose(0, 1) for x in [query_layer, key_layer, value_layer]
            ]
3834
3835
3836
3837
3838
3839
3840
3841
3842
3843
3844
3845
3846
3847
3848
        if qkv_format == "sbhd_2bshd":
            key_layer, value_layer = [x.transpose(0, 1) for x in [key_layer, value_layer]]

        total_tokens, batch_size = None, None
        if qkv_format == "thd_2bshd":
            total_tokens, batch_size = query_layer.shape[0], key_layer.shape[0]
            query_layer = tex.convert_thd_to_bshd(
                query_layer,
                cu_seqlens_q,
                batch_size,
                inference_params.max_ctx_len,
            )
            query_layer, key_layer, value_layer = [
                x.transpose(0, 1) for x in [query_layer, key_layer, value_layer]
            ]
3849
3850
3851
3852
3853
        batch_size, max_seqlen_q, max_seqlen_kv = (
            query_layer.shape[1],
            query_layer.shape[0],
            key_layer.shape[0],
        )
3854

3855
3856
3857
3858
        if "padding" in attn_mask_type and attention_mask is None:
            attention_mask = dpa_utils.get_padding_mask(
                batch_size, cu_seqlens_q, cu_seqlens_kv, max_seqlen_q, max_seqlen_kv
            )
3859
3860
3861
3862
3863
3864
3865
3866
3867
        attn_mask_type, attention_mask, actual_seqlens_q, actual_seqlens_kv = (
            dpa_utils.get_full_mask(
                max_seqlen_q,
                max_seqlen_kv,
                attn_mask_type=attn_mask_type,
                attention_mask=attention_mask,
                window_size=window_size,
                attention_type=self.attention_type,
            )
3868
        )
3869

3870
        batch_size, seqlen = query_layer.shape[1], query_layer.shape[0]
3871
        apply_qk_layer_scaling = self.apply_qk_layer_scaling and key_layer.dtype == torch.float16
3872
3873
3874
3875
3876
3877
3878
3879
3880

        # [b, np, sq, sk]
        output_size = (
            query_layer.size(1),
            query_layer.size(2),
            query_layer.size(0),
            key_layer.size(0),
        )

3881
        if key_layer.shape[2] != query_layer.shape[2]:
3882
3883
3884
            assert (
                query_layer.shape[2] % key_layer.shape[2] == 0
            ), "The number of attention heads must be divisible by the number of GQA groups!"
3885
            key_layer = key_layer.repeat_interleave(
3886
3887
                int(query_layer.shape[2] / key_layer.shape[2]), dim=2
            )
3888
            value_layer = value_layer.repeat_interleave(
3889
3890
                int(query_layer.shape[2] / value_layer.shape[2]), dim=2
            )
3891

3892
        # [sq, b, np, hn] -> [sq, b * np, hn]
3893
        query_layer = query_layer.reshape(output_size[2], output_size[0] * output_size[1], -1)
3894
3895
3896
3897
3898
3899
3900
3901
        # [sk, b, np, hn] -> [sk, b * np, hn]
        key_layer = key_layer.reshape(output_size[3], output_size[0] * output_size[1], -1)

        # preallocting result tensor: [b * np, sq, sk]
        matmul_result = torch.empty(
            output_size[0] * output_size[1],
            output_size[2],
            output_size[3],
3902
            dtype=query_layer.dtype,
3903
3904
3905
            device=torch.cuda.current_device(),
        )

3906
        scale = self.softmax_scale
3907
        if apply_qk_layer_scaling:
3908
            scale /= self.layer_number
3909
3910

        # Raw attention scores. [b * np, sq, sk]
3911
3912
3913
3914
3915
3916
        if core_attention_bias_type == "no_bias":
            matmul_result = torch.baddbmm(
                matmul_result,
                query_layer.transpose(0, 1),  # [b * np, sq, hn]
                key_layer.transpose(0, 1).transpose(1, 2),  # [b * np, hn, sk]
                beta=0.0,
3917
                alpha=scale,
3918
            ).view(*output_size)
3919
3920
3921
3922
3923
3924
3925

        elif core_attention_bias_type == "pre_scale_bias":
            assert core_attention_bias is not None, "core_attention_bias should not be None!"
            matmul_result = torch.bmm(
                query_layer.transpose(0, 1),  # [b * np, sq, hn]
                key_layer.transpose(0, 1).transpose(1, 2),  # [b * np, hn, sk]
            )
3926
            matmul_result = matmul_result.view(*output_size) + core_attention_bias
3927
            matmul_result *= scale
3928

3929
3930
3931
3932
        elif core_attention_bias_type in ["post_scale_bias", "alibi"]:
            if core_attention_bias_type == "post_scale_bias":
                assert core_attention_bias is not None, "core_attention_bias should not be None!"
            if core_attention_bias_type == "alibi":
3933
3934
                _, core_attention_bias = dpa_utils.get_alibi(
                    _alibi_cache,
3935
3936
3937
                    output_size[1],
                    output_size[2],
                    output_size[3],
3938
3939
                    actual_seqlens_q=actual_seqlens_q if "padding" in attn_mask_type else None,
                    actual_seqlens_kv=actual_seqlens_kv if "padding" in attn_mask_type else None,
3940
3941
                    alibi_slopes=alibi_slopes,
                    bottom_right_alignment=attn_mask_type not in ["causal", "padding_causal"],
3942
                )
3943
3944
3945
3946
3947
            matmul_result = torch.baddbmm(
                matmul_result,
                query_layer.transpose(0, 1),  # [b * np, sq, hn]
                key_layer.transpose(0, 1).transpose(1, 2),  # [b * np, hn, sk]
                beta=0.0,
3948
                alpha=scale,
3949
            )
3950
3951
            matmul_result = (matmul_result.view(*output_size) + core_attention_bias).to(
                dtype=query_layer.dtype
3952
            )
3953
3954
3955

        # attention scores and attention mask [b, np, sq, sk]
        softmax_scale = self.layer_number if apply_qk_layer_scaling else None
3956
        attention_probs = self.scale_mask_softmax(
3957
            matmul_result, attention_mask, attn_mask_type, softmax_scale
3958
        )
3959

3960
3961
3962
3963
3964
        # mask out the pad positions in softmax results, mostly for the rows (pad tokens from q)
        # the columns (pad tokens from k) are already zeroed out during softmax
        if "padding" in attn_mask_type:
            attention_probs = attention_probs.masked_fill(attention_mask, 0)

3965
3966
3967
3968
3969
3970
3971
3972
3973
3974
3975
3976
3977
3978
3979
        # This is actually dropping out entire tokens to attend to, which might
        # seem a bit unusual, but is taken from the original Transformer paper.
        with self.attention_dropout_ctx():
            attention_probs = self.attention_dropout(attention_probs)

        # value_layer -> context layer.
        # [sk, b, np, hn] --> [b, np, sq, hn]
        output_size = (
            value_layer.size(1),
            value_layer.size(2),
            query_layer.size(0),
            value_layer.size(3),
        )

        # change view [sk, b * np, hn]
3980
        value_layer = value_layer.reshape(value_layer.size(0), output_size[0] * output_size[1], -1)
3981
3982

        # change view [b * np, sq, sk]
3983
        attention_probs = attention_probs.view(output_size[0] * output_size[1], output_size[2], -1)
3984
3985
3986
3987
3988
3989
3990

        # matmul: [b * np, sq, hn]
        context_layer = torch.bmm(attention_probs, value_layer.transpose(0, 1))

        # change view [b, np, sq, hn]
        context_layer = context_layer.view(*output_size)

3991
        if q_format == "sbhd":
3992
3993
            # [b, np, sq, hn] --> [sq, b, np, hn]
            context_layer = context_layer.permute(2, 0, 1, 3).contiguous()
3994

3995
3996
3997
            # [sq, b, np, hn] --> [sq, b, hp]
            context_layer = context_layer.view(seqlen, batch_size, -1)

3998
        if q_format == "bshd":
3999
4000
4001
4002
4003
            # [b, np, sq, hn] --> [b, sq, np, hn]
            context_layer = context_layer.permute(0, 2, 1, 3).contiguous()

            # [b, sq, np, hn] --> [b, sq, hp]
            context_layer = context_layer.view(batch_size, seqlen, -1)
4004

4005
4006
4007
4008
4009
4010
4011
4012
4013
4014
4015
4016
4017
4018
        if q_format == "thd":
            # [b, np, sq, hn] --> [b, sq, np, hn]
            context_layer = context_layer.permute(0, 2, 1, 3).contiguous()

            # [b, sq, np, hn] --> [tq, np, hn]
            context_layer = tex.convert_bshd_to_thd(
                context_layer,
                cu_seqlens_q,
                total_tokens,
            )

            # [tq, np, hn] --> [tq, hp]
            context_layer = context_layer.view(total_tokens, -1)

4019
4020
4021
4022
4023
        return context_layer


class _PrepareQKVForFA(torch.autograd.Function):
    """This class converts QKV from interleaved (s, b, ...) layout
4024
    to separate contiguous q, k, v tensors in (b, s, ...) layout."""
4025
4026

    @staticmethod
4027
4028
4029
4030
    def forward(
        _ctx: torch.autograd.function.FunctionCtx,  # unused
        query_layer: torch.Tensor,
        key_layer: torch.Tensor,
4031
        value_layer: torch.Tensor,
4032
    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
4033
        # pylint: disable=missing-function-docstring
4034
4035
4036
4037
4038
4039
4040
4041
4042
4043
4044
        # All inputs received are non-contiguous tensors.
        # The `query_layer` tensor is used to access the
        # full memory region of the QKV tensor.
        qkv = tex.fa_prepare_fwd(query_layer)
        q, k, v = split_tensor_along_dim(qkv, 0, 3)
        query_layer = torch.squeeze(q, 0)
        key_layer = torch.squeeze(k, 0)
        value_layer = torch.squeeze(v, 0)
        return query_layer, key_layer, value_layer

    @staticmethod
4045
4046
4047
4048
    def backward(
        _ctx: torch.autograd.function.FunctionCtx,  # unused
        dq: torch.Tensor,
        dk: torch.Tensor,
4049
        dv: torch.Tensor,
4050
    ) -> Tuple[Union[torch.Tensor, None], ...]:
4051
        # pylint: disable=missing-function-docstring
4052
4053
4054
4055
        dqkv = tex.fa_prepare_bwd(dq, dk, dv)
        dq, dk, dv = split_tensor_along_dim(dqkv, -1, 3)
        return dq, dk, dv

4056

4057
class FlashAttention(torch.nn.Module):
4058
    """Dot product attention, using HazyResearch flash-attn package:
4059
    https://github.com/Dao-AILab/flash-attention
4060
4061
4062
4063
    """

    def __init__(
        self,
4064
        softmax_scale: float,
4065
4066
        attention_dropout: float = 0.0,
        attention_dropout_ctx: Optional[Callable] = nullcontext,
4067
4068
        attention_type: str = "self",
        layer_number: Optional[int] = None,
4069
        deterministic: bool = False,
4070
4071
4072
    ) -> None:
        super().__init__()

4073
        if fa_utils.is_installed:
4074
            assert (
4075
4076
                fa_utils.version >= fa_utils.version_required
            ), f"FlashAttention minimum version {fa_utils.version_required} is required."
4077
            assert (
4078
4079
                fa_utils.version <= fa_utils.max_version
            ), f"FlashAttention maximum version {fa_utils.max_version} is supported."
4080

4081
        self.softmax_scale = softmax_scale
4082
4083
        self.attention_dropout_ctx = attention_dropout_ctx
        self.attention_dropout = attention_dropout
4084
4085
        self.attention_type = attention_type
        self.layer_number = 1 if layer_number is None else layer_number
4086
        self.deterministic = deterministic
4087
        self.logger = logging.getLogger("FlashAttention")
4088
        self.logger.setLevel(attn_log._log_level)
4089
        if not self.logger.hasHandlers():
4090
            self.logger.addHandler(attn_log._stream_handler)
4091
4092
4093
4094
4095
4096

    def forward(
        self,
        query_layer: torch.Tensor,
        key_layer: torch.Tensor,
        value_layer: torch.Tensor,
4097
        attention_mask: Optional[Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]] = None,
4098
4099
4100
        qkv_layout: str = "sbh3d",
        cu_seqlens_q: Optional[torch.Tensor] = None,
        cu_seqlens_kv: Optional[torch.Tensor] = None,
4101
4102
        max_seqlen_q: Optional[int] = None,
        max_seqlen_kv: Optional[int] = None,
4103
        attn_mask_type: str = "causal",
4104
        window_size: Optional[Tuple[int, int]] = None,
4105
        alibi_slopes: Optional[torch.Tensor] = None,
4106
        cp_group: Optional[Union[dist_group_type, List[dist_group_type]]] = None,
4107
        cp_global_ranks: List[int] = None,
4108
        cp_stream: torch.cuda.Stream = None,
4109
        cp_comm_type: str = "p2p",
4110
4111
        fp8: bool = False,
        fp8_meta: Optional[Dict[str, Any]] = None,
4112
        quantizers=None,
4113
4114
        inference_params: Optional[InferenceParams] = None,
        flash_attention_backend: Optional[PkgVersion] = PkgVersion("0"),
4115
4116
4117
    ) -> torch.Tensor:
        """flash-attn fprop"""

4118
4119
4120
4121
        assert all(
            x.dtype in [torch.float16, torch.bfloat16] or isinstance(x, Float8Tensor)
            for x in [query_layer, key_layer, value_layer]
        ), "FlashAttention only supports FP16 and BF16 data types, or Float8Tensors."
4122
4123
        assert (
            query_layer.is_cuda and key_layer.is_cuda and value_layer.is_cuda
4124
        ), "FlashAttention currently only supports CUDA tensors."
4125
4126
        assert (
            qkv_layout in QKVLayouts
4127
        ), f"FlashAttention does not support qkv_layout = {qkv_layout}!"
4128

4129
4130
4131
4132
4133
4134
        cp_size = 1
        if isinstance(cp_group, dist_group_type):
            cp_size = get_distributed_world_size(cp_group)
        elif isinstance(cp_group, list):
            for group in cp_group:
                cp_size *= get_distributed_world_size(group)
4135
        context_parallel = cp_size > 1
4136

4137
4138
        # get q_format and kv_format for training and inference
        qkv_format, q_format, kv_format = dpa_utils.get_qkv_format(qkv_layout, inference_params)
4139

4140
        # convert q, k, v to bshd if they are in sbhd; qkv_format doesn't change
4141
4142
4143
4144
4145
4146
4147
4148
4149
4150
4151
4152
4153
        if all(not isinstance(x, Float8Tensor) for x in [query_layer, key_layer, value_layer]):
            if qkv_format == "sbhd":
                # For now just 128, will make it more general in the future
                if (
                    query_layer.shape[-1] == 128
                    and query_layer.shape[0] * query_layer.shape[1] >= 512
                    and qkv_layout == "sbh3d"
                ):
                    query_layer, key_layer, value_layer = _PrepareQKVForFA.apply(
                        query_layer, key_layer, value_layer
                    )
                else:
                    query_layer, key_layer, value_layer = [
4154
4155
                        x.transpose(0, 1).contiguous()
                        for x in (query_layer, key_layer, value_layer)
4156
                    ]
4157
4158
            elif q_format == "sbhd" and kv_format == "bshd":
                query_layer = query_layer.transpose(0, 1).contiguous()
4159
            if context_parallel:
4160
                query_layer, key_layer, value_layer = [
4161
4162
4163
4164
4165
                    x.contiguous() for x in (query_layer, key_layer, value_layer)
                ]
        else:
            if qkv_format == "sbhd":
                query_layer._data, key_layer._data, value_layer._data = [
4166
                    x.transpose(0, 1).contiguous()
4167
4168
                    for x in (query_layer._data, key_layer._data, value_layer._data)
                ]
4169
                query_layer, key_layer, value_layer = [
4170
                    Float8Tensor.make_like(x, data=x._data, shape=x._data.shape)
4171
4172
                    for x in (query_layer, key_layer, value_layer)
                ]
4173
4174
4175
4176
4177
            elif q_format == "sbhd" and kv_format == "bshd":
                query_layer._data = query_layer._data.transpose(0, 1).contiguous()
                query_layer = Float8Tensor.make_like(
                    query_layer, data=query_layer._data, shape=query_layer._data.shape
                )
4178
            if context_parallel:
4179
4180
                query_layer._data, key_layer._data, value_layer._data = [
                    x.contiguous() for x in (query_layer._data, key_layer._data, value_layer._data)
4181
                ]
4182

4183
4184
4185
4186
4187
4188
4189
4190
        # get batch_size, max_seqlen and cu_seqlens
        batch_size, context_len = None, None
        if inference_params is None:
            if qkv_format in ["sbhd", "bshd"]:
                batch_size = query_layer.shape[0]
                max_seqlen_q, max_seqlen_kv = query_layer.shape[1], key_layer.shape[1]
                max_seqlen_q *= cp_size
                max_seqlen_kv *= cp_size
4191

4192
4193
4194
4195
                if "padding" in attn_mask_type:
                    assert (
                        not context_parallel
                    ), "Padding mask not supported with context parallelism!"
4196

4197
4198
4199
4200
4201
                    # [b * s, h, d]
                    query_layer, key_layer, value_layer = [
                        x.reshape(x.shape[0] * x.shape[1], *x.shape[2:])
                        for x in [query_layer, key_layer, value_layer]
                    ]
4202

4203
                    if self.attention_type == "self":
4204
                        assert (
4205
4206
4207
4208
4209
4210
4211
4212
4213
4214
4215
4216
4217
4218
                            max_seqlen_q == max_seqlen_kv
                        ), "Maximum sequence length for Q and KV should be the same."
                        if cu_seqlens_q is None:
                            assert (
                                attention_mask is not None
                            ), "Please provide attention_mask for padding!"
                            cu_seqlens_q, indices_q = dpa_utils.get_cu_seqlens_and_indices(
                                attention_mask
                            )
                        else:
                            indices_q = dpa_utils.get_indices(max_seqlen_q, cu_seqlens_q)
                        cu_seqlens_kv = cu_seqlens_q
                        query_layer, key_layer, value_layer = dpa_utils.PackTensors.apply(
                            indices_q, query_layer, key_layer, value_layer
4219
                        )
4220
                    else:
4221
4222
4223
4224
4225
4226
4227
4228
4229
4230
4231
4232
4233
4234
4235
4236
4237
                        if cu_seqlens_q is None or cu_seqlens_kv is None:
                            assert (
                                attention_mask is not None
                            ), "Please provide attention_mask for padding!"
                            cu_seqlens_q, indices_q = dpa_utils.get_cu_seqlens_and_indices(
                                attention_mask[0]
                            )
                            cu_seqlens_kv, indices_kv = dpa_utils.get_cu_seqlens_and_indices(
                                attention_mask[1]
                            )
                        else:
                            indices_q = dpa_utils.get_indices(max_seqlen_q, cu_seqlens_q)
                            indices_kv = dpa_utils.get_indices(max_seqlen_kv, cu_seqlens_kv)
                        query_layer = dpa_utils.PackTensors.apply(indices_q, query_layer)
                        key_layer, value_layer = dpa_utils.PackTensors.apply(
                            indices_kv, key_layer, value_layer
                        )
4238
                else:
4239
4240
4241
4242
4243
4244
                    # Cumulative sequence lengths for unpadded data
                    if cu_seqlens_q is None:
                        cu_seqlens_q = dpa_utils.get_full_cu_seqlens(
                            batch_size,
                            max_seqlen_q,
                            query_layer.device,
4245
                        )
4246
4247
4248
4249
4250
                    if cu_seqlens_kv is None:
                        cu_seqlens_kv = dpa_utils.get_full_cu_seqlens(
                            batch_size,
                            max_seqlen_kv,
                            key_layer.device,
4251
                        )
4252
4253
4254
4255
4256
4257
4258
4259
4260
4261
4262
4263
4264
4265
4266
4267
4268
4269
4270
4271
4272
4273
4274
            elif qkv_format == "thd":
                assert (
                    cu_seqlens_q is not None and cu_seqlens_kv is not None
                ), "cu_seqlens_q and cu_seqlens_kv can not be None when qkv_format = thd!"
                if max_seqlen_q is None:
                    seqlens_q = cu_seqlens_q[1:] - cu_seqlens_q[:-1]
                    max_seqlen_q = seqlens_q.max().item()
                if max_seqlen_kv is None:
                    seqlens_kv = cu_seqlens_kv[1:] - cu_seqlens_kv[:-1]
                    max_seqlen_kv = seqlens_kv.max().item()
        else:
            if qkv_format in ["sbhd_2bshd", "bshd"]:
                # q is in bshd in both cases from conversion above or the original input
                batch_size, context_len = query_layer.shape[:2]
                cu_seqlens_q = cu_seqlens_q[: batch_size + 1]
                cu_seqlens_kv = cu_seqlens_kv[: batch_size + 1]
                # convert from bshd to thd_2bshd for flash_attn_varlen_func/_with_kvcache;
                # kernel assumes tensor is contiguous
                if isinstance(query_layer, Float8Tensor):
                    query_layer._data = tex.convert_bshd_to_thd(
                        query_layer._data,
                        cu_seqlens_q,
                        batch_size * context_len,
4275
                    )
4276
4277
                    query_layer = Float8Tensor.make_like(
                        query_layer, data=query_layer._data, shape=query_layer._data.shape
4278
                    )
4279
4280
4281
4282
4283
                else:
                    query_layer = tex.convert_bshd_to_thd(
                        query_layer,
                        cu_seqlens_q,
                        batch_size * context_len,
4284
                    )
4285

4286
4287
4288
        use_flash_attn_3 = False
        if flash_attention_backend is not None and flash_attention_backend > PkgVersion("3.0.0b"):
            use_flash_attn_3 = True
4289
4290
4291
        if context_parallel and all(
            not isinstance(x, Float8Tensor) for x in [query_layer, key_layer, value_layer]
        ):
4292
4293
4294
            assert (
                alibi_slopes is None
            ), "Alibi slope bias addition is not supported with context parallelism."
4295
            with self.attention_dropout_ctx():
4296
                output = attn_forward_func_with_cp(
4297
4298
4299
4300
4301
4302
4303
4304
                    self.training,
                    query_layer,
                    key_layer,
                    value_layer,
                    cu_seqlens_q,
                    cu_seqlens_kv,
                    max_seqlen_q,
                    max_seqlen_kv,
4305
4306
                    cu_seqlens_q if qkv_format == "thd" else None,
                    cu_seqlens_kv if qkv_format == "thd" else None,
4307
                    self.attention_dropout if self.training else 0.0,
4308
4309
4310
                    cp_group,
                    cp_global_ranks,
                    cp_stream,
4311
                    cp_comm_type,
4312
                    softmax_scale=self.softmax_scale,
4313
                    qkv_format="bshd" if qkv_format == "sbhd" else qkv_format,
4314
                    attn_mask_type=attn_mask_type,
4315
                    deterministic=self.deterministic,
4316
                    window_size=window_size,
4317
                    quantizers=quantizers,
4318
                    pad_between_seqs=False,
4319
                    use_flash_attn_3=use_flash_attn_3,
4320
4321
                )
        else:
4322
4323

            from .cpu_offload import CPUOffloadEnabled
4324

4325
4326
4327
4328
4329
4330
            if CPUOffloadEnabled:
                tensor_list = [query_layer, key_layer, value_layer, cu_seqlens_q, cu_seqlens_kv]
                for tensor in tensor_list:
                    if tensor is not None:
                        tensor.activation_offloading = True

4331
            with self.attention_dropout_ctx():
4332
4333
4334
4335
4336
4337
4338
4339
4340
4341
4342
4343
                #       | API                     | use cases
                # ----------------------------------------------------------------------
                # FA v2 | flash_attn_func         | bshd/sbhd + not padding
                #       | flash_attn_varlen_func  | bshd/sbhd + padding
                #       |                         | thd + padding
                #       |                         | KV cache (not-paged/paged), i.e.
                #       |                         |     bshd/sbhd/thd + padding
                # FA v3 | flash_attn_func         | bshd/sbhd + not padding
                #       | flash_attn_varlen_func  | bshd/sbhd + padding
                #       |                         | thd + padding
                #       | flash_attn_with_kvcache | KV cache (not-paged/paged), i.e.
                #       |                         |     bshd/sbhd/thd + padding
4344
4345
4346
                fa_optional_forward_args_thd = []
                if qkv_format in ["bshd", "sbhd"] and "padding" not in attn_mask_type:
                    func = (
4347
4348
4349
4350
4351
4352
4353
4354
4355
4356
4357
4358
4359
4360
4361
4362
4363
4364
4365
4366
4367
4368
4369
4370
4371
4372
4373
4374
4375
4376
4377
4378
4379
4380
4381
4382
4383
4384
4385
4386
                        flash_attn_func if not use_flash_attn_3 else flash_attn_func_v3
                    )  # pylint: disable=possibly-used-before-assignment
                else:
                    if not use_flash_attn_3:
                        func = flash_attn_varlen_func
                    elif inference_params is None:
                        func = flash_attn_varlen_func_v3  # pylint: disable=possibly-used-before-assignment
                    else:
                        func = flash_attn_with_kvcache_v3  # pylint: disable=possibly-used-before-assignment
                    if not use_flash_attn_3 or inference_params is None:
                        fa_optional_forward_args_thd.append(cu_seqlens_q)
                        fa_optional_forward_args_thd.append(cu_seqlens_kv)
                        fa_optional_forward_args_thd.append(max_seqlen_q)
                        fa_optional_forward_args_thd.append(max_seqlen_kv)
                if not use_flash_attn_3:
                    fa_optional_forward_kwargs = {}
                    if fa_utils.v2_3_plus:
                        fa_optional_forward_kwargs["window_size"] = window_size
                    if fa_utils.v2_4_plus:
                        fa_optional_forward_kwargs["alibi_slopes"] = alibi_slopes
                    if fa_utils.v2_4_1_plus:
                        fa_optional_forward_kwargs["deterministic"] = self.deterministic
                    if inference_params is not None:
                        # use block_table kwarg to support thd_2bshd for non-paged
                        fa_optional_forward_kwargs["block_table"] = (
                            inference_params.cache_manager.page_table[:batch_size]
                            if inference_params.is_paged
                            else inference_params.cache_manager.batch_indices_post_step.unsqueeze(
                                1
                            )[:batch_size]
                        )
                    output = func(
                        query_layer,
                        key_layer,
                        value_layer,
                        *fa_optional_forward_args_thd,
                        self.attention_dropout if self.training else 0.0,
                        softmax_scale=self.softmax_scale,
                        causal="causal" in attn_mask_type,
                        **fa_optional_forward_kwargs,
4387
                    )
4388
                else:
4389
4390
                    fa_3_optional_forward_kwargs = {}
                    fa_3_optional_forward_kwargs["window_size"] = window_size
4391
4392
4393
4394
4395
4396
4397
4398
4399
4400
4401
4402
                    if inference_params is None:
                        fa_3_optional_forward_kwargs["deterministic"] = self.deterministic
                    else:
                        fa_3_optional_forward_kwargs["cu_seqlens_q"] = cu_seqlens_q
                        fa_3_optional_forward_kwargs["max_seqlen_q"] = max_seqlen_q
                        cache_seqlens = cu_seqlens_kv[1:] - cu_seqlens_kv[:-1]
                        fa_3_optional_forward_kwargs["cache_seqlens"] = cache_seqlens
                        # flash_attn_with_kvcache accepts thd_2bshd for non-paged
                        if inference_params.is_paged:
                            fa_3_optional_forward_kwargs["page_table"] = (
                                inference_params.cache_manager.page_table[:batch_size]
                            )
4403
                    if fp8:
4404
                        QKV_quantizer = quantizers["scaling_fwd"][META_QKV]
4405
                        torch_dtype = get_fp8_torch_dtype(fp8_meta["recipe"], fprop_tensor=True)
4406
                        torch_orig_dtype = query_layer.dtype
4407
4408
4409
4410
4411
4412
4413
4414
4415
4416
4417

                        def convert_to_torch_float8(tensor, dtype):
                            out = torch.Tensor().to(device=tensor.device, dtype=dtype)
                            out.set_(
                                tensor._data.untyped_storage(),
                                tensor._data.storage_offset(),
                                tensor._data.shape,
                                tensor._data.stride(),
                            )
                            return out

4418
4419
4420
4421
4422
                        # "fp8_mha" decides outputs in fp8, while inputs are inferred from
                        # the real dtype
                        assert isinstance(key_layer, query_layer.__class__) and isinstance(
                            value_layer, query_layer.__class__
                        ), "q, k, and v must have the same type."
4423
                        if not isinstance(query_layer, Float8Tensor):
4424
                            query_layer, key_layer, value_layer = (
4425
                                QKV_quantizer(x) for x in [query_layer, key_layer, value_layer]
4426
                            )
4427
4428
4429
4430
                        batch_size = cu_seqlens_q.shape[0] - 1
                        num_heads_k = key_layer.shape[-2]
                        fa_3_optional_forward_kwargs["q_descale"] = (
                            query_layer._scale_inv.unsqueeze(0).repeat(batch_size, num_heads_k)
4431
                        )
4432
                        fa_3_optional_forward_kwargs["k_descale"] = key_layer._scale_inv.unsqueeze(
4433
                            0
4434
4435
4436
                        ).repeat(batch_size, num_heads_k)
                        fa_3_optional_forward_kwargs["v_descale"] = (
                            value_layer._scale_inv.unsqueeze(0).repeat(batch_size, num_heads_k)
4437
                        )
4438
4439
4440
                        query_layer, key_layer, value_layer = (
                            convert_to_torch_float8(x, torch_dtype)
                            for x in [query_layer, key_layer, value_layer]
4441
                        )
4442
                    try:
4443
                        output = func(
4444
4445
4446
4447
4448
4449
4450
4451
                            query_layer,
                            key_layer,
                            value_layer,
                            *fa_optional_forward_args_thd,
                            softmax_scale=self.softmax_scale,
                            causal="causal" in attn_mask_type,
                            **fa_3_optional_forward_kwargs,
                        )
4452
4453
                        if isinstance(output, (List, Tuple)):
                            output = output[0]
4454
                    except TypeError as e:
4455
                        if fa_utils.v3_0_0_beta:
4456
4457
4458
4459
                            e.args = (
                                e.args[0]
                                + ". Please update your flash-attn v3 (beta) installation as it "
                                + "may have added more supported arguments to its API. \n"
4460
                                + fa_utils.v3_installation_steps,
4461
4462
4463
4464
4465
4466
4467
4468
                            ) + e.args[1:]
                        raise

                    if fp8:
                        output = output.to(dtype=torch_orig_dtype)
                    if fp8 and fp8_meta["recipe"].fp8_mha:
                        O_quantizer = quantizers["scaling_fwd"][META_O]
                        output = O_quantizer(output)
4469

4470
4471
4472
4473
4474
4475
4476
4477
4478
4479
4480
4481
4482
4483
4484
4485
4486
4487
4488
4489
4490
        if inference_params is None:
            if qkv_format in ["sbhd", "bshd"] and "padding" in attn_mask_type:
                output = dpa_utils.UnpackTensor.apply(indices_q, batch_size * max_seqlen_q, output)
        elif qkv_format in ["bshd", "sbhd_2bshd"]:
            # all KV caching cases use thd_2bshd for calculation
            # convert results back to bshd from thd_2bshd
            if isinstance(query_layer, Float8Tensor):
                output._data = tex.convert_thd_to_bshd(
                    output._data,
                    cu_seqlens_q,
                    batch_size,
                    context_len,
                )
                output = Float8Tensor.make_like(output, data=output._data, shape=output._data.shape)
            else:
                output = tex.convert_thd_to_bshd(
                    output,
                    cu_seqlens_q,
                    batch_size,
                    context_len,
                )
4491

4492
        if q_format == "sbhd":
4493
4494
4495
4496
4497
4498
4499
4500
4501
4502
4503
4504
4505
4506
            # (bs)hd -> bs(hd) -> sb(hd)
            if fp8 and fp8_meta["recipe"].fp8_mha:
                output_data = (
                    output._data.reshape(batch_size, max_seqlen_q // cp_size, -1)
                    .transpose(0, 1)
                    .contiguous()
                )
                output = Float8Tensor.make_like(
                    output,
                    data=output_data,
                    shape=output_data.shape,
                )
            else:
                output = output.view(batch_size, max_seqlen_q // cp_size, -1).transpose(0, 1)
4507
        elif q_format == "bshd":
4508
4509
            # (bs)hd -> bs(hd)
            output = output.reshape(batch_size, max_seqlen_q // cp_size, -1)
4510
        elif q_format == "thd":
4511
4512
4513
4514
4515
4516
4517
4518
4519
4520
4521
4522
4523
4524
4525
4526
4527
4528
4529
4530
4531
4532
4533
4534
4535
4536
4537
4538
4539
4540
4541
4542
            # thd -> t(hd)
            output = output.reshape(output.shape[0], -1)

        return output.contiguous()


def _combine_tensors(
    tensors: List[torch.Tensor],
    dim: int,
) -> torch.Tensor:
    """Combine tensors along a particular dimension"""

    num_tensors = len(tensors)
    new_shape = list(tensors[0].shape)
    new_shape.insert(dim, num_tensors)
    if isinstance(tensors[0], Float8Tensor):
        new_stride = list(tensors[0]._data.stride())
        new_stride.insert(dim, int(new_stride[dim - 1] / num_tensors))
        combined_tensor = torch.Tensor().to(device=tensors[0].device, dtype=tensors[0]._data.dtype)
        combined_tensor.set_(
            tensors[0]._data.untyped_storage(),
            tensors[0]._data.storage_offset(),
            new_shape,
            new_stride,
        )
        combined_tensor = Float8Tensor.make_like(tensors[0], data=combined_tensor, shape=new_shape)
    else:
        new_stride = list(tensors[0].stride())
        new_stride.insert(dim, int(new_stride[dim - 1] / num_tensors))
        combined_tensor = torch.Tensor().to(device=tensors[0].device, dtype=tensors[0].dtype)
        combined_tensor.set_(
            tensors[0].untyped_storage(), tensors[0].storage_offset(), new_shape, new_stride
4543
4544
        )

4545
4546
    return combined_tensor

4547

4548
4549
4550
4551
class FusedAttnFunc(torch.autograd.Function):
    """Function for FusedAttention with separate Q, K, V tensors"""

    @staticmethod
4552
4553
4554
4555
4556
4557
4558
    def forward(
        ctx,
        is_training,
        max_seqlen_q,
        max_seqlen_kv,
        cu_seqlens_q,
        cu_seqlens_kv,
4559
4560
        cu_seqlens_q_padded,
        cu_seqlens_kv_padded,
4561
4562
        page_table_k,
        page_table_v,
4563
4564
4565
4566
4567
4568
4569
4570
4571
4572
        q,
        k,
        v,
        attn_bias,
        attn_scale,
        dropout_p,
        fast_zero_fill,
        qkv_layout,
        attn_bias_type,
        attn_mask_type,
4573
        window_size,
4574
4575
4576
4577
4578
        rng_gen,
        fused_attention_backend,
        use_FAv2_bwd,
        fp8,
        fp8_meta,
4579
        quantizers,
4580
        deterministic,
4581
    ):
4582
        # pylint: disable=missing-function-docstring
4583
        # "fp8_mha" decides outputs in fp8, while inputs are inferred from the real dtype
4584
        is_input_fp8 = False
4585
        is_output_fp8 = fp8_meta["recipe"].fp8_mha if "recipe" in fp8_meta else False
4586
4587
4588
4589

        # FP16/BF16 attn:                  fake_dtype = torch.float16 or torch.bfloat16
        # FP8 attn, is_output_fp8 = False: fake_dtype = torch.float16 or torch.bfloat16
        # FP8 attn, is_output_fp8 = True:  fake_dtype = torch.float8_e4m3fn
4590
4591
4592
        fake_dtype = q.dtype

        QKV_quantizer, O_quantizer, S_quantizer, dQKV_quantizer, dO_quantizer, dP_quantizer = (
4593
            dpa_utils.get_attention_quantizers(fp8, quantizers, cp_specific_quantizers=False)
4594
        )
4595
4596
        if fp8:
            fused_attention_backend = FusedAttnBackend["FP8"]
4597
4598
4599
            assert isinstance(k, q.__class__) and isinstance(
                v, q.__class__
            ), "q, k, and v must have the same type."
4600

4601
            is_input_fp8 = isinstance(q, Float8Tensor)
4602
            q_fp8, k_fp8, v_fp8 = None, None, None
4603
            if is_input_fp8:
4604
                q_fp8, k_fp8, v_fp8 = q, k, v
4605
4606
            else:
                # 1: qkv packed, 2: kv packed, 3: qkv separate
4607
                qkv_group = len(qkv_layout.replace("paged_kv_", "").split("_"))
4608
4609
4610
4611
4612
4613
4614
4615
4616
4617
4618
4619
4620
4621
4622
4623
4624
4625
4626
4627
                match qkv_group:
                    case 1:
                        dim = qkv_layout.find("3")
                        qkv = _combine_tensors([q, k, v], dim)
                        qkv_c = qkv.view(-1, qkv.shape[-3] * qkv.shape[-2] * qkv.shape[-1])
                        qkv_fp8 = QKV_quantizer(qkv)
                        q_fp8, k_fp8, v_fp8 = _SplitAlongDim.apply(qkv_fp8, dim, [1, 1, 1], True)
                    case 2:
                        q_fp8 = QKV_quantizer(q)
                        dim = qkv_layout.split("_")[1].find("2")
                        kv = _combine_tensors([k, v], dim)
                        kv_c = kv.view(-1, kv.shape[-3] * kv.shape[-2] * kv.shape[-1])
                        kv_fp8 = QKV_quantizer(kv_c)
                        k_fp8, v_fp8 = _SplitAlongDim.apply(kv_fp8, dim, [1, 1], True)
                    case 3:
                        q_fp8 = QKV_quantizer(q)
                        k_fp8 = QKV_quantizer(k)
                        v_fp8 = QKV_quantizer(v)
                    case _:
                        raise "Invalid qkv_layout " + qkv_layout
4628
            # q_fp8, k_fp8, v_fp8, out_fp8: torch.float8_e4m3fn
4629
            out_fp8, aux_ctx_tensors = fused_attn_fwd(
4630
4631
4632
4633
4634
4635
4636
4637
                is_training,
                max_seqlen_q,
                max_seqlen_kv,
                cu_seqlens_q,
                cu_seqlens_kv,
                q_fp8,
                k_fp8,
                v_fp8,
4638
                fake_dtype,
4639
4640
                fused_attention_backend,
                attn_bias,
4641
4642
                cu_seqlens_q_padded,
                cu_seqlens_kv_padded,
4643
4644
                None,
                None,
4645
4646
                S_quantizer,
                O_quantizer,
4647
4648
4649
4650
4651
4652
                attn_scale,
                dropout_p,
                fast_zero_fill,
                qkv_layout,
                attn_bias_type,
                attn_mask_type,
4653
                window_size,
4654
4655
                rng_gen,
            )
4656
            if is_output_fp8:
4657
                out_ret = out_fp8
4658
            else:
4659
                out_ret = out_fp8.dequantize().view(out_fp8.shape)
4660
4661
            # is_output_fp8 = False: out_save.dtype = torch.float16 or torch.bfloat16
            # is_output_fp8 = True:  out_save.dtype = torch.float8_e4m3fn
4662
4663
            out_save = out_ret

4664
            if not int(os.getenv("NVTE_FP8_DPA_BWD", "1")):
4665
                # 1: qkv packed, 2: kv packed, 3: qkv separate
4666
                if is_input_fp8:
4667
                    qkv_group = len(qkv_layout.replace("paged_kv_", "").split("_"))
4668
4669
4670
4671
                    if qkv_group == 1:
                        dim = qkv_layout.find("3")
                        qkv = _combine_tensors([q, k, v], dim)
                        qkv_c = qkv.view(-1, qkv.shape[-3] * qkv.shape[-2] * qkv.shape[-1])
4672
4673
                        qkv_no_fp8 = qkv_c.dequantize().view(qkv.shape)
                        q, k, v = _SplitAlongDim.apply(qkv_no_fp8, dim, [1, 1, 1], True)
4674
                    if qkv_group == 2:
4675
                        q = q.dequantize()
4676
                        dim = qkv_layout.replace("paged_kv_", "").split("_")[1].find("2")
4677
4678
                        kv = _combine_tensors([k, v], dim)
                        kv_c = kv.view(-1, kv.shape[-3] * kv.shape[-2] * kv.shape[-1])
4679
4680
                        kv_no_fp8 = kv.dequantize()
                        k, v = _SplitAlongDim.apply(kv_no_fp8, dim, [1, 1], True)
4681
                    if qkv_group == 3:
4682
4683
4684
                        q = q.dequantize()
                        k = k.dequantize()
                        v = v.dequantize()
4685
                if is_output_fp8:
4686
4687
4688
                    out_save = out_fp8.dequantize()

            fp8_tensors = (q_fp8, k_fp8, v_fp8, out_fp8)
4689
        else:
4690
            # q, k, v, out_ret: torch.float16 or torch.bfloat16
4691
            out_ret, aux_ctx_tensors = fused_attn_fwd(
4692
4693
4694
4695
4696
4697
4698
4699
                is_training,
                max_seqlen_q,
                max_seqlen_kv,
                cu_seqlens_q,
                cu_seqlens_kv,
                q,
                k,
                v,
4700
                fake_dtype,
4701
4702
                fused_attention_backend,
                attn_bias,
4703
4704
                cu_seqlens_q_padded,
                cu_seqlens_kv_padded,
4705
4706
                page_table_k,
                page_table_v,
4707
4708
                None,  # s_quantizer
                None,  # o_quantizer
4709
4710
4711
4712
4713
4714
                attn_scale,
                dropout_p,
                fast_zero_fill,
                qkv_layout,
                attn_bias_type,
                attn_mask_type,
4715
                window_size,
4716
4717
                rng_gen,
            )
4718
            out_save = out_ret
4719
            fp8_tensors = (None, None, None, None)
4720

4721
4722
        ctx.fp8 = fp8 and int(os.getenv("NVTE_FP8_DPA_BWD", "1"))

4723
        from .cpu_offload import CPUOffloadEnabled
4724

4725
        if CPUOffloadEnabled:
4726
4727
4728
4729
4730
4731
4732
            if ctx.fp8:
                tensor_list = fp8_tensors
            else:
                tensor_list = [q, k, v, out_save]

            tensor_list.extend(aux_ctx_tensors)

4733
            qkv_layout = "sbhd_sbhd_sbhd"
4734
4735
4736
4737
            for tensor in tensor_list:
                if tensor is not None:
                    tensor.activation_offloading = True

4738
4739
        ctx.is_input_fp8 = is_input_fp8
        ctx.is_output_fp8 = is_output_fp8
4740
        qkvo_tensors = (q, k, v, out_save) if not ctx.fp8 else (None, None, None, None)
4741
4742
        tensors_to_save, tensor_objects = prepare_for_saving(
            *fp8_tensors,
4743
4744
4745
            *qkvo_tensors,
            cu_seqlens_q,
            cu_seqlens_kv,
4746
4747
            cu_seqlens_q_padded,
            cu_seqlens_kv_padded,
4748
4749
            *aux_ctx_tensors,
        )
4750
4751
        ctx.save_for_backward(*tensors_to_save)
        ctx.tensor_objects = tensor_objects
4752
        ctx.fp8_meta = fp8_meta
4753
4754
4755
4756
4757

        ctx.dQKV_quantizer = dQKV_quantizer
        ctx.dO_quantizer = dO_quantizer
        ctx.dP_quantizer = dP_quantizer
        ctx.S_quantizer = S_quantizer
4758
4759
4760
        if ctx.fp8:
            ctx.S_quantizer = S_quantizer.copy()
            ctx.S_quantizer.scale = S_quantizer.scale.clone()
4761

4762
4763
4764
4765
4766
4767
4768
4769
        ctx.max_seqlen_q = max_seqlen_q
        ctx.max_seqlen_kv = max_seqlen_kv
        ctx.attn_scale = attn_scale
        ctx.dropout_p = dropout_p
        ctx.fast_zero_fill = fast_zero_fill
        ctx.qkv_layout = qkv_layout
        ctx.attn_bias_type = attn_bias_type
        ctx.attn_mask_type = attn_mask_type
4770
        ctx.window_size = window_size
4771
        ctx.fused_attention_backend = (
4772
            fused_attention_backend if ctx.fp8 else FusedAttnBackend["F16_arbitrary_seqlen"]
4773
        )
4774
        ctx.use_FAv2_bwd = use_FAv2_bwd
4775
        ctx.deterministic = deterministic
4776

4777
        return out_ret
4778
4779
4780

    @staticmethod
    def backward(ctx, d_out):
4781
        # pylint: disable=missing-function-docstring
4782
        if ctx.is_output_fp8:
4783
4784
4785
            assert isinstance(
                d_out, Float8Tensor
            ), "Gradient of the DPA output must be in Float8Tensor type for FP8 MHA."
4786

4787
4788
4789
4790
4791
        # FP16/BF16 attn:                  fake_dtype = torch.float16 or torch.bfloat16
        # FP8 attn, is_output_fp8 = False: fake_dtype = torch.float16 or torch.bfloat16
        # FP8 attn, is_output_fp8 = True:  fake_dtype = torch.float8_e5m2
        fake_dtype = d_out.dtype

4792
        d_out = d_out.contiguous()
4793
        (
4794
4795
4796
4797
            q_fp8,
            k_fp8,
            v_fp8,
            out_fp8,
4798
4799
4800
4801
4802
4803
            q,
            k,
            v,
            out,
            cu_seqlens_q,
            cu_seqlens_kv,
4804
4805
            cu_seqlens_q_padded,
            cu_seqlens_kv_padded,
4806
4807
4808
4809
4810
            *other_tensors,
        ) = restore_from_saved(ctx.tensor_objects, ctx.saved_tensors)

        aux_ctx_tensors = other_tensors

4811
4812
        if not aux_ctx_tensors[0].is_contiguous():
            aux_ctx_tensors[0] = aux_ctx_tensors[0].contiguous()
4813
        rest = [None]
4814
        if ctx.use_FAv2_bwd:
4815
            softmax_lse, rng_state = aux_ctx_tensors
4816
4817
4818
            dq = torch.empty_like(q)
            dk = torch.empty_like(k)
            dv = torch.empty_like(v)
4819
            d_out, q, k, v, out = [maybe_contiguous(x) for x in (d_out, q, k, v, out)]
4820
            flash_attn_cuda_bwd(
4821
4822
4823
4824
4825
4826
4827
4828
4829
4830
4831
4832
4833
4834
4835
4836
4837
4838
4839
                d_out,
                q,
                k,
                v,
                out,
                softmax_lse,
                dq,
                dk,
                dv,
                cu_seqlens_q,
                cu_seqlens_kv,
                ctx.max_seqlen_q,
                ctx.max_seqlen_kv,
                ctx.dropout_p,
                ctx.attn_scale,
                False,
                "causal" in ctx.attn_mask_type,
                None,
                rng_state,
4840
            )
4841
4842
4843
            dq = dq[..., : d_out.shape[-1]]
            dk = dk[..., : d_out.shape[-1]]
            dv = dv[..., : d_out.shape[-1]]
4844
        else:
4845
4846
            with torch.cuda.nvtx.range("_FusedAttn"):
                if ctx.fp8:
4847
                    if ctx.is_output_fp8:
4848
4849
                        d_out_fp8 = d_out
                    else:
4850
                        d_out_fp8 = ctx.dO_quantizer(d_out)
4851
4852
4853
                    dqkv_dtype = TE_DType[d_out_fp8._data.dtype]
                    # q_fp8, k_fp8, v_fp8, out_fp8:      torch.float8_e4m3fn
                    # d_out_fp8, dq_fp8, dk_fp8, dv_fp8: torch.float8_e5m2
4854
                    dq_fp8, dk_fp8, dv_fp8, *rest = fused_attn_bwd(
4855
4856
4857
4858
4859
4860
4861
4862
4863
                        ctx.max_seqlen_q,
                        ctx.max_seqlen_kv,
                        cu_seqlens_q,
                        cu_seqlens_kv,
                        q_fp8,
                        k_fp8,
                        v_fp8,
                        out_fp8,
                        d_out_fp8,
4864
4865
                        fake_dtype,
                        dqkv_dtype,
4866
                        aux_ctx_tensors,
4867
                        ctx.fused_attention_backend,
4868
4869
                        cu_seqlens_q_padded,
                        cu_seqlens_kv_padded,
4870
4871
4872
                        ctx.S_quantizer,
                        ctx.dP_quantizer,
                        ctx.dQKV_quantizer,
4873
4874
4875
4876
4877
4878
                        ctx.attn_scale,
                        ctx.dropout_p,
                        ctx.fast_zero_fill,
                        ctx.qkv_layout,
                        ctx.attn_bias_type,
                        ctx.attn_mask_type,
4879
4880
                        ctx.window_size,
                        ctx.deterministic,
4881
                    )
4882

4883
4884
                    # is_input_fp8 = False: dq, dk, dv: torch.float16 or torch.bfloat16
                    # is_input_fp8 = True:  dq, dk, dv: torch.float8_e5m2
4885
                    if not ctx.is_input_fp8:
4886
                        qkv_group = len(ctx.qkv_layout.replace("paged_kv_", "").split("_"))
4887
                        if qkv_group == 1:
4888
                            dim = ctx.qkv_layout.find("3")
4889
4890
                            dqkv_fp8_data = _combine_tensors(
                                [dq_fp8._data, dk_fp8._data, dv_fp8._data], dim
4891
                            )
4892
4893
4894
4895
4896
                            dqkv_fp8 = dq_fp8.make_like(
                                tensor=dq_fp8, data=dqkv_fp8_data, shape=dqkv_fp8_data.shape
                            )
                            dqkv = dqkv_fp8.dequantize()
                            dq, dk, dv = _SplitAlongDim.apply(dqkv, dim, [1, 1, 1], True)
4897
                        if qkv_group == 2:
4898
                            dq = dq_fp8.dequantize()
4899
4900
4901
4902
4903
                            dim = ctx.qkv_layout.split("_")[1].find("2")
                            dkv_fp8 = _combine_tensors([dk_fp8, dv_fp8], dim)
                            dkv_c_fp8 = dkv_fp8.view(
                                -1, dkv_fp8.shape[-3] * dkv_fp8.shape[-2] * dkv_fp8.shape[-1]
                            )
4904
4905
                            dkv = dkv_c_fp8.dequantize()
                            dk, dv = _SplitAlongDim.apply(dkv, dim, [1, 1], True)
4906
                        if qkv_group == 3:
4907
4908
4909
4910
4911
                            dq = dq_fp8.dequantize()
                            dk = dk_fp8.dequantize()
                            dv = dv_fp8.dequantize()
                    else:
                        dq, dk, dv = dq_fp8, dk_fp8, dv_fp8
4912
                else:
4913
4914
                    if isinstance(d_out, QuantizedTensor):
                        d_out = d_out.dequantize()
4915
4916
                    dqkv_dtype = TE_DType[d_out.dtype]
                    # q, k, v, out, d_out, dq, dk, dv: torch.float16 or torch.bfloat16
4917
                    dq, dk, dv, *rest = fused_attn_bwd(
4918
4919
4920
4921
4922
4923
4924
4925
4926
                        ctx.max_seqlen_q,
                        ctx.max_seqlen_kv,
                        cu_seqlens_q,
                        cu_seqlens_kv,
                        q,
                        k,
                        v,
                        out,
                        d_out,
4927
4928
                        fake_dtype,
                        dqkv_dtype,
4929
                        aux_ctx_tensors,
4930
                        ctx.fused_attention_backend,
4931
4932
                        cu_seqlens_q_padded,
                        cu_seqlens_kv_padded,
4933
4934
4935
4936
4937
4938
4939
4940
4941
                        None,
                        None,
                        None,
                        ctx.attn_scale,
                        ctx.dropout_p,
                        ctx.fast_zero_fill,
                        ctx.qkv_layout,
                        ctx.attn_bias_type,
                        ctx.attn_mask_type,
4942
4943
                        ctx.window_size,
                        ctx.deterministic,
4944
                    )
4945

4946
4947
        # if no_bias or alibi, return dqkv
        if ctx.attn_bias_type in ["no_bias", "alibi"]:
4948
4949
4950
4951
4952
4953
4954
4955
            return (
                None,
                None,
                None,
                None,
                None,
                None,
                None,
4956
4957
                None,
                None,
4958
4959
4960
4961
4962
4963
4964
4965
4966
4967
4968
4969
4970
4971
4972
4973
4974
4975
4976
                dq,
                dk,
                dv,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
            )
4977
        # else, return (dqkv, dbias)
4978
4979
4980
4981
4982
4983
4984
4985
        return (
            None,
            None,
            None,
            None,
            None,
            None,
            None,
4986
4987
            None,
            None,
4988
4989
4990
4991
4992
4993
4994
4995
4996
4997
4998
4999
5000
5001
5002
5003
5004
            dq,
            dk,
            dv,
            rest[0],
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
5005
            None,
5006
        )
5007

5008

5009
class FusedAttention(torch.nn.Module):
5010
5011
5012
5013
5014
5015
5016
5017
5018
    """Dot product attention, with multiple backends:

    1. FusedAttnBackend["F16_max512_seqlen"]
       cuDNN based fused attention for FP16/BF16 and <=512 sequence length.
    2. FusedAttnBackend["F16_arbitrary_seqlen"]
       cuDNN based fused attention for FP16/BF16 and any sequence length.

    Support matrix:

5019
5020
5021
5022
    | backend       | 1                       | 2                              |
    | flash based   | no                      | yes                            |
    | cuDNN based   | yes                     | yes                            |
    | qkv dtype     | fp16/bf16               | fp16/bf16                      |
5023
    | attn_type     | self/cross              | self/cross                     |
5024
    | qkv_layout    |                         |                                |
5025
    |  - (q,k,v)    | sb3hd, bs3hd            | sb3hd, bs3hd, sbh3d, bsh3d     |
5026
    |               | sbhd_sb2hd, bshd_bs2hd  | sbhd_sb2hd, bshd_bs2hd         |
5027
5028
    |               | bshd_bshd_bshd          | sbhd_sbh2d, bshd_bsh2d         |
    |               |                         | sbhd_sbhd_sbhd, bshd_bshd_bshd |
5029
5030
    | mask_type     | causal/padding/no_mask  | causal/padding/no_mask         |
    | bias_type     | post_scale_bias/no_bias | post_scale_bias/alibi/no_bias  |
5031
    | dropout       | yes                     | yes                            |
5032
5033
    | max_seqlen    | <=512, multiple of 64   | any, multiple of 64            |
    | head_dim      | 64                      | <=128, multiple of 8           |
5034
    | output dtype  | fp16/bf16               | fp16/bf16                      |
5035
5036
5037
5038
    """

    def __init__(
        self,
5039
        softmax_scale: float,
5040
5041
5042
        attention_dropout: float = 0.0,
        attention_dropout_ctx: Optional[Callable] = nullcontext,
        attention_type: str = "self",
5043
5044
        layer_number: Optional[int] = None,
        deterministic: bool = False,
5045
5046
5047
    ) -> None:
        super().__init__()

5048
        self.softmax_scale = softmax_scale
5049
5050
5051
        self.attention_dropout = attention_dropout
        self.attention_dropout_ctx = attention_dropout_ctx
        self.attention_type = attention_type
5052
5053
5054
        self.use_FAv2_bwd = os.getenv(
            "NVTE_FUSED_ATTN_USE_FAv2_BWD", "0"
        ) == "1" and get_device_compute_capability() == (9, 0)
5055
        self.layer_number = 1 if layer_number is None else layer_number
5056
        self.deterministic = deterministic
5057

5058
        def remove_extra_states_check(self, incompatible_keys):  # pylint: disable=unused-argument
5059
5060
            """
            Temporarily remove fused_attention._extra_state as a missing key
5061
            or an unexpected key when loading Transformer Engine checkpoints.
5062
5063
            Please store FP8 metadata as DotProductAttention's _extra_state,
            rather than FusedAttention's _extra_state. This hook will be
5064
            phased out in Transformer Engine 2.0.
5065
5066
            """
            for key in incompatible_keys.missing_keys:
5067
                if "fused_attention._extra_state" in key:
5068
                    incompatible_keys.missing_keys.remove(key)
5069
5070
5071
5072
5073
5074
5075
            for key in incompatible_keys.unexpected_keys:
                if "fused_attention._extra_state" in key:
                    incompatible_keys.unexpected_keys.remove(key)
                    warnings.warn(
                        "fused_attention._extra_state is not loaded from checkpoint. Please map "
                        "FusedAttention's _extra_state to DotProductAttention's _extra_state."
                    )
5076

5077
5078
        self.register_load_state_dict_post_hook(remove_extra_states_check)

5079
    @no_torch_dynamo()
5080
5081
5082
5083
5084
    def forward(
        self,
        query_layer: torch.Tensor,
        key_layer: torch.Tensor,
        value_layer: torch.Tensor,
5085
5086
5087
        qkv_layout: str = "sbh3d",
        cu_seqlens_q: Optional[torch.Tensor] = None,
        cu_seqlens_kv: Optional[torch.Tensor] = None,
5088
5089
        cu_seqlens_q_padded: Optional[torch.Tensor] = None,
        cu_seqlens_kv_padded: Optional[torch.Tensor] = None,
5090
5091
        max_seqlen_q: Optional[int] = None,
        max_seqlen_kv: Optional[int] = None,
5092
        attn_mask_type: str = "causal",
5093
        attention_mask: Optional[Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]] = None,
5094
        window_size: Optional[Tuple[int, int]] = None,
5095
        fused_attention_backend: tex.NVTE_Fused_Attn_Backend = tex.NVTE_Fused_Attn_Backend.NVTE_No_Backend,
5096
5097
5098
        core_attention_bias_type: str = "no_bias",
        core_attention_bias: Optional[torch.Tensor] = None,
        fast_zero_fill: bool = True,
5099
        cp_group: Optional[Union[dist_group_type, List[dist_group_type]]] = None,
5100
5101
        cp_global_ranks: List[int] = None,
        cp_stream: torch.cuda.Stream = None,
5102
        cp_comm_type: str = "p2p",
5103
5104
        fp8: bool = False,
        fp8_meta: Optional[Dict[str, Any]] = None,
5105
        quantizers=None,
5106
        pad_between_seqs: bool = False,
5107
        inference_params: Optional[InferenceParams] = None,
5108
5109
    ) -> torch.Tensor:
        """fused attention fprop"""
5110
5111
5112
        assert (
            fused_attention_backend != tex.NVTE_Fused_Attn_Backend.NVTE_No_Backend
        ), "No fused attention backend supports this input combination!"
5113
5114
5115
5116
        assert all(
            x.dtype in [torch.float16, torch.bfloat16] or isinstance(x, Float8Tensor)
            for x in [query_layer, key_layer, value_layer]
        ), "FusedAttention only supports FP16 and BF16 data types, or Float8Tensors."
5117
5118
        assert (
            query_layer.is_cuda and key_layer.is_cuda and value_layer.is_cuda
5119
        ), "FusedAttention only supports CUDA tensors."
5120
5121
        assert (
            qkv_layout in QKVLayouts
5122
        ), f"FusedAttention does not support qkv_layout = {qkv_layout}!"
5123

5124
5125
5126
5127
5128
5129
        cp_size = 1
        if isinstance(cp_group, dist_group_type):
            cp_size = get_distributed_world_size(cp_group)
        elif isinstance(cp_group, list):
            for group in cp_group:
                cp_size *= get_distributed_world_size(group)
5130
        context_parallel = cp_size > 1
5131

5132
5133
        # get q_format and kv_format for training and inference
        qkv_format, q_format, kv_format = dpa_utils.get_qkv_format(qkv_layout, inference_params)
5134

5135
5136
5137
5138
5139
5140
5141
5142
5143
5144
        # cuDNN can work with 0-length sequences in the batch for both bshd/sbhd and thd formats
        # however, for bshd/sbhd, q/k/v tensors need to have the same batch size as indicated by
        # cu_seqlens, whereas thd does not have this requirement
        # e.g. if q_format = bshd, and q.shape = [3, 1, 16, 64], we should have k.shape[0] =
        # v.shape[0] = q.shape[0], and cu_seqlens_q.shape = cu_seqlens_kv.shape = [4]
        if q_format in ["bshd", "sbhd"] or kv_format in ["bshd", "sbhd"]:
            batch_size = query_layer.shape[0] if q_format == "bshd" else query_layer.shape[1]
            cu_seqlens_q = cu_seqlens_q[: batch_size + 1]
            cu_seqlens_kv = cu_seqlens_kv[: batch_size + 1]

5145
5146
5147
5148
5149
5150
5151
5152
5153
5154
5155
5156
5157
5158
5159
5160
5161
5162
5163
5164
5165
5166
5167
5168
5169
5170
5171
5172
5173
5174
5175
5176
5177
5178
        page_table = None
        if inference_params is None:
            if qkv_format in ["sbhd", "bshd"]:
                if qkv_format == "sbhd":
                    batch_size = query_layer.shape[1]
                    max_seqlen_q = query_layer.shape[0]
                    max_seqlen_kv = key_layer.shape[0]
                if qkv_format == "bshd":
                    batch_size = query_layer.shape[0]
                    max_seqlen_q = query_layer.shape[1]
                    max_seqlen_kv = key_layer.shape[1]
                max_seqlen_q *= cp_size
                max_seqlen_kv *= cp_size
                if "padding" in attn_mask_type:
                    assert (
                        not context_parallel
                    ), "Padding mask not supported with context parallelism!"
                    if cu_seqlens_q is None or cu_seqlens_kv is None:
                        if attention_mask is None:
                            raise RuntimeError(
                                "Please provide attention_mask or cu_seqlens for padding!"
                            )
                        if self.attention_type == "self":
                            cu_seqlens_q = dpa_utils.get_cu_seqlens(attention_mask)
                            cu_seqlens_kv = cu_seqlens_q
                        else:
                            cu_seqlens_q = dpa_utils.get_cu_seqlens(attention_mask[0])
                            cu_seqlens_kv = dpa_utils.get_cu_seqlens(attention_mask[1])
                else:
                    if cu_seqlens_q is None:
                        cu_seqlens_q = dpa_utils.get_full_cu_seqlens(
                            batch_size,
                            max_seqlen_q,
                            query_layer.device,
5179
                        )
5180
5181
5182
5183
5184
5185
5186
5187
5188
5189
5190
5191
5192
5193
5194
5195
5196
                    if cu_seqlens_kv is None:
                        cu_seqlens_kv = dpa_utils.get_full_cu_seqlens(
                            batch_size,
                            max_seqlen_kv,
                            key_layer.device,
                        )
            if qkv_format == "thd":
                assert (
                    max_seqlen_q is not None
                    and max_seqlen_kv is not None
                    and cu_seqlens_q is not None
                    and cu_seqlens_kv is not None
                ), "max_seqlen_q/kv and cu_seqlens_q/kv can not be None when qkv_format is thd!"
        elif inference_params.is_paged:
            page_table = inference_params.cache_manager.page_table

        if (q_format == "thd" or "padding" in attn_mask_type) and cu_seqlens_q_padded is None:
5197
            cu_seqlens_q_padded = cu_seqlens_q
5198
        if (kv_format == "thd" or "padding" in attn_mask_type) and cu_seqlens_kv_padded is None:
5199
            cu_seqlens_kv_padded = cu_seqlens_kv
5200

5201
5202
5203
5204
5205
        use_FAv2_bwd = (
            self.use_FAv2_bwd
            and (core_attention_bias_type == "no_bias")
            and (fused_attention_backend == tex.NVTE_Fused_Attn_Backend.NVTE_F16_arbitrary_seqlen)
        )
5206

5207
5208
5209
5210
5211
5212
5213
5214
5215
5216
5217
        if fp8:
            assert fused_attention_backend == tex.NVTE_Fused_Attn_Backend.NVTE_FP8, (
                f"cuDNN attention sub-backend {int(tex.NVTE_Fused_Attn_Backend.NVTE_FP8)}"
                " is required for FP8 attention!"
            )
            assert fp8_meta is not None, "FP8 metadata fp8_meta is required for FP8 attention!"
            assert not context_parallel or fp8_meta["recipe"].reduce_amax, (
                "Amax reduction across TP+CP group is necessary when using context parallelism with"
                " FP8!"
            )

5218
        if context_parallel:
5219
            assert (
5220
5221
                fp8
                or fused_attention_backend == tex.NVTE_Fused_Attn_Backend.NVTE_F16_arbitrary_seqlen
5222
5223
5224
5225
5226
5227
5228
            ), f"{fused_attention_backend} does not work with context parallelism!"
            assert core_attention_bias_type not in [
                "alibi"
            ], f"{core_attention_bias_type} is not supported with context parallelism!"
            query_layer, key_layer, value_layer = [
                x.contiguous() for x in (query_layer, key_layer, value_layer)
            ]
5229
5230
5231
            with self.attention_dropout_ctx():
                output = attn_forward_func_with_cp(
                    self.training,
5232
5233
5234
5235
5236
5237
5238
                    query_layer,
                    key_layer,
                    value_layer,
                    cu_seqlens_q,
                    cu_seqlens_kv,
                    max_seqlen_q,
                    max_seqlen_kv,
5239
5240
                    cu_seqlens_q_padded,
                    cu_seqlens_kv_padded,
5241
                    self.attention_dropout if self.training else 0.0,
5242
5243
5244
                    cp_group,
                    cp_global_ranks,
                    cp_stream,
5245
                    cp_comm_type,
5246
                    softmax_scale=self.softmax_scale,
5247
                    qkv_format=qkv_format,
5248
                    attn_mask_type=attn_mask_type,
5249
5250
                    attn_bias_type=core_attention_bias_type,
                    attn_bias=core_attention_bias,
5251
                    deterministic=self.deterministic,
5252
                    use_fused_attention=True,
5253
                    window_size=window_size,
5254
5255
                    fp8=fp8,
                    fp8_meta=fp8_meta,
5256
                    quantizers=quantizers,
5257
                    pad_between_seqs=pad_between_seqs,
5258
5259
                )
        else:
5260
5261
5262
5263
5264
5265
5266
            with self.attention_dropout_ctx():
                output = FusedAttnFunc.apply(
                    self.training,
                    max_seqlen_q,
                    max_seqlen_kv,
                    cu_seqlens_q,
                    cu_seqlens_kv,
5267
5268
                    cu_seqlens_q_padded,
                    cu_seqlens_kv_padded,
5269
5270
                    page_table,
                    page_table,
5271
5272
5273
5274
5275
5276
5277
5278
5279
5280
                    query_layer,
                    key_layer,
                    value_layer,
                    core_attention_bias,
                    self.softmax_scale,
                    self.attention_dropout if self.training else 0.0,
                    fast_zero_fill,
                    qkv_layout,
                    core_attention_bias_type,
                    attn_mask_type,
5281
                    window_size,
5282
5283
5284
5285
5286
                    None,  # rng_gen
                    fused_attention_backend,
                    use_FAv2_bwd,
                    fp8,
                    fp8_meta,
5287
                    quantizers,
5288
                    self.deterministic,
5289
                )
5290

5291
5292
        # ...hd -> ...(hd)
        return output.view(*output.shape[:-2], -1)
5293
5294


5295
class DotProductAttention(TransformerEngineBaseModule):
5296
5297
5298
5299
5300
5301
    """Allows the model to jointly attend to information from different
    representation subspaces as described in the paper:
    `Attention Is All You Need <https://arxiv.org/abs/1706.03762>`_.

    .. note::

5302
        Argument :attr:`attention_mask` in the `forward` call is only used when
5303
        :attr:`attn_mask_type` includes '"padding"' or `"arbitrary"`.
5304
5305
5306

    .. warning::

5307
        FlashAttention uses a non-deterministic algorithm for optimal performance. To observe
5308
        deterministic behavior at the cost of performance, use FlashAttention version >= `2.4.1`
5309
5310
        and set the environment variable :attr:`NVTE_ALLOW_NONDETERMINISTIC_ALGO=0`. In order
        to disable`flash-attn` entirely, set :attr:`NVTE_FLASH_ATTN=0`.
5311

5312
5313
5314
5315
5316
5317
5318
    .. note::

        Transformer Engine stores the FP8 metadata under a `._extra_state` key when checkpointing.
        As the FP8 attention support expands from one backend to multiple backends, the location
        of that key has also shifted (see `FP8 checkpoint compatibility <https://docs.nvidia.com/deeplearning/transformer-engine/user-guide/faq.html#fp8-checkpoint-compatibility>`_).


5319
5320
5321
5322
    Parameters
    ----------
    num_attention_heads : int
                         number of attention heads in the transformer layer.
5323
5324
5325
    kv_channels : Union[int, Tuple[int, int]]
                the head size in key and value tensors. If the same, :attr:`kv_channels` can be
                an integer; if not, :attr:`kv_channels` should be a tuple of two integers.
5326
5327
5328
5329
5330
5331
5332
5333
    num_gqa_groups : Optional[int] = None
                    number of GQA groups in the transformer layer.
                    Grouped Query Attention is described in
                    `this paper <https://arxiv.org/pdf/2305.13245.pdf>`_.
                    This only affects the keys and values, not the queries.
                    GQA-1 is equivalent to Multi-Query Attention
                    (`MQA <https://arxiv.org/pdf/1911.02150.pdf>`_), while GQA-H
                    is equivalent to MHA, i.e. `num_gqa_groups = num_attention_heads`.
5334
5335
    attention_dropout: float, default = 0.0
                      dropout probability for the dropout op during multi-head attention.
5336
    attn_mask_type: str, default = `causal`
5337
                   type of attention mask passed into softmax operation, options are "`no_mask`",
5338
5339
5340
5341
5342
5343
5344
5345
5346
                   "`padding`", "`causal`", "`padding,causal`", "`causal,padding`",
                   "`padding_causal`", "`causal_bottom_right`", "`padding_causal_bottom_right`", and
                   "`arbitrary`", where "`padding,causal`", "`causal,padding`" and "`padding_causal`"
                   are equivalent. This arg can be overridden by :attr:`attn_mask_type` in the
                   `forward` method. It is useful for cases involving compilation/tracing, e.g.
                   ONNX export, and the forward arg is useful for dynamically changing mask types,
                   e.g. a different mask for training and inference.
                   1. For "`no_mask`", no attention mask is applied.
                   2. For "`causal`", "`causal_bottom_right`", or the causal mask in
5347
                   "`padding_causal`" and "`padding_causal_bottom_right`", Transformer Engine
5348
5349
5350
5351
5352
5353
5354
5355
5356
5357
5358
5359
5360
5361
                   calculates and applies an upper triangular mask to the softmax input.
                   No user input is needed. Causal masks without the "`bottom_right`" appendix align
                   the diagonal line to the top left corner of the softmax matrix. With
                   "`bottom_right`", the causal mask is aligned to the bottom right corner, which is
                   often used in inference/KV caching.
                   3. For "`padding`", or the padding mask in "`padding_causal`" and
                   "`padding_causal_bottom_right`", users need to provide the locations of padded
                   tokens, either via :attr:`cu_seqlens_q` and :attr:`cu_seqlens_kv` (both in shape
                   [batch_size + 1]), or via :attr:`attention_mask` (one tensor for self-attention
                   in shape [batch_size, 1, 1, max_seqlen_q], or two tensors in a tuple for
                   cross-attention in shapes [batch_size, 1, 1, max_seqlen_q] and
                   [batch_size, 1, 1, max_seqlen_kv]).
                   4. For "`arbitrary`", users need to provide a mask that is broadcastable to
                   the shape of softmax input [batch_size, num_heads, max_seqlen_q, max_seqlen_kv].
5362
5363
5364
5365
    window_size: Optional[Tuple[int, int]], default = `None`
                sliding window size for local attention, where query at position i attends to keys
                in [i + seqlen_k - seqlen_q - window_size[0], i + seqlen_k - seqlen_q
                + window_size[1]] inclusive. Special cases (-1, -1) and (-1, 0) mean no sliding
5366
5367
5368
                window and causal mask specifically. Both `causal` and `causal_bottom_right` masks
                map to `window_size = (-1, 0)` and Transformer Engine distinguishes them based on
                `attn_mask_type`. Similar to :attr:`attn_mask_type`, `window_size` can
5369
                be overridden by :attr:`window_size` in `forward` as well.
5370
5371
    attention_type: str, default = `self`
                   type of attention, either "`self`" and "`cross`".
5372
5373
5374
    layer_number: int, default = `None`
                 layer number of the current `DotProductAttention` when multiple such modules
                 are concatenated, for instance in consecutive transformer blocks.
5375
5376
5377
    qkv_format: str, default = `sbhd`
               dimension format for `query_layer`, `key_layer` and `value_layer`,
               {`sbhd`, `bshd`, `thd`}. `s` stands for the sequence length, `b` batch size,
5378
               `h` the number of heads, `d` head size, and `t` the total number of tokens
5379
5380
5381
5382
5383
               in a batch, with `t = sum(s_i), for i = 0...b-1`. `sbhd` and `bshd` formats
               are used for when sequences in a batch are of equal length or padded to
               equal length, and the `thd` format is used for when sequences in a batch
               have different lengths. Please note that these formats do not reflect how
               tensors `query_layer`, `key_layer`, `value_layer` are laid out in memory.
5384
               For that, please use `get_qkv_layout` to gain the layout information.
5385
5386
    softmax_scale: Optional[float], default = `None`
                softmax scale for the attention scores. If `None`, defaults to
5387
                `1.0/math.sqrt(kv_channels if isinstance(kv_channels, int) else kv_channels[0])`.
5388
5389
5390
5391
5392
5393
5394
5395
5396

    Parallelism parameters
    ----------------------
    sequence_parallel : bool, default = `False`
                       if set to `True`, uses sequence parallelism.
    tp_size : int, default = 1
             tensor parallel world size.
    tp_group : ProcessGroup, default = `None`
              tensor parallel process group.
5397
    cp_group : Union[ProcessGroup, List[ProcessGroup]], default = `None`
5398
              context parallel process group.
5399
5400
5401
              ProcessGroup is for cp_comm_type of "p2p", "all_gather", and "a2a".
              List[ProcessGroup] is for cp_comm_type of "a2a+p2p", where cp_group[0]
              and cp_group[1] are for a2a and p2p communications respectively.
5402
5403
5404
5405
5406
5407
5408
    cp_global_ranks : list of global rank IDs, default = `None`
                     global rank IDs of GPUs that are in cp_group.
    cp_stream : CUDA stream, default = `None`
               context parallelism splits flash attention into multiple steps for
               compute and communication overlapping. To address the wave quantization
               issue of each split step, we add an additional CUDA stream so that we
               can overlap two flash attention kernels.
5409
    cp_comm_type : str, default = `p2p`
5410
                  inter-gpu communication type for context parallelism.
5411
                  Can be "p2p" or "all_gather" or "a2a" or "a2a+p2p".
5412
5413
5414
5415
5416
5417
                  "p2p": Exchange KV chunks with P2P communications in ring topology.
                         P2P is async and can be overlapped with attention compute.
                  "all_gather": All-gather to get full sequence of KV before attention.
                                The all-gather is not async, and cannot be overlapped.
                  "a2a": Like DeepSpeed Ulysses, scatter attention heads across the CP
                         group, and gather to get full sequence of QKV.
5418
5419
5420
                  "a2a+p2p": hierarchical CP implementation. First applying a2a to QKV
                  across each CP sub-group (e.g., via NVLink), then exchanging KV with
                  p2p between sub-groups (e.g., via IBLink).
5421
5422
5423
5424
5425
    """

    def __init__(
        self,
        num_attention_heads: int,
5426
        kv_channels: Union[int, Tuple[int, int]],
5427
        num_gqa_groups: Optional[int] = None,
5428
        attention_dropout: float = 0.0,
5429
        qkv_format: str = "sbhd",
5430
        attn_mask_type: str = "causal",
5431
        window_size: Optional[Tuple[int, int]] = None,
5432
5433
5434
5435
5436
        sequence_parallel: bool = False,
        tp_size: int = 1,
        get_rng_state_tracker: Optional[Callable] = None,
        tp_group: Optional[dist_group_type] = None,
        layer_number: Optional[int] = None,
5437
        attention_type: str = "self",
5438
        cp_group: Optional[Union[dist_group_type, List[dist_group_type]]] = None,
5439
        cp_global_ranks: List[int] = None,
5440
        cp_stream: torch.cuda.Stream = None,
5441
        cp_comm_type: str = "p2p",
5442
        softmax_scale: Optional[float] = None,
5443
5444
5445
    ) -> None:
        super().__init__()

5446
        self.logger = logging.getLogger("DotProductAttention")
5447
        self.logger.setLevel(attn_log._log_level)
5448
        if not self.logger.hasHandlers():
5449
            self.logger.addHandler(attn_log._stream_handler)
5450
        self.qkv_format = qkv_format
5451
        attn_mask_type = attn_mask_type.replace(",", "_")
5452
5453
        if attn_mask_type == "causal_padding":
            attn_mask_type = "padding_causal"
5454
        self.attn_mask_type = attn_mask_type
5455
        self.window_size = dpa_utils.check_set_window_size(attn_mask_type, window_size)
5456
5457
5458
5459
5460
5461
5462
        if tp_group is None:
            self.tp_size = tp_size
            if tp_size == 1:
                self.set_tensor_parallel_group(tp_group)
        else:
            self.tp_size = get_distributed_world_size(tp_group)
            self.set_tensor_parallel_group(tp_group)
5463
        self.get_rng_state_tracker = get_rng_state_tracker
5464
        self.num_attention_heads = num_attention_heads
5465
        self.layer_number = 1 if layer_number is None else layer_number
5466
5467
5468
        self.cp_group = cp_group
        self.cp_global_ranks = cp_global_ranks
        self.cp_stream = cp_stream
5469
        self.cp_comm_type = cp_comm_type
5470

5471
5472
5473
5474
5475
5476
        self.hidden_size_per_attention_head_k = (
            kv_channels if isinstance(kv_channels, int) else kv_channels[0]
        )
        self.hidden_size_per_attention_head_v = (
            kv_channels if isinstance(kv_channels, int) else kv_channels[1]
        )
5477

5478
        self.num_gqa_groups = num_attention_heads if num_gqa_groups is None else num_gqa_groups
5479
        self.num_gqa_groups_per_partition = int(self.num_gqa_groups // self.tp_size)
5480

5481
5482
5483
        assert (
            num_attention_heads % self.num_gqa_groups == 0
        ), "The number of attention heads must be divisible by the number of GQA groups!"
5484

5485
        self.rng_states_tracker = None
5486
5487
5488
        if sequence_parallel or get_rng_state_tracker is None:
            attention_dropout_ctx = nullcontext
        else:
5489
5490
5491
            self.rng_states_tracker = get_rng_state_tracker()
            set_all_rng_states(self.rng_states_tracker.get_states())
            attention_dropout_ctx = self.rng_states_tracker.fork
5492

5493
        if softmax_scale is None:
5494
5495
5496
            softmax_scale = 1.0 / math.sqrt(
                kv_channels if isinstance(kv_channels, int) else kv_channels[0]
            )
5497

5498
5499
5500
        self.deterministic = (
            not bool(int(os.getenv("NVTE_ALLOW_NONDETERMINISTIC_ALGO", "1")))
            or torch.are_deterministic_algorithms_enabled()
5501
        )
5502
5503
5504
5505
5506
5507
5508
5509
5510
5511
5512
5513
5514
5515
5516
5517
5518
5519
5520
        # To use the workspace optimization path for determinism, please
        # set NVTE_FUSED_ATTN_FORCE_WORKSPACE_OPT=1 for cuDNN >=8.9.5 and <9.0.0,
        # and set NVTE_ALLOW_NONDETERMINISTIC_ALGO=0 for cuDNN >=9.0.0.
        cudnn_version = get_cudnn_version()
        if (8, 9, 5) <= cudnn_version < (9, 0, 0):
            if self.deterministic:
                os.environ["NVTE_FUSED_ATTN_FORCE_WORKSPACE_OPT"] = "1"

            # CUDNN_FRONTEND_ATTN_DP_WORKSPACE_LIMIT
            # - unset:       enables workspace optimization when required workspace is <= 256MB
            #                or when bias gradient needs to be computed
            # - n:           enables workspace optimization when required workspace is <= n bytes
            # - -1:          enables workspace optimization always
            # - 0:           disables workspace optimization always
            if "NVTE_FUSED_ATTN_FORCE_WORKSPACE_OPT" in os.environ:
                if os.environ["NVTE_FUSED_ATTN_FORCE_WORKSPACE_OPT"] == "0":
                    os.environ["CUDNN_FRONTEND_ATTN_DP_WORKSPACE_LIMIT"] = "0"
                if os.environ["NVTE_FUSED_ATTN_FORCE_WORKSPACE_OPT"] == "1":
                    os.environ["CUDNN_FRONTEND_ATTN_DP_WORKSPACE_LIMIT"] = "-1"
5521

5522
        assert attention_type in AttnTypes, f"attention_type {attention_type} not supported"
5523
5524
5525
5526

        self.attention_type = attention_type
        self.attention_dropout = attention_dropout

5527
5528
5529
5530
5531
        attn_kwargs = {
            "attention_dropout": attention_dropout,
            "attention_dropout_ctx": attention_dropout_ctx,
        }

5532
5533
5534
5535
5536
5537
5538
        self.flash_attention = FlashAttention(
            softmax_scale,
            attention_type=attention_type,
            layer_number=layer_number,
            deterministic=self.deterministic,
            **attn_kwargs,
        )
5539

5540
        # Instantiating three types since use of flash-attn and FusedAttention
5541
        # might be ruled out due to forward inputs.
5542
5543
5544
5545
5546
5547
5548
        self.fused_attention = FusedAttention(
            softmax_scale,
            attention_type=attention_type,
            layer_number=layer_number,
            deterministic=self.deterministic,
            **attn_kwargs,
        )
5549

5550
        self.unfused_attention = UnfusedDotProductAttention(
5551
5552
5553
5554
            softmax_scale,
            attention_type=attention_type,
            **attn_kwargs,
            layer_number=layer_number,
5555
        )
5556

5557
5558
5559
        def remove_extra_states_check(self, incompatible_keys):  # pylint: disable=unused-argument
            """
            Temporarily remove core_attention._extra_state as a missing key
5560
5561
            when loading older Transformer Engine checkpoints. Will phase out
            this hook in Transformer Engine 2.0.
5562
5563
5564
5565
5566
5567
5568
            """
            for key in incompatible_keys.missing_keys:
                if "core_attention._extra_state" in key:
                    incompatible_keys.missing_keys.remove(key)

        self.register_load_state_dict_post_hook(remove_extra_states_check)

5569
5570
5571
5572
5573
5574
5575
5576
5577
5578
5579
5580
5581
5582
5583
5584
5585
5586
5587
5588
5589
5590
    def _load_from_state_dict(
        self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs
    ):
        """
        This function helps to load Transformer Engine 1.6 and 1.7 checkpoints, where FP8 attention
        metadata is stored under the `core_attention.fused_attention._extra_state` key and not the
        `core_attention._extra_state` key. Please see `FP8 checkpoint compatibility
        <https://docs.nvidia.com/deeplearning/transformer-engine/user-guide/faq.html#fp8-checkpoint-compatibility>`_ for more details.
        """
        fused_attn_key = False
        dot_product_attn_key = False
        for k in state_dict.keys():
            if "core_attention.fused_attention._extra_state" in k:
                fused_attn_key = True
            if "core_attention._extra_state" in k:
                dot_product_attn_key = True
        if fused_attn_key and not dot_product_attn_key:
            prefix = prefix + "fused_attention."
        super()._load_from_state_dict(
            state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs
        )

5591
5592
5593
5594
    def _checkpointed_attention_forward(
        self,
        attention_func: Callable,
        *forward_args: Tuple[torch.Tensor, ...],
5595
        **forward_kwargs: Dict[str, Any],
5596
5597
5598
    ) -> torch.Tensor:
        """Forward method with activation checkpointing."""

5599
5600
        def custom_forward(*input_args, **input_kwargs):
            return attention_func(*input_args, **input_kwargs)
5601
5602
5603

        hidden_states = checkpoint(
            custom_forward,
5604
5605
5606
            distribute_saved_activations=False,
            get_rng_state_tracker=self.get_rng_state_tracker,
            tp_group=self.tp_group,
5607
            *forward_args,
5608
            **forward_kwargs,
5609
5610
5611
5612
        )

        return hidden_states

5613
5614
    def set_context_parallel_group(
        self,
5615
        cp_group: Union[dist_group_type, List[dist_group_type], None],
5616
5617
        cp_global_ranks: List[int],
        cp_stream: torch.cuda.Stream,
5618
        cp_comm_type: str = "p2p",
5619
    ) -> None:
5620
5621
5622
5623
5624
5625
        """
        Set the context parallel attributes for the given
        module before executing the forward pass.

        Parameters
        ----------
5626
        cp_group : Union[ProcessGroup, List[ProcessGroup]]
5627
                  context parallel process group.
5628
5629
5630
                  ProcessGroup is for cp_comm_type of "p2p", "all_gather", and "a2a".
                  List[ProcessGroup] is for cp_comm_type of "a2a+p2p", where cp_group[0]
                  and cp_group[1] are for a2a and p2p communications respectively.
5631
5632
5633
5634
        cp_global_ranks : List[int]
                         list of global ranks in the context group.
        cp_stream : torch.cuda.Stream
                   cuda stream for context parallel execution.
5635
        cp_comm_type : str, default = `p2p`
5636
                      inter-gpu communication type for context parallelism.
5637
                      Can be "p2p" or "all_gather" or "a2a" or "a2a+p2p".
5638
5639
5640
5641
5642
5643
                      "p2p": Exchange KV chunks with P2P communications in ring topology.
                             P2P is async and can be overlapped with attention compute.
                      "all_gather": All-gather to get full sequence of KV before attention.
                                    The all-gather is not async, and cannot be overlapped.
                      "a2a": Like DeepSpeed Ulysses, scatter attention heads across the CP
                             group, and gather to get full sequence of QKV.
5644
5645
5646
                      "a2a+p2p": hierarchical CP implementation. First applying a2a to QKV
                      across each CP sub-group (e.g., via NVLink), then exchanging KV with
                      p2p between sub-groups (e.g., via IBLink).
5647
        """
5648
5649
5650
        self.cp_group = cp_group
        self.cp_global_ranks = cp_global_ranks
        self.cp_stream = cp_stream
5651
        self.cp_comm_type = cp_comm_type
5652

5653
    @no_torch_dynamo(recursive=False)
5654
5655
5656
5657
5658
    def forward(
        self,
        query_layer: torch.Tensor,
        key_layer: torch.Tensor,
        value_layer: torch.Tensor,
5659
5660
5661
5662
5663
5664
5665
5666
        attention_mask: Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]] = None,
        qkv_format: str = None,
        cu_seqlens_q: torch.Tensor = None,
        cu_seqlens_kv: torch.Tensor = None,
        cu_seqlens_q_padded: torch.Tensor = None,
        cu_seqlens_kv_padded: torch.Tensor = None,
        max_seqlen_q: int = None,
        max_seqlen_kv: int = None,
5667
        attn_mask_type: Optional[str] = None,
5668
        window_size: Optional[Tuple[int, int]] = None,
5669
        checkpoint_core_attention: bool = False,
5670
5671
        core_attention_bias_type: str = "no_bias",
        core_attention_bias: Optional[torch.Tensor] = None,
5672
        alibi_slopes: Optional[torch.Tensor] = None,
5673
        fast_zero_fill: bool = True,
5674
        inference_params: Optional[InferenceParams] = None,
5675
        pad_between_seqs: Optional[bool] = None,
5676
5677
5678
5679
5680
5681
    ) -> torch.Tensor:
        """
        Dot Product Attention Layer.

        .. note::

5682
5683
            Argument :attr:`attention_mask` is only used when :attr:`attn_mask_type`
            includes '"padding"' or `"arbitrary"`.
5684

5685
5686
        .. note::

5687
5688
5689
5690
5691
5692
5693
5694
5695
5696
5697
5698
5699
            DotProductAttention supports three backends: 1) FlashAttention which calls
            HazyResearch/Dao-AILab's `flash-attn <https://arxiv.org/pdf/2305.13245.pdf>`_
            PyTorch API, 2) FusedAttention which has multiple fused attention implementations
            based on `cuDNN Graph API
            <https://docs.nvidia.com/deeplearning/cudnn/developer-guide/index.html#op-fusion>`_
            (see :attr:`FusedAttention` for more details on FusedAttention backends), and 3)
            UnfusedDotProductAttention which is the native PyTorch implementation
            with fused scaled masked softmax.

        .. note::

            Users can use environment variables :attr:`NVTE_FLASH_ATTN`, :attr:`NVTE_FUSED_ATTN`,
            and :attr:`NVTE_FUSED_ATTN_BACKEND` to control which DotProductAttention backend,
5700
            and FusedAttention backend if applicable, to use. Transformer Engine prioritizes
5701
5702
5703
5704
            FlashAttention over FusedAttention and over UnfusedDotProductAttention.
            If FusedAttention is being used, users can also choose to switch to flash-attn's
            implementation for backward by setting :attr:`NVTE_FUSED_ATTN_USE_FAv2_BWD=1`
            (default: 0), because of the performance differences between various versions of
5705
5706
            flash-attn and FusedAttention. Further, :attr:`NVTE_FUSED_ATTN_FORCE_WORKSPACE_OPT`
            can be used to enable (:attr:`1`) or disable (:attr:`0`) the workspace related
5707
            optimizations in FusedAttention. When unset, Transformer Engine determines the code path
5708
5709
            based on its internal logic. These optimizations trade memory for performance
            and should be used with care.
5710

5711
5712
5713
5714
5715
5716
5717
5718
5719
5720
5721
5722
5723
5724
5725
5726
5727
5728
5729
5730
5731
5732
5733
5734
5735
5736
5737
5738
5739
5740
5741
5742
5743
5744
5745
5746
5747
5748
5749
5750
5751
5752
5753
5754
5755
5756
5757
5758
5759
5760
5761
5762
5763
5764
        .. note::
            .. _cu_seqlens note:

            When training data has variable sequence lengths, users have two options.

            1. Manipulate the data and pad all sequences to the same length. Use
               :attr:`qkv_format` = {"bshd", "sbhd"} and
               :attr:`attn_mask_type` = {"padding", "padding_causal", "padding_causal_bottom_right"}.
               Pass in :attr:`cu_seqlens_q` and :attr:`cu_seqlens_kv`, or :attr:`attention_mask`
               (which will be converted to :attr:`cu_seqlens_q` and :attr:`cu_seqlens_kv`), to provide
               the real sequence length information. For example, a batch of 3 sequences
               [a a a b b c c c c] can be padded to [a a a PAD b b PAD PAD c c c c], and the cumulative
               sequence length tensors would be
               :attr:`cu_seqlens_q` = :attr:`cu_seqlens_kv` = [0, 3, 5, 9] for self-attention.

            2. Do not perform padding on training data. Use :attr:`qkv_format` = "thd" and
               :attr:`attn_mask_type` = {"padding", "padding_causal", "padding_causal_bottom_right"}.
               Pass in :attr:`cu_seqlens_q` and :attr:`cu_seqlens_kv`, or :attr:`attention_mask`,
               as in option 1. For example, a batch of 3 sequences [a a a b b c c c c] can be processed
               without any padding, and the sequence length tensors would be
               :attr:`cu_seqlens_q` = :attr:`cu_seqlens_kv` = [0, 3, 5, 9] for self-attention.

               In certain use cases, a varying number of identifier tokens are inserted between
               sequences. These tokens do not participate in the attention calculation.
               :attr:`cu_seqlens_q_padded` and :attr:`cu_seqlens_kv_padded` must be specified
               in such cases to correctly identify the start and end of each sequence in a batch.
               For example, a batch of 3 sequences [a a a 1 b b 2 2 c c c c 3] would have
               :attr:`cu_seqlens_q` = :attr:`cu_seqlens_kv` = [0, 3, 5, 9], and
               :attr:`cu_seqlens_q_padded` = :attr:`cu_seqlens_kv_padded` = [0, 4, 8, 13]
               for self-attention.

        .. note::
            .. _max_seqlen note:

            When :attr:`qkv_format` = {"bshd", "sbhd"}, sequences are of equal length in a batch.
            :attr:`max_seqlen_q` and :attr:`max_seqlen_kv` should be the same as the "s" dimension of
            :attr:`query_layer` and :attr:`key_layer` tensors. When unset, Transformer Engine will
            infer them as such.

            When :attr:`qkv_format` = "thd", sequences have varying lengths. :attr:`max_seqlen_q` and
            :attr:`max_seqlen_kv` should be the maximum query and key/value sequence length in a batch.
            When unset, Transformer Engine deduces them from :attr:`cu_seqlens_q` and :attr:`cu_seqlens_kv`.
            This deduction costs a small kernel and some CPU-GPU synchronization, and to avoid this
            overhead, users are recommended to obtain the maximum sequence lengths from the data loaders
            and pass them in.

            - As the maximum sequence lengths, batch size, and number of tokens change from batch to batch,
              dynamic shapes need to be supported for tensor construction. FlashAttention and
              UnfusedDotProductAttention naturally do so, while FusedAttention requires parameters to be static
              to create graphs before performance heuristics analysis. To reduce the number of graphs created
              per run, Transformer Engine 1.13+ quantizes relevant parameters: for cuDNN < 9.6, {batch size,
              :attr:`max_seqlen_q`, :attr:`max_seqlen_kv`}, and for cuDNN >= 9.6, {"t" dimension of
              :attr:`query_layer`, "t" dimension of :attr:`key_layer`}.

5765
5766
5767
5768
5769
5770
5771
5772
        Parameters
        ----------
        query_layer : torch.Tensor
                     Query tensor.
        key_layer : torch.Tensor
                   Key tensor.
        value_layer : torch.Tensor
                     Value tensor.
5773
5774
        attention_mask: Optional[Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]],
             default = `None`. Boolean tensor(s) used to mask out attention softmax input.
5775
             It should be `None` for causal masks and "`no_mask`". For padding masks, it should be
5776
5777
             a single tensor of [batch_size, 1, 1, seqlen_q] for self-attention, and a tuple of
             two tensors in shapes [batch_size, 1, 1, seqlen_q] and [batch_size, 1, 1, seqlen_kv]
5778
5779
5780
5781
             for cross-attention. For "`arbitrary`" mask, it should be in a shape broadcastable
             to [batch_size, num_heads, max_seqlen_q, max_seqlen_kv]. A `True` value means
             the corresponding position is masked out and a `False` means that position
             is allowed to participate in attention.
5782
5783
5784
        qkv_format: str, default = `None`
                   If provided, overrides :attr:`qkv_format` from initialization.
        cu_seqlens_q: Optional[torch.Tensor], default = `None`
5785
                   Cumulative sum of sequence lengths (without offset) in a batch for `query_layer`,
5786
                   with shape [batch_size + 1] and dtype torch.int32.
5787
                   See :ref:`note<cu_seqlens note>` for more details.
5788
        cu_seqlens_kv: Optional[torch.Tensor], default = `None`
5789
5790
                   Cumulative sum of sequence lengths (without offset) in a batch for `key_layer`
                   and `value_layer`, with shape [batch_size + 1] and dtype torch.int32.
5791
                   See :ref:`note<cu_seqlens note>` for more details.
5792
5793
5794
5795
5796
        cu_seqlens_q_padded: Optional[torch.Tensor], default = `None`
                   Cumulative sum of sequence lengths (with offset) in a batch for
                   `query_layer`, with shape [batch_size + 1] and dtype torch.int32.
                   When there is no padding between sequences in a batch,
                   `cu_seqlens_q_padded = cu_seqlens_q`.
5797
                   See :ref:`note<cu_seqlens note>` for more details.
5798
5799
5800
5801
5802
        cu_seqlens_kv_padded: Optional[torch.Tensor], default = `None`
                   Cumulative sum of sequence lengths (with offset) in a batch for `key_layer`
                   and `value_layer`, with shape [batch_size + 1] and dtype torch.int32.
                   When there is no padding between sequences in a batch,
                   `cu_seqlens_kv_padded = cu_seqlens_kv`.
5803
                   See :ref:`note<cu_seqlens note>` for more details.
5804
5805
        max_seqlen_q: Optional[int], default = `None`
                      Maximum sequence length in `query_layer`.
5806
                      See :ref:`note<max_seqlen note>` for more details.
5807
5808
        max_seqlen_kv: Optional[int], default = `None`
                       Maximum sequence length in `key_layer` and `value_layer`.
5809
                       See :ref:`note<max_seqlen note>` for more details.
5810
5811
5812
5813
5814
5815
5816
        attn_mask_type: {'no_mask', 'padding', 'causal', 'padding,causal', 'causal,padding',
                       'padding_causal', 'causal_bottom_right', 'padding_causal_bottom_right',
                       'arbitrary'}, default = `None`. Type of attention mask passed into
                       softmax operation. 'padding,causal', 'causal,padding' and 'padding_causal'
                       are equivalent. By default, causal masks are aligned to the top left corner
                       of the softmax matrix. When "`bottom_right`" is specified in the mask type,
                       causal masks are aligned to the bottom right corner.
5817
        window_size: Optional[Tuple[int, int]], default = `None`
5818
                    Sliding window size for local attention.
5819
5820
5821
5822
5823
        checkpoint_core_attention : bool, default = `False`
                                   If true, forward activations for attention are recomputed
                                   during the backward pass in order to save memory that would
                                   otherwise be occupied to store the forward activations until
                                   backprop.
5824
        core_attention_bias_type: str, default = `no_bias`
5825
                    Bias type, {`no_bias`, `pre_scale_bias`, `post_scale_bias`, `alibi`}
5826
        core_attention_bias: Optional[torch.Tensor], default = `None`
5827
5828
                    Bias tensor for Q * K.T, shape [1, num_head, max_seqlen_q, max_seqlen_kv].
                    It should be 'None' for 'no_bias' and 'alibi' bias types.
5829
5830
5831
5832
        alibi_slopes: Optional[torch.Tensor], default = `None`
                     ALiBi slopes in FP32 and shape [nheads] or [batch_size, nheads].
                     It adds a bias of (-alibi_slope * (i + seqlen_k - seqlen_q - j))
                     to the attention score of query i and key j.
5833
        fast_zero_fill: bool, default = `True`
5834
                    Whether to use the fast path to set output tensors to 0 or not.
5835
5836
5837
5838
5839
5840
5841
5842
5843
5844
        inference_params: Optional[InferenceParams], default = `None`
            Optimizes execution performance during inference by caching Keys and Values of the
            current decoding iteration. These cached values are appended to the K and V values
            computed in previous iterations, eliminating the need to recalculate them for the
            entire sequence.
            Initialization of `inference_params` is required prior to use to ensure sufficient
            memory allocation.
            Adjustments of the sequence_len_offset should be done after a complete forward pass.
            If rotary positional embeddings (RoPE) are utilized, they must be prepared beforehand.
            Supports "sbhd" and "bshd" layouts, with the "sbhd" layout being more efficient.
5845
5846
5847
        pad_between_seqs: Optional[bool], default = `None`
            If None, inferred from qkv_format, cu_seqlens and cu_seqlens_padded.
            If true, there are padding tokens between individual sequences in a packed batch.
5848
        """
5849

5850
5851
5852
5853
5854
        with self.prepare_forward(
            query_layer,
            num_gemms=3,
            allow_non_contiguous=True,
        ) as query_layer:
5855
5856
5857
5858
5859
5860
5861
5862
5863
5864
            # checks for RNG
            if self.rng_states_tracker is not None and is_graph_capturing():
                assert isinstance(
                    self.rng_states_tracker, CudaRNGStatesTracker
                ), "Unsupported RNG states tracker."
                assert (
                    graph_safe_rng_available()
                ), "Upgrade PyTorch version to get RNG manipulation support for cuda graph capture."

            # checks for FP8
5865
5866
5867
5868
            if self.fp8:
                if self.fp8_meta["recipe"].fp8_mha:
                    if not self.fp8_meta["recipe"].fp8_dpa:
                        self.fp8_meta["recipe"].fp8_dpa = True
5869
                        self.logger.warning(
5870
5871
5872
                            """Forcing fp8_meta["recipe"].fp8_dpa=True due to """
                            """fp8_meta["recipe"].fp8_mha=True"""
                        )
5873
5874
5875
5876
5877
5878
5879
5880
5881
5882
            if self.fp8 and self.fp8_meta["recipe"].fp8_dpa:
                forward_dtype = get_fp8_te_dtype(self.fp8_meta["recipe"], fprop_tensor=True)
                backward_dtype = get_fp8_te_dtype(self.fp8_meta["recipe"], fprop_tensor=False)
                assert forward_dtype in [
                    tex.DType.kFloat8E4M3,
                    tex.DType.kFloat8E5M2,
                ] and backward_dtype in [
                    tex.DType.kFloat8E4M3,
                    tex.DType.kFloat8E5M2,
                ], """DotProductAttention only supports "E4M3" and "E5M2" FP8 data types."""
5883

5884
            # checks for q/k/v shapes
5885
5886
5887
            assert (
                query_layer.is_cuda and key_layer.is_cuda and value_layer.is_cuda
            ), "DotProductAttention only supports CUDA tensors."
5888
5889
5890
            assert (
                query_layer.dtype == key_layer.dtype and query_layer.dtype == value_layer.dtype
            ), "Queries, keys and values must have the same data type!"
5891
5892
5893
            assert (
                key_layer.shape[:-1] == value_layer.shape[:-1]
            ), "Keys and values must have the same batch size, sequence length and number of heads!"
5894
5895
            num_attention_heads = query_layer.shape[-2]
            num_gqa_groups = key_layer.shape[-2]
5896
            assert (
5897
5898
5899
                query_layer.shape[-1] == key_layer.shape[-1]
            ), "Queries and keys must have the same head dimension!"
            head_dim_qk, head_dim_v = query_layer.shape[-1], value_layer.shape[-1]
5900
            assert (
5901
5902
                head_dim_qk == self.hidden_size_per_attention_head_k
            ), f"Keys have head_dim = {head_dim_qk}, "
5903
5904
            "but expected head_dim = {self.hidden_size_per_attention_head_k}!"
            assert (
5905
5906
                head_dim_v == self.hidden_size_per_attention_head_v
            ), f"Values have head_dim = {head_dim_v}, "
5907
            "but expected head_dim = {self.hidden_size_per_attention_head_v}!"
5908
5909
5910
5911
            assert num_gqa_groups == self.num_gqa_groups_per_partition, (
                "Keys and values must have num_gqa_group ="
                f" {self.num_gqa_groups_per_partition} heads! Found {num_gqa_groups}."
            )
5912

5913
            # checks for attention mask
5914
5915
5916
5917
5918
5919
            if attn_mask_type is None:
                attn_mask_type = self.attn_mask_type
            else:
                attn_mask_type = attn_mask_type.replace(",", "_")
                if attn_mask_type == "causal_padding":
                    attn_mask_type = "padding_causal"
5920
            assert (
5921
5922
                attn_mask_type in AttnMaskTypes
            ), f"Attention mask type {attn_mask_type} is not supported!"
5923

5924
            # checks for sliding window
5925
5926
            if window_size is None:
                window_size = self.window_size
5927
            window_size = dpa_utils.check_set_window_size(attn_mask_type, window_size)
5928

5929
5930
5931
            # checks for qkv_format
            if qkv_format is None:
                qkv_format = self.qkv_format
5932
5933
5934
5935
5936
            assert qkv_format in [
                "sbhd",
                "bshd",
                "thd",
            ], "DotProductAttention only supports qkv_format = {'sbhd', 'bshd', 'thd'}!"
5937
5938
5939
5940
5941
5942
5943
5944
5945
5946
5947
5948
5949
            batch_size = None
            if qkv_format in ["sbhd", "bshd"]:
                assert all(
                    len(x.shape) == 4 for x in (query_layer, key_layer, value_layer)
                ), f"Queries, keys and values must be 4D tensors when {qkv_format=}!"
                if qkv_format == "sbhd":
                    batch_size = query_layer.shape[1]
                    max_seqlen_q = query_layer.shape[0] if max_seqlen_q is None else max_seqlen_q
                    max_seqlen_kv = key_layer.shape[0] if max_seqlen_kv is None else max_seqlen_kv
                else:
                    batch_size = query_layer.shape[0]
                    max_seqlen_q = query_layer.shape[1] if max_seqlen_q is None else max_seqlen_q
                    max_seqlen_kv = key_layer.shape[1] if max_seqlen_kv is None else max_seqlen_kv
5950
            if qkv_format == "thd":
5951
                assert all(
5952
5953
                    len(x.shape) == 3 for x in (query_layer, key_layer, value_layer)
                ), "Queries, keys and values must be 3D tensors when qkv_format = thd!"
5954
5955
5956
                assert (
                    "padding" in attn_mask_type
                ), "Attention mask type must be padding or padding_causal for qkv_format=thd!"
5957
5958
5959
5960
5961
5962
5963
5964
5965
5966
5967
                assert (
                    cu_seqlens_q is not None and cu_seqlens_kv is not None
                ), "cu_seqlens_q and cu_seqlens_kv can not be None when qkv_format = thd!"
                assert (
                    cu_seqlens_q.shape == cu_seqlens_kv.shape
                    and len(cu_seqlens_q.shape) == 1
                    and len(cu_seqlens_kv.shape) == 1
                ), "cu_seqlens_q and cu_seqlens_q must both have shape [batch_size + 1]!"
                assert (
                    cu_seqlens_q.dtype == torch.int32 and cu_seqlens_kv.dtype == torch.int32
                ), "cu_seqlens_q and cu_seqlens_q must both be in dtype torch.int32!"
5968
                batch_size = len(cu_seqlens_q) - 1
5969
                if max_seqlen_q is None:
5970
5971
5972
5973
                    if cu_seqlens_q_padded is not None:
                        seqlens_q = cu_seqlens_q_padded[1:] - cu_seqlens_q_padded[:-1]
                    else:
                        seqlens_q = cu_seqlens_q[1:] - cu_seqlens_q[:-1]
5974
                    max_seqlen_q = int((seqlens_q.max().item() + 63) // 64 * 64)
5975
                if max_seqlen_kv is None:
5976
5977
5978
5979
                    if cu_seqlens_kv_padded is not None:
                        seqlens_kv = cu_seqlens_kv_padded[1:] - cu_seqlens_kv_padded[:-1]
                    else:
                        seqlens_kv = cu_seqlens_kv[1:] - cu_seqlens_kv[:-1]
5980
                    max_seqlen_kv = int((seqlens_kv.max().item() + 63) // 64 * 64)
5981

5982
5983
5984
5985
5986
5987
5988
5989
5990
5991
5992
5993
5994
5995
5996
5997
5998
5999
6000
6001
6002
6003
6004
6005
6006
6007
6008
6009
6010
6011
6012
6013
6014
6015
6016
6017
6018
6019
6020
6021
6022
6023
6024
6025
6026
6027
6028
6029
6030
6031
6032
6033
6034
6035
6036
6037
6038
6039
6040
6041
6042
6043
6044
6045
6046
6047
6048
6049
6050
6051
            # update KV cache and retrieve saved tokens from cache for inference
            if inference_params is not None:
                assert self.layer_number is not None, "Layer number must be set!"

                # convert top-left causal to bottom-right causal due to KV caching
                # users can still use the same attention mask for inference as for training
                assert "padding" in attn_mask_type, "KV caching requires padding mask!"
                if attn_mask_type == "padding_causal":
                    attn_mask_type = attn_mask_type + "_bottom_right"

                self.attention_type = "cross"
                self.flash_attention.attention_type = self.attention_type
                self.fused_attention.attention_type = self.attention_type
                self.unfused_attention.attention_type = self.attention_type

                query_layer, key_layer, value_layer = [
                    x.contiguous() if not x.is_contiguous() else x
                    for x in [query_layer, key_layer, value_layer]
                ]

                # get full K/V tensors from cache and adjust cu_seqlens, qkv_format based on the cache
                (
                    key_layer,
                    value_layer,
                    cu_seqlens_q,
                    cu_seqlens_kv,
                    max_seqlen_kv,
                    qkv_format,
                ) = inference_params.step(
                    self.layer_number,
                    key_layer,
                    value_layer,
                    qkv_format,
                )
                cu_seqlens_q_padded = None
                cu_seqlens_kv_padded = None

            # get qkv's memory layout
            if all(isinstance(x, Float8Tensor) for x in [query_layer, key_layer, value_layer]):
                (
                    qkv_layout,
                    query_layer._data,
                    key_layer._data,
                    value_layer._data,
                    q_format,
                    kv_format,
                ) = dpa_utils.get_qkv_layout(
                    query_layer._data,
                    key_layer._data,
                    value_layer._data,
                    qkv_format=qkv_format,
                    inference_params=inference_params,
                )
            else:
                (
                    qkv_layout,
                    query_layer,
                    key_layer,
                    value_layer,
                    q_format,
                    kv_format,
                ) = dpa_utils.get_qkv_layout(
                    query_layer,
                    key_layer,
                    value_layer,
                    qkv_format=qkv_format,
                    inference_params=inference_params,
                )

            # adjust max_seqlen and cu_seqlens for CP
6052
6053
6054
6055
6056
6057
            cp_size = 1
            if isinstance(self.cp_group, dist_group_type):
                cp_size = get_distributed_world_size(self.cp_group)
            elif isinstance(self.cp_group, list):
                for group in self.cp_group:
                    cp_size *= get_distributed_world_size(group)
6058
            context_parallel = cp_size > 1
6059
            if q_format in ["sbhd", "bshd"]:
6060
                max_seqlen_q *= cp_size
6061
                if cu_seqlens_q is None:
6062
6063
6064
6065
                    if "padding" in attn_mask_type:
                        assert (
                            attention_mask is not None
                        ), "Please provide attention_mask for padding!"
6066
                        if self.attention_type == "self":
6067
                            cu_seqlens_q = dpa_utils.get_cu_seqlens(attention_mask)
6068
                        else:
6069
                            cu_seqlens_q = dpa_utils.get_cu_seqlens(attention_mask[0])
6070
                    else:
6071
                        cu_seqlens_q = dpa_utils.get_full_cu_seqlens(
6072
6073
6074
6075
                            batch_size,
                            max_seqlen_q,
                            query_layer.device,
                        )
6076
6077
6078
6079
6080
6081
6082
6083
6084
6085
6086
6087
            if kv_format in ["sbhd", "bshd"]:
                max_seqlen_kv *= cp_size
                if cu_seqlens_kv is None:
                    if "padding" in attn_mask_type:
                        assert (
                            attention_mask is not None
                        ), "Please provide attention_mask for padding!"
                        if self.attention_type == "self":
                            cu_seqlens_kv = dpa_utils.get_cu_seqlens(attention_mask)
                        else:
                            cu_seqlens_kv = dpa_utils.get_cu_seqlens(attention_mask[1])
                    else:
6088
                        cu_seqlens_kv = dpa_utils.get_full_cu_seqlens(
6089
6090
6091
6092
                            batch_size,
                            max_seqlen_kv,
                            key_layer.device,
                        )
6093

6094
            # set ALiBi attributes
6095
6096
6097
6098
6099
6100
6101
6102
            global _alibi_cache
            if alibi_slopes is not None:
                assert (
                    core_attention_bias_type == "alibi"
                ), "core_attention_bias_type must be alibi in order to use alibi_slopes!"
                if self.layer_number == 1:
                    _alibi_cache["_alibi_slopes_require_update"] = True
                    _alibi_cache["_alibi_bias_require_update"] = True
6103
            bottom_right_alignment = (attn_mask_type not in ["causal", "padding_causal"],)
6104
6105
6106
6107
6108
6109
6110
6111
            if core_attention_bias_type == "alibi":
                assert (
                    core_attention_bias is None
                ), "core_attention_bias must be None when core_attention_bias_type is alibi!"
                if (
                    _alibi_cache["_num_heads"] != query_layer.shape[-2]
                    or _alibi_cache["_max_seqlen_q"] != max_seqlen_q
                    or _alibi_cache["_max_seqlen_kv"] != max_seqlen_kv
6112
                    or _alibi_cache["_bottom_right_alignment"] != bottom_right_alignment
6113
6114
6115
6116
6117
                    or _alibi_cache["_alibi_slopes"] is None
                ):
                    _alibi_cache["_alibi_slopes_require_update"] = True
                    _alibi_cache["_alibi_bias_require_update"] = True

6118
            # detect bias shape
6119
6120
            core_attention_bias_shape = None
            if core_attention_bias is not None:
6121
                if (
6122
6123
                    core_attention_bias.shape[0] == batch_size
                    and core_attention_bias.shape[1] == query_layer.shape[-2]
6124
                ):
6125
6126
6127
6128
6129
6130
6131
6132
6133
6134
6135
6136
6137
6138
6139
6140
6141
                    core_attention_bias_shape = "bhss"
                elif (
                    core_attention_bias.shape[0] == 1
                    and core_attention_bias.shape[1] == query_layer.shape[-2]
                ):
                    core_attention_bias_shape = "1hss"
                elif (
                    core_attention_bias.shape[0] == batch_size and core_attention_bias.shape[1] == 1
                ):
                    core_attention_bias_shape = "b1ss"
                elif core_attention_bias.shape[0] == 1 and core_attention_bias.shape[1] == 1:
                    core_attention_bias_shape = "11ss"
                else:
                    assert (
                        False
                    ), "core_attention_bias must be in one of {bhss, 1hss, b1ss, 11ss} shapes"

6142
6143
6144
6145
6146
6147
6148
6149
6150
6151
6152
            if pad_between_seqs is None:
                if qkv_format == "thd":
                    pad_between_seqs = (
                        cu_seqlens_q_padded is not None
                        and not torch.equal(cu_seqlens_q_padded[:-1], cu_seqlens_q[:-1])
                    ) or (
                        cu_seqlens_kv_padded is not None
                        and not torch.equal(cu_seqlens_kv_padded[:-1], cu_seqlens_kv[:-1])
                    )
                else:
                    pad_between_seqs = False
6153

6154
            # gather attention params for get_attention_backend
6155
            attention_params = dpa_utils.AttentionParams(
6156
6157
6158
6159
                qkv_type=type(query_layer),
                qkv_dtype=query_layer.dtype,
                qkv_layout=qkv_layout,
                batch_size=batch_size,
6160
6161
                num_heads=num_attention_heads,
                num_gqa_groups=num_gqa_groups,
6162
6163
                max_seqlen_q=max_seqlen_q,
                max_seqlen_kv=max_seqlen_kv,
6164
6165
                head_dim_qk=head_dim_qk,
                head_dim_v=head_dim_v,
6166
6167
6168
6169
6170
6171
6172
6173
6174
6175
6176
                attn_mask_type=attn_mask_type,
                window_size=window_size,
                alibi_slopes_shape=alibi_slopes.shape if alibi_slopes is not None else None,
                core_attention_bias_type=core_attention_bias_type,
                core_attention_bias_shape=core_attention_bias_shape,
                core_attention_bias_requires_grad=(
                    core_attention_bias.requires_grad if core_attention_bias is not None else False
                ),
                pad_between_seqs=pad_between_seqs,
                attention_dropout=self.attention_dropout,
                context_parallel=context_parallel,
6177
6178
                deterministic=self.deterministic,
                is_training=self.training,
6179
6180
                fp8=self.fp8,
                fp8_meta=self.fp8_meta,
6181
                inference_params=inference_params,
6182
            )
6183
            global _attention_backends
6184
6185
6186
6187
6188
6189
6190
6191
6192
            if (
                _attention_backends["attention_params"] is None
                or attention_params != _attention_backends["attention_params"]
            ):
                _attention_backends["attention_params"] = attention_params
                _attention_backends["backend_selection_requires_update"] = True
            if _attention_backends["backend_selection_requires_update"]:
                (
                    use_flash_attention,
6193
                    flash_attention_backend,
6194
6195
6196
6197
                    use_fused_attention,
                    fused_attention_backend,
                    use_unfused_attention,
                    _,
6198
6199
6200
6201
                ) = dpa_utils.get_attention_backend(attention_params)
                # Set global _attention_backends var using return value
                # from get_attention_backend()
                _attention_backends["use_flash_attention"] = use_flash_attention
6202
                _attention_backends["flash_attention_backend"] = flash_attention_backend
6203
6204
6205
6206
                _attention_backends["use_fused_attention"] = use_fused_attention
                _attention_backends["fused_attention_backend"] = fused_attention_backend
                _attention_backends["use_unfused_attention"] = use_unfused_attention
                _attention_backends["backend_selection_requires_update"] = False
6207
                if use_flash_attention:
6208
6209
                    self.logger.info(
                        "Running with FlashAttention backend (version %s)",
6210
                        flash_attention_backend,
6211
                    )
6212
6213
6214
6215
                elif use_fused_attention:
                    self.logger.info(
                        "Running with FusedAttention backend (sub-backend %s)",
                        int(fused_attention_backend),
6216
                    )
6217
6218
6219
6220
                elif use_unfused_attention:
                    self.logger.info("Running with UnfusedDotProductAttention backend")
            else:
                use_flash_attention = _attention_backends["use_flash_attention"]
6221
                flash_attention_backend = _attention_backends["flash_attention_backend"]
6222
6223
6224
                use_fused_attention = _attention_backends["use_fused_attention"]
                fused_attention_backend = _attention_backends["fused_attention_backend"]
                use_unfused_attention = _attention_backends["use_unfused_attention"]
6225

6226
6227
            # raise exception if no backend is available
            if sum([use_flash_attention, use_fused_attention, use_unfused_attention]) == 0:
6228
6229
6230
6231
6232
                raise ValueError(
                    "No dot product attention backend is available for the provided inputs. Please"
                    " run with NVTE_DEBUG=1 NVTE_DEBUG_LEVEL=2 to find out the reasons for"
                    " disabling all backends."
                )
6233
6234

            # run attention
6235
6236
            if use_flash_attention:
                if core_attention_bias_type == "alibi":
6237
6238
                    alibi_slopes, _ = dpa_utils.get_alibi(
                        _alibi_cache,
6239
6240
6241
6242
6243
6244
6245
6246
6247
6248
6249
6250
6251
6252
6253
6254
6255
6256
6257
                        query_layer.shape[-2],
                        max_seqlen_q,
                        max_seqlen_kv,
                        alibi_slopes=alibi_slopes,
                    )
                return self.flash_attention(
                    query_layer,
                    key_layer,
                    value_layer,
                    attention_mask=attention_mask,
                    qkv_layout=qkv_layout,
                    cu_seqlens_q=cu_seqlens_q,
                    cu_seqlens_kv=cu_seqlens_kv,
                    attn_mask_type=attn_mask_type,
                    window_size=window_size,
                    alibi_slopes=alibi_slopes,
                    cp_group=self.cp_group,
                    cp_global_ranks=self.cp_global_ranks,
                    cp_stream=self.cp_stream,
6258
                    cp_comm_type=self.cp_comm_type,
6259
6260
                    max_seqlen_q=max_seqlen_q,
                    max_seqlen_kv=max_seqlen_kv,
6261
6262
                    fp8=self.fp8 and self.fp8_meta["recipe"].fp8_dpa,
                    fp8_meta=self.fp8_meta,
6263
                    quantizers=self.quantizers,
6264
6265
                    inference_params=inference_params,
                    flash_attention_backend=flash_attention_backend,
6266
                )
6267

6268
            if use_fused_attention:
6269
6270
                fu_core_attention_bias_type = core_attention_bias_type
                fu_core_attention_bias = core_attention_bias
6271
6272
6273
                if core_attention_bias_type == "alibi" and (
                    alibi_slopes is not None or max_seqlen_q != max_seqlen_kv
                ):
6274
                    fu_core_attention_bias_type = "post_scale_bias"
6275
6276
                    _, fu_core_attention_bias = dpa_utils.get_alibi(
                        _alibi_cache,
6277
6278
6279
6280
6281
                        query_layer.shape[-2],
                        max_seqlen_q,
                        max_seqlen_kv,
                        alibi_slopes=alibi_slopes,
                        bias_dtype=query_layer.dtype,
6282
                        bottom_right_alignment=attn_mask_type not in ["causal", "padding_causal"],
6283
                    )
6284
                # checkpoint_core_attention=False
6285
6286
6287
6288
6289
6290
6291
6292
6293
                if checkpoint_core_attention:
                    return self._checkpointed_attention_forward(
                        self.fused_attention,
                        query_layer,
                        key_layer,
                        value_layer,
                        qkv_layout=qkv_layout,
                        cu_seqlens_q=cu_seqlens_q,
                        cu_seqlens_kv=cu_seqlens_kv,
6294
6295
                        cu_seqlens_q_padded=cu_seqlens_q_padded,
                        cu_seqlens_kv_padded=cu_seqlens_kv_padded,
6296
6297
6298
6299
                        max_seqlen_q=max_seqlen_q,
                        max_seqlen_kv=max_seqlen_kv,
                        attn_mask_type=attn_mask_type,
                        attention_mask=attention_mask,
6300
                        window_size=window_size,
6301
6302
6303
6304
6305
6306
6307
                        fused_attention_backend=fused_attention_backend,
                        core_attention_bias_type=fu_core_attention_bias_type,
                        core_attention_bias=fu_core_attention_bias,
                        fast_zero_fill=fast_zero_fill,
                        cp_group=self.cp_group,
                        cp_global_ranks=self.cp_global_ranks,
                        cp_stream=self.cp_stream,
6308
                        cp_comm_type=self.cp_comm_type,
6309
6310
                        fp8=self.fp8 and self.fp8_meta["recipe"].fp8_dpa,
                        fp8_meta=self.fp8_meta,
6311
                        quantizers=self.quantizers,
6312
                        pad_between_seqs=pad_between_seqs,
6313
                        inference_params=inference_params,
6314
6315
                    )
                return self.fused_attention(
6316
6317
6318
6319
6320
6321
                    query_layer,
                    key_layer,
                    value_layer,
                    qkv_layout=qkv_layout,
                    cu_seqlens_q=cu_seqlens_q,
                    cu_seqlens_kv=cu_seqlens_kv,
6322
6323
                    cu_seqlens_q_padded=cu_seqlens_q_padded,
                    cu_seqlens_kv_padded=cu_seqlens_kv_padded,
6324
6325
                    max_seqlen_q=max_seqlen_q,
                    max_seqlen_kv=max_seqlen_kv,
6326
6327
                    attn_mask_type=attn_mask_type,
                    attention_mask=attention_mask,
6328
                    window_size=window_size,
6329
                    fused_attention_backend=fused_attention_backend,
6330
6331
                    core_attention_bias_type=fu_core_attention_bias_type,
                    core_attention_bias=fu_core_attention_bias,
6332
6333
6334
6335
                    fast_zero_fill=fast_zero_fill,
                    cp_group=self.cp_group,
                    cp_global_ranks=self.cp_global_ranks,
                    cp_stream=self.cp_stream,
6336
                    cp_comm_type=self.cp_comm_type,
6337
6338
                    fp8=self.fp8 and self.fp8_meta["recipe"].fp8_dpa,
                    fp8_meta=self.fp8_meta,
6339
                    quantizers=self.quantizers,
6340
                    pad_between_seqs=pad_between_seqs,
6341
                    inference_params=inference_params,
6342
                )
6343

6344
            from .cpu_offload import CPUOffloadEnabled
6345

6346
6347
6348
6349
6350
            if CPUOffloadEnabled:
                warnings.warn(
                    "Attention activation Offloading is only implemented"
                    "with Flash Attention and Fused Attention!"
                )
6351

6352
6353
6354
6355
6356
6357
6358
6359
6360
6361
6362
6363
            if use_unfused_attention:
                if checkpoint_core_attention:
                    return self._checkpointed_attention_forward(
                        self.unfused_attention,
                        query_layer,
                        key_layer,
                        value_layer,
                        qkv_layout=qkv_layout,
                        cu_seqlens_q=cu_seqlens_q,
                        cu_seqlens_kv=cu_seqlens_kv,
                        attn_mask_type=attn_mask_type,
                        attention_mask=attention_mask,
6364
                        window_size=window_size,
6365
6366
6367
                        core_attention_bias_type=core_attention_bias_type,
                        core_attention_bias=core_attention_bias,
                        alibi_slopes=alibi_slopes,
6368
                        inference_params=inference_params,
6369
6370
                    )
                return self.unfused_attention(
6371
6372
6373
                    query_layer,
                    key_layer,
                    value_layer,
6374
6375
6376
6377
6378
                    qkv_layout=qkv_layout,
                    cu_seqlens_q=cu_seqlens_q,
                    cu_seqlens_kv=cu_seqlens_kv,
                    attn_mask_type=attn_mask_type,
                    attention_mask=attention_mask,
6379
                    window_size=window_size,
6380
6381
6382
                    core_attention_bias_type=core_attention_bias_type,
                    core_attention_bias=core_attention_bias,
                    alibi_slopes=alibi_slopes,
6383
                    inference_params=inference_params,
6384
                )
6385
            return None
6386
6387


6388
6389
6390
6391
6392
6393
6394
class MultiheadAttention(torch.nn.Module):
    r"""
    Multi-head Attention (MHA), including Query,
    Key, Value and Output projection.

    .. note::

6395
6396
        Argument :attr:`attention_mask` in the `forward` call is only used when
        :attr:`attn_mask_type` includes '"padding"' or `"arbitrary"`.
6397

6398
6399
6400
6401
6402
6403
6404
6405
6406
6407
6408
6409
6410
6411
6412
6413
6414
6415
6416
6417
6418
6419
6420
6421
6422
    Parameters
    ----------
    hidden_size : int
                 size of each input sample.
    num_attention_heads : int
                         number of attention heads in the transformer layer.
    kv_channels: int, default = `None`
                number of key-value channels. defaults to
                :attr:`hidden_size` / :attr:`num_attention_heads` if `None`.
    attention_dropout: float, default = 0.1
                      dropout probability for the dropout op during multi-head attention.
    layernorm_epsilon : float, default = 1e-5
                       a value added to the denominator of layer normalization
                       for numerical stability.
    init_method : Callable, default = `None`
                 used for initializing weights of QKV and FC1 weights in the following way:
                 `init_method(weight)`. When set to `None`, defaults to
                 `torch.nn.init.normal_(mean=0.0, std=0.023)`.
    output_layer_init_method : Callable, default = `None`
                              used for initializing weights of PROJ and FC2 in the following way:
                              `output_layer_init_method(weight)`. When set to `None`, defaults to
                              `torch.nn.init.normal_(mean=0.0, std=0.023)`.
    layer_number: int, default = `None`
                 layer number of the current `TransformerLayer` when multiple such modules are
                 concatenated to form a transformer block.
6423
6424
    attn_mask_type: {'no_mask', 'padding', 'causal', 'padding_causal', 'causal_bottom_right',
                   'padding_causal_bottom_right','arbitrary'},
6425
                   default = `causal`
6426
6427
6428
6429
6430
                   type of attention mask passed into softmax operation. Overridden by
                   :attr:`attn_mask_type` in the `forward` method. The forward
                   arg is useful for dynamically changing mask types, e.g. a different
                   mask for training and inference. The init arg is useful for cases
                   involving compilation/tracing, e.g. ONNX export.
6431
6432
6433
6434
    window_size: Optional[Tuple[int, int]], default = `None`
                sliding window size for local attention, where query at position i attends to keys
                in [i + seqlen_k - seqlen_q - window_size[0], i + seqlen_k - seqlen_q
                + window_size[1]] inclusive. Special cases (-1, -1) and (-1, 0) mean no sliding
6435
6436
6437
                window and causal mask specifically. Both `causal` and `causal_bottom_right` masks
                map to `window_size = (-1, 0)` and Transformer Engine distinguishes them based on
                `attn_mask_type`. Similar to :attr:`attn_mask_type`, `window_size` can
6438
                be overridden by :attr:`window_size` in `forward` as well.
6439
6440
6441
6442
6443
6444
6445
6446
6447
6448
6449
6450
6451
    num_gqa_groups : int, default = `None`
                         number of GQA groups in the transformer layer.
                         Grouped Query Attention is described in
                         `this paper <https://arxiv.org/pdf/2305.13245.pdf>`_.
                         This only affects the keys and values, not the querys.
                         GQA-1 is equivalent to Multi-Query Attention
                         (`MQA <https://arxiv.org/pdf/1911.02150.pdf>`_), while GQA-H
                         is equivalent to MHA, i.e. `num_gqa_groups = num_attention_heads`.
    return_layernorm_output : bool, default = `False`
                             if set to `True`, output of layernorm is returned from the forward
                             together with the output of the linear transformation.
                             Example use case: residual connection for transformer module is
                             taken post layernorm.
6452
6453
    input_layernorm: bool, default = `False`
                     if set to `True`, layer normalization to the input is applied.
6454
6455
6456
6457
6458
6459
6460
6461
6462
6463
6464
6465
6466
6467
6468
6469
6470
6471
6472
6473
    attention_type: { 'self', 'cross' }, default = 'self'
                   type of attention applied.
    zero_centered_gamma : bool, default = 'False'
                         if set to 'True', gamma parameter in LayerNorm is initialized to 0 and
                         the LayerNorm formula changes to

                         .. math::
                            y = \frac{x - \mathrm{E}[x]}{ \sqrt{\mathrm{Var}[x] + \varepsilon}} *
                            (1 + \gamma) + \beta
    normalization : { 'LayerNorm', 'RMSNorm' }, default = 'LayerNorm'
                   type of normalization applied.
    qkv_weight_interleaved : bool, default = `True`
                            if set to `False`, the QKV weight is interpreted as a concatenation of
                            query, key, and value weights along the `0th` dimension. The default
                            interpretation is that the individual `q`, `k`, and `v` weights for each
                            attention head are interleaved. This parameter is set to `False` when
                            using :attr:`fuse_qkv_params=False`.
    bias : bool, default = `True`
          if set to `False`, the transformer layer will not learn any additive biases.
    device : Union[torch.device, str], default = "cuda"
6474
          The device on which the parameters of the model will be allocated. It is the user's
6475
6476
          responsibility to ensure all parameters are moved to the GPU before running the
          forward pass.
6477
6478
6479
6480
6481
6482
6483
    qkv_format: str, default = `sbhd`
            dimension format for `query_layer`, `key_layer` and `value_layer`,
            {`sbhd`, `bshd`}. `s` stands for the sequence length, `b` batch size,
            `h` the number of heads and `d` head size. `sbhd` and `bshd` formats
            are used for when sequences in a batch are of equal length or padded to
            equal length. Please note that these formats do not reflect how
            tensors `query_layer`, `key_layer`, `value_layer` are laid out in memory.
6484
            For that, please use `get_qkv_layout` to gain the layout information.
6485
6486
6487
6488
6489
6490
6491
6492
6493
6494
6495
6496
6497
6498
6499
6500
6501
6502
6503
6504
6505
6506
6507
6508
6509
6510
6511
6512
6513
6514
6515
6516
6517
6518
6519
6520
6521
6522
6523
6524

    Parallelism parameters
    ----------------------
    set_parallel_mode : bool, default = `False`
                      if set to `True`, QKV and FC1 layers are used as Column Parallel
                      whereas PROJ and FC2 is used as Row Parallel as described
                      `here <https://arxiv.org/pdf/1909.08053.pdf>`_.
    sequence_parallel : bool, default = `False`
                       if set to `True`, uses sequence parallelism.
    tp_group : ProcessGroup, default = `None`
              tensor parallel process group.
    tp_size : int, default = 1
             used as TP (tensor parallel) world size when TP groups are not formed during
             initialization. In this case, users must call the
             `set_tensor_parallel_group(tp_group)` method on the initialized module before the
             forward pass to supply the tensor parallel group needed for tensor and sequence
             parallel collectives.

    Optimization parameters
    -----------------------
    fuse_wgrad_accumulation : bool, default = 'False'
                             if set to `True`, enables fusing of creation and accumulation of
                             the weight gradient. When enabled, it is assumed that the weights
                             have an additional `main_grad` attribute (used instead of the
                             regular `grad`) which is a pre-allocated buffer of the correct
                             size to accumulate gradients in.
    params_dtype : torch.dtype, default = `torch.get_default_dtype()`
                  it controls the type used to allocate the initial parameters. Useful when
                  the model is trained with lower precision and the original FP32 parameters
                  would not fit in GPU memory.
    return_bias : bool, default = `False`
                 when set to `True`, this module will not apply the additive bias itself, but
                 instead return the bias value during the forward pass together with the
                 output of the linear transformation :math:`y = xA^T`. This is useful when
                 the bias addition can be fused to subsequent operations.
    fuse_qkv_params: bool, default = 'False'
                    if set to `True`, `TransformerLayer` module exposes a single fused
                    parameter for query-key-value. This enables optimizations such as QKV
                    fusion without concatentations/splits and also enables the argument
                    `fuse_wgrad_accumulation`.
6525
6526
6527
6528
6529
6530
    """

    def __init__(
        self,
        hidden_size: int,
        num_attention_heads: int,
6531
6532
6533
6534
6535
        kv_channels: Optional[int] = None,
        attention_dropout: float = 0.1,
        layernorm_epsilon: float = 1e-5,
        init_method: Optional[Callable] = None,
        output_layer_init_method: Optional[Callable] = None,
6536
        layer_number: Optional[int] = None,
6537
        attn_mask_type: str = "causal",
6538
        window_size: Optional[Tuple[int, int]] = None,
6539
6540
        tp_group: Optional[dist_group_type] = None,
        tp_size: int = 1,
6541
        num_gqa_groups: Optional[int] = None,
6542
6543
6544
        fuse_wgrad_accumulation: bool = False,
        get_rng_state_tracker: Optional[Callable] = None,
        sequence_parallel: bool = False,
6545
        params_dtype: Optional[torch.dtype] = None,
6546
        return_bias: bool = False,
6547
6548
6549
6550
6551
6552
6553
        return_layernorm_output: bool = False,
        input_layernorm: bool = False,
        attention_type: str = "self",
        set_parallel_mode: bool = False,
        fuse_qkv_params: bool = False,
        zero_centered_gamma: bool = False,
        qkv_weight_interleaved: bool = True,
6554
        ub_overlap_ag: bool = False,
6555
6556
6557
6558
        ub_overlap_rs: bool = False,
        ub_overlap_rs_dgrad: bool = False,
        ub_bulk_dgrad: bool = False,
        ub_bulk_wgrad: bool = False,
6559
        bias: bool = True,
6560
        normalization: str = "LayerNorm",
6561
        device: Union[torch.device, str] = "cuda",
6562
        qkv_format: str = "sbhd",
6563
6564
    ) -> None:
        super().__init__()
6565

6566
        self.qkv_format = qkv_format
6567
        self.attn_mask_type = attn_mask_type
6568
        self.window_size = dpa_utils.check_set_window_size(attn_mask_type, window_size)
6569
        self.layer_number = 1 if layer_number is None else layer_number
6570
6571
6572
6573
6574
        self.input_layernorm = input_layernorm
        self.attention_type = attention_type
        self.get_rng_state_tracker = get_rng_state_tracker
        self.tp_group = tp_group
        self.return_layernorm_output = return_layernorm_output
6575
        self.params_dtype = torch.get_default_dtype() if params_dtype is None else params_dtype
6576
        self.num_attention_heads = num_attention_heads
6577
        self.return_bias = return_bias
6578
6579
        self.cp_size = 1
        self.cp_rank = 0
6580
6581
6582
6583
6584
6585
6586

        kv_channels = kv_channels if kv_channels else (hidden_size // num_attention_heads)

        if init_method is None:
            init_method = get_default_init_method()
        if output_layer_init_method is None:
            output_layer_init_method = get_default_init_method()
6587
6588
6589
6590
6591

        if not fuse_qkv_params:
            qkv_weight_interleaved = False
        self.qkv_weight_interleaved = qkv_weight_interleaved

6592
6593
6594
        assert attention_type in AttnTypes, f"attention_type {attention_type} not supported"
        if layer_number is not None:
            assert layer_number > 0, "layer_number must be a positive integer"
6595
6596
6597
6598
6599
6600

        tp_size = tp_size if tp_group is None else get_distributed_world_size(tp_group)
        self.tp_size = tp_size
        self.sequence_parallel = (tp_size > 1) and sequence_parallel

        self.num_attention_heads_per_partition = divide(num_attention_heads, tp_size)
6601
6602
6603
6604
6605
6606
6607
        self.num_gqa_groups = num_attention_heads if num_gqa_groups is None else num_gqa_groups
        assert (
            num_attention_heads % self.num_gqa_groups == 0
        ), "The number of attention heads must be divisible by the number of GQA groups!"
        assert (
            self.num_gqa_groups % tp_size == 0
        ), "The number of GQA groups must be divisible by tensor parallel size!"
6608
        self.num_gqa_groups_per_partition = int(self.num_gqa_groups // tp_size)
6609
6610
6611
6612

        self.hidden_size_per_attention_head = kv_channels
        self.hidden_size_q = self.hidden_size_per_attention_head * num_attention_heads
        self.hidden_size_kv = self.hidden_size_per_attention_head * self.num_gqa_groups
6613
6614
6615
6616
6617
6618
6619

        common_gemm_kwargs = {
            "fuse_wgrad_accumulation": fuse_wgrad_accumulation,
            "tp_group": tp_group,
            "tp_size": tp_size,
            "get_rng_state_tracker": get_rng_state_tracker,
            "sequence_parallel": sequence_parallel,
6620
            "params_dtype": self.params_dtype,
6621
            "device": device,
6622
6623
6624
6625
        }

        qkv_parallel_mode = "column" if set_parallel_mode else None

cyanguwa's avatar
cyanguwa committed
6626
        if self.attention_type == "self":
6627
6628
            parameters_split = None
            if not fuse_qkv_params:
6629
6630
6631
6632
6633
6634
6635
                parameters_split = collections.OrderedDict(
                    [
                        ("query", self.hidden_size_q),
                        ("key", self.hidden_size_kv),
                        ("value", self.hidden_size_kv),
                    ]
                )
6636
6637
6638
            if self.input_layernorm:
                self.layernorm_qkv = LayerNormLinear(
                    hidden_size,
6639
                    self.hidden_size_q + 2 * self.hidden_size_kv,
6640
6641
6642
6643
6644
6645
                    eps=layernorm_epsilon,
                    init_method=init_method,
                    bias=bias,
                    return_bias=False,
                    parallel_mode=qkv_parallel_mode,
                    return_layernorm_output=return_layernorm_output,
cyanguwa's avatar
cyanguwa committed
6646
                    parameters_split=parameters_split,
6647
6648
6649
                    zero_centered_gamma=zero_centered_gamma,
                    ub_bulk_wgrad=ub_bulk_wgrad,
                    ub_bulk_dgrad=ub_bulk_dgrad,
Jaemin Choi's avatar
Jaemin Choi committed
6650
                    ub_overlap_rs_dgrad=ub_overlap_rs_dgrad,
6651
                    ub_overlap_ag=ub_overlap_ag,
6652
                    normalization=normalization,
Przemyslaw Tredak's avatar
Przemyslaw Tredak committed
6653
                    ub_name="qkv",
6654
6655
6656
6657
6658
                    **common_gemm_kwargs,
                )
            else:
                self.qkv = Linear(
                    hidden_size,
6659
                    self.hidden_size_q + 2 * self.hidden_size_kv,
6660
6661
6662
6663
                    init_method=init_method,
                    bias=bias,
                    return_bias=False,
                    parallel_mode=qkv_parallel_mode,
cyanguwa's avatar
cyanguwa committed
6664
                    parameters_split=parameters_split,
6665
6666
                    **common_gemm_kwargs,
                )
cyanguwa's avatar
cyanguwa committed
6667
        elif self.attention_type == "cross":
6668
6669
6670
            if self.input_layernorm:
                self.layernorm_query = LayerNormLinear(
                    hidden_size,
6671
                    self.hidden_size_q,
6672
6673
6674
6675
6676
                    eps=layernorm_epsilon,
                    init_method=init_method,
                    bias=bias,
                    return_bias=False,
                    parallel_mode=qkv_parallel_mode,
6677
                    parameters_split=("query",) if not fuse_qkv_params else None,
6678
6679
6680
6681
                    return_layernorm_output=return_layernorm_output,
                    zero_centered_gamma=zero_centered_gamma,
                    ub_bulk_wgrad=ub_bulk_wgrad,
                    ub_bulk_dgrad=ub_bulk_dgrad,
Jaemin Choi's avatar
Jaemin Choi committed
6682
                    ub_overlap_rs_dgrad=ub_overlap_rs_dgrad,
6683
                    ub_overlap_ag=ub_overlap_ag,
6684
                    normalization=normalization,
Przemyslaw Tredak's avatar
Przemyslaw Tredak committed
6685
                    ub_name="qkv",
6686
6687
6688
6689
6690
                    **common_gemm_kwargs,
                )
            else:
                self.query_layer = Linear(
                    hidden_size,
6691
                    self.hidden_size_q,
6692
6693
6694
6695
6696
6697
6698
6699
                    init_method=init_method,
                    bias=bias,
                    return_bias=False,
                    parallel_mode=qkv_parallel_mode,
                    **common_gemm_kwargs,
                )
            self.key_value = Linear(
                hidden_size,
6700
                2 * self.hidden_size_kv,
6701
6702
6703
6704
                init_method=init_method,
                bias=bias,
                return_bias=False,
                parallel_mode=qkv_parallel_mode,
6705
                parameters_split=("key", "value") if not fuse_qkv_params else None,
6706
6707
6708
6709
6710
6711
                **common_gemm_kwargs,
            )

        # Attention.
        self.core_attention = DotProductAttention(
            num_attention_heads,
6712
            self.hidden_size_per_attention_head,
6713
6714
            num_gqa_groups=self.num_gqa_groups,
            attention_dropout=attention_dropout,
6715
            qkv_format=self.qkv_format,
6716
6717
6718
6719
            tp_size=tp_size,
            get_rng_state_tracker=get_rng_state_tracker,
            sequence_parallel=sequence_parallel,
            tp_group=tp_group,
6720
            layer_number=self.layer_number,
6721
            attention_type=self.attention_type,
6722
6723
6724
6725
        )

        # Linear
        self.proj = Linear(
6726
            self.hidden_size_q,
6727
6728
6729
            hidden_size,
            init_method=output_layer_init_method,
            bias=bias,
6730
            return_bias=return_bias,
6731
            parallel_mode="row" if set_parallel_mode else None,
6732
6733
            ub_overlap_rs=ub_overlap_rs,
            ub_overlap_ag=ub_overlap_ag,
Przemyslaw Tredak's avatar
Przemyslaw Tredak committed
6734
            ub_name="proj",
6735
6736
6737
6738
            **common_gemm_kwargs,
        )

    def set_tensor_parallel_group(self, tp_group: Union[dist_group_type, None]) -> None:
6739
6740
6741
6742
6743
6744
6745
6746
6747
        """
        Set the tensor parallel group for the given
        module before executing the forward pass.

        Parameters
        ----------
        tp_group : ProcessGroup, default = `None`
                  tensor parallel process group.
        """
6748
6749
        self.tp_group = tp_group

6750
    def set_context_parallel_group(
6751
        self,
6752
        cp_group: Union[dist_group_type, List[dist_group_type], None],
6753
        cp_global_ranks: List[int],
6754
        cp_stream: torch.cuda.Stream,
6755
        cp_comm_type: str = "p2p",
6756
    ) -> None:
6757
6758
6759
6760
6761
6762
        """
        Set the context parallel attributes for the given
        module before executing the forward pass.

        Parameters
        ----------
6763
        cp_group : Union[ProcessGroup, List[ProcessGroup]]
6764
                  context parallel process group.
6765
6766
6767
                  ProcessGroup is for cp_comm_type of "p2p", "all_gather", and "a2a".
                  List[ProcessGroup] is for cp_comm_type of "a2a+p2p", where cp_group[0]
                  and cp_group[1] are for a2a and p2p communications respectively.
6768
6769
6770
6771
        cp_global_ranks : List[int]
                         list of global ranks in the context group.
        cp_stream : torch.cuda.Stream
                   cuda stream for context parallel execution.
6772
        cp_comm_type : str, default = `p2p`
6773
                      inter-gpu communication type for context parallelism.
6774
                      Can be "p2p" or "all_gather" or "a2a", "a2a+p2p".
6775
6776
6777
6778
6779
6780
                      "p2p": Exchange KV chunks with P2P communications in ring topology.
                             P2P is async and can be overlapped with attention compute.
                      "all_gather": All-gather to get full sequence of KV before attention.
                                    The all-gather is not async, and cannot be overlapped.
                      "a2a": Like DeepSpeed Ulysses, scatter attention heads across the CP
                             group, and gather to get full sequence of QKV.
6781
6782
6783
                      "a2a+p2p": hierarchical CP implementation. First applying a2a to QKV
                      across each CP sub-group (e.g., via NVLink), then exchanging KV with
                      p2p between sub-groups (e.g., via IBLink).
6784
        """
6785
6786
6787
6788
6789
6790
6791
6792
6793
6794
6795
6796
6797
6798
6799
        if isinstance(cp_group, dist_group_type):
            self.cp_size = get_distributed_world_size(cp_group)
            self.cp_rank = get_distributed_rank(cp_group)
        elif isinstance(cp_group, list):
            assert len(cp_group) == 2, "Current implementation only supports two-level CP groups!"
            assert (
                cp_comm_type == "a2a+p2p"
            ), "Only cp_comm_type of a2a+p2p requires hierarchical CP groups!"
            cp_size_a2a = get_distributed_world_size(cp_group[0])
            cp_rank_a2a = get_distributed_rank(cp_group[0])
            cp_size_p2p = get_distributed_world_size(cp_group[1])
            cp_rank_p2p = get_distributed_rank(cp_group[1])
            self.cp_size = cp_size_a2a * cp_size_p2p
            self.cp_rank = cp_size_a2a * cp_rank_p2p + cp_rank_a2a

6800
6801
6802
6803
6804
        # Deep iterate but skip self to avoid infinite recursion.
        for index, child in enumerate(self.modules()):
            if index == 0:
                continue
            if hasattr(child, "set_context_parallel_group"):
6805
                child.set_context_parallel_group(cp_group, cp_global_ranks, cp_stream, cp_comm_type)
6806

6807
6808
6809
    def forward(
        self,
        hidden_states: torch.Tensor,
6810
        attention_mask: Optional[Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]] = None,
6811
        encoder_output: Optional[torch.Tensor] = None,
6812
        attn_mask_type: Optional[str] = None,
6813
        window_size: Optional[Tuple[int, int]] = None,
6814
6815
        is_first_microbatch: Optional[bool] = None,
        checkpoint_core_attention: bool = False,
6816
        inference_params: Optional[InferenceParams] = None,
6817
        rotary_pos_emb: Optional[Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]] = None,
6818
6819
        core_attention_bias_type: str = "no_bias",
        core_attention_bias: Optional[torch.Tensor] = None,
6820
        alibi_slopes: Optional[torch.Tensor] = None,
6821
6822
6823
6824
        cu_seqlens_q: Optional[torch.Tensor] = None,
        cu_seqlens_kv: Optional[torch.Tensor] = None,
        max_seqlen_q: Optional[int] = None,
        max_seqlen_kv: Optional[int] = None,
6825
        fast_zero_fill: bool = True,
6826
        pad_between_seqs: Optional[bool] = None,
6827
    ) -> Tuple[Union[torch.Tensor, None], ...]:
6828
6829
6830
6831
6832
        """
        Forward propagation for MultiheadAttention layer.

        .. note::

6833
6834
            Argument :attr:`attention_mask` is only used when :attr:`attn_mask_type`
            includes `"padding"` or `"arbitrary"`.
6835
6836
6837
6838
6839

        Parameters
        ----------
        hidden_states : torch.Tensor
             Input tensor.
6840
6841
        attention_mask: Optional[Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]],
             default = `None`. Boolean tensor(s) used to mask out attention softmax input.
6842
             It should be `None` for causal masks and "`no_mask`". For padding masks, it should be
6843
6844
             a single tensor of [batch_size, 1, 1, seqlen_q] for self-attention, and a tuple of
             two tensors in shapes [batch_size, 1, 1, seqlen_q] and [batch_size, 1, 1, seqlen_kv]
6845
6846
6847
6848
6849
6850
             for cross-attention. For "`arbitrary`" mask, it should be in a shape broadcastable to
             [batch_size, num_heads, max_seqlen_q, max_seqlen_kv]. A `True` value means
             the corresponding position is masked out and a `False` means that position
             is allowed to participate in attention.
        attn_mask_type: {'no_mask', 'padding', 'causal', 'padding_causal', 'causal_bottom_right',
                       'padding_causal_bottom_right','arbitrary'},
6851
                       default = `None`
6852
6853
6854
6855
                       type of attention mask passed into softmax operation. By default,
                       causal masks are aligned to the top left corner of the softmax matrix.
                       When "`bottom_right`" is specified in the mask type, causal masks are
                       aligned to the bottom right corner.
6856
6857
        window_size: Optional[Tuple[int, int]], default = `None`
                    sliding window size for local attention.
6858
6859
6860
6861
6862
6863
6864
6865
6866
6867
6868
6869
6870
6871
6872
6873
6874
6875
6876
6877
6878
6879
6880
6881
6882
        encoder_output : Optional[torch.Tensor], default = `None`
             Output of the encoder block to be fed into the decoder block if using
             `layer_type="decoder"`.
        is_first_microbatch : {True, False, None}, default = None
                             During training using either gradient accumulation or
                             pipeline parallelism a minibatch of data is further split
                             into microbatches. Between the microbatches of the same minibatch
                             the model weights are not updated. Setting this parameter indicates
                             whether the current microbatch is the first in a minibatch or not.
                             When set, this parameter enables additional optimizations:

                             * during FP8 training, it allows caching of the FP8 versions of
                               the weights
                             * it also allows skipping gradient accumulation during the
                               first microbatch (since it is the first gradient being
                               produced)
        checkpoint_core_attention: bool, default = `False`
                                  If true, forward activations for core attention are recomputed
                                  during the backward pass in order to save memory that would
                                  otherwise be occupied to store the forward activations until
                                  backprop.
        rotary_pos_emb: Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]], default = `None`
                       Embeddings for query and key tensors for applying rotary position
                       embedding. By default no input embedding is applied.
        core_attention_bias_type: str, default = `no_bias`
6883
                    Bias type, {`no_bias`, `pre_scale_bias`, 'post_scale_bias`, `alibi`}
6884
        core_attention_bias: Optional[torch.Tensor], default = `None`
6885
6886
                    Bias tensor for Q * K.T, shape [1, num_head, max_seqlen_q, max_seqlen_kv].
                    It should be 'None' for 'no_bias' and 'alibi' bias types.
6887
6888
6889
6890
        alibi_slopes: Optional[torch.Tensor], default = `None`
                     ALiBi slopes in FP32 and shape [nheads] or [batch_size, nheads].
                     It adds a bias of (-alibi_slope * (i + seqlen_k - seqlen_q - j))
                     to the attention score of query i and key j.
6891
6892
6893
6894
6895
6896
6897
6898
6899
6900
6901
6902
        cu_seqlens_q: Optional[torch.Tensor], default = `None`
                   Cumulative sum of sequence lengths (without offset) in a batch for `query_layer`,
                   with shape [batch_size + 1] and dtype torch.int32.
        cu_seqlens_kv: Optional[torch.Tensor], default = `None`
                   Cumulative sum of sequence lengths (without offset) in a batch for `key_layer`
                   and `value_layer`, with shape [batch_size + 1] and dtype torch.int32.
        max_seqlen_q: Optional[int], default = `None`
                      Maximum sequence length in `query_layer`.
                      Calculated from `cu_seqlens_q` if not provided.
        max_seqlen_kv: Optional[int], default = `None`
                       Maximum sequence length in `key_layer` and `value_layer`.
                       Calculated from `cu_seqlens_kv` if not provided.
6903
6904
        fast_zero_fill: bool, default = `True`
                    Whether to set output tensors to 0 or not before use.
6905
6906
6907
        pad_between_seqs: Optional[bool], default = `None`
            If None, inferred from qkv_format, cu_seqlens and cu_seqlens_padded.
            If true, there are padding tokens between individual sequences in a packed batch.
6908
        """
6909
6910
        # hidden_states: [sq, b, h]

6911
        if attn_mask_type is None:
6912
            attn_mask_type = self.attn_mask_type
6913
6914
        if window_size is None:
            window_size = self.window_size
6915
        window_size = dpa_utils.check_set_window_size(attn_mask_type, window_size)
6916

6917
        if "padding" in attn_mask_type and attention_mask is not None:
6918
6919
            for mask in attention_mask:
                assert mask.dtype == torch.bool, "Attention mask must be in boolean type!"
6920

6921
6922
6923
        assert (
            core_attention_bias_type in AttnBiasTypes
        ), f"core_attention_bias_type {core_attention_bias_type} is not supported!"
6924

6925
        # =================================================
6926
        # Pre-allocate memory for key-value cache for inference
6927
6928
        # =================================================

6929
6930
6931
6932
6933
        if (
            inference_params is not None
            and self.layer_number not in inference_params.cache_manager.cache
        ):
            inference_params.allocate_memory(self.layer_number)
6934

6935
        # ======================
6936
        # Query, Key, and Value
6937
        # ======================
6938

6939
6940
6941
6942
6943
        fp8_mha = (
            FP8GlobalStateManager.is_fp8_enabled()
            and FP8GlobalStateManager.get_fp8_recipe().fp8_mha
        )

6944
        layernorm_output = None
cyanguwa's avatar
cyanguwa committed
6945
6946
        if self.attention_type == "self":
            # Attention heads [sq, b, h] --> [sq, b, ng * (np/ng + 2) * hn]
6947
6948
6949
6950
            if self.input_layernorm:
                layernorm_qkv_outputs = self.layernorm_qkv(
                    hidden_states,
                    is_first_microbatch=is_first_microbatch,
6951
                    fp8_output=fp8_mha and rotary_pos_emb is None,
6952
6953
6954
6955
6956
6957
6958
6959
6960
                )
                if self.return_layernorm_output:
                    mixed_x_layer, layernorm_output = layernorm_qkv_outputs
                else:
                    mixed_x_layer = layernorm_qkv_outputs
            else:
                mixed_x_layer = self.qkv(
                    hidden_states,
                    is_first_microbatch=is_first_microbatch,
6961
                    fp8_output=fp8_mha and rotary_pos_emb is None,
6962
6963
                )

6964
6965
6966
            num_queries_per_key_value = (
                self.num_attention_heads_per_partition // self.num_gqa_groups_per_partition
            )
6967
            if self.qkv_weight_interleaved:
cyanguwa's avatar
cyanguwa committed
6968
                # [sq, b, ng * (np/ng + 2) * hn] --> [sq, b, ng, (np/ng + 2), hn]
6969
                new_tensor_shape = mixed_x_layer.size()[:-1] + (
cyanguwa's avatar
cyanguwa committed
6970
6971
                    self.num_gqa_groups_per_partition,
                    (num_queries_per_key_value + 2),
6972
6973
6974
6975
                    self.hidden_size_per_attention_head,
                )
                # split along second last dimension
                split_dim = -2
cyanguwa's avatar
cyanguwa committed
6976
6977
6978
6979
6980
            else:
                # [sq, b, ng * (np/ng + 2) * hn] --> [sq, b, (np/ng + 2), ng, hn]
                new_tensor_shape = mixed_x_layer.size()[:-1] + (
                    (num_queries_per_key_value + 2),
                    self.num_gqa_groups_per_partition,
6981
                    self.hidden_size_per_attention_head,
cyanguwa's avatar
cyanguwa committed
6982
6983
6984
                )
                # split along third last dimension
                split_dim = -3
6985
6986
6987

            mixed_x_layer = mixed_x_layer.view(*new_tensor_shape)

cyanguwa's avatar
cyanguwa committed
6988
6989
6990
6991
6992
6993
            # qkv_weight_interleaved:
            #  [sq, b, ng, (np/ng + 2), hn]
            #  --> [sq, b, ng, np/ng, hn], [sq, b, ng, 1, hn], [sq, b, ng, 1, hn]
            # not qkv_weight_interleaved:
            #  [sq, b, (np/ng + 2), ng, hn]
            #  --> [sq, b, np/ng, np, hn], [sq, b, 1, ng, hn], [sq, b, 1, ng, hn]
6994
6995
6996
            query_layer, key_layer, value_layer = _SplitAlongDim.apply(
                mixed_x_layer, split_dim, (num_queries_per_key_value, 1, 1)
            )
cyanguwa's avatar
cyanguwa committed
6997

6998
6999
7000
7001
7002
7003
7004
7005
7006
7007
7008
7009
            if self.qkv_format == "thd":
                query_layer, key_layer, value_layer = (
                    x.reshape(x.size(0), -1, self.hidden_size_per_attention_head)
                    for x in (query_layer, key_layer, value_layer)
                )
            else:
                # query: -> [sq, b, np, hn]
                # key, value: -> [sq, b, ng, hn]
                query_layer, key_layer, value_layer = (
                    x.reshape(x.size(0), x.size(1), -1, self.hidden_size_per_attention_head)
                    for x in (query_layer, key_layer, value_layer)
                )
cyanguwa's avatar
cyanguwa committed
7010
7011
        elif self.attention_type == "cross":
            # Attention heads [sk, b, h] --> [sk, b, (ng * 2 * hn)]
7012
            mixed_kv_layer = self.key_value(
cyanguwa's avatar
cyanguwa committed
7013
                encoder_output,
7014
                is_first_microbatch=is_first_microbatch,
7015
                fp8_output=fp8_mha and rotary_pos_emb is None,
7016
7017
7018
            )

            if self.qkv_weight_interleaved:
cyanguwa's avatar
cyanguwa committed
7019
                # [sq, b, (ng * 2 * hn)] --> [sq, b, ng, 2 * hn]
7020
                new_tensor_shape = mixed_kv_layer.size()[:-1] + (
7021
                    self.num_gqa_groups_per_partition,
7022
7023
7024
7025
7026
                    2 * self.hidden_size_per_attention_head,
                )
                # split along last dimension
                split_dim = -1
            else:
cyanguwa's avatar
cyanguwa committed
7027
                # [sq, b, (ng * 2 * hn)] --> [sq, b, 2 * ng, hn]
7028
                new_tensor_shape = mixed_kv_layer.size()[:-1] + (
7029
                    2 * self.num_gqa_groups_per_partition,
7030
7031
7032
7033
7034
7035
7036
                    self.hidden_size_per_attention_head,
                )
                # split along second last dimension
                split_dim = -2

            mixed_kv_layer = mixed_kv_layer.view(*new_tensor_shape)

cyanguwa's avatar
cyanguwa committed
7037
            # mixed_kv_layer --> 2 [sk, b, ng, hn]
7038
7039
7040
7041
7042
            key_layer, value_layer = _SplitAlongDim.apply(
                mixed_kv_layer,
                split_dim,
                mixed_kv_layer.shape[split_dim] // 2,
            )
7043
7044
7045
7046
7047
7048
7049
7050
7051
            key_layer, value_layer = (
                x.reshape(
                    x.size(0),
                    x.size(1),
                    -1,
                    self.hidden_size_per_attention_head,
                )
                for x in (key_layer, value_layer)
            )
7052
7053
7054
7055
7056
7057

            # Attention head [sq, b, h] --> [sq, b, hp]
            if self.input_layernorm:
                layernorm_query_outputs = self.layernorm_query(
                    hidden_states,
                    is_first_microbatch=is_first_microbatch,
7058
                    fp8_output=fp8_mha and rotary_pos_emb is None,
7059
7060
7061
7062
7063
7064
7065
7066
7067
                )
                if self.return_layernorm_output:
                    query_layer, layernorm_output = layernorm_query_outputs
                else:
                    query_layer = layernorm_query_outputs
            else:
                query_layer = self.query_layer(
                    hidden_states,
                    is_first_microbatch=is_first_microbatch,
7068
                    fp8_output=fp8_mha and rotary_pos_emb is None,
7069
7070
7071
7072
7073
7074
7075
7076
7077
                )

            # [sq, b, hp] --> [sq, b, np, hn]
            new_tensor_shape = query_layer.size()[:-1] + (
                self.num_attention_heads_per_partition,
                self.hidden_size_per_attention_head,
            )
            query_layer = query_layer.view(*new_tensor_shape)

7078
7079
7080
        # ======================================================
        # Apply relative positional encoding (rotary embedding)
        # ======================================================
7081

7082
        if rotary_pos_emb is not None:
7083
7084
7085
            assert not isinstance(query_layer, Float8Tensor) and not isinstance(
                key_layer, Float8Tensor
            ), "RoPE is not supported for Float8Tensors!"
7086
            # duplicate the pos_emb for self attention
7087
            if not isinstance(rotary_pos_emb, tuple):
7088
                rotary_pos_emb = (rotary_pos_emb,) * 2
7089
7090

            q_pos_emb, k_pos_emb = rotary_pos_emb
7091
7092
7093
7094
7095
7096
7097

            # adjust key and value for inference
            if inference_params is not None:
                if self.qkv_format == "sbhd":
                    sequence_length = key_layer.size(0)
                elif self.qkv_format == "bshd":
                    sequence_length = key_layer.size(1)
7098
                else:
7099
7100
7101
                    raise ValueError(
                        f"qkv_format={self.qkv_format} not supported for KV caching and RoPE."
                    )
7102

7103
7104
                sequence_start = inference_params.get_seqlens_pre_step()
                # sequence_start = inference_params.seqlens[0]
7105
7106
7107
7108
7109
                sequence_end = sequence_start + sequence_length

                q_pos_emb = q_pos_emb[sequence_start:sequence_end, ...]
                k_pos_emb = k_pos_emb[sequence_start:sequence_end, ...]

7110
7111
7112
7113
7114
7115
7116
7117
7118
7119
7120
7121
7122
7123
7124
7125
7126
7127
            query_layer = apply_rotary_pos_emb(
                query_layer,
                q_pos_emb,
                self.qkv_format,
                fused=True,
                cu_seqlens=cu_seqlens_q,
                cp_size=self.cp_size,
                cp_rank=self.cp_rank,
            )
            key_layer = apply_rotary_pos_emb(
                key_layer,
                k_pos_emb,
                self.qkv_format,
                fused=True,
                cu_seqlens=cu_seqlens_kv,
                cp_size=self.cp_size,
                cp_rank=self.cp_rank,
            )
7128

7129
7130
7131
7132
        # ===========================
        # Core attention computation
        # ===========================

7133
7134
7135
7136
        context_layer = self.core_attention(
            query_layer,
            key_layer,
            value_layer,
7137
            qkv_format=self.qkv_format,
7138
7139
7140
7141
            cu_seqlens_q=cu_seqlens_q,
            cu_seqlens_kv=cu_seqlens_kv,
            max_seqlen_q=max_seqlen_q,
            max_seqlen_kv=max_seqlen_kv,
7142
7143
            attention_mask=attention_mask,
            attn_mask_type=attn_mask_type,
7144
            window_size=window_size,
7145
7146
7147
            checkpoint_core_attention=checkpoint_core_attention,
            core_attention_bias_type=core_attention_bias_type,
            core_attention_bias=core_attention_bias,
7148
            alibi_slopes=alibi_slopes,
7149
            fast_zero_fill=fast_zero_fill,
7150
            inference_params=inference_params,
7151
            pad_between_seqs=pad_between_seqs,
7152
7153
        )

7154
        # ===================
7155
        # Output. [sq, b, h]
7156
        # ===================
7157
        projection_output = self.proj(
7158
7159
            context_layer,
            is_first_microbatch=is_first_microbatch,
7160
            fp8_grad=isinstance(context_layer, QuantizedTensor),
7161
7162
        )

7163
7164
7165
7166
7167
7168
7169
7170
        if self.return_bias:
            attention_output, attention_bias = projection_output
        else:
            attention_output, attention_bias = projection_output, None

        outputs = (attention_output,)
        if self.return_bias:
            outputs += (attention_bias,)
7171
        if self.input_layernorm and self.return_layernorm_output:
7172
7173
            outputs += (layernorm_output,)
        return outputs if len(outputs) > 1 else outputs[0]