attention.py 270 KB
Newer Older
1
# Copyright (c) 2022-2024, 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
import math
10
import os
11
from typing import Any, Callable, Dict, List, Optional, Tuple, Union
12
import warnings
13
import logging
14

cyanguwa's avatar
cyanguwa committed
15
import numpy as np
16
from packaging.version import Version as PkgVersion
17
18

import torch
19
import torch.nn.functional as F
20

21
import transformer_engine_torch as tex
22
23
import transformer_engine as te
from transformer_engine.pytorch.utils import get_cudnn_version
24
25
26
27
from transformer_engine.pytorch.cpp_extensions import (
    cast_to_fp8,
    cast_from_fp8,
)
28
29
30
31
32
from transformer_engine.pytorch.cpp_extensions.fused_attn import (
    fused_attn_fwd_qkvpacked,
    fused_attn_bwd_qkvpacked,
    fused_attn_fwd_kvpacked,
    fused_attn_bwd_kvpacked,
33
34
    fused_attn_fwd,
    fused_attn_bwd,
35
36
37
38
39
    QKVLayout,
    AttnBiasType,
    AttnMaskType,
    FusedAttnBackend,
)
40
41
from transformer_engine.pytorch.fp8 import get_fp8_te_dtype
from transformer_engine.pytorch.float8_tensor import Float8Tensor
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
)
from transformer_engine.pytorch.export import is_in_onnx_export_mode
69
from transformer_engine.pytorch.jit import jit_fuser, no_torch_dynamo
70
71
from transformer_engine.pytorch.graph import is_graph_capturing

72

73
74
75
_flash_attn_version = PkgVersion(get_pkg_version("flash-attn"))
_flash_attn_version_required = PkgVersion("2.0.6")
_flash_attn_max_version = PkgVersion("2.5.8")
76
_flash_attn_2_plus = _flash_attn_version >= PkgVersion("2")
77
78
79
80
_flash_attn_2_1_plus = _flash_attn_version >= PkgVersion("2.1")
_flash_attn_2_3_plus = _flash_attn_version >= PkgVersion("2.3")
_flash_attn_2_4_plus = _flash_attn_version >= PkgVersion("2.4")
_flash_attn_2_4_1_plus = _flash_attn_version >= PkgVersion("2.4.1")
81

82
if _flash_attn_version >= _flash_attn_version_required:
83
84
85
86
    from flash_attn.flash_attn_interface import flash_attn_varlen_func as flash_attn_forward_func
    from flash_attn.flash_attn_interface import _flash_attn_varlen_forward as _flash_attn_forward
    from flash_attn.flash_attn_interface import _flash_attn_varlen_backward as _flash_attn_backward
    from flash_attn_2_cuda import varlen_bwd as flash_attn_cuda_bwd
87

88
META_QKV = tex.FP8FwdTensors.GEMM1_OUTPUT
89
META_DQKV = tex.FP8BwdTensors.GRAD_OUTPUT1
90
91
92
93
META_O = tex.FP8FwdTensors.GEMM2_INPUT
META_DO = tex.FP8BwdTensors.GRAD_INPUT2
META_S = tex.FP8FwdTensors.GEMM3_OUTPUT
META_DP = tex.FP8BwdTensors.GRAD_INPUT3
94

95
# NVTE_DEBUG = 0/1 # disables/enables debug mode, default = 0
96
_NVTE_DEBUG = int(os.getenv("NVTE_DEBUG", "0"))
97
98
99
100
101
# NVTE_DEBUG_LEVEL = 0/1/2 # enables more and more verbose debug mode, default = 0
_NVTE_DEBUG_LEVEL = int(os.getenv("NVTE_DEBUG_LEVEL", "0"))
log_level = _NVTE_DEBUG * _NVTE_DEBUG_LEVEL
log_levels = {0: logging.WARNING, 1: logging.INFO, 2: logging.DEBUG}
logging.basicConfig(
102
103
    format="[%(levelname)-8s | %(name)-19s]: %(message)s",
    level=log_levels[log_level if log_level in [0, 1, 2] else 2],
104
105
)

106
107
108
109
110
111
112
113
_alibi_cache = {
    "_num_heads": None,
    "_alibi_slopes": None,
    "_max_seqlen_q": None,
    "_max_seqlen_kv": None,
    "_alibi_bias": None,
    "_alibi_slopes_require_update": False,
    "_alibi_bias_require_update": False,
114
}
115
116


117
118
__all__ = ["DotProductAttention", "InferenceParams", "MultiheadAttention"]

119

120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
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
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
407
408
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
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
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
def get_attention_backend(
    qkv_type: Union[torch.Tensor, Float8Tensor] = torch.Tensor,
    qkv_dtype: torch.dtype = torch.bfloat16,
    qkv_layout: str = "sbh3d",
    batch_size: int = 1,
    num_heads: int = 16,
    num_gqa_groups: int = 16,
    max_seqlen_q: int = 128,
    max_seqlen_kv: int = 128,
    head_dim: int = 64,
    attn_mask_type: str = "no_mask",
    window_size: Tuple[int, int] = (-1, -1),
    alibi_slopes_shape: Optional[Union[torch.Size, List]] = None,
    core_attention_bias_type: str = "no_bias",
    core_attention_bias_shape: str = "1hss",
    core_attention_bias_requires_grad: bool = True,
    pad_between_seqs: bool = False,
    attention_dropout: float = 0.0,
    context_parallel: bool = False,
    deterministic: bool = False,
    fp8: bool = False,
    fp8_meta: Optional[Dict[str, Any]] = None,
):
    """
    Select an attention backend based on the user input and runtime environment.

    Parameters
    ----------
    qkv_type: Union[torch.Tensor, Float8Tensor], default = `torch.Tensor`
        Type of query/key/value tensors, {`torch.Tensor`, `Float8Tensor`}.
    qkv_dtype: torch.dtype, default = `torch.bfloat16`
        Data type of query/key/value tensors.
    qkv_layout: str, default = "sbh3d"
        Query/key/value tensor memory layout.
    batch_size: int, default = 1
        Batch size.
    num_heads: int, default = 16
        Number of attention heads in the query tensor.
    num_gqa_groups: int, default = 16
        Number of attention heads in key and value tensors.
    max_seqlen_q: int, default = 128
        Maximum sequence length of the query tensor.
    max_seqlen_kv: int, default = 128
        Maximum sequence length of the key and value tensors.
    head_dim: int, default = 64
        The size of each attention head.
    attn_mask_type: str, default = `no_mask`
        Attention mask type, {`no_mask`, `padding`, `causal`, `padding_causal`,
        `causal_bottom_right`, `padding_causal_bottom_right`, `arbitrary`}
    window_size: Tuple[int, int], default = (-1, -1)
        Sliding window attention size.
    alibi_slopes_shape: Optional[Union[torch.Size, List]], default = `None`
        Tensor shape of :attr:`alibi_slopes` in `DotProductAttention`.
    core_attention_bias_type: str, default = `no_bias`
        Attention bias type, {`no_bias`, `pre_scale_bias`, `post_scale_bias`, `alibi`}.
    core_attention_bias_shape: str, default = `1hss`
        Attention bias shape, {`1hss`, `b1ss`, `bhss`}.
    core_attention_bias_requires_grad: bool, default = `True`
        Whether attention bias requires gradient.
    pad_between_seqs: bool, default = `False`
        Whether there is padding between sequences in a batch.
        This only applies to `qkv_format=thd`.
    attention_dropout: float, default = 0.0
        Attention dropout.
    context_parallel: bool, default = `False`
        Whether context parallelism is used or not.
    deterministic: bool, default = `False`
        Whether to run `DotProductAttention` with determinism or not.
    fp8: bool, default = `False`
        Whether `DotProductAttention` is in an `fp8_autocast` region.
    fp8_meta: Optional[Dict[str Any]], default = `None`
        The FP8 metadata tensor of `DotProductAttention`.

    Returns
    ----------
    use_flash_attention: bool
        Whether the `FlashAttention` backend has been selected.
    use_fused_attention: bool
        Whether the `FusedAttention` backend has been selected.
    use_unfused_attention: bool
        Whether the `UnfusedDotProductAttention` backend has been selected.
    fused_attention_backend: tex.NVTE_Fused_Attn_Backend
        If `use_fused_attention = True`, the `FusedAttention` sub-backend, else `None`.
    available_backends: List[bool]
        All available backends that could support the provided input. A list of Booleans
        in the form of [use_flash_attention, use_fused_attention, use_unfused_attention].
    """
    logger = logging.getLogger("DotProductAttention")

    # Filter: Environment variables
    use_flash_attention = int(os.getenv("NVTE_FLASH_ATTN", "1"))
    use_fused_attention = int(os.getenv("NVTE_FUSED_ATTN", "1"))
    use_unfused_attention = int(os.getenv("NVTE_UNFUSED_ATTN", "1"))
    if not use_flash_attention:
        logger.debug("Disabling FlashAttention due to NVTE_FLASH_ATTN=0")
    if not use_fused_attention:
        logger.debug("Disabling FusedAttention due to NVTE_FUSED_ATTN=0")
    if not use_unfused_attention:
        logger.debug("Disabling UnfusedDotProductAttention due to NVTE_UNFUSED_ATTN=0")

    # Filter: ONNX mode
    if is_in_onnx_export_mode():
        if use_flash_attention:
            logger.debug("Disabling FlashAttention due to ONNX mode")
        use_flash_attention = False
        if use_fused_attention:
            logger.debug("Disabling FusedAttention due to ONNX mode")
        use_fused_attention = False

    # Filter: Compute capability
    device_compute_capability = get_device_compute_capability()
    if device_compute_capability < (8, 0):
        if use_flash_attention:
            logger.debug("Disabling FlashAttention as it requires compute capability sm80+")
            use_flash_attention = False
        if use_fused_attention:
            logger.debug("Disabling FusedAttention as it requires compute capability sm80+")
            use_fused_attention = False

    # Filter: Context parallelism
    if context_parallel and use_unfused_attention:
        logger.debug(
            "Disabling UnfusedDotProductAttention as it does not support context parallelism"
        )
        use_unfused_attention = False

    # Filter: Data type
    if use_flash_attention and (
        qkv_dtype not in [torch.bfloat16, torch.float16] or qkv_type == Float8Tensor
    ):
        logger.debug(
            "Disabling FlashAttention due to unsupported QKV data type. "
            "Supported: qkv_type = torch.Tensor, qkv_dtype = {torch.bfloat16, torch.float16}. "
            "Found: qkv_type = %s, qkv_dtype = %s.",
            qkv_type,
            qkv_dtype,
        )
        use_flash_attention = False
    if use_fused_attention and (qkv_dtype not in [torch.bfloat16, torch.float16]):
        logger.debug(
            "Disabling FusedAttention due to unsupported QKV data type. "
            "Supported: qkv_dtype = {torch.bfloat16, torch.float16}. "
            "Found: qkv_dtype = %s.",
            qkv_dtype,
        )
        use_fused_attention = False

    # Filter: Execution type
    if fp8 and fp8_meta["recipe"].fp8_dpa:
        if use_flash_attention:
            logger.debug("Disabling FlashAttention as it does not support FP8")
            use_flash_attention = False
        if use_unfused_attention:
            logger.debug("Disabling UnfusedDotProductAttention as it does not support FP8")
            use_unfused_attention = False

    # Filter: Head dimension
    if use_flash_attention and (
        head_dim > 256
        or head_dim % 8 != 0
        or (head_dim > 192 and device_compute_capability not in ((8, 0), (9, 0)))
    ):
        logger.debug(
            "Disabling FlashAttention due to unsupported head_dim. "
            "Supported: head_dim %%8 = 0, head_dim <= 256 (>192 requires sm80/90). "
            "Found: head_dim = %s on sm%s.",
            head_dim,
            ".".join([str(i) for i in device_compute_capability]),
        )
        use_flash_attention = False

    # Filter: QKV layout
    qkv_format = "".join([i for i in qkv_layout.split("_")[0] if i.isalpha()])
    if qkv_format == "thd":
        if use_unfused_attention:
            logger.debug("Disabling UnfusedDotProductAttention for qkv_format = thd")
            use_unfused_attention = False
        if use_flash_attention and pad_between_seqs:
            logger.debug(
                "Disabling FlashAttention for qkv_format = thd when there is "
                "padding between sequences, i.e. [a, a, PAD, b, b, b, PAD, c, PAD]"
            )
            use_flash_attention = False

    # Filter: Attention mask
    # attn_mask_type               |     supported backends
    # -------------------------------------------------------------------
    # no_mask                      |     All
    # padding                      |     FlashAttention, FusedAttention
    # causal                       |
    #     self-attention           |     All
    #     cross-attention          |     FusedAttention
    # padding_causal               |
    #     self-attention           |     FlashAttention, FusedAttention
    #     cross-attention          |     FusedAttention
    # causal_bottom_right          |     All
    # padding_causal_bottom_right  |     FlashAttention, FusedAttention
    # arbitrary                    |     UnfusedDotProductAttention
    if attn_mask_type == "arbitrary":
        if use_flash_attention:
            logger.debug("Disabling FlashAttention for arbitrary mask")
        use_flash_attention = False
        if use_fused_attention:
            logger.debug("Disabling FusedAttention for arbitrary mask")
        use_fused_attention = False
    if use_unfused_attention and "padding" in attn_mask_type:
        logger.debug("Disabling UnfusedDotProductAttention for %s mask", attn_mask_type)
        use_unfused_attention = False
    if use_unfused_attention and attn_mask_type == "causal" and max_seqlen_q != max_seqlen_kv:
        logger.debug(
            "Disabling UnfusedDotProductAttention for "
            "top-left-diagonal causal masks for cross-attention"
        )
        use_unfused_attention = False
    if (
        use_flash_attention
        and _flash_attn_2_1_plus
        and attn_mask_type in ["causal", "padding_causal"]
        and max_seqlen_q != max_seqlen_kv
    ):
        logger.warning(
            "Disabling FlashAttention as it only supports bottom-right-diagonal "
            "causal mask since flash-attn 2.1. See "
            "https://github.com/Dao-AILab/flash-attention#21-change-behavior-of-causal-flag"
        )
        use_flash_attention = False
    if (
        use_flash_attention
        and not _flash_attn_2_1_plus
        and attn_mask_type in ["causal_bottom_right", "padding_causal_bottom_right"]
        and max_seqlen_q != max_seqlen_kv
    ):
        logger.warning(
            "Disabling FlashAttention as it only supports top-left-diagonal "
            "causal mask before flash-attn 2.1. See "
            "https://github.com/Dao-AILab/flash-attention#21-change-behavior-of-causal-flag"
        )
        use_flash_attention = False

    # Filter: Sliding window attention
    if window_size is not None and window_size[0] != -1 and window_size[1] not in [-1, 0]:
        if use_unfused_attention:
            logger.debug(
                "Disabling UnfusedDotProductAttention as "
                "it does not support sliding window attention"
            )
            use_unfused_attention = False
        if use_fused_attention:
            logger.debug("Disabling FusedAttention as it does not support sliding window attention")
            use_fused_attention = False
        if use_flash_attention and (not _flash_attn_2_3_plus or context_parallel):
            logger.debug(
                "Disabling FlashAttention as sliding window attention requires "
                "flash-attn 2.3+ and no context parallelism"
            )
            use_flash_attention = False

    # Filter: Attention bias
    if use_flash_attention and (
        core_attention_bias_type not in ["no_bias", "alibi"]
        or core_attention_bias_shape is not None
    ):
        logger.debug("Disabling FlashAttention for pre/post_scale_bias")
        use_flash_attention = False

    fu_core_attention_bias_type = core_attention_bias_type
    fu_core_attention_bias_shape = core_attention_bias_shape
    fu_core_attention_bias_requires_grad = core_attention_bias_requires_grad
    if (
        use_fused_attention
        and core_attention_bias_type == "alibi"
        and alibi_slopes_shape is not None
    ):
        fu_core_attention_bias_type = "post_scale_bias"
        fu_core_attention_bias_requires_grad = False
        if (
            len(alibi_slopes_shape) == 2
            and alibi_slopes_shape[0] == batch_size
            and alibi_slopes_shape[1] == num_heads
        ):
            fu_core_attention_bias_shape = "bhss"
        elif len(alibi_slopes_shape) == 1 and alibi_slopes_shape[0] == num_heads:
            fu_core_attention_bias_shape = "1hss"

    if (
        use_fused_attention
        and fu_core_attention_bias_type == "post_scale_bias"
        and fu_core_attention_bias_shape != "1hss"
    ):
        if fu_core_attention_bias_requires_grad:
            # remove this line when cuDNN adds bwd support for
            # [1, 1, s, s], [b, 1, s, s] and [b, h, s, s]
            logger.debug("Disabling FusedAttention for dBias in [1, H, S, S] shape")
            use_fused_attention = False
        else:
            # max512 backend will only support [1, h, s, s]
            os.environ["NVTE_FUSED_ATTN_BACKEND"] = "1"

    # Filter: cuDNN support
    fused_attention_backend = None
    if use_fused_attention:
        q_type = TE_DType[qkv_dtype]
        kv_type = q_type
        if fp8 and fp8_meta["recipe"].fp8_dpa:
            q_type = get_fp8_te_dtype(fp8_meta["recipe"], fprop_tensor=True)
            kv_type = q_type
        fused_attention_backend = tex.get_fused_attn_backend(
            q_type,
            kv_type,
            QKVLayout[qkv_layout],
            AttnBiasType[fu_core_attention_bias_type],
            AttnMaskType[attn_mask_type],
            attention_dropout,
            num_heads,
            num_gqa_groups,
            max_seqlen_q,
            max_seqlen_kv,
            head_dim,
        )
        if fused_attention_backend == FusedAttnBackend["No_Backend"] or (
            context_parallel and fused_attention_backend != FusedAttnBackend["F16_arbitrary_seqlen"]
        ):
            logger.debug("Disabling FusedAttention as no backend supports the provided input")
            use_fused_attention = False
        elif (
            fused_attention_backend == FusedAttnBackend["F16_max512_seqlen"]
            and fu_core_attention_bias_type == "post_scale_bias"
            and fu_core_attention_bias_shape != "1hss"
        ):
            logger.debug(
                "Disabling FusedAttention as cuDNN sub-backend 0 only supports post_scale_bias in"
                " [1, H, S, S] shape"
            )
            use_fused_attention = False

    # Filter: Determinism
    # backend                      | deterministic
    # ---------------------------------------------
    # FlashAttention               |
    #     flash-attn >=2.0, <2.4.1 | no
    #     flash-attn >=2.4.1       | yes
    # FusedAttention               |
    #     sub-backend 0            | yes
    #     sub-backend 1            | workspace optimization path and sm90+: yes;
    #                              | otherwise: no
    #     sub-backend 2            | no
    # UnfusedDotProductAttention   | yes
    if use_flash_attention and deterministic and not _flash_attn_2_4_1_plus:
        logger.warning(
            "Disabling FlashAttention as version <2.4.1 does not support deterministic "
            "execution. To use FlashAttention with deterministic behavior, "
            "please install flash-attn >= 2.4.1."
        )
        use_flash_attention = False
    if (
        use_fused_attention
        and (
            fused_attention_backend == FusedAttnBackend["FP8"]
            or (
                fused_attention_backend == FusedAttnBackend["F16_arbitrary_seqlen"]
                and device_compute_capability < (9, 0)
            )
        )
        and deterministic
    ):
        logger.debug("Disabling FusedAttention for determinism reasons")
        use_fused_attention = False

    # All available backends
    available_backends = [use_flash_attention, use_fused_attention, use_unfused_attention]

    # Select FusedAttention for performance
    if (
        use_flash_attention
        and use_fused_attention
        and fused_attention_backend == FusedAttnBackend["F16_arbitrary_seqlen"]
    ):
        if device_compute_capability == (9, 0):
            logger.debug(
                "Disabling FlashAttention to give FusedAttention preference on Hopper+ "
                "for performance reasons"
            )
            use_flash_attention = False

    # Selected backend
    if use_flash_attention:
        use_fused_attention = False
        use_unfused_attention = False
    elif use_fused_attention:
        use_unfused_attention = False
    if use_flash_attention:
        selected_backend = "FlashAttention"
    elif use_fused_attention:
        selected_backend = f"FusedAttention (sub-backend {int(fused_attention_backend)})"
    elif use_unfused_attention:
        selected_backend = "UnfusedDotProductAttention"
    else:
        selected_backend = "NoBackend"

    logger.debug(
        "Available backends: FlashAttention=%s, FusedAttention=%s, UnfusedDotProductAttention=%s",
        bool(available_backends[0]),
        bool(available_backends[1]),
        bool(available_backends[2]),
    )
    logger.debug("Selected backend: %s", selected_backend)

    return (
        use_flash_attention,
        use_fused_attention,
        use_unfused_attention,
        available_backends,
        fused_attention_backend,
    )


536
class InferenceParams:  # pylint: disable=too-few-public-methods
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
    """
    Inference parameters that are passed to the main model in order
    to efficienly calculate and store the context during inference.

    Parameters
    ----------
    max_batch_size : int
                    maximum batch size during inference.
    max_sequence_length : int
                         maximum sequence length during inference.
    """

    def __init__(self, max_batch_size, max_sequence_length):
        self.max_sequence_length = max_sequence_length
        self.max_batch_size = max_batch_size
        self.sequence_len_offset = 0
        self.batch_size_offset = 0
        self.key_value_memory_dict = {}

    def swap_key_value_dict(self, batch_indices):
        """
        Reorders the KV cache using the specified batch indices.

        Parameters
        ----------
        batch_indices : List[int]
                       Sequence of indices to reorder along the batch dimensions of
                       the KV cache. Must have a length equal to the batch size.
        """
        if len(self.key_value_memory_dict) == 0:
            raise ValueError("should not swap when dict in empty")

        for layer_number, inference_memory in self.key_value_memory_dict.items():
            inference_key_memory, inference_value_memory = inference_memory
            assert (
                len(batch_indices) == inference_key_memory.shape[1]
            )  # make sure batch size is the same
            new_inference_key_memory = inference_key_memory[:, batch_indices]
            new_inference_value_memory = inference_value_memory[:, batch_indices]
            self.key_value_memory_dict[layer_number] = (
                new_inference_key_memory,
                new_inference_value_memory,
            )
580

581

582
583
584
585
586
@torch.no_grad()
def get_alibi(
    num_heads: int,
    max_seqlen_q: int,
    max_seqlen_kv: int,
587
588
589
    alibi_slopes: Optional[torch.Tensor] = None,
    bias_dtype: Optional[torch.dtype] = None,
) -> Tuple[torch.Tensor, torch.Tensor]:
590
    """
591
592
593
594
595
596
597
598
599
600
601
602
    Parameters
    ----------
    num_heads: int
        Number of heads.
    max_seqlen_q: int
        Maximum sequence length for queries.
    max_seqlen_kv: int
        Maximum sequence length for keys and values.
    alibi_slopes: Optional[torch.Tensor], default = `None`
        Custom ALiBi slopes, FP32, CUDA tensor, in shape [num_heads] or [batch_size, num_heads].
    bias_dtype: Optional[torch.dtype], default = `None`
        Dtype of the generated ALiBi bias. If None, use torch.float32.
603

604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
    Returns
    ----------
    alibi_slopes: torch.Tensor
        ALiBi slopes in FP32 and shape [num_heads] or [batch_size, num_heads].
    alibi_bias: torch.Tensor
        ALiBi bias in FP32 or `bias_dtype`. If `alibi_slopes` is in [num_heads] shape,
        then `alibi_bias` is in [1, num_heads, max_seqlen_q, max_seqlen_kv], and if
        `alibi_slopes` is in [batch_size, num_heads], then the bias is in
        [batch_size, num_heads, max_seqlen_q, max_seqlen_kv].
    """
    global _alibi_cache
    if _alibi_cache["_alibi_slopes_require_update"]:
        if alibi_slopes is not None:
            _alibi_cache["_alibi_slopes"] = alibi_slopes
        else:
            n = 2 ** math.floor(math.log2(num_heads))
            m_0 = 2.0 ** (-8.0 / n)
            m = torch.pow(m_0, torch.arange(1, 1 + n))

            if n < num_heads:
                m_hat_0 = 2.0 ** (-4.0 / n)
                m_hat = torch.pow(m_hat_0, torch.arange(1, 1 + 2 * (num_heads - n), 2))
                m = torch.cat([m, m_hat])

            _alibi_cache["_alibi_slopes"] = m.to(dtype=torch.float32, device="cuda")
        _alibi_cache["_num_heads"] = num_heads
        _alibi_cache["_alibi_slopes_require_update"] = False

    if _alibi_cache["_alibi_bias_require_update"]:
        assert _alibi_cache["_alibi_slopes"] is not None, "ALiBi slopes can not be None!"
        if _alibi_cache["_alibi_slopes"].dim() == 1:
            slopes_shape = torch.Size([1, _alibi_cache["_alibi_slopes"].shape[0], 1, 1])
        if _alibi_cache["_alibi_slopes"].dim() == 2:
            slopes_shape = torch.Size([*_alibi_cache["_alibi_slopes"].shape[:], 1, 1])
638
639
640
641
642
643
        bias = torch.arange(1 - max_seqlen_kv, 1, dtype=torch.int32, device="cuda").view(
            1, 1, 1, max_seqlen_kv
        )
        bias = bias - torch.arange(1 - max_seqlen_q, 1, dtype=torch.int32, device="cuda").view(
            1, 1, max_seqlen_q, 1
        )
644
645
646
647
648
649
650
651
        bias = bias.abs().mul(-1)
        bias = bias * _alibi_cache["_alibi_slopes"].view(slopes_shape)
        _alibi_cache["_max_seqlen_q"], _alibi_cache["_max_seqlen_kv"] = max_seqlen_q, max_seqlen_kv
        bias_dtype = torch.float32 if bias_dtype is None else bias_dtype
        _alibi_cache["_alibi_bias"] = bias.contiguous().to(dtype=bias_dtype, device="cuda")
        _alibi_cache["_alibi_bias_require_update"] = False

    return _alibi_cache["_alibi_slopes"], _alibi_cache["_alibi_bias"]
652
653
654
655
656
657
658
659
660


def get_cu_seqlens(mask: torch.Tensor) -> torch.Tensor:
    """
    Given a padding mask of shape [batch_size, 1, 1, max_seqlen], returns an int32
    tensor of shape [batch_size + 1] containing the cumulative sequence lengths of
    the samples in a batch.
    """
    mask = mask.squeeze(1).squeeze(1)
661
    reduced_mask = mask.logical_not().sum(dim=1)
662
663
664
665
666
667
    cu_seqlens = reduced_mask.cumsum(dim=0).to(torch.int32)
    zero = torch.zeros(1, dtype=torch.int32, device="cuda")
    cu_seqlens = torch.cat((zero, cu_seqlens))

    return cu_seqlens

668

669
670
671
def get_cu_seqlens_and_indices(mask: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
    """
    Given a padding mask of shape [batch_size, 1, 1, max_seqlen], returns an int32
672
673
674
    tensor of shape [batch_size + 1] containing the cumulative sequence lengths of
    the samples in a batch, and another int32 tensor of shape [batch_size * max_seqlen, 1, 1]
    containing the indices for the valid tokens.
675
676
677
678
    """
    mask = mask.squeeze(1).squeeze(1)
    bs, seqlen = mask.shape

679
    reduced_mask = mask.logical_not().sum(dim=1)
680
681
682
683
684
    cu_seqlens = reduced_mask.cumsum(dim=0).to(torch.int32)
    zero = torch.zeros(1, dtype=torch.int32, device="cuda")
    cu_seqlens = torch.cat((zero, cu_seqlens))

    mask = mask.reshape(-1)
685
    indices = mask.logical_not().nonzero()
686
687
688
689
    indices = indices.unsqueeze(-1)

    num_nonzeros = indices.shape[0]
    pad_amount = bs * seqlen - num_nonzeros
690
691
692
    indices = F.pad(
        input=indices, pad=(0, 0, 0, 0, 0, pad_amount), mode="constant", value=float(bs * seqlen)
    )
693
694
695
696

    return cu_seqlens, indices


697
698
699
700
701
702
703
704
def get_indices(max_seqlen: int, cu_seqlens: torch.Tensor) -> torch.Tensor:
    """
    Given max_seqlen and cu_seqlens of shape [batch_size + 1], returns an int32
    tensor of shape [batch_size * max_seqlen, 1, 1] containing the indices for
    the valid tokens in a batch.
    """
    bs = len(cu_seqlens) - 1
    seqlens = cu_seqlens[1:] - cu_seqlens[:-1]
705
706
    indices = [i * max_seqlen + ii for i, j in enumerate(seqlens) for ii in range(j)]
    indices = torch.Tensor(indices).unsqueeze(1).unsqueeze(1).to(dtype=torch.int64, device="cuda")
707
708
709

    num_nonzeros = indices.shape[0]
    pad_amount = bs * max_seqlen - num_nonzeros
710
711
712
713
714
715
    indices = F.pad(
        input=indices,
        pad=(0, 0, 0, 0, 0, pad_amount),
        mode="constant",
        value=float(bs * max_seqlen),
    )
716
717
718

    return indices

719

720
_cu_seqlens_cache = {}
721
722


723
724
725
726
727
728
729
730
731
732
def _get_full_cu_seqlens(
    batch_size: int,
    max_seqlen: int,
    device: torch.device,
) -> torch.Tensor:
    """Cumulative sequence lengths in full data batch

    All sequences in batch have the maximum sequence length.

    """
733
734
735
736
737
738
739
740
741
742
    global _cu_seqlens_cache
    if (batch_size, max_seqlen) not in _cu_seqlens_cache:
        _cu_seqlens_cache[(batch_size, max_seqlen)] = torch.arange(
            0,
            (batch_size + 1) * max_seqlen,
            step=max_seqlen,
            dtype=torch.int32,
            device=device,
        )
    return _cu_seqlens_cache[(batch_size, max_seqlen)]
743
744


745
746
747
748
749
750
751
752
753
@jit_fuser
def pack_tensor(
    indices: torch.Tensor,
    tensor: torch.Tensor,
) -> torch.Tensor:
    """
    Packs the given tensor using the `indices`.
    """
    padding_indice = torch.zeros(
754
755
        1, tensor.shape[1], tensor.shape[2], dtype=tensor.dtype, device=tensor.device
    )
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
    tensor = torch.cat((tensor, padding_indice), dim=0)

    indices = indices.repeat(1, tensor.shape[1], tensor.shape[2])
    packed = torch.gather(tensor, 0, indices)
    return packed


@jit_fuser
def pack_2_tensors(
    indices: torch.Tensor,
    t1: torch.Tensor,
    t2: torch.Tensor,
) -> Tuple[torch.Tensor, torch.Tensor]:
    """
    Packs the given 2 tensors using the `indices`.
    """
    t1_packed = pack_tensor(indices, t1)
    t2_packed = pack_tensor(indices, t2)
    return t1_packed, t2_packed


@jit_fuser
def pack_3_tensors(
    indices: torch.Tensor,
    t1: torch.Tensor,
    t2: torch.Tensor,
    t3: torch.Tensor,
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
    """
    Packs the given 3 tensors using the `indices`.
    """
    t1_packed = pack_tensor(indices, t1)
    t2_packed = pack_tensor(indices, t2)
    t3_packed = pack_tensor(indices, t3)
    return t1_packed, t2_packed, t3_packed


@jit_fuser
def unpack_tensor(
    indices: torch.Tensor,
    dim0: int,
    tensor: torch.Tensor,
) -> torch.Tensor:
    """
    Inverse of `pack_tensor`.
    """
    indices = indices.repeat(1, tensor.shape[1], tensor.shape[2])
    unpacked = torch.zeros(
804
805
        dim0 + 1, tensor.shape[1], tensor.shape[2], dtype=tensor.dtype, device=tensor.device
    )
806
    unpacked.scatter_(0, indices, tensor)
807
    unpacked = unpacked[0:-1, :, :]
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
    return unpacked


@jit_fuser
def unpack_2_tensors(
    indices: torch.Tensor,
    dim0: int,
    t1: torch.Tensor,
    t2: torch.Tensor,
) -> Tuple[torch.Tensor, torch.Tensor]:
    """
    Inverse of `pack_2_tensors`.
    """
    t1_unpacked = unpack_tensor(indices, dim0, t1)
    t2_unpacked = unpack_tensor(indices, dim0, t2)
    return t1_unpacked, t2_unpacked


@jit_fuser
def unpack_3_tensors(
    indices: torch.Tensor,
    dim0: int,
    t1: torch.Tensor,
    t2: torch.Tensor,
    t3: torch.Tensor,
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
    """
    Inverse of `pack_3_tensors`.
    """
    t1_unpacked = unpack_tensor(indices, dim0, t1)
    t2_unpacked = unpack_tensor(indices, dim0, t2)
    t3_unpacked = unpack_tensor(indices, dim0, t3)
    return t1_unpacked, t2_unpacked, t3_unpacked


class PackTensors(torch.autograd.Function):
    """
    Autograd function to pack tensors.
    """
847

848
849
    @staticmethod
    def forward(
850
        ctx, indices: torch.Tensor, *tensors: Tuple[torch.Tensor, ...]
851
852
    ) -> Union[Tuple[torch.Tensor, ...], torch.Tensor]:
        assert 1 <= len(tensors) <= 3, f"Packing {len(tensors)} tensors not supported."
853
        ctx.save_for_backward(indices)
854
855
856
857
858
859
860
861
862
        ctx.dim0 = tensors[0].shape[0]
        if len(tensors) == 1:
            return pack_tensor(indices, *tensors)
        if len(tensors) == 2:
            return pack_2_tensors(indices, *tensors)
        return pack_3_tensors(indices, *tensors)

    @staticmethod
    def backward(ctx, *grad_outputs: Tuple[torch.Tensor, ...]):
863
        (indices,) = ctx.saved_tensors
864
        if len(grad_outputs) == 1:
865
            return None, unpack_tensor(indices, ctx.dim0, *grad_outputs)
866
        if len(grad_outputs) == 2:
867
868
            return None, *unpack_2_tensors(indices, ctx.dim0, *grad_outputs)
        return None, *unpack_3_tensors(indices, ctx.dim0, *grad_outputs)
869
870
871
872
873
874


class UnpackTensor(torch.autograd.Function):
    """
    Autograd function to unpack a tensor.
    """
875

876
877
878
879
880
881
882
    @staticmethod
    def forward(
        ctx,
        indices: torch.Tensor,
        dim0: int,
        tensor: torch.Tensor,
    ) -> torch.Tensor:
883
        ctx.save_for_backward(indices)
884
885
886
887
        return unpack_tensor(indices, dim0, tensor)

    @staticmethod
    def backward(ctx, grad_output):
888
889
        (indices,) = ctx.saved_tensors
        return None, None, pack_tensor(indices, grad_output)
890
891


892
893
894
def flash_attn_p2p_communicate(
    rank, send_tensor, send_dst, recv_tensor, recv_src, cp_group, batch_p2p_comm
):
895
    """Point-to-point communications of KV and dKV in Attention with context parallelism"""
896
897
898
899
    send_recv_ops = []

    if batch_p2p_comm:
        if rank % 2 == 0:
900
901
902
903
904
905
            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
            )
906
907
908
            send_recv_ops.append(send_op)
            send_recv_ops.append(recv_op)
        else:
909
910
911
912
913
914
            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
            )
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
            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


934
@jit_fuser
935
def flash_attn_fwd_out_correction(out, out_per_step, seq_dim, softmax_lse, softmax_lse_per_step):
936
    """Merge partial outputs of each step in Attention with context parallelism"""
937
    softmax_lse_corrected_exp = torch.exp(softmax_lse_per_step - softmax_lse).movedim(2, seq_dim)
938
    softmax_lse_corrected_exp = softmax_lse_corrected_exp.unsqueeze(-1)
939
    out_corrected = out_per_step * softmax_lse_corrected_exp
940
941
942
    out.add_(out_corrected)


943
@jit_fuser
944
def flash_attn_fwd_softmax_lse_correction(softmax_lse, softmax_lse_per_step):
945
    """Merge softmax stats of each step in Attention with context parallelism"""
946
947
948
949
    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.log(1 + torch.exp(min_scale - max_scale))
    softmax_lse.copy_(new_scale)
950
951


952
class AttnFuncWithCP(torch.autograd.Function):
953
    """
954
955
    Attention implementation with context parallelism.
    Split attention compute into multiple steps, and overlap current-step
956
957
958
959
    compute with next-step communication.
    """

    @staticmethod
960
961
962
963
964
965
966
967
968
969
    def forward(
        ctx,
        is_training,
        q,
        k,
        v,
        cu_seqlens_q,
        cu_seqlens_k,
        max_seqlen_q,
        max_seqlen_k,
970
971
        cu_seqlens_q_padded,
        cu_seqlens_kv_padded,
972
973
974
975
976
977
978
979
980
981
982
983
        dropout_p,
        cp_group,
        cp_global_ranks,
        cp_stream,
        softmax_scale,
        qkv_format,
        attn_mask_type,
        attn_bias_type,
        attn_bias,
        deterministic,
        use_fused_attention,
    ):
984
985
986
987
988
989
        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)
        send_dst = cp_global_ranks[(rank + 1) % cp_size]
990
        recv_src = cp_global_ranks[(rank - 1) % cp_size]
991
992
        batch_p2p_comm = int(os.getenv("NVTE_BATCH_MHA_P2P_COMM", "0")) or (cp_size == 2)

993
994
        causal = "causal" in attn_mask_type
        padding = "padding" in attn_mask_type
995

996
997
        qkv_layout = qkv_format + "_" + qkv_format + "_" + qkv_format

998
        if causal:
999
1000
            if qkv_format == "bshd":
                # [b, s, np, hn] -> [b, 2, s//2, np, hn]
1001
                q, k, v = [x.view(x.shape[0], 2, x.shape[1] // 2, *x.shape[2:]) for x in [q, k, v]]
1002
1003
            elif qkv_format == "sbhd":
                # [s, b, np, hn] -> [2, s//2, b, np, hn]
1004
                q, k, v = [x.view(2, x.shape[0] // 2, *x.shape[1:]) for x in [q, k, v]]
1005
        if attn_bias is not None:
1006
            assert len(attn_bias.shape) == 4, (
1007
1008
1009
1010
                "Only support bias shape of [b, h, sq, sk] for forward, "
                "and [1, h, sq, sk] for backward!"
            )
            # [b, np, sq, sk] -> [b, np, 2, sq//2, 2*cp, sk//(2*cp)]
1011
1012
1013
1014
1015
1016
            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),
1017
1018
            )
            # [b, np, sq, sk] -> [b, np, sq, 2*cp, sk//(2*cp)]
1019
1020
            attn_bias = attn_bias.view(
                *attn_bias.shape[:-1], 2 * cp_size, attn_bias.shape[-1] // (2 * cp_size)
1021
            )
1022
        assert q.shape[-1] % 8 == 0, "hidden size per attention head should be multiple of 8"
1023
1024
1025
1026
1027
        fa_optional_forward_kwargs = {}
        if _flash_attn_2_3_plus:
            fa_optional_forward_kwargs["window_size"] = [-1, 0] if causal else [-1, -1]
        if _flash_attn_2_4_plus:
            fa_optional_forward_kwargs["alibi_slopes"] = None
1028

1029
1030
1031
        # Flash Attn inputs
        q_inputs = [None, None]
        kv_inputs = [None, None]
1032
        attn_bias_inputs = [None, None]
1033
1034
1035
1036
        # 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)]
1037
        attn_biases = [None for _ in range(cp_size)]
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047

        # 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)]
        p2p_comm_buffers[0] = torch.cat((k.unsqueeze(0), v.unsqueeze(0)), dim=0)
        send_recv_reqs = [[], []]

1048
        for i in range(cp_size + 1):
1049
            if i < cp_size:
1050
                with torch.cuda.stream(flash_attn_streams[i % 2]):
1051
                    # wait until KV is received
1052
                    for req in send_recv_reqs[(i + 1) % 2]:
1053
1054
                        req.wait()

1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
                    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,
                        )

                    kv_inputs[i % 2] = p2p_comm_buffers[i]
1068
1069
                    if causal:
                        if i == 0:
1070
                            if use_fused_attention:
1071
1072
                                if qkv_format == "bshd":
                                    # [b, 2, sq//2, np, hn] -> [b, sq, np, hn]
1073
                                    q_inputs[i % 2] = q.view(q.shape[0], -1, *q.shape[-2:])
1074
                                    # [2, b, 2, sk//2, np, hn] -> [2, b, sk, np, hn]
1075
1076
1077
                                    kv_inputs[i % 2] = kv_inputs[i % 2].view(
                                        2, k.shape[0], -1, *k.shape[-2:]
                                    )
1078
1079
                                elif qkv_format == "sbhd":
                                    # [2, sq//2, b, np, hn] -> [sq, b, np, hn]
1080
                                    q_inputs[i % 2] = q.view(-1, *q.shape[-3:])
1081
                                    # [2, 2, sk//2, b, np, hn] -> [2, sk, b, np, hn]
1082
                                    kv_inputs[i % 2] = kv_inputs[i % 2].view(2, -1, *k.shape[-3:])
1083
                                elif qkv_format == "thd":
1084
                                    q_inputs[i % 2] = q
1085
1086
                                if attn_bias is not None:
                                    idx = (rank - i) % cp_size
1087
1088
1089
1090
1091
1092
                                    attn_bias_inputs[i % 2] = torch.cat(
                                        (
                                            attn_bias[..., idx, :],
                                            attn_bias[..., (2 * cp_size - idx - 1), :],
                                        ),
                                        dim=-1,
1093
                                    ).contiguous()
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
                                out_per_step[i], [softmax_lse_per_step[i], rng_states[i], *rest] = (
                                    fused_attn_fwd(
                                        is_training,
                                        max_seqlen_q,
                                        max_seqlen_k,
                                        cu_seqlens_q,
                                        cu_seqlens_k,
                                        q_inputs[i % 2],
                                        kv_inputs[i % 2][0],
                                        kv_inputs[i % 2][1],
                                        TE_DType[q.dtype],
                                        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_inputs[i % 2],
1112
1113
                                        cu_seqlens_q_padded=cu_seqlens_q_padded,
                                        cu_seqlens_kv_padded=cu_seqlens_kv_padded,
1114
                                    )
1115
                                )
1116
1117
                                if len(rest) > 0:
                                    attn_biases[i] = rest[0]
1118
1119
                            else:
                                # [b, 2, sq//2, np, hn] -> [b*sq, np, hn]
1120
                                q_inputs[i % 2] = q.view(-1, *q.shape[-2:])
1121
                                # [2, b, 2, sk//2, np, hn] -> [2, b*sk, np, hn]
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
                                kv_inputs[i % 2] = kv_inputs[i % 2].view(2, -1, *k.shape[-2:])
                                (
                                    _,
                                    _,
                                    _,
                                    _,
                                    out_per_step[i],
                                    softmax_lse_per_step[i],
                                    _,
                                    rng_states[i],
                                ) = _flash_attn_forward(
                                    q_inputs[i % 2],
                                    kv_inputs[i % 2][0],
                                    kv_inputs[i % 2][1],
                                    cu_seqlens_q,
                                    cu_seqlens_k,
                                    max_seqlen_q,
                                    max_seqlen_k,
                                    dropout_p,
                                    softmax_scale,
                                    causal=True,
                                    return_softmax=False,
                                    **fa_optional_forward_kwargs,
1145
                                )
1146
                        elif i <= rank:
1147
                            if use_fused_attention:
1148
1149
                                if qkv_format == "bshd":
                                    # [b, 2, sq//2, np, hn] -> [b, sq, np, hn]
1150
                                    q_inputs[i % 2] = q.view(q.shape[0], -1, *q.shape[-2:])
1151
                                    # [2, b, 2, sk//2, np, hn] -> [2, b, sk//2, np, hn]
1152
                                    kv_inputs[i % 2] = kv_inputs[i % 2][:, :, 0, ...].contiguous()
1153
1154
                                elif qkv_format == "sbhd":
                                    # [2, sq//2, b, np, hn] -> [sq, b, np, hn]
1155
                                    q_inputs[i % 2] = q.view(-1, *q.shape[-3:])
1156
                                    # [2, 2, sk//2, b, np, hn] -> [2, sk//2, b, np, hn]
1157
                                    kv_inputs[i % 2] = kv_inputs[i % 2][:, 0, ...].contiguous()
1158
                                elif qkv_format == "thd":
1159
                                    q_inputs[i % 2] = q
1160
                                    # [2, t, np, hn] -> [2, t/2, np, hn]
1161
1162
1163
                                    kv_inputs[i % 2] = tex.thd_read_half_tensor(
                                        kv_inputs[i % 2], cu_seqlens_k, 0
                                    )
1164
1165
                                if attn_bias is not None:
                                    idx = (rank - i) % cp_size
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
                                    attn_bias_inputs[i % 2] = attn_bias[..., idx, :].contiguous()
                                out_per_step[i], [softmax_lse_per_step[i], rng_states[i], *rest] = (
                                    fused_attn_fwd(
                                        is_training,
                                        max_seqlen_q,
                                        max_seqlen_k // 2,
                                        cu_seqlens_q,
                                        cu_seqlens_k // 2,
                                        q_inputs[i % 2],
                                        kv_inputs[i % 2][0],
                                        kv_inputs[i % 2][1],
                                        TE_DType[q.dtype],
                                        tex.NVTE_Fused_Attn_Backend.NVTE_F16_arbitrary_seqlen,
                                        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],
1185
1186
1187
1188
1189
                                        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
1190
1191
                                        ),
                                    )
1192
                                )
1193
1194
                                if len(rest) > 0:
                                    attn_biases[i] = rest[0]
1195
1196
                            else:
                                # [b, 2, sq//2, np, hn] -> [b*sq, np, hn]
1197
                                q_inputs[i % 2] = q.view(-1, *q.shape[-2:])
1198
1199
                                if qkv_format == "thd":
                                    # [2, t, np, hn] -> [2, t/2, np, hn]
1200
1201
1202
                                    kv_inputs[i % 2] = tex.thd_read_half_tensor(
                                        kv_inputs[i % 2], cu_seqlens_k, 0
                                    )
1203
1204
                                else:
                                    # [2, b, 2, sk//2, np, hn] -> [2, b, sk//2, np, hn]
1205
                                    kv_inputs[i % 2] = kv_inputs[i % 2][:, :, 0, ...].contiguous()
1206
                                # [2, b, sk//2, np, hn] -> [2, b*sk//2, np, hn]
1207
                                kv_inputs[i % 2] = kv_inputs[i % 2].view(2, -1, *k.shape[-2:])
1208
1209
                                if _flash_attn_2_3_plus:
                                    fa_optional_forward_kwargs["window_size"] = [-1, -1]
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
                                (
                                    _,
                                    _,
                                    _,
                                    _,
                                    out_per_step[i],
                                    softmax_lse_per_step[i],
                                    _,
                                    rng_states[i],
                                ) = _flash_attn_forward(
                                    q_inputs[i % 2],
                                    kv_inputs[i % 2][0],
                                    kv_inputs[i % 2][1],
                                    cu_seqlens_q,
                                    cu_seqlens_k // 2,
                                    max_seqlen_q,
                                    max_seqlen_k // 2,
                                    dropout_p,
                                    softmax_scale,
                                    causal=False,
                                    return_softmax=False,
                                    **fa_optional_forward_kwargs,
1232
1233
1234
                                )
                        else:
                            if use_fused_attention:
1235
1236
                                if qkv_format == "bshd":
                                    # [b, 2, sq//2, np, hn] -> [b, sq//2, np, hn]
1237
                                    q_inputs[i % 2] = q[:, 1, ...].contiguous()
1238
                                    # [2, b, 2, sk//2, np, hn] -> [2, b, sk, np, hn]
1239
1240
1241
                                    kv_inputs[i % 2] = kv_inputs[i % 2].view(
                                        2, k.shape[0], -1, *k.shape[-2:]
                                    )
1242
1243
                                elif qkv_format == "sbhd":
                                    # [2, sq//2, b, np, hn] -> [sq//2, b, np, hn]
1244
                                    q_inputs[i % 2] = q[1].contiguous()
1245
                                    # [2, 2, sk//2, b, np, hn] -> [2, sk, b, np, hn]
1246
                                    kv_inputs[i % 2] = kv_inputs[i % 2].view(2, -1, *k.shape[-3:])
1247
1248
                                elif qkv_format == "thd":
                                    # [t, np, hn] -> [t/2, np, hn]
1249
                                    q_inputs[i % 2] = tex.thd_read_half_tensor(q, cu_seqlens_q, 1)
1250
1251
                                if attn_bias is not None:
                                    idx = (rank - i) % cp_size
1252
1253
1254
1255
1256
1257
                                    attn_bias_inputs[i % 2] = torch.cat(
                                        (
                                            attn_bias_[..., 1, :, idx, :],
                                            attn_bias_[..., 1, :, (2 * cp_size - idx - 1), :],
                                        ),
                                        dim=-1,
1258
                                    ).contiguous()
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
                                out_per_step[i], [softmax_lse_per_step[i], rng_states[i], *rest] = (
                                    fused_attn_fwd(
                                        is_training,
                                        max_seqlen_q // 2,
                                        max_seqlen_k,
                                        cu_seqlens_q // 2,
                                        cu_seqlens_k,
                                        q_inputs[i % 2],
                                        kv_inputs[i % 2][0],
                                        kv_inputs[i % 2][1],
                                        TE_DType[q.dtype],
                                        tex.NVTE_Fused_Attn_Backend.NVTE_F16_arbitrary_seqlen,
                                        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],
1277
1278
1279
1280
                                        cu_seqlens_q_padded=(
                                            None
                                            if cu_seqlens_q_padded is None
                                            else cu_seqlens_q_padded // 2
1281
                                        ),
1282
                                        cu_seqlens_kv_padded=cu_seqlens_kv_padded,
1283
                                    )
1284
                                )
1285
1286
                                if len(rest) > 0:
                                    attn_biases[i] = rest[0]
1287
                            else:
1288
1289
                                if qkv_format == "thd":
                                    # [t, np, hn] -> [t/2, np, hn]
1290
                                    q_inputs[i % 2] = tex.thd_read_half_tensor(q, cu_seqlens_q, 1)
1291
1292
                                else:
                                    # [b, 2, sq//2, np, hn]->[b, sq//2, np, hn]->[b*sq//2, np, hn]
1293
                                    q_inputs[i % 2] = (
1294
                                        q[:, 1, ...].contiguous().view(-1, *q.shape[-2:])
1295
                                    )
1296
                                # [2, b, 2, sk//2, np, hn] -> [2, b*sk, np, hn]
1297
                                kv_inputs[i % 2] = kv_inputs[i % 2].view(2, -1, *k.shape[-2:])
1298
1299
                                if _flash_attn_2_3_plus:
                                    fa_optional_forward_kwargs["window_size"] = [-1, -1]
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
                                (
                                    _,
                                    _,
                                    _,
                                    _,
                                    out_per_step[i],
                                    softmax_lse_per_step[i],
                                    _,
                                    rng_states[i],
                                ) = _flash_attn_forward(
                                    q_inputs[i % 2],
                                    kv_inputs[i % 2][0],
                                    kv_inputs[i % 2][1],
                                    cu_seqlens_q // 2,
                                    cu_seqlens_k,
                                    max_seqlen_q // 2,
                                    max_seqlen_k,
                                    dropout_p,
                                    softmax_scale,
                                    causal=False,
                                    return_softmax=False,
                                    **fa_optional_forward_kwargs,
1322
1323
1324
                                )
                    else:
                        if use_fused_attention:
1325
1326
                            if attn_bias is not None:
                                idx = (rank - i) % cp_size
1327
1328
1329
1330
1331
1332
                                attn_bias_inputs[i % 2] = torch.cat(
                                    (
                                        attn_bias[..., idx, :],
                                        attn_bias[..., (2 * cp_size - idx - 1), :],
                                    ),
                                    dim=-1,
1333
                                ).contiguous()
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
                            out_per_step[i], [softmax_lse_per_step[i], rng_states[i], *rest] = (
                                fused_attn_fwd(
                                    is_training,
                                    max_seqlen_q,
                                    max_seqlen_k,
                                    cu_seqlens_q,
                                    cu_seqlens_k,
                                    q,
                                    kv_inputs[i % 2][0],
                                    kv_inputs[i % 2][1],
                                    TE_DType[q.dtype],
                                    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_inputs[i % 2],
1352
1353
                                    cu_seqlens_q_padded=cu_seqlens_q_padded,
                                    cu_seqlens_kv_padded=cu_seqlens_kv_padded,
1354
                                )
1355
                            )
1356
1357
                            if len(rest) > 0:
                                attn_biases[i] = rest[0]
1358
                        else:
1359
                            # [b, sq, np, hn] -> [b*sq, np, hn]
1360
                            q_inputs[i % 2] = q.view(-1, *q.shape[-2:])
1361
                            # [2, b, sk, np, hn] -> [2, b*sk, np, hn]
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
                            kv_inputs[i % 2] = kv_inputs[i % 2].view(2, -1, *k.shape[-2:])
                            (
                                _,
                                _,
                                _,
                                _,
                                out_per_step[i],
                                softmax_lse_per_step[i],
                                _,
                                rng_states[i],
                            ) = _flash_attn_forward(
                                q_inputs[i % 2],
                                kv_inputs[i % 2][0],
                                kv_inputs[i % 2][1],
                                cu_seqlens_q,
                                cu_seqlens_k,
                                max_seqlen_q,
                                max_seqlen_k,
                                dropout_p,
                                softmax_scale,
                                causal=False,
                                return_softmax=False,
                                **fa_optional_forward_kwargs,
1385
                            )
1386
1387
1388
1389

            if i > 0:
                # wait until fwd restuls correction of last step is done
                if i > 1:
1390
                    flash_attn_streams[(i - 1) % 2].wait_event(fwd_results_correction_done)
1391

1392
1393
                if use_fused_attention:
                    # [b, np, sq, 1] -> [b, np, sq]
1394
                    softmax_lse_per_step[i - 1].squeeze_(-1)
1395

1396
                with torch.cuda.stream(flash_attn_streams[(i - 1) % 2]):
1397
1398
1399
                    if i == 1:
                        out = torch.empty_like(q).zero_()
                        softmax_lse = torch.clone(softmax_lse_per_step[0]).to(torch.double)
1400
                        if causal and qkv_format != "thd":
1401
1402
                            # [b, np, sq] -> [b, np, 2, sq//2]
                            softmax_lse_ = softmax_lse.view(
1403
                                *softmax_lse.shape[:-1], 2, softmax_lse.shape[-1] // 2
1404
                            )
1405
1406
1407
1408
                    elif (i - 1) <= rank or not causal:
                        flash_attn_fwd_softmax_lse_correction(
                            softmax_lse, softmax_lse_per_step[i - 1]
                        )
1409
                    else:
1410
                        if qkv_format == "thd":
1411
1412
1413
                            tex.thd_second_half_lse_correction(
                                softmax_lse, softmax_lse_per_step[i - 1], cu_seqlens_q, q.size(0)
                            )
1414
                        else:
1415
1416
1417
                            flash_attn_fwd_softmax_lse_correction(
                                softmax_lse_[..., 1, :], softmax_lse_per_step[i - 1]
                            )
1418
1419

                if i < cp_size:
1420
                    flash_attn_streams[(i - 1) % 2].record_event(fwd_results_correction_done)
1421
1422
1423
1424

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

        softmax_lse = softmax_lse.to(torch.float)
1425
1426
        if qkv_format in ["bshd", "sbhd"]:
            seq_dim = qkv_format.index("s")
1427
        for i in range(cp_size):
1428
1429
1430
1431
1432
1433
            if qkv_format == "bshd":
                out_per_step[i] = out_per_step[i].view(out.shape[0], -1, *out.shape[-2:])
                out_ = out[:, 1, ...]
            elif qkv_format == "sbhd":
                out_per_step[i] = out_per_step[i].view(-1, *out.shape[-3:])
                out_ = out[1]
1434

1435
            if i <= rank or not causal:
1436
                if qkv_format in ["bshd", "sbhd"]:
1437
1438
1439
1440
1441
1442
1443
                    flash_attn_fwd_out_correction(
                        out.view(*out_per_step[i].shape),
                        out_per_step[i],
                        seq_dim,
                        softmax_lse,
                        softmax_lse_per_step[i],
                    )
1444
                elif qkv_format == "thd":
1445
1446
1447
1448
1449
1450
1451
1452
                    tex.thd_out_correction(
                        out,
                        out_per_step[i],
                        softmax_lse,
                        softmax_lse_per_step[i],
                        cu_seqlens_q,
                        False,
                    )
1453
1454
                else:
                    assert False, f"{qkv_format} is an unsupported qkv_format!"
1455
            else:
1456
                if qkv_format in ["bshd", "sbhd"]:
1457
1458
1459
1460
1461
1462
1463
                    flash_attn_fwd_out_correction(
                        out_,
                        out_per_step[i],
                        seq_dim,
                        softmax_lse_[..., 1, :],
                        softmax_lse_per_step[i],
                    )
1464
                elif qkv_format == "thd":
1465
1466
1467
1468
1469
1470
1471
1472
                    tex.thd_out_correction(
                        out,
                        out_per_step[i],
                        softmax_lse,
                        softmax_lse_per_step[i],
                        cu_seqlens_q,
                        True,
                    )
1473
1474
                else:
                    assert False, f"{qkv_format} is an unsupported qkv_format!"
1475
1476

        kv = p2p_comm_buffers[-1]
1477
        if use_fused_attention:
1478
1479
1480
1481
            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:])
1482
1483
        else:
            out = out.view(-1, *out.shape[-2:])
1484

1485
        ctx.save_for_backward(
1486
1487
1488
1489
1490
1491
            q,
            kv,
            out,
            softmax_lse,
            cu_seqlens_q,
            cu_seqlens_k,
1492
1493
            cu_seqlens_q_padded,
            cu_seqlens_kv_padded,
1494
1495
            *rng_states,
            *attn_biases,
1496
        )
1497
1498
1499
1500
1501
1502
        ctx.cp_group = cp_group
        ctx.cp_global_ranks = cp_global_ranks
        ctx.dropout_p = dropout_p
        ctx.max_seqlen_q = max_seqlen_q
        ctx.max_seqlen_k = max_seqlen_k
        ctx.softmax_scale = softmax_scale
1503
        ctx.qkv_format = qkv_format
1504
        ctx.attn_mask_type = attn_mask_type
1505
1506
        ctx.attn_bias_type = attn_bias_type
        ctx.attn_bias_shape = None if attn_bias is None else attn_bias.shape
1507
        ctx.deterministic = deterministic
1508
        ctx.use_fused_attention = use_fused_attention
1509
1510
1511
1512
        return out

    @staticmethod
    def backward(ctx, dout):
1513
        (q, kv, out, softmax_lse, cu_seqlens_q, cu_seqlens_k) = ctx.saved_tensors[:6]
1514
        (cu_seqlens_q_padded, cu_seqlens_kv_padded) = ctx.saved_tensors[6:8]
1515
        cp_size = get_distributed_world_size(ctx.cp_group)
1516
1517
        rng_states = ctx.saved_tensors[8 : 8 + cp_size]
        attn_biases = ctx.saved_tensors[8 + cp_size : 8 + cp_size * 2]
1518

1519
        rank = get_distributed_rank(ctx.cp_group)
1520
        send_dst = ctx.cp_global_ranks[(rank - 1) % cp_size]
1521
1522
1523
        recv_src = ctx.cp_global_ranks[(rank + 1) % cp_size]
        batch_p2p_comm = int(os.getenv("NVTE_BATCH_MHA_P2P_COMM", "0")) or (cp_size == 2)

1524
1525
        causal = "causal" in ctx.attn_mask_type
        padding = "padding" in ctx.attn_mask_type
1526
1527
        qkv_layout = ctx.qkv_format + "_" + ctx.qkv_format + "_" + ctx.qkv_format

1528
        if attn_biases[0] is not None:
1529
1530
            # [b, np, sq, 2*cp, sk//(2*cp)]
            attn_dbias = torch.zeros(
1531
                *ctx.attn_bias_shape, dtype=attn_biases[0].dtype, device=attn_biases[0].device
1532
1533
1534
            )
            # [b, np, sq, 2*cp, sk//(2*cp)] -> [b, np, 2, sq//2, 2*cp, sk//(2*cp)]
            attn_dbias_ = attn_dbias.view(
1535
                *attn_dbias.shape[:-3], 2, attn_dbias.shape[-3] // 2, *attn_dbias.shape[-2:]
1536
1537
1538
1539
            )
        else:
            attn_dbias = None

1540
        if causal:
1541
1542
1543
1544
            if ctx.qkv_format == "thd":
                softmax_lse_ = tex.thd_read_second_half_lse(softmax_lse, cu_seqlens_q, q.size(0))
            else:
                # [b, np, sq] -> [b, np, 2, sq//2]
1545
1546
1547
                softmax_lse_ = softmax_lse.view(
                    *softmax_lse.shape[:-1], 2, softmax_lse.shape[-1] // 2
                )
1548
1549
1550
1551
1552
                softmax_lse_ = softmax_lse_[..., 1, :].contiguous()
                if ctx.use_fused_attention:
                    # [b, np, sq//2] -> [b, np, sq//2, 1]
                    softmax_lse_.unsqueeze_(-1)

1553
1554
1555
        if ctx.use_fused_attention:
            # [b, np, sq] -> [b, np, sq, 1]
            softmax_lse.unsqueeze_(-1)
1556
1557
1558
1559
1560
        out = out.view(*q.shape)
        dout = dout.view(*q.shape)
        # Flash Attn outputs
        dq = torch.empty_like(q)

1561
1562
1563
1564
        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),
        ]
1565
1566
1567
        p2p_comm_buffers[0][0].copy_(kv)
        send_recv_reqs = []

1568
1569
1570
1571
1572
1573
        fa_optional_backward_kwargs = {}
        if _flash_attn_2_4_plus:
            fa_optional_backward_kwargs["alibi_slopes"] = None
        if _flash_attn_2_4_1_plus:
            fa_optional_backward_kwargs["deterministic"] = ctx.deterministic

1574
1575
1576
1577
1578
        for i in range(cp_size):
            # wait until KV is received
            for req in send_recv_reqs:
                req.wait()

1579
1580
            send_tensor = p2p_comm_buffers[i % 2]
            recv_tensor = p2p_comm_buffers[(i + 1) % 2]
1581
1582
1583
            if i == 0:
                send_tensor = send_tensor[0]
                recv_tensor = recv_tensor[0]
1584
            if i == (cp_size - 1):
1585
1586
1587
                send_tensor = send_tensor[1]
                recv_tensor = recv_tensor[1]

1588
1589
1590
            send_recv_reqs = flash_attn_p2p_communicate(
                rank, send_tensor, send_dst, recv_tensor, recv_src, ctx.cp_group, batch_p2p_comm
            )
1591

1592
            kv = p2p_comm_buffers[i % 2][0]
1593
            # In reversed order of fwd
1594
            if causal:
1595
                if i == (cp_size - 1):
1596
                    if ctx.use_fused_attention:
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
                        if ctx.qkv_format == "bshd":
                            # [b, 2, sq//2, np, hn] -> [b, sq, np, hn]
                            q_ = q.view(q.shape[0], -1, *q.shape[-2:])
                            # [2, b, 2, sk//2, np, hn] -> [2, b, sk, np, hn]
                            kv_ = kv.view(*kv.shape[0:2], -1, *kv.shape[-2:])
                            # [b, 2, sq//2, np, hn] -> [b, sq, np, hn]
                            out_ = out.view(out.shape[0], -1, *out.shape[-2:])
                            dout_ = dout.view(dout.shape[0], -1, *dout.shape[-2:])
                        elif ctx.qkv_format == "sbhd":
                            # [2, sq//2, b, np, hn] -> [sq, b, np, hn]
                            q_ = q.view(-1, *q.shape[-3:])
                            # [2, 2, sk//2, b, np, hn] -> [2, sk, b, np, hn]
                            kv_ = kv.view(kv.shape[0], -1, *kv.shape[-3:])
                            # [2, sq//2, b, np, hn] -> [sq, b, np, hn]
                            out_ = out.view(-1, *out.shape[-3:])
                            dout_ = dout.view(-1, *dout.shape[-3:])
1613
1614
                        elif ctx.qkv_format == "thd":
                            q_, kv_, out_, dout_ = q, kv, out, dout
1615
                        aux_ctx_tensors = [softmax_lse, rng_states[cp_size - i - 1]]
1616
                        if attn_dbias is not None:
1617
                            aux_ctx_tensors += [attn_biases[cp_size - i - 1]]
1618
                        dq_, dk_, dv_, dbias_ = fused_attn_bwd(
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
                            ctx.max_seqlen_q,
                            ctx.max_seqlen_k,
                            cu_seqlens_q,
                            cu_seqlens_k,
                            q_,
                            kv_[0],
                            kv_[1],
                            out_,
                            dout_,
                            TE_DType[q.dtype],
                            TE_DType[kv.dtype],
                            aux_ctx_tensors,
1631
                            tex.NVTE_Fused_Attn_Backend.NVTE_F16_arbitrary_seqlen,
1632
1633
                            cu_seqlens_q_padded=cu_seqlens_q_padded,
                            cu_seqlens_kv_padded=cu_seqlens_kv_padded,
1634
1635
                            attn_scale=ctx.softmax_scale,
                            dropout=ctx.dropout_p,
1636
                            qkv_layout=qkv_layout,
1637
                            attn_mask_type=ctx.attn_mask_type,
1638
                            attn_bias_type=ctx.attn_bias_type,
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
                        )
                    else:
                        # [b, 2, sq//2, np, hn] -> [b*sq, np, hn]
                        q_ = q.view(-1, *q.shape[-2:])
                        dq_ = torch.empty_like(q_)
                        # [2, b, 2, sk//2, np, hn] -> [2, b*sk, np, hn]
                        kv_ = kv.view(2, -1, *kv.shape[-2:])
                        dkv_ = torch.empty_like(kv_)
                        # [b, 2, sq//2, np, hn] -> [b*sq, np, hn]
                        out_ = out.view(-1, *out.shape[-2:])
                        dout_ = dout.view(-1, *dout.shape[-2:])
                        if _flash_attn_2_3_plus:
                            fa_optional_backward_kwargs["window_size"] = [-1, 0]
                        _flash_attn_backward(
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
                            dout_,
                            q_,
                            kv_[0],
                            kv_[1],
                            out_,
                            softmax_lse,
                            dq_,
                            dkv_[0],
                            dkv_[1],
                            cu_seqlens_q,
                            cu_seqlens_k,
                            ctx.max_seqlen_q,
                            ctx.max_seqlen_k,
                            ctx.dropout_p,
                            ctx.softmax_scale,
                            True,
                            rng_state=rng_states[cp_size - i - 1],
                            **fa_optional_backward_kwargs,
1671
                        )
1672
                elif i >= (cp_size - rank - 1):
1673
                    if ctx.use_fused_attention:
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
                        if ctx.qkv_format == "bshd":
                            # [b, 2, sq//2, np, hn] -> [b, sq, np, hn]
                            q_ = q.view(q.shape[0], -1, *q.shape[-2:])
                            # [2, b, 2, sk//2, np, hn] -> [2, b, sk//2, np, hn]
                            kv_ = kv[:, :, 0, ...].contiguous()
                            # [b, 2, sq//2, np, hn] -> [b, sq, np, hn]
                            out_ = out.view(out.shape[0], -1, *out.shape[-2:])
                            dout_ = dout.view(dout.shape[0], -1, *dout.shape[-2:])
                        elif ctx.qkv_format == "sbhd":
                            # [2, sq//2, b, np, hn] -> [sq, b, np, hn]
                            q_ = q.view(-1, *q.shape[-3:])
                            # [2, 2, sk//2, b, np, hn] -> [2, sk//2, b, np, hn]
                            kv_ = kv[:, 0, ...].contiguous()
                            # [2, sq//2, b, np, hn] -> [sq, b, np, hn]
                            out_ = out.view(-1, *out.shape[-3:])
                            dout_ = dout.view(-1, *dout.shape[-3:])
1690
1691
1692
1693
                        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_k, 0)
1694
                        aux_ctx_tensors = [softmax_lse, rng_states[cp_size - i - 1]]
1695
                        if attn_dbias is not None:
1696
                            aux_ctx_tensors += [attn_biases[cp_size - i - 1]]
1697
                        dq_, dk_, dv_, dbias_ = fused_attn_bwd(
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
                            ctx.max_seqlen_q,
                            ctx.max_seqlen_k // 2,
                            cu_seqlens_q,
                            cu_seqlens_k // 2,
                            q_,
                            kv_[0],
                            kv_[1],
                            out_,
                            dout_,
                            TE_DType[q.dtype],
                            TE_DType[kv.dtype],
                            aux_ctx_tensors,
1710
                            tex.NVTE_Fused_Attn_Backend.NVTE_F16_arbitrary_seqlen,
1711
1712
1713
1714
                            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
                            ),
1715
1716
                            attn_scale=ctx.softmax_scale,
                            dropout=ctx.dropout_p,
1717
                            qkv_layout=qkv_layout,
1718
                            attn_mask_type="padding" if padding else "no_mask",
1719
                            attn_bias_type=ctx.attn_bias_type,
1720
1721
1722
1723
1724
                        )
                    else:
                        # [b, 2, sq//2, np, hn] -> [b*sq, np, hn]
                        q_ = q.view(-1, *q.shape[-2:])
                        dq_ = torch.empty_like(q_)
1725
1726
1727
1728
1729
1730
                        if ctx.qkv_format == "thd":
                            # [2, t, np, hn] -> [2, t/2, np, hn]
                            kv_ = tex.thd_read_half_tensor(kv, cu_seqlens_k, 0)
                        else:
                            # [2, b, 2, sk//2, np, hn]->[2, b, sk//2, np, hn]->[2, b*sk//2, np, hn]
                            kv_ = kv[:, :, 0, ...].contiguous().view(2, -1, *kv.shape[-2:])
1731
1732
1733
1734
1735
1736
1737
                        dkv_ = torch.empty_like(kv_)
                        # [b, 2, sq//2, np, hn] -> [b*sq, np, hn]
                        out_ = out.view(-1, *out.shape[-2:])
                        dout_ = dout.view(-1, *dout.shape[-2:])
                        if _flash_attn_2_3_plus:
                            fa_optional_backward_kwargs["window_size"] = [-1, -1]
                        _flash_attn_backward(
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
                            dout_,
                            q_,
                            kv_[0],
                            kv_[1],
                            out_,
                            softmax_lse,
                            dq_,
                            dkv_[0],
                            dkv_[1],
                            cu_seqlens_q,
                            cu_seqlens_k // 2,
                            ctx.max_seqlen_q,
                            ctx.max_seqlen_k // 2,
                            ctx.dropout_p,
                            ctx.softmax_scale,
                            False,
                            rng_state=rng_states[cp_size - i - 1],
                            **fa_optional_backward_kwargs,
1756
1757
1758
                        )
                else:
                    if ctx.use_fused_attention:
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
                        if ctx.qkv_format == "bshd":
                            # [b, 2, sq//2, np, hn] -> [b, sq//2, np, hn]
                            q_ = q[:, 1, ...].contiguous()
                            # [2, b, 2, sk//2, np, hn] -> [2, b, sk, np, hn]
                            kv_ = kv.view(*kv.shape[0:2], -1, *kv.shape[-2:])
                            # [b, 2, sq//2, np, hn] -> [b, sq//2, np, hn]
                            out_ = out[:, 1, ...].contiguous()
                            dout_ = dout[:, 1, ...].contiguous()
                        elif ctx.qkv_format == "sbhd":
                            # [2, sq//2, b, np, hn] -> [sq//2, b, np, hn]
                            q_ = q[1].contiguous()
                            # [2, 2, sk//2, b, np, hn] -> [2, sk, b, np, hn]
                            kv_ = kv.view(kv.shape[0], -1, *kv.shape[-3:])
                            # [2, sq//2, b, np, hn] -> [sq//2, b, np, hn]
                            out_ = out[1].contiguous()
                            dout_ = dout[1].contiguous()
1775
1776
1777
1778
1779
1780
                        elif ctx.qkv_format == "thd":
                            # [t, np, hn] -> [t/2, np, hn]
                            q_ = tex.thd_read_half_tensor(q, cu_seqlens_q, 1)
                            out_ = tex.thd_read_half_tensor(out, cu_seqlens_q, 1)
                            dout_ = tex.thd_read_half_tensor(dout, cu_seqlens_q, 1)
                            kv_ = kv
1781
                        aux_ctx_tensors = [softmax_lse_, rng_states[cp_size - i - 1]]
1782
                        if attn_dbias is not None:
1783
                            aux_ctx_tensors += [attn_biases[cp_size - i - 1]]
1784
                        dq_, dk_, dv_, dbias_ = fused_attn_bwd(
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
                            ctx.max_seqlen_q // 2,
                            ctx.max_seqlen_k,
                            cu_seqlens_q // 2,
                            cu_seqlens_k,
                            q_,
                            kv_[0],
                            kv_[1],
                            out_,
                            dout_,
                            TE_DType[q.dtype],
                            TE_DType[kv.dtype],
                            aux_ctx_tensors,
1797
                            tex.NVTE_Fused_Attn_Backend.NVTE_F16_arbitrary_seqlen,
1798
1799
1800
1801
                            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,
1802
1803
                            attn_scale=ctx.softmax_scale,
                            dropout=ctx.dropout_p,
1804
                            qkv_layout=qkv_layout,
1805
                            attn_mask_type="padding" if padding else "no_mask",
1806
                            attn_bias_type=ctx.attn_bias_type,
1807
1808
                        )
                    else:
1809
1810
1811
1812
1813
1814
                        if ctx.qkv_format == "thd":
                            # [t, np, hn] -> [t/2, np, hn]
                            q_ = tex.thd_read_half_tensor(q, cu_seqlens_q, 1)
                        else:
                            # [b, 2, sq//2, np, hn] -> [b, sq//2, np, hn] -> [b*sq//2, np, hn]
                            q_ = q[:, 1, ...].contiguous().view(-1, *q.shape[-2:])
1815
1816
1817
1818
                        dq_ = torch.empty_like(q_)
                        # [2, b, 2, sk//2, np, hn] -> [2, b*sk, np, hn]
                        kv_ = kv.view(2, -1, *kv.shape[-2:])
                        dkv_ = torch.empty_like(kv_)
1819
1820
1821
1822
1823
1824
1825
                        if ctx.qkv_format == "thd":
                            out_ = tex.thd_read_half_tensor(out, cu_seqlens_q, 1)
                            dout_ = tex.thd_read_half_tensor(dout, cu_seqlens_q, 1)
                        else:
                            # [b, 2, sq//2, np, hn] -> [b, sq//2, np, hn] -> [b*sq//2, np, hn]
                            out_ = out[:, 1, ...].contiguous().view(-1, *out.shape[-2:])
                            dout_ = dout[:, 1, ...].contiguous().view(-1, *dout.shape[-2:])
1826
1827
1828
                        if _flash_attn_2_3_plus:
                            fa_optional_backward_kwargs["window_size"] = [-1, -1]
                        _flash_attn_backward(
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
                            dout_,
                            q_,
                            kv_[0],
                            kv_[1],
                            out_,
                            softmax_lse_,
                            dq_,
                            dkv_[0],
                            dkv_[1],
                            cu_seqlens_q // 2,
                            cu_seqlens_k,
                            ctx.max_seqlen_q // 2,
                            ctx.max_seqlen_k,
                            ctx.dropout_p,
                            ctx.softmax_scale,
                            False,
                            rng_state=rng_states[cp_size - i - 1],
                            **fa_optional_backward_kwargs,
1847
1848
1849
                        )
            else:
                if ctx.use_fused_attention:
1850
                    aux_ctx_tensors = [softmax_lse, rng_states[cp_size - i - 1]]
1851
                    if attn_dbias is not None:
1852
                        aux_ctx_tensors += [attn_biases[cp_size - i - 1]]
1853
                    dq_, dk_, dv_, dbias_ = fused_attn_bwd(
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
                        ctx.max_seqlen_q,
                        ctx.max_seqlen_k,
                        cu_seqlens_q,
                        cu_seqlens_k,
                        q,
                        kv[0],
                        kv[1],
                        out,
                        dout,
                        TE_DType[q.dtype],
                        TE_DType[kv.dtype],
                        aux_ctx_tensors,
1866
                        tex.NVTE_Fused_Attn_Backend.NVTE_F16_arbitrary_seqlen,
1867
1868
                        cu_seqlens_q_padded=cu_seqlens_q_padded,
                        cu_seqlens_kv_padded=cu_seqlens_kv_padded,
1869
1870
                        attn_scale=ctx.softmax_scale,
                        dropout=ctx.dropout_p,
1871
                        qkv_layout=qkv_layout,
1872
                        attn_mask_type=ctx.attn_mask_type,
1873
                        attn_bias_type=ctx.attn_bias_type,
1874
1875
1876
                    )
                else:
                    # [b, sq, np, hn] -> [b*sq, np, hn]
1877
1878
                    q_ = q.view(-1, *q.shape[-2:])
                    dq_ = torch.empty_like(q_)
1879
                    # [2, b, sk, np, hn] -> [2, b*sk, np, hn]
1880
1881
                    kv_ = kv.view(2, -1, *kv.shape[-2:])
                    dkv_ = torch.empty_like(kv_)
1882
                    # [b, sq, np, hn] -> [b*sq, np, hn]
1883
1884
                    out_ = out.view(-1, *out.shape[-2:])
                    dout_ = dout.view(-1, *dout.shape[-2:])
1885
1886
                    if _flash_attn_2_3_plus:
                        fa_optional_backward_kwargs["window_size"] = [-1, -1]
1887
                    _flash_attn_backward(
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
                        dout_,
                        q_,
                        kv_[0],
                        kv_[1],
                        out_,
                        softmax_lse,
                        dq_,
                        dkv_[0],
                        dkv_[1],
                        cu_seqlens_q,
                        cu_seqlens_k,
                        ctx.max_seqlen_q,
                        ctx.max_seqlen_k,
                        ctx.dropout_p,
                        ctx.softmax_scale,
                        False,
                        **fa_optional_backward_kwargs,
1905
1906
                    )

1907
            if i >= (cp_size - rank - 1) or not causal:
1908
1909
1910
1911
                # [b*sq, np, hn] -> [b, 2, sq//2, np, hn] if causal
                # [b*sq, np, hn] -> [b, sq, np, hn] if not causal
                dq_ = dq_.view(*dq.shape)
            else:
1912
1913
1914
1915
1916
1917
                if ctx.qkv_format == "bshd":
                    # [b*sq//2, np, hn] -> [b, sq//2, np, hn]
                    dq_ = dq_.view(dq.shape[0], *dq.shape[2:])
                elif ctx.qkv_format == "sbhd":
                    # [b*sq//2, np, hn] -> [sq//2, b, np, hn]
                    dq_ = dq_.view(-1, *dq.shape[-3:])
1918

1919
            if causal:
1920
                if i > (cp_size - rank - 1):
1921
                    dq.add_(dq_)
1922
1923
                elif i == (cp_size - rank - 1):
                    if rank == (cp_size - 1):
1924
1925
                        dq.copy_(dq_)
                    else:
1926
1927
1928
1929
1930
1931
                        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])
1932
1933
                        elif ctx.qkv_format == "thd":
                            tex.thd_grad_correction(dq, dq_, cu_seqlens_q, "copy", "add")
1934
                elif i > 0:
1935
1936
1937
1938
                    if ctx.qkv_format == "bshd":
                        dq[:, 1, ...].add_(dq_)
                    elif ctx.qkv_format == "sbhd":
                        dq[1].add_(dq_)
1939
1940
                    elif ctx.qkv_format == "thd":
                        tex.thd_grad_correction(dq, dq_, cu_seqlens_q, "none", "add")
1941
                else:
1942
1943
1944
1945
                    if ctx.qkv_format == "bshd":
                        dq[:, 1, ...].copy_(dq_)
                    elif ctx.qkv_format == "sbhd":
                        dq[1].copy_(dq_)
1946
1947
                    elif ctx.qkv_format == "thd":
                        tex.thd_grad_correction(dq, dq_, cu_seqlens_q, "none", "copy")
1948
1949
1950
1951
1952
            else:
                if i == 0:
                    dq.copy_(dq_)
                else:
                    dq.add_(dq_)
1953

1954
            if attn_dbias is not None:
1955
                idx = (rank + i + 1) % cp_size
1956
                if i == (cp_size - 1) or not causal:
1957
                    # [b, np, sq, sk//cp] -> [b, np, sq, 2, sk//(2*cp)]
1958
                    dbias_ = dbias_.view(*dbias_.shape[:-1], 2, dbias_.shape[-1] // 2)
1959
                    attn_dbias[..., idx, :].copy_(dbias_[..., 0, :])
1960
1961
                    attn_dbias[..., (2 * cp_size - idx - 1), :].copy_(dbias_[..., 1, :])
                elif i >= (cp_size - rank - 1):
1962
1963
1964
1965
                    # [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)]
1966
                    dbias_ = dbias_.view(*dbias_.shape[:-1], 2, dbias_.shape[-1] // 2)
1967
                    attn_dbias_[..., 1, :, idx, :].copy_(dbias_[..., 0, :])
1968
                    attn_dbias_[..., 1, :, (2 * cp_size - idx - 1), :].copy_(dbias_[..., 1, :])
1969

1970
1971
1972
            # wait until dKV is received
            for req in send_recv_reqs:
                req.wait()
1973

1974
            dkv = p2p_comm_buffers[(i + 1) % 2][1]
1975
1976
            if ctx.use_fused_attention:
                dkv_ = torch.cat((dk_.unsqueeze(0), dv_.unsqueeze(0)), dim=0)
1977
            if causal and i >= (cp_size - rank - 1) and i != (cp_size - 1):
1978
1979
1980
1981
1982
1983
                if ctx.qkv_format == "bshd":
                    # [2, b*sk//2, np, hn] -> [2, b, sk//2, np, hn]
                    dkv_ = dkv_.view(*dkv.shape[0:2], *dkv.shape[3:])
                elif ctx.qkv_format == "sbhd":
                    # [2, b*sk//2, np, hn] -> [2, sk//2, b, np, hn]
                    dkv_ = dkv_.view(dkv.shape[0], -1, *dkv.shape[-3:])
1984
1985
1986
1987
            else:
                # [2, b*sk, np, hn] -> [2, b, 2, sk//2, np, hn] if causal
                # [2, b*sk, np, hn] -> [2, b, sk, np, hn] if not causal
                dkv_ = dkv_.view(*dkv.shape)
1988

1989
            if causal:
1990
                if i == (cp_size - 1):
1991
                    if rank == 0:
1992
1993
1994
1995
1996
1997
                        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, ...])
1998
1999
                        elif ctx.qkv_format == "thd":
                            tex.thd_grad_correction(dkv, dkv_, cu_seqlens_k, "add", "copy")
2000
2001
                    else:
                        dkv.add_(dkv_)
2002
2003
                elif i >= (cp_size - rank - 1):
                    if i == 0 and rank == (cp_size - 1):
2004
2005
2006
2007
                        if ctx.qkv_format == "bshd":
                            dkv[:, :, 0, ...].copy_(dkv_)
                        elif ctx.qkv_format == "sbhd":
                            dkv[:, 0, ...].copy_(dkv_)
2008
2009
                        elif ctx.qkv_format == "thd":
                            tex.thd_grad_correction(dkv, dkv_, cu_seqlens_k, "copy", "none")
2010
                    else:
2011
2012
2013
2014
                        if ctx.qkv_format == "bshd":
                            dkv[:, :, 0, ...].add_(dkv_)
                        elif ctx.qkv_format == "sbhd":
                            dkv[:, 0, ...].add_(dkv_)
2015
2016
                        elif ctx.qkv_format == "thd":
                            tex.thd_grad_correction(dkv, dkv_, cu_seqlens_k, "add", "none")
2017
2018
2019
2020
2021
                elif i > 0:
                    dkv.add_(dkv_)
                else:
                    dkv.copy_(dkv_)
            else:
2022
2023
2024
2025
2026
                if i == 0:
                    dkv.copy_(dkv_)
                else:
                    dkv.add_(dkv_)

2027
        if causal:
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
            if ctx.qkv_format == "bshd":
                # [b, 2, sq//2, np, hn] -> [b, sq, np, hn]
                dq = dq.view(q.shape[0], -1, *q.shape[-2:])
                # [2, b, 2, sk//2, np, hn] -> [2, b, sk, np, hn]
                dkv = dkv.view(*kv.shape[0:2], -1, *kv.shape[-2:])
            elif ctx.qkv_format == "sbhd":
                # [2, sq//2, b, np, hn] -> [sq, b, np, hn]
                dq = dq.view(-1, *q.shape[-3:])
                # [2, 2, sk//2, b, np, hn] -> [2, sk, b, np, hn]
                dkv = dkv.view(kv.shape[0], -1, *kv.shape[-3:])

        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)

2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
        return (
            None,
            dq,
            dkv[0],
            dkv[1],
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            attn_dbias,
            None,
            None,
        )
2066
2067
2068


def attn_forward_func_with_cp(
2069
2070
2071
2072
2073
2074
2075
2076
    is_training,
    q,
    k,
    v,
    cu_seqlens_q,
    cu_seqlens_k,
    max_seqlen_q,
    max_seqlen_k,
2077
2078
    cu_seqlens_q_padded,
    cu_seqlens_kv_padded,
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
    dropout_p,
    cp_group,
    cp_global_ranks,
    cp_stream,
    softmax_scale=None,
    qkv_format="bshd",
    attn_mask_type="causal",
    attn_bias_type="no_bias",
    attn_bias=None,
    deterministic=False,
    use_fused_attention=False,
2090
2091
) -> torch.Tensor:
    """Attention implementation with context parallelism"""
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
    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 (
        qkv_format != "thd"
        or not use_fused_attention
        or attn_mask_type in ["padding", "padding_causal"]
    ), (
        f"Context parallelism is not supported for {attn_mask_type} mask type and "
        f"{qkv_format} format with {'FusedAttention' if use_fused_attention else 'FlashAttention'}!"
    )
    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!"""
    )
2112
    out = AttnFuncWithCP.apply(
2113
2114
2115
2116
2117
2118
2119
2120
        is_training,
        q,
        k,
        v,
        cu_seqlens_q,
        cu_seqlens_k,
        max_seqlen_q,
        max_seqlen_k,
2121
2122
        cu_seqlens_q_padded,
        cu_seqlens_kv_padded,
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
        dropout_p,
        cp_group,
        cp_global_ranks,
        cp_stream,
        softmax_scale,
        qkv_format,
        attn_mask_type,
        attn_bias_type,
        attn_bias,
        deterministic,
        use_fused_attention,
2134
2135
2136
2137
    )
    return out


2138
2139
2140
2141
class RotaryPositionEmbedding(torch.nn.Module):
    """
    Implements Rotary Position Embedding from https://arxiv.org/abs/2104.09864.
    """
2142

2143
2144
2145
    def __init__(
        self,
        dim: int,
2146
        rotary_percent: float = 1.0,
2147
2148
2149
2150
2151
2152
2153
2154
        seq_len_interpolation_factor: Optional[int] = None,
        pretrained_max_position_embeddings: Optional[int] = None,
    ):
        """
        Parameters
        ----------
        dim: int
            rotary embedding dimension
2155
2156
        rotary_percent: float
            Percent of rotary dimension to use for rotary position embeddings.
2157
2158
2159
2160
2161
2162
2163
        seq_len_interpolation_factor: int
            if not None, discrete positions will be interpolated by this factor via the trick in
            https://arxiv.org/abs/2306.15595
        pretrained_max_position_embeddings: int
            pre-trained max_position_embeddings before position interpolation
        """
        super().__init__()
2164
2165
        if rotary_percent < 1.0:
            dim = int(dim * rotary_percent)
2166
        self.seq_len_interpolation_factor = seq_len_interpolation_factor
2167
2168
2169
2170
2171
2172
2173
        inv_freq = 1.0 / (
            10000
            ** (
                torch.arange(0, dim, 2, dtype=torch.float32, device=torch.cuda.current_device())
                / dim
            )
        )
2174
        self.register_buffer("inv_freq", inv_freq)
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
        self.pretrained_max_position_embeddings = pretrained_max_position_embeddings

    def forward(self, max_seq_len: int, offset: int = 0):
        """
        Create rotary position embedding frequencies

        Parameters
        ----------
        max_seq_len: int
            sequence length of a sample
        offset: int, default = 0
            fixed offset for freqencies
        """
2188
2189
2190
2191
        seq = (
            torch.arange(max_seq_len, device=self.inv_freq.device, dtype=self.inv_freq.dtype)
            + offset
        )
2192

2193
2194
2195
2196
2197
2198
2199
2200
        if (
            self.pretrained_max_position_embeddings is not None
            and self.seq_len_interpolation_factor is not None
        ):
            if (
                max_seq_len
                > self.pretrained_max_position_embeddings * self.seq_len_interpolation_factor
            ):
2201
2202
2203
2204
2205
2206
                # dynamic linear scaling (length > position we have learned)
                seq *= 1 / (max_seq_len / self.pretrained_max_position_embeddings)
            else:
                # fixed linear scaling
                seq *= 1 / self.seq_len_interpolation_factor

2207
        freqs = torch.einsum("i , j -> i j", seq, self.inv_freq)
2208
2209
2210
2211
2212
2213
        # first part even vector components, second part odd vector components,
        #  2 * dim in dimension size
        emb = torch.cat((freqs, freqs), dim=-1)
        # emb [seq_length, .., dim]
        return emb.reshape(emb.size(0), 1, 1, emb.size(1))

2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231

class FusedRoPEFunc(torch.autograd.Function):
    """
    Function for FusedRoPE

    This implementation assumes the input tensor to be in `sbhd`, `bshd` or `thd` format and
    the RoPE tensor to be of shape (s, 1, 1, d). It accepts arbitrary memory layouts to avoid
    the expensive `.contiguous()` calls, thus it may not achieve the best memory access pattern.
    """

    @staticmethod
    def forward(
        ctx,
        t: torch.Tensor,
        freqs: torch.Tensor,
        tensor_format: str = "sbhd",
        cu_seqlens: Union[torch.Tensor, None] = None,
    ) -> torch.Tensor:
2232
2233
        if freqs.dtype != torch.float32:
            freqs = freqs.float()
2234
2235
2236
        if tensor_format == "sbhd":
            output = tex.fused_rope_forward(t, freqs, False)
        elif tensor_format == "bshd":
2237
            output = tex.fused_rope_forward(t.transpose(0, 1), freqs, True).transpose(0, 1)
2238
2239
2240
2241
2242
2243
2244
2245
2246
2247
        elif tensor_format == "thd":
            output = tex.fused_rope_thd_forward(t, cu_seqlens, freqs)
        else:
            raise ValueError(f"Unsupported tensor_format: {tensor_format}.")
        ctx.save_for_backward(freqs, cu_seqlens)
        ctx.tensor_format = tensor_format

        return output

    @staticmethod
2248
    def backward(ctx, grad_output: torch.Tensor) -> Tuple[Union[torch.Tensor, None], ...]:
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261
2262
2263
        freqs, cu_seqlens = ctx.saved_tensors
        if ctx.tensor_format == "sbhd":
            grad_input = tex.fused_rope_backward(grad_output, freqs, False)
        elif ctx.tensor_format == "bshd":
            grad_input = tex.fused_rope_backward(
                grad_output.transpose(0, 1), freqs, True
            ).transpose(0, 1)
        elif ctx.tensor_format == "thd":
            grad_input = tex.fused_rope_thd_backward(grad_output, cu_seqlens, freqs)
        else:
            raise ValueError(f"Unsupported tensor_format: {ctx.tensor_format}.")

        return grad_input, None, None, None, None


2264
2265
2266
2267
2268
2269
2270
2271
2272
def _rotate_half(x: torch.Tensor) -> torch.Tensor:
    """
    change sign so the last dimension becomes [-odd, +even]
    """
    x = x.view(x.shape[:-1] + torch.Size((2, x.shape[-1] // 2)))
    x1, x2 = x.unbind(dim=-2)
    return torch.cat((-x2, x1), dim=-1)


2273
def apply_rotary_pos_emb(
2274
2275
2276
2277
2278
2279
    t: torch.Tensor,
    freqs: torch.Tensor,
    tensor_format: str = "sbhd",
    fused: bool = False,
    cu_seqlens: Union[torch.Tensor, None] = None,
) -> torch.Tensor:
2280
    """
2281
    Apply rotary positional embedding tensor to the input tensor.
2282

2283
2284
2285
    Parameters
    ----------
    t: torch.Tensor
2286
        Input tensor of shape `[s, b, h, d]`, `[b, s, h, d]` or `[t, h, d]`, on which
2287
2288
2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
        rotary positional embedding will be applied.
    freqs: torch.Tensor
        Rotary positional embedding tensor of shape `[s2, 1, 1, d2]` and dtype 'float',
        with `s2 >= s` and `d2 <= d`.
    fused: bool, default = False
        Whether to use a fused applying RoPE implementation.
    tensor_format: {'sbhd', 'bshd', 'thd'}, default = 'sbhd'
        is `bshd` if `t` is of shape `[bs, seq, ...]`, or `sbhd` if `t` is
        of shape `[seq, bs, ...]`. 'thd' is only supported when `fused` is True.
    cu_seqlens: torch.Tensor, default = None.
        Cumulative sum of sequence lengths in a batch for `t`, with shape [b + 1] and
        dtype torch.int32. Only valid when `tensor_format` is 'thd'.
2299
    """
2300
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310
    if fused:
        assert (
            tensor_format != "thd" or cu_seqlens is not None
        ), "cu_seqlens must not be None when tensor_format is 'thd'."
        return FusedRoPEFunc.apply(t, freqs, tensor_format, cu_seqlens)

    assert tensor_format in ("sbhd", "bshd"), (
        "Only formats `sbhd` or `bshd` are supported for input tensor `t` "
        f"when fused is False, got {tensor_format}."
    )

2311
2312
2313
2314
2315
    max_seq_len = freqs.shape[0]
    cur_seq_len = t.shape[1] if tensor_format == "bshd" else t.shape[0]

    # Only apply the rotary embeddings up to the sequence length of the running
    # input.
2316
2317
2318
    assert (
        cur_seq_len <= max_seq_len
    ), f"Rotary Embeddings only supported up to {max_seq_len} sequence length!"
2319
    freqs = freqs[:cur_seq_len]
2320
    if tensor_format == "bshd":
2321
2322
2323
2324
        freqs = freqs.transpose(0, 1)  # [seq, 1, 1, dim] -> [1, seq, 1, dim]
    # cos/sin first then dtype conversion for better precision
    cos_ = torch.cos(freqs).to(t.dtype)
    sin_ = torch.sin(freqs).to(t.dtype)
2325

2326
2327
2328
2329
2330
2331
    rot_dim = freqs.shape[-1]
    # ideally t_pass is empty so rotary pos embedding is applied to all tensor t
    t, t_pass = t[..., :rot_dim], t[..., rot_dim:]

    # first part is cosine component
    # second part is sine component, need to change signs with _rotate_half method
2332
    t = (t * cos_) + (_rotate_half(t) * sin_)
2333
2334
2335
    return torch.cat((t, t_pass), dim=-1)


cyanguwa's avatar
cyanguwa committed
2336
class _SplitAlongDim(torch.autograd.Function):
2337
2338
2339
    """"""

    @staticmethod
2340
2341
2342
2343
2344
    def forward(
        ctx,
        mixed_x_layer: torch.Tensor,
        split_dim: int,
        split_size_or_sections: Union[int, List[int], Tuple[int]],
2345
    ) -> Tuple[torch.Tensor, ...]:
cyanguwa's avatar
cyanguwa committed
2346
2347
        ctx.split_dim = split_dim
        ctx.split_size_or_sections = split_size_or_sections
2348
        if isinstance(mixed_x_layer, Float8Tensor):
2349
2350
2351
2352
2353
2354
            return tuple(
                Float8Tensor.make_like(
                    mixed_x_layer,
                    data=x,
                )
                for x in torch.split(
2355
2356
                    mixed_x_layer._data,
                    split_size_or_sections=split_size_or_sections,
2357
2358
2359
2360
                    dim=split_dim,
                )
            )
        return torch.split(mixed_x_layer, split_size_or_sections, dim=split_dim)
2361
2362

    @staticmethod
2363
    def backward(ctx, *grad_outputs):
2364
2365
        assert len(grad_outputs) > 0, "No gradients received for backprop!"

cyanguwa's avatar
cyanguwa committed
2366
2367
        if isinstance(ctx.split_size_or_sections, (list, tuple)):
            split_sizes = ctx.split_size_or_sections
2368
2369
2370
            assert len(grad_outputs) == len(
                split_sizes
            ), "Unequal number of gradients vs split sections for backprop!"
cyanguwa's avatar
cyanguwa committed
2371
2372
2373
2374
2375
        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

2376
2377
2378
2379
2380
2381
2382
2383
        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]
2384
2385
2386
2387
2388
2389
2390
                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
                ):
2391
2392
2393
                    noop_ok = False
                    break
            if noop_ok:
2394
2395
2396
                ret = torch.Tensor().to(
                    device=grad_outputs[0].device, dtype=grad_outputs[0]._data.dtype
                )
2397
2398
                new_shape = list(shape)
                new_shape[split_dim] = sum(split_sizes)
2399
2400
2401
2402
2403
                ret.set_(
                    grad_outputs[0]._data.untyped_storage(),
                    grad_outputs[0]._data.storage_offset(),
                    new_shape,
                    strides,
2404
2405
2406
2407
                )
                return Float8Tensor.make_like(grad_outputs[0], data=ret), None, None

            grad_outputs_data = [x._data for x in grad_outputs]
2408
2409
2410
2411
2412
2413
2414
            return (
                Float8Tensor.make_like(
                    grad_outputs[0], data=torch.cat(grad_outputs_data, dim=split_dim)
                ),
                None,
                None,
            )
2415
2416
        noop_ok = True
        strides = grad_outputs[0].stride()
2417
        data_ptr = grad_outputs[0].untyped_storage().data_ptr()
cyanguwa's avatar
cyanguwa committed
2418
        shape = list(grad_outputs[0].shape)
2419
        for i, tensor in enumerate(grad_outputs):
cyanguwa's avatar
cyanguwa committed
2420
2421
            shape_i = shape
            shape_i[split_dim] = split_sizes[i]
2422
2423
2424
2425
2426
2427
2428
            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
            ):
2429
2430
2431
                noop_ok = False
                break
        if noop_ok:
2432
            ret = torch.Tensor().to(device=grad_outputs[0].device, dtype=grad_outputs[0].dtype)
2433
            new_shape = list(shape)
cyanguwa's avatar
cyanguwa committed
2434
            new_shape[split_dim] = sum(split_sizes)
2435
2436
2437
2438
2439
            ret.set_(
                grad_outputs[0].untyped_storage(),
                grad_outputs[0].storage_offset(),
                new_shape,
                strides,
2440
            )
cyanguwa's avatar
cyanguwa committed
2441
            return ret, None, None
2442

2443
        return torch.cat(grad_outputs, dim=split_dim), None, None
2444
2445
2446
2447
2448
2449
2450
2451
2452


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

    def __init__(
        self,
2453
        softmax_scale: float,
2454
2455
2456
2457
2458
2459
        attention_dropout: float = 0.0,
        attention_dropout_ctx: Optional[Callable] = nullcontext,
        layer_number: Optional[int] = None,
    ) -> None:
        super().__init__()

2460
        self.softmax_scale = softmax_scale
2461
2462
2463
        self.attention_dropout_ctx = attention_dropout_ctx
        self.layer_number = layer_number

2464
        self.scale_mask_softmax = FusedScaleMaskSoftmax(attention_mask_func)
2465
2466
2467
2468
2469
2470

        # 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)

2471
2472
        # An FP16 training trick required for certain GPT-like models.
        self.apply_qk_layer_scaling = (
2473
2474
            bool(int(os.getenv("NVTE_APPLY_QK_LAYER_SCALING", "0"))) and layer_number is not None
        )
2475

2476
2477
2478
2479
2480
    def forward(
        self,
        query_layer: torch.Tensor,
        key_layer: torch.Tensor,
        value_layer: torch.Tensor,
2481
        qkv_layout: str = "sbh3d",
2482
2483
        cu_seqlens_q: Optional[torch.Tensor] = None,  # pylint: disable=unused-argument
        cu_seqlens_kv: Optional[torch.Tensor] = None,  # pylint: disable=unused-argument
2484
        attn_mask_type: str = "causal",
2485
        attention_mask: Optional[Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]] = None,
2486
2487
        core_attention_bias_type: str = "no_bias",
        core_attention_bias: Optional[torch.Tensor] = None,
2488
        alibi_slopes: Optional[torch.Tensor] = None,
2489
    ) -> torch.Tensor:
2490
        """Unfused attention fprop"""
2491
2492
2493
2494
2495
        assert (
            qkv_layout in QKVLayouts
        ), f"UnfusedDotProductAttention does not support qkv_layout = {qkv_layout}!"
        qkv_format = "".join([i for i in qkv_layout.split("_")[0] if i.isalpha()])
        if qkv_format == "bshd":
2496
            # convert to sbhd and use sbhd implementation for now
2497
2498
2499
            query_layer, key_layer, value_layer = [
                x.transpose(0, 1) for x in [query_layer, key_layer, value_layer]
            ]
2500

2501
        batch_size, seqlen = query_layer.shape[1], query_layer.shape[0]
2502
        apply_qk_layer_scaling = self.apply_qk_layer_scaling and key_layer.dtype == torch.float16
2503
2504
2505
2506
2507
2508
2509
2510
2511

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

2512
        if key_layer.shape[2] != query_layer.shape[2]:
2513
2514
2515
            assert (
                query_layer.shape[2] % key_layer.shape[2] == 0
            ), "The number of attention heads must be divisible by the number of GQA groups!"
2516
            key_layer = key_layer.repeat_interleave(
2517
2518
                int(query_layer.shape[2] / key_layer.shape[2]), dim=2
            )
2519
            value_layer = value_layer.repeat_interleave(
2520
2521
                int(query_layer.shape[2] / value_layer.shape[2]), dim=2
            )
2522

2523
        # [sq, b, np, hn] -> [sq, b * np, hn]
2524
        query_layer = query_layer.reshape(output_size[2], output_size[0] * output_size[1], -1)
2525
2526
2527
2528
        # [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]
2529
2530
        # WAR to set dtype to FP32 as ONNX lacks BF16 support for ConstantOfShape operator
        is_bf16 = query_layer.dtype == torch.bfloat16
2531
2532
2533
2534
        matmul_result = torch.empty(
            output_size[0] * output_size[1],
            output_size[2],
            output_size[3],
2535
            dtype=torch.float32 if is_in_onnx_export_mode() and is_bf16 else query_layer.dtype,
2536
2537
2538
            device=torch.cuda.current_device(),
        )

2539
2540
2541
        if is_in_onnx_export_mode() and is_bf16:
            matmul_result = matmul_result.bfloat16()

2542
        scale = self.softmax_scale
2543
        if apply_qk_layer_scaling:
2544
            scale /= self.layer_number
2545
2546

        # Raw attention scores. [b * np, sq, sk]
2547
2548
2549
2550
2551
2552
        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,
2553
                alpha=scale,
2554
2555
2556
2557
2558
2559
2560
2561
            )

        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]
            )
2562
2563
2564
2565
            matmul_result = (
                matmul_result.view(output_size[0], output_size[1], output_size[2], output_size[3])
                + core_attention_bias
            ).view(-1, output_size[2], output_size[3])
2566
            matmul_result *= scale
2567

2568
2569
2570
2571
        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":
2572
                _, core_attention_bias = get_alibi(
2573
2574
                    output_size[1], output_size[2], output_size[3], alibi_slopes=alibi_slopes
                )
2575
2576
2577
2578
2579
            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,
2580
                alpha=scale,
2581
            )
2582
2583
2584
2585
2586
2587
2588
2589
2590
2591
            matmul_result = (
                (
                    matmul_result.view(
                        output_size[0], output_size[1], output_size[2], output_size[3]
                    )
                    + core_attention_bias
                )
                .view(-1, output_size[2], output_size[3])
                .to(dtype=query_layer.dtype)
            )
2592
2593
2594
2595
2596
2597

        # change view to [b, np, sq, sk]
        attention_scores = matmul_result.view(*output_size)

        # attention scores and attention mask [b, np, sq, sk]
        softmax_scale = self.layer_number if apply_qk_layer_scaling else None
2598
        attention_probs = self.scale_mask_softmax(
2599
2600
            attention_scores, attention_mask, attn_mask_type, softmax_scale
        )
2601
2602
2603
2604
2605
2606
2607
2608
2609
2610
2611
2612
2613
2614
2615
2616

        # 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]
2617
        value_layer = value_layer.reshape(value_layer.size(0), output_size[0] * output_size[1], -1)
2618
2619

        # change view [b * np, sq, sk]
2620
        attention_probs = attention_probs.view(output_size[0] * output_size[1], output_size[2], -1)
2621
2622
2623
2624
2625
2626
2627

        # 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)

2628
        if qkv_format == "sbhd":
2629
2630
            # [b, np, sq, hn] --> [sq, b, np, hn]
            context_layer = context_layer.permute(2, 0, 1, 3).contiguous()
2631

2632
2633
2634
            # [sq, b, np, hn] --> [sq, b, hp]
            context_layer = context_layer.view(seqlen, batch_size, -1)

2635
        if qkv_format == "bshd":
2636
2637
2638
2639
2640
            # [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)
2641
2642
2643
2644
2645
2646

        return context_layer


class _PrepareQKVForFA(torch.autograd.Function):
    """This class converts QKV from interleaved (s, b, ...) layout
2647
    to separate contiguous q, k, v tensors in (b, s, ...) layout."""
2648
2649

    @staticmethod
2650
2651
2652
2653
    def forward(
        _ctx: torch.autograd.function.FunctionCtx,  # unused
        query_layer: torch.Tensor,
        key_layer: torch.Tensor,
2654
        value_layer: torch.Tensor,
2655
    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
2656
2657
2658
2659
2660
2661
2662
2663
2664
2665
2666
        # 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
2667
2668
2669
2670
    def backward(
        _ctx: torch.autograd.function.FunctionCtx,  # unused
        dq: torch.Tensor,
        dk: torch.Tensor,
2671
        dv: torch.Tensor,
2672
2673
2674
2675
2676
    ) -> Tuple[Union[torch.Tensor, None], ...]:
        dqkv = tex.fa_prepare_bwd(dq, dk, dv)
        dq, dk, dv = split_tensor_along_dim(dqkv, -1, 3)
        return dq, dk, dv

2677

2678
def get_qkv_layout(
2679
2680
2681
2682
2683
    q: torch.Tensor,
    k: torch.Tensor,
    v: torch.Tensor,
    qkv_format: str = "sbhd",
) -> str:
2684
    """Get qkv layout.
2685

2686
2687
2688
2689
2690
2691
2692
2693
2694
2695
2696
2697
2698
2699
2700
2701
2702
2703
2704
2705
2706
2707
2708
2709
2710
2711
2712
2713
    Parameters
    ----------
    q: torch.Tensor
        Query tensor.
    k: torch.Tensor
        Key tensor.
    v: torch.Tensor
        Value tensor.
    qkv_format: str, default = `sbhd`
        Dimension format for `q`, `k` and `v`, {`sbhd`, `bshd`, `thd`}. `s` stands for
        the sequence length dimension, `b` batch size, `h` the number of attention heads,
        `d` head size, and `t` the total number of sequences in a batch, i.e.
        `t = sum(s_i) for i = 0...b-1`.

    Returns
    ----------
    qkv_layout: str
       Memory layout of `q`, `k` and `v`. Each `qkv_format` can be mapped to one of five
       memory layouts. For example, `sb3hd` means `q`, `k`, `v` are created as one chunk
       of memory and that they are interleaved in the `2`nd dimension. `sbhd_sbh2d` means
       `q` and `kv` are created in two chunks and that `q` itself is contiguous and `k`, `v`
       are interleaved with each other in the `3`rd dimension, `k = kv[:,:,:,0,:]` and
       `v = kv[:,:,:,1,:]`.
       Mapping:
       `sbhd`: {`sb3hd`, `sbh3d`, `sbhd_sb2hd`, `sbhd_sbh2d`, `sbhd_sbhd_sbhd`}
       `bshd`: {`bs3hd`, `bsh3d`, `bshd_bs2hd`, `bshd_bsh2d`, `bshd_bshd_bshd`}
       `thd` : {`t3hd`, `th3d`, `thd_t2hd`, `thd_th2d`, `thd_thd_thd`}
    """
2714

2715
2716
    check_last_dim_contiguous = all(x.stride(-1) == 1 for x in [q, k, v])
    assert check_last_dim_contiguous, "q, k and v must have stride 1 in their last dimension!"
2717

2718
2719
2720
2721
2722
2723
2724
2725
2726
2727
2728
2729
2730
2731
2732
2733
2734
    def run_iteratively(q, k, v):
        data_ptr = q.untyped_storage().data_ptr()
        check_ptrs_qkv = all(x.untyped_storage().data_ptr() == data_ptr for x in [q, k, v])
        data_ptr = k.untyped_storage().data_ptr()
        check_ptrs_kv = all(x.untyped_storage().data_ptr() == data_ptr for x in [k, v])

        stride = q.stride()
        check_strides_qkv = all(stride == x.stride() for x in [q, k, v])
        stride = k.stride()
        check_strides_kv = all(stride == x.stride() for x in [k, v])

        shape = q.shape
        check_shapes_qkv = all(shape == x.shape for x in [q, k, v])
        shape = k.shape
        check_shapes_kv = all(shape == x.shape for x in [k, v])

        last_dim_size = q.shape[-1]
2735
2736
2737
        check_last_dim_offsets_qkv = all(
            i * last_dim_size == x.storage_offset() for i, x in enumerate([q, k, v])
        )
2738
        last_dim_size = k.shape[-1]
2739
2740
2741
        check_last_dim_offsets_kv = all(
            i * last_dim_size == x.storage_offset() for i, x in enumerate([k, v])
        )
2742
2743

        last_two_dims_size = q.shape[-1] * q.shape[-2]
2744
2745
2746
        check_last_two_dims_offsets_qkv = all(
            i * last_two_dims_size == x.storage_offset() for i, x in enumerate([q, k, v])
        )
2747
        last_two_dims_size = k.shape[-1] * k.shape[-2]
2748
2749
2750
        check_last_two_dims_offsets_kv = all(
            i * last_two_dims_size == x.storage_offset() for i, x in enumerate([k, v])
        )
2751

2752
2753
2754
2755
        if (
            check_ptrs_qkv
            and check_strides_qkv
            and check_shapes_qkv
2756
            and check_last_two_dims_offsets_qkv
2757
2758
            and not check_last_dim_offsets_qkv
        ):
2759
            # sb3hd, bs3hd, t3hd
2760
2761
2762
2763
            qkv_layout = qkv_format[:-2] + "3" + qkv_format[-2:]
        elif (
            check_ptrs_qkv and check_strides_qkv and check_shapes_qkv and check_last_dim_offsets_qkv
        ):
2764
            # sbh3d, bsh3d, th3d
2765
2766
2767
2768
2769
            qkv_layout = qkv_format[:-1] + "3" + qkv_format[-1:]
        elif (
            check_ptrs_kv
            and check_strides_kv
            and check_shapes_kv
2770
            and check_last_two_dims_offsets_kv
2771
2772
            and not check_last_dim_offsets_kv
        ):
2773
            # sbhd_sb2hd, bshd_bs2hd, thd_t2hd
2774
2775
            qkv_layout = qkv_format + "_" + qkv_format[:-2] + "2" + qkv_format[-2:]
        elif check_ptrs_kv and check_strides_kv and check_shapes_kv and check_last_dim_offsets_kv:
2776
            # sbhd_sbh2d, bshd_bsh2d, thd_th2d
2777
            qkv_layout = qkv_format + "_" + qkv_format[:-1] + "2" + qkv_format[-1:]
2778
2779
        elif check_strides_kv and check_shapes_kv:
            # sbhd_sbhd_sbhd, bshd_bshd_bshd, thd_thd_thd
2780
            qkv_layout = "_".join(list([qkv_format]) * 3)
2781
        else:
2782
            qkv_layout = "not_supported"
2783
2784
2785
2786

        return qkv_layout

    qkv_layout = run_iteratively(q, k, v)
2787
    if qkv_layout == "not_supported":
2788
2789
2790
        # force q,k,v to be contiguous and run get_layout again
        q, k, v = [x.contiguous() for x in [q, k, v]]
        qkv_layout = run_iteratively(q, k, v)
2791
    if qkv_layout == "not_supported":
2792
2793
        raise Exception("The provided qkv memory layout is not supported!")

2794
    return qkv_layout, q, k, v
2795

2796

2797
def check_set_window_size(
2798
2799
2800
    attn_mask_type: str,
    window_size: Tuple[int, int] = None,
):
2801
2802
2803
2804
2805
2806
2807
2808
    """Check if sliding window size is compliant with attention mask type.
    If not, set it to the appropriate size.

         attn_mask_type                              |   window_size
    -------------------------------------------------------------------------
    no_mask, padding, arbitrary                      | (-1, -1) or (>=0, >=0)
    causal, padding_causal                           | (-1,  0) or (>=0, 0)
    causal_bottom_right, padding_causal_bottom_right | (-1,  0) or (>=0, 0)
2809
    """
2810
    orig_window_size = window_size
2811
    if "causal" in attn_mask_type:
2812
2813
2814
        if orig_window_size is None or (
            orig_window_size[0] == -1 and orig_window_size[1] in [-1, 0]
        ):
2815
            window_size = (-1, 0)
2816
2817
2818
2819
2820
2821
2822
2823
            warnings.warn(
                "window_size should be (-1, 0) or (>=0, 0) for attn_mask_type=" + attn_mask_type
            )
        elif orig_window_size[0] >= 0:
            window_size = (orig_window_size[0], 0)
            warnings.warn(
                "window_size should be (-1, 0) or (>=0, 0) for attn_mask_type=" + attn_mask_type
            )
2824
        else:
2825
2826
2827
2828
2829
2830
2831
            assert False, (
                "window_size should be (-1, 0) or (>=0, 0) for attn_mask_type=" + attn_mask_type
            )
    elif attn_mask_type in ["no_mask", "padding", "arbitrary"]:
        if orig_window_size is None or (
            orig_window_size[0] == -1 and orig_window_size[1] in [-1, 0]
        ):
2832
            window_size = (-1, -1)
2833
2834
2835
2836
2837
2838
2839
2840
2841
            warnings.warn(
                "window_size should be (-1, -1) or (>=0, >=0) for attn_mask_type=" + attn_mask_type
            )
        elif orig_window_size[0] < 0 or orig_window_size[0] < 0:
            assert False, (
                "window_size should be (-1, -1) or (>=0, >=0) for attn_mask_type=" + attn_mask_type
            )
    else:
        assert False, "Invalid attn_mask_type: " + attn_mask_type
2842
    return window_size
2843

2844

2845
class FlashAttention(torch.nn.Module):
2846
    """Dot product attention, using HazyResearch flash-attn package:
2847
    https://github.com/Dao-AILab/flash-attention
2848
2849
2850
2851
    """

    def __init__(
        self,
2852
        softmax_scale: float,
2853
2854
        attention_dropout: float = 0.0,
        attention_dropout_ctx: Optional[Callable] = nullcontext,
2855
2856
        attention_type: str = "self",
        layer_number: Optional[int] = None,
2857
        deterministic: bool = False,
2858
2859
2860
2861
2862
2863
    ) -> None:
        super().__init__()

        assert (
            _flash_attn_version >= _flash_attn_version_required
        ), f"FlashAttention minimum version {_flash_attn_version_required} is required."
2864
2865
2866
        assert (
            _flash_attn_version <= _flash_attn_max_version
        ), f"FlashAttention maximum version {_flash_attn_max_version} is supported."
2867

2868
        self.softmax_scale = softmax_scale
2869
2870
        self.attention_dropout_ctx = attention_dropout_ctx
        self.attention_dropout = attention_dropout
2871
2872
        self.attention_type = attention_type
        self.layer_number = 1 if layer_number is None else layer_number
2873
        self.deterministic = deterministic
2874
2875
2876
2877
2878
2879

    def forward(
        self,
        query_layer: torch.Tensor,
        key_layer: torch.Tensor,
        value_layer: torch.Tensor,
2880
        attention_mask: Optional[Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]] = None,
2881
2882
2883
        qkv_layout: str = "sbh3d",
        cu_seqlens_q: Optional[torch.Tensor] = None,
        cu_seqlens_kv: Optional[torch.Tensor] = None,
2884
2885
        max_seqlen_q: Optional[int] = None,
        max_seqlen_kv: Optional[int] = None,
2886
        attn_mask_type: str = "causal",
2887
        window_size: Optional[Tuple[int, int]] = None,
2888
        alibi_slopes: Optional[torch.Tensor] = None,
2889
        cp_group: Optional[dist_group_type] = None,
2890
        cp_global_ranks: List[int] = None,
2891
        cp_stream: torch.cuda.Stream = None,
2892
2893
2894
    ) -> torch.Tensor:
        """flash-attn fprop"""

2895
2896
        window_size = check_set_window_size(attn_mask_type, window_size)

2897
        assert (
2898
2899
2900
            query_layer.dtype in [torch.float16, torch.bfloat16]
            and key_layer.dtype in [torch.float16, torch.bfloat16]
            and value_layer.dtype in [torch.float16, torch.bfloat16]
2901
        ), "FlashAttention currently only supports FP16 and BF16."
2902
2903
        assert (
            query_layer.is_cuda and key_layer.is_cuda and value_layer.is_cuda
2904
        ), "FlashAttention currently only supports CUDA tensors."
2905
2906
        assert (
            qkv_layout in QKVLayouts
2907
        ), f"FlashAttention does not support qkv_layout = {qkv_layout}!"
2908

2909
2910
        context_parallel = (cp_group is not None) and (get_distributed_world_size(cp_group) != 1)

2911
        qkv_format = "".join([i for i in qkv_layout.split("_")[0] if i.isalpha()])
2912

2913
        if qkv_format == "sbhd":
2914
            # For now just 128, will make it more general in the future
2915
2916
2917
2918
2919
2920
2921
2922
            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
                )
2923
            else:
2924
2925
2926
2927
2928
2929
2930
                query_layer, key_layer, value_layer = [
                    x.transpose(0, 1).contiguous() for x in (query_layer, key_layer, value_layer)
                ]
        elif qkv_format in ["bshd", "thd"]:
            query_layer, key_layer, value_layer = [
                x.contiguous() for x in (query_layer, key_layer, value_layer)
            ]
2931

2932
        batch_size = query_layer.shape[0]
2933

2934
        if qkv_format in ["sbhd", "bshd"]:
2935
            max_seqlen_q, max_seqlen_kv = query_layer.shape[1], key_layer.shape[1]
2936
2937
2938
2939
2940
2941
2942
            if not context_parallel:
                # [b * s, h, d]
                query_layer, key_layer, value_layer = [
                    x.view(x.shape[0] * x.shape[1], *x.shape[2:])
                    for x in [query_layer, key_layer, value_layer]
                ]

2943
            if "padding" in attn_mask_type:
2944
                assert not context_parallel, "Padding mask not supported with context parallelism!"
2945
2946
2947
2948
2949

                if self.attention_type == "self":
                    assert (
                        max_seqlen_q == max_seqlen_kv
                    ), "Maximum sequence length for Q and KV should be the same."
2950
                    if cu_seqlens_q is None:
2951
2952
2953
                        assert (
                            attention_mask is not None
                        ), "Please provide attention_mask for padding!"
2954
2955
2956
2957
2958
2959
                        cu_seqlens_q, indices_q = get_cu_seqlens_and_indices(attention_mask)
                    else:
                        indices_q = get_indices(max_seqlen_q, cu_seqlens_q)
                    cu_seqlens_kv = cu_seqlens_q
                    query_layer, key_layer, value_layer = PackTensors.apply(
                        indices_q, query_layer, key_layer, value_layer
2960
2961
                    )
                else:
2962
                    if cu_seqlens_q is None or cu_seqlens_kv is None:
2963
2964
2965
2966
2967
                        assert (
                            attention_mask is not None
                        ), "Please provide attention_mask for padding!"
                        cu_seqlens_q, indices_q = get_cu_seqlens_and_indices(attention_mask[0])
                        cu_seqlens_kv, indices_kv = get_cu_seqlens_and_indices(attention_mask[1])
2968
2969
2970
2971
                    else:
                        indices_q = get_indices(max_seqlen_q, cu_seqlens_q)
                        indices_kv = get_indices(max_seqlen_kv, cu_seqlens_kv)
                    query_layer = PackTensors.apply(indices_q, query_layer)
2972
                    key_layer, value_layer = PackTensors.apply(indices_kv, key_layer, value_layer)
2973
            else:
2974
2975
2976
2977
2978
2979
2980
2981
2982
2983
2984
2985
2986
                # Cumulative sequence lengths for unpadded data
                if cu_seqlens_q is None:
                    cu_seqlens_q = _get_full_cu_seqlens(
                        batch_size,
                        max_seqlen_q,
                        query_layer.device,
                    )
                if cu_seqlens_kv is None:
                    cu_seqlens_kv = _get_full_cu_seqlens(
                        batch_size,
                        max_seqlen_kv,
                        key_layer.device,
                    )
2987
2988
2989
2990
        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!"
2991
2992
2993
2994
2995
2996
            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()
2997

2998
        if context_parallel:
2999
3000
3001
3002
            assert window_size in (
                (-1, -1),
                (-1, 0),
            ), "Sliding window attention is not supported with context parallelism."
3003
3004
3005
            assert (
                alibi_slopes is None
            ), "Alibi slope bias addition is not supported with context parallelism."
3006
            with self.attention_dropout_ctx():
3007
                output = attn_forward_func_with_cp(
3008
3009
3010
3011
3012
3013
3014
3015
3016
3017
                    self.training,
                    query_layer,
                    key_layer,
                    value_layer,
                    cu_seqlens_q,
                    cu_seqlens_kv,
                    max_seqlen_q,
                    max_seqlen_kv,
                    None,
                    None,
3018
                    self.attention_dropout if self.training else 0.0,
3019
3020
3021
                    cp_group,
                    cp_global_ranks,
                    cp_stream,
3022
                    softmax_scale=self.softmax_scale,
3023
                    qkv_format="bshd" if qkv_format == "sbhd" else qkv_format,
3024
                    attn_mask_type=attn_mask_type,
3025
                    deterministic=self.deterministic,
3026
3027
                )
        else:
3028
3029

            from .cpu_offload import CPUOffloadEnabled
3030

3031
3032
3033
3034
3035
3036
            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

3037
            with self.attention_dropout_ctx():
3038
                fa_optional_forward_kwargs = {}
3039
3040
                if _flash_attn_2_3_plus:
                    fa_optional_forward_kwargs["window_size"] = window_size
3041
3042
3043
3044
                if _flash_attn_2_4_plus:
                    fa_optional_forward_kwargs["alibi_slopes"] = alibi_slopes
                if _flash_attn_2_4_1_plus:
                    fa_optional_forward_kwargs["deterministic"] = self.deterministic
3045
                output = flash_attn_forward_func(
3046
3047
3048
3049
3050
3051
3052
                    query_layer,
                    key_layer,
                    value_layer,
                    cu_seqlens_q,
                    cu_seqlens_kv,
                    max_seqlen_q,
                    max_seqlen_kv,
3053
                    self.attention_dropout if self.training else 0.0,
3054
3055
                    softmax_scale=self.softmax_scale,
                    causal="causal" in attn_mask_type,
3056
                    **fa_optional_forward_kwargs,
3057
                )
3058

3059
        if qkv_format in ["sbhd", "bshd"] and "padding" in attn_mask_type:
3060
            output = UnpackTensor.apply(indices_q, batch_size * max_seqlen_q, output)
3061

3062
        if qkv_format == "sbhd":
3063
3064
            # (bs)hd -> bs(hd) -> sb(hd)
            output = output.view(batch_size, max_seqlen_q, -1).transpose(0, 1).contiguous()
3065
        elif qkv_format == "bshd":
3066
3067
            # (bs)hd -> bs(hd)
            output = output.view(batch_size, max_seqlen_q, -1).contiguous()
3068
        elif qkv_format == "thd":
3069
3070
            # thd -> t(hd)
            output = output.view(output.shape[0], -1).contiguous()
3071
3072

        return output
3073

3074

3075
def _combine_tensors(
3076
3077
3078
    tensors: List[torch.Tensor],
    dim: int,
) -> torch.Tensor:
3079
3080
3081
3082
3083
3084
    """Combine tensors along a particular dimension"""

    num_tensors = len(tensors)
    new_shape = list(tensors[0].shape)
    new_shape.insert(dim, num_tensors)
    new_stride = list(tensors[0].stride())
3085
    new_stride.insert(dim, int(new_stride[dim - 1] / num_tensors))
3086
    if isinstance(tensors[0], Float8Tensor):
3087
        combined_tensor = torch.Tensor().to(device=tensors[0].device, dtype=tensors[0]._data.dtype)
3088
3089
3090
        combined_tensor.set_(
            tensors[0]._data.untyped_storage(),
            tensors[0]._data.storage_offset(),
3091
3092
3093
3094
            new_shape,
            new_stride,
        )
        combined_tensor = Float8Tensor.make_like(tensors[0], data=combined_tensor)
3095
    else:
3096
        combined_tensor = torch.Tensor().to(device=tensors[0].device, dtype=tensors[0].dtype)
3097
        combined_tensor.set_(
3098
3099
            tensors[0].untyped_storage(), tensors[0].storage_offset(), new_shape, new_stride
        )
3100
3101

    return combined_tensor
3102

3103

3104
3105
3106
3107
class FusedAttnFunc_qkvpacked(torch.autograd.Function):
    """Function for FusedAttention with packed QKV input"""

    @staticmethod
3108
3109
3110
3111
3112
    def forward(
        ctx,
        is_training,
        max_seqlen,
        cu_seqlens,
3113
        cu_seqlens_padded,
3114
3115
3116
3117
3118
3119
3120
3121
3122
3123
3124
3125
3126
3127
3128
        qkv,
        qkv_dtype,
        attn_bias,
        attn_scale,
        dropout_p,
        fast_zero_fill,
        qkv_layout,
        attn_bias_type,
        attn_mask_type,
        rng_gen,
        fused_attention_backend,
        use_FAv2_bwd,
        fp8,
        fp8_meta,
    ):
3129
        logger = logging.getLogger("FusedAttnFunc_qkvpacked")
3130
        if fp8:
3131
            logger.debug("Running forward in FP8")
3132
            if fp8_meta["recipe"].fp8_mha:
3133
                assert isinstance(qkv, Float8Tensor), "qkv must be Float8Tensors for FP8 MHA."
3134
3135
3136
3137
                fp8_meta["scaling_fwd"].scale_inv[META_QKV] = qkv._scale_inv
            fused_attention_backend = FusedAttnBackend["FP8"]
            fp8_dtype_forward = get_fp8_te_dtype(fp8_meta["recipe"], fprop_tensor=True)
            # 1: qkv packed, 2: kv packed, 3: qkv separate
3138
3139
3140
3141
3142
            qkv_group = len(qkv_layout.split("_"))
            assert qkv_group == 1, (
                "qkv layout should conform to 3hd or h3d, e.g. sb3hd,                 but found"
                f" {qkv_layout}."
            )
3143
3144
3145
3146
            if fp8_meta["recipe"].fp8_mha:
                qkv_fp8 = qkv._data
            else:
                qkv_c = qkv.view(-1, qkv.shape[-3] * qkv.shape[-2] * qkv.shape[-1])
3147
3148
3149
                qkv_fp8 = cast_to_fp8(
                    qkv_c, fp8_meta["scaling_fwd"], META_QKV, fp8_dtype_forward
                ).view(qkv.shape)
3150
            out_fp8, aux_ctx_tensors = fused_attn_fwd_qkvpacked(
3151
3152
3153
3154
3155
3156
3157
                is_training,
                max_seqlen,
                cu_seqlens,
                qkv_fp8,
                fp8_dtype_forward,
                fused_attention_backend,
                attn_bias,
3158
                cu_seqlens_padded,
3159
3160
3161
3162
3163
3164
                fp8_meta["scaling_fwd"].scale_inv[META_QKV],
                fp8_meta["scaling_fwd"].scale_inv[META_S],
                fp8_meta["scaling_fwd"].scale[META_S],
                fp8_meta["scaling_fwd"].scale[META_O],
                fp8_meta["scaling_fwd"].amax_history[0][META_S],
                fp8_meta["scaling_fwd"].amax_history[0][META_O],
3165
3166
3167
3168
3169
3170
3171
3172
                attn_scale,
                dropout_p,
                fast_zero_fill,
                qkv_layout,
                attn_bias_type,
                attn_mask_type,
                rng_gen,
            )
3173
            if fp8_meta["recipe"].fp8_mha:
3174
3175
                out_ret = Float8Tensor(
                    data=out_fp8,
3176
3177
3178
3179
3180
3181
3182
3183
3184
                    fp8_meta=fp8_meta,
                    fp8_meta_forward=True,
                    fp8_meta_index=META_O,
                    fp8_dtype=fp8_dtype_forward,
                    dtype=qkv.dtype,
                )
            else:
                out_ret = cast_from_fp8(
                    out_fp8.view(-1, out_fp8.shape[-2] * out_fp8.shape[-1]),
3185
3186
3187
3188
3189
                    fp8_meta["scaling_fwd"],
                    META_O,
                    fp8_dtype_forward,
                    qkv_dtype,
                ).view(out_fp8.shape)
3190
3191
3192
            out_save = out_ret
            if fp8_meta["recipe"].fp8_mha and not int(os.getenv("NVTE_FP8_DPA_BWD", "1")):
                qkv_c = qkv.view(-1, qkv.shape[-3] * qkv.shape[-2] * qkv.shape[-1])
3193
3194
                qkv = cast_from_fp8(
                    qkv_c._data,
3195
                    fp8_meta["scaling_fwd"],
3196
3197
3198
3199
                    META_QKV,
                    fp8_dtype_forward,
                    TE_DType[qkv.dtype],
                ).view(qkv.shape)
3200
3201
                out_save = cast_from_fp8(
                    out_fp8.view(-1, out_fp8.shape[-2] * out_fp8.shape[-1]),
3202
3203
3204
3205
3206
3207
3208
3209
                    fp8_meta["scaling_fwd"],
                    META_O,
                    fp8_dtype_forward,
                    qkv_dtype,
                ).view(out_fp8.shape)
            fp8_tensors = (
                qkv_fp8,
                out_fp8,
3210
                fp8_meta["scaling_fwd"].scale.clone(),
3211
3212
                fp8_meta["scaling_fwd"].scale_inv.clone(),
            )
3213
        else:
3214
            logger.debug("Running forward in %s", qkv.dtype)
3215
            out_ret, aux_ctx_tensors = fused_attn_fwd_qkvpacked(
3216
3217
3218
3219
3220
3221
3222
                is_training,
                max_seqlen,
                cu_seqlens,
                qkv,
                qkv_dtype,
                fused_attention_backend,
                attn_bias,
3223
                cu_seqlens_padded,
3224
3225
3226
3227
3228
3229
3230
3231
3232
3233
3234
3235
3236
3237
                None,
                None,
                None,
                None,
                None,
                None,
                attn_scale,
                dropout_p,
                fast_zero_fill,
                qkv_layout,
                attn_bias_type,
                attn_mask_type,
                rng_gen,
            )
3238
3239
3240
3241
3242
            fp8_tensors = (None, None, None, None)
            out_save = out_ret

        ctx.fp8 = fp8 and int(os.getenv("NVTE_FP8_DPA_BWD", "1"))
        qkvo_tensors = (qkv, out_save) if not ctx.fp8 else (None, None)
3243
        ctx.save_for_backward(
3244
            *qkvo_tensors, cu_seqlens, cu_seqlens_padded, *fp8_tensors, *aux_ctx_tensors
3245
        )
3246
        ctx.fp8_meta = fp8_meta
3247
3248
3249
3250
3251
3252
3253
3254
        ctx.max_seqlen = max_seqlen
        ctx.qkv_dtype = qkv_dtype
        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
3255
        ctx.fused_attention_backend = (
3256
            fused_attention_backend if ctx.fp8 else FusedAttnBackend["F16_arbitrary_seqlen"]
3257
        )
3258
        ctx.use_FAv2_bwd = use_FAv2_bwd
3259

3260
        return out_ret
3261
3262
3263

    @staticmethod
    def backward(ctx, d_out):
3264
        logger = logging.getLogger("FusedAttnFunc_qkvpacked")
3265
        if ctx.fp8_meta["recipe"].fp8_mha:
3266
3267
3268
            assert isinstance(
                d_out, Float8Tensor
            ), "Gradient of the DPA output must be in Float8Tensor type for FP8 MHA."
3269
3270
3271
            d_out_f8tensor = d_out
            d_out = d_out._data

3272
        d_out = d_out.contiguous()
3273
3274
3275
3276
        (
            qkv,
            out,
            cu_seqlens,
3277
            cu_seqlens_padded,
3278
3279
3280
3281
3282
3283
            qkv_fp8,
            out_fp8,
            fwd_scales,
            fwd_scale_invs,
            *aux_ctx_tensors,
        ) = ctx.saved_tensors
3284
3285
        if not aux_ctx_tensors[0].is_contiguous():
            aux_ctx_tensors[0] = aux_ctx_tensors[0].contiguous()
3286
        if ctx.use_FAv2_bwd:
3287
            softmax_lse, rng_state = aux_ctx_tensors
3288
3289
            dqkv = torch.empty_like(qkv)
            maybe_contiguous = lambda x: x.contiguous() if x.stride(-1) != 1 else x
3290
3291
3292
            d_out, q, k, v, out = [
                maybe_contiguous(x) for x in (d_out, qkv[:, 0], qkv[:, 1], qkv[:, 2], out)
            ]
3293
            flash_attn_cuda_bwd(
3294
3295
3296
3297
3298
3299
3300
3301
3302
3303
3304
3305
3306
3307
3308
3309
3310
3311
3312
                d_out,
                q,
                k,
                v,
                out,
                softmax_lse,
                dqkv[:, 0],
                dqkv[:, 1],
                dqkv[:, 2],
                cu_seqlens,
                cu_seqlens,
                ctx.max_seqlen,
                ctx.max_seqlen,
                ctx.dropout_p,
                ctx.attn_scale,
                False,
                "causal" in ctx.attn_mask_type,
                None,
                rng_state,
3313
            )
3314
            dqkv = dqkv[..., : d_out.shape[-1]]
3315
        else:
3316
3317
            with torch.cuda.nvtx.range("_FusedAttn_qkvpacked"):
                if ctx.fp8:
3318
                    logger.debug("Running backward in FP8")
3319
                    fp8_dtype_forward = get_fp8_te_dtype(ctx.fp8_meta["recipe"], fprop_tensor=True)
3320
                    fp8_dtype_backward = get_fp8_te_dtype(
3321
3322
                        ctx.fp8_meta["recipe"], fprop_tensor=False
                    )
3323
3324
                    if ctx.fp8_meta["recipe"].fp8_mha:
                        d_out_fp8 = d_out
3325
                        ctx.fp8_meta["scaling_bwd"].scale_inv[META_DO] = d_out_f8tensor._scale_inv
3326
3327
3328
                    else:
                        d_out_fp8 = cast_to_fp8(
                            d_out.view(-1, d_out.shape[-2] * d_out.shape[-1]),
3329
3330
3331
3332
                            ctx.fp8_meta["scaling_bwd"],
                            META_DO,
                            fp8_dtype_backward,
                        ).view(d_out.shape)
3333
                    dqkv_fp8, *rest = fused_attn_bwd_qkvpacked(
3334
3335
3336
3337
3338
3339
3340
3341
                        ctx.max_seqlen,
                        cu_seqlens,
                        qkv_fp8,
                        out_fp8,
                        d_out_fp8,
                        fp8_dtype_forward,
                        fp8_dtype_backward,
                        aux_ctx_tensors,
3342
                        ctx.fused_attention_backend,
3343
                        cu_seqlens_padded,
3344
3345
3346
3347
3348
3349
3350
3351
3352
3353
3354
3355
3356
3357
3358
3359
3360
                        fwd_scale_invs[META_QKV],  # d_scale_qkv,
                        fwd_scale_invs[META_S],  # d_scale_s,
                        fwd_scale_invs[META_O],  # d_scale_o,
                        ctx.fp8_meta["scaling_bwd"].scale_inv[META_DO],  # d_scale_do
                        ctx.fp8_meta["scaling_bwd"].scale_inv[META_DP],  # d_scale_dp
                        fwd_scales[META_S],  # q_scale_s
                        ctx.fp8_meta["scaling_bwd"].scale[META_DP],  # q_scale_dp
                        ctx.fp8_meta["scaling_bwd"].scale[META_DQKV],  # q_scale_dqkv
                        ctx.fp8_meta["scaling_bwd"].amax_history[0][META_DP],  # amax_dp
                        ctx.fp8_meta["scaling_bwd"].amax_history[0][META_DQKV],  # amax_dqkv
                        ctx.attn_scale,
                        ctx.dropout_p,
                        ctx.fast_zero_fill,
                        ctx.qkv_layout,
                        ctx.attn_bias_type,
                        ctx.attn_mask_type,
                    )
3361
                    if ctx.fp8_meta["recipe"].fp8_mha:
3362
3363
                        dqkv = Float8Tensor(
                            data=dqkv_fp8,
3364
3365
3366
3367
3368
                            fp8_meta=ctx.fp8_meta,
                            fp8_meta_forward=False,
                            fp8_meta_index=META_DQKV,
                            fp8_dtype=fp8_dtype_backward,
                            dtype=d_out_f8tensor.dtype,
3369
                        )
3370
                    else:
3371
3372
3373
3374
3375
3376
3377
3378
3379
3380
                        dqkv_c_fp8 = dqkv_fp8.view(
                            -1, dqkv_fp8.shape[-3] * dqkv_fp8.shape[-2] * dqkv_fp8.shape[-1]
                        )
                        dqkv = cast_from_fp8(
                            dqkv_c_fp8,
                            ctx.fp8_meta["scaling_bwd"],
                            META_DQKV,
                            fp8_dtype_backward,
                            ctx.qkv_dtype,
                        ).view(dqkv_fp8.shape)
3381
                else:
3382
                    logger.debug("Running backward in %s", qkv.dtype)
3383
3384
3385
                    if d_out.dtype == torch.uint8:
                        d_out = d_out_f8tensor.from_float8(qkv.dtype)
                    dqkv, *rest = fused_attn_bwd_qkvpacked(
3386
3387
3388
3389
3390
3391
3392
3393
                        ctx.max_seqlen,
                        cu_seqlens,
                        qkv,
                        out,
                        d_out,
                        ctx.qkv_dtype,
                        ctx.qkv_dtype,
                        aux_ctx_tensors,
3394
                        ctx.fused_attention_backend,
3395
                        cu_seqlens_padded,
3396
3397
3398
3399
3400
3401
3402
3403
3404
3405
3406
3407
3408
3409
3410
3411
3412
                        None,
                        None,
                        None,
                        None,
                        None,
                        None,
                        None,
                        None,
                        None,
                        None,
                        ctx.attn_scale,
                        ctx.dropout_p,
                        ctx.fast_zero_fill,
                        ctx.qkv_layout,
                        ctx.attn_bias_type,
                        ctx.attn_mask_type,
                    )
3413

3414
3415
        # if no_bias or alibi, return dqkv
        if ctx.attn_bias_type in ["no_bias", "alibi"]:
3416
3417
3418
3419
3420
3421
3422
3423
3424
3425
3426
3427
3428
3429
3430
3431
3432
3433
3434
3435
3436
3437
            return (
                None,
                None,
                None,
                None,
                dqkv,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
            )
3438
        # else, return (dqkv, dbias)
3439
3440
3441
3442
3443
3444
3445
3446
3447
3448
3449
3450
3451
3452
3453
3454
3455
3456
3457
3458
3459
3460
        return (
            None,
            None,
            None,
            None,
            dqkv,
            None,
            rest[0],
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
        )
3461

3462

3463
3464
3465
3466
class FusedAttnFunc_kvpacked(torch.autograd.Function):
    """Function for FusedAttention with packed KV input"""

    @staticmethod
3467
3468
3469
3470
3471
3472
3473
    def forward(
        ctx,
        is_training,
        max_seqlen_q,
        max_seqlen_kv,
        cu_seqlens_q,
        cu_seqlens_kv,
3474
3475
        cu_seqlens_q_padded,
        cu_seqlens_kv_padded,
3476
3477
3478
3479
3480
3481
3482
3483
3484
3485
3486
3487
3488
3489
3490
3491
        q,
        kv,
        qkv_dtype,
        attn_bias,
        attn_scale,
        dropout_p,
        fast_zero_fill,
        qkv_layout,
        attn_bias_type,
        attn_mask_type,
        rng_gen,
        fused_attention_backend,
        use_FAv2_bwd,
        fp8,
        fp8_meta,
    ):
3492
        logger = logging.getLogger("FusedAttnFunc_kvpacked")
3493
        if fp8:
3494
            logger.debug("Running forward in FP8")
3495
            if fp8_meta["recipe"].fp8_mha:
3496
3497
3498
                assert isinstance(q, Float8Tensor) and isinstance(
                    kv, Float8Tensor
                ), "q/kv must be Float8Tensors for FP8 MHA."
3499
3500
3501
3502
3503
3504
3505
                fp8_meta["scaling_fwd"].scale_inv[META_QKV] = q._scale_inv
            fused_attention_backend = FusedAttnBackend["FP8"]
            fp8_dtype_forward = get_fp8_te_dtype(fp8_meta["recipe"], fprop_tensor=True)
            if fp8_meta["recipe"].fp8_mha:
                q_fp8, kv_fp8 = q._data, kv._data
            else:
                # 1: qkv packed, 2: kv packed, 3: qkv separate
3506
3507
3508
3509
3510
3511
3512
3513
                qkv_group = len(qkv_layout.split("_"))
                assert qkv_group == 2, (
                    "qkv layout should conform to hd_2hd or hd_h2d, e.g. sbhd_sb2hd,              "
                    f"       but found {qkv_layout}."
                )
                q_fp8 = cast_to_fp8(q, fp8_meta["scaling_fwd"], META_QKV, fp8_dtype_forward).view(
                    q.shape
                )
3514
                kv_c = kv.view(-1, kv.shape[-3] * kv.shape[-2] * kv.shape[-1])
3515
3516
3517
                kv_fp8 = cast_to_fp8(
                    kv_c, fp8_meta["scaling_fwd"], META_QKV, fp8_dtype_forward
                ).view(kv.shape)
3518
            out_fp8, aux_ctx_tensors = fused_attn_fwd_kvpacked(
3519
3520
3521
3522
3523
3524
3525
3526
3527
3528
                is_training,
                max_seqlen_q,
                max_seqlen_kv,
                cu_seqlens_q,
                cu_seqlens_kv,
                q_fp8,
                kv_fp8,
                fp8_dtype_forward,
                fused_attention_backend,
                attn_bias,
3529
3530
                cu_seqlens_q_padded,
                cu_seqlens_kv_padded,
3531
3532
3533
3534
3535
3536
                fp8_meta["scaling_fwd"].scale_inv[META_QKV],
                fp8_meta["scaling_fwd"].scale_inv[META_S],
                fp8_meta["scaling_fwd"].scale[META_S],
                fp8_meta["scaling_fwd"].scale[META_O],
                fp8_meta["scaling_fwd"].amax_history[0][META_S],
                fp8_meta["scaling_fwd"].amax_history[0][META_O],
3537
3538
3539
3540
3541
3542
3543
3544
                attn_scale,
                dropout_p,
                fast_zero_fill,
                qkv_layout,
                attn_bias_type,
                attn_mask_type,
                rng_gen,
            )
3545
            if fp8_meta["recipe"].fp8_mha:
3546
3547
                out_ret = Float8Tensor(
                    data=out_fp8,
3548
3549
3550
3551
3552
3553
3554
3555
3556
                    fp8_meta=fp8_meta,
                    fp8_meta_forward=True,
                    fp8_meta_index=META_O,
                    fp8_dtype=fp8_dtype_forward,
                    dtype=q.dtype,
                )
            else:
                out_ret = cast_from_fp8(
                    out_fp8.view(-1, out_fp8.shape[-2] * out_fp8.shape[-1]),
3557
3558
3559
3560
3561
                    fp8_meta["scaling_fwd"],
                    META_O,
                    fp8_dtype_forward,
                    qkv_dtype,
                ).view(out_fp8.shape)
3562
3563
            out_save = out_ret
            if fp8_meta["recipe"].fp8_mha and not int(os.getenv("NVTE_FP8_DPA_BWD", "1")):
3564
3565
3566
                q = cast_from_fp8(
                    q._data, fp8_meta["scaling_fwd"], META_QKV, fp8_dtype_forward, TE_DType[q.dtype]
                ).view(q.shape)
3567
                kv_c = kv.view(-1, kv.shape[-3] * kv.shape[-2] * kv.shape[-1])
3568
3569
                kv = cast_from_fp8(
                    kv_c._data,
3570
                    fp8_meta["scaling_fwd"],
3571
3572
3573
3574
                    META_QKV,
                    fp8_dtype_forward,
                    TE_DType[kv.dtype],
                ).view(kv.shape)
3575
3576
                out_save = cast_from_fp8(
                    out_fp8.view(-1, out_fp8.shape[-2] * out_fp8.shape[-1]),
3577
3578
3579
3580
3581
3582
3583
3584
3585
                    fp8_meta["scaling_fwd"],
                    META_O,
                    fp8_dtype_forward,
                    qkv_dtype,
                ).view(out_fp8.shape)
            fp8_tensors = (
                q_fp8,
                kv_fp8,
                out_fp8,
3586
                fp8_meta["scaling_fwd"].scale.clone(),
3587
3588
                fp8_meta["scaling_fwd"].scale_inv.clone(),
            )
3589
        else:
3590
            logger.debug("Running forward in %s", q.dtype)
3591
            out_ret, aux_ctx_tensors = fused_attn_fwd_kvpacked(
3592
3593
3594
3595
3596
3597
3598
3599
3600
3601
                is_training,
                max_seqlen_q,
                max_seqlen_kv,
                cu_seqlens_q,
                cu_seqlens_kv,
                q,
                kv,
                qkv_dtype,
                fused_attention_backend,
                attn_bias,
3602
3603
                cu_seqlens_q_padded,
                cu_seqlens_kv_padded,
3604
3605
3606
3607
3608
3609
3610
3611
3612
3613
3614
3615
3616
3617
                None,
                None,
                None,
                None,
                None,
                None,
                attn_scale,
                dropout_p,
                fast_zero_fill,
                qkv_layout,
                attn_bias_type,
                attn_mask_type,
                rng_gen,
            )
3618
3619
3620
3621
3622
            out_save = out_ret
            fp8_tensors = (None, None, None, None, None)

        ctx.fp8 = fp8 and int(os.getenv("NVTE_FP8_DPA_BWD", "1"))
        qkvo_tensors = (q, kv, out_save) if not ctx.fp8 else (None, None, None)
3623
3624
3625
3626
        ctx.save_for_backward(
            *qkvo_tensors,
            cu_seqlens_q,
            cu_seqlens_kv,
3627
3628
            cu_seqlens_q_padded,
            cu_seqlens_kv_padded,
3629
3630
3631
            *fp8_tensors,
            *aux_ctx_tensors,
        )
3632
        ctx.fp8_meta = fp8_meta
3633
3634
3635
3636
3637
3638
3639
3640
3641
        ctx.max_seqlen_q = max_seqlen_q
        ctx.max_seqlen_kv = max_seqlen_kv
        ctx.qkv_dtype = qkv_dtype
        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
3642
        ctx.fused_attention_backend = (
3643
            fused_attention_backend if ctx.fp8 else FusedAttnBackend["F16_arbitrary_seqlen"]
3644
        )
3645
        ctx.use_FAv2_bwd = use_FAv2_bwd
3646

3647
        return out_ret
3648
3649
3650

    @staticmethod
    def backward(ctx, d_out):
3651
        logger = logging.getLogger("FusedAttnFunc_kvpacked")
3652
        if ctx.fp8_meta["recipe"].fp8_mha:
3653
3654
3655
            assert isinstance(
                d_out, Float8Tensor
            ), "Gradient of the DPA output must be in Float8Tensor type for FP8 MHA."
3656
3657
3658
            d_out_f8tensor = d_out
            d_out = d_out._data

3659
        d_out = d_out.contiguous()
3660
3661
3662
3663
3664
3665
        (
            q,
            kv,
            out,
            cu_seqlens_q,
            cu_seqlens_kv,
3666
3667
            cu_seqlens_q_padded,
            cu_seqlens_kv_padded,
3668
3669
3670
3671
3672
3673
3674
            q_fp8,
            kv_fp8,
            out_fp8,
            fwd_scales,
            fwd_scale_invs,
            *aux_ctx_tensors,
        ) = ctx.saved_tensors
3675
3676
        if not aux_ctx_tensors[0].is_contiguous():
            aux_ctx_tensors[0] = aux_ctx_tensors[0].contiguous()
3677
        if ctx.use_FAv2_bwd:
3678
            softmax_lse, rng_state = aux_ctx_tensors
3679
3680
3681
            dq = torch.empty_like(q)
            dkv = torch.empty_like(kv)
            maybe_contiguous = lambda x: x.contiguous() if x.stride(-1) != 1 else x
3682
            d_out, q, k, v, out = [maybe_contiguous(x) for x in (d_out, q, kv[:, 0], kv[:, 1], out)]
3683
            flash_attn_cuda_bwd(
3684
3685
3686
3687
3688
3689
3690
3691
3692
3693
3694
3695
3696
3697
3698
3699
3700
3701
3702
                d_out,
                q,
                k,
                v,
                out,
                softmax_lse,
                dq,
                dkv[:, 0],
                dkv[:, 1],
                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,
3703
            )
3704
3705
            dq = dq[..., : d_out.shape[-1]]
            dkv = dkv[..., : d_out.shape[-1]]
3706
        else:
3707
3708
            with torch.cuda.nvtx.range("_FusedAttn_kvpacked"):
                if ctx.fp8:
3709
                    logger.debug("Running backward in FP8")
3710
                    fp8_dtype_forward = get_fp8_te_dtype(ctx.fp8_meta["recipe"], fprop_tensor=True)
3711
                    fp8_dtype_backward = get_fp8_te_dtype(
3712
3713
                        ctx.fp8_meta["recipe"], fprop_tensor=False
                    )
3714
3715
                    if ctx.fp8_meta["recipe"].fp8_mha:
                        d_out_fp8 = d_out
3716
                        ctx.fp8_meta["scaling_bwd"].scale_inv[META_DO] = d_out_f8tensor._scale_inv
3717
3718
3719
                    else:
                        d_out_fp8 = cast_to_fp8(
                            d_out.view(-1, d_out.shape[-2] * d_out.shape[-1]),
3720
3721
3722
3723
                            ctx.fp8_meta["scaling_bwd"],
                            META_DO,
                            fp8_dtype_backward,
                        ).view(d_out.shape)
3724
                    dq_fp8, dkv_fp8, *rest = fused_attn_bwd_kvpacked(
3725
3726
3727
3728
3729
3730
3731
3732
3733
3734
3735
                        ctx.max_seqlen_q,
                        ctx.max_seqlen_kv,
                        cu_seqlens_q,
                        cu_seqlens_kv,
                        q_fp8,
                        kv_fp8,
                        out_fp8,
                        d_out_fp8,
                        fp8_dtype_forward,
                        fp8_dtype_backward,
                        aux_ctx_tensors,
3736
                        ctx.fused_attention_backend,
3737
3738
                        cu_seqlens_q_padded,
                        cu_seqlens_kv_padded,
3739
3740
3741
3742
3743
3744
3745
3746
3747
3748
3749
3750
3751
3752
3753
3754
3755
                        fwd_scale_invs[META_QKV],  # d_scale_qkv,
                        fwd_scale_invs[META_S],  # d_scale_s,
                        fwd_scale_invs[META_O],  # d_scale_o,
                        ctx.fp8_meta["scaling_bwd"].scale_inv[META_DO],  # d_scale_do
                        ctx.fp8_meta["scaling_bwd"].scale_inv[META_DP],  # d_scale_dp
                        fwd_scales[META_S],  # q_scale_s
                        ctx.fp8_meta["scaling_bwd"].scale[META_DP],  # q_scale_dp
                        ctx.fp8_meta["scaling_bwd"].scale[META_DQKV],  # q_scale_dqkv
                        ctx.fp8_meta["scaling_bwd"].amax_history[0][META_DP],  # amax_dp
                        ctx.fp8_meta["scaling_bwd"].amax_history[0][META_DQKV],  # amax_dqkv
                        ctx.attn_scale,
                        ctx.dropout_p,
                        ctx.fast_zero_fill,
                        ctx.qkv_layout,
                        ctx.attn_bias_type,
                        ctx.attn_mask_type,
                    )
3756
                    if ctx.fp8_meta["recipe"].fp8_mha:
3757
3758
                        dq = Float8Tensor(
                            data=dq_fp8,
3759
3760
3761
3762
3763
                            fp8_meta=ctx.fp8_meta,
                            fp8_meta_forward=False,
                            fp8_meta_index=META_DQKV,
                            fp8_dtype=fp8_dtype_backward,
                            dtype=d_out_f8tensor.dtype,
3764
3765
3766
                        )
                        dkv = Float8Tensor(
                            data=dkv_fp8,
3767
3768
3769
3770
3771
                            fp8_meta=ctx.fp8_meta,
                            fp8_meta_forward=False,
                            fp8_meta_index=META_DQKV,
                            fp8_dtype=fp8_dtype_backward,
                            dtype=d_out_f8tensor.dtype,
3772
                        )
3773
3774
3775
                    else:
                        dq = cast_from_fp8(
                            dq_fp8.view(-1, dq_fp8.shape[-2] * dq_fp8.shape[-1]),
3776
3777
3778
3779
3780
3781
3782
3783
3784
3785
3786
3787
3788
3789
3790
                            ctx.fp8_meta["scaling_bwd"],
                            META_DQKV,
                            fp8_dtype_backward,
                            ctx.qkv_dtype,
                        ).view(dq_fp8.shape)
                        dkv_c_fp8 = dkv_fp8.view(
                            -1, dkv_fp8.shape[-3] * dkv_fp8.shape[-2] * dkv_fp8.shape[-1]
                        )
                        dkv = cast_from_fp8(
                            dkv_c_fp8,
                            ctx.fp8_meta["scaling_bwd"],
                            META_DQKV,
                            fp8_dtype_backward,
                            ctx.qkv_dtype,
                        ).view(dkv_fp8.shape)
3791
                else:
3792
                    logger.debug("Running backward in %s", q.dtype)
3793
3794
3795
                    if d_out.dtype == torch.uint8:
                        d_out = d_out_f8tensor.from_float8(q.dtype)
                    dq, dkv, *rest = fused_attn_bwd_kvpacked(
3796
3797
3798
3799
3800
3801
3802
3803
3804
3805
3806
                        ctx.max_seqlen_q,
                        ctx.max_seqlen_kv,
                        cu_seqlens_q,
                        cu_seqlens_kv,
                        q,
                        kv,
                        out,
                        d_out,
                        ctx.qkv_dtype,
                        ctx.qkv_dtype,
                        aux_ctx_tensors,
3807
                        ctx.fused_attention_backend,
3808
3809
                        cu_seqlens_q_padded,
                        cu_seqlens_kv_padded,
3810
3811
3812
3813
3814
3815
3816
3817
3818
3819
3820
3821
3822
3823
3824
3825
3826
                        None,
                        None,
                        None,
                        None,
                        None,
                        None,
                        None,
                        None,
                        None,
                        None,
                        ctx.attn_scale,
                        ctx.dropout_p,
                        ctx.fast_zero_fill,
                        ctx.qkv_layout,
                        ctx.attn_bias_type,
                        ctx.attn_mask_type,
                    )
3827

3828
3829
        # if no_bias or alibi, return dqkv
        if ctx.attn_bias_type in ["no_bias", "alibi"]:
3830
3831
3832
3833
3834
3835
3836
3837
3838
3839
3840
3841
3842
3843
3844
3845
3846
3847
3848
3849
3850
3851
3852
3853
3854
3855
            return (
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                dq,
                dkv,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
            )
3856
        # else, return (dqkv, dbias)
3857
3858
3859
3860
3861
3862
3863
3864
3865
3866
3867
3868
3869
3870
3871
3872
3873
3874
3875
3876
3877
3878
3879
3880
3881
3882
3883
        return (
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            dq,
            dkv,
            None,
            rest[0],
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
        )

3884

3885
3886
3887
3888
class FusedAttnFunc(torch.autograd.Function):
    """Function for FusedAttention with separate Q, K, V tensors"""

    @staticmethod
3889
3890
3891
3892
3893
3894
3895
    def forward(
        ctx,
        is_training,
        max_seqlen_q,
        max_seqlen_kv,
        cu_seqlens_q,
        cu_seqlens_kv,
3896
3897
        cu_seqlens_q_padded,
        cu_seqlens_kv_padded,
3898
3899
3900
3901
3902
3903
3904
3905
3906
3907
3908
3909
3910
3911
3912
3913
3914
        q,
        k,
        v,
        qkv_dtype,
        attn_bias,
        attn_scale,
        dropout_p,
        fast_zero_fill,
        qkv_layout,
        attn_bias_type,
        attn_mask_type,
        rng_gen,
        fused_attention_backend,
        use_FAv2_bwd,
        fp8,
        fp8_meta,
    ):
3915
        logger = logging.getLogger("FusedAttnFunc")
3916
        if fp8:
3917
            logger.debug("Running forward in FP8")
3918
3919
3920
            fused_attention_backend = FusedAttnBackend["FP8"]
            fp8_dtype_forward = get_fp8_te_dtype(fp8_meta["recipe"], fprop_tensor=True)
            if fp8_meta["recipe"].fp8_mha:
3921
3922
                assert (
                    isinstance(q, Float8Tensor)
3923
                    and isinstance(k, Float8Tensor)
3924
3925
                    and isinstance(v, Float8Tensor)
                ), "q/k/v must be Float8Tensors for FP8 MHA."
3926
3927
3928
3929
                fp8_meta["scaling_fwd"].scale_inv[META_QKV] = q._scale_inv
                q_fp8, k_fp8, v_fp8 = q._data, k._data, v._data
            else:
                # 1: qkv packed, 2: kv packed, 3: qkv separate
3930
                qkv_group = len(qkv_layout.split("_"))
3931
                if qkv_group == 1:
3932
3933
                    dim = qkv_layout.find("3")
                    qkv = _combine_tensors([q, k, v], dim)
3934
                    qkv_c = qkv.view(-1, qkv.shape[-3] * qkv.shape[-2] * qkv.shape[-1])
3935
3936
3937
3938
                    qkv_fp8 = cast_to_fp8(
                        qkv_c, fp8_meta["scaling_fwd"], META_QKV, fp8_dtype_forward
                    ).view(qkv.shape)
                    q_fp8, k_fp8, v_fp8 = _SplitAlongDim.apply(qkv_fp8, dim, [1, 1, 1])
3939
3940
                    q_fp8, k_fp8, v_fp8 = [x.squeeze(dim) for x in [q_fp8, k_fp8, v_fp8]]
                if qkv_group == 2:
3941
3942
3943
3944
3945
                    q_fp8 = cast_to_fp8(
                        q, fp8_meta["scaling_fwd"], META_QKV, fp8_dtype_forward
                    ).view(q.shape)
                    dim = qkv_layout.split("_")[1].find("2")
                    kv = _combine_tensors([k, v], dim)
3946
                    kv_c = kv.view(-1, kv.shape[-3] * kv.shape[-2] * kv.shape[-1])
3947
3948
3949
3950
                    kv_fp8 = cast_to_fp8(
                        kv_c, fp8_meta["scaling_fwd"], META_QKV, fp8_dtype_forward
                    ).view(kv.shape)
                    k_fp8, v_fp8 = _SplitAlongDim.apply(kv_fp8, dim, [1, 1])
3951
3952
                    k_fp8, v_fp8 = [x.squeeze(dim) for x in [k_fp8, v_fp8]]
                if qkv_group == 3:
3953
3954
3955
3956
3957
3958
3959
3960
3961
                    q_fp8 = cast_to_fp8(
                        q, fp8_meta["scaling_fwd"], META_QKV, fp8_dtype_forward
                    ).view(q.shape)
                    k_fp8 = cast_to_fp8(
                        k, fp8_meta["scaling_fwd"], META_QKV, fp8_dtype_forward
                    ).view(k.shape)
                    v_fp8 = cast_to_fp8(
                        v, fp8_meta["scaling_fwd"], META_QKV, fp8_dtype_forward
                    ).view(v.shape)
3962
            out_fp8, aux_ctx_tensors = fused_attn_fwd(
3963
3964
3965
3966
3967
3968
3969
3970
3971
3972
3973
                is_training,
                max_seqlen_q,
                max_seqlen_kv,
                cu_seqlens_q,
                cu_seqlens_kv,
                q_fp8,
                k_fp8,
                v_fp8,
                fp8_dtype_forward,
                fused_attention_backend,
                attn_bias,
3974
3975
                cu_seqlens_q_padded,
                cu_seqlens_kv_padded,
3976
3977
3978
3979
3980
3981
                fp8_meta["scaling_fwd"].scale_inv[META_QKV],
                fp8_meta["scaling_fwd"].scale_inv[META_S],
                fp8_meta["scaling_fwd"].scale[META_S],
                fp8_meta["scaling_fwd"].scale[META_O],
                fp8_meta["scaling_fwd"].amax_history[0][META_S],
                fp8_meta["scaling_fwd"].amax_history[0][META_O],
3982
3983
3984
3985
3986
3987
3988
3989
                attn_scale,
                dropout_p,
                fast_zero_fill,
                qkv_layout,
                attn_bias_type,
                attn_mask_type,
                rng_gen,
            )
3990
            if fp8_meta["recipe"].fp8_mha:
3991
3992
                out_ret = Float8Tensor(
                    data=out_fp8,
3993
3994
3995
3996
3997
3998
3999
4000
4001
                    fp8_meta=fp8_meta,
                    fp8_meta_forward=True,
                    fp8_meta_index=META_O,
                    fp8_dtype=fp8_dtype_forward,
                    dtype=q.dtype,
                )
            else:
                out_ret = cast_from_fp8(
                    out_fp8.view(-1, out_fp8.shape[-2] * out_fp8.shape[-1]),
4002
4003
4004
4005
4006
                    fp8_meta["scaling_fwd"],
                    META_O,
                    fp8_dtype_forward,
                    qkv_dtype,
                ).view(out_fp8.shape)
4007
4008
4009
4010
            out_save = out_ret

            if fp8_meta["recipe"].fp8_mha and not int(os.getenv("NVTE_FP8_DPA_BWD", "1")):
                # 1: qkv packed, 2: kv packed, 3: qkv separate
4011
                qkv_group = len(qkv_layout.split("_"))
4012
                if qkv_group == 1:
4013
4014
                    dim = qkv_layout.find("3")
                    qkv = _combine_tensors([q, k, v], dim)
4015
                    qkv_c = qkv.view(-1, qkv.shape[-3] * qkv.shape[-2] * qkv.shape[-1])
4016
4017
                    qkv_no_fp8 = cast_from_fp8(
                        qkv_c._data,
4018
                        fp8_meta["scaling_fwd"],
4019
4020
4021
4022
4023
                        META_QKV,
                        fp8_dtype_forward,
                        TE_DType[qkv.dtype],
                    ).view(qkv.shape)
                    q, k, v = _SplitAlongDim.apply(qkv_no_fp8, dim, [1, 1, 1])
4024
4025
                    q, k, v = [x.squeeze(dim) for x in [q, k, v]]
                if qkv_group == 2:
4026
4027
                    q = cast_from_fp8(
                        q._data,
4028
                        fp8_meta["scaling_fwd"],
4029
4030
4031
4032
4033
4034
                        META_QKV,
                        fp8_dtype_forward,
                        TE_DType[q.dtype],
                    ).view(q.shape)
                    dim = qkv_layout.split("_")[1].find("2")
                    kv = _combine_tensors([k, v], dim)
4035
                    kv_c = kv.view(-1, kv.shape[-3] * kv.shape[-2] * kv.shape[-1])
4036
4037
                    kv_no_fp8 = cast_from_fp8(
                        kv_c._data,
4038
                        fp8_meta["scaling_fwd"],
4039
4040
4041
4042
4043
                        META_QKV,
                        fp8_dtype_forward,
                        TE_DType[kv.dtype],
                    ).view(kv.shape)
                    k, v = _SplitAlongDim.apply(kv_no_fp8, dim, [1, 1])
4044
4045
                    k, v = [x.squeeze(dim) for x in [k, v]]
                if qkv_group == 3:
4046
4047
                    q = cast_from_fp8(
                        q._data,
4048
                        fp8_meta["scaling_fwd"],
4049
4050
4051
4052
4053
4054
                        META_QKV,
                        fp8_dtype_forward,
                        TE_DType[q.dtype],
                    ).view(q.shape)
                    k = cast_from_fp8(
                        k._data,
4055
                        fp8_meta["scaling_fwd"],
4056
4057
4058
4059
4060
4061
                        META_QKV,
                        fp8_dtype_forward,
                        TE_DType[k.dtype],
                    ).view(k.shape)
                    v = cast_from_fp8(
                        v._data,
4062
                        fp8_meta["scaling_fwd"],
4063
4064
4065
4066
                        META_QKV,
                        fp8_dtype_forward,
                        TE_DType[v.dtype],
                    ).view(v.shape)
4067
4068
                out_save = cast_from_fp8(
                    out_fp8.view(-1, out_fp8.shape[-2] * out_fp8.shape[-1]),
4069
4070
4071
4072
4073
4074
4075
4076
4077
4078
4079
                    fp8_meta["scaling_fwd"],
                    META_O,
                    fp8_dtype_forward,
                    qkv_dtype,
                ).view(out_fp8.shape)

            fp8_tensors = (
                q_fp8,
                k_fp8,
                v_fp8,
                out_fp8,
4080
                fp8_meta["scaling_fwd"].scale.clone(),
4081
4082
                fp8_meta["scaling_fwd"].scale_inv.clone(),
            )
4083
        else:
4084
            logger.debug("Running forward in %s", q.dtype)
4085
            out_ret, aux_ctx_tensors = fused_attn_fwd(
4086
4087
4088
4089
4090
4091
4092
4093
4094
4095
4096
                is_training,
                max_seqlen_q,
                max_seqlen_kv,
                cu_seqlens_q,
                cu_seqlens_kv,
                q,
                k,
                v,
                qkv_dtype,
                fused_attention_backend,
                attn_bias,
4097
4098
                cu_seqlens_q_padded,
                cu_seqlens_kv_padded,
4099
4100
4101
4102
4103
4104
4105
4106
4107
4108
4109
4110
4111
4112
                None,
                None,
                None,
                None,
                None,
                None,
                attn_scale,
                dropout_p,
                fast_zero_fill,
                qkv_layout,
                attn_bias_type,
                attn_mask_type,
                rng_gen,
            )
4113
4114
            out_save = out_ret
            fp8_tensors = (None, None, None, None, None, None)
4115

4116
        from .cpu_offload import CPUOffloadEnabled
4117

4118
        if CPUOffloadEnabled:
4119
            tensor_list = [q, k, v, out_save, cu_seqlens_q, cu_seqlens_kv]
4120
            qkv_layout = "sbhd_sbhd_sbhd"
4121
4122
4123
4124
            for tensor in tensor_list:
                if tensor is not None:
                    tensor.activation_offloading = True

4125
4126
        ctx.fp8 = fp8 and int(os.getenv("NVTE_FP8_DPA_BWD", "1"))
        qkvo_tensors = (q, k, v, out_save) if not ctx.fp8 else (None, None, None, None)
4127
4128
4129
4130
        ctx.save_for_backward(
            *qkvo_tensors,
            cu_seqlens_q,
            cu_seqlens_kv,
4131
4132
            cu_seqlens_q_padded,
            cu_seqlens_kv_padded,
4133
4134
4135
            *fp8_tensors,
            *aux_ctx_tensors,
        )
4136
        ctx.fp8_meta = fp8_meta
4137
4138
4139
4140
4141
4142
4143
4144
4145
        ctx.max_seqlen_q = max_seqlen_q
        ctx.max_seqlen_kv = max_seqlen_kv
        ctx.qkv_dtype = qkv_dtype
        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
4146
        ctx.fused_attention_backend = (
4147
            fused_attention_backend if ctx.fp8 else FusedAttnBackend["F16_arbitrary_seqlen"]
4148
        )
4149
4150
        ctx.use_FAv2_bwd = use_FAv2_bwd

4151
        return out_ret
4152
4153
4154

    @staticmethod
    def backward(ctx, d_out):
4155
        logger = logging.getLogger("FusedAttnFunc")
4156
        if ctx.fp8_meta["recipe"].fp8_mha:
4157
4158
4159
            assert isinstance(
                d_out, Float8Tensor
            ), "Gradient of the DPA output must be in Float8Tensor type for FP8 MHA."
4160
4161
4162
            d_out_f8tensor = d_out
            d_out = d_out._data

4163
        d_out = d_out.contiguous()
4164
4165
4166
4167
4168
4169
4170
        (
            q,
            k,
            v,
            out,
            cu_seqlens_q,
            cu_seqlens_kv,
4171
4172
            cu_seqlens_q_padded,
            cu_seqlens_kv_padded,
4173
4174
4175
4176
4177
4178
4179
4180
            q_fp8,
            k_fp8,
            v_fp8,
            out_fp8,
            fwd_scales,
            fwd_scale_invs,
            *aux_ctx_tensors,
        ) = ctx.saved_tensors
4181
4182
        if not aux_ctx_tensors[0].is_contiguous():
            aux_ctx_tensors[0] = aux_ctx_tensors[0].contiguous()
4183
        if ctx.use_FAv2_bwd:
4184
            softmax_lse, rng_state = aux_ctx_tensors
4185
4186
4187
4188
            dq = torch.empty_like(q)
            dk = torch.empty_like(k)
            dv = torch.empty_like(v)
            maybe_contiguous = lambda x: x.contiguous() if x.stride(-1) != 1 else x
4189
            d_out, q, k, v, out = [maybe_contiguous(x) for x in (d_out, q, k, v, out)]
4190
            flash_attn_cuda_bwd(
4191
4192
4193
4194
4195
4196
4197
4198
4199
4200
4201
4202
4203
4204
4205
4206
4207
4208
4209
                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,
4210
            )
4211
4212
4213
            dq = dq[..., : d_out.shape[-1]]
            dk = dk[..., : d_out.shape[-1]]
            dv = dv[..., : d_out.shape[-1]]
4214
        else:
4215
4216
            with torch.cuda.nvtx.range("_FusedAttn"):
                if ctx.fp8:
4217
                    logger.debug("Running backward in FP8")
4218
4219
                    fp8_dtype_forward = get_fp8_te_dtype(ctx.fp8_meta["recipe"], fprop_tensor=True)
                    fp8_dtype_backward = get_fp8_te_dtype(
4220
4221
                        ctx.fp8_meta["recipe"], fprop_tensor=False
                    )
4222
4223
                    if ctx.fp8_meta["recipe"].fp8_mha:
                        d_out_fp8 = d_out
4224
                        ctx.fp8_meta["scaling_bwd"].scale_inv[META_DO] = d_out_f8tensor._scale_inv
4225
4226
4227
                    else:
                        d_out_fp8 = cast_to_fp8(
                            d_out.view(-1, d_out.shape[-2] * d_out.shape[-1]),
4228
4229
4230
4231
                            ctx.fp8_meta["scaling_bwd"],
                            META_DO,
                            fp8_dtype_backward,
                        ).view(d_out.shape)
4232
                    dq_fp8, dk_fp8, dv_fp8, *rest = fused_attn_bwd(
4233
4234
4235
4236
4237
4238
4239
4240
4241
4242
4243
4244
                        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,
                        fp8_dtype_forward,
                        fp8_dtype_backward,
                        aux_ctx_tensors,
4245
                        ctx.fused_attention_backend,
4246
4247
                        cu_seqlens_q_padded,
                        cu_seqlens_kv_padded,
4248
4249
4250
4251
4252
4253
4254
4255
4256
4257
4258
4259
4260
4261
4262
4263
4264
                        fwd_scale_invs[META_QKV],  # d_scale_qkv,
                        fwd_scale_invs[META_S],  # d_scale_s,
                        fwd_scale_invs[META_O],  # d_scale_o,
                        ctx.fp8_meta["scaling_bwd"].scale_inv[META_DO],  # d_scale_do
                        ctx.fp8_meta["scaling_bwd"].scale_inv[META_DP],  # d_scale_dp
                        fwd_scales[META_S],  # q_scale_s
                        ctx.fp8_meta["scaling_bwd"].scale[META_DP],  # q_scale_dp
                        ctx.fp8_meta["scaling_bwd"].scale[META_DQKV],  # q_scale_dqkv
                        ctx.fp8_meta["scaling_bwd"].amax_history[0][META_DP],  # amax_dp
                        ctx.fp8_meta["scaling_bwd"].amax_history[0][META_DQKV],  # amax_dqkv
                        ctx.attn_scale,
                        ctx.dropout_p,
                        ctx.fast_zero_fill,
                        ctx.qkv_layout,
                        ctx.attn_bias_type,
                        ctx.attn_mask_type,
                    )
4265

4266
                    if ctx.fp8_meta["recipe"].fp8_mha:
4267
4268
                        dq = Float8Tensor(
                            data=dq_fp8,
4269
4270
4271
4272
4273
                            fp8_meta=ctx.fp8_meta,
                            fp8_meta_forward=False,
                            fp8_meta_index=META_DQKV,
                            fp8_dtype=fp8_dtype_backward,
                            dtype=d_out_f8tensor.dtype,
4274
4275
4276
                        )
                        dk = Float8Tensor(
                            data=dk_fp8,
4277
4278
4279
4280
4281
                            fp8_meta=ctx.fp8_meta,
                            fp8_meta_forward=False,
                            fp8_meta_index=META_DQKV,
                            fp8_dtype=fp8_dtype_backward,
                            dtype=d_out_f8tensor.dtype,
4282
4283
4284
                        )
                        dv = Float8Tensor(
                            data=dv_fp8,
4285
4286
4287
4288
4289
                            fp8_meta=ctx.fp8_meta,
                            fp8_meta_forward=False,
                            fp8_meta_index=META_DQKV,
                            fp8_dtype=fp8_dtype_backward,
                            dtype=d_out_f8tensor.dtype,
4290
                        )
4291
                    else:
4292
                        qkv_group = len(ctx.qkv_layout.split("_"))
4293
                        if qkv_group == 1:
4294
4295
4296
4297
4298
4299
4300
4301
4302
4303
4304
4305
4306
                            dim = ctx.qkv_layout.find("3")
                            dqkv_fp8 = _combine_tensors([dq_fp8, dk_fp8, dv_fp8], dim)
                            dqkv_c_fp8 = dqkv_fp8.view(
                                -1, dqkv_fp8.shape[-3] * dqkv_fp8.shape[-2] * dqkv_fp8.shape[-1]
                            )
                            dqkv = cast_from_fp8(
                                dqkv_c_fp8,
                                ctx.fp8_meta["scaling_bwd"],
                                META_DQKV,
                                fp8_dtype_backward,
                                ctx.qkv_dtype,
                            ).view(dqkv_fp8.shape)
                            dq, dk, dv = _SplitAlongDim.apply(dqkv, dim, [1, 1, 1])
4307
4308
4309
4310
                            dq, dk, dv = [x.squeeze(dim) for x in [dq, dk, dv]]
                        if qkv_group == 2:
                            dq = cast_from_fp8(
                                dq_fp8.view(-1, dq_fp8.shape[-2] * dq_fp8.shape[-1]),
4311
4312
4313
4314
4315
4316
4317
4318
4319
4320
4321
4322
4323
4324
4325
4326
4327
4328
                                ctx.fp8_meta["scaling_bwd"],
                                META_DQKV,
                                fp8_dtype_backward,
                                ctx.qkv_dtype,
                            ).view(dq_fp8.shape)
                            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]
                            )
                            dkv = cast_from_fp8(
                                dkv_c_fp8,
                                ctx.fp8_meta["scaling_bwd"],
                                META_DQKV,
                                fp8_dtype_backward,
                                ctx.qkv_dtype,
                            ).view(dkv_fp8.shape)
                            dk, dv = _SplitAlongDim.apply(dkv, dim, [1, 1])
4329
4330
4331
4332
                            dk, dv = [x.squeeze(dim) for x in [dk, dv]]
                        if qkv_group == 3:
                            dq = cast_from_fp8(
                                dq_fp8.view(-1, dq_fp8.shape[-2] * dq_fp8.shape[-1]),
4333
4334
4335
4336
4337
                                ctx.fp8_meta["scaling_bwd"],
                                META_DQKV,
                                fp8_dtype_backward,
                                ctx.qkv_dtype,
                            ).view(dq_fp8.shape)
4338
4339
                            dk = cast_from_fp8(
                                dk_fp8.view(-1, dk_fp8.shape[-2] * dk_fp8.shape[-1]),
4340
4341
4342
4343
4344
                                ctx.fp8_meta["scaling_bwd"],
                                META_DQKV,
                                fp8_dtype_backward,
                                ctx.qkv_dtype,
                            ).view(dk_fp8.shape)
4345
4346
                            dv = cast_from_fp8(
                                dv_fp8.view(-1, dv_fp8.shape[-2] * dv_fp8.shape[-1]),
4347
4348
4349
4350
4351
                                ctx.fp8_meta["scaling_bwd"],
                                META_DQKV,
                                fp8_dtype_backward,
                                ctx.qkv_dtype,
                            ).view(dv_fp8.shape)
4352
                else:
4353
                    logger.debug("Running backward in %s", q.dtype)
4354
4355
4356
                    if d_out.dtype == torch.uint8:
                        d_out = d_out_f8tensor.from_float8(q.dtype)
                    dq, dk, dv, *rest = fused_attn_bwd(
4357
4358
4359
4360
4361
4362
4363
4364
4365
4366
4367
4368
                        ctx.max_seqlen_q,
                        ctx.max_seqlen_kv,
                        cu_seqlens_q,
                        cu_seqlens_kv,
                        q,
                        k,
                        v,
                        out,
                        d_out,
                        ctx.qkv_dtype,
                        ctx.qkv_dtype,
                        aux_ctx_tensors,
4369
                        ctx.fused_attention_backend,
4370
4371
                        cu_seqlens_q_padded,
                        cu_seqlens_kv_padded,
4372
4373
4374
4375
4376
4377
4378
4379
4380
4381
4382
4383
4384
4385
4386
4387
4388
                        None,
                        None,
                        None,
                        None,
                        None,
                        None,
                        None,
                        None,
                        None,
                        None,
                        ctx.attn_scale,
                        ctx.dropout_p,
                        ctx.fast_zero_fill,
                        ctx.qkv_layout,
                        ctx.attn_bias_type,
                        ctx.attn_mask_type,
                    )
4389

4390
4391
        # if no_bias or alibi, return dqkv
        if ctx.attn_bias_type in ["no_bias", "alibi"]:
4392
4393
4394
4395
4396
4397
4398
4399
4400
4401
4402
4403
4404
4405
4406
4407
4408
4409
4410
4411
4412
4413
4414
4415
4416
4417
4418
            return (
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                dq,
                dk,
                dv,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
            )
4419
        # else, return (dqkv, dbias)
4420
4421
4422
4423
4424
4425
4426
4427
4428
4429
4430
4431
4432
4433
4434
4435
4436
4437
4438
4439
4440
4441
4442
4443
4444
4445
4446
        return (
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            dq,
            dk,
            dv,
            None,
            rest[0],
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
        )
4447

4448

4449
class FusedAttention(torch.nn.Module):
4450
4451
4452
4453
4454
4455
4456
4457
4458
    """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:

4459
4460
4461
4462
    | backend       | 1                       | 2                              |
    | flash based   | no                      | yes                            |
    | cuDNN based   | yes                     | yes                            |
    | qkv dtype     | fp16/bf16               | fp16/bf16                      |
4463
    | attn_type     | self/cross              | self/cross                     |
4464
    | qkv_layout    |                         |                                |
4465
    |  - (q,k,v)    | sb3hd, bs3hd            | sb3hd, bs3hd, sbh3d, bsh3d     |
4466
    |               | sbhd_sb2hd, bshd_bs2hd  | sbhd_sb2hd, bshd_bs2hd         |
4467
4468
    |               | bshd_bshd_bshd          | sbhd_sbh2d, bshd_bsh2d         |
    |               |                         | sbhd_sbhd_sbhd, bshd_bshd_bshd |
4469
4470
    | mask_type     | causal/padding/no_mask  | causal/padding/no_mask         |
    | bias_type     | post_scale_bias/no_bias | post_scale_bias/alibi/no_bias  |
4471
    | dropout       | yes                     | yes                            |
4472
4473
    | max_seqlen    | <=512, multiple of 64   | any, multiple of 64            |
    | head_dim      | 64                      | <=128, multiple of 8           |
4474
    | output dtype  | fp16/bf16               | fp16/bf16                      |
4475
4476
4477
4478
    """

    def __init__(
        self,
4479
        softmax_scale: float,
4480
4481
4482
        attention_dropout: float = 0.0,
        attention_dropout_ctx: Optional[Callable] = nullcontext,
        attention_type: str = "self",
4483
4484
        layer_number: Optional[int] = None,
        deterministic: bool = False,
4485
4486
4487
    ) -> None:
        super().__init__()

4488
        self.logger = logging.getLogger("FusedAttention")
4489
        self.softmax_scale = softmax_scale
4490
4491
4492
        self.attention_dropout = attention_dropout
        self.attention_dropout_ctx = attention_dropout_ctx
        self.attention_type = attention_type
4493
4494
4495
        self.use_FAv2_bwd = os.getenv(
            "NVTE_FUSED_ATTN_USE_FAv2_BWD", "0"
        ) == "1" and get_device_compute_capability() == (9, 0)
4496
4497
4498
4499
4500
4501
4502
4503
4504
4505
4506
4507
4508
4509
4510
4511
        self.layer_number = 1 if layer_number is None else layer_number
        if deterministic:
            # workspace optimization path is deterministic
            os.environ["CUDNN_FRONTEND_ATTN_DP_WORKSPACE_LIMIT"] = "-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"
4512

4513
        def remove_extra_states_check(self, incompatible_keys):  # pylint: disable=unused-argument
4514
4515
            """
            Temporarily remove fused_attention._extra_state as a missing key
4516
4517
4518
4519
            or an unexpected key when loading TransformerEngine checkpoints.
            Please store FP8 metadata as DotProductAttention's _extra_state,
            rather than FusedAttention's _extra_state. This hook will be
            phased out in TransformerEngine 2.0.
4520
4521
            """
            for key in incompatible_keys.missing_keys:
4522
                if "fused_attention._extra_state" in key:
4523
                    incompatible_keys.missing_keys.remove(key)
4524
4525
4526
4527
4528
4529
4530
            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."
                    )
4531

4532
4533
        self.register_load_state_dict_post_hook(remove_extra_states_check)

4534
    @no_torch_dynamo()
4535
4536
4537
4538
4539
    def forward(
        self,
        query_layer: torch.Tensor,
        key_layer: torch.Tensor,
        value_layer: torch.Tensor,
4540
4541
4542
        qkv_layout: str = "sbh3d",
        cu_seqlens_q: Optional[torch.Tensor] = None,
        cu_seqlens_kv: Optional[torch.Tensor] = None,
4543
4544
        cu_seqlens_q_padded: Optional[torch.Tensor] = None,
        cu_seqlens_kv_padded: Optional[torch.Tensor] = None,
4545
4546
        max_seqlen_q: Optional[int] = None,
        max_seqlen_kv: Optional[int] = None,
4547
        attn_mask_type: str = "causal",
4548
        attention_mask: Optional[Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]] = None,
4549
        fused_attention_backend: tex.NVTE_Fused_Attn_Backend = tex.NVTE_Fused_Attn_Backend.NVTE_No_Backend,
4550
4551
4552
        core_attention_bias_type: str = "no_bias",
        core_attention_bias: Optional[torch.Tensor] = None,
        fast_zero_fill: bool = True,
4553
4554
4555
        cp_group: Optional[dist_group_type] = None,
        cp_global_ranks: List[int] = None,
        cp_stream: torch.cuda.Stream = None,
4556
4557
        fp8: bool = False,
        fp8_meta: Optional[Dict[str, Any]] = None,
4558
4559
    ) -> torch.Tensor:
        """fused attention fprop"""
4560
4561
4562
        assert (
            fused_attention_backend != tex.NVTE_Fused_Attn_Backend.NVTE_No_Backend
        ), "No fused attention backend supports this input combination!"
4563
        assert (
4564
4565
4566
            (query_layer.dtype in [torch.float16, torch.bfloat16, torch.uint8])
            and (key_layer.dtype in [torch.float16, torch.bfloat16, torch.uint8])
            and (value_layer.dtype in [torch.float16, torch.bfloat16, torch.uint8])
4567
        ), "FusedAttention only supports FP16 and BF16 data types."
4568
4569
        assert (
            query_layer.is_cuda and key_layer.is_cuda and value_layer.is_cuda
4570
        ), "FusedAttention only supports CUDA tensors."
4571
4572
        assert (
            qkv_layout in QKVLayouts
4573
        ), f"FusedAttention does not support qkv_layout = {qkv_layout}!"
4574

4575
4576
        context_parallel = (cp_group is not None) and (get_distributed_world_size(cp_group) != 1)

4577
        qkv_format = "".join([i for i in qkv_layout.split("_")[0] if i.isalpha()])
4578

4579
4580
        if qkv_format in ["sbhd", "bshd"]:
            if qkv_format == "sbhd":
4581
                batch_size, max_seqlen_q, max_seqlen_kv = (
4582
4583
4584
4585
4586
                    query_layer.shape[1],
                    query_layer.shape[0],
                    key_layer.shape[0],
                )
            if qkv_format == "bshd":
4587
                batch_size, max_seqlen_q, max_seqlen_kv = (
4588
4589
4590
4591
4592
                    query_layer.shape[0],
                    query_layer.shape[1],
                    key_layer.shape[1],
                )
            if "padding" in attn_mask_type:
4593
4594
                assert not context_parallel, "Padding mask not supported with context parallelism!"

4595
4596
4597
4598
4599
                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!"
                        )
4600
                    if self.attention_type == "self":
4601
4602
                        cu_seqlens_q = get_cu_seqlens(attention_mask)
                        cu_seqlens_kv = cu_seqlens_q
4603
                    else:
4604
4605
                        cu_seqlens_q = get_cu_seqlens(attention_mask[0])
                        cu_seqlens_kv = get_cu_seqlens(attention_mask[1])
4606
            else:
4607
4608
4609
4610
4611
4612
4613
4614
4615
4616
4617
4618
                if cu_seqlens_q is None:
                    cu_seqlens_q = _get_full_cu_seqlens(
                        batch_size,
                        max_seqlen_q,
                        query_layer.device,
                    )
                if cu_seqlens_kv is None:
                    cu_seqlens_kv = _get_full_cu_seqlens(
                        batch_size,
                        max_seqlen_kv,
                        key_layer.device,
                    )
4619
4620
4621
        if qkv_format == "thd":
            assert (
                max_seqlen_q is not None
4622
4623
4624
                and max_seqlen_kv is not None
                and cu_seqlens_q is not None
                and cu_seqlens_kv is not None
4625
            ), "max_seqlen_q/kv and cu_seqlens_q/kv can not be None when qkv_format is thd!"
4626
4627
4628
4629

        if cu_seqlens_q_padded is None or cu_seqlens_kv_padded is None:
            cu_seqlens_q_padded = cu_seqlens_q
            cu_seqlens_kv_padded = cu_seqlens_kv
4630
4631
4632

        qkv_dtype = TE_DType[query_layer.dtype]

4633
4634
4635
4636
4637
        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)
        )
4638
4639

        if context_parallel:
4640
            assert (
4641
4642
4643
4644
4645
4646
4647
4648
                fused_attention_backend == tex.NVTE_Fused_Attn_Backend.NVTE_F16_arbitrary_seqlen
            ), 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)
            ]
4649
4650
4651
            with self.attention_dropout_ctx():
                output = attn_forward_func_with_cp(
                    self.training,
4652
4653
4654
4655
4656
4657
4658
                    query_layer,
                    key_layer,
                    value_layer,
                    cu_seqlens_q,
                    cu_seqlens_kv,
                    max_seqlen_q,
                    max_seqlen_kv,
4659
4660
                    cu_seqlens_q_padded,
                    cu_seqlens_kv_padded,
4661
                    self.attention_dropout if self.training else 0.0,
4662
4663
4664
                    cp_group,
                    cp_global_ranks,
                    cp_stream,
4665
                    softmax_scale=self.softmax_scale,
4666
                    qkv_format=qkv_format,
4667
                    attn_mask_type=attn_mask_type,
4668
4669
                    attn_bias_type=core_attention_bias_type,
                    attn_bias=core_attention_bias,
4670
4671
4672
                    use_fused_attention=True,
                )
        else:
4673
4674
4675
4676
4677
            with self.attention_dropout_ctx():
                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!"
4678
                    )
4679
4680
4681
4682
4683
4684
4685
4686
4687
                    assert (
                        fp8_meta is not None
                    ), "FP8 metadata fp8_meta is required for FP8 attention!"
                output = FusedAttnFunc.apply(
                    self.training,
                    max_seqlen_q,
                    max_seqlen_kv,
                    cu_seqlens_q,
                    cu_seqlens_kv,
4688
4689
                    cu_seqlens_q_padded,
                    cu_seqlens_kv_padded,
4690
4691
4692
4693
4694
4695
4696
4697
4698
4699
4700
4701
4702
4703
4704
4705
4706
                    query_layer,
                    key_layer,
                    value_layer,
                    qkv_dtype,
                    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,
                    None,  # rng_gen
                    fused_attention_backend,
                    use_FAv2_bwd,
                    fp8,
                    fp8_meta,
                )
4707

4708
4709
        # ...hd -> ...(hd)
        return output.view(*output.shape[:-2], -1)
4710
4711


4712
class DotProductAttention(TransformerEngineBaseModule):
4713
4714
4715
4716
4717
4718
    """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::

4719
        Argument :attr:`attention_mask` in the `forward` call is only used when
4720
        :attr:`attn_mask_type` includes '"padding"' or `"arbitrary"`.
4721
4722
4723

    .. warning::

4724
        FlashAttention uses a non-deterministic algorithm for optimal performance. To observe
4725
        deterministic behavior at the cost of performance, use FlashAttention version >= `2.4.1`
4726
4727
        and set the environment variable :attr:`NVTE_ALLOW_NONDETERMINISTIC_ALGO=0`. In order
        to disable`flash-attn` entirely, set :attr:`NVTE_FLASH_ATTN=0`.
4728
4729
4730
4731
4732
4733

    Parameters
    ----------
    num_attention_heads : int
                         number of attention heads in the transformer layer.
    kv_channels : int
4734
                number of key-query-value channels per attention head.
4735
4736
4737
4738
4739
4740
4741
4742
    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`.
4743
4744
    attention_dropout: float, default = 0.0
                      dropout probability for the dropout op during multi-head attention.
4745
    attn_mask_type: str, default = `causal`
4746
                   type of attention mask passed into softmax operation, options are "`no_mask`",
4747
4748
4749
4750
4751
4752
4753
4754
4755
4756
4757
4758
4759
4760
4761
4762
4763
4764
4765
4766
4767
4768
4769
4770
                   "`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
                   "`padding_causal`" and "`padding_causal_bottom_right`", TransformerEngine
                   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].
4771
4772
4773
4774
    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
4775
                window and "`causal`" mask specifically. Similar to :attr:`attn_mask_type`, it can
4776
                be overridden by :attr:`window_size` in `forward` as well.
4777
4778
    attention_type: str, default = `self`
                   type of attention, either "`self`" and "`cross`".
4779
4780
4781
    layer_number: int, default = `None`
                 layer number of the current `DotProductAttention` when multiple such modules
                 are concatenated, for instance in consecutive transformer blocks.
4782
4783
4784
4785
4786
4787
4788
4789
4790
    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,
               `h` the number of heads, `d` head size, and `t` the total number of sequences
               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.
4791
               For that, please use `get_qkv_layout` to gain the layout information.
4792
4793
4794
    softmax_scale: Optional[float], default = `None`
                softmax scale for the attention scores. If `None`, defaults to
                `1.0 / math.sqrt(kv_channels)`.
4795
4796
4797
4798
4799
4800
4801
4802
4803

    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.
4804
4805
4806
4807
4808
4809
4810
4811
4812
    cp_group : ProcessGroup, default = `None`
              context parallel process group.
    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.
4813
4814
4815
4816
4817
4818
    """

    def __init__(
        self,
        num_attention_heads: int,
        kv_channels: int,
4819
        num_gqa_groups: Optional[int] = None,
4820
        attention_dropout: float = 0.0,
4821
        qkv_format: str = "sbhd",
4822
        attn_mask_type: str = "causal",
4823
        window_size: Optional[Tuple[int, int]] = None,
4824
4825
4826
4827
4828
        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,
4829
        attention_type: str = "self",
4830
        cp_group: Optional[dist_group_type] = None,
4831
        cp_global_ranks: List[int] = None,
4832
        cp_stream: torch.cuda.Stream = None,
4833
        softmax_scale: Optional[float] = None,
4834
4835
4836
    ) -> None:
        super().__init__()

4837
        self.logger = logging.getLogger("DotProductAttention")
4838
        self.qkv_format = qkv_format
4839
        attn_mask_type = attn_mask_type.replace(",", "_")
4840
4841
        if attn_mask_type == "causal_padding":
            attn_mask_type = "padding_causal"
4842
        self.attn_mask_type = attn_mask_type
4843
        self.window_size = check_set_window_size(attn_mask_type, window_size)
4844
4845
4846
4847
4848
4849
4850
        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)
4851
        self.get_rng_state_tracker = get_rng_state_tracker
4852
        self.num_attention_heads = num_attention_heads
4853
        self.layer_number = 1 if layer_number is None else layer_number
4854
4855
4856
        self.cp_group = cp_group
        self.cp_global_ranks = cp_global_ranks
        self.cp_stream = cp_stream
4857

4858
        self.hidden_size_per_attention_head = kv_channels
4859

4860
        self.num_gqa_groups = num_attention_heads if num_gqa_groups is None else num_gqa_groups
4861
4862
        self.num_gqa_groups_per_partition = int(self.num_gqa_groups // tp_size)

4863
4864
4865
        assert (
            num_attention_heads % self.num_gqa_groups == 0
        ), "The number of attention heads must be divisible by the number of GQA groups!"
4866

4867
        self.rng_states_tracker = None
4868
4869
4870
        if sequence_parallel or get_rng_state_tracker is None:
            attention_dropout_ctx = nullcontext
        else:
4871
4872
4873
            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
4874

4875
4876
        if softmax_scale is None:
            softmax_scale = 1.0 / math.sqrt(kv_channels)
4877
4878

        self.device_compute_capability = get_device_compute_capability()
4879
4880
4881
        self.deterministic = (
            not bool(int(os.getenv("NVTE_ALLOW_NONDETERMINISTIC_ALGO", "1")))
            or torch.are_deterministic_algorithms_enabled()
4882
        )
4883
4884
4885
4886

        self.use_flash_attention = int(
            os.getenv("NVTE_FLASH_ATTN", "1")
        ) and self.device_compute_capability >= (8, 0)
4887
4888
4889
4890
4891
        if int(os.getenv("NVTE_FLASH_ATTN", "1")) == 0:
            self.logger.debug("Disabling FlashAttention due to NVTE_FLASH_ATTN=0")
        if self.device_compute_capability < (8, 0):
            self.logger.debug("Disabling FlashAttention for compute capability < sm80")

4892
        if not _flash_attn_2_4_1_plus and self.deterministic:
4893
            self.use_flash_attention = False
4894
            self.logger.warning(
4895
4896
4897
                "Disabling usage of FlashAttention since version <2.4.1 does not support "
                "deterministic execution. In order to use FA with deterministic behavior,"
                " please install FlashAttention version >=2.4.1."
4898
4899
            )

4900
4901
4902
        self.use_fused_attention = int(
            os.getenv("NVTE_FUSED_ATTN", "1")
        ) and self.device_compute_capability >= (8, 0)
4903
4904
4905
4906
        if int(os.getenv("NVTE_FUSED_ATTN", "1")) == 0:
            self.logger.debug("Disabling FusedAttention due to NVTE_FUSED_ATTN=0")
        if self.device_compute_capability < (8, 0):
            self.logger.debug("Disabling FusedAttention for compute capability < sm80")
4907

4908
        assert attention_type in AttnTypes, f"attention_type {attention_type} not supported"
4909
4910
4911
4912

        self.attention_type = attention_type
        self.attention_dropout = attention_dropout

4913
4914
4915
4916
4917
4918
        attn_kwargs = {
            "attention_dropout": attention_dropout,
            "attention_dropout_ctx": attention_dropout_ctx,
        }

        if self.use_flash_attention:
4919
4920
4921
4922
4923
4924
4925
            self.flash_attention = FlashAttention(
                softmax_scale,
                attention_type=attention_type,
                layer_number=layer_number,
                deterministic=self.deterministic,
                **attn_kwargs,
            )
4926

4927
        # Instantiating three types since use of flash-attn and FusedAttention
4928
        # might be ruled out due to forward inputs.
4929
        if self.use_fused_attention:
4930
4931
4932
4933
4934
4935
4936
            self.fused_attention = FusedAttention(
                softmax_scale,
                attention_type=attention_type,
                layer_number=layer_number,
                deterministic=self.deterministic,
                **attn_kwargs,
            )
4937

4938
        self.unfused_attention = UnfusedDotProductAttention(
4939
4940
            softmax_scale, **attn_kwargs, layer_number=layer_number
        )
4941

4942
4943
4944
4945
4946
4947
4948
4949
4950
4951
4952
4953
        def remove_extra_states_check(self, incompatible_keys):  # pylint: disable=unused-argument
            """
            Temporarily remove core_attention._extra_state as a missing key
            when loading older TransformerEngine checkpoints. Will phase out
            this hook in TransformerEngine 2.0.
            """
            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)

4954
4955
4956
4957
    def _checkpointed_attention_forward(
        self,
        attention_func: Callable,
        *forward_args: Tuple[torch.Tensor, ...],
4958
        **forward_kwargs: Dict[str, Any],
4959
4960
4961
    ) -> torch.Tensor:
        """Forward method with activation checkpointing."""

4962
4963
        def custom_forward(*input_args, **input_kwargs):
            return attention_func(*input_args, **input_kwargs)
4964
4965
4966

        hidden_states = checkpoint(
            custom_forward,
4967
4968
4969
            distribute_saved_activations=False,
            get_rng_state_tracker=self.get_rng_state_tracker,
            tp_group=self.tp_group,
4970
            *forward_args,
4971
            **forward_kwargs,
4972
4973
4974
4975
        )

        return hidden_states

4976
4977
4978
4979
4980
4981
    def set_context_parallel_group(
        self,
        cp_group: Union[dist_group_type, None],
        cp_global_ranks: List[int],
        cp_stream: torch.cuda.Stream,
    ) -> None:
4982
4983
4984
4985
4986
4987
4988
4989
4990
4991
4992
4993
4994
        """
        Set the context parallel attributes for the given
        module before executing the forward pass.

        Parameters
        ----------
        cp_group : ProcessGroup
                  context parallel process group.
        cp_global_ranks : List[int]
                         list of global ranks in the context group.
        cp_stream : torch.cuda.Stream
                   cuda stream for context parallel execution.
        """
4995
4996
4997
4998
        self.cp_group = cp_group
        self.cp_global_ranks = cp_global_ranks
        self.cp_stream = cp_stream

4999
    @no_torch_dynamo(recursive=False)
5000
5001
5002
5003
5004
    def forward(
        self,
        query_layer: torch.Tensor,
        key_layer: torch.Tensor,
        value_layer: torch.Tensor,
5005
        attention_mask: Optional[Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]] = None,
5006
5007
5008
        qkv_format: Optional[str] = None,
        cu_seqlens_q: Optional[torch.Tensor] = None,
        cu_seqlens_kv: Optional[torch.Tensor] = None,
5009
5010
        cu_seqlens_q_padded: Optional[torch.Tensor] = None,
        cu_seqlens_kv_padded: Optional[torch.Tensor] = None,
5011
5012
        max_seqlen_q: Optional[int] = None,
        max_seqlen_kv: Optional[int] = None,
5013
        attn_mask_type: Optional[str] = None,
5014
        window_size: Optional[Tuple[int, int]] = None,
5015
        checkpoint_core_attention: bool = False,
5016
5017
        core_attention_bias_type: str = "no_bias",
        core_attention_bias: Optional[torch.Tensor] = None,
5018
        alibi_slopes: Optional[torch.Tensor] = None,
5019
        fast_zero_fill: bool = True,
5020
        inference_params: Optional[InferenceParams] = None,
5021
        is_first_microbatch: Optional[bool] = None,
5022
5023
5024
5025
5026
5027
    ) -> torch.Tensor:
        """
        Dot Product Attention Layer.

        .. note::

5028
5029
            Argument :attr:`attention_mask` is only used when :attr:`attn_mask_type`
            includes '"padding"' or `"arbitrary"`.
5030
5031
5032

        .. note::

5033
5034
5035
            Input tensor :attr:`query_layer` must be of shape
            (:attr:`sequence_length`, :attr:`batch_size`, :attr:`num_attention_heads`,
            :attr:`kv_channels`) and the tensors :attr:`key_layer` and :attr:`value_layer`
5036
            must each be of shape (:attr:`sequence_length`, :attr:`batch_size`,
5037
            :attr:`num_gqa_groups`, :attr:`kv_channels`). Output of shape
5038
5039
5040
            (:attr:`sequence_length`, :attr:`batch_size`, :attr:`num_attention_heads`
            * :attr:`kv_channels`) is returned.

5041
5042
        .. note::

5043
5044
5045
5046
5047
5048
5049
5050
5051
5052
5053
5054
5055
5056
5057
5058
5059
5060
            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,
            and FusedAttention backend if applicable, to use. TransformerEngine prioritizes
            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
5061
5062
5063
5064
5065
            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
            optimizations in FusedAttention. When unset, TransformerEngine determines the code path
            based on its internal logic. These optimizations trade memory for performance
            and should be used with care.
5066

5067
5068
5069
5070
5071
5072
5073
5074
        Parameters
        ----------
        query_layer : torch.Tensor
                     Query tensor.
        key_layer : torch.Tensor
                   Key tensor.
        value_layer : torch.Tensor
                     Value tensor.
5075
5076
        attention_mask: Optional[Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]],
             default = `None`. Boolean tensor(s) used to mask out attention softmax input.
5077
             It should be `None` for causal masks and "`no_mask`". For padding masks, it should be
5078
5079
             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]
5080
5081
5082
5083
             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.
5084
5085
5086
        qkv_format: str, default = `None`
                   If provided, overrides :attr:`qkv_format` from initialization.
        cu_seqlens_q: Optional[torch.Tensor], default = `None`
5087
                   Cumulative sum of sequence lengths (without offset) in a batch for `query_layer`,
5088
5089
                   with shape [batch_size + 1] and dtype torch.int32.
        cu_seqlens_kv: Optional[torch.Tensor], default = `None`
5090
5091
5092
5093
5094
5095
5096
5097
5098
5099
5100
5101
                   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.
        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`.
        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`.
5102
5103
5104
5105
5106
5107
        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.
5108
5109
5110
5111
5112
5113
5114
        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.
5115
        window_size: Optional[Tuple[int, int]], default = `None`
5116
                    Sliding window size for local attention.
5117
5118
5119
5120
5121
        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.
5122
        core_attention_bias_type: str, default = `no_bias`
5123
                    Bias type, {`no_bias`, `pre_scale_bias`, `post_scale_bias`, `alibi`}
5124
        core_attention_bias: Optional[torch.Tensor], default = `None`
5125
5126
                    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.
5127
5128
5129
5130
        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.
5131
        fast_zero_fill: bool, default = `True`
5132
                    Whether to use the fast path to set output tensors to 0 or not.
5133
5134
5135
5136
5137
5138
5139
5140
5141
5142
        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.
5143
5144
5145
5146
5147
5148
5149
5150
5151
5152
5153
5154
5155
        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)
5156
        """
5157
5158
5159
5160
5161
5162
5163
5164
5165
5166
5167
5168
5169
5170
5171
5172
5173
5174
5175
5176
5177
5178
5179
5180
        with self.prepare_forward(
            query_layer,
            is_first_microbatch,
            num_gemms=3,
            allow_non_contiguous=True,
        ) as query_layer:

            if self.fp8:
                forced_fp8_dpa = ""
                if self.fp8_meta["recipe"].fp8_mha:
                    if not self.fp8_meta["recipe"].fp8_dpa:
                        self.fp8_meta["recipe"].fp8_dpa = True
                        forced_fp8_dpa = " (forced)"

            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."""
5181

5182
5183
5184
            assert (
                query_layer.is_cuda and key_layer.is_cuda and value_layer.is_cuda
            ), "DotProductAttention only supports CUDA tensors."
5185
5186
5187
            assert (
                query_layer.dtype == key_layer.dtype and query_layer.dtype == value_layer.dtype
            ), "Queries, keys and values must have the same data type!"
5188
            assert key_layer.shape == value_layer.shape, "Keys and values must have the same shape!"
5189

5190
5191
5192
5193
5194
5195
            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"
5196
            assert (
5197
5198
5199
5200
5201
5202
                attn_mask_type in AttnMaskTypes
            ), f"Attention mask type {attn_mask_type} is not supported!"
            if qkv_format == "thd":
                assert (
                    "padding" in attn_mask_type
                ), "Attention mask type must be padding or padding_causal for qkv_format=thd!"
5203

5204
5205
5206
5207
            if window_size is None:
                window_size = self.window_size
            window_size = check_set_window_size(attn_mask_type, window_size)

5208
5209
5210
5211
5212
5213
5214
            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."
5215

5216
5217
            if qkv_format is None:
                qkv_format = self.qkv_format
5218

5219
5220
            if inference_params is not None:
                assert self.layer_number is not None, "Layer number must be set!"
5221

5222
5223
5224
                if qkv_format == "bshd":
                    key_layer = key_layer.transpose(0, 1)
                    value_layer = value_layer.transpose(0, 1)
5225

5226
5227
5228
5229
                (
                    inference_key_memory,
                    inference_value_memory,
                ) = inference_params.key_value_memory_dict[self.layer_number]
5230

5231
5232
5233
                batch_start = inference_params.batch_size_offset
                batch_end = batch_start + key_layer.size(1)
                assert batch_end <= inference_key_memory.size(1)
5234

5235
5236
5237
                sequence_start = inference_params.sequence_len_offset
                sequence_end = sequence_start + key_layer.size(0)
                assert sequence_end <= inference_key_memory.size(0)
5238

5239
5240
5241
5242
5243
5244
5245
5246
5247
                # Copy keys and values into KV-cache
                inference_key_memory[sequence_start:sequence_end, batch_start:batch_end, ...] = (
                    key_layer
                )
                inference_value_memory[sequence_start:sequence_end, batch_start:batch_end, ...] = (
                    value_layer
                )
                key_layer = inference_key_memory[:sequence_end, batch_start:batch_end, ...]
                value_layer = inference_value_memory[:sequence_end, batch_start:batch_end, ...]
5248

5249
5250
5251
                if qkv_format == "bshd":
                    key_layer = key_layer.transpose(0, 1)
                    value_layer = value_layer.transpose(0, 1)
5252

5253
5254
                key_layer = key_layer.contiguous()
                value_layer = value_layer.contiguous()
5255
5256

            assert (
5257
5258
5259
5260
5261
5262
5263
5264
5265
5266
                key_layer.shape[-2] == self.num_gqa_groups_per_partition
                and value_layer.shape[-2] == self.num_gqa_groups_per_partition
            ), f"Keys and values must have num_gqa_group = {self.num_gqa_groups} heads!"
            assert qkv_format in [
                "sbhd",
                "bshd",
                "thd",
            ], "DotProductAttention only supports qkv_format = {'sbhd', 'bshd', 'thd'}!"

            if qkv_format == "thd":
5267
                assert all(
5268
5269
5270
5271
5272
5273
5274
5275
5276
5277
5278
5279
5280
5281
5282
5283
5284
5285
5286
                    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!"
                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!"
                if max_seqlen_q is None:
                    seqlens_q = cu_seqlens_q[1:] - cu_seqlens_q[:-1]
                    max_seqlen_q = pow(2, math.ceil(math.log2(seqlens_q.max().item())))
                if max_seqlen_kv is None:
                    seqlens_kv = cu_seqlens_kv[1:] - cu_seqlens_kv[:-1]
                    max_seqlen_kv = pow(2, math.ceil(math.log2(seqlens_kv.max().item())))
5287
                batch_size = len(cu_seqlens_q) - 1
5288
5289

            if qkv_format in ["sbhd", "bshd"]:
5290
                assert all(
5291
5292
5293
5294
                    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 = {qkv_format}!"
                if qkv_format == "sbhd":
                    max_seqlen_q, max_seqlen_kv = (query_layer.shape[0], key_layer.shape[0])
5295
                    batch_size = query_layer.shape[1]
5296
5297
                if qkv_format == "bshd":
                    max_seqlen_q, max_seqlen_kv = (query_layer.shape[1], key_layer.shape[1])
5298
                    batch_size = query_layer.shape[0]
5299
5300
5301
5302
5303
5304
5305
5306
5307
5308
5309
5310
                if cu_seqlens_q is not None:
                    seqlens_q = cu_seqlens_q[1:] - cu_seqlens_q[:-1]
                    assert all(
                        seqlens_q <= max_seqlen_q
                    ), """Sequence lengths indicated by cu_seqlens_q must be no greater than
                        the sequence dimention in 'query_layer'!"""
                if cu_seqlens_kv is not None:
                    seqlens_kv = cu_seqlens_kv[1:] - cu_seqlens_kv[:-1]
                    assert all(
                        seqlens_kv <= max_seqlen_kv
                    ), """Sequence lengths indicated by cu_seqlens_kv must be no greater than
                        the sequence dimention in 'key_layer' and 'value_layer'!"""
5311

5312
5313
5314
5315
5316
            if (
                isinstance(query_layer, Float8Tensor)
                and isinstance(key_layer, Float8Tensor)
                and isinstance(value_layer, Float8Tensor)
            ):
5317
                qkv_layout, query_layer._data, key_layer._data, value_layer._data = get_qkv_layout(
5318
5319
5320
                    query_layer._data, key_layer._data, value_layer._data, qkv_format=qkv_format
                )
            else:
5321
                qkv_layout, query_layer, key_layer, value_layer = get_qkv_layout(
5322
5323
                    query_layer, key_layer, value_layer, qkv_format=qkv_format
                )
5324

5325
5326
5327
5328
5329
5330
5331
5332
5333
5334
5335
5336
5337
5338
5339
5340
5341
5342
5343
5344
5345
            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
            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
                    or _alibi_cache["_alibi_slopes"] is None
                ):
                    _alibi_cache["_alibi_slopes_require_update"] = True
                    _alibi_cache["_alibi_bias_require_update"] = True

5346
5347
5348
5349
            deterministic = (
                not bool(int(os.getenv("NVTE_ALLOW_NONDETERMINISTIC_ALGO", "1")))
                or torch.are_deterministic_algorithms_enabled()
            )
5350

5351
5352
5353
            context_parallel = (
                self.cp_group is not None and get_distributed_world_size(self.cp_group) != 1
            )
5354

5355
5356
            core_attention_bias_shape = None
            if core_attention_bias is not None:
5357
                if (
5358
5359
                    core_attention_bias.shape[0] == batch_size
                    and core_attention_bias.shape[1] == query_layer.shape[-2]
5360
                ):
5361
5362
5363
5364
5365
5366
5367
5368
5369
5370
5371
5372
5373
5374
5375
5376
5377
5378
5379
5380
5381
5382
5383
5384
                    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"

            pad_between_seqs = (
                cu_seqlens_q_padded is not None
                and not torch.equal(cu_seqlens_q_padded, cu_seqlens_q)
            ) or (
                cu_seqlens_kv_padded is not None
                and not torch.equal(cu_seqlens_kv_padded, cu_seqlens_kv)
            )
5385

5386
5387
5388
5389
5390
5391
5392
5393
5394
5395
5396
5397
5398
5399
5400
5401
5402
5403
5404
5405
5406
5407
5408
5409
5410
5411
5412
5413
5414
5415
5416
            (
                use_flash_attention,
                use_fused_attention,
                use_unfused_attention,
                _,
                fused_attention_backend,
            ) = get_attention_backend(
                qkv_type=type(query_layer),
                qkv_dtype=query_layer.dtype,
                qkv_layout=qkv_layout,
                batch_size=batch_size,
                num_heads=query_layer.shape[-2],
                num_gqa_groups=key_layer.shape[-2],
                max_seqlen_q=max_seqlen_q,
                max_seqlen_kv=max_seqlen_kv,
                head_dim=query_layer.shape[-1],
                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,
                deterministic=deterministic,
                fp8=self.fp8,
                fp8_meta=self.fp8_meta,
            )
5417

5418
5419
5420
5421
5422
5423
5424
5425
5426
5427
5428
5429
5430
5431
5432
5433
5434
5435
5436
5437
5438
5439
5440
5441
5442
5443
5444
            run_config = {
                "compute_capability": "sm"
                + str(
                    (lambda x, y: x * 10 + y)(
                        self.device_compute_capability[0], self.device_compute_capability[1]
                    )
                ),
                "q_dtype": query_layer.dtype,
                "k_dtype": key_layer.dtype,
                "v_dtype": value_layer.dtype,
                "q_shape": list(query_layer.shape),
                "k_shape": list(key_layer.shape),
                "v_shape": list(value_layer.shape),
                "qkv_format": qkv_format,
                "qkv_layout": qkv_layout,
                "mask_type": attn_mask_type,
                "bias_type": core_attention_bias_type,
                "bias_shape": (
                    core_attention_bias.shape if core_attention_bias is not None else None
                ),
                "dropout": self.attention_dropout,
                "context_parallel": context_parallel,
                "is_training": self.training,
                "transformer_engine_version": te.__version__,
                "flash_attn_version": _flash_attn_version,
                "cudnn_version": ".".join([str(i) for i in get_cudnn_version()]),
            }
5445

5446
5447
5448
5449
5450
5451
5452
5453
5454
5455
5456
5457
5458
5459
5460
5461
5462
5463
5464
5465
5466
5467
5468
5469
5470
5471
            if use_flash_attention:
                self.logger.info("Running with FlashAttention backend ")
                self.logger.debug("Running with config=%s", run_config)
                if core_attention_bias_type == "alibi":
                    alibi_slopes, _ = get_alibi(
                        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,
                    max_seqlen_q=max_seqlen_q,
                    max_seqlen_kv=max_seqlen_kv,
5472
                )
5473

5474
5475
5476
5477
            if use_fused_attention:
                self.logger.info(
                    "Running with FusedAttention backend (sub-backend %s)",
                    int(fused_attention_backend),
5478
                )
5479
5480
5481
5482
5483
5484
5485
5486
5487
5488
                if self.fp8:
                    self.logger.debug(
                        "Running with fp8_recipe.fp8_mha=%s, "
                        "fp8_recipe.fp8_dpa=%s%s, and NVTE_FP8_DPA_BWD=%s",
                        self.fp8_meta["recipe"].fp8_mha,
                        self.fp8_meta["recipe"].fp8_dpa,
                        forced_fp8_dpa,
                        int(os.getenv("NVTE_FP8_DPA_BWD", "1")),
                    )
                self.logger.debug("Running with config=%s", run_config)
5489
5490
5491
5492
5493
5494
5495
5496
5497
5498
5499
                fu_core_attention_bias_type = core_attention_bias_type
                fu_core_attention_bias = core_attention_bias
                if core_attention_bias_type == "alibi" and alibi_slopes is not None:
                    fu_core_attention_bias_type = "post_scale_bias"
                    _, fu_core_attention_bias = get_alibi(
                        query_layer.shape[-2],
                        max_seqlen_q,
                        max_seqlen_kv,
                        alibi_slopes=alibi_slopes,
                        bias_dtype=query_layer.dtype,
                    )
5500
5501
5502
5503
5504
5505
5506
5507
5508
                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,
5509
5510
                        cu_seqlens_q_padded=cu_seqlens_q_padded,
                        cu_seqlens_kv_padded=cu_seqlens_kv_padded,
5511
5512
5513
5514
5515
5516
5517
5518
5519
5520
5521
5522
5523
5524
5525
                        max_seqlen_q=max_seqlen_q,
                        max_seqlen_kv=max_seqlen_kv,
                        attn_mask_type=attn_mask_type,
                        attention_mask=attention_mask,
                        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,
                        fp8=self.fp8 and self.fp8_meta["recipe"].fp8_dpa,
                        fp8_meta=self.fp8_meta,
                    )
                return self.fused_attention(
5526
5527
5528
5529
5530
5531
                    query_layer,
                    key_layer,
                    value_layer,
                    qkv_layout=qkv_layout,
                    cu_seqlens_q=cu_seqlens_q,
                    cu_seqlens_kv=cu_seqlens_kv,
5532
5533
                    cu_seqlens_q_padded=cu_seqlens_q_padded,
                    cu_seqlens_kv_padded=cu_seqlens_kv_padded,
5534
5535
                    max_seqlen_q=max_seqlen_q,
                    max_seqlen_kv=max_seqlen_kv,
5536
5537
5538
                    attn_mask_type=attn_mask_type,
                    attention_mask=attention_mask,
                    fused_attention_backend=fused_attention_backend,
5539
5540
                    core_attention_bias_type=fu_core_attention_bias_type,
                    core_attention_bias=fu_core_attention_bias,
5541
5542
5543
5544
                    fast_zero_fill=fast_zero_fill,
                    cp_group=self.cp_group,
                    cp_global_ranks=self.cp_global_ranks,
                    cp_stream=self.cp_stream,
5545
5546
                    fp8=self.fp8 and self.fp8_meta["recipe"].fp8_dpa,
                    fp8_meta=self.fp8_meta,
5547
                )
5548

5549
            from .cpu_offload import CPUOffloadEnabled
5550

5551
5552
5553
5554
5555
            if CPUOffloadEnabled:
                warnings.warn(
                    "Attention activation Offloading is only implemented"
                    "with Flash Attention and Fused Attention!"
                )
5556

5557
5558
5559
5560
5561
5562
5563
5564
5565
5566
5567
5568
5569
5570
5571
5572
5573
5574
5575
            if use_unfused_attention:
                self.logger.info("Running with UnfusedDotProductAttention backend")
                self.logger.debug("Running with config=%s", run_config)
                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,
                        core_attention_bias_type=core_attention_bias_type,
                        core_attention_bias=core_attention_bias,
                        alibi_slopes=alibi_slopes,
                    )
                return self.unfused_attention(
5576
5577
5578
                    query_layer,
                    key_layer,
                    value_layer,
5579
5580
5581
5582
5583
5584
5585
5586
5587
                    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,
                    core_attention_bias_type=core_attention_bias_type,
                    core_attention_bias=core_attention_bias,
                    alibi_slopes=alibi_slopes,
                )
5588

5589
            raise Exception("No dot product attention support for the provided inputs!")
5590
5591


5592
5593
5594
5595
5596
5597
5598
class MultiheadAttention(torch.nn.Module):
    r"""
    Multi-head Attention (MHA), including Query,
    Key, Value and Output projection.

    .. note::

5599
5600
        Argument :attr:`attention_mask` in the `forward` call is only used when
        :attr:`attn_mask_type` includes '"padding"' or `"arbitrary"`.
5601

5602
5603
5604
5605
5606
5607
5608
5609
5610
5611
5612
5613
5614
5615
5616
5617
5618
5619
5620
5621
5622
5623
5624
5625
5626
    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.
5627
5628
    attn_mask_type: {'no_mask', 'padding', 'causal', 'padding_causal', 'causal_bottom_right',
                   'padding_causal_bottom_right','arbitrary'},
5629
                   default = `causal`
5630
5631
5632
5633
5634
                   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.
5635
5636
5637
5638
    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
5639
                window and "`causal`" mask specifically. Similar to :attr:`attn_mask_type`, it can
5640
                be overridden by :attr:`window_size` in `forward` as well.
5641
5642
5643
5644
5645
5646
5647
5648
5649
5650
5651
5652
5653
    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.
5654
5655
    input_layernorm: bool, default = `False`
                     if set to `True`, layer normalization to the input is applied.
5656
5657
5658
5659
5660
5661
5662
5663
5664
5665
5666
5667
5668
5669
5670
5671
5672
5673
5674
5675
5676
5677
5678
    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"
          The device on which the parameters of the model will allocated. It is the user's
          responsibility to ensure all parameters are moved to the GPU before running the
          forward pass.
5679
5680
5681
5682
5683
5684
5685
    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.
5686
            For that, please use `get_qkv_layout` to gain the layout information.
5687
5688
5689
5690
5691
5692
5693
5694
5695
5696
5697
5698
5699
5700
5701
5702
5703
5704
5705
5706
5707
5708
5709
5710
5711
5712
5713
5714
5715
5716
5717
5718
5719
5720
5721
5722
5723
5724
5725
5726

    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`.
5727
5728
5729
5730
5731
5732
    """

    def __init__(
        self,
        hidden_size: int,
        num_attention_heads: int,
5733
5734
5735
5736
5737
        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,
5738
        layer_number: Optional[int] = None,
5739
        attn_mask_type: str = "causal",
5740
        window_size: Optional[Tuple[int, int]] = None,
5741
5742
        tp_group: Optional[dist_group_type] = None,
        tp_size: int = 1,
5743
        num_gqa_groups: Optional[int] = None,
5744
5745
5746
        fuse_wgrad_accumulation: bool = False,
        get_rng_state_tracker: Optional[Callable] = None,
        sequence_parallel: bool = False,
5747
        params_dtype: Optional[torch.dtype] = None,
5748
        return_bias: bool = False,
5749
5750
5751
5752
5753
5754
5755
5756
5757
        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,
        ub_bulk_wgrad: bool = False,
        ub_bulk_dgrad: bool = False,
Jaemin Choi's avatar
Jaemin Choi committed
5758
        ub_overlap_rs_dgrad: bool = False,
5759
5760
        ub_overlap_rs: bool = False,
        ub_overlap_ag: bool = False,
5761
        bias: bool = True,
5762
        normalization: str = "LayerNorm",
5763
        device: Union[torch.device, str] = "cuda",
5764
        qkv_format: str = "sbhd",
5765
5766
    ) -> None:
        super().__init__()
5767

5768
        self.qkv_format = qkv_format
5769
        self.attn_mask_type = attn_mask_type
5770
        self.window_size = check_set_window_size(attn_mask_type, window_size)
5771
        self.layer_number = layer_number
5772
5773
5774
5775
5776
        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
5777
        self.params_dtype = torch.get_default_dtype() if params_dtype is None else params_dtype
5778
        self.num_attention_heads = num_attention_heads
5779
5780
5781
5782
5783
5784
5785
5786
        self.return_bias = return_bias

        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()
5787
5788
5789
5790
5791

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

5792
5793
5794
        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"
5795
5796
5797
5798
5799
5800

        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)
5801
5802
5803
5804
5805
5806
5807
        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!"
5808
        self.num_gqa_groups_per_partition = int(self.num_gqa_groups // tp_size)
5809
5810
5811
5812

        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
5813
5814
5815
5816
5817
5818
5819

        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,
5820
            "params_dtype": self.params_dtype,
5821
            "device": device,
5822
5823
5824
5825
        }

        qkv_parallel_mode = "column" if set_parallel_mode else None

cyanguwa's avatar
cyanguwa committed
5826
        if self.attention_type == "self":
5827
5828
            parameters_split = None
            if not fuse_qkv_params:
5829
5830
5831
5832
5833
5834
5835
                parameters_split = collections.OrderedDict(
                    [
                        ("query", self.hidden_size_q),
                        ("key", self.hidden_size_kv),
                        ("value", self.hidden_size_kv),
                    ]
                )
5836
5837
5838
            if self.input_layernorm:
                self.layernorm_qkv = LayerNormLinear(
                    hidden_size,
5839
                    self.hidden_size_q + 2 * self.hidden_size_kv,
5840
5841
5842
5843
5844
5845
                    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
5846
                    parameters_split=parameters_split,
5847
5848
5849
                    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
5850
                    ub_overlap_rs_dgrad=ub_overlap_rs_dgrad,
5851
                    ub_overlap_ag=ub_overlap_ag,
5852
                    normalization=normalization,
Przemyslaw Tredak's avatar
Przemyslaw Tredak committed
5853
                    ub_name="qkv",
5854
5855
5856
5857
5858
                    **common_gemm_kwargs,
                )
            else:
                self.qkv = Linear(
                    hidden_size,
5859
                    self.hidden_size_q + 2 * self.hidden_size_kv,
5860
5861
5862
5863
                    init_method=init_method,
                    bias=bias,
                    return_bias=False,
                    parallel_mode=qkv_parallel_mode,
cyanguwa's avatar
cyanguwa committed
5864
                    parameters_split=parameters_split,
5865
5866
                    **common_gemm_kwargs,
                )
cyanguwa's avatar
cyanguwa committed
5867
        elif self.attention_type == "cross":
5868
5869
5870
            if self.input_layernorm:
                self.layernorm_query = LayerNormLinear(
                    hidden_size,
5871
                    self.hidden_size_q,
5872
5873
5874
5875
5876
                    eps=layernorm_epsilon,
                    init_method=init_method,
                    bias=bias,
                    return_bias=False,
                    parallel_mode=qkv_parallel_mode,
5877
                    parameters_split=("query",) if not fuse_qkv_params else None,
5878
5879
5880
5881
                    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
5882
                    ub_overlap_rs_dgrad=ub_overlap_rs_dgrad,
5883
                    ub_overlap_ag=ub_overlap_ag,
5884
                    normalization=normalization,
Przemyslaw Tredak's avatar
Przemyslaw Tredak committed
5885
                    ub_name="qkv",
5886
5887
5888
5889
5890
                    **common_gemm_kwargs,
                )
            else:
                self.query_layer = Linear(
                    hidden_size,
5891
                    self.hidden_size_q,
5892
5893
5894
5895
5896
5897
5898
5899
                    init_method=init_method,
                    bias=bias,
                    return_bias=False,
                    parallel_mode=qkv_parallel_mode,
                    **common_gemm_kwargs,
                )
            self.key_value = Linear(
                hidden_size,
5900
                2 * self.hidden_size_kv,
5901
5902
5903
5904
                init_method=init_method,
                bias=bias,
                return_bias=False,
                parallel_mode=qkv_parallel_mode,
5905
                parameters_split=("key", "value") if not fuse_qkv_params else None,
5906
5907
5908
5909
5910
5911
                **common_gemm_kwargs,
            )

        # Attention.
        self.core_attention = DotProductAttention(
            num_attention_heads,
5912
            self.hidden_size_per_attention_head,
5913
5914
            num_gqa_groups=self.num_gqa_groups,
            attention_dropout=attention_dropout,
5915
            qkv_format=self.qkv_format,
5916
5917
5918
5919
            tp_size=tp_size,
            get_rng_state_tracker=get_rng_state_tracker,
            sequence_parallel=sequence_parallel,
            tp_group=tp_group,
5920
            layer_number=self.layer_number,
5921
            attention_type=self.attention_type,
5922
5923
5924
5925
        )

        # Linear
        self.proj = Linear(
5926
            self.hidden_size_q,
5927
5928
5929
            hidden_size,
            init_method=output_layer_init_method,
            bias=bias,
5930
            return_bias=return_bias,
5931
            parallel_mode="row" if set_parallel_mode else None,
5932
5933
            ub_overlap_rs=ub_overlap_rs,
            ub_overlap_ag=ub_overlap_ag,
Przemyslaw Tredak's avatar
Przemyslaw Tredak committed
5934
            ub_name="proj",
5935
5936
5937
5938
            **common_gemm_kwargs,
        )

    def _allocate_memory(
5939
        self, inference_max_sequence_len: int, batch_size: int, dtype: torch.dtype
5940
5941
5942
5943
    ) -> torch.Tensor:
        return torch.empty(
            inference_max_sequence_len,
            batch_size,
5944
            self.num_gqa_groups_per_partition,
5945
            self.hidden_size_per_attention_head,
5946
            dtype=dtype,
5947
5948
5949
5950
            device=torch.cuda.current_device(),
        )

    def set_tensor_parallel_group(self, tp_group: Union[dist_group_type, None]) -> None:
5951
5952
5953
5954
5955
5956
5957
5958
5959
        """
        Set the tensor parallel group for the given
        module before executing the forward pass.

        Parameters
        ----------
        tp_group : ProcessGroup, default = `None`
                  tensor parallel process group.
        """
5960
5961
        self.tp_group = tp_group

5962
    def set_context_parallel_group(
5963
5964
        self,
        cp_group: Union[dist_group_type, None],
5965
        cp_global_ranks: List[int],
5966
5967
        cp_stream: torch.cuda.Stream,
    ) -> None:
5968
5969
5970
5971
5972
5973
5974
5975
5976
5977
5978
5979
5980
        """
        Set the context parallel attributes for the given
        module before executing the forward pass.

        Parameters
        ----------
        cp_group : ProcessGroup
                  context parallel process group.
        cp_global_ranks : List[int]
                         list of global ranks in the context group.
        cp_stream : torch.cuda.Stream
                   cuda stream for context parallel execution.
        """
5981
5982
5983
5984
5985
5986
        # 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"):
                child.set_context_parallel_group(cp_group, cp_global_ranks, cp_stream)
5987

5988
5989
5990
    def forward(
        self,
        hidden_states: torch.Tensor,
5991
        attention_mask: Optional[Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]] = None,
5992
        encoder_output: Optional[torch.Tensor] = None,
5993
        attn_mask_type: Optional[str] = None,
5994
        window_size: Optional[Tuple[int, int]] = None,
5995
5996
        is_first_microbatch: Optional[bool] = None,
        checkpoint_core_attention: bool = False,
5997
        inference_params: Optional[InferenceParams] = None,
5998
        rotary_pos_emb: Optional[Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]] = None,
5999
6000
        core_attention_bias_type: str = "no_bias",
        core_attention_bias: Optional[torch.Tensor] = None,
6001
        alibi_slopes: Optional[torch.Tensor] = None,
6002
        fast_zero_fill: bool = True,
6003
    ) -> Tuple[Union[torch.Tensor, None], ...]:
6004
6005
6006
6007
6008
        """
        Forward propagation for MultiheadAttention layer.

        .. note::

6009
6010
            Argument :attr:`attention_mask` is only used when :attr:`attn_mask_type`
            includes `"padding"` or `"arbitrary"`.
6011
6012
6013
6014
6015

        Parameters
        ----------
        hidden_states : torch.Tensor
             Input tensor.
6016
6017
        attention_mask: Optional[Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]],
             default = `None`. Boolean tensor(s) used to mask out attention softmax input.
6018
             It should be `None` for causal masks and "`no_mask`". For padding masks, it should be
6019
6020
             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]
6021
6022
6023
6024
6025
6026
             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'},
6027
                       default = `None`
6028
6029
6030
6031
                       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.
6032
6033
        window_size: Optional[Tuple[int, int]], default = `None`
                    sliding window size for local attention.
6034
6035
6036
6037
6038
6039
6040
6041
6042
6043
6044
6045
6046
6047
6048
6049
6050
6051
6052
6053
6054
6055
6056
6057
6058
        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`
6059
                    Bias type, {`no_bias`, `pre_scale_bias`, 'post_scale_bias`, `alibi`}
6060
        core_attention_bias: Optional[torch.Tensor], default = `None`
6061
6062
                    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.
6063
6064
6065
6066
        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.
6067
6068
6069
        fast_zero_fill: bool, default = `True`
                    Whether to set output tensors to 0 or not before use.
        """
6070
6071
        # hidden_states: [sq, b, h]

6072
        if attn_mask_type is None:
6073
            attn_mask_type = self.attn_mask_type
6074
6075
        if window_size is None:
            window_size = self.window_size
6076
        window_size = check_set_window_size(attn_mask_type, window_size)
6077

6078
        if "padding" in attn_mask_type and attention_mask is not None:
6079
            for i, _ in enumerate(attention_mask):
6080
6081
6082
                assert (
                    attention_mask[i].dtype == torch.bool
                ), "Attention mask must be in boolean type!"
6083

6084
6085
6086
        assert (
            core_attention_bias_type in AttnBiasTypes
        ), f"core_attention_bias_type {core_attention_bias_type} is not supported!"
6087

6088
        # =================================================
6089
        # Pre-allocate memory for key-values for inference
6090
6091
6092
6093
        # =================================================

        if inference_params and self.layer_number is not None:
            if self.layer_number not in inference_params.key_value_memory_dict:
6094
                inf_max_seq_len = inference_params.max_sequence_length
6095
6096
                inf_max_batch_size = inference_params.max_batch_size
                inference_key_memory = self._allocate_memory(
6097
                    inf_max_seq_len, inf_max_batch_size, hidden_states.dtype
6098
6099
                )
                inference_value_memory = self._allocate_memory(
6100
                    inf_max_seq_len, inf_max_batch_size, hidden_states.dtype
6101
6102
6103
6104
6105
6106
6107
6108
6109
6110
6111
                )
                inference_params.key_value_memory_dict[self.layer_number] = (
                    inference_key_memory,
                    inference_value_memory,
                )
            else:
                (
                    inference_key_memory,
                    inference_value_memory,
                ) = inference_params.key_value_memory_dict[self.layer_number]

6112
        # ======================
6113
        # Query, Key, and Value
6114
        # ======================
6115

cyanguwa's avatar
cyanguwa committed
6116
6117
        if self.attention_type == "self":
            # Attention heads [sq, b, h] --> [sq, b, ng * (np/ng + 2) * hn]
6118
6119
6120
6121
6122
6123
6124
6125
6126
6127
6128
6129
6130
            if self.input_layernorm:
                layernorm_qkv_outputs = self.layernorm_qkv(
                    hidden_states,
                    is_first_microbatch=is_first_microbatch,
                )
                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,
6131
                    is_first_module_in_mha=True,  # specific to FP8 MHA
6132
6133
                )

6134
6135
6136
            num_queries_per_key_value = (
                self.num_attention_heads_per_partition // self.num_gqa_groups_per_partition
            )
6137
            if self.qkv_weight_interleaved:
cyanguwa's avatar
cyanguwa committed
6138
                # [sq, b, ng * (np/ng + 2) * hn] --> [sq, b, ng, (np/ng + 2), hn]
6139
                new_tensor_shape = mixed_x_layer.size()[:-1] + (
cyanguwa's avatar
cyanguwa committed
6140
6141
                    self.num_gqa_groups_per_partition,
                    (num_queries_per_key_value + 2),
6142
6143
6144
6145
                    self.hidden_size_per_attention_head,
                )
                # split along second last dimension
                split_dim = -2
cyanguwa's avatar
cyanguwa committed
6146
6147
6148
6149
6150
            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,
6151
                    self.hidden_size_per_attention_head,
cyanguwa's avatar
cyanguwa committed
6152
6153
6154
                )
                # split along third last dimension
                split_dim = -3
6155
6156
6157

            mixed_x_layer = mixed_x_layer.view(*new_tensor_shape)

cyanguwa's avatar
cyanguwa committed
6158
6159
6160
6161
6162
6163
6164
6165
6166
            # 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]
            if not is_in_onnx_export_mode():
                query_layer, key_layer, value_layer = _SplitAlongDim.apply(
                    mixed_x_layer, split_dim, (num_queries_per_key_value, 1, 1)
6167
                )
6168
            else:
cyanguwa's avatar
cyanguwa committed
6169
                query_layer, key_layer, value_layer = torch.split(
6170
6171
6172
6173
                    mixed_x_layer,
                    (num_queries_per_key_value, 1, 1),
                    dim=split_dim,
                )
cyanguwa's avatar
cyanguwa committed
6174
6175
6176

            # query: -> [sq, b, np, hn]
            # key, value: -> [sq, b, ng, hn]
6177
6178
6179
6180
            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
6181
6182
6183

        elif self.attention_type == "cross":
            # Attention heads [sk, b, h] --> [sk, b, (ng * 2 * hn)]
6184
            mixed_kv_layer = self.key_value(
cyanguwa's avatar
cyanguwa committed
6185
                encoder_output,
6186
                is_first_microbatch=is_first_microbatch,
6187
                is_first_module_in_mha=True,  # specific to FP8 MHA
6188
6189
6190
            )

            if self.qkv_weight_interleaved:
cyanguwa's avatar
cyanguwa committed
6191
                # [sq, b, (ng * 2 * hn)] --> [sq, b, ng, 2 * hn]
6192
                new_tensor_shape = mixed_kv_layer.size()[:-1] + (
6193
                    self.num_gqa_groups_per_partition,
6194
6195
6196
6197
6198
                    2 * self.hidden_size_per_attention_head,
                )
                # split along last dimension
                split_dim = -1
            else:
cyanguwa's avatar
cyanguwa committed
6199
                # [sq, b, (ng * 2 * hn)] --> [sq, b, 2 * ng, hn]
6200
                new_tensor_shape = mixed_kv_layer.size()[:-1] + (
6201
                    2 * self.num_gqa_groups_per_partition,
6202
6203
6204
6205
6206
6207
6208
                    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
6209
6210
6211
            # mixed_kv_layer --> 2 [sk, b, ng, hn]
            if not is_in_onnx_export_mode():
                key_layer, value_layer = _SplitAlongDim.apply(
6212
6213
6214
                    mixed_kv_layer,
                    split_dim,
                    mixed_kv_layer.shape[split_dim] // 2,
cyanguwa's avatar
cyanguwa committed
6215
                )
6216
            else:
cyanguwa's avatar
cyanguwa committed
6217
                key_layer, value_layer = torch.split(
6218
6219
6220
                    mixed_kv_layer,
                    mixed_kv_layer.shape[split_dim] // 2,
                    dim=split_dim,
cyanguwa's avatar
cyanguwa committed
6221
                )
6222
6223
6224
6225
6226
6227
6228
6229
6230
            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)
            )
6231
6232
6233
6234
6235
6236
6237
6238
6239
6240
6241
6242
6243
6244
6245

            # 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,
                )
                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,
6246
                    is_first_module_in_mha=True,  # specific to FP8 MHA
6247
6248
6249
6250
6251
6252
6253
6254
6255
                )

            # [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)

6256
6257
6258
        # ======================================================
        # Apply relative positional encoding (rotary embedding)
        # ======================================================
6259

6260
        if rotary_pos_emb is not None:
6261
6262
6263
            assert not isinstance(query_layer, Float8Tensor) and not isinstance(
                key_layer, Float8Tensor
            ), "RoPE is not supported for Float8Tensors!"
6264
            # duplicate the pos_emb for self attention
6265
            if not isinstance(rotary_pos_emb, tuple):
6266
                rotary_pos_emb = (rotary_pos_emb,) * 2
6267
6268

            q_pos_emb, k_pos_emb = rotary_pos_emb
6269
6270
6271
6272
6273
6274
6275
6276
6277
6278
6279
6280
6281
6282

            # 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)

                sequence_start = inference_params.sequence_len_offset
                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, ...]

6283
6284
            query_layer = apply_rotary_pos_emb(query_layer, q_pos_emb, self.qkv_format, fused=True)
            key_layer = apply_rotary_pos_emb(key_layer, k_pos_emb, self.qkv_format, fused=True)
6285

6286
6287
6288
6289
        # ===========================
        # Core attention computation
        # ===========================

6290
6291
6292
6293
        context_layer = self.core_attention(
            query_layer,
            key_layer,
            value_layer,
6294
            qkv_format=self.qkv_format,
6295
6296
            cu_seqlens_q=None,
            cu_seqlens_kv=None,
6297
6298
            attention_mask=attention_mask,
            attn_mask_type=attn_mask_type,
6299
            window_size=window_size,
6300
6301
6302
            checkpoint_core_attention=checkpoint_core_attention,
            core_attention_bias_type=core_attention_bias_type,
            core_attention_bias=core_attention_bias,
6303
            alibi_slopes=alibi_slopes,
6304
            fast_zero_fill=fast_zero_fill,
6305
            inference_params=inference_params,
6306
6307
        )

6308
        # ===================
6309
        # Output. [sq, b, h]
6310
        # ===================
6311

6312
        projection_output = self.proj(
6313
6314
            context_layer,
            is_first_microbatch=is_first_microbatch,
6315
6316
        )

6317
6318
6319
6320
6321
6322
6323
6324
        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,)
6325
        if self.input_layernorm and self.return_layernorm_output:
6326
6327
            outputs += (layernorm_output,)
        return outputs if len(outputs) > 1 else outputs[0]