attention.py 401 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
from importlib.metadata import PackageNotFoundError
10
import math
11
import os
12
from typing import Any, Callable, Dict, List, Optional, Tuple, Union
13
import warnings
14
import logging
15
import functools
16

17
from dataclasses import dataclass, fields
cyanguwa's avatar
cyanguwa committed
18
import numpy as np
19
from packaging.version import Version as PkgVersion
20
21

import torch
22
import torch.nn.functional as F
23

24
import transformer_engine_torch as tex
25
26
import transformer_engine as te
from transformer_engine.pytorch.utils import get_cudnn_version
27
28
29
30
from transformer_engine.pytorch.cpp_extensions import (
    cast_to_fp8,
    cast_from_fp8,
)
31
32
33
34
35
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,
36
37
    fused_attn_fwd,
    fused_attn_bwd,
38
39
40
41
    QKVLayout,
    AttnBiasType,
    AttnMaskType,
    FusedAttnBackend,
42
43
44
45
46
47
48
49
50
51
52
53
54
    META_QKV,
    META_DQKV,
    META_O,
    META_DO,
    META_S,
    META_DP,
    META_O_CP,
    META_DQKV_CP,
)
from transformer_engine.pytorch.fp8 import (
    FP8GlobalStateManager,
    get_fp8_te_dtype,
    get_fp8_torch_dtype,
55
)
56
from transformer_engine.pytorch.float8_tensor import Float8Tensor
57
from transformer_engine.pytorch.module import LayerNormLinear, Linear
58
from transformer_engine.pytorch.module.base import TransformerEngineBaseModule
59
60
61
62
63
from transformer_engine.pytorch.utils import (
    divide,
    attention_mask_func,
    split_tensor_along_dim,
    get_device_compute_capability,
64
    get_default_init_method,
65
66
67
68
)
from transformer_engine.pytorch.constants import (
    AttnMaskTypes,
    AttnTypes,
69
    AttnBiasTypes,
70
    QKVLayouts,
71
    dist_group_type,
72
    TE_DType,
73
74
75
76
)
from transformer_engine.pytorch.softmax import FusedScaleMaskSoftmax
from transformer_engine.pytorch.distributed import (
    get_distributed_world_size,
77
    get_distributed_rank,
78
    checkpoint,
79
80
81
    set_all_rng_states,
    CudaRNGStatesTracker,
    graph_safe_rng_available,
82
83
    gather_along_first_dim,
    reduce_scatter_along_first_dim,
84
85
)
from transformer_engine.pytorch.export import is_in_onnx_export_mode
86
from transformer_engine.pytorch.jit import jit_fuser, no_torch_dynamo
87
88
from transformer_engine.pytorch.graph import is_graph_capturing

89

90
91
92
93
94
95
96
97
98
99
# NVTE_DEBUG = 0/1 # disables/enables debug mode, default = 0
_NVTE_DEBUG = int(os.getenv("NVTE_DEBUG", "0"))
# 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}
_log_level = _log_levels[_log_level if _log_level in [0, 1, 2] else 2]
_formatter = logging.Formatter("[%(levelname)-8s | %(name)-19s]: %(message)s")
_stream_handler = logging.StreamHandler()
_stream_handler.setFormatter(_formatter)
100
101
102
103
104
105
106
107
108
109
fa_logger = logging.getLogger()
fa_logger.setLevel(_log_level)
if not fa_logger.hasHandlers():
    fa_logger.addHandler(_stream_handler)


@functools.lru_cache(maxsize=None)
def _get_supported_versions(version_min, version_max):
    return ">= " + str(version_min) + ", " + "<= " + str(version_max)

110

111
112
113
_NVTE_FLASH_ATTN = int(os.getenv("NVTE_FLASH_ATTN", "1"))
_NVTE_FUSED_ATTN = int(os.getenv("NVTE_FUSED_ATTN", "1"))
_NVTE_UNFUSED_ATTN = int(os.getenv("NVTE_UNFUSED_ATTN", "1"))
114
115
116
117
118

# Detect flash-attn v2 in the environment
_flash_attn_is_installed = False
_flash_attn_version = PkgVersion("0")
_flash_attn_version_required = PkgVersion("2.1.1")
119
_flash_attn_max_version = PkgVersion("2.6.3")
120
121
122
123
124
125
126
_flash_attn_2_plus = False
_flash_attn_2_1_plus = False
_flash_attn_2_3_plus = False
_flash_attn_2_4_plus = False
_flash_attn_2_4_1_plus = False
_flash_attn_2_5_7_plus = False
_flash_attn_2_6_0_plus = False
127
128
129
130
131
132
133

flash_attn_func = None
flash_attn_varlen_func = None
flash_attn_varlen_fwd = None
flash_attn_varlen_bwd = None
flash_attn_cuda_bwd = None

134
135
136
try:
    _flash_attn_version = PkgVersion(get_pkg_version("flash-attn"))
except PackageNotFoundError:
137
    if torch.cuda.is_available() and get_device_compute_capability() >= (8, 0) and _NVTE_FLASH_ATTN:
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
        fa_logger.debug(
            "flash-attn v2 is not installed. To use, please install it by"
            """ "pip install flash-attn".""",
        )
else:
    if _flash_attn_version_required <= _flash_attn_version <= _flash_attn_max_version:
        from flash_attn.flash_attn_interface import flash_attn_func, flash_attn_varlen_func
        from flash_attn.flash_attn_interface import (
            _flash_attn_varlen_forward as flash_attn_varlen_fwd,
        )
        from flash_attn.flash_attn_interface import (
            _flash_attn_varlen_backward as flash_attn_varlen_bwd,
        )
        from flash_attn_2_cuda import varlen_bwd as flash_attn_cuda_bwd

        _flash_attn_is_installed = True
        _flash_attn_2_plus = _flash_attn_version >= PkgVersion("2")
        _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")
        _flash_attn_2_5_7_plus = _flash_attn_version >= PkgVersion("2.5.7")
        _flash_attn_2_6_0_plus = _flash_attn_version >= PkgVersion("2.6.0")
161
162
163
    elif (
        torch.cuda.is_available() and get_device_compute_capability() >= (8, 0) and _NVTE_FLASH_ATTN
    ):
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
        fa_logger.warning(
            "Supported flash-attn versions are %s. Found flash-attn %s.",
            _get_supported_versions(
                _flash_attn_version_required,
                _flash_attn_max_version,
            ),
            _flash_attn_version,
        )

# Detect flash-attn v3 in the environment
# This section will be removed when FA3 is released as a regular FA package,
# i.e. flashattn-hopper 3.0.0 as flash-attn 3.0.0
_flash_attn_3_is_installed = False
_flash_attn_3_version = PkgVersion("0")
_flash_attn_3_0_0_beta = False
179
_use_flash_attn_3 = False
180
181
182
183
184
_flash_attn_3_installation_steps = """\
(1) pip install "git+https://github.com/Dao-AILab/flash-attention.git#egg=flashattn-hopper&subdirectory=hopper"
(2) python_path=`python -c "import site; print(site.getsitepackages()[0])"`
(3) mkdir -p $python_path/flashattn_hopper
(4) wget -P $python_path/flashattn_hopper https://raw.githubusercontent.com/Dao-AILab/flash-attention/main/hopper/flash_attn_interface.py"""
185
try:
186
    _flash_attn_3_version = PkgVersion(get_pkg_version("flashattn-hopper"))
187
except PackageNotFoundError:
188
    if torch.cuda.is_available() and get_device_compute_capability() >= (9, 0) and _NVTE_FLASH_ATTN:
189
190
        fa_logger.debug(
            "flash-attn v3 is not installed. To use, please install it by \n%s",
191
            _flash_attn_3_installation_steps,
192
        )
193
194
195
196
197
else:
    from flashattn_hopper.flash_attn_interface import flash_attn_func as flash_attn_func_v3
    from flashattn_hopper.flash_attn_interface import (
        flash_attn_varlen_func as flash_attn_varlen_func_v3,
    )
198
199
200
201
202
203
    from flashattn_hopper.flash_attn_interface import (
        _flash_attn_varlen_forward as flash_attn_varlen_fwd_v3,
    )
    from flashattn_hopper.flash_attn_interface import (
        _flash_attn_varlen_backward as flash_attn_varlen_bwd_v3,
    )
204

205
206
    _flash_attn_3_is_installed = True
    _flash_attn_3_0_0_beta = PkgVersion("3.0.0b") < _flash_attn_3_version < PkgVersion("3.0.0")
207
    _use_flash_attn_3 = True
208

209
210
211
212
213
214
215
_attention_backends = {
    "attention_params": None,
    "use_flash_attention": None,
    "use_fused_attention": None,
    "fused_attention_backend": None,
    "use_unfused_attention": None,
    "backend_selection_requires_update": False,
216
}
217
218


219
220
@dataclass(eq=True)
class AttentionParams:
221
    """
222
    Attention parameters used to determine which backend to be used.
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241

    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.
242
243
244
245
    head_dim_qk: int, default = 64
        The size of each attention head in query and key tensors.
    head_dim_v: int, default = 64
        The size of each attention head in the value tensor.
246
247
248
    attn_mask_type: str, default = `no_mask`
        Attention mask type, {`no_mask`, `padding`, `causal`, `padding_causal`,
        `causal_bottom_right`, `padding_causal_bottom_right`, `arbitrary`}
249
    window_size: Tuple[int, int], default = None
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
        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.
268
269
    is_training: bool, default = `True`
        Whether in training mode (`True`) or inference mode (`False`)
270
271
272
273
    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`.
274
275
276
277
278
279
280
281
282
283
    """

    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
284
285
    head_dim_qk: int = 64
    head_dim_v: int = 64
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
    attn_mask_type: str = "no_mask"
    window_size: Union[Tuple[int, int], None] = None
    alibi_slopes_shape: Union[torch.Size, List, None] = 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
    is_training: bool = True
    fp8: bool = False
    fp8_meta: Union[Dict[str, Any], None] = None


_alibi_cache = {
    "_num_heads": None,
    "_alibi_slopes": None,
    "_max_seqlen_q": None,
    "_max_seqlen_kv": None,
    "_bottom_right_alignment": True,
    "_alibi_bias": None,
    "_alibi_slopes_require_update": False,
    "_alibi_bias_require_update": False,
}


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


316
317
318
319
320
def maybe_contiguous(tensor: torch.Tensor) -> torch.Tensor:
    """Make tensor contiguous if final stride is not 1."""
    return tensor.contiguous() if tensor.stride(-1) != 1 else tensor


321
322
323
324
325
326
327
328
329
def get_attention_backend(
    attention_params: AttentionParams = None,
):
    """
    Select the appropriate attention backend/sub-backend based on user input and runtime environment.

    Parameters
    ----------
    See `AttentionParams`.
330
331
332
333
334
335
336

    Returns
    ----------
    use_flash_attention: bool
        Whether the `FlashAttention` backend has been selected.
    use_fused_attention: bool
        Whether the `FusedAttention` backend has been selected.
337
338
    fused_attention_backend: tex.NVTE_Fused_Attn_Backend
        If `use_fused_attention = True`, one of `FusedAttention` three sub-backends, else `None`.
339
340
341
342
343
344
    use_unfused_attention: bool
        Whether the `UnfusedDotProductAttention` backend has been selected.
    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].
    """
345
346
347
348
349
350
351
352
    qkv_type = attention_params.qkv_type
    qkv_dtype = attention_params.qkv_dtype
    qkv_layout = attention_params.qkv_layout
    batch_size = attention_params.batch_size
    num_heads = attention_params.num_heads
    num_gqa_groups = attention_params.num_gqa_groups
    max_seqlen_q = attention_params.max_seqlen_q
    max_seqlen_kv = attention_params.max_seqlen_kv
353
354
    head_dim_qk = attention_params.head_dim_qk
    head_dim_v = attention_params.head_dim_v
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
    attn_mask_type = attention_params.attn_mask_type
    window_size = attention_params.window_size
    alibi_slopes_shape = attention_params.alibi_slopes_shape
    core_attention_bias_type = attention_params.core_attention_bias_type
    core_attention_bias_shape = attention_params.core_attention_bias_shape
    core_attention_bias_requires_grad = attention_params.core_attention_bias_requires_grad
    pad_between_seqs = attention_params.pad_between_seqs
    attention_dropout = attention_params.attention_dropout
    context_parallel = attention_params.context_parallel
    deterministic = attention_params.deterministic
    is_training = attention_params.is_training
    fp8 = attention_params.fp8
    fp8_meta = attention_params.fp8_meta

    # Run config
370
    logger = logging.getLogger("DotProductAttention")
371
372
373
    logger.setLevel(_log_level)
    if not logger.hasHandlers():
        logger.addHandler(_stream_handler)
374
375
376
377
378
    device_compute_capability = get_device_compute_capability()
    cudnn_version = get_cudnn_version()
    run_config = {
        "transformer_engine_version": te.__version__,
        "compute_capability": "sm"
379
        + str(10 * device_compute_capability[0] + device_compute_capability[1]),
380
381
382
383
384
385
        "flash_attn_version": (
            str(_flash_attn_version) if _flash_attn_is_installed else "not installed"
        ),
        "flash_attn_3_version": (
            str(_flash_attn_3_version) if _flash_attn_3_is_installed else "not installed"
        ),
386
387
388
389
390
391
392
393
394
        "cudnn_version": ".".join([str(i) for i in cudnn_version]),
    }
    attention_params_dict = {
        field.name: getattr(attention_params, field.name) for field in fields(attention_params)
    }
    run_config.update(attention_params_dict)
    if fp8:
        run_config["NVTE_FP8_DPA_BWD"] = int(os.getenv("NVTE_FP8_DPA_BWD", "1"))
    logger.debug("Running with config=%s", run_config)
395

396
397
398
399
400
401
    # The following sections check if `FlashAttention` supports the provided attention params,
    # regardless of whether FA2 or FA3 is installed. If FA2 or FA3 is not installed but is
    # necessary for performance/functionality, a warning will be issued to prompt users to
    # install an appropriate FA version.
    global _flash_attn_version_required, _flash_attn_max_version, _use_flash_attn_3

402
    # Filter: Environment variables
403
404
405
406
    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 and _flash_attn_is_installed:
407
408
409
410
411
412
413
414
        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():
415
        if use_flash_attention and _flash_attn_is_installed:
416
417
418
419
420
421
422
423
            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
    if device_compute_capability < (8, 0):
424
        if use_flash_attention and _flash_attn_is_installed:
425
            logger.debug("Disabling FlashAttention as it requires compute capability sm80+")
426
        use_flash_attention = False
427
428
429
        if use_fused_attention:
            logger.debug("Disabling FusedAttention as it requires compute capability sm80+")
            use_fused_attention = False
430
    if device_compute_capability < (9, 0):
431
        if use_flash_attention and _flash_attn_3_is_installed:
432
            logger.debug("Disabling FlashAttention 3 as it requires compute capability sm90+")
433
        _use_flash_attn_3 = False
434
435

    # Filter: Data type
436
437
438
439
    if qkv_dtype not in [torch.bfloat16, torch.float16] or qkv_type not in [
        torch.Tensor,
        Float8Tensor,
    ]:
440
        if use_flash_attention and _flash_attn_is_installed:
441
442
443
444
445
446
            logger.debug(
                "Disabling FlashAttention due to unsupported QKV data type. "
                "Supported: qkv_dtype = {torch.bfloat16, torch.float16}. "
                "Found: qkv_dtype = %s.",
                qkv_dtype,
            )
447
        use_flash_attention = False
448
449
450
451
452
453
454
455
        if use_fused_attention:
            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
456
457
458

    # Filter: Execution type
    if fp8 and fp8_meta["recipe"].fp8_dpa:
459
        if use_flash_attention and not _use_flash_attn_3:
460
461
            if _flash_attn_is_installed:
                logger.debug("Disabling FlashAttention as FlashAttention 2 does not support FP8")
462
463
464
465
466
            use_flash_attention = False
        if use_flash_attention and _use_flash_attn_3 and is_training:
            logger.debug(
                "Disabling FlashAttention as FlashAttention 3 does not support FP8 training"
            )
467
468
469
470
471
472
            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
473
    if use_flash_attention and head_dim_qk != head_dim_v:
474
475
        if _flash_attn_is_installed:
            logger.debug("Disabling FlashAttention as it does not support MLA.")
476
        use_flash_attention = False
477
    if use_flash_attention and (
478
479
480
        head_dim_qk > 256
        or head_dim_qk % 8 != 0
        or (head_dim_qk > 192 and device_compute_capability not in ((8, 0), (9, 0)))
481
    ):
482
483
484
485
486
487
488
489
490
491
        if _flash_attn_is_installed:
            logger.debug(
                "Disabling FlashAttention due to unsupported head_dim_qk and head_dim_v. "
                "Supported: head_dim_qk = head_dim_v, head_dim_qk %%8 = 0, "
                "head_dim_qk <= 256 (>192 requires sm80/90). "
                "Found: head_dim_qk = %s, head_dim_v = %s, on sm%s.",
                head_dim_qk,
                head_dim_v,
                ".".join([str(i) for i in device_compute_capability]),
            )
492
        use_flash_attention = False
493
494
495
496
497
498
499
    qkv_layout_group = qkv_layout.replace("b", "").replace("s", "").replace("t", "")
    if use_fused_attention and head_dim_qk != head_dim_v and qkv_layout_group != "hd_hd_hd":
        logger.debug(
            "Disabling FusedAttention as MLA is not supported with qkv_layout = %s",
            qkv_layout,
        )
        use_fused_attention = False
500
501
502
503
504
505
506
507

    # 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:
508
509
510
511
512
            if _flash_attn_is_installed:
                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]"
                )
513
514
            use_flash_attention = False

515
    # Filter: Dropout
516
517
518
    if attention_dropout != 0.0 and use_flash_attention and _use_flash_attn_3:
        logger.debug("Disabling FlashAttention 3 for dropout")
        _use_flash_attn_3 = False
519

520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
    # Filter: Context parallelism
    # qkv_format | attn_mask_type              | attn_bias_type           | supported backends
    # ----------------------------------------------------------------------------------------------------
    # bshd, sbhd | self-attention:             | no_bias, post_scale_bias | FlashAttention, FusedAttention
    #            |     no_mask, causal         |                          |
    #            | cross-attention:            |                          |
    #            |     no_mask                 |                          |
    # thd        | self-attention:             | no_bias                  | FlashAttention, FusedAttention
    #            |     padding, padding_causal |                          | if no padding between sequences,
    #            | cross-attention:            |                          | FusedAttention
    #            |     padding                 |                          | if there is padding between sequences
    # Note: context parallelism requires seq_len % (cp_size * 2) == 0 for each sequence in q, k, v.
    if context_parallel and use_unfused_attention:
        logger.debug(
            "Disabling UnfusedDotProductAttention as it does not support context parallelism"
        )
        use_unfused_attention = False
    if context_parallel and use_flash_attention:
538
        if fp8 and fp8_meta["recipe"].fp8_dpa:
539
540
541
542
            if _flash_attn_is_installed:
                logger.debug(
                    "Disabling FlashAttention as it does not support context parallelism with FP8"
                )
543
            use_flash_attention = False
544
        if "bottom_right" in attn_mask_type:
545
546
547
548
549
            if _flash_attn_is_installed:
                logger.debug(
                    "Disabling FlashAttention as it does not support context parallelism with"
                    " causal_bottom_right masking"
                )
550
551
            use_flash_attention = False
        elif "causal" in attn_mask_type and max_seqlen_q != max_seqlen_kv:
552
553
554
555
556
            if _flash_attn_is_installed:
                logger.debug(
                    "Disabling FlashAttention as it does not support context parallelism with"
                    " causal masking for cross-attention"
                )
557
558
            use_flash_attention = False
        elif core_attention_bias_type not in ["no_bias", "post_scale_bias"]:
559
560
561
562
563
564
            if _flash_attn_is_installed:
                logger.debug(
                    "Disabling FlashAttention as it does not support context parallelism with bias"
                    " type of %s",
                    core_attention_bias_type,
                )
565
566
            use_flash_attention = False
        elif qkv_format == "thd" and core_attention_bias_type != "no_bias":
567
568
569
570
571
            if _flash_attn_is_installed:
                logger.debug(
                    "Disabling FlashAttention as it does not support context parallelism with"
                    " attention bias for THD format"
                )
572
            use_flash_attention = False
573

574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
    if context_parallel and use_fused_attention:
        if "bottom_right" in attn_mask_type:
            logger.debug(
                "Disabling FusedAttention as it does not support context parallelism with"
                " causal_bottom_right masking"
            )
            use_fused_attention = False
        elif "causal" in attn_mask_type and max_seqlen_q != max_seqlen_kv:
            logger.debug(
                "Disabling FusedAttention as it does not support context parallelism with causal"
                " masking for cross-attention"
            )
            use_fused_attention = False
        elif core_attention_bias_type not in ["no_bias", "post_scale_bias"]:
            logger.debug(
                "Disabling FusedAttention as it does not support context parallelism with bias type"
                " of %s",
                core_attention_bias_type,
            )
            use_fused_attention = False
        elif qkv_format == "thd" and core_attention_bias_type != "no_bias":
            logger.debug(
                "Disabling FusedAttention as it does not support context parallelism with attention"
                " bias for THD format"
            )
            use_fused_attention = False
        elif head_dim_qk != head_dim_v:
            logger.debug(
                "Disabling FusedAttention as it does not support context parallelism with MLA"
            )
            use_fused_attention = False

606
    # Filter: Attention mask
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
    # attn_mask_type              | attention_mask                       | supported backends
    # ----------------------------------------------------------------------------------------
    # no_mask                     | None                                 | All
    # padding                     |                                      | All
    #     self-attention          | One tensor in shape [b, 1, 1, sq]    |
    #     cross-attention         | Tuple of two tensors in shapes       |
    #                             | [b, 1, 1, sq] and [b, 1, 1, skv]     |
    # causal                      | None                                 |
    #     self-attention          |                                      | All
    #     cross-attention         |                                      | FusedAttention, UnfusedDotProductAttention
    # padding_causal              | Same as "padding"                    |
    #     self-attention          |                                      | All
    #     cross-attention         |                                      | FusedAttention, UnfusedDotProductAttention
    # causal_bottom_right         | None                                 | All
    # padding_causal_bottom_right | Same as "padding"                    |
    #     self-attention          |                                      | All
    #     cross-attention         |                                      | FlashAttention, UnfusedDotProductAttention
    # arbitrary                   | One tensor in shape broadcastable to | UnfusedDotProductAttention
    #                             | [b, h, sq, skv]                      |
626
    if attn_mask_type == "arbitrary":
627
        if use_flash_attention and _flash_attn_is_installed:
628
629
630
631
632
            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
633
634
    if (
        use_flash_attention
635
        and _use_flash_attn_3
636
637
638
639
640
641
642
643
644
        and attn_mask_type in ["causal", "padding_causal"]
        and max_seqlen_q != max_seqlen_kv
    ):
        logger.warning(
            "Disabling FlashAttention 3 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_attn_3 = False
645
646
647
648
649
    if (
        use_flash_attention
        and attn_mask_type in ["causal", "padding_causal"]
        and max_seqlen_q != max_seqlen_kv
    ):
650
651
652
653
654
655
656
657
658
        if _flash_attn_2_1_plus:
            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 not _flash_attn_is_installed:
            _flash_attn_max_version = PkgVersion("2.1")
659
660
661
662
663
    if (
        use_flash_attention
        and attn_mask_type in ["causal_bottom_right", "padding_causal_bottom_right"]
        and max_seqlen_q != max_seqlen_kv
    ):
664
665
666
667
668
669
670
671
672
        if not _flash_attn_is_installed:
            _flash_attn_version_required = PkgVersion("2.1")
        elif not _flash_attn_2_1_plus and not _use_flash_attn_3:
            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
673
674
675
676
677
678
679
680
681
    if (
        use_flash_attention
        and _use_flash_attn_3
        and fp8
        and fp8_meta["recipe"].fp8_dpa
        and "padding" in attn_mask_type
    ):
        logger.debug("Disabling FlashAttention 3 for FP8 and padding masks")
        _use_flash_attn_3 = False
682
683

    # Filter: Sliding window attention
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
    #    backend                 |      window_size       | diagonal alignment
    # ---------------------------------------------------------------------------------
    # FlashAttention             | (-1, -1) or (>=0, >=0) | bottom right
    # FusedAttention             | (-1,  0) or (>=0, 0)   | top left
    # UnfusedDotProductAttention | (-1, -1) or (>=0, >=0) | both;
    #                            |                        | converts window_size to an 'arbitrary' mask
    if window_size is None:
        window_size = check_set_window_size(attn_mask_type, window_size)
    else:
        if use_fused_attention and (window_size[0] != -1 or window_size[1] not in [-1, 0]):
            if fp8 and (fp8_meta["recipe"].fp8_dpa or fp8_meta["recipe"].fp8_mha):
                logger.debug(
                    "Disabling FusedAttention as it does not support sliding window attention"
                    " for FP8"
                )
                use_fused_attention = False
            elif window_size[1] != 0 or attention_dropout != 0.0 or qkv_format == "thd":
                logger.debug(
                    "Disabling FusedAttention as it only supports sliding window attention "
                    "with causal mask, no dropout, and qkv_format = bshd/sbhd"
                )
                use_fused_attention = False
            elif max_seqlen_q != max_seqlen_kv and attn_mask_type in [
                "no_mask",
                "padding",
                "causal_bottom_right",
                "padding_causal_bottom_right",
            ]:
                logger.debug(
                    "Disabling FusedAttention as it does not support sliding window attention "
                    "with attn_mask_type = %s for cross-attention",
                    attn_mask_type,
                )
                use_fused_attention = False
            elif "padding" in attn_mask_type:
                logger.debug(
                    "Disabling FusedAttention as it does not support sliding window attention "
                    "with attn_mask_type = %s",
                    attn_mask_type,
                )
                use_fused_attention = False
725
726
727
728
729
730
731
732
733
734
735
736
737
        if use_flash_attention and (window_size[0] != -1 or window_size[1] not in [-1, 0]):
            if _use_flash_attn_3:
                logger.debug(
                    "Disabling FlashAttention 3 as it does not support sliding window attention"
                )
                _use_flash_attn_3 = False
            if not _flash_attn_is_installed:
                _flash_attn_version_required = PkgVersion("2.3")
            elif not _flash_attn_2_3_plus:
                logger.debug(
                    "Disabling FlashAttention as sliding window attention requires flash-attn 2.3+"
                )
                use_flash_attention = False
738
739

    # Filter: Attention bias
740
741
742
743
744
745
746
747
    #    backend                 |      bias types              | ALiBi diagonal alignment
    # ---------------------------------------------------------------------------------
    # FlashAttention             | no_bias, alibi/alibi_slopes  | bottom right
    # FusedAttention             | no_bias, post_scale_bias     |
    #                            | alibi/alibi_slopes           | top left,
    #                            |                              | bottom_right (converts to a 'post_scale_bias' bias)
    # UnfusedDotProductAttention | no_bias, pre/post_scale_bias |
    #                            | alibi/alibi_slopes           | both; converts to a 'post_scale_bias' bias
748
    if use_flash_attention and core_attention_bias_type == "alibi":
749
        if _use_flash_attn_3:
750
751
            logger.debug("Disabling FlashAttention 3 for ALiBi")
            _use_flash_attn_3 = False
752
753
754
        if not _flash_attn_is_installed:
            _flash_attn_version_required = PkgVersion("2.4")
        elif not _flash_attn_2_4_plus:
755
756
            logger.debug("Disabling FlashAttention as ALiBi requires flash-attn 2.4+")
            use_flash_attention = False
757

758
759
760
761
    if use_flash_attention and (
        core_attention_bias_type not in ["no_bias", "alibi"]
        or core_attention_bias_shape is not None
    ):
762
763
        if _flash_attn_is_installed:
            logger.debug("Disabling FlashAttention for pre/post_scale_bias")
764
765
766
767
768
769
770
771
        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"
772
        and (alibi_slopes_shape is not None or max_seqlen_q != max_seqlen_kv)
773
774
775
    ):
        fu_core_attention_bias_type = "post_scale_bias"
        fu_core_attention_bias_requires_grad = False
776
777
778
779
780
        if alibi_slopes_shape is None:
            fu_core_attention_bias_shape = "1hss"
        elif len(alibi_slopes_shape) == 1 and alibi_slopes_shape[0] == num_heads:
            fu_core_attention_bias_shape = "1hss"
        elif (
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
            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"

    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,
820
821
            head_dim_qk,
            head_dim_v,
822
823
            window_size[0],
            window_size[1],
824
        )
825
        if fused_attention_backend == FusedAttnBackend["No_Backend"]:
826
827
            logger.debug("Disabling FusedAttention as no backend supports the provided input")
            use_fused_attention = False
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
            fused_attention_backend = None
        if (
            use_fused_attention
            and window_size is not None
            and window_size[0] != -1
            and fused_attention_backend != FusedAttnBackend["F16_arbitrary_seqlen"]
        ):
            logger.debug(
                "Disabling FusedAttention as only sub-backend %s does not support "
                "slidng window attention",
                int(fused_attention_backend),
            )
            use_fused_attention = False
            fused_attention_backend = None
        if (
            use_fused_attention
            and fused_attention_backend == FusedAttnBackend["F16_max512_seqlen"]
845
846
847
848
849
850
851
852
            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
853
            fused_attention_backend = None
854
855
856
857
858
859
860
861
862
863
864
865
866

    # 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
867
868
869
870
871
872
873
874
875
876
    if use_flash_attention and deterministic:
        if not _flash_attn_is_installed:
            _flash_attn_version_required = PkgVersion("2.4.1")
        elif not _flash_attn_2_4_1_plus and not _use_flash_attn_3:
            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
877
878
879
880
881
882
883
884
885
886
887
    if use_fused_attention and deterministic:
        if fused_attention_backend == FusedAttnBackend["FP8"] and is_training:
            logger.debug("Disabling FusedAttention for determinism reasons")
            use_fused_attention = False
        if (
            fused_attention_backend == FusedAttnBackend["F16_arbitrary_seqlen"]
            and is_training
            and (
                device_compute_capability < (9, 0)
                or core_attention_bias_requires_grad
                or cudnn_version < (8, 9, 5)
888
            )
889
890
891
        ):
            logger.debug("Disabling FusedAttention for determinism reasons")
            use_fused_attention = False
892
893
894

    # All available backends
    available_backends = [use_flash_attention, use_fused_attention, use_unfused_attention]
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911

    # `FusedAttention` and `FlashAttention` are faster backends than `UnfusedDotProductAttention`.
    # When `FusedAttention` does not support the provided attention params, and `FlashAttention`
    # does, we recommend users to install flash-attn if not installed already.
    if not use_fused_attention and use_flash_attention and not _flash_attn_is_installed:
        logger.warning(
            "flash-attn may provide important feature support or performance improvement."
            " Please install flash-attn %s.",
            _get_supported_versions(
                _flash_attn_version_required,
                _flash_attn_max_version,
            ),
        )
    if use_flash_attention and not _flash_attn_is_installed:
        use_flash_attention = False
        available_backends[0] = False

912
913
914
915
916
917
918
919
920
921
922
923
    logger.debug(
        "Available backends = {FlashAttention=%s, FusedAttention=%s%s,"
        " UnfusedDotProductAttention=%s}",
        bool(available_backends[0]),
        bool(available_backends[1]),
        (
            f" (sub-backend {int(fused_attention_backend)})"
            if fused_attention_backend is not None
            else ""
        ),
        bool(available_backends[2]),
    )
924
925
926
927
928
929
930
931
932
933
934
935
936

    # 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
937
938
939
940
941
942
943
    if (
        use_flash_attention
        and use_fused_attention
        and fused_attention_backend == FusedAttnBackend["FP8"]
        and _use_flash_attn_3
    ):
        logger.debug(
944
945
            "Disabling FlashAttention 3 to give FusedAttention preference for performance reasons "
            "in FP8 execution"
946
947
948
        )
        use_flash_attention = False

949
950
951
952
953
954
    # Selected backend
    if use_flash_attention:
        use_fused_attention = False
        use_unfused_attention = False
    elif use_fused_attention:
        use_unfused_attention = False
955
    selected_backend = "NoBackend"
956
957
958
959
960
961
    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"
962
    logger.debug("Selected backend = %s", selected_backend)
963

964
965
966
967
968
969
    global _attention_backends
    _attention_backends["use_flash_attention"] = use_flash_attention
    _attention_backends["use_fused_attention"] = use_fused_attention
    _attention_backends["fused_attention_backend"] = fused_attention_backend
    _attention_backends["use_unfused_attention"] = use_unfused_attention
    _attention_backends["backend_selection_requires_update"] = False
970
971
972
973

    return (
        use_flash_attention,
        use_fused_attention,
974
        fused_attention_backend,
975
976
977
978
979
        use_unfused_attention,
        available_backends,
    )


980
class InferenceParams:  # pylint: disable=too-few-public-methods
981
982
    """
    Inference parameters that are passed to the main model in order
983
    to efficiently calculate and store the context during inference.
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023

    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,
            )
1024

1025

1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
@torch.no_grad()
def get_swa_mask(
    window_size: Tuple[int, int],
    max_seqlen_q: int,
    max_seqlen_kv: int,
    attn_mask_type: str = "no_mask",
    attention_mask: Optional[Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]] = None,
) -> torch.Tensor:
    """
    Convert sliding window `window_size` to an equivalent "`arbitrary`" mask.
    For "`causal`" mask type, the sliding window diagonal is aligned to the top left corner,
    and for other mask types, the bottom right corner.

    Parameters
    ----------
    window_size: Tuple[int, int]
        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
        window and causal mask specifically. Both `causal` and `causal_bottom_right` masks
        map to `window_size = (-1, 0)` and Transformer Engine distinguishes them based on
        `attn_mask_type`.
    max_seqlen_q: int
        Maximum sequence length for queries.
    max_seqlen_kv: int
        Maximum sequence length for keys and values.
    attn_mask_type: str, default = `no_mask`
        Attention mask type, {"`no_mask`", "`padding`", "`causal`", "`padding_causal`",
        "`causal_bottom_right`", "`padding_causal_bottom_right`", "`arbitrary`"}
    attention_mask: Optional[Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]],
        default = `None`
        Boolean tensor(s) used to mask out attention softmax input.

    Returns
    ----------
    attention_mask: torch.Tensor
        Combined `attention_mask` (input) and sliding window attention mask.
        The shape is [max_seqlen_q, max_seqlen_kv] when input `attention_mask` is None;
        else, the same shape as input `attention_mask`.
    """
    mask = torch.ones(max_seqlen_q, max_seqlen_kv, dtype=torch.bool, device="cuda")
    if attn_mask_type in ["causal"]:
        left = window_size[0] if window_size[0] != -1 else max_seqlen_q
        right = window_size[1] if window_size[1] != -1 else max_seqlen_q
        mask_upper = torch.triu(mask, diagonal=-left)
        mask_lower = torch.tril(mask_upper, diagonal=right)
    else:
        left = window_size[0] if window_size[0] != -1 else max_seqlen_kv
        right = window_size[1] if window_size[1] != -1 else max_seqlen_kv
        mask_upper = torch.triu(mask, diagonal=max_seqlen_kv - max_seqlen_q - left)
        mask_lower = torch.tril(mask_upper, diagonal=max_seqlen_kv - max_seqlen_q + right)
    attn_mask_type = "arbitrary"
    mask = mask_lower.logical_not()
    if attention_mask is not None:
        mask = torch.logical_and(attention_mask, mask)
    return attn_mask_type, mask


1084
1085
1086
1087
1088
@torch.no_grad()
def get_alibi(
    num_heads: int,
    max_seqlen_q: int,
    max_seqlen_kv: int,
1089
1090
    actual_seqlens_q: Optional[torch.Tensor] = None,
    actual_seqlens_kv: Optional[torch.Tensor] = None,
1091
1092
    alibi_slopes: Optional[torch.Tensor] = None,
    bias_dtype: Optional[torch.dtype] = None,
1093
    bottom_right_alignment: bool = True,
1094
) -> Tuple[torch.Tensor, torch.Tensor]:
1095
    """
1096
1097
1098
1099
1100
1101
1102
1103
    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.
1104
1105
1106
1107
    actual_seqlens_q: Optional[torch.Tensor], default = `None`
        Actual sequence lengths for queries, in shape [batch_size].
    actual_seqlens_kv: Optional[torch.Tensor], default = `None`
        Actual sequence lengths for keys and values, in shape [batch_size].
1108
1109
1110
1111
    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.
1112
1113
1114
    bottom_right_alignment: bool, default = `True`
        Whether to align the diagonal of the ALiBi bias to the bottom right corner of
        the matrix (`True`) or top left (`False`).
1115

1116
1117
1118
1119
1120
    Returns
    ----------
    alibi_slopes: torch.Tensor
        ALiBi slopes in FP32 and shape [num_heads] or [batch_size, num_heads].
    alibi_bias: torch.Tensor
1121
1122
1123
1124
1125
1126
        ALiBi bias in FP32 or `bias_dtype`. Its shape is
        (1) [1, num_heads, max_seqlen_q, max_seqlen_kv] if `alibi_slopes` is in [num_heads] shape,
        and `actual_seqlens_q` and `actual_seqlens_kv` are `None`; or
        (2) [batch_size, num_heads, max_seqlen_q, max_seqlen_kv] if `alibi_slopes` is in
        [batch_size, num_heads] shape, or, if `alibi_slopes` is in [num_heads] shape and
        `actual_seqlens_q` and `actual_seqlens_kv` are not `None`.
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
    """
    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])
1150
        elif _alibi_cache["_alibi_slopes"].dim() == 2:
1151
            slopes_shape = torch.Size([*_alibi_cache["_alibi_slopes"].shape[:], 1, 1])
1152
1153
1154
        else:
            raise ValueError("ALiBi slopes cannot exceed 2 dimensions.")

1155
        bias = torch.arange(max_seqlen_q, dtype=torch.int32, device="cuda").view(
1156
            1, 1, max_seqlen_q, 1
1157
1158
        ) - torch.arange(max_seqlen_kv, dtype=torch.int32, device="cuda").view(
            1, 1, 1, max_seqlen_kv
1159
        )
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
        if actual_seqlens_q is None and actual_seqlens_kv is None:
            if bottom_right_alignment:
                bias = bias + max_seqlen_kv - max_seqlen_q
        elif actual_seqlens_q is not None and actual_seqlens_kv is not None:
            batch_size = actual_seqlens_q.shape[0]
            bias = bias.expand(batch_size, 1, max_seqlen_q, max_seqlen_kv)
            if bottom_right_alignment:
                bias = bias + (actual_seqlens_kv - actual_seqlens_q).view(batch_size, 1, 1, 1)
        else:
            assert (
                False
            ), "actual_seqlens_q and actual_seqlens_kv need to be both None or torch.Tensors!"
1172
1173
1174
        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
1175
        _alibi_cache["_bottom_right_alignment"] = bottom_right_alignment
1176
1177
1178
1179
1180
        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"]
1181
1182
1183
1184
1185
1186
1187
1188
1189


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)
1190
    reduced_mask = mask.logical_not().sum(dim=1)
1191
1192
1193
1194
1195
1196
    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

1197

1198
1199
1200
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
1201
1202
1203
    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.
1204
1205
1206
1207
    """
    mask = mask.squeeze(1).squeeze(1)
    bs, seqlen = mask.shape

1208
    reduced_mask = mask.logical_not().sum(dim=1)
1209
1210
1211
1212
1213
    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)
1214
    indices = mask.logical_not().nonzero()
1215
1216
1217
1218
    indices = indices.unsqueeze(-1)

    num_nonzeros = indices.shape[0]
    pad_amount = bs * seqlen - num_nonzeros
1219
1220
1221
    indices = F.pad(
        input=indices, pad=(0, 0, 0, 0, 0, pad_amount), mode="constant", value=float(bs * seqlen)
    )
1222
1223
1224
1225

    return cu_seqlens, indices


1226
1227
1228
1229
1230
1231
1232
1233
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]
1234
1235
    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")
1236
1237
1238

    num_nonzeros = indices.shape[0]
    pad_amount = bs * max_seqlen - num_nonzeros
1239
1240
1241
1242
1243
1244
    indices = F.pad(
        input=indices,
        pad=(0, 0, 0, 0, 0, pad_amount),
        mode="constant",
        value=float(bs * max_seqlen),
    )
1245
1246
1247

    return indices

1248

1249
_cu_seqlens_cache = {}
1250
1251


1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
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.

    """
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
    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)]
1272
1273


1274
@torch.compile
1275
1276
1277
1278
1279
1280
1281
1282
def pack_tensor(
    indices: torch.Tensor,
    tensor: torch.Tensor,
) -> torch.Tensor:
    """
    Packs the given tensor using the `indices`.
    """
    padding_indice = torch.zeros(
1283
1284
        1, tensor.shape[1], tensor.shape[2], dtype=tensor.dtype, device=tensor.device
    )
1285
    indices = indices.repeat(1, tensor.shape[1], tensor.shape[2])
1286
1287
1288
1289
1290
1291
1292
1293
    if isinstance(tensor, Float8Tensor):
        tensor_data = torch.cat((tensor._data, padding_indice), dim=0)

        packed = Float8Tensor.make_like(tensor, data=torch.gather(tensor_data, 0, indices))
    else:
        tensor = torch.cat((tensor, padding_indice), dim=0)

        packed = torch.gather(tensor, 0, indices)
1294
1295
1296
    return packed


1297
@torch.compile
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
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


1311
@torch.compile
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
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


1327
@torch.compile
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
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(
1338
1339
        dim0 + 1, tensor.shape[1], tensor.shape[2], dtype=tensor.dtype, device=tensor.device
    )
1340
1341
1342
1343
1344
1345
    if isinstance(tensor, Float8Tensor):
        unpacked.scatter_(0, indices, tensor._data)
        unpacked = Float8Tensor.make_like(tensor, data=unpacked[0:-1, :, :])
    else:
        unpacked.scatter_(0, indices, tensor)
        unpacked = unpacked[0:-1, :, :]
1346
1347
1348
    return unpacked


1349
@torch.compile
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
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


1364
@torch.compile
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
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.
    """
1385

1386
1387
    @staticmethod
    def forward(
1388
        ctx, indices: torch.Tensor, *tensors: Tuple[torch.Tensor, ...]
1389
    ) -> Union[Tuple[torch.Tensor, ...], torch.Tensor]:
1390
        # pylint: disable=missing-function-docstring
1391
        assert 1 <= len(tensors) <= 3, f"Packing {len(tensors)} tensors not supported."
1392
        ctx.save_for_backward(indices)
1393
1394
1395
1396
1397
1398
1399
1400
1401
        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, ...]):
1402
        # pylint: disable=missing-function-docstring
1403
        (indices,) = ctx.saved_tensors
1404
        if len(grad_outputs) == 1:
1405
            return None, unpack_tensor(indices, ctx.dim0, *grad_outputs)
1406
        if len(grad_outputs) == 2:
1407
1408
            return None, *unpack_2_tensors(indices, ctx.dim0, *grad_outputs)
        return None, *unpack_3_tensors(indices, ctx.dim0, *grad_outputs)
1409
1410
1411
1412
1413
1414


class UnpackTensor(torch.autograd.Function):
    """
    Autograd function to unpack a tensor.
    """
1415

1416
1417
1418
1419
1420
1421
1422
    @staticmethod
    def forward(
        ctx,
        indices: torch.Tensor,
        dim0: int,
        tensor: torch.Tensor,
    ) -> torch.Tensor:
1423
        # pylint: disable=missing-function-docstring
1424
        ctx.save_for_backward(indices)
1425
1426
1427
1428
        return unpack_tensor(indices, dim0, tensor)

    @staticmethod
    def backward(ctx, grad_output):
1429
        # pylint: disable=missing-function-docstring
1430
1431
        (indices,) = ctx.saved_tensors
        return None, None, pack_tensor(indices, grad_output)
1432
1433


1434
1435
1436
def flash_attn_p2p_communicate(
    rank, send_tensor, send_dst, recv_tensor, recv_src, cp_group, batch_p2p_comm
):
1437
    """Point-to-point communications of KV and dKV in Attention with context parallelism"""
1438
1439
1440
1441
    send_recv_ops = []

    if batch_p2p_comm:
        if rank % 2 == 0:
1442
1443
1444
1445
1446
1447
            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
            )
1448
1449
1450
            send_recv_ops.append(send_op)
            send_recv_ops.append(recv_op)
        else:
1451
1452
1453
1454
1455
1456
            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
            )
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
            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


1476
@jit_fuser
1477
1478
1479
1480
1481
def flash_attn_fwd_out_correction(
    out: torch.Tensor,
    out_per_step: torch.Tensor,
    softmax_lse: torch.Tensor,
    softmax_lse_per_step: torch.Tensor,
1482
1483
    movedim_src: int,
    movedim_dst: int,
1484
):
1485
    """Merge partial outputs of each step in Attention with context parallelism"""
1486
1487
1488
    softmax_lse_corrected_exp = torch.exp(softmax_lse_per_step - softmax_lse).movedim(
        movedim_src, movedim_dst
    )
1489
    softmax_lse_corrected_exp = softmax_lse_corrected_exp.unsqueeze(-1)
1490
    out_corrected = out_per_step * softmax_lse_corrected_exp
1491
1492
1493
    out.add_(out_corrected)


1494
@jit_fuser
1495
1496
1497
1498
def flash_attn_fwd_softmax_lse_correction(
    softmax_lse: torch.Tensor,
    softmax_lse_per_step: torch.Tensor,
):
1499
    """Merge softmax stats of each step in Attention with context parallelism"""
1500
1501
1502
1503
    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)
1504
1505


1506
1507
@jit_fuser
def get_cu_seqlens_on_cp_rank(
1508
1509
1510
1511
1512
1513
    cu_seqlens: torch.Tensor,
    cu_seqlens_padded_on_cp_rank: torch.Tensor,
    cp_size: int,
    cp_rank: int,
    first_half: bool,
    second_half: bool,
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
):
    """Compute cu_seqlens of a context parallelism rank"""
    seqlens = cu_seqlens[1:] - cu_seqlens[:-1]
    seqlens_padded = (cu_seqlens_padded_on_cp_rank[1:] - cu_seqlens_padded_on_cp_rank[:-1]) // 2
    zeros = torch.zeros_like(seqlens)
    cu_seqlens_on_cp_rank = torch.zeros_like(cu_seqlens)
    if first_half:
        seqlens_1 = seqlens - cp_rank * seqlens_padded
        seqlens_1 = seqlens_1.clamp(zeros, seqlens_padded)
        cu_seqlens_on_cp_rank[1:].add_(seqlens_1)
    if second_half:
        seqlens_2 = seqlens - (2 * cp_size - cp_rank - 1) * seqlens_padded
        seqlens_2 = seqlens_2.clamp(zeros, seqlens_padded)
        cu_seqlens_on_cp_rank[1:].add_(seqlens_2)
    cu_seqlens_on_cp_rank.cumsum_(dim=0)
    return cu_seqlens_on_cp_rank


1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
@torch.compile
def get_seq_chunk_ids_for_reordering(cp_size, device, to_contiguous):
    """
    Context parallelism assigns two discontiguous sequence chunks to each GPU for load balancing.
    To make sure tokens are ordered correctly for compute, we need to reorder sequence chunks
    before or after CP communications (e.g., all-gather, all-to-all). This function is to compute
    sequence chunk ids for reordering.
    """
    chunk_ids = torch.empty(2 * cp_size, dtype=torch.int32, device=device)
    if to_contiguous:
        for rank in range(cp_size):
            chunk_ids[rank] = 2 * rank
            chunk_ids[rank + cp_size] = 2 * cp_size - 2 * rank - 1
    else:
        for rank in range(cp_size):
            chunk_ids[2 * rank] = rank
            chunk_ids[2 * rank + 1] = 2 * cp_size - rank - 1
    return chunk_ids


@torch.compile
def reorder_seq_chunks_for_a2a(x, chunk_ids_for_a2a, seq_dim, cp_size, before_attn):
    """Reorder sequence chunk for A2A communication."""
    if before_attn:
        # [cp, b, s, np//cp, hn] -> [b, cp, s, np//cp, hn]
        # or [cp, s, b, np//cp, hn] -> [cp, s, b, np//cp, hn]
        x = x.movedim(0, seq_dim).contiguous()
        # [b, cp, s, np//cp, hn] -> [b, cp*2, s//2, np//cp, hn]
        # or [cp, s, b, np//cp, hn] -> [cp*2, s//2, b, np//cp, hn]
        x = x.view(*x.shape[:seq_dim], cp_size * 2, -1, *x.shape[(seq_dim + 2) :])
        # reorder the sequence chunks
        x = torch.index_select(x, dim=seq_dim, index=chunk_ids_for_a2a)
    else:
        # [b, cp*2, s//2, np//cp, hn] -> [cp*2, b, s//2, np//cp, hn]
        # or [cp*2, s//2, b, np//cp, hn] -> [cp*2, s//2, b, np//cp, hn]
        x = x.movedim(seq_dim, 0).contiguous()
        # reorder the sequence chunks
        x = torch.index_select(x, dim=0, index=chunk_ids_for_a2a)
        # [cp*2, b, s//2, np//cp, hn] -> [cp, 2, b, s//2, np//cp, hn]
        # or [cp*2, s//2, b, np//cp, hn] -> [cp, 2, s//2, b, np//cp, hn]
        x = x.view(cp_size, 2, *x.shape[1:])
    return x


def flash_attn_a2a_communicate(
    a2a_inputs: Union[torch.Tensor, List[torch.Tensor]],
    chunk_ids_for_a2a: torch.Tensor,
    seq_dim: int,
    cp_size: int,
    cp_group: dist_group_type,
    cp_stream: torch.cuda.Stream,
    before_attn: bool,
) -> Union[torch.Tensor, List[torch.Tensor]]:
    """A2A communication for context parallelism."""
    a2a_inputs = [a2a_inputs] if not isinstance(a2a_inputs, list) else a2a_inputs
    a2a_outputs, a2a_reqs = [None] * len(a2a_inputs), [None] * len(a2a_inputs)
    if before_attn:
        for i in range(len(a2a_inputs) + 2):
            if 0 < i < len(a2a_inputs) + 1:
                a2a_outputs[i - 1] = torch.empty_like(a2a_inputs[i - 1])
                a2a_reqs[i - 1] = torch.distributed.all_to_all_single(
                    a2a_outputs[i - 1], a2a_inputs[i - 1], group=cp_group, async_op=True
                )
            if i > 1:
                with torch.cuda.stream(cp_stream):
                    a2a_reqs[i - 2].wait()
                    x = a2a_outputs[i - 2]
                    # reorder the sequence chunks
                    x = reorder_seq_chunks_for_a2a(
                        x, chunk_ids_for_a2a, seq_dim, cp_size, before_attn
                    )
                    # [b, cp*2, s//2, np//cp, hn] -> [b, cp*s, np//cp, hn]
                    # or [cp*2, s//2, b, np//cp, hn] -> [cp*s, b, np//cp, hn]
                    a2a_outputs[i - 2] = x.view(*x.shape[:seq_dim], -1, *x.shape[(seq_dim + 2) :])
            if i < len(a2a_inputs):
                x = a2a_inputs[i]
                # [b, s, np, hn] -> [b, s, cp, np//cp, hn]
                # or [s, b, np, hn] -> [s, b, cp, np//cp, hn]
                x = x.view(*x.shape[:-2], cp_size, x.shape[-2] // cp_size, x.shape[-1])
                # [b, s, cp, np//cp, hn] -> [cp, b, s, np//cp, hn]
                # or [s, b, cp, np//cp, hn] -> [cp, s, b, np//cp, hn]
                a2a_inputs[i] = x.movedim(-3, 0).contiguous()
    else:
        for i in range(len(a2a_inputs) + 2):
            if 0 < i < len(a2a_inputs) + 1:
                a2a_outputs[i - 1] = torch.empty_like(a2a_inputs[i - 1])
                a2a_reqs[i - 1] = torch.distributed.all_to_all_single(
                    a2a_outputs[i - 1], a2a_inputs[i - 1], group=cp_group, async_op=True
                )
            if i < len(a2a_inputs):
                x = a2a_inputs[i]
                # [b, cp*s, np//cp, hn] -> [b, cp*2, s//2, np//cp, hn]
                # or [cp*s, b, np//cp, hn] -> [cp*2, s//2, b, np//cp, hn]
                x = x.view(*x.shape[:seq_dim], cp_size * 2, -1, *x.shape[(seq_dim + 1) :])
                # reorder the sequence chunks
                a2a_inputs[i] = reorder_seq_chunks_for_a2a(
                    x, chunk_ids_for_a2a, seq_dim, cp_size, before_attn
                )
            if i > 1:
                with torch.cuda.stream(cp_stream):
                    a2a_reqs[i - 2].wait()
                    x = a2a_outputs[i - 2]
                    # [cp, 2, b, s//2, np//cp, hn] -> [b, 2, s//2, cp, np//cp, hn]
                    # or [cp, 2, s//2, b, np//cp, hn] -> [2, s//2, b, cp, np//cp, hn]
                    x = x.movedim(0, -3).movedim(0, seq_dim).contiguous()
                    # [b, 2, s//2, cp, np//cp, hn] -> [b*s, np, hn]
                    # or [2, s//2, b, cp, np//cp, hn] -> [s*b, np, hn]
                    a2a_outputs[i - 2] = x.view(-1, x.shape[-3] * x.shape[-2], x.shape[-1])
    torch.cuda.current_stream().wait_stream(cp_stream)
    return a2a_outputs[0] if len(a2a_inputs) == 1 else a2a_outputs


1644
class AttnFuncWithCPAndKVP2P(torch.autograd.Function):
1645
    """
1646
1647
1648
    Attention implementation with context parallelism. Exchange KV between CP ranks
    with P2P in ring topology. Split attention compute into multiple steps, and overlap
    current-step compute with next-step communication.
1649
1650
1651
1652
1653

    This implementation also supports hierarchical CP, which parallelizes attention
    heads in low-level CP groups and parallelizes sequence dimension in high-level CP
    groups. For more details, please refer to `LongVILA <https://arxiv.org/abs/2408.10188>`_
    and `USP <https://arxiv.org/abs/2405.07719>`_.
1654
1655
1656
    """

    @staticmethod
1657
1658
1659
1660
1661
1662
1663
    def forward(
        ctx,
        is_training,
        q,
        k,
        v,
        cu_seqlens_q,
1664
        cu_seqlens_kv,
1665
        max_seqlen_q,
1666
        max_seqlen_kv,
1667
1668
        cu_seqlens_q_padded,
        cu_seqlens_kv_padded,
1669
1670
1671
1672
1673
1674
1675
1676
        dropout_p,
        softmax_scale,
        qkv_format,
        attn_mask_type,
        attn_bias_type,
        attn_bias,
        deterministic,
        use_fused_attention,
1677
1678
        fp8,
        fp8_meta,
1679
1680
1681
        cp_group,
        cp_global_ranks,
        cp_stream,
1682
    ):
1683
        # pylint: disable=missing-function-docstring
1684
1685
1686
        if softmax_scale is None:
            softmax_scale = q.shape[-1] ** (-0.5)

1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
        if isinstance(cp_group, list):
            assert (
                qkv_format != "thd"
            ), f"{qkv_format} format is not supported with hierarchical CP implementation yet!"
            assert attn_bias_type == "no_bias", (
                f"{attn_bias_type} bias type is not supported with hierarchical CP implementation"
                " yet!"
            )
            cp_group_a2a = cp_group[0]
            cp_size_a2a = get_distributed_world_size(cp_group_a2a)
            rank_a2a = get_distributed_rank(cp_group_a2a)
            cp_group = cp_group[1]
        else:
            cp_group_a2a = None
            cp_size_a2a = 1
            rank_a2a = 0

1704
1705
        cp_size = get_distributed_world_size(cp_group)
        rank = get_distributed_rank(cp_group)
1706
1707
        send_dst = cp_global_ranks[(rank + 1) % cp_size * cp_size_a2a + rank_a2a]
        recv_src = cp_global_ranks[(rank - 1) % cp_size * cp_size_a2a + rank_a2a]
1708
1709
        batch_p2p_comm = int(os.getenv("NVTE_BATCH_MHA_P2P_COMM", "0")) or (cp_size == 2)

1710
1711
        causal = "causal" in attn_mask_type
        padding = "padding" in attn_mask_type
1712

1713
        seq_dim = None
1714
        if qkv_format in ["bshd", "sbhd"]:
1715
            seq_dim = qkv_format.index("s")
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
            qkv_layout = qkv_format + "_" + qkv_format[:-2] + "2" + qkv_format[-2:]
        else:
            qkv_layout = qkv_format + "_" + qkv_format + "_" + qkv_format

        pad_between_seqs_q = not torch.equal(cu_seqlens_q_padded, cu_seqlens_q)
        pad_between_seqs_kv = not torch.equal(cu_seqlens_kv_padded, cu_seqlens_kv)
        max_seqlen_q = max_seqlen_q // cp_size
        max_seqlen_kv = max_seqlen_kv // cp_size
        cu_seqlens_q_padded = cu_seqlens_q_padded // cp_size
        cu_seqlens_kv_padded = cu_seqlens_kv_padded // cp_size
        cu_seqlens_q_per_step = [None for _ in range(cp_size)]
        cu_seqlens_kv_per_step = [None for _ in range(cp_size)]
1728

1729
1730
1731
        fused_attn_qkv_dtype = None
        fused_attn_backend = None
        amax_per_step = None
1732
1733
1734
1735
        qkv_dtype = q.dtype
        # "fp8_mha" decides outputs in fp8, while inputs are inferred from the real dtype
        is_input_fp8 = False
        is_output_fp8 = fp8_meta is not None and fp8_meta["recipe"].fp8_mha
1736
1737
1738
1739
1740
        if fp8:
            if use_fused_attention:
                fp8_dtype_forward = get_fp8_te_dtype(fp8_meta["recipe"], fprop_tensor=True)
                fused_attn_qkv_dtype = fp8_dtype_forward
                fused_attn_backend = FusedAttnBackend["FP8"]
1741
1742
1743
1744
1745
                assert isinstance(k, q.__class__) and isinstance(
                    v, q.__class__
                ), "q, k, and v must have the same type."
                is_input_fp8 = isinstance(q, Float8Tensor)
                if is_input_fp8:
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
                    fp8_meta["scaling_fwd"].scale_inv[META_QKV] = q._scale_inv
                    q_fp8, k_fp8, v_fp8 = q, k, v
                    q, k, v = q_fp8._data, k_fp8._data, v_fp8._data
                else:
                    q_f16, k_f16, v_f16 = q, k, v
                    if cp_size_a2a == 1 or int(os.getenv("NVTE_FP8_DPA_BWD", "1")):
                        q = cast_to_fp8(q_f16, fp8_meta["scaling_fwd"], META_QKV, fp8_dtype_forward)
                    if int(os.getenv("NVTE_FP8_DPA_BWD", "1")):
                        k, v = [
                            cast_to_fp8(x, fp8_meta["scaling_fwd"], META_QKV, fp8_dtype_forward)
                            for x in [k_f16, v_f16]
                        ]
                fp8_meta_kwargs = {}
                fp8_meta_kwargs["d_scale_qkv"] = fp8_meta["scaling_fwd"].scale_inv
                fp8_meta_kwargs["d_scale_qkv_offset"] = META_QKV
                fp8_meta_kwargs["d_scale_s"] = fp8_meta["scaling_fwd"].scale_inv
                fp8_meta_kwargs["d_scale_s_offset"] = META_S
                fp8_meta_kwargs["q_scale_s"] = fp8_meta["scaling_fwd"].scale
                fp8_meta_kwargs["q_scale_s_offset"] = META_S
                fp8_meta_kwargs["q_scale_o"] = fp8_meta["scaling_fwd"].scale
                fp8_meta_kwargs["q_scale_o_offset"] = META_O_CP
                amax_per_step = torch.zeros((2, cp_size), dtype=torch.float32, device=q.device)
            else:
                assert False, "FP8 is only supported with Fused Attention!"
        else:
            q_f16 = q
            if use_fused_attention:
                fp8_meta_kwargs = {}
                fused_attn_qkv_dtype = TE_DType[q.dtype]
                fused_attn_backend = FusedAttnBackend["F16_arbitrary_seqlen"]

        if cp_size_a2a > 1:
            chunk_ids_for_a2a = get_seq_chunk_ids_for_reordering(cp_size_a2a, q.device, True)
            q, k, v = flash_attn_a2a_communicate(
                [q, k, v], chunk_ids_for_a2a, seq_dim, cp_size_a2a, cp_group_a2a, cp_stream, True
            )
            if not fp8:
                q_f16 = q
1784
            elif not is_input_fp8 and not int(os.getenv("NVTE_FP8_DPA_BWD", "1")):
1785
1786
1787
                q_f16 = q
                q = cast_to_fp8(q_f16, fp8_meta["scaling_fwd"], META_QKV, fp8_dtype_forward)

1788
1789
1790
        assert qkv_format == "thd" or (
            q.shape[seq_dim] % 2 == 0 and k.shape[seq_dim] % 2 == 0
        ), "Sequence length per GPU needs to be divisible by 2!"
1791
        if causal:
1792
1793
            if qkv_format == "bshd":
                # [b, s, np, hn] -> [b, 2, s//2, np, hn]
1794
                q, k, v = [x.view(x.shape[0], 2, x.shape[1] // 2, *x.shape[2:]) for x in [q, k, v]]
1795
1796
            elif qkv_format == "sbhd":
                # [s, b, np, hn] -> [2, s//2, b, np, hn]
1797
                q, k, v = [x.view(2, x.shape[0] // 2, *x.shape[1:]) for x in [q, k, v]]
1798
1799
1800
        total_tokens_kv = None if qkv_format != "thd" else k.shape[0]
        # remove padded tokens at the end
        k, v = [x if qkv_format != "thd" else x[: cu_seqlens_kv_padded[-1]] for x in [k, v]]
1801
        if attn_bias is not None:
1802
            assert len(attn_bias.shape) == 4, (
1803
1804
1805
                "Only support bias shape of [b, h, sq, sk] for forward, "
                "and [1, h, sq, sk] for backward!"
            )
1806
1807
1808
            assert (
                attn_bias.shape[-2] % 2 == 0 and attn_bias.shape[-1] % (2 * cp_size) == 0
            ), "Sequence length does not meet divisible requirements!"
1809
            # [b, np, sq, sk] -> [b, np, 2, sq//2, 2*cp, sk//(2*cp)]
1810
1811
1812
1813
1814
1815
            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),
1816
1817
            )
            # [b, np, sq, sk] -> [b, np, sq, 2*cp, sk//(2*cp)]
1818
1819
            attn_bias = attn_bias.view(
                *attn_bias.shape[:-1], 2 * cp_size, attn_bias.shape[-1] // (2 * cp_size)
1820
            )
1821
        assert q.shape[-1] % 8 == 0, "hidden size per attention head should be multiple of 8"
1822
1823
1824
1825

        softmax_lse_in_packed_format = not use_fused_attention and (
            _flash_attn_2_6_0_plus or _use_flash_attn_3
        )
1826
        flash_attn_fwd = None
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
        if not use_fused_attention:
            fa_forward_kwargs = {"softmax_scale": softmax_scale}
            if _use_flash_attn_3:
                flash_attn_fwd = flash_attn_varlen_fwd_v3
                fa_forward_kwargs["window_size"] = (-1, 0) if causal else (-1, -1)
            else:
                flash_attn_fwd = flash_attn_varlen_fwd
                fa_forward_kwargs["dropout_p"] = dropout_p
                fa_forward_kwargs["return_softmax"] = False
                if _flash_attn_2_3_plus:
                    fa_forward_kwargs["window_size"] = (-1, 0) if causal else (-1, -1)
                if _flash_attn_2_4_plus:
                    fa_forward_kwargs["alibi_slopes"] = None
                if _flash_attn_2_5_7_plus:
                    fa_forward_kwargs["block_table"] = None
1842

1843
1844
1845
        # Flash Attn inputs
        q_inputs = [None, None]
        kv_inputs = [None, None]
1846
        attn_bias_inputs = [None, None]
1847
1848
1849
1850
        # 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)]
1851
        attn_biases = [None for _ in range(cp_size)]
1852
1853
1854
1855
1856
1857
1858

        # 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)]
1859
1860
1861
1862
        if use_fused_attention and qkv_format in ["bshd", "sbhd"]:
            p2p_comm_buffers[0] = torch.cat((k.unsqueeze(-3), v.unsqueeze(-3)), dim=-3)
        else:
            p2p_comm_buffers[0] = torch.cat((k.unsqueeze(0), v.unsqueeze(0)), dim=0)
1863
1864
        send_recv_reqs = [[], []]

1865
1866
        softmax_lse_ = None
        out = None
1867
        for i in range(cp_size + 1):
1868
            if i < cp_size:
1869
                with torch.cuda.stream(flash_attn_streams[i % 2]):
1870
                    # wait until KV is received
1871
                    for req in send_recv_reqs[(i + 1) % 2]:
1872
1873
                        req.wait()

1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
                    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,
                        )

1886
                    if not fp8 or is_input_fp8 or int(os.getenv("NVTE_FP8_DPA_BWD", "1")):
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
                        kv_inputs[i % 2] = p2p_comm_buffers[i]
                    else:
                        # KV exchange is in BF16/FP16, cast received KV in each step
                        kv_inputs[i % 2] = cast_to_fp8(
                            p2p_comm_buffers[i],
                            fp8_meta["scaling_fwd"],
                            META_QKV,
                            fp8_dtype_forward,
                        )
                    if fp8 and use_fused_attention:
1897
1898
1899
1900
                        fp8_meta_kwargs["amax_s"] = amax_per_step
                        fp8_meta_kwargs["amax_s_offset"] = i
                        fp8_meta_kwargs["amax_o"] = amax_per_step
                        fp8_meta_kwargs["amax_o_offset"] = cp_size + i
1901
1902
                    if causal:
                        if i == 0:
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
                            if pad_between_seqs_q:
                                cu_seqlens_q_per_step[i] = get_cu_seqlens_on_cp_rank(
                                    cu_seqlens_q, cu_seqlens_q_padded, cp_size, rank, True, True
                                )
                            else:
                                cu_seqlens_q_per_step[i] = cu_seqlens_q // cp_size
                            if pad_between_seqs_kv:
                                cu_seqlens_kv_per_step[i] = get_cu_seqlens_on_cp_rank(
                                    cu_seqlens_kv, cu_seqlens_kv_padded, cp_size, rank, True, True
                                )
                            else:
                                cu_seqlens_kv_per_step[i] = cu_seqlens_kv // cp_size
1915
                            if use_fused_attention:
1916
1917
                                if qkv_format == "bshd":
                                    # [b, 2, sq//2, np, hn] -> [b, sq, np, hn]
1918
                                    q_inputs[i % 2] = q.view(q.shape[0], -1, *q.shape[-2:])
1919
                                    # [b, 2, sk//2, 2, np, hn] -> [b, sk, 2, np, hn]
1920
                                    kv_inputs[i % 2] = kv_inputs[i % 2].view(
1921
                                        k.shape[0], -1, 2, *k.shape[-2:]
1922
                                    )
1923
1924
                                elif qkv_format == "sbhd":
                                    # [2, sq//2, b, np, hn] -> [sq, b, np, hn]
1925
                                    q_inputs[i % 2] = q.view(-1, *q.shape[-3:])
1926
1927
1928
1929
                                    # [2, sk//2, b, 2, np, hn] -> [sk, b, 2, np, hn]
                                    kv_inputs[i % 2] = kv_inputs[i % 2].view(
                                        -1, k.shape[2], 2, *k.shape[-2:]
                                    )
1930
                                elif qkv_format == "thd":
1931
                                    q_inputs[i % 2] = q
1932
1933
                                if attn_bias is not None:
                                    idx = (rank - i) % cp_size
1934
1935
1936
1937
1938
1939
                                    attn_bias_inputs[i % 2] = torch.cat(
                                        (
                                            attn_bias[..., idx, :],
                                            attn_bias[..., (2 * cp_size - idx - 1), :],
                                        ),
                                        dim=-1,
1940
                                    ).contiguous()
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
                                out_per_step[i], aux_ctx_tensors = fused_attn_fwd(
                                    is_training,
                                    max_seqlen_q,
                                    max_seqlen_kv,
                                    cu_seqlens_q_per_step[i],
                                    cu_seqlens_kv_per_step[i],
                                    q_inputs[i % 2],
                                    (
                                        kv_inputs[i % 2][..., 0, :, :]
                                        if qkv_format in ["bshd", "sbhd"]
                                        else kv_inputs[i % 2][0]
                                    ),
                                    (
                                        kv_inputs[i % 2][..., 1, :, :]
                                        if qkv_format in ["bshd", "sbhd"]
                                        else kv_inputs[i % 2][1]
                                    ),
                                    fused_attn_qkv_dtype,
                                    fused_attn_backend,
                                    attn_scale=softmax_scale,
                                    dropout=dropout_p,
                                    qkv_layout=qkv_layout,
                                    attn_mask_type=attn_mask_type,
                                    attn_bias_type=attn_bias_type,
                                    attn_bias=attn_bias_inputs[i % 2],
                                    cu_seqlens_q_padded=cu_seqlens_q_padded,
                                    cu_seqlens_kv_padded=cu_seqlens_kv_padded,
                                    **fp8_meta_kwargs,
1969
                                )
1970
1971
1972
1973
1974
                                if fp8:
                                    softmax_lse_per_step[i], _, rng_states[i] = aux_ctx_tensors
                                else:
                                    softmax_lse_per_step[i], rng_states[i], *rest = aux_ctx_tensors
                                    attn_biases[i] = rest[0] if len(rest) > 0 else None
1975
1976
                            else:
                                # [b, 2, sq//2, np, hn] -> [b*sq, np, hn]
1977
                                q_inputs[i % 2] = q.view(-1, *q.shape[-2:])
1978
                                # [2, b, 2, sk//2, np, hn] -> [2, b*sk, np, hn]
1979
                                kv_inputs[i % 2] = kv_inputs[i % 2].view(2, -1, *k.shape[-2:])
1980
                                fa_outputs = flash_attn_fwd(
1981
1982
1983
                                    q_inputs[i % 2],
                                    kv_inputs[i % 2][0],
                                    kv_inputs[i % 2][1],
1984
1985
                                    cu_seqlens_q_per_step[i],
                                    cu_seqlens_kv_per_step[i],
1986
                                    max_seqlen_q,
1987
                                    max_seqlen_kv,
1988
                                    causal=True,
1989
                                    **fa_forward_kwargs,
1990
                                )
1991
1992
1993
1994
                                out_per_step[i] = fa_outputs[4]
                                softmax_lse_per_step[i] = fa_outputs[5]
                                if not _use_flash_attn_3:
                                    rng_states[i] = fa_outputs[7]
1995
                        elif i <= rank:
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
                            if pad_between_seqs_q:
                                cu_seqlens_q_per_step[i] = get_cu_seqlens_on_cp_rank(
                                    cu_seqlens_q, cu_seqlens_q_padded, cp_size, rank, True, True
                                )
                            else:
                                cu_seqlens_q_per_step[i] = cu_seqlens_q // cp_size
                            if pad_between_seqs_kv:
                                cu_seqlens_kv_per_step[i] = get_cu_seqlens_on_cp_rank(
                                    cu_seqlens_kv,
                                    cu_seqlens_kv_padded,
                                    cp_size,
                                    (rank - i) % cp_size,
                                    True,
                                    False,
                                )
                            else:
                                cu_seqlens_kv_per_step[i] = cu_seqlens_kv // (cp_size * 2)
2013
                            if use_fused_attention:
2014
2015
                                if qkv_format == "bshd":
                                    # [b, 2, sq//2, np, hn] -> [b, sq, np, hn]
2016
                                    q_inputs[i % 2] = q.view(q.shape[0], -1, *q.shape[-2:])
2017
2018
                                    # [b, 2, sk//2, 2, np, hn] -> [b, sk//2, 2, np, hn]
                                    kv_inputs[i % 2] = kv_inputs[i % 2][:, 0, ...].contiguous()
2019
2020
                                elif qkv_format == "sbhd":
                                    # [2, sq//2, b, np, hn] -> [sq, b, np, hn]
2021
                                    q_inputs[i % 2] = q.view(-1, *q.shape[-3:])
2022
2023
                                    # [2, sk//2, b, 2, np, hn] -> [sk//2, b, 2, np, hn]
                                    kv_inputs[i % 2] = kv_inputs[i % 2][0].contiguous()
2024
                                elif qkv_format == "thd":
2025
                                    q_inputs[i % 2] = q
2026
                                    # [2, t, np, hn] -> [2, t/2, np, hn]
2027
                                    kv_inputs[i % 2] = tex.thd_read_half_tensor(
2028
                                        kv_inputs[i % 2], cu_seqlens_kv_padded, 0
2029
                                    )
2030
2031
                                if attn_bias is not None:
                                    idx = (rank - i) % cp_size
2032
                                    attn_bias_inputs[i % 2] = attn_bias[..., idx, :].contiguous()
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
                                out_per_step[i], aux_ctx_tensors = fused_attn_fwd(
                                    is_training,
                                    max_seqlen_q,
                                    max_seqlen_kv // 2,
                                    cu_seqlens_q_per_step[i],
                                    cu_seqlens_kv_per_step[i],
                                    q_inputs[i % 2],
                                    (
                                        kv_inputs[i % 2][..., 0, :, :]
                                        if qkv_format in ["bshd", "sbhd"]
                                        else kv_inputs[i % 2][0]
                                    ),
                                    (
                                        kv_inputs[i % 2][..., 1, :, :]
                                        if qkv_format in ["bshd", "sbhd"]
                                        else kv_inputs[i % 2][1]
                                    ),
                                    fused_attn_qkv_dtype,
                                    fused_attn_backend,
                                    attn_scale=softmax_scale,
                                    dropout=dropout_p,
                                    qkv_layout=qkv_layout,
                                    attn_mask_type="padding" if padding else "no_mask",
                                    attn_bias_type=attn_bias_type,
                                    attn_bias=attn_bias_inputs[i % 2],
                                    cu_seqlens_q_padded=cu_seqlens_q_padded,
                                    cu_seqlens_kv_padded=(
                                        None
                                        if cu_seqlens_kv_padded is None
                                        else cu_seqlens_kv_padded // 2
                                    ),
                                    **fp8_meta_kwargs,
2065
                                )
2066
2067
2068
2069
2070
                                if fp8:
                                    softmax_lse_per_step[i], _, rng_states[i] = aux_ctx_tensors
                                else:
                                    softmax_lse_per_step[i], rng_states[i], *rest = aux_ctx_tensors
                                    attn_biases[i] = rest[0] if len(rest) > 0 else None
2071
2072
                            else:
                                # [b, 2, sq//2, np, hn] -> [b*sq, np, hn]
2073
                                q_inputs[i % 2] = q.view(-1, *q.shape[-2:])
2074
2075
                                if qkv_format == "thd":
                                    # [2, t, np, hn] -> [2, t/2, np, hn]
2076
                                    kv_inputs[i % 2] = tex.thd_read_half_tensor(
2077
                                        kv_inputs[i % 2], cu_seqlens_kv_padded, 0
2078
                                    )
2079
2080
                                else:
                                    # [2, b, 2, sk//2, np, hn] -> [2, b, sk//2, np, hn]
2081
                                    kv_inputs[i % 2] = kv_inputs[i % 2][:, :, 0, ...].contiguous()
2082
                                # [2, b, sk//2, np, hn] -> [2, b*sk//2, np, hn]
2083
                                kv_inputs[i % 2] = kv_inputs[i % 2].view(2, -1, *k.shape[-2:])
2084
2085
2086
                                if _use_flash_attn_3 or _flash_attn_2_3_plus:
                                    fa_forward_kwargs["window_size"] = (-1, -1)
                                fa_outputs = flash_attn_fwd(
2087
2088
2089
                                    q_inputs[i % 2],
                                    kv_inputs[i % 2][0],
                                    kv_inputs[i % 2][1],
2090
2091
                                    cu_seqlens_q_per_step[i],
                                    cu_seqlens_kv_per_step[i],
2092
                                    max_seqlen_q,
2093
                                    max_seqlen_kv // 2,
2094
                                    causal=False,
2095
                                    **fa_forward_kwargs,
2096
                                )
2097
2098
2099
2100
                                out_per_step[i] = fa_outputs[4]
                                softmax_lse_per_step[i] = fa_outputs[5]
                                if not _use_flash_attn_3:
                                    rng_states[i] = fa_outputs[7]
2101
                        else:
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
                            if pad_between_seqs_q:
                                cu_seqlens_q_per_step[i] = get_cu_seqlens_on_cp_rank(
                                    cu_seqlens_q, cu_seqlens_q_padded, cp_size, rank, False, True
                                )
                            else:
                                cu_seqlens_q_per_step[i] = cu_seqlens_q // (cp_size * 2)
                            if pad_between_seqs_kv:
                                cu_seqlens_kv_per_step[i] = get_cu_seqlens_on_cp_rank(
                                    cu_seqlens_kv,
                                    cu_seqlens_kv_padded,
                                    cp_size,
                                    (rank - i) % cp_size,
                                    True,
                                    True,
                                )
                            else:
                                cu_seqlens_kv_per_step[i] = cu_seqlens_kv // cp_size
2119
                            if use_fused_attention:
2120
2121
                                if qkv_format == "bshd":
                                    # [b, 2, sq//2, np, hn] -> [b, sq//2, np, hn]
2122
                                    q_inputs[i % 2] = q[:, 1, ...].contiguous()
2123
                                    # [b, 2, sk//2, 2, np, hn] -> [b, sk, 2, np, hn]
2124
                                    kv_inputs[i % 2] = kv_inputs[i % 2].view(
2125
                                        k.shape[0], -1, 2, *k.shape[-2:]
2126
                                    )
2127
2128
                                elif qkv_format == "sbhd":
                                    # [2, sq//2, b, np, hn] -> [sq//2, b, np, hn]
2129
                                    q_inputs[i % 2] = q[1].contiguous()
2130
2131
2132
2133
                                    # [2, sk//2, b, 2, np, hn] -> [sk, b, 2, np, hn]
                                    kv_inputs[i % 2] = kv_inputs[i % 2].view(
                                        -1, k.shape[2], 2, *k.shape[-2:]
                                    )
2134
2135
                                elif qkv_format == "thd":
                                    # [t, np, hn] -> [t/2, np, hn]
2136
2137
2138
                                    q_inputs[i % 2] = tex.thd_read_half_tensor(
                                        q, cu_seqlens_q_padded, 1
                                    )
2139
2140
                                if attn_bias is not None:
                                    idx = (rank - i) % cp_size
2141
2142
2143
2144
2145
2146
                                    attn_bias_inputs[i % 2] = torch.cat(
                                        (
                                            attn_bias_[..., 1, :, idx, :],
                                            attn_bias_[..., 1, :, (2 * cp_size - idx - 1), :],
                                        ),
                                        dim=-1,
2147
                                    ).contiguous()
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179
                                out_per_step[i], aux_ctx_tensors = fused_attn_fwd(
                                    is_training,
                                    max_seqlen_q // 2,
                                    max_seqlen_kv,
                                    cu_seqlens_q_per_step[i],
                                    cu_seqlens_kv_per_step[i],
                                    q_inputs[i % 2],
                                    (
                                        kv_inputs[i % 2][..., 0, :, :]
                                        if qkv_format in ["bshd", "sbhd"]
                                        else kv_inputs[i % 2][0]
                                    ),
                                    (
                                        kv_inputs[i % 2][..., 1, :, :]
                                        if qkv_format in ["bshd", "sbhd"]
                                        else kv_inputs[i % 2][1]
                                    ),
                                    fused_attn_qkv_dtype,
                                    fused_attn_backend,
                                    attn_scale=softmax_scale,
                                    dropout=dropout_p,
                                    qkv_layout=qkv_layout,
                                    attn_mask_type="padding" if padding else "no_mask",
                                    attn_bias_type=attn_bias_type,
                                    attn_bias=attn_bias_inputs[i % 2],
                                    cu_seqlens_q_padded=(
                                        None
                                        if cu_seqlens_q_padded is None
                                        else cu_seqlens_q_padded // 2
                                    ),
                                    cu_seqlens_kv_padded=cu_seqlens_kv_padded,
                                    **fp8_meta_kwargs,
2180
                                )
2181
2182
2183
2184
2185
                                if fp8:
                                    softmax_lse_per_step[i], _, rng_states[i] = aux_ctx_tensors
                                else:
                                    softmax_lse_per_step[i], rng_states[i], *rest = aux_ctx_tensors
                                    attn_biases[i] = rest[0] if len(rest) > 0 else None
2186
                            else:
2187
2188
                                if qkv_format == "thd":
                                    # [t, np, hn] -> [t/2, np, hn]
2189
2190
2191
                                    q_inputs[i % 2] = tex.thd_read_half_tensor(
                                        q, cu_seqlens_q_padded, 1
                                    )
2192
2193
                                else:
                                    # [b, 2, sq//2, np, hn]->[b, sq//2, np, hn]->[b*sq//2, np, hn]
2194
                                    q_inputs[i % 2] = (
2195
                                        q[:, 1, ...].contiguous().view(-1, *q.shape[-2:])
2196
                                    )
2197
                                # [2, b, 2, sk//2, np, hn] -> [2, b*sk, np, hn]
2198
                                kv_inputs[i % 2] = kv_inputs[i % 2].view(2, -1, *k.shape[-2:])
2199
2200
2201
                                if _use_flash_attn_3 or _flash_attn_2_3_plus:
                                    fa_forward_kwargs["window_size"] = (-1, -1)
                                fa_outputs = flash_attn_fwd(
2202
2203
2204
                                    q_inputs[i % 2],
                                    kv_inputs[i % 2][0],
                                    kv_inputs[i % 2][1],
2205
2206
                                    cu_seqlens_q_per_step[i],
                                    cu_seqlens_kv_per_step[i],
2207
                                    max_seqlen_q // 2,
2208
                                    max_seqlen_kv,
2209
                                    causal=False,
2210
                                    **fa_forward_kwargs,
2211
                                )
2212
2213
2214
2215
                                out_per_step[i] = fa_outputs[4]
                                softmax_lse_per_step[i] = fa_outputs[5]
                                if not _use_flash_attn_3:
                                    rng_states[i] = fa_outputs[7]
2216
                    else:
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
2233
                        if pad_between_seqs_q:
                            cu_seqlens_q_per_step[i] = get_cu_seqlens_on_cp_rank(
                                cu_seqlens_q, cu_seqlens_q_padded, cp_size, rank, True, True
                            )
                        else:
                            cu_seqlens_q_per_step[i] = cu_seqlens_q // cp_size
                        if pad_between_seqs_kv:
                            cu_seqlens_kv_per_step[i] = get_cu_seqlens_on_cp_rank(
                                cu_seqlens_kv,
                                cu_seqlens_kv_padded,
                                cp_size,
                                (rank - i) % cp_size,
                                True,
                                True,
                            )
                        else:
                            cu_seqlens_kv_per_step[i] = cu_seqlens_kv // cp_size
2234
                        if use_fused_attention:
2235
2236
                            if attn_bias is not None:
                                idx = (rank - i) % cp_size
2237
2238
2239
2240
2241
2242
                                attn_bias_inputs[i % 2] = torch.cat(
                                    (
                                        attn_bias[..., idx, :],
                                        attn_bias[..., (2 * cp_size - idx - 1), :],
                                    ),
                                    dim=-1,
2243
                                ).contiguous()
2244
2245
2246
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261
2262
2263
2264
2265
2266
2267
2268
2269
2270
2271
                            out_per_step[i], aux_ctx_tensors = fused_attn_fwd(
                                is_training,
                                max_seqlen_q,
                                max_seqlen_kv,
                                cu_seqlens_q_per_step[i],
                                cu_seqlens_kv_per_step[i],
                                q,
                                (
                                    kv_inputs[i % 2][..., 0, :, :]
                                    if qkv_format in ["bshd", "sbhd"]
                                    else kv_inputs[i % 2][0]
                                ),
                                (
                                    kv_inputs[i % 2][..., 1, :, :]
                                    if qkv_format in ["bshd", "sbhd"]
                                    else kv_inputs[i % 2][1]
                                ),
                                fused_attn_qkv_dtype,
                                fused_attn_backend,
                                attn_scale=softmax_scale,
                                dropout=dropout_p,
                                qkv_layout=qkv_layout,
                                attn_mask_type=attn_mask_type,
                                attn_bias_type=attn_bias_type,
                                attn_bias=attn_bias_inputs[i % 2],
                                cu_seqlens_q_padded=cu_seqlens_q_padded,
                                cu_seqlens_kv_padded=cu_seqlens_kv_padded,
                                **fp8_meta_kwargs,
2272
                            )
2273
2274
2275
2276
2277
                            if fp8:
                                softmax_lse_per_step[i], _, rng_states[i] = aux_ctx_tensors
                            else:
                                softmax_lse_per_step[i], rng_states[i], *rest = aux_ctx_tensors
                                attn_biases[i] = rest[0] if len(rest) > 0 else None
2278
                        else:
2279
                            # [b, sq, np, hn] -> [b*sq, np, hn]
2280
                            q_inputs[i % 2] = q.view(-1, *q.shape[-2:])
2281
                            # [2, b, sk, np, hn] -> [2, b*sk, np, hn]
2282
                            kv_inputs[i % 2] = kv_inputs[i % 2].view(2, -1, *k.shape[-2:])
2283
                            fa_outputs = flash_attn_fwd(
2284
2285
2286
                                q_inputs[i % 2],
                                kv_inputs[i % 2][0],
                                kv_inputs[i % 2][1],
2287
2288
                                cu_seqlens_q_per_step[i],
                                cu_seqlens_kv_per_step[i],
2289
                                max_seqlen_q,
2290
                                max_seqlen_kv,
2291
                                causal=False,
2292
                                **fa_forward_kwargs,
2293
                            )
2294
2295
2296
2297
                            out_per_step[i] = fa_outputs[4]
                            softmax_lse_per_step[i] = fa_outputs[5]
                            if not _use_flash_attn_3:
                                rng_states[i] = fa_outputs[7]
2298
2299
2300
2301

            if i > 0:
                # wait until fwd restuls correction of last step is done
                if i > 1:
2302
                    flash_attn_streams[(i - 1) % 2].wait_event(fwd_results_correction_done)
2303

2304
2305
                if use_fused_attention:
                    # [b, np, sq, 1] -> [b, np, sq]
2306
                    softmax_lse_per_step[i - 1].squeeze_(-1)
2307
2308
2309
2310
2311
                if qkv_format != "thd" and softmax_lse_in_packed_format:
                    # [np, t] -> [np, b, sq]
                    softmax_lse_per_step[i - 1] = softmax_lse_per_step[i - 1].view(
                        q.shape[-2], q.shape[0], -1
                    )
2312

2313
                with torch.cuda.stream(flash_attn_streams[(i - 1) % 2]):
2314
2315
2316
2317
2318
2319
2320
2321
                    if fp8:
                        out_per_step[i - 1] = cast_from_fp8(
                            out_per_step[i - 1],
                            fp8_meta["scaling_fwd"],
                            META_O_CP,
                            fp8_dtype_forward,
                            TE_DType[torch.float32],
                        )
2322
                    if i == 1:
2323
                        out = torch.zeros_like(q if not fp8 else out_per_step[0]).view(q.shape)
2324
                        softmax_lse = torch.clone(softmax_lse_per_step[0]).to(torch.double)
2325
                        if causal and qkv_format != "thd":
2326
2327
                            # [b, np, sq] -> [b, np, 2, sq//2] lse not in packed format
                            # [np, b, sq] -> [np, b, 2, sq//2] lse in packed format
2328
                            softmax_lse_ = softmax_lse.view(
2329
                                *softmax_lse.shape[:-1], 2, softmax_lse.shape[-1] // 2
2330
                            )
2331
2332
2333
2334
                    elif (i - 1) <= rank or not causal:
                        flash_attn_fwd_softmax_lse_correction(
                            softmax_lse, softmax_lse_per_step[i - 1]
                        )
2335
                    else:
2336
                        if qkv_format == "thd":
2337
                            tex.thd_second_half_lse_correction(
2338
2339
2340
                                softmax_lse,
                                softmax_lse_per_step[i - 1],
                                cu_seqlens_q_padded,
2341
                                softmax_lse_in_packed_format,
2342
                            )
2343
                        else:
2344
2345
2346
                            flash_attn_fwd_softmax_lse_correction(
                                softmax_lse_[..., 1, :], softmax_lse_per_step[i - 1]
                            )
2347
2348

                if i < cp_size:
2349
                    flash_attn_streams[(i - 1) % 2].record_event(fwd_results_correction_done)
2350
2351
2352
2353
2354

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

        softmax_lse = softmax_lse.to(torch.float)
        for i in range(cp_size):
2355
            out_ = None
2356
            if qkv_format == "bshd":
2357
2358
2359
                out_per_step[i] = out_per_step[i].view(
                    out.shape[0], -1, *out.shape[-2:]
                )  # pylint: disable=used-before-assignment
2360
2361
2362
2363
                out_ = out[:, 1, ...]
            elif qkv_format == "sbhd":
                out_per_step[i] = out_per_step[i].view(-1, *out.shape[-3:])
                out_ = out[1]
2364

2365
            if i <= rank or not causal:
2366
                if qkv_format in ["bshd", "sbhd"]:
2367
2368
2369
2370
2371
                    flash_attn_fwd_out_correction(
                        out.view(*out_per_step[i].shape),
                        out_per_step[i],
                        softmax_lse,
                        softmax_lse_per_step[i],
2372
2373
                        0 if softmax_lse_in_packed_format else 2,
                        2 if softmax_lse_in_packed_format else seq_dim,
2374
                    )
2375
                elif qkv_format == "thd":
2376
2377
2378
2379
2380
                    tex.thd_out_correction(
                        out,
                        out_per_step[i],
                        softmax_lse,
                        softmax_lse_per_step[i],
2381
                        cu_seqlens_q_padded,
2382
                        False,
2383
                        softmax_lse_in_packed_format,
2384
                    )
2385
            else:
2386
                if qkv_format in ["bshd", "sbhd"]:
2387
2388
2389
2390
2391
                    flash_attn_fwd_out_correction(
                        out_,
                        out_per_step[i],
                        softmax_lse_[..., 1, :],
                        softmax_lse_per_step[i],
2392
2393
                        0 if softmax_lse_in_packed_format else 2,
                        2 if softmax_lse_in_packed_format else seq_dim,
2394
                    )
2395
                elif qkv_format == "thd":
2396
2397
2398
2399
2400
                    tex.thd_out_correction(
                        out,
                        out_per_step[i],
                        softmax_lse,
                        softmax_lse_per_step[i],
2401
                        cu_seqlens_q_padded,
2402
                        True,
2403
                        softmax_lse_in_packed_format,
2404
                    )
2405

2406
2407
2408
        if qkv_format != "thd" and softmax_lse_in_packed_format:
            # [np, b, sq] -> [np, t]
            softmax_lse = softmax_lse.view(softmax_lse.shape[0], -1)
2409
        kv = p2p_comm_buffers[-1]
2410
2411
2412
2413
2414
2415
2416
2417
2418
2419
2420
2421
2422
2423
2424
2425
2426
2427
2428
2429
        if qkv_format == "bshd":
            out = out.view(out.shape[0], -1, *out.shape[-2:])
            ctx.batch_size = out.shape[0]
        elif qkv_format == "sbhd":
            out = out.view(-1, *out.shape[-3:])
            ctx.batch_size = out.shape[1]

        if cp_size_a2a > 1:
            chunk_ids_for_a2a = get_seq_chunk_ids_for_reordering(cp_size_a2a, out.device, False)
            out = flash_attn_a2a_communicate(
                out, chunk_ids_for_a2a, seq_dim, cp_size_a2a, cp_group_a2a, cp_stream, False
            )
            if use_fused_attention:
                if qkv_format == "bshd":
                    # [b*s, np, hn] -> [b, s, np, hn]
                    out = out.view(ctx.batch_size, -1, *out.shape[-2:])
                elif qkv_format == "sbhd":
                    # [s*b, np, hn] -> [s, b, np, hn]
                    out = out.view(-1, ctx.batch_size, *out.shape[-2:])
        elif not use_fused_attention:
2430
            out = out.view(-1, *out.shape[-2:])
2431

2432
2433
2434
2435
2436
        if fp8 and use_fused_attention:
            amax_cp_fwd = amax_per_step.amax(dim=1)
            fp8_meta["scaling_fwd"].amax_history[0][META_S] = amax_cp_fwd[0]
            fp8_meta["scaling_fwd"].amax_history[0][META_O_CP] = amax_cp_fwd[1]

2437
        out_fp8 = None
2438
2439
        out_f16 = out.to(qkv_dtype)
        if fp8 and (is_output_fp8 or int(os.getenv("NVTE_FP8_DPA_BWD", "1"))):
2440
2441
            out_fp8 = cast_to_fp8(out_f16, fp8_meta["scaling_fwd"], META_O, fp8_dtype_forward)

2442
        if fp8 and is_output_fp8:
2443
2444
2445
2446
2447
2448
            out_ret = Float8Tensor(
                data=out_fp8,
                fp8_meta=fp8_meta,
                fp8_meta_forward=True,
                fp8_meta_index=META_O,
                fp8_dtype=fp8_dtype_forward,
2449
                dtype=qkv_dtype,
2450
2451
2452
2453
2454
2455
2456
2457
            )
        else:
            out_ret = out_f16

        if fp8 and int(os.getenv("NVTE_FP8_DPA_BWD", "1")):
            q_save, kv_save, out_save = q, kv, out_fp8
            fp8_fwd_scales = fp8_meta["scaling_fwd"].scale.clone()
            fp8_fwd_scale_invs = fp8_meta["scaling_fwd"].scale_inv.clone()
2458
        elif fp8 and is_input_fp8:
2459
2460
2461
2462
2463
2464
2465
2466
            q_fp8 = Float8Tensor(
                data=q,
                fp8_meta=fp8_meta,
                fp8_meta_forward=True,
                fp8_meta_index=META_QKV,
                fp8_dtype=fp8_dtype_forward,
                dtype=q_fp8.dtype,
            )
2467
2468
2469
2470
2471
2472
2473
2474
2475
2476
2477
            kv_fp8 = Float8Tensor(
                data=kv,
                fp8_meta=fp8_meta,
                fp8_meta_forward=True,
                fp8_meta_index=META_QKV,
                fp8_dtype=fp8_dtype_forward,
                dtype=k_fp8.dtype,
            )
            q_save, kv_save, out_save = q_fp8, kv_fp8, out_f16
            fp8_fwd_scales, fp8_fwd_scale_invs = None, None
        else:
2478
            q_f16 = q_f16.view(q.shape)
2479
2480
2481
            q_save, kv_save, out_save = q_f16, kv, out_f16
            fp8_fwd_scales, fp8_fwd_scale_invs = None, None

2482
        ctx.save_for_backward(
2483
2484
2485
            q_save,
            kv_save,
            out_save,
2486
            softmax_lse,
2487
2488
            cu_seqlens_q_padded,
            cu_seqlens_kv_padded,
2489
2490
            fp8_fwd_scales,
            fp8_fwd_scale_invs,
2491
2492
            *cu_seqlens_q_per_step,
            *cu_seqlens_kv_per_step,
2493
2494
            *rng_states,
            *attn_biases,
2495
        )
2496
2497
2498
        ctx.cp_group_a2a = cp_group_a2a
        ctx.cp_size_a2a = cp_size_a2a
        ctx.rank_a2a = rank_a2a
2499
2500
        ctx.cp_group = cp_group
        ctx.cp_global_ranks = cp_global_ranks
2501
        ctx.cp_stream = cp_stream
2502
        ctx.dropout_p = dropout_p
2503
        ctx.total_tokens_kv = total_tokens_kv
2504
        ctx.max_seqlen_q = max_seqlen_q
2505
        ctx.max_seqlen_kv = max_seqlen_kv
2506
        ctx.softmax_scale = softmax_scale
2507
        ctx.qkv_format = qkv_format
2508
        ctx.attn_mask_type = attn_mask_type
2509
2510
        ctx.attn_bias_type = attn_bias_type
        ctx.attn_bias_shape = None if attn_bias is None else attn_bias.shape
2511
        ctx.deterministic = deterministic
2512
        ctx.use_fused_attention = use_fused_attention
2513
2514
        ctx.fp8 = fp8 and int(os.getenv("NVTE_FP8_DPA_BWD", "1"))
        ctx.fp8_meta = fp8_meta
2515
2516
        ctx.is_input_fp8 = is_input_fp8
        ctx.is_output_fp8 = is_output_fp8
2517
        return out_ret
2518
2519
2520

    @staticmethod
    def backward(ctx, dout):
2521
        # pylint: disable=missing-function-docstring
2522
2523
2524
        cp_size_a2a = ctx.cp_size_a2a
        rank_a2a = ctx.rank_a2a

2525
2526
        cp_size = get_distributed_world_size(ctx.cp_group)
        rank = get_distributed_rank(ctx.cp_group)
2527
2528
        send_dst = ctx.cp_global_ranks[(rank - 1) % cp_size * cp_size_a2a + rank_a2a]
        recv_src = ctx.cp_global_ranks[(rank + 1) % cp_size * cp_size_a2a + rank_a2a]
2529
2530
        batch_p2p_comm = int(os.getenv("NVTE_BATCH_MHA_P2P_COMM", "0")) or (cp_size == 2)

2531
        (q, kv, out, softmax_lse, cu_seqlens_q_padded, cu_seqlens_kv_padded) = ctx.saved_tensors[:6]
2532
2533
2534
2535
2536
        (fp8_fwd_scales, fp8_fwd_scale_invs) = ctx.saved_tensors[6:8]
        cu_seqlens_q_per_step = ctx.saved_tensors[8 : 8 + cp_size]
        cu_seqlens_kv_per_step = ctx.saved_tensors[8 + cp_size : 8 + cp_size * 2]
        rng_states = ctx.saved_tensors[8 + cp_size * 2 : 8 + cp_size * 3]
        attn_biases = ctx.saved_tensors[8 + cp_size * 3 : 8 + cp_size * 4]
2537

2538
2539
        causal = "causal" in ctx.attn_mask_type
        padding = "padding" in ctx.attn_mask_type
2540
2541

        seq_dim = None
2542
        if ctx.qkv_format in ["bshd", "sbhd"]:
2543
            seq_dim = ctx.qkv_format.index("s")
2544
2545
2546
            qkv_layout = ctx.qkv_format + "_" + ctx.qkv_format[:-2] + "2" + ctx.qkv_format[-2:]
        else:
            qkv_layout = ctx.qkv_format + "_" + ctx.qkv_format + "_" + ctx.qkv_format
2547

2548
        if attn_biases[0] is not None:
2549
2550
            # [b, np, sq, 2*cp, sk//(2*cp)]
            attn_dbias = torch.zeros(
2551
                *ctx.attn_bias_shape, dtype=attn_biases[0].dtype, device=attn_biases[0].device
2552
2553
2554
            )
            # [b, np, sq, 2*cp, sk//(2*cp)] -> [b, np, 2, sq//2, 2*cp, sk//(2*cp)]
            attn_dbias_ = attn_dbias.view(
2555
                *attn_dbias.shape[:-3], 2, attn_dbias.shape[-3] // 2, *attn_dbias.shape[-2:]
2556
2557
2558
            )
        else:
            attn_dbias = None
2559
            attn_dbias_ = None
2560

2561
2562
2563
2564
        softmax_lse_in_packed_format = not ctx.use_fused_attention and (
            _flash_attn_2_6_0_plus or _use_flash_attn_3
        )

2565
        if causal:
2566
            if ctx.qkv_format == "thd" or softmax_lse_in_packed_format:
2567
                softmax_lse_ = tex.thd_read_second_half_lse(
2568
                    softmax_lse, cu_seqlens_q_padded, softmax_lse_in_packed_format
2569
                )
2570
2571
            else:
                # [b, np, sq] -> [b, np, 2, sq//2]
2572
2573
2574
                softmax_lse_ = softmax_lse.view(
                    *softmax_lse.shape[:-1], 2, softmax_lse.shape[-1] // 2
                )
2575
2576
2577
2578
                softmax_lse_ = softmax_lse_[..., 1, :].contiguous()
                if ctx.use_fused_attention:
                    # [b, np, sq//2] -> [b, np, sq//2, 1]
                    softmax_lse_.unsqueeze_(-1)
2579
2580
2581
        if ctx.use_fused_attention:
            # [b, np, sq] -> [b, np, sq, 1]
            softmax_lse.unsqueeze_(-1)
2582

2583
        dout_dtype = dout.dtype
2584
2585
2586
2587
2588
        fused_attn_backend = None
        fused_attn_qkv_dtype = None
        fused_attn_dqkv_dtype = None
        amax_per_step = None
        dout_fp8_dtype = None
2589
2590
        if ctx.fp8:
            if ctx.use_fused_attention:
2591
                fp8_dtype_forward = get_fp8_te_dtype(ctx.fp8_meta["recipe"], fprop_tensor=True)
2592
                fp8_dtype_backward = get_fp8_te_dtype(ctx.fp8_meta["recipe"], fprop_tensor=False)
2593
                fused_attn_qkv_dtype = fp8_dtype_forward
2594
2595
2596
2597
2598
                fused_attn_dqkv_dtype = fp8_dtype_backward
                fused_attn_backend = FusedAttnBackend["FP8"]
                dq_fp8 = torch.empty((cp_size, *q.shape), dtype=q.dtype, device=q.device)
                dkv_fp8 = torch.empty((cp_size, *kv.shape), dtype=kv.dtype, device=kv.device)
                dkv_fp8_ = torch.empty_like(dkv_fp8)
2599
                if ctx.is_output_fp8:
2600
2601
2602
2603
2604
2605
2606
2607
2608
2609
2610
2611
2612
2613
2614
2615
2616
2617
2618
2619
2620
                    assert isinstance(dout, Float8Tensor), "dout must be Float8Tensors for FP8 MHA!"
                    ctx.fp8_meta["scaling_bwd"].scale_inv[META_DO] = dout._scale_inv
                    dout = dout._data
                else:
                    dout = cast_to_fp8(
                        dout, ctx.fp8_meta["scaling_bwd"], META_DO, fp8_dtype_backward
                    )
                p2p_comm_buffers = [[kv, dkv_fp8], [torch.empty_like(kv), dkv_fp8_]]
                fp8_meta_kwargs = {}
                fp8_meta_kwargs["d_scale_qkv"] = fp8_fwd_scale_invs[META_QKV]
                fp8_meta_kwargs["d_scale_s"] = fp8_fwd_scale_invs[META_S]
                fp8_meta_kwargs["d_scale_o"] = fp8_fwd_scale_invs[META_O]
                fp8_meta_kwargs["d_scale_do"] = ctx.fp8_meta["scaling_bwd"].scale_inv[META_DO]
                fp8_meta_kwargs["d_scale_dp"] = ctx.fp8_meta["scaling_bwd"].scale_inv[META_DP]
                fp8_meta_kwargs["q_scale_s"] = fp8_fwd_scales[META_S]
                fp8_meta_kwargs["q_scale_dp"] = ctx.fp8_meta["scaling_bwd"].scale[META_DP]
                fp8_meta_kwargs["q_scale_dqkv"] = ctx.fp8_meta["scaling_bwd"].scale[META_DQKV_CP]
                amax_per_step = torch.zeros((2, cp_size), dtype=torch.float32, device=q.device)
            else:
                assert False, "FP8 is only supported with Fused Attention!"
        else:
2621
            if ctx.fp8_meta is not None and ctx.is_input_fp8:
2622
2623
2624
2625
2626
2627
2628
                q, kv = [x.from_float8(x.dtype) for x in [q, kv]]
                if cp_size_a2a == 1:
                    dout = dout.from_float8(dout_dtype)
                else:
                    dout_fp8_dtype = dout._fp8_dtype
                    dout_scale_inv = dout._scale_inv
                    dout = dout._data
2629
2630
2631
2632
2633
2634
2635
2636
2637
2638
2639
            dq = torch.empty_like(q)
            if ctx.qkv_format == "thd" and causal:
                dq[cu_seqlens_q_padded[-1] :].fill_(0)
            p2p_comm_buffers = [
                torch.empty((2, *kv.shape), dtype=kv.dtype, device=kv.device),
                torch.empty((2, *kv.shape), dtype=kv.dtype, device=kv.device),
            ]
            p2p_comm_buffers[0][0].copy_(kv)
            if ctx.use_fused_attention:
                fp8_meta_kwargs = {}
                fused_attn_qkv_dtype = TE_DType[q.dtype]
2640
                fused_attn_dqkv_dtype = TE_DType[dout_dtype]
2641
2642
                fused_attn_backend = FusedAttnBackend["F16_arbitrary_seqlen"]

2643
2644
2645
2646
2647
2648
2649
2650
2651
2652
2653
2654
2655
2656
        if cp_size_a2a > 1:
            if not ctx.use_fused_attention:
                out = out.view(ctx.batch_size, -1, *out.shape[-2:])
                dout = dout.view(*out.shape)
            chunk_ids_for_a2a = get_seq_chunk_ids_for_reordering(cp_size_a2a, out.device, True)
            out, dout = flash_attn_a2a_communicate(
                [out, dout],
                chunk_ids_for_a2a,
                seq_dim,
                cp_size_a2a,
                ctx.cp_group_a2a,
                ctx.cp_stream,
                True,
            )
2657
            if not ctx.fp8 and ctx.fp8_meta is not None and ctx.is_output_fp8:
2658
                dout = cast_from_fp8(
2659
2660
2661
2662
2663
2664
                    dout,
                    None,
                    None,
                    dout_fp8_dtype,
                    TE_DType[dout_dtype],
                    scale_inv=dout_scale_inv,  # pylint: disable=used-before-assignment
2665
2666
                )

2667
2668
2669
2670
        out = out.view(*q.shape)
        dout = dout.view(*q.shape)
        send_recv_reqs = []

2671
        flash_attn_bwd = None
2672
2673
2674
2675
2676
2677
2678
2679
2680
2681
2682
2683
        if not ctx.use_fused_attention:
            fa_backward_kwargs = {"softmax_scale": ctx.softmax_scale}
            if _use_flash_attn_3:
                flash_attn_bwd = flash_attn_varlen_bwd_v3
                fa_backward_kwargs["deterministic"] = ctx.deterministic
            else:
                flash_attn_bwd = flash_attn_varlen_bwd
                fa_backward_kwargs["dropout_p"] = ctx.dropout_p
                if _flash_attn_2_4_plus:
                    fa_backward_kwargs["alibi_slopes"] = None
                if _flash_attn_2_4_1_plus:
                    fa_backward_kwargs["deterministic"] = ctx.deterministic
2684

2685
2686
2687
2688
2689
        for i in range(cp_size):
            # wait until KV is received
            for req in send_recv_reqs:
                req.wait()

2690
2691
            send_tensor = p2p_comm_buffers[i % 2]
            recv_tensor = p2p_comm_buffers[(i + 1) % 2]
2692
2693
2694
2695
2696
2697
2698
2699
2700
2701
2702
2703
2704
2705
2706
2707
2708
2709
2710
2711
2712
2713
2714
2715
2716
2717
2718
2719
2720
            if ctx.fp8:
                if i < cp_size - 1:
                    send_recv_reqs = flash_attn_p2p_communicate(
                        rank,
                        send_tensor[0],
                        send_dst,
                        recv_tensor[0],
                        recv_src,
                        ctx.cp_group,
                        batch_p2p_comm,
                    )
                else:
                    dkv_a2a_req = torch.distributed.all_to_all_single(
                        dkv_fp8,
                        dkv_fp8_,
                        group=ctx.cp_group,
                        async_op=True,
                    )
                    send_recv_reqs = [dkv_a2a_req]
            else:
                if i == 0:
                    send_tensor = send_tensor[0]
                    recv_tensor = recv_tensor[0]
                if i == (cp_size - 1):
                    send_tensor = send_tensor[1]
                    recv_tensor = recv_tensor[1]
                send_recv_reqs = flash_attn_p2p_communicate(
                    rank, send_tensor, send_dst, recv_tensor, recv_src, ctx.cp_group, batch_p2p_comm
                )
2721

2722
            kv = p2p_comm_buffers[i % 2][0]
2723
            dk_, dv_ = None, None
2724
2725
2726
            if ctx.fp8 and ctx.use_fused_attention:
                fp8_meta_kwargs["amax_dp"] = amax_per_step[0][i]
                fp8_meta_kwargs["amax_dqkv"] = amax_per_step[0][i]
2727
            # In reversed order of fwd
2728
            if causal:
2729
                if i == (cp_size - 1):
2730
                    if ctx.use_fused_attention:
2731
2732
2733
                        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:])
2734
2735
                            # [b, 2, sk//2, 2, np, hn] -> [b, sk, 2, np, hn]
                            kv_ = kv.view(kv.shape[0], -1, *kv.shape[-3:])
2736
2737
2738
2739
2740
2741
                            # [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:])
2742
2743
                            # [2, sk//2, b, 2, np, hn] -> [sk, b, 2, np, hn]
                            kv_ = kv.view(-1, *kv.shape[-4:])
2744
2745
2746
                            # [2, sq//2, b, np, hn] -> [sq, b, np, hn]
                            out_ = out.view(-1, *out.shape[-3:])
                            dout_ = dout.view(-1, *dout.shape[-3:])
2747
2748
                        elif ctx.qkv_format == "thd":
                            q_, kv_, out_, dout_ = q, kv, out, dout
2749
2750
2751
2752
2753
2754
2755
2756
                        if ctx.fp8:
                            aux_ctx_tensors = [
                                softmax_lse,
                                softmax_lse,
                                rng_states[cp_size - i - 1],
                            ]
                        else:
                            aux_ctx_tensors = [softmax_lse, rng_states[cp_size - i - 1]]
2757
                        if attn_dbias is not None:
2758
                            aux_ctx_tensors += [attn_biases[cp_size - i - 1]]
2759
                        dq_, dk_, dv_, dbias_ = fused_attn_bwd(
2760
                            ctx.max_seqlen_q,
2761
2762
2763
                            ctx.max_seqlen_kv,
                            cu_seqlens_q_per_step[cp_size - i - 1],
                            cu_seqlens_kv_per_step[cp_size - i - 1],
2764
                            q_,
2765
2766
                            kv_[..., 0, :, :] if ctx.qkv_format in ["bshd", "sbhd"] else kv_[0],
                            kv_[..., 1, :, :] if ctx.qkv_format in ["bshd", "sbhd"] else kv_[1],
2767
2768
                            out_,
                            dout_,
2769
2770
                            fused_attn_qkv_dtype,
                            fused_attn_dqkv_dtype,
2771
                            aux_ctx_tensors,
2772
                            fused_attn_backend,
2773
2774
                            cu_seqlens_q_padded=cu_seqlens_q_padded,
                            cu_seqlens_kv_padded=cu_seqlens_kv_padded,
2775
2776
                            attn_scale=ctx.softmax_scale,
                            dropout=ctx.dropout_p,
2777
                            qkv_layout=qkv_layout,
2778
                            attn_mask_type=ctx.attn_mask_type,
2779
                            attn_bias_type=ctx.attn_bias_type,
2780
2781
                            deterministic=ctx.deterministic,
                            **fp8_meta_kwargs,
2782
2783
2784
2785
                        )
                    else:
                        # [b, 2, sq//2, np, hn] -> [b*sq, np, hn]
                        q_ = q.view(-1, *q.shape[-2:])
2786
                        dq_ = torch.zeros_like(q_)
2787
2788
2789
2790
2791
2792
                        # [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:])
2793
2794
2795
2796
2797
                        if _use_flash_attn_3 or _flash_attn_2_3_plus:
                            fa_backward_kwargs["window_size"] = (-1, 0)
                        if not _use_flash_attn_3:
                            fa_backward_kwargs["rng_state"] = rng_states[cp_size - i - 1]
                        flash_attn_bwd(
2798
2799
2800
2801
2802
2803
2804
2805
2806
                            dout_,
                            q_,
                            kv_[0],
                            kv_[1],
                            out_,
                            softmax_lse,
                            dq_,
                            dkv_[0],
                            dkv_[1],
2807
2808
                            cu_seqlens_q_per_step[cp_size - i - 1],
                            cu_seqlens_kv_per_step[cp_size - i - 1],
2809
                            ctx.max_seqlen_q,
2810
                            ctx.max_seqlen_kv,
2811
2812
                            causal=True,
                            **fa_backward_kwargs,
2813
                        )
2814
                elif i >= (cp_size - rank - 1):
2815
                    if ctx.use_fused_attention:
2816
2817
2818
                        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:])
2819
2820
                            # [b, 2, sk//2, 2, np, hn] -> [b, sk//2, 2, np, hn]
                            kv_ = kv[:, 0, ...].contiguous()
2821
2822
2823
2824
2825
2826
                            # [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:])
2827
2828
                            # [2, sk//2, b, 2, np, hn] -> [sk//2, b, 2, np, hn]
                            kv_ = kv[0].contiguous()
2829
2830
2831
                            # [2, sq//2, b, np, hn] -> [sq, b, np, hn]
                            out_ = out.view(-1, *out.shape[-3:])
                            dout_ = dout.view(-1, *dout.shape[-3:])
2832
2833
2834
                        elif ctx.qkv_format == "thd":
                            q_, out_, dout_ = q, out, dout
                            # [2, t, np, hn] -> [2, t/2, np, hn]
2835
                            kv_ = tex.thd_read_half_tensor(kv, cu_seqlens_kv_padded, 0)
2836
2837
2838
2839
2840
2841
2842
2843
                        if ctx.fp8:
                            aux_ctx_tensors = [
                                softmax_lse,
                                softmax_lse,
                                rng_states[cp_size - i - 1],
                            ]
                        else:
                            aux_ctx_tensors = [softmax_lse, rng_states[cp_size - i - 1]]
2844
                        if attn_dbias is not None:
2845
                            aux_ctx_tensors += [attn_biases[cp_size - i - 1]]
2846
                        dq_, dk_, dv_, dbias_ = fused_attn_bwd(
2847
                            ctx.max_seqlen_q,
2848
2849
2850
                            ctx.max_seqlen_kv // 2,
                            cu_seqlens_q_per_step[cp_size - i - 1],
                            cu_seqlens_kv_per_step[cp_size - i - 1],
2851
                            q_,
2852
2853
                            kv_[..., 0, :, :] if ctx.qkv_format in ["bshd", "sbhd"] else kv_[0],
                            kv_[..., 1, :, :] if ctx.qkv_format in ["bshd", "sbhd"] else kv_[1],
2854
2855
                            out_,
                            dout_,
2856
2857
                            fused_attn_qkv_dtype,
                            fused_attn_dqkv_dtype,
2858
                            aux_ctx_tensors,
2859
                            fused_attn_backend,
2860
2861
2862
2863
                            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
                            ),
2864
2865
                            attn_scale=ctx.softmax_scale,
                            dropout=ctx.dropout_p,
2866
                            qkv_layout=qkv_layout,
2867
                            attn_mask_type="padding" if padding else "no_mask",
2868
                            attn_bias_type=ctx.attn_bias_type,
2869
2870
                            deterministic=ctx.deterministic,
                            **fp8_meta_kwargs,
2871
2872
2873
2874
                        )
                    else:
                        # [b, 2, sq//2, np, hn] -> [b*sq, np, hn]
                        q_ = q.view(-1, *q.shape[-2:])
2875
                        dq_ = torch.zeros_like(q_)
2876
2877
                        if ctx.qkv_format == "thd":
                            # [2, t, np, hn] -> [2, t/2, np, hn]
2878
                            kv_ = tex.thd_read_half_tensor(kv, cu_seqlens_kv_padded, 0)
2879
2880
2881
                        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:])
2882
2883
2884
2885
                        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:])
2886
2887
2888
2889
2890
                        if _use_flash_attn_3 or _flash_attn_2_3_plus:
                            fa_backward_kwargs["window_size"] = (-1, -1)
                        if not _use_flash_attn_3:
                            fa_backward_kwargs["rng_state"] = rng_states[cp_size - i - 1]
                        flash_attn_bwd(
2891
2892
2893
2894
2895
2896
2897
2898
2899
                            dout_,
                            q_,
                            kv_[0],
                            kv_[1],
                            out_,
                            softmax_lse,
                            dq_,
                            dkv_[0],
                            dkv_[1],
2900
2901
                            cu_seqlens_q_per_step[cp_size - i - 1],
                            cu_seqlens_kv_per_step[cp_size - i - 1],
2902
                            ctx.max_seqlen_q,
2903
                            ctx.max_seqlen_kv // 2,
2904
2905
                            causal=False,
                            **fa_backward_kwargs,
2906
2907
2908
                        )
                else:
                    if ctx.use_fused_attention:
2909
2910
2911
                        if ctx.qkv_format == "bshd":
                            # [b, 2, sq//2, np, hn] -> [b, sq//2, np, hn]
                            q_ = q[:, 1, ...].contiguous()
2912
2913
                            # [b, 2, sk//2, 2, np, hn] -> [b, sk, 2, np, hn]
                            kv_ = kv.view(kv.shape[0], -1, *kv.shape[-3:])
2914
2915
2916
2917
2918
2919
                            # [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()
2920
2921
                            # [2, sk//2, b, 2, np, hn] -> [sk, b, 2, np, hn]
                            kv_ = kv.view(-1, *kv.shape[-4:])
2922
2923
2924
                            # [2, sq//2, b, np, hn] -> [sq//2, b, np, hn]
                            out_ = out[1].contiguous()
                            dout_ = dout[1].contiguous()
2925
2926
                        elif ctx.qkv_format == "thd":
                            # [t, np, hn] -> [t/2, np, hn]
2927
2928
2929
                            q_ = tex.thd_read_half_tensor(q, cu_seqlens_q_padded, 1)
                            out_ = tex.thd_read_half_tensor(out, cu_seqlens_q_padded, 1)
                            dout_ = tex.thd_read_half_tensor(dout, cu_seqlens_q_padded, 1)
2930
                            kv_ = kv
2931
2932
2933
2934
2935
2936
2937
2938
                        if ctx.fp8:
                            aux_ctx_tensors = [
                                softmax_lse_,
                                softmax_lse_,
                                rng_states[cp_size - i - 1],
                            ]
                        else:
                            aux_ctx_tensors = [softmax_lse_, rng_states[cp_size - i - 1]]
2939
                        if attn_dbias is not None:
2940
                            aux_ctx_tensors += [attn_biases[cp_size - i - 1]]
2941
                        dq_, dk_, dv_, dbias_ = fused_attn_bwd(
2942
                            ctx.max_seqlen_q // 2,
2943
2944
2945
                            ctx.max_seqlen_kv,
                            cu_seqlens_q_per_step[cp_size - i - 1],
                            cu_seqlens_kv_per_step[cp_size - i - 1],
2946
                            q_,
2947
2948
                            kv_[..., 0, :, :] if ctx.qkv_format in ["bshd", "sbhd"] else kv_[0],
                            kv_[..., 1, :, :] if ctx.qkv_format in ["bshd", "sbhd"] else kv_[1],
2949
2950
                            out_,
                            dout_,
2951
2952
                            fused_attn_qkv_dtype,
                            fused_attn_dqkv_dtype,
2953
                            aux_ctx_tensors,
2954
                            fused_attn_backend,
2955
2956
2957
2958
                            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,
2959
2960
                            attn_scale=ctx.softmax_scale,
                            dropout=ctx.dropout_p,
2961
                            qkv_layout=qkv_layout,
2962
                            attn_mask_type="padding" if padding else "no_mask",
2963
                            attn_bias_type=ctx.attn_bias_type,
2964
2965
                            deterministic=ctx.deterministic,
                            **fp8_meta_kwargs,
2966
2967
                        )
                    else:
2968
2969
                        if ctx.qkv_format == "thd":
                            # [t, np, hn] -> [t/2, np, hn]
2970
                            q_ = tex.thd_read_half_tensor(q, cu_seqlens_q_padded, 1)
2971
2972
2973
                        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:])
2974
                        dq_ = torch.zeros_like(q_)
2975
2976
2977
                        # [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_)
2978
                        if ctx.qkv_format == "thd":
2979
2980
                            out_ = tex.thd_read_half_tensor(out, cu_seqlens_q_padded, 1)
                            dout_ = tex.thd_read_half_tensor(dout, cu_seqlens_q_padded, 1)
2981
2982
2983
2984
                        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:])
2985
2986
2987
2988
2989
                        if _use_flash_attn_3 or _flash_attn_2_3_plus:
                            fa_backward_kwargs["window_size"] = (-1, -1)
                        if not _use_flash_attn_3:
                            fa_backward_kwargs["rng_state"] = rng_states[cp_size - i - 1]
                        flash_attn_bwd(
2990
2991
2992
2993
2994
2995
2996
2997
2998
                            dout_,
                            q_,
                            kv_[0],
                            kv_[1],
                            out_,
                            softmax_lse_,
                            dq_,
                            dkv_[0],
                            dkv_[1],
2999
3000
                            cu_seqlens_q_per_step[cp_size - i - 1],
                            cu_seqlens_kv_per_step[cp_size - i - 1],
3001
                            ctx.max_seqlen_q // 2,
3002
                            ctx.max_seqlen_kv,
3003
3004
                            causal=False,
                            **fa_backward_kwargs,
3005
3006
3007
                        )
            else:
                if ctx.use_fused_attention:
3008
3009
3010
3011
                    if ctx.fp8:
                        aux_ctx_tensors = [softmax_lse, softmax_lse, rng_states[cp_size - i - 1]]
                    else:
                        aux_ctx_tensors = [softmax_lse, rng_states[cp_size - i - 1]]
3012
                    if attn_dbias is not None:
3013
                        aux_ctx_tensors += [attn_biases[cp_size - i - 1]]
3014
                    dq_, dk_, dv_, dbias_ = fused_attn_bwd(
3015
                        ctx.max_seqlen_q,
3016
3017
3018
                        ctx.max_seqlen_kv,
                        cu_seqlens_q_per_step[cp_size - i - 1],
                        cu_seqlens_kv_per_step[cp_size - i - 1],
3019
                        q,
3020
3021
                        kv[..., 0, :, :] if ctx.qkv_format in ["bshd", "sbhd"] else kv[0],
                        kv[..., 1, :, :] if ctx.qkv_format in ["bshd", "sbhd"] else kv[1],
3022
3023
                        out,
                        dout,
3024
3025
                        fused_attn_qkv_dtype,
                        fused_attn_dqkv_dtype,
3026
                        aux_ctx_tensors,
3027
                        fused_attn_backend,
3028
3029
                        cu_seqlens_q_padded=cu_seqlens_q_padded,
                        cu_seqlens_kv_padded=cu_seqlens_kv_padded,
3030
3031
                        attn_scale=ctx.softmax_scale,
                        dropout=ctx.dropout_p,
3032
                        qkv_layout=qkv_layout,
3033
                        attn_mask_type=ctx.attn_mask_type,
3034
                        attn_bias_type=ctx.attn_bias_type,
3035
3036
                        deterministic=ctx.deterministic,
                        **fp8_meta_kwargs,
3037
3038
3039
                    )
                else:
                    # [b, sq, np, hn] -> [b*sq, np, hn]
3040
                    q_ = q.view(-1, *q.shape[-2:])
3041
                    dq_ = torch.zeros_like(q_)
3042
                    # [2, b, sk, np, hn] -> [2, b*sk, np, hn]
3043
3044
                    kv_ = kv.view(2, -1, *kv.shape[-2:])
                    dkv_ = torch.empty_like(kv_)
3045
                    # [b, sq, np, hn] -> [b*sq, np, hn]
3046
3047
                    out_ = out.view(-1, *out.shape[-2:])
                    dout_ = dout.view(-1, *dout.shape[-2:])
3048
3049
3050
3051
3052
                    if _use_flash_attn_3 or _flash_attn_2_3_plus:
                        fa_backward_kwargs["window_size"] = (-1, -1)
                    if not _use_flash_attn_3:
                        fa_backward_kwargs["rng_state"] = rng_states[cp_size - i - 1]
                    flash_attn_bwd(
3053
3054
3055
3056
3057
3058
3059
3060
3061
                        dout_,
                        q_,
                        kv_[0],
                        kv_[1],
                        out_,
                        softmax_lse,
                        dq_,
                        dkv_[0],
                        dkv_[1],
3062
3063
                        cu_seqlens_q_per_step[cp_size - i - 1],
                        cu_seqlens_kv_per_step[cp_size - i - 1],
3064
                        ctx.max_seqlen_q,
3065
                        ctx.max_seqlen_kv,
3066
3067
                        causal=False,
                        **fa_backward_kwargs,
3068
3069
                    )

3070
3071
            if ctx.fp8:
                dq = dq_fp8[(rank + i + 1) % cp_size]
3072
            if i >= (cp_size - rank - 1) or not causal:
3073
3074
3075
3076
                # [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:
3077
3078
3079
3080
3081
3082
                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:])
3083

3084
3085
3086
3087
3088
3089
3090
3091
3092
3093
3094
            if ctx.fp8:
                if i >= (cp_size - rank - 1) or not causal:
                    dq.copy_(dq_)
                else:
                    if ctx.qkv_format == "bshd":
                        dq[:, 0, ...].fill_(0)
                        dq[:, 1, ...].copy_(dq_)
                    elif ctx.qkv_format == "sbhd":
                        dq[0].fill_(0)
                        dq[1].copy_(dq_)
            elif causal:
3095
                if i > (cp_size - rank - 1):
3096
                    dq.add_(dq_)
3097
3098
                elif i == (cp_size - rank - 1):
                    if rank == (cp_size - 1):
3099
3100
                        dq.copy_(dq_)
                    else:
3101
3102
3103
3104
3105
3106
                        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])
3107
                        elif ctx.qkv_format == "thd":
3108
                            tex.thd_grad_correction(dq, dq_, cu_seqlens_q_padded, "copy", "add")
3109
                elif i > 0:
3110
3111
3112
3113
                    if ctx.qkv_format == "bshd":
                        dq[:, 1, ...].add_(dq_)
                    elif ctx.qkv_format == "sbhd":
                        dq[1].add_(dq_)
3114
                    elif ctx.qkv_format == "thd":
3115
                        tex.thd_grad_correction(dq, dq_, cu_seqlens_q_padded, "none", "add")
3116
                else:
3117
3118
3119
3120
                    if ctx.qkv_format == "bshd":
                        dq[:, 1, ...].copy_(dq_)
                    elif ctx.qkv_format == "sbhd":
                        dq[1].copy_(dq_)
3121
                    elif ctx.qkv_format == "thd":
3122
                        tex.thd_grad_correction(dq, dq_, cu_seqlens_q_padded, "none", "copy")
3123
3124
3125
3126
3127
            else:
                if i == 0:
                    dq.copy_(dq_)
                else:
                    dq.add_(dq_)
3128

3129
            if attn_dbias is not None:
3130
                idx = (rank + i + 1) % cp_size
3131
                if i == (cp_size - 1) or not causal:
3132
                    # [b, np, sq, sk//cp] -> [b, np, sq, 2, sk//(2*cp)]
3133
                    dbias_ = dbias_.view(*dbias_.shape[:-1], 2, dbias_.shape[-1] // 2)
3134
                    attn_dbias[..., idx, :].copy_(dbias_[..., 0, :])
3135
3136
                    attn_dbias[..., (2 * cp_size - idx - 1), :].copy_(dbias_[..., 1, :])
                elif i >= (cp_size - rank - 1):
3137
3138
3139
3140
                    # [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)]
3141
                    dbias_ = dbias_.view(*dbias_.shape[:-1], 2, dbias_.shape[-1] // 2)
3142
                    attn_dbias_[..., 1, :, idx, :].copy_(dbias_[..., 0, :])
3143
                    attn_dbias_[..., 1, :, (2 * cp_size - idx - 1), :].copy_(dbias_[..., 1, :])
3144

3145
3146
3147
            # wait until dKV is received
            for req in send_recv_reqs:
                req.wait()
3148

3149
3150
3151
3152
3153
3154
3155
            if ctx.fp8:
                if i < cp_size - 1:
                    dkv = dkv_fp8_[(rank + i + 1) % cp_size]
                else:
                    dkv = dkv_fp8[(rank + i + 1) % cp_size]
            else:
                dkv = p2p_comm_buffers[(i + 1) % 2][1]
3156
            if ctx.use_fused_attention:
3157
3158
3159
                dkv_ = torch.cat(
                    (dk_.unsqueeze(0), dv_.unsqueeze(0)), dim=0
                )  # pylint: disable=used-before-assignment
3160
3161
3162
3163
                if ctx.qkv_format in ["bshd", "sbhd"]:
                    # [b, 2, sk//2, 2, np, hn] -> [2, b, 2, sk//2, np, hn] or
                    # [2, sk//2, b, 2, np, hn] -> [2, 2, sk//2, b, np, hn]
                    dkv = dkv.view(2, *dkv.shape[0:-3], *dkv.shape[-2:])
3164
            if causal and i >= (cp_size - rank - 1) and i != (cp_size - 1):
3165
3166
3167
3168
3169
3170
                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:])
3171
3172
3173
3174
            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)
3175

3176
3177
3178
3179
3180
3181
3182
3183
3184
3185
3186
            if ctx.fp8:
                if causal and i >= (cp_size - rank - 1) and i != (cp_size - 1):
                    if ctx.qkv_format == "bshd":
                        dkv[:, :, 0, ...].copy_(dkv_)
                        dkv[:, :, 1, ...].fill_(0)
                    elif ctx.qkv_format == "sbhd":
                        dkv[:, 0, ...].copy_(dkv_)
                        dkv[:, 1, ...].fill_(0)
                else:
                    dkv.copy_(dkv_)
            elif causal:
3187
                if i == (cp_size - 1):
3188
                    if rank == 0:
3189
3190
3191
3192
3193
3194
                        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, ...])
3195
                        elif ctx.qkv_format == "thd":
3196
                            tex.thd_grad_correction(dkv, dkv_, cu_seqlens_kv_padded, "add", "copy")
3197
3198
                    else:
                        dkv.add_(dkv_)
3199
3200
                elif i >= (cp_size - rank - 1):
                    if i == 0 and rank == (cp_size - 1):
3201
3202
3203
3204
                        if ctx.qkv_format == "bshd":
                            dkv[:, :, 0, ...].copy_(dkv_)
                        elif ctx.qkv_format == "sbhd":
                            dkv[:, 0, ...].copy_(dkv_)
3205
                        elif ctx.qkv_format == "thd":
3206
                            tex.thd_grad_correction(dkv, dkv_, cu_seqlens_kv_padded, "copy", "none")
3207
                    else:
3208
3209
3210
3211
                        if ctx.qkv_format == "bshd":
                            dkv[:, :, 0, ...].add_(dkv_)
                        elif ctx.qkv_format == "sbhd":
                            dkv[:, 0, ...].add_(dkv_)
3212
                        elif ctx.qkv_format == "thd":
3213
                            tex.thd_grad_correction(dkv, dkv_, cu_seqlens_kv_padded, "add", "none")
3214
3215
3216
3217
3218
                elif i > 0:
                    dkv.add_(dkv_)
                else:
                    dkv.copy_(dkv_)
            else:
3219
3220
3221
3222
3223
                if i == 0:
                    dkv.copy_(dkv_)
                else:
                    dkv.add_(dkv_)

3224
3225
3226
3227
3228
3229
3230
3231
3232
3233
3234
3235
3236
3237
3238
3239
3240
3241
3242
3243
        if ctx.fp8 and ctx.use_fused_attention:
            amax_cp_bwd = amax_per_step.amax(dim=1)
            ctx.fp8_meta["scaling_bwd"].amax_history[0][META_DP] = amax_cp_bwd[0]
            ctx.fp8_meta["scaling_bwd"].amax_history[0][META_DQKV_CP] = amax_cp_bwd[1]
            if ctx.qkv_format in ["bshd", "sbhd"]:
                # [cp, b, 2, sk//2, 2, np, hn] -> [cp, 2, b, 2, sk//2, np, hn] or
                # [cp, 2, sk//2, b, 2, np, hn] -> [cp, 2, 2, sk//2, b, np, hn]
                dkv_fp8 = dkv_fp8.view(cp_size, 2, *dkv_fp8.shape[1:-3], *dkv_fp8.shape[-2:])
            dq, dkv = [
                cast_from_fp8(
                    x,
                    ctx.fp8_meta["scaling_bwd"],
                    META_DQKV_CP,
                    fp8_dtype_backward,
                    TE_DType[torch.float32],
                )
                for x in [dq_fp8, dkv_fp8]
            ]
            dq, dkv = [x.sum(dim=0).to(dout_dtype) for x in [dq, dkv]]

3244
        if causal:
3245
3246
            if ctx.qkv_format == "bshd":
                # [b, 2, sq//2, np, hn] -> [b, sq, np, hn]
3247
                dq = dq.view(dq.shape[0], -1, *dq.shape[-2:])
3248
                # [2, b, 2, sk//2, np, hn] -> [2, b, sk, np, hn]
3249
                dkv = dkv.view(*dkv.shape[0:2], -1, *dkv.shape[-2:])
3250
3251
            elif ctx.qkv_format == "sbhd":
                # [2, sq//2, b, np, hn] -> [sq, b, np, hn]
3252
                dq = dq.view(-1, *dq.shape[-3:])
3253
                # [2, 2, sk//2, b, np, hn] -> [2, sk, b, np, hn]
3254
3255
3256
3257
3258
3259
3260
3261
3262
                dkv = dkv.view(dkv.shape[0], -1, *dkv.shape[-3:])

        if ctx.qkv_format == "thd":
            dkv_ = torch.empty(
                2, ctx.total_tokens_kv, *dkv.shape[-2:], dtype=dkv.dtype, device=dkv.device
            )
            dkv_[:, : cu_seqlens_kv_padded[-1]].copy_(dkv)
            dkv_[:, cu_seqlens_kv_padded[-1] :].fill_(0)
            dkv = dkv_
3263

3264
        if ctx.fp8 and ctx.is_input_fp8:
3265
3266
3267
3268
            dq, dkv = [
                cast_to_fp8(x, ctx.fp8_meta["scaling_bwd"], META_DQKV, fp8_dtype_backward)
                for x in [dq, dkv]
            ]
3269
3270
3271
3272
3273
3274
3275
3276
3277
3278
3279
3280
3281
3282
3283
3284
3285
3286
        dk, dv = dkv[0], dkv[1]

        if cp_size_a2a > 1:
            chunk_ids_for_a2a = get_seq_chunk_ids_for_reordering(cp_size_a2a, q.device, False)
            dq, dk, dv = flash_attn_a2a_communicate(
                [dq, dk, dv],
                chunk_ids_for_a2a,
                seq_dim,
                cp_size_a2a,
                ctx.cp_group_a2a,
                ctx.cp_stream,
                False,
            )
            if ctx.qkv_format == "bshd":
                dq, dk, dv = [x.view(ctx.batch_size, -1, *x.shape[-2:]) for x in [dq, dk, dv]]
            elif ctx.qkv_format == "sbhd":
                dq, dk, dv = [x.view(-1, ctx.batch_size, *x.shape[-2:]) for x in [dq, dk, dv]]

3287
        if ctx.fp8 and ctx.is_input_fp8:
3288
3289
3290
3291
3292
3293
3294
3295
3296
            dq, dk, dv = [
                Float8Tensor(
                    data=x,
                    fp8_meta=ctx.fp8_meta,
                    fp8_meta_forward=False,
                    fp8_meta_index=META_DQKV,
                    fp8_dtype=fp8_dtype_backward,
                    dtype=dout_dtype,
                )
3297
                for x in [dq, dk, dv]
3298
3299
            ]

3300
3301
3302
3303
        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)

3304
3305
3306
        return (
            None,
            dq,
3307
3308
            dk,
            dv,
3309
3310
3311
3312
3313
3314
3315
3316
3317
3318
3319
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
3320
            attn_dbias,
3321
3322
3323
3324
3325
            None,
            None,
            None,
            None,
            None,
3326
3327
            None,
            None,
3328
        )
3329
3330


3331
3332
def get_kv_seq_info_after_all_gather(
    local_chunk_id, cp_size, max_seqlen_q, max_seqlen_kv, window_size, causal
3333
):
3334
3335
3336
3337
3338
3339
3340
3341
3342
3343
3344
3345
3346
3347
3348
3349
3350
3351
3352
3353
3354
3355
    """Compute KV sequence index range and update window size after all-gather."""
    local_chunk_end_idx = (local_chunk_id + 1) * max_seqlen_kv
    full_seq_end_idx = max_seqlen_kv * cp_size * 2

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

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

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

    return (seq_start_idx, seq_end_idx), (window_size_left, window_size_right)
3356
3357
3358
3359


class AttnFuncWithCPAndKVAllGather(torch.autograd.Function):
    """
3360
3361
    Attention implementation with context parallelism. KV all-gather between CP ranks is exposed.
    Refer section 3.3.2 of `The Llama 3 Herd of Models <https://arxiv.org/abs/2407.21783>`_.
3362
3363
3364
3365
3366
3367
3368
3369
3370
3371
3372
3373
3374
3375
3376
3377
3378
3379
3380
3381
3382
3383
    """

    @staticmethod
    def forward(
        ctx,
        is_training,
        q,
        k,
        v,
        cu_seqlens_q,
        max_seqlen_q,
        max_seqlen_kv,
        cu_seqlens_q_padded,
        dropout_p,
        softmax_scale,
        qkv_format,
        attn_mask_type,
        attn_bias_type,
        attn_bias,
        deterministic,
        use_fused_attention,
        window_size,
3384
3385
        cp_group,
        cp_stream,
3386
    ):
3387
        # pylint: disable=missing-function-docstring
3388
3389
3390
3391
3392
3393
3394
3395
        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)

        causal = "causal" in attn_mask_type
        padding = "padding" in attn_mask_type
3396
        assert not padding, f"{attn_mask_type} mask type is not supported!"
3397
3398
3399
3400
3401
3402
3403
        if use_fused_attention and causal and "bottom_right" not in attn_mask_type:
            attn_mask_type = attn_mask_type + "_bottom_right"
        assert attn_bias_type == "no_bias", f"{attn_bias_type} bias type is not supported!"
        assert q.shape[-1] % 8 == 0, "Hidden size per attention head should be multiple of 8!"
        assert (
            use_fused_attention or _flash_attn_2_3_plus
        ), "Sliding window attention only can work with FusedAttention or FlashAttention >= 2.3!"
3404

3405
        flash_attn_fwd = None
3406
3407
3408
3409
3410
3411
3412
3413
3414
3415
3416
3417
        if not use_fused_attention:
            fa_forward_kwargs = {"softmax_scale": softmax_scale}
            if _use_flash_attn_3:
                flash_attn_fwd = flash_attn_varlen_fwd_v3
            else:
                flash_attn_fwd = flash_attn_varlen_fwd
                fa_forward_kwargs["dropout_p"] = dropout_p
                fa_forward_kwargs["return_softmax"] = False
                if _flash_attn_2_4_plus:
                    fa_forward_kwargs["alibi_slopes"] = None
                if _flash_attn_2_5_7_plus:
                    fa_forward_kwargs["block_table"] = None
3418
3419
3420
3421
3422
3423
3424
3425
3426
3427
3428
3429
3430
3431

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

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

        max_seqlen_q = max_seqlen_q // (2 * cp_size)
        max_seqlen_kv = max_seqlen_kv // (2 * cp_size)
        cu_seqlens_q = cu_seqlens_q // (2 * cp_size)
        cu_seqlens_q_padded = cu_seqlens_q_padded // (2 * cp_size)

3432
3433
3434
3435
        # [b, s, np, hn] -> [b, 2, s//2, np, hn] or [s, b, np, hn] -> [2, s//2, b, np, hn]
        q = q.view(*q.shape[:seq_dim], 2, q.shape[seq_dim] // 2, *q.shape[(seq_dim + 1) :])
        # [b, s, np, hn] or [s, b, np, hn] -> [s, b, np, hn]
        k, v = [x.movedim(seq_dim, 0).contiguous() for x in [k, v]]
3436

3437
        # [s, b, np, hn] -> [cp, s, b, np, hn]
3438
3439
        k_ag, _ = gather_along_first_dim(k, cp_group)
        v_ag, _ = gather_along_first_dim(v, cp_group)
3440
3441

        # [cp, s, b, np, hn] -> [cp*2, s//2, b, np, hn]
3442
3443
        k_ag = k_ag.view(2 * cp_size, k.shape[0] // 2, *k.shape[1:])
        v_ag = v_ag.view(2 * cp_size, v.shape[0] // 2, *v.shape[1:])
3444
3445
3446
3447
3448
3449
3450
3451
3452
3453
        chunk_ids_for_kv_ag = get_seq_chunk_ids_for_reordering(cp_size, k.device, True)
        k_ag = torch.index_select(k_ag, dim=0, index=chunk_ids_for_kv_ag)
        v_ag = torch.index_select(v_ag, dim=0, index=chunk_ids_for_kv_ag)
        # [cp*2, s//2, b, np, hn] -> [cp*s, b, np, hn]
        k_ag = k_ag.view(-1, *k.shape[1:])
        v_ag = v_ag.view(-1, *v.shape[1:])
        cp_stream.wait_stream(torch.cuda.current_stream())

        # create two streams to resolve wave quantization issue of Flash Attn in each step
        flash_attn_streams = [torch.cuda.current_stream(), cp_stream]
3454
3455

        local_seq_chunk_ids = [rank, 2 * cp_size - rank - 1]
3456
3457
3458
        kv_seq_range_per_step = [None, None]
        window_size_per_step = [None, None]
        cu_seqlens_kv_per_step = [None, None]
3459
3460
3461
3462
3463
3464
3465
3466
        out_per_step = [None, None]
        softmax_lse_per_step = [None, None]
        rng_states = [None, None]
        out = torch.empty_like(q)

        for i in range(len(local_seq_chunk_ids) + 1):
            if i < len(local_seq_chunk_ids):
                with torch.cuda.stream(flash_attn_streams[i]):
3467
3468
                    # [b, 2, sq//2, np, hn] -> [b, sq//2, np, hn]
                    # or [2, sq//2, b, np, hn] -> [sq//2, b, np, hn]
3469
3470
3471
3472
3473
3474
3475
3476
3477
                    q_ = q.select(seq_dim, i).contiguous()
                    kv_seq_range_per_step[i], window_size_per_step[i] = (
                        get_kv_seq_info_after_all_gather(
                            local_seq_chunk_ids[i],
                            cp_size,
                            max_seqlen_q,
                            max_seqlen_kv,
                            window_size,
                            causal,
3478
                        )
3479
3480
3481
3482
3483
3484
3485
3486
3487
3488
3489
3490
                    )
                    seq_start_idx, seq_end_idx = (
                        kv_seq_range_per_step[i][0],
                        kv_seq_range_per_step[i][1],
                    )
                    max_seqlen_kv_ = seq_end_idx - seq_start_idx
                    cu_seqlens_kv_per_step[i] = _get_full_cu_seqlens(
                        k.shape[1], max_seqlen_kv_, k.device
                    )
                    k_, v_ = [x[seq_start_idx:seq_end_idx] for x in [k_ag, v_ag]]
                    # [s_range, b, np, hn] -> [b, s_range, np, hn] or [s_range, b, np, hn]
                    k_, v_ = [x.movedim(0, seq_dim).contiguous() for x in [k_, v_]]
3491
3492
3493
3494
                    if use_fused_attention:
                        out_per_step[i], [softmax_lse_per_step[i], rng_states[i]] = fused_attn_fwd(
                            is_training,
                            max_seqlen_q,
3495
                            max_seqlen_kv_,
3496
                            cu_seqlens_q,
3497
                            cu_seqlens_kv_per_step[i],
3498
3499
3500
3501
3502
3503
3504
3505
3506
3507
3508
3509
                            q_,
                            k_,
                            v_,
                            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,
                            cu_seqlens_q_padded=cu_seqlens_q_padded,
3510
3511
                            cu_seqlens_kv_padded=cu_seqlens_kv_per_step[i],
                            window_size=window_size_per_step[i],
3512
3513
3514
                        )
                    else:
                        q_, k_, v_ = [x.view(-1, *x.shape[-2:]) for x in [q_, k_, v_]]
3515
3516
3517
3518
3519
3520
3521
3522
3523
3524
3525
                        fa_outputs = flash_attn_fwd(
                            q_,
                            k_,
                            v_,
                            cu_seqlens_q,
                            cu_seqlens_kv_per_step[i],
                            max_seqlen_q,
                            max_seqlen_kv_,
                            causal=causal,
                            window_size=window_size_per_step[i],
                            **fa_forward_kwargs,
3526
                        )
3527
3528
3529
3530
                        out_per_step[i] = fa_outputs[4]
                        softmax_lse_per_step[i] = fa_outputs[5]
                        if not _use_flash_attn_3:
                            rng_states[i] = fa_outputs[7]
3531
3532
3533
3534

            if i > 0:
                with torch.cuda.stream(flash_attn_streams[i - 1]):
                    if qkv_format == "bshd":
3535
                        out[:, i - 1].copy_(out_per_step[i - 1].view(out[:, i - 1].shape))
3536
                    elif qkv_format == "sbhd":
3537
                        out[i - 1].copy_(out_per_step[i - 1].view(out[i - 1].shape))
3538
3539
3540
3541
3542
3543
3544
3545
3546
3547
3548
3549
3550
3551
3552
3553
3554

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

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

        ctx.save_for_backward(
            q,
            k,
            v,
            cu_seqlens_q,
            cu_seqlens_q_padded,
3555
            *cu_seqlens_kv_per_step,
3556
3557
3558
3559
            *out_per_step,
            *softmax_lse_per_step,
            *rng_states,
        )
3560
3561
        ctx.kv_seq_range_per_step = kv_seq_range_per_step
        ctx.window_size_per_step = window_size_per_step
3562
3563
3564
3565
3566
3567
3568
        ctx.cp_group = cp_group
        ctx.cp_stream = cp_stream
        ctx.dropout_p = dropout_p
        ctx.max_seqlen_q = max_seqlen_q
        ctx.softmax_scale = softmax_scale
        ctx.qkv_format = qkv_format
        ctx.attn_bias_type = attn_bias_type
3569
        ctx.attn_mask_type = attn_mask_type
3570
3571
3572
3573
3574
3575
        ctx.deterministic = deterministic
        ctx.use_fused_attention = use_fused_attention
        return out

    @staticmethod
    def backward(ctx, dout):
3576
        # pylint: disable=missing-function-docstring
3577
3578
3579
        cp_size = get_distributed_world_size(ctx.cp_group)
        rank = get_distributed_rank(ctx.cp_group)

3580
3581
3582
3583
3584
3585
3586
        (q, k, v, cu_seqlens_q, cu_seqlens_q_padded) = ctx.saved_tensors[:5]
        cu_seqlens_kv_per_step = ctx.saved_tensors[5:7]
        out_per_step = ctx.saved_tensors[7:9]
        softmax_lse_per_step = ctx.saved_tensors[9:11]
        rng_states = ctx.saved_tensors[11:13]
        kv_seq_range_per_step = ctx.kv_seq_range_per_step
        window_size_per_step = ctx.window_size_per_step
3587

3588
        seq_dim = ctx.qkv_format.index("s")
3589
3590
        qkv_layout = ctx.qkv_format + "_" + ctx.qkv_format + "_" + ctx.qkv_format

3591
        dout = dout.view(q.shape)
3592
        dq = torch.empty_like(q)
3593
        dk = torch.zeros((k.shape[0] * cp_size, *k.shape[1:]), dtype=k.dtype, device=k.device)
3594
3595
3596
3597
3598
3599
3600
3601
3602
3603
        dv = torch.zeros_like(dk)
        dq_per_step = [None, None]
        dk_per_step = [None, None]
        dv_per_step = [None, None]

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

3604
        # [s, b, np, hn] -> [cp, s, b, np, hn]
3605
3606
        k_ag, _ = gather_along_first_dim(k, ctx.cp_group)
        v_ag, _ = gather_along_first_dim(v, ctx.cp_group)
3607
3608

        # [cp, s, b, np, hn] -> [cp*2, s//2, b, np, hn]
3609
3610
        k_ag = k_ag.view(2 * cp_size, k.shape[0] // 2, *k.shape[1:])
        v_ag = v_ag.view(2 * cp_size, v.shape[0] // 2, *v.shape[1:])
3611
3612
3613
3614
3615
3616
3617
        chunk_ids_for_kv_ag = get_seq_chunk_ids_for_reordering(cp_size, k.device, True)
        k_ag = torch.index_select(k_ag, dim=0, index=chunk_ids_for_kv_ag)
        v_ag = torch.index_select(v_ag, dim=0, index=chunk_ids_for_kv_ag)
        # [cp*2, s//2, b, np, hn] -> [cp*s, b, np, hn]
        k_ag = k_ag.view(-1, *k.shape[1:])
        v_ag = v_ag.view(-1, *v.shape[1:])
        ctx.cp_stream.wait_stream(torch.cuda.current_stream())
3618
3619
3620

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

3621
        flash_attn_bwd = None
3622
3623
3624
3625
3626
3627
3628
3629
3630
3631
3632
3633
        if not ctx.use_fused_attention:
            fa_backward_kwargs = {"softmax_scale": ctx.softmax_scale}
            if _use_flash_attn_3:
                flash_attn_bwd = flash_attn_varlen_bwd_v3
                fa_backward_kwargs["deterministic"] = ctx.deterministic
            else:
                flash_attn_bwd = flash_attn_varlen_bwd
                fa_backward_kwargs["dropout_p"] = ctx.dropout_p
                if _flash_attn_2_4_plus:
                    fa_backward_kwargs["alibi_slopes"] = None
                if _flash_attn_2_4_1_plus:
                    fa_backward_kwargs["deterministic"] = ctx.deterministic
3634
3635
3636
3637

        for i in range(len(local_seq_chunk_ids) + 1):
            if i < len(local_seq_chunk_ids):
                with torch.cuda.stream(flash_attn_streams[i]):
3638
3639
                    # [b, 2, sq//2, np, hn] -> [b, sq//2, np, hn]
                    # or [2, sq//2, b, np, hn] -> [sq//2, b, np, hn]
3640
3641
3642
3643
3644
3645
3646
3647
3648
                    q_ = q.select(seq_dim, i).contiguous()
                    seq_start_idx, seq_end_idx = (
                        kv_seq_range_per_step[i][0],
                        kv_seq_range_per_step[i][1],
                    )
                    max_seqlen_kv = seq_end_idx - seq_start_idx
                    k_, v_ = [x[seq_start_idx:seq_end_idx] for x in [k_ag, v_ag]]
                    # [cp*s, b, np, hn] -> [b, s_range, np, hn] or [s_range, b, np, hn]
                    k_, v_ = [x.movedim(0, seq_dim).contiguous() for x in [k_, v_]]
3649
                    out_ = out_per_step[i]
3650
                    dout_ = dout.select(seq_dim, i).contiguous().view(out_.shape)
3651
3652
3653
3654
                    if ctx.use_fused_attention:
                        aux_ctx_tensors = [softmax_lse_per_step[i], rng_states[i]]
                        dq_per_step[i], dk_per_step[i], dv_per_step[i], _ = fused_attn_bwd(
                            ctx.max_seqlen_q,
3655
                            max_seqlen_kv,
3656
                            cu_seqlens_q,
3657
                            cu_seqlens_kv_per_step[i],
3658
3659
3660
3661
3662
3663
                            q_,
                            k_,
                            v_,
                            out_,
                            dout_,
                            TE_DType[q.dtype],
3664
                            TE_DType[dout.dtype],
3665
3666
3667
                            aux_ctx_tensors,
                            tex.NVTE_Fused_Attn_Backend.NVTE_F16_arbitrary_seqlen,
                            cu_seqlens_q_padded=cu_seqlens_q_padded,
3668
                            cu_seqlens_kv_padded=cu_seqlens_kv_per_step[i],
3669
3670
3671
3672
3673
                            attn_scale=ctx.softmax_scale,
                            dropout=ctx.dropout_p,
                            qkv_layout=qkv_layout,
                            attn_mask_type=ctx.attn_mask_type,
                            attn_bias_type=ctx.attn_bias_type,
3674
3675
                            window_size=window_size_per_step[i],
                            deterministic=ctx.deterministic,
3676
3677
                        )
                    else:
3678
                        batch_size = k_.shape[0]
3679
3680
3681
3682
                        q_, k_, v_ = [x.view(-1, *x.shape[-2:]) for x in [q_, k_, v_]]
                        dq_per_step[i], dk_per_step[i], dv_per_step[i] = [
                            torch.empty_like(x) for x in [q_, k_, v_]
                        ]
3683
3684
3685
                        if not _use_flash_attn_3:
                            fa_backward_kwargs["rng_state"] = rng_states[i]
                        flash_attn_bwd(
3686
3687
3688
3689
3690
3691
3692
3693
3694
3695
                            dout_,
                            q_,
                            k_,
                            v_,
                            out_,
                            softmax_lse_per_step[i],
                            dq_per_step[i],
                            dk_per_step[i],
                            dv_per_step[i],
                            cu_seqlens_q,
3696
                            cu_seqlens_kv_per_step[i],
3697
                            ctx.max_seqlen_q,
3698
                            max_seqlen_kv,
3699
                            causal="causal" in ctx.attn_mask_type,
3700
                            window_size=window_size_per_step[i],
3701
                            **fa_backward_kwargs,
3702
                        )
3703
3704
3705
3706
3707
3708
3709
                        # [b*sq//2, np, hn] -> [b, sq//2, np, hn]
                        dq_per_step[i] = dq_per_step[i].view(dq[:, i].shape)
                        # [b*s_range, np, hn] -> [b, s_range, np, hn]
                        dk_per_step[i], dv_per_step[i] = [
                            x.view(batch_size, -1, *x.shape[-2:])
                            for x in [dk_per_step[i], dv_per_step[i]]
                        ]
3710
3711
3712
3713

            if i > 0:
                with torch.cuda.stream(flash_attn_streams[i - 1]):
                    if ctx.qkv_format == "bshd":
3714
                        dq[:, i - 1].copy_(dq_per_step[i - 1])
3715
                    elif ctx.qkv_format == "sbhd":
3716
3717
3718
3719
3720
3721
                        dq[i - 1].copy_(dq_per_step[i - 1])
                    # [b, s_range, np, hn] or [s_range, b, np, hn] -> [s_range, b, np, hn]
                    dk_per_step[i - 1], dv_per_step[i - 1] = [
                        x.movedim(seq_dim, 0).contiguous()
                        for x in [dk_per_step[i - 1], dv_per_step[i - 1]]
                    ]
3722
3723
3724
                    # wait until dkv update of last step is done
                    if i > 1:
                        flash_attn_streams[i - 1].wait_event(dkv_update_done)
3725
3726
3727
3728
3729
3730
                    seq_start_idx, seq_end_idx = (
                        kv_seq_range_per_step[i - 1][0],
                        kv_seq_range_per_step[i - 1][1],
                    )
                    dk[seq_start_idx:seq_end_idx].add_(dk_per_step[i - 1])
                    dv[seq_start_idx:seq_end_idx].add_(dv_per_step[i - 1])
3731
3732
3733
3734
3735
                    if i < len(local_seq_chunk_ids):
                        flash_attn_streams[i - 1].record_event(dkv_update_done)

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

3736
3737
3738
3739
3740
3741
3742
        # [cp*s, b, np, hn] -> [cp*2, s//2, b, np, hn]
        dk = dk.view(2 * cp_size, -1, *dk.shape[-3:])
        dv = dv.view(2 * cp_size, -1, *dv.shape[-3:])
        chunk_ids_for_kv_ag = get_seq_chunk_ids_for_reordering(cp_size, dk.device, False)
        dk = torch.index_select(dk, dim=0, index=chunk_ids_for_kv_ag)
        dv = torch.index_select(dv, dim=0, index=chunk_ids_for_kv_ag)
        # [cp*2, s//2, b, np, hn] -> [cp*s, b, np, hn]
3743
3744
3745
3746
3747
        dk = dk.view(-1, *dk.shape[-3:])
        dv = dv.view(-1, *dv.shape[-3:])
        dk, _ = reduce_scatter_along_first_dim(dk, ctx.cp_group)
        dv, _ = reduce_scatter_along_first_dim(dv, ctx.cp_group)

3748
3749
3750
3751
3752
3753
3754
3755
3756
3757
3758
3759
3760
3761
3762
3763
3764
3765
3766
3767
3768
3769
3770
3771
3772
3773
3774
3775
3776
3777
3778
3779
3780
3781
3782
3783
3784
3785
3786
3787
3788
3789
3790
3791
3792
3793
3794
3795
3796
3797
3798
3799
3800
3801
3802
3803
3804
3805
3806
3807
        dq = dq.view(*dq.shape[:seq_dim], -1, *dq.shape[(seq_dim + 2) :])
        dk = dk.movedim(0, seq_dim).contiguous()
        dv = dv.movedim(0, seq_dim).contiguous()

        return (
            None,
            dq,
            dk,
            dv,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
        )


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

    @staticmethod
    def forward(
        ctx,
        is_training,
        q,
        k,
        v,
        cu_seqlens_q,
        cu_seqlens_kv,
        max_seqlen_q,
        max_seqlen_kv,
        cu_seqlens_q_padded,
        cu_seqlens_kv_padded,
        dropout_p,
        softmax_scale,
        qkv_format,
        attn_mask_type,
        attn_bias_type,
        attn_bias,
        deterministic,
        use_fused_attention,
        window_size,
        fp8,
        fp8_meta,
        cp_group,
        cp_stream,
    ):
3808
        # pylint: disable=missing-function-docstring
3809
3810
3811
3812
3813
3814
3815
3816
3817
3818
3819
3820
3821
3822
3823
3824
        if softmax_scale is None:
            softmax_scale = q.shape[-1] ** (-0.5)

        cp_size = get_distributed_world_size(cp_group)

        causal = "causal" in attn_mask_type
        padding = "padding" in attn_mask_type
        assert not padding, f"{attn_mask_type} mask type is not supported!"
        assert attn_bias_type == "no_bias", f"{attn_bias_type} bias type is not supported!"
        assert q.shape[-1] % 8 == 0, "Hidden size per attention head should be multiple of 8!"
        assert (
            window_size == (-1, 0)
            or window_size == (-1, -1)
            or use_fused_attention
            or _flash_attn_2_3_plus
        ), "Sliding window attention only can work with FusedAttention or FlashAttention >= 2.3!"
3825

3826
        flash_attn_fwd = None
3827
3828
3829
3830
3831
3832
3833
3834
3835
3836
3837
3838
3839
3840
3841
        if not use_fused_attention:
            fa_forward_kwargs = {"softmax_scale": softmax_scale}
            if _use_flash_attn_3:
                flash_attn_fwd = flash_attn_varlen_fwd_v3
                fa_forward_kwargs["window_size"] = window_size
            else:
                flash_attn_fwd = flash_attn_varlen_fwd
                fa_forward_kwargs["dropout_p"] = dropout_p
                fa_forward_kwargs["return_softmax"] = False
                if _flash_attn_2_3_plus:
                    fa_forward_kwargs["window_size"] = window_size
                if _flash_attn_2_4_plus:
                    fa_forward_kwargs["alibi_slopes"] = None
                if _flash_attn_2_5_7_plus:
                    fa_forward_kwargs["block_table"] = None
3842
3843
3844
3845
3846
3847
3848
3849
3850
3851
3852
3853
3854
3855

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

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

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

3856
        qkv_dtype = q.dtype
3857
3858
        fused_attn_backend = None
        fused_attn_qkv_dtype = None
3859
3860
3861
        # "fp8_mha" decides outputs in fp8, while inputs are inferred from the real dtype
        is_input_fp8 = False
        is_output_fp8 = fp8_meta is not None and fp8_meta["recipe"].fp8_mha
3862
3863
3864
3865
3866
        if fp8:
            if use_fused_attention:
                fp8_dtype_forward = get_fp8_te_dtype(fp8_meta["recipe"], fprop_tensor=True)
                fused_attn_qkv_dtype = fp8_dtype_forward
                fused_attn_backend = FusedAttnBackend["FP8"]
3867
3868
3869
3870
3871
                assert isinstance(k, q.__class__) and isinstance(
                    v, q.__class__
                ), "q, k, and v must have the same type."
                is_input_fp8 = isinstance(q, Float8Tensor)
                if is_input_fp8:
3872
3873
3874
3875
3876
3877
3878
3879
3880
3881
3882
3883
3884
3885
3886
3887
3888
3889
3890
3891
3892
3893
3894
3895
3896
3897
3898
3899
3900
3901
3902
3903
3904
3905
3906
                    fp8_meta["scaling_fwd"].scale_inv[META_QKV] = q._scale_inv
                    q_fp8, k_fp8, v_fp8 = q, k, v
                    q, k, v = q_fp8._data, k_fp8._data, v_fp8._data
                elif int(os.getenv("NVTE_FP8_DPA_BWD", "1")):
                    q_f16, k_f16, v_f16 = q, k, v
                    q, k, v = [
                        cast_to_fp8(x, fp8_meta["scaling_fwd"], META_QKV, fp8_dtype_forward)
                        for x in [q_f16, k_f16, v_f16]
                    ]
                fp8_meta_kwargs = {}
                fp8_meta_kwargs["d_scale_qkv"] = fp8_meta["scaling_fwd"].scale_inv
                fp8_meta_kwargs["d_scale_qkv_offset"] = META_QKV
                fp8_meta_kwargs["d_scale_s"] = fp8_meta["scaling_fwd"].scale_inv
                fp8_meta_kwargs["d_scale_s_offset"] = META_S
                fp8_meta_kwargs["q_scale_s"] = fp8_meta["scaling_fwd"].scale
                fp8_meta_kwargs["q_scale_s_offset"] = META_S
                fp8_meta_kwargs["q_scale_o"] = fp8_meta["scaling_fwd"].scale
                fp8_meta_kwargs["q_scale_o_offset"] = META_O
                fp8_meta_kwargs["amax_s"] = fp8_meta["scaling_fwd"].amax_history
                fp8_meta_kwargs["amax_s_offset"] = META_S
                fp8_meta_kwargs["amax_o"] = fp8_meta["scaling_fwd"].amax_history
                fp8_meta_kwargs["amax_o_offset"] = META_O
            else:
                assert False, "FP8 is only supported with Fused Attention!"
        else:
            if use_fused_attention:
                fp8_meta_kwargs = {}
                fused_attn_qkv_dtype = TE_DType[q.dtype]
                fused_attn_backend = FusedAttnBackend["F16_arbitrary_seqlen"]

        chunk_ids_for_a2a = get_seq_chunk_ids_for_reordering(cp_size, q.device, True)
        q, k, v = flash_attn_a2a_communicate(
            [q, k, v], chunk_ids_for_a2a, seq_dim, cp_size, cp_group, cp_stream, True
        )

3907
        if fp8 and not is_input_fp8 and not int(os.getenv("NVTE_FP8_DPA_BWD", "1")):
3908
3909
3910
3911
3912
3913
3914
3915
3916
3917
3918
3919
3920
3921
3922
3923
3924
3925
3926
3927
3928
3929
3930
3931
3932
3933
3934
3935
3936
3937
3938
3939
3940
            q_f16, k_f16, v_f16 = q, k, v
            q, k, v = [
                cast_to_fp8(x, fp8_meta["scaling_fwd"], META_QKV, fp8_dtype_forward)
                for x in [q_f16, k_f16, v_f16]
            ]

        batch_size = q.shape[batch_dim]
        if use_fused_attention:
            out, aux_ctx_tensors = fused_attn_fwd(
                is_training,
                max_seqlen_q,
                max_seqlen_kv,
                cu_seqlens_q,
                cu_seqlens_kv,
                q,
                k,
                v,
                fused_attn_qkv_dtype,
                fused_attn_backend,
                attn_scale=softmax_scale,
                dropout=dropout_p,
                qkv_layout=qkv_layout,
                attn_mask_type=attn_mask_type,
                attn_bias_type=attn_bias_type,
                attn_bias=attn_bias,
                cu_seqlens_q_padded=cu_seqlens_q_padded,
                cu_seqlens_kv_padded=cu_seqlens_kv_padded,
                window_size=window_size,
                **fp8_meta_kwargs,
            )
        else:
            # [b, cp*s, np//cp, hn] -> [b*cp*s, np//cp, hn]
            q, k, v = [x.view(-1, *x.shape[-2:]) for x in [q, k, v]]
3941
            fa_outputs = flash_attn_fwd(
3942
3943
3944
3945
3946
3947
3948
3949
                q,
                k,
                v,
                cu_seqlens_q,
                cu_seqlens_kv,
                max_seqlen_q,
                max_seqlen_kv,
                causal=causal,
3950
                **fa_forward_kwargs,
3951
            )
3952
3953
            out, softmax_lse = fa_outputs[4], fa_outputs[5]
            rng_state = fa_outputs[7] if not _use_flash_attn_3 else None
3954
3955
3956
3957
3958
3959
3960
3961
3962
3963
3964
3965
3966
3967
3968
3969
3970
3971
            aux_ctx_tensors = [softmax_lse, rng_state]
            # [b*cp*s, np//cp, hn] -> [b, cp*s, np//cp, hn]
            out = out.view(batch_size, -1, *out.shape[-2:])

        chunk_ids_for_a2a = get_seq_chunk_ids_for_reordering(cp_size, out.device, False)
        out = flash_attn_a2a_communicate(
            out, chunk_ids_for_a2a, seq_dim, cp_size, cp_group, cp_stream, False
        )

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

        if fp8:
3972
            if is_output_fp8:
3973
3974
3975
3976
3977
3978
                out_fp8 = Float8Tensor(
                    data=out,
                    fp8_meta=fp8_meta,
                    fp8_meta_forward=True,
                    fp8_meta_index=META_O,
                    fp8_dtype=fp8_dtype_forward,
3979
                    dtype=qkv_dtype,
3980
3981
3982
3983
3984
3985
3986
3987
3988
3989
3990
3991
3992
3993
3994
3995
3996
3997
                )
                out = out_fp8._data
                out_ret = out_fp8
            else:
                out_f16 = cast_from_fp8(
                    out,
                    fp8_meta["scaling_fwd"],
                    META_O,
                    fp8_dtype_forward,
                    TE_DType[q_f16.dtype],
                )
                out_ret = out_f16
        else:
            out_ret = out

        if fp8:
            if int(os.getenv("NVTE_FP8_DPA_BWD", "1")):
                q_save, k_save, v_save, out_save = q, k, v, out
3998
            elif is_input_fp8:
3999
4000
4001
4002
4003
4004
4005
4006
4007
4008
4009
4010
4011
4012
4013
4014
4015
4016
4017
4018
4019
4020
4021
4022
4023
4024
4025
4026
4027
4028
4029
4030
4031
4032
4033
4034
4035
4036
4037
4038
4039
4040
4041
4042
4043
4044
4045
4046
4047
4048
4049
                q_fp8, k_fp8, v_fp8 = [
                    Float8Tensor(
                        data=x,
                        fp8_meta=fp8_meta,
                        fp8_meta_forward=True,
                        fp8_meta_index=META_QKV,
                        fp8_dtype=fp8_dtype_forward,
                        dtype=out_fp8.dtype,
                    )
                    for x in [q, k, v]
                ]
                q_save, k_save, v_save, out_save = q_fp8, k_fp8, v_fp8, out_fp8
            else:
                q_save, k_save, v_save, out_save = q_f16, k_f16, v_f16, out_f16
        else:
            q_save, k_save, v_save, out_save = q, k, v, out

        if fp8 and int(os.getenv("NVTE_FP8_DPA_BWD", "1")):
            fp8_fwd_scales = fp8_meta["scaling_fwd"].scale.clone()
            fp8_fwd_scale_invs = fp8_meta["scaling_fwd"].scale_inv.clone()
        else:
            fp8_fwd_scales, fp8_fwd_scale_invs = None, None

        ctx.save_for_backward(
            q_save,
            k_save,
            v_save,
            out_save,
            cu_seqlens_q,
            cu_seqlens_kv,
            cu_seqlens_q_padded,
            cu_seqlens_kv_padded,
            fp8_fwd_scales,
            fp8_fwd_scale_invs,
            *aux_ctx_tensors,
        )
        ctx.batch_size = batch_size
        ctx.cp_group = cp_group
        ctx.cp_stream = cp_stream
        ctx.dropout_p = dropout_p
        ctx.max_seqlen_q = max_seqlen_q
        ctx.max_seqlen_kv = max_seqlen_kv
        ctx.softmax_scale = softmax_scale
        ctx.qkv_format = qkv_format
        ctx.attn_mask_type = attn_mask_type
        ctx.attn_bias_type = attn_bias_type
        ctx.deterministic = deterministic
        ctx.window_size = window_size
        ctx.use_fused_attention = use_fused_attention
        ctx.fp8 = fp8 and int(os.getenv("NVTE_FP8_DPA_BWD", "1"))
        ctx.fp8_meta = fp8_meta
4050
4051
        ctx.is_input_fp8 = is_input_fp8
        ctx.is_output_fp8 = is_output_fp8
4052
4053
4054
4055
        return out_ret

    @staticmethod
    def backward(ctx, dout):
4056
        # pylint: disable=missing-function-docstring
4057
4058
4059
4060
4061
4062
4063
4064
4065
4066
4067
4068
4069
        cp_size = get_distributed_world_size(ctx.cp_group)

        q, k, v, out = ctx.saved_tensors[:4]
        cu_seqlens_q, cu_seqlens_kv, cu_seqlens_q_padded, cu_seqlens_kv_padded = ctx.saved_tensors[
            4:8
        ]
        fp8_fwd_scales, fp8_fwd_scale_invs = ctx.saved_tensors[8:10]
        aux_ctx_tensors = ctx.saved_tensors[10:]

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

4070
4071
4072
        fused_attn_backend = None
        fused_attn_dqkv_dtype = None
        fused_attn_qkv_dtype = None
4073
        dout_dtype = dout.dtype
4074
4075
4076
4077
4078
4079
4080
        if ctx.fp8:
            if ctx.use_fused_attention:
                fp8_dtype_forward = get_fp8_te_dtype(ctx.fp8_meta["recipe"], fprop_tensor=True)
                fp8_dtype_backward = get_fp8_te_dtype(ctx.fp8_meta["recipe"], fprop_tensor=False)
                fused_attn_qkv_dtype = fp8_dtype_forward
                fused_attn_dqkv_dtype = fp8_dtype_backward
                fused_attn_backend = FusedAttnBackend["FP8"]
4081
                if ctx.is_output_fp8:
4082
4083
4084
4085
4086
4087
4088
4089
4090
4091
4092
4093
4094
4095
4096
4097
4098
4099
4100
4101
4102
4103
4104
4105
4106
                    assert isinstance(dout, Float8Tensor), "dout must be Float8Tensors for FP8 MHA!"
                    ctx.fp8_meta["scaling_bwd"].scale_inv[META_DO] = dout._scale_inv
                    dout_fp8 = dout
                    dout = dout_fp8._data
                else:
                    dout_f16 = dout
                    dout = cast_to_fp8(
                        dout_f16, ctx.fp8_meta["scaling_bwd"], META_DO, fp8_dtype_backward
                    )
                fp8_meta_kwargs = {}
                fp8_meta_kwargs["d_scale_qkv"] = fp8_fwd_scale_invs[META_QKV]
                fp8_meta_kwargs["d_scale_s"] = fp8_fwd_scale_invs[META_S]
                fp8_meta_kwargs["d_scale_o"] = fp8_fwd_scale_invs[META_O]
                fp8_meta_kwargs["d_scale_do"] = ctx.fp8_meta["scaling_bwd"].scale_inv[META_DO]
                fp8_meta_kwargs["d_scale_dp"] = ctx.fp8_meta["scaling_bwd"].scale_inv[META_DP]
                fp8_meta_kwargs["q_scale_s"] = fp8_fwd_scales[META_S]
                fp8_meta_kwargs["q_scale_dp"] = ctx.fp8_meta["scaling_bwd"].scale[META_DP]
                fp8_meta_kwargs["q_scale_dqkv"] = ctx.fp8_meta["scaling_bwd"].scale[META_DQKV]
                fp8_meta_kwargs["amax_dp"] = ctx.fp8_meta["scaling_bwd"].amax_history[0][META_DP]
                fp8_meta_kwargs["amax_dqkv"] = ctx.fp8_meta["scaling_bwd"].amax_history[0][
                    META_DQKV
                ]
            else:
                assert False, "FP8 is only supported with Fused Attention!"
        else:
4107
            if ctx.fp8_meta is not None and ctx.is_output_fp8:
4108
4109
4110
4111
4112
4113
4114
4115
4116
4117
4118
4119
4120
4121
4122
4123
4124
                assert isinstance(dout, Float8Tensor), "dout must be Float8Tensors for FP8 MHA!"
                q, k, v, out, dout = [x.from_float8(x.dtype) for x in [q, k, v, out, dout]]
            if ctx.use_fused_attention:
                fp8_meta_kwargs = {}
                fused_attn_qkv_dtype = TE_DType[q.dtype]
                fused_attn_dqkv_dtype = TE_DType[dout.dtype]
                fused_attn_backend = FusedAttnBackend["F16_arbitrary_seqlen"]

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

        chunk_ids_for_a2a = get_seq_chunk_ids_for_reordering(cp_size, out.device, True)
        out, dout = flash_attn_a2a_communicate(
            [out, dout], chunk_ids_for_a2a, seq_dim, cp_size, ctx.cp_group, ctx.cp_stream, True
        )

4125
        flash_attn_bwd = None
4126
4127
4128
4129
4130
4131
4132
4133
4134
4135
4136
4137
4138
4139
4140
        if not ctx.use_fused_attention:
            fa_backward_kwargs = {"softmax_scale": ctx.softmax_scale}
            if _use_flash_attn_3:
                flash_attn_bwd = flash_attn_varlen_bwd_v3
                fa_backward_kwargs["window_size"] = ctx.window_size
                fa_backward_kwargs["deterministic"] = ctx.deterministic
            else:
                flash_attn_bwd = flash_attn_varlen_bwd
                fa_backward_kwargs["dropout_p"] = ctx.dropout_p
                if _flash_attn_2_3_plus:
                    fa_backward_kwargs["window_size"] = ctx.window_size
                if _flash_attn_2_4_plus:
                    fa_backward_kwargs["alibi_slopes"] = None
                if _flash_attn_2_4_1_plus:
                    fa_backward_kwargs["deterministic"] = ctx.deterministic
4141
4142
4143
4144
4145
4146
4147
4148
4149
4150
4151
4152
4153
4154
4155
4156
4157
4158
4159
4160
4161
4162
4163
4164
4165
4166
4167
4168
4169
4170
4171

        if ctx.use_fused_attention:
            dq, dk, dv, _ = fused_attn_bwd(
                ctx.max_seqlen_q,
                ctx.max_seqlen_kv,
                cu_seqlens_q,
                cu_seqlens_kv,
                q,
                k,
                v,
                out,
                dout,
                fused_attn_qkv_dtype,
                fused_attn_dqkv_dtype,
                aux_ctx_tensors,
                fused_attn_backend,
                cu_seqlens_q_padded=cu_seqlens_q_padded,
                cu_seqlens_kv_padded=cu_seqlens_kv_padded,
                attn_scale=ctx.softmax_scale,
                dropout=ctx.dropout_p,
                qkv_layout=qkv_layout,
                attn_mask_type=ctx.attn_mask_type,
                attn_bias_type=ctx.attn_bias_type,
                window_size=ctx.window_size,
                deterministic=ctx.deterministic,
                **fp8_meta_kwargs,
            )
        else:
            softmax_lse, rng_state = aux_ctx_tensors
            out, dout = [x.view(-1, *x.shape[-2:]) for x in [out, dout]]
            dq, dk, dv = [torch.empty_like(x) for x in [q, k, v]]
4172
4173
4174
            if not _use_flash_attn_3:
                fa_backward_kwargs["rng_state"] = rng_state
            flash_attn_bwd(
4175
4176
4177
4178
4179
4180
4181
4182
4183
4184
4185
4186
4187
                dout,
                q,
                k,
                v,
                out,
                softmax_lse,
                dq,
                dk,
                dv,
                cu_seqlens_q,
                cu_seqlens_kv,
                ctx.max_seqlen_q,
                ctx.max_seqlen_kv,
4188
4189
                causal=causal,
                **fa_backward_kwargs,
4190
4191
4192
4193
4194
4195
4196
4197
            )
            dq, dk, dv = [x.view(ctx.batch_size, -1, *x.shape[-2:]) for x in [dq, dk, dv]]

        chunk_ids_for_a2a = get_seq_chunk_ids_for_reordering(cp_size, q.device, False)
        dq, dk, dv = flash_attn_a2a_communicate(
            [dq, dk, dv], chunk_ids_for_a2a, seq_dim, cp_size, ctx.cp_group, ctx.cp_stream, False
        )

4198
        if ctx.qkv_format == "bshd":
4199
            dq, dk, dv = [x.view(ctx.batch_size, -1, *x.shape[-2:]) for x in [dq, dk, dv]]
4200
        elif ctx.qkv_format == "sbhd":
4201
4202
4203
            dq, dk, dv = [x.view(-1, ctx.batch_size, *x.shape[-2:]) for x in [dq, dk, dv]]

        if ctx.fp8:
4204
            if ctx.is_input_fp8:
4205
4206
4207
4208
4209
4210
4211
                dq, dk, dv = [
                    Float8Tensor(
                        data=x,
                        fp8_meta=ctx.fp8_meta,
                        fp8_meta_forward=False,
                        fp8_meta_index=META_DQKV,
                        fp8_dtype=fp8_dtype_backward,
4212
                        dtype=dout_dtype,
4213
4214
4215
4216
4217
4218
4219
4220
4221
4222
                    )
                    for x in [dq, dk, dv]
                ]
            else:
                dq, dk, dv = [
                    cast_from_fp8(
                        x,
                        ctx.fp8_meta["scaling_bwd"],
                        META_DQKV,
                        fp8_dtype_backward,
4223
                        TE_DType[dout_dtype],
4224
4225
4226
                    )
                    for x in [dq, dk, dv]
                ]
4227
4228
4229
4230
4231
4232
4233
4234
4235
4236
4237
4238
4239
4240
4241
4242
4243
4244
4245
4246
4247
4248
4249

        return (
            None,
            dq,
            dk,
            dv,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
4250
4251
4252
            None,
            None,
            None,
4253
4254
4255
        )


4256
def attn_forward_func_with_cp(
4257
4258
4259
4260
4261
    is_training,
    q,
    k,
    v,
    cu_seqlens_q,
4262
    cu_seqlens_kv,
4263
    max_seqlen_q,
4264
    max_seqlen_kv,
4265
4266
    cu_seqlens_q_padded,
    cu_seqlens_kv_padded,
4267
4268
4269
4270
    dropout_p,
    cp_group,
    cp_global_ranks,
    cp_stream,
4271
    cp_comm_type,
4272
4273
4274
4275
4276
4277
4278
    softmax_scale=None,
    qkv_format="bshd",
    attn_mask_type="causal",
    attn_bias_type="no_bias",
    attn_bias=None,
    deterministic=False,
    use_fused_attention=False,
4279
    window_size=None,
4280
4281
    fp8=False,
    fp8_meta=None,
4282
) -> torch.Tensor:
4283
4284
4285
4286
    """
    Attention implementation with context parallelism.
    """

4287
4288
4289
4290
4291
4292
4293
4294
4295
4296
4297
4298
4299
4300
4301
4302
    if cp_comm_type == "a2a+p2p":
        assert isinstance(
            cp_group, list
        ), "Hierarchical CP implementation needs multi-level CP groups!"
        assert len(cp_group) == 2, "Current implementation only supports two-level CP groups!"
        if get_distributed_world_size(cp_group[0]) == 1:
            cp_group = cp_group[1]
            cp_comm_type = "p2p"
        elif get_distributed_world_size(cp_group[1]) == 1:
            cp_group = cp_group[0]
            cp_comm_type = "a2a"
    else:
        assert isinstance(
            cp_group, dist_group_type
        ), f"Unsupported process group for CP communication type {cp_comm_type}!"

4303
4304
4305
4306
4307
4308
4309
4310
4311
4312
4313
4314
4315
4316
4317
4318
4319
4320
4321
4322
    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!"""
    )
4323
4324
4325
    assert (
        cu_seqlens_q_padded is not None and cu_seqlens_kv_padded is not None
    ), "cu_seqlens_q_padded and cu_seqlens_kv_padded cannot be None with context parallelism!"
4326
4327
4328

    sliding_window_attn = (
        window_size is not None and window_size != (-1, 0) and window_size != (-1, -1)
4329
    )
4330
4331
4332
4333
4334
    assert (
        not sliding_window_attn
        or cp_comm_type == "a2a"
        or (cp_comm_type == "all_gather" and not use_fused_attention)
    ), "The context parallel running configs cannot support sliding window attetnion!"
4335

4336
4337
4338
4339
4340
4341
4342
4343
4344
4345
4346
4347
4348
4349
4350
4351
4352
4353
4354
4355
4356
    args = [
        is_training,
        q,
        k,
        v,
        cu_seqlens_q,
        cu_seqlens_kv,
        max_seqlen_q,
        max_seqlen_kv,
        cu_seqlens_q_padded,
        cu_seqlens_kv_padded,
        dropout_p,
        softmax_scale,
        qkv_format,
        attn_mask_type,
        attn_bias_type,
        attn_bias,
        deterministic,
        use_fused_attention,
    ]

4357
    if cp_comm_type in ["p2p", "a2a+p2p"]:
4358
4359
4360
4361
4362
4363
4364
4365
4366
4367
        args += [fp8, fp8_meta, cp_group, cp_global_ranks, cp_stream]
        out = AttnFuncWithCPAndKVP2P.apply(*args)
    elif cp_comm_type == "all_gather":
        args.pop(5)
        args.pop(8)
        args += [window_size, cp_group, cp_stream]
        out = AttnFuncWithCPAndKVAllGather.apply(*args)
    elif cp_comm_type == "a2a":
        args += [window_size, fp8, fp8_meta, cp_group, cp_stream]
        out = AttnFuncWithCPAndQKVOA2A.apply(*args)
4368
4369
4370
    else:
        raise ValueError(f"Unsupported communication type: {cp_comm_type}!")

4371
4372
4373
    return out


4374
4375
4376
4377
class RotaryPositionEmbedding(torch.nn.Module):
    """
    Implements Rotary Position Embedding from https://arxiv.org/abs/2104.09864.
    """
4378

4379
4380
4381
    def __init__(
        self,
        dim: int,
4382
        rotary_percent: float = 1.0,
4383
4384
        seq_len_interpolation_factor: Optional[int] = None,
        pretrained_max_position_embeddings: Optional[int] = None,
4385
        rotary_base: float = 10000.0,
4386
4387
4388
4389
4390
4391
    ):
        """
        Parameters
        ----------
        dim: int
            rotary embedding dimension
4392
4393
        rotary_percent: float
            Percent of rotary dimension to use for rotary position embeddings.
4394
4395
4396
4397
4398
4399
4400
        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__()
4401
4402
        if rotary_percent < 1.0:
            dim = int(dim * rotary_percent)
4403
        self.seq_len_interpolation_factor = seq_len_interpolation_factor
4404
        self.rotary_base = rotary_base
4405
        inv_freq = 1.0 / (
4406
            self.rotary_base
4407
4408
4409
4410
4411
            ** (
                torch.arange(0, dim, 2, dtype=torch.float32, device=torch.cuda.current_device())
                / dim
            )
        )
4412
        self.register_buffer("inv_freq", inv_freq)
4413
4414
4415
4416
4417
4418
4419
4420
4421
4422
4423
4424
4425
        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
        """
4426
4427
4428
4429
        seq = (
            torch.arange(max_seq_len, device=self.inv_freq.device, dtype=self.inv_freq.dtype)
            + offset
        )
4430

4431
4432
4433
4434
4435
4436
4437
4438
        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
            ):
4439
4440
4441
4442
4443
4444
                # 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

4445
        freqs = torch.einsum("i , j -> i j", seq, self.inv_freq)
4446
4447
4448
4449
4450
4451
        # 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))

4452
4453
4454
4455
4456
4457
4458
4459
4460
4461
4462
4463
4464
4465
4466
4467
4468

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,
4469
4470
        cp_size: int = 1,
        cp_rank: int = 0,
4471
    ) -> torch.Tensor:
4472
        # pylint: disable=missing-function-docstring
4473
4474
        if freqs.dtype != torch.float32:
            freqs = freqs.float()
4475
4476
4477
        if tensor_format == "sbhd":
            output = tex.fused_rope_forward(t, freqs, False)
        elif tensor_format == "bshd":
4478
            output = tex.fused_rope_forward(t.transpose(0, 1), freqs, True).transpose(0, 1)
4479
        elif tensor_format == "thd":
4480
            output = tex.fused_rope_thd_forward(t, cu_seqlens, freqs, cp_size, cp_rank)
4481
4482
4483
4484
        else:
            raise ValueError(f"Unsupported tensor_format: {tensor_format}.")
        ctx.save_for_backward(freqs, cu_seqlens)
        ctx.tensor_format = tensor_format
4485
4486
        ctx.cp_size = cp_size
        ctx.cp_rank = cp_rank
4487
4488
4489
4490

        return output

    @staticmethod
4491
    def backward(ctx, grad_output: torch.Tensor) -> Tuple[Union[torch.Tensor, None], ...]:
4492
        # pylint: disable=missing-function-docstring
4493
4494
4495
4496
4497
4498
4499
4500
        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":
4501
4502
4503
            grad_input = tex.fused_rope_thd_backward(
                grad_output, cu_seqlens, freqs, ctx.cp_size, ctx.cp_rank
            )
4504
4505
4506
        else:
            raise ValueError(f"Unsupported tensor_format: {ctx.tensor_format}.")

4507
        return grad_input, None, None, None, None, None
4508
4509


4510
4511
4512
4513
4514
4515
4516
4517
4518
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)


4519
def apply_rotary_pos_emb(
4520
4521
4522
4523
4524
    t: torch.Tensor,
    freqs: torch.Tensor,
    tensor_format: str = "sbhd",
    fused: bool = False,
    cu_seqlens: Union[torch.Tensor, None] = None,
4525
4526
    cp_size: int = 1,
    cp_rank: int = 0,
4527
) -> torch.Tensor:
4528
    """
4529
    Apply rotary positional embedding tensor to the input tensor.
4530

4531
4532
4533
    Parameters
    ----------
    t: torch.Tensor
4534
        Input tensor of shape `[s, b, h, d]`, `[b, s, h, d]` or `[t, h, d]`, on which
4535
4536
4537
4538
4539
4540
4541
4542
4543
4544
4545
4546
        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'.
4547
4548
4549
4550
4551
        Should be `cu_seqlens_padded` when cp_size > 1.
    cp_size: int, default = 1.
        Context parallel world size. Only valid when `tensor_format` is 'thd' and `fused` is True.
    cp_rank: int, default = 0.
        Context parallel rank. Only valid when `tensor_format` is 'thd' and `fused` is True.
4552
    """
4553
4554
4555
4556
    if fused:
        assert (
            tensor_format != "thd" or cu_seqlens is not None
        ), "cu_seqlens must not be None when tensor_format is 'thd'."
4557
        return FusedRoPEFunc.apply(t, freqs, tensor_format, cu_seqlens, cp_size, cp_rank)
4558
4559
4560
4561
4562
4563

    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}."
    )

4564
4565
4566
4567
4568
    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.
4569
4570
4571
    assert (
        cur_seq_len <= max_seq_len
    ), f"Rotary Embeddings only supported up to {max_seq_len} sequence length!"
4572
    freqs = freqs[:cur_seq_len]
4573
    if tensor_format == "bshd":
4574
4575
4576
4577
        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)
4578

4579
4580
4581
4582
4583
4584
    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
4585
    t = (t * cos_) + (_rotate_half(t) * sin_)
4586
4587
4588
    return torch.cat((t, t_pass), dim=-1)


cyanguwa's avatar
cyanguwa committed
4589
class _SplitAlongDim(torch.autograd.Function):
4590
4591
4592
    """"""

    @staticmethod
4593
4594
4595
4596
4597
    def forward(
        ctx,
        mixed_x_layer: torch.Tensor,
        split_dim: int,
        split_size_or_sections: Union[int, List[int], Tuple[int]],
4598
    ) -> Tuple[torch.Tensor, ...]:
4599
        # pylint: disable=missing-function-docstring
cyanguwa's avatar
cyanguwa committed
4600
4601
        ctx.split_dim = split_dim
        ctx.split_size_or_sections = split_size_or_sections
4602
        if isinstance(mixed_x_layer, Float8Tensor):
4603
4604
4605
4606
4607
4608
            return tuple(
                Float8Tensor.make_like(
                    mixed_x_layer,
                    data=x,
                )
                for x in torch.split(
4609
4610
                    mixed_x_layer._data,
                    split_size_or_sections=split_size_or_sections,
4611
4612
4613
4614
                    dim=split_dim,
                )
            )
        return torch.split(mixed_x_layer, split_size_or_sections, dim=split_dim)
4615
4616

    @staticmethod
4617
    def backward(ctx, *grad_outputs):
4618
        # pylint: disable=missing-function-docstring
4619
4620
        assert len(grad_outputs) > 0, "No gradients received for backprop!"

cyanguwa's avatar
cyanguwa committed
4621
4622
        if isinstance(ctx.split_size_or_sections, (list, tuple)):
            split_sizes = ctx.split_size_or_sections
4623
4624
4625
            assert len(grad_outputs) == len(
                split_sizes
            ), "Unequal number of gradients vs split sections for backprop!"
cyanguwa's avatar
cyanguwa committed
4626
4627
4628
4629
4630
        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

4631
4632
4633
4634
4635
4636
4637
4638
        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]
4639
4640
4641
4642
4643
4644
4645
                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
                ):
4646
4647
4648
                    noop_ok = False
                    break
            if noop_ok:
4649
4650
4651
                ret = torch.Tensor().to(
                    device=grad_outputs[0].device, dtype=grad_outputs[0]._data.dtype
                )
4652
4653
                new_shape = list(shape)
                new_shape[split_dim] = sum(split_sizes)
4654
4655
4656
4657
4658
                ret.set_(
                    grad_outputs[0]._data.untyped_storage(),
                    grad_outputs[0]._data.storage_offset(),
                    new_shape,
                    strides,
4659
4660
4661
4662
                )
                return Float8Tensor.make_like(grad_outputs[0], data=ret), None, None

            grad_outputs_data = [x._data for x in grad_outputs]
4663
4664
4665
4666
4667
4668
4669
            return (
                Float8Tensor.make_like(
                    grad_outputs[0], data=torch.cat(grad_outputs_data, dim=split_dim)
                ),
                None,
                None,
            )
4670
4671
        noop_ok = True
        strides = grad_outputs[0].stride()
4672
        data_ptr = grad_outputs[0].untyped_storage().data_ptr()
cyanguwa's avatar
cyanguwa committed
4673
        shape = list(grad_outputs[0].shape)
4674
        for i, tensor in enumerate(grad_outputs):
cyanguwa's avatar
cyanguwa committed
4675
4676
            shape_i = shape
            shape_i[split_dim] = split_sizes[i]
4677
4678
4679
4680
4681
4682
4683
            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
            ):
4684
4685
4686
                noop_ok = False
                break
        if noop_ok:
4687
            ret = torch.Tensor().to(device=grad_outputs[0].device, dtype=grad_outputs[0].dtype)
4688
            new_shape = list(shape)
cyanguwa's avatar
cyanguwa committed
4689
            new_shape[split_dim] = sum(split_sizes)
4690
4691
4692
4693
4694
            ret.set_(
                grad_outputs[0].untyped_storage(),
                grad_outputs[0].storage_offset(),
                new_shape,
                strides,
4695
            )
cyanguwa's avatar
cyanguwa committed
4696
            return ret, None, None
4697

4698
        return torch.cat(grad_outputs, dim=split_dim), None, None
4699
4700
4701
4702
4703
4704
4705
4706
4707


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

    def __init__(
        self,
4708
        softmax_scale: float,
4709
        attention_type: str = "self",
4710
4711
4712
4713
4714
4715
        attention_dropout: float = 0.0,
        attention_dropout_ctx: Optional[Callable] = nullcontext,
        layer_number: Optional[int] = None,
    ) -> None:
        super().__init__()

4716
        self.softmax_scale = softmax_scale
4717
        self.attention_type = attention_type
4718
4719
4720
        self.attention_dropout_ctx = attention_dropout_ctx
        self.layer_number = layer_number

4721
        self.scale_mask_softmax = FusedScaleMaskSoftmax(attention_mask_func)
4722
4723
4724
4725
4726
4727

        # 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)

4728
4729
        # An FP16 training trick required for certain GPT-like models.
        self.apply_qk_layer_scaling = (
4730
4731
            bool(int(os.getenv("NVTE_APPLY_QK_LAYER_SCALING", "0"))) and layer_number is not None
        )
4732

4733
4734
4735
4736
4737
    def forward(
        self,
        query_layer: torch.Tensor,
        key_layer: torch.Tensor,
        value_layer: torch.Tensor,
4738
        qkv_layout: str = "sbh3d",
4739
4740
        cu_seqlens_q: Optional[torch.Tensor] = None,  # pylint: disable=unused-argument
        cu_seqlens_kv: Optional[torch.Tensor] = None,  # pylint: disable=unused-argument
4741
        attn_mask_type: str = "causal",
4742
        attention_mask: Optional[Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]] = None,
4743
4744
        core_attention_bias_type: str = "no_bias",
        core_attention_bias: Optional[torch.Tensor] = None,
4745
        alibi_slopes: Optional[torch.Tensor] = None,
4746
    ) -> torch.Tensor:
4747
        """Unfused attention fprop"""
4748
4749
4750
4751
4752
        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":
4753
            # convert to sbhd and use sbhd implementation for now
4754
4755
4756
            query_layer, key_layer, value_layer = [
                x.transpose(0, 1) for x in [query_layer, key_layer, value_layer]
            ]
4757
4758
4759
4760
4761
4762
4763
4764
4765
4766
4767
4768
4769
4770
4771
4772
4773
4774
4775
4776
4777
4778
4779
4780
4781
4782
4783
4784
4785
4786
4787
4788
4789
4790
4791
4792
4793
4794
4795
4796
4797
4798
4799
4800
4801
4802
4803
4804
4805
4806
4807
4808
        batch_size, max_seqlen_q, max_seqlen_kv = (
            query_layer.shape[1],
            query_layer.shape[0],
            key_layer.shape[0],
        )
        if "padding" in attn_mask_type:
            if self.attention_type == "self":
                assert attention_mask.shape == (
                    batch_size,
                    1,
                    1,
                    max_seqlen_q,
                ), "attention_mask should be a single tensor with [b, 1, 1, sq] shape!"
                attention_mask = torch.logical_or(
                    attention_mask.squeeze(1).unsqueeze(3), attention_mask
                )
            else:
                assert (
                    len(attention_mask) == 2
                    and attention_mask[0].shape == (batch_size, 1, 1, max_seqlen_q)
                    and attention_mask[1].shape == (batch_size, 1, 1, max_seqlen_kv)
                ), (
                    "attention_mask should be a tuple of two tensors with shapes "
                    "[b, 1, 1, sq] and [b, 1, 1, skv]!"
                )
                attention_mask = torch.logical_or(
                    attention_mask[0].squeeze(1).unsqueeze(3), attention_mask[1]
                )
            mask = attention_mask.squeeze(1).logical_not()
            actual_seqlens_q = mask[:, :, 0].sum(dim=1)
            actual_seqlens_kv = mask[:, 0, :].sum(dim=1)
            mask = torch.arange(max_seqlen_q, dtype=torch.int32, device="cuda").view(
                1, 1, max_seqlen_q, 1
            ) - torch.arange(max_seqlen_kv, dtype=torch.int32, device="cuda").view(
                1, 1, 1, max_seqlen_kv
            )
            if attn_mask_type == "padding_causal":
                attention_mask = torch.logical_or(
                    torch.where(mask.view(1, 1, max_seqlen_q, max_seqlen_kv) < 0, 1, 0),
                    attention_mask,
                )
            if attn_mask_type == "padding_causal_bottom_right":
                attention_mask = torch.logical_or(
                    torch.where(
                        mask.expand(batch_size, 1, max_seqlen_q, max_seqlen_kv)
                        + (actual_seqlens_kv - actual_seqlens_q).view(batch_size, 1, 1, 1)
                        < 0,
                        1,
                        0,
                    ),
                    attention_mask,
                )
4809

4810
        batch_size, seqlen = query_layer.shape[1], query_layer.shape[0]
4811
        apply_qk_layer_scaling = self.apply_qk_layer_scaling and key_layer.dtype == torch.float16
4812
4813
4814
4815
4816
4817
4818
4819
4820

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

4821
        if key_layer.shape[2] != query_layer.shape[2]:
4822
4823
4824
            assert (
                query_layer.shape[2] % key_layer.shape[2] == 0
            ), "The number of attention heads must be divisible by the number of GQA groups!"
4825
            key_layer = key_layer.repeat_interleave(
4826
4827
                int(query_layer.shape[2] / key_layer.shape[2]), dim=2
            )
4828
            value_layer = value_layer.repeat_interleave(
4829
4830
                int(query_layer.shape[2] / value_layer.shape[2]), dim=2
            )
4831

4832
        # [sq, b, np, hn] -> [sq, b * np, hn]
4833
        query_layer = query_layer.reshape(output_size[2], output_size[0] * output_size[1], -1)
4834
4835
4836
4837
        # [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]
4838
4839
        # WAR to set dtype to FP32 as ONNX lacks BF16 support for ConstantOfShape operator
        is_bf16 = query_layer.dtype == torch.bfloat16
4840
4841
4842
4843
        matmul_result = torch.empty(
            output_size[0] * output_size[1],
            output_size[2],
            output_size[3],
4844
            dtype=torch.float32 if is_in_onnx_export_mode() and is_bf16 else query_layer.dtype,
4845
4846
4847
            device=torch.cuda.current_device(),
        )

4848
4849
4850
        if is_in_onnx_export_mode() and is_bf16:
            matmul_result = matmul_result.bfloat16()

4851
        scale = self.softmax_scale
4852
        if apply_qk_layer_scaling:
4853
            scale /= self.layer_number
4854
4855

        # Raw attention scores. [b * np, sq, sk]
4856
4857
4858
4859
4860
4861
        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,
4862
                alpha=scale,
4863
            ).view(*output_size)
4864
4865
4866
4867
4868
4869
4870

        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]
            )
4871
            matmul_result = matmul_result.view(*output_size) + core_attention_bias
4872
            matmul_result *= scale
4873

4874
4875
4876
4877
        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":
4878
                _, core_attention_bias = get_alibi(
4879
4880
4881
                    output_size[1],
                    output_size[2],
                    output_size[3],
4882
4883
                    actual_seqlens_q=actual_seqlens_q if "padding" in attn_mask_type else None,
                    actual_seqlens_kv=actual_seqlens_kv if "padding" in attn_mask_type else None,
4884
4885
                    alibi_slopes=alibi_slopes,
                    bottom_right_alignment=attn_mask_type not in ["causal", "padding_causal"],
4886
                )
4887
4888
4889
4890
4891
            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,
4892
                alpha=scale,
4893
            )
4894
4895
            matmul_result = (matmul_result.view(*output_size) + core_attention_bias).to(
                dtype=query_layer.dtype
4896
            )
4897
4898
4899

        # attention scores and attention mask [b, np, sq, sk]
        softmax_scale = self.layer_number if apply_qk_layer_scaling else None
4900
        attention_probs = self.scale_mask_softmax(
4901
            matmul_result, attention_mask, attn_mask_type, softmax_scale
4902
        )
4903

4904
4905
4906
4907
4908
        # mask out the pad positions in softmax results, mostly for the rows (pad tokens from q)
        # the columns (pad tokens from k) are already zeroed out during softmax
        if "padding" in attn_mask_type:
            attention_probs = attention_probs.masked_fill(attention_mask, 0)

4909
4910
4911
4912
4913
4914
4915
4916
4917
4918
4919
4920
4921
4922
4923
        # 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]
4924
        value_layer = value_layer.reshape(value_layer.size(0), output_size[0] * output_size[1], -1)
4925
4926

        # change view [b * np, sq, sk]
4927
        attention_probs = attention_probs.view(output_size[0] * output_size[1], output_size[2], -1)
4928
4929
4930
4931
4932
4933
4934

        # 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)

4935
        if qkv_format == "sbhd":
4936
4937
            # [b, np, sq, hn] --> [sq, b, np, hn]
            context_layer = context_layer.permute(2, 0, 1, 3).contiguous()
4938

4939
4940
4941
            # [sq, b, np, hn] --> [sq, b, hp]
            context_layer = context_layer.view(seqlen, batch_size, -1)

4942
        if qkv_format == "bshd":
4943
4944
4945
4946
4947
            # [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)
4948
4949
4950
4951
4952
4953

        return context_layer


class _PrepareQKVForFA(torch.autograd.Function):
    """This class converts QKV from interleaved (s, b, ...) layout
4954
    to separate contiguous q, k, v tensors in (b, s, ...) layout."""
4955
4956

    @staticmethod
4957
4958
4959
4960
    def forward(
        _ctx: torch.autograd.function.FunctionCtx,  # unused
        query_layer: torch.Tensor,
        key_layer: torch.Tensor,
4961
        value_layer: torch.Tensor,
4962
    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
4963
        # pylint: disable=missing-function-docstring
4964
4965
4966
4967
4968
4969
4970
4971
4972
4973
4974
        # 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
4975
4976
4977
4978
    def backward(
        _ctx: torch.autograd.function.FunctionCtx,  # unused
        dq: torch.Tensor,
        dk: torch.Tensor,
4979
        dv: torch.Tensor,
4980
    ) -> Tuple[Union[torch.Tensor, None], ...]:
4981
        # pylint: disable=missing-function-docstring
4982
4983
4984
4985
        dqkv = tex.fa_prepare_bwd(dq, dk, dv)
        dq, dk, dv = split_tensor_along_dim(dqkv, -1, 3)
        return dq, dk, dv

4986

4987
def get_qkv_layout(
4988
4989
4990
4991
4992
    q: torch.Tensor,
    k: torch.Tensor,
    v: torch.Tensor,
    qkv_format: str = "sbhd",
) -> str:
4993
    """Get qkv layout.
4994

4995
4996
4997
4998
4999
5000
5001
5002
5003
5004
5005
    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,
5006
        `d` head size, and `t` the total number of tokens in a batch, i.e.
5007
5008
5009
5010
5011
5012
5013
5014
5015
5016
5017
5018
5019
5020
5021
        `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`}
5022
5023
5024
5025
5026
5027
5028
5029
5030
    q: torch.Tensor
        Query tensor. It may be different from input `q` as we try to fit tensors to
        a supported layout.
    k: torch.Tensor
        Key tensor. It may be different from input `k` as we try to fit tensors to
        a supported layout.
    v: torch.Tensor
        Value tensor. It may be different from input `v` as we try to fit tensors to
        a supported layout.
5031
    """
5032

5033
5034
    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!"
5035

5036
    def run_iteratively(q, k, v):
5037
        # check data pointers
5038
5039
        data_ptr = q.untyped_storage().data_ptr()
        check_ptrs_qkv = all(x.untyped_storage().data_ptr() == data_ptr for x in [q, k, v])
5040
        check_ptrs_qk = all(x.untyped_storage().data_ptr() == data_ptr for x in [q, k])
5041
5042
5043
        data_ptr = k.untyped_storage().data_ptr()
        check_ptrs_kv = all(x.untyped_storage().data_ptr() == data_ptr for x in [k, v])

5044
5045
5046
5047
5048
5049
5050
        # check tensor shapes
        shape = q.shape
        check_shapes_qkv = all(shape == x.shape for x in [q, k, v])
        shape = k.shape
        check_shapes_kv = shape[:-1] == v.shape[:-1]

        # check tensor strides
5051
5052
        stride = q.stride()
        check_strides_qkv = all(stride == x.stride() for x in [q, k, v])
5053
5054
        check_strides_kv = tuple(sk / k.shape[-1] for sk in k.stride()[:-1]) == tuple(
            sv / v.shape[-1] for sv in v.stride()[:-1]
5055
        )
5056

5057
5058
5059
5060
5061
5062
        # check tensor offsets for h3d and 3hd layouts
        prod_h_d = q.shape[-1] * q.shape[-2]
        check_3hd_offsets = all(x.storage_offset() == i * prod_h_d for i, x in enumerate([q, k, v]))
        check_h3d_offsets = all(
            x.storage_offset() == i * q.shape[-1] for i, x in enumerate([q, k, v])
        )
5063

5064
5065
5066
5067
5068
5069
        # check tensor offsets for hd_h2d and hd_2hd layouts
        prod_all_dims = [np.prod(x.shape) for x in [q, k]]
        offset = prod_all_dims[0] if check_ptrs_qkv else 0
        prod_h_d = k.shape[-1] * k.shape[-2]
        check_2hd_offsets = all(
            x.storage_offset() == (offset + i * prod_h_d) for i, x in enumerate([k, v])
5070
        )
5071
5072
        check_h2d_offsets = all(
            x.storage_offset() == (offset + i * k.shape[-1]) for i, x in enumerate([k, v])
5073
        )
5074

5075
5076
5077
5078
5079
5080
5081
5082
5083
5084
        # check tensor offsets for hd_hd_hd layouts
        check_hd_offsets_qkv = (
            all(x.storage_offset() == sum(prod_all_dims[:i]) for i, x in enumerate([q, k, v]))
            if check_ptrs_qkv
            else all(x.storage_offset() == 0 for i, x in enumerate([q, k, v]))
        )
        check_hd_offsets_qk = (
            all(x.storage_offset() == sum(prod_all_dims[:i]) for i, x in enumerate([q, k]))
            if not check_ptrs_qkv and check_ptrs_qk
            else all(x.storage_offset() == 0 for i, x in enumerate([q, k]))
5085
        )
5086
5087
5088
5089
        check_hd_offsets_kv = (
            all(x.storage_offset() == sum(prod_all_dims[1 : i + 1]) for i, x in enumerate([k, v]))
            if not check_ptrs_qkv and check_ptrs_kv
            else all(x.storage_offset() == 0 for i, x in enumerate([k, v]))
5090
        )
5091

5092
        if check_ptrs_qkv and check_strides_qkv and check_shapes_qkv and check_3hd_offsets:
5093
            # sb3hd, bs3hd, t3hd
5094
            # one chunk of memory, qkv, with q, k, v interleaved at dim=-3 in qkv
5095
            qkv_layout = qkv_format[:-2] + "3" + qkv_format[-2:]
5096
        elif check_ptrs_qkv and check_strides_qkv and check_shapes_qkv and check_h3d_offsets:
5097
            # sbh3d, bsh3d, th3d
5098
            # one chunk of memory, qkv, with q, k, v interleaved at dim=-2 in qkv
5099
            qkv_layout = qkv_format[:-1] + "3" + qkv_format[-1:]
5100
        elif check_ptrs_kv and check_strides_kv and check_shapes_kv and check_2hd_offsets:
5101
            # sbhd_sb2hd, bshd_bs2hd, thd_t2hd
5102
5103
5104
            # two chunks of memory, q and kv, with k, v interleaved at dim=-3 in kv
            # q and kv may be disjoint or consecutive in memory, and when consecutive, they may
            # have the same data pointer, i.e. check_ptrs_qkv=True
5105
            qkv_layout = qkv_format + "_" + qkv_format[:-2] + "2" + qkv_format[-2:]
5106
        elif check_ptrs_kv and check_strides_kv and check_shapes_kv and check_h2d_offsets:
5107
            # sbhd_sbh2d, bshd_bsh2d, thd_th2d
5108
5109
5110
            # two chunks of memory, q and kv, with k, v interleaved at dim=-2 in kv
            # q and kv may be disjoint or consecutive in memory, and when consecutive, they may
            # have the same data pointer, i.e. check_ptrs_qkv=True
5111
            qkv_layout = qkv_format + "_" + qkv_format[:-1] + "2" + qkv_format[-1:]
5112
5113
5114
5115
5116
        elif (
            check_strides_kv
            and check_shapes_kv
            and (check_hd_offsets_qkv or check_hd_offsets_kv or check_hd_offsets_qk)
        ):
5117
            # sbhd_sbhd_sbhd, bshd_bshd_bshd, thd_thd_thd
5118
5119
5120
            # three chunks of memory, q, k and v, which may be disjoint or consecutive, and
            # when consecutive, they may have the same data pointer, i.e. check_ptrs_qkv=True or
            # check_ptrs_qk=True or check_ptrs_kv=True
5121
            qkv_layout = "_".join(list([qkv_format]) * 3)
5122
        else:
5123
            qkv_layout = "not_supported"
5124
5125
5126
5127

        return qkv_layout

    qkv_layout = run_iteratively(q, k, v)
5128
    if qkv_layout == "not_supported":
5129
5130
5131
        # 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)
5132
    if qkv_layout == "not_supported":
5133
        raise RuntimeError("The provided qkv memory layout is not supported!")
5134

5135
    return qkv_layout, q, k, v
5136

5137

5138
def check_set_window_size(
5139
5140
5141
    attn_mask_type: str,
    window_size: Tuple[int, int] = None,
):
5142
5143
5144
5145
5146
5147
5148
5149
    """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)
5150
    """
5151
    orig_window_size = window_size
5152
    if "causal" in attn_mask_type:
5153
        if orig_window_size is None:
5154
            window_size = (-1, 0)
5155
5156
5157
        elif orig_window_size == (-1, -1) or (
            orig_window_size[0] >= 0 and orig_window_size[1] != 0
        ):
5158
5159
5160
5161
            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
            )
5162
        elif orig_window_size != (-1, 0) and (orig_window_size[0] < 0 or orig_window_size[1] != 0):
5163
5164
5165
5166
            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"]:
5167
5168
5169
        if orig_window_size is None:
            window_size = (-1, -1)
        elif orig_window_size == (-1, 0):
5170
            window_size = (-1, -1)
5171
5172
5173
            warnings.warn(
                "window_size should be (-1, -1) or (>=0, >=0) for attn_mask_type=" + attn_mask_type
            )
5174
        elif orig_window_size != (-1, -1) and (orig_window_size[0] < 0 or orig_window_size[1] < 0):
5175
5176
5177
5178
5179
            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
5180
    return window_size
5181

5182

5183
class FlashAttention(torch.nn.Module):
5184
    """Dot product attention, using HazyResearch flash-attn package:
5185
    https://github.com/Dao-AILab/flash-attention
5186
5187
5188
5189
    """

    def __init__(
        self,
5190
        softmax_scale: float,
5191
5192
        attention_dropout: float = 0.0,
        attention_dropout_ctx: Optional[Callable] = nullcontext,
5193
5194
        attention_type: str = "self",
        layer_number: Optional[int] = None,
5195
        deterministic: bool = False,
5196
5197
5198
    ) -> None:
        super().__init__()

5199
5200
5201
5202
5203
5204
5205
        if _flash_attn_is_installed:
            assert (
                _flash_attn_version >= _flash_attn_version_required
            ), f"FlashAttention minimum version {_flash_attn_version_required} is required."
            assert (
                _flash_attn_version <= _flash_attn_max_version
            ), f"FlashAttention maximum version {_flash_attn_max_version} is supported."
5206

5207
        self.softmax_scale = softmax_scale
5208
5209
        self.attention_dropout_ctx = attention_dropout_ctx
        self.attention_dropout = attention_dropout
5210
5211
        self.attention_type = attention_type
        self.layer_number = 1 if layer_number is None else layer_number
5212
        self.deterministic = deterministic
5213
5214
5215
5216
        self.logger = logging.getLogger("FlashAttention")
        self.logger.setLevel(_log_level)
        if not self.logger.hasHandlers():
            self.logger.addHandler(_stream_handler)
5217
5218
5219
5220
5221
5222

    def forward(
        self,
        query_layer: torch.Tensor,
        key_layer: torch.Tensor,
        value_layer: torch.Tensor,
5223
        attention_mask: Optional[Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]] = None,
5224
5225
5226
        qkv_layout: str = "sbh3d",
        cu_seqlens_q: Optional[torch.Tensor] = None,
        cu_seqlens_kv: Optional[torch.Tensor] = None,
5227
5228
        max_seqlen_q: Optional[int] = None,
        max_seqlen_kv: Optional[int] = None,
5229
        attn_mask_type: str = "causal",
5230
        window_size: Optional[Tuple[int, int]] = None,
5231
        alibi_slopes: Optional[torch.Tensor] = None,
5232
        cp_group: Optional[Union[dist_group_type, List[dist_group_type]]] = None,
5233
        cp_global_ranks: List[int] = None,
5234
        cp_stream: torch.cuda.Stream = None,
5235
        cp_comm_type: str = "p2p",
5236
5237
        fp8: bool = False,
        fp8_meta: Optional[Dict[str, Any]] = None,
5238
5239
5240
    ) -> torch.Tensor:
        """flash-attn fprop"""

5241
5242
5243
5244
        assert all(
            x.dtype in [torch.float16, torch.bfloat16] or isinstance(x, Float8Tensor)
            for x in [query_layer, key_layer, value_layer]
        ), "FlashAttention only supports FP16 and BF16 data types, or Float8Tensors."
5245
5246
        assert (
            query_layer.is_cuda and key_layer.is_cuda and value_layer.is_cuda
5247
        ), "FlashAttention currently only supports CUDA tensors."
5248
5249
        assert (
            qkv_layout in QKVLayouts
5250
        ), f"FlashAttention does not support qkv_layout = {qkv_layout}!"
5251

5252
5253
5254
5255
5256
5257
        cp_size = 1
        if isinstance(cp_group, dist_group_type):
            cp_size = get_distributed_world_size(cp_group)
        elif isinstance(cp_group, list):
            for group in cp_group:
                cp_size *= get_distributed_world_size(group)
5258
        context_parallel = cp_size > 1
5259

5260
        qkv_format = "".join([i for i in qkv_layout.split("_")[0] if i.isalpha()])
5261

5262
5263
5264
5265
5266
5267
5268
5269
5270
5271
5272
5273
5274
        if all(not isinstance(x, Float8Tensor) for x in [query_layer, key_layer, value_layer]):
            if qkv_format == "sbhd":
                # For now just 128, will make it more general in the future
                if (
                    query_layer.shape[-1] == 128
                    and query_layer.shape[0] * query_layer.shape[1] >= 512
                    and qkv_layout == "sbh3d"
                ):
                    query_layer, key_layer, value_layer = _PrepareQKVForFA.apply(
                        query_layer, key_layer, value_layer
                    )
                else:
                    query_layer, key_layer, value_layer = [
5275
                        x.transpose(0, 1) for x in (query_layer, key_layer, value_layer)
5276
                    ]
5277
            if context_parallel:
5278
                query_layer, key_layer, value_layer = [
5279
5280
5281
5282
5283
                    x.contiguous() for x in (query_layer, key_layer, value_layer)
                ]
        else:
            if qkv_format == "sbhd":
                query_layer._data, key_layer._data, value_layer._data = [
5284
                    x.transpose(0, 1)
5285
5286
                    for x in (query_layer._data, key_layer._data, value_layer._data)
                ]
5287
5288
5289
5290
                query_layer, key_layer, value_layer = [
                    Float8Tensor.make_like(x, data=x._data)
                    for x in (query_layer, key_layer, value_layer)
                ]
5291
            if context_parallel:
5292
5293
                query_layer._data, key_layer._data, value_layer._data = [
                    x.contiguous() for x in (query_layer._data, key_layer._data, value_layer._data)
5294
                ]
5295

5296
        batch_size = query_layer.shape[0]
5297

5298
        if qkv_format in ["sbhd", "bshd"]:
5299
            max_seqlen_q, max_seqlen_kv = query_layer.shape[1], key_layer.shape[1]
5300
5301
            max_seqlen_q *= cp_size
            max_seqlen_kv *= cp_size
5302
5303
5304

            if "padding" in attn_mask_type:
                assert not context_parallel, "Padding mask not supported with context parallelism!"
5305
5306
                # [b * s, h, d]
                query_layer, key_layer, value_layer = [
5307
                    x.reshape(x.shape[0] * x.shape[1], *x.shape[2:])
5308
5309
5310
5311
5312
5313
5314
                    for x in [query_layer, key_layer, value_layer]
                ]

                if self.attention_type == "self":
                    assert (
                        max_seqlen_q == max_seqlen_kv
                    ), "Maximum sequence length for Q and KV should be the same."
5315
                    if cu_seqlens_q is None:
5316
5317
5318
                        assert (
                            attention_mask is not None
                        ), "Please provide attention_mask for padding!"
5319
5320
5321
5322
5323
5324
                        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
5325
5326
                    )
                else:
5327
                    if cu_seqlens_q is None or cu_seqlens_kv is None:
5328
5329
5330
5331
5332
                        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])
5333
5334
5335
5336
                    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)
5337
                    key_layer, value_layer = PackTensors.apply(indices_kv, key_layer, value_layer)
5338
            else:
5339
5340
5341
5342
5343
5344
5345
5346
5347
5348
5349
5350
5351
                # 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,
                    )
5352
5353
5354
5355
        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!"
5356
5357
5358
5359
5360
5361
            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()
5362

5363
5364
5365
        if context_parallel and all(
            not isinstance(x, Float8Tensor) for x in [query_layer, key_layer, value_layer]
        ):
5366
5367
5368
            assert (
                alibi_slopes is None
            ), "Alibi slope bias addition is not supported with context parallelism."
5369
            with self.attention_dropout_ctx():
5370
                output = attn_forward_func_with_cp(
5371
5372
5373
5374
5375
5376
5377
5378
                    self.training,
                    query_layer,
                    key_layer,
                    value_layer,
                    cu_seqlens_q,
                    cu_seqlens_kv,
                    max_seqlen_q,
                    max_seqlen_kv,
5379
5380
                    cu_seqlens_q,
                    cu_seqlens_kv,
5381
                    self.attention_dropout if self.training else 0.0,
5382
5383
5384
                    cp_group,
                    cp_global_ranks,
                    cp_stream,
5385
                    cp_comm_type,
5386
                    softmax_scale=self.softmax_scale,
5387
                    qkv_format="bshd" if qkv_format == "sbhd" else qkv_format,
5388
                    attn_mask_type=attn_mask_type,
5389
                    deterministic=self.deterministic,
5390
                    window_size=window_size,
5391
5392
                )
        else:
5393
5394

            from .cpu_offload import CPUOffloadEnabled
5395

5396
5397
5398
5399
5400
5401
            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

5402
            with self.attention_dropout_ctx():
5403
                fa_optional_forward_kwargs = {}
5404
5405
                if _flash_attn_2_3_plus:
                    fa_optional_forward_kwargs["window_size"] = window_size
5406
5407
5408
5409
                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
5410
5411
5412
5413
                fa_optional_forward_args_thd = []
                if qkv_format in ["bshd", "sbhd"] and "padding" not in attn_mask_type:
                    func = flash_attn_func if not _use_flash_attn_3 else flash_attn_func_v3
                else:
5414
5415
                    if _flash_attn_2_5_7_plus:
                        fa_optional_forward_kwargs["block_table"] = None
5416
5417
5418
5419
5420
5421
5422
5423
5424
5425
                    func = (
                        flash_attn_varlen_func
                        if not _use_flash_attn_3
                        else flash_attn_varlen_func_v3
                    )
                    fa_optional_forward_args_thd.append(cu_seqlens_q)
                    fa_optional_forward_args_thd.append(cu_seqlens_kv)
                    fa_optional_forward_args_thd.append(max_seqlen_q)
                    fa_optional_forward_args_thd.append(max_seqlen_kv)
                if _use_flash_attn_3:
5426
5427
5428
                    fa_3_optional_forward_kwargs = {}
                    fa_3_optional_forward_kwargs["window_size"] = window_size
                    fa_3_optional_forward_kwargs["deterministic"] = self.deterministic
5429
                    activation_dtype = query_layer.dtype
5430
5431
5432
                    if fp8:
                        fp8_dtype_forward = get_fp8_te_dtype(fp8_meta["recipe"], fprop_tensor=True)
                        torch_dtype = get_fp8_torch_dtype(fp8_meta["recipe"], fprop_tensor=True)
5433
5434
5435
5436
5437
5438
5439
5440
5441
5442
5443

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

5444
5445
5446
5447
5448
5449
                        # "fp8_mha" decides outputs in fp8, while inputs are inferred from
                        # the real dtype
                        assert isinstance(key_layer, query_layer.__class__) and isinstance(
                            value_layer, query_layer.__class__
                        ), "q, k, and v must have the same type."
                        if isinstance(query_layer, Float8Tensor):
5450
5451
5452
                            fp8_meta["scaling_fwd"].scale_inv[META_QKV] = query_layer._scale_inv
                        else:
                            query_layer, key_layer, value_layer = (
5453
5454
                                Float8Tensor.to_float8(x, fp8_dtype=fp8_dtype_forward)
                                for x in [query_layer, key_layer, value_layer]
5455
                            )
5456
5457
5458
5459
5460
5461
5462
5463
5464
5465
5466
5467
5468
5469
5470
5471
5472
5473
5474
5475
5476
                        fa_3_optional_forward_kwargs["descale_q"] = query_layer._scale_inv
                        fa_3_optional_forward_kwargs["descale_k"] = key_layer._scale_inv
                        fa_3_optional_forward_kwargs["descale_v"] = value_layer._scale_inv
                        query_layer, key_layer, value_layer = (
                            convert_to_torch_float8(x, torch_dtype)
                            for x in [query_layer, key_layer, value_layer]
                        )
                    try:
                        output, _ = func(
                            query_layer,
                            key_layer,
                            value_layer,
                            *fa_optional_forward_args_thd,
                            softmax_scale=self.softmax_scale,
                            causal="causal" in attn_mask_type,
                            **fa_3_optional_forward_kwargs,
                        )
                    except TypeError as e:
                        if _flash_attn_3_0_0_beta:
                            e.args = (
                                e.args[0]
5477
                                + ". Please update your flash-attn v3 (beta) installation as it "
5478
5479
5480
5481
5482
                                + "may have added more supported arguments to its API. \n"
                                + _flash_attn_3_installation_steps,
                            ) + e.args[1:]
                        raise

5483
5484
5485
5486
5487
5488
5489
5490
5491
5492
5493
5494
5495
5496
5497
5498
5499
5500
5501
5502
5503
5504
5505
5506
5507
5508
                    if fp8 and fp8_meta["recipe"].fp8_mha:
                        output = cast_to_fp8(
                            output,
                            fp8_meta["scaling_fwd"],
                            META_O,
                            fp8_dtype_forward,
                        )
                        output = Float8Tensor(
                            data=output,
                            fp8_meta=fp8_meta,
                            fp8_meta_forward=True,
                            fp8_meta_index=META_O,
                            fp8_dtype=fp8_dtype_forward,
                            dtype=activation_dtype,
                        )
                else:
                    output = func(
                        query_layer,
                        key_layer,
                        value_layer,
                        *fa_optional_forward_args_thd,
                        self.attention_dropout if self.training else 0.0,
                        softmax_scale=self.softmax_scale,
                        causal="causal" in attn_mask_type,
                        **fa_optional_forward_kwargs,
                    )
5509

5510
        if qkv_format in ["sbhd", "bshd"] and "padding" in attn_mask_type:
5511
            output = UnpackTensor.apply(indices_q, batch_size * max_seqlen_q, output)
5512

5513
        if qkv_format == "sbhd":
5514
            # (bs)hd -> bs(hd) -> sb(hd)
5515
            if fp8 and fp8_meta["recipe"].fp8_mha:
5516
5517
5518
5519
5520
5521
                output = Float8Tensor.make_like(
                    output,
                    data=output._data.reshape(batch_size, max_seqlen_q // cp_size, -1)
                    .transpose(0, 1)
                    .contiguous(),
                )
5522
            else:
5523
                output = output.view(batch_size, max_seqlen_q // cp_size, -1).transpose(0, 1)
5524
        elif qkv_format == "bshd":
5525
            # (bs)hd -> bs(hd)
5526
            output = output.reshape(batch_size, max_seqlen_q // cp_size, -1)
5527
        elif qkv_format == "thd":
5528
            # thd -> t(hd)
5529
            output = output.reshape(output.shape[0], -1)
5530

5531
        return output.contiguous()
5532

5533

5534
def _combine_tensors(
5535
5536
5537
    tensors: List[torch.Tensor],
    dim: int,
) -> torch.Tensor:
5538
5539
5540
5541
5542
5543
    """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())
5544
    new_stride.insert(dim, int(new_stride[dim - 1] / num_tensors))
5545
    if isinstance(tensors[0], Float8Tensor):
5546
        combined_tensor = torch.Tensor().to(device=tensors[0].device, dtype=tensors[0]._data.dtype)
5547
5548
5549
        combined_tensor.set_(
            tensors[0]._data.untyped_storage(),
            tensors[0]._data.storage_offset(),
5550
5551
5552
5553
            new_shape,
            new_stride,
        )
        combined_tensor = Float8Tensor.make_like(tensors[0], data=combined_tensor)
5554
    else:
5555
        combined_tensor = torch.Tensor().to(device=tensors[0].device, dtype=tensors[0].dtype)
5556
        combined_tensor.set_(
5557
5558
            tensors[0].untyped_storage(), tensors[0].storage_offset(), new_shape, new_stride
        )
5559
5560

    return combined_tensor
5561

5562

5563
5564
5565
5566
class FusedAttnFunc_qkvpacked(torch.autograd.Function):
    """Function for FusedAttention with packed QKV input"""

    @staticmethod
5567
5568
5569
5570
5571
    def forward(
        ctx,
        is_training,
        max_seqlen,
        cu_seqlens,
5572
        cu_seqlens_padded,
5573
5574
5575
5576
5577
5578
5579
5580
5581
        qkv,
        qkv_dtype,
        attn_bias,
        attn_scale,
        dropout_p,
        fast_zero_fill,
        qkv_layout,
        attn_bias_type,
        attn_mask_type,
5582
        window_size,
5583
5584
5585
5586
5587
        rng_gen,
        fused_attention_backend,
        use_FAv2_bwd,
        fp8,
        fp8_meta,
5588
        deterministic,
5589
    ):
5590
        # pylint: disable=missing-function-docstring
5591
        # "fp8_mha" decides outputs in fp8, while inputs are inferred from the real dtype
5592
5593
        is_input_fp8 = False
        is_output_fp8 = fp8_meta["recipe"].fp8_mha
5594
        if fp8:
5595
5596
            is_input_fp8 = isinstance(qkv, Float8Tensor)
            if is_input_fp8:
5597
5598
5599
5600
                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
5601
            qkv_group = len(qkv_layout.split("_"))
5602
5603
5604
5605
            assert (
                qkv_group == 1
            ), f"qkv layout should conform to 3hd or h3d, e.g. sb3hd, but found {qkv_layout}."
            if is_input_fp8:
5606
5607
5608
                qkv_fp8 = qkv._data
            else:
                qkv_c = qkv.view(-1, qkv.shape[-3] * qkv.shape[-2] * qkv.shape[-1])
5609
5610
5611
                qkv_fp8 = cast_to_fp8(
                    qkv_c, fp8_meta["scaling_fwd"], META_QKV, fp8_dtype_forward
                ).view(qkv.shape)
5612
            out_fp8, aux_ctx_tensors = fused_attn_fwd_qkvpacked(
5613
5614
5615
5616
5617
5618
5619
                is_training,
                max_seqlen,
                cu_seqlens,
                qkv_fp8,
                fp8_dtype_forward,
                fused_attention_backend,
                attn_bias,
5620
                cu_seqlens_padded,
5621
5622
5623
5624
5625
5626
5627
5628
5629
5630
5631
5632
                fp8_meta["scaling_fwd"].scale_inv,  # d_scale_qkv
                META_QKV,  # d_scale_qkv_offset
                fp8_meta["scaling_fwd"].scale_inv,  # d_scale_s
                META_S,  # d_scale_s_offset
                fp8_meta["scaling_fwd"].scale,  # q_scale_s
                META_S,  # q_scale_s_offset
                fp8_meta["scaling_fwd"].scale,  # q_scale_o
                META_O,  # q_scale_o_offset
                fp8_meta["scaling_fwd"].amax_history,  # amax_s
                META_S,  # amax_s_offset
                fp8_meta["scaling_fwd"].amax_history,  # amax_o
                META_O,  # amax_o_offset
5633
5634
5635
5636
5637
5638
                attn_scale,
                dropout_p,
                fast_zero_fill,
                qkv_layout,
                attn_bias_type,
                attn_mask_type,
5639
                window_size,
5640
5641
                rng_gen,
            )
5642
            if is_output_fp8:
5643
5644
                out_ret = Float8Tensor(
                    data=out_fp8,
5645
5646
5647
5648
5649
5650
5651
5652
5653
                    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]),
5654
5655
5656
5657
5658
                    fp8_meta["scaling_fwd"],
                    META_O,
                    fp8_dtype_forward,
                    qkv_dtype,
                ).view(out_fp8.shape)
5659
            out_save = out_ret
5660
5661
5662
5663
5664
5665
5666
5667
5668
5669
5670
5671
5672
5673
5674
5675
5676
5677
            if not int(os.getenv("NVTE_FP8_DPA_BWD", "1")):
                if is_input_fp8:
                    qkv_c = qkv.view(-1, qkv.shape[-3] * qkv.shape[-2] * qkv.shape[-1])
                    qkv = cast_from_fp8(
                        qkv_c._data,
                        fp8_meta["scaling_fwd"],
                        META_QKV,
                        fp8_dtype_forward,
                        TE_DType[qkv.dtype],
                    ).view(qkv.shape)
                if is_output_fp8:
                    out_save = cast_from_fp8(
                        out_fp8.view(-1, out_fp8.shape[-2] * out_fp8.shape[-1]),
                        fp8_meta["scaling_fwd"],
                        META_O,
                        fp8_dtype_forward,
                        qkv_dtype,
                    ).view(out_fp8.shape)
5678
5679
5680
            fp8_tensors = (
                qkv_fp8,
                out_fp8,
5681
                fp8_meta["scaling_fwd"].scale.clone(),
5682
5683
                fp8_meta["scaling_fwd"].scale_inv.clone(),
            )
5684
5685
        else:
            out_ret, aux_ctx_tensors = fused_attn_fwd_qkvpacked(
5686
5687
5688
5689
5690
5691
5692
                is_training,
                max_seqlen,
                cu_seqlens,
                qkv,
                qkv_dtype,
                fused_attention_backend,
                attn_bias,
5693
                cu_seqlens_padded,
5694
5695
5696
5697
5698
5699
5700
5701
5702
5703
5704
5705
                None,  # d_scale_qkv
                0,  # d_scale_qkv_offset
                None,  # d_scale_s
                0,  # d_scale_s_offset
                None,  # q_scale_s
                0,  # q_scale_s_offset
                None,  # q_scale_o
                0,  # q_scale_o_offset
                None,  # amax_s
                0,  # amax_s_offset
                None,  # amax_o
                0,  # amax_o_offset
5706
5707
5708
5709
5710
5711
                attn_scale,
                dropout_p,
                fast_zero_fill,
                qkv_layout,
                attn_bias_type,
                attn_mask_type,
5712
                window_size,
5713
5714
                rng_gen,
            )
5715
5716
5717
5718
            fp8_tensors = (None, None, None, None)
            out_save = out_ret

        ctx.fp8 = fp8 and int(os.getenv("NVTE_FP8_DPA_BWD", "1"))
5719
5720
        ctx.is_input_fp8 = is_input_fp8
        ctx.is_output_fp8 = is_output_fp8
5721
        qkvo_tensors = (qkv, out_save) if not ctx.fp8 else (None, None)
5722
        ctx.save_for_backward(
5723
            *qkvo_tensors, cu_seqlens, cu_seqlens_padded, *fp8_tensors, *aux_ctx_tensors
5724
        )
5725
        ctx.fp8_meta = fp8_meta
5726
5727
5728
5729
5730
5731
5732
5733
        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
5734
        ctx.window_size = window_size
5735
        ctx.fused_attention_backend = (
5736
            fused_attention_backend if ctx.fp8 else FusedAttnBackend["F16_arbitrary_seqlen"]
5737
        )
5738
        ctx.use_FAv2_bwd = use_FAv2_bwd
5739
        ctx.deterministic = deterministic
5740

5741
        return out_ret
5742
5743
5744

    @staticmethod
    def backward(ctx, d_out):
5745
        # pylint: disable=missing-function-docstring
5746
        if ctx.is_output_fp8:
5747
5748
5749
            assert isinstance(
                d_out, Float8Tensor
            ), "Gradient of the DPA output must be in Float8Tensor type for FP8 MHA."
5750
5751
5752
            d_out_f8tensor = d_out
            d_out = d_out._data

5753
        d_out = d_out.contiguous()
5754
5755
5756
5757
        (
            qkv,
            out,
            cu_seqlens,
5758
            cu_seqlens_padded,
5759
5760
5761
5762
5763
5764
            qkv_fp8,
            out_fp8,
            fwd_scales,
            fwd_scale_invs,
            *aux_ctx_tensors,
        ) = ctx.saved_tensors
5765
        rest = [None]
5766
5767
        if not aux_ctx_tensors[0].is_contiguous():
            aux_ctx_tensors[0] = aux_ctx_tensors[0].contiguous()
5768
        if ctx.use_FAv2_bwd:
5769
            softmax_lse, rng_state = aux_ctx_tensors
5770
            dqkv = torch.empty_like(qkv)
5771
5772
5773
            d_out, q, k, v, out = [
                maybe_contiguous(x) for x in (d_out, qkv[:, 0], qkv[:, 1], qkv[:, 2], out)
            ]
5774
            flash_attn_cuda_bwd(
5775
5776
5777
5778
5779
5780
5781
5782
5783
5784
5785
5786
5787
5788
5789
5790
5791
5792
5793
                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,
5794
            )
5795
            dqkv = dqkv[..., : d_out.shape[-1]]
5796
        else:
5797
5798
            with torch.cuda.nvtx.range("_FusedAttn_qkvpacked"):
                if ctx.fp8:
5799
                    fp8_dtype_forward = get_fp8_te_dtype(ctx.fp8_meta["recipe"], fprop_tensor=True)
5800
                    fp8_dtype_backward = get_fp8_te_dtype(
5801
5802
                        ctx.fp8_meta["recipe"], fprop_tensor=False
                    )
5803
                    if ctx.is_output_fp8:
5804
                        d_out_fp8 = d_out
5805
                        ctx.fp8_meta["scaling_bwd"].scale_inv[META_DO] = d_out_f8tensor._scale_inv
5806
5807
5808
                    else:
                        d_out_fp8 = cast_to_fp8(
                            d_out.view(-1, d_out.shape[-2] * d_out.shape[-1]),
5809
5810
5811
5812
                            ctx.fp8_meta["scaling_bwd"],
                            META_DO,
                            fp8_dtype_backward,
                        ).view(d_out.shape)
5813
                    dqkv_fp8, *rest = fused_attn_bwd_qkvpacked(
5814
5815
5816
5817
5818
5819
5820
5821
                        ctx.max_seqlen,
                        cu_seqlens,
                        qkv_fp8,
                        out_fp8,
                        d_out_fp8,
                        fp8_dtype_forward,
                        fp8_dtype_backward,
                        aux_ctx_tensors,
5822
                        ctx.fused_attention_backend,
5823
                        cu_seqlens_padded,
5824
5825
5826
5827
5828
5829
5830
5831
5832
5833
5834
5835
5836
5837
5838
5839
                        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,
5840
5841
                        ctx.window_size,
                        ctx.deterministic,
5842
                    )
5843
                    if ctx.is_input_fp8:
5844
5845
                        dqkv = Float8Tensor(
                            data=dqkv_fp8,
5846
5847
5848
5849
5850
                            fp8_meta=ctx.fp8_meta,
                            fp8_meta_forward=False,
                            fp8_meta_index=META_DQKV,
                            fp8_dtype=fp8_dtype_backward,
                            dtype=d_out_f8tensor.dtype,
5851
                        )
5852
                    else:
5853
5854
5855
5856
5857
5858
5859
5860
5861
5862
                        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)
5863
5864
5865
5866
                else:
                    if d_out.dtype == torch.uint8:
                        d_out = d_out_f8tensor.from_float8(qkv.dtype)
                    dqkv, *rest = fused_attn_bwd_qkvpacked(
5867
5868
5869
5870
5871
5872
5873
5874
                        ctx.max_seqlen,
                        cu_seqlens,
                        qkv,
                        out,
                        d_out,
                        ctx.qkv_dtype,
                        ctx.qkv_dtype,
                        aux_ctx_tensors,
5875
                        ctx.fused_attention_backend,
5876
                        cu_seqlens_padded,
5877
5878
5879
5880
5881
5882
5883
5884
5885
5886
5887
5888
5889
5890
5891
5892
                        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,
5893
5894
                        ctx.window_size,
                        ctx.deterministic,
5895
                    )
5896

5897
5898
        # if no_bias or alibi, return dqkv
        if ctx.attn_bias_type in ["no_bias", "alibi"]:
5899
5900
5901
5902
5903
5904
5905
5906
5907
5908
5909
5910
5911
5912
5913
5914
5915
5916
5917
5918
5919
            return (
                None,
                None,
                None,
                None,
                dqkv,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
5920
5921
                None,
                None,
5922
            )
5923
        # else, return (dqkv, dbias)
5924
5925
5926
5927
5928
5929
5930
5931
5932
5933
5934
5935
5936
5937
5938
5939
5940
5941
5942
5943
5944
        return (
            None,
            None,
            None,
            None,
            dqkv,
            None,
            rest[0],
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
5945
5946
            None,
            None,
5947
        )
5948

5949

5950
5951
5952
5953
class FusedAttnFunc_kvpacked(torch.autograd.Function):
    """Function for FusedAttention with packed KV input"""

    @staticmethod
5954
5955
5956
5957
5958
5959
5960
    def forward(
        ctx,
        is_training,
        max_seqlen_q,
        max_seqlen_kv,
        cu_seqlens_q,
        cu_seqlens_kv,
5961
5962
        cu_seqlens_q_padded,
        cu_seqlens_kv_padded,
5963
5964
5965
5966
5967
5968
5969
5970
5971
5972
        q,
        kv,
        qkv_dtype,
        attn_bias,
        attn_scale,
        dropout_p,
        fast_zero_fill,
        qkv_layout,
        attn_bias_type,
        attn_mask_type,
5973
        window_size,
5974
5975
5976
5977
5978
        rng_gen,
        fused_attention_backend,
        use_FAv2_bwd,
        fp8,
        fp8_meta,
5979
        deterministic,
5980
    ):
5981
        # pylint: disable=missing-function-docstring
5982
        # "fp8_mha" decides outputs in fp8, while inputs are inferred from the real dtype
5983
5984
        is_input_fp8 = False
        is_output_fp8 = fp8_meta["recipe"].fp8_mha
5985
        if fp8:
5986
5987
5988
            assert isinstance(kv, q.__class__), "q and kv must have the same type."
            is_input_fp8 = isinstance(q, Float8Tensor)
            if is_input_fp8:
5989
5990
5991
                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)
5992
            if is_input_fp8:
5993
5994
5995
                q_fp8, kv_fp8 = q._data, kv._data
            else:
                # 1: qkv packed, 2: kv packed, 3: qkv separate
5996
5997
                qkv_group = len(qkv_layout.split("_"))
                assert qkv_group == 2, (
5998
5999
                    "qkv layout should conform to hd_2hd or hd_h2d, e.g. sbhd_sb2hd, "
                    f"but found {qkv_layout}."
6000
6001
6002
6003
                )
                q_fp8 = cast_to_fp8(q, fp8_meta["scaling_fwd"], META_QKV, fp8_dtype_forward).view(
                    q.shape
                )
6004
                kv_c = kv.view(-1, kv.shape[-3] * kv.shape[-2] * kv.shape[-1])
6005
6006
6007
                kv_fp8 = cast_to_fp8(
                    kv_c, fp8_meta["scaling_fwd"], META_QKV, fp8_dtype_forward
                ).view(kv.shape)
6008
            out_fp8, aux_ctx_tensors = fused_attn_fwd_kvpacked(
6009
6010
6011
6012
6013
6014
6015
6016
6017
6018
                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,
6019
6020
                cu_seqlens_q_padded,
                cu_seqlens_kv_padded,
6021
6022
6023
6024
6025
6026
6027
6028
6029
6030
6031
6032
                fp8_meta["scaling_fwd"].scale_inv,  # d_scale_qkv
                META_QKV,  # d_scale_qkv_offset
                fp8_meta["scaling_fwd"].scale_inv,  # d_scale_s
                META_S,  # d_scale_s_offset
                fp8_meta["scaling_fwd"].scale,  # q_scale_s
                META_S,  # q_scale_s_offset
                fp8_meta["scaling_fwd"].scale,  # q_scale_o
                META_O,  # q_scale_o_offset
                fp8_meta["scaling_fwd"].amax_history,  # amax_s
                META_S,  # amax_s_offset
                fp8_meta["scaling_fwd"].amax_history,  # amax_o
                META_O,  # amax_o_offset
6033
6034
6035
6036
6037
6038
                attn_scale,
                dropout_p,
                fast_zero_fill,
                qkv_layout,
                attn_bias_type,
                attn_mask_type,
6039
                window_size,
6040
6041
                rng_gen,
            )
6042
            if is_output_fp8:
6043
6044
                out_ret = Float8Tensor(
                    data=out_fp8,
6045
6046
6047
6048
6049
6050
6051
6052
6053
                    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]),
6054
6055
6056
6057
6058
                    fp8_meta["scaling_fwd"],
                    META_O,
                    fp8_dtype_forward,
                    qkv_dtype,
                ).view(out_fp8.shape)
6059
            out_save = out_ret
6060
6061
6062
6063
6064
6065
6066
6067
6068
6069
6070
6071
6072
6073
6074
6075
6076
6077
6078
6079
6080
6081
6082
6083
6084
            if not int(os.getenv("NVTE_FP8_DPA_BWD", "1")):
                if is_input_fp8:
                    q = cast_from_fp8(
                        q._data,
                        fp8_meta["scaling_fwd"],
                        META_QKV,
                        fp8_dtype_forward,
                        TE_DType[q.dtype],
                    ).view(q.shape)
                    kv_c = kv.view(-1, kv.shape[-3] * kv.shape[-2] * kv.shape[-1])
                    kv = cast_from_fp8(
                        kv_c._data,
                        fp8_meta["scaling_fwd"],
                        META_QKV,
                        fp8_dtype_forward,
                        TE_DType[kv.dtype],
                    ).view(kv.shape)
                if is_output_fp8:
                    out_save = cast_from_fp8(
                        out_fp8.view(-1, out_fp8.shape[-2] * out_fp8.shape[-1]),
                        fp8_meta["scaling_fwd"],
                        META_O,
                        fp8_dtype_forward,
                        qkv_dtype,
                    ).view(out_fp8.shape)
6085
6086
6087
6088
            fp8_tensors = (
                q_fp8,
                kv_fp8,
                out_fp8,
6089
                fp8_meta["scaling_fwd"].scale.clone(),
6090
6091
                fp8_meta["scaling_fwd"].scale_inv.clone(),
            )
6092
6093
        else:
            out_ret, aux_ctx_tensors = fused_attn_fwd_kvpacked(
6094
6095
6096
6097
6098
6099
6100
6101
6102
6103
                is_training,
                max_seqlen_q,
                max_seqlen_kv,
                cu_seqlens_q,
                cu_seqlens_kv,
                q,
                kv,
                qkv_dtype,
                fused_attention_backend,
                attn_bias,
6104
6105
                cu_seqlens_q_padded,
                cu_seqlens_kv_padded,
6106
6107
6108
6109
6110
6111
6112
6113
6114
6115
6116
6117
                None,  # d_scale_qkv
                0,  # d_scale_qkv_offset
                None,  # d_scale_s
                0,  # d_scale_s_offset
                None,  # q_scale_s
                0,  # q_scale_s_offset
                None,  # q_scale_o
                0,  # q_scale_o_offset
                None,  # amax_s
                0,  # amax_s_offset
                None,  # amax_o
                0,  # amax_o_offset
6118
6119
6120
6121
6122
6123
                attn_scale,
                dropout_p,
                fast_zero_fill,
                qkv_layout,
                attn_bias_type,
                attn_mask_type,
6124
                window_size,
6125
6126
                rng_gen,
            )
6127
6128
6129
6130
            out_save = out_ret
            fp8_tensors = (None, None, None, None, None)

        ctx.fp8 = fp8 and int(os.getenv("NVTE_FP8_DPA_BWD", "1"))
6131
6132
        ctx.is_input_fp8 = is_input_fp8
        ctx.is_output_fp8 = is_output_fp8
6133
        qkvo_tensors = (q, kv, out_save) if not ctx.fp8 else (None, None, None)
6134
6135
6136
6137
        ctx.save_for_backward(
            *qkvo_tensors,
            cu_seqlens_q,
            cu_seqlens_kv,
6138
6139
            cu_seqlens_q_padded,
            cu_seqlens_kv_padded,
6140
6141
6142
            *fp8_tensors,
            *aux_ctx_tensors,
        )
6143
        ctx.fp8_meta = fp8_meta
6144
6145
6146
6147
6148
6149
6150
6151
6152
        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
6153
        ctx.window_size = window_size
6154
        ctx.fused_attention_backend = (
6155
            fused_attention_backend if ctx.fp8 else FusedAttnBackend["F16_arbitrary_seqlen"]
6156
        )
6157
        ctx.use_FAv2_bwd = use_FAv2_bwd
6158
        ctx.deterministic = deterministic
6159

6160
        return out_ret
6161
6162
6163

    @staticmethod
    def backward(ctx, d_out):
6164
        # pylint: disable=missing-function-docstring
6165
        if ctx.is_output_fp8:
6166
6167
6168
            assert isinstance(
                d_out, Float8Tensor
            ), "Gradient of the DPA output must be in Float8Tensor type for FP8 MHA."
6169
6170
6171
            d_out_f8tensor = d_out
            d_out = d_out._data

6172
        d_out = d_out.contiguous()
6173
6174
6175
6176
6177
6178
        (
            q,
            kv,
            out,
            cu_seqlens_q,
            cu_seqlens_kv,
6179
6180
            cu_seqlens_q_padded,
            cu_seqlens_kv_padded,
6181
6182
6183
6184
6185
6186
6187
            q_fp8,
            kv_fp8,
            out_fp8,
            fwd_scales,
            fwd_scale_invs,
            *aux_ctx_tensors,
        ) = ctx.saved_tensors
6188
        rest = [None]
6189
6190
        if not aux_ctx_tensors[0].is_contiguous():
            aux_ctx_tensors[0] = aux_ctx_tensors[0].contiguous()
6191
        if ctx.use_FAv2_bwd:
6192
            softmax_lse, rng_state = aux_ctx_tensors
6193
6194
            dq = torch.empty_like(q)
            dkv = torch.empty_like(kv)
6195
            d_out, q, k, v, out = [maybe_contiguous(x) for x in (d_out, q, kv[:, 0], kv[:, 1], out)]
6196
            flash_attn_cuda_bwd(
6197
6198
6199
6200
6201
6202
6203
6204
6205
6206
6207
6208
6209
6210
6211
6212
6213
6214
6215
                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,
6216
            )
6217
6218
            dq = dq[..., : d_out.shape[-1]]
            dkv = dkv[..., : d_out.shape[-1]]
6219
        else:
6220
6221
            with torch.cuda.nvtx.range("_FusedAttn_kvpacked"):
                if ctx.fp8:
6222
                    fp8_dtype_forward = get_fp8_te_dtype(ctx.fp8_meta["recipe"], fprop_tensor=True)
6223
                    fp8_dtype_backward = get_fp8_te_dtype(
6224
6225
                        ctx.fp8_meta["recipe"], fprop_tensor=False
                    )
6226
                    if ctx.is_output_fp8:
6227
                        d_out_fp8 = d_out
6228
                        ctx.fp8_meta["scaling_bwd"].scale_inv[META_DO] = d_out_f8tensor._scale_inv
6229
6230
6231
                    else:
                        d_out_fp8 = cast_to_fp8(
                            d_out.view(-1, d_out.shape[-2] * d_out.shape[-1]),
6232
6233
6234
6235
                            ctx.fp8_meta["scaling_bwd"],
                            META_DO,
                            fp8_dtype_backward,
                        ).view(d_out.shape)
6236
                    dq_fp8, dkv_fp8, *rest = fused_attn_bwd_kvpacked(
6237
6238
6239
6240
6241
6242
6243
6244
6245
6246
6247
                        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,
6248
                        ctx.fused_attention_backend,
6249
6250
                        cu_seqlens_q_padded,
                        cu_seqlens_kv_padded,
6251
6252
6253
6254
6255
6256
6257
6258
6259
6260
6261
6262
6263
6264
6265
6266
                        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,
6267
6268
                        ctx.window_size,
                        ctx.deterministic,
6269
                    )
6270
                    if ctx.is_input_fp8:
6271
6272
                        dq = Float8Tensor(
                            data=dq_fp8,
6273
6274
6275
6276
6277
                            fp8_meta=ctx.fp8_meta,
                            fp8_meta_forward=False,
                            fp8_meta_index=META_DQKV,
                            fp8_dtype=fp8_dtype_backward,
                            dtype=d_out_f8tensor.dtype,
6278
6279
6280
                        )
                        dkv = Float8Tensor(
                            data=dkv_fp8,
6281
6282
6283
6284
6285
                            fp8_meta=ctx.fp8_meta,
                            fp8_meta_forward=False,
                            fp8_meta_index=META_DQKV,
                            fp8_dtype=fp8_dtype_backward,
                            dtype=d_out_f8tensor.dtype,
6286
                        )
6287
6288
6289
                    else:
                        dq = cast_from_fp8(
                            dq_fp8.view(-1, dq_fp8.shape[-2] * dq_fp8.shape[-1]),
6290
6291
6292
6293
6294
6295
6296
6297
6298
6299
6300
6301
6302
6303
6304
                            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)
6305
6306
6307
6308
                else:
                    if d_out.dtype == torch.uint8:
                        d_out = d_out_f8tensor.from_float8(q.dtype)
                    dq, dkv, *rest = fused_attn_bwd_kvpacked(
6309
6310
6311
6312
6313
6314
6315
6316
6317
6318
6319
                        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,
6320
                        ctx.fused_attention_backend,
6321
6322
                        cu_seqlens_q_padded,
                        cu_seqlens_kv_padded,
6323
6324
6325
6326
6327
6328
6329
6330
6331
6332
6333
6334
6335
6336
6337
6338
                        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,
6339
6340
                        ctx.window_size,
                        ctx.deterministic,
6341
                    )
6342

6343
6344
        # if no_bias or alibi, return dqkv
        if ctx.attn_bias_type in ["no_bias", "alibi"]:
6345
6346
6347
6348
6349
6350
6351
6352
6353
6354
6355
6356
6357
6358
6359
6360
6361
6362
6363
6364
6365
6366
6367
6368
6369
            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,
6370
6371
                None,
                None,
6372
            )
6373
        # else, return (dqkv, dbias)
6374
6375
6376
6377
6378
6379
6380
6381
6382
6383
6384
6385
6386
6387
6388
6389
6390
6391
6392
6393
6394
6395
6396
6397
6398
        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,
6399
6400
            None,
            None,
6401
6402
        )

6403

6404
6405
6406
6407
class FusedAttnFunc(torch.autograd.Function):
    """Function for FusedAttention with separate Q, K, V tensors"""

    @staticmethod
6408
6409
6410
6411
6412
6413
6414
    def forward(
        ctx,
        is_training,
        max_seqlen_q,
        max_seqlen_kv,
        cu_seqlens_q,
        cu_seqlens_kv,
6415
6416
        cu_seqlens_q_padded,
        cu_seqlens_kv_padded,
6417
6418
6419
6420
6421
6422
6423
6424
6425
6426
6427
        q,
        k,
        v,
        qkv_dtype,
        attn_bias,
        attn_scale,
        dropout_p,
        fast_zero_fill,
        qkv_layout,
        attn_bias_type,
        attn_mask_type,
6428
        window_size,
6429
6430
6431
6432
6433
        rng_gen,
        fused_attention_backend,
        use_FAv2_bwd,
        fp8,
        fp8_meta,
6434
        deterministic,
6435
    ):
6436
        # pylint: disable=missing-function-docstring
6437
        # "fp8_mha" decides outputs in fp8, while inputs are inferred from the real dtype
6438
6439
        is_input_fp8 = False
        is_output_fp8 = fp8_meta["recipe"].fp8_mha
6440
6441
6442
        if fp8:
            fused_attention_backend = FusedAttnBackend["FP8"]
            fp8_dtype_forward = get_fp8_te_dtype(fp8_meta["recipe"], fprop_tensor=True)
6443
6444
6445
6446
6447
            assert isinstance(k, q.__class__) and isinstance(
                v, q.__class__
            ), "q, k, and v must have the same type."
            is_input_fp8 = isinstance(q, Float8Tensor)
            if is_input_fp8:
6448
6449
6450
6451
                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
6452
                qkv_group = len(qkv_layout.split("_"))
6453
                if qkv_group == 1:
6454
6455
                    dim = qkv_layout.find("3")
                    qkv = _combine_tensors([q, k, v], dim)
6456
                    qkv_c = qkv.view(-1, qkv.shape[-3] * qkv.shape[-2] * qkv.shape[-1])
6457
6458
6459
6460
                    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])
6461
6462
                    q_fp8, k_fp8, v_fp8 = [x.squeeze(dim) for x in [q_fp8, k_fp8, v_fp8]]
                if qkv_group == 2:
6463
6464
6465
6466
6467
                    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)
6468
                    kv_c = kv.view(-1, kv.shape[-3] * kv.shape[-2] * kv.shape[-1])
6469
6470
6471
6472
                    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])
6473
6474
                    k_fp8, v_fp8 = [x.squeeze(dim) for x in [k_fp8, v_fp8]]
                if qkv_group == 3:
6475
6476
6477
6478
6479
6480
6481
6482
6483
                    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)
6484
            out_fp8, aux_ctx_tensors = fused_attn_fwd(
6485
6486
6487
6488
6489
6490
6491
6492
6493
6494
6495
                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,
6496
6497
                cu_seqlens_q_padded,
                cu_seqlens_kv_padded,
6498
6499
6500
6501
6502
6503
6504
6505
6506
6507
6508
6509
                fp8_meta["scaling_fwd"].scale_inv,  # d_scale_qkv
                META_QKV,  # d_scale_qkv_offset
                fp8_meta["scaling_fwd"].scale_inv,  # d_scale_s
                META_S,  # d_scale_s_offset
                fp8_meta["scaling_fwd"].scale,  # q_scale_s
                META_S,  # q_scale_s_offset
                fp8_meta["scaling_fwd"].scale,  # q_scale_o
                META_O,  # q_scale_o_offset
                fp8_meta["scaling_fwd"].amax_history,  # amax_s
                META_S,  # amax_s_offset
                fp8_meta["scaling_fwd"].amax_history,  # amax_o
                META_O,  # amax_o_offset
6510
6511
6512
6513
6514
6515
                attn_scale,
                dropout_p,
                fast_zero_fill,
                qkv_layout,
                attn_bias_type,
                attn_mask_type,
6516
                window_size,
6517
6518
                rng_gen,
            )
6519
            if is_output_fp8:
6520
6521
                out_ret = Float8Tensor(
                    data=out_fp8,
6522
6523
6524
6525
6526
6527
6528
6529
6530
                    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]),
6531
6532
6533
6534
6535
                    fp8_meta["scaling_fwd"],
                    META_O,
                    fp8_dtype_forward,
                    qkv_dtype,
                ).view(out_fp8.shape)
6536
6537
            out_save = out_ret

6538
            if not int(os.getenv("NVTE_FP8_DPA_BWD", "1")):
6539
                # 1: qkv packed, 2: kv packed, 3: qkv separate
6540
6541
6542
6543
6544
6545
6546
6547
6548
6549
6550
6551
6552
6553
6554
6555
6556
6557
6558
6559
6560
6561
6562
6563
6564
6565
6566
6567
6568
6569
6570
6571
6572
6573
6574
6575
6576
6577
6578
6579
6580
6581
6582
6583
6584
6585
6586
6587
6588
6589
6590
6591
6592
6593
6594
6595
6596
6597
6598
6599
                if is_input_fp8:
                    qkv_group = len(qkv_layout.split("_"))
                    if qkv_group == 1:
                        dim = qkv_layout.find("3")
                        qkv = _combine_tensors([q, k, v], dim)
                        qkv_c = qkv.view(-1, qkv.shape[-3] * qkv.shape[-2] * qkv.shape[-1])
                        qkv_no_fp8 = cast_from_fp8(
                            qkv_c._data,
                            fp8_meta["scaling_fwd"],
                            META_QKV,
                            fp8_dtype_forward,
                            TE_DType[qkv.dtype],
                        ).view(qkv.shape)
                        q, k, v = _SplitAlongDim.apply(qkv_no_fp8, dim, [1, 1, 1])
                        q, k, v = [x.squeeze(dim) for x in [q, k, v]]
                    if qkv_group == 2:
                        q = cast_from_fp8(
                            q._data,
                            fp8_meta["scaling_fwd"],
                            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)
                        kv_c = kv.view(-1, kv.shape[-3] * kv.shape[-2] * kv.shape[-1])
                        kv_no_fp8 = cast_from_fp8(
                            kv_c._data,
                            fp8_meta["scaling_fwd"],
                            META_QKV,
                            fp8_dtype_forward,
                            TE_DType[kv.dtype],
                        ).view(kv.shape)
                        k, v = _SplitAlongDim.apply(kv_no_fp8, dim, [1, 1])
                        k, v = [x.squeeze(dim) for x in [k, v]]
                    if qkv_group == 3:
                        q = cast_from_fp8(
                            q._data,
                            fp8_meta["scaling_fwd"],
                            META_QKV,
                            fp8_dtype_forward,
                            TE_DType[q.dtype],
                        ).view(q.shape)
                        k = cast_from_fp8(
                            k._data,
                            fp8_meta["scaling_fwd"],
                            META_QKV,
                            fp8_dtype_forward,
                            TE_DType[k.dtype],
                        ).view(k.shape)
                        v = cast_from_fp8(
                            v._data,
                            fp8_meta["scaling_fwd"],
                            META_QKV,
                            fp8_dtype_forward,
                            TE_DType[v.dtype],
                        ).view(v.shape)
                if is_output_fp8:
                    out_save = cast_from_fp8(
                        out_fp8.view(-1, out_fp8.shape[-2] * out_fp8.shape[-1]),
6600
                        fp8_meta["scaling_fwd"],
6601
                        META_O,
6602
                        fp8_dtype_forward,
6603
6604
                        qkv_dtype,
                    ).view(out_fp8.shape)
6605
6606
6607
6608
6609
6610

            fp8_tensors = (
                q_fp8,
                k_fp8,
                v_fp8,
                out_fp8,
6611
                fp8_meta["scaling_fwd"].scale.clone(),
6612
6613
                fp8_meta["scaling_fwd"].scale_inv.clone(),
            )
6614
6615
        else:
            out_ret, aux_ctx_tensors = fused_attn_fwd(
6616
6617
6618
6619
6620
6621
6622
6623
6624
6625
6626
                is_training,
                max_seqlen_q,
                max_seqlen_kv,
                cu_seqlens_q,
                cu_seqlens_kv,
                q,
                k,
                v,
                qkv_dtype,
                fused_attention_backend,
                attn_bias,
6627
6628
                cu_seqlens_q_padded,
                cu_seqlens_kv_padded,
6629
6630
6631
6632
6633
6634
6635
6636
6637
6638
6639
6640
                None,  # d_scale_qkv
                0,  # d_scale_qkv_offset
                None,  # d_scale_s
                0,  # d_scale_s_offset
                None,  # q_scale_s
                0,  # q_scale_s_offset
                None,  # q_scale_o
                0,  # q_scale_o_offset
                None,  # amax_s
                0,  # amax_s_offset
                None,  # amax_o
                0,  # amax_o_offset
6641
6642
6643
6644
6645
6646
                attn_scale,
                dropout_p,
                fast_zero_fill,
                qkv_layout,
                attn_bias_type,
                attn_mask_type,
6647
                window_size,
6648
6649
                rng_gen,
            )
6650
6651
            out_save = out_ret
            fp8_tensors = (None, None, None, None, None, None)
6652

6653
6654
        ctx.fp8 = fp8 and int(os.getenv("NVTE_FP8_DPA_BWD", "1"))

6655
        from .cpu_offload import CPUOffloadEnabled
6656

6657
        if CPUOffloadEnabled:
6658
6659
6660
6661
6662
6663
6664
            if ctx.fp8:
                tensor_list = fp8_tensors
            else:
                tensor_list = [q, k, v, out_save]

            tensor_list.extend(aux_ctx_tensors)

6665
            qkv_layout = "sbhd_sbhd_sbhd"
6666
6667
6668
6669
            for tensor in tensor_list:
                if tensor is not None:
                    tensor.activation_offloading = True

6670
6671
        ctx.is_input_fp8 = is_input_fp8
        ctx.is_output_fp8 = is_output_fp8
6672
        qkvo_tensors = (q, k, v, out_save) if not ctx.fp8 else (None, None, None, None)
6673
6674
6675
6676
        ctx.save_for_backward(
            *qkvo_tensors,
            cu_seqlens_q,
            cu_seqlens_kv,
6677
6678
            cu_seqlens_q_padded,
            cu_seqlens_kv_padded,
6679
6680
6681
            *fp8_tensors,
            *aux_ctx_tensors,
        )
6682
        ctx.fp8_meta = fp8_meta
6683
6684
6685
6686
6687
6688
6689
6690
6691
        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
6692
        ctx.window_size = window_size
6693
        ctx.fused_attention_backend = (
6694
            fused_attention_backend if ctx.fp8 else FusedAttnBackend["F16_arbitrary_seqlen"]
6695
        )
6696
        ctx.use_FAv2_bwd = use_FAv2_bwd
6697
        ctx.deterministic = deterministic
6698

6699
        return out_ret
6700
6701
6702

    @staticmethod
    def backward(ctx, d_out):
6703
        # pylint: disable=missing-function-docstring
6704
        if ctx.is_output_fp8:
6705
6706
6707
            assert isinstance(
                d_out, Float8Tensor
            ), "Gradient of the DPA output must be in Float8Tensor type for FP8 MHA."
6708
6709
6710
            d_out_f8tensor = d_out
            d_out = d_out._data

6711
        d_out = d_out.contiguous()
6712
6713
6714
6715
6716
6717
6718
        (
            q,
            k,
            v,
            out,
            cu_seqlens_q,
            cu_seqlens_kv,
6719
6720
            cu_seqlens_q_padded,
            cu_seqlens_kv_padded,
6721
6722
6723
6724
6725
6726
6727
6728
            q_fp8,
            k_fp8,
            v_fp8,
            out_fp8,
            fwd_scales,
            fwd_scale_invs,
            *aux_ctx_tensors,
        ) = ctx.saved_tensors
6729
6730
        if not aux_ctx_tensors[0].is_contiguous():
            aux_ctx_tensors[0] = aux_ctx_tensors[0].contiguous()
6731
        rest = [None]
6732
        if ctx.use_FAv2_bwd:
6733
            softmax_lse, rng_state = aux_ctx_tensors
6734
6735
6736
            dq = torch.empty_like(q)
            dk = torch.empty_like(k)
            dv = torch.empty_like(v)
6737
            d_out, q, k, v, out = [maybe_contiguous(x) for x in (d_out, q, k, v, out)]
6738
            flash_attn_cuda_bwd(
6739
6740
6741
6742
6743
6744
6745
6746
6747
6748
6749
6750
6751
6752
6753
6754
6755
6756
6757
                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,
6758
            )
6759
6760
6761
            dq = dq[..., : d_out.shape[-1]]
            dk = dk[..., : d_out.shape[-1]]
            dv = dv[..., : d_out.shape[-1]]
6762
        else:
6763
6764
6765
6766
            with torch.cuda.nvtx.range("_FusedAttn"):
                if ctx.fp8:
                    fp8_dtype_forward = get_fp8_te_dtype(ctx.fp8_meta["recipe"], fprop_tensor=True)
                    fp8_dtype_backward = get_fp8_te_dtype(
6767
6768
                        ctx.fp8_meta["recipe"], fprop_tensor=False
                    )
6769
                    if ctx.is_output_fp8:
6770
                        d_out_fp8 = d_out
6771
                        ctx.fp8_meta["scaling_bwd"].scale_inv[META_DO] = d_out_f8tensor._scale_inv
6772
6773
6774
                    else:
                        d_out_fp8 = cast_to_fp8(
                            d_out.view(-1, d_out.shape[-2] * d_out.shape[-1]),
6775
6776
6777
6778
                            ctx.fp8_meta["scaling_bwd"],
                            META_DO,
                            fp8_dtype_backward,
                        ).view(d_out.shape)
6779
                    dq_fp8, dk_fp8, dv_fp8, *rest = fused_attn_bwd(
6780
6781
6782
6783
6784
6785
6786
6787
6788
6789
6790
6791
                        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,
6792
                        ctx.fused_attention_backend,
6793
6794
                        cu_seqlens_q_padded,
                        cu_seqlens_kv_padded,
6795
6796
6797
6798
6799
6800
6801
6802
6803
6804
6805
6806
6807
6808
6809
6810
                        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,
6811
6812
                        ctx.window_size,
                        ctx.deterministic,
6813
                    )
6814

6815
                    if ctx.is_input_fp8:
6816
6817
                        dq = Float8Tensor(
                            data=dq_fp8,
6818
6819
6820
6821
6822
                            fp8_meta=ctx.fp8_meta,
                            fp8_meta_forward=False,
                            fp8_meta_index=META_DQKV,
                            fp8_dtype=fp8_dtype_backward,
                            dtype=d_out_f8tensor.dtype,
6823
6824
6825
                        )
                        dk = Float8Tensor(
                            data=dk_fp8,
6826
6827
6828
6829
6830
                            fp8_meta=ctx.fp8_meta,
                            fp8_meta_forward=False,
                            fp8_meta_index=META_DQKV,
                            fp8_dtype=fp8_dtype_backward,
                            dtype=d_out_f8tensor.dtype,
6831
6832
6833
                        )
                        dv = Float8Tensor(
                            data=dv_fp8,
6834
6835
6836
6837
6838
                            fp8_meta=ctx.fp8_meta,
                            fp8_meta_forward=False,
                            fp8_meta_index=META_DQKV,
                            fp8_dtype=fp8_dtype_backward,
                            dtype=d_out_f8tensor.dtype,
6839
                        )
6840
                    else:
6841
                        qkv_group = len(ctx.qkv_layout.split("_"))
6842
                        if qkv_group == 1:
6843
6844
6845
6846
6847
6848
6849
6850
6851
6852
6853
6854
6855
                            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])
6856
6857
6858
6859
                            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]),
6860
6861
6862
6863
6864
6865
6866
6867
6868
6869
6870
6871
6872
6873
6874
6875
6876
6877
                                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])
6878
6879
6880
6881
                            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]),
6882
6883
6884
6885
6886
                                ctx.fp8_meta["scaling_bwd"],
                                META_DQKV,
                                fp8_dtype_backward,
                                ctx.qkv_dtype,
                            ).view(dq_fp8.shape)
6887
6888
                            dk = cast_from_fp8(
                                dk_fp8.view(-1, dk_fp8.shape[-2] * dk_fp8.shape[-1]),
6889
6890
6891
6892
6893
                                ctx.fp8_meta["scaling_bwd"],
                                META_DQKV,
                                fp8_dtype_backward,
                                ctx.qkv_dtype,
                            ).view(dk_fp8.shape)
6894
6895
                            dv = cast_from_fp8(
                                dv_fp8.view(-1, dv_fp8.shape[-2] * dv_fp8.shape[-1]),
6896
6897
6898
6899
6900
                                ctx.fp8_meta["scaling_bwd"],
                                META_DQKV,
                                fp8_dtype_backward,
                                ctx.qkv_dtype,
                            ).view(dv_fp8.shape)
6901
6902
6903
6904
                else:
                    if d_out.dtype == torch.uint8:
                        d_out = d_out_f8tensor.from_float8(q.dtype)
                    dq, dk, dv, *rest = fused_attn_bwd(
6905
6906
6907
6908
6909
6910
6911
6912
6913
6914
6915
6916
                        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,
6917
                        ctx.fused_attention_backend,
6918
6919
                        cu_seqlens_q_padded,
                        cu_seqlens_kv_padded,
6920
6921
6922
6923
6924
6925
6926
6927
6928
6929
6930
6931
6932
6933
6934
6935
                        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,
6936
6937
                        ctx.window_size,
                        ctx.deterministic,
6938
                    )
6939

6940
6941
        # if no_bias or alibi, return dqkv
        if ctx.attn_bias_type in ["no_bias", "alibi"]:
6942
6943
6944
6945
6946
6947
6948
6949
6950
6951
6952
6953
6954
6955
6956
6957
6958
6959
6960
6961
6962
6963
6964
6965
6966
6967
            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,
6968
6969
                None,
                None,
6970
            )
6971
        # else, return (dqkv, dbias)
6972
6973
6974
6975
6976
6977
6978
6979
6980
6981
6982
6983
6984
6985
6986
6987
6988
6989
6990
6991
6992
6993
6994
6995
6996
6997
        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,
6998
6999
            None,
            None,
7000
        )
7001

7002

7003
class FusedAttention(torch.nn.Module):
7004
7005
7006
7007
7008
7009
7010
7011
7012
    """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:

7013
7014
7015
7016
    | backend       | 1                       | 2                              |
    | flash based   | no                      | yes                            |
    | cuDNN based   | yes                     | yes                            |
    | qkv dtype     | fp16/bf16               | fp16/bf16                      |
7017
    | attn_type     | self/cross              | self/cross                     |
7018
    | qkv_layout    |                         |                                |
7019
    |  - (q,k,v)    | sb3hd, bs3hd            | sb3hd, bs3hd, sbh3d, bsh3d     |
7020
    |               | sbhd_sb2hd, bshd_bs2hd  | sbhd_sb2hd, bshd_bs2hd         |
7021
7022
    |               | bshd_bshd_bshd          | sbhd_sbh2d, bshd_bsh2d         |
    |               |                         | sbhd_sbhd_sbhd, bshd_bshd_bshd |
7023
7024
    | mask_type     | causal/padding/no_mask  | causal/padding/no_mask         |
    | bias_type     | post_scale_bias/no_bias | post_scale_bias/alibi/no_bias  |
7025
    | dropout       | yes                     | yes                            |
7026
7027
    | max_seqlen    | <=512, multiple of 64   | any, multiple of 64            |
    | head_dim      | 64                      | <=128, multiple of 8           |
7028
    | output dtype  | fp16/bf16               | fp16/bf16                      |
7029
7030
7031
7032
    """

    def __init__(
        self,
7033
        softmax_scale: float,
7034
7035
7036
        attention_dropout: float = 0.0,
        attention_dropout_ctx: Optional[Callable] = nullcontext,
        attention_type: str = "self",
7037
7038
        layer_number: Optional[int] = None,
        deterministic: bool = False,
7039
7040
7041
    ) -> None:
        super().__init__()

7042
        self.softmax_scale = softmax_scale
7043
7044
7045
        self.attention_dropout = attention_dropout
        self.attention_dropout_ctx = attention_dropout_ctx
        self.attention_type = attention_type
7046
7047
7048
        self.use_FAv2_bwd = os.getenv(
            "NVTE_FUSED_ATTN_USE_FAv2_BWD", "0"
        ) == "1" and get_device_compute_capability() == (9, 0)
7049
        self.layer_number = 1 if layer_number is None else layer_number
7050
        self.deterministic = deterministic
7051

7052
        def remove_extra_states_check(self, incompatible_keys):  # pylint: disable=unused-argument
7053
7054
            """
            Temporarily remove fused_attention._extra_state as a missing key
7055
            or an unexpected key when loading Transformer Engine checkpoints.
7056
7057
            Please store FP8 metadata as DotProductAttention's _extra_state,
            rather than FusedAttention's _extra_state. This hook will be
7058
            phased out in Transformer Engine 2.0.
7059
7060
            """
            for key in incompatible_keys.missing_keys:
7061
                if "fused_attention._extra_state" in key:
7062
                    incompatible_keys.missing_keys.remove(key)
7063
7064
7065
7066
7067
7068
7069
            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."
                    )
7070

7071
7072
        self.register_load_state_dict_post_hook(remove_extra_states_check)

7073
    @no_torch_dynamo()
7074
7075
7076
7077
7078
    def forward(
        self,
        query_layer: torch.Tensor,
        key_layer: torch.Tensor,
        value_layer: torch.Tensor,
7079
7080
7081
        qkv_layout: str = "sbh3d",
        cu_seqlens_q: Optional[torch.Tensor] = None,
        cu_seqlens_kv: Optional[torch.Tensor] = None,
7082
7083
        cu_seqlens_q_padded: Optional[torch.Tensor] = None,
        cu_seqlens_kv_padded: Optional[torch.Tensor] = None,
7084
7085
        max_seqlen_q: Optional[int] = None,
        max_seqlen_kv: Optional[int] = None,
7086
        attn_mask_type: str = "causal",
7087
        attention_mask: Optional[Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]] = None,
7088
        window_size: Optional[Tuple[int, int]] = None,
7089
        fused_attention_backend: tex.NVTE_Fused_Attn_Backend = tex.NVTE_Fused_Attn_Backend.NVTE_No_Backend,
7090
7091
7092
        core_attention_bias_type: str = "no_bias",
        core_attention_bias: Optional[torch.Tensor] = None,
        fast_zero_fill: bool = True,
7093
        cp_group: Optional[Union[dist_group_type, List[dist_group_type]]] = None,
7094
7095
        cp_global_ranks: List[int] = None,
        cp_stream: torch.cuda.Stream = None,
7096
        cp_comm_type: str = "p2p",
7097
7098
        fp8: bool = False,
        fp8_meta: Optional[Dict[str, Any]] = None,
7099
7100
    ) -> torch.Tensor:
        """fused attention fprop"""
7101
7102
7103
        assert (
            fused_attention_backend != tex.NVTE_Fused_Attn_Backend.NVTE_No_Backend
        ), "No fused attention backend supports this input combination!"
7104
7105
7106
7107
        assert all(
            x.dtype in [torch.float16, torch.bfloat16] or isinstance(x, Float8Tensor)
            for x in [query_layer, key_layer, value_layer]
        ), "FusedAttention only supports FP16 and BF16 data types, or Float8Tensors."
7108
7109
        assert (
            query_layer.is_cuda and key_layer.is_cuda and value_layer.is_cuda
7110
        ), "FusedAttention only supports CUDA tensors."
7111
7112
        assert (
            qkv_layout in QKVLayouts
7113
        ), f"FusedAttention does not support qkv_layout = {qkv_layout}!"
7114

7115
7116
7117
7118
7119
7120
        cp_size = 1
        if isinstance(cp_group, dist_group_type):
            cp_size = get_distributed_world_size(cp_group)
        elif isinstance(cp_group, list):
            for group in cp_group:
                cp_size *= get_distributed_world_size(group)
7121
        context_parallel = cp_size > 1
7122

7123
        qkv_format = "".join([i for i in qkv_layout.split("_")[0] if i.isalpha()])
7124

7125
7126
        if qkv_format in ["sbhd", "bshd"]:
            if qkv_format == "sbhd":
7127
                batch_size, max_seqlen_q, max_seqlen_kv = (
7128
7129
7130
7131
7132
                    query_layer.shape[1],
                    query_layer.shape[0],
                    key_layer.shape[0],
                )
            if qkv_format == "bshd":
7133
                batch_size, max_seqlen_q, max_seqlen_kv = (
7134
7135
7136
7137
                    query_layer.shape[0],
                    query_layer.shape[1],
                    key_layer.shape[1],
                )
7138
7139
            max_seqlen_q *= cp_size
            max_seqlen_kv *= cp_size
7140
            if "padding" in attn_mask_type:
7141
7142
                assert not context_parallel, "Padding mask not supported with context parallelism!"

7143
7144
7145
7146
7147
                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!"
                        )
7148
                    if self.attention_type == "self":
7149
7150
                        cu_seqlens_q = get_cu_seqlens(attention_mask)
                        cu_seqlens_kv = cu_seqlens_q
7151
                    else:
7152
7153
                        cu_seqlens_q = get_cu_seqlens(attention_mask[0])
                        cu_seqlens_kv = get_cu_seqlens(attention_mask[1])
7154
            else:
7155
7156
7157
7158
7159
7160
7161
7162
7163
7164
7165
7166
                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,
                    )
7167
7168
7169
        if qkv_format == "thd":
            assert (
                max_seqlen_q is not None
7170
7171
7172
                and max_seqlen_kv is not None
                and cu_seqlens_q is not None
                and cu_seqlens_kv is not None
7173
            ), "max_seqlen_q/kv and cu_seqlens_q/kv can not be None when qkv_format is thd!"
7174
7175
7176
7177

        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
7178
7179
7180

        qkv_dtype = TE_DType[query_layer.dtype]

7181
7182
7183
7184
7185
        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)
        )
7186

7187
7188
7189
7190
7191
7192
7193
7194
7195
7196
7197
        if fp8:
            assert fused_attention_backend == tex.NVTE_Fused_Attn_Backend.NVTE_FP8, (
                f"cuDNN attention sub-backend {int(tex.NVTE_Fused_Attn_Backend.NVTE_FP8)}"
                " is required for FP8 attention!"
            )
            assert fp8_meta is not None, "FP8 metadata fp8_meta is required for FP8 attention!"
            assert not context_parallel or fp8_meta["recipe"].reduce_amax, (
                "Amax reduction across TP+CP group is necessary when using context parallelism with"
                " FP8!"
            )

7198
        if context_parallel:
7199
            assert (
7200
7201
                fp8
                or fused_attention_backend == tex.NVTE_Fused_Attn_Backend.NVTE_F16_arbitrary_seqlen
7202
7203
7204
7205
7206
7207
7208
            ), 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)
            ]
7209
7210
7211
            with self.attention_dropout_ctx():
                output = attn_forward_func_with_cp(
                    self.training,
7212
7213
7214
7215
7216
7217
7218
                    query_layer,
                    key_layer,
                    value_layer,
                    cu_seqlens_q,
                    cu_seqlens_kv,
                    max_seqlen_q,
                    max_seqlen_kv,
7219
7220
                    cu_seqlens_q_padded,
                    cu_seqlens_kv_padded,
7221
                    self.attention_dropout if self.training else 0.0,
7222
7223
7224
                    cp_group,
                    cp_global_ranks,
                    cp_stream,
7225
                    cp_comm_type,
7226
                    softmax_scale=self.softmax_scale,
7227
                    qkv_format=qkv_format,
7228
                    attn_mask_type=attn_mask_type,
7229
7230
                    attn_bias_type=core_attention_bias_type,
                    attn_bias=core_attention_bias,
7231
                    deterministic=self.deterministic,
7232
                    use_fused_attention=True,
7233
                    window_size=window_size,
7234
7235
                    fp8=fp8,
                    fp8_meta=fp8_meta,
7236
7237
                )
        else:
7238
7239
7240
7241
7242
7243
7244
            with self.attention_dropout_ctx():
                output = FusedAttnFunc.apply(
                    self.training,
                    max_seqlen_q,
                    max_seqlen_kv,
                    cu_seqlens_q,
                    cu_seqlens_kv,
7245
7246
                    cu_seqlens_q_padded,
                    cu_seqlens_kv_padded,
7247
7248
7249
7250
7251
7252
7253
7254
7255
7256
7257
                    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,
7258
                    window_size,
7259
7260
7261
7262
7263
                    None,  # rng_gen
                    fused_attention_backend,
                    use_FAv2_bwd,
                    fp8,
                    fp8_meta,
7264
                    self.deterministic,
7265
                )
7266

7267
7268
        # ...hd -> ...(hd)
        return output.view(*output.shape[:-2], -1)
7269
7270


7271
class DotProductAttention(TransformerEngineBaseModule):
7272
7273
7274
7275
7276
7277
    """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::

7278
        Argument :attr:`attention_mask` in the `forward` call is only used when
7279
        :attr:`attn_mask_type` includes '"padding"' or `"arbitrary"`.
7280
7281
7282

    .. warning::

7283
        FlashAttention uses a non-deterministic algorithm for optimal performance. To observe
7284
        deterministic behavior at the cost of performance, use FlashAttention version >= `2.4.1`
7285
7286
        and set the environment variable :attr:`NVTE_ALLOW_NONDETERMINISTIC_ALGO=0`. In order
        to disable`flash-attn` entirely, set :attr:`NVTE_FLASH_ATTN=0`.
7287

7288
7289
7290
7291
7292
7293
7294
    .. note::

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


7295
7296
7297
7298
    Parameters
    ----------
    num_attention_heads : int
                         number of attention heads in the transformer layer.
7299
7300
7301
    kv_channels : Union[int, Tuple[int, int]]
                the head size in key and value tensors. If the same, :attr:`kv_channels` can be
                an integer; if not, :attr:`kv_channels` should be a tuple of two integers.
7302
7303
7304
7305
7306
7307
7308
7309
    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`.
7310
7311
    attention_dropout: float, default = 0.0
                      dropout probability for the dropout op during multi-head attention.
7312
    attn_mask_type: str, default = `causal`
7313
                   type of attention mask passed into softmax operation, options are "`no_mask`",
7314
7315
7316
7317
7318
7319
7320
7321
7322
                   "`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
7323
                   "`padding_causal`" and "`padding_causal_bottom_right`", Transformer Engine
7324
7325
7326
7327
7328
7329
7330
7331
7332
7333
7334
7335
7336
7337
                   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].
7338
7339
7340
7341
    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
7342
7343
7344
                window and causal mask specifically. Both `causal` and `causal_bottom_right` masks
                map to `window_size = (-1, 0)` and Transformer Engine distinguishes them based on
                `attn_mask_type`. Similar to :attr:`attn_mask_type`, `window_size` can
7345
                be overridden by :attr:`window_size` in `forward` as well.
7346
7347
    attention_type: str, default = `self`
                   type of attention, either "`self`" and "`cross`".
7348
7349
7350
    layer_number: int, default = `None`
                 layer number of the current `DotProductAttention` when multiple such modules
                 are concatenated, for instance in consecutive transformer blocks.
7351
7352
7353
    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,
7354
               `h` the number of heads, `d` head size, and `t` the total number of tokens
7355
7356
7357
7358
7359
               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.
7360
               For that, please use `get_qkv_layout` to gain the layout information.
7361
7362
    softmax_scale: Optional[float], default = `None`
                softmax scale for the attention scores. If `None`, defaults to
7363
                `1.0/math.sqrt(kv_channels if isinstance(kv_channels, int) else kv_channels[0])`.
7364
7365
7366
7367
7368
7369
7370
7371
7372

    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.
7373
    cp_group : Union[ProcessGroup, List[ProcessGroup]], default = `None`
7374
              context parallel process group.
7375
7376
7377
              ProcessGroup is for cp_comm_type of "p2p", "all_gather", and "a2a".
              List[ProcessGroup] is for cp_comm_type of "a2a+p2p", where cp_group[0]
              and cp_group[1] are for a2a and p2p communications respectively.
7378
7379
7380
7381
7382
7383
7384
    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.
7385
    cp_comm_type : str, default = `p2p`
7386
                  inter-gpu communication type for context parallelism.
7387
                  Can be "p2p" or "all_gather" or "a2a" or "a2a+p2p".
7388
7389
7390
7391
7392
7393
                  "p2p": Exchange KV chunks with P2P communications in ring topology.
                         P2P is async and can be overlapped with attention compute.
                  "all_gather": All-gather to get full sequence of KV before attention.
                                The all-gather is not async, and cannot be overlapped.
                  "a2a": Like DeepSpeed Ulysses, scatter attention heads across the CP
                         group, and gather to get full sequence of QKV.
7394
7395
7396
                  "a2a+p2p": hierarchical CP implementation. First applying a2a to QKV
                  across each CP sub-group (e.g., via NVLink), then exchanging KV with
                  p2p between sub-groups (e.g., via IBLink).
7397
7398
7399
7400
7401
    """

    def __init__(
        self,
        num_attention_heads: int,
7402
        kv_channels: Union[int, Tuple[int, int]],
7403
        num_gqa_groups: Optional[int] = None,
7404
        attention_dropout: float = 0.0,
7405
        qkv_format: str = "sbhd",
7406
        attn_mask_type: str = "causal",
7407
        window_size: Optional[Tuple[int, int]] = None,
7408
7409
7410
7411
7412
        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,
7413
        attention_type: str = "self",
7414
        cp_group: Optional[Union[dist_group_type, List[dist_group_type]]] = None,
7415
        cp_global_ranks: List[int] = None,
7416
        cp_stream: torch.cuda.Stream = None,
7417
        cp_comm_type: str = "p2p",
7418
        softmax_scale: Optional[float] = None,
7419
7420
7421
    ) -> None:
        super().__init__()

7422
        self.logger = logging.getLogger("DotProductAttention")
7423
7424
7425
        self.logger.setLevel(_log_level)
        if not self.logger.hasHandlers():
            self.logger.addHandler(_stream_handler)
7426
        self.qkv_format = qkv_format
7427
        attn_mask_type = attn_mask_type.replace(",", "_")
7428
7429
        if attn_mask_type == "causal_padding":
            attn_mask_type = "padding_causal"
7430
        self.attn_mask_type = attn_mask_type
7431
        self.window_size = check_set_window_size(attn_mask_type, window_size)
7432
7433
7434
7435
7436
7437
7438
        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)
7439
        self.get_rng_state_tracker = get_rng_state_tracker
7440
        self.num_attention_heads = num_attention_heads
7441
        self.layer_number = 1 if layer_number is None else layer_number
7442
7443
7444
        self.cp_group = cp_group
        self.cp_global_ranks = cp_global_ranks
        self.cp_stream = cp_stream
7445
        self.cp_comm_type = cp_comm_type
7446

7447
7448
7449
7450
7451
7452
        self.hidden_size_per_attention_head_k = (
            kv_channels if isinstance(kv_channels, int) else kv_channels[0]
        )
        self.hidden_size_per_attention_head_v = (
            kv_channels if isinstance(kv_channels, int) else kv_channels[1]
        )
7453

7454
        self.num_gqa_groups = num_attention_heads if num_gqa_groups is None else num_gqa_groups
7455
        self.num_gqa_groups_per_partition = int(self.num_gqa_groups // self.tp_size)
7456

7457
7458
7459
        assert (
            num_attention_heads % self.num_gqa_groups == 0
        ), "The number of attention heads must be divisible by the number of GQA groups!"
7460

7461
        self.rng_states_tracker = None
7462
7463
7464
        if sequence_parallel or get_rng_state_tracker is None:
            attention_dropout_ctx = nullcontext
        else:
7465
7466
7467
            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
7468

7469
        if softmax_scale is None:
7470
7471
7472
            softmax_scale = 1.0 / math.sqrt(
                kv_channels if isinstance(kv_channels, int) else kv_channels[0]
            )
7473

7474
7475
7476
        self.deterministic = (
            not bool(int(os.getenv("NVTE_ALLOW_NONDETERMINISTIC_ALGO", "1")))
            or torch.are_deterministic_algorithms_enabled()
7477
        )
7478
7479
7480
7481
7482
7483
7484
7485
7486
7487
7488
7489
7490
7491
7492
7493
7494
7495
7496
        # To use the workspace optimization path for determinism, please
        # set NVTE_FUSED_ATTN_FORCE_WORKSPACE_OPT=1 for cuDNN >=8.9.5 and <9.0.0,
        # and set NVTE_ALLOW_NONDETERMINISTIC_ALGO=0 for cuDNN >=9.0.0.
        cudnn_version = get_cudnn_version()
        if (8, 9, 5) <= cudnn_version < (9, 0, 0):
            if self.deterministic:
                os.environ["NVTE_FUSED_ATTN_FORCE_WORKSPACE_OPT"] = "1"

            # CUDNN_FRONTEND_ATTN_DP_WORKSPACE_LIMIT
            # - unset:       enables workspace optimization when required workspace is <= 256MB
            #                or when bias gradient needs to be computed
            # - n:           enables workspace optimization when required workspace is <= n bytes
            # - -1:          enables workspace optimization always
            # - 0:           disables workspace optimization always
            if "NVTE_FUSED_ATTN_FORCE_WORKSPACE_OPT" in os.environ:
                if os.environ["NVTE_FUSED_ATTN_FORCE_WORKSPACE_OPT"] == "0":
                    os.environ["CUDNN_FRONTEND_ATTN_DP_WORKSPACE_LIMIT"] = "0"
                if os.environ["NVTE_FUSED_ATTN_FORCE_WORKSPACE_OPT"] == "1":
                    os.environ["CUDNN_FRONTEND_ATTN_DP_WORKSPACE_LIMIT"] = "-1"
7497

7498
        assert attention_type in AttnTypes, f"attention_type {attention_type} not supported"
7499
7500
7501
7502

        self.attention_type = attention_type
        self.attention_dropout = attention_dropout

7503
7504
7505
7506
7507
        attn_kwargs = {
            "attention_dropout": attention_dropout,
            "attention_dropout_ctx": attention_dropout_ctx,
        }

7508
7509
7510
7511
7512
7513
7514
        self.flash_attention = FlashAttention(
            softmax_scale,
            attention_type=attention_type,
            layer_number=layer_number,
            deterministic=self.deterministic,
            **attn_kwargs,
        )
7515

7516
        # Instantiating three types since use of flash-attn and FusedAttention
7517
        # might be ruled out due to forward inputs.
7518
7519
7520
7521
7522
7523
7524
        self.fused_attention = FusedAttention(
            softmax_scale,
            attention_type=attention_type,
            layer_number=layer_number,
            deterministic=self.deterministic,
            **attn_kwargs,
        )
7525

7526
        self.unfused_attention = UnfusedDotProductAttention(
7527
7528
7529
7530
            softmax_scale,
            attention_type=attention_type,
            **attn_kwargs,
            layer_number=layer_number,
7531
        )
7532

7533
7534
7535
        def remove_extra_states_check(self, incompatible_keys):  # pylint: disable=unused-argument
            """
            Temporarily remove core_attention._extra_state as a missing key
7536
7537
            when loading older Transformer Engine checkpoints. Will phase out
            this hook in Transformer Engine 2.0.
7538
7539
7540
7541
7542
7543
7544
            """
            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)

7545
7546
7547
7548
7549
7550
7551
7552
7553
7554
7555
7556
7557
7558
7559
7560
7561
7562
7563
7564
7565
7566
    def _load_from_state_dict(
        self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs
    ):
        """
        This function helps to load Transformer Engine 1.6 and 1.7 checkpoints, where FP8 attention
        metadata is stored under the `core_attention.fused_attention._extra_state` key and not the
        `core_attention._extra_state` key. Please see `FP8 checkpoint compatibility
        <https://docs.nvidia.com/deeplearning/transformer-engine/user-guide/faq.html#fp8-checkpoint-compatibility>`_ for more details.
        """
        fused_attn_key = False
        dot_product_attn_key = False
        for k in state_dict.keys():
            if "core_attention.fused_attention._extra_state" in k:
                fused_attn_key = True
            if "core_attention._extra_state" in k:
                dot_product_attn_key = True
        if fused_attn_key and not dot_product_attn_key:
            prefix = prefix + "fused_attention."
        super()._load_from_state_dict(
            state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs
        )

7567
7568
7569
7570
    def _checkpointed_attention_forward(
        self,
        attention_func: Callable,
        *forward_args: Tuple[torch.Tensor, ...],
7571
        **forward_kwargs: Dict[str, Any],
7572
7573
7574
    ) -> torch.Tensor:
        """Forward method with activation checkpointing."""

7575
7576
        def custom_forward(*input_args, **input_kwargs):
            return attention_func(*input_args, **input_kwargs)
7577
7578
7579

        hidden_states = checkpoint(
            custom_forward,
7580
7581
7582
            distribute_saved_activations=False,
            get_rng_state_tracker=self.get_rng_state_tracker,
            tp_group=self.tp_group,
7583
            *forward_args,
7584
            **forward_kwargs,
7585
7586
7587
7588
        )

        return hidden_states

7589
7590
    def set_context_parallel_group(
        self,
7591
        cp_group: Union[dist_group_type, List[dist_group_type], None],
7592
7593
        cp_global_ranks: List[int],
        cp_stream: torch.cuda.Stream,
7594
        cp_comm_type: str = "p2p",
7595
    ) -> None:
7596
7597
7598
7599
7600
7601
        """
        Set the context parallel attributes for the given
        module before executing the forward pass.

        Parameters
        ----------
7602
        cp_group : Union[ProcessGroup, List[ProcessGroup]]
7603
                  context parallel process group.
7604
7605
7606
                  ProcessGroup is for cp_comm_type of "p2p", "all_gather", and "a2a".
                  List[ProcessGroup] is for cp_comm_type of "a2a+p2p", where cp_group[0]
                  and cp_group[1] are for a2a and p2p communications respectively.
7607
7608
7609
7610
        cp_global_ranks : List[int]
                         list of global ranks in the context group.
        cp_stream : torch.cuda.Stream
                   cuda stream for context parallel execution.
7611
        cp_comm_type : str, default = `p2p`
7612
                      inter-gpu communication type for context parallelism.
7613
                      Can be "p2p" or "all_gather" or "a2a" or "a2a+p2p".
7614
7615
7616
7617
7618
7619
                      "p2p": Exchange KV chunks with P2P communications in ring topology.
                             P2P is async and can be overlapped with attention compute.
                      "all_gather": All-gather to get full sequence of KV before attention.
                                    The all-gather is not async, and cannot be overlapped.
                      "a2a": Like DeepSpeed Ulysses, scatter attention heads across the CP
                             group, and gather to get full sequence of QKV.
7620
7621
7622
                      "a2a+p2p": hierarchical CP implementation. First applying a2a to QKV
                      across each CP sub-group (e.g., via NVLink), then exchanging KV with
                      p2p between sub-groups (e.g., via IBLink).
7623
        """
7624
7625
7626
        self.cp_group = cp_group
        self.cp_global_ranks = cp_global_ranks
        self.cp_stream = cp_stream
7627
        self.cp_comm_type = cp_comm_type
7628

7629
    @no_torch_dynamo(recursive=False)
7630
7631
7632
7633
7634
    def forward(
        self,
        query_layer: torch.Tensor,
        key_layer: torch.Tensor,
        value_layer: torch.Tensor,
7635
        attention_mask: Optional[Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]] = None,
7636
7637
7638
        qkv_format: Optional[str] = None,
        cu_seqlens_q: Optional[torch.Tensor] = None,
        cu_seqlens_kv: Optional[torch.Tensor] = None,
7639
7640
        cu_seqlens_q_padded: Optional[torch.Tensor] = None,
        cu_seqlens_kv_padded: Optional[torch.Tensor] = None,
7641
7642
        max_seqlen_q: Optional[int] = None,
        max_seqlen_kv: Optional[int] = None,
7643
        attn_mask_type: Optional[str] = None,
7644
        window_size: Optional[Tuple[int, int]] = None,
7645
        checkpoint_core_attention: bool = False,
7646
7647
        core_attention_bias_type: str = "no_bias",
        core_attention_bias: Optional[torch.Tensor] = None,
7648
        alibi_slopes: Optional[torch.Tensor] = None,
7649
        fast_zero_fill: bool = True,
7650
        inference_params: Optional[InferenceParams] = None,
7651
        is_first_microbatch: Optional[bool] = None,
7652
7653
7654
7655
7656
7657
    ) -> torch.Tensor:
        """
        Dot Product Attention Layer.

        .. note::

7658
7659
            Argument :attr:`attention_mask` is only used when :attr:`attn_mask_type`
            includes '"padding"' or `"arbitrary"`.
7660

7661
7662
        .. note::

7663
7664
7665
7666
7667
7668
7669
7670
7671
7672
7673
7674
7675
            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,
7676
            and FusedAttention backend if applicable, to use. Transformer Engine prioritizes
7677
7678
7679
7680
            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
7681
7682
            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
7683
            optimizations in FusedAttention. When unset, Transformer Engine determines the code path
7684
7685
            based on its internal logic. These optimizations trade memory for performance
            and should be used with care.
7686

7687
7688
7689
7690
7691
7692
7693
7694
7695
7696
7697
7698
7699
7700
7701
7702
7703
7704
7705
7706
7707
7708
7709
7710
7711
7712
7713
7714
7715
7716
7717
7718
7719
7720
7721
7722
7723
7724
7725
7726
7727
7728
7729
7730
7731
7732
7733
7734
7735
7736
7737
7738
7739
7740
        .. note::
            .. _cu_seqlens note:

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

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

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

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

        .. note::
            .. _max_seqlen note:

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

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

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

7741
7742
7743
7744
7745
7746
7747
7748
        Parameters
        ----------
        query_layer : torch.Tensor
                     Query tensor.
        key_layer : torch.Tensor
                   Key tensor.
        value_layer : torch.Tensor
                     Value tensor.
7749
7750
        attention_mask: Optional[Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]],
             default = `None`. Boolean tensor(s) used to mask out attention softmax input.
7751
             It should be `None` for causal masks and "`no_mask`". For padding masks, it should be
7752
7753
             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]
7754
7755
7756
7757
             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.
7758
7759
7760
        qkv_format: str, default = `None`
                   If provided, overrides :attr:`qkv_format` from initialization.
        cu_seqlens_q: Optional[torch.Tensor], default = `None`
7761
                   Cumulative sum of sequence lengths (without offset) in a batch for `query_layer`,
7762
                   with shape [batch_size + 1] and dtype torch.int32.
7763
                   See :ref:`note<cu_seqlens note>` for more details.
7764
        cu_seqlens_kv: Optional[torch.Tensor], default = `None`
7765
7766
                   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.
7767
                   See :ref:`note<cu_seqlens note>` for more details.
7768
7769
7770
7771
7772
        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`.
7773
                   See :ref:`note<cu_seqlens note>` for more details.
7774
7775
7776
7777
7778
        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`.
7779
                   See :ref:`note<cu_seqlens note>` for more details.
7780
7781
        max_seqlen_q: Optional[int], default = `None`
                      Maximum sequence length in `query_layer`.
7782
                      See :ref:`note<max_seqlen note>` for more details.
7783
7784
        max_seqlen_kv: Optional[int], default = `None`
                       Maximum sequence length in `key_layer` and `value_layer`.
7785
                       See :ref:`note<max_seqlen note>` for more details.
7786
7787
7788
7789
7790
7791
7792
        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.
7793
        window_size: Optional[Tuple[int, int]], default = `None`
7794
                    Sliding window size for local attention.
7795
7796
7797
7798
7799
        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.
7800
        core_attention_bias_type: str, default = `no_bias`
7801
                    Bias type, {`no_bias`, `pre_scale_bias`, `post_scale_bias`, `alibi`}
7802
        core_attention_bias: Optional[torch.Tensor], default = `None`
7803
7804
                    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.
7805
7806
7807
7808
        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.
7809
        fast_zero_fill: bool, default = `True`
7810
                    Whether to use the fast path to set output tensors to 0 or not.
7811
7812
7813
7814
7815
7816
7817
7818
7819
7820
        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.
7821
7822
7823
7824
7825
7826
7827
7828
7829
7830
7831
7832
7833
        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)
7834
        """
7835
7836
7837
7838
7839
7840
7841
7842
7843
7844
7845
        with self.prepare_forward(
            query_layer,
            is_first_microbatch,
            num_gemms=3,
            allow_non_contiguous=True,
        ) as query_layer:

            if self.fp8:
                if self.fp8_meta["recipe"].fp8_mha:
                    if not self.fp8_meta["recipe"].fp8_dpa:
                        self.fp8_meta["recipe"].fp8_dpa = True
7846
                        self.logger.warning(
7847
7848
7849
                            """Forcing fp8_meta["recipe"].fp8_dpa=True due to """
                            """fp8_meta["recipe"].fp8_mha=True"""
                        )
7850
7851
7852
7853
7854
7855
7856
7857
7858
7859
7860

            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."""
7861

7862
7863
7864
            assert (
                query_layer.is_cuda and key_layer.is_cuda and value_layer.is_cuda
            ), "DotProductAttention only supports CUDA tensors."
7865
7866
7867
            assert (
                query_layer.dtype == key_layer.dtype and query_layer.dtype == value_layer.dtype
            ), "Queries, keys and values must have the same data type!"
7868
7869
7870
            assert (
                key_layer.shape[:-1] == value_layer.shape[:-1]
            ), "Keys and values must have the same batch size, sequence length and number of heads!"
7871
7872
7873
7874
7875
7876
7877
7878
            assert (
                key_layer.shape[-1] == self.hidden_size_per_attention_head_k
            ), f"Keys have head_dim = {key_layer.shape[-1]}, "
            "but expected head_dim = {self.hidden_size_per_attention_head_k}!"
            assert (
                value_layer.shape[-1] == self.hidden_size_per_attention_head_v
            ), f"Values have head_dim = {value_layer.shape[-1]}, "
            "but expected head_dim = {self.hidden_size_per_attention_head_v}!"
7879

7880
7881
7882
7883
7884
7885
            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"
7886
            assert (
7887
7888
7889
7890
7891
7892
                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!"
7893

7894
7895
7896
7897
            if window_size is None:
                window_size = self.window_size
            window_size = check_set_window_size(attn_mask_type, window_size)

7898
7899
7900
7901
7902
7903
7904
            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."
7905

7906
7907
            if qkv_format is None:
                qkv_format = self.qkv_format
7908

7909
7910
            if inference_params is not None:
                assert self.layer_number is not None, "Layer number must be set!"
7911

7912
7913
7914
7915
7916
                # convert causal to causal_bottom_right in inference when KV-caching is in use
                # so users can run with the same attn_mask_type for training and inference
                if attn_mask_type in ["causal", "padding_causal"]:
                    attn_mask_type = attn_mask_type + "_bottom_right"

7917
7918
7919
                if qkv_format == "bshd":
                    key_layer = key_layer.transpose(0, 1)
                    value_layer = value_layer.transpose(0, 1)
7920

7921
7922
7923
7924
                (
                    inference_key_memory,
                    inference_value_memory,
                ) = inference_params.key_value_memory_dict[self.layer_number]
7925

7926
7927
7928
                batch_start = inference_params.batch_size_offset
                batch_end = batch_start + key_layer.size(1)
                assert batch_end <= inference_key_memory.size(1)
7929

7930
7931
7932
                sequence_start = inference_params.sequence_len_offset
                sequence_end = sequence_start + key_layer.size(0)
                assert sequence_end <= inference_key_memory.size(0)
7933

7934
7935
7936
7937
7938
7939
7940
7941
7942
                # 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, ...]
7943

7944
7945
7946
                if qkv_format == "bshd":
                    key_layer = key_layer.transpose(0, 1)
                    value_layer = value_layer.transpose(0, 1)
7947

7948
7949
                key_layer = key_layer.contiguous()
                value_layer = value_layer.contiguous()
7950
7951

            assert (
7952
7953
7954
7955
7956
7957
7958
7959
7960
7961
                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":
7962
                assert all(
7963
7964
7965
7966
7967
7968
7969
7970
7971
7972
7973
7974
7975
                    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!"
7976
                batch_size = len(cu_seqlens_q) - 1
7977
                if max_seqlen_q is None:
7978
7979
7980
7981
                    if cu_seqlens_q_padded is not None:
                        seqlens_q = cu_seqlens_q_padded[1:] - cu_seqlens_q_padded[:-1]
                    else:
                        seqlens_q = cu_seqlens_q[1:] - cu_seqlens_q[:-1]
7982
                    max_seqlen_q = int((seqlens_q.max().item() + 63) // 64 * 64)
7983
                if max_seqlen_kv is None:
7984
7985
7986
7987
                    if cu_seqlens_kv_padded is not None:
                        seqlens_kv = cu_seqlens_kv_padded[1:] - cu_seqlens_kv_padded[:-1]
                    else:
                        seqlens_kv = cu_seqlens_kv[1:] - cu_seqlens_kv[:-1]
7988
                    max_seqlen_kv = int((seqlens_kv.max().item() + 63) // 64 * 64)
7989

7990
7991
7992
7993
7994
7995
            cp_size = 1
            if isinstance(self.cp_group, dist_group_type):
                cp_size = get_distributed_world_size(self.cp_group)
            elif isinstance(self.cp_group, list):
                for group in self.cp_group:
                    cp_size *= get_distributed_world_size(group)
7996
7997
            context_parallel = cp_size > 1

7998
            if qkv_format in ["sbhd", "bshd"]:
7999
                assert all(
8000
8001
8002
                    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":
8003
8004
                    max_seqlen_q = query_layer.shape[0] if max_seqlen_q is None else max_seqlen_q
                    max_seqlen_kv = key_layer.shape[0] if max_seqlen_kv is None else max_seqlen_kv
8005
                    batch_size = query_layer.shape[1]
8006
                else:
8007
8008
                    max_seqlen_q = query_layer.shape[1] if max_seqlen_q is None else max_seqlen_q
                    max_seqlen_kv = key_layer.shape[1] if max_seqlen_kv is None else max_seqlen_kv
8009
                    batch_size = query_layer.shape[0]
8010
8011
                max_seqlen_q *= cp_size
                max_seqlen_kv *= cp_size
8012
8013
8014
8015
8016
                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
8017
                        the sequence dimension in 'query_layer'!"""
8018
8019
8020
8021
8022
                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
8023
                        the sequence dimension in 'key_layer' and 'value_layer'!"""
8024
8025
8026
8027
8028
                if cu_seqlens_q is None or cu_seqlens_kv is None:
                    if "padding" in attn_mask_type:
                        assert (
                            attention_mask is not None
                        ), "Please provide attention_mask for padding!"
8029
                        if self.attention_type == "self":
8030
8031
8032
8033
8034
8035
8036
8037
8038
8039
8040
8041
8042
8043
8044
8045
                            cu_seqlens_q = get_cu_seqlens(attention_mask)
                            cu_seqlens_kv = cu_seqlens_q
                        else:
                            cu_seqlens_q = get_cu_seqlens(attention_mask[0])
                            cu_seqlens_kv = get_cu_seqlens(attention_mask[1])
                    else:
                        cu_seqlens_q = _get_full_cu_seqlens(
                            batch_size,
                            max_seqlen_q,
                            query_layer.device,
                        )
                        cu_seqlens_kv = _get_full_cu_seqlens(
                            batch_size,
                            max_seqlen_kv,
                            key_layer.device,
                        )
8046

8047
8048
8049
8050
8051
            if (
                isinstance(query_layer, Float8Tensor)
                and isinstance(key_layer, Float8Tensor)
                and isinstance(value_layer, Float8Tensor)
            ):
8052
                qkv_layout, query_layer._data, key_layer._data, value_layer._data = get_qkv_layout(
8053
8054
8055
                    query_layer._data, key_layer._data, value_layer._data, qkv_format=qkv_format
                )
            else:
8056
                qkv_layout, query_layer, key_layer, value_layer = get_qkv_layout(
8057
8058
                    query_layer, key_layer, value_layer, qkv_format=qkv_format
                )
8059

8060
8061
8062
8063
8064
8065
8066
8067
            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
8068
            bottom_right_alignment = (attn_mask_type not in ["causal", "padding_causal"],)
8069
8070
8071
8072
8073
8074
8075
8076
            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
8077
                    or _alibi_cache["_bottom_right_alignment"] != bottom_right_alignment
8078
8079
8080
8081
8082
                    or _alibi_cache["_alibi_slopes"] is None
                ):
                    _alibi_cache["_alibi_slopes_require_update"] = True
                    _alibi_cache["_alibi_bias_require_update"] = True

8083
8084
            core_attention_bias_shape = None
            if core_attention_bias is not None:
8085
                if (
8086
8087
                    core_attention_bias.shape[0] == batch_size
                    and core_attention_bias.shape[1] == query_layer.shape[-2]
8088
                ):
8089
8090
8091
8092
8093
8094
8095
8096
8097
8098
8099
8100
8101
8102
8103
8104
8105
8106
8107
8108
8109
8110
8111
8112
                    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)
            )
8113

8114
            attention_params = AttentionParams(
8115
8116
8117
8118
8119
8120
8121
8122
                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,
8123
8124
                head_dim_qk=query_layer.shape[-1],
                head_dim_v=value_layer.shape[-1],
8125
8126
8127
8128
8129
8130
8131
8132
8133
8134
8135
                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,
8136
8137
                deterministic=self.deterministic,
                is_training=self.training,
8138
8139
8140
                fp8=self.fp8,
                fp8_meta=self.fp8_meta,
            )
8141
            global _attention_backends, _use_flash_attn_3
8142
8143
8144
8145
8146
8147
8148
            if (
                _attention_backends["attention_params"] is None
                or attention_params != _attention_backends["attention_params"]
            ):
                _attention_backends["attention_params"] = attention_params
                _attention_backends["backend_selection_requires_update"] = True
            if _attention_backends["backend_selection_requires_update"]:
8149
                _use_flash_attn_3 = _flash_attn_3_is_installed
8150
8151
8152
8153
8154
8155
8156
8157
                (
                    use_flash_attention,
                    use_fused_attention,
                    fused_attention_backend,
                    use_unfused_attention,
                    _,
                ) = get_attention_backend(attention_params)
                if use_flash_attention:
8158
8159
                    self.logger.info(
                        "Running with FlashAttention backend (version %s)",
8160
                        _flash_attn_version if not _use_flash_attn_3 else _flash_attn_3_version,
8161
                    )
8162
8163
8164
8165
                elif use_fused_attention:
                    self.logger.info(
                        "Running with FusedAttention backend (sub-backend %s)",
                        int(fused_attention_backend),
8166
                    )
8167
8168
8169
8170
8171
8172
8173
                elif use_unfused_attention:
                    self.logger.info("Running with UnfusedDotProductAttention backend")
            else:
                use_flash_attention = _attention_backends["use_flash_attention"]
                use_fused_attention = _attention_backends["use_fused_attention"]
                fused_attention_backend = _attention_backends["fused_attention_backend"]
                use_unfused_attention = _attention_backends["use_unfused_attention"]
8174

8175
8176
8177
8178
8179
8180
8181
8182
8183
8184
8185
8186
8187
8188
8189
8190
8191
8192
8193
8194
8195
8196
            if use_flash_attention:
                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,
8197
                    cp_comm_type=self.cp_comm_type,
8198
8199
                    max_seqlen_q=max_seqlen_q,
                    max_seqlen_kv=max_seqlen_kv,
8200
8201
                    fp8=self.fp8 and self.fp8_meta["recipe"].fp8_dpa,
                    fp8_meta=self.fp8_meta,
8202
                )
8203

8204
            if use_fused_attention:
8205
8206
                fu_core_attention_bias_type = core_attention_bias_type
                fu_core_attention_bias = core_attention_bias
8207
8208
8209
                if core_attention_bias_type == "alibi" and (
                    alibi_slopes is not None or max_seqlen_q != max_seqlen_kv
                ):
8210
8211
8212
8213
8214
8215
8216
                    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,
8217
                        bottom_right_alignment=attn_mask_type not in ["causal", "padding_causal"],
8218
                    )
8219
8220
8221
8222
8223
8224
8225
8226
8227
                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,
8228
8229
                        cu_seqlens_q_padded=cu_seqlens_q_padded,
                        cu_seqlens_kv_padded=cu_seqlens_kv_padded,
8230
8231
8232
8233
                        max_seqlen_q=max_seqlen_q,
                        max_seqlen_kv=max_seqlen_kv,
                        attn_mask_type=attn_mask_type,
                        attention_mask=attention_mask,
8234
                        window_size=window_size,
8235
8236
8237
8238
8239
8240
8241
                        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,
8242
                        cp_comm_type=self.cp_comm_type,
8243
8244
8245
8246
                        fp8=self.fp8 and self.fp8_meta["recipe"].fp8_dpa,
                        fp8_meta=self.fp8_meta,
                    )
                return self.fused_attention(
8247
8248
8249
8250
8251
8252
                    query_layer,
                    key_layer,
                    value_layer,
                    qkv_layout=qkv_layout,
                    cu_seqlens_q=cu_seqlens_q,
                    cu_seqlens_kv=cu_seqlens_kv,
8253
8254
                    cu_seqlens_q_padded=cu_seqlens_q_padded,
                    cu_seqlens_kv_padded=cu_seqlens_kv_padded,
8255
8256
                    max_seqlen_q=max_seqlen_q,
                    max_seqlen_kv=max_seqlen_kv,
8257
8258
                    attn_mask_type=attn_mask_type,
                    attention_mask=attention_mask,
8259
                    window_size=window_size,
8260
                    fused_attention_backend=fused_attention_backend,
8261
8262
                    core_attention_bias_type=fu_core_attention_bias_type,
                    core_attention_bias=fu_core_attention_bias,
8263
8264
8265
8266
                    fast_zero_fill=fast_zero_fill,
                    cp_group=self.cp_group,
                    cp_global_ranks=self.cp_global_ranks,
                    cp_stream=self.cp_stream,
8267
                    cp_comm_type=self.cp_comm_type,
8268
8269
                    fp8=self.fp8 and self.fp8_meta["recipe"].fp8_dpa,
                    fp8_meta=self.fp8_meta,
8270
                )
8271

8272
            from .cpu_offload import CPUOffloadEnabled
8273

8274
8275
8276
8277
8278
            if CPUOffloadEnabled:
                warnings.warn(
                    "Attention activation Offloading is only implemented"
                    "with Flash Attention and Fused Attention!"
                )
8279

8280
            if use_unfused_attention:
8281
8282
8283
8284
8285
8286
                if window_size is not None and (
                    window_size[0] != -1 or window_size[1] not in [-1, 0]
                ):
                    attn_mask_type, attention_mask = get_swa_mask(
                        window_size, max_seqlen_q, max_seqlen_kv, attn_mask_type, attention_mask
                    )
8287
8288
8289
8290
8291
8292
8293
8294
8295
8296
8297
8298
8299
8300
8301
8302
                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(
8303
8304
8305
                    query_layer,
                    key_layer,
                    value_layer,
8306
8307
8308
8309
8310
8311
8312
8313
8314
                    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,
                )
8315

8316
            raise ValueError("No dot product attention support for the provided inputs!")
8317
8318


8319
8320
8321
8322
8323
8324
8325
class MultiheadAttention(torch.nn.Module):
    r"""
    Multi-head Attention (MHA), including Query,
    Key, Value and Output projection.

    .. note::

8326
8327
        Argument :attr:`attention_mask` in the `forward` call is only used when
        :attr:`attn_mask_type` includes '"padding"' or `"arbitrary"`.
8328

8329
8330
8331
8332
8333
8334
8335
8336
8337
8338
8339
8340
8341
8342
8343
8344
8345
8346
8347
8348
8349
8350
8351
8352
8353
    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.
8354
8355
    attn_mask_type: {'no_mask', 'padding', 'causal', 'padding_causal', 'causal_bottom_right',
                   'padding_causal_bottom_right','arbitrary'},
8356
                   default = `causal`
8357
8358
8359
8360
8361
                   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.
8362
8363
8364
8365
    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
8366
8367
8368
                window and causal mask specifically. Both `causal` and `causal_bottom_right` masks
                map to `window_size = (-1, 0)` and Transformer Engine distinguishes them based on
                `attn_mask_type`. Similar to :attr:`attn_mask_type`, `window_size` can
8369
                be overridden by :attr:`window_size` in `forward` as well.
8370
8371
8372
8373
8374
8375
8376
8377
8378
8379
8380
8381
8382
    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.
8383
8384
    input_layernorm: bool, default = `False`
                     if set to `True`, layer normalization to the input is applied.
8385
8386
8387
8388
8389
8390
8391
8392
8393
8394
8395
8396
8397
8398
8399
8400
8401
8402
8403
8404
    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"
8405
          The device on which the parameters of the model will be allocated. It is the user's
8406
8407
          responsibility to ensure all parameters are moved to the GPU before running the
          forward pass.
8408
8409
8410
8411
8412
8413
8414
    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.
8415
            For that, please use `get_qkv_layout` to gain the layout information.
8416
8417
8418
8419
8420
8421
8422
8423
8424
8425
8426
8427
8428
8429
8430
8431
8432
8433
8434
8435
8436
8437
8438
8439
8440
8441
8442
8443
8444
8445
8446
8447
8448
8449
8450
8451
8452
8453
8454
8455

    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`.
8456
8457
8458
8459
8460
8461
    """

    def __init__(
        self,
        hidden_size: int,
        num_attention_heads: int,
8462
8463
8464
8465
8466
        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,
8467
        layer_number: Optional[int] = None,
8468
        attn_mask_type: str = "causal",
8469
        window_size: Optional[Tuple[int, int]] = None,
8470
8471
        tp_group: Optional[dist_group_type] = None,
        tp_size: int = 1,
8472
        num_gqa_groups: Optional[int] = None,
8473
8474
8475
        fuse_wgrad_accumulation: bool = False,
        get_rng_state_tracker: Optional[Callable] = None,
        sequence_parallel: bool = False,
8476
        params_dtype: Optional[torch.dtype] = None,
8477
        return_bias: bool = False,
8478
8479
8480
8481
8482
8483
8484
8485
8486
        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
8487
        ub_overlap_rs_dgrad: bool = False,
8488
8489
        ub_overlap_rs: bool = False,
        ub_overlap_ag: bool = False,
8490
        bias: bool = True,
8491
        normalization: str = "LayerNorm",
8492
        device: Union[torch.device, str] = "cuda",
8493
        qkv_format: str = "sbhd",
8494
8495
    ) -> None:
        super().__init__()
8496

8497
        self.qkv_format = qkv_format
8498
        self.attn_mask_type = attn_mask_type
8499
        self.window_size = check_set_window_size(attn_mask_type, window_size)
8500
        self.layer_number = layer_number
8501
8502
8503
8504
8505
        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
8506
        self.params_dtype = torch.get_default_dtype() if params_dtype is None else params_dtype
8507
        self.num_attention_heads = num_attention_heads
8508
        self.return_bias = return_bias
8509
8510
        self.cp_size = 1
        self.cp_rank = 0
8511
8512
8513
8514
8515
8516
8517

        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()
8518
8519
8520
8521
8522

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

8523
8524
8525
        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"
8526
8527
8528
8529
8530
8531

        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)
8532
8533
8534
8535
8536
8537
8538
        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!"
8539
        self.num_gqa_groups_per_partition = int(self.num_gqa_groups // tp_size)
8540
8541
8542
8543

        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
8544
8545
8546
8547
8548
8549
8550

        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,
8551
            "params_dtype": self.params_dtype,
8552
            "device": device,
8553
8554
8555
8556
        }

        qkv_parallel_mode = "column" if set_parallel_mode else None

cyanguwa's avatar
cyanguwa committed
8557
        if self.attention_type == "self":
8558
8559
            parameters_split = None
            if not fuse_qkv_params:
8560
8561
8562
8563
8564
8565
8566
                parameters_split = collections.OrderedDict(
                    [
                        ("query", self.hidden_size_q),
                        ("key", self.hidden_size_kv),
                        ("value", self.hidden_size_kv),
                    ]
                )
8567
8568
8569
            if self.input_layernorm:
                self.layernorm_qkv = LayerNormLinear(
                    hidden_size,
8570
                    self.hidden_size_q + 2 * self.hidden_size_kv,
8571
8572
8573
8574
8575
8576
                    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
8577
                    parameters_split=parameters_split,
8578
8579
8580
                    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
8581
                    ub_overlap_rs_dgrad=ub_overlap_rs_dgrad,
8582
                    ub_overlap_ag=ub_overlap_ag,
8583
                    normalization=normalization,
Przemyslaw Tredak's avatar
Przemyslaw Tredak committed
8584
                    ub_name="qkv",
8585
8586
8587
8588
8589
                    **common_gemm_kwargs,
                )
            else:
                self.qkv = Linear(
                    hidden_size,
8590
                    self.hidden_size_q + 2 * self.hidden_size_kv,
8591
8592
8593
8594
                    init_method=init_method,
                    bias=bias,
                    return_bias=False,
                    parallel_mode=qkv_parallel_mode,
cyanguwa's avatar
cyanguwa committed
8595
                    parameters_split=parameters_split,
8596
8597
                    **common_gemm_kwargs,
                )
cyanguwa's avatar
cyanguwa committed
8598
        elif self.attention_type == "cross":
8599
8600
8601
            if self.input_layernorm:
                self.layernorm_query = LayerNormLinear(
                    hidden_size,
8602
                    self.hidden_size_q,
8603
8604
8605
8606
8607
                    eps=layernorm_epsilon,
                    init_method=init_method,
                    bias=bias,
                    return_bias=False,
                    parallel_mode=qkv_parallel_mode,
8608
                    parameters_split=("query",) if not fuse_qkv_params else None,
8609
8610
8611
8612
                    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
8613
                    ub_overlap_rs_dgrad=ub_overlap_rs_dgrad,
8614
                    ub_overlap_ag=ub_overlap_ag,
8615
                    normalization=normalization,
Przemyslaw Tredak's avatar
Przemyslaw Tredak committed
8616
                    ub_name="qkv",
8617
8618
8619
8620
8621
                    **common_gemm_kwargs,
                )
            else:
                self.query_layer = Linear(
                    hidden_size,
8622
                    self.hidden_size_q,
8623
8624
8625
8626
8627
8628
8629
8630
                    init_method=init_method,
                    bias=bias,
                    return_bias=False,
                    parallel_mode=qkv_parallel_mode,
                    **common_gemm_kwargs,
                )
            self.key_value = Linear(
                hidden_size,
8631
                2 * self.hidden_size_kv,
8632
8633
8634
8635
                init_method=init_method,
                bias=bias,
                return_bias=False,
                parallel_mode=qkv_parallel_mode,
8636
                parameters_split=("key", "value") if not fuse_qkv_params else None,
8637
8638
8639
8640
8641
8642
                **common_gemm_kwargs,
            )

        # Attention.
        self.core_attention = DotProductAttention(
            num_attention_heads,
8643
            self.hidden_size_per_attention_head,
8644
8645
            num_gqa_groups=self.num_gqa_groups,
            attention_dropout=attention_dropout,
8646
            qkv_format=self.qkv_format,
8647
8648
8649
8650
            tp_size=tp_size,
            get_rng_state_tracker=get_rng_state_tracker,
            sequence_parallel=sequence_parallel,
            tp_group=tp_group,
8651
            layer_number=self.layer_number,
8652
            attention_type=self.attention_type,
8653
8654
8655
8656
        )

        # Linear
        self.proj = Linear(
8657
            self.hidden_size_q,
8658
8659
8660
            hidden_size,
            init_method=output_layer_init_method,
            bias=bias,
8661
            return_bias=return_bias,
8662
            parallel_mode="row" if set_parallel_mode else None,
8663
8664
            ub_overlap_rs=ub_overlap_rs,
            ub_overlap_ag=ub_overlap_ag,
Przemyslaw Tredak's avatar
Przemyslaw Tredak committed
8665
            ub_name="proj",
8666
8667
8668
8669
            **common_gemm_kwargs,
        )

    def _allocate_memory(
8670
        self, inference_max_sequence_len: int, batch_size: int, dtype: torch.dtype
8671
    ) -> torch.Tensor:
8672
        """Allocates memory for KV cache."""
8673
8674
8675
        return torch.empty(
            inference_max_sequence_len,
            batch_size,
8676
            self.num_gqa_groups_per_partition,
8677
            self.hidden_size_per_attention_head,
8678
            dtype=dtype,
8679
8680
8681
8682
            device=torch.cuda.current_device(),
        )

    def set_tensor_parallel_group(self, tp_group: Union[dist_group_type, None]) -> None:
8683
8684
8685
8686
8687
8688
8689
8690
8691
        """
        Set the tensor parallel group for the given
        module before executing the forward pass.

        Parameters
        ----------
        tp_group : ProcessGroup, default = `None`
                  tensor parallel process group.
        """
8692
8693
        self.tp_group = tp_group

8694
    def set_context_parallel_group(
8695
        self,
8696
        cp_group: Union[dist_group_type, List[dist_group_type], None],
8697
        cp_global_ranks: List[int],
8698
        cp_stream: torch.cuda.Stream,
8699
        cp_comm_type: str = "p2p",
8700
    ) -> None:
8701
8702
8703
8704
8705
8706
        """
        Set the context parallel attributes for the given
        module before executing the forward pass.

        Parameters
        ----------
8707
        cp_group : Union[ProcessGroup, List[ProcessGroup]]
8708
                  context parallel process group.
8709
8710
8711
                  ProcessGroup is for cp_comm_type of "p2p", "all_gather", and "a2a".
                  List[ProcessGroup] is for cp_comm_type of "a2a+p2p", where cp_group[0]
                  and cp_group[1] are for a2a and p2p communications respectively.
8712
8713
8714
8715
        cp_global_ranks : List[int]
                         list of global ranks in the context group.
        cp_stream : torch.cuda.Stream
                   cuda stream for context parallel execution.
8716
        cp_comm_type : str, default = `p2p`
8717
                      inter-gpu communication type for context parallelism.
8718
                      Can be "p2p" or "all_gather" or "a2a", "a2a+p2p".
8719
8720
8721
8722
8723
8724
                      "p2p": Exchange KV chunks with P2P communications in ring topology.
                             P2P is async and can be overlapped with attention compute.
                      "all_gather": All-gather to get full sequence of KV before attention.
                                    The all-gather is not async, and cannot be overlapped.
                      "a2a": Like DeepSpeed Ulysses, scatter attention heads across the CP
                             group, and gather to get full sequence of QKV.
8725
8726
8727
                      "a2a+p2p": hierarchical CP implementation. First applying a2a to QKV
                      across each CP sub-group (e.g., via NVLink), then exchanging KV with
                      p2p between sub-groups (e.g., via IBLink).
8728
        """
8729
8730
8731
8732
8733
8734
8735
8736
8737
8738
8739
8740
8741
8742
8743
        if isinstance(cp_group, dist_group_type):
            self.cp_size = get_distributed_world_size(cp_group)
            self.cp_rank = get_distributed_rank(cp_group)
        elif isinstance(cp_group, list):
            assert len(cp_group) == 2, "Current implementation only supports two-level CP groups!"
            assert (
                cp_comm_type == "a2a+p2p"
            ), "Only cp_comm_type of a2a+p2p requires hierarchical CP groups!"
            cp_size_a2a = get_distributed_world_size(cp_group[0])
            cp_rank_a2a = get_distributed_rank(cp_group[0])
            cp_size_p2p = get_distributed_world_size(cp_group[1])
            cp_rank_p2p = get_distributed_rank(cp_group[1])
            self.cp_size = cp_size_a2a * cp_size_p2p
            self.cp_rank = cp_size_a2a * cp_rank_p2p + cp_rank_a2a

8744
8745
8746
8747
8748
        # 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"):
8749
                child.set_context_parallel_group(cp_group, cp_global_ranks, cp_stream, cp_comm_type)
8750

8751
8752
8753
    def forward(
        self,
        hidden_states: torch.Tensor,
8754
        attention_mask: Optional[Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]] = None,
8755
        encoder_output: Optional[torch.Tensor] = None,
8756
        attn_mask_type: Optional[str] = None,
8757
        window_size: Optional[Tuple[int, int]] = None,
8758
8759
        is_first_microbatch: Optional[bool] = None,
        checkpoint_core_attention: bool = False,
8760
        inference_params: Optional[InferenceParams] = None,
8761
        rotary_pos_emb: Optional[Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]] = None,
8762
8763
        core_attention_bias_type: str = "no_bias",
        core_attention_bias: Optional[torch.Tensor] = None,
8764
        alibi_slopes: Optional[torch.Tensor] = None,
8765
8766
8767
8768
        cu_seqlens_q: Optional[torch.Tensor] = None,
        cu_seqlens_kv: Optional[torch.Tensor] = None,
        max_seqlen_q: Optional[int] = None,
        max_seqlen_kv: Optional[int] = None,
8769
        fast_zero_fill: bool = True,
8770
    ) -> Tuple[Union[torch.Tensor, None], ...]:
8771
8772
8773
8774
8775
        """
        Forward propagation for MultiheadAttention layer.

        .. note::

8776
8777
            Argument :attr:`attention_mask` is only used when :attr:`attn_mask_type`
            includes `"padding"` or `"arbitrary"`.
8778
8779
8780
8781
8782

        Parameters
        ----------
        hidden_states : torch.Tensor
             Input tensor.
8783
8784
        attention_mask: Optional[Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]],
             default = `None`. Boolean tensor(s) used to mask out attention softmax input.
8785
             It should be `None` for causal masks and "`no_mask`". For padding masks, it should be
8786
8787
             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]
8788
8789
8790
8791
8792
8793
             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'},
8794
                       default = `None`
8795
8796
8797
8798
                       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.
8799
8800
        window_size: Optional[Tuple[int, int]], default = `None`
                    sliding window size for local attention.
8801
8802
8803
8804
8805
8806
8807
8808
8809
8810
8811
8812
8813
8814
8815
8816
8817
8818
8819
8820
8821
8822
8823
8824
8825
        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`
8826
                    Bias type, {`no_bias`, `pre_scale_bias`, 'post_scale_bias`, `alibi`}
8827
        core_attention_bias: Optional[torch.Tensor], default = `None`
8828
8829
                    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.
8830
8831
8832
8833
        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.
8834
8835
8836
8837
8838
8839
8840
8841
8842
8843
8844
8845
        cu_seqlens_q: Optional[torch.Tensor], default = `None`
                   Cumulative sum of sequence lengths (without offset) in a batch for `query_layer`,
                   with shape [batch_size + 1] and dtype torch.int32.
        cu_seqlens_kv: Optional[torch.Tensor], default = `None`
                   Cumulative sum of sequence lengths (without offset) in a batch for `key_layer`
                   and `value_layer`, with shape [batch_size + 1] and dtype torch.int32.
        max_seqlen_q: Optional[int], default = `None`
                      Maximum sequence length in `query_layer`.
                      Calculated from `cu_seqlens_q` if not provided.
        max_seqlen_kv: Optional[int], default = `None`
                       Maximum sequence length in `key_layer` and `value_layer`.
                       Calculated from `cu_seqlens_kv` if not provided.
8846
8847
8848
        fast_zero_fill: bool, default = `True`
                    Whether to set output tensors to 0 or not before use.
        """
8849
8850
        # hidden_states: [sq, b, h]

8851
        if attn_mask_type is None:
8852
            attn_mask_type = self.attn_mask_type
8853
8854
        if window_size is None:
            window_size = self.window_size
8855
        window_size = check_set_window_size(attn_mask_type, window_size)
8856

8857
        if "padding" in attn_mask_type and attention_mask is not None:
8858
8859
            for mask in attention_mask:
                assert mask.dtype == torch.bool, "Attention mask must be in boolean type!"
8860

8861
8862
8863
        assert (
            core_attention_bias_type in AttnBiasTypes
        ), f"core_attention_bias_type {core_attention_bias_type} is not supported!"
8864

8865
        # =================================================
8866
        # Pre-allocate memory for key-values for inference
8867
8868
8869
        # =================================================

        if inference_params and self.layer_number is not None:
8870
8871
8872
            assert (
                self.qkv_format != "thd"
            ), "qkv_format == thd is not supported for an inference with KV-cache!"
8873
            if self.layer_number not in inference_params.key_value_memory_dict:
8874
                inf_max_seq_len = inference_params.max_sequence_length
8875
8876
                inf_max_batch_size = inference_params.max_batch_size
                inference_key_memory = self._allocate_memory(
8877
                    inf_max_seq_len, inf_max_batch_size, hidden_states.dtype
8878
8879
                )
                inference_value_memory = self._allocate_memory(
8880
                    inf_max_seq_len, inf_max_batch_size, hidden_states.dtype
8881
8882
8883
8884
8885
8886
8887
8888
8889
8890
8891
                )
                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]

8892
        # ======================
8893
        # Query, Key, and Value
8894
        # ======================
8895

8896
8897
8898
8899
8900
        fp8_mha = (
            FP8GlobalStateManager.is_fp8_enabled()
            and FP8GlobalStateManager.get_fp8_recipe().fp8_mha
        )

8901
        layernorm_output = None
cyanguwa's avatar
cyanguwa committed
8902
8903
        if self.attention_type == "self":
            # Attention heads [sq, b, h] --> [sq, b, ng * (np/ng + 2) * hn]
8904
8905
8906
8907
            if self.input_layernorm:
                layernorm_qkv_outputs = self.layernorm_qkv(
                    hidden_states,
                    is_first_microbatch=is_first_microbatch,
8908
                    fp8_output=fp8_mha and rotary_pos_emb is None,
8909
8910
8911
8912
8913
8914
8915
8916
8917
                )
                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,
8918
                    fp8_output=fp8_mha and rotary_pos_emb is None,
8919
8920
                )

8921
8922
8923
            num_queries_per_key_value = (
                self.num_attention_heads_per_partition // self.num_gqa_groups_per_partition
            )
8924
            if self.qkv_weight_interleaved:
cyanguwa's avatar
cyanguwa committed
8925
                # [sq, b, ng * (np/ng + 2) * hn] --> [sq, b, ng, (np/ng + 2), hn]
8926
                new_tensor_shape = mixed_x_layer.size()[:-1] + (
cyanguwa's avatar
cyanguwa committed
8927
8928
                    self.num_gqa_groups_per_partition,
                    (num_queries_per_key_value + 2),
8929
8930
8931
8932
                    self.hidden_size_per_attention_head,
                )
                # split along second last dimension
                split_dim = -2
cyanguwa's avatar
cyanguwa committed
8933
8934
8935
8936
8937
            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,
8938
                    self.hidden_size_per_attention_head,
cyanguwa's avatar
cyanguwa committed
8939
8940
8941
                )
                # split along third last dimension
                split_dim = -3
8942
8943
8944

            mixed_x_layer = mixed_x_layer.view(*new_tensor_shape)

cyanguwa's avatar
cyanguwa committed
8945
8946
8947
8948
8949
8950
8951
8952
8953
            # 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)
8954
                )
8955
            else:
cyanguwa's avatar
cyanguwa committed
8956
                query_layer, key_layer, value_layer = torch.split(
8957
8958
8959
8960
                    mixed_x_layer,
                    (num_queries_per_key_value, 1, 1),
                    dim=split_dim,
                )
cyanguwa's avatar
cyanguwa committed
8961

8962
8963
8964
8965
8966
8967
8968
8969
8970
8971
8972
8973
            if self.qkv_format == "thd":
                query_layer, key_layer, value_layer = (
                    x.reshape(x.size(0), -1, self.hidden_size_per_attention_head)
                    for x in (query_layer, key_layer, value_layer)
                )
            else:
                # query: -> [sq, b, np, hn]
                # key, value: -> [sq, b, ng, hn]
                query_layer, key_layer, value_layer = (
                    x.reshape(x.size(0), x.size(1), -1, self.hidden_size_per_attention_head)
                    for x in (query_layer, key_layer, value_layer)
                )
cyanguwa's avatar
cyanguwa committed
8974
8975
        elif self.attention_type == "cross":
            # Attention heads [sk, b, h] --> [sk, b, (ng * 2 * hn)]
8976
            mixed_kv_layer = self.key_value(
cyanguwa's avatar
cyanguwa committed
8977
                encoder_output,
8978
                is_first_microbatch=is_first_microbatch,
8979
                fp8_output=fp8_mha and rotary_pos_emb is None,
8980
8981
8982
            )

            if self.qkv_weight_interleaved:
cyanguwa's avatar
cyanguwa committed
8983
                # [sq, b, (ng * 2 * hn)] --> [sq, b, ng, 2 * hn]
8984
                new_tensor_shape = mixed_kv_layer.size()[:-1] + (
8985
                    self.num_gqa_groups_per_partition,
8986
8987
8988
8989
8990
                    2 * self.hidden_size_per_attention_head,
                )
                # split along last dimension
                split_dim = -1
            else:
cyanguwa's avatar
cyanguwa committed
8991
                # [sq, b, (ng * 2 * hn)] --> [sq, b, 2 * ng, hn]
8992
                new_tensor_shape = mixed_kv_layer.size()[:-1] + (
8993
                    2 * self.num_gqa_groups_per_partition,
8994
8995
8996
8997
8998
8999
9000
                    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
9001
9002
9003
            # mixed_kv_layer --> 2 [sk, b, ng, hn]
            if not is_in_onnx_export_mode():
                key_layer, value_layer = _SplitAlongDim.apply(
9004
9005
9006
                    mixed_kv_layer,
                    split_dim,
                    mixed_kv_layer.shape[split_dim] // 2,
cyanguwa's avatar
cyanguwa committed
9007
                )
9008
            else:
cyanguwa's avatar
cyanguwa committed
9009
                key_layer, value_layer = torch.split(
9010
9011
9012
                    mixed_kv_layer,
                    mixed_kv_layer.shape[split_dim] // 2,
                    dim=split_dim,
cyanguwa's avatar
cyanguwa committed
9013
                )
9014
9015
9016
9017
9018
9019
9020
9021
9022
            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)
            )
9023
9024
9025
9026
9027
9028

            # 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,
9029
                    fp8_output=fp8_mha and rotary_pos_emb is None,
9030
9031
9032
9033
9034
9035
9036
9037
9038
                )
                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,
9039
                    fp8_output=fp8_mha and rotary_pos_emb is None,
9040
9041
9042
9043
9044
9045
9046
9047
9048
                )

            # [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)

9049
9050
9051
        # ======================================================
        # Apply relative positional encoding (rotary embedding)
        # ======================================================
9052

9053
        if rotary_pos_emb is not None:
9054
9055
9056
            assert not isinstance(query_layer, Float8Tensor) and not isinstance(
                key_layer, Float8Tensor
            ), "RoPE is not supported for Float8Tensors!"
9057
            # duplicate the pos_emb for self attention
9058
            if not isinstance(rotary_pos_emb, tuple):
9059
                rotary_pos_emb = (rotary_pos_emb,) * 2
9060
9061

            q_pos_emb, k_pos_emb = rotary_pos_emb
9062
9063
9064
9065
9066
9067
9068

            # 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)
9069
9070
                else:
                    raise ValueError(f"QKV format {self.qkv_format} not supported for KV caching.")
9071
9072
9073
9074
9075
9076
9077

                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, ...]

9078
9079
9080
9081
9082
9083
9084
9085
9086
9087
9088
9089
9090
9091
9092
9093
9094
9095
            query_layer = apply_rotary_pos_emb(
                query_layer,
                q_pos_emb,
                self.qkv_format,
                fused=True,
                cu_seqlens=cu_seqlens_q,
                cp_size=self.cp_size,
                cp_rank=self.cp_rank,
            )
            key_layer = apply_rotary_pos_emb(
                key_layer,
                k_pos_emb,
                self.qkv_format,
                fused=True,
                cu_seqlens=cu_seqlens_kv,
                cp_size=self.cp_size,
                cp_rank=self.cp_rank,
            )
9096

9097
9098
9099
9100
        # ===========================
        # Core attention computation
        # ===========================

9101
9102
9103
9104
        context_layer = self.core_attention(
            query_layer,
            key_layer,
            value_layer,
9105
            qkv_format=self.qkv_format,
9106
9107
9108
9109
            cu_seqlens_q=cu_seqlens_q,
            cu_seqlens_kv=cu_seqlens_kv,
            max_seqlen_q=max_seqlen_q,
            max_seqlen_kv=max_seqlen_kv,
9110
9111
            attention_mask=attention_mask,
            attn_mask_type=attn_mask_type,
9112
            window_size=window_size,
9113
9114
9115
            checkpoint_core_attention=checkpoint_core_attention,
            core_attention_bias_type=core_attention_bias_type,
            core_attention_bias=core_attention_bias,
9116
            alibi_slopes=alibi_slopes,
9117
            fast_zero_fill=fast_zero_fill,
9118
            inference_params=inference_params,
9119
9120
        )

9121
        # ===================
9122
        # Output. [sq, b, h]
9123
        # ===================
9124

9125
        projection_output = self.proj(
9126
9127
            context_layer,
            is_first_microbatch=is_first_microbatch,
9128
9129
        )

9130
9131
9132
9133
9134
9135
9136
9137
        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,)
9138
        if self.input_layernorm and self.return_layernorm_output:
9139
9140
            outputs += (layernorm_output,)
        return outputs if len(outputs) > 1 else outputs[0]