attention.py 386 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

16
from dataclasses import dataclass, fields
cyanguwa's avatar
cyanguwa committed
17
import numpy as np
18
from packaging.version import Version as PkgVersion
19
20

import torch
21
import torch.nn.functional as F
22

23
import transformer_engine_torch as tex
24
25
import transformer_engine as te
from transformer_engine.pytorch.utils import get_cudnn_version
26
27
28
29
from transformer_engine.pytorch.cpp_extensions import (
    cast_to_fp8,
    cast_from_fp8,
)
30
31
32
33
34
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,
35
36
    fused_attn_fwd,
    fused_attn_bwd,
37
38
39
40
    QKVLayout,
    AttnBiasType,
    AttnMaskType,
    FusedAttnBackend,
41
42
43
44
45
46
47
48
49
50
51
52
53
    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,
54
)
55
from transformer_engine.pytorch.float8_tensor import Float8Tensor
56
from transformer_engine.pytorch.module import LayerNormLinear, Linear
57
from transformer_engine.pytorch.module.base import TransformerEngineBaseModule
58
59
60
61
62
from transformer_engine.pytorch.utils import (
    divide,
    attention_mask_func,
    split_tensor_along_dim,
    get_device_compute_capability,
63
    get_default_init_method,
64
65
66
67
)
from transformer_engine.pytorch.constants import (
    AttnMaskTypes,
    AttnTypes,
68
    AttnBiasTypes,
69
    QKVLayouts,
70
    dist_group_type,
71
    TE_DType,
72
73
74
75
)
from transformer_engine.pytorch.softmax import FusedScaleMaskSoftmax
from transformer_engine.pytorch.distributed import (
    get_distributed_world_size,
76
    get_distributed_rank,
77
    checkpoint,
78
79
80
    set_all_rng_states,
    CudaRNGStatesTracker,
    graph_safe_rng_available,
81
82
    gather_along_first_dim,
    reduce_scatter_along_first_dim,
83
84
)
from transformer_engine.pytorch.export import is_in_onnx_export_mode
85
from transformer_engine.pytorch.jit import jit_fuser, no_torch_dynamo
86
87
from transformer_engine.pytorch.graph import is_graph_capturing

88
89
90
91
92
93
94
95
96
97
# 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)
98

99
100
101
_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"))
102
103
_flash_attn_version = PkgVersion(get_pkg_version("flash-attn"))
_flash_attn_version_required = PkgVersion("2.0.6")
104
_flash_attn_max_version = PkgVersion("2.6.3")
105
_flash_attn_2_plus = _flash_attn_version >= PkgVersion("2")
106
107
108
109
_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")
110
_flash_attn_2_5_7_plus = _flash_attn_version >= PkgVersion("2.5.7")
111
112
_flash_attn_3_plus = False
_use_flash_attn_3 = False
113
114
115
116
117
_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"""
118
119
try:
    _flash_attn_v3_version = PkgVersion(get_pkg_version("flashattn-hopper"))
120
121
    _flash_attn_3_plus = _flash_attn_v3_version >= PkgVersion("2.9")
    _flash_attn_3_0_0_beta = _flash_attn_3_plus and _flash_attn_v3_version < PkgVersion("3.0.0")
122
except PackageNotFoundError:
123
    if get_device_compute_capability() == (9, 0) and _NVTE_FLASH_ATTN:
124
125
126
127
128
129
130
131
        fa3_logger = logging.getLogger()
        fa3_logger.setLevel(_log_level)
        if not fa3_logger.hasHandlers():
            fa3_logger.addHandler(_stream_handler)
        fa3_logger.debug(
            "To use flash-attn v3, please follow these steps to install the flashattn-hopper "
            "package: \n%s",
            _flash_attn_3_installation_steps,
132
        )
133
134
135
136
137
138
139
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,
    )

    _use_flash_attn_3 = True
140

141
if _flash_attn_version >= _flash_attn_version_required:
142
    from flash_attn.flash_attn_interface import flash_attn_func, flash_attn_varlen_func
143
144
145
    from flash_attn.flash_attn_interface import _flash_attn_varlen_forward as _flash_attn_forward
    from flash_attn.flash_attn_interface import _flash_attn_varlen_backward as _flash_attn_backward
    from flash_attn_2_cuda import varlen_bwd as flash_attn_cuda_bwd
146

147
148
149
150
151
152
153
_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,
154
}
155
156


157
158
@dataclass(eq=True)
class AttentionParams:
159
    """
160
    Attention parameters used to determine which backend to be used.
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179

    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.
180
181
182
183
    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.
184
185
186
    attn_mask_type: str, default = `no_mask`
        Attention mask type, {`no_mask`, `padding`, `causal`, `padding_causal`,
        `causal_bottom_right`, `padding_causal_bottom_right`, `arbitrary`}
187
    window_size: Tuple[int, int], default = None
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
        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.
206
207
    is_training: bool, default = `True`
        Whether in training mode (`True`) or inference mode (`False`)
208
209
210
211
    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`.
212
213
214
215
216
217
218
219
220
221
    """

    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
222
223
    head_dim_qk: int = 64
    head_dim_v: int = 64
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
    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"]


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`.
263
264
265
266
267
268
269

    Returns
    ----------
    use_flash_attention: bool
        Whether the `FlashAttention` backend has been selected.
    use_fused_attention: bool
        Whether the `FusedAttention` backend has been selected.
270
271
    fused_attention_backend: tex.NVTE_Fused_Attn_Backend
        If `use_fused_attention = True`, one of `FusedAttention` three sub-backends, else `None`.
272
273
274
275
276
277
    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].
    """
278
279
280
281
282
283
284
285
    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
286
287
    head_dim_qk = attention_params.head_dim_qk
    head_dim_v = attention_params.head_dim_v
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
    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
303
    logger = logging.getLogger("DotProductAttention")
304
305
306
    logger.setLevel(_log_level)
    if not logger.hasHandlers():
        logger.addHandler(_stream_handler)
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
    device_compute_capability = get_device_compute_capability()
    cudnn_version = get_cudnn_version()
    run_config = {
        "transformer_engine_version": te.__version__,
        "compute_capability": "sm"
        + str(
            (lambda x, y: x * 10 + y)(device_compute_capability[0], device_compute_capability[1])
        ),
        "flash_attn_version": _flash_attn_version,
        "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)
325
326

    # Filter: Environment variables
327
328
329
330
331
332
333
    global _NVTE_FLASH_ATTN, _NVTE_FUSED_ATTN, _NVTE_UNFUSED_ATTN
    _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"))
    use_flash_attention = _NVTE_FLASH_ATTN
    use_fused_attention = _NVTE_FUSED_ATTN
    use_unfused_attention = _NVTE_UNFUSED_ATTN
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
    if not use_flash_attention:
        logger.debug("Disabling FlashAttention due to NVTE_FLASH_ATTN=0")
    if not use_fused_attention:
        logger.debug("Disabling FusedAttention due to NVTE_FUSED_ATTN=0")
    if not use_unfused_attention:
        logger.debug("Disabling UnfusedDotProductAttention due to NVTE_UNFUSED_ATTN=0")

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

    # Filter: Compute capability
351
    global _use_flash_attn_3
352
353
354
355
356
357
358
    if device_compute_capability < (8, 0):
        if use_flash_attention:
            logger.debug("Disabling FlashAttention as it requires compute capability sm80+")
            use_flash_attention = False
        if use_fused_attention:
            logger.debug("Disabling FusedAttention as it requires compute capability sm80+")
            use_fused_attention = False
359
    if device_compute_capability < (9, 0):
360
        if use_flash_attention and _use_flash_attn_3:
361
362
            logger.debug("Disabling FlashAttention 3 as it requires compute capability sm90+")
            _use_flash_attn_3 = False
363
364

    # Filter: Data type
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
    if qkv_dtype not in [torch.bfloat16, torch.float16] or qkv_type not in [
        torch.Tensor,
        Float8Tensor,
    ]:
        if use_flash_attention:
            logger.debug(
                "Disabling FlashAttention due to unsupported QKV data type. "
                "Supported: qkv_dtype = {torch.bfloat16, torch.float16}. "
                "Found: qkv_dtype = %s.",
                qkv_dtype,
            )
            use_flash_attention = False
        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
385
386
387

    # Filter: Execution type
    if fp8 and fp8_meta["recipe"].fp8_dpa:
388
389
390
391
392
393
394
        if use_flash_attention and not _use_flash_attn_3:
            logger.debug("Disabling FlashAttention as FlashAttention 2 does not support FP8")
            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"
            )
395
396
397
398
399
400
            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
401
402
403
    if use_flash_attention and head_dim_qk != head_dim_v:
        logger.debug("Disabling FlashAttention as it does not support MLA.")
        use_flash_attention = False
404
    if use_flash_attention and (
405
406
407
        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)))
408
409
    ):
        logger.debug(
410
411
412
413
414
415
            "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,
416
417
418
            ".".join([str(i) for i in device_compute_capability]),
        )
        use_flash_attention = False
419
420
421
422
423
424
425
    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
426
427
428
429
430
431
432
433
434
435
436
437
438
439

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

440
    # Filter: Dropout
441
442
443
    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
444

445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
    # 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:
463
        if _use_flash_attn_3:
464
465
466
467
468
469
470
            logger.debug("Disabling FlashAttention 3 for context parallelism")
            _use_flash_attn_3 = False
        if fp8 and fp8_meta["recipe"].fp8_dpa:
            logger.debug(
                "Disabling FlashAttention as it does not support context parallelism with FP8"
            )
            use_flash_attention = False
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
        if "bottom_right" in attn_mask_type:
            logger.debug(
                "Disabling FlashAttention as it does not support context parallelism with"
                " causal_bottom_right masking"
            )
            use_flash_attention = False
        elif "causal" in attn_mask_type and max_seqlen_q != max_seqlen_kv:
            logger.debug(
                "Disabling FlashAttention as it does not support context parallelism with causal"
                " masking for cross-attention"
            )
            use_flash_attention = False
        elif core_attention_bias_type not in ["no_bias", "post_scale_bias"]:
            logger.debug(
                "Disabling FlashAttention as it does not support context parallelism with bias type"
                " of %s",
                core_attention_bias_type,
            )
            use_flash_attention = False
        elif qkv_format == "thd" and core_attention_bias_type != "no_bias":
            logger.debug(
                "Disabling FlashAttention as it does not support context parallelism with attention"
                " bias for THD format"
            )
            use_flash_attention = False
496

497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
    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

529
    # Filter: Attention mask
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
    # 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]                      |
549
550
551
552
553
554
555
    if attn_mask_type == "arbitrary":
        if use_flash_attention:
            logger.debug("Disabling FlashAttention for arbitrary mask")
        use_flash_attention = False
        if use_fused_attention:
            logger.debug("Disabling FusedAttention for arbitrary mask")
        use_fused_attention = False
556
557
    if (
        use_flash_attention
558
        and _use_flash_attn_3
559
560
561
562
563
564
565
566
567
        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
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
    if (
        use_flash_attention
        and _flash_attn_2_1_plus
        and attn_mask_type in ["causal", "padding_causal"]
        and max_seqlen_q != max_seqlen_kv
    ):
        logger.warning(
            "Disabling FlashAttention as it only supports bottom-right-diagonal "
            "causal mask since flash-attn 2.1. See "
            "https://github.com/Dao-AILab/flash-attention#21-change-behavior-of-causal-flag"
        )
        use_flash_attention = False
    if (
        use_flash_attention
        and not _flash_attn_2_1_plus
        and attn_mask_type in ["causal_bottom_right", "padding_causal_bottom_right"]
        and max_seqlen_q != max_seqlen_kv
    ):
        logger.warning(
            "Disabling FlashAttention as it only supports top-left-diagonal "
            "causal mask before flash-attn 2.1. See "
            "https://github.com/Dao-AILab/flash-attention#21-change-behavior-of-causal-flag"
        )
        use_flash_attention = False
592
593
594
595
596
597
598
599
600
    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
601
602

    # Filter: Sliding window attention
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
    #    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
        if (
            use_flash_attention
            and (window_size[0] != -1 or window_size[1] not in [-1, 0])
647
            and not _flash_attn_2_3_plus
648
        ):
649
            logger.debug(
650
                "Disabling FlashAttention as sliding window attention requires flash-attn 2.3+"
651
652
653
654
            )
            use_flash_attention = False

    # Filter: Attention bias
655
656
657
658
659
660
661
662
    #    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
663
    if use_flash_attention and core_attention_bias_type == "alibi":
664
        if _use_flash_attn_3:
665
666
            logger.debug("Disabling FlashAttention 3 for ALiBi")
            _use_flash_attn_3 = False
667
668
669
        if not _use_flash_attn_3 and not _flash_attn_2_4_plus:
            logger.debug("Disabling FlashAttention as ALiBi requires flash-attn 2.4+")
            use_flash_attention = False
670

671
672
673
674
675
676
677
678
679
680
681
682
683
    if use_flash_attention and (
        core_attention_bias_type not in ["no_bias", "alibi"]
        or core_attention_bias_shape is not None
    ):
        logger.debug("Disabling FlashAttention for pre/post_scale_bias")
        use_flash_attention = False

    fu_core_attention_bias_type = core_attention_bias_type
    fu_core_attention_bias_shape = core_attention_bias_shape
    fu_core_attention_bias_requires_grad = core_attention_bias_requires_grad
    if (
        use_fused_attention
        and core_attention_bias_type == "alibi"
684
        and (alibi_slopes_shape is not None or max_seqlen_q != max_seqlen_kv)
685
686
687
    ):
        fu_core_attention_bias_type = "post_scale_bias"
        fu_core_attention_bias_requires_grad = False
688
689
690
691
692
        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 (
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
725
726
727
728
729
730
731
            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,
732
733
            head_dim_qk,
            head_dim_v,
734
735
            window_size[0],
            window_size[1],
736
        )
737
        if fused_attention_backend == FusedAttnBackend["No_Backend"]:
738
739
            logger.debug("Disabling FusedAttention as no backend supports the provided input")
            use_fused_attention = False
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
            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"]
757
758
759
760
761
762
763
764
            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
765
            fused_attention_backend = None
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785

    # Filter: Determinism
    # backend                      | deterministic
    # ---------------------------------------------
    # FlashAttention               |
    #     flash-attn >=2.0, <2.4.1 | no
    #     flash-attn >=2.4.1       | yes
    # FusedAttention               |
    #     sub-backend 0            | yes
    #     sub-backend 1            | workspace optimization path and sm90+: yes;
    #                              | otherwise: no
    #     sub-backend 2            | no
    # UnfusedDotProductAttention   | yes
    if use_flash_attention and deterministic and not _flash_attn_2_4_1_plus:
        logger.warning(
            "Disabling FlashAttention as version <2.4.1 does not support deterministic "
            "execution. To use FlashAttention with deterministic behavior, "
            "please install flash-attn >= 2.4.1."
        )
        use_flash_attention = False
786
787
788
789
790
791
792
793
794
795
796
    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)
797
            )
798
799
800
        ):
            logger.debug("Disabling FusedAttention for determinism reasons")
            use_fused_attention = False
801
802
803

    # All available backends
    available_backends = [use_flash_attention, use_fused_attention, use_unfused_attention]
804
805
806
807
808
809
810
811
812
813
814
815
    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]),
    )
816
817
818
819
820
821
822
823
824
825
826
827
828

    # 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
829
830
831
832
833
834
835
    if (
        use_flash_attention
        and use_fused_attention
        and fused_attention_backend == FusedAttnBackend["FP8"]
        and _use_flash_attn_3
    ):
        logger.debug(
836
837
            "Disabling FlashAttention 3 to give FusedAttention preference for performance reasons "
            "in FP8 execution"
838
839
840
        )
        use_flash_attention = False

841
842
843
844
845
846
    # Selected backend
    if use_flash_attention:
        use_fused_attention = False
        use_unfused_attention = False
    elif use_fused_attention:
        use_unfused_attention = False
847
    selected_backend = "NoBackend"
848
849
850
851
852
853
    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"
854
    logger.debug("Selected backend = %s", selected_backend)
855

856
857
858
859
860
861
    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
862
863
864
865

    return (
        use_flash_attention,
        use_fused_attention,
866
        fused_attention_backend,
867
868
869
870
871
        use_unfused_attention,
        available_backends,
    )


872
class InferenceParams:  # pylint: disable=too-few-public-methods
873
874
    """
    Inference parameters that are passed to the main model in order
875
    to efficiently calculate and store the context during inference.
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915

    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,
            )
916

917

918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
@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


976
977
978
979
980
@torch.no_grad()
def get_alibi(
    num_heads: int,
    max_seqlen_q: int,
    max_seqlen_kv: int,
981
982
    actual_seqlens_q: Optional[torch.Tensor] = None,
    actual_seqlens_kv: Optional[torch.Tensor] = None,
983
984
    alibi_slopes: Optional[torch.Tensor] = None,
    bias_dtype: Optional[torch.dtype] = None,
985
    bottom_right_alignment: bool = True,
986
) -> Tuple[torch.Tensor, torch.Tensor]:
987
    """
988
989
990
991
992
993
994
995
    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.
996
997
998
999
    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].
1000
1001
1002
1003
    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.
1004
1005
1006
    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`).
1007

1008
1009
1010
1011
1012
    Returns
    ----------
    alibi_slopes: torch.Tensor
        ALiBi slopes in FP32 and shape [num_heads] or [batch_size, num_heads].
    alibi_bias: torch.Tensor
1013
1014
1015
1016
1017
1018
        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`.
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
    """
    global _alibi_cache
    if _alibi_cache["_alibi_slopes_require_update"]:
        if alibi_slopes is not None:
            _alibi_cache["_alibi_slopes"] = alibi_slopes
        else:
            n = 2 ** math.floor(math.log2(num_heads))
            m_0 = 2.0 ** (-8.0 / n)
            m = torch.pow(m_0, torch.arange(1, 1 + n))

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

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

    if _alibi_cache["_alibi_bias_require_update"]:
        assert _alibi_cache["_alibi_slopes"] is not None, "ALiBi slopes can not be None!"
        if _alibi_cache["_alibi_slopes"].dim() == 1:
            slopes_shape = torch.Size([1, _alibi_cache["_alibi_slopes"].shape[0], 1, 1])
        if _alibi_cache["_alibi_slopes"].dim() == 2:
            slopes_shape = torch.Size([*_alibi_cache["_alibi_slopes"].shape[:], 1, 1])
1044
        bias = torch.arange(max_seqlen_q, dtype=torch.int32, device="cuda").view(
1045
            1, 1, max_seqlen_q, 1
1046
1047
        ) - torch.arange(max_seqlen_kv, dtype=torch.int32, device="cuda").view(
            1, 1, 1, max_seqlen_kv
1048
        )
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
        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!"
1061
1062
1063
        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
1064
        _alibi_cache["_bottom_right_alignment"] = bottom_right_alignment
1065
1066
1067
1068
1069
        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"]
1070
1071
1072
1073
1074
1075
1076
1077
1078


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)
1079
    reduced_mask = mask.logical_not().sum(dim=1)
1080
1081
1082
1083
1084
1085
    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

1086

1087
1088
1089
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
1090
1091
1092
    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.
1093
1094
1095
1096
    """
    mask = mask.squeeze(1).squeeze(1)
    bs, seqlen = mask.shape

1097
    reduced_mask = mask.logical_not().sum(dim=1)
1098
1099
1100
1101
1102
    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)
1103
    indices = mask.logical_not().nonzero()
1104
1105
1106
1107
    indices = indices.unsqueeze(-1)

    num_nonzeros = indices.shape[0]
    pad_amount = bs * seqlen - num_nonzeros
1108
1109
1110
    indices = F.pad(
        input=indices, pad=(0, 0, 0, 0, 0, pad_amount), mode="constant", value=float(bs * seqlen)
    )
1111
1112
1113
1114

    return cu_seqlens, indices


1115
1116
1117
1118
1119
1120
1121
1122
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]
1123
1124
    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")
1125
1126
1127

    num_nonzeros = indices.shape[0]
    pad_amount = bs * max_seqlen - num_nonzeros
1128
1129
1130
1131
1132
1133
    indices = F.pad(
        input=indices,
        pad=(0, 0, 0, 0, 0, pad_amount),
        mode="constant",
        value=float(bs * max_seqlen),
    )
1134
1135
1136

    return indices

1137

1138
_cu_seqlens_cache = {}
1139
1140


1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
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.

    """
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
    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)]
1161
1162


1163
@torch.compile
1164
1165
1166
1167
1168
1169
1170
1171
def pack_tensor(
    indices: torch.Tensor,
    tensor: torch.Tensor,
) -> torch.Tensor:
    """
    Packs the given tensor using the `indices`.
    """
    padding_indice = torch.zeros(
1172
1173
        1, tensor.shape[1], tensor.shape[2], dtype=tensor.dtype, device=tensor.device
    )
1174
    indices = indices.repeat(1, tensor.shape[1], tensor.shape[2])
1175
1176
1177
1178
1179
1180
1181
1182
    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)
1183
1184
1185
    return packed


1186
@torch.compile
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
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


1200
@torch.compile
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
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


1216
@torch.compile
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
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(
1227
1228
        dim0 + 1, tensor.shape[1], tensor.shape[2], dtype=tensor.dtype, device=tensor.device
    )
1229
1230
1231
1232
1233
1234
    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, :, :]
1235
1236
1237
    return unpacked


1238
@torch.compile
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
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


1253
@torch.compile
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
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.
    """
1274

1275
1276
    @staticmethod
    def forward(
1277
        ctx, indices: torch.Tensor, *tensors: Tuple[torch.Tensor, ...]
1278
1279
    ) -> Union[Tuple[torch.Tensor, ...], torch.Tensor]:
        assert 1 <= len(tensors) <= 3, f"Packing {len(tensors)} tensors not supported."
1280
        ctx.save_for_backward(indices)
1281
1282
1283
1284
1285
1286
1287
1288
1289
        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, ...]):
1290
        (indices,) = ctx.saved_tensors
1291
        if len(grad_outputs) == 1:
1292
            return None, unpack_tensor(indices, ctx.dim0, *grad_outputs)
1293
        if len(grad_outputs) == 2:
1294
1295
            return None, *unpack_2_tensors(indices, ctx.dim0, *grad_outputs)
        return None, *unpack_3_tensors(indices, ctx.dim0, *grad_outputs)
1296
1297
1298
1299
1300
1301


class UnpackTensor(torch.autograd.Function):
    """
    Autograd function to unpack a tensor.
    """
1302

1303
1304
1305
1306
1307
1308
1309
    @staticmethod
    def forward(
        ctx,
        indices: torch.Tensor,
        dim0: int,
        tensor: torch.Tensor,
    ) -> torch.Tensor:
1310
        ctx.save_for_backward(indices)
1311
1312
1313
1314
        return unpack_tensor(indices, dim0, tensor)

    @staticmethod
    def backward(ctx, grad_output):
1315
1316
        (indices,) = ctx.saved_tensors
        return None, None, pack_tensor(indices, grad_output)
1317
1318


1319
1320
1321
def flash_attn_p2p_communicate(
    rank, send_tensor, send_dst, recv_tensor, recv_src, cp_group, batch_p2p_comm
):
1322
    """Point-to-point communications of KV and dKV in Attention with context parallelism"""
1323
1324
1325
1326
    send_recv_ops = []

    if batch_p2p_comm:
        if rank % 2 == 0:
1327
1328
1329
1330
1331
1332
            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
            )
1333
1334
1335
            send_recv_ops.append(send_op)
            send_recv_ops.append(recv_op)
        else:
1336
1337
1338
1339
1340
1341
            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
            )
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
            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


1361
@jit_fuser
1362
1363
1364
1365
1366
1367
1368
def flash_attn_fwd_out_correction(
    out: torch.Tensor,
    out_per_step: torch.Tensor,
    seq_dim: int,
    softmax_lse: torch.Tensor,
    softmax_lse_per_step: torch.Tensor,
):
1369
    """Merge partial outputs of each step in Attention with context parallelism"""
1370
    softmax_lse_corrected_exp = torch.exp(softmax_lse_per_step - softmax_lse).movedim(2, seq_dim)
1371
    softmax_lse_corrected_exp = softmax_lse_corrected_exp.unsqueeze(-1)
1372
    out_corrected = out_per_step * softmax_lse_corrected_exp
1373
1374
1375
    out.add_(out_corrected)


1376
@jit_fuser
1377
1378
1379
1380
def flash_attn_fwd_softmax_lse_correction(
    softmax_lse: torch.Tensor,
    softmax_lse_per_step: torch.Tensor,
):
1381
    """Merge softmax stats of each step in Attention with context parallelism"""
1382
1383
1384
1385
    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)
1386
1387


1388
1389
@jit_fuser
def get_cu_seqlens_on_cp_rank(
1390
1391
1392
1393
1394
1395
    cu_seqlens: torch.Tensor,
    cu_seqlens_padded_on_cp_rank: torch.Tensor,
    cp_size: int,
    cp_rank: int,
    first_half: bool,
    second_half: bool,
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
):
    """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


1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
@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


1526
class AttnFuncWithCPAndKVP2P(torch.autograd.Function):
1527
    """
1528
1529
1530
    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.
1531
1532
1533
1534
1535

    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>`_.
1536
1537
1538
    """

    @staticmethod
1539
1540
1541
1542
1543
1544
1545
    def forward(
        ctx,
        is_training,
        q,
        k,
        v,
        cu_seqlens_q,
1546
        cu_seqlens_kv,
1547
        max_seqlen_q,
1548
        max_seqlen_kv,
1549
1550
        cu_seqlens_q_padded,
        cu_seqlens_kv_padded,
1551
1552
1553
1554
1555
1556
1557
1558
        dropout_p,
        softmax_scale,
        qkv_format,
        attn_mask_type,
        attn_bias_type,
        attn_bias,
        deterministic,
        use_fused_attention,
1559
1560
        fp8,
        fp8_meta,
1561
1562
1563
        cp_group,
        cp_global_ranks,
        cp_stream,
1564
    ):
1565
1566
1567
        if softmax_scale is None:
            softmax_scale = q.shape[-1] ** (-0.5)

1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
        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

1585
1586
        cp_size = get_distributed_world_size(cp_group)
        rank = get_distributed_rank(cp_group)
1587
1588
        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]
1589
1590
        batch_p2p_comm = int(os.getenv("NVTE_BATCH_MHA_P2P_COMM", "0")) or (cp_size == 2)

1591
1592
        causal = "causal" in attn_mask_type
        padding = "padding" in attn_mask_type
1593

1594
        if qkv_format in ["bshd", "sbhd"]:
1595
            seq_dim = qkv_format.index("s")
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
            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)]
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
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
        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"]
                if fp8_meta["recipe"].fp8_mha:
                    assert (
                        isinstance(q, Float8Tensor)
                        and isinstance(k, Float8Tensor)
                        and isinstance(v, Float8Tensor)
                    ), "q/k/v must be Float8Tensors for FP8 MHA!"
                    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
            elif not fp8_meta["recipe"].fp8_mha and not int(os.getenv("NVTE_FP8_DPA_BWD", "1")):
                q_f16 = q
                q = cast_to_fp8(q_f16, fp8_meta["scaling_fwd"], META_QKV, fp8_dtype_forward)

1662
1663
1664
        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!"
1665
        if causal:
1666
1667
            if qkv_format == "bshd":
                # [b, s, np, hn] -> [b, 2, s//2, np, hn]
1668
                q, k, v = [x.view(x.shape[0], 2, x.shape[1] // 2, *x.shape[2:]) for x in [q, k, v]]
1669
1670
            elif qkv_format == "sbhd":
                # [s, b, np, hn] -> [2, s//2, b, np, hn]
1671
                q, k, v = [x.view(2, x.shape[0] // 2, *x.shape[1:]) for x in [q, k, v]]
1672
1673
1674
        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]]
1675
        if attn_bias is not None:
1676
            assert len(attn_bias.shape) == 4, (
1677
1678
1679
                "Only support bias shape of [b, h, sq, sk] for forward, "
                "and [1, h, sq, sk] for backward!"
            )
1680
1681
1682
            assert (
                attn_bias.shape[-2] % 2 == 0 and attn_bias.shape[-1] % (2 * cp_size) == 0
            ), "Sequence length does not meet divisible requirements!"
1683
            # [b, np, sq, sk] -> [b, np, 2, sq//2, 2*cp, sk//(2*cp)]
1684
1685
1686
1687
1688
1689
            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),
1690
1691
            )
            # [b, np, sq, sk] -> [b, np, sq, 2*cp, sk//(2*cp)]
1692
1693
            attn_bias = attn_bias.view(
                *attn_bias.shape[:-1], 2 * cp_size, attn_bias.shape[-1] // (2 * cp_size)
1694
            )
1695
        assert q.shape[-1] % 8 == 0, "hidden size per attention head should be multiple of 8"
1696
1697
        fa_optional_forward_kwargs = {}
        if _flash_attn_2_3_plus:
1698
            fa_optional_forward_kwargs["window_size"] = (-1, 0) if causal else (-1, -1)
1699
1700
        if _flash_attn_2_4_plus:
            fa_optional_forward_kwargs["alibi_slopes"] = None
1701
1702
        if _flash_attn_2_5_7_plus:
            fa_optional_forward_kwargs["block_table"] = None
1703

1704
1705
1706
        # Flash Attn inputs
        q_inputs = [None, None]
        kv_inputs = [None, None]
1707
        attn_bias_inputs = [None, None]
1708
1709
1710
1711
        # 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)]
1712
        attn_biases = [None for _ in range(cp_size)]
1713
1714
1715
1716
1717
1718
1719

        # 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)]
1720
1721
1722
1723
        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)
1724
1725
        send_recv_reqs = [[], []]

1726
        for i in range(cp_size + 1):
1727
            if i < cp_size:
1728
                with torch.cuda.stream(flash_attn_streams[i % 2]):
1729
                    # wait until KV is received
1730
                    for req in send_recv_reqs[(i + 1) % 2]:
1731
1732
                        req.wait()

1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
                    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,
                        )

1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
                    if (
                        not fp8
                        or fp8_meta["recipe"].fp8_mha
                        or int(os.getenv("NVTE_FP8_DPA_BWD", "1"))
                    ):
                        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:
1760
1761
1762
1763
                        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
1764
1765
                    if causal:
                        if i == 0:
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
                            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
1778
                            if use_fused_attention:
1779
1780
                                if qkv_format == "bshd":
                                    # [b, 2, sq//2, np, hn] -> [b, sq, np, hn]
1781
                                    q_inputs[i % 2] = q.view(q.shape[0], -1, *q.shape[-2:])
1782
                                    # [b, 2, sk//2, 2, np, hn] -> [b, sk, 2, np, hn]
1783
                                    kv_inputs[i % 2] = kv_inputs[i % 2].view(
1784
                                        k.shape[0], -1, 2, *k.shape[-2:]
1785
                                    )
1786
1787
                                elif qkv_format == "sbhd":
                                    # [2, sq//2, b, np, hn] -> [sq, b, np, hn]
1788
                                    q_inputs[i % 2] = q.view(-1, *q.shape[-3:])
1789
1790
1791
1792
                                    # [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:]
                                    )
1793
                                elif qkv_format == "thd":
1794
                                    q_inputs[i % 2] = q
1795
1796
                                if attn_bias is not None:
                                    idx = (rank - i) % cp_size
1797
1798
1799
1800
1801
1802
                                    attn_bias_inputs[i % 2] = torch.cat(
                                        (
                                            attn_bias[..., idx, :],
                                            attn_bias[..., (2 * cp_size - idx - 1), :],
                                        ),
                                        dim=-1,
1803
                                    ).contiguous()
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
                                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,
1832
                                )
1833
1834
1835
1836
1837
                                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
1838
1839
                            else:
                                # [b, 2, sq//2, np, hn] -> [b*sq, np, hn]
1840
                                q_inputs[i % 2] = q.view(-1, *q.shape[-2:])
1841
                                # [2, b, 2, sk//2, np, hn] -> [2, b*sk, np, hn]
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
                                kv_inputs[i % 2] = kv_inputs[i % 2].view(2, -1, *k.shape[-2:])
                                (
                                    _,
                                    _,
                                    _,
                                    _,
                                    out_per_step[i],
                                    softmax_lse_per_step[i],
                                    _,
                                    rng_states[i],
                                ) = _flash_attn_forward(
                                    q_inputs[i % 2],
                                    kv_inputs[i % 2][0],
                                    kv_inputs[i % 2][1],
1856
1857
                                    cu_seqlens_q_per_step[i],
                                    cu_seqlens_kv_per_step[i],
1858
                                    max_seqlen_q,
1859
                                    max_seqlen_kv,
1860
1861
1862
1863
1864
                                    dropout_p,
                                    softmax_scale,
                                    causal=True,
                                    return_softmax=False,
                                    **fa_optional_forward_kwargs,
1865
                                )
1866
                        elif i <= rank:
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
                            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)
1884
                            if use_fused_attention:
1885
1886
                                if qkv_format == "bshd":
                                    # [b, 2, sq//2, np, hn] -> [b, sq, np, hn]
1887
                                    q_inputs[i % 2] = q.view(q.shape[0], -1, *q.shape[-2:])
1888
1889
                                    # [b, 2, sk//2, 2, np, hn] -> [b, sk//2, 2, np, hn]
                                    kv_inputs[i % 2] = kv_inputs[i % 2][:, 0, ...].contiguous()
1890
1891
                                elif qkv_format == "sbhd":
                                    # [2, sq//2, b, np, hn] -> [sq, b, np, hn]
1892
                                    q_inputs[i % 2] = q.view(-1, *q.shape[-3:])
1893
1894
                                    # [2, sk//2, b, 2, np, hn] -> [sk//2, b, 2, np, hn]
                                    kv_inputs[i % 2] = kv_inputs[i % 2][0].contiguous()
1895
                                elif qkv_format == "thd":
1896
                                    q_inputs[i % 2] = q
1897
                                    # [2, t, np, hn] -> [2, t/2, np, hn]
1898
                                    kv_inputs[i % 2] = tex.thd_read_half_tensor(
1899
                                        kv_inputs[i % 2], cu_seqlens_kv_padded, 0
1900
                                    )
1901
1902
                                if attn_bias is not None:
                                    idx = (rank - i) % cp_size
1903
                                    attn_bias_inputs[i % 2] = attn_bias[..., idx, :].contiguous()
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
                                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,
1936
                                )
1937
1938
1939
1940
1941
                                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
1942
1943
                            else:
                                # [b, 2, sq//2, np, hn] -> [b*sq, np, hn]
1944
                                q_inputs[i % 2] = q.view(-1, *q.shape[-2:])
1945
1946
                                if qkv_format == "thd":
                                    # [2, t, np, hn] -> [2, t/2, np, hn]
1947
                                    kv_inputs[i % 2] = tex.thd_read_half_tensor(
1948
                                        kv_inputs[i % 2], cu_seqlens_kv_padded, 0
1949
                                    )
1950
1951
                                else:
                                    # [2, b, 2, sk//2, np, hn] -> [2, b, sk//2, np, hn]
1952
                                    kv_inputs[i % 2] = kv_inputs[i % 2][:, :, 0, ...].contiguous()
1953
                                # [2, b, sk//2, np, hn] -> [2, b*sk//2, np, hn]
1954
                                kv_inputs[i % 2] = kv_inputs[i % 2].view(2, -1, *k.shape[-2:])
1955
                                if _flash_attn_2_3_plus:
1956
                                    fa_optional_forward_kwargs["window_size"] = (-1, -1)
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
                                (
                                    _,
                                    _,
                                    _,
                                    _,
                                    out_per_step[i],
                                    softmax_lse_per_step[i],
                                    _,
                                    rng_states[i],
                                ) = _flash_attn_forward(
                                    q_inputs[i % 2],
                                    kv_inputs[i % 2][0],
                                    kv_inputs[i % 2][1],
1970
1971
                                    cu_seqlens_q_per_step[i],
                                    cu_seqlens_kv_per_step[i],
1972
                                    max_seqlen_q,
1973
                                    max_seqlen_kv // 2,
1974
1975
1976
1977
1978
                                    dropout_p,
                                    softmax_scale,
                                    causal=False,
                                    return_softmax=False,
                                    **fa_optional_forward_kwargs,
1979
1980
                                )
                        else:
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
                            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
1998
                            if use_fused_attention:
1999
2000
                                if qkv_format == "bshd":
                                    # [b, 2, sq//2, np, hn] -> [b, sq//2, np, hn]
2001
                                    q_inputs[i % 2] = q[:, 1, ...].contiguous()
2002
                                    # [b, 2, sk//2, 2, np, hn] -> [b, sk, 2, np, hn]
2003
                                    kv_inputs[i % 2] = kv_inputs[i % 2].view(
2004
                                        k.shape[0], -1, 2, *k.shape[-2:]
2005
                                    )
2006
2007
                                elif qkv_format == "sbhd":
                                    # [2, sq//2, b, np, hn] -> [sq//2, b, np, hn]
2008
                                    q_inputs[i % 2] = q[1].contiguous()
2009
2010
2011
2012
                                    # [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:]
                                    )
2013
2014
                                elif qkv_format == "thd":
                                    # [t, np, hn] -> [t/2, np, hn]
2015
2016
2017
                                    q_inputs[i % 2] = tex.thd_read_half_tensor(
                                        q, cu_seqlens_q_padded, 1
                                    )
2018
2019
                                if attn_bias is not None:
                                    idx = (rank - i) % cp_size
2020
2021
2022
2023
2024
2025
                                    attn_bias_inputs[i % 2] = torch.cat(
                                        (
                                            attn_bias_[..., 1, :, idx, :],
                                            attn_bias_[..., 1, :, (2 * cp_size - idx - 1), :],
                                        ),
                                        dim=-1,
2026
                                    ).contiguous()
2027
2028
2029
2030
2031
2032
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
                                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,
2059
                                )
2060
2061
2062
2063
2064
                                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
2065
                            else:
2066
2067
                                if qkv_format == "thd":
                                    # [t, np, hn] -> [t/2, np, hn]
2068
2069
2070
                                    q_inputs[i % 2] = tex.thd_read_half_tensor(
                                        q, cu_seqlens_q_padded, 1
                                    )
2071
2072
                                else:
                                    # [b, 2, sq//2, np, hn]->[b, sq//2, np, hn]->[b*sq//2, np, hn]
2073
                                    q_inputs[i % 2] = (
2074
                                        q[:, 1, ...].contiguous().view(-1, *q.shape[-2:])
2075
                                    )
2076
                                # [2, b, 2, sk//2, np, hn] -> [2, b*sk, np, hn]
2077
                                kv_inputs[i % 2] = kv_inputs[i % 2].view(2, -1, *k.shape[-2:])
2078
                                if _flash_attn_2_3_plus:
2079
                                    fa_optional_forward_kwargs["window_size"] = (-1, -1)
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
                                (
                                    _,
                                    _,
                                    _,
                                    _,
                                    out_per_step[i],
                                    softmax_lse_per_step[i],
                                    _,
                                    rng_states[i],
                                ) = _flash_attn_forward(
                                    q_inputs[i % 2],
                                    kv_inputs[i % 2][0],
                                    kv_inputs[i % 2][1],
2093
2094
                                    cu_seqlens_q_per_step[i],
                                    cu_seqlens_kv_per_step[i],
2095
                                    max_seqlen_q // 2,
2096
                                    max_seqlen_kv,
2097
2098
2099
2100
2101
                                    dropout_p,
                                    softmax_scale,
                                    causal=False,
                                    return_softmax=False,
                                    **fa_optional_forward_kwargs,
2102
2103
                                )
                    else:
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
                        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
2121
                        if use_fused_attention:
2122
2123
                            if attn_bias is not None:
                                idx = (rank - i) % cp_size
2124
2125
2126
2127
2128
2129
                                attn_bias_inputs[i % 2] = torch.cat(
                                    (
                                        attn_bias[..., idx, :],
                                        attn_bias[..., (2 * cp_size - idx - 1), :],
                                    ),
                                    dim=-1,
2130
                                ).contiguous()
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
                            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,
2159
                            )
2160
2161
2162
2163
2164
                            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
2165
                        else:
2166
                            # [b, sq, np, hn] -> [b*sq, np, hn]
2167
                            q_inputs[i % 2] = q.view(-1, *q.shape[-2:])
2168
                            # [2, b, sk, np, hn] -> [2, b*sk, np, hn]
2169
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
2182
                            kv_inputs[i % 2] = kv_inputs[i % 2].view(2, -1, *k.shape[-2:])
                            (
                                _,
                                _,
                                _,
                                _,
                                out_per_step[i],
                                softmax_lse_per_step[i],
                                _,
                                rng_states[i],
                            ) = _flash_attn_forward(
                                q_inputs[i % 2],
                                kv_inputs[i % 2][0],
                                kv_inputs[i % 2][1],
2183
2184
                                cu_seqlens_q_per_step[i],
                                cu_seqlens_kv_per_step[i],
2185
                                max_seqlen_q,
2186
                                max_seqlen_kv,
2187
2188
2189
2190
2191
                                dropout_p,
                                softmax_scale,
                                causal=False,
                                return_softmax=False,
                                **fa_optional_forward_kwargs,
2192
                            )
2193
2194
2195
2196

            if i > 0:
                # wait until fwd restuls correction of last step is done
                if i > 1:
2197
                    flash_attn_streams[(i - 1) % 2].wait_event(fwd_results_correction_done)
2198

2199
2200
                if use_fused_attention:
                    # [b, np, sq, 1] -> [b, np, sq]
2201
                    softmax_lse_per_step[i - 1].squeeze_(-1)
2202

2203
                with torch.cuda.stream(flash_attn_streams[(i - 1) % 2]):
2204
2205
2206
2207
2208
2209
2210
2211
                    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],
                        )
2212
                    if i == 1:
2213
                        out = torch.zeros_like(q if not fp8 else out_per_step[0]).view(q.shape)
2214
                        softmax_lse = torch.clone(softmax_lse_per_step[0]).to(torch.double)
2215
                        if causal and qkv_format != "thd":
2216
2217
                            # [b, np, sq] -> [b, np, 2, sq//2]
                            softmax_lse_ = softmax_lse.view(
2218
                                *softmax_lse.shape[:-1], 2, softmax_lse.shape[-1] // 2
2219
                            )
2220
2221
2222
2223
                    elif (i - 1) <= rank or not causal:
                        flash_attn_fwd_softmax_lse_correction(
                            softmax_lse, softmax_lse_per_step[i - 1]
                        )
2224
                    else:
2225
                        if qkv_format == "thd":
2226
                            tex.thd_second_half_lse_correction(
2227
2228
2229
2230
                                softmax_lse,
                                softmax_lse_per_step[i - 1],
                                cu_seqlens_q_padded,
                                max_seqlen_q,
2231
                            )
2232
                        else:
2233
2234
2235
                            flash_attn_fwd_softmax_lse_correction(
                                softmax_lse_[..., 1, :], softmax_lse_per_step[i - 1]
                            )
2236
2237

                if i < cp_size:
2238
                    flash_attn_streams[(i - 1) % 2].record_event(fwd_results_correction_done)
2239
2240
2241
2242
2243

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

        softmax_lse = softmax_lse.to(torch.float)
        for i in range(cp_size):
2244
2245
2246
2247
2248
2249
            if qkv_format == "bshd":
                out_per_step[i] = out_per_step[i].view(out.shape[0], -1, *out.shape[-2:])
                out_ = out[:, 1, ...]
            elif qkv_format == "sbhd":
                out_per_step[i] = out_per_step[i].view(-1, *out.shape[-3:])
                out_ = out[1]
2250

2251
            if i <= rank or not causal:
2252
                if qkv_format in ["bshd", "sbhd"]:
2253
2254
2255
2256
2257
2258
2259
                    flash_attn_fwd_out_correction(
                        out.view(*out_per_step[i].shape),
                        out_per_step[i],
                        seq_dim,
                        softmax_lse,
                        softmax_lse_per_step[i],
                    )
2260
                elif qkv_format == "thd":
2261
2262
2263
2264
2265
                    tex.thd_out_correction(
                        out,
                        out_per_step[i],
                        softmax_lse,
                        softmax_lse_per_step[i],
2266
                        cu_seqlens_q_padded,
2267
2268
                        False,
                    )
2269
            else:
2270
                if qkv_format in ["bshd", "sbhd"]:
2271
2272
2273
2274
2275
2276
2277
                    flash_attn_fwd_out_correction(
                        out_,
                        out_per_step[i],
                        seq_dim,
                        softmax_lse_[..., 1, :],
                        softmax_lse_per_step[i],
                    )
2278
                elif qkv_format == "thd":
2279
2280
2281
2282
2283
                    tex.thd_out_correction(
                        out,
                        out_per_step[i],
                        softmax_lse,
                        softmax_lse_per_step[i],
2284
                        cu_seqlens_q_padded,
2285
2286
                        True,
                    )
2287
2288

        kv = p2p_comm_buffers[-1]
2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
2308
        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:
2309
            out = out.view(-1, *out.shape[-2:])
2310

2311
2312
2313
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
2335
2336
        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]

        out_f16 = out.to(q_fp8.dtype if fp8 and fp8_meta["recipe"].fp8_mha else q_f16.dtype)
        if fp8 and (fp8_meta["recipe"].fp8_mha or int(os.getenv("NVTE_FP8_DPA_BWD", "1"))):
            out_fp8 = cast_to_fp8(out_f16, fp8_meta["scaling_fwd"], META_O, fp8_dtype_forward)

        if fp8 and fp8_meta["recipe"].fp8_mha:
            out_ret = Float8Tensor(
                data=out_fp8,
                fp8_meta=fp8_meta,
                fp8_meta_forward=True,
                fp8_meta_index=META_O,
                fp8_dtype=fp8_dtype_forward,
                dtype=q_fp8.dtype,
            )
        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()
        elif fp8 and fp8_meta["recipe"].fp8_mha:
2337
2338
2339
2340
2341
2342
2343
2344
            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,
            )
2345
2346
2347
2348
2349
2350
2351
2352
2353
2354
2355
            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:
2356
            q_f16 = q_f16.view(q.shape)
2357
2358
2359
            q_save, kv_save, out_save = q_f16, kv, out_f16
            fp8_fwd_scales, fp8_fwd_scale_invs = None, None

2360
        ctx.save_for_backward(
2361
2362
2363
            q_save,
            kv_save,
            out_save,
2364
            softmax_lse,
2365
2366
            cu_seqlens_q_padded,
            cu_seqlens_kv_padded,
2367
2368
            fp8_fwd_scales,
            fp8_fwd_scale_invs,
2369
2370
            *cu_seqlens_q_per_step,
            *cu_seqlens_kv_per_step,
2371
2372
            *rng_states,
            *attn_biases,
2373
        )
2374
2375
2376
        ctx.cp_group_a2a = cp_group_a2a
        ctx.cp_size_a2a = cp_size_a2a
        ctx.rank_a2a = rank_a2a
2377
2378
        ctx.cp_group = cp_group
        ctx.cp_global_ranks = cp_global_ranks
2379
        ctx.cp_stream = cp_stream
2380
        ctx.dropout_p = dropout_p
2381
        ctx.total_tokens_kv = total_tokens_kv
2382
        ctx.max_seqlen_q = max_seqlen_q
2383
        ctx.max_seqlen_kv = max_seqlen_kv
2384
        ctx.softmax_scale = softmax_scale
2385
        ctx.qkv_format = qkv_format
2386
        ctx.attn_mask_type = attn_mask_type
2387
2388
        ctx.attn_bias_type = attn_bias_type
        ctx.attn_bias_shape = None if attn_bias is None else attn_bias.shape
2389
        ctx.deterministic = deterministic
2390
        ctx.use_fused_attention = use_fused_attention
2391
2392
2393
        ctx.fp8 = fp8 and int(os.getenv("NVTE_FP8_DPA_BWD", "1"))
        ctx.fp8_meta = fp8_meta
        return out_ret
2394
2395
2396

    @staticmethod
    def backward(ctx, dout):
2397
2398
2399
        cp_size_a2a = ctx.cp_size_a2a
        rank_a2a = ctx.rank_a2a

2400
2401
        cp_size = get_distributed_world_size(ctx.cp_group)
        rank = get_distributed_rank(ctx.cp_group)
2402
2403
        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]
2404
2405
        batch_p2p_comm = int(os.getenv("NVTE_BATCH_MHA_P2P_COMM", "0")) or (cp_size == 2)

2406
        (q, kv, out, softmax_lse, cu_seqlens_q_padded, cu_seqlens_kv_padded) = ctx.saved_tensors[:6]
2407
2408
2409
2410
2411
        (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]
2412

2413
2414
        causal = "causal" in ctx.attn_mask_type
        padding = "padding" in ctx.attn_mask_type
2415
        if ctx.qkv_format in ["bshd", "sbhd"]:
2416
            seq_dim = ctx.qkv_format.index("s")
2417
2418
2419
            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
2420

2421
        if attn_biases[0] is not None:
2422
2423
            # [b, np, sq, 2*cp, sk//(2*cp)]
            attn_dbias = torch.zeros(
2424
                *ctx.attn_bias_shape, dtype=attn_biases[0].dtype, device=attn_biases[0].device
2425
2426
2427
            )
            # [b, np, sq, 2*cp, sk//(2*cp)] -> [b, np, 2, sq//2, 2*cp, sk//(2*cp)]
            attn_dbias_ = attn_dbias.view(
2428
                *attn_dbias.shape[:-3], 2, attn_dbias.shape[-3] // 2, *attn_dbias.shape[-2:]
2429
2430
2431
2432
            )
        else:
            attn_dbias = None

2433
        if causal:
2434
            if ctx.qkv_format == "thd":
2435
2436
2437
                softmax_lse_ = tex.thd_read_second_half_lse(
                    softmax_lse, cu_seqlens_q_padded, ctx.max_seqlen_q
                )
2438
2439
            else:
                # [b, np, sq] -> [b, np, 2, sq//2]
2440
2441
2442
                softmax_lse_ = softmax_lse.view(
                    *softmax_lse.shape[:-1], 2, softmax_lse.shape[-1] // 2
                )
2443
2444
2445
2446
                softmax_lse_ = softmax_lse_[..., 1, :].contiguous()
                if ctx.use_fused_attention:
                    # [b, np, sq//2] -> [b, np, sq//2, 1]
                    softmax_lse_.unsqueeze_(-1)
2447
2448
2449
        if ctx.use_fused_attention:
            # [b, np, sq] -> [b, np, sq, 1]
            softmax_lse.unsqueeze_(-1)
2450

2451
        dout_dtype = dout.dtype
2452
2453
        if ctx.fp8:
            if ctx.use_fused_attention:
2454
                fp8_dtype_forward = get_fp8_te_dtype(ctx.fp8_meta["recipe"], fprop_tensor=True)
2455
                fp8_dtype_backward = get_fp8_te_dtype(ctx.fp8_meta["recipe"], fprop_tensor=False)
2456
                fused_attn_qkv_dtype = fp8_dtype_forward
2457
2458
2459
2460
2461
2462
2463
2464
2465
2466
2467
2468
2469
2470
2471
2472
2473
2474
2475
2476
2477
2478
2479
2480
2481
2482
2483
2484
                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)
                if ctx.fp8_meta["recipe"].fp8_mha:
                    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:
            if ctx.fp8_meta is not None and ctx.fp8_meta["recipe"].fp8_mha:
2485
2486
2487
2488
2489
2490
2491
                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
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
            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]
2503
                fused_attn_dqkv_dtype = TE_DType[dout_dtype]
2504
2505
                fused_attn_backend = FusedAttnBackend["F16_arbitrary_seqlen"]

2506
2507
2508
2509
2510
2511
2512
2513
2514
2515
2516
2517
2518
2519
2520
2521
2522
2523
2524
        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,
            )
            if not ctx.fp8 and ctx.fp8_meta is not None and ctx.fp8_meta["recipe"].fp8_mha:
                dout = cast_from_fp8(
                    dout, None, None, dout_fp8_dtype, TE_DType[dout_dtype], scale_inv=dout_scale_inv
                )

2525
2526
2527
2528
        out = out.view(*q.shape)
        dout = dout.view(*q.shape)
        send_recv_reqs = []

2529
2530
2531
2532
2533
2534
        fa_optional_backward_kwargs = {}
        if _flash_attn_2_4_plus:
            fa_optional_backward_kwargs["alibi_slopes"] = None
        if _flash_attn_2_4_1_plus:
            fa_optional_backward_kwargs["deterministic"] = ctx.deterministic

2535
2536
2537
2538
2539
        for i in range(cp_size):
            # wait until KV is received
            for req in send_recv_reqs:
                req.wait()

2540
2541
            send_tensor = p2p_comm_buffers[i % 2]
            recv_tensor = p2p_comm_buffers[(i + 1) % 2]
2542
2543
2544
2545
2546
2547
2548
2549
2550
2551
2552
2553
2554
2555
2556
2557
2558
2559
2560
2561
2562
2563
2564
2565
2566
2567
2568
2569
2570
            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
                )
2571

2572
            kv = p2p_comm_buffers[i % 2][0]
2573
2574
2575
            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]
2576
            # In reversed order of fwd
2577
            if causal:
2578
                if i == (cp_size - 1):
2579
                    if ctx.use_fused_attention:
2580
2581
2582
                        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:])
2583
2584
                            # [b, 2, sk//2, 2, np, hn] -> [b, sk, 2, np, hn]
                            kv_ = kv.view(kv.shape[0], -1, *kv.shape[-3:])
2585
2586
2587
2588
2589
2590
                            # [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:])
2591
2592
                            # [2, sk//2, b, 2, np, hn] -> [sk, b, 2, np, hn]
                            kv_ = kv.view(-1, *kv.shape[-4:])
2593
2594
2595
                            # [2, sq//2, b, np, hn] -> [sq, b, np, hn]
                            out_ = out.view(-1, *out.shape[-3:])
                            dout_ = dout.view(-1, *dout.shape[-3:])
2596
2597
                        elif ctx.qkv_format == "thd":
                            q_, kv_, out_, dout_ = q, kv, out, dout
2598
2599
2600
2601
2602
2603
2604
2605
                        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]]
2606
                        if attn_dbias is not None:
2607
                            aux_ctx_tensors += [attn_biases[cp_size - i - 1]]
2608
                        dq_, dk_, dv_, dbias_ = fused_attn_bwd(
2609
                            ctx.max_seqlen_q,
2610
2611
2612
                            ctx.max_seqlen_kv,
                            cu_seqlens_q_per_step[cp_size - i - 1],
                            cu_seqlens_kv_per_step[cp_size - i - 1],
2613
                            q_,
2614
2615
                            kv_[..., 0, :, :] if ctx.qkv_format in ["bshd", "sbhd"] else kv_[0],
                            kv_[..., 1, :, :] if ctx.qkv_format in ["bshd", "sbhd"] else kv_[1],
2616
2617
                            out_,
                            dout_,
2618
2619
                            fused_attn_qkv_dtype,
                            fused_attn_dqkv_dtype,
2620
                            aux_ctx_tensors,
2621
                            fused_attn_backend,
2622
2623
                            cu_seqlens_q_padded=cu_seqlens_q_padded,
                            cu_seqlens_kv_padded=cu_seqlens_kv_padded,
2624
2625
                            attn_scale=ctx.softmax_scale,
                            dropout=ctx.dropout_p,
2626
                            qkv_layout=qkv_layout,
2627
                            attn_mask_type=ctx.attn_mask_type,
2628
                            attn_bias_type=ctx.attn_bias_type,
2629
2630
                            deterministic=ctx.deterministic,
                            **fp8_meta_kwargs,
2631
2632
2633
2634
                        )
                    else:
                        # [b, 2, sq//2, np, hn] -> [b*sq, np, hn]
                        q_ = q.view(-1, *q.shape[-2:])
2635
                        dq_ = torch.zeros_like(q_)
2636
2637
2638
2639
2640
2641
2642
                        # [2, b, 2, sk//2, np, hn] -> [2, b*sk, np, hn]
                        kv_ = kv.view(2, -1, *kv.shape[-2:])
                        dkv_ = torch.empty_like(kv_)
                        # [b, 2, sq//2, np, hn] -> [b*sq, np, hn]
                        out_ = out.view(-1, *out.shape[-2:])
                        dout_ = dout.view(-1, *dout.shape[-2:])
                        if _flash_attn_2_3_plus:
2643
                            fa_optional_backward_kwargs["window_size"] = (-1, 0)
2644
                        _flash_attn_backward(
2645
2646
2647
2648
2649
2650
2651
2652
2653
                            dout_,
                            q_,
                            kv_[0],
                            kv_[1],
                            out_,
                            softmax_lse,
                            dq_,
                            dkv_[0],
                            dkv_[1],
2654
2655
                            cu_seqlens_q_per_step[cp_size - i - 1],
                            cu_seqlens_kv_per_step[cp_size - i - 1],
2656
                            ctx.max_seqlen_q,
2657
                            ctx.max_seqlen_kv,
2658
2659
2660
2661
2662
                            ctx.dropout_p,
                            ctx.softmax_scale,
                            True,
                            rng_state=rng_states[cp_size - i - 1],
                            **fa_optional_backward_kwargs,
2663
                        )
2664
                elif i >= (cp_size - rank - 1):
2665
                    if ctx.use_fused_attention:
2666
2667
2668
                        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:])
2669
2670
                            # [b, 2, sk//2, 2, np, hn] -> [b, sk//2, 2, np, hn]
                            kv_ = kv[:, 0, ...].contiguous()
2671
2672
2673
2674
2675
2676
                            # [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:])
2677
2678
                            # [2, sk//2, b, 2, np, hn] -> [sk//2, b, 2, np, hn]
                            kv_ = kv[0].contiguous()
2679
2680
2681
                            # [2, sq//2, b, np, hn] -> [sq, b, np, hn]
                            out_ = out.view(-1, *out.shape[-3:])
                            dout_ = dout.view(-1, *dout.shape[-3:])
2682
2683
2684
                        elif ctx.qkv_format == "thd":
                            q_, out_, dout_ = q, out, dout
                            # [2, t, np, hn] -> [2, t/2, np, hn]
2685
                            kv_ = tex.thd_read_half_tensor(kv, cu_seqlens_kv_padded, 0)
2686
2687
2688
2689
2690
2691
2692
2693
                        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]]
2694
                        if attn_dbias is not None:
2695
                            aux_ctx_tensors += [attn_biases[cp_size - i - 1]]
2696
                        dq_, dk_, dv_, dbias_ = fused_attn_bwd(
2697
                            ctx.max_seqlen_q,
2698
2699
2700
                            ctx.max_seqlen_kv // 2,
                            cu_seqlens_q_per_step[cp_size - i - 1],
                            cu_seqlens_kv_per_step[cp_size - i - 1],
2701
                            q_,
2702
2703
                            kv_[..., 0, :, :] if ctx.qkv_format in ["bshd", "sbhd"] else kv_[0],
                            kv_[..., 1, :, :] if ctx.qkv_format in ["bshd", "sbhd"] else kv_[1],
2704
2705
                            out_,
                            dout_,
2706
2707
                            fused_attn_qkv_dtype,
                            fused_attn_dqkv_dtype,
2708
                            aux_ctx_tensors,
2709
                            fused_attn_backend,
2710
2711
2712
2713
                            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
                            ),
2714
2715
                            attn_scale=ctx.softmax_scale,
                            dropout=ctx.dropout_p,
2716
                            qkv_layout=qkv_layout,
2717
                            attn_mask_type="padding" if padding else "no_mask",
2718
                            attn_bias_type=ctx.attn_bias_type,
2719
2720
                            deterministic=ctx.deterministic,
                            **fp8_meta_kwargs,
2721
2722
2723
2724
                        )
                    else:
                        # [b, 2, sq//2, np, hn] -> [b*sq, np, hn]
                        q_ = q.view(-1, *q.shape[-2:])
2725
                        dq_ = torch.zeros_like(q_)
2726
2727
                        if ctx.qkv_format == "thd":
                            # [2, t, np, hn] -> [2, t/2, np, hn]
2728
                            kv_ = tex.thd_read_half_tensor(kv, cu_seqlens_kv_padded, 0)
2729
2730
2731
                        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:])
2732
2733
2734
2735
2736
                        dkv_ = torch.empty_like(kv_)
                        # [b, 2, sq//2, np, hn] -> [b*sq, np, hn]
                        out_ = out.view(-1, *out.shape[-2:])
                        dout_ = dout.view(-1, *dout.shape[-2:])
                        if _flash_attn_2_3_plus:
2737
                            fa_optional_backward_kwargs["window_size"] = (-1, -1)
2738
                        _flash_attn_backward(
2739
2740
2741
2742
2743
2744
2745
2746
2747
                            dout_,
                            q_,
                            kv_[0],
                            kv_[1],
                            out_,
                            softmax_lse,
                            dq_,
                            dkv_[0],
                            dkv_[1],
2748
2749
                            cu_seqlens_q_per_step[cp_size - i - 1],
                            cu_seqlens_kv_per_step[cp_size - i - 1],
2750
                            ctx.max_seqlen_q,
2751
                            ctx.max_seqlen_kv // 2,
2752
2753
2754
2755
2756
                            ctx.dropout_p,
                            ctx.softmax_scale,
                            False,
                            rng_state=rng_states[cp_size - i - 1],
                            **fa_optional_backward_kwargs,
2757
2758
2759
                        )
                else:
                    if ctx.use_fused_attention:
2760
2761
2762
                        if ctx.qkv_format == "bshd":
                            # [b, 2, sq//2, np, hn] -> [b, sq//2, np, hn]
                            q_ = q[:, 1, ...].contiguous()
2763
2764
                            # [b, 2, sk//2, 2, np, hn] -> [b, sk, 2, np, hn]
                            kv_ = kv.view(kv.shape[0], -1, *kv.shape[-3:])
2765
2766
2767
2768
2769
2770
                            # [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()
2771
2772
                            # [2, sk//2, b, 2, np, hn] -> [sk, b, 2, np, hn]
                            kv_ = kv.view(-1, *kv.shape[-4:])
2773
2774
2775
                            # [2, sq//2, b, np, hn] -> [sq//2, b, np, hn]
                            out_ = out[1].contiguous()
                            dout_ = dout[1].contiguous()
2776
2777
                        elif ctx.qkv_format == "thd":
                            # [t, np, hn] -> [t/2, np, hn]
2778
2779
2780
                            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)
2781
                            kv_ = kv
2782
2783
2784
2785
2786
2787
2788
2789
                        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]]
2790
                        if attn_dbias is not None:
2791
                            aux_ctx_tensors += [attn_biases[cp_size - i - 1]]
2792
                        dq_, dk_, dv_, dbias_ = fused_attn_bwd(
2793
                            ctx.max_seqlen_q // 2,
2794
2795
2796
                            ctx.max_seqlen_kv,
                            cu_seqlens_q_per_step[cp_size - i - 1],
                            cu_seqlens_kv_per_step[cp_size - i - 1],
2797
                            q_,
2798
2799
                            kv_[..., 0, :, :] if ctx.qkv_format in ["bshd", "sbhd"] else kv_[0],
                            kv_[..., 1, :, :] if ctx.qkv_format in ["bshd", "sbhd"] else kv_[1],
2800
2801
                            out_,
                            dout_,
2802
2803
                            fused_attn_qkv_dtype,
                            fused_attn_dqkv_dtype,
2804
                            aux_ctx_tensors,
2805
                            fused_attn_backend,
2806
2807
2808
2809
                            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,
2810
2811
                            attn_scale=ctx.softmax_scale,
                            dropout=ctx.dropout_p,
2812
                            qkv_layout=qkv_layout,
2813
                            attn_mask_type="padding" if padding else "no_mask",
2814
                            attn_bias_type=ctx.attn_bias_type,
2815
2816
                            deterministic=ctx.deterministic,
                            **fp8_meta_kwargs,
2817
2818
                        )
                    else:
2819
2820
                        if ctx.qkv_format == "thd":
                            # [t, np, hn] -> [t/2, np, hn]
2821
                            q_ = tex.thd_read_half_tensor(q, cu_seqlens_q_padded, 1)
2822
2823
2824
                        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:])
2825
                        dq_ = torch.zeros_like(q_)
2826
2827
2828
                        # [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_)
2829
                        if ctx.qkv_format == "thd":
2830
2831
                            out_ = tex.thd_read_half_tensor(out, cu_seqlens_q_padded, 1)
                            dout_ = tex.thd_read_half_tensor(dout, cu_seqlens_q_padded, 1)
2832
2833
2834
2835
                        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:])
2836
                        if _flash_attn_2_3_plus:
2837
                            fa_optional_backward_kwargs["window_size"] = (-1, -1)
2838
                        _flash_attn_backward(
2839
2840
2841
2842
2843
2844
2845
2846
2847
                            dout_,
                            q_,
                            kv_[0],
                            kv_[1],
                            out_,
                            softmax_lse_,
                            dq_,
                            dkv_[0],
                            dkv_[1],
2848
2849
                            cu_seqlens_q_per_step[cp_size - i - 1],
                            cu_seqlens_kv_per_step[cp_size - i - 1],
2850
                            ctx.max_seqlen_q // 2,
2851
                            ctx.max_seqlen_kv,
2852
2853
2854
2855
2856
                            ctx.dropout_p,
                            ctx.softmax_scale,
                            False,
                            rng_state=rng_states[cp_size - i - 1],
                            **fa_optional_backward_kwargs,
2857
2858
2859
                        )
            else:
                if ctx.use_fused_attention:
2860
2861
2862
2863
                    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]]
2864
                    if attn_dbias is not None:
2865
                        aux_ctx_tensors += [attn_biases[cp_size - i - 1]]
2866
                    dq_, dk_, dv_, dbias_ = fused_attn_bwd(
2867
                        ctx.max_seqlen_q,
2868
2869
2870
                        ctx.max_seqlen_kv,
                        cu_seqlens_q_per_step[cp_size - i - 1],
                        cu_seqlens_kv_per_step[cp_size - i - 1],
2871
                        q,
2872
2873
                        kv[..., 0, :, :] if ctx.qkv_format in ["bshd", "sbhd"] else kv[0],
                        kv[..., 1, :, :] if ctx.qkv_format in ["bshd", "sbhd"] else kv[1],
2874
2875
                        out,
                        dout,
2876
2877
                        fused_attn_qkv_dtype,
                        fused_attn_dqkv_dtype,
2878
                        aux_ctx_tensors,
2879
                        fused_attn_backend,
2880
2881
                        cu_seqlens_q_padded=cu_seqlens_q_padded,
                        cu_seqlens_kv_padded=cu_seqlens_kv_padded,
2882
2883
                        attn_scale=ctx.softmax_scale,
                        dropout=ctx.dropout_p,
2884
                        qkv_layout=qkv_layout,
2885
                        attn_mask_type=ctx.attn_mask_type,
2886
                        attn_bias_type=ctx.attn_bias_type,
2887
2888
                        deterministic=ctx.deterministic,
                        **fp8_meta_kwargs,
2889
2890
2891
                    )
                else:
                    # [b, sq, np, hn] -> [b*sq, np, hn]
2892
                    q_ = q.view(-1, *q.shape[-2:])
2893
                    dq_ = torch.zeros_like(q_)
2894
                    # [2, b, sk, np, hn] -> [2, b*sk, np, hn]
2895
2896
                    kv_ = kv.view(2, -1, *kv.shape[-2:])
                    dkv_ = torch.empty_like(kv_)
2897
                    # [b, sq, np, hn] -> [b*sq, np, hn]
2898
2899
                    out_ = out.view(-1, *out.shape[-2:])
                    dout_ = dout.view(-1, *dout.shape[-2:])
2900
                    if _flash_attn_2_3_plus:
2901
                        fa_optional_backward_kwargs["window_size"] = (-1, -1)
2902
                    _flash_attn_backward(
2903
2904
2905
2906
2907
2908
2909
2910
2911
                        dout_,
                        q_,
                        kv_[0],
                        kv_[1],
                        out_,
                        softmax_lse,
                        dq_,
                        dkv_[0],
                        dkv_[1],
2912
2913
                        cu_seqlens_q_per_step[cp_size - i - 1],
                        cu_seqlens_kv_per_step[cp_size - i - 1],
2914
                        ctx.max_seqlen_q,
2915
                        ctx.max_seqlen_kv,
2916
2917
2918
                        ctx.dropout_p,
                        ctx.softmax_scale,
                        False,
2919
                        rng_state=rng_states[cp_size - i - 1],
2920
                        **fa_optional_backward_kwargs,
2921
2922
                    )

2923
2924
            if ctx.fp8:
                dq = dq_fp8[(rank + i + 1) % cp_size]
2925
            if i >= (cp_size - rank - 1) or not causal:
2926
2927
2928
2929
                # [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:
2930
2931
2932
2933
2934
2935
                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:])
2936

2937
2938
2939
2940
2941
2942
2943
2944
2945
2946
2947
            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:
2948
                if i > (cp_size - rank - 1):
2949
                    dq.add_(dq_)
2950
2951
                elif i == (cp_size - rank - 1):
                    if rank == (cp_size - 1):
2952
2953
                        dq.copy_(dq_)
                    else:
2954
2955
2956
2957
2958
2959
                        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])
2960
                        elif ctx.qkv_format == "thd":
2961
                            tex.thd_grad_correction(dq, dq_, cu_seqlens_q_padded, "copy", "add")
2962
                elif i > 0:
2963
2964
2965
2966
                    if ctx.qkv_format == "bshd":
                        dq[:, 1, ...].add_(dq_)
                    elif ctx.qkv_format == "sbhd":
                        dq[1].add_(dq_)
2967
                    elif ctx.qkv_format == "thd":
2968
                        tex.thd_grad_correction(dq, dq_, cu_seqlens_q_padded, "none", "add")
2969
                else:
2970
2971
2972
2973
                    if ctx.qkv_format == "bshd":
                        dq[:, 1, ...].copy_(dq_)
                    elif ctx.qkv_format == "sbhd":
                        dq[1].copy_(dq_)
2974
                    elif ctx.qkv_format == "thd":
2975
                        tex.thd_grad_correction(dq, dq_, cu_seqlens_q_padded, "none", "copy")
2976
2977
2978
2979
2980
            else:
                if i == 0:
                    dq.copy_(dq_)
                else:
                    dq.add_(dq_)
2981

2982
            if attn_dbias is not None:
2983
                idx = (rank + i + 1) % cp_size
2984
                if i == (cp_size - 1) or not causal:
2985
                    # [b, np, sq, sk//cp] -> [b, np, sq, 2, sk//(2*cp)]
2986
                    dbias_ = dbias_.view(*dbias_.shape[:-1], 2, dbias_.shape[-1] // 2)
2987
                    attn_dbias[..., idx, :].copy_(dbias_[..., 0, :])
2988
2989
                    attn_dbias[..., (2 * cp_size - idx - 1), :].copy_(dbias_[..., 1, :])
                elif i >= (cp_size - rank - 1):
2990
2991
2992
2993
                    # [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)]
2994
                    dbias_ = dbias_.view(*dbias_.shape[:-1], 2, dbias_.shape[-1] // 2)
2995
                    attn_dbias_[..., 1, :, idx, :].copy_(dbias_[..., 0, :])
2996
                    attn_dbias_[..., 1, :, (2 * cp_size - idx - 1), :].copy_(dbias_[..., 1, :])
2997

2998
2999
3000
            # wait until dKV is received
            for req in send_recv_reqs:
                req.wait()
3001

3002
3003
3004
3005
3006
3007
3008
            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]
3009
3010
            if ctx.use_fused_attention:
                dkv_ = torch.cat((dk_.unsqueeze(0), dv_.unsqueeze(0)), dim=0)
3011
3012
3013
3014
                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:])
3015
            if causal and i >= (cp_size - rank - 1) and i != (cp_size - 1):
3016
3017
3018
3019
3020
3021
                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:])
3022
3023
3024
3025
            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)
3026

3027
3028
3029
3030
3031
3032
3033
3034
3035
3036
3037
            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:
3038
                if i == (cp_size - 1):
3039
                    if rank == 0:
3040
3041
3042
3043
3044
3045
                        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, ...])
3046
                        elif ctx.qkv_format == "thd":
3047
                            tex.thd_grad_correction(dkv, dkv_, cu_seqlens_kv_padded, "add", "copy")
3048
3049
                    else:
                        dkv.add_(dkv_)
3050
3051
                elif i >= (cp_size - rank - 1):
                    if i == 0 and rank == (cp_size - 1):
3052
3053
3054
3055
                        if ctx.qkv_format == "bshd":
                            dkv[:, :, 0, ...].copy_(dkv_)
                        elif ctx.qkv_format == "sbhd":
                            dkv[:, 0, ...].copy_(dkv_)
3056
                        elif ctx.qkv_format == "thd":
3057
                            tex.thd_grad_correction(dkv, dkv_, cu_seqlens_kv_padded, "copy", "none")
3058
                    else:
3059
3060
3061
3062
                        if ctx.qkv_format == "bshd":
                            dkv[:, :, 0, ...].add_(dkv_)
                        elif ctx.qkv_format == "sbhd":
                            dkv[:, 0, ...].add_(dkv_)
3063
                        elif ctx.qkv_format == "thd":
3064
                            tex.thd_grad_correction(dkv, dkv_, cu_seqlens_kv_padded, "add", "none")
3065
3066
3067
3068
3069
                elif i > 0:
                    dkv.add_(dkv_)
                else:
                    dkv.copy_(dkv_)
            else:
3070
3071
3072
3073
3074
                if i == 0:
                    dkv.copy_(dkv_)
                else:
                    dkv.add_(dkv_)

3075
3076
3077
3078
3079
3080
3081
3082
3083
3084
3085
3086
3087
3088
3089
3090
3091
3092
3093
3094
        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]]

3095
        if causal:
3096
3097
            if ctx.qkv_format == "bshd":
                # [b, 2, sq//2, np, hn] -> [b, sq, np, hn]
3098
                dq = dq.view(dq.shape[0], -1, *dq.shape[-2:])
3099
                # [2, b, 2, sk//2, np, hn] -> [2, b, sk, np, hn]
3100
                dkv = dkv.view(*dkv.shape[0:2], -1, *dkv.shape[-2:])
3101
3102
            elif ctx.qkv_format == "sbhd":
                # [2, sq//2, b, np, hn] -> [sq, b, np, hn]
3103
                dq = dq.view(-1, *dq.shape[-3:])
3104
                # [2, 2, sk//2, b, np, hn] -> [2, sk, b, np, hn]
3105
3106
3107
3108
3109
3110
3111
3112
3113
                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_
3114

3115
3116
3117
3118
3119
        if ctx.fp8 and ctx.fp8_meta["recipe"].fp8_mha:
            dq, dkv = [
                cast_to_fp8(x, ctx.fp8_meta["scaling_bwd"], META_DQKV, fp8_dtype_backward)
                for x in [dq, dkv]
            ]
3120
3121
3122
3123
3124
3125
3126
3127
3128
3129
3130
3131
3132
3133
3134
3135
3136
3137
3138
        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]]

        if ctx.fp8 and ctx.fp8_meta["recipe"].fp8_mha:
3139
3140
3141
3142
3143
3144
3145
3146
3147
            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,
                )
3148
                for x in [dq, dk, dv]
3149
3150
            ]

3151
3152
3153
3154
        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)

3155
3156
3157
        return (
            None,
            dq,
3158
3159
            dk,
            dv,
3160
3161
3162
3163
3164
3165
3166
3167
3168
3169
3170
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
3171
            attn_dbias,
3172
3173
3174
3175
3176
            None,
            None,
            None,
            None,
            None,
3177
3178
            None,
            None,
3179
        )
3180
3181


3182
3183
def get_kv_seq_info_after_all_gather(
    local_chunk_id, cp_size, max_seqlen_q, max_seqlen_kv, window_size, causal
3184
):
3185
3186
3187
3188
3189
3190
3191
3192
3193
3194
3195
3196
3197
3198
3199
3200
3201
3202
3203
3204
3205
3206
    """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)
3207
3208
3209
3210


class AttnFuncWithCPAndKVAllGather(torch.autograd.Function):
    """
3211
3212
    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>`_.
3213
3214
3215
3216
3217
3218
3219
3220
3221
3222
3223
3224
3225
3226
3227
3228
3229
3230
3231
3232
3233
3234
    """

    @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,
3235
3236
        cp_group,
        cp_stream,
3237
3238
3239
3240
3241
3242
3243
3244
3245
    ):
        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
3246
        assert not padding, f"{attn_mask_type} mask type is not supported!"
3247
3248
3249
3250
3251
3252
3253
3254
3255
3256
        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!"
        fa_optional_forward_kwargs = {}
        if _flash_attn_2_4_plus:
            fa_optional_forward_kwargs["alibi_slopes"] = None
3257
3258
        if _flash_attn_2_5_7_plus:
            fa_optional_forward_kwargs["block_table"] = None
3259
3260
3261
3262
3263
3264
3265
3266
3267
3268
3269
3270
3271
3272

        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)

3273
3274
3275
3276
        # [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]]
3277

3278
        # [s, b, np, hn] -> [cp, s, b, np, hn]
3279
3280
        k_ag, _ = gather_along_first_dim(k, cp_group)
        v_ag, _ = gather_along_first_dim(v, cp_group)
3281
3282

        # [cp, s, b, np, hn] -> [cp*2, s//2, b, np, hn]
3283
3284
        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:])
3285
3286
3287
3288
3289
3290
3291
3292
3293
3294
        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]
3295
3296

        local_seq_chunk_ids = [rank, 2 * cp_size - rank - 1]
3297
3298
3299
        kv_seq_range_per_step = [None, None]
        window_size_per_step = [None, None]
        cu_seqlens_kv_per_step = [None, None]
3300
3301
3302
3303
3304
3305
3306
3307
        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]):
3308
3309
                    # [b, 2, sq//2, np, hn] -> [b, sq//2, np, hn]
                    # or [2, sq//2, b, np, hn] -> [sq//2, b, np, hn]
3310
3311
3312
3313
3314
3315
3316
3317
3318
                    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,
3319
                        )
3320
3321
3322
3323
3324
3325
3326
3327
3328
3329
3330
3331
                    )
                    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_]]
3332
3333
3334
3335
                    if use_fused_attention:
                        out_per_step[i], [softmax_lse_per_step[i], rng_states[i]] = fused_attn_fwd(
                            is_training,
                            max_seqlen_q,
3336
                            max_seqlen_kv_,
3337
                            cu_seqlens_q,
3338
                            cu_seqlens_kv_per_step[i],
3339
3340
3341
3342
3343
3344
3345
3346
3347
3348
3349
3350
                            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,
3351
3352
                            cu_seqlens_kv_padded=cu_seqlens_kv_per_step[i],
                            window_size=window_size_per_step[i],
3353
3354
3355
3356
3357
3358
3359
3360
3361
                        )
                    else:
                        q_, k_, v_ = [x.view(-1, *x.shape[-2:]) for x in [q_, k_, v_]]
                        _, _, _, _, out_per_step[i], softmax_lse_per_step[i], _, rng_states[i] = (
                            _flash_attn_forward(
                                q_,
                                k_,
                                v_,
                                cu_seqlens_q,
3362
                                cu_seqlens_kv_per_step[i],
3363
                                max_seqlen_q,
3364
                                max_seqlen_kv_,
3365
3366
                                dropout_p,
                                softmax_scale,
3367
                                causal=causal,
3368
                                return_softmax=False,
3369
                                window_size=window_size_per_step[i],
3370
3371
3372
3373
3374
3375
3376
                                **fa_optional_forward_kwargs,
                            )
                        )

            if i > 0:
                with torch.cuda.stream(flash_attn_streams[i - 1]):
                    if qkv_format == "bshd":
3377
                        out[:, i - 1].copy_(out_per_step[i - 1].view(out[:, i - 1].shape))
3378
                    elif qkv_format == "sbhd":
3379
                        out[i - 1].copy_(out_per_step[i - 1].view(out[i - 1].shape))
3380
3381
3382
3383
3384
3385
3386
3387
3388
3389
3390
3391
3392
3393
3394
3395
3396

        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,
3397
            *cu_seqlens_kv_per_step,
3398
3399
3400
3401
            *out_per_step,
            *softmax_lse_per_step,
            *rng_states,
        )
3402
3403
        ctx.kv_seq_range_per_step = kv_seq_range_per_step
        ctx.window_size_per_step = window_size_per_step
3404
3405
3406
3407
3408
3409
3410
        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
3411
        ctx.attn_mask_type = attn_mask_type
3412
3413
3414
3415
3416
3417
3418
3419
3420
        ctx.deterministic = deterministic
        ctx.use_fused_attention = use_fused_attention
        return out

    @staticmethod
    def backward(ctx, dout):
        cp_size = get_distributed_world_size(ctx.cp_group)
        rank = get_distributed_rank(ctx.cp_group)

3421
3422
3423
3424
3425
3426
3427
        (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
3428

3429
        seq_dim = ctx.qkv_format.index("s")
3430
3431
        qkv_layout = ctx.qkv_format + "_" + ctx.qkv_format + "_" + ctx.qkv_format

3432
        dout = dout.view(q.shape)
3433
        dq = torch.empty_like(q)
3434
        dk = torch.zeros((k.shape[0] * cp_size, *k.shape[1:]), dtype=k.dtype, device=k.device)
3435
3436
3437
3438
3439
3440
3441
3442
3443
3444
        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()

3445
        # [s, b, np, hn] -> [cp, s, b, np, hn]
3446
3447
        k_ag, _ = gather_along_first_dim(k, ctx.cp_group)
        v_ag, _ = gather_along_first_dim(v, ctx.cp_group)
3448
3449

        # [cp, s, b, np, hn] -> [cp*2, s//2, b, np, hn]
3450
3451
        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:])
3452
3453
3454
3455
3456
3457
3458
        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())
3459
3460
3461
3462
3463
3464
3465
3466
3467
3468
3469
3470

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

        fa_optional_backward_kwargs = {}
        if _flash_attn_2_4_plus:
            fa_optional_backward_kwargs["alibi_slopes"] = None
        if _flash_attn_2_4_1_plus:
            fa_optional_backward_kwargs["deterministic"] = ctx.deterministic

        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]):
3471
3472
                    # [b, 2, sq//2, np, hn] -> [b, sq//2, np, hn]
                    # or [2, sq//2, b, np, hn] -> [sq//2, b, np, hn]
3473
3474
3475
3476
3477
3478
3479
3480
3481
                    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_]]
3482
                    out_ = out_per_step[i]
3483
                    dout_ = dout.select(seq_dim, i).contiguous().view(out_.shape)
3484
3485
3486
3487
                    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,
3488
                            max_seqlen_kv,
3489
                            cu_seqlens_q,
3490
                            cu_seqlens_kv_per_step[i],
3491
3492
3493
3494
3495
3496
                            q_,
                            k_,
                            v_,
                            out_,
                            dout_,
                            TE_DType[q.dtype],
3497
                            TE_DType[dout.dtype],
3498
3499
3500
                            aux_ctx_tensors,
                            tex.NVTE_Fused_Attn_Backend.NVTE_F16_arbitrary_seqlen,
                            cu_seqlens_q_padded=cu_seqlens_q_padded,
3501
                            cu_seqlens_kv_padded=cu_seqlens_kv_per_step[i],
3502
3503
3504
3505
3506
                            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,
3507
3508
                            window_size=window_size_per_step[i],
                            deterministic=ctx.deterministic,
3509
3510
                        )
                    else:
3511
                        batch_size = k_.shape[0]
3512
3513
3514
3515
3516
3517
3518
3519
3520
3521
3522
3523
3524
3525
3526
                        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_]
                        ]
                        _flash_attn_backward(
                            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,
3527
                            cu_seqlens_kv_per_step[i],
3528
                            ctx.max_seqlen_q,
3529
                            max_seqlen_kv,
3530
3531
                            ctx.dropout_p,
                            ctx.softmax_scale,
3532
3533
                            "causal" in ctx.attn_mask_type,
                            window_size=window_size_per_step[i],
3534
3535
3536
                            rng_state=rng_states[i],
                            **fa_optional_backward_kwargs,
                        )
3537
3538
3539
3540
3541
3542
3543
                        # [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]]
                        ]
3544
3545
3546
3547

            if i > 0:
                with torch.cuda.stream(flash_attn_streams[i - 1]):
                    if ctx.qkv_format == "bshd":
3548
                        dq[:, i - 1].copy_(dq_per_step[i - 1])
3549
                    elif ctx.qkv_format == "sbhd":
3550
3551
3552
3553
3554
3555
                        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]]
                    ]
3556
3557
3558
                    # wait until dkv update of last step is done
                    if i > 1:
                        flash_attn_streams[i - 1].wait_event(dkv_update_done)
3559
3560
3561
3562
3563
3564
                    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])
3565
3566
3567
3568
3569
                    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)

3570
3571
3572
3573
3574
3575
3576
        # [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]
3577
3578
3579
3580
3581
        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)

3582
3583
3584
3585
3586
3587
3588
3589
3590
3591
3592
3593
3594
3595
3596
3597
3598
3599
3600
3601
3602
3603
3604
3605
3606
3607
3608
3609
3610
3611
3612
3613
3614
3615
3616
3617
3618
3619
3620
3621
3622
3623
3624
3625
3626
3627
3628
3629
3630
3631
3632
3633
3634
3635
3636
3637
3638
3639
3640
3641
3642
3643
3644
3645
3646
3647
3648
3649
3650
3651
3652
3653
3654
3655
3656
3657
3658
3659
3660
3661
3662
3663
3664
3665
3666
3667
3668
3669
3670
3671
3672
3673
3674
3675
3676
3677
3678
3679
3680
3681
3682
3683
3684
3685
3686
3687
3688
3689
3690
3691
3692
3693
3694
3695
3696
3697
3698
3699
3700
3701
3702
3703
3704
3705
3706
3707
3708
3709
3710
3711
3712
3713
3714
3715
3716
3717
3718
3719
3720
3721
3722
3723
3724
3725
3726
3727
3728
3729
3730
3731
3732
3733
3734
3735
3736
3737
3738
3739
3740
3741
3742
3743
3744
3745
3746
3747
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
3808
3809
3810
3811
3812
3813
3814
3815
3816
3817
3818
3819
3820
3821
3822
3823
3824
3825
3826
3827
3828
3829
3830
3831
3832
3833
3834
3835
3836
3837
3838
3839
3840
3841
3842
3843
3844
3845
3846
3847
3848
3849
3850
3851
3852
3853
3854
3855
3856
3857
3858
3859
3860
3861
3862
3863
3864
3865
3866
3867
3868
3869
3870
3871
3872
3873
3874
3875
3876
3877
3878
3879
3880
3881
3882
3883
3884
3885
3886
3887
3888
3889
3890
3891
3892
3893
3894
3895
3896
3897
3898
3899
3900
3901
3902
3903
3904
3905
3906
3907
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
3941
3942
3943
3944
3945
3946
3947
3948
3949
3950
3951
3952
3953
3954
3955
3956
3957
3958
3959
3960
3961
3962
3963
3964
3965
3966
3967
3968
3969
3970
3971
3972
3973
3974
3975
3976
3977
3978
3979
3980
3981
3982
3983
3984
3985
3986
3987
3988
3989
3990
3991
3992
3993
3994
3995
3996
3997
3998
3999
4000
4001
4002
4003
4004
4005
4006
4007
4008
4009
4010
        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,
    ):
        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!"
        fa_optional_forward_kwargs = {}
        if _flash_attn_2_3_plus:
            fa_optional_forward_kwargs["window_size"] = window_size
        if _flash_attn_2_4_plus:
            fa_optional_forward_kwargs["alibi_slopes"] = None
        if _flash_attn_2_5_7_plus:
            fa_optional_forward_kwargs["block_table"] = None

        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!"

        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"]
                if fp8_meta["recipe"].fp8_mha:
                    assert (
                        isinstance(q, Float8Tensor)
                        and isinstance(k, Float8Tensor)
                        and isinstance(v, Float8Tensor)
                    ), "q/k/v must be Float8Tensors for FP8 MHA!"
                    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
        )

        if fp8 and not fp8_meta["recipe"].fp8_mha and not 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]
            ]

        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]]
            (
                _,
                _,
                _,
                _,
                out,
                softmax_lse,
                _,
                rng_state,
            ) = _flash_attn_forward(
                q,
                k,
                v,
                cu_seqlens_q,
                cu_seqlens_kv,
                max_seqlen_q,
                max_seqlen_kv,
                dropout_p,
                softmax_scale,
                causal=causal,
                return_softmax=False,
                **fa_optional_forward_kwargs,
            )
            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:
            if fp8_meta["recipe"].fp8_mha:
                out_fp8 = Float8Tensor(
                    data=out,
                    fp8_meta=fp8_meta,
                    fp8_meta_forward=True,
                    fp8_meta_index=META_O,
                    fp8_dtype=fp8_dtype_forward,
                    dtype=q_fp8.dtype,
                )
                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
            elif fp8_meta["recipe"].fp8_mha:
                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
        return out_ret

    @staticmethod
    def backward(ctx, dout):
        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")

        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"]
                if ctx.fp8_meta["recipe"].fp8_mha:
                    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:
            if ctx.fp8_meta is not None and ctx.fp8_meta["recipe"].fp8_mha:
                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
        )

        fa_optional_backward_kwargs = {}
        if _flash_attn_2_3_plus:
            fa_optional_backward_kwargs["window_size"] = ctx.window_size
        if _flash_attn_2_4_plus:
            fa_optional_backward_kwargs["alibi_slopes"] = None
        if _flash_attn_2_4_1_plus:
            fa_optional_backward_kwargs["deterministic"] = ctx.deterministic

        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]]
            _flash_attn_backward(
                dout,
                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.softmax_scale,
                causal,
                rng_state=rng_state,
                **fa_optional_backward_kwargs,
            )
            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
        )

4011
        if ctx.qkv_format == "bshd":
4012
            dq, dk, dv = [x.view(ctx.batch_size, -1, *x.shape[-2:]) for x in [dq, dk, dv]]
4013
        elif ctx.qkv_format == "sbhd":
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
            dq, dk, dv = [x.view(-1, ctx.batch_size, *x.shape[-2:]) for x in [dq, dk, dv]]

        if ctx.fp8:
            if ctx.fp8_meta["recipe"].fp8_mha:
                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_fp8.dtype,
                    )
                    for x in [dq, dk, dv]
                ]
            else:
                dq, dk, dv = [
                    cast_from_fp8(
                        x,
                        ctx.fp8_meta["scaling_bwd"],
                        META_DQKV,
                        fp8_dtype_backward,
                        TE_DType[dout_f16.dtype],
                    )
                    for x in [dq, dk, dv]
                ]
4040
4041
4042
4043
4044
4045
4046
4047
4048
4049
4050
4051
4052
4053
4054
4055
4056
4057
4058
4059
4060
4061
4062

        return (
            None,
            dq,
            dk,
            dv,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
4063
4064
4065
            None,
            None,
            None,
4066
4067
4068
        )


4069
def attn_forward_func_with_cp(
4070
4071
4072
4073
4074
    is_training,
    q,
    k,
    v,
    cu_seqlens_q,
4075
    cu_seqlens_kv,
4076
    max_seqlen_q,
4077
    max_seqlen_kv,
4078
4079
    cu_seqlens_q_padded,
    cu_seqlens_kv_padded,
4080
4081
4082
4083
    dropout_p,
    cp_group,
    cp_global_ranks,
    cp_stream,
4084
    cp_comm_type,
4085
4086
4087
4088
4089
4090
4091
    softmax_scale=None,
    qkv_format="bshd",
    attn_mask_type="causal",
    attn_bias_type="no_bias",
    attn_bias=None,
    deterministic=False,
    use_fused_attention=False,
4092
    window_size=None,
4093
4094
    fp8=False,
    fp8_meta=None,
4095
) -> torch.Tensor:
4096
4097
4098
4099
    """
    Attention implementation with context parallelism.
    """

4100
4101
4102
4103
4104
4105
4106
4107
4108
4109
4110
4111
4112
4113
4114
4115
    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}!"

4116
4117
4118
4119
4120
4121
4122
4123
4124
4125
4126
4127
4128
4129
4130
4131
4132
4133
4134
4135
    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!"""
    )
4136
4137
4138
    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!"
4139
4140
4141

    sliding_window_attn = (
        window_size is not None and window_size != (-1, 0) and window_size != (-1, -1)
4142
    )
4143
4144
4145
4146
4147
    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!"
4148

4149
4150
4151
4152
4153
4154
4155
4156
4157
4158
4159
4160
4161
4162
4163
4164
4165
4166
4167
4168
4169
    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,
    ]

4170
    if cp_comm_type in ["p2p", "a2a+p2p"]:
4171
4172
4173
4174
4175
4176
4177
4178
4179
4180
        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)
4181
4182
4183
    else:
        raise ValueError(f"Unsupported communication type: {cp_comm_type}!")

4184
4185
4186
    return out


4187
4188
4189
4190
class RotaryPositionEmbedding(torch.nn.Module):
    """
    Implements Rotary Position Embedding from https://arxiv.org/abs/2104.09864.
    """
4191

4192
4193
4194
    def __init__(
        self,
        dim: int,
4195
        rotary_percent: float = 1.0,
4196
4197
        seq_len_interpolation_factor: Optional[int] = None,
        pretrained_max_position_embeddings: Optional[int] = None,
4198
        rotary_base: float = 10000.0,
4199
4200
4201
4202
4203
4204
    ):
        """
        Parameters
        ----------
        dim: int
            rotary embedding dimension
4205
4206
        rotary_percent: float
            Percent of rotary dimension to use for rotary position embeddings.
4207
4208
4209
4210
4211
4212
4213
        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__()
4214
4215
        if rotary_percent < 1.0:
            dim = int(dim * rotary_percent)
4216
        self.seq_len_interpolation_factor = seq_len_interpolation_factor
4217
        self.rotary_base = rotary_base
4218
        inv_freq = 1.0 / (
4219
            self.rotary_base
4220
4221
4222
4223
4224
            ** (
                torch.arange(0, dim, 2, dtype=torch.float32, device=torch.cuda.current_device())
                / dim
            )
        )
4225
        self.register_buffer("inv_freq", inv_freq)
4226
4227
4228
4229
4230
4231
4232
4233
4234
4235
4236
4237
4238
        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
        """
4239
4240
4241
4242
        seq = (
            torch.arange(max_seq_len, device=self.inv_freq.device, dtype=self.inv_freq.dtype)
            + offset
        )
4243

4244
4245
4246
4247
4248
4249
4250
4251
        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
            ):
4252
4253
4254
4255
4256
4257
                # 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

4258
        freqs = torch.einsum("i , j -> i j", seq, self.inv_freq)
4259
4260
4261
4262
4263
4264
        # 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))

4265
4266
4267
4268
4269
4270
4271
4272
4273
4274
4275
4276
4277
4278
4279
4280
4281
4282

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

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

    @staticmethod
    def forward(
        ctx,
        t: torch.Tensor,
        freqs: torch.Tensor,
        tensor_format: str = "sbhd",
        cu_seqlens: Union[torch.Tensor, None] = None,
    ) -> torch.Tensor:
4283
4284
        if freqs.dtype != torch.float32:
            freqs = freqs.float()
4285
4286
4287
        if tensor_format == "sbhd":
            output = tex.fused_rope_forward(t, freqs, False)
        elif tensor_format == "bshd":
4288
            output = tex.fused_rope_forward(t.transpose(0, 1), freqs, True).transpose(0, 1)
4289
4290
4291
4292
4293
4294
4295
4296
4297
4298
        elif tensor_format == "thd":
            output = tex.fused_rope_thd_forward(t, cu_seqlens, freqs)
        else:
            raise ValueError(f"Unsupported tensor_format: {tensor_format}.")
        ctx.save_for_backward(freqs, cu_seqlens)
        ctx.tensor_format = tensor_format

        return output

    @staticmethod
4299
    def backward(ctx, grad_output: torch.Tensor) -> Tuple[Union[torch.Tensor, None], ...]:
4300
4301
4302
4303
4304
4305
4306
4307
4308
4309
4310
4311
4312
4313
4314
        freqs, cu_seqlens = ctx.saved_tensors
        if ctx.tensor_format == "sbhd":
            grad_input = tex.fused_rope_backward(grad_output, freqs, False)
        elif ctx.tensor_format == "bshd":
            grad_input = tex.fused_rope_backward(
                grad_output.transpose(0, 1), freqs, True
            ).transpose(0, 1)
        elif ctx.tensor_format == "thd":
            grad_input = tex.fused_rope_thd_backward(grad_output, cu_seqlens, freqs)
        else:
            raise ValueError(f"Unsupported tensor_format: {ctx.tensor_format}.")

        return grad_input, None, None, None, None


4315
4316
4317
4318
4319
4320
4321
4322
4323
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)


4324
def apply_rotary_pos_emb(
4325
4326
4327
4328
4329
4330
    t: torch.Tensor,
    freqs: torch.Tensor,
    tensor_format: str = "sbhd",
    fused: bool = False,
    cu_seqlens: Union[torch.Tensor, None] = None,
) -> torch.Tensor:
4331
    """
4332
    Apply rotary positional embedding tensor to the input tensor.
4333

4334
4335
4336
    Parameters
    ----------
    t: torch.Tensor
4337
        Input tensor of shape `[s, b, h, d]`, `[b, s, h, d]` or `[t, h, d]`, on which
4338
4339
4340
4341
4342
4343
4344
4345
4346
4347
4348
4349
        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'.
4350
    """
4351
4352
4353
4354
4355
4356
4357
4358
4359
4360
4361
    if fused:
        assert (
            tensor_format != "thd" or cu_seqlens is not None
        ), "cu_seqlens must not be None when tensor_format is 'thd'."
        return FusedRoPEFunc.apply(t, freqs, tensor_format, cu_seqlens)

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

4362
4363
4364
4365
4366
    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.
4367
4368
4369
    assert (
        cur_seq_len <= max_seq_len
    ), f"Rotary Embeddings only supported up to {max_seq_len} sequence length!"
4370
    freqs = freqs[:cur_seq_len]
4371
    if tensor_format == "bshd":
4372
4373
4374
4375
        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)
4376

4377
4378
4379
4380
4381
4382
    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
4383
    t = (t * cos_) + (_rotate_half(t) * sin_)
4384
4385
4386
    return torch.cat((t, t_pass), dim=-1)


cyanguwa's avatar
cyanguwa committed
4387
class _SplitAlongDim(torch.autograd.Function):
4388
4389
4390
    """"""

    @staticmethod
4391
4392
4393
4394
4395
    def forward(
        ctx,
        mixed_x_layer: torch.Tensor,
        split_dim: int,
        split_size_or_sections: Union[int, List[int], Tuple[int]],
4396
    ) -> Tuple[torch.Tensor, ...]:
cyanguwa's avatar
cyanguwa committed
4397
4398
        ctx.split_dim = split_dim
        ctx.split_size_or_sections = split_size_or_sections
4399
        if isinstance(mixed_x_layer, Float8Tensor):
4400
4401
4402
4403
4404
4405
            return tuple(
                Float8Tensor.make_like(
                    mixed_x_layer,
                    data=x,
                )
                for x in torch.split(
4406
4407
                    mixed_x_layer._data,
                    split_size_or_sections=split_size_or_sections,
4408
4409
4410
4411
                    dim=split_dim,
                )
            )
        return torch.split(mixed_x_layer, split_size_or_sections, dim=split_dim)
4412
4413

    @staticmethod
4414
    def backward(ctx, *grad_outputs):
4415
4416
        assert len(grad_outputs) > 0, "No gradients received for backprop!"

cyanguwa's avatar
cyanguwa committed
4417
4418
        if isinstance(ctx.split_size_or_sections, (list, tuple)):
            split_sizes = ctx.split_size_or_sections
4419
4420
4421
            assert len(grad_outputs) == len(
                split_sizes
            ), "Unequal number of gradients vs split sections for backprop!"
cyanguwa's avatar
cyanguwa committed
4422
4423
4424
4425
4426
        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

4427
4428
4429
4430
4431
4432
4433
4434
        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]
4435
4436
4437
4438
4439
4440
4441
                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
                ):
4442
4443
4444
                    noop_ok = False
                    break
            if noop_ok:
4445
4446
4447
                ret = torch.Tensor().to(
                    device=grad_outputs[0].device, dtype=grad_outputs[0]._data.dtype
                )
4448
4449
                new_shape = list(shape)
                new_shape[split_dim] = sum(split_sizes)
4450
4451
4452
4453
4454
                ret.set_(
                    grad_outputs[0]._data.untyped_storage(),
                    grad_outputs[0]._data.storage_offset(),
                    new_shape,
                    strides,
4455
4456
4457
4458
                )
                return Float8Tensor.make_like(grad_outputs[0], data=ret), None, None

            grad_outputs_data = [x._data for x in grad_outputs]
4459
4460
4461
4462
4463
4464
4465
            return (
                Float8Tensor.make_like(
                    grad_outputs[0], data=torch.cat(grad_outputs_data, dim=split_dim)
                ),
                None,
                None,
            )
4466
4467
        noop_ok = True
        strides = grad_outputs[0].stride()
4468
        data_ptr = grad_outputs[0].untyped_storage().data_ptr()
cyanguwa's avatar
cyanguwa committed
4469
        shape = list(grad_outputs[0].shape)
4470
        for i, tensor in enumerate(grad_outputs):
cyanguwa's avatar
cyanguwa committed
4471
4472
            shape_i = shape
            shape_i[split_dim] = split_sizes[i]
4473
4474
4475
4476
4477
4478
4479
            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
            ):
4480
4481
4482
                noop_ok = False
                break
        if noop_ok:
4483
            ret = torch.Tensor().to(device=grad_outputs[0].device, dtype=grad_outputs[0].dtype)
4484
            new_shape = list(shape)
cyanguwa's avatar
cyanguwa committed
4485
            new_shape[split_dim] = sum(split_sizes)
4486
4487
4488
4489
4490
            ret.set_(
                grad_outputs[0].untyped_storage(),
                grad_outputs[0].storage_offset(),
                new_shape,
                strides,
4491
            )
cyanguwa's avatar
cyanguwa committed
4492
            return ret, None, None
4493

4494
        return torch.cat(grad_outputs, dim=split_dim), None, None
4495
4496
4497
4498
4499
4500
4501
4502
4503


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

    def __init__(
        self,
4504
        softmax_scale: float,
4505
        attention_type: str = "self",
4506
4507
4508
4509
4510
4511
        attention_dropout: float = 0.0,
        attention_dropout_ctx: Optional[Callable] = nullcontext,
        layer_number: Optional[int] = None,
    ) -> None:
        super().__init__()

4512
        self.softmax_scale = softmax_scale
4513
        self.attention_type = attention_type
4514
4515
4516
        self.attention_dropout_ctx = attention_dropout_ctx
        self.layer_number = layer_number

4517
        self.scale_mask_softmax = FusedScaleMaskSoftmax(attention_mask_func)
4518
4519
4520
4521
4522
4523

        # 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)

4524
4525
        # An FP16 training trick required for certain GPT-like models.
        self.apply_qk_layer_scaling = (
4526
4527
            bool(int(os.getenv("NVTE_APPLY_QK_LAYER_SCALING", "0"))) and layer_number is not None
        )
4528

4529
4530
4531
4532
4533
    def forward(
        self,
        query_layer: torch.Tensor,
        key_layer: torch.Tensor,
        value_layer: torch.Tensor,
4534
        qkv_layout: str = "sbh3d",
4535
4536
        cu_seqlens_q: Optional[torch.Tensor] = None,  # pylint: disable=unused-argument
        cu_seqlens_kv: Optional[torch.Tensor] = None,  # pylint: disable=unused-argument
4537
        attn_mask_type: str = "causal",
4538
        attention_mask: Optional[Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]] = None,
4539
4540
        core_attention_bias_type: str = "no_bias",
        core_attention_bias: Optional[torch.Tensor] = None,
4541
        alibi_slopes: Optional[torch.Tensor] = None,
4542
    ) -> torch.Tensor:
4543
        """Unfused attention fprop"""
4544
4545
4546
4547
4548
        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":
4549
            # convert to sbhd and use sbhd implementation for now
4550
4551
4552
            query_layer, key_layer, value_layer = [
                x.transpose(0, 1) for x in [query_layer, key_layer, value_layer]
            ]
4553
4554
4555
4556
4557
4558
4559
4560
4561
4562
4563
4564
4565
4566
4567
4568
4569
4570
4571
4572
4573
4574
4575
4576
4577
4578
4579
4580
4581
4582
4583
4584
4585
4586
4587
4588
4589
4590
4591
4592
4593
4594
4595
4596
4597
4598
4599
4600
4601
4602
4603
4604
        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,
                )
4605

4606
        batch_size, seqlen = query_layer.shape[1], query_layer.shape[0]
4607
        apply_qk_layer_scaling = self.apply_qk_layer_scaling and key_layer.dtype == torch.float16
4608
4609
4610
4611
4612
4613
4614
4615
4616

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

4617
        if key_layer.shape[2] != query_layer.shape[2]:
4618
4619
4620
            assert (
                query_layer.shape[2] % key_layer.shape[2] == 0
            ), "The number of attention heads must be divisible by the number of GQA groups!"
4621
            key_layer = key_layer.repeat_interleave(
4622
4623
                int(query_layer.shape[2] / key_layer.shape[2]), dim=2
            )
4624
            value_layer = value_layer.repeat_interleave(
4625
4626
                int(query_layer.shape[2] / value_layer.shape[2]), dim=2
            )
4627

4628
        # [sq, b, np, hn] -> [sq, b * np, hn]
4629
        query_layer = query_layer.reshape(output_size[2], output_size[0] * output_size[1], -1)
4630
4631
4632
4633
        # [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]
4634
4635
        # WAR to set dtype to FP32 as ONNX lacks BF16 support for ConstantOfShape operator
        is_bf16 = query_layer.dtype == torch.bfloat16
4636
4637
4638
4639
        matmul_result = torch.empty(
            output_size[0] * output_size[1],
            output_size[2],
            output_size[3],
4640
            dtype=torch.float32 if is_in_onnx_export_mode() and is_bf16 else query_layer.dtype,
4641
4642
4643
            device=torch.cuda.current_device(),
        )

4644
4645
4646
        if is_in_onnx_export_mode() and is_bf16:
            matmul_result = matmul_result.bfloat16()

4647
        scale = self.softmax_scale
4648
        if apply_qk_layer_scaling:
4649
            scale /= self.layer_number
4650
4651

        # Raw attention scores. [b * np, sq, sk]
4652
4653
4654
4655
4656
4657
        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,
4658
                alpha=scale,
4659
            ).view(*output_size)
4660
4661
4662
4663
4664
4665
4666

        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]
            )
4667
            matmul_result = matmul_result.view(*output_size) + core_attention_bias
4668
            matmul_result *= scale
4669

4670
4671
4672
4673
        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":
4674
                _, core_attention_bias = get_alibi(
4675
4676
4677
                    output_size[1],
                    output_size[2],
                    output_size[3],
4678
4679
                    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,
4680
4681
                    alibi_slopes=alibi_slopes,
                    bottom_right_alignment=attn_mask_type not in ["causal", "padding_causal"],
4682
                )
4683
4684
4685
4686
4687
            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,
4688
                alpha=scale,
4689
            )
4690
4691
            matmul_result = (matmul_result.view(*output_size) + core_attention_bias).to(
                dtype=query_layer.dtype
4692
            )
4693
4694
4695

        # attention scores and attention mask [b, np, sq, sk]
        softmax_scale = self.layer_number if apply_qk_layer_scaling else None
4696
        attention_probs = self.scale_mask_softmax(
4697
            matmul_result, attention_mask, attn_mask_type, softmax_scale
4698
        )
4699

4700
4701
4702
4703
4704
        # 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)

4705
4706
4707
4708
4709
4710
4711
4712
4713
4714
4715
4716
4717
4718
4719
        # 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]
4720
        value_layer = value_layer.reshape(value_layer.size(0), output_size[0] * output_size[1], -1)
4721
4722

        # change view [b * np, sq, sk]
4723
        attention_probs = attention_probs.view(output_size[0] * output_size[1], output_size[2], -1)
4724
4725
4726
4727
4728
4729
4730

        # 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)

4731
        if qkv_format == "sbhd":
4732
4733
            # [b, np, sq, hn] --> [sq, b, np, hn]
            context_layer = context_layer.permute(2, 0, 1, 3).contiguous()
4734

4735
4736
4737
            # [sq, b, np, hn] --> [sq, b, hp]
            context_layer = context_layer.view(seqlen, batch_size, -1)

4738
        if qkv_format == "bshd":
4739
4740
4741
4742
4743
            # [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)
4744
4745
4746
4747
4748
4749

        return context_layer


class _PrepareQKVForFA(torch.autograd.Function):
    """This class converts QKV from interleaved (s, b, ...) layout
4750
    to separate contiguous q, k, v tensors in (b, s, ...) layout."""
4751
4752

    @staticmethod
4753
4754
4755
4756
    def forward(
        _ctx: torch.autograd.function.FunctionCtx,  # unused
        query_layer: torch.Tensor,
        key_layer: torch.Tensor,
4757
        value_layer: torch.Tensor,
4758
    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
4759
4760
4761
4762
4763
4764
4765
4766
4767
4768
4769
        # 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
4770
4771
4772
4773
    def backward(
        _ctx: torch.autograd.function.FunctionCtx,  # unused
        dq: torch.Tensor,
        dk: torch.Tensor,
4774
        dv: torch.Tensor,
4775
4776
4777
4778
4779
    ) -> Tuple[Union[torch.Tensor, None], ...]:
        dqkv = tex.fa_prepare_bwd(dq, dk, dv)
        dq, dk, dv = split_tensor_along_dim(dqkv, -1, 3)
        return dq, dk, dv

4780

4781
def get_qkv_layout(
4782
4783
4784
4785
4786
    q: torch.Tensor,
    k: torch.Tensor,
    v: torch.Tensor,
    qkv_format: str = "sbhd",
) -> str:
4787
    """Get qkv layout.
4788

4789
4790
4791
4792
4793
4794
4795
4796
4797
4798
4799
    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,
4800
        `d` head size, and `t` the total number of tokens in a batch, i.e.
4801
4802
4803
4804
4805
4806
4807
4808
4809
4810
4811
4812
4813
4814
4815
        `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`}
4816
4817
4818
4819
4820
4821
4822
4823
4824
    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.
4825
    """
4826

4827
4828
    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!"
4829

4830
    def run_iteratively(q, k, v):
4831
        # check data pointers
4832
4833
        data_ptr = q.untyped_storage().data_ptr()
        check_ptrs_qkv = all(x.untyped_storage().data_ptr() == data_ptr for x in [q, k, v])
4834
        check_ptrs_qk = all(x.untyped_storage().data_ptr() == data_ptr for x in [q, k])
4835
4836
4837
        data_ptr = k.untyped_storage().data_ptr()
        check_ptrs_kv = all(x.untyped_storage().data_ptr() == data_ptr for x in [k, v])

4838
4839
4840
4841
4842
4843
4844
        # 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
4845
4846
        stride = q.stride()
        check_strides_qkv = all(stride == x.stride() for x in [q, k, v])
4847
4848
        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]
4849
        )
4850

4851
4852
4853
4854
4855
4856
        # 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])
        )
4857

4858
4859
4860
4861
4862
4863
        # 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])
4864
        )
4865
4866
        check_h2d_offsets = all(
            x.storage_offset() == (offset + i * k.shape[-1]) for i, x in enumerate([k, v])
4867
        )
4868

4869
4870
4871
4872
4873
4874
4875
4876
4877
4878
        # 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]))
4879
        )
4880
4881
4882
4883
        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]))
4884
        )
4885

4886
        if check_ptrs_qkv and check_strides_qkv and check_shapes_qkv and check_3hd_offsets:
4887
            # sb3hd, bs3hd, t3hd
4888
            # one chunk of memory, qkv, with q, k, v interleaved at dim=-3 in qkv
4889
            qkv_layout = qkv_format[:-2] + "3" + qkv_format[-2:]
4890
        elif check_ptrs_qkv and check_strides_qkv and check_shapes_qkv and check_h3d_offsets:
4891
            # sbh3d, bsh3d, th3d
4892
            # one chunk of memory, qkv, with q, k, v interleaved at dim=-2 in qkv
4893
            qkv_layout = qkv_format[:-1] + "3" + qkv_format[-1:]
4894
        elif check_ptrs_kv and check_strides_kv and check_shapes_kv and check_2hd_offsets:
4895
            # sbhd_sb2hd, bshd_bs2hd, thd_t2hd
4896
4897
4898
            # 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
4899
            qkv_layout = qkv_format + "_" + qkv_format[:-2] + "2" + qkv_format[-2:]
4900
        elif check_ptrs_kv and check_strides_kv and check_shapes_kv and check_h2d_offsets:
4901
            # sbhd_sbh2d, bshd_bsh2d, thd_th2d
4902
4903
4904
            # 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
4905
            qkv_layout = qkv_format + "_" + qkv_format[:-1] + "2" + qkv_format[-1:]
4906
4907
4908
4909
4910
        elif (
            check_strides_kv
            and check_shapes_kv
            and (check_hd_offsets_qkv or check_hd_offsets_kv or check_hd_offsets_qk)
        ):
4911
            # sbhd_sbhd_sbhd, bshd_bshd_bshd, thd_thd_thd
4912
4913
4914
            # 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
4915
            qkv_layout = "_".join(list([qkv_format]) * 3)
4916
        else:
4917
            qkv_layout = "not_supported"
4918
4919
4920
4921

        return qkv_layout

    qkv_layout = run_iteratively(q, k, v)
4922
    if qkv_layout == "not_supported":
4923
4924
4925
        # 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)
4926
    if qkv_layout == "not_supported":
4927
4928
        raise Exception("The provided qkv memory layout is not supported!")

4929
    return qkv_layout, q, k, v
4930

4931

4932
def check_set_window_size(
4933
4934
4935
    attn_mask_type: str,
    window_size: Tuple[int, int] = None,
):
4936
4937
4938
4939
4940
4941
4942
4943
    """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)
4944
    """
4945
    orig_window_size = window_size
4946
    if "causal" in attn_mask_type:
4947
        if orig_window_size is None:
4948
            window_size = (-1, 0)
4949
4950
4951
        elif orig_window_size == (-1, -1) or (
            orig_window_size[0] >= 0 and orig_window_size[1] != 0
        ):
4952
4953
4954
4955
            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
            )
4956
        elif orig_window_size != (-1, 0) and (orig_window_size[0] < 0 or orig_window_size[1] != 0):
4957
4958
4959
4960
            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"]:
4961
4962
4963
        if orig_window_size is None:
            window_size = (-1, -1)
        elif orig_window_size == (-1, 0):
4964
            window_size = (-1, -1)
4965
4966
4967
            warnings.warn(
                "window_size should be (-1, -1) or (>=0, >=0) for attn_mask_type=" + attn_mask_type
            )
4968
        elif orig_window_size != (-1, -1) and (orig_window_size[0] < 0 or orig_window_size[1] < 0):
4969
4970
4971
4972
4973
            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
4974
    return window_size
4975

4976

4977
class FlashAttention(torch.nn.Module):
4978
    """Dot product attention, using HazyResearch flash-attn package:
4979
    https://github.com/Dao-AILab/flash-attention
4980
4981
4982
4983
    """

    def __init__(
        self,
4984
        softmax_scale: float,
4985
4986
        attention_dropout: float = 0.0,
        attention_dropout_ctx: Optional[Callable] = nullcontext,
4987
4988
        attention_type: str = "self",
        layer_number: Optional[int] = None,
4989
        deterministic: bool = False,
4990
4991
4992
4993
4994
4995
    ) -> None:
        super().__init__()

        assert (
            _flash_attn_version >= _flash_attn_version_required
        ), f"FlashAttention minimum version {_flash_attn_version_required} is required."
4996
4997
4998
        assert (
            _flash_attn_version <= _flash_attn_max_version
        ), f"FlashAttention maximum version {_flash_attn_max_version} is supported."
4999

5000
        self.softmax_scale = softmax_scale
5001
5002
        self.attention_dropout_ctx = attention_dropout_ctx
        self.attention_dropout = attention_dropout
5003
5004
        self.attention_type = attention_type
        self.layer_number = 1 if layer_number is None else layer_number
5005
        self.deterministic = deterministic
5006
5007
5008
5009
        self.logger = logging.getLogger("FlashAttention")
        self.logger.setLevel(_log_level)
        if not self.logger.hasHandlers():
            self.logger.addHandler(_stream_handler)
5010
5011
5012
5013
5014
5015

    def forward(
        self,
        query_layer: torch.Tensor,
        key_layer: torch.Tensor,
        value_layer: torch.Tensor,
5016
        attention_mask: Optional[Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]] = None,
5017
5018
5019
        qkv_layout: str = "sbh3d",
        cu_seqlens_q: Optional[torch.Tensor] = None,
        cu_seqlens_kv: Optional[torch.Tensor] = None,
5020
5021
        max_seqlen_q: Optional[int] = None,
        max_seqlen_kv: Optional[int] = None,
5022
        attn_mask_type: str = "causal",
5023
        window_size: Optional[Tuple[int, int]] = None,
5024
        alibi_slopes: Optional[torch.Tensor] = None,
5025
        cp_group: Optional[Union[dist_group_type, List[dist_group_type]]] = None,
5026
        cp_global_ranks: List[int] = None,
5027
        cp_stream: torch.cuda.Stream = None,
5028
        cp_comm_type: str = "p2p",
5029
5030
        fp8: bool = False,
        fp8_meta: Optional[Dict[str, Any]] = None,
5031
5032
5033
    ) -> torch.Tensor:
        """flash-attn fprop"""

5034
5035
5036
5037
        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."
5038
5039
        assert (
            query_layer.is_cuda and key_layer.is_cuda and value_layer.is_cuda
5040
        ), "FlashAttention currently only supports CUDA tensors."
5041
5042
        assert (
            qkv_layout in QKVLayouts
5043
        ), f"FlashAttention does not support qkv_layout = {qkv_layout}!"
5044

5045
5046
5047
5048
5049
5050
        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)
5051
        context_parallel = cp_size > 1
5052

5053
        qkv_format = "".join([i for i in qkv_layout.split("_")[0] if i.isalpha()])
5054

5055
5056
5057
5058
5059
5060
5061
5062
5063
5064
5065
5066
5067
        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 = [
5068
                        x.transpose(0, 1) for x in (query_layer, key_layer, value_layer)
5069
                    ]
5070
            if context_parallel:
5071
                query_layer, key_layer, value_layer = [
5072
5073
5074
5075
5076
                    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 = [
5077
                    x.transpose(0, 1)
5078
5079
                    for x in (query_layer._data, key_layer._data, value_layer._data)
                ]
5080
5081
5082
5083
                query_layer, key_layer, value_layer = [
                    Float8Tensor.make_like(x, data=x._data)
                    for x in (query_layer, key_layer, value_layer)
                ]
5084
            if context_parallel:
5085
5086
                query_layer._data, key_layer._data, value_layer._data = [
                    x.contiguous() for x in (query_layer._data, key_layer._data, value_layer._data)
5087
                ]
5088

5089
        batch_size = query_layer.shape[0]
5090

5091
        if qkv_format in ["sbhd", "bshd"]:
5092
            max_seqlen_q, max_seqlen_kv = query_layer.shape[1], key_layer.shape[1]
5093
5094
            max_seqlen_q *= cp_size
            max_seqlen_kv *= cp_size
5095
5096
5097

            if "padding" in attn_mask_type:
                assert not context_parallel, "Padding mask not supported with context parallelism!"
5098
5099
                # [b * s, h, d]
                query_layer, key_layer, value_layer = [
5100
                    x.reshape(x.shape[0] * x.shape[1], *x.shape[2:])
5101
5102
5103
5104
5105
5106
5107
                    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."
5108
                    if cu_seqlens_q is None:
5109
5110
5111
                        assert (
                            attention_mask is not None
                        ), "Please provide attention_mask for padding!"
5112
5113
5114
5115
5116
5117
                        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
5118
5119
                    )
                else:
5120
                    if cu_seqlens_q is None or cu_seqlens_kv is None:
5121
5122
5123
5124
5125
                        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])
5126
5127
5128
5129
                    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)
5130
                    key_layer, value_layer = PackTensors.apply(indices_kv, key_layer, value_layer)
5131
            else:
5132
5133
5134
5135
5136
5137
5138
5139
5140
5141
5142
5143
5144
                # 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,
                    )
5145
5146
5147
5148
        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!"
5149
5150
5151
5152
5153
5154
            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()
5155

5156
5157
5158
        if context_parallel and all(
            not isinstance(x, Float8Tensor) for x in [query_layer, key_layer, value_layer]
        ):
5159
5160
5161
            assert (
                alibi_slopes is None
            ), "Alibi slope bias addition is not supported with context parallelism."
5162
            with self.attention_dropout_ctx():
5163
                output = attn_forward_func_with_cp(
5164
5165
5166
5167
5168
5169
5170
5171
                    self.training,
                    query_layer,
                    key_layer,
                    value_layer,
                    cu_seqlens_q,
                    cu_seqlens_kv,
                    max_seqlen_q,
                    max_seqlen_kv,
5172
5173
                    cu_seqlens_q,
                    cu_seqlens_kv,
5174
                    self.attention_dropout if self.training else 0.0,
5175
5176
5177
                    cp_group,
                    cp_global_ranks,
                    cp_stream,
5178
                    cp_comm_type,
5179
                    softmax_scale=self.softmax_scale,
5180
                    qkv_format="bshd" if qkv_format == "sbhd" else qkv_format,
5181
                    attn_mask_type=attn_mask_type,
5182
                    deterministic=self.deterministic,
5183
                    window_size=window_size,
5184
5185
                )
        else:
5186
5187

            from .cpu_offload import CPUOffloadEnabled
5188

5189
5190
5191
5192
5193
5194
            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

5195
            with self.attention_dropout_ctx():
5196
                fa_optional_forward_kwargs = {}
5197
5198
                if _flash_attn_2_3_plus:
                    fa_optional_forward_kwargs["window_size"] = window_size
5199
5200
5201
5202
                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
5203
5204
5205
5206
                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:
5207
5208
                    if _flash_attn_2_5_7_plus:
                        fa_optional_forward_kwargs["block_table"] = None
5209
5210
5211
5212
5213
5214
5215
5216
5217
5218
                    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:
5219
5220
5221
                    fa_3_optional_forward_kwargs = {}
                    fa_3_optional_forward_kwargs["window_size"] = window_size
                    fa_3_optional_forward_kwargs["deterministic"] = self.deterministic
5222
5223
5224
5225
                    if fp8:
                        fp8_dtype_forward = get_fp8_te_dtype(fp8_meta["recipe"], fprop_tensor=True)
                        activation_dtype = query_layer.dtype
                        torch_dtype = get_fp8_torch_dtype(fp8_meta["recipe"], fprop_tensor=True)
5226
5227
5228
5229
5230
5231
5232
5233
5234
5235
5236

                        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

5237
5238
5239
5240
5241
5242
5243
5244
                        if fp8_meta["recipe"].fp8_mha:
                            assert all(
                                isinstance(x, Float8Tensor)
                                for x in [query_layer, key_layer, value_layer]
                            ), "q/k/v must be Float8Tensors for FP8 MHA."
                            fp8_meta["scaling_fwd"].scale_inv[META_QKV] = query_layer._scale_inv
                        else:
                            query_layer, key_layer, value_layer = (
5245
5246
                                Float8Tensor.to_float8(x, fp8_dtype=fp8_dtype_forward)
                                for x in [query_layer, key_layer, value_layer]
5247
                            )
5248
5249
5250
5251
5252
5253
5254
5255
5256
5257
5258
5259
5260
5261
5262
5263
5264
5265
5266
5267
5268
5269
5270
5271
5272
5273
5274
                        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]
                                + ". Please update your FlashAttention 3 (beta) installation as it "
                                + "may have added more supported arguments to its API. \n"
                                + _flash_attn_3_installation_steps,
                            ) + e.args[1:]
                        raise

5275
5276
5277
5278
5279
5280
5281
5282
5283
5284
5285
5286
5287
5288
5289
5290
5291
5292
5293
5294
5295
5296
5297
5298
5299
5300
                    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,
                    )
5301

5302
        if qkv_format in ["sbhd", "bshd"] and "padding" in attn_mask_type:
5303
            output = UnpackTensor.apply(indices_q, batch_size * max_seqlen_q, output)
5304

5305
        if qkv_format == "sbhd":
5306
            # (bs)hd -> bs(hd) -> sb(hd)
5307
            if fp8 and fp8_meta["recipe"].fp8_mha:
5308
5309
5310
5311
5312
5313
                output = Float8Tensor.make_like(
                    output,
                    data=output._data.reshape(batch_size, max_seqlen_q // cp_size, -1)
                    .transpose(0, 1)
                    .contiguous(),
                )
5314
            else:
5315
                output = output.view(batch_size, max_seqlen_q // cp_size, -1).transpose(0, 1)
5316
        elif qkv_format == "bshd":
5317
            # (bs)hd -> bs(hd)
5318
            output = output.reshape(batch_size, max_seqlen_q // cp_size, -1)
5319
        elif qkv_format == "thd":
5320
            # thd -> t(hd)
5321
            output = output.reshape(output.shape[0], -1)
5322

5323
        return output.contiguous()
5324

5325

5326
def _combine_tensors(
5327
5328
5329
    tensors: List[torch.Tensor],
    dim: int,
) -> torch.Tensor:
5330
5331
5332
5333
5334
5335
    """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())
5336
    new_stride.insert(dim, int(new_stride[dim - 1] / num_tensors))
5337
    if isinstance(tensors[0], Float8Tensor):
5338
        combined_tensor = torch.Tensor().to(device=tensors[0].device, dtype=tensors[0]._data.dtype)
5339
5340
5341
        combined_tensor.set_(
            tensors[0]._data.untyped_storage(),
            tensors[0]._data.storage_offset(),
5342
5343
5344
5345
            new_shape,
            new_stride,
        )
        combined_tensor = Float8Tensor.make_like(tensors[0], data=combined_tensor)
5346
    else:
5347
        combined_tensor = torch.Tensor().to(device=tensors[0].device, dtype=tensors[0].dtype)
5348
        combined_tensor.set_(
5349
5350
            tensors[0].untyped_storage(), tensors[0].storage_offset(), new_shape, new_stride
        )
5351
5352

    return combined_tensor
5353

5354

5355
5356
5357
5358
class FusedAttnFunc_qkvpacked(torch.autograd.Function):
    """Function for FusedAttention with packed QKV input"""

    @staticmethod
5359
5360
5361
5362
5363
    def forward(
        ctx,
        is_training,
        max_seqlen,
        cu_seqlens,
5364
        cu_seqlens_padded,
5365
5366
5367
5368
5369
5370
5371
5372
5373
        qkv,
        qkv_dtype,
        attn_bias,
        attn_scale,
        dropout_p,
        fast_zero_fill,
        qkv_layout,
        attn_bias_type,
        attn_mask_type,
5374
        window_size,
5375
5376
5377
5378
5379
        rng_gen,
        fused_attention_backend,
        use_FAv2_bwd,
        fp8,
        fp8_meta,
5380
        deterministic,
5381
    ):
5382
5383
        is_input_fp8 = False
        is_output_fp8 = fp8_meta["recipe"].fp8_mha
5384
        if fp8:
5385
5386
            is_input_fp8 = isinstance(qkv, Float8Tensor)
            if is_input_fp8:
5387
5388
5389
5390
                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
5391
            qkv_group = len(qkv_layout.split("_"))
5392
5393
5394
5395
            assert (
                qkv_group == 1
            ), f"qkv layout should conform to 3hd or h3d, e.g. sb3hd, but found {qkv_layout}."
            if is_input_fp8:
5396
5397
5398
                qkv_fp8 = qkv._data
            else:
                qkv_c = qkv.view(-1, qkv.shape[-3] * qkv.shape[-2] * qkv.shape[-1])
5399
5400
5401
                qkv_fp8 = cast_to_fp8(
                    qkv_c, fp8_meta["scaling_fwd"], META_QKV, fp8_dtype_forward
                ).view(qkv.shape)
5402
            out_fp8, aux_ctx_tensors = fused_attn_fwd_qkvpacked(
5403
5404
5405
5406
5407
5408
5409
                is_training,
                max_seqlen,
                cu_seqlens,
                qkv_fp8,
                fp8_dtype_forward,
                fused_attention_backend,
                attn_bias,
5410
                cu_seqlens_padded,
5411
5412
5413
5414
5415
5416
5417
5418
5419
5420
5421
5422
                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
5423
5424
5425
5426
5427
5428
                attn_scale,
                dropout_p,
                fast_zero_fill,
                qkv_layout,
                attn_bias_type,
                attn_mask_type,
5429
                window_size,
5430
5431
                rng_gen,
            )
5432
            if is_output_fp8:
5433
5434
                out_ret = Float8Tensor(
                    data=out_fp8,
5435
5436
5437
5438
5439
5440
5441
5442
5443
                    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]),
5444
5445
5446
5447
5448
                    fp8_meta["scaling_fwd"],
                    META_O,
                    fp8_dtype_forward,
                    qkv_dtype,
                ).view(out_fp8.shape)
5449
            out_save = out_ret
5450
5451
5452
5453
5454
5455
5456
5457
5458
5459
5460
5461
5462
5463
5464
5465
5466
5467
            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)
5468
5469
5470
            fp8_tensors = (
                qkv_fp8,
                out_fp8,
5471
                fp8_meta["scaling_fwd"].scale.clone(),
5472
5473
                fp8_meta["scaling_fwd"].scale_inv.clone(),
            )
5474
5475
        else:
            out_ret, aux_ctx_tensors = fused_attn_fwd_qkvpacked(
5476
5477
5478
5479
5480
5481
5482
                is_training,
                max_seqlen,
                cu_seqlens,
                qkv,
                qkv_dtype,
                fused_attention_backend,
                attn_bias,
5483
                cu_seqlens_padded,
5484
5485
5486
5487
5488
5489
5490
5491
5492
5493
5494
5495
                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
5496
5497
5498
5499
5500
5501
                attn_scale,
                dropout_p,
                fast_zero_fill,
                qkv_layout,
                attn_bias_type,
                attn_mask_type,
5502
                window_size,
5503
5504
                rng_gen,
            )
5505
5506
5507
5508
            fp8_tensors = (None, None, None, None)
            out_save = out_ret

        ctx.fp8 = fp8 and int(os.getenv("NVTE_FP8_DPA_BWD", "1"))
5509
5510
        ctx.is_input_fp8 = is_input_fp8
        ctx.is_output_fp8 = is_output_fp8
5511
        qkvo_tensors = (qkv, out_save) if not ctx.fp8 else (None, None)
5512
        ctx.save_for_backward(
5513
            *qkvo_tensors, cu_seqlens, cu_seqlens_padded, *fp8_tensors, *aux_ctx_tensors
5514
        )
5515
        ctx.fp8_meta = fp8_meta
5516
5517
5518
5519
5520
5521
5522
5523
        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
5524
        ctx.window_size = window_size
5525
        ctx.fused_attention_backend = (
5526
            fused_attention_backend if ctx.fp8 else FusedAttnBackend["F16_arbitrary_seqlen"]
5527
        )
5528
        ctx.use_FAv2_bwd = use_FAv2_bwd
5529
        ctx.deterministic = deterministic
5530

5531
        return out_ret
5532
5533
5534

    @staticmethod
    def backward(ctx, d_out):
5535
        if ctx.is_output_fp8:
5536
5537
5538
            assert isinstance(
                d_out, Float8Tensor
            ), "Gradient of the DPA output must be in Float8Tensor type for FP8 MHA."
5539
5540
5541
            d_out_f8tensor = d_out
            d_out = d_out._data

5542
        d_out = d_out.contiguous()
5543
5544
5545
5546
        (
            qkv,
            out,
            cu_seqlens,
5547
            cu_seqlens_padded,
5548
5549
5550
5551
5552
5553
            qkv_fp8,
            out_fp8,
            fwd_scales,
            fwd_scale_invs,
            *aux_ctx_tensors,
        ) = ctx.saved_tensors
5554
5555
        if not aux_ctx_tensors[0].is_contiguous():
            aux_ctx_tensors[0] = aux_ctx_tensors[0].contiguous()
5556
        if ctx.use_FAv2_bwd:
5557
            softmax_lse, rng_state = aux_ctx_tensors
5558
5559
            dqkv = torch.empty_like(qkv)
            maybe_contiguous = lambda x: x.contiguous() if x.stride(-1) != 1 else x
5560
5561
5562
            d_out, q, k, v, out = [
                maybe_contiguous(x) for x in (d_out, qkv[:, 0], qkv[:, 1], qkv[:, 2], out)
            ]
5563
            flash_attn_cuda_bwd(
5564
5565
5566
5567
5568
5569
5570
5571
5572
5573
5574
5575
5576
5577
5578
5579
5580
5581
5582
                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,
5583
            )
5584
            dqkv = dqkv[..., : d_out.shape[-1]]
5585
        else:
5586
5587
            with torch.cuda.nvtx.range("_FusedAttn_qkvpacked"):
                if ctx.fp8:
5588
                    fp8_dtype_forward = get_fp8_te_dtype(ctx.fp8_meta["recipe"], fprop_tensor=True)
5589
                    fp8_dtype_backward = get_fp8_te_dtype(
5590
5591
                        ctx.fp8_meta["recipe"], fprop_tensor=False
                    )
5592
                    if ctx.is_output_fp8:
5593
                        d_out_fp8 = d_out
5594
                        ctx.fp8_meta["scaling_bwd"].scale_inv[META_DO] = d_out_f8tensor._scale_inv
5595
5596
5597
                    else:
                        d_out_fp8 = cast_to_fp8(
                            d_out.view(-1, d_out.shape[-2] * d_out.shape[-1]),
5598
5599
5600
5601
                            ctx.fp8_meta["scaling_bwd"],
                            META_DO,
                            fp8_dtype_backward,
                        ).view(d_out.shape)
5602
                    dqkv_fp8, *rest = fused_attn_bwd_qkvpacked(
5603
5604
5605
5606
5607
5608
5609
5610
                        ctx.max_seqlen,
                        cu_seqlens,
                        qkv_fp8,
                        out_fp8,
                        d_out_fp8,
                        fp8_dtype_forward,
                        fp8_dtype_backward,
                        aux_ctx_tensors,
5611
                        ctx.fused_attention_backend,
5612
                        cu_seqlens_padded,
5613
5614
5615
5616
5617
5618
5619
5620
5621
5622
5623
5624
5625
5626
5627
5628
                        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,
5629
5630
                        ctx.window_size,
                        ctx.deterministic,
5631
                    )
5632
                    if ctx.is_input_fp8:
5633
5634
                        dqkv = Float8Tensor(
                            data=dqkv_fp8,
5635
5636
5637
5638
5639
                            fp8_meta=ctx.fp8_meta,
                            fp8_meta_forward=False,
                            fp8_meta_index=META_DQKV,
                            fp8_dtype=fp8_dtype_backward,
                            dtype=d_out_f8tensor.dtype,
5640
                        )
5641
                    else:
5642
5643
5644
5645
5646
5647
5648
5649
5650
5651
                        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)
5652
5653
5654
5655
                else:
                    if d_out.dtype == torch.uint8:
                        d_out = d_out_f8tensor.from_float8(qkv.dtype)
                    dqkv, *rest = fused_attn_bwd_qkvpacked(
5656
5657
5658
5659
5660
5661
5662
5663
                        ctx.max_seqlen,
                        cu_seqlens,
                        qkv,
                        out,
                        d_out,
                        ctx.qkv_dtype,
                        ctx.qkv_dtype,
                        aux_ctx_tensors,
5664
                        ctx.fused_attention_backend,
5665
                        cu_seqlens_padded,
5666
5667
5668
5669
5670
5671
5672
5673
5674
5675
5676
5677
5678
5679
5680
5681
                        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,
5682
5683
                        ctx.window_size,
                        ctx.deterministic,
5684
                    )
5685

5686
5687
        # if no_bias or alibi, return dqkv
        if ctx.attn_bias_type in ["no_bias", "alibi"]:
5688
5689
5690
5691
5692
5693
5694
5695
5696
5697
5698
5699
5700
5701
5702
5703
5704
5705
5706
5707
5708
            return (
                None,
                None,
                None,
                None,
                dqkv,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
                None,
5709
5710
                None,
                None,
5711
            )
5712
        # else, return (dqkv, dbias)
5713
5714
5715
5716
5717
5718
5719
5720
5721
5722
5723
5724
5725
5726
5727
5728
5729
5730
5731
5732
5733
        return (
            None,
            None,
            None,
            None,
            dqkv,
            None,
            rest[0],
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
5734
5735
            None,
            None,
5736
        )
5737

5738

5739
5740
5741
5742
class FusedAttnFunc_kvpacked(torch.autograd.Function):
    """Function for FusedAttention with packed KV input"""

    @staticmethod
5743
5744
5745
5746
5747
5748
5749
    def forward(
        ctx,
        is_training,
        max_seqlen_q,
        max_seqlen_kv,
        cu_seqlens_q,
        cu_seqlens_kv,
5750
5751
        cu_seqlens_q_padded,
        cu_seqlens_kv_padded,
5752
5753
5754
5755
5756
5757
5758
5759
5760
5761
        q,
        kv,
        qkv_dtype,
        attn_bias,
        attn_scale,
        dropout_p,
        fast_zero_fill,
        qkv_layout,
        attn_bias_type,
        attn_mask_type,
5762
        window_size,
5763
5764
5765
5766
5767
        rng_gen,
        fused_attention_backend,
        use_FAv2_bwd,
        fp8,
        fp8_meta,
5768
        deterministic,
5769
    ):
5770
5771
        is_input_fp8 = False
        is_output_fp8 = fp8_meta["recipe"].fp8_mha
5772
        if fp8:
5773
5774
5775
            assert isinstance(kv, q.__class__), "q and kv must have the same type."
            is_input_fp8 = isinstance(q, Float8Tensor)
            if is_input_fp8:
5776
5777
5778
                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)
5779
            if is_input_fp8:
5780
5781
5782
                q_fp8, kv_fp8 = q._data, kv._data
            else:
                # 1: qkv packed, 2: kv packed, 3: qkv separate
5783
5784
                qkv_group = len(qkv_layout.split("_"))
                assert qkv_group == 2, (
5785
5786
                    "qkv layout should conform to hd_2hd or hd_h2d, e.g. sbhd_sb2hd, "
                    f"but found {qkv_layout}."
5787
5788
5789
5790
                )
                q_fp8 = cast_to_fp8(q, fp8_meta["scaling_fwd"], META_QKV, fp8_dtype_forward).view(
                    q.shape
                )
5791
                kv_c = kv.view(-1, kv.shape[-3] * kv.shape[-2] * kv.shape[-1])
5792
5793
5794
                kv_fp8 = cast_to_fp8(
                    kv_c, fp8_meta["scaling_fwd"], META_QKV, fp8_dtype_forward
                ).view(kv.shape)
5795
            out_fp8, aux_ctx_tensors = fused_attn_fwd_kvpacked(
5796
5797
5798
5799
5800
5801
5802
5803
5804
5805
                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,
5806
5807
                cu_seqlens_q_padded,
                cu_seqlens_kv_padded,
5808
5809
5810
5811
5812
5813
5814
5815
5816
5817
5818
5819
                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
5820
5821
5822
5823
5824
5825
                attn_scale,
                dropout_p,
                fast_zero_fill,
                qkv_layout,
                attn_bias_type,
                attn_mask_type,
5826
                window_size,
5827
5828
                rng_gen,
            )
5829
            if is_output_fp8:
5830
5831
                out_ret = Float8Tensor(
                    data=out_fp8,
5832
5833
5834
5835
5836
5837
5838
5839
5840
                    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]),
5841
5842
5843
5844
5845
                    fp8_meta["scaling_fwd"],
                    META_O,
                    fp8_dtype_forward,
                    qkv_dtype,
                ).view(out_fp8.shape)
5846
            out_save = out_ret
5847
5848
5849
5850
5851
5852
5853
5854
5855
5856
5857
5858
5859
5860
5861
5862
5863
5864
5865
5866
5867
5868
5869
5870
5871
            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)
5872
5873
5874
5875
            fp8_tensors = (
                q_fp8,
                kv_fp8,
                out_fp8,
5876
                fp8_meta["scaling_fwd"].scale.clone(),
5877
5878
                fp8_meta["scaling_fwd"].scale_inv.clone(),
            )
5879
5880
        else:
            out_ret, aux_ctx_tensors = fused_attn_fwd_kvpacked(
5881
5882
5883
5884
5885
5886
5887
5888
5889
5890
                is_training,
                max_seqlen_q,
                max_seqlen_kv,
                cu_seqlens_q,
                cu_seqlens_kv,
                q,
                kv,
                qkv_dtype,
                fused_attention_backend,
                attn_bias,
5891
5892
                cu_seqlens_q_padded,
                cu_seqlens_kv_padded,
5893
5894
5895
5896
5897
5898
5899
5900
5901
5902
5903
5904
                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
5905
5906
5907
5908
5909
5910
                attn_scale,
                dropout_p,
                fast_zero_fill,
                qkv_layout,
                attn_bias_type,
                attn_mask_type,
5911
                window_size,
5912
5913
                rng_gen,
            )
5914
5915
5916
5917
            out_save = out_ret
            fp8_tensors = (None, None, None, None, None)

        ctx.fp8 = fp8 and int(os.getenv("NVTE_FP8_DPA_BWD", "1"))
5918
5919
        ctx.is_input_fp8 = is_input_fp8
        ctx.is_output_fp8 = is_output_fp8
5920
        qkvo_tensors = (q, kv, out_save) if not ctx.fp8 else (None, None, None)
5921
5922
5923
5924
        ctx.save_for_backward(
            *qkvo_tensors,
            cu_seqlens_q,
            cu_seqlens_kv,
5925
5926
            cu_seqlens_q_padded,
            cu_seqlens_kv_padded,
5927
5928
5929
            *fp8_tensors,
            *aux_ctx_tensors,
        )
5930
        ctx.fp8_meta = fp8_meta
5931
5932
5933
5934
5935
5936
5937
5938
5939
        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
5940
        ctx.window_size = window_size
5941
        ctx.fused_attention_backend = (
5942
            fused_attention_backend if ctx.fp8 else FusedAttnBackend["F16_arbitrary_seqlen"]
5943
        )
5944
        ctx.use_FAv2_bwd = use_FAv2_bwd
5945
        ctx.deterministic = deterministic
5946

5947
        return out_ret
5948
5949
5950

    @staticmethod
    def backward(ctx, d_out):
5951
        if ctx.is_output_fp8:
5952
5953
5954
            assert isinstance(
                d_out, Float8Tensor
            ), "Gradient of the DPA output must be in Float8Tensor type for FP8 MHA."
5955
5956
5957
            d_out_f8tensor = d_out
            d_out = d_out._data

5958
        d_out = d_out.contiguous()
5959
5960
5961
5962
5963
5964
        (
            q,
            kv,
            out,
            cu_seqlens_q,
            cu_seqlens_kv,
5965
5966
            cu_seqlens_q_padded,
            cu_seqlens_kv_padded,
5967
5968
5969
5970
5971
5972
5973
            q_fp8,
            kv_fp8,
            out_fp8,
            fwd_scales,
            fwd_scale_invs,
            *aux_ctx_tensors,
        ) = ctx.saved_tensors
5974
5975
        if not aux_ctx_tensors[0].is_contiguous():
            aux_ctx_tensors[0] = aux_ctx_tensors[0].contiguous()
5976
        if ctx.use_FAv2_bwd:
5977
            softmax_lse, rng_state = aux_ctx_tensors
5978
5979
5980
            dq = torch.empty_like(q)
            dkv = torch.empty_like(kv)
            maybe_contiguous = lambda x: x.contiguous() if x.stride(-1) != 1 else x
5981
            d_out, q, k, v, out = [maybe_contiguous(x) for x in (d_out, q, kv[:, 0], kv[:, 1], out)]
5982
            flash_attn_cuda_bwd(
5983
5984
5985
5986
5987
5988
5989
5990
5991
5992
5993
5994
5995
5996
5997
5998
5999
6000
6001
                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,
6002
            )
6003
6004
            dq = dq[..., : d_out.shape[-1]]
            dkv = dkv[..., : d_out.shape[-1]]
6005
        else:
6006
6007
            with torch.cuda.nvtx.range("_FusedAttn_kvpacked"):
                if ctx.fp8:
6008
                    fp8_dtype_forward = get_fp8_te_dtype(ctx.fp8_meta["recipe"], fprop_tensor=True)
6009
                    fp8_dtype_backward = get_fp8_te_dtype(
6010
6011
                        ctx.fp8_meta["recipe"], fprop_tensor=False
                    )
6012
                    if ctx.is_output_fp8:
6013
                        d_out_fp8 = d_out
6014
                        ctx.fp8_meta["scaling_bwd"].scale_inv[META_DO] = d_out_f8tensor._scale_inv
6015
6016
6017
                    else:
                        d_out_fp8 = cast_to_fp8(
                            d_out.view(-1, d_out.shape[-2] * d_out.shape[-1]),
6018
6019
6020
6021
                            ctx.fp8_meta["scaling_bwd"],
                            META_DO,
                            fp8_dtype_backward,
                        ).view(d_out.shape)
6022
                    dq_fp8, dkv_fp8, *rest = fused_attn_bwd_kvpacked(
6023
6024
6025
6026
6027
6028
6029
6030
6031
6032
6033
                        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,
6034
                        ctx.fused_attention_backend,
6035
6036
                        cu_seqlens_q_padded,
                        cu_seqlens_kv_padded,
6037
6038
6039
6040
6041
6042
6043
6044
6045
6046
6047
6048
6049
6050
6051
6052
                        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,
6053
6054
                        ctx.window_size,
                        ctx.deterministic,
6055
                    )
6056
                    if ctx.is_input_fp8:
6057
6058
                        dq = Float8Tensor(
                            data=dq_fp8,
6059
6060
6061
6062
6063
                            fp8_meta=ctx.fp8_meta,
                            fp8_meta_forward=False,
                            fp8_meta_index=META_DQKV,
                            fp8_dtype=fp8_dtype_backward,
                            dtype=d_out_f8tensor.dtype,
6064
6065
6066
                        )
                        dkv = Float8Tensor(
                            data=dkv_fp8,
6067
6068
6069
6070
6071
                            fp8_meta=ctx.fp8_meta,
                            fp8_meta_forward=False,
                            fp8_meta_index=META_DQKV,
                            fp8_dtype=fp8_dtype_backward,
                            dtype=d_out_f8tensor.dtype,
6072
                        )
6073
6074
6075
                    else:
                        dq = cast_from_fp8(
                            dq_fp8.view(-1, dq_fp8.shape[-2] * dq_fp8.shape[-1]),
6076
6077
6078
6079
6080
6081
6082
6083
6084
6085
6086
6087
6088
6089
6090
                            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)
6091
6092
6093
6094
                else:
                    if d_out.dtype == torch.uint8:
                        d_out = d_out_f8tensor.from_float8(q.dtype)
                    dq, dkv, *rest = fused_attn_bwd_kvpacked(
6095
6096
6097
6098
6099
6100
6101
6102
6103
6104
6105
                        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,
6106
                        ctx.fused_attention_backend,
6107
6108
                        cu_seqlens_q_padded,
                        cu_seqlens_kv_padded,
6109
6110
6111
6112
6113
6114
6115
6116
6117
6118
6119
6120
6121
6122
6123
6124
                        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,
6125
6126
                        ctx.window_size,
                        ctx.deterministic,
6127
                    )
6128

6129
6130
        # if no_bias or alibi, return dqkv
        if ctx.attn_bias_type in ["no_bias", "alibi"]:
6131
6132
6133
6134
6135
6136
6137
6138
6139
6140
6141
6142
6143
6144
6145
6146
6147
6148
6149
6150
6151
6152
6153
6154
6155
            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,
6156
6157
                None,
                None,
6158
            )
6159
        # else, return (dqkv, dbias)
6160
6161
6162
6163
6164
6165
6166
6167
6168
6169
6170
6171
6172
6173
6174
6175
6176
6177
6178
6179
6180
6181
6182
6183
6184
        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,
6185
6186
            None,
            None,
6187
6188
        )

6189

6190
6191
6192
6193
class FusedAttnFunc(torch.autograd.Function):
    """Function for FusedAttention with separate Q, K, V tensors"""

    @staticmethod
6194
6195
6196
6197
6198
6199
6200
    def forward(
        ctx,
        is_training,
        max_seqlen_q,
        max_seqlen_kv,
        cu_seqlens_q,
        cu_seqlens_kv,
6201
6202
        cu_seqlens_q_padded,
        cu_seqlens_kv_padded,
6203
6204
6205
6206
6207
6208
6209
6210
6211
6212
6213
        q,
        k,
        v,
        qkv_dtype,
        attn_bias,
        attn_scale,
        dropout_p,
        fast_zero_fill,
        qkv_layout,
        attn_bias_type,
        attn_mask_type,
6214
        window_size,
6215
6216
6217
6218
6219
        rng_gen,
        fused_attention_backend,
        use_FAv2_bwd,
        fp8,
        fp8_meta,
6220
        deterministic,
6221
    ):
6222
6223
        is_input_fp8 = False
        is_output_fp8 = fp8_meta["recipe"].fp8_mha
6224
6225
6226
        if fp8:
            fused_attention_backend = FusedAttnBackend["FP8"]
            fp8_dtype_forward = get_fp8_te_dtype(fp8_meta["recipe"], fprop_tensor=True)
6227
6228
6229
6230
6231
            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:
6232
6233
6234
6235
                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
6236
                qkv_group = len(qkv_layout.split("_"))
6237
                if qkv_group == 1:
6238
6239
                    dim = qkv_layout.find("3")
                    qkv = _combine_tensors([q, k, v], dim)
6240
                    qkv_c = qkv.view(-1, qkv.shape[-3] * qkv.shape[-2] * qkv.shape[-1])
6241
6242
6243
6244
                    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])
6245
6246
                    q_fp8, k_fp8, v_fp8 = [x.squeeze(dim) for x in [q_fp8, k_fp8, v_fp8]]
                if qkv_group == 2:
6247
6248
6249
6250
6251
                    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)
6252
                    kv_c = kv.view(-1, kv.shape[-3] * kv.shape[-2] * kv.shape[-1])
6253
6254
6255
6256
                    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])
6257
6258
                    k_fp8, v_fp8 = [x.squeeze(dim) for x in [k_fp8, v_fp8]]
                if qkv_group == 3:
6259
6260
6261
6262
6263
6264
6265
6266
6267
                    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)
6268
            out_fp8, aux_ctx_tensors = fused_attn_fwd(
6269
6270
6271
6272
6273
6274
6275
6276
6277
6278
6279
                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,
6280
6281
                cu_seqlens_q_padded,
                cu_seqlens_kv_padded,
6282
6283
6284
6285
6286
6287
6288
6289
6290
6291
6292
6293
                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
6294
6295
6296
6297
6298
6299
                attn_scale,
                dropout_p,
                fast_zero_fill,
                qkv_layout,
                attn_bias_type,
                attn_mask_type,
6300
                window_size,
6301
6302
                rng_gen,
            )
6303
            if is_output_fp8:
6304
6305
                out_ret = Float8Tensor(
                    data=out_fp8,
6306
6307
6308
6309
6310
6311
6312
6313
6314
                    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]),
6315
6316
6317
6318
6319
                    fp8_meta["scaling_fwd"],
                    META_O,
                    fp8_dtype_forward,
                    qkv_dtype,
                ).view(out_fp8.shape)
6320
6321
            out_save = out_ret

6322
            if not int(os.getenv("NVTE_FP8_DPA_BWD", "1")):
6323
                # 1: qkv packed, 2: kv packed, 3: qkv separate
6324
6325
6326
6327
6328
6329
6330
6331
6332
6333
6334
6335
6336
6337
6338
6339
6340
6341
6342
6343
6344
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
6370
6371
6372
6373
6374
6375
6376
6377
6378
6379
6380
6381
6382
6383
                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]),
6384
                        fp8_meta["scaling_fwd"],
6385
                        META_O,
6386
                        fp8_dtype_forward,
6387
6388
                        qkv_dtype,
                    ).view(out_fp8.shape)
6389
6390
6391
6392
6393
6394

            fp8_tensors = (
                q_fp8,
                k_fp8,
                v_fp8,
                out_fp8,
6395
                fp8_meta["scaling_fwd"].scale.clone(),
6396
6397
                fp8_meta["scaling_fwd"].scale_inv.clone(),
            )
6398
6399
        else:
            out_ret, aux_ctx_tensors = fused_attn_fwd(
6400
6401
6402
6403
6404
6405
6406
6407
6408
6409
6410
                is_training,
                max_seqlen_q,
                max_seqlen_kv,
                cu_seqlens_q,
                cu_seqlens_kv,
                q,
                k,
                v,
                qkv_dtype,
                fused_attention_backend,
                attn_bias,
6411
6412
                cu_seqlens_q_padded,
                cu_seqlens_kv_padded,
6413
6414
6415
6416
6417
6418
6419
6420
6421
6422
6423
6424
                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
6425
6426
6427
6428
6429
6430
                attn_scale,
                dropout_p,
                fast_zero_fill,
                qkv_layout,
                attn_bias_type,
                attn_mask_type,
6431
                window_size,
6432
6433
                rng_gen,
            )
6434
6435
            out_save = out_ret
            fp8_tensors = (None, None, None, None, None, None)
6436

6437
6438
        ctx.fp8 = fp8 and int(os.getenv("NVTE_FP8_DPA_BWD", "1"))

6439
        from .cpu_offload import CPUOffloadEnabled
6440

6441
        if CPUOffloadEnabled:
6442
6443
6444
6445
6446
6447
6448
            if ctx.fp8:
                tensor_list = fp8_tensors
            else:
                tensor_list = [q, k, v, out_save]

            tensor_list.extend(aux_ctx_tensors)

6449
            qkv_layout = "sbhd_sbhd_sbhd"
6450
6451
6452
6453
            for tensor in tensor_list:
                if tensor is not None:
                    tensor.activation_offloading = True

6454
6455
        ctx.is_input_fp8 = is_input_fp8
        ctx.is_output_fp8 = is_output_fp8
6456
        qkvo_tensors = (q, k, v, out_save) if not ctx.fp8 else (None, None, None, None)
6457
6458
6459
6460
        ctx.save_for_backward(
            *qkvo_tensors,
            cu_seqlens_q,
            cu_seqlens_kv,
6461
6462
            cu_seqlens_q_padded,
            cu_seqlens_kv_padded,
6463
6464
6465
            *fp8_tensors,
            *aux_ctx_tensors,
        )
6466
        ctx.fp8_meta = fp8_meta
6467
6468
6469
6470
6471
6472
6473
6474
6475
        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
6476
        ctx.window_size = window_size
6477
        ctx.fused_attention_backend = (
6478
            fused_attention_backend if ctx.fp8 else FusedAttnBackend["F16_arbitrary_seqlen"]
6479
        )
6480
        ctx.use_FAv2_bwd = use_FAv2_bwd
6481
        ctx.deterministic = deterministic
6482

6483
        return out_ret
6484
6485
6486

    @staticmethod
    def backward(ctx, d_out):
6487
        if ctx.is_output_fp8:
6488
6489
6490
            assert isinstance(
                d_out, Float8Tensor
            ), "Gradient of the DPA output must be in Float8Tensor type for FP8 MHA."
6491
6492
6493
            d_out_f8tensor = d_out
            d_out = d_out._data

6494
        d_out = d_out.contiguous()
6495
6496
6497
6498
6499
6500
6501
        (
            q,
            k,
            v,
            out,
            cu_seqlens_q,
            cu_seqlens_kv,
6502
6503
            cu_seqlens_q_padded,
            cu_seqlens_kv_padded,
6504
6505
6506
6507
6508
6509
6510
6511
            q_fp8,
            k_fp8,
            v_fp8,
            out_fp8,
            fwd_scales,
            fwd_scale_invs,
            *aux_ctx_tensors,
        ) = ctx.saved_tensors
6512
6513
        if not aux_ctx_tensors[0].is_contiguous():
            aux_ctx_tensors[0] = aux_ctx_tensors[0].contiguous()
6514
        if ctx.use_FAv2_bwd:
6515
            softmax_lse, rng_state = aux_ctx_tensors
6516
6517
6518
6519
            dq = torch.empty_like(q)
            dk = torch.empty_like(k)
            dv = torch.empty_like(v)
            maybe_contiguous = lambda x: x.contiguous() if x.stride(-1) != 1 else x
6520
            d_out, q, k, v, out = [maybe_contiguous(x) for x in (d_out, q, k, v, out)]
6521
            flash_attn_cuda_bwd(
6522
6523
6524
6525
6526
6527
6528
6529
6530
6531
6532
6533
6534
6535
6536
6537
6538
6539
6540
                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,
6541
            )
6542
6543
6544
            dq = dq[..., : d_out.shape[-1]]
            dk = dk[..., : d_out.shape[-1]]
            dv = dv[..., : d_out.shape[-1]]
6545
        else:
6546
6547
6548
6549
            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(
6550
6551
                        ctx.fp8_meta["recipe"], fprop_tensor=False
                    )
6552
                    if ctx.is_output_fp8:
6553
                        d_out_fp8 = d_out
6554
                        ctx.fp8_meta["scaling_bwd"].scale_inv[META_DO] = d_out_f8tensor._scale_inv
6555
6556
6557
                    else:
                        d_out_fp8 = cast_to_fp8(
                            d_out.view(-1, d_out.shape[-2] * d_out.shape[-1]),
6558
6559
6560
6561
                            ctx.fp8_meta["scaling_bwd"],
                            META_DO,
                            fp8_dtype_backward,
                        ).view(d_out.shape)
6562
                    dq_fp8, dk_fp8, dv_fp8, *rest = fused_attn_bwd(
6563
6564
6565
6566
6567
6568
6569
6570
6571
6572
6573
6574
                        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,
6575
                        ctx.fused_attention_backend,
6576
6577
                        cu_seqlens_q_padded,
                        cu_seqlens_kv_padded,
6578
6579
6580
6581
6582
6583
6584
6585
6586
6587
6588
6589
6590
6591
6592
6593
                        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,
6594
6595
                        ctx.window_size,
                        ctx.deterministic,
6596
                    )
6597

6598
                    if ctx.is_input_fp8:
6599
6600
                        dq = Float8Tensor(
                            data=dq_fp8,
6601
6602
6603
6604
6605
                            fp8_meta=ctx.fp8_meta,
                            fp8_meta_forward=False,
                            fp8_meta_index=META_DQKV,
                            fp8_dtype=fp8_dtype_backward,
                            dtype=d_out_f8tensor.dtype,
6606
6607
6608
                        )
                        dk = Float8Tensor(
                            data=dk_fp8,
6609
6610
6611
6612
6613
                            fp8_meta=ctx.fp8_meta,
                            fp8_meta_forward=False,
                            fp8_meta_index=META_DQKV,
                            fp8_dtype=fp8_dtype_backward,
                            dtype=d_out_f8tensor.dtype,
6614
6615
6616
                        )
                        dv = Float8Tensor(
                            data=dv_fp8,
6617
6618
6619
6620
6621
                            fp8_meta=ctx.fp8_meta,
                            fp8_meta_forward=False,
                            fp8_meta_index=META_DQKV,
                            fp8_dtype=fp8_dtype_backward,
                            dtype=d_out_f8tensor.dtype,
6622
                        )
6623
                    else:
6624
                        qkv_group = len(ctx.qkv_layout.split("_"))
6625
                        if qkv_group == 1:
6626
6627
6628
6629
6630
6631
6632
6633
6634
6635
6636
6637
6638
                            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])
6639
6640
6641
6642
                            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]),
6643
6644
6645
6646
6647
6648
6649
6650
6651
6652
6653
6654
6655
6656
6657
6658
6659
6660
                                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])
6661
6662
6663
6664
                            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]),
6665
6666
6667
6668
6669
                                ctx.fp8_meta["scaling_bwd"],
                                META_DQKV,
                                fp8_dtype_backward,
                                ctx.qkv_dtype,
                            ).view(dq_fp8.shape)
6670
6671
                            dk = cast_from_fp8(
                                dk_fp8.view(-1, dk_fp8.shape[-2] * dk_fp8.shape[-1]),
6672
6673
6674
6675
6676
                                ctx.fp8_meta["scaling_bwd"],
                                META_DQKV,
                                fp8_dtype_backward,
                                ctx.qkv_dtype,
                            ).view(dk_fp8.shape)
6677
6678
                            dv = cast_from_fp8(
                                dv_fp8.view(-1, dv_fp8.shape[-2] * dv_fp8.shape[-1]),
6679
6680
6681
6682
6683
                                ctx.fp8_meta["scaling_bwd"],
                                META_DQKV,
                                fp8_dtype_backward,
                                ctx.qkv_dtype,
                            ).view(dv_fp8.shape)
6684
6685
6686
6687
                else:
                    if d_out.dtype == torch.uint8:
                        d_out = d_out_f8tensor.from_float8(q.dtype)
                    dq, dk, dv, *rest = fused_attn_bwd(
6688
6689
6690
6691
6692
6693
6694
6695
6696
6697
6698
6699
                        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,
6700
                        ctx.fused_attention_backend,
6701
6702
                        cu_seqlens_q_padded,
                        cu_seqlens_kv_padded,
6703
6704
6705
6706
6707
6708
6709
6710
6711
6712
6713
6714
6715
6716
6717
6718
                        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,
6719
6720
                        ctx.window_size,
                        ctx.deterministic,
6721
                    )
6722

6723
6724
        # if no_bias or alibi, return dqkv
        if ctx.attn_bias_type in ["no_bias", "alibi"]:
6725
6726
6727
6728
6729
6730
6731
6732
6733
6734
6735
6736
6737
6738
6739
6740
6741
6742
6743
6744
6745
6746
6747
6748
6749
6750
            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,
6751
6752
                None,
                None,
6753
            )
6754
        # else, return (dqkv, dbias)
6755
6756
6757
6758
6759
6760
6761
6762
6763
6764
6765
6766
6767
6768
6769
6770
6771
6772
6773
6774
6775
6776
6777
6778
6779
6780
        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,
6781
6782
            None,
            None,
6783
        )
6784

6785

6786
class FusedAttention(torch.nn.Module):
6787
6788
6789
6790
6791
6792
6793
6794
6795
    """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:

6796
6797
6798
6799
    | backend       | 1                       | 2                              |
    | flash based   | no                      | yes                            |
    | cuDNN based   | yes                     | yes                            |
    | qkv dtype     | fp16/bf16               | fp16/bf16                      |
6800
    | attn_type     | self/cross              | self/cross                     |
6801
    | qkv_layout    |                         |                                |
6802
    |  - (q,k,v)    | sb3hd, bs3hd            | sb3hd, bs3hd, sbh3d, bsh3d     |
6803
    |               | sbhd_sb2hd, bshd_bs2hd  | sbhd_sb2hd, bshd_bs2hd         |
6804
6805
    |               | bshd_bshd_bshd          | sbhd_sbh2d, bshd_bsh2d         |
    |               |                         | sbhd_sbhd_sbhd, bshd_bshd_bshd |
6806
6807
    | mask_type     | causal/padding/no_mask  | causal/padding/no_mask         |
    | bias_type     | post_scale_bias/no_bias | post_scale_bias/alibi/no_bias  |
6808
    | dropout       | yes                     | yes                            |
6809
6810
    | max_seqlen    | <=512, multiple of 64   | any, multiple of 64            |
    | head_dim      | 64                      | <=128, multiple of 8           |
6811
    | output dtype  | fp16/bf16               | fp16/bf16                      |
6812
6813
6814
6815
    """

    def __init__(
        self,
6816
        softmax_scale: float,
6817
6818
6819
        attention_dropout: float = 0.0,
        attention_dropout_ctx: Optional[Callable] = nullcontext,
        attention_type: str = "self",
6820
6821
        layer_number: Optional[int] = None,
        deterministic: bool = False,
6822
6823
6824
    ) -> None:
        super().__init__()

6825
        self.softmax_scale = softmax_scale
6826
6827
6828
        self.attention_dropout = attention_dropout
        self.attention_dropout_ctx = attention_dropout_ctx
        self.attention_type = attention_type
6829
6830
6831
        self.use_FAv2_bwd = os.getenv(
            "NVTE_FUSED_ATTN_USE_FAv2_BWD", "0"
        ) == "1" and get_device_compute_capability() == (9, 0)
6832
        self.layer_number = 1 if layer_number is None else layer_number
6833
        self.deterministic = deterministic
6834

6835
        def remove_extra_states_check(self, incompatible_keys):  # pylint: disable=unused-argument
6836
6837
            """
            Temporarily remove fused_attention._extra_state as a missing key
6838
            or an unexpected key when loading Transformer Engine checkpoints.
6839
6840
            Please store FP8 metadata as DotProductAttention's _extra_state,
            rather than FusedAttention's _extra_state. This hook will be
6841
            phased out in Transformer Engine 2.0.
6842
6843
            """
            for key in incompatible_keys.missing_keys:
6844
                if "fused_attention._extra_state" in key:
6845
                    incompatible_keys.missing_keys.remove(key)
6846
6847
6848
6849
6850
6851
6852
            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."
                    )
6853

6854
6855
        self.register_load_state_dict_post_hook(remove_extra_states_check)

6856
    @no_torch_dynamo()
6857
6858
6859
6860
6861
    def forward(
        self,
        query_layer: torch.Tensor,
        key_layer: torch.Tensor,
        value_layer: torch.Tensor,
6862
6863
6864
        qkv_layout: str = "sbh3d",
        cu_seqlens_q: Optional[torch.Tensor] = None,
        cu_seqlens_kv: Optional[torch.Tensor] = None,
6865
6866
        cu_seqlens_q_padded: Optional[torch.Tensor] = None,
        cu_seqlens_kv_padded: Optional[torch.Tensor] = None,
6867
6868
        max_seqlen_q: Optional[int] = None,
        max_seqlen_kv: Optional[int] = None,
6869
        attn_mask_type: str = "causal",
6870
        attention_mask: Optional[Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]] = None,
6871
        window_size: Optional[Tuple[int, int]] = None,
6872
        fused_attention_backend: tex.NVTE_Fused_Attn_Backend = tex.NVTE_Fused_Attn_Backend.NVTE_No_Backend,
6873
6874
6875
        core_attention_bias_type: str = "no_bias",
        core_attention_bias: Optional[torch.Tensor] = None,
        fast_zero_fill: bool = True,
6876
        cp_group: Optional[Union[dist_group_type, List[dist_group_type]]] = None,
6877
6878
        cp_global_ranks: List[int] = None,
        cp_stream: torch.cuda.Stream = None,
6879
        cp_comm_type: str = "p2p",
6880
6881
        fp8: bool = False,
        fp8_meta: Optional[Dict[str, Any]] = None,
6882
6883
    ) -> torch.Tensor:
        """fused attention fprop"""
6884
6885
6886
        assert (
            fused_attention_backend != tex.NVTE_Fused_Attn_Backend.NVTE_No_Backend
        ), "No fused attention backend supports this input combination!"
6887
6888
6889
6890
        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."
6891
6892
        assert (
            query_layer.is_cuda and key_layer.is_cuda and value_layer.is_cuda
6893
        ), "FusedAttention only supports CUDA tensors."
6894
6895
        assert (
            qkv_layout in QKVLayouts
6896
        ), f"FusedAttention does not support qkv_layout = {qkv_layout}!"
6897

6898
6899
6900
6901
6902
6903
        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)
6904
        context_parallel = cp_size > 1
6905

6906
        qkv_format = "".join([i for i in qkv_layout.split("_")[0] if i.isalpha()])
6907

6908
6909
        if qkv_format in ["sbhd", "bshd"]:
            if qkv_format == "sbhd":
6910
                batch_size, max_seqlen_q, max_seqlen_kv = (
6911
6912
6913
6914
6915
                    query_layer.shape[1],
                    query_layer.shape[0],
                    key_layer.shape[0],
                )
            if qkv_format == "bshd":
6916
                batch_size, max_seqlen_q, max_seqlen_kv = (
6917
6918
6919
6920
                    query_layer.shape[0],
                    query_layer.shape[1],
                    key_layer.shape[1],
                )
6921
6922
            max_seqlen_q *= cp_size
            max_seqlen_kv *= cp_size
6923
            if "padding" in attn_mask_type:
6924
6925
                assert not context_parallel, "Padding mask not supported with context parallelism!"

6926
6927
6928
6929
6930
                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!"
                        )
6931
                    if self.attention_type == "self":
6932
6933
                        cu_seqlens_q = get_cu_seqlens(attention_mask)
                        cu_seqlens_kv = cu_seqlens_q
6934
                    else:
6935
6936
                        cu_seqlens_q = get_cu_seqlens(attention_mask[0])
                        cu_seqlens_kv = get_cu_seqlens(attention_mask[1])
6937
            else:
6938
6939
6940
6941
6942
6943
6944
6945
6946
6947
6948
6949
                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,
                    )
6950
6951
6952
        if qkv_format == "thd":
            assert (
                max_seqlen_q is not None
6953
6954
6955
                and max_seqlen_kv is not None
                and cu_seqlens_q is not None
                and cu_seqlens_kv is not None
6956
            ), "max_seqlen_q/kv and cu_seqlens_q/kv can not be None when qkv_format is thd!"
6957
6958
6959
6960

        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
6961
6962
6963

        qkv_dtype = TE_DType[query_layer.dtype]

6964
6965
6966
6967
6968
        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)
        )
6969

6970
6971
6972
6973
6974
6975
6976
6977
6978
6979
6980
        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!"
            )

6981
        if context_parallel:
6982
            assert (
6983
6984
                fp8
                or fused_attention_backend == tex.NVTE_Fused_Attn_Backend.NVTE_F16_arbitrary_seqlen
6985
6986
6987
6988
6989
6990
6991
            ), 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)
            ]
6992
6993
6994
            with self.attention_dropout_ctx():
                output = attn_forward_func_with_cp(
                    self.training,
6995
6996
6997
6998
6999
7000
7001
                    query_layer,
                    key_layer,
                    value_layer,
                    cu_seqlens_q,
                    cu_seqlens_kv,
                    max_seqlen_q,
                    max_seqlen_kv,
7002
7003
                    cu_seqlens_q_padded,
                    cu_seqlens_kv_padded,
7004
                    self.attention_dropout if self.training else 0.0,
7005
7006
7007
                    cp_group,
                    cp_global_ranks,
                    cp_stream,
7008
                    cp_comm_type,
7009
                    softmax_scale=self.softmax_scale,
7010
                    qkv_format=qkv_format,
7011
                    attn_mask_type=attn_mask_type,
7012
7013
                    attn_bias_type=core_attention_bias_type,
                    attn_bias=core_attention_bias,
7014
                    deterministic=self.deterministic,
7015
                    use_fused_attention=True,
7016
                    window_size=window_size,
7017
7018
                    fp8=fp8,
                    fp8_meta=fp8_meta,
7019
7020
                )
        else:
7021
7022
7023
7024
7025
7026
7027
            with self.attention_dropout_ctx():
                output = FusedAttnFunc.apply(
                    self.training,
                    max_seqlen_q,
                    max_seqlen_kv,
                    cu_seqlens_q,
                    cu_seqlens_kv,
7028
7029
                    cu_seqlens_q_padded,
                    cu_seqlens_kv_padded,
7030
7031
7032
7033
7034
7035
7036
7037
7038
7039
7040
                    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,
7041
                    window_size,
7042
7043
7044
7045
7046
                    None,  # rng_gen
                    fused_attention_backend,
                    use_FAv2_bwd,
                    fp8,
                    fp8_meta,
7047
                    self.deterministic,
7048
                )
7049

7050
7051
        # ...hd -> ...(hd)
        return output.view(*output.shape[:-2], -1)
7052
7053


7054
class DotProductAttention(TransformerEngineBaseModule):
7055
7056
7057
7058
7059
7060
    """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::

7061
        Argument :attr:`attention_mask` in the `forward` call is only used when
7062
        :attr:`attn_mask_type` includes '"padding"' or `"arbitrary"`.
7063
7064
7065

    .. warning::

7066
        FlashAttention uses a non-deterministic algorithm for optimal performance. To observe
7067
        deterministic behavior at the cost of performance, use FlashAttention version >= `2.4.1`
7068
7069
        and set the environment variable :attr:`NVTE_ALLOW_NONDETERMINISTIC_ALGO=0`. In order
        to disable`flash-attn` entirely, set :attr:`NVTE_FLASH_ATTN=0`.
7070

7071
7072
7073
7074
7075
7076
7077
    .. 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>`_).


7078
7079
7080
7081
    Parameters
    ----------
    num_attention_heads : int
                         number of attention heads in the transformer layer.
7082
7083
7084
    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.
7085
7086
7087
7088
7089
7090
7091
7092
    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`.
7093
7094
    attention_dropout: float, default = 0.0
                      dropout probability for the dropout op during multi-head attention.
7095
    attn_mask_type: str, default = `causal`
7096
                   type of attention mask passed into softmax operation, options are "`no_mask`",
7097
7098
7099
7100
7101
7102
7103
7104
7105
                   "`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
7106
                   "`padding_causal`" and "`padding_causal_bottom_right`", Transformer Engine
7107
7108
7109
7110
7111
7112
7113
7114
7115
7116
7117
7118
7119
7120
                   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].
7121
7122
7123
7124
    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
7125
7126
7127
                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
7128
                be overridden by :attr:`window_size` in `forward` as well.
7129
7130
    attention_type: str, default = `self`
                   type of attention, either "`self`" and "`cross`".
7131
7132
7133
    layer_number: int, default = `None`
                 layer number of the current `DotProductAttention` when multiple such modules
                 are concatenated, for instance in consecutive transformer blocks.
7134
7135
7136
    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,
7137
               `h` the number of heads, `d` head size, and `t` the total number of tokens
7138
7139
7140
7141
7142
               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.
7143
               For that, please use `get_qkv_layout` to gain the layout information.
7144
7145
    softmax_scale: Optional[float], default = `None`
                softmax scale for the attention scores. If `None`, defaults to
7146
                `1.0/math.sqrt(kv_channels if isinstance(kv_channels, int) else kv_channels[0])`.
7147
7148
7149
7150
7151
7152
7153
7154
7155

    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.
7156
    cp_group : Union[ProcessGroup, List[ProcessGroup]], default = `None`
7157
              context parallel process group.
7158
7159
7160
              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.
7161
7162
7163
7164
7165
7166
7167
    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.
7168
    cp_comm_type : str, default = `p2p`
7169
                  inter-gpu communication type for context parallelism.
7170
                  Can be "p2p" or "all_gather" or "a2a" or "a2a+p2p".
7171
7172
7173
7174
7175
7176
                  "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.
7177
7178
7179
                  "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).
7180
7181
7182
7183
7184
    """

    def __init__(
        self,
        num_attention_heads: int,
7185
        kv_channels: Union[int, Tuple[int, int]],
7186
        num_gqa_groups: Optional[int] = None,
7187
        attention_dropout: float = 0.0,
7188
        qkv_format: str = "sbhd",
7189
        attn_mask_type: str = "causal",
7190
        window_size: Optional[Tuple[int, int]] = None,
7191
7192
7193
7194
7195
        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,
7196
        attention_type: str = "self",
7197
        cp_group: Optional[Union[dist_group_type, List[dist_group_type]]] = None,
7198
        cp_global_ranks: List[int] = None,
7199
        cp_stream: torch.cuda.Stream = None,
7200
        cp_comm_type: str = "p2p",
7201
        softmax_scale: Optional[float] = None,
7202
7203
7204
    ) -> None:
        super().__init__()

7205
        self.logger = logging.getLogger("DotProductAttention")
7206
7207
7208
        self.logger.setLevel(_log_level)
        if not self.logger.hasHandlers():
            self.logger.addHandler(_stream_handler)
7209
        self.qkv_format = qkv_format
7210
        attn_mask_type = attn_mask_type.replace(",", "_")
7211
7212
        if attn_mask_type == "causal_padding":
            attn_mask_type = "padding_causal"
7213
        self.attn_mask_type = attn_mask_type
7214
        self.window_size = check_set_window_size(attn_mask_type, window_size)
7215
7216
7217
7218
7219
7220
7221
        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)
7222
        self.get_rng_state_tracker = get_rng_state_tracker
7223
        self.num_attention_heads = num_attention_heads
7224
        self.layer_number = 1 if layer_number is None else layer_number
7225
7226
7227
        self.cp_group = cp_group
        self.cp_global_ranks = cp_global_ranks
        self.cp_stream = cp_stream
7228
        self.cp_comm_type = cp_comm_type
7229

7230
7231
7232
7233
7234
7235
        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]
        )
7236

7237
        self.num_gqa_groups = num_attention_heads if num_gqa_groups is None else num_gqa_groups
7238
        self.num_gqa_groups_per_partition = int(self.num_gqa_groups // self.tp_size)
7239

7240
7241
7242
        assert (
            num_attention_heads % self.num_gqa_groups == 0
        ), "The number of attention heads must be divisible by the number of GQA groups!"
7243

7244
        self.rng_states_tracker = None
7245
7246
7247
        if sequence_parallel or get_rng_state_tracker is None:
            attention_dropout_ctx = nullcontext
        else:
7248
7249
7250
            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
7251

7252
        if softmax_scale is None:
7253
7254
7255
            softmax_scale = 1.0 / math.sqrt(
                kv_channels if isinstance(kv_channels, int) else kv_channels[0]
            )
7256

7257
7258
7259
        self.deterministic = (
            not bool(int(os.getenv("NVTE_ALLOW_NONDETERMINISTIC_ALGO", "1")))
            or torch.are_deterministic_algorithms_enabled()
7260
        )
7261
7262
7263
7264
7265
7266
7267
7268
7269
7270
7271
7272
7273
7274
7275
7276
7277
7278
7279
        # 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"
7280

7281
        assert attention_type in AttnTypes, f"attention_type {attention_type} not supported"
7282
7283
7284
7285

        self.attention_type = attention_type
        self.attention_dropout = attention_dropout

7286
7287
7288
7289
7290
        attn_kwargs = {
            "attention_dropout": attention_dropout,
            "attention_dropout_ctx": attention_dropout_ctx,
        }

7291
7292
7293
7294
7295
7296
7297
        self.flash_attention = FlashAttention(
            softmax_scale,
            attention_type=attention_type,
            layer_number=layer_number,
            deterministic=self.deterministic,
            **attn_kwargs,
        )
7298

7299
        # Instantiating three types since use of flash-attn and FusedAttention
7300
        # might be ruled out due to forward inputs.
7301
7302
7303
7304
7305
7306
7307
        self.fused_attention = FusedAttention(
            softmax_scale,
            attention_type=attention_type,
            layer_number=layer_number,
            deterministic=self.deterministic,
            **attn_kwargs,
        )
7308

7309
        self.unfused_attention = UnfusedDotProductAttention(
7310
7311
7312
7313
            softmax_scale,
            attention_type=attention_type,
            **attn_kwargs,
            layer_number=layer_number,
7314
        )
7315

7316
7317
7318
        def remove_extra_states_check(self, incompatible_keys):  # pylint: disable=unused-argument
            """
            Temporarily remove core_attention._extra_state as a missing key
7319
7320
            when loading older Transformer Engine checkpoints. Will phase out
            this hook in Transformer Engine 2.0.
7321
7322
7323
7324
7325
7326
7327
            """
            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)

7328
7329
7330
7331
7332
7333
7334
7335
7336
7337
7338
7339
7340
7341
7342
7343
7344
7345
7346
7347
7348
7349
    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
        )

7350
7351
7352
7353
    def _checkpointed_attention_forward(
        self,
        attention_func: Callable,
        *forward_args: Tuple[torch.Tensor, ...],
7354
        **forward_kwargs: Dict[str, Any],
7355
7356
7357
    ) -> torch.Tensor:
        """Forward method with activation checkpointing."""

7358
7359
        def custom_forward(*input_args, **input_kwargs):
            return attention_func(*input_args, **input_kwargs)
7360
7361
7362

        hidden_states = checkpoint(
            custom_forward,
7363
7364
7365
            distribute_saved_activations=False,
            get_rng_state_tracker=self.get_rng_state_tracker,
            tp_group=self.tp_group,
7366
            *forward_args,
7367
            **forward_kwargs,
7368
7369
7370
7371
        )

        return hidden_states

7372
7373
    def set_context_parallel_group(
        self,
7374
        cp_group: Union[dist_group_type, List[dist_group_type], None],
7375
7376
        cp_global_ranks: List[int],
        cp_stream: torch.cuda.Stream,
7377
        cp_comm_type: str = "p2p",
7378
    ) -> None:
7379
7380
7381
7382
7383
7384
        """
        Set the context parallel attributes for the given
        module before executing the forward pass.

        Parameters
        ----------
7385
        cp_group : Union[ProcessGroup, List[ProcessGroup]]
7386
                  context parallel process group.
7387
7388
7389
                  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.
7390
7391
7392
7393
        cp_global_ranks : List[int]
                         list of global ranks in the context group.
        cp_stream : torch.cuda.Stream
                   cuda stream for context parallel execution.
7394
        cp_comm_type : str, default = `p2p`
7395
                      inter-gpu communication type for context parallelism.
7396
                      Can be "p2p" or "all_gather" or "a2a" or "a2a+p2p".
7397
7398
7399
7400
7401
7402
                      "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.
7403
7404
7405
                      "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).
7406
        """
7407
7408
7409
        self.cp_group = cp_group
        self.cp_global_ranks = cp_global_ranks
        self.cp_stream = cp_stream
7410
        self.cp_comm_type = cp_comm_type
7411

7412
    @no_torch_dynamo(recursive=False)
7413
7414
7415
7416
7417
    def forward(
        self,
        query_layer: torch.Tensor,
        key_layer: torch.Tensor,
        value_layer: torch.Tensor,
7418
        attention_mask: Optional[Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]] = None,
7419
7420
7421
        qkv_format: Optional[str] = None,
        cu_seqlens_q: Optional[torch.Tensor] = None,
        cu_seqlens_kv: Optional[torch.Tensor] = None,
7422
7423
        cu_seqlens_q_padded: Optional[torch.Tensor] = None,
        cu_seqlens_kv_padded: Optional[torch.Tensor] = None,
7424
7425
        max_seqlen_q: Optional[int] = None,
        max_seqlen_kv: Optional[int] = None,
7426
        attn_mask_type: Optional[str] = None,
7427
        window_size: Optional[Tuple[int, int]] = None,
7428
        checkpoint_core_attention: bool = False,
7429
7430
        core_attention_bias_type: str = "no_bias",
        core_attention_bias: Optional[torch.Tensor] = None,
7431
        alibi_slopes: Optional[torch.Tensor] = None,
7432
        fast_zero_fill: bool = True,
7433
        inference_params: Optional[InferenceParams] = None,
7434
        is_first_microbatch: Optional[bool] = None,
7435
7436
7437
7438
7439
7440
    ) -> torch.Tensor:
        """
        Dot Product Attention Layer.

        .. note::

7441
7442
            Argument :attr:`attention_mask` is only used when :attr:`attn_mask_type`
            includes '"padding"' or `"arbitrary"`.
7443

7444
7445
        .. note::

7446
7447
7448
7449
7450
7451
7452
7453
7454
7455
7456
7457
7458
            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,
7459
            and FusedAttention backend if applicable, to use. Transformer Engine prioritizes
7460
7461
7462
7463
            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
7464
7465
            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
7466
            optimizations in FusedAttention. When unset, Transformer Engine determines the code path
7467
7468
            based on its internal logic. These optimizations trade memory for performance
            and should be used with care.
7469

7470
7471
7472
7473
7474
7475
7476
7477
        Parameters
        ----------
        query_layer : torch.Tensor
                     Query tensor.
        key_layer : torch.Tensor
                   Key tensor.
        value_layer : torch.Tensor
                     Value tensor.
7478
7479
        attention_mask: Optional[Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]],
             default = `None`. Boolean tensor(s) used to mask out attention softmax input.
7480
             It should be `None` for causal masks and "`no_mask`". For padding masks, it should be
7481
7482
             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]
7483
7484
7485
7486
             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.
7487
7488
7489
        qkv_format: str, default = `None`
                   If provided, overrides :attr:`qkv_format` from initialization.
        cu_seqlens_q: Optional[torch.Tensor], default = `None`
7490
                   Cumulative sum of sequence lengths (without offset) in a batch for `query_layer`,
7491
7492
                   with shape [batch_size + 1] and dtype torch.int32.
        cu_seqlens_kv: Optional[torch.Tensor], default = `None`
7493
7494
7495
7496
7497
7498
7499
7500
7501
7502
7503
7504
                   Cumulative sum of sequence lengths (without offset) in a batch for `key_layer`
                   and `value_layer`, with shape [batch_size + 1] and dtype torch.int32.
        cu_seqlens_q_padded: Optional[torch.Tensor], default = `None`
                   Cumulative sum of sequence lengths (with offset) in a batch for
                   `query_layer`, with shape [batch_size + 1] and dtype torch.int32.
                   When there is no padding between sequences in a batch,
                   `cu_seqlens_q_padded = cu_seqlens_q`.
        cu_seqlens_kv_padded: Optional[torch.Tensor], default = `None`
                   Cumulative sum of sequence lengths (with offset) in a batch for `key_layer`
                   and `value_layer`, with shape [batch_size + 1] and dtype torch.int32.
                   When there is no padding between sequences in a batch,
                   `cu_seqlens_kv_padded = cu_seqlens_kv`.
7505
7506
7507
7508
7509
7510
        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.
7511
7512
7513
7514
7515
7516
7517
        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.
7518
        window_size: Optional[Tuple[int, int]], default = `None`
7519
                    Sliding window size for local attention.
7520
7521
7522
7523
7524
        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.
7525
        core_attention_bias_type: str, default = `no_bias`
7526
                    Bias type, {`no_bias`, `pre_scale_bias`, `post_scale_bias`, `alibi`}
7527
        core_attention_bias: Optional[torch.Tensor], default = `None`
7528
7529
                    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.
7530
7531
7532
7533
        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.
7534
        fast_zero_fill: bool, default = `True`
7535
                    Whether to use the fast path to set output tensors to 0 or not.
7536
7537
7538
7539
7540
7541
7542
7543
7544
7545
        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.
7546
7547
7548
7549
7550
7551
7552
7553
7554
7555
7556
7557
7558
        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)
7559
        """
7560
7561
7562
7563
7564
7565
7566
7567
7568
7569
7570
        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
7571
                        self.logger.warning(
7572
7573
7574
                            """Forcing fp8_meta["recipe"].fp8_dpa=True due to """
                            """fp8_meta["recipe"].fp8_mha=True"""
                        )
7575
7576
7577
7578
7579
7580
7581
7582
7583
7584
7585

            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."""
7586

7587
7588
7589
            assert (
                query_layer.is_cuda and key_layer.is_cuda and value_layer.is_cuda
            ), "DotProductAttention only supports CUDA tensors."
7590
7591
7592
            assert (
                query_layer.dtype == key_layer.dtype and query_layer.dtype == value_layer.dtype
            ), "Queries, keys and values must have the same data type!"
7593
7594
7595
            assert (
                key_layer.shape[:-1] == value_layer.shape[:-1]
            ), "Keys and values must have the same batch size, sequence length and number of heads!"
7596
7597
7598
7599
7600
7601
7602
7603
            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}!"
7604

7605
7606
7607
7608
7609
7610
            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"
7611
            assert (
7612
7613
7614
7615
7616
7617
                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!"
7618

7619
7620
7621
7622
            if window_size is None:
                window_size = self.window_size
            window_size = check_set_window_size(attn_mask_type, window_size)

7623
7624
7625
7626
7627
7628
7629
            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."
7630

7631
7632
            if qkv_format is None:
                qkv_format = self.qkv_format
7633

7634
7635
            if inference_params is not None:
                assert self.layer_number is not None, "Layer number must be set!"
7636

7637
7638
7639
7640
7641
                # 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"

7642
7643
7644
                if qkv_format == "bshd":
                    key_layer = key_layer.transpose(0, 1)
                    value_layer = value_layer.transpose(0, 1)
7645

7646
7647
7648
7649
                (
                    inference_key_memory,
                    inference_value_memory,
                ) = inference_params.key_value_memory_dict[self.layer_number]
7650

7651
7652
7653
                batch_start = inference_params.batch_size_offset
                batch_end = batch_start + key_layer.size(1)
                assert batch_end <= inference_key_memory.size(1)
7654

7655
7656
7657
                sequence_start = inference_params.sequence_len_offset
                sequence_end = sequence_start + key_layer.size(0)
                assert sequence_end <= inference_key_memory.size(0)
7658

7659
7660
7661
7662
7663
7664
7665
7666
7667
                # 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, ...]
7668

7669
7670
7671
                if qkv_format == "bshd":
                    key_layer = key_layer.transpose(0, 1)
                    value_layer = value_layer.transpose(0, 1)
7672

7673
7674
                key_layer = key_layer.contiguous()
                value_layer = value_layer.contiguous()
7675
7676

            assert (
7677
7678
7679
7680
7681
7682
7683
7684
7685
7686
                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":
7687
                assert all(
7688
7689
7690
7691
7692
7693
7694
7695
7696
7697
7698
7699
7700
7701
                    len(x.shape) == 3 for x in (query_layer, key_layer, value_layer)
                ), "Queries, keys and values must be 3D tensors when qkv_format = thd!"
                assert (
                    cu_seqlens_q is not None and cu_seqlens_kv is not None
                ), "cu_seqlens_q and cu_seqlens_kv can not be None when qkv_format = thd!"
                assert (
                    cu_seqlens_q.shape == cu_seqlens_kv.shape
                    and len(cu_seqlens_q.shape) == 1
                    and len(cu_seqlens_kv.shape) == 1
                ), "cu_seqlens_q and cu_seqlens_q must both have shape [batch_size + 1]!"
                assert (
                    cu_seqlens_q.dtype == torch.int32 and cu_seqlens_kv.dtype == torch.int32
                ), "cu_seqlens_q and cu_seqlens_q must both be in dtype torch.int32!"
                if max_seqlen_q is None:
7702
7703
7704
7705
                    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]
7706
                    max_seqlen_q = int((seqlens_q.max().item() + 63) // 64 * 64)
7707
                if max_seqlen_kv is None:
7708
7709
7710
7711
                    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]
7712
                    max_seqlen_kv = int((seqlens_kv.max().item() + 63) // 64 * 64)
7713
                batch_size = len(cu_seqlens_q) - 1
7714

7715
7716
7717
7718
7719
7720
            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)
7721
7722
            context_parallel = cp_size > 1

7723
            if qkv_format in ["sbhd", "bshd"]:
7724
                assert all(
7725
7726
7727
7728
                    len(x.shape) == 4 for x in (query_layer, key_layer, value_layer)
                ), f"Queries, keys and values must be 4D tensors when qkv_format = {qkv_format}!"
                if qkv_format == "sbhd":
                    max_seqlen_q, max_seqlen_kv = (query_layer.shape[0], key_layer.shape[0])
7729
                    batch_size = query_layer.shape[1]
7730
7731
                if qkv_format == "bshd":
                    max_seqlen_q, max_seqlen_kv = (query_layer.shape[1], key_layer.shape[1])
7732
                    batch_size = query_layer.shape[0]
7733
7734
                max_seqlen_q *= cp_size
                max_seqlen_kv *= cp_size
7735
7736
7737
7738
7739
7740
7741
7742
7743
7744
7745
7746
                if cu_seqlens_q is not None:
                    seqlens_q = cu_seqlens_q[1:] - cu_seqlens_q[:-1]
                    assert all(
                        seqlens_q <= max_seqlen_q
                    ), """Sequence lengths indicated by cu_seqlens_q must be no greater than
                        the sequence dimention in 'query_layer'!"""
                if cu_seqlens_kv is not None:
                    seqlens_kv = cu_seqlens_kv[1:] - cu_seqlens_kv[:-1]
                    assert all(
                        seqlens_kv <= max_seqlen_kv
                    ), """Sequence lengths indicated by cu_seqlens_kv must be no greater than
                        the sequence dimention in 'key_layer' and 'value_layer'!"""
7747
7748
7749
7750
7751
                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!"
7752
                        if self.attention_type == "self":
7753
7754
7755
7756
7757
7758
7759
7760
7761
7762
7763
7764
7765
7766
7767
7768
                            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,
                        )
7769

7770
7771
7772
7773
7774
            if (
                isinstance(query_layer, Float8Tensor)
                and isinstance(key_layer, Float8Tensor)
                and isinstance(value_layer, Float8Tensor)
            ):
7775
                qkv_layout, query_layer._data, key_layer._data, value_layer._data = get_qkv_layout(
7776
7777
7778
                    query_layer._data, key_layer._data, value_layer._data, qkv_format=qkv_format
                )
            else:
7779
                qkv_layout, query_layer, key_layer, value_layer = get_qkv_layout(
7780
7781
                    query_layer, key_layer, value_layer, qkv_format=qkv_format
                )
7782

7783
7784
7785
7786
7787
7788
7789
7790
            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
7791
            bottom_right_alignment = (attn_mask_type not in ["causal", "padding_causal"],)
7792
7793
7794
7795
7796
7797
7798
7799
            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
7800
                    or _alibi_cache["_bottom_right_alignment"] != bottom_right_alignment
7801
7802
7803
7804
7805
                    or _alibi_cache["_alibi_slopes"] is None
                ):
                    _alibi_cache["_alibi_slopes_require_update"] = True
                    _alibi_cache["_alibi_bias_require_update"] = True

7806
7807
            core_attention_bias_shape = None
            if core_attention_bias is not None:
7808
                if (
7809
7810
                    core_attention_bias.shape[0] == batch_size
                    and core_attention_bias.shape[1] == query_layer.shape[-2]
7811
                ):
7812
7813
7814
7815
7816
7817
7818
7819
7820
7821
7822
7823
7824
7825
7826
7827
7828
7829
7830
7831
7832
7833
7834
7835
                    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)
            )
7836

7837
            attention_params = AttentionParams(
7838
7839
7840
7841
7842
7843
7844
7845
                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,
7846
7847
                head_dim_qk=query_layer.shape[-1],
                head_dim_v=value_layer.shape[-1],
7848
7849
7850
7851
7852
7853
7854
7855
7856
7857
7858
                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,
7859
7860
                deterministic=self.deterministic,
                is_training=self.training,
7861
7862
7863
                fp8=self.fp8,
                fp8_meta=self.fp8_meta,
            )
7864
            global _attention_backends, _flash_attn_3_plus, _use_flash_attn_3
7865
7866
7867
7868
7869
7870
7871
            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"]:
7872
                _use_flash_attn_3 = _flash_attn_3_plus
7873
7874
7875
7876
7877
7878
7879
7880
                (
                    use_flash_attention,
                    use_fused_attention,
                    fused_attention_backend,
                    use_unfused_attention,
                    _,
                ) = get_attention_backend(attention_params)
                if use_flash_attention:
7881
7882
7883
7884
                    self.logger.info(
                        "Running with FlashAttention backend (version %s)",
                        _flash_attn_version if not _use_flash_attn_3 else _flash_attn_v3_version,
                    )
7885
7886
7887
7888
                elif use_fused_attention:
                    self.logger.info(
                        "Running with FusedAttention backend (sub-backend %s)",
                        int(fused_attention_backend),
7889
                    )
7890
7891
7892
7893
7894
7895
7896
                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"]
7897

7898
7899
7900
7901
7902
7903
7904
7905
7906
7907
7908
7909
7910
7911
7912
7913
7914
7915
7916
7917
7918
7919
            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,
7920
                    cp_comm_type=self.cp_comm_type,
7921
7922
                    max_seqlen_q=max_seqlen_q,
                    max_seqlen_kv=max_seqlen_kv,
7923
7924
                    fp8=self.fp8 and self.fp8_meta["recipe"].fp8_dpa,
                    fp8_meta=self.fp8_meta,
7925
                )
7926

7927
            if use_fused_attention:
7928
7929
                fu_core_attention_bias_type = core_attention_bias_type
                fu_core_attention_bias = core_attention_bias
7930
7931
7932
                if core_attention_bias_type == "alibi" and (
                    alibi_slopes is not None or max_seqlen_q != max_seqlen_kv
                ):
7933
7934
7935
7936
7937
7938
7939
                    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,
7940
                        bottom_right_alignment=attn_mask_type not in ["causal", "padding_causal"],
7941
                    )
7942
7943
7944
7945
7946
7947
7948
7949
7950
                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,
7951
7952
                        cu_seqlens_q_padded=cu_seqlens_q_padded,
                        cu_seqlens_kv_padded=cu_seqlens_kv_padded,
7953
7954
7955
7956
                        max_seqlen_q=max_seqlen_q,
                        max_seqlen_kv=max_seqlen_kv,
                        attn_mask_type=attn_mask_type,
                        attention_mask=attention_mask,
7957
                        window_size=window_size,
7958
7959
7960
7961
7962
7963
7964
                        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,
7965
                        cp_comm_type=self.cp_comm_type,
7966
7967
7968
7969
                        fp8=self.fp8 and self.fp8_meta["recipe"].fp8_dpa,
                        fp8_meta=self.fp8_meta,
                    )
                return self.fused_attention(
7970
7971
7972
7973
7974
7975
                    query_layer,
                    key_layer,
                    value_layer,
                    qkv_layout=qkv_layout,
                    cu_seqlens_q=cu_seqlens_q,
                    cu_seqlens_kv=cu_seqlens_kv,
7976
7977
                    cu_seqlens_q_padded=cu_seqlens_q_padded,
                    cu_seqlens_kv_padded=cu_seqlens_kv_padded,
7978
7979
                    max_seqlen_q=max_seqlen_q,
                    max_seqlen_kv=max_seqlen_kv,
7980
7981
                    attn_mask_type=attn_mask_type,
                    attention_mask=attention_mask,
7982
                    window_size=window_size,
7983
                    fused_attention_backend=fused_attention_backend,
7984
7985
                    core_attention_bias_type=fu_core_attention_bias_type,
                    core_attention_bias=fu_core_attention_bias,
7986
7987
7988
7989
                    fast_zero_fill=fast_zero_fill,
                    cp_group=self.cp_group,
                    cp_global_ranks=self.cp_global_ranks,
                    cp_stream=self.cp_stream,
7990
                    cp_comm_type=self.cp_comm_type,
7991
7992
                    fp8=self.fp8 and self.fp8_meta["recipe"].fp8_dpa,
                    fp8_meta=self.fp8_meta,
7993
                )
7994

7995
            from .cpu_offload import CPUOffloadEnabled
7996

7997
7998
7999
8000
8001
            if CPUOffloadEnabled:
                warnings.warn(
                    "Attention activation Offloading is only implemented"
                    "with Flash Attention and Fused Attention!"
                )
8002

8003
            if use_unfused_attention:
8004
8005
8006
8007
8008
8009
                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
                    )
8010
8011
8012
8013
8014
8015
8016
8017
8018
8019
8020
8021
8022
8023
8024
8025
                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(
8026
8027
8028
                    query_layer,
                    key_layer,
                    value_layer,
8029
8030
8031
8032
8033
8034
8035
8036
8037
                    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,
                )
8038

8039
            raise Exception("No dot product attention support for the provided inputs!")
8040
8041


8042
8043
8044
8045
8046
8047
8048
class MultiheadAttention(torch.nn.Module):
    r"""
    Multi-head Attention (MHA), including Query,
    Key, Value and Output projection.

    .. note::

8049
8050
        Argument :attr:`attention_mask` in the `forward` call is only used when
        :attr:`attn_mask_type` includes '"padding"' or `"arbitrary"`.
8051

8052
8053
8054
8055
8056
8057
8058
8059
8060
8061
8062
8063
8064
8065
8066
8067
8068
8069
8070
8071
8072
8073
8074
8075
8076
    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.
8077
8078
    attn_mask_type: {'no_mask', 'padding', 'causal', 'padding_causal', 'causal_bottom_right',
                   'padding_causal_bottom_right','arbitrary'},
8079
                   default = `causal`
8080
8081
8082
8083
8084
                   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.
8085
8086
8087
8088
    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
8089
8090
8091
                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
8092
                be overridden by :attr:`window_size` in `forward` as well.
8093
8094
8095
8096
8097
8098
8099
8100
8101
8102
8103
8104
8105
    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.
8106
8107
    input_layernorm: bool, default = `False`
                     if set to `True`, layer normalization to the input is applied.
8108
8109
8110
8111
8112
8113
8114
8115
8116
8117
8118
8119
8120
8121
8122
8123
8124
8125
8126
8127
    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"
8128
          The device on which the parameters of the model will be allocated. It is the user's
8129
8130
          responsibility to ensure all parameters are moved to the GPU before running the
          forward pass.
8131
8132
8133
8134
8135
8136
8137
    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.
8138
            For that, please use `get_qkv_layout` to gain the layout information.
8139
8140
8141
8142
8143
8144
8145
8146
8147
8148
8149
8150
8151
8152
8153
8154
8155
8156
8157
8158
8159
8160
8161
8162
8163
8164
8165
8166
8167
8168
8169
8170
8171
8172
8173
8174
8175
8176
8177
8178

    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`.
8179
8180
8181
8182
8183
8184
    """

    def __init__(
        self,
        hidden_size: int,
        num_attention_heads: int,
8185
8186
8187
8188
8189
        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,
8190
        layer_number: Optional[int] = None,
8191
        attn_mask_type: str = "causal",
8192
        window_size: Optional[Tuple[int, int]] = None,
8193
8194
        tp_group: Optional[dist_group_type] = None,
        tp_size: int = 1,
8195
        num_gqa_groups: Optional[int] = None,
8196
8197
8198
        fuse_wgrad_accumulation: bool = False,
        get_rng_state_tracker: Optional[Callable] = None,
        sequence_parallel: bool = False,
8199
        params_dtype: Optional[torch.dtype] = None,
8200
        return_bias: bool = False,
8201
8202
8203
8204
8205
8206
8207
8208
8209
        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
8210
        ub_overlap_rs_dgrad: bool = False,
8211
8212
        ub_overlap_rs: bool = False,
        ub_overlap_ag: bool = False,
8213
        bias: bool = True,
8214
        normalization: str = "LayerNorm",
8215
        device: Union[torch.device, str] = "cuda",
8216
        qkv_format: str = "sbhd",
8217
8218
    ) -> None:
        super().__init__()
8219

8220
        self.qkv_format = qkv_format
8221
        self.attn_mask_type = attn_mask_type
8222
        self.window_size = check_set_window_size(attn_mask_type, window_size)
8223
        self.layer_number = layer_number
8224
8225
8226
8227
8228
        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
8229
        self.params_dtype = torch.get_default_dtype() if params_dtype is None else params_dtype
8230
        self.num_attention_heads = num_attention_heads
8231
8232
8233
8234
8235
8236
8237
8238
        self.return_bias = return_bias

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

        if init_method is None:
            init_method = get_default_init_method()
        if output_layer_init_method is None:
            output_layer_init_method = get_default_init_method()
8239
8240
8241
8242
8243

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

8244
8245
8246
        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"
8247
8248
8249
8250
8251
8252

        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)
8253
8254
8255
8256
8257
8258
8259
        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!"
8260
        self.num_gqa_groups_per_partition = int(self.num_gqa_groups // tp_size)
8261
8262
8263
8264

        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
8265
8266
8267
8268
8269
8270
8271

        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,
8272
            "params_dtype": self.params_dtype,
8273
            "device": device,
8274
8275
8276
8277
        }

        qkv_parallel_mode = "column" if set_parallel_mode else None

cyanguwa's avatar
cyanguwa committed
8278
        if self.attention_type == "self":
8279
8280
            parameters_split = None
            if not fuse_qkv_params:
8281
8282
8283
8284
8285
8286
8287
                parameters_split = collections.OrderedDict(
                    [
                        ("query", self.hidden_size_q),
                        ("key", self.hidden_size_kv),
                        ("value", self.hidden_size_kv),
                    ]
                )
8288
8289
8290
            if self.input_layernorm:
                self.layernorm_qkv = LayerNormLinear(
                    hidden_size,
8291
                    self.hidden_size_q + 2 * self.hidden_size_kv,
8292
8293
8294
8295
8296
8297
                    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
8298
                    parameters_split=parameters_split,
8299
8300
8301
                    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
8302
                    ub_overlap_rs_dgrad=ub_overlap_rs_dgrad,
8303
                    ub_overlap_ag=ub_overlap_ag,
8304
                    normalization=normalization,
Przemyslaw Tredak's avatar
Przemyslaw Tredak committed
8305
                    ub_name="qkv",
8306
8307
8308
8309
8310
                    **common_gemm_kwargs,
                )
            else:
                self.qkv = Linear(
                    hidden_size,
8311
                    self.hidden_size_q + 2 * self.hidden_size_kv,
8312
8313
8314
8315
                    init_method=init_method,
                    bias=bias,
                    return_bias=False,
                    parallel_mode=qkv_parallel_mode,
cyanguwa's avatar
cyanguwa committed
8316
                    parameters_split=parameters_split,
8317
8318
                    **common_gemm_kwargs,
                )
cyanguwa's avatar
cyanguwa committed
8319
        elif self.attention_type == "cross":
8320
8321
8322
            if self.input_layernorm:
                self.layernorm_query = LayerNormLinear(
                    hidden_size,
8323
                    self.hidden_size_q,
8324
8325
8326
8327
8328
                    eps=layernorm_epsilon,
                    init_method=init_method,
                    bias=bias,
                    return_bias=False,
                    parallel_mode=qkv_parallel_mode,
8329
                    parameters_split=("query",) if not fuse_qkv_params else None,
8330
8331
8332
8333
                    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
8334
                    ub_overlap_rs_dgrad=ub_overlap_rs_dgrad,
8335
                    ub_overlap_ag=ub_overlap_ag,
8336
                    normalization=normalization,
Przemyslaw Tredak's avatar
Przemyslaw Tredak committed
8337
                    ub_name="qkv",
8338
8339
8340
8341
8342
                    **common_gemm_kwargs,
                )
            else:
                self.query_layer = Linear(
                    hidden_size,
8343
                    self.hidden_size_q,
8344
8345
8346
8347
8348
8349
8350
8351
                    init_method=init_method,
                    bias=bias,
                    return_bias=False,
                    parallel_mode=qkv_parallel_mode,
                    **common_gemm_kwargs,
                )
            self.key_value = Linear(
                hidden_size,
8352
                2 * self.hidden_size_kv,
8353
8354
8355
8356
                init_method=init_method,
                bias=bias,
                return_bias=False,
                parallel_mode=qkv_parallel_mode,
8357
                parameters_split=("key", "value") if not fuse_qkv_params else None,
8358
8359
8360
8361
8362
8363
                **common_gemm_kwargs,
            )

        # Attention.
        self.core_attention = DotProductAttention(
            num_attention_heads,
8364
            self.hidden_size_per_attention_head,
8365
8366
            num_gqa_groups=self.num_gqa_groups,
            attention_dropout=attention_dropout,
8367
            qkv_format=self.qkv_format,
8368
8369
8370
8371
            tp_size=tp_size,
            get_rng_state_tracker=get_rng_state_tracker,
            sequence_parallel=sequence_parallel,
            tp_group=tp_group,
8372
            layer_number=self.layer_number,
8373
            attention_type=self.attention_type,
8374
8375
8376
8377
        )

        # Linear
        self.proj = Linear(
8378
            self.hidden_size_q,
8379
8380
8381
            hidden_size,
            init_method=output_layer_init_method,
            bias=bias,
8382
            return_bias=return_bias,
8383
            parallel_mode="row" if set_parallel_mode else None,
8384
8385
            ub_overlap_rs=ub_overlap_rs,
            ub_overlap_ag=ub_overlap_ag,
Przemyslaw Tredak's avatar
Przemyslaw Tredak committed
8386
            ub_name="proj",
8387
8388
8389
8390
            **common_gemm_kwargs,
        )

    def _allocate_memory(
8391
        self, inference_max_sequence_len: int, batch_size: int, dtype: torch.dtype
8392
8393
8394
8395
    ) -> torch.Tensor:
        return torch.empty(
            inference_max_sequence_len,
            batch_size,
8396
            self.num_gqa_groups_per_partition,
8397
            self.hidden_size_per_attention_head,
8398
            dtype=dtype,
8399
8400
8401
8402
            device=torch.cuda.current_device(),
        )

    def set_tensor_parallel_group(self, tp_group: Union[dist_group_type, None]) -> None:
8403
8404
8405
8406
8407
8408
8409
8410
8411
        """
        Set the tensor parallel group for the given
        module before executing the forward pass.

        Parameters
        ----------
        tp_group : ProcessGroup, default = `None`
                  tensor parallel process group.
        """
8412
8413
        self.tp_group = tp_group

8414
    def set_context_parallel_group(
8415
        self,
8416
        cp_group: Union[dist_group_type, List[dist_group_type], None],
8417
        cp_global_ranks: List[int],
8418
        cp_stream: torch.cuda.Stream,
8419
        cp_comm_type: str = "p2p",
8420
    ) -> None:
8421
8422
8423
8424
8425
8426
        """
        Set the context parallel attributes for the given
        module before executing the forward pass.

        Parameters
        ----------
8427
        cp_group : Union[ProcessGroup, List[ProcessGroup]]
8428
                  context parallel process group.
8429
8430
8431
                  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.
8432
8433
8434
8435
        cp_global_ranks : List[int]
                         list of global ranks in the context group.
        cp_stream : torch.cuda.Stream
                   cuda stream for context parallel execution.
8436
        cp_comm_type : str, default = `p2p`
8437
                      inter-gpu communication type for context parallelism.
8438
                      Can be "p2p" or "all_gather" or "a2a", "a2a+p2p".
8439
8440
8441
8442
8443
8444
                      "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.
8445
8446
8447
                      "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).
8448
        """
8449
8450
8451
8452
8453
        # 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"):
8454
                child.set_context_parallel_group(cp_group, cp_global_ranks, cp_stream, cp_comm_type)
8455

8456
8457
8458
    def forward(
        self,
        hidden_states: torch.Tensor,
8459
        attention_mask: Optional[Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]] = None,
8460
        encoder_output: Optional[torch.Tensor] = None,
8461
        attn_mask_type: Optional[str] = None,
8462
        window_size: Optional[Tuple[int, int]] = None,
8463
8464
        is_first_microbatch: Optional[bool] = None,
        checkpoint_core_attention: bool = False,
8465
        inference_params: Optional[InferenceParams] = None,
8466
        rotary_pos_emb: Optional[Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]] = None,
8467
8468
        core_attention_bias_type: str = "no_bias",
        core_attention_bias: Optional[torch.Tensor] = None,
8469
        alibi_slopes: Optional[torch.Tensor] = None,
8470
8471
8472
8473
        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,
8474
        fast_zero_fill: bool = True,
8475
    ) -> Tuple[Union[torch.Tensor, None], ...]:
8476
8477
8478
8479
8480
        """
        Forward propagation for MultiheadAttention layer.

        .. note::

8481
8482
            Argument :attr:`attention_mask` is only used when :attr:`attn_mask_type`
            includes `"padding"` or `"arbitrary"`.
8483
8484
8485
8486
8487

        Parameters
        ----------
        hidden_states : torch.Tensor
             Input tensor.
8488
8489
        attention_mask: Optional[Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]],
             default = `None`. Boolean tensor(s) used to mask out attention softmax input.
8490
             It should be `None` for causal masks and "`no_mask`". For padding masks, it should be
8491
8492
             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]
8493
8494
8495
8496
8497
8498
             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'},
8499
                       default = `None`
8500
8501
8502
8503
                       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.
8504
8505
        window_size: Optional[Tuple[int, int]], default = `None`
                    sliding window size for local attention.
8506
8507
8508
8509
8510
8511
8512
8513
8514
8515
8516
8517
8518
8519
8520
8521
8522
8523
8524
8525
8526
8527
8528
8529
8530
        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`
8531
                    Bias type, {`no_bias`, `pre_scale_bias`, 'post_scale_bias`, `alibi`}
8532
        core_attention_bias: Optional[torch.Tensor], default = `None`
8533
8534
                    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.
8535
8536
8537
8538
        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.
8539
8540
8541
8542
8543
8544
8545
8546
8547
8548
8549
8550
        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.
8551
8552
8553
        fast_zero_fill: bool, default = `True`
                    Whether to set output tensors to 0 or not before use.
        """
8554
8555
        # hidden_states: [sq, b, h]

8556
        if attn_mask_type is None:
8557
            attn_mask_type = self.attn_mask_type
8558
8559
        if window_size is None:
            window_size = self.window_size
8560
        window_size = check_set_window_size(attn_mask_type, window_size)
8561

8562
        if "padding" in attn_mask_type and attention_mask is not None:
8563
            for i, _ in enumerate(attention_mask):
8564
8565
8566
                assert (
                    attention_mask[i].dtype == torch.bool
                ), "Attention mask must be in boolean type!"
8567

8568
8569
8570
        assert (
            core_attention_bias_type in AttnBiasTypes
        ), f"core_attention_bias_type {core_attention_bias_type} is not supported!"
8571

8572
        # =================================================
8573
        # Pre-allocate memory for key-values for inference
8574
8575
8576
        # =================================================

        if inference_params and self.layer_number is not None:
8577
8578
8579
            assert (
                self.qkv_format != "thd"
            ), "qkv_format == thd is not supported for an inference with KV-cache!"
8580
            if self.layer_number not in inference_params.key_value_memory_dict:
8581
                inf_max_seq_len = inference_params.max_sequence_length
8582
8583
                inf_max_batch_size = inference_params.max_batch_size
                inference_key_memory = self._allocate_memory(
8584
                    inf_max_seq_len, inf_max_batch_size, hidden_states.dtype
8585
8586
                )
                inference_value_memory = self._allocate_memory(
8587
                    inf_max_seq_len, inf_max_batch_size, hidden_states.dtype
8588
8589
8590
8591
8592
8593
8594
8595
8596
8597
8598
                )
                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]

8599
        # ======================
8600
        # Query, Key, and Value
8601
        # ======================
8602

8603
8604
8605
8606
8607
        fp8_mha = (
            FP8GlobalStateManager.is_fp8_enabled()
            and FP8GlobalStateManager.get_fp8_recipe().fp8_mha
        )

cyanguwa's avatar
cyanguwa committed
8608
8609
        if self.attention_type == "self":
            # Attention heads [sq, b, h] --> [sq, b, ng * (np/ng + 2) * hn]
8610
8611
8612
8613
            if self.input_layernorm:
                layernorm_qkv_outputs = self.layernorm_qkv(
                    hidden_states,
                    is_first_microbatch=is_first_microbatch,
8614
                    fp8_output=fp8_mha and rotary_pos_emb is None,
8615
8616
8617
8618
8619
8620
8621
8622
8623
                )
                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,
8624
                    fp8_output=fp8_mha and rotary_pos_emb is None,
8625
8626
                )

8627
8628
8629
            num_queries_per_key_value = (
                self.num_attention_heads_per_partition // self.num_gqa_groups_per_partition
            )
8630
            if self.qkv_weight_interleaved:
cyanguwa's avatar
cyanguwa committed
8631
                # [sq, b, ng * (np/ng + 2) * hn] --> [sq, b, ng, (np/ng + 2), hn]
8632
                new_tensor_shape = mixed_x_layer.size()[:-1] + (
cyanguwa's avatar
cyanguwa committed
8633
8634
                    self.num_gqa_groups_per_partition,
                    (num_queries_per_key_value + 2),
8635
8636
8637
8638
                    self.hidden_size_per_attention_head,
                )
                # split along second last dimension
                split_dim = -2
cyanguwa's avatar
cyanguwa committed
8639
8640
8641
8642
8643
            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,
8644
                    self.hidden_size_per_attention_head,
cyanguwa's avatar
cyanguwa committed
8645
8646
8647
                )
                # split along third last dimension
                split_dim = -3
8648
8649
8650

            mixed_x_layer = mixed_x_layer.view(*new_tensor_shape)

cyanguwa's avatar
cyanguwa committed
8651
8652
8653
8654
8655
8656
8657
8658
8659
            # 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)
8660
                )
8661
            else:
cyanguwa's avatar
cyanguwa committed
8662
                query_layer, key_layer, value_layer = torch.split(
8663
8664
8665
8666
                    mixed_x_layer,
                    (num_queries_per_key_value, 1, 1),
                    dim=split_dim,
                )
cyanguwa's avatar
cyanguwa committed
8667

8668
8669
8670
8671
8672
8673
8674
8675
8676
8677
8678
8679
            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
8680
8681
        elif self.attention_type == "cross":
            # Attention heads [sk, b, h] --> [sk, b, (ng * 2 * hn)]
8682
            mixed_kv_layer = self.key_value(
cyanguwa's avatar
cyanguwa committed
8683
                encoder_output,
8684
                is_first_microbatch=is_first_microbatch,
8685
                fp8_output=fp8_mha and rotary_pos_emb is None,
8686
8687
8688
            )

            if self.qkv_weight_interleaved:
cyanguwa's avatar
cyanguwa committed
8689
                # [sq, b, (ng * 2 * hn)] --> [sq, b, ng, 2 * hn]
8690
                new_tensor_shape = mixed_kv_layer.size()[:-1] + (
8691
                    self.num_gqa_groups_per_partition,
8692
8693
8694
8695
8696
                    2 * self.hidden_size_per_attention_head,
                )
                # split along last dimension
                split_dim = -1
            else:
cyanguwa's avatar
cyanguwa committed
8697
                # [sq, b, (ng * 2 * hn)] --> [sq, b, 2 * ng, hn]
8698
                new_tensor_shape = mixed_kv_layer.size()[:-1] + (
8699
                    2 * self.num_gqa_groups_per_partition,
8700
8701
8702
8703
8704
8705
8706
                    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
8707
8708
8709
            # mixed_kv_layer --> 2 [sk, b, ng, hn]
            if not is_in_onnx_export_mode():
                key_layer, value_layer = _SplitAlongDim.apply(
8710
8711
8712
                    mixed_kv_layer,
                    split_dim,
                    mixed_kv_layer.shape[split_dim] // 2,
cyanguwa's avatar
cyanguwa committed
8713
                )
8714
            else:
cyanguwa's avatar
cyanguwa committed
8715
                key_layer, value_layer = torch.split(
8716
8717
8718
                    mixed_kv_layer,
                    mixed_kv_layer.shape[split_dim] // 2,
                    dim=split_dim,
cyanguwa's avatar
cyanguwa committed
8719
                )
8720
8721
8722
8723
8724
8725
8726
8727
8728
            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)
            )
8729
8730
8731
8732
8733
8734

            # 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,
8735
                    fp8_output=fp8_mha and rotary_pos_emb is None,
8736
8737
8738
8739
8740
8741
8742
8743
8744
                )
                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,
8745
                    fp8_output=fp8_mha and rotary_pos_emb is None,
8746
8747
8748
8749
8750
8751
8752
8753
8754
                )

            # [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)

8755
8756
8757
        # ======================================================
        # Apply relative positional encoding (rotary embedding)
        # ======================================================
8758

8759
        if rotary_pos_emb is not None:
8760
8761
8762
            assert not isinstance(query_layer, Float8Tensor) and not isinstance(
                key_layer, Float8Tensor
            ), "RoPE is not supported for Float8Tensors!"
8763
            # duplicate the pos_emb for self attention
8764
            if not isinstance(rotary_pos_emb, tuple):
8765
                rotary_pos_emb = (rotary_pos_emb,) * 2
8766
8767

            q_pos_emb, k_pos_emb = rotary_pos_emb
8768
8769
8770
8771
8772
8773
8774
8775
8776
8777
8778
8779
8780
8781

            # adjust key and value for inference
            if inference_params is not None:
                if self.qkv_format == "sbhd":
                    sequence_length = key_layer.size(0)
                elif self.qkv_format == "bshd":
                    sequence_length = key_layer.size(1)

                sequence_start = inference_params.sequence_len_offset
                sequence_end = sequence_start + sequence_length

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

8782
8783
            query_layer = apply_rotary_pos_emb(query_layer, q_pos_emb, self.qkv_format, fused=True)
            key_layer = apply_rotary_pos_emb(key_layer, k_pos_emb, self.qkv_format, fused=True)
8784

8785
8786
8787
8788
        # ===========================
        # Core attention computation
        # ===========================

8789
8790
8791
8792
        context_layer = self.core_attention(
            query_layer,
            key_layer,
            value_layer,
8793
            qkv_format=self.qkv_format,
8794
8795
8796
8797
            cu_seqlens_q=cu_seqlens_q,
            cu_seqlens_kv=cu_seqlens_kv,
            max_seqlen_q=max_seqlen_q,
            max_seqlen_kv=max_seqlen_kv,
8798
8799
            attention_mask=attention_mask,
            attn_mask_type=attn_mask_type,
8800
            window_size=window_size,
8801
8802
8803
            checkpoint_core_attention=checkpoint_core_attention,
            core_attention_bias_type=core_attention_bias_type,
            core_attention_bias=core_attention_bias,
8804
            alibi_slopes=alibi_slopes,
8805
            fast_zero_fill=fast_zero_fill,
8806
            inference_params=inference_params,
8807
8808
        )

8809
        # ===================
8810
        # Output. [sq, b, h]
8811
        # ===================
8812

8813
        projection_output = self.proj(
8814
8815
            context_layer,
            is_first_microbatch=is_first_microbatch,
8816
8817
        )

8818
8819
8820
8821
8822
8823
8824
8825
        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,)
8826
        if self.input_layernorm and self.return_layernorm_output:
8827
8828
            outputs += (layernorm_output,)
        return outputs if len(outputs) > 1 else outputs[0]