transformer.py 86.3 KB
Newer Older
1
# Copyright (c) 2022-2024, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
2
3
4
5
6
7
8
#
# See LICENSE for license information.
"""
Wrapper module for Transformer related layers with FP8 support.
"""
import functools
from enum import Enum
9
from math import sqrt
10
import os
11
from typing import Any, Callable, Optional, Sequence, Tuple, Union
12
import warnings
13

14
import jax
15
16
17
18
import jax.numpy as jnp
import numpy as np
from flax import linen as nn
from flax.linen import partitioning as nn_partitioning
19
from flax.linen.attention import combine_masks
20
21
22
from jax import nn as jax_nn
from jax import random as jax_random
from jax import lax, vmap
23
from jax.ad_checkpoint import checkpoint_name
24
25
26

from .module import DenseGeneral, LayerNormDenseGeneral, LayerNormMLP
from .module import LayerNorm, Softmax
27
from ..attention import AttnBiasType, AttnMaskType, QKVLayout
28
from ..attention import is_fused_attn_kernel_available, make_swa_mask, canonicalize_attn_mask_type
29
from ..attention import fused_attn
30
from ..softmax import SoftmaxType
31
32
33
34
35
36
from ..sharding import num_of_devices
from ..sharding import get_sharding_map_logic_axis_to_mesh_axis
from ..sharding import with_sharding_constraint_by_logical_axes
from ..sharding import BATCH_AXES, SEQLEN_AXES, SEQLEN_TP_AXES, HEAD_AXES
from ..sharding import HIDDEN_AXES, HIDDEN_TP_AXES, JOINED_AXES
from ..sharding import W_NO_SHARD_AXES, W_FSDP_AXES, W_TP_AXES, W_JOINED_AXES
37
38
39
40
41

PRNGKey = Any
Shape = Tuple[int, ...]
DType = jnp.dtype
Array = jnp.ndarray
42
43
44
PrecisionLike = Union[
    None, str, lax.Precision, Tuple[str, str], Tuple[lax.Precision, lax.Precision]
]
45
46
47
48
49
50
51
52
53
54
55
56
57
Initializer = Callable[[PRNGKey, Shape, DType], Array]
LogicalRules = Sequence[Tuple[str, Union[str, None]]]


def _generate_drop_path_shape(shape: Sequence[int], batch_dim: int) -> Sequence[int]:
    # Generate broadcast dims for drop_path.
    drop_path_shape = list(range(0, len(shape)))
    drop_path_shape.pop(batch_dim)
    return drop_path_shape


def extend_logical_axis_rules(rules: LogicalRules) -> LogicalRules:
    """
58
    Extend the given Flax logical axis rules with the predefined TransformerLayer's
59
60
61
    logical axis rules.

    .. note::
Ming-Xu Huang's avatar
Ming-Xu Huang committed
62
63
64
65
        We currently only support logical axis rules for single GPU training, data parallel
        training and 1D-sharding tensor parallel training.
        Refer to `Figure 3 in` `Megatron-LM tensor parallel <https://arxiv.org/pdf/1909.08053.pdf>`_
        for 1D-sharding tensor parallelism.
66
67
68
69

    .. warning::
        Please make sure ShardingResource is set via fp8_autocast before calling this function.

Ming-Xu Huang's avatar
Ming-Xu Huang committed
70
71
72
73
    .. note::
        This function is only needed when using TransformerLayer. For  other modules, such as
        DenseGeneral, please properly set axes of kernels and bias.

74
75
    Parameters
    ----------
76
    rules: Sequence[Tuple[str, Union[str, None]]]
77
78
79
80
        the base Flax logical axis rules to extend.

    Returns
    -------
81
    extended_rules: Sequence[Tuple[str, Union[str, None]]]
82
83
84
85
        the extended Flax logical axis rules.
    """
    rules_map = {}
    for item in rules:
86
        assert len(item) == 2, "The logical axis rule should be like (axis_name, mesh_axis_name)."
87
88
        key = item[0]
        val = item[1]
89
90
91
92
        assert isinstance(key, str), f"Thie axis_name should be str, but got {type(key)}."
        assert isinstance(val, str) or (
            val is None
        ), f"Thie mesh_axis_name should be str or None, but got {type(val)}."
Ming-Xu Huang's avatar
Ming-Xu Huang committed
93
94
95
96
        if key in rules_map:
            rules_map[key].append(val)
        else:
            rules_map[key] = [val]
97
98

    extended_rules = [*rules]
99
    for item in get_sharding_map_logic_axis_to_mesh_axis().items():
100
101
102
        key = item[0]
        val = item[1]
        if key in rules_map:
103
104
105
            assert len(rules_map[key]) == 1 and rules_map[key][0] == val, (
                "The rule diverged between TE and given rule."
                f"Axis:{key} map to {rules_map[key]} in the given"
106
                f" rules, but {val} in TE's rules."
107
            )
108
109
110
111
112
        else:
            extended_rules.append(item)
    return tuple(extended_rules)


113
114
class _UnfusedDotProductAttention(nn.Module):  # pylint: disable=too-few-public-methods
    attention_dropout: float = 0.0
115
116
117
118
119
120
    attn_mask_type: AttnMaskType = AttnMaskType.CAUSAL_MASK
    attn_bias_type: Optional[AttnBiasType] = None
    dtype: DType = jnp.float32
    float32_logits: bool = False
    scale_factor: Optional[float] = None
    transpose_batch_sequence: bool = True
121
    window_size: Optional[Tuple[int, int]] = None
122
123

    @nn.compact
124
125
126
127
128
129
130
131
132
133
134
135
    def __call__(
        self,
        query: Array,
        key: Array,
        value: Array,
        mask: Optional[Array] = None,
        bias: Optional[Array] = None,
        *,
        dropout_rng: Optional[PRNGKey] = None,
        deterministic: bool = False,
    ) -> Array:
        assert key.ndim == query.ndim == value.ndim, "q, k, v must have same rank."
136
        batch_dim = 1 if self.transpose_batch_sequence else 0
137
138
139
        assert (
            query.shape[batch_dim] == key.shape[batch_dim] == value.shape[batch_dim]
        ), "q, k, v batch dims must match."
140
        sequence_dim = 0 if self.transpose_batch_sequence else 1
141
142
143
        assert key.shape[sequence_dim] == value.shape[sequence_dim], "k, v lengths must match."
        assert key.shape[-2] == value.shape[-2], "k, v num_attention_heads must match."
        assert query.shape[-1] == key.shape[-1], "q, k head_dim must match."
144
145
146
147
148
149
150
151
152
153
154
155
156

        if self.scale_factor is None:
            scale_factor = 1.0 / sqrt(query.shape[-1])
        else:
            scale_factor = self.scale_factor
        del self.scale_factor

        if self.float32_logits:
            query = query.astype(jnp.float32)
            key = key.astype(jnp.float32)
        h_q, h_kv = query.shape[-2], key.shape[-2]
        # The generated GQA kernels are slower than normal MHA kernels even when h_q == h_kv.
        # Therefore, we have to maintain two code paths.
157
        is_gqa = h_q != h_kv
158

159
        if is_gqa:
160
161
162
163
164
165
            assert (h_q % h_kv == 0) and (h_q >= h_kv)
            group_size = h_q // h_kv
            grouped_query = query.reshape((*query.shape[:2], h_kv, group_size, query.shape[-1]))

        if self.transpose_batch_sequence:
            if is_gqa:
166
                attn_weights = jnp.einsum("qbhgd,kbhd->bhgqk", grouped_query, key)
167
            else:
168
                attn_weights = jnp.einsum("qbhd,kbhd->bhqk", query, key)
169
        else:
170
            if is_gqa:
171
                attn_weights = jnp.einsum("bqhgd,bkhd->bhgqk", grouped_query, key)
172
            else:
173
                attn_weights = jnp.einsum("bqhd,bkhd->bhqk", query, key)
174

175
        attn_weights = checkpoint_name(attn_weights, "logits")
176

177
        if is_gqa:
178
179
180
181
182
            b, h, g, q, k = attn_weights_with_groups_shape = attn_weights.shape
            attn_weights_without_groups_shape = (b, h * g, q, k)
            attn_weights = attn_weights.reshape(attn_weights_without_groups_shape)

        attn_weights = with_sharding_constraint_by_logical_axes(
183
184
            attn_weights, (BATCH_AXES, HEAD_AXES, SEQLEN_AXES, SEQLEN_AXES)
        )
185
186
187
188
189

        # When post_scale_bias is present, the computation is Softmax(attn_weights * scale + bias)
        # In this case, the scale can not fused into the Softmax module.
        if self.attn_bias_type == AttnBiasType.POST_SCALE_BIAS:
            attn_weights = attn_weights * scale_factor
190
            fused_scale_factor = 1.0
191
        else:
192
193
194
195
196
            # If not post_scale_bias, the scale can be fused into Softmax module
            fused_scale_factor = scale_factor
            if self.attn_bias_type == AttnBiasType.PRE_SCALE_BIAS:
                attn_weights += bias

197
198
199
200
201
202
203
204
205
206
207
        def apply_swa_mask(attn_mask_type: AttnMaskType, original_mask: Array) -> Array:
            """Apply the sliding window mask to a given mask"""
            max_seqlen_q = original_mask.shape[-2]
            max_seqlen_kv = original_mask.shape[-1]
            swa_mask = make_swa_mask(max_seqlen_q, max_seqlen_kv, self.window_size, attn_mask_type)
            # In swa_mask 0 is masked out, in original_mask 1 is masked out
            swa_mask = 1 - swa_mask.astype(original_mask.dtype)
            swa_mask_bcast = jnp.broadcast_to(swa_mask, original_mask.shape)
            new_mask = jnp.where(original_mask == 0, swa_mask_bcast, original_mask)
            return new_mask

208
209
        def convert_to_softmax_type(attn_mask_type, mask):
            """Convert the attn_mask_type to SoftmaxType"""
210
211
212
213
            # mask is ignored for no_mask and causal_mask without sliding window
            if attn_mask_type == AttnMaskType.NO_MASK:
                mask = None
            if attn_mask_type == AttnMaskType.CAUSAL_MASK and self.window_size is None:
214
                mask = None
215
216
217
            if mask is not None:
                mask = apply_swa_mask(attn_mask_type, mask)
            # Currently cuDNN backend only supports SWA for causal/padding_causal, follow this
218
            if attn_mask_type in [AttnMaskType.CAUSAL_MASK, AttnMaskType.PADDING_CAUSAL_MASK]:
219
                return SoftmaxType.SCALED_UPPER_TRIANG_MASKED, mask
220
221
            if attn_mask_type in [AttnMaskType.NO_MASK, AttnMaskType.PADDING_MASK]:
                if mask is not None:
222
223
                    return SoftmaxType.SCALED_MASKED, mask
                return SoftmaxType.SCALED, mask
224
225
226
227
            raise ValueError(
                f"Unsupported {attn_mask_type=}, supported attn_mask_type="
                "{'no_mask', 'padding', 'causal', 'padding_causal', 'causal_padding'}"
            )
228

229
        softmax_type, mask = convert_to_softmax_type(self.attn_mask_type, mask)
230

231
232
233
        attn_weights = Softmax(softmax_type=softmax_type, scale_factor=fused_scale_factor)(
            attn_weights, mask, bias
        ).astype(self.dtype)
234

235
        if is_gqa:
236
237
            attn_weights = attn_weights.reshape(attn_weights_with_groups_shape)

238
        if not deterministic and self.attention_dropout > 0.0:
239
240
241
242
            keep_prob = 1.0 - self.attention_dropout
            dropout_shape = list(attn_weights.shape)
            # TODO(rewang): add attention dropout broadcast dimension arguments for users
            keep = jax_random.bernoulli(dropout_rng, keep_prob, dropout_shape)
243
            multiplier = keep.astype(attn_weights.dtype) / jnp.asarray(keep_prob, dtype=self.dtype)
244
245
246
247
            attn_weights = attn_weights * multiplier

        if self.transpose_batch_sequence:
            if is_gqa:
248
249
                return jnp.einsum("bhgqk,kbhd->qbhgd", attn_weights, value).reshape(query.shape)
            return jnp.einsum("bhqk,kbhd->qbhd", attn_weights, value)
250
251

        if is_gqa:
252
253
            return jnp.einsum("bhgqk,bkhd->bqhgd", attn_weights, value).reshape(query.shape)
        return jnp.einsum("bhqk,bkhd->bqhd", attn_weights, value)
254
255


256
257
class _FusedDotProductAttention(nn.Module):  # pylint: disable=too-few-public-methods
    attention_dropout: float = 0.0
258
259
260
261
262
263
    attn_mask_type: AttnMaskType = AttnMaskType.CAUSAL_MASK
    attn_bias_type: Optional[AttnBiasType] = None
    dtype: DType = jnp.float32
    qkv_layout: QKVLayout = QKVLayout.BSHD_BSHD_BSHD
    scale_factor: Optional[float] = None
    transpose_batch_sequence: bool = False
264
    window_size: Optional[Tuple[int, int]] = None
265
266

    @nn.compact
267
268
269
270
271
272
273
274
275
276
277
    def __call__(
        self,
        query: Array,
        key: Array,
        value: Array,
        mask: Optional[Array] = None,
        bias: Optional[Array] = None,
        *,
        dropout_rng: Optional[PRNGKey] = None,
        deterministic: bool = False,
    ) -> Array:
278
279
280
281
282
283
284
285
286
287
288

        seed = None
        if dropout_rng is not None:
            seed = jax.random.split(dropout_rng, num_of_devices())

        if self.scale_factor is None:
            scale_factor = 1.0 / sqrt(query.shape[-1])
        else:
            scale_factor = self.scale_factor
        del self.scale_factor

289
        # TODO(rewang): integrate THD format
290
291
292
293
294
295
296
297
298
        if self.qkv_layout == QKVLayout.BS3HD:
            """qkvpacked format, treat
            query: qkvpacked tensor, shape = [..., 3, h, d]
            key: ignore
            value: ignore
            """
            qkv_packed = query
            if self.transpose_batch_sequence:
                qkv_packed = qkv_packed.transpose([1, 0, 2, 3, 4])
299
300
            x = fused_attn(
                (qkv_packed,),
301
302
303
304
305
                bias,
                mask,
                seed,
                attn_mask_type=self.attn_mask_type,
                attn_bias_type=self.attn_bias_type,
306
                qkv_layout=self.qkv_layout,
307
308
309
                scaling_factor=scale_factor,
                dropout_probability=self.attention_dropout,
                is_training=not deterministic,
310
                window_size=self.window_size,
311
            )
312
313
314
315
316
317
318
319
320
321
        elif self.qkv_layout == QKVLayout.BSHD_BS2HD:
            """kvpacked format, treat
            query: query tensor, shape = [..., h, d]
            key: kvpacked tensor, shape = [..., 2, h, d]
            value: ignore
            """
            kv_packed = key
            if self.transpose_batch_sequence:
                query = query.transpose([1, 0, 2, 3])
                kv_packed = kv_packed.transpose([1, 0, 2, 3, 4])
322
323
            x = fused_attn(
                (query, kv_packed),
324
325
326
327
328
                bias,
                mask,
                seed,
                attn_mask_type=self.attn_mask_type,
                attn_bias_type=self.attn_bias_type,
329
                qkv_layout=self.qkv_layout,
330
331
332
                scaling_factor=scale_factor,
                dropout_probability=self.attention_dropout,
                is_training=not deterministic,
333
                window_size=self.window_size,
334
            )
335
336
337
338
339
        elif self.qkv_layout == QKVLayout.BSHD_BSHD_BSHD:
            if self.transpose_batch_sequence:
                query = query.transpose([1, 0, 2, 3])
                key = key.transpose([1, 0, 2, 3])
                value = value.transpose([1, 0, 2, 3])
340
            x = fused_attn(
341
                (query, key, value),
342
343
344
345
346
                bias,
                mask,
                seed,
                attn_mask_type=self.attn_mask_type,
                attn_bias_type=self.attn_bias_type,
347
                qkv_layout=self.qkv_layout,
348
349
350
                scaling_factor=scale_factor,
                dropout_probability=self.attention_dropout,
                is_training=not deterministic,
351
                window_size=self.window_size,
352
            )
353
354
355
356
357
358
359
360
361
        else:
            raise ValueError(f"Unsupported {self.qkv_layout=}.")

        if self.transpose_batch_sequence:
            x = x.transpose([1, 0, 2, 3])

        return x


362
class DotProductAttention(nn.Module):  # pylint: disable=too-few-public-methods
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
    r"""
    Dot Product Attention (DPA). 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::
        The DotProductAttention module supports two backends: the unfused and the fused attention
        mechanisms. The unfused attention is implemented using JAX native operations, providing
        broad compatibility and flexibility. In contrast, the fused attention uses `cuDNN fused
        attention
        <https://github.com/NVIDIA/cudnn-frontend/blob/main/docs/operations/Attention.md>`_ for
        higher performance and lower memory usage on the supported hardwares.
        Users can select between these two backends via the :attr:`NVTE_FUSED_ATTN` environment
        variable:

        * Set :attr:`NVTE_FUSED_ATTN=0` for unfused attention (default).
        * Set :attr:`NVTE_FUSED_ATTN=1` for fused attention. If the required cuDNN fused attention
          kernel is not available on the system, a warning will be issued, and the module will
          automatically fall back to the unfused backend.

383
384
385
386
387
388
389
390
    .. note::
        The DotProductAttention default setting enables non-deterministic kernels for reduced
        workspace requirements and faster computation. Users can disable the non-deterministic
        kernels via the :attr:`NVTE_ALLOW_NONDETERMINISTIC_ALGO` environment variable:

        * :attr:`NVTE_ALLOW_NONDETERMINISTIC_ALGO=0` to allow only deterministic kernels.
        * :attr:`NVTE_ALLOW_NONDETERMINISTIC_ALGO=1` to allow non-deterministic kernels (default).

391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
    Parameters
    ----------
    head_dim: int
        The hidden dimension of each attention head.
    num_attention_heads: int
        The number of attention heads.
    num_gqa_groups: int, default = `None`
        Number of GQA groups. When `None` is present, it is equal to num_attention_heads.
        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`.
    attention_dropout: float, default = 0.0
        Dropout probability for the dropout op after the softmax.
    attn_mask_type: str, default = 'causal'
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
        This parameter specifies the type of attention mask to be applied during the softmax
        operation.
        Available options are {'no_mask', 'padding', 'causal', 'causal_padding', 'padding_causal'}

        Each described below:

        * no_mask: No attention mask is applied. This means the attention will consider the
          full sequence without any restrictions.
        * padding: Indicates the presence of padding at the end of each sequence.
          Users must provide a mask with the shape [batch, 1, max_seqlen_q, max_seqlen_kv] in the
          :attr:`__call__` method to specify the padding positions.
        * causal: An upper triangular mask is applied to the softmax inputs,
          ensuring that the prediction for a certain position is only dependent on known outputs
          from positions before it.
        * causal_padding / padding_causal: A combination of both causal and padding masks.
          Both 'causal_padding' and 'padding_causal' are acceptable and have the same effect.

        .. note:: :attr:`mask` in :attr:`__call__` is ignored for 'no_mask' and 'causal'.

427
    attn_bias_type: Optional[str], default = None
428
        Type of the attention bias passed in the attention.
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
        Available options: {'no_bias', 'pre_scale_bias', 'post_scale_bias'}.
        When default is present, the type is automatically decided by the MHA's bias parameter.
        Where it is :attr:`post_scale_bias` if there is bias. Otherwise :attr:`no_bias` is used.
    dropout_rng_name: str, default = 'dropout'
        The key in given RNGs via flax.linen.Module.apply that is used
        to generate Dropout masks in the core attention.
    float32_logits: bool, default = False
        Whether to compute attention logits in float32 for the unfused attention backend.
        For fused attention backend, the accumulation is always float32 without the perf overhead.
    qkv_layout: str, default = 'bshd_bshd_bshd'
        Specifies the dimensional layout format for the query, key, and value tensors in __call__().
        It indicates how the inputs are processed.
        Available options: {'bs3hd', 'bshd_bs2hd', 'bshd_bshd_bshd'}. Where

        * bs3hd: query tensor is treated as a qkvpacked tensor with shape = [b, s, 3, h, d].
          key and value arguments in :attr:`__call__()` are ignored in this layout.
        * bshd_bs2hd: query tensor with shape = [b, s, h, d]. key tensor is treaded as a kvpacked
          tensor with shape = [b, s, 2, h, d]. `value` argument in :attr:`__call__()` is ignored.
        * bshd_bshd_bshd: query, key, and value are seperated with shape = [b, s, h, d].

        Explanation of denotations:

        * b: batch size
        * s: seqeuence length
        * h: num_attention_heads or num_gqa_groups
        * d: head dimension

    scale_factor: Optional[float], default = None
        Scale factor to apply on query. When :attr:`None` is present, the scale factor is equal
        to :math:`\frac{1}{\sqrt{head\_dim}}`. This is useful for model like T5X, which doesn't
        need to apply scale on query, which is to set :attr:`scale_factor=1.`.
    transpose_batch_sequence: bool, default = True
        Indicate whether the input tensors were switched axis of batch
        and sequence length dimension. if set to True, the input tensors
        should be in (seqlen, batch, ...), otherwise (batch, seqlen, ...).
464
465
    window_size: Optional[Tuple[int, int]], default = None
        Sliding window size. The default value is no sliding window.
466
467
468
469
470
471

    Optimization parameters
    -----------------------
    dtype: jax.numpy.dtype, default = jax.numpy.float32
        The data type used to allocate the initial parameters.
    """
472

473
474
475
    head_dim: int
    num_attention_heads: int
    num_gqa_groups: Optional[int] = None
476
477
    attention_dropout: float = 0.0
    attn_mask_type: AttnMaskType = "causal"
478
479
    attn_bias_type: AttnBiasType = None
    dtype: DType = jnp.float32
480
    dropout_rng_name: str = "dropout"
481
    float32_logits: bool = False
482
    qkv_layout: str = "bshd_bshd_bshd"
483
484
    scale_factor: Optional[float] = None
    transpose_batch_sequence: bool = True
485
    window_size: Optional[Tuple[int, int]] = None
486
487

    @nn.compact
488
489
490
491
492
493
494
495
496
497
    def __call__(
        self,
        query: Array,
        key: Array,
        value: Array,
        mask: Optional[Array] = None,
        bias: Optional[Array] = None,
        *,
        deterministic: bool = False,
    ) -> Array:
498
499
500
501
502
503
504
505
506
507
508
509
        """
        Parameters
        ----------
        query: jax.numpy.ndarray
            The details of query tensor representation is described in :attr:`qkv_layout`.
        key: jax.numpy.ndarrary
            The details of kery tensor representation is described in :attr:`qkv_layout`.
        value: jax.numpy.ndarrary
            The details of value tensor representation is described in :attr:`qkv_layout`.
        mask: jax.numpy.ndarray, default = None
            Boolean tensor used to mask out the attention softmax input.
            :attr:`True` means to mask out the corresponding values.
510
            Ignored when :attr:`self.attn_mask_type` is either 'no_mask' or 'causal'.
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
        bias: jax.numpy.ndarray, default = None
            A tensor used to shift attention softmax input.
        *:
            Below parameters are keyword only
        deterministic: bool, default = False
            Disable dropout layers if set to True.

        Returns
        -------
        outputs: jax.numpy.ndarray
            Output tensors.
        """

        # For internal API, we use enum to maintain
        if self.attn_bias_type is None:
            attn_bias_type = AttnBiasType.NO_BIAS if bias is None else AttnBiasType.POST_SCALE_BIAS
        else:
            attn_bias_type = AttnBiasType[self.attn_bias_type.upper()]
        attn_mask_type = canonicalize_attn_mask_type(self.attn_mask_type)
        qkv_layout = QKVLayout[self.qkv_layout.upper()]
        del self.attn_bias_type, self.attn_mask_type, self.qkv_layout

        if attn_bias_type == AttnBiasType.NO_BIAS:
            assert bias is None
        else:
            assert bias is not None

        enable_fused_attn = int(os.getenv("NVTE_FUSED_ATTN", "0"))

        sequence_dim = 0 if self.transpose_batch_sequence else 1
        seqlen_q = query.shape[sequence_dim]
        if qkv_layout == QKVLayout.BS3HD:
            seqlen_kv = seqlen_q
        else:
            seqlen_kv = key.shape[sequence_dim]

547
548
549
550
551
552
553
554
555
556
557
558
        has_fused_attn_kernel = is_fused_attn_kernel_available(
            self.dtype,
            self.dtype,
            qkv_layout,
            attn_bias_type,
            attn_mask_type,
            self.attention_dropout,
            self.num_attention_heads,
            self.num_gqa_groups,
            seqlen_q,
            seqlen_kv,
            self.head_dim,
559
            self.window_size,
560
        )
561

562
        use_fused_attn = enable_fused_attn and has_fused_attn_kernel
563
564

        if enable_fused_attn and not has_fused_attn_kernel:
565
566
567
568
569
570
571
572
            warnings.warn(
                "Fused attention is not enabled because there is no available kernel.\n"
                "Fall back to the unfused attention.\n"
                "Please try to update the cuDNN and TE to the latest version.\n"
                f"{self.dtype=}\n{qkv_layout=}\n{attn_bias_type=}\n{attn_mask_type=}\n"
                f"{self.attention_dropout=}\n{self.num_attention_heads=}\n"
                f"{self.num_gqa_groups=}\n{seqlen_q=}\n{seqlen_kv=}\n{self.head_dim=}\n"
            )
573
574

        dropout_rng = None
575
        if not deterministic and self.attention_dropout > 0.0:
576
577
578
579
580
581
582
583
584
585
586
587
            dropout_rng = self.make_rng(self.dropout_rng_name)

        if self.scale_factor is None:
            scale_factor = 1.0 / sqrt(self.head_dim)
        else:
            scale_factor = self.scale_factor
        del self.scale_factor

        if not use_fused_attn:
            # unfused attention only supports splitted query, key, value
            if qkv_layout == QKVLayout.BS3HD:
                query, key, value = jnp.split(query, [1, 2], axis=-3)
588
589
590
                query, key, value = map(
                    functools.partial(jnp.squeeze, axis=-3), [query, key, value]
                )
591
592
593
594
595
596
            elif qkv_layout == QKVLayout.BSHD_BS2HD:
                key, value = jnp.split(key, [1], axis=-3)
                key, value = map(functools.partial(jnp.squeeze, axis=-3), [key, value])
            else:
                assert qkv_layout == QKVLayout.BSHD_BSHD_BSHD

597
598
599
600
601
602
603
604
            x = _UnfusedDotProductAttention(
                attention_dropout=self.attention_dropout,
                attn_mask_type=attn_mask_type,
                attn_bias_type=attn_bias_type,
                dtype=self.dtype,
                float32_logits=self.float32_logits,
                scale_factor=scale_factor,
                transpose_batch_sequence=self.transpose_batch_sequence,
605
                window_size=self.window_size,
606
            )(query, key, value, mask, bias, dropout_rng=dropout_rng, deterministic=deterministic)
607
608
609
610
611
612
613
614
615
        else:
            x = _FusedDotProductAttention(
                attention_dropout=self.attention_dropout,
                attn_mask_type=attn_mask_type,
                attn_bias_type=attn_bias_type,
                dtype=self.dtype,
                scale_factor=scale_factor,
                transpose_batch_sequence=self.transpose_batch_sequence,
                qkv_layout=qkv_layout,
616
                window_size=self.window_size,
617
            )(query, key, value, mask, bias, dropout_rng=dropout_rng, deterministic=deterministic)
618

619
        return x
620
621


622
623
624
625
626
627
def rotary_pos_emb(
    x: Array,
    windows: Tuple[int, int],
    transpose_batch_sequence: bool,
    group_method: str = "consecutive",
):
628
629
630
    """
    Rotary Positional Embedding
    x should be in shape of
631
632
    [Batch, Seqlen, ..., Heads, Hidden] if transpose_batch_sequence is False, or
    [Seqlen, Batch, ..., Heads, Hidden] if transpose_batch_sequence is True.
633
    """
634
635
    hidden_dim = x.shape[-1]
    half_hidden_dim = hidden_dim // 2
636
637
638
    min_window = windows[0]
    max_window = windows[1]

639
    fraction = 2 * jnp.arange(0, half_hidden_dim) / hidden_dim
640
    time_scales = min_window * (max_window / min_window) ** fraction
641
642
643
644
645
646
647
648
    time_scales = jnp.expand_dims(time_scales, axis=tuple(range(x.ndim - 1)))

    batch_dim = 1 if transpose_batch_sequence else 0
    seq_dim = 1 - batch_dim

    positions = jnp.expand_dims(jnp.arange(x.shape[seq_dim]), axis=batch_dim)
    positions = jnp.expand_dims(positions, axis=tuple(range(2, x.ndim)))

649
650
651
652
653
    def generate_sin_cos(timescales):
        sinusoidal_positions = positions / timescales
        sin = jnp.sin(sinusoidal_positions)
        cos = jnp.cos(sinusoidal_positions)
        return sin, cos
654

655
656
    def alternate_impl():
        sin, cos = generate_sin_cos(time_scales)
657

658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
        x1, x2 = jnp.split(x, 2, axis=-1)
        part_1 = (x1 * cos - x2 * sin).astype(x.dtype)
        part_2 = (x2 * cos + x1 * sin).astype(x.dtype)

        output = jnp.concatenate([part_1, part_2], axis=-1)
        return output

    def consecutive_impl():
        sin, cos = generate_sin_cos(jnp.repeat(time_scales, 2, axis=-1))

        x_shifted_left = jnp.roll(x, -1, axis=-1)
        x_shifted_right = jnp.roll(x, 1, axis=-1)
        x_shifted = jax.lax.select(
            jnp.tile(
                jnp.mod(jnp.arange(hidden_dim, dtype=jnp.int32), 2),
                x.shape[:-1] + (1,),
            ),
            x_shifted_right,
            x_shifted_left,
        )

        sign = jnp.sign(jnp.mod(jnp.arange(hidden_dim, dtype=jnp.int32), 2) - 0.5)

        output = x * cos + x_shifted * sin * sign
        output = output.astype(x.dtype)
        return output

    def canonicalize_group_method(gm):
686
687
688
        canonicalized_gm = gm.lower().strip().replace("-", "").replace("_", "")
        assert canonicalized_gm in ["consecutive", "alternate"], (
            "Invalid relative positional embedding group method. "
689
            f"Expect to be in []'alternate' or 'consecutive'], but got {gm}."
690
        )
691
692
693
694
695

        return canonicalized_gm

    group_method = canonicalize_group_method(group_method)

696
    if group_method == "alternate":
697
698
        return alternate_impl()
    return consecutive_impl()
699
700


701
class LoRAScope:  # pylint: disable=too-few-public-methods
702
703
704
705
706
707
708
709
    """LoRA Scope"""

    def __init__(self, qkv_proj=False, output_proj=False, mlp=False):
        self.qkv_proj = qkv_proj
        self.output_proj = output_proj
        self.mlp = mlp

    def __eq__(self, other):
710
711
712
713
714
        return (self.qkv_proj, self.output_proj, self.mlp) == (
            other.qkv_proj,
            other.output_proj,
            other.mlp,
        )
715
716
717
718


def _canonicalize_lora_scope(scope):

719
720
721
722
723
724
725
726
    SCOPE_NONE = "none"
    SCOPE_ALL = "all"
    SCOPE_QKV_PROJ = "qkv_proj"
    SCOPE_OUTPUT_PROJ = "output_proj"
    SCOPE_MLP = "mlp"
    SCOPE_EX_QKV_PROJ = "exclude_qkv_proj"
    SCOPE_EX_OUTPUT_PROJ = "exclude_output_proj"
    SCOPE_EX_MLP = "exclude_mlp"
727
728
729
730
731
732

    scope = SCOPE_NONE if scope is None else scope

    scope = scope.lower()

    assert scope in [
733
734
735
736
737
738
739
740
        SCOPE_NONE,
        SCOPE_ALL,
        SCOPE_QKV_PROJ,
        SCOPE_OUTPUT_PROJ,
        SCOPE_MLP,
        SCOPE_EX_QKV_PROJ,
        SCOPE_EX_OUTPUT_PROJ,
        SCOPE_EX_MLP,
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
    ]

    lora_scope = LoRAScope()

    if scope in [SCOPE_ALL, SCOPE_QKV_PROJ, SCOPE_EX_OUTPUT_PROJ, SCOPE_EX_MLP]:
        lora_scope.qkv_proj = True

    if scope in [SCOPE_ALL, SCOPE_OUTPUT_PROJ, SCOPE_EX_QKV_PROJ, SCOPE_EX_MLP]:
        lora_scope.output_proj = True

    if scope in [SCOPE_ALL, SCOPE_MLP, SCOPE_EX_QKV_PROJ, SCOPE_EX_OUTPUT_PROJ]:
        lora_scope.mlp = True

    return lora_scope


757
class MultiHeadAttention(nn.Module):  # pylint: disable=too-few-public-methods
758
759
760
761
762
763
    r"""
    Multi-head Attention (MHA), including Query,
    Key, Value and Output projection.

    Parameters
    ----------
764
    head_dim: int
765
        The hidden dimension of each attention head.
766
767
768
769
    num_attention_heads: int
        The number of attention heads.
    num_gqa_groups: int, default = `None`
        Number of GQA groups. When `None` is present, it is equal to num_attention_heads.
zlsh80826's avatar
zlsh80826 committed
770
771
772
773
774
775
        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`.
776
777
778
    attention_dropout: float, default = 0.0
        Dropout probability for the dropout op after the softmax.
    attn_mask_type: str, default = 'causal'
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
        This parameter specifies the type of attention mask to be applied during the softmax
        operation.
        Available options are {'no_mask', 'padding', 'causal', 'causal_padding', 'padding_causal'}

        Each described below:

        * no_mask: No attention mask is applied. This means the attention will consider the
          full sequence without any restrictions.
        * padding: Indicates the presence of padding at the end of each sequence.
          Users must provide a mask with the shape [batch, 1, max_seqlen_q, max_seqlen_kv] in the
          :attr:`__call__` method to specify the padding positions.
        * causal: An upper triangular mask is applied to the softmax inputs,
          ensuring that the prediction for a certain position is only dependent on known outputs
          from positions before it.
        * causal_padding / padding_causal: A combination of both causal and padding masks.
          Both 'causal_padding' and 'padding_causal' are acceptable and have the same effect.

        .. note:: :attr:`mask` in :attr:`__call__` is ignored for 'no_mask' and 'causal'.

798
799
800
801
802
    attn_bias_type: Optional[str], default = None
        Type of the attention bias passed in the attention.
        Available options: {'no_bias', 'pre_scale_bias', 'post_scale_bias'}.
        When default is present, the type is automatically decided by the MHA's bias parameter.
        Where it is `post_scale_bias` if there is bias. Otherwise `no_bias` is used.
803
    dropout_rng_name: str, default = 'dropout'
804
805
        The key in given RNGs via flax.linen.Module.apply that is used
        to generate Dropout masks in the core attention.
806
    layernorm_type: {'layernorm', 'rmsnorm'}, default = 'layernorm'
807
        Indicate the type of layer normalization.
808
    layernorm_epsilon: float, default = 1e-6
809
        A value added to the denominator of layer normalization for numerical stability.
810
    zero_centered_gamma: bool, default = False
811
812
813
814
815
816
817
        If set to `True`, the LayerNorm formula changes to

        .. math::
            y = \frac{x - \mathrm{E}[x]}{ \sqrt{\mathrm{Var}[x] + \epsilon}} *
            (1 + \gamma) + \beta

        This parameter is only applicable for 'layernorm'.
818
819
    kernel_init: Initializer, default =
        flax.linen.initializers.variance_scaling(1.0, 'fan_in', 'normal')
820
        Used for initializing the QKV and output projection weights.
821
        It should be a callable object with three arguments (jax.random.PRNGKey, shape, dtype).
822
    use_bias: bool, default = False
823
        Indicate whether or not to enable bias shifting for QKV and output projections.
824
        If set to False, the layer will not learn additive biases.
825
    bias_init: Initializer, default = flax.linen.initializers.zeros
826
827
        Used for initializing bias of QKVO projections, only used when :attr:`use_bias=True`.
        It should be a callable object with three arguments (jax.random.PRNGKey, shape, dtype).
828
829
830
831
832
833
    input_layernorm: bool, default = True
        If set to False, layer normalization to the input is not applied.
    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.
834
835
836
837
838
    enable_rotary_pos_emb: bool, default = False
        Whether to enable rotary position embedding to projected query and key.
    rotary_pos_emb_windows: Tuple[int, int], default = (1, 10000)
        Indicate the min and max time-scales of rotary position embedding,
        only used when :attr:`enable_rotary_pos_emb=True`
839
840
841
842
    rotary_pos_emb_group_method: str, default = 'consecutive'
        Indicate the method to coupled the coordinates. It should be one of
        ['consecutive', 'alternate']. 'alternate' is to pair index :math:`i` with :math:`i + d/2`
        , d is the hidden dimension. 'consecutive' pairs index :math:`i` with :math:`i + 1`.
843
844
845
846
847
848
849
850
851
    low_rank_adaptation_scope: str, default = 'none'
        Indicate the scope to apply low rank adaptation. It should be one of
        ['none', 'all', 'qkv_proj', 'output_proj', 'exclude_qkv_proj', 'exclude_output_proj']
    low_rank_adaptation_dim: int, default = 32
        The dimension for low rank adaptation, only used when
        :attr:`enable_low_rank_adaptation=True`
    low_rank_adaptation_alpha: float, default = None
        The alpha for computing the scaling factor of LoRA output.
        :math:`\frac{alpha}{rank} * lora_output`. None means no scaling.
852
853
    enable_sequence_parallel: bool, default = False
        Whether to enable sequence parallelism to operations except dot.
854
855
856
857
858
859
860
861
    num_heads: int, default = None
        Deprecated. Please refer `num_attention_heads`.
    dropout_rate: float, default = None
        Deprecated. Please refer `attention_dropout`.
    output_layernorm: bool, default = None
        Deprecated. Please refer `input_layernorm`
    apply_residual_connection_post_layernorm: bool, default = None
        Deprecated. Please refer `return_layernorm_output`.
862
863
864

    Optimization parameters
    -----------------------
865
    dtype: jax.numpy.dtype, default = jax.numpy.float32
866
        The data type used to allocate the initial parameters.
867
    fuse_qkv_params: bool, default = True
868
        If set to True, this module exposes a single fused
869
870
        parameter for query-key-value for self-attention and key-value for
        cross-attention.
871
    transpose_batch_sequence: bool, default = True
872
        Indicate whether the input tensors were switched axis of batch
873
874
875
        and sequence length dimension. if set to True, the input tensors
        should be in (seqlen, batch, hidden), otherwise (batch, seqlen, hidden).
    scale_attn_logits: bool, default = False
876
        Indicate whether to scale attention logits.
877
        If set to True, :math:`\frac{Q}{\sqrt{head\_dim}*K}`,
878
        else :math:`Q*K`
879
880
881
882
883
884
885
    scaled_query_init: bool, default = True
        Whether to scale WQ on initialization by :math:`\frac{1}{\sqrt{head\_dim}}`
    float32_logits: bool, default = False
        Whether to compute attention logits in float32 for the unfused attention backend.
        For fused attention backend, the accumulation is always float32 without the perf overhead.
    fuse_qkv: bool, default = None
        Deprecated. Please refer `fuse_qkv_params`
886
887
    window_size: Optional[Tuple[int, int]], default = None
        Sliding window size. Default value is no sliding window.
888
889
890
    """

    head_dim: int
891
892
    num_attention_heads: int
    num_gqa_groups: Optional[int] = None
893
894
    attention_dropout: float = 0.0
    dropout_rng_name: str = "dropout"
895
    input_layernorm: bool = True
896
897
    layernorm_type: str = "layernorm"
    layernorm_epsilon: float = 1e-6
898
    return_layernorm_output: bool = False
899
    zero_centered_gamma: bool = False
900
901
902
    kernel_init: Initializer = None
    use_bias: bool = False
    bias_init: Initializer = nn.initializers.zeros
903
    attn_mask_type: str = "causal"
904
    attn_bias_type: Optional[str] = None
905
906
    enable_rotary_pos_emb: bool = False
    rotary_pos_emb_windows: Tuple[int, int] = (1, 10000)
907
908
    rotary_pos_emb_group_method: str = "consecutive"
    low_rank_adaptation_scope: str = "none"
909
910
    low_rank_adaptation_dim: int = 32
    low_rank_adaptation_alpha: float = None
911
    dtype: DType = jnp.float32
912
    fuse_qkv_params: bool = True
913
    transpose_batch_sequence: bool = True
914
    enable_sequence_parallel: bool = False
915
916
    scale_attn_logits: bool = False
    scaled_query_init: bool = True
917
    float32_logits: bool = False
918
    window_size: Optional[Tuple[int, int]] = None
919
920
921
922
923
924
925

    # Deprecated parameters
    num_heads: Optional[int] = None
    dropout_rate: Optional[float] = None
    output_layernorm: Optional[bool] = None
    apply_residual_connection_post_layernorm: Optional[bool] = None
    fuse_qkv: Optional[bool] = None
926
927

    def __post_init__(self):
928
929
930
931
932
        # Deal with the deprecated parameters
        if self.num_heads is not None:
            self.num_attention_heads = self.num_heads
            warnings.warn(
                f"{__class__}.num_heads is deprecated. It will be removed recently. "
933
934
935
                f"Please uses {__class__}.num_attention_heads as the new API.",
                DeprecationWarning,
            )
936
937
938
939
        if self.dropout_rate is not None:
            self.attention_dropout = self.dropout_rate
            warnings.warn(
                f"{__class__}.dropout_rate is deprecated. It will be removed recently. "
940
941
942
                f"Please use {__class__}.attention_dropout as the new API.",
                DeprecationWarning,
            )
943
944
945
946
        if self.apply_residual_connection_post_layernorm is not None:
            warnings.warn(
                f"{__class__}.apply_residual_connection_post_layernorm is deprecated. "
                f"It will be removed recently, please use {__class__}.return_layernorm_output.",
947
948
                DeprecationWarning,
            )
949
950
951
        if self.fuse_qkv is not None:
            warnings.warn(
                f"{__class__}.fuse_qkv is deprecated. It will be removed recently. "
952
953
954
                f"Please use {__class__}.fuse_qkv_params as the new API.",
                DeprecationWarning,
            )
955
956
        assert self.output_layernorm is None, (
            f"{__class__}.output_layernorm is deprecated. It will be removed recently. "
957
958
            f"Please use {__class__}.input_layernorm for controlling whether to apply layernorm."
        )
959

960
        if self.kernel_init is None:
961
            self.kernel_init = nn.initializers.variance_scaling(1.0, "fan_in", "normal")
zlsh80826's avatar
zlsh80826 committed
962
        if self.num_gqa_groups is None:
963
            self.num_gqa_groups = self.num_attention_heads
964
965
966
        super().__post_init__()

    @nn.compact
967
968
969
970
971
972
973
974
975
976
    def __call__(
        self,
        inputs_q: Array,
        inputs_kv: Array,
        mask: Optional[Array] = None,
        bias: Optional[Array] = None,
        *,
        decode: bool = False,
        deterministic: bool = False,
    ) -> Array:
977
978
979
980
981
982
        """
        MultiHeadAttention Layer:
        [Query, Key, Value projection] -> Dot Product Attention -> Output projection.

        Parameters
        ----------
983
        inputs_q: jax.numpy.ndarray
984
            Input tensor for query projection.
985
        inputs_kv: jax.numpy.ndarray
986
            Input tensor for key/value projection.
987
988
989
        mask: jax.numpy.ndarray, default = None
            Boolean tensor used to mask out the attention softmax input.
            :attr:`True` means mask out the corresponding values.
990
            Ignored when :attr:`self.attn_mask_type` is either 'no_mask' or 'causal'.
991
992
        bias: jax.numpy.ndarray, default = None
            A tensor used to shift the attention softmax input.
993
        *
994
        decode: bool, default = False
995
            Indicate whether to prepare and use an autoregressive cache.
996
        deterministic: bool, default = False
997
998
999
1000
            Disable dropout layers if set to True.

        Returns
        -------
1001
        outputs: jax.numpy.ndarray
1002
1003
            Output tensors.
        """
1004
1005

        def query_init(*args):
1006
            depth_scaling = jnp.sqrt(self.head_dim).astype(self.dtype)
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
            return self.kernel_init(*args) / (depth_scaling if self.scaled_query_init else 1.0)

        def qkv_init(key, shape, dtype):
            assert len(shape) == 3
            assert shape[-2] == 3

            q_key, k_key, v_key = jax_random.split(key, num=3)

            q_shape = (shape[0], shape[-1])
            k_shape = (shape[0], shape[-1])
            v_shape = (shape[0], shape[-1])

            q_kernel = query_init(q_key, q_shape, dtype)
            k_kernel = self.kernel_init(k_key, k_shape, dtype)
            v_kernel = self.kernel_init(v_key, v_shape, dtype)

            return jnp.stack([q_kernel, k_kernel, v_kernel], axis=-2, dtype=dtype)

        def kv_init(key, shape, dtype):
            assert len(shape) == 3
            assert shape[-2] == 2

            k_key, v_key = jax_random.split(key)

            k_shape = (shape[0], shape[-1])
            v_shape = (shape[0], shape[-1])

            k_kernel = self.kernel_init(k_key, k_shape, dtype)
            v_kernel = self.kernel_init(v_key, v_shape, dtype)

            return jnp.stack([k_kernel, v_kernel], axis=-2, dtype=dtype)

1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
        def generate_batch_seqlen_logical_axes(is_sharded_seq):
            sequence_dim = 0 if self.transpose_batch_sequence else 1
            batch_dim = 1 - sequence_dim

            axes = [None, None]

            axes[batch_dim] = BATCH_AXES
            axes[sequence_dim] = SEQLEN_TP_AXES if is_sharded_seq else SEQLEN_AXES
            return tuple(axes)

1049
1050
1051
        is_self_attn = inputs_q is inputs_kv
        is_gqa = self.num_attention_heads != self.num_gqa_groups
        is_qkvpack = is_self_attn and not is_gqa
1052

1053
1054
1055
1056
        inputs_logical_axes_maybe_sp = (
            *generate_batch_seqlen_logical_axes(self.enable_sequence_parallel),
            HIDDEN_AXES,
        )
1057
1058
1059
1060
        inputs_logical_axes_no_sp = (*generate_batch_seqlen_logical_axes(False), HIDDEN_AXES)

        inputs_q = with_sharding_constraint_by_logical_axes(inputs_q, inputs_logical_axes_maybe_sp)

1061
1062
        lora_scope = _canonicalize_lora_scope(self.low_rank_adaptation_scope)

1063
        if self.fuse_qkv_params:
zlsh80826's avatar
zlsh80826 committed
1064
            if is_qkvpack:
1065
                qkv_proj, ln_out = LayerNormDenseGeneral(
1066
                    enable_layernorm=self.input_layernorm,
1067
                    layernorm_type=self.layernorm_type,
1068
                    zero_centered_gamma=self.zero_centered_gamma,
1069
1070
                    epsilon=self.layernorm_epsilon,
                    axis=-1,
1071
                    features=(3, self.num_attention_heads * self.head_dim),
1072
                    transpose_batch_sequence=self.transpose_batch_sequence,
1073
                    return_layernorm_output=self.return_layernorm_output,
1074
1075
1076
                    scale_axes=(W_NO_SHARD_AXES,),
                    ln_bias_axes=(W_NO_SHARD_AXES,),
                    kernel_axes=(W_FSDP_AXES, W_JOINED_AXES, W_TP_AXES),
1077
1078
1079
                    kernel_init=qkv_init,
                    use_bias=self.use_bias,
                    bias_init=self.bias_init,
1080
                    bias_axes=(W_JOINED_AXES, W_TP_AXES),
1081
1082
1083
                    enable_low_rank_adaptation=lora_scope.qkv_proj,
                    low_rank_adaptation_dim=self.low_rank_adaptation_dim,
                    low_rank_adaptation_alpha=self.low_rank_adaptation_alpha,
1084
1085
                    layernorm_input_axes=inputs_logical_axes_maybe_sp,
                    dot_input_axes=inputs_logical_axes_no_sp,
1086
1087
1088
1089
                    name="qkv",
                    dtype=self.dtype,
                )(inputs_q)
                qkv_proj = checkpoint_name(qkv_proj, "combined_qkv_proj")
1090
                qkv_layout = QKVLayout.BS3HD
1091
1092
            else:
                query, ln_out = LayerNormDenseGeneral(
1093
                    enable_layernorm=self.input_layernorm,
1094
                    layernorm_type=self.layernorm_type,
1095
                    zero_centered_gamma=self.zero_centered_gamma,
1096
1097
                    epsilon=self.layernorm_epsilon,
                    axis=-1,
1098
                    features=self.num_attention_heads * self.head_dim,
1099
                    transpose_batch_sequence=self.transpose_batch_sequence,
1100
                    return_layernorm_output=(self.return_layernorm_output or is_self_attn),
1101
1102
1103
                    scale_axes=(W_NO_SHARD_AXES,),
                    ln_bias_axes=(W_NO_SHARD_AXES,),
                    kernel_axes=(W_FSDP_AXES, W_TP_AXES),
1104
1105
                    use_bias=self.use_bias,
                    bias_init=self.bias_init,
1106
                    bias_axes=(W_TP_AXES,),
1107
1108
1109
                    enable_low_rank_adaptation=lora_scope.qkv_proj,
                    low_rank_adaptation_dim=self.low_rank_adaptation_dim,
                    low_rank_adaptation_alpha=self.low_rank_adaptation_alpha,
1110
1111
                    dtype=self.dtype,
                    kernel_init=query_init,
1112
1113
                    layernorm_input_axes=inputs_logical_axes_maybe_sp,
                    dot_input_axes=inputs_logical_axes_no_sp,
1114
1115
                    name="query",
                )(inputs_q)
zlsh80826's avatar
zlsh80826 committed
1116
1117
1118
1119
1120

                if is_self_attn:
                    assert ln_out is not None
                    inputs_kv = ln_out

1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
                kv_proj = DenseGeneral(
                    axis=-1,
                    features=(2, self.num_gqa_groups * self.head_dim),
                    transpose_batch_sequence=self.transpose_batch_sequence,
                    kernel_axes=(W_FSDP_AXES, W_JOINED_AXES, W_TP_AXES),
                    kernel_init=kv_init,
                    use_bias=self.use_bias,
                    bias_init=self.bias_init,
                    bias_axes=(W_JOINED_AXES, W_TP_AXES),
                    enable_low_rank_adaptation=lora_scope.qkv_proj,
                    low_rank_adaptation_dim=self.low_rank_adaptation_dim,
                    low_rank_adaptation_alpha=self.low_rank_adaptation_alpha,
                    name="kv",
                    dtype=self.dtype,
                )(inputs_kv)
                kv_proj = checkpoint_name(kv_proj, "combined_kv_proj")
1137
                qkv_layout = QKVLayout.BSHD_BS2HD
1138
1139
1140
1141
        else:
            kv_projection = functools.partial(
                DenseGeneral,
                axis=-1,
zlsh80826's avatar
zlsh80826 committed
1142
                features=self.num_gqa_groups * self.head_dim,
1143
                transpose_batch_sequence=self.transpose_batch_sequence,
1144
                kernel_axes=(W_FSDP_AXES, W_TP_AXES),
1145
1146
                use_bias=self.use_bias,
                bias_init=self.bias_init,
1147
                bias_axes=(W_TP_AXES,),
1148
1149
1150
                enable_low_rank_adaptation=lora_scope.qkv_proj,
                low_rank_adaptation_dim=self.low_rank_adaptation_dim,
                low_rank_adaptation_alpha=self.low_rank_adaptation_alpha,
1151
1152
                dtype=self.dtype,
            )
1153
            query, ln_out = LayerNormDenseGeneral(
1154
                enable_layernorm=self.input_layernorm,
1155
                layernorm_type=self.layernorm_type,
1156
                zero_centered_gamma=self.zero_centered_gamma,
1157
1158
                epsilon=self.layernorm_epsilon,
                axis=-1,
1159
                features=self.num_attention_heads * self.head_dim,
1160
1161
                transpose_batch_sequence=self.transpose_batch_sequence,
                return_layernorm_output=True,
1162
1163
1164
                scale_axes=(W_NO_SHARD_AXES,),
                ln_bias_axes=(W_NO_SHARD_AXES,),
                kernel_axes=(W_FSDP_AXES, W_TP_AXES),
1165
1166
                use_bias=self.use_bias,
                bias_init=self.bias_init,
1167
                bias_axes=(W_TP_AXES,),
1168
1169
1170
                enable_low_rank_adaptation=lora_scope.qkv_proj,
                low_rank_adaptation_dim=self.low_rank_adaptation_dim,
                low_rank_adaptation_alpha=self.low_rank_adaptation_alpha,
1171
1172
                dtype=self.dtype,
                kernel_init=query_init,
1173
1174
                layernorm_input_axes=inputs_logical_axes_maybe_sp,
                dot_input_axes=inputs_logical_axes_no_sp,
1175
1176
                name="query",
            )(inputs_q)
1177

1178
            if is_self_attn:
1179
1180
1181
                assert ln_out is not None
                inputs_kv = ln_out

1182
1183
1184
1185
1186
            key = kv_projection(kernel_init=self.kernel_init, name="key")(inputs_kv)
            value = kv_projection(kernel_init=self.kernel_init, name="value")(inputs_kv)
            query = checkpoint_name(query, "query_proj")
            key = checkpoint_name(key, "key_proj")
            value = checkpoint_name(value, "value_proj")
1187
            qkv_layout = QKVLayout.BSHD_BSHD_BSHD
1188

1189
        if self.enable_rotary_pos_emb:
1190
1191
1192
1193
1194
1195
            if qkv_layout == QKVLayout.BS3HD:
                query, key, value = jnp.split(qkv_proj, [1, 2], axis=-2)
            elif qkv_layout == QKVLayout.BSHD_BS2HD:
                key, value = jnp.split(kv_proj, [1], axis=-2)
            else:
                assert qkv_layout == QKVLayout.BSHD_BSHD_BSHD
1196

1197
            # No changes to memory layout, should trigger bitcast only (Ideally no Perf impact)
1198
1199
1200
            query = query.reshape((*query.shape[:2], self.num_attention_heads, self.head_dim))
            key = key.reshape((*key.shape[:2], self.num_gqa_groups, self.head_dim))

1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
            query = rotary_pos_emb(
                query,
                self.rotary_pos_emb_windows,
                self.transpose_batch_sequence,
                self.rotary_pos_emb_group_method,
            )
            key = rotary_pos_emb(
                key,
                self.rotary_pos_emb_windows,
                self.transpose_batch_sequence,
                self.rotary_pos_emb_group_method,
            )
1213
            qkv_layout = QKVLayout.BSHD_BSHD_BSHD
1214

1215
1216
        if qkv_layout == QKVLayout.BSHD_BSHD_BSHD:
            query = query.reshape((*query.shape[:2], self.num_attention_heads, self.head_dim))
zlsh80826's avatar
zlsh80826 committed
1217
1218
            key = key.reshape((*key.shape[:2], self.num_gqa_groups, self.head_dim))
            value = value.reshape((*value.shape[:2], self.num_gqa_groups, self.head_dim))
1219
1220

        if decode:
1221
            assert qkv_layout == QKVLayout.BSHD_BSHD_BSHD
1222
1223
1224
1225
1226
1227
1228
1229
1230
            is_initialized = self.has_variable("cache", "cached_key")

            cached_key = self.variable("cache", "cached_key", jnp.zeros, key.shape, key.dtype)
            cached_value = self.variable(
                "cache", "cached_value", jnp.zeros, value.shape, value.dtype
            )
            cache_index = self.variable(
                "cache", "cache_index", lambda: jnp.array(0, dtype=jnp.int32)
            )
1231
            if is_initialized:
Ming-Xu Huang's avatar
Ming-Xu Huang committed
1232
                if self.transpose_batch_sequence:
1233
1234
                    length, batch, num_attention_heads, head_dim = cached_key.value.shape
                    expected_shape = (1, batch, num_attention_heads, head_dim)
Ming-Xu Huang's avatar
Ming-Xu Huang committed
1235
1236
                    one_hot_indices_shape = (length, 1, 1, 1)
                else:
1237
1238
                    batch, length, num_attention_heads, head_dim = cached_key.value.shape
                    expected_shape = (batch, 1, num_attention_heads, head_dim)
Ming-Xu Huang's avatar
Ming-Xu Huang committed
1239
                    one_hot_indices_shape = (1, length, 1, 1)
1240
1241
1242
1243

                # Sanity shape check of cached key against input query.
                if expected_shape != query.shape:
                    raise ValueError(
1244
1245
1246
                        "Autoregressive cache shape error, "
                        f"expected query shape {expected_shape} instead got {query.shape}."
                    )
1247
1248
1249

                cur_index = cache_index.value
                one_hot_indices = jax_nn.one_hot(cur_index, length, dtype=key.dtype)
Ming-Xu Huang's avatar
Ming-Xu Huang committed
1250
1251
1252
                one_hot_indices = jnp.reshape(one_hot_indices, one_hot_indices_shape)
                key = cached_key.value + key * one_hot_indices
                value = cached_value.value + value * one_hot_indices
1253
1254
1255
1256
1257
                cached_key.value = key
                cached_value.value = value
                cache_index.value = cache_index.value + 1

                mask = combine_masks(
1258
1259
                    mask, jnp.broadcast_to(jnp.arange(length) > cur_index, (batch, 1, 1, length))
                )
1260
1261

                if bias is not None:
1262
1263
1264
1265
1266
1267
                    dynamic_vector_slice_in_dim = vmap(
                        lax.dynamic_slice_in_dim, in_axes=(None, 0, None, None)
                    )
                    bias = dynamic_vector_slice_in_dim(
                        jnp.squeeze(bias, axis=0), jnp.reshape(cur_index, (-1)), 1, -2
                    )
1268

1269
1270
1271
1272
1273
        LEADING_AXES = (BATCH_AXES, SEQLEN_AXES)
        if self.transpose_batch_sequence:
            LEADING_AXES = (SEQLEN_AXES, BATCH_AXES)

        if qkv_layout == QKVLayout.BS3HD:
1274
1275
1276
            qkv_proj = qkv_proj.reshape(
                *qkv_proj.shape[:2], 3, self.num_attention_heads, self.head_dim
            )
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
            qkv_sharding_constraint = (*LEADING_AXES, JOINED_AXES, HEAD_AXES, HIDDEN_AXES)
            qkv_proj = with_sharding_constraint_by_logical_axes(qkv_proj, qkv_sharding_constraint)
            dpa_args = [qkv_proj, None, None]
        elif qkv_layout == QKVLayout.BSHD_BS2HD:
            query = query.reshape(*query.shape[:2], self.num_attention_heads, self.head_dim)
            kv_proj = kv_proj.reshape(*kv_proj.shape[:2], 2, self.num_gqa_groups, self.head_dim)
            q_sharding_constraint = (*LEADING_AXES, HEAD_AXES, HIDDEN_AXES)
            kv_sharding_constraint = (*LEADING_AXES, JOINED_AXES, HEAD_AXES, HIDDEN_AXES)
            query = with_sharding_constraint_by_logical_axes(query, q_sharding_constraint)
            kv_proj = with_sharding_constraint_by_logical_axes(kv_proj, kv_sharding_constraint)
            dpa_args = [query, kv_proj, None]
1288
        else:
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
            assert qkv_layout == QKVLayout.BSHD_BSHD_BSHD
            query = query.reshape((*query.shape[:2], self.num_attention_heads, self.head_dim))
            key = key.reshape((*key.shape[:2], self.num_gqa_groups, self.head_dim))
            value = value.reshape((*value.shape[:2], self.num_gqa_groups, self.head_dim))
            qkv_sharding_constraint = (*LEADING_AXES, HEAD_AXES, HIDDEN_AXES)
            query = with_sharding_constraint_by_logical_axes(query, qkv_sharding_constraint)
            key = with_sharding_constraint_by_logical_axes(key, qkv_sharding_constraint)
            value = with_sharding_constraint_by_logical_axes(value, qkv_sharding_constraint)
            dpa_args = [query, key, value]

1299
        scale_factor = 1.0 / sqrt(self.head_dim) if self.scale_attn_logits else 1.0
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
        x = DotProductAttention(
            head_dim=self.head_dim,
            num_attention_heads=self.num_attention_heads,
            num_gqa_groups=self.num_gqa_groups,
            attn_mask_type=self.attn_mask_type,
            attn_bias_type=self.attn_bias_type,
            attention_dropout=self.attention_dropout,
            dtype=self.dtype,
            dropout_rng_name=self.dropout_rng_name,
            float32_logits=self.float32_logits,
            qkv_layout=qkv_layout.name,
            scale_factor=scale_factor,
            transpose_batch_sequence=self.transpose_batch_sequence,
1313
            window_size=self.window_size,
1314
        )(*dpa_args, mask, bias, deterministic=deterministic)
1315
1316
        x = x.reshape((x.shape[0], x.shape[1], x.shape[2] * x.shape[3]))

1317
        attn_context_sharding_constraint = (*LEADING_AXES, HIDDEN_TP_AXES)
1318
        x = with_sharding_constraint_by_logical_axes(x, attn_context_sharding_constraint)
1319

1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
        out = DenseGeneral(
            features=inputs_q.shape[-1],
            transpose_batch_sequence=self.transpose_batch_sequence,
            axis=-1,
            kernel_init=self.kernel_init,
            kernel_axes=(W_TP_AXES, W_FSDP_AXES),
            use_bias=self.use_bias,
            bias_init=self.bias_init,
            bias_axes=(W_NO_SHARD_AXES,),
            enable_low_rank_adaptation=lora_scope.output_proj,
            low_rank_adaptation_dim=self.low_rank_adaptation_dim,
            low_rank_adaptation_alpha=self.low_rank_adaptation_alpha,
            dtype=self.dtype,
            name="out",
        )(x)
        out = checkpoint_name(out, "out_proj")
1336
1337

        return out, ln_out
1338
1339


1340
class RelativePositionBiases(nn.Module):  # pylint: disable=too-few-public-methods
1341
1342
1343
1344
1345
    """
    T5-style relative positional embeddings to the attention logits.

    Parameters
    ----------
1346
    num_buckets: int
1347
        The number of buckets to bucket distances between key and query positions into.
1348
    max_distance: int
1349
        The maximum distance before everything is lumped into the last
1350
        distance bucket.
1351
    num_attention_heads: int
1352
        Number of attention heads in the transformer layer.
1353
    embedding_init: Initializer, default = flax.linen.linear.default_embed_init
1354
        Used for initializing relative embedding tables.
1355
    embedding_axes: Tuple[str, ...], default = ('heads', 'relpos_buckets')
1356
        The name of axes used to shard embedding attention bias with a corresponding mesh.
1357
1358
1359

    Optimization parameters
    -----------------------
1360
    dtype: jax.numpy.dtype, default  = jax.numpy.float32
1361
        The data type used to allocate the initial parameters.
1362
    """
1363

1364
1365
1366
1367
    num_buckets: int
    max_distance: int
    num_attention_heads: int
    embedding_init: Callable[..., Array] = nn.linear.default_embed_init
1368
    embedding_axes: Tuple[str, ...] = ("heads", "relpos_buckets")
1369
1370
1371
1372
1373
1374
1375
1376
1377
    dtype: DType = jnp.float32

    @nn.compact
    def __call__(self, q_seqlen, k_seqlen, bidirectional=True):
        """
        Generate relative position embedding attention biases.

        Parameters
        ----------
1378
        q_seqlen: int
1379
            The sequence length of query.
1380
        k_seqlen: int
1381
            The sequence length of key.
1382
        bidirectional: bool, default = True
1383
            Indicate whether to allow positive memory-query relative position
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
            embeddings.

        Returns
        -------
        output: jax.numpy.ndarray
            An attention bias with shape `(1, num_attention_heads, q_seqlen, k_seqlen)`.
        """
        context_position = np.arange(q_seqlen, dtype=jnp.int32)[:, None]
        memory_position = np.arange(k_seqlen, dtype=jnp.int32)[None, :]
        relative_position = memory_position - context_position

        # Compute relative position bucket
        rp_bucket = 0
        negative_rp = -relative_position
        rpb_num_buckets = self.num_buckets

        if bidirectional:
            rpb_num_buckets //= 2
            rp_bucket += (negative_rp < 0).astype(np.int32) * rpb_num_buckets
            negative_rp = np.abs(negative_rp)
        else:
            negative_rp = np.maximum(negative_rp, 0)

        rpb_max_exact = rpb_num_buckets // 2
        rpb_is_small = negative_rp < rpb_max_exact
        rpb_val_if_large = rpb_max_exact + (
1410
1411
1412
1413
            np.log(negative_rp.astype(np.float32) / rpb_max_exact + np.finfo(np.float32).eps)
            / np.log(self.max_distance / rpb_max_exact)
            * (rpb_num_buckets - rpb_max_exact)
        ).astype(np.int32)
1414
1415
1416
1417
1418
        rpb_val_if_large = np.minimum(rpb_val_if_large, rpb_num_buckets - 1)
        rp_bucket += np.where(rpb_is_small, negative_rp, rpb_val_if_large)

        # Compute relative attention bias
        relative_attention_bias = nn_partitioning.param_with_axes(
1419
1420
1421
            "rel_embedding",
            self.embedding_init,
            (self.num_attention_heads, self.num_buckets),
1422
            jnp.float32,
1423
1424
            axes=self.embedding_axes,
        )
1425
1426
1427
1428
1429
1430

        relative_attention_bias = jnp.asarray(relative_attention_bias, self.dtype)

        bcast_iota = lax.broadcasted_iota(jnp.int32, (self.num_buckets, 1, 1), 0)
        rp_bucket_one_hot = jnp.array(rp_bucket[jnp.newaxis, ...] == bcast_iota, dtype=self.dtype)

1431
1432
1433
        values = lax.dot_general(
            relative_attention_bias, rp_bucket_one_hot, (((1,), (0,)), ((), ()))
        )
1434
1435
1436
1437
        return values[jnp.newaxis, ...]


class TransformerLayerType(Enum):
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
    r"""
    TransformerLayerType is an Enum class to specify a type of TransformerLayer

    Values
    ----------
    ENCODER:
        Encoder type of TransformerLayer.
    DECODER:
        Decoder type of TransformerLayer.
    """
1448

1449
1450
1451
1452
    ENCODER = "encoder"
    DECODER = "decoder"


1453
class TransformerLayer(nn.Module):  # pylint: disable=too-few-public-methods
1454
1455
1456
1457
1458
1459
1460
1461
    r"""
    TransformerLayer is made up of a relative embedding,
    an attention block and a feedforward network (MLP).
    This standard layer is based on the paper “Attention Is All You Need”.

    Parameters
    ----------
    hidden_size: int, default = 512
1462
        The hidden size of each input sample.
1463
    mlp_hidden_size: int, default = 2048
1464
        Intermediate size to which input samples are projected.
1465
    num_attention_heads: int, default = 8
1466
        Number of attention heads in the transformer layer.
1467
    num_gqa_groups: int, default = `None`
zlsh80826's avatar
zlsh80826 committed
1468
1469
1470
1471
1472
1473
1474
        Number of GQA groups. When `None` is present, it is equal to num_attention_heads.
        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`.
1475
    layernorm_type: {'layernorm', 'rmsnorm'}, default = 'layernorm'
1476
        Indicate the type of layer normalization.
1477
    layernorm_epsilon: float, default = 1e-6
1478
        A value added to the denominator of layer normalization for numerical stability.
1479
    zero_centered_gamma: bool, default = False
1480
1481
1482
1483
1484
1485
1486
        If set to `True`, the LayerNorm formula changes to

        .. math::
            y = \frac{x - \mathrm{E}[x]}{ \sqrt{\mathrm{Var}[x] + \epsilon}} *
            (1 + \gamma) + \beta

        This parameter is only applicable for 'layernorm'.
1487
    hidden_dropout: float, default = 0.1
1488
        Dropout probability for the dropout op after FC2 layer.
1489
    hidden_dropout_dims: Sequence[int], default = ()
1490
        Dimensions that will share the same dropout mask for hidden
1491
    attention_dropout: float, default = 0.1
1492
        Dropout probability for the dropout op during multi-head attention.
1493
1494
1495
1496
    intermediate_dropout: float, default = 0.1
        Dropout probability for the dropout op after FC1 layer.
    intermediate_dropout_dims: Sequence[int], default = ()
        Dimensions that will share the same dropout mask for hidden after FC1 layer.
1497
    dropout_rng_name: str, default = 'dropout'
1498
1499
        The key in given RNGs via flax.linen.Module.apply that for
        generating Dropout masks in the Multi-Head Attention.
1500
1501
    mha_kernel_init: Initializer, default =
        flax.linen.initializers.variance_scaling(1.0, 'fan_in', 'normal')
1502
1503
        Used for initializing weights of QKV and Output projection weights.
        It should be a callable object with three arguments (jax.random.PRNGKey, shape, dtype).
1504
1505
    mlp_kernel_init: Initializer, default =
        flax.linen.initializers.variance_scaling(1.0, 'fan_in', 'truncated_normal')
1506
1507
        Used for initializing weights of FC1 and FC2 layers.
        It should be a callable object with three arguments (jax.random.PRNGKey, shape, dtype).
1508
    mlp_activations: Sequence[str], default = ('relu', )
1509
        The sequence of activation functions to apply after the first linear transformation.
1510
1511
        Each activation has its own transformation layer.
    use_bias: bool, default = False
1512
1513
        Indicate whether to enable bias shifting for QKVO projections, FC1 and FC2.
        If set to False, the layer will not learn additive biases.
1514
    bias_init: Initializer, default = flax.linen.initializers.zeros
1515
1516
1517
        Used for initializing bias of QKVO projections,
        FC1 and FC2. It is only used when :attr:`use_bias=True`.
        It should be a callable object with three arguments (jax.random.PRNGKey, shape, dtype).
1518
    apply_residual_connection_post_layernorm: bool, default = False
1519
        If set to True, residual connections are taken from the output
1520
1521
        of layer norm (default is taken from input of layer norm)
    output_layernorm: bool, default = False
1522
        If set to True, layer normalization is applied on the output side,
1523
1524
1525
        after the final dropout-add. default behavior is to apply layer
        normalization on the input side, before the QKV transformation.
    float32_attention_logits: bool, default = False
1526
1527
        Whether to compute attention logits in float32 for the unfused attention backend.
        For fused attention backend, the accumulation is always float32 without the perf overhead.
1528
    layer_type: TransformerLayerType, default = TransformerLayerType.ENCODER
1529
        If set to TransformerLayerType.DECODER, an additional cross-attention block
1530
1531
        is added after self-attention.this can be used for structures like `T5`
        Transformer in conjunction with the TransformerLayerType.ENCODER option.
1532
    self_attn_mask_type: str, default = 'causal'
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
        This parameter specifies the type of attention mask to be applied during the softmax
        operation in the self attention.
        Available options are {'no_mask', 'padding', 'causal', 'causal_padding', 'padding_causal'}

        Each described below:

        * no_mask: No attention mask is applied. This means the self attention will consider the
          full sequence without any restrictions.
        * padding: Indicates the presence of padding at the end of each sequence.
          Users must provide a mask with the shape [batch, 1, max_seqlen_q, max_seqlen_kv] in the
          :attr:`__call__` method to specify the padding positions.
        * causal: An upper triangular mask is applied to the softmax inputs,
          ensuring that the prediction for a certain position is only dependent on known outputs
          from positions before it.
        * causal_padding / padding_causal: A combination of both causal and padding masks.
          Both 'causal_padding' and 'padding_causal' are acceptable and have the same effect.

        .. note:: :attr:`attention_mask` in :attr:`__call__` is ignored for 'no_mask' and 'causal'.

1552
1553
1554
1555
1556
    self_attn_bias_type: Optional[str], default = None
        Type of the attention bias passed into the self attention.
        Available options: {'no_bias', 'pre_scale_bias', 'post_scale_bias'}.
        When default is present, the type is automatically decided by the MHA's bias parameter.
        Where it is `post_scale_bias` if there is bias. Otherwise `no_bias` is used.
1557
    enable_relative_embedding: bool, default = True
1558
        Whether to enable relative embedding as shifting of attention logits.
1559
    relative_embedding: flax.linen.Module, default = None
1560
        The module for relative embedding execution, only used when
1561
1562
1563
1564
1565
1566
        :attr:`enable_relative_embedding=True`. Default is None, which will create
        an instance of RelativePositionBiases if :attr:`enable_relative_embedding=True`.
        Default: RelativePositionBiases( num_buckets=32, max_distance=128,
        num_attention_heads=self.num_attention_heads, dtype=self.dtype,
        embedding_init=flax.linen.initializers.variance_scaling(1.0, 'fan_avg', 'uniform'),
        name='relpos_bias')
1567
1568
1569
1570
1571
    enable_rotary_pos_emb: bool, default = False
        Whether to enable rotary position embedding to projected query and key in MHA.
    rotary_pos_emb_windows: Tuple[int, int], default = (1, 10000)
        Indicate the min and max time-scales of rotary position embedding,
        only used when :attr:`enable_rotary_pos_emb=True`
1572
    rotary_pos_emb_group_method: str, default = 'consecutive'
1573
1574
1575
1576
        Indicate the method to couple the coordinates. It should be one of
        ['consecutive', 'alternate']. 'alternate' is to pair index :math:`i` with :math:`i + d/2`,
        where :math:`d` is the hidden dimension. 'consecutive' pairs index :math:`i` with
        :math:`i + 1`.
1577
1578
1579
    low_rank_adaptation_scope: str, default = 'none'
        Indicate the scope to apply low rank adaptation. It should be one of
        ['none', 'all', 'qkv_proj', 'output_proj', 'mlp', 'exclude_qkv_proj',
1580
        'exclude_output_proj', 'exclude_mlp']
1581
1582
1583
1584
1585
    low_rank_adaptation_dim: int, default = 32
        The dimension for low rank adaptation, only used when
        :attr:`enable_low_rank_adaptation=True`
    low_rank_adaptation_alpha: float, default = None
        The alpha for computing the scaling factor of LoRA output.
1586
        :math:`\frac{alpha}{rank} * lora\_output`. None means no scaling.
1587
1588
    enable_sequence_parallel: bool, default = False
        Whether to enable sequence parallelism to operations except dot.
1589
1590
    window_size: Optional[Tuple[int, int]], default = None
        Sliding window size. Default value is no sliding window.
1591
1592
1593

    Optimization parameters
    -----------------------
1594
    dtype: jax.numpy.dtype, default  = jax.numpy.float32
1595
        The data type used to allocate the initial parameters.
1596
    drop_path: float, default = 0.0
1597
        When > 0.0, applies stochastic depth per sample in the main
1598
1599
        path of the residual block.
    fuse_qkv_params: bool, default = True
1600
        If set to True, `TransformerLayer` module exposes a single fused
1601
1602
        parameter for query-key-value for self-attention and key-value for
        cross-attention.
1603
    transpose_batch_sequence: bool, default = False
1604
        Indicate whether the input tensors were switched axis of batch
1605
1606
1607
        and sequence length dimension. if set to True, the input tensors
        should be in (seqlen, batch, hidden), otherwise (batch, seqlen, hidden).
    scale_attn_logits: bool, default = False
1608
        Indicate whether to scale attention logits.
1609
1610
1611
        if set to True, :math:`\frac{Q}{\sqrt{head_dim}*K}`,
        else :math:`Q*K`
    scaled_query_init: bool, default = `True`
1612
        Whether to scale WQ on initialization by :math:`\sqrt{head_dim}`
1613
1614
1615
1616
1617
    """

    hidden_size: int = 512
    mlp_hidden_size: int = 2048
    num_attention_heads: int = 8
1618
    num_gqa_groups: Optional[int] = None
1619
    layernorm_type: str = "layernorm"
1620
    layernorm_epsilon: float = 1e-6
1621
    zero_centered_gamma: bool = False
1622
1623
1624
    hidden_dropout: float = 0.1
    hidden_dropout_dims: Sequence[int] = ()
    attention_dropout: float = 0.1
1625
1626
    intermediate_dropout: float = 0.1
    intermediate_dropout_dims: Sequence[int] = ()
1627
    dropout_rng_name: str = "dropout"
1628
1629
    mha_kernel_init: Initializer = None
    mlp_kernel_init: Initializer = None
1630
    mlp_activations: Sequence[str] = ("relu",)
1631
1632
1633
1634
1635
1636
    use_bias: bool = False
    bias_init: Initializer = nn.initializers.zeros
    apply_residual_connection_post_layernorm: bool = False
    output_layernorm: bool = False
    float32_attention_logits: bool = False
    layer_type: TransformerLayerType = TransformerLayerType.ENCODER
1637
    self_attn_mask_type: str = "causal"
1638
    self_attn_bias_type: Optional[str] = None
1639
1640
    enable_relative_embedding: bool = True
    relative_embedding: nn.Module = None
1641
1642
    enable_rotary_pos_emb: bool = False
    rotary_pos_emb_windows: Tuple[int, int] = (1, 10000)
1643
1644
    rotary_pos_emb_group_method: str = "consecutive"
    low_rank_adaptation_scope: str = "none"
1645
1646
    low_rank_adaptation_dim: int = 32
    low_rank_adaptation_alpha: float = None
1647
1648
1649
    dtype: DType = jnp.float32
    drop_path: float = 0.0
    fuse_qkv_params: bool = True
Jeng Bai-Cheng's avatar
Jeng Bai-Cheng committed
1650
    transpose_batch_sequence: bool = False
1651
    enable_sequence_parallel: bool = False
1652
1653
    scale_attn_logits: bool = False
    scaled_query_init: bool = True
1654
    window_size: Optional[Tuple[int, int]] = None
1655
1656
1657

    def __post_init__(self):
        if self.mha_kernel_init is None:
1658
            self.mha_kernel_init = nn.initializers.variance_scaling(1.0, "fan_in", "normal")
1659
        if self.mlp_kernel_init is None:
1660
1661
1662
            self.mlp_kernel_init = nn.initializers.variance_scaling(
                1.0, "fan_in", "truncated_normal"
            )
zlsh80826's avatar
zlsh80826 committed
1663
1664
        if self.num_gqa_groups is None:
            self.num_gqa_groups = self.num_attention_heads
1665
1666
1667
        super().__post_init__()

    @nn.compact
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
    def __call__(
        self,
        inputs: Array,
        encoded: Array = None,
        attention_mask: Array = None,
        encoder_decoder_mask: Array = None,
        deterministic: bool = False,
        decode: bool = False,
        max_decode_length: bool = None,
    ):
1678
1679
1680
1681
1682
        """
        Transformer Layer: attention block and a feedforward network (MLP)

        Parameters
        ----------
1683
        inputs: jax.numpy.ndarray
1684
            Input tensor.
1685
        encoded: jax.numpy.ndarray, default = None
1686
1687
1688
1689
            Output tensors of the encoder block to be fed into the decoder block if using
            :attr:`layer_type=TransformerLayerType.DECODER`.
        attention_mask : jax.numpy.ndarray, default = None
            Boolean tensor used to mask out self-attention softmax input.
1690
1691
            :attr:`True` means mask out the corresponding values.
            Ignored when :attr:`self.self_attn_mask_type` is either 'no_mask' or 'causal'.
1692
        encoder_decoder_mask: jax.numpy.ndarray, default = None
1693
1694
            Boolean tensor used to mask out cross-attention softmax input when
            :attr:`layer_type=TransformerLayerType.DECODER`.
1695
            :attr:`True` means mask out the corresponding values.
1696
        deterministic: bool, default = False
1697
            Disable dropout layers if set to True.
1698
        decode: bool, default = False
1699
1700
            Indicate whether to prepare and use an autoregressive cache
            in Multi-head attention (MHA).
1701
        max_decode_length: bool, default = None
1702
1703
1704
1705
1706
1707
            The maximum length to generate relative embedding biases when
            :attr:`layer_type=TransformerLayerType.DECODER` and
            :attr:`enable_relative_embedding=True`.

        Returns
        -------
1708
        outputs: jax.numpy.ndarray
1709
            Output tensors.
1710
        """
1711
1712
1713
        assert (
            self.layer_type in TransformerLayerType
        ), f"layer_type should be one of TransformerLayerType, but got {self.layer_type}."
1714

1715
1716
1717
1718
        assert self.hidden_size % self.num_attention_heads == 0, (
            "hidden_size should be multiples of num_attention_heads"
            f", but got {self.hidden_size=} and {self.num_attention_heads=}."
        )
1719

1720
1721
1722
        assert self.layer_type == TransformerLayerType.DECODER or (
            self.layer_type == TransformerLayerType.ENCODER and decode is False
        ), "decode should be False when layer_type == TransformerLayerType.ENCODER."
1723
1724
1725
1726
1727
1728

        head_dim = self.hidden_size // self.num_attention_heads

        sequence_dim = 0 if self.transpose_batch_sequence else 1
        batch_dim = 1 - sequence_dim

1729
1730
1731
        def generate_batch_seqlen_logical_axes(is_shared_seq=None):
            axes = [None, None]

1732
1733
1734
            is_shared_seq = (
                self.enable_sequence_parallel if is_shared_seq is None else is_shared_seq
            )
1735
1736
1737
1738
1739

            axes[batch_dim] = BATCH_AXES
            axes[sequence_dim] = SEQLEN_TP_AXES if is_shared_seq else SEQLEN_AXES
            return tuple(axes)

1740
1741
1742
        attn_bias = None
        if self.enable_relative_embedding:
            if self.relative_embedding is None:
1743
1744
1745
1746
1747
1748
1749
1750
                rel_emb = RelativePositionBiases(
                    num_buckets=32,
                    max_distance=128,
                    num_attention_heads=self.num_attention_heads,
                    dtype=self.dtype,
                    embedding_init=nn.initializers.variance_scaling(1.0, "fan_avg", "uniform"),
                    name="relpos_bias",
                )
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
            else:
                rel_emb = self.relative_embedding

            if self.layer_type == TransformerLayerType.ENCODER:
                attn_bias = rel_emb(inputs.shape[sequence_dim], inputs.shape[sequence_dim], True)
            else:
                if decode and max_decode_length:
                    l = max_decode_length
                else:
                    l = inputs.shape[sequence_dim]
                attn_bias = rel_emb(l, l, False)

        assert inputs.ndim == 3

        # Make name be the exactly same as T5X, since names would affect
        # RNGKey during init and apply. Myabe no need in the feature.
        if self.layer_type == TransformerLayerType.ENCODER:
1768
            mha_name = "attention"
1769
        else:
1770
            mha_name = "self_attention"
1771

1772
        inputs = with_sharding_constraint_by_logical_axes(
1773
1774
            inputs, (*generate_batch_seqlen_logical_axes(), HIDDEN_AXES)
        )
1775

1776
        # [batch, length, emb_dim] -> [batch, length, emb_dim]
1777
1778
1779
        residual = inputs
        x, ln_out = MultiHeadAttention(
            num_attention_heads=self.num_attention_heads,
1780
1781
            dtype=self.dtype,
            head_dim=head_dim,
zlsh80826's avatar
zlsh80826 committed
1782
            num_gqa_groups=self.num_gqa_groups,
1783
            transpose_batch_sequence=self.transpose_batch_sequence,
1784
            enable_sequence_parallel=self.enable_sequence_parallel,
1785
            attention_dropout=self.attention_dropout,
1786
1787
1788
1789
1790
1791
            dropout_rng_name=self.dropout_rng_name,
            float32_logits=self.float32_attention_logits,
            scale_attn_logits=self.scale_attn_logits,
            scaled_query_init=self.scaled_query_init,
            layernorm_type=self.layernorm_type,
            layernorm_epsilon=self.layernorm_epsilon,
1792
            zero_centered_gamma=self.zero_centered_gamma,
1793
1794
            return_layernorm_output=self.apply_residual_connection_post_layernorm,
            input_layernorm=not self.output_layernorm,
1795
            attn_mask_type=self.self_attn_mask_type,
1796
            attn_bias_type=self.self_attn_bias_type,
1797
1798
            enable_rotary_pos_emb=self.enable_rotary_pos_emb,
            rotary_pos_emb_windows=self.rotary_pos_emb_windows,
1799
            rotary_pos_emb_group_method=self.rotary_pos_emb_group_method,
1800
1801
1802
            low_rank_adaptation_scope=self.low_rank_adaptation_scope,
            low_rank_adaptation_dim=self.low_rank_adaptation_dim,
            low_rank_adaptation_alpha=self.low_rank_adaptation_alpha,
1803
            fuse_qkv_params=self.fuse_qkv_params,
1804
1805
1806
            kernel_init=self.mha_kernel_init,
            use_bias=self.use_bias,
            bias_init=self.bias_init,
1807
            name=mha_name,
1808
            window_size=self.window_size,
1809
        )(inputs, inputs, attention_mask, attn_bias, deterministic=deterministic, decode=decode)
1810
1811
1812
1813
1814

        def hidden_dropout(x, deterministic):
            assert isinstance(self.hidden_dropout_dims, Sequence)
            x_shape_len = len(x.shape)
            for dims in self.hidden_dropout_dims:
Ming-Xu Huang's avatar
Ming-Xu Huang committed
1815
                assert -x_shape_len <= dims < x_shape_len
1816

1817
1818
1819
1820
1821
            return nn.Dropout(
                rate=self.hidden_dropout,
                broadcast_dims=self.hidden_dropout_dims,
                rng_collection=self.dropout_rng_name,
            )(x, deterministic=deterministic)
1822

1823
        x = with_sharding_constraint_by_logical_axes(
1824
1825
            x, (*generate_batch_seqlen_logical_axes(), HIDDEN_AXES)
        )
1826
        residual = with_sharding_constraint_by_logical_axes(
1827
1828
            residual, (*generate_batch_seqlen_logical_axes(), HIDDEN_AXES)
        )
1829

1830
1831
1832
        x = hidden_dropout(x, deterministic)
        if self.drop_path > 0.0:
            drop_path_shape = _generate_drop_path_shape(x.shape, batch_dim)
1833
1834
1835
1836
1837
            x = nn.Dropout(
                rate=self.drop_path,
                broadcast_dims=drop_path_shape,
                rng_collection=self.dropout_rng_name,
            )(x, deterministic=deterministic)
1838
1839
1840
1841
1842

        if self.apply_residual_connection_post_layernorm:
            assert ln_out is not None
            residual = ln_out

1843
1844
1845
1846
        x = x + residual

        mlp_input = x
        if self.layer_type == TransformerLayerType.DECODER:
1847
1848
1849
            assert (
                encoded is not None
            ), "encoded is required when layer_type == TransformerLayerType.DECODER."
1850

1851
            x = with_sharding_constraint_by_logical_axes(
1852
1853
                x, (*generate_batch_seqlen_logical_axes(), HIDDEN_AXES)
            )
1854

1855
1856
1857
            residual = x
            y, ln_out = MultiHeadAttention(
                num_attention_heads=self.num_attention_heads,
1858
1859
                dtype=self.dtype,
                head_dim=head_dim,
zlsh80826's avatar
zlsh80826 committed
1860
                num_gqa_groups=self.num_gqa_groups,
1861
                transpose_batch_sequence=self.transpose_batch_sequence,
1862
                enable_sequence_parallel=self.enable_sequence_parallel,
1863
                attention_dropout=self.attention_dropout,
1864
1865
1866
                dropout_rng_name=self.dropout_rng_name,
                layernorm_type=self.layernorm_type,
                layernorm_epsilon=self.layernorm_epsilon,
1867
                zero_centered_gamma=self.zero_centered_gamma,
1868
                return_layernorm_output=self.apply_residual_connection_post_layernorm,
1869
1870
1871
                input_layernorm=True,  # Must do LayerNorm before MHA.
                attn_mask_type="padding",
                attn_bias_type="no_bias",
1872
1873
                enable_rotary_pos_emb=self.enable_rotary_pos_emb,
                rotary_pos_emb_windows=self.rotary_pos_emb_windows,
1874
                rotary_pos_emb_group_method=self.rotary_pos_emb_group_method,
1875
1876
1877
                low_rank_adaptation_scope=self.low_rank_adaptation_scope,
                low_rank_adaptation_dim=self.low_rank_adaptation_dim,
                low_rank_adaptation_alpha=self.low_rank_adaptation_alpha,
1878
1879
1880
                float32_logits=self.float32_attention_logits,
                scale_attn_logits=self.scale_attn_logits,
                scaled_query_init=self.scaled_query_init,
1881
                fuse_qkv_params=self.fuse_qkv_params,
1882
1883
1884
                kernel_init=self.mha_kernel_init,
                use_bias=self.use_bias,
                bias_init=self.bias_init,
1885
                name="encoder_decoder_attention",
1886
                window_size=self.window_size,
1887
            )(x, encoded, encoder_decoder_mask, deterministic=deterministic)
1888
1889

            y = with_sharding_constraint_by_logical_axes(
1890
1891
                y, (*generate_batch_seqlen_logical_axes(), HIDDEN_AXES)
            )
1892
            residual = with_sharding_constraint_by_logical_axes(
1893
1894
                residual, (*generate_batch_seqlen_logical_axes(), HIDDEN_AXES)
            )
1895

1896
            y = hidden_dropout(y, deterministic)
1897
1898
1899
1900
1901

            if self.apply_residual_connection_post_layernorm:
                assert ln_out is not None
                residual = ln_out

1902
1903
            mlp_input = y + residual

1904
        mlp_input = with_sharding_constraint_by_logical_axes(
1905
1906
            mlp_input, (*generate_batch_seqlen_logical_axes(), HIDDEN_AXES)
        )
1907

1908
1909
        lora_scope = _canonicalize_lora_scope(self.low_rank_adaptation_scope)

1910
1911
1912
1913
        # MlpBlock
        residual = mlp_input
        z, ln_out = LayerNormMLP(
            layernorm_type=self.layernorm_type,
1914
            zero_centered_gamma=self.zero_centered_gamma,
1915
1916
1917
1918
1919
            epsilon=self.layernorm_epsilon,
            transpose_batch_sequence=self.transpose_batch_sequence,
            return_layernorm_output=self.apply_residual_connection_post_layernorm,
            intermediate_dim=self.mlp_hidden_size,
            activations=self.mlp_activations,
1920
1921
1922
            intermediate_dropout_rng_name=self.dropout_rng_name,
            intermediate_dropout_rate=self.intermediate_dropout,
            intermediate_hidden_dropout_dims=self.intermediate_dropout_dims,
1923
            dtype=self.dtype,
1924
1925
            scale_axes=(W_NO_SHARD_AXES,),
            ln_bias_axes=(W_NO_SHARD_AXES,),
1926
            kernel_init=self.mlp_kernel_init,
1927
1928
            kernel_axes_1=(W_FSDP_AXES, W_JOINED_AXES, W_TP_AXES),
            kernel_axes_2=(W_TP_AXES, W_FSDP_AXES),
1929
1930
            use_bias=self.use_bias,
            bias_init=self.bias_init,
1931
1932
            bias_axes_1=(W_JOINED_AXES, W_TP_AXES),
            bias_axes_2=(W_NO_SHARD_AXES,),
1933
1934
1935
            enable_low_rank_adaptation=lora_scope.mlp,
            low_rank_adaptation_dim=self.low_rank_adaptation_dim,
            low_rank_adaptation_alpha=self.low_rank_adaptation_alpha,
1936
1937
1938
            layernorm_input_axes=(*generate_batch_seqlen_logical_axes(), HIDDEN_AXES),
            dot_1_input_axes=(*generate_batch_seqlen_logical_axes(False), HIDDEN_AXES),
            dot_2_input_axes=(*generate_batch_seqlen_logical_axes(False), HIDDEN_TP_AXES),
1939
            name="mlp",
1940
1941
1942
1943
1944
1945
        )(mlp_input, deterministic=deterministic)

        if self.apply_residual_connection_post_layernorm:
            assert ln_out is not None
            residual = ln_out

1946
        z = with_sharding_constraint_by_logical_axes(
1947
1948
            z, (*generate_batch_seqlen_logical_axes(), HIDDEN_AXES)
        )
1949
        residual = with_sharding_constraint_by_logical_axes(
1950
1951
            residual, (*generate_batch_seqlen_logical_axes(), HIDDEN_AXES)
        )
1952

1953
1954
1955
        z = hidden_dropout(z, deterministic)
        if self.drop_path > 0.0:
            drop_path_shape = _generate_drop_path_shape(z.shape, batch_dim)
1956
1957
1958
            z = nn.Dropout(rate=self.drop_path, broadcast_dims=drop_path_shape)(
                z, deterministic=deterministic
            )
1959
1960
1961
        z = z + residual

        if self.output_layernorm:
1962
            z = with_sharding_constraint_by_logical_axes(
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
                z, (*generate_batch_seqlen_logical_axes(), HIDDEN_AXES)
            )
            z = LayerNorm(
                layernorm_type=self.layernorm_type,
                zero_centered_gamma=self.zero_centered_gamma,
                epsilon=self.layernorm_epsilon,
                scale_axes=(W_NO_SHARD_AXES,),
                bias_axes=(W_NO_SHARD_AXES,),
                transpose_batch_sequence=self.transpose_batch_sequence,
                dtype=self.dtype,
                name="output_layernorm",
            )(z)
1975
1976

        return z