transformer.py 53.6 KB
Newer Older
1
2
3
4
5
6
7
8
# Copyright (c) 2022-2023, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
#
# 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 jax import dtypes
20
21
22
23
24
25
from jax import nn as jax_nn
from jax import random as jax_random
from jax import lax, vmap

from .module import DenseGeneral, LayerNormDenseGeneral, LayerNormMLP
from .module import LayerNorm, Softmax
26
from ..fused_attn import AttnBiasType, AttnMaskType, QKVLayout
27
28
from ..fused_attn import is_fused_attn_kernel_available
from ..fused_attn import self_fused_attn, cross_fused_attn
29
30
from ..softmax import SoftmaxType
from ..sharding import infer_major_sharding_type, infer_sharding_type
31
32
from ..sharding import global_shard_resource, with_sharding_constraint
from ..sharding import ShardingType
33
34
35
36
37
38
39
40
41
42

PRNGKey = Any
Shape = Tuple[int, ...]
DType = jnp.dtype
Array = jnp.ndarray
PrecisionLike = Union[None, str, lax.Precision, Tuple[str, str], Tuple[lax.Precision,
                                                                       lax.Precision]]
Initializer = Callable[[PRNGKey, Shape, DType], Array]
LogicalRules = Sequence[Tuple[str, Union[str, None]]]

43
44
45
46
47
48
49
50
51
52
53
BATCH_AXES = 'nvte_batch'
SEQLEN_AXES = 'nvte_seqlen'
HEAD_AXES = 'nvte_head'
HIDDEN_AXES = 'nvte_hidden'
HIDDEN_TP_AXES = 'nvte_hidden_tp'
JOINED_AXES = 'nvte_joined'
W_NO_SHARD_AXES = 'nvte_w_no_shard'
W_FSDP_AXES = 'nvte_w_fsdp'
W_TP_AXES = 'nvte_w_tp'
W_JOINED_AXES = 'nvte_w_joined'

54
55
56
57
58
59
60
61
62
63

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:
    """
64
    Extend the given Flax logical axis rules with the predefined TransformerLayer's
65
66
67
    logical axis rules.

    .. note::
Ming-Xu Huang's avatar
Ming-Xu Huang committed
68
69
70
71
        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.
72
73
74
75

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

Ming-Xu Huang's avatar
Ming-Xu Huang committed
76
77
78
79
    .. note::
        This function is only needed when using TransformerLayer. For  other modules, such as
        DenseGeneral, please properly set axes of kernels and bias.

80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
    Parameters
    ----------
    rules : Sequence[Tuple[str, Union[str, None]]]
        the base Flax logical axis rules to extend.

    Returns
    -------
    extended_rules : Sequence[Tuple[str, Union[str, None]]]
        the extended Flax logical axis rules.
    """
    rules_map = {}
    for item in rules:
        assert len(item) == 2, \
            "The logical axis rule should be like (axis_name, mesh_axis_name)."
        key = item[0]
        val = item[1]
        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
100
101
102
103
        if key in rules_map:
            rules_map[key].append(val)
        else:
            rules_map[key] = [val]
104
105

    gsr = global_shard_resource()
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131

    batch_dim_rule = []
    if gsr.dp_resource is not None:
        batch_dim_rule.append(gsr.dp_resource)
    if gsr.fsdp_resource is not None and gsr.dp_resource != gsr.fsdp_resource:
        batch_dim_rule.append(gsr.fsdp_resource)

    if len(batch_dim_rule) <= 0:
        batch_dim_rule = None
    elif len(batch_dim_rule) == 1:
        batch_dim_rule = batch_dim_rule[0]
    else:
        batch_dim_rule = tuple(batch_dim_rule)

    te_logical_axis_rules = (
        (BATCH_AXES, batch_dim_rule),
        (SEQLEN_AXES, None),
        (HEAD_AXES, gsr.tp_resource),
        (HIDDEN_AXES, None),
        (HIDDEN_TP_AXES, gsr.tp_resource),
        (JOINED_AXES, None),
        (W_NO_SHARD_AXES, None),
        (W_FSDP_AXES, gsr.fsdp_resource),
        (W_TP_AXES, gsr.tp_resource),
        (W_JOINED_AXES, None),
    )
132
133
134
135
136
137

    extended_rules = [*rules]
    for item in te_logical_axis_rules:
        key = item[0]
        val = item[1]
        if key in rules_map:
Ming-Xu Huang's avatar
Ming-Xu Huang committed
138
            assert len(rules_map[key]) == 1 and rules_map[key][0] == val, \
139
140
141
142
143
144
145
146
                f"The rule diverged between TE and given rule." \
                f"Axis:{key} map to {rules_map[key]} in the given" \
                f" rules, but {val} in TE's rules."
        else:
            extended_rules.append(item)
    return tuple(extended_rules)


147
148
149
150
151
152
153
154
155
156
157
158
def _with_sharding_constraint(x: Array, logical_axis_names: Shape):
    assert len(x.shape) == len(logical_axis_names)
    rules = extend_logical_axis_rules(tuple())
    rules_dict = {}
    for key, value in rules:
        rules_dict[key] = value

    mesh_axis_names = [rules_dict[name] for name in logical_axis_names]
    pspec = jax.sharding.PartitionSpec(*mesh_axis_names)
    return with_sharding_constraint(x, pspec)


159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
def _merge_mask(func, *masks: Optional[Array]):
    masks = [m for m in masks if m is not None]
    if not masks:
        return None
    assert all(map(lambda x: x.ndim == masks[0].ndim,
                   masks)), (f'masks must have same rank: {tuple(map(lambda x: x.ndim, masks))}')
    mask, *other_masks = masks
    for other_mask in other_masks:
        mask = func(mask, other_mask)
    return mask


def combine_masks(*masks: Optional[Array], dtype: DType = jnp.float32):
    """Combine attention masks."""
    func = jnp.logical_and
    return _merge_mask(func, *masks).astype(dtype)


def combine_biases(*masks: Optional[Array]):
    """Combine attention biases."""
    func = lambda a, b: a + b
    return _merge_mask(func, *masks)


def core_attention(query: Array,
                   key: Array,
                   value: Array,
186
                   scale_factor: float,
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
                   transpose_batch_sequence: bool,
                   softmax_type: SoftmaxType = SoftmaxType.SCALED,
                   softmax_sharding_type: ShardingType = ShardingType.SINGLE,
                   mask: Optional[Array] = None,
                   bias: Optional[Array] = None,
                   dropout_rng: Optional[PRNGKey] = None,
                   dropout_rate: float = 0.,
                   deterministic: bool = False,
                   dtype: DType = jnp.float32,
                   float32_logits: bool = False):
    """Core attention"""
    assert key.ndim == query.ndim == value.ndim, 'q, k, v must have same rank.'
    batch_dim = 1 if transpose_batch_sequence else 0
    assert query.shape[batch_dim] == key.shape[batch_dim] == value.shape[batch_dim], (
        'q, k, v batch dims must match.')
    assert query.shape[-2] == key.shape[-2] == value.shape[-2], ('q, k, v num_heads must match.')
    sequence_dim = 0 if transpose_batch_sequence else 1
    assert key.shape[sequence_dim] == value.shape[sequence_dim], 'k, v lengths must match.'
    assert query.shape[-1] == key.shape[-1], 'q, k depths must match.'

    if float32_logits:
        query = query.astype(jnp.float32)
        key = key.astype(jnp.float32)

    if transpose_batch_sequence:
        attn_weights = jnp.einsum('qbhd,kbhd->bhqk', query, key)
    else:
        attn_weights = jnp.einsum('bqhd,bkhd->bhqk', query, key)

216
217
218
    attn_weights = _with_sharding_constraint(attn_weights,
                                             (BATCH_AXES, HEAD_AXES, SEQLEN_AXES, SEQLEN_AXES))

219
220
221
222
223
224
225
226
227
    # When a bias is present, the computation is performed as Softmax(attn_weights * scale + bias).
    # In this case, the scale can not fused into the Softmax module.
    if bias is not None:
        attn_weights = attn_weights * scale_factor
        fused_scale_factor = 1.
    else:
        # If no bias, the scale can be fused into Softmax module
        fused_scale_factor = scale_factor

228
    attn_weights = Softmax(softmax_type=softmax_type,
229
                           scale_factor=fused_scale_factor,
230
231
                           sharding_type=softmax_sharding_type)(attn_weights, mask,
                                                                bias).astype(dtype)
232
233
234
235

    if not deterministic and dropout_rate > 0.:
        keep_prob = 1.0 - dropout_rate
        dropout_shape = list(attn_weights.shape)
236
        # TODO(rewang): add attention dropout broadcast dimension arguments for users
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
        keep = jax_random.bernoulli(dropout_rng, keep_prob, dropout_shape)
        multiplier = (keep.astype(attn_weights.dtype) / jnp.asarray(keep_prob, dtype=dtype))
        attn_weights = attn_weights * multiplier

    if transpose_batch_sequence:
        return jnp.einsum('bhqk,kbhd->qbhd', attn_weights, value)

    return jnp.einsum('bhqk,bkhd->bqhd', attn_weights, value)


dynamic_vector_slice_in_dim = vmap(lax.dynamic_slice_in_dim, in_axes=(None, 0, None, None))


class MultiHeadAttention(nn.Module):
    r"""
    Multi-head Attention (MHA), including Query,
    Key, Value and Output projection.

255
    .. note::
256

257
258
        Argument :attr:`mask` will be ignored when
        :attr:`attn_mask_type` is set to `"causal"`.
259

260
261
262
    Parameters
    ----------
    head_dim : int
263
        The hidden dimension of each attention head.
264
    num_heads : int
265
        The number of attention heads
266
    dropout_rate : float, default = 0.0
267
        Dropout probability for the dropout op during multi-head attention.
268
    dropout_rng_name: str, default = 'dropout'
269
270
        The key in given RNGs via flax.linen.Module.apply that is used
        to generate Dropout masks in the core attention.
271
    layernorm_type : {'layernorm', 'rmsnorm'}, default = 'layernorm'
272
        Indicate the type of layer normalization.
273
    layernorm_epsilon: float, default = 1e-6
274
        A value added to the denominator of layer normalization for numerical stability.
275
276
277
278
279
280
281
282
    zero_centered_gamma : bool, default = False
        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'.
283
284
    kernel_init: Initializer, default =
        flax.linen.initializers.variance_scaling(1.0, 'fan_in', 'normal')
285
286
        Used for initializing the QKV and Output projection weights.
        It should be a callable object with three arguments (jax.random.PRNGKey, shape, dtype).
287
    use_bias: bool, default = False
288
289
        Indicate whether or not to enable bias shifting for QKVO projections.
        If set to False, the layer will not learn additive biases.
290
    bias_init: Initializer, default = flax.linen.initializers.zeros
291
292
        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).
293
    apply_residual_connection_post_layernorm : bool, default = False
294
        Indicate if apply residual connection with the output of layer normalization.
295
    output_layernorm : bool, default = False
296
        Indicate if apply a layer normalization at the end of MHA.
297
298
    attn_mask_type: {'causal', 'padding'}, default = 'causal'
        Type of attention mask passed into softmax operation.
299
        Introduced in v0.10.0.
300
301
302
303

    Optimization parameters
    -----------------------
    dtype :jax.numpy.dtype, default  = jax.numpy.float32
304
        The data type used to allocate the initial parameters.
305
    fuse_qkv: bool, default = True
306
        If set to True, this module exposes a single fused
307
308
309
        parameter for query-key-value for self-attention and key-value for
        cross-attention.
    transpose_batch_sequence : bool, default = True
310
        Indicate whether the input tensors were switched axis of batch
311
312
313
        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
314
315
        Indicate whether to scale attention logits.
        If set to True, :math:`\frac{Q}{\sqrt{head_dim}*K}`,
316
317
        else :math:`Q*K`
    scaled_query_init: bool, default = `True`
318
        Whether to scale WQ on initialization by :math:`\sqrt{head_dim}`
319
    float32_logits : bool, default = False
320
        Whether to compute attention logits in float32.
321
322
323
324
325
326
327
328
    """

    head_dim: int
    num_heads: int
    dropout_rate: float = 0.
    dropout_rng_name: str = 'dropout'
    layernorm_type: str = "layernorm"
    layernorm_epsilon: float = 1e-6
329
    zero_centered_gamma: bool = False
330
331
332
333
334
    kernel_init: Initializer = None
    use_bias: bool = False
    bias_init: Initializer = nn.initializers.zeros
    apply_residual_connection_post_layernorm: bool = False
    output_layernorm: bool = False
335
    attn_mask_type: str = 'causal'
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
    dtype: DType = jnp.float32
    fuse_qkv: bool = True
    transpose_batch_sequence: bool = True
    scale_attn_logits: bool = False
    scaled_query_init: bool = True
    float32_logits: bool = False    # computes logits in float32 for stability.

    def __post_init__(self):
        if self.kernel_init is None:
            self.kernel_init = nn.initializers.variance_scaling(1.0, 'fan_in', 'normal')
        super().__post_init__()

    @nn.compact
    def __call__(self,
                 inputs_q: Array,
                 inputs_kv: Array,
                 mask: Optional[Array] = None,
                 bias: Optional[Array] = None,
                 *,
                 decode: bool = False,
                 deterministic: bool = False) -> Array:
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
        """
        MultiHeadAttention Layer:
        [Query, Key, Value projection] -> Dot Product Attention -> Output projection.

        Parameters
        ----------
        inputs_q : jax.numpy.ndarray
            Input tensor for query projection.
        inputs_kv : jax.numpy.ndarray
            Input tensor for key/value projection.
        mask : jax.numpy.ndarray, default = None
            Boolean tensor used to mask out self-attention softmax input.
        bias : jax.numpy.ndarray, default = None
            A tensor used to shift self-attention softmax input.
        *
        decode : bool,default = False
            Indicate whether to prepare and use an autoregressive cache.
        deterministic : bool,default = False
            Disable dropout layers if set to True.

        Returns
        -------
        outputs : jax.numpy.ndarray
            Output tensors.
        """
382
383

        def query_init(*args):
384
            depth_scaling = jnp.sqrt(self.head_dim).astype(self.dtype)
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
            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)

417
418
419
420
421
422
423
424
425
426
427
428
429
430
        # TODO(rewang): make it configurable for pre_scale_bias
        attn_bias_type = AttnBiasType.NO_BIAS if bias is None else AttnBiasType.POST_SCALE_BIAS

        def canonicalize_attn_mask_type(attn_mask_type):
            """
            Convert the string to AttnMaskType
            """
            if attn_mask_type == 'causal':
                return AttnMaskType.CAUSAL_MASK
            if attn_mask_type == 'padding':
                return AttnMaskType.PADDING_MASK
            raise ValueError(f"Unsupported {attn_mask_type=}, "
                             "supported attn_mask_type = {'causal', 'padding'}")

431
432
        is_self_attn = (inputs_q is inputs_kv)
        qkv_layout = QKVLayout.BS3HD if is_self_attn else QKVLayout.BSHD_BS2HD
433
        attn_mask_type = canonicalize_attn_mask_type(self.attn_mask_type)
434

435
436
437
        canonicalize_dtype = dtypes.canonicalize_dtype(self.dtype)
        q_seqlen = inputs_q.shape[0] if self.transpose_batch_sequence else inputs_q.shape[1]
        kv_seqlen = inputs_kv.shape[0] if self.transpose_batch_sequence else inputs_kv.shape[1]
438
        enable_fused_attn = int(os.getenv("NVTE_FUSED_ATTN", "0"))
439
440
441
442
443
444
445

        def _check_seqlen(seqlen):
            return seqlen % 64 == 0

        def _check_head_dim(head_dim):
            return head_dim in [64, 128]

446
        has_fused_attn_kernel = is_fused_attn_kernel_available(self.dtype, self.dtype, qkv_layout,
447
448
449
450
                                                               attn_bias_type, attn_mask_type,
                                                               self.dropout_rate, q_seqlen,
                                                               kv_seqlen, self.head_dim)

451
        use_fused_attn = not decode and not self.transpose_batch_sequence and self.fuse_qkv and \
452
            canonicalize_dtype in [jnp.bfloat16, jnp.float16] and \
453
454
            _check_seqlen(q_seqlen) and _check_seqlen(kv_seqlen) and \
            _check_head_dim(self.head_dim) and \
455
            has_fused_attn_kernel and \
456
            enable_fused_attn
457

458
        if enable_fused_attn and not use_fused_attn:
459
460
461
462
463
464
465
466
467
468
469
            reason = ""
            if decode:
                reason += f"decode=False is required but got {decode}, "
            if self.transpose_batch_sequence:
                reason += f"transpose_batch_sequence=False is required " \
                          f"but got {self.transpose_batch_sequence}, "
            if not self.fuse_qkv:
                reason += f"fuse_qkv=True is required but got {self.fuse_qkv}, "
            if canonicalize_dtype not in [jnp.bfloat16, jnp.float16]:
                reason += f"dtype in [BF16, FP16] is required " \
                          f"but got dtype={canonicalize_dtype}, "
470
471
            if not _check_seqlen(q_seqlen):
                reason += f"q_seqlen % 64 == 0 is required " \
472
                          f"but got {q_seqlen=}, "
473
474
            if not _check_seqlen(kv_seqlen):
                reason += f"kv_seqlen % 64 == 0 is required " \
475
                          f"but got {kv_seqlen=}, "
476
477
            if not _check_head_dim(self.head_dim):
                reason += f"head_dim should be 64 or 128 but got {self.head_dim}, "
478
479
            if not has_fused_attn_kernel:
                reason += "no fused attention kernel is available, "
480
481

            warnings.warn(
482
483
484
485
                f"Fused attention is not enabled. Because " \
                f"{reason}fall back to unfused attention.")

        first_sharding_type, second_sharding_type = infer_sharding_type()
486

487
488
        residual = inputs_q
        if self.fuse_qkv:
489
            if is_self_attn:
490
491
492
                qkv_proj, ln_out = LayerNormDenseGeneral(
                    enable_layernorm=not self.output_layernorm,
                    layernorm_type=self.layernorm_type,
493
                    zero_centered_gamma=self.zero_centered_gamma,
494
495
496
497
498
499
                    epsilon=self.layernorm_epsilon,
                    axis=-1,
                    features=(3, self.num_heads * self.head_dim),
                    sharding_type=first_sharding_type,
                    transpose_batch_sequence=self.transpose_batch_sequence,
                    return_layernorm_output=self.apply_residual_connection_post_layernorm,
500
501
502
                    scale_axes=(W_NO_SHARD_AXES,),
                    ln_bias_axes=(W_NO_SHARD_AXES,),
                    kernel_axes=(W_FSDP_AXES, W_JOINED_AXES, W_TP_AXES),
503
504
505
                    kernel_init=qkv_init,
                    use_bias=self.use_bias,
                    bias_init=self.bias_init,
506
                    bias_axes=(W_JOINED_AXES, W_TP_AXES),
507
508
                    name='qkv',
                    dtype=self.dtype)(inputs_q)
509
510
                if not use_fused_attn:
                    query, key, value = jnp.split(qkv_proj, [1, 2], axis=-2)
511
512
513
514
            else:
                query, ln_out = LayerNormDenseGeneral(
                    enable_layernorm=not self.output_layernorm,
                    layernorm_type=self.layernorm_type,
515
                    zero_centered_gamma=self.zero_centered_gamma,
516
517
518
519
520
521
                    epsilon=self.layernorm_epsilon,
                    axis=-1,
                    features=self.num_heads * self.head_dim,
                    sharding_type=first_sharding_type,
                    transpose_batch_sequence=self.transpose_batch_sequence,
                    return_layernorm_output=self.apply_residual_connection_post_layernorm,
522
523
524
                    scale_axes=(W_NO_SHARD_AXES,),
                    ln_bias_axes=(W_NO_SHARD_AXES,),
                    kernel_axes=(W_FSDP_AXES, W_TP_AXES),
525
526
                    use_bias=self.use_bias,
                    bias_init=self.bias_init,
527
                    bias_axes=(W_TP_AXES,),
528
529
530
531
532
533
534
                    dtype=self.dtype,
                    kernel_init=query_init,
                    name='query')(inputs_q)
                kv_proj = DenseGeneral(axis=-1,
                                       features=(2, self.num_heads * self.head_dim),
                                       sharding_type=first_sharding_type,
                                       transpose_batch_sequence=self.transpose_batch_sequence,
535
                                       kernel_axes=(W_FSDP_AXES, W_JOINED_AXES, W_TP_AXES),
536
537
538
                                       kernel_init=kv_init,
                                       use_bias=self.use_bias,
                                       bias_init=self.bias_init,
539
                                       bias_axes=(W_JOINED_AXES, W_TP_AXES),
540
541
                                       name='kv',
                                       dtype=self.dtype)(inputs_kv)
542
543
                if not use_fused_attn:
                    key, value = jnp.split(kv_proj, [1], axis=-2)
544
545
546
547
548
549
550
        else:
            kv_projection = functools.partial(
                DenseGeneral,
                axis=-1,
                features=self.num_heads * self.head_dim,
                sharding_type=first_sharding_type,
                transpose_batch_sequence=self.transpose_batch_sequence,
551
                kernel_axes=(W_FSDP_AXES, W_TP_AXES),
552
553
                use_bias=self.use_bias,
                bias_init=self.bias_init,
554
                bias_axes=(W_TP_AXES,),
555
556
557
558
                dtype=self.dtype)
            query, ln_out = LayerNormDenseGeneral(
                enable_layernorm=not self.output_layernorm,
                layernorm_type=self.layernorm_type,
559
                zero_centered_gamma=self.zero_centered_gamma,
560
561
562
563
564
565
                epsilon=self.layernorm_epsilon,
                axis=-1,
                features=self.num_heads * self.head_dim,
                sharding_type=first_sharding_type,
                transpose_batch_sequence=self.transpose_batch_sequence,
                return_layernorm_output=True,
566
567
568
                scale_axes=(W_NO_SHARD_AXES,),
                ln_bias_axes=(W_NO_SHARD_AXES,),
                kernel_axes=(W_FSDP_AXES, W_TP_AXES),
569
570
                use_bias=self.use_bias,
                bias_init=self.bias_init,
571
                bias_axes=(W_TP_AXES,),
572
573
574
575
                dtype=self.dtype,
                kernel_init=query_init,
                name='query')(inputs_q)

576
            if is_self_attn:
577
578
579
580
581
582
583
584
585
586
                assert ln_out is not None
                inputs_kv = ln_out

            key = kv_projection(kernel_init=self.kernel_init, name='key')(inputs_kv)
            value = kv_projection(kernel_init=self.kernel_init, name='value')(inputs_kv)

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

587
588
589
590
591
        if not use_fused_attn:
            query = query.reshape((query.shape[0], query.shape[1], self.num_heads, self.head_dim))
            key = key.reshape((key.shape[0], key.shape[1], self.num_heads, self.head_dim))
            value = value.reshape((value.shape[0], value.shape[1], self.num_heads, self.head_dim))
            qkv_sharding_constraint = \
592
                (SEQLEN_AXES, BATCH_AXES, HEAD_AXES, HIDDEN_AXES) \
593
                if self.transpose_batch_sequence \
594
595
596
597
                else (BATCH_AXES, SEQLEN_AXES, HEAD_AXES, HIDDEN_AXES)
            query = _with_sharding_constraint(query, qkv_sharding_constraint)
            key = _with_sharding_constraint(key, qkv_sharding_constraint)
            value = _with_sharding_constraint(value, qkv_sharding_constraint)
598
599
600
601

        if decode:
            is_initialized = self.has_variable('cache', 'cached_key')

Ming-Xu Huang's avatar
Ming-Xu Huang committed
602
603
            cached_key = self.variable('cache', 'cached_key', jnp.zeros, key.shape, key.dtype)
            cached_value = self.variable('cache', 'cached_value', jnp.zeros, value.shape,
604
605
606
607
                                         value.dtype)
            cache_index = self.variable('cache', 'cache_index',
                                        lambda: jnp.array(0, dtype=jnp.int32))
            if is_initialized:
Ming-Xu Huang's avatar
Ming-Xu Huang committed
608
609
610
611
612
613
614
615
                if self.transpose_batch_sequence:
                    length, batch, num_heads, head_dim = cached_key.value.shape
                    expected_shape = (1, batch, num_heads, head_dim)
                    one_hot_indices_shape = (length, 1, 1, 1)
                else:
                    batch, length, num_heads, head_dim = cached_key.value.shape
                    expected_shape = (batch, 1, num_heads, head_dim)
                    one_hot_indices_shape = (1, length, 1, 1)
616
617
618
619
620
621
622
623
624

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

                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
625
626
627
                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
628
629
630
631
632
                cached_key.value = key
                cached_value.value = value
                cache_index.value = cache_index.value + 1

                mask = combine_masks(
Ming-Xu Huang's avatar
Ming-Xu Huang committed
633
                    mask, jnp.broadcast_to(jnp.arange(length) > cur_index, (batch, 1, 1, length)))
634
635
636
637
638

                if bias is not None:
                    bias = dynamic_vector_slice_in_dim(jnp.squeeze(bias, axis=0),
                                                       jnp.reshape(cur_index, (-1)), 1, -2)

639
640
        scale_factor = 1.0 / sqrt(self.head_dim) if self.scale_attn_logits else 1.0

641
642
643
644
        dropout_rng = None
        if not deterministic and self.dropout_rate > 0.:
            dropout_rng = self.make_rng(self.dropout_rng_name)

645
646
647
        if use_fused_attn:
            assert mask is not None and mask.ndim == 4    # (b, 1, s_q, s_kv)
            assert not self.transpose_batch_sequence
648

649
650
651
652
653
654
            seed = None
            if dropout_rng is not None:
                seed = jax.random.split(dropout_rng, len(jax.devices()))
                # ensure the old key never used
                del dropout_rng

655
            if is_self_attn:
656
                qkv_proj = qkv_proj.reshape((*qkv_proj.shape[:-1], self.num_heads, self.head_dim))
657
658
659
                qkv_sharding_constraint = (BATCH_AXES, SEQLEN_AXES, JOINED_AXES, HEAD_AXES,
                                           HIDDEN_AXES)
                qkv_proj = _with_sharding_constraint(qkv_proj, qkv_sharding_constraint)
660
661
662
                x = self_fused_attn(qkv_proj,
                                    bias,
                                    mask,
663
                                    seed,
664
                                    attn_bias_type=attn_bias_type,
665
                                    attn_mask_type=attn_mask_type,
666
667
668
669
670
671
672
673
                                    scaling_factor=scale_factor,
                                    dropout_probability=self.dropout_rate,
                                    is_training=not deterministic,
                                    sharding_type=first_sharding_type)
            else:
                assert bias is None
                query = query.reshape((*query.shape[:-1], self.num_heads, self.head_dim))
                kv_proj = kv_proj.reshape((*kv_proj.shape[:-1], self.num_heads, self.head_dim))
674
675
676
677
678
                q_sharding_constraint = (BATCH_AXES, SEQLEN_AXES, HEAD_AXES, HIDDEN_AXES)
                kv_sharding_constraint = (BATCH_AXES, SEQLEN_AXES, JOINED_AXES, HEAD_AXES,
                                          HIDDEN_AXES)
                query = _with_sharding_constraint(query, q_sharding_constraint)
                kv_proj = _with_sharding_constraint(kv_proj, kv_sharding_constraint)
679
680
681
682

                x = cross_fused_attn(query,
                                     kv_proj,
                                     mask,
683
                                     seed,
684
                                     attn_bias_type=attn_bias_type,
685
                                     attn_mask_type=attn_mask_type,
686
687
688
689
690
                                     scaling_factor=scale_factor,
                                     dropout_probability=self.dropout_rate,
                                     is_training=not deterministic,
                                     sharding_type=first_sharding_type)
        else:
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705

            def convert_to_softmax_type(attn_mask_type, mask):
                """
                Convert the string to SoftmaxType
                """
                if attn_mask_type == 'causal':
                    return SoftmaxType.SCALED_UPPER_TRIANG_MASKED
                if attn_mask_type == 'padding':
                    if mask is not None:
                        return SoftmaxType.SCALED_MASKED
                    return SoftmaxType.SCALED
                raise ValueError(f"Unsupported {attn_mask_type=}, "
                                 "supported attn_mask_type = {'causal', 'padding'}")

            softmax_type = convert_to_softmax_type(self.attn_mask_type, mask)
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720

            x = core_attention(query,
                               key,
                               value,
                               scale_factor=scale_factor,
                               transpose_batch_sequence=self.transpose_batch_sequence,
                               softmax_type=softmax_type,
                               softmax_sharding_type=first_sharding_type,
                               mask=mask,
                               bias=bias,
                               dropout_rng=dropout_rng,
                               dropout_rate=self.dropout_rate,
                               deterministic=deterministic,
                               dtype=self.dtype,
                               float32_logits=self.float32_logits)
721
722
723
724

        x = x.reshape((x.shape[0], x.shape[1], x.shape[2] * x.shape[3]))

        attn_context_sharding_constraint = \
725
            (SEQLEN_AXES, BATCH_AXES, HIDDEN_TP_AXES) \
726
            if self.transpose_batch_sequence \
727
728
            else (BATCH_AXES, SEQLEN_AXES, HIDDEN_TP_AXES)
        x = _with_sharding_constraint(x, attn_context_sharding_constraint)
729
730
731
732
733
734

        out = DenseGeneral(features=inputs_q.shape[-1],
                           sharding_type=second_sharding_type,
                           transpose_batch_sequence=self.transpose_batch_sequence,
                           axis=-1,
                           kernel_init=self.kernel_init,
735
                           kernel_axes=(W_TP_AXES, W_FSDP_AXES),
736
737
                           use_bias=self.use_bias,
                           bias_init=self.bias_init,
738
                           bias_axes=(W_NO_SHARD_AXES,),
739
740
741
742
743
744
745
746
747
748
749
750
                           dtype=self.dtype,
                           name='out')(x)
        return out, residual


class RelativePositionBiases(nn.Module):
    """
    T5-style relative positional embeddings to the attention logits.

    Parameters
    ----------
    num_buckets : int
751
        The number of buckets to bucket distances between key and query positions into.
752
    max_distance : int
753
        The maximum distance before everything is lumped into the last
754
755
        distance bucket.
    num_attention_heads : int
756
        Number of attention heads in the transformer layer.
757
    embedding_init : Initializer, default = flax.linen.linear.default_embed_init
758
        Used for initializing relative embedding tables.
759
    embedding_axes : Tuple[str, ...], default = ('heads', 'relpos_buckets')
760
        The name of axes used to shard embedding attention bias with a corresponding mesh.
761
762
763
764

    Optimization parameters
    -----------------------
    dtype : jax.numpy.dtype, default  = jax.numpy.float32
765
        The data type used to allocate the initial parameters.
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
    """
    num_buckets: int
    max_distance: int
    num_attention_heads: int
    embedding_init: Callable[..., Array] = nn.linear.default_embed_init
    embedding_axes: Tuple[str, ...] = ('heads', 'relpos_buckets')
    dtype: DType = jnp.float32

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

        Parameters
        ----------
        q_seqlen : int
782
            The sequence length of query.
783
        k_seqlen : int
784
            The sequence length of key.
785
        bidirectional : bool, default = True
786
            Indicate whether to allow positive memory-query relative position
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
            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 + (
            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)
        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(
            'rel_embedding',
            self.embedding_init, (self.num_attention_heads, self.num_buckets),
            jnp.float32,
            axes=self.embedding_axes)

        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)

        values = lax.dot_general(relative_attention_bias, rp_bucket_one_hot,
                                 (((1,), (0,)), ((), ())))
        return values[jnp.newaxis, ...]


class TransformerLayerType(Enum):
837
838
839
840
841
842
843
844
845
846
    r"""
    TransformerLayerType is an Enum class to specify a type of TransformerLayer

    Values
    ----------
    ENCODER:
        Encoder type of TransformerLayer.
    DECODER:
        Decoder type of TransformerLayer.
    """
847
848
849
850
851
852
853
854
855
856
    ENCODER = "encoder"
    DECODER = "decoder"


class TransformerLayer(nn.Module):
    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”.

857
858
859
860
861
    .. note::

        Argument :attr:`attention_mask` will be ignored when
        :attr:`self_attn_mask_type` is set to `"causal"`.

862
863
864
    Parameters
    ----------
    hidden_size: int, default = 512
865
        The hidden size of each input sample.
866
    mlp_hidden_size: int, default = 2048
867
        Intermediate size to which input samples are projected.
868
    num_attention_heads: int, default = 8
869
        Number of attention heads in the transformer layer.
870
    layernorm_type : {'layernorm', 'rmsnorm'}, default = 'layernorm'
871
        Indicate the type of layer normalization.
872
    layernorm_epsilon: float, default = 1e-6
873
        A value added to the denominator of layer normalization for numerical stability.
874
875
876
877
878
879
880
881
    zero_centered_gamma : bool, default = False
        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'.
882
    hidden_dropout: float, default = 0.1
883
        Dropout probability for the dropout op after FC2 layer.
884
    hidden_dropout_dims: Sequence[int], default = ()
885
        Dimensions that will share the same dropout mask for hidden
886
    attention_dropout: float, default = 0.1
887
        Dropout probability for the dropout op during multi-head attention.
888
889
890
891
    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.
892
    dropout_rng_name: str, default = 'dropout'
893
894
        The key in given RNGs via flax.linen.Module.apply that for
        generating Dropout masks in the Multi-Head Attention.
895
896
    mha_kernel_init: Initializer, default =
        flax.linen.initializers.variance_scaling(1.0, 'fan_in', 'normal')
897
898
        Used for initializing weights of QKV and Output projection weights.
        It should be a callable object with three arguments (jax.random.PRNGKey, shape, dtype).
899
900
    mlp_kernel_init: Initializer, default =
        flax.linen.initializers.variance_scaling(1.0, 'fan_in', 'truncated_normal')
901
902
        Used for initializing weights of FC1 and FC2 layers.
        It should be a callable object with three arguments (jax.random.PRNGKey, shape, dtype).
903
    mlp_activations: Sequence[str], default = ('relu', )
904
        The sequence of activation functions to apply after the first linear transformation.
905
906
        Each activation has its own transformation layer.
    use_bias: bool, default = False
907
908
        Indicate whether to enable bias shifting for QKVO projections, FC1 and FC2.
        If set to False, the layer will not learn additive biases.
909
    bias_init: Initializer, default = flax.linen.initializers.zeros
910
911
912
        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).
913
    apply_residual_connection_post_layernorm: bool, default = False
914
        If set to True, residual connections are taken from the output
915
916
        of layer norm (default is taken from input of layer norm)
    output_layernorm: bool, default = False
917
        If set to True, layer normalization is applied on the output side,
918
919
920
        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
921
        If set to True, attention logits are executed in jax.numpy.float32.
922
    layer_type: TransformerLayerType, default = TransformerLayerType.ENCODER
923
        If set to TransformerLayerType.DECODER, an additional cross-attention block
924
925
        is added after self-attention.this can be used for structures like `T5`
        Transformer in conjunction with the TransformerLayerType.ENCODER option.
926
927
    self_attn_mask_type: {'causal', 'padding'}, default = 'causal'
        Type of attention mask passed into softmax operation.
928
        Introduced in v0.10.0.
929
    enable_relative_embedding: bool, default = True
930
        Whether to enable relative embedding as shifting of attention logits.
931
    relative_embedding: flax.linen.Module, default = None
932
        The module for relative embedding execution, only used when
933
934
935
936
937
938
939
940
941
942
        :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')

    Optimization parameters
    -----------------------
    dtype :jax.numpy.dtype, default  = jax.numpy.float32
943
        The data type used to allocate the initial parameters.
944
    drop_path: float, default = 0.0
945
        When > 0.0, applies stochastic depth per sample in the main
946
947
        path of the residual block.
    fuse_qkv_params: bool, default = True
948
        If set to True, `TransformerLayer` module exposes a single fused
949
950
        parameter for query-key-value for self-attention and key-value for
        cross-attention.
Jeng Bai-Cheng's avatar
Jeng Bai-Cheng committed
951
    transpose_batch_sequence : bool, default = False
952
        Indicate whether the input tensors were switched axis of batch
953
954
955
        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
956
        Indicate whether to scale attention logits.
957
958
959
        if set to True, :math:`\frac{Q}{\sqrt{head_dim}*K}`,
        else :math:`Q*K`
    scaled_query_init: bool, default = `True`
960
        Whether to scale WQ on initialization by :math:`\sqrt{head_dim}`
961
962
963
964
965
966
967
    """

    hidden_size: int = 512
    mlp_hidden_size: int = 2048
    num_attention_heads: int = 8
    layernorm_type: str = 'layernorm'
    layernorm_epsilon: float = 1e-6
968
    zero_centered_gamma: bool = False
969
970
971
    hidden_dropout: float = 0.1
    hidden_dropout_dims: Sequence[int] = ()
    attention_dropout: float = 0.1
972
973
    intermediate_dropout: float = 0.1
    intermediate_dropout_dims: Sequence[int] = ()
974
975
976
977
978
979
980
981
982
983
    dropout_rng_name: str = 'dropout'
    mha_kernel_init: Initializer = None
    mlp_kernel_init: Initializer = None
    mlp_activations: Sequence[str] = ('relu',)
    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
984
    self_attn_mask_type: str = 'causal'
985
986
987
988
989
    enable_relative_embedding: bool = True
    relative_embedding: nn.Module = None
    dtype: DType = jnp.float32
    drop_path: float = 0.0
    fuse_qkv_params: bool = True
Jeng Bai-Cheng's avatar
Jeng Bai-Cheng committed
990
    transpose_batch_sequence: bool = False
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
    scale_attn_logits: bool = False
    scaled_query_init: bool = True

    def __post_init__(self):
        if self.mha_kernel_init is None:
            self.mha_kernel_init = nn.initializers.variance_scaling(1.0, 'fan_in', 'normal')
        if self.mlp_kernel_init is None:
            self.mlp_kernel_init = nn.initializers.variance_scaling(1.0, 'fan_in',
                                                                    'truncated_normal')
        super().__post_init__()

    @nn.compact
    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):
        """
        Transformer Layer: attention block and a feedforward network (MLP)

        Parameters
        ----------
        inputs : jax.numpy.ndarray
            Input tensor.
        encoded : jax.numpy.ndarray, default = None
            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.
        encoder_decoder_mask : jax.numpy.ndarray, default = None
            Boolean tensor used to mask out cross-attention softmax input when
            :attr:`layer_type=TransformerLayerType.DECODER`.
        deterministic: bool, default = False
1027
            Disable dropout layers if set to True.
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
        decode: bool,default = False
            Indicate whether to prepare and use an autoregressive cache
            in Multi-head attention (MHA).
        max_decode_length : bool, default = None
            The maximum length to generate relative embedding biases when
            :attr:`layer_type=TransformerLayerType.DECODER` and
            :attr:`enable_relative_embedding=True`.

        Returns
        -------
        outputs : jax.numpy.ndarray
1039
            Output tensors.
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
        """
        assert self.layer_type in TransformerLayerType, \
                "layer_type should be one of TransformerLayerType" \
                f", but got {self.layer_type}."

        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=}."

        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."

        head_dim = self.hidden_size // self.num_attention_heads

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

        attn_bias = None
        if self.enable_relative_embedding:
            if self.relative_embedding is None:
                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')
            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:
            mha_name = 'attention'
        else:
            mha_name = 'self_attention'

1089
1090
        inputs = _with_sharding_constraint(inputs, (BATCH_AXES, SEQLEN_AXES, HIDDEN_AXES))

1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
        # [batch, length, emb_dim] -> [batch, length, emb_dim]
        x, residual = MultiHeadAttention(
            num_heads=self.num_attention_heads,
            dtype=self.dtype,
            head_dim=head_dim,
            transpose_batch_sequence=self.transpose_batch_sequence,
            dropout_rate=self.attention_dropout,
            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,
1104
            zero_centered_gamma=self.zero_centered_gamma,
1105
1106
            apply_residual_connection_post_layernorm=self.apply_residual_connection_post_layernorm,
            output_layernorm=self.output_layernorm,
1107
            attn_mask_type=self.self_attn_mask_type,
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
            fuse_qkv=self.fuse_qkv_params,
            kernel_init=self.mha_kernel_init,
            use_bias=self.use_bias,
            bias_init=self.bias_init,
            name=mha_name)(inputs,
                           inputs,
                           attention_mask,
                           attn_bias,
                           deterministic=deterministic,
                           decode=decode)

        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
1123
                assert -x_shape_len <= dims < x_shape_len
1124
1125

            return nn.Dropout(rate=self.hidden_dropout,
1126
1127
                              broadcast_dims=self.hidden_dropout_dims,
                              rng_collection=self.dropout_rng_name)(x, deterministic=deterministic)
1128
1129
1130
1131
1132

        x = hidden_dropout(x, deterministic)
        if self.drop_path > 0.0:
            drop_path_shape = _generate_drop_path_shape(x.shape, batch_dim)
            x = nn.Dropout(rate=self.drop_path,
1133
1134
                           broadcast_dims=drop_path_shape,
                           rng_collection=self.dropout_rng_name)(x, deterministic=deterministic)
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
        x = x + residual

        mlp_input = x
        if self.layer_type == TransformerLayerType.DECODER:
            assert encoded is not None, \
                "encoded is required when layer_type == TransformerLayerType.DECODER."

            y, residual = MultiHeadAttention(
                num_heads=self.num_attention_heads,
                dtype=self.dtype,
                head_dim=head_dim,
                transpose_batch_sequence=self.transpose_batch_sequence,
                dropout_rate=self.attention_dropout,
                dropout_rng_name=self.dropout_rng_name,
                layernorm_type=self.layernorm_type,
                layernorm_epsilon=self.layernorm_epsilon,
1151
                zero_centered_gamma=self.zero_centered_gamma,
1152
1153
1154
                apply_residual_connection_post_layernorm=self.
                apply_residual_connection_post_layernorm,
                output_layernorm=False,    # Must do LayerNorm before MHA.
1155
                attn_mask_type='padding',
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
                float32_logits=self.float32_attention_logits,
                scale_attn_logits=self.scale_attn_logits,
                scaled_query_init=self.scaled_query_init,
                fuse_qkv=self.fuse_qkv_params,
                kernel_init=self.mha_kernel_init,
                use_bias=self.use_bias,
                bias_init=self.bias_init,
                name='encoder_decoder_attention')(x,
                                                  encoded,
                                                  encoder_decoder_mask,
                                                  deterministic=deterministic)
            y = hidden_dropout(y, deterministic)
            mlp_input = y + residual

1170
1171
        mlp_input = _with_sharding_constraint(mlp_input, (BATCH_AXES, SEQLEN_AXES, HIDDEN_AXES))

1172
1173
1174
1175
        # MlpBlock
        residual = mlp_input
        z, ln_out = LayerNormMLP(
            layernorm_type=self.layernorm_type,
1176
            zero_centered_gamma=self.zero_centered_gamma,
1177
1178
1179
1180
1181
1182
            epsilon=self.layernorm_epsilon,
            major_sharding_type=infer_major_sharding_type(),
            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,
1183
1184
1185
            intermediate_dropout_rng_name=self.dropout_rng_name,
            intermediate_dropout_rate=self.intermediate_dropout,
            intermediate_hidden_dropout_dims=self.intermediate_dropout_dims,
1186
            dtype=self.dtype,
1187
1188
            scale_axes=(W_NO_SHARD_AXES,),
            ln_bias_axes=(W_NO_SHARD_AXES,),
1189
            kernel_init=self.mlp_kernel_init,
1190
1191
            kernel_axes_1=(W_FSDP_AXES, W_JOINED_AXES, W_TP_AXES),
            kernel_axes_2=(W_TP_AXES, W_FSDP_AXES),
1192
1193
            use_bias=self.use_bias,
            bias_init=self.bias_init,
1194
1195
            bias_axes_1=(W_JOINED_AXES, W_TP_AXES),
            bias_axes_2=(W_NO_SHARD_AXES,),
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
            name='mlp',
        )(mlp_input, deterministic=deterministic)

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

        z = hidden_dropout(z, deterministic)
        if self.drop_path > 0.0:
            drop_path_shape = _generate_drop_path_shape(z.shape, batch_dim)
            z = nn.Dropout(rate=self.drop_path,
                           broadcast_dims=drop_path_shape)(z, deterministic=deterministic)
        z = z + residual

        if self.output_layernorm:
            ln_sharding_type, _ = infer_sharding_type()
            z = LayerNorm(layernorm_type=self.layernorm_type,
1213
1214
                          zero_centered_gamma=self.zero_centered_gamma,
                          epsilon=self.layernorm_epsilon,
1215
1216
                          scale_axes=(W_NO_SHARD_AXES,),
                          bias_axes=(W_NO_SHARD_AXES,),
1217
1218
1219
1220
1221
1222
                          transpose_batch_sequence=self.transpose_batch_sequence,
                          dtype=self.dtype,
                          sharding_type=ln_sharding_type,
                          name="output_layer_norm")(z)

        return z