cpp_extensions.py 214 KB
Newer Older
1
# Copyright (c) 2022-2024, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
2
3
4
5
6
7
8
9
#
# See LICENSE for license information.
"""JAX te custom call"""
from abc import ABCMeta, abstractmethod
from dataclasses import dataclass
from typing import Tuple
from functools import partial, reduce
import operator
10
import os
11
12
import warnings

13
14
15
16
17
import numpy as np
import jax.numpy as jnp
from jax.lib import xla_client
from jax import core, dtypes
from jax.interpreters import xla, mlir
18
from jax.experimental.custom_partitioning import custom_partitioning
19
from jax.interpreters.mlir import ir, dtype_to_ir_type
20
21
from jax.sharding import PartitionSpec, NamedSharding
from jax._src.interpreters import batching
22
23
from jax._src import dispatch

24
25
26
27
28
29
30
import transformer_engine_jax
from transformer_engine_jax import DType as TEDType
from transformer_engine_jax import NVTE_Bias_Type
from transformer_engine_jax import NVTE_Mask_Type
from transformer_engine_jax import NVTE_QKV_Layout
from transformer_engine_jax import NVTE_Fused_Attn_Backend

31
32
33
34
from .sharding import all_reduce_max_along_all_axes_except_PP
from .sharding import all_reduce_sum_along_dp_fsdp
from .sharding import get_all_mesh_axes, num_of_devices
from .sharding import get_padded_spec as te_get_padded_spec
35

36
37
38
39
40
41
42
try:
    from jaxlib.hlo_helpers import custom_call
except ImportError:
    # Newer JAX changed its API. But we want to support a few JAX
    # version, so we still need this import.
    pass

43
44
45
46
47
48
49
50
51
for _name, _value in transformer_engine_jax.registrations().items():
    xla_client.register_custom_call_target(_name, _value, platform="CUDA")


def te_dtype_to_jax_dtype(te_dtype):
    """
    convert TE dtype to jax dtype
    """
    assert isinstance(te_dtype, TEDType)
52
53
54
55
56
57
58
59
60

    converter = {
        TEDType.kFloat32: jnp.float32,
        TEDType.kFloat16: jnp.float16,
        TEDType.kBFloat16: jnp.bfloat16,
        TEDType.kInt32: jnp.int32,
        TEDType.kInt64: jnp.int64,
        TEDType.kFloat8E4M3: jnp.float8_e4m3fn,
        TEDType.kFloat8E5M2: jnp.float8_e5m2,
61
        TEDType.kByte: jnp.uint8
62
63
64
65
66
67
    }

    if te_dtype not in converter:
        raise ValueError(f"Unsupported {te_dtype=}")

    return converter.get(te_dtype)
68
69
70
71
72
73
74
75
76


def te_dtype_to_ir_dtype(te_dtype):
    """
    convert TE dtype to MLIR dtype
    """
    return dtype_to_ir_type(np.dtype(te_dtype_to_jax_dtype(te_dtype)))


77
78
79
80
81
82
83
def jax_dtype_to_ir_dtype(jax_dtype):
    """
    convert Jax dtype to MLIR dtype
    """
    return dtype_to_ir_type(np.dtype(jax_dtype))


84
85
86
87
def jax_dtype_to_te_dtype(jax_dtype):
    """
    convert jax dtype to TE dtype
    """
88
    jax_dtype = dtypes.canonicalize_dtype(jax_dtype)
89

90
91
92
93
94
95
96
97
    converter = {
        jnp.float32.dtype: TEDType.kFloat32,
        jnp.float16.dtype: TEDType.kFloat16,
        jnp.bfloat16.dtype: TEDType.kBFloat16,
        jnp.int32.dtype: TEDType.kInt32,
        jnp.int64.dtype: TEDType.kInt64,
        jnp.float8_e4m3fn.dtype: TEDType.kFloat8E4M3,
        jnp.float8_e5m2.dtype: TEDType.kFloat8E5M2,
98
        jnp.uint8.dtype: TEDType.kByte,
99
    }
100

101
102
    if jax_dtype not in converter:
        raise ValueError(f"Unsupported {jax_dtype=}")
103

104
    return converter.get(jax_dtype)
105
106


107
108
109
110
111
112
113
114
def get_padded_spec(arg_info):
    """
    Get padded spec for partitioning from arguments' information
    """
    if arg_info.sharding is None:
        return te_get_padded_spec(None, arg_info.ndim)
    ndim, spec = arg_info.ndim, arg_info.sharding.spec
    return te_get_padded_spec(spec, ndim)
115
116


117
def _check_valid_batch_dims(bdims):
118
    """
119
    Assert out non-supported bath dims
120
    """
121
122
123
124
    for dim in bdims:
        assert dim in [0, None], \
            "Currently only support batch_dim in [0, None], " \
            f"but got {dim=}"
125
126
127
128


class BasePrimitive(metaclass=ABCMeta):
    """
129
    jax primitive
130
131
132
133
134
135
136
137
138
139
    """

    @staticmethod
    @abstractmethod
    def abstract():
        """
        to describe computing graph
        """
        return NotImplemented

140
141
142
143
144
145
146
    @classmethod
    def outer_abstract(cls, *args, **kwargs):
        """
        optional abstract wrapper to eliminate workspace tensors
        """
        return cls.abstract(*args, **kwargs)

147
148
149
150
151
152
153
154
    @staticmethod
    @abstractmethod
    def lowering():
        """
        to describe MLIR
        """
        return NotImplemented

155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
    @staticmethod
    @abstractmethod
    def impl():
        """
        to describe implementation
        """
        return NotImplemented

    @staticmethod
    @abstractmethod
    def batcher():
        """
        to describe batch rules for vmap
        """
        return NotImplemented

    @staticmethod
    @abstractmethod
    def infer_sharding_from_operands():
        """
        to describe infer_sharding_from_operands for custom_partitioning
        """
        return NotImplemented

    @staticmethod
    @abstractmethod
    def partition():
        """
        to describe partition for custom_partitioning
        """
        return NotImplemented

187
188
189
190
191

def register_primitive(cls):
    """
    register jax primitive
    """
192
193
194
195
196

    def name_of_wrapper_p():
        return cls.name + "_wrapper"

    inner_p = core.Primitive(cls.name)
197
    dispatch.prim_requires_devices_during_lowering.add(inner_p)
198
199
200
201
202
203
204
    inner_p.multiple_results = cls.multiple_results
    inner_p.def_impl(partial(xla.apply_primitive, inner_p))
    inner_p.def_abstract_eval(cls.abstract)
    mlir.register_lowering(inner_p, cls.lowering, platform='cuda')
    cls.inner_primitive = inner_p

    outer_p = core.Primitive(name_of_wrapper_p())
205
    dispatch.prim_requires_devices_during_lowering.add(outer_p)
206
207
    outer_p.multiple_results = cls.multiple_results
    outer_p.def_impl(cls.impl)
208
    outer_p.def_abstract_eval(cls.outer_abstract)
209
210
211
212
213
214
215
    batching.primitive_batchers[outer_p] = cls.batcher
    outer_p_lower = custom_partitioning(cls.impl, static_argnums=cls.impl_static_args)
    outer_p_lower.def_partition(infer_sharding_from_operands=cls.infer_sharding_from_operands,
                                partition=cls.partition)
    mlir.register_lowering(outer_p,
                           mlir.lower_fun(outer_p_lower, multiple_results=cls.multiple_results))
    cls.outer_primitive = outer_p
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262


@dataclass
class CustomCallArgsWrapper:
    """
    wrapper of XLA custom call args
    """

    def __init__(self,
                 output_types,
                 operands,
                 operand_shapes,
                 operand_specific_layouts=None,
                 output_specific_layouts=None):
        self.output_types = output_types
        self.operands = operands
        self.operand_layouts = CustomCallArgsWrapper.generate_layouts(operand_shapes,
                                                                      operand_specific_layouts)
        output_shapes = [x.shape for x in output_types]
        self.output_layouts = CustomCallArgsWrapper.generate_layouts(output_shapes,
                                                                     output_specific_layouts)

    @staticmethod
    def generate_layouts(shapes, specific_layouts):
        """
        setup layouts for XLA custom call
        """

        def default_layout(shape):
            return range(len(shape) - 1, -1, -1)

        if specific_layouts is None:
            specific_layouts = {}

        layouts = []
        for idx, shape in enumerate(shapes):
            if idx in specific_layouts:
                layouts.append(specific_layouts[idx])
            else:
                layouts.append(default_layout(shape))
        return layouts


def custom_caller(name, args, opaque, has_side_effect, **kwargs):
    """
    XLA custom call warpper
    """
263
264
265
266
267
268
269
270
271
272
273
274
275
    if hasattr(mlir, "custom_call"):
        out = mlir.custom_call(name,
                               result_types=args.output_types,
                               operands=args.operands,
                               operand_layouts=args.operand_layouts,
                               result_layouts=args.output_layouts,
                               backend_config=opaque,
                               has_side_effect=has_side_effect,
                               **kwargs).results
    else:
        # Need to disable one pylint error as the second function
        # parameter name recenctly in JAX. Otherwise we won't be
        # compatible with multiple JAX version.
276
277
278
279
280
281
282
283
284
        out = custom_call(    # pylint: disable=too-many-function-args
            name,
            args.output_types,
            operands=args.operands,
            operand_layouts=args.operand_layouts,
            result_layouts=args.output_layouts,
            backend_config=opaque,
            has_side_effect=has_side_effect,
            **kwargs)
285
286
287
    return out


288
class LayerNormFwdPrimitive(BasePrimitive):
289
    """
290
    Layer Normalization Forward Primitive
291
    """
292
293
294
295
296
    name = "te_layernorm_forward"
    multiple_results = True
    impl_static_args = (3, 4)    # zero_centered_gamma, epsilon
    inner_primitive = None
    outer_primitive = None
297
298

    @staticmethod
299
    def abstract(x_aval, gamma_aval, beta_aval, **kwargs):
300
        """
301
        LayerNorm fwd inner primitive abstract
302
        """
303
304
305
306
        x_dtype = dtypes.canonicalize_dtype(x_aval.dtype)
        assert x_dtype in [jnp.float32, jnp.float16, jnp.bfloat16]

        mu_rsigama_dtype = jnp.float32
307

308
309
        out_aval = core.raise_to_shaped(x_aval)
        mu_aval = rsigma_aval = out_aval.update(shape=out_aval.shape[:-1], dtype=mu_rsigama_dtype)
310

311
312
313
314
        assert gamma_aval.size == beta_aval.size
        hidden_size = gamma_aval.size
        assert x_aval.size % hidden_size == 0

315
        wkspace_info, barrier_info = transformer_engine_jax.get_layernorm_fwd_workspace_sizes(
316
            x_aval.size // hidden_size,    # batch size
317
            hidden_size,
318
319
320
321
322
323
            jax_dtype_to_te_dtype(x_aval.dtype),    # in te_dtype
            jax_dtype_to_te_dtype(gamma_aval.dtype),    # weight te_dtype
            jax_dtype_to_te_dtype(x_aval.dtype),    # out te_dtype (same as input for Fp16/Bf16)
            True,
            kwargs['zero_centered_gamma'],
            kwargs['epsilon'])
324
325
326
327
328
329
330
331
332
333
334
335
336
337
        wkspace_aval = out_aval.update(shape=wkspace_info[0],
                                       dtype=te_dtype_to_jax_dtype(wkspace_info[1]))
        barrier_aval = out_aval.update(shape=barrier_info[0],
                                       dtype=te_dtype_to_jax_dtype(barrier_info[1]))

        return out_aval, mu_aval, rsigma_aval, wkspace_aval, barrier_aval

    @staticmethod
    def outer_abstract(*args, **kwargs):
        """
        LayerNorm fwd outer primitive abstract
        """
        out_aval, mu_aval, rsigma_aval, _, _ = \
            LayerNormFwdPrimitive.abstract(*args, **kwargs)
338
        return out_aval, mu_aval, rsigma_aval
339
340

    @staticmethod
341
    def lowering(ctx, x, gamma, beta, *, zero_centered_gamma, epsilon):
342
        """
343
        LayerNorm fwd lowering rules
344
        """
345
346
347
348
349
350
351
352
        x_aval, gamma_aval, beta_aval = ctx.avals_in
        assert gamma_aval.dtype == beta_aval.dtype
        x_type = ir.RankedTensorType(x.type)
        x_shape = x_type.shape
        g_type = ir.RankedTensorType(gamma.type)
        g_shape = g_type.shape
        b_type = ir.RankedTensorType(beta.type)
        b_shape = b_type.shape
353

354
355
        assert g_type == b_type
        assert g_shape == b_shape
356

357
358
359
360
361
        # Output shape is same as the input shape, but the output type is same as the weight type.
        # See ln_api.cpp
        output_type = g_type.element_type
        ir_mu_dtype = ir.F32Type.get()
        ir_rsigma_dtype = ir.F32Type.get()
362

363
364
365
366
        out_shape = x_shape
        hidden_size = reduce(operator.mul, g_shape)
        batch_shape = out_shape[:-1]
        batch_size = reduce(operator.mul, x_shape) // hidden_size
367

368
369
        wkspace_aval, barrier_aval = ctx.avals_out[-2:]

370
371
372
373
        out_types = [
            ir.RankedTensorType.get(out_shape, output_type),
            ir.RankedTensorType.get(batch_shape, ir_mu_dtype),
            ir.RankedTensorType.get(batch_shape, ir_rsigma_dtype),
374
375
            ir.RankedTensorType.get(wkspace_aval.shape, jax_dtype_to_ir_dtype(wkspace_aval.dtype)),
            ir.RankedTensorType.get(barrier_aval.shape, jax_dtype_to_ir_dtype(barrier_aval.dtype))
376
377
378
379
        ]
        operands = [x, gamma, beta]
        operand_shapes = [x_shape, g_shape, b_shape]
        args = CustomCallArgsWrapper(out_types, operands, operand_shapes)
380

381
382
        sm_margin = int(os.getenv("NVTE_FWD_LAYERNORM_SM_MARGIN", "0"))

383
384
385
        opaque = transformer_engine_jax.pack_norm_descriptor(
            batch_size,
            hidden_size,
386
387
            wkspace_aval.size,
            barrier_aval.size,
388
389
            0,    # no dgamma_part in FWD pass
            0,    # no dbeta_part in BWD pass
390
391
            jax_dtype_to_te_dtype(x_aval.dtype),
            jax_dtype_to_te_dtype(gamma_aval.dtype),
392
393
            jax_dtype_to_te_dtype(wkspace_aval.dtype),
            jax_dtype_to_te_dtype(barrier_aval.dtype),
394
395
            TEDType.kByte,    # dummy dgamma_part te_dtype
            TEDType.kByte,    # dummy dbeta_part te_dtype
396
397
            zero_centered_gamma,
            epsilon,
398
            sm_margin,
399
        )
400

401
        out = custom_caller(LayerNormFwdPrimitive.name, args, opaque, False)
402

403
        return out
404
405

    @staticmethod
406
    def impl(x, gamma, beta, zero_centered_gamma, epsilon):
407
        """
408
        to describe implementation
409
        """
410
        assert LayerNormFwdPrimitive.inner_primitive is not None
411
        out, mu, rsigma, _, _ = LayerNormFwdPrimitive.inner_primitive.bind(
412
413
414
415
416
            x, gamma, beta, zero_centered_gamma=zero_centered_gamma, epsilon=epsilon)
        return out, mu, rsigma

    @staticmethod
    def batcher(batched_args, batch_dims, *, zero_centered_gamma, epsilon):
417
        """
418
        to describe batch rules for vmap
419
        """
420
421
422
423
        _check_valid_batch_dims(batch_dims)
        assert LayerNormFwdPrimitive.outer_primitive is not None
        x, gamma, beta = batched_args
        x_bdim, _, _ = batch_dims
424

425
426
427
428
429
430
        out_bdims = x_bdim, x_bdim, x_bdim
        return LayerNormFwdPrimitive.outer_primitive.bind(x,
                                                          gamma,
                                                          beta,
                                                          zero_centered_gamma=zero_centered_gamma,
                                                          epsilon=epsilon), out_bdims
431

432
433
434
435
436
437
438
439
440
441
442
443
444
    @staticmethod
    def infer_sharding_from_operands(zero_centered_gamma, epsilon, mesh, arg_infos, result_infos):
        del zero_centered_gamma, epsilon, result_infos
        x_spec = get_padded_spec(arg_infos[0])
        if x_spec[-1] is not None:
            warnings.warn(
                f"Does not support to shard hidden dim in {LayerNormFwdPrimitive.name}! " \
                f"Force to not shard the hidden dim, which might introduce extra collective ops, " \
                f"and hurt performance."
            )
        out_sharding = NamedSharding(mesh, PartitionSpec(*x_spec[:-1], None))
        mu_sharding = rsigma_sharding = NamedSharding(mesh, PartitionSpec(*x_spec[:-1]))
        return (out_sharding, mu_sharding, rsigma_sharding)
445

446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
    @staticmethod
    def partition(zero_centered_gamma, epsilon, mesh, arg_infos, result_infos):
        del result_infos
        x_spec, g_spec, b_spec = map(get_padded_spec, arg_infos)
        if x_spec[-1] is not None:
            warnings.warn(
                f"Does not support to shard hidden dim in {LayerNormFwdPrimitive.name}! " \
                f"Force to not shard the hidden dim, which might introduce extra collective ops, " \
                f"and hurt performance."
            )
        x_sharding = NamedSharding(mesh, PartitionSpec(*x_spec[:-1], None))
        g_sharding = NamedSharding(mesh, PartitionSpec(*g_spec))
        b_sharding = NamedSharding(mesh, PartitionSpec(*b_spec))
        out_sharding = x_sharding
        mu_sharding = rsigma_sharding = NamedSharding(mesh, PartitionSpec(*x_spec[:-1]))
461

462
463
464
465
466
467
        arg_shardings = (x_sharding, g_sharding, b_sharding)
        out_shardings = (out_sharding, mu_sharding, rsigma_sharding)
        impl = partial(LayerNormFwdPrimitive.impl,
                       zero_centered_gamma=zero_centered_gamma,
                       epsilon=epsilon)
        return mesh, impl, out_shardings, arg_shardings
468
469


470
register_primitive(LayerNormFwdPrimitive)
471
472


473
474
def layernorm_fwd(x: jnp.ndarray, gamma: jnp.ndarray, beta: jnp.ndarray, zero_centered_gamma: bool,
                  epsilon: float):
475
    """
476
    Wrapper for TE layernorm fwd
477
    """
478
479
480
481
482
    return LayerNormFwdPrimitive.outer_primitive.bind(x,
                                                      gamma,
                                                      beta,
                                                      zero_centered_gamma=zero_centered_gamma,
                                                      epsilon=epsilon)
483
484


485
class LayerNormBwdPrimitive(BasePrimitive):
486
    """
487
    Layer Normalization Backward Primitive
488
    """
489
490
491
492
493
    name = "te_layernorm_backward"
    multiple_results = True
    impl_static_args = (5, 6)    # zero_centered_gamma, epsilon
    inner_primitive = None
    outer_primitive = None
494
495

    @staticmethod
496
    def abstract(dz_aval, x_aval, mu_aval, rsigma_aval, gamma_aval, **kwargs):
497
        """
498
        Layernorm bwd inner primitive abstract
499
        """
500
501
502
503
504
505
506
507
        w_dtype = dtypes.canonicalize_dtype(gamma_aval.dtype)
        mu_dtype = dtypes.canonicalize_dtype(mu_aval.dtype)
        rsigma_dtype = dtypes.canonicalize_dtype(rsigma_aval.dtype)

        assert dtypes.canonicalize_dtype(dz_aval.dtype) == w_dtype
        assert dz_aval.shape == x_aval.shape
        assert mu_aval.shape == rsigma_aval.shape == x_aval.shape[:-1]
        assert mu_dtype == rsigma_dtype == jnp.float32
508

509
510
        dx_aval = core.raise_to_shaped(dz_aval)
        dgamma_aval = dbeta_aval = core.raise_to_shaped(gamma_aval)
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526

        wkspace_info, barrier_info, dgamma_part_info, dbeta_part_info = \
            transformer_engine_jax.get_layernorm_bwd_workspace_sizes(
                x_aval.size // gamma_aval.size,           # batch size
                gamma_aval.size,                          # hidden size
                jax_dtype_to_te_dtype(x_aval.dtype),      # input te_dtype
                jax_dtype_to_te_dtype(gamma_aval.dtype),  # weight te_dtype
                True, kwargs['zero_centered_gamma'], kwargs['epsilon']
            )
        wkspace_aval = dx_aval.update(shape=wkspace_info[0],
                                      dtype=te_dtype_to_jax_dtype(wkspace_info[1]))
        barrier_aval = dx_aval.update(shape=barrier_info[0],
                                      dtype=te_dtype_to_jax_dtype(barrier_info[1]))
        dgamma_part_aval = dgamma_aval.update(shape=dgamma_part_info[0],
                                              dtype=te_dtype_to_jax_dtype(dgamma_part_info[1]))
        dbeta_part_aval = dbeta_aval.update(shape=dbeta_part_info[0],
527
                                            dtype=te_dtype_to_jax_dtype(dbeta_part_info[1]))
528
529
530
531
532
533
534
535
536
537
538

        return dx_aval, dgamma_aval, dbeta_aval, wkspace_aval, barrier_aval, \
               dgamma_part_aval, dbeta_part_aval

    @staticmethod
    def outer_abstract(*args, **kwargs):
        """
        LayerNorm bwd outer primitive abstract
        """
        dx_aval, dgamma_aval, dbeta_aval, _, _, _, _ = \
            LayerNormBwdPrimitive.abstract(*args, **kwargs)
539
        return dx_aval, dgamma_aval, dbeta_aval
540
541

    @staticmethod
542
    def lowering(ctx, dz, x, mu, rsigma, gamma, *, zero_centered_gamma, epsilon):
543
        """
544
        Layernorm bwd lowering rules
545
        """
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
        _, x_aval, _, _, gamma_aval = ctx.avals_in
        x_type = ir.RankedTensorType(x.type)
        x_shape = x_type.shape
        g_type = ir.RankedTensorType(gamma.type)
        g_shape = g_type.shape
        b_type = ir.RankedTensorType(gamma.type)
        b_shape = b_type.shape
        assert g_type == b_type
        assert g_shape == b_shape

        dz_shape = ir.RankedTensorType(dz.type).shape
        mu_shape = ir.RankedTensorType(mu.type).shape
        rsigma_shape = ir.RankedTensorType(rsigma.type).shape

        hidden_size = reduce(operator.mul, g_shape)
        batch_size = reduce(operator.mul, x_shape) // hidden_size
562
563

        out_types = [
564
565
            ir.RankedTensorType.get(output.shape, mlir.dtype_to_ir_type(output.dtype))
            for output in ctx.avals_out
566
        ]
567

568
569
        operands = [dz, mu, rsigma, x, gamma]
        operand_shapes = [dz_shape, mu_shape, rsigma_shape, x_shape, g_shape]
570
571
        args = CustomCallArgsWrapper(out_types, operands, operand_shapes)

572
573
        sm_margin = int(os.getenv("NVTE_BWD_LAYERNORM_SM_MARGIN", "0"))

574
        wkspace_aval, barrier_aval, dgamma_part_aval, dbeta_part_aval = ctx.avals_out[-4:]
575
576
577
        opaque = transformer_engine_jax.pack_norm_descriptor(
            batch_size,
            hidden_size,
578
579
580
581
            wkspace_aval.size,
            barrier_aval.size,
            dgamma_part_aval.size,
            dbeta_part_aval.size,
582
583
            jax_dtype_to_te_dtype(x_aval.dtype),
            jax_dtype_to_te_dtype(gamma_aval.dtype),
584
585
586
587
            jax_dtype_to_te_dtype(wkspace_aval.dtype),
            jax_dtype_to_te_dtype(barrier_aval.dtype),
            jax_dtype_to_te_dtype(dgamma_part_aval.dtype),
            jax_dtype_to_te_dtype(dbeta_part_aval.dtype),
588
589
            zero_centered_gamma,
            epsilon,
590
            sm_margin,
591
        )
592

593
        out = custom_caller(LayerNormBwdPrimitive.name, args, opaque, False)
594

595
        return out
596

597
598
599
    @staticmethod
    def impl(dz, x, mu, rsigma, gamma, zero_centered_gamma, epsilon):
        assert LayerNormBwdPrimitive.inner_primitive is not None
600
        dx, dgamma, dbeta, _, _, _, _ = LayerNormBwdPrimitive.inner_primitive.bind(
601
602
            dz, x, mu, rsigma, gamma, zero_centered_gamma=zero_centered_gamma, epsilon=epsilon)
        return dx, dgamma, dbeta
603

604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
    @staticmethod
    def batcher(batched_args, batch_dims, *, zero_centered_gamma, epsilon):
        _check_valid_batch_dims(batch_dims)
        assert LayerNormBwdPrimitive.outer_primitive is not None
        dz, x, mu, rsigma, gamma = batched_args
        _, x_bdim, _, _, gamma_bdim = batch_dims

        out_bdims = x_bdim, gamma_bdim, gamma_bdim
        return LayerNormBwdPrimitive.outer_primitive.bind(dz,
                                                          x,
                                                          mu,
                                                          rsigma,
                                                          gamma,
                                                          zero_centered_gamma=zero_centered_gamma,
                                                          epsilon=epsilon), out_bdims
619

620
621
622
623
624
625
626
627
628
629
630
631
632
633
    @staticmethod
    def infer_sharding_from_operands(zero_centered_gamma, epsilon, mesh, arg_infos, result_infos):
        del zero_centered_gamma, epsilon, result_infos
        x_spec = get_padded_spec(arg_infos[1])
        if x_spec[-1] is not None:
            warnings.warn(
                f"Does not support to shard hidden dim in {LayerNormBwdPrimitive.name}! " \
                f"Force to not shard the hidden dim, which might introduce extra collective ops, " \
                f"and hurt performance."
            )
        g_b_spec = get_padded_spec(arg_infos[4])
        dx_sharding = NamedSharding(mesh, PartitionSpec(*x_spec[:-1], None))
        dgamma_sharding = dbeta_sharding = NamedSharding(mesh, PartitionSpec(*g_b_spec))
        return dx_sharding, dgamma_sharding, dbeta_sharding
634

635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
    @staticmethod
    def partition(zero_centered_gamma, epsilon, mesh, arg_infos, result_infos):
        del result_infos
        x_spec = get_padded_spec(arg_infos[1])
        if x_spec[-1] is not None:
            warnings.warn(
                f"Does not support to shard hidden dim in {LayerNormBwdPrimitive.name}! " \
                f"Force to not shard the hidden dim, which might introduce extra collective ops, " \
                f"and hurt performance."
            )
        g_b_spec = get_padded_spec(arg_infos[4])
        dx_sharding = NamedSharding(mesh, PartitionSpec(*x_spec[:-1], None))
        dgamma_sharding = dbeta_sharding = NamedSharding(mesh, PartitionSpec(*g_b_spec))
        out_shardings = dx_sharding, dgamma_sharding, dbeta_sharding
        x_shardings = (dx_sharding,) * 2    # dz and x should have the same sharding.
        mu_shardings = (NamedSharding(mesh, PartitionSpec(*x_spec[:-1])),) * 2
        arg_shardings = (*x_shardings, *mu_shardings, NamedSharding(mesh, PartitionSpec(*g_b_spec)))

        def sharded_impl(dz, x, mu, rsigma, gamma):
            local_dx, local_dgamma, local_dbeta = \
                LayerNormBwdPrimitive.impl(dz, x, mu, rsigma, gamma,
                     zero_centered_gamma=zero_centered_gamma,
                     epsilon=epsilon)
            global_dgamma = all_reduce_sum_along_dp_fsdp(local_dgamma)
            global_dbeta = all_reduce_sum_along_dp_fsdp(local_dbeta)
            return local_dx, global_dgamma, global_dbeta

        return mesh, sharded_impl, out_shardings, arg_shardings


register_primitive(LayerNormBwdPrimitive)


def layernorm_bwd(dz: jnp.ndarray, x: jnp.ndarray, mu: jnp.ndarray, rsigma: jnp.ndarray,
                  gamma: jnp.ndarray, zero_centered_gamma: bool, epsilon: float):
670
    """
671
    Wrapper for TE layernorm bwd
672
    """
673
674
675
676
677
678
679
    return LayerNormBwdPrimitive.outer_primitive.bind(dz,
                                                      x,
                                                      mu,
                                                      rsigma,
                                                      gamma,
                                                      zero_centered_gamma=zero_centered_gamma,
                                                      epsilon=epsilon)
680
681


682
class RmsNormFwdPrimitive(BasePrimitive):
683
    """
684
    RMS Normalization Forward Primitive
685
    """
686
    name = "te_rmsnorm_forward"
687
    multiple_results = True
688
689
690
    impl_static_args = (2,)    # epsilon
    inner_primitive = None
    outer_primitive = None
691
692

    @staticmethod
693
    def abstract(x_aval, gamma_aval, **kwargs):
694
        """
695
        RMSNorm fwd inner primitive abstract
696
        """
697
698
699
700
701
702
703
        x_dtype = dtypes.canonicalize_dtype(x_aval.dtype)
        assert x_dtype in [jnp.float32, jnp.float16, jnp.bfloat16]

        rsigama_dtype = jnp.float32

        out_aval = core.raise_to_shaped(x_aval)
        rsigma_aval = out_aval.update(shape=out_aval.shape[:-1], dtype=rsigama_dtype)
704

705
706
        hidden_size = gamma_aval.size
        assert x_aval.size % hidden_size == 0
707

708
        wkspace_info, barrier_info = transformer_engine_jax.get_layernorm_fwd_workspace_sizes(
709
            x_aval.size // hidden_size,    # batch size
710
            hidden_size,
711
712
713
714
715
716
            jax_dtype_to_te_dtype(x_aval.dtype),    # in te_dtype
            jax_dtype_to_te_dtype(gamma_aval.dtype),    # weight te_dtype
            jax_dtype_to_te_dtype(x_aval.dtype),    # out te_dtype (same as input for Fp16/Bf16)
            False,
            False,
            kwargs['epsilon'])
717
718
719
720
721
722
723
724
725
726
727
728
729
        wkspace_aval = out_aval.update(shape=wkspace_info[0],
                                       dtype=te_dtype_to_jax_dtype(wkspace_info[1]))
        barrier_aval = out_aval.update(shape=barrier_info[0],
                                       dtype=te_dtype_to_jax_dtype(barrier_info[1]))

        return out_aval, rsigma_aval, wkspace_aval, barrier_aval

    @staticmethod
    def outer_abstract(*args, **kwargs):
        """
        RMSNorm fwd outer primitive abstract
        """
        out_aval, rsigma_aval, _, _ = RmsNormFwdPrimitive.abstract(*args, **kwargs)
730
        return out_aval, rsigma_aval
731
732

    @staticmethod
733
    def lowering(ctx, x, gamma, *, epsilon):
734
        """
735
        RMSNorm fwd lowering rules
736
        """
737
738
739
740
741
742
743
744
745
746
747
        x_aval, gamma_aval = ctx.avals_in
        x_type = ir.RankedTensorType(x.type)
        x_shape = x_type.shape
        g_type = ir.RankedTensorType(gamma.type)
        g_shape = g_type.shape
        rsigma_element_type = ir.F32Type.get()

        out_shape = x_shape
        hidden_size = reduce(operator.mul, g_shape)
        batch_shape = out_shape[:-1]
        batch_size = reduce(operator.mul, x_shape) // hidden_size
748

749
750
        wkspace_aval, barrier_aval = ctx.avals_out[-2:]

751
        out_types = [
752
753
            ir.RankedTensorType.get(out_shape, x_type.element_type),
            ir.RankedTensorType.get(batch_shape, rsigma_element_type),
754
755
            ir.RankedTensorType.get(wkspace_aval.shape, jax_dtype_to_ir_dtype(wkspace_aval.dtype)),
            ir.RankedTensorType.get(barrier_aval.shape, jax_dtype_to_ir_dtype(barrier_aval.dtype))
756
        ]
757
758
        operands = [x, gamma]
        operand_shapes = [x_shape, g_shape]
759
760
        args = CustomCallArgsWrapper(out_types, operands, operand_shapes)

761
762
        sm_margin = int(os.getenv("NVTE_FWD_LAYERNORM_SM_MARGIN", "0"))

763
764
765
        opaque = transformer_engine_jax.pack_norm_descriptor(
            batch_size,
            hidden_size,
766
767
            wkspace_aval.size,
            barrier_aval.size,
768
769
            0,    # no dgamma_part in FWD pass
            0,    # no dbeta_part in BWD pass
770
771
            jax_dtype_to_te_dtype(x_aval.dtype),
            jax_dtype_to_te_dtype(gamma_aval.dtype),
772
773
            jax_dtype_to_te_dtype(wkspace_aval.dtype),
            jax_dtype_to_te_dtype(barrier_aval.dtype),
774
775
            TEDType.kByte,    # dummy dgamma_part te_dtype
            TEDType.kByte,    # dummy dbeta_part te_dtype
776
777
            False,    # RMSNorm doesn't support zero_centered_gamma
            epsilon,
778
            sm_margin,
779
        )
780

781
        out = custom_caller(RmsNormFwdPrimitive.name, args, opaque, False)
782
783
784
785

        return out

    @staticmethod
786
    def impl(x, gamma, epsilon):
787
        """
788
        to describe implementation
789
        """
790
        assert RmsNormFwdPrimitive.inner_primitive is not None
791
        out, rsigma, _, _ = RmsNormFwdPrimitive.inner_primitive.bind(x, gamma, epsilon=epsilon)
792
        return out, rsigma
793
794

    @staticmethod
795
    def batcher(batched_args, batch_dims, *, epsilon):
796
        """
797
        to describe batch rules for vmap
798
        """
799
800
801
802
        _check_valid_batch_dims(batch_dims)
        assert RmsNormFwdPrimitive.outer_primitive is not None
        x, gamma = batched_args
        x_bdim, _ = batch_dims
803

804
805
        out_bdims = x_bdim, x_bdim
        return RmsNormFwdPrimitive.outer_primitive.bind(x, gamma, epsilon=epsilon), out_bdims
806

807
808
809
810
811
812
813
814
815
816
817
818
819
    @staticmethod
    def infer_sharding_from_operands(epsilon, mesh, arg_infos, result_infos):
        del epsilon, result_infos
        x_spec = get_padded_spec(arg_infos[0])
        if x_spec[-1] is not None:
            warnings.warn(
                f"Does not support to shard hidden dim in {RmsNormFwdPrimitive.name}! " \
                f"Force to not shard the hidden dim, which might introduce extra collective ops, " \
                f"and hurt performance."
            )
        out_sharding = NamedSharding(mesh, PartitionSpec(*x_spec[:-1], None))
        rsigma_sharding = NamedSharding(mesh, PartitionSpec(*x_spec[:-1]))
        return (out_sharding, rsigma_sharding)
820

821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
    @staticmethod
    def partition(epsilon, mesh, arg_infos, result_infos):
        del result_infos
        x_spec, g_spec = map(get_padded_spec, arg_infos)
        if x_spec[-1] is not None:
            warnings.warn(
                f"Does not support to shard hidden dim in {RmsNormFwdPrimitive.name}! " \
                f"Force to not shard the hidden dim, which might introduce extra collective ops, " \
                f"and hurt performance."
            )
        x_sharding = NamedSharding(mesh, PartitionSpec(*x_spec[:-1], None))
        g_sharding = NamedSharding(mesh, PartitionSpec(*g_spec))
        out_sharding = x_sharding
        rsigma_sharding = NamedSharding(mesh, PartitionSpec(*x_spec[:-1]))
        arg_shardings = (x_sharding, g_sharding)
        out_shardings = (out_sharding, rsigma_sharding)
        impl = partial(RmsNormFwdPrimitive.impl, epsilon=epsilon)
        return mesh, impl, out_shardings, arg_shardings
839
840


841
register_primitive(RmsNormFwdPrimitive)
842
843


844
def rmsnorm_fwd(x: jnp.ndarray, gamma: jnp.ndarray, epsilon: float):
845
    """
846
    Wrapper for TE rmsnorm fwd
847
    """
848
    return RmsNormFwdPrimitive.outer_primitive.bind(x, gamma, epsilon=epsilon)
849
850


851
class RmsNormBwdPrimitive(BasePrimitive):
852
    """
853
    RMS Normalization Backward Primitive
854
    """
855
    name = "te_rmsnorm_backward"
856
    multiple_results = True
857
858
859
    impl_static_args = (4,)    # epsilon
    inner_primitive = None
    outer_primitive = None
860
861

    @staticmethod
862
    def abstract(dz_aval, x_aval, rsigma_aval, gamma_aval, **kwargs):
863
        """
864
        RMSNorm bwd inner primitive abstract
865
        """
866
867
868
869
870
871
872
873
874
875
        w_dtype = dtypes.canonicalize_dtype(gamma_aval.dtype)
        rsigma_dtype = dtypes.canonicalize_dtype(rsigma_aval.dtype)

        assert dtypes.canonicalize_dtype(dz_aval.dtype) == w_dtype
        assert dz_aval.shape == x_aval.shape
        assert rsigma_aval.shape == x_aval.shape[:-1]
        assert rsigma_dtype == jnp.float32

        dx_aval = core.raise_to_shaped(dz_aval)
        dgamma_aval = core.raise_to_shaped(gamma_aval)
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899

        wkspace_info, barrier_info, dgamma_part_info, _ = \
            transformer_engine_jax.get_layernorm_bwd_workspace_sizes(
                x_aval.size // gamma_aval.size,           # batch size
                gamma_aval.size,                          # hidden size
                jax_dtype_to_te_dtype(x_aval.dtype),      # in te_dtype
                jax_dtype_to_te_dtype(gamma_aval.dtype),  # weight te_dtype
                False, False, kwargs['epsilon']
            )
        wkspace_aval = dx_aval.update(shape=wkspace_info[0],
                                      dtype=te_dtype_to_jax_dtype(wkspace_info[1]))
        barrier_aval = dx_aval.update(shape=barrier_info[0],
                                      dtype=te_dtype_to_jax_dtype(barrier_info[1]))
        dgamma_part_aval = dgamma_aval.update(shape=dgamma_part_info[0],
                                              dtype=te_dtype_to_jax_dtype(dgamma_part_info[1]))

        return dx_aval, dgamma_aval, wkspace_aval, barrier_aval, dgamma_part_aval

    @staticmethod
    def outer_abstract(*args, **kwargs):
        """
        RMSNorm bwd outer primitive abstract
        """
        dx_aval, dgamma_aval, _, _, _ = RmsNormBwdPrimitive.abstract(*args, **kwargs)
900
901
902
903
        return dx_aval, dgamma_aval

    @staticmethod
    def lowering(ctx, dz, x, rsigma, gamma, *, epsilon):
904
        """
905
        RMSNorm bwd lowering rules
906
        """
907
908
909
910
911
912
913
914
915
916
        _, x_aval, _, gamma_aval = ctx.avals_in
        x_type = ir.RankedTensorType(x.type)
        x_shape = x_type.shape
        g_type = ir.RankedTensorType(gamma.type)
        g_shape = g_type.shape
        dz_shape = ir.RankedTensorType(dz.type).shape
        rsigma_shape = ir.RankedTensorType(rsigma.type).shape

        hidden_size = reduce(operator.mul, g_shape)
        batch_size = reduce(operator.mul, x_shape) // hidden_size
917

918
919
        wkspace_aval, barrier_aval, dgamma_part_aval = ctx.avals_out[-3:]

920
        out_types = [
921
922
            ir.RankedTensorType.get(x_shape, x_type.element_type),
            ir.RankedTensorType.get(g_shape, g_type.element_type),
923
924
925
926
            ir.RankedTensorType.get(wkspace_aval.shape, jax_dtype_to_ir_dtype(wkspace_aval.dtype)),
            ir.RankedTensorType.get(barrier_aval.shape, jax_dtype_to_ir_dtype(barrier_aval.dtype)),
            ir.RankedTensorType.get(dgamma_part_aval.shape,
                                    jax_dtype_to_ir_dtype(dgamma_part_aval.dtype))
927
        ]
928
929
        operands = [dz, rsigma, x, gamma]
        operand_shapes = [dz_shape, rsigma_shape, x_shape, g_shape]
930
931
        args = CustomCallArgsWrapper(out_types, operands, operand_shapes)

932
933
        sm_margin = int(os.getenv("NVTE_BWD_LAYERNORM_SM_MARGIN", "0"))

934
935
936
        opaque = transformer_engine_jax.pack_norm_descriptor(
            batch_size,
            hidden_size,
937
938
939
            wkspace_aval.size,
            barrier_aval.size,
            dgamma_part_aval.size,
940
            0,    # no dbeta_part for RMSnorm
941
942
            jax_dtype_to_te_dtype(x_aval.dtype),
            jax_dtype_to_te_dtype(gamma_aval.dtype),
943
944
945
            jax_dtype_to_te_dtype(wkspace_aval.dtype),
            jax_dtype_to_te_dtype(barrier_aval.dtype),
            jax_dtype_to_te_dtype(dgamma_part_aval.dtype),
946
            TEDType.kByte,    # dummy dbeta_part te_dtype
947
948
            False,    # RMSNorm doesn't support zero_centered_gamma
            epsilon,
949
            sm_margin,
950
        )
951

952
        out = custom_caller(RmsNormBwdPrimitive.name, args, opaque, False)
953
954
955

        return out

956
957
958
    @staticmethod
    def impl(dz, x, rsigma, gamma, epsilon):
        assert RmsNormBwdPrimitive.inner_primitive is not None
959
960
        dx, dgamma, _, _, _ = \
            RmsNormBwdPrimitive.inner_primitive.bind(dz, x, rsigma, gamma, epsilon=epsilon)
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
        return dx, dgamma

    @staticmethod
    def batcher(batched_args, batch_dims, *, epsilon):
        _check_valid_batch_dims(batch_dims)
        assert RmsNormBwdPrimitive.outer_primitive is not None
        dz, x, rsigma, gamma = batched_args
        _, x_bdim, _, gamma_bdim = batch_dims

        out_bdims = x_bdim, gamma_bdim
        return RmsNormBwdPrimitive.outer_primitive.bind(dz, x, rsigma, gamma,
                                                        epsilon=epsilon), out_bdims

    @staticmethod
    def infer_sharding_from_operands(epsilon, mesh, arg_infos, result_infos):
        del epsilon, result_infos
        x_spec = get_padded_spec(arg_infos[1])
        if x_spec[-1] is not None:
            warnings.warn(
                f"Does not support to shard hidden dim in {RmsNormBwdPrimitive.name}! " \
                f"Force to not shard the hidden dim, which might introduce extra collective ops, " \
                f"and hurt performance."
            )
        g_spec = get_padded_spec(arg_infos[3])
        dx_sharding = NamedSharding(mesh, PartitionSpec(*x_spec[:-1], None))
        dgamma_sharding = NamedSharding(mesh, PartitionSpec(*g_spec))
        return dx_sharding, dgamma_sharding

    @staticmethod
    def partition(epsilon, mesh, arg_infos, result_infos):
        del result_infos
        x_spec = get_padded_spec(arg_infos[1])
        if x_spec[-1] is not None:
            warnings.warn(
                f"Does not support to shard hidden dim in {RmsNormBwdPrimitive.name}! " \
                f"Force to not shard the hidden dim, which might introduce extra collective ops, " \
                f"and hurt performance."
            )
        g_spec = get_padded_spec(arg_infos[3])
        dx_sharding = NamedSharding(mesh, PartitionSpec(*x_spec[:-1], None))
        dgamma_sharding = NamedSharding(mesh, PartitionSpec(*g_spec))
        out_shardings = dx_sharding, dgamma_sharding
        x_shardings = (dx_sharding,) * 2    # dz and x should have the same sharding.
        rsigma_sharding = NamedSharding(mesh, PartitionSpec(*x_spec[:-1]))
        arg_shardings = (*x_shardings, rsigma_sharding, NamedSharding(mesh, PartitionSpec(*g_spec)))

        def sharded_impl(dz, x, rsigma, gamma):
            local_dx, local_dgamma = \
                RmsNormBwdPrimitive.impl(dz, x, rsigma, gamma, epsilon=epsilon)
            global_dgamma = all_reduce_sum_along_dp_fsdp(local_dgamma)
            return local_dx, global_dgamma

        return mesh, sharded_impl, out_shardings, arg_shardings

1015

1016
register_primitive(RmsNormBwdPrimitive)
1017
1018


1019
1020
def rmsnorm_bwd(dz: jnp.ndarray, x: jnp.ndarray, rsigma: jnp.ndarray, gamma: jnp.ndarray,
                epsilon: float):
1021
    """
1022
    Wrapper for TE layernorm bwd
1023
    """
1024
    return RmsNormBwdPrimitive.outer_primitive.bind(dz, x, rsigma, gamma, epsilon=epsilon)
1025
1026


1027
class SoftmaxPrimitive(BasePrimitive):
1028
    """
1029
    Softmax Primitive
1030
    """
1031
    max_k_seqlen_supported = 4096
1032
    name = "te_softmax_internal_placeholder"
1033
1034

    @staticmethod
1035
1036
1037
1038
1039
    @abstractmethod
    def is_kernel_available(batch: int, heads: int, q_seqlen: int, k_seqlen: int,
                            dtype: jnp.dtype) -> bool:
        """Check Softmax kernel availability based on size"""
        raise NotImplementedError
1040

1041
1042
1043
1044
1045
    @staticmethod
    def get_batch_per_block(k_seqlen: int) -> int:
        """Get batch per CTA in Softmax kernels"""
        threads_per_warp = 32
        threads_per_block = 128    # Depends on the kernel implmentation
1046

1047
1048
1049
1050
1051
1052
        pow2 = 1 << (k_seqlen - 1).bit_length()
        warp_size = pow2 if pow2 < threads_per_warp else threads_per_warp
        batches_per_warp = 2 if pow2 <= 128 else 1
        warps_per_block = threads_per_block // warp_size
        batches_per_block = warps_per_block * batches_per_warp
        return batches_per_block
1053
1054

    @staticmethod
1055
    def forward_abstract(logits_aval, scale_factor):
1056
        """
1057
        softmax_forward abstract
1058
        """
1059
1060
1061
1062
1063
1064
1065
1066
1067
        del scale_factor
        i_dtype = dtypes.canonicalize_dtype(logits_aval.dtype)
        assert i_dtype in [jnp.float16, jnp.bfloat16]
        i_shape = logits_aval.shape
        # Assume [...Batch, Head, Q_Seqlen, K_Seqlen]
        q_seqlen = i_shape[-2]
        k_seqlen = i_shape[-1]
        assert k_seqlen <= SoftmaxPrimitive.max_k_seqlen_supported
        assert q_seqlen > 1
1068

1069
1070
        out_aval = core.raise_to_shaped(logits_aval)
        return out_aval
1071

1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
    @staticmethod
    def forward_lowering(name, ctx, logits, *, scale_factor):
        """
        softmax_forward lowering rules
        """
        i_aval, = ctx.avals_in
        i_type = ir.RankedTensorType(logits.type)
        i_shape = i_type.shape
        # Assume [...Batch, Head, Q_Seqlen, K_Seqlen]
        batch = reduce(operator.mul, i_shape[:-3])
        pad_batch = batch
        heads = i_shape[-3]
        q_seqlen = i_shape[-2]
        k_seqlen = i_shape[-1]

        out_types = [ir.RankedTensorType.get(i_shape, i_type.element_type)]
        operands = [logits]
        operand_shapes = [i_shape]
1090
1091
        args = CustomCallArgsWrapper(out_types, operands, operand_shapes)

1092
1093
1094
1095
        opaque = transformer_engine_jax.pack_softmax_descriptor(batch, pad_batch, heads, q_seqlen,
                                                                k_seqlen,
                                                                jax_dtype_to_te_dtype(i_aval.dtype),
                                                                scale_factor)
1096

1097
        out = custom_caller(name, args, opaque, False)
1098
1099
1100

        return [out]

1101
1102
1103
1104
1105
1106
1107
1108
    @staticmethod
    def forward_impl(primitive, logits, scale_factor):
        """
        softmax_forward implementation
        """
        assert primitive is not None
        output = primitive.bind(logits, scale_factor=scale_factor)
        return output
1109

1110
1111
1112
1113
1114
1115
1116
1117
    @staticmethod
    def forward_batcher(primitive, batched_args, batch_dims, *, scale_factor):
        """
        softmax_forward batcher
        """
        assert primitive is not None
        logits, = batched_args
        logits_bdim, = batch_dims
1118

1119
1120
        out_bdims = logits_bdim
        return primitive.bind(logits, scale_factor=scale_factor), out_bdims
1121

1122
1123
    @classmethod
    def forward_infer_sharding_from_operands(cls, scale_factor, mesh, arg_infos, result_infos):
1124
1125
1126
1127
1128
        """
        softmax_forward infer_sharding_from_operands
        """
        del scale_factor, result_infos    # Unused.
        logits_spec = get_padded_spec(arg_infos[0])
1129
1130
1131
1132
1133
1134
1135
        if logits_spec[-1] is not None:
            warnings.warn(
                f"Sharding the hidden dimension is not supported in {cls.name}! " \
                f"Forcing XLA to not shard the hidden dim, which might introduce extra " \
                f"collective ops and hurt performance."
            )
        out_sharding = NamedSharding(mesh, PartitionSpec(*logits_spec[:-1], None))
1136
        return out_sharding
1137

1138
1139
    @classmethod
    def forward_partition(cls, impl, scale_factor, mesh, arg_infos, result_infos):
1140
        """
1141
        softmax_forward partitioning
1142
        """
1143
        del result_infos
1144
1145
1146
1147
1148
1149
1150
1151
1152
        logits_spec = get_padded_spec(arg_infos[0])
        if logits_spec[-1] is not None:
            warnings.warn(
                f"Sharding the hidden dimension is not supported in {cls.name}! " \
                f"Forcing XLA to not shard the hidden dim, which might introduce extra " \
                f"collective ops and hurt performance."
            )
        out_shardings = NamedSharding(mesh, PartitionSpec(*logits_spec[:-1], None))
        arg_shardings = (out_shardings,)
1153
1154
        impl = partial(impl, scale_factor=scale_factor)
        return mesh, impl, out_shardings, arg_shardings
1155

1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
    @staticmethod
    def backward_abstract(dz_aval, softmax_out_aval, scale_factor=None):    # pylint: disable=unused-argument
        """
        softmax_backward abstract
        """
        dz_dtype = dtypes.canonicalize_dtype(dz_aval.dtype)
        softmax_out_dtype = dtypes.canonicalize_dtype(softmax_out_aval.dtype)
        assert dz_dtype == softmax_out_dtype
        assert dz_dtype in [jnp.float16, jnp.bfloat16]
        assert softmax_out_dtype in [jnp.float16, jnp.bfloat16]
1166

1167
        assert dz_aval.shape == softmax_out_aval.shape
1168

1169
        dx_aval = core.raise_to_shaped(dz_aval)
1170
        return dx_aval
1171
1172

    @staticmethod
1173
    def backward_lowering(name, ctx, dz, softmax_out, *, scale_factor):
1174
        """
1175
        softmax_backward lowering rules
1176
        """
1177
        dz_aval, _ = ctx.avals_in
1178

1179
1180
        dz_type = ir.RankedTensorType(dz.type)
        dz_shape = dz_type.shape
1181

1182
1183
1184
1185
1186
1187
        # Assume [...Batch, Head, Q_Seqlen, K_Seqlen]
        batch = reduce(operator.mul, dz_shape[:-3])
        pad_batch = batch    # unused
        heads = dz_shape[-3]
        q_seqlen = dz_shape[-2]
        k_seqlen = dz_shape[-1]
1188

1189
1190
        softmax_out_type = ir.RankedTensorType(softmax_out.type)
        softmax_out_shape = softmax_out_type.shape
1191

1192
        out_types = [ir.RankedTensorType.get(dz_shape, dz_type.element_type)]
1193
1194
        operands = [dz, softmax_out]
        operand_shapes = [dz_shape, softmax_out_shape]
1195
1196
        args = CustomCallArgsWrapper(out_types, operands, operand_shapes)

1197
1198
1199
        opaque = transformer_engine_jax.pack_softmax_descriptor(
            batch, pad_batch, heads, q_seqlen, k_seqlen, jax_dtype_to_te_dtype(dz_aval.dtype),
            scale_factor)
1200

1201
        out = custom_caller(name, args, opaque, False)
1202

1203
        return [out]
1204
1205

    @staticmethod
1206
    def backward_impl(primitive, dz, softmax_out, scale_factor):
1207
        """
1208
        softmax_backward implementation
1209
        """
1210
1211
1212
        assert primitive is not None
        dx = primitive.bind(dz, softmax_out, scale_factor=scale_factor)
        return dx
1213

1214
1215
1216
1217
1218
1219
1220
1221
    @staticmethod
    def backward_batcher(primitive, batched_args, batch_dims, *, scale_factor):
        """
        softmax_backward batcher
        """
        assert primitive is not None
        dz, softmax_out = batched_args
        _, softmax_out_bdim = batch_dims
1222

1223
1224
        out_bdims = softmax_out_bdim
        return primitive.bind(dz, softmax_out, scale_factor=scale_factor), out_bdims
1225

1226
1227
    @classmethod
    def backward_infer_sharding_from_operands(cls, scale_factor, mesh, arg_infos, result_infos):
1228
        """
1229
        softmax_backward infer_sharding_from_operands
1230
        """
1231
        del scale_factor, result_infos    # Unused.
1232
1233
1234
1235
1236
1237
1238
1239
        dz_spec = get_padded_spec(arg_infos[0])
        if dz_spec[-1] is not None:
            warnings.warn(
                f"Sharding the hidden dimension is not supported in {cls.name}! " \
                f"Forcing XLA to not shard the hidden dim, which might introduce extra " \
                f"collective ops and hurt performance."
            )
        dx_sharding = NamedSharding(mesh, PartitionSpec(*dz_spec[:-1], None))
1240
        return dx_sharding
1241

1242
1243
    @classmethod
    def backward_partition(cls, impl, scale_factor, mesh, arg_infos, result_infos):
1244
1245
1246
1247
        """
        softmax_backward partition
        """
        del result_infos
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263

        dz_spec = get_padded_spec(arg_infos[0])
        softmax_out_spec = get_padded_spec(arg_infos[1])
        if dz_spec[-1] is not None or softmax_out_spec[-1] is not None:
            warnings.warn(
                f"Sharding the hidden dimension is not supported in {cls.name}! " \
                f"Forcing XLA to not shard the hidden dim, which might introduce extra " \
                f"collective ops and hurt performance."
            )

        dz_sharding = NamedSharding(mesh, PartitionSpec(*dz_spec[:-1], None))
        softmax_out_sharding = NamedSharding(mesh, PartitionSpec(*softmax_out_spec[:-1], None))
        dx_sharding = dz_sharding
        arg_shardings = (dz_sharding, softmax_out_sharding)
        out_shardings = dx_sharding

1264
1265
        impl = partial(impl, scale_factor=scale_factor)
        return mesh, impl, out_shardings, arg_shardings
1266
1267


1268
1269
1270
1271
1272
1273
1274
1275
1276
class ScaledSoftmaxFwdPrimitive(SoftmaxPrimitive):
    """
    Scaled Softmax Fwd Primitive
    """
    name = "te_scaled_softmax_forward"
    multiple_results = False
    impl_static_args = (1,)    # scale_factor
    inner_primitive = None
    outer_primitive = None
1277

1278
1279
1280
1281
1282
    @staticmethod
    def is_kernel_available(batch: int, heads: int, q_seqlen: int, k_seqlen: int,
                            dtype: jnp.dtype) -> bool:
        """Check Softmax kernel availability based on size"""
        attn_batches = batch * heads
1283

1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
        dtype = dtypes.canonicalize_dtype(dtype)
        if (dtype in [jnp.float16, jnp.bfloat16]
                and 16 < k_seqlen <= SoftmaxPrimitive.max_k_seqlen_supported
        # k_seqlen must be 16 ~ 4096
                and q_seqlen % 4 == 0    # q_seqlen must be divisor of 4
                and attn_batches % 4 == 0    # batch * heads must be divisor of 4
           ):
            if 0 <= k_seqlen <= SoftmaxPrimitive.max_k_seqlen_supported:
                batch_per_block = SoftmaxPrimitive.get_batch_per_block(k_seqlen)
                return q_seqlen % batch_per_block == 0
        return False
1295

1296
1297
1298
1299
1300
1301
    @staticmethod
    def abstract(logits_aval, scale_factor):    # pylint: disable=unused-argument
        """
        te_scaled_softmax_forward abstract
        """
        return SoftmaxPrimitive.forward_abstract(logits_aval, scale_factor)
1302

1303
1304
1305
1306
1307
1308
1309
1310
1311
    @staticmethod
    def lowering(ctx, logits, *, scale_factor):
        """
        te_scaled_softmax_forward lowering rules
        """
        return SoftmaxPrimitive.forward_lowering(ScaledSoftmaxFwdPrimitive.name,
                                                 ctx,
                                                 logits,
                                                 scale_factor=scale_factor)
1312

1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
    @staticmethod
    def impl(logits, scale_factor):
        return SoftmaxPrimitive.forward_impl(ScaledSoftmaxFwdPrimitive.inner_primitive, logits,
                                             scale_factor)

    @staticmethod
    def batcher(batched_args, batch_dims, *, scale_factor):
        _check_valid_batch_dims(batch_dims)
        return SoftmaxPrimitive.forward_batcher(ScaledSoftmaxFwdPrimitive.outer_primitive,
                                                batched_args,
                                                batch_dims,
                                                scale_factor=scale_factor)
1325

1326
1327
    @staticmethod
    def infer_sharding_from_operands(scale_factor, mesh, arg_infos, result_infos):
1328
1329
1330
        return ScaledSoftmaxFwdPrimitive.forward_infer_sharding_from_operands(
            scale_factor, mesh, arg_infos, result_infos
        )
1331
1332
1333

    @staticmethod
    def partition(scale_factor, mesh, arg_infos, result_infos):
1334
1335
1336
        return ScaledSoftmaxFwdPrimitive.forward_partition(
            ScaledSoftmaxFwdPrimitive.impl, scale_factor, mesh, arg_infos, result_infos
        )
1337
1338


1339
register_primitive(ScaledSoftmaxFwdPrimitive)
1340

1341
1342

def scaled_softmax_fwd(logits: jnp.ndarray, scale_factor: float) -> jnp.ndarray:
1343
    """
1344
1345
    scaled_softmax_forward wrapper
    Return FP16/BF16 tensor
1346
    """
1347
    return ScaledSoftmaxFwdPrimitive.outer_primitive.bind(logits, scale_factor=scale_factor)
1348
1349


1350
class ScaledSoftmaxBwdPrimitive(SoftmaxPrimitive):
1351
    """
1352
    Scaled Softmax Bwd Primitive
1353
    """
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
    name = "te_scaled_softmax_backward"
    multiple_results = False
    impl_static_args = (2,)    # scale_factor
    inner_primitive = None
    outer_primitive = None

    @staticmethod
    def is_kernel_available(batch: int, heads: int, q_seqlen: int, k_seqlen: int,
                            dtype: jnp.dtype) -> bool:
        """Check Softmax kernel availability based on size"""
        return ScaledSoftmaxFwdPrimitive.is_kernel_available(batch, heads, q_seqlen, k_seqlen,
                                                             dtype)
1366
1367

    @staticmethod
1368
    def abstract(dz_aval, softmax_out_aval, scale_factor):
1369
        """
1370
        te_scaled_softmax_backward abstract
1371
        """
1372
        return SoftmaxPrimitive.backward_abstract(dz_aval, softmax_out_aval, scale_factor)
1373

1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
    @staticmethod
    def lowering(ctx, dz, softmax_out, *, scale_factor):
        """
        te_scaled_softmax_backward lowering rules
        """
        out = SoftmaxPrimitive.backward_lowering(ScaledSoftmaxBwdPrimitive.name,
                                                 ctx,
                                                 dz,
                                                 softmax_out,
                                                 scale_factor=scale_factor)
1384

1385
        return out
1386

1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
    @staticmethod
    def impl(dz, softmax_out, scale_factor):
        return SoftmaxPrimitive.backward_impl(ScaledSoftmaxBwdPrimitive.inner_primitive,
                                              dz,
                                              softmax_out,
                                              scale_factor=scale_factor)

    @staticmethod
    def batcher(batched_args, batch_dims, *, scale_factor):
        _check_valid_batch_dims(batch_dims)
        return SoftmaxPrimitive.backward_batcher(ScaledSoftmaxBwdPrimitive.outer_primitive,
                                                 batched_args,
                                                 batch_dims,
                                                 scale_factor=scale_factor)

    @staticmethod
    def infer_sharding_from_operands(scale_factor, mesh, arg_infos, result_infos):
1404
1405
1406
        return ScaledSoftmaxBwdPrimitive.backward_infer_sharding_from_operands(
            scale_factor, mesh, arg_infos, result_infos
        )
1407
1408

    @staticmethod
1409
    def partition(scale_factor, mesh, arg_infos, result_infos):
1410
1411
1412
        return ScaledSoftmaxBwdPrimitive.backward_partition(
            ScaledSoftmaxBwdPrimitive.impl, scale_factor, mesh, arg_infos, result_infos
        )
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458


register_primitive(ScaledSoftmaxBwdPrimitive)


def scaled_softmax_bwd(dz: jnp.ndarray, softmax_out: jnp.ndarray,
                       scale_factor: float) -> jnp.ndarray:
    """
    scaled_backward wrapper
    Return FP16/BF16 tensor
    """
    return ScaledSoftmaxBwdPrimitive.outer_primitive.bind(dz,
                                                          softmax_out,
                                                          scale_factor=scale_factor)


class ScaledMaskedSoftmaxFwdPrimitive(SoftmaxPrimitive):
    """
    Scaled Masked Softmax Fwd Primitive
    """
    name = "te_scaled_masked_softmax_forward"
    multiple_results = False
    impl_static_args = (2,)    # scale_factor
    inner_primitive = None
    outer_primitive = None

    @staticmethod
    def is_kernel_available(batch: int, heads: int, q_seqlen: int, k_seqlen: int,
                            dtype: jnp.dtype) -> bool:
        """Check Softmax kernel availability based on size"""
        attn_batches = batch * heads

        dtype = dtypes.canonicalize_dtype(dtype)
        if (dtype in [jnp.float16, jnp.bfloat16]
                and 16 < k_seqlen <= SoftmaxPrimitive.max_k_seqlen_supported
        # k_seqlen must be 16 ~ 4096
                and q_seqlen % 4 == 0    # q_seqlen must be divisor of 4
                and attn_batches % 4 == 0    # batch * heads must be divisor of 4
           ):
            if 0 <= k_seqlen <= SoftmaxPrimitive.max_k_seqlen_supported:
                batch_per_block = SoftmaxPrimitive.get_batch_per_block(k_seqlen)
                return q_seqlen % batch_per_block == 0
        return False

    @staticmethod
    def abstract(logits_aval, mask_aval, scale_factor):    # pylint: disable=unused-argument
1459
        """
1460
        te_scaled_masked_softmax_forward abstract
1461
1462
        """

1463
1464
1465
        i_dtype = dtypes.canonicalize_dtype(logits_aval.dtype)
        assert i_dtype in [jnp.float16, jnp.bfloat16]
        i_shape = logits_aval.shape
1466

1467
1468
1469
1470
1471
1472
        # Assume [...Batch, Head, Q_Seqlen, K_Seqlen]
        batch = reduce(operator.mul, i_shape[:-3])
        q_seqlen = i_shape[-2]
        k_seqlen = i_shape[-1]
        assert k_seqlen <= SoftmaxPrimitive.max_k_seqlen_supported
        assert q_seqlen > 1
1473

1474
1475
1476
        mask_dtype = dtypes.canonicalize_dtype(mask_aval.dtype)
        assert mask_dtype in [
            jnp.uint8,
1477
        ]
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
        mask_shape = mask_aval.shape
        pad_batch = batch = reduce(operator.mul, mask_shape[:-3])
        assert pad_batch in (1, batch)    # 1 means broadcast
        assert mask_shape[-3] == 1    # 1 means broadcast
        assert mask_shape[-2] == q_seqlen
        assert mask_shape[-1] == k_seqlen

        out_aval = core.raise_to_shaped(logits_aval)
        return out_aval

    @staticmethod
    def lowering(ctx, logits, mask, *, scale_factor):
        """
        te_scaled_masked_softmax_forward lowering rules
        """

        logits_aval, _ = ctx.avals_in
        i_type = ir.RankedTensorType(logits.type)
        i_shape = i_type.shape
        # Assume [...Batch, Head, Q_Seqlen, K_Seqlen]
        batch = reduce(operator.mul, i_shape[:-3])
        heads = i_shape[-3]
        q_seqlen = i_shape[-2]
        k_seqlen = i_shape[-1]

        mask_type = ir.RankedTensorType(mask.type)
        mask_shape = mask_type.shape
        pad_batch = reduce(operator.mul, mask_shape[:-3])

        out_types = [ir.RankedTensorType.get(i_shape, i_type.element_type)]
        operands = [logits, mask]
        operand_shapes = [i_shape, mask_shape]
1510
1511
        args = CustomCallArgsWrapper(out_types, operands, operand_shapes)

1512
1513
1514
        opaque = transformer_engine_jax.pack_softmax_descriptor(
            batch, pad_batch, heads, q_seqlen, k_seqlen, jax_dtype_to_te_dtype(logits_aval.dtype),
            scale_factor)
1515

1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
        out = custom_caller(ScaledMaskedSoftmaxFwdPrimitive.name, args, opaque, False)

        return [out]

    @staticmethod
    def impl(logits, mask, scale_factor):
        assert ScaledMaskedSoftmaxFwdPrimitive.inner_primitive is not None
        output = ScaledMaskedSoftmaxFwdPrimitive.inner_primitive.bind(logits,
                                                                      mask,
                                                                      scale_factor=scale_factor)
        return output

    @staticmethod
    def batcher(batched_args, batch_dims, *, scale_factor):
        _check_valid_batch_dims(batch_dims)
        assert ScaledMaskedSoftmaxFwdPrimitive.outer_primitive is not None
        logits, mask = batched_args
        logits_bdim, _ = batch_dims

        out_bdims = logits_bdim
        return ScaledMaskedSoftmaxFwdPrimitive.outer_primitive.bind(
            logits, mask, scale_factor=scale_factor), out_bdims

    @staticmethod
    def infer_sharding_from_operands(scale_factor, mesh, arg_infos, result_infos):
1541
1542
1543
        return ScaledMaskedSoftmaxFwdPrimitive.forward_infer_sharding_from_operands(
            scale_factor, mesh, arg_infos,result_infos
        )
1544
1545
1546

    @staticmethod
    def partition(scale_factor, mesh, arg_infos, result_infos):
1547
1548
1549
        return ScaledMaskedSoftmaxFwdPrimitive.backward_partition(
            ScaledMaskedSoftmaxFwdPrimitive.impl, scale_factor, mesh, arg_infos, result_infos
        )
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619


register_primitive(ScaledMaskedSoftmaxFwdPrimitive)


def scaled_masked_softmax_fwd(logits: jnp.ndarray, mask: jnp.ndarray,
                              scale_factor: float) -> jnp.ndarray:
    """
    scaled_masked_softmax_forward wrapper
    Return FP16/BF16 tensor
    """
    return ScaledMaskedSoftmaxFwdPrimitive.outer_primitive.bind(logits,
                                                                mask,
                                                                scale_factor=scale_factor)


class ScaledMaskedSoftmaxBwdPrimitive(SoftmaxPrimitive):
    """
    Scaled Masked Softmax Bwd Primitive
    """
    name = "te_scaled_masked_softmax_backward"
    multiple_results = False
    impl_static_args = (2,)    # scale_factor
    inner_primitive = None
    outer_primitive = None

    @staticmethod
    def is_kernel_available(batch: int, heads: int, q_seqlen: int, k_seqlen: int,
                            dtype: jnp.dtype) -> bool:
        """Check Softmax kernel availability based on size"""
        return ScaledSoftmaxFwdPrimitive.is_kernel_available(batch, heads, q_seqlen, k_seqlen,
                                                             dtype)

    @staticmethod
    def abstract(dz_aval, softmax_out_aval, *, scale_factor):
        """
        te_scaled_upper_triang_masked_backward abstract
        """
        return SoftmaxPrimitive.backward_abstract(dz_aval, softmax_out_aval, scale_factor)

    @staticmethod
    def lowering(ctx, dz, softmax_out, *, scale_factor):
        """
        te_scaled_upper_triang_masked_backward lowering rules
        """
        out = SoftmaxPrimitive.backward_lowering(ScaledMaskedSoftmaxBwdPrimitive.name,
                                                 ctx,
                                                 dz,
                                                 softmax_out,
                                                 scale_factor=scale_factor)

        return out

    @staticmethod
    def impl(dz, softmax_out, scale_factor):
        return SoftmaxPrimitive.backward_impl(ScaledMaskedSoftmaxBwdPrimitive.inner_primitive,
                                              dz,
                                              softmax_out,
                                              scale_factor=scale_factor)

    @staticmethod
    def batcher(batched_args, batch_dims, *, scale_factor):
        _check_valid_batch_dims(batch_dims)
        return SoftmaxPrimitive.backward_batcher(ScaledMaskedSoftmaxBwdPrimitive.outer_primitive,
                                                 batched_args,
                                                 batch_dims,
                                                 scale_factor=scale_factor)

    @staticmethod
    def infer_sharding_from_operands(scale_factor, mesh, arg_infos, result_infos):
1620
1621
1622
        return ScaledMaskedSoftmaxBwdPrimitive.backward_infer_sharding_from_operands(
            scale_factor, mesh, arg_infos, result_infos
        )
1623
1624
1625

    @staticmethod
    def partition(scale_factor, mesh, arg_infos, result_infos):
1626
1627
1628
        return ScaledMaskedSoftmaxBwdPrimitive.backward_partition(
            ScaledMaskedSoftmaxBwdPrimitive.impl, scale_factor, mesh, arg_infos, result_infos
        )
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677


register_primitive(ScaledMaskedSoftmaxBwdPrimitive)


def scaled_masked_softmax_bwd(dz: jnp.ndarray, softmax_out: jnp.ndarray,
                              scale_factor: float) -> jnp.ndarray:
    """
    scaled_masked_backward wrapper
    Return FP16/BF16 tensor
    """
    return ScaledMaskedSoftmaxBwdPrimitive.outer_primitive.bind(dz,
                                                                softmax_out,
                                                                scale_factor=scale_factor)


class ScaledUpperTriangMaskedSoftmaxFwdPrimitive(SoftmaxPrimitive):
    """
    Scaled Upper Triang Masked Softmax Fwd Primitive
    """
    name = "te_scaled_upper_triang_masked_softmax_forward"
    multiple_results = False
    impl_static_args = (1,)    # scale_factor
    inner_primitive = None
    outer_primitive = None

    @staticmethod
    def is_kernel_available(batch: int, heads: int, q_seqlen: int, k_seqlen: int,
                            dtype: jnp.dtype) -> bool:
        """Check Softmax kernel availability based on size"""
        attn_batches = batch * heads

        dtype = dtypes.canonicalize_dtype(dtype)
        if (dtype in [jnp.float16, jnp.bfloat16]
                and 16 < k_seqlen <= SoftmaxPrimitive.max_k_seqlen_supported
        # k_seqlen must be 16 ~ 4096
                and q_seqlen % 4 == 0    # q_seqlen must be divisor of 4
                and attn_batches % 4 == 0    # batch * heads must be divisor of 4
           ):
            if 0 <= k_seqlen <= SoftmaxPrimitive.max_k_seqlen_supported:
                batch_per_block = SoftmaxPrimitive.get_batch_per_block(k_seqlen)
                return attn_batches % batch_per_block == 0
        return False

    @staticmethod
    def abstract(logits_aval, scale_factor):    # pylint: disable=unused-argument
        """
        te_scaled_upper_triang_masked_softmax_forward abstract
        """
1678
1679
        q_seqlen = logits_aval.shape[-2]
        k_seqlen = logits_aval.shape[-1]
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
        assert q_seqlen == k_seqlen
        return SoftmaxPrimitive.forward_abstract(logits_aval, scale_factor)

    @staticmethod
    def lowering(ctx, logits, *, scale_factor):
        """
        te_scaled_upper_triang_masked_softmax_forward lowering rules
        """
        return SoftmaxPrimitive.forward_lowering(ScaledUpperTriangMaskedSoftmaxFwdPrimitive.name,
                                                 ctx,
                                                 logits,
                                                 scale_factor=scale_factor)

    @staticmethod
    def impl(logits, scale_factor):
        return SoftmaxPrimitive.forward_impl(
            ScaledUpperTriangMaskedSoftmaxFwdPrimitive.inner_primitive, logits, scale_factor)

    @staticmethod
    def batcher(batched_args, batch_dims, *, scale_factor):
        _check_valid_batch_dims(batch_dims)
        return SoftmaxPrimitive.forward_batcher(
            ScaledUpperTriangMaskedSoftmaxFwdPrimitive.outer_primitive,
            batched_args,
            batch_dims,
            scale_factor=scale_factor)

    @staticmethod
    def infer_sharding_from_operands(scale_factor, mesh, arg_infos, result_infos):
1709
1710
1711
        return ScaledUpperTriangMaskedSoftmaxFwdPrimitive.forward_infer_sharding_from_operands(
            scale_factor, mesh, arg_infos, result_infos
        )
1712
1713
1714

    @staticmethod
    def partition(scale_factor, mesh, arg_infos, result_infos):
1715
1716
1717
1718
        return ScaledUpperTriangMaskedSoftmaxFwdPrimitive.forward_partition(
            ScaledUpperTriangMaskedSoftmaxFwdPrimitive.impl, scale_factor, mesh,
            arg_infos, result_infos
        )
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766


register_primitive(ScaledUpperTriangMaskedSoftmaxFwdPrimitive)


def scaled_upper_triang_masked_softmax_fwd(logits: jnp.ndarray, scale_factor: float) -> jnp.ndarray:
    """
    scaled_upper_triang_masked_softmax_forward wrapper
    Return FP16/BF16 tensor
    """
    return ScaledUpperTriangMaskedSoftmaxFwdPrimitive.outer_primitive.bind(
        logits, scale_factor=scale_factor)


class ScaledUpperTriangMaskedSoftmaxBwdPrimitive(SoftmaxPrimitive):
    """
    Scaled Upper Triang Masked Softmax Bwd Primitive
    """
    name = "te_scaled_upper_triang_masked_softmax_backward"
    multiple_results = False
    impl_static_args = (2,)    # scale_factor
    inner_primitive = None
    outer_primitive = None

    @staticmethod
    def is_kernel_available(batch: int, heads: int, q_seqlen: int, k_seqlen: int,
                            dtype: jnp.dtype) -> bool:
        """Check Softmax kernel availability based on size"""
        return ScaledUpperTriangMaskedSoftmaxFwdPrimitive.is_kernel_available(
            batch, heads, q_seqlen, k_seqlen, dtype)

    @staticmethod
    def abstract(dz_aval, softmax_out_aval, *, scale_factor):
        """
        te_scaled_upper_triang_masked_backward abstract
        """
        return SoftmaxPrimitive.backward_abstract(dz_aval, softmax_out_aval, scale_factor)

    @staticmethod
    def lowering(ctx, dz, softmax_out, *, scale_factor):
        """
        te_scaled_upper_triang_masked_backward lowering rules
        """
        out = SoftmaxPrimitive.backward_lowering(ScaledUpperTriangMaskedSoftmaxBwdPrimitive.name,
                                                 ctx,
                                                 dz,
                                                 softmax_out,
                                                 scale_factor=scale_factor)
1767
1768
1769

        return out

1770
1771
1772
1773
1774
1775
1776
    @staticmethod
    def impl(dz, softmax_out, scale_factor):
        return SoftmaxPrimitive.backward_impl(
            ScaledUpperTriangMaskedSoftmaxBwdPrimitive.inner_primitive,
            dz,
            softmax_out,
            scale_factor=scale_factor)
1777

1778
1779
1780
1781
1782
1783
1784
1785
    @staticmethod
    def batcher(batched_args, batch_dims, *, scale_factor):
        _check_valid_batch_dims(batch_dims)
        return SoftmaxPrimitive.backward_batcher(
            ScaledUpperTriangMaskedSoftmaxBwdPrimitive.outer_primitive,
            batched_args,
            batch_dims,
            scale_factor=scale_factor)
1786

1787
1788
    @staticmethod
    def infer_sharding_from_operands(scale_factor, mesh, arg_infos, result_infos):
1789
1790
1791
        return ScaledUpperTriangMaskedSoftmaxBwdPrimitive.backward_infer_sharding_from_operands(
            scale_factor, mesh, arg_infos, result_infos
        )
1792

1793
1794
    @staticmethod
    def partition(scale_factor, mesh, arg_infos, result_infos):
1795
1796
1797
1798
        return ScaledUpperTriangMaskedSoftmaxBwdPrimitive.backward_partition(
            ScaledUpperTriangMaskedSoftmaxBwdPrimitive.impl, scale_factor, mesh,
            arg_infos, result_infos
        )
1799
1800
1801
1802
1803
1804
1805


register_primitive(ScaledUpperTriangMaskedSoftmaxBwdPrimitive)


def scaled_upper_triang_masked_softmax_bwd(dz: jnp.ndarray, softmax_out: jnp.ndarray,
                                           scale_factor: float) -> jnp.ndarray:
1806
    """
1807
1808
    scaled_upper_triang_masked_backward wrapper
    Return FP16/BF16 tensor
1809
    """
1810
1811
    return ScaledUpperTriangMaskedSoftmaxBwdPrimitive.outer_primitive.bind(
        dz, softmax_out, scale_factor=scale_factor)
1812
1813


1814
1815
@dataclass(frozen=True)
class FusedAttnHelper:
1816
    """
1817
    Helper for the fused attention backend
1818
    """
1819
1820
1821
1822
1823
1824
1825

    q_type: jnp.dtype
    kv_type: jnp.dtype
    qkv_layout: NVTE_QKV_Layout
    attn_bias_type: NVTE_Bias_Type
    attn_mask_type: NVTE_Mask_Type
    dropout_probability: float
1826
1827
    num_heads_q: int
    num_heads_kv: int
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
    max_seqlen_q: int
    max_seqlen_kv: int
    head_dim: int

    def is_fused_attn_kernel_available(self):
        """Check if there is available fused attention kernel"""
        return self.get_fused_attn_backend() != NVTE_Fused_Attn_Backend.NVTE_No_Backend

    def get_fused_attn_backend(self):
        """Get the fused attention kernel backend"""
zlsh80826's avatar
zlsh80826 committed
1838
1839
1840
1841
1842
        return transformer_engine_jax.get_fused_attn_backend(
            jax_dtype_to_te_dtype(self.q_type), jax_dtype_to_te_dtype(self.kv_type),
            self.qkv_layout, self.attn_bias_type, self.attn_mask_type, self.dropout_probability,
            self.num_heads_q, self.num_heads_kv, self.max_seqlen_q, self.max_seqlen_kv,
            self.head_dim)
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884


@dataclass(frozen=True)
class _FusedAttnRNGStateChecker:
    """
    Checker for guarding the fused attention rng state.
    The fused attention backend requires a 64 bits seed and a 64 bits offset.
    However, JAX doesn't enable 64 bits by default,
    so we have to emulate seed as two 32 bits array.
    The offset calculation is maintained in the backend.
    """
    rng_state_dtype: jnp.dtype = jnp.uint32
    # (seed,) with internal dtype int64
    seed_size: int = 2
    # (seed, offset) with internal dtype int64
    rng_state_size: int = 2 * 2

    def check_seed(self, seed, dropout_probability, is_training):
        """
        Check the seed and convert the data type of seed if possible.
        """
        # Jax can't bind None, create a dummy tensor for None
        if seed is None:
            dropout_enabled = dropout_probability > 0 and is_training
            assert not dropout_enabled, "seed is not allowed to be None when dropout is enabled."
            seed = jnp.zeros(2, dtype=self.rng_state_dtype)
            seed = jnp.repeat(seed, num_of_devices())

        if seed.dtype != self.rng_state_dtype:
            warnings.warn(
                f"Requested {seed.dtype=} is not available, and will be "
                f"casted to dtype {self.rng_state_dtype}. "
                f"Please use threefry/rbg/unsafe_rbg PRNG implementations to remove this warning.")
            seed = seed.astype(self.rng_state_dtype)

        assert seed.dtype == self.rng_state_dtype
        # Backend takes an int64_t seed, so only the first two u32 elements are taken
        assert seed.size >= self.seed_size

        return seed


zlsh80826's avatar
zlsh80826 committed
1885
def generate_cu_seqlen(actual_seqlen):
1886
1887
1888
    """
    Generating cumsum seqlen for a batch
    """
zlsh80826's avatar
zlsh80826 committed
1889
    cu_seqlen = jnp.cumsum(actual_seqlen)
1890
1891
1892
1893
1894
1895
1896
1897
1898
    cu_seqlen = jnp.hstack((0, cu_seqlen))
    return cu_seqlen


class SelfFusedAttnFwdPrimitive(BasePrimitive):
    """
    Self Fused Attention Forward Primitive
    """
    name = "te_self_fused_attn_forward"
1899
    multiple_results = True
1900
1901
1902
    impl_static_args = (4, 5, 6, 7, 8)
    inner_primitive = None
    outer_primitive = None
1903
1904

    @staticmethod
zlsh80826's avatar
zlsh80826 committed
1905
    def abstract(qkv_aval, bias_aval, seqlen_or_cu_seqlen_aval, seed_aval, *, attn_bias_type,
1906
                 attn_mask_type, scaling_factor, dropout_probability, is_training):
1907
        """
1908
        Self fused attention fwd inner primitive abstract
1909
        """
1910
1911
        # outer_primitve is squeezed_mask, inner_primitive is cu_seqlen
        del seqlen_or_cu_seqlen_aval
1912
        qkv_dtype = dtypes.canonicalize_dtype(qkv_aval.dtype)
1913
        *input_batch_shape, max_seqlen, nqkv, attn_heads, head_dim = qkv_aval.shape
1914
1915
        assert nqkv == 3
        assert qkv_aval.dtype == bias_aval.dtype
1916

1917
        output_shape = (*input_batch_shape, max_seqlen, attn_heads, head_dim)
1918
        out_aval = qkv_aval.update(shape=output_shape, dtype=qkv_dtype)
1919

1920
        # backend determines the softmax buffer shape/dtype
1921
        backend = FusedAttnHelper(qkv_dtype, qkv_dtype, NVTE_QKV_Layout.NVTE_BS3HD, attn_bias_type,
1922
                                  attn_mask_type, dropout_probability, attn_heads, attn_heads,
1923
                                  max_seqlen, max_seqlen, head_dim).get_fused_attn_backend()
1924

1925
        if backend == NVTE_Fused_Attn_Backend.NVTE_F16_max512_seqlen:
1926
            softmax_shape = (*input_batch_shape, attn_heads, max_seqlen, max_seqlen)
1927
1928
            softmax_dtype = qkv_dtype
        elif backend == NVTE_Fused_Attn_Backend.NVTE_F16_arbitrary_seqlen:
1929
            softmax_shape = (*input_batch_shape, attn_heads, max_seqlen, 1)
1930
1931
            softmax_dtype = dtypes.canonicalize_dtype(jnp.float32)
        else:
1932
            raise ValueError(f'Unsupported {backend=}')
1933
        softmax_aux_aval = qkv_aval.update(shape=softmax_shape, dtype=softmax_dtype)
1934

1935
1936
        # JAX does not enable 64-bit int by default so we get XLA to allocate x8 memory with
        # 32-bit unsigned int to get the buffer size we need in the C++ kernel
1937
1938
1939
1940
        checker = _FusedAttnRNGStateChecker()
        seed_dtype = dtypes.canonicalize_dtype(seed_aval.dtype)
        assert seed_dtype == checker.rng_state_dtype
        rng_state_shape = (seed_aval.shape[0], checker.rng_state_size)
1941
1942
        rng_state_aval = seed_aval.update(shape=rng_state_shape, dtype=checker.rng_state_dtype)

1943
1944
1945
1946
1947
1948
        if attn_bias_type == NVTE_Bias_Type.NVTE_NO_BIAS:
            bias_batch = bias_heads = 0
        else:
            *bias_batch_shape, bias_heads, _, _ = bias_aval.shape
            bias_batch = reduce(operator.mul, bias_batch_shape)

1949
1950
        # do a dummy kernel call here to get workspace buffer shapes/dtypes that XLA needs to
        # prepare for the active fused-attn backend
1951
        input_batch = reduce(operator.mul, input_batch_shape)
1952
        wkspace_info = transformer_engine_jax.get_self_fused_attn_fwd_workspace_sizes(
1953
1954
1955
            input_batch, bias_batch, max_seqlen, attn_heads, bias_heads, head_dim,
            scaling_factor, dropout_probability, attn_bias_type, attn_mask_type,
            jax_dtype_to_te_dtype(qkv_aval.dtype), is_training)
1956
1957
1958
1959
        wkspace_aval = qkv_aval.update(shape=wkspace_info[0],
                                       dtype=te_dtype_to_jax_dtype(wkspace_info[1]))

        return out_aval, softmax_aux_aval, rng_state_aval, wkspace_aval
1960

1961
1962
1963
1964
1965
1966
1967
    @staticmethod
    def outer_abstract(*args, **kwargs):
        """
        Self fused attention fwd outer primitive abstract
        """
        out_aval, softmax_aux_aval, rng_state_aval, _ = \
            SelfFusedAttnFwdPrimitive.abstract(*args, **kwargs)
1968
        return out_aval, softmax_aux_aval, rng_state_aval
1969
1970

    @staticmethod
1971
1972
    def lowering(ctx, qkv, bias, cu_seqlen, seed, *, attn_bias_type, attn_mask_type, scaling_factor,
                 dropout_probability, is_training):
1973
        """
1974
        Self fused attention fwd lowering rules
1975
        """
1976
1977
        operands = [qkv, bias, cu_seqlen, seed]
        operand_shapes = map(lambda x: x.type.shape, operands)
1978
        out_types = [
1979
1980
            ir.RankedTensorType.get(output.shape, mlir.dtype_to_ir_type(output.dtype))
            for output in ctx.avals_out
1981
1982
        ]
        args = CustomCallArgsWrapper(out_types, operands, operand_shapes)
1983

1984
1985
1986
1987
1988
1989
1990
1991
1992
        qkv_aval, bias_aval, *_ = ctx.avals_in
        *input_batch_shape, max_seqlen, _, attn_heads, head_dim = qkv_aval.shape
        input_batch = reduce(operator.mul, input_batch_shape)

        if attn_bias_type == NVTE_Bias_Type.NVTE_NO_BIAS:
            bias_batch = bias_heads = 0
        else:
            *bias_batch_shape, bias_heads, _, _ = bias_aval.shape
            bias_batch = reduce(operator.mul, bias_batch_shape)
1993
1994
1995

        wkspace_aval = ctx.avals_out[-1]

1996
        opaque = transformer_engine_jax.pack_fused_attn_descriptor(
1997
1998
            input_batch, bias_batch, max_seqlen, max_seqlen,
            attn_heads, attn_heads, bias_heads, head_dim, wkspace_aval.size,
1999
2000
2001
            scaling_factor, dropout_probability, attn_bias_type, attn_mask_type,
            jax_dtype_to_te_dtype(qkv_aval.dtype), jax_dtype_to_te_dtype(wkspace_aval.dtype),
            is_training)
2002

2003
        out = custom_caller(SelfFusedAttnFwdPrimitive.name, args, opaque, has_side_effect=False)
2004

2005
2006
2007
        return out

    @staticmethod
zlsh80826's avatar
zlsh80826 committed
2008
    def impl(qkv, bias, seqlen, seed, attn_bias_type, attn_mask_type, scaling_factor,
2009
2010
             dropout_probability, is_training):
        assert SelfFusedAttnFwdPrimitive.inner_primitive is not None
2011

zlsh80826's avatar
zlsh80826 committed
2012
        cu_seqlen = generate_cu_seqlen(seqlen)
2013

2014
        output, softmax_aux, rng_state, _ = SelfFusedAttnFwdPrimitive.inner_primitive.bind(
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
            qkv,
            bias,
            cu_seqlen,
            seed,
            attn_bias_type=attn_bias_type,
            attn_mask_type=attn_mask_type,
            scaling_factor=scaling_factor,
            dropout_probability=dropout_probability,
            is_training=is_training)
        return output, softmax_aux, rng_state
2025

2026
2027
2028
2029
2030
2031
2032
2033
2034
    @staticmethod
    def batcher(batched_args, batch_dims, *, attn_bias_type, attn_mask_type, scaling_factor,
                dropout_probability, is_training):
        _check_valid_batch_dims(batch_dims)
        assert SelfFusedAttnFwdPrimitive.outer_primitive is not None
        qkv_bdim, _, _, seed_bdim = batch_dims

        out_bdims = qkv_bdim, qkv_bdim, seed_bdim
        return SelfFusedAttnFwdPrimitive.outer_primitive.bind(
2035
            *batched_args,
2036
2037
2038
2039
2040
            attn_bias_type=attn_bias_type,
            attn_mask_type=attn_mask_type,
            scaling_factor=scaling_factor,
            dropout_probability=dropout_probability,
            is_training=is_training), out_bdims
2041

2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
    @staticmethod
    def infer_sharding_from_operands(attn_bias_type, attn_mask_type, scaling_factor,
                                     dropout_probability, is_training, mesh, arg_infos,
                                     result_infos):
        del attn_bias_type, attn_mask_type, scaling_factor
        del dropout_probability, is_training, result_infos
        x_spec = get_padded_spec(arg_infos[0])    # (...batch, seqlen, 3, head, hidden)
        out_sharding = NamedSharding(mesh, PartitionSpec(*x_spec[:-3], *x_spec[-2:]))
        softmax_aux_sharding = NamedSharding(
            mesh, PartitionSpec(*x_spec[:-4], x_spec[-2], x_spec[-4], None))
        rng_state_sharding = NamedSharding(mesh, PartitionSpec(get_all_mesh_axes(), None))
        return (out_sharding, softmax_aux_sharding, rng_state_sharding)
2054

2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
    @staticmethod
    def partition(attn_bias_type, attn_mask_type, scaling_factor, dropout_probability, is_training,
                  mesh, arg_infos, result_infos):
        del result_infos
        x_spec = get_padded_spec(arg_infos[0])    # (...batch, seqlen, 3, head, hidden)
        out_sharding = NamedSharding(mesh, PartitionSpec(*x_spec[:-3], *x_spec[-2:]))
        softmax_aux_sharding = NamedSharding(
            mesh, PartitionSpec(*x_spec[:-4], x_spec[-2], x_spec[-4], None))
        rng_state_sharding = NamedSharding(mesh, PartitionSpec(get_all_mesh_axes(), None))
        arg_shardings = tuple([arg_i.sharding for arg_i in arg_infos[:-1]] + [rng_state_sharding])
        out_shardings = (out_sharding, softmax_aux_sharding, rng_state_sharding)
        impl = partial(SelfFusedAttnFwdPrimitive.impl,
                       attn_bias_type=attn_bias_type,
                       attn_mask_type=attn_mask_type,
                       scaling_factor=scaling_factor,
                       dropout_probability=dropout_probability,
                       is_training=is_training)
        return mesh, impl, out_shardings, arg_shardings


register_primitive(SelfFusedAttnFwdPrimitive)


zlsh80826's avatar
zlsh80826 committed
2078
2079
2080
def self_fused_attn_fwd(qkv: jnp.ndarray, bias: jnp.ndarray, seqlen: jnp.ndarray, seed: jnp.ndarray,
                        attn_bias_type: NVTE_Bias_Type, attn_mask_type: NVTE_Mask_Type,
                        scaling_factor: float, dropout_probability: float, is_training: bool):
2081
    """
2082
    Wrapper for TE self fused attention fwd
2083
    Return BMM1 -> (PreScaleBias) -> Scale -> (PostScaleBias) -> Softmax -> (Dropout) -> BMM2
2084
    """
2085
2086
2087
2088
2089
2090
    checker = _FusedAttnRNGStateChecker()
    seed = checker.check_seed(seed, dropout_probability, is_training)

    if attn_bias_type == NVTE_Bias_Type.NVTE_NO_BIAS:
        assert bias is None
        bias = jnp.zeros(0, dtype=qkv.dtype)
2091

2092
2093
    return SelfFusedAttnFwdPrimitive.outer_primitive.bind(qkv,
                                                          bias,
zlsh80826's avatar
zlsh80826 committed
2094
                                                          seqlen,
2095
2096
2097
2098
2099
2100
                                                          seed,
                                                          attn_bias_type=attn_bias_type,
                                                          attn_mask_type=attn_mask_type,
                                                          scaling_factor=scaling_factor,
                                                          dropout_probability=dropout_probability,
                                                          is_training=is_training)
2101
2102


2103
class SelfFusedAttnBwdPrimitive(BasePrimitive):
2104
    """
2105
    Self Fused Attention Backward Primitive
2106
    """
2107
    name = "te_self_fused_attn_backward"
2108
    multiple_results = True
2109
    impl_static_args = (7, 8, 9, 10, 11)
2110
2111
    inner_primitive = None
    outer_primitive = None
2112
2113

    @staticmethod
2114
    def abstract(qkv_aval, bias_aval, softmax_aux_aval, rng_state_aval, output_aval, doutput_aval,
zlsh80826's avatar
zlsh80826 committed
2115
                 seqlen_or_cu_seqlen_aval, *, attn_bias_type, attn_mask_type, scaling_factor,
2116
                 dropout_probability, is_training):
2117
        """
2118
        Self fused attention bwd abstract
2119
        """
2120
2121
2122
        del softmax_aux_aval, rng_state_aval, seqlen_or_cu_seqlen_aval

        assert qkv_aval.dtype == bias_aval.dtype == output_aval.dtype == doutput_aval.dtype
2123
        *input_batch_shape, max_seqlen, nqkv, attn_heads, head_dim = qkv_aval.shape
2124
        assert nqkv == 3
2125
        qkv_dtype = dtypes.canonicalize_dtype(qkv_aval.dtype)
2126
        bias_dtype = dtypes.canonicalize_dtype(bias_aval.dtype)
2127

2128
2129
2130
2131
2132
2133
2134
        if attn_bias_type == NVTE_Bias_Type.NVTE_NO_BIAS:
            bias_batch = bias_heads = 0
        else:
            *bias_batch_shape, bias_heads, _, _ = bias_aval.shape
            bias_batch = reduce(operator.mul, bias_batch_shape)

        input_batch = reduce(operator.mul, input_batch_shape)
2135
2136
        wkspace_shape, wkspace_dtype = \
            transformer_engine_jax.get_self_fused_attn_bwd_workspace_sizes(
2137
                input_batch, bias_batch, max_seqlen, attn_heads, bias_heads, head_dim,
2138
2139
2140
                scaling_factor, dropout_probability, attn_bias_type, attn_mask_type,
                jax_dtype_to_te_dtype(qkv_aval.dtype), is_training
            )
2141

2142
        dqkv_aval = qkv_aval.update(shape=qkv_aval.shape, dtype=qkv_dtype)
2143
        dbias_aval = bias_aval.update(shape=bias_aval.shape, dtype=bias_dtype)
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153
2154
        wkspace_aval = qkv_aval.update(shape=wkspace_shape,
                                       dtype=te_dtype_to_jax_dtype(wkspace_dtype))

        return dqkv_aval, dbias_aval, wkspace_aval

    @staticmethod
    def outer_abstract(*args, **kwargs):
        """
        Self fused attention bwd outer primitive abstract
        """
        dqkv_aval, dbias_aval, _ = SelfFusedAttnBwdPrimitive.abstract(*args, **kwargs)
2155
        return dqkv_aval, dbias_aval
2156
2157

    @staticmethod
2158
2159
    def lowering(ctx, qkv, bias, softmax_aux, rng_state, output, doutput, cu_seqlen, *,
                 attn_bias_type, attn_mask_type, scaling_factor, dropout_probability, is_training):
2160
        """
2161
        Self fused attention bwd lowering rules
2162
        """
2163
        operands = [qkv, bias, softmax_aux, rng_state, output, doutput, cu_seqlen]
2164
        operand_shapes = map(lambda x: x.type.shape, operands)
2165
        out_types = [
2166
2167
            ir.RankedTensorType.get(output.shape, mlir.dtype_to_ir_type(output.dtype))
            for output in ctx.avals_out
2168
2169
2170
        ]
        args = CustomCallArgsWrapper(out_types, operands, operand_shapes)

2171
2172
2173
2174
2175
2176
2177
2178
2179
        qkv_aval, bias_aval, *_ = ctx.avals_in
        *input_batch_shape, max_seqlen, _, attn_heads, head_dim = qkv_aval.shape
        input_batch = reduce(operator.mul, input_batch_shape)

        if attn_bias_type == NVTE_Bias_Type.NVTE_NO_BIAS:
            bias_batch = bias_heads = 0
        else:
            *bias_batch_shape, bias_heads, _, _ = bias_aval.shape
            bias_batch = reduce(operator.mul, bias_batch_shape)
2180
2181
2182

        wkspace_aval = ctx.avals_out[-1]

2183
        opaque = transformer_engine_jax.pack_fused_attn_descriptor(
2184
2185
            input_batch, bias_batch, max_seqlen, max_seqlen,
            attn_heads, attn_heads, bias_heads, head_dim, wkspace_aval.size,
2186
2187
2188
            scaling_factor, dropout_probability, attn_bias_type, attn_mask_type,
            jax_dtype_to_te_dtype(qkv_aval.dtype), jax_dtype_to_te_dtype(wkspace_aval.dtype),
            is_training)
2189

2190
        out = custom_caller(SelfFusedAttnBwdPrimitive.name, args, opaque, has_side_effect=False)
2191
2192
2193

        return out

2194
    @staticmethod
zlsh80826's avatar
zlsh80826 committed
2195
    def impl(qkv, bias, softmax_aux, rng_state, output, doutput, seqlen, attn_bias_type,
2196
2197
2198
             attn_mask_type, scaling_factor, dropout_probability, is_training):
        assert SelfFusedAttnBwdPrimitive.inner_primitive is not None

zlsh80826's avatar
zlsh80826 committed
2199
        cu_seqlen = generate_cu_seqlen(seqlen)
2200

2201
        dqkv, dbias, _ = SelfFusedAttnBwdPrimitive.inner_primitive.bind(
2202
            qkv,
2203
            bias,
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
            softmax_aux,
            rng_state,
            output,
            doutput,
            cu_seqlen,
            attn_bias_type=attn_bias_type,
            attn_mask_type=attn_mask_type,
            scaling_factor=scaling_factor,
            dropout_probability=dropout_probability,
            is_training=is_training)
        return dqkv, dbias
2215

2216
2217
2218
2219
2220
2221
2222
2223
2224
    @staticmethod
    def batcher(batched_args, batch_dims, *, attn_bias_type, attn_mask_type, scaling_factor,
                dropout_probability, is_training):
        _check_valid_batch_dims(batch_dims)
        assert SelfFusedAttnBwdPrimitive.outer_primitive is not None
        qkv_bdim, *_ = batch_dims

        out_bdims = qkv_bdim, qkv_bdim
        return SelfFusedAttnBwdPrimitive.outer_primitive.bind(
2225
            *batched_args,
2226
2227
2228
2229
2230
            attn_bias_type=attn_bias_type,
            attn_mask_type=attn_mask_type,
            scaling_factor=scaling_factor,
            dropout_probability=dropout_probability,
            is_training=is_training), out_bdims
2231

2232
2233
2234
2235
    @staticmethod
    def infer_sharding_from_operands(attn_bias_type, attn_mask_type, scaling_factor,
                                     dropout_probability, is_training, mesh, arg_infos,
                                     result_infos):
2236
        del attn_bias_type, attn_mask_type, scaling_factor, dropout_probability,
2237
2238
        del is_training, result_infos
        x_spec = get_padded_spec(arg_infos[0])
2239
        bias_spec = get_padded_spec(arg_infos[1])
2240
        dx_sharding = NamedSharding(mesh, PartitionSpec(*x_spec))
2241
        dbias_sharding = NamedSharding(mesh, PartitionSpec(*bias_spec))
2242
2243
2244
2245
2246
2247
2248
        return (dx_sharding, dbias_sharding)

    @staticmethod
    def partition(attn_bias_type, attn_mask_type, scaling_factor, dropout_probability, is_training,
                  mesh, arg_infos, result_infos):
        del result_infos
        x_spec = get_padded_spec(arg_infos[0])
2249
        bias_spec = get_padded_spec(arg_infos[1])
2250
        dx_sharding = NamedSharding(mesh, PartitionSpec(*x_spec))
2251
        dbias_sharding = NamedSharding(mesh, PartitionSpec(*bias_spec))
2252
2253
2254
        arg_shardings = tuple(arg_i.sharding for arg_i in arg_infos)
        out_shardings = (dx_sharding, dbias_sharding)

2255
        def sharded_impl(qkv, bias, softmax_aux, rng_state, output, doutput, cu_seqlen):
2256
2257
            local_dx, local_dbias = SelfFusedAttnBwdPrimitive.impl(
                qkv,
2258
                bias,
2259
2260
2261
2262
2263
2264
2265
2266
2267
2268
2269
2270
2271
2272
2273
2274
2275
2276
2277
2278
2279
                softmax_aux,
                rng_state,
                output,
                doutput,
                cu_seqlen,
                attn_bias_type=attn_bias_type,
                attn_mask_type=attn_mask_type,
                scaling_factor=scaling_factor,
                dropout_probability=dropout_probability,
                is_training=is_training)
            global_dbias = local_dbias
            if attn_bias_type is not NVTE_Bias_Type.NVTE_NO_BIAS:
                global_dbias = all_reduce_sum_along_dp_fsdp(local_dbias)
            return local_dx, global_dbias

        return mesh, sharded_impl, out_shardings, arg_shardings


register_primitive(SelfFusedAttnBwdPrimitive)


2280
2281
def self_fused_attn_bwd(qkv: jnp.ndarray, bias: jnp.ndarray, softmax_aux: jnp.ndarray,
                        rng_state: jnp.ndarray, output: jnp.ndarray, doutput: jnp.ndarray,
zlsh80826's avatar
zlsh80826 committed
2282
                        seqlen: jnp.ndarray, attn_bias_type: NVTE_Bias_Type,
2283
2284
                        attn_mask_type: NVTE_Mask_Type, scaling_factor: float,
                        dropout_probability: float, is_training: bool):
2285
    """
2286
2287
    Wrapper for TE self fused attention bwd
    Return the gradients of self fused attention with packed qkv input
2288
    """
2289
2290
2291
    if attn_bias_type == NVTE_Bias_Type.NVTE_NO_BIAS:
        assert bias is None
        bias = jnp.zeros(0, dtype=qkv.dtype)
2292

2293
    return SelfFusedAttnBwdPrimitive.outer_primitive.bind(qkv,
2294
                                                          bias,
2295
2296
2297
2298
                                                          softmax_aux,
                                                          rng_state,
                                                          output,
                                                          doutput,
zlsh80826's avatar
zlsh80826 committed
2299
                                                          seqlen,
2300
2301
2302
2303
2304
                                                          attn_bias_type=attn_bias_type,
                                                          attn_mask_type=attn_mask_type,
                                                          scaling_factor=scaling_factor,
                                                          dropout_probability=dropout_probability,
                                                          is_training=is_training)
2305
2306


2307
class CrossFusedAttnFwdPrimitive(BasePrimitive):
2308
    """
2309
    Cross Fused Attention Forward Primitive
2310
    """
2311
    name = "te_cross_fused_attn_forward"
2312
    multiple_results = True
2313
    impl_static_args = (6, 7, 8, 9, 10)
2314
2315
    inner_primitive = None
    outer_primitive = None
2316
2317

    @staticmethod
zlsh80826's avatar
zlsh80826 committed
2318
2319
2320
    def abstract(q_aval, kv_aval, bias_aval, q_seqlen_or_cu_seqlen_aval,
                 kv_seqlen_or_cu_seqlen_aval, seed_aval, *, attn_bias_type, attn_mask_type,
                 scaling_factor, dropout_probability, is_training):
2321
        """
2322
        Cross fused attention fwd abstract
2323
        """
2324
2325
        q_dtype = dtypes.canonicalize_dtype(q_aval.dtype)
        kv_dtype = dtypes.canonicalize_dtype(kv_aval.dtype)
2326
2327
        bias_dtype = dtypes.canonicalize_dtype(bias_aval.dtype)
        assert q_dtype == kv_dtype == bias_dtype
2328
2329
        assert q_seqlen_or_cu_seqlen_aval.dtype == kv_seqlen_or_cu_seqlen_aval.dtype

2330
        *q_batch_shape, q_max_seqlen, attn_heads, q_head_dim = q_aval.shape
2331
        *kv_batch_shape, kv_max_seqlen, nkv, num_gqa_groups, kv_head_dim = kv_aval.shape
2332
2333
2334
        assert q_batch_shape == kv_batch_shape
        assert q_head_dim == kv_head_dim
        assert nkv == 2
2335
        out_aval = q_aval.update(shape=q_aval.shape, dtype=q_dtype)
2336

2337
        # backend determines the softmax buffer shape/dtype
2338
        backend = FusedAttnHelper(q_dtype, kv_dtype, NVTE_QKV_Layout.NVTE_BSHD_BS2HD,
2339
                                  attn_bias_type, attn_mask_type, dropout_probability, attn_heads,
2340
                                  num_gqa_groups, q_max_seqlen, kv_max_seqlen,
zlsh80826's avatar
zlsh80826 committed
2341
                                  q_head_dim).get_fused_attn_backend()
2342

2343
        if backend == NVTE_Fused_Attn_Backend.NVTE_F16_max512_seqlen:
2344
            softmax_shape = (*q_batch_shape, attn_heads, q_max_seqlen, kv_max_seqlen)
2345
            softmax_dtype = q_dtype
2346
        elif backend == NVTE_Fused_Attn_Backend.NVTE_F16_arbitrary_seqlen:
2347
            softmax_shape = (*q_batch_shape, attn_heads, q_max_seqlen, 1)
2348
            softmax_dtype = dtypes.canonicalize_dtype(jnp.float32)
2349
2350
        else:
            raise ValueError(f'Unsupported {backend=}')
2351
        softmax_aux_aval = q_aval.update(shape=softmax_shape, dtype=softmax_dtype)
2352

2353
2354
        # JAX does not enable 64-bit int by default so we get XLA to allocate x8 memory with
        # 32-bit unsigned int to get the buffer size we need in the C++ kernel
2355
2356
2357
2358
        checker = _FusedAttnRNGStateChecker()
        seed_dtype = dtypes.canonicalize_dtype(seed_aval.dtype)
        assert seed_dtype == checker.rng_state_dtype
        rng_state_shape = (seed_aval.shape[0], checker.rng_state_size)
2359
2360
        rng_state_aval = seed_aval.update(shape=rng_state_shape, dtype=checker.rng_state_dtype)

2361
2362
2363
2364
2365
2366
        if attn_bias_type == NVTE_Bias_Type.NVTE_NO_BIAS:
            bias_batch = bias_heads = 0
        else:
            *bias_batch_shape, bias_heads, _, _ = bias_aval.shape
            bias_batch = reduce(operator.mul, bias_batch_shape)

2367
2368
        # do a dummy kernel call here to get workspace buffer shapes/dtypes that XLA needs to
        # prepare for the active fused-attn backend
2369
        input_batch = reduce(operator.mul, q_batch_shape)
2370
        wkspace_info = transformer_engine_jax.get_cross_fused_attn_fwd_workspace_sizes(
2371
2372
            input_batch, bias_batch, q_max_seqlen, kv_max_seqlen,
            attn_heads, num_gqa_groups, bias_heads, q_head_dim,
2373
            scaling_factor, dropout_probability, attn_bias_type, attn_mask_type,
2374
            jax_dtype_to_te_dtype(q_aval.dtype), is_training)
2375
2376
        wkspace_aval = q_aval.update(shape=wkspace_info[0],
                                     dtype=te_dtype_to_jax_dtype(wkspace_info[1]))
2377

2378
        return out_aval, softmax_aux_aval, rng_state_aval, wkspace_aval
2379

2380
2381
2382
2383
2384
2385
2386
    @staticmethod
    def outer_abstract(*args, **kwargs):
        """
        Cross fused attention fwd outer primitive abstract
        """
        out_aval, softmax_aux_aval, rng_state_aval, _ = \
            CrossFusedAttnFwdPrimitive.abstract(*args, **kwargs)
2387
        return out_aval, softmax_aux_aval, rng_state_aval
2388
2389

    @staticmethod
2390
2391
    def lowering(ctx, q, kv, bias, q_cu_seqlen, kv_cu_seqlen, seed, *, attn_bias_type,
                 attn_mask_type, scaling_factor, dropout_probability, is_training):
2392
        """
2393
        Cross fused attention fwd lowering rules
2394
        """
2395
        operands = [q, kv, bias, q_cu_seqlen, kv_cu_seqlen, seed]
2396
        operand_shapes = map(lambda x: x.type.shape, operands)
2397
        out_types = [
2398
2399
            ir.RankedTensorType.get(output.shape, mlir.dtype_to_ir_type(output.dtype))
            for output in ctx.avals_out
2400
        ]
2401
        args = CustomCallArgsWrapper(out_types, operands, operand_shapes)
2402

2403
2404
        q_aval, kv_aval, bias_aval, *_ = ctx.avals_in
        *input_batch_shape, q_max_seqlen, attn_heads, head_dim = q_aval.shape
2405
        *_, kv_max_seqlen, _, num_gqa_groups, _ = kv_aval.shape
2406
2407
2408
2409
2410
2411
2412
        input_batch = reduce(operator.mul, input_batch_shape)

        if attn_bias_type == NVTE_Bias_Type.NVTE_NO_BIAS:
            bias_batch = bias_heads = 0
        else:
            *bias_batch_shape, bias_heads, _, _ = bias_aval.shape
            bias_batch = reduce(operator.mul, bias_batch_shape)
2413
2414
2415

        wkspace_aval = ctx.avals_out[-1]

2416
        opaque = transformer_engine_jax.pack_fused_attn_descriptor(
2417
2418
            input_batch, bias_batch, q_max_seqlen, kv_max_seqlen,
            attn_heads, num_gqa_groups, bias_heads, head_dim,
2419
            wkspace_aval.size, scaling_factor, dropout_probability, attn_bias_type, attn_mask_type,
2420
2421
            jax_dtype_to_te_dtype(q_aval.dtype), jax_dtype_to_te_dtype(wkspace_aval.dtype),
            is_training)
2422

2423
        out = custom_caller(CrossFusedAttnFwdPrimitive.name, args, opaque, has_side_effect=False)
2424
2425
2426

        return out

2427
    @staticmethod
zlsh80826's avatar
zlsh80826 committed
2428
2429
    def impl(q, kv, bias, q_seqlen, kv_seqlen, seed, attn_bias_type, attn_mask_type, scaling_factor,
             dropout_probability, is_training):
2430
        assert CrossFusedAttnFwdPrimitive.inner_primitive is not None
2431

zlsh80826's avatar
zlsh80826 committed
2432
2433
        q_cu_seqlen = generate_cu_seqlen(q_seqlen)
        kv_cu_seqlen = generate_cu_seqlen(kv_seqlen)
2434

2435
        output, softmax_aux, rng_state, _ = CrossFusedAttnFwdPrimitive.inner_primitive.bind(
2436
2437
            q,
            kv,
2438
            bias,
2439
2440
2441
2442
2443
2444
2445
2446
            q_cu_seqlen,
            kv_cu_seqlen,
            seed,
            attn_bias_type=attn_bias_type,
            attn_mask_type=attn_mask_type,
            scaling_factor=scaling_factor,
            dropout_probability=dropout_probability,
            is_training=is_training)
2447
        return output, softmax_aux, rng_state
2448

2449
2450
2451
2452
2453
    @staticmethod
    def batcher(batched_args, batch_dims, *, attn_bias_type, attn_mask_type, scaling_factor,
                dropout_probability, is_training):
        _check_valid_batch_dims(batch_dims)
        assert CrossFusedAttnFwdPrimitive.outer_primitive is not None
2454
        q_bdim, *_, seed_bdim = batch_dims
2455

2456
        out_bdims = q_bdim, q_bdim, seed_bdim
2457
        return CrossFusedAttnFwdPrimitive.outer_primitive.bind(
2458
            *batched_args,
2459
2460
2461
2462
2463
2464
2465
2466
2467
2468
2469
2470
2471
2472
2473
2474
2475
            attn_bias_type=attn_bias_type,
            attn_mask_type=attn_mask_type,
            scaling_factor=scaling_factor,
            dropout_probability=dropout_probability,
            is_training=is_training), out_bdims

    @staticmethod
    def infer_sharding_from_operands(attn_bias_type, attn_mask_type, scaling_factor,
                                     dropout_probability, is_training, mesh, arg_infos,
                                     result_infos):
        del attn_bias_type, attn_mask_type, scaling_factor
        del dropout_probability, is_training, result_infos
        q_spec = get_padded_spec(arg_infos[0])    # (...batch, q_seqlen, head, hidden)
        kv_spec = get_padded_spec(arg_infos[1])    # (...batch, kv_seqlen, 2, head, hidden)
        out_sharding = NamedSharding(mesh, PartitionSpec(*q_spec))
        softmax_aux_sharding = NamedSharding(
            mesh, PartitionSpec(*q_spec[:-3], q_spec[-2], q_spec[-3], kv_spec[-4]))
2476
2477
        rng_state_sharding = NamedSharding(mesh, PartitionSpec(get_all_mesh_axes(), None))
        return (out_sharding, softmax_aux_sharding, rng_state_sharding)
2478
2479
2480
2481
2482
2483
2484
2485
2486
2487

    @staticmethod
    def partition(attn_bias_type, attn_mask_type, scaling_factor, dropout_probability, is_training,
                  mesh, arg_infos, result_infos):
        del result_infos
        q_spec = get_padded_spec(arg_infos[0])    # (...batch, q_seqlen, head, hidden)
        kv_spec = get_padded_spec(arg_infos[1])    # (...batch, kv_seqlen, 2, head, hidden)
        out_sharding = NamedSharding(mesh, PartitionSpec(*q_spec))
        softmax_aux_sharding = NamedSharding(
            mesh, PartitionSpec(*q_spec[:-3], q_spec[-2], q_spec[-3], kv_spec[-4]))
2488
2489
        rng_state_sharding = seed_sharding = NamedSharding(mesh,
                                                           PartitionSpec(get_all_mesh_axes(), None))
2490
        arg_shardings = tuple([arg_i.sharding for arg_i in arg_infos[:-1]] + [seed_sharding])
2491
        out_shardings = (out_sharding, softmax_aux_sharding, rng_state_sharding)
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503
        impl = partial(CrossFusedAttnFwdPrimitive.impl,
                       attn_bias_type=attn_bias_type,
                       attn_mask_type=attn_mask_type,
                       scaling_factor=scaling_factor,
                       dropout_probability=dropout_probability,
                       is_training=is_training)
        return mesh, impl, out_shardings, arg_shardings


register_primitive(CrossFusedAttnFwdPrimitive)


zlsh80826's avatar
zlsh80826 committed
2504
2505
def cross_fused_attn_fwd(q: jnp.ndarray, kv: jnp.ndarray, bias: jnp.ndarray, q_seqlen: jnp.ndarray,
                         kv_seqlen: jnp.ndarray, seed: jnp.ndarray, attn_bias_type: NVTE_Bias_Type,
2506
2507
                         attn_mask_type: NVTE_Mask_Type, scaling_factor: float,
                         dropout_probability: float, is_training: bool):
2508
    """
2509
    Wrapper for TE cross fused attention fwd
2510
    Return BMM1 -> (PreScaleBias) -> Scale -> (PostScaleBias) -> Softmax -> (Dropout) -> BMM2
2511
    """
2512
2513
2514
    checker = _FusedAttnRNGStateChecker()
    seed = checker.check_seed(seed, dropout_probability, is_training)

2515
2516
2517
2518
    if attn_bias_type == NVTE_Bias_Type.NVTE_NO_BIAS:
        assert bias is None
        bias = jnp.zeros(0, dtype=q.dtype)

2519
2520
    return CrossFusedAttnFwdPrimitive.outer_primitive.bind(q,
                                                           kv,
2521
                                                           bias,
zlsh80826's avatar
zlsh80826 committed
2522
2523
                                                           q_seqlen,
                                                           kv_seqlen,
2524
2525
2526
2527
2528
2529
                                                           seed,
                                                           attn_bias_type=attn_bias_type,
                                                           attn_mask_type=attn_mask_type,
                                                           scaling_factor=scaling_factor,
                                                           dropout_probability=dropout_probability,
                                                           is_training=is_training)
2530
2531


2532
class CrossFusedAttnBwdPrimitive(BasePrimitive):
2533
    """
2534
    Cross Fused Attention Backward Primitive
2535
    """
2536
    name = "te_cross_fused_attn_backward"
2537
    multiple_results = True
2538
    impl_static_args = (9, 10, 11, 12, 13)
2539
2540
    inner_primitive = None
    outer_primitive = None
2541
2542

    @staticmethod
2543
2544
2545
    def abstract(q_aval, kv_aval, bias_aval, softmax_aux_aval, rng_state_aval, output_aval,
                 doutput_aval, q_cu_seqlen_aval, kv_cu_seqlen_aval, *, attn_bias_type,
                 attn_mask_type, scaling_factor, dropout_probability, is_training):
2546
        """
2547
        Cross fused attention bwd abstract
2548
        """
2549
        del softmax_aux_aval, rng_state_aval, output_aval
2550

2551
2552
        q_dtype = dtypes.canonicalize_dtype(q_aval.dtype)
        kv_dtype = dtypes.canonicalize_dtype(kv_aval.dtype)
2553
        bias_dtype = dtypes.canonicalize_dtype(bias_aval.dtype)
2554
        doutput_dtype = dtypes.canonicalize_dtype(doutput_aval.dtype)
2555
        assert q_dtype == kv_dtype == bias_dtype == doutput_dtype
2556
        assert q_cu_seqlen_aval.dtype == kv_cu_seqlen_aval.dtype
2557

2558
        *q_batch_shape, q_max_seqlen, attn_heads, q_head_dim = q_aval.shape
2559
2560
2561
2562
2563
        *kv_batch_shape, kv_max_seqlen, nkv, num_gqa_groups, kv_head_dim = kv_aval.shape
        assert q_batch_shape == kv_batch_shape
        assert q_head_dim == kv_head_dim
        assert nkv == 2

2564
2565
2566
2567
2568
2569
2570
        if attn_bias_type == NVTE_Bias_Type.NVTE_NO_BIAS:
            bias_batch = bias_heads = 0
        else:
            *bias_batch_shape, bias_heads, _, _ = bias_aval.shape
            bias_batch = reduce(operator.mul, bias_batch_shape)

        input_batch = reduce(operator.mul, q_batch_shape)
2571
2572
        wkspace_shape, wkspace_dtype = \
            transformer_engine_jax.get_cross_fused_attn_bwd_workspace_sizes(
2573
2574
                input_batch, bias_batch, q_max_seqlen, kv_max_seqlen,
                attn_heads, num_gqa_groups, bias_heads, q_head_dim,
2575
2576
2577
2578
                scaling_factor, dropout_probability, attn_bias_type, attn_mask_type,
                jax_dtype_to_te_dtype(q_aval.dtype), is_training
            )

2579
2580
        dq_aval = q_aval.update(shape=q_aval.shape, dtype=q_dtype)
        dkv_aval = kv_aval.update(shape=kv_aval.shape, dtype=kv_dtype)
2581
        dbias_aval = bias_aval.update(shape=bias_aval.shape, dtype=bias_dtype)
2582
2583
2584
2585
2586
2587
2588
2589
2590
2591
2592
2593
        wkspace_aval = q_aval.update(shape=wkspace_shape,
                                     dtype=te_dtype_to_jax_dtype(wkspace_dtype))

        return dq_aval, dkv_aval, dbias_aval, wkspace_aval

    @staticmethod
    def outer_abstract(*args, **kwargs):
        """
        Cross fused attention fwd outer primitive abstract
        """
        dq_aval, dkv_aval, dbias_aval, _ = \
            CrossFusedAttnBwdPrimitive.abstract(*args, **kwargs)
2594
        return dq_aval, dkv_aval, dbias_aval
2595
2596

    @staticmethod
2597
2598
2599
    def lowering(ctx, q, kv, bias, softmax_aux, rng_state, output, doutput, q_cu_seqlen,
                 kv_cu_seqlen, *, attn_bias_type, attn_mask_type, scaling_factor,
                 dropout_probability, is_training):
2600
        """
2601
        Cross fused attention bwd lowering rules
2602
        """
2603
        operands = [q, kv, bias, softmax_aux, rng_state, output, doutput, q_cu_seqlen, kv_cu_seqlen]
2604
        operand_shapes = map(lambda x: x.type.shape, operands)
2605
        out_types = [
2606
2607
            ir.RankedTensorType.get(output.shape, mlir.dtype_to_ir_type(output.dtype))
            for output in ctx.avals_out
2608
        ]
2609

2610
2611
        args = CustomCallArgsWrapper(out_types, operands, operand_shapes)

2612
2613
        q_aval, kv_aval, bias_aval, *_ = ctx.avals_in
        *input_batch_shape, q_max_seqlen, attn_heads, head_dim = q_aval.shape
2614
        *_, kv_max_seqlen, _, num_gqa_groups, _ = kv_aval.shape
2615
2616
2617
2618
2619
2620
2621
        input_batch = reduce(operator.mul, input_batch_shape)

        if attn_bias_type == NVTE_Bias_Type.NVTE_NO_BIAS:
            bias_batch = bias_heads = 0
        else:
            *bias_batch_shape, bias_heads, _, _ = bias_aval.shape
            bias_batch = reduce(operator.mul, bias_batch_shape)
2622
2623
2624

        wkspace_aval = ctx.avals_out[-1]

2625
        opaque = transformer_engine_jax.pack_fused_attn_descriptor(
2626
2627
            input_batch, bias_batch, q_max_seqlen, kv_max_seqlen,
            attn_heads, num_gqa_groups, bias_heads, head_dim,
2628
            wkspace_aval.size, scaling_factor, dropout_probability, attn_bias_type, attn_mask_type,
2629
2630
            jax_dtype_to_te_dtype(q_aval.dtype), jax_dtype_to_te_dtype(wkspace_aval.dtype),
            is_training)
2631

2632
        out = custom_caller(CrossFusedAttnBwdPrimitive.name, args, opaque, has_side_effect=False)
2633
2634
2635

        return out

2636
    @staticmethod
zlsh80826's avatar
zlsh80826 committed
2637
2638
    def impl(q, kv, bias, softmax_aux, rng_state, output, doutput, q_seqlen, kv_seqlen,
             attn_bias_type, attn_mask_type, scaling_factor, dropout_probability, is_training):
2639
2640
        assert CrossFusedAttnBwdPrimitive.inner_primitive is not None

zlsh80826's avatar
zlsh80826 committed
2641
2642
        q_cu_seqlen = generate_cu_seqlen(q_seqlen)
        kv_cu_seqlen = generate_cu_seqlen(kv_seqlen)
2643

2644
        dq, dkv, dbias, _ = CrossFusedAttnBwdPrimitive.inner_primitive.bind(
2645
2646
            q,
            kv,
2647
            bias,
2648
            softmax_aux,
2649
2650
            rng_state,
            output,
2651
2652
2653
2654
2655
2656
2657
2658
            doutput,
            q_cu_seqlen,
            kv_cu_seqlen,
            attn_bias_type=attn_bias_type,
            attn_mask_type=attn_mask_type,
            scaling_factor=scaling_factor,
            dropout_probability=dropout_probability,
            is_training=is_training)
2659
        return dq, dkv, dbias
2660
2661
2662
2663
2664
2665
2666
2667

    @staticmethod
    def batcher(batched_args, batch_dims, *, attn_bias_type, attn_mask_type, scaling_factor,
                dropout_probability, is_training):
        _check_valid_batch_dims(batch_dims)
        assert CrossFusedAttnBwdPrimitive.outer_primitive is not None
        q_bdim, kv_bdim, *_ = batch_dims

2668
        out_bdims = q_bdim, kv_bdim, q_bdim
2669
        return CrossFusedAttnBwdPrimitive.outer_primitive.bind(
2670
            *batched_args,
2671
2672
2673
2674
2675
2676
2677
2678
2679
2680
2681
2682
2683
2684
            attn_bias_type=attn_bias_type,
            attn_mask_type=attn_mask_type,
            scaling_factor=scaling_factor,
            dropout_probability=dropout_probability,
            is_training=is_training), out_bdims

    @staticmethod
    def infer_sharding_from_operands(attn_bias_type, attn_mask_type, scaling_factor,
                                     dropout_probability, is_training, mesh, arg_infos,
                                     result_infos):
        del attn_bias_type, attn_mask_type, scaling_factor
        del dropout_probability, is_training, result_infos
        q_spec = get_padded_spec(arg_infos[0])
        kv_spec = get_padded_spec(arg_infos[1])
2685
        bias_spec = get_padded_spec(arg_infos[2])
2686
2687
        dq_sharding = NamedSharding(mesh, PartitionSpec(*q_spec))
        dkv_sharding = NamedSharding(mesh, PartitionSpec(*kv_spec))
2688
2689
        dbias_sharding = NamedSharding(mesh, PartitionSpec(*bias_spec))
        return (dq_sharding, dkv_sharding, dbias_sharding)
2690
2691
2692
2693
2694
2695
2696

    @staticmethod
    def partition(attn_bias_type, attn_mask_type, scaling_factor, dropout_probability, is_training,
                  mesh, arg_infos, result_infos):
        del result_infos
        q_spec = get_padded_spec(arg_infos[0])
        kv_spec = get_padded_spec(arg_infos[1])
2697
        bias_spec = get_padded_spec(arg_infos[2])
2698
2699
        dq_sharding = NamedSharding(mesh, PartitionSpec(*q_spec))
        dkv_sharding = NamedSharding(mesh, PartitionSpec(*kv_spec))
2700
        dbias_sharding = NamedSharding(mesh, PartitionSpec(*bias_spec))
2701
        arg_shardings = tuple(arg_i.sharding for arg_i in arg_infos)
2702
2703
2704
2705
2706
2707
2708
2709
2710
2711
2712
2713
2714
2715
2716
2717
2718
2719
2720
2721
2722
2723
2724
        out_shardings = (dq_sharding, dkv_sharding, dbias_sharding)

        def sharded_impl(q, kv, bias, softmax_aux, rng_state, output, doutput, q_cu_seqlen,
                         kv_cu_seqlen):
            local_dq, local_dkv, local_dbias = CrossFusedAttnBwdPrimitive.impl(
                q,
                kv,
                bias,
                softmax_aux,
                rng_state,
                output,
                doutput,
                q_cu_seqlen,
                kv_cu_seqlen,
                attn_bias_type=attn_bias_type,
                attn_mask_type=attn_mask_type,
                scaling_factor=scaling_factor,
                dropout_probability=dropout_probability,
                is_training=is_training)
            global_dbias = local_dbias
            if attn_bias_type is not NVTE_Bias_Type.NVTE_NO_BIAS:
                global_dbias = all_reduce_sum_along_dp_fsdp(local_dbias)
            return local_dq, local_dkv, global_dbias
2725

2726
        return mesh, sharded_impl, out_shardings, arg_shardings
2727

2728

2729
register_primitive(CrossFusedAttnBwdPrimitive)
2730
2731


2732
2733
def cross_fused_attn_bwd(q: jnp.ndarray, kv: jnp.ndarray, bias: jnp.ndarray,
                         softmax_aux: jnp.ndarray, rng_state: jnp.ndarray, output: jnp.ndarray,
zlsh80826's avatar
zlsh80826 committed
2734
2735
2736
                         doutput: jnp.ndarray, q_seqlen: jnp.ndarray, kv_seqlen: jnp.ndarray,
                         attn_bias_type: NVTE_Bias_Type, attn_mask_type: NVTE_Mask_Type,
                         scaling_factor: float, dropout_probability: float, is_training: bool):
2737
    """
2738
2739
    Wrapper for TE cross fused attention bwd
    Return the gradients of cross fused attention with packed kv input
2740
    """
2741
2742
2743
    if attn_bias_type == NVTE_Bias_Type.NVTE_NO_BIAS:
        assert bias is None
        bias = jnp.zeros(0, dtype=q.dtype)
2744

2745
2746
    return CrossFusedAttnBwdPrimitive.outer_primitive.bind(q,
                                                           kv,
2747
                                                           bias,
2748
                                                           softmax_aux,
2749
2750
                                                           rng_state,
                                                           output,
2751
                                                           doutput,
zlsh80826's avatar
zlsh80826 committed
2752
2753
                                                           q_seqlen,
                                                           kv_seqlen,
2754
2755
2756
2757
2758
                                                           attn_bias_type=attn_bias_type,
                                                           attn_mask_type=attn_mask_type,
                                                           scaling_factor=scaling_factor,
                                                           dropout_probability=dropout_probability,
                                                           is_training=is_training)
2759
2760


2761
2762
2763
2764
2765
2766
2767
2768
2769
2770
2771
2772
2773
2774
2775
2776
2777
2778
2779
2780
2781
2782
2783
2784
2785
class FusedAttnFwdPrimitive(BasePrimitive):
    """
    Fused Attention Forward Primitive
    Query, key, value are seperated tensors
    """
    name = "te_fused_attn_forward"
    multiple_results = True
    impl_static_args = (7, 8, 9, 10, 11)
    inner_primitive = None
    outer_primitive = None

    @staticmethod
    def abstract(q_aval, k_aval, v_aval, bias_aval, q_seqlen_or_cu_seqlen_aval,
                 kv_seqlen_or_cu_seqlen_aval, seed_aval, *, attn_bias_type, attn_mask_type,
                 scaling_factor, dropout_probability, is_training):
        """
        Fused attention fwd abstract
        """
        q_dtype = dtypes.canonicalize_dtype(q_aval.dtype)
        k_dtype = dtypes.canonicalize_dtype(k_aval.dtype)
        v_dtype = dtypes.canonicalize_dtype(v_aval.dtype)
        bias_dtype = dtypes.canonicalize_dtype(bias_aval.dtype)
        assert q_dtype == k_dtype == v_dtype == bias_dtype
        assert q_seqlen_or_cu_seqlen_aval.dtype == kv_seqlen_or_cu_seqlen_aval.dtype

2786
        *q_batch_shape, q_max_seqlen, attn_heads, q_head_dim = q_aval.shape
2787
2788
2789
2790
2791
2792
2793
2794
        *kv_batch_shape, kv_max_seqlen, num_gqa_groups, kv_head_dim = k_aval.shape
        assert q_batch_shape == kv_batch_shape
        assert q_head_dim == kv_head_dim
        assert k_aval.shape == v_aval.shape
        out_aval = q_aval.update(shape=q_aval.shape, dtype=q_dtype)

        # backend determines the softmax buffer shape/dtype
        backend = FusedAttnHelper(q_dtype, k_dtype, NVTE_QKV_Layout.NVTE_BSHD_BS2HD, attn_bias_type,
2795
                                  attn_mask_type, dropout_probability, attn_heads, num_gqa_groups,
2796
2797
2798
                                  q_max_seqlen, kv_max_seqlen, q_head_dim).get_fused_attn_backend()

        if backend == NVTE_Fused_Attn_Backend.NVTE_F16_max512_seqlen:
2799
            softmax_shape = (*q_batch_shape, attn_heads, q_max_seqlen, kv_max_seqlen)
2800
2801
            softmax_dtype = q_dtype
        elif backend == NVTE_Fused_Attn_Backend.NVTE_F16_arbitrary_seqlen:
2802
            softmax_shape = (*q_batch_shape, attn_heads, q_max_seqlen, 1)
2803
2804
2805
2806
2807
2808
2809
2810
2811
2812
2813
2814
2815
            softmax_dtype = dtypes.canonicalize_dtype(jnp.float32)
        else:
            raise ValueError(f'Unsupported {backend=}')
        softmax_aux_aval = q_aval.update(shape=softmax_shape, dtype=softmax_dtype)

        # JAX does not enable 64-bit int by default so we get XLA to allocate x8 memory with
        # 32-bit unsigned int to get the buffer size we need in the C++ kernel
        checker = _FusedAttnRNGStateChecker()
        seed_dtype = dtypes.canonicalize_dtype(seed_aval.dtype)
        assert seed_dtype == checker.rng_state_dtype
        rng_state_shape = (seed_aval.shape[0], checker.rng_state_size)
        rng_state_aval = seed_aval.update(shape=rng_state_shape, dtype=checker.rng_state_dtype)

2816
2817
2818
2819
2820
2821
        if attn_bias_type == NVTE_Bias_Type.NVTE_NO_BIAS:
            bias_batch = bias_heads = 0
        else:
            *bias_batch_shape, bias_heads, _, _ = bias_aval.shape
            bias_batch = reduce(operator.mul, bias_batch_shape)

2822
2823
        # do a dummy kernel call here to get workspace buffer shapes/dtypes that XLA needs to
        # prepare for the active fused-attn backend
2824
        input_batch = reduce(operator.mul, q_batch_shape)
2825
        wkspace_info = transformer_engine_jax.get_fused_attn_fwd_workspace_sizes(
2826
2827
            input_batch, bias_batch, q_max_seqlen, kv_max_seqlen,
            attn_heads, num_gqa_groups, bias_heads, q_head_dim,
2828
2829
2830
2831
2832
2833
2834
2835
2836
2837
2838
2839
2840
2841
2842
2843
2844
2845
2846
2847
2848
2849
2850
2851
2852
2853
2854
2855
2856
2857
            scaling_factor, dropout_probability, attn_bias_type, attn_mask_type,
            jax_dtype_to_te_dtype(q_aval.dtype), is_training)
        wkspace_aval = q_aval.update(shape=wkspace_info[0],
                                     dtype=te_dtype_to_jax_dtype(wkspace_info[1]))

        return out_aval, softmax_aux_aval, rng_state_aval, wkspace_aval

    @staticmethod
    def outer_abstract(*args, **kwargs):
        """
        Fused attention fwd outer primitive abstract
        """
        out_aval, softmax_aux_aval, rng_state_aval, _ = \
            FusedAttnFwdPrimitive.abstract(*args, **kwargs)
        return out_aval, softmax_aux_aval, rng_state_aval

    @staticmethod
    def lowering(ctx, q, k, v, bias, q_cu_seqlen, kv_cu_seqlen, seed, *, attn_bias_type,
                 attn_mask_type, scaling_factor, dropout_probability, is_training):
        """
        Fused attention fwd lowering rules
        """
        operands = [q, k, v, bias, q_cu_seqlen, kv_cu_seqlen, seed]
        operand_shapes = map(lambda x: x.type.shape, operands)
        out_types = [
            ir.RankedTensorType.get(output.shape, mlir.dtype_to_ir_type(output.dtype))
            for output in ctx.avals_out
        ]
        args = CustomCallArgsWrapper(out_types, operands, operand_shapes)

2858
2859
        q_aval, k_aval, v_aval, bias_aval, *_ = ctx.avals_in
        *batch_shape, q_max_seqlen, attn_heads, head_dim = q_aval.shape
2860
2861
        *_, kv_max_seqlen, num_gqa_groups, _ = k_aval.shape
        assert k_aval.shape == v_aval.shape
2862
2863
2864
2865
2866
2867
2868
        input_batch = reduce(operator.mul, batch_shape)

        if attn_bias_type == NVTE_Bias_Type.NVTE_NO_BIAS:
            bias_batch = bias_heads = 0
        else:
            *bias_batch_shape, bias_heads, _, _ = bias_aval.shape
            bias_batch = reduce(operator.mul, bias_batch_shape)
2869
2870
2871
2872

        wkspace_aval = ctx.avals_out[-1]

        opaque = transformer_engine_jax.pack_fused_attn_descriptor(
2873
2874
            input_batch, bias_batch, q_max_seqlen, kv_max_seqlen,
            attn_heads, num_gqa_groups, bias_heads, head_dim,
2875
2876
2877
2878
2879
2880
2881
2882
2883
2884
2885
2886
2887
2888
2889
2890
2891
2892
2893
2894
2895
2896
2897
2898
2899
2900
2901
2902
2903
2904
2905
2906
2907
2908
2909
2910
2911
2912
2913
2914
2915
2916
2917
2918
2919
2920
2921
2922
2923
2924
2925
2926
2927
2928
2929
2930
2931
2932
2933
2934
2935
2936
2937
2938
2939
2940
2941
2942
2943
2944
2945
2946
2947
2948
2949
2950
2951
2952
2953
2954
2955
2956
2957
2958
2959
2960
2961
2962
2963
2964
2965
2966
2967
2968
2969
2970
2971
2972
2973
2974
2975
2976
2977
2978
2979
2980
2981
2982
2983
2984
2985
2986
2987
2988
2989
2990
2991
2992
2993
2994
2995
2996
2997
2998
2999
3000
3001
3002
3003
3004
3005
3006
3007
3008
3009
3010
3011
3012
3013
3014
3015
            wkspace_aval.size, scaling_factor, dropout_probability, attn_bias_type, attn_mask_type,
            jax_dtype_to_te_dtype(q_aval.dtype), jax_dtype_to_te_dtype(wkspace_aval.dtype),
            is_training)

        out = custom_caller(FusedAttnFwdPrimitive.name, args, opaque, has_side_effect=False)

        return out

    @staticmethod
    def impl(q, k, v, bias, q_seqlen, kv_seqlen, seed, attn_bias_type, attn_mask_type,
             scaling_factor, dropout_probability, is_training):
        assert FusedAttnFwdPrimitive.inner_primitive is not None

        q_cu_seqlen = generate_cu_seqlen(q_seqlen)
        kv_cu_seqlen = generate_cu_seqlen(kv_seqlen)

        output, softmax_aux, rng_state, _ = FusedAttnFwdPrimitive.inner_primitive.bind(
            q,
            k,
            v,
            bias,
            q_cu_seqlen,
            kv_cu_seqlen,
            seed,
            attn_bias_type=attn_bias_type,
            attn_mask_type=attn_mask_type,
            scaling_factor=scaling_factor,
            dropout_probability=dropout_probability,
            is_training=is_training)
        return output, softmax_aux, rng_state

    @staticmethod
    def batcher(batched_args, batch_dims, *, attn_bias_type, attn_mask_type, scaling_factor,
                dropout_probability, is_training):
        _check_valid_batch_dims(batch_dims)
        assert FusedAttnFwdPrimitive.outer_primitive is not None
        q_bdim, *_, seed_bdim = batch_dims

        out_bdims = q_bdim, q_bdim, seed_bdim
        return FusedAttnFwdPrimitive.outer_primitive.bind(*batched_args,
                                                          attn_bias_type=attn_bias_type,
                                                          attn_mask_type=attn_mask_type,
                                                          scaling_factor=scaling_factor,
                                                          dropout_probability=dropout_probability,
                                                          is_training=is_training), out_bdims

    @staticmethod
    def infer_sharding_from_operands(attn_bias_type, attn_mask_type, scaling_factor,
                                     dropout_probability, is_training, mesh, arg_infos,
                                     result_infos):
        del attn_bias_type, attn_mask_type, scaling_factor
        del dropout_probability, is_training, result_infos
        q_spec = get_padded_spec(arg_infos[0])    # (...batch, q_seqlen, head, hidden)
        k_spec = get_padded_spec(arg_infos[1])    # (...batch, kv_seqlen, head, hidden)
        out_sharding = NamedSharding(mesh, PartitionSpec(*q_spec))
        softmax_aux_sharding = NamedSharding(
            mesh, PartitionSpec(*q_spec[:-3], q_spec[-2], q_spec[-3], k_spec[-3]))
        rng_state_sharding = NamedSharding(mesh, PartitionSpec(get_all_mesh_axes(), None))
        return (out_sharding, softmax_aux_sharding, rng_state_sharding)

    @staticmethod
    def partition(attn_bias_type, attn_mask_type, scaling_factor, dropout_probability, is_training,
                  mesh, arg_infos, result_infos):
        del result_infos
        q_spec = get_padded_spec(arg_infos[0])    # (...batch, q_seqlen, head, hidden)
        k_spec = get_padded_spec(arg_infos[1])    # (...batch, kv_seqlen, head, hidden)
        out_sharding = NamedSharding(mesh, PartitionSpec(*q_spec))
        softmax_aux_sharding = NamedSharding(
            mesh, PartitionSpec(*q_spec[:-3], q_spec[-2], q_spec[-3], k_spec[-3]))
        rng_state_sharding = seed_sharding = NamedSharding(mesh,
                                                           PartitionSpec(get_all_mesh_axes(), None))
        arg_shardings = tuple([arg_i.sharding for arg_i in arg_infos[:-1]] + [seed_sharding])
        out_shardings = (out_sharding, softmax_aux_sharding, rng_state_sharding)
        impl = partial(FusedAttnFwdPrimitive.impl,
                       attn_bias_type=attn_bias_type,
                       attn_mask_type=attn_mask_type,
                       scaling_factor=scaling_factor,
                       dropout_probability=dropout_probability,
                       is_training=is_training)
        return mesh, impl, out_shardings, arg_shardings


register_primitive(FusedAttnFwdPrimitive)


def fused_attn_fwd(q: jnp.ndarray, k: jnp.ndarray, v: jnp.ndarray, bias: jnp.ndarray,
                   q_seqlen: jnp.ndarray, kv_seqlen: jnp.ndarray, seed: jnp.ndarray,
                   attn_bias_type: NVTE_Bias_Type, attn_mask_type: NVTE_Mask_Type,
                   scaling_factor: float, dropout_probability: float, is_training: bool):
    """
    Wrapper for TE fused attention fwd, where query, key, value are seperated tensors
    Return BMM1 -> (PreScaleBias) -> Scale -> (PostScaleBias) -> Softmax -> (Dropout) -> BMM2
    """
    checker = _FusedAttnRNGStateChecker()
    seed = checker.check_seed(seed, dropout_probability, is_training)

    if attn_bias_type == NVTE_Bias_Type.NVTE_NO_BIAS:
        assert bias is None
        bias = jnp.zeros(0, dtype=q.dtype)

    return FusedAttnFwdPrimitive.outer_primitive.bind(q,
                                                      k,
                                                      v,
                                                      bias,
                                                      q_seqlen,
                                                      kv_seqlen,
                                                      seed,
                                                      attn_bias_type=attn_bias_type,
                                                      attn_mask_type=attn_mask_type,
                                                      scaling_factor=scaling_factor,
                                                      dropout_probability=dropout_probability,
                                                      is_training=is_training)


class FusedAttnBwdPrimitive(BasePrimitive):
    """
    Fused Attention Backward Primitive
    """
    name = "te_fused_attn_backward"
    multiple_results = True
    impl_static_args = (10, 11, 12, 13, 14)
    inner_primitive = None
    outer_primitive = None

    @staticmethod
    def abstract(q_aval, k_aval, v_aval, bias_aval, softmax_aux_aval, rng_state_aval, output_aval,
                 doutput_aval, q_cu_seqlen_aval, kv_cu_seqlen_aval, *, attn_bias_type,
                 attn_mask_type, scaling_factor, dropout_probability, is_training):
        """
        Fused attention bwd abstract
        """
        del softmax_aux_aval, rng_state_aval, output_aval

        q_dtype = dtypes.canonicalize_dtype(q_aval.dtype)
        k_dtype = dtypes.canonicalize_dtype(k_aval.dtype)
        v_dtype = dtypes.canonicalize_dtype(v_aval.dtype)
        bias_dtype = dtypes.canonicalize_dtype(bias_aval.dtype)
        doutput_dtype = dtypes.canonicalize_dtype(doutput_aval.dtype)
        assert q_dtype == k_dtype == v_dtype == bias_dtype == doutput_dtype
        assert q_cu_seqlen_aval.dtype == kv_cu_seqlen_aval.dtype

3016
        *q_batch_shape, q_max_seqlen, attn_heads, q_head_dim = q_aval.shape
3017
3018
3019
3020
3021
        *kv_batch_shape, kv_max_seqlen, num_gqa_groups, kv_head_dim = k_aval.shape
        assert q_batch_shape == kv_batch_shape
        assert q_head_dim == kv_head_dim
        assert k_aval.shape == v_aval.shape

3022
3023
3024
3025
3026
3027
3028
        if attn_bias_type == NVTE_Bias_Type.NVTE_NO_BIAS:
            bias_batch = bias_heads = 0
        else:
            *bias_batch_shape, bias_heads, _, _ = bias_aval.shape
            bias_batch = reduce(operator.mul, bias_batch_shape)

        input_batch = reduce(operator.mul, q_batch_shape)
3029
3030
        wkspace_shape, wkspace_dtype = \
            transformer_engine_jax.get_fused_attn_bwd_workspace_sizes(
3031
3032
                input_batch, bias_batch, q_max_seqlen, kv_max_seqlen,
                attn_heads, num_gqa_groups, bias_heads, q_head_dim,
3033
3034
3035
3036
3037
3038
3039
3040
3041
3042
3043
3044
3045
3046
3047
3048
3049
3050
3051
3052
3053
3054
3055
3056
3057
3058
3059
3060
3061
3062
3063
3064
3065
3066
3067
3068
3069
3070
3071
3072
                scaling_factor, dropout_probability, attn_bias_type, attn_mask_type,
                jax_dtype_to_te_dtype(q_aval.dtype), is_training
            )

        dq_aval = q_aval.update(shape=q_aval.shape, dtype=q_dtype)
        dk_aval = k_aval.update(shape=k_aval.shape, dtype=k_dtype)
        dv_aval = v_aval.update(shape=v_aval.shape, dtype=v_dtype)
        dbias_aval = bias_aval.update(shape=bias_aval.shape, dtype=bias_dtype)
        wkspace_aval = q_aval.update(shape=wkspace_shape,
                                     dtype=te_dtype_to_jax_dtype(wkspace_dtype))

        return dq_aval, dk_aval, dv_aval, dbias_aval, wkspace_aval

    @staticmethod
    def outer_abstract(*args, **kwargs):
        """
        Fused attention fwd outer primitive abstract
        """
        dq_aval, dk_aval, dv_aval, dbias_aval, _ = \
            FusedAttnBwdPrimitive.abstract(*args, **kwargs)
        return dq_aval, dk_aval, dv_aval, dbias_aval

    @staticmethod
    def lowering(ctx, q, k, v, bias, softmax_aux, rng_state, output, doutput, q_cu_seqlen,
                 kv_cu_seqlen, *, attn_bias_type, attn_mask_type, scaling_factor,
                 dropout_probability, is_training):
        """
        Fused attention bwd lowering rules
        """
        operands = [
            q, k, v, bias, softmax_aux, rng_state, output, doutput, q_cu_seqlen, kv_cu_seqlen
        ]
        operand_shapes = map(lambda x: x.type.shape, operands)
        out_types = [
            ir.RankedTensorType.get(output.shape, mlir.dtype_to_ir_type(output.dtype))
            for output in ctx.avals_out
        ]

        args = CustomCallArgsWrapper(out_types, operands, operand_shapes)

3073
3074
        q_aval, k_aval, v_aval, bias_aval, *_ = ctx.avals_in
        *batch_shape, q_max_seqlen, attn_heads, head_dim = q_aval.shape
3075
3076
        *_, kv_max_seqlen, num_gqa_groups, _ = k_aval.shape
        assert k_aval.shape == v_aval.shape
3077
3078
3079
3080
3081
3082
3083
        input_batch = reduce(operator.mul, batch_shape)

        if attn_bias_type == NVTE_Bias_Type.NVTE_NO_BIAS:
            bias_batch = bias_heads = 0
        else:
            *bias_batch_shape, bias_heads, _, _ = bias_aval.shape
            bias_batch = reduce(operator.mul, bias_batch_shape)
3084
3085
3086
3087

        wkspace_aval = ctx.avals_out[-1]

        opaque = transformer_engine_jax.pack_fused_attn_descriptor(
3088
3089
            input_batch, bias_batch, q_max_seqlen, kv_max_seqlen,
            attn_heads, num_gqa_groups, bias_heads, head_dim,
3090
3091
3092
3093
3094
3095
3096
3097
3098
3099
3100
3101
3102
3103
3104
3105
3106
3107
3108
3109
3110
3111
3112
3113
3114
3115
3116
3117
3118
3119
3120
3121
3122
3123
3124
3125
3126
3127
3128
3129
3130
3131
3132
3133
3134
3135
3136
3137
3138
3139
3140
3141
3142
3143
3144
3145
3146
3147
3148
3149
3150
3151
3152
3153
3154
3155
3156
3157
3158
3159
3160
3161
3162
3163
3164
3165
3166
3167
3168
3169
3170
3171
3172
3173
3174
3175
3176
3177
3178
3179
3180
3181
3182
3183
3184
3185
3186
3187
3188
3189
3190
3191
3192
3193
3194
3195
3196
3197
3198
3199
3200
3201
3202
3203
3204
3205
3206
3207
3208
3209
3210
            wkspace_aval.size, scaling_factor, dropout_probability, attn_bias_type, attn_mask_type,
            jax_dtype_to_te_dtype(q_aval.dtype), jax_dtype_to_te_dtype(wkspace_aval.dtype),
            is_training)

        out = custom_caller(FusedAttnBwdPrimitive.name, args, opaque, has_side_effect=False)

        return out

    @staticmethod
    def impl(q, k, v, bias, softmax_aux, rng_state, output, doutput, q_seqlen, kv_seqlen,
             attn_bias_type, attn_mask_type, scaling_factor, dropout_probability, is_training):
        assert FusedAttnBwdPrimitive.inner_primitive is not None

        q_cu_seqlen = generate_cu_seqlen(q_seqlen)
        kv_cu_seqlen = generate_cu_seqlen(kv_seqlen)

        dq, dk, dv, dbias, _ = FusedAttnBwdPrimitive.inner_primitive.bind(
            q,
            k,
            v,
            bias,
            softmax_aux,
            rng_state,
            output,
            doutput,
            q_cu_seqlen,
            kv_cu_seqlen,
            attn_bias_type=attn_bias_type,
            attn_mask_type=attn_mask_type,
            scaling_factor=scaling_factor,
            dropout_probability=dropout_probability,
            is_training=is_training)
        return dq, dk, dv, dbias

    @staticmethod
    def batcher(batched_args, batch_dims, *, attn_bias_type, attn_mask_type, scaling_factor,
                dropout_probability, is_training):
        _check_valid_batch_dims(batch_dims)
        assert FusedAttnBwdPrimitive.outer_primitive is not None
        q_bdim, k_bdim, v_bdim, *_ = batch_dims

        out_bdims = q_bdim, k_bdim, v_bdim, q_bdim
        return FusedAttnBwdPrimitive.outer_primitive.bind(*batched_args,
                                                          attn_bias_type=attn_bias_type,
                                                          attn_mask_type=attn_mask_type,
                                                          scaling_factor=scaling_factor,
                                                          dropout_probability=dropout_probability,
                                                          is_training=is_training), out_bdims

    @staticmethod
    def infer_sharding_from_operands(attn_bias_type, attn_mask_type, scaling_factor,
                                     dropout_probability, is_training, mesh, arg_infos,
                                     result_infos):
        del attn_bias_type, attn_mask_type, scaling_factor
        del dropout_probability, is_training, result_infos
        q_spec = get_padded_spec(arg_infos[0])
        k_spec = get_padded_spec(arg_infos[1])
        v_spec = get_padded_spec(arg_infos[2])
        bias_spec = get_padded_spec(arg_infos[3])
        dq_sharding = NamedSharding(mesh, PartitionSpec(*q_spec))
        dk_sharding = NamedSharding(mesh, PartitionSpec(*k_spec))
        dv_sharding = NamedSharding(mesh, PartitionSpec(*v_spec))
        dbias_sharding = NamedSharding(mesh, PartitionSpec(*bias_spec))
        return (dq_sharding, dk_sharding, dv_sharding, dbias_sharding)

    @staticmethod
    def partition(attn_bias_type, attn_mask_type, scaling_factor, dropout_probability, is_training,
                  mesh, arg_infos, result_infos):
        del result_infos
        q_spec = get_padded_spec(arg_infos[0])
        k_spec = get_padded_spec(arg_infos[1])
        v_spec = get_padded_spec(arg_infos[2])
        bias_spec = get_padded_spec(arg_infos[3])
        dq_sharding = NamedSharding(mesh, PartitionSpec(*q_spec))
        dk_sharding = NamedSharding(mesh, PartitionSpec(*k_spec))
        dv_sharding = NamedSharding(mesh, PartitionSpec(*v_spec))
        dbias_sharding = NamedSharding(mesh, PartitionSpec(*bias_spec))
        arg_shardings = tuple(arg_i.sharding for arg_i in arg_infos)
        out_shardings = (dq_sharding, dk_sharding, dv_sharding, dbias_sharding)

        def sharded_impl(q, k, v, bias, softmax_aux, rng_state, output, doutput, q_cu_seqlen,
                         kv_cu_seqlen):
            local_dq, local_dk, local_dv, local_dbias = FusedAttnBwdPrimitive.impl(
                q,
                k,
                v,
                bias,
                softmax_aux,
                rng_state,
                output,
                doutput,
                q_cu_seqlen,
                kv_cu_seqlen,
                attn_bias_type=attn_bias_type,
                attn_mask_type=attn_mask_type,
                scaling_factor=scaling_factor,
                dropout_probability=dropout_probability,
                is_training=is_training)
            global_dbias = local_dbias
            if attn_bias_type is not NVTE_Bias_Type.NVTE_NO_BIAS:
                global_dbias = all_reduce_sum_along_dp_fsdp(local_dbias)
            return local_dq, local_dk, local_dv, global_dbias

        return mesh, sharded_impl, out_shardings, arg_shardings


register_primitive(FusedAttnBwdPrimitive)


def fused_attn_bwd(q: jnp.ndarray, k: jnp.ndarray, v: jnp.ndarray, bias: jnp.ndarray,
                   softmax_aux: jnp.ndarray, rng_state: jnp.ndarray, output: jnp.ndarray,
                   doutput: jnp.ndarray, q_seqlen: jnp.ndarray, kv_seqlen: jnp.ndarray,
                   attn_bias_type: NVTE_Bias_Type, attn_mask_type: NVTE_Mask_Type,
                   scaling_factor: float, dropout_probability: float, is_training: bool):
    """
    Wrapper for TE fused attention bwd
    Return the gradients of fused attention with seperated query, key, value tensors
    """
    if attn_bias_type == NVTE_Bias_Type.NVTE_NO_BIAS:
        assert bias is None
        bias = jnp.zeros(0, dtype=q.dtype)
3211

3212
3213
3214
3215
3216
3217
3218
3219
3220
3221
3222
3223
3224
3225
3226
3227
3228
    return FusedAttnBwdPrimitive.outer_primitive.bind(q,
                                                      k,
                                                      v,
                                                      bias,
                                                      softmax_aux,
                                                      rng_state,
                                                      output,
                                                      doutput,
                                                      q_seqlen,
                                                      kv_seqlen,
                                                      attn_bias_type=attn_bias_type,
                                                      attn_mask_type=attn_mask_type,
                                                      scaling_factor=scaling_factor,
                                                      dropout_probability=dropout_probability,
                                                      is_training=is_training)


3229
3230
3231
3232
3233
3234
3235
3236
3237
3238
3239
3240
3241
3242
3243
3244
3245
3246
3247
3248
3249
3250
3251
3252
3253
3254
3255
3256
3257
3258
3259
3260
3261
3262
3263
3264
3265
3266
3267
3268
3269
3270
3271
3272
3273
3274
3275
3276
3277
3278
3279
3280
3281
3282
3283
3284
3285
3286
3287
3288
3289
3290
3291
3292
3293
3294
3295
3296
3297
3298
3299
3300
3301
3302
3303
3304
3305
3306
3307
3308
3309
3310
3311
3312
3313
3314
3315
3316
3317
3318
3319
3320
3321
3322
3323
3324
3325
3326
3327
3328
3329
3330
3331
3332
3333
3334
3335
3336
3337
3338
3339
3340
3341
3342
3343
3344
3345
3346
3347
3348
3349
3350
3351
3352
3353
3354
3355
3356
3357
3358
3359
3360
3361
3362
3363
3364
3365
3366
3367
3368
3369
3370
3371
3372
3373
3374
3375
3376
3377
3378
3379
3380
3381
3382
3383
3384
3385
3386
3387
3388
3389
3390
3391
3392
3393
3394
3395
3396
3397
3398
3399
3400
3401
3402
3403
3404
3405
3406
3407
3408
3409
3410
3411
3412
3413
3414
3415
3416
3417
3418
3419
3420
3421
3422
3423
3424
3425
3426
3427
3428
3429
3430
3431
3432
3433
3434
3435
3436
3437
3438
3439
3440
3441
3442
3443
3444
class GeluPrimitive(BasePrimitive):
    """
    Gelu Froward Primitive
    """
    name = "te_gelu"
    multiple_results = False
    inner_primitive = None
    outer_primitive = None
    impl_static_args = ()

    @staticmethod
    def abstract(x_aval):
        """
        gated_gelu abstract
        """
        dtype = dtypes.canonicalize_dtype(x_aval.dtype)
        assert dtype in [jnp.float32, jnp.float16, jnp.bfloat16]

        out_aval = core.raise_to_shaped(x_aval)
        return out_aval

    @staticmethod
    def lowering(ctx, x):
        """
        gated_gelu lowering rules
        """
        (x_aval,) = ctx.avals_in
        assert x_aval.dtype in [jnp.float32, jnp.float16, jnp.bfloat16]
        ir_x_type = ir.RankedTensorType(x.type)
        ir_x_shape = ir_x_type.shape
        out_shape = ir_x_shape

        out_types = [
            ir.RankedTensorType.get(out_shape, ir_x_type.element_type),
        ]
        operands = [x]
        operand_shapes = [ir_x_shape]
        args = CustomCallArgsWrapper(out_types, operands, operand_shapes)

        hidden_size = ir_x_shape[-1]
        batch_size = reduce(operator.mul, ir_x_shape[:-1])
        in_dtype = jax_dtype_to_te_dtype(x_aval.dtype)
        opaque = transformer_engine_jax.pack_common_descriptor((batch_size, hidden_size), in_dtype,
                                                               in_dtype)

        out = custom_caller(GeluPrimitive.name, args, opaque, False)

        return [out]

    @staticmethod
    def impl(x):
        assert GeluPrimitive.inner_primitive is not None
        out = GeluPrimitive.inner_primitive.bind(x)
        return out

    @staticmethod
    def batcher(batched_args, batch_dims):
        """
        gated_gelu batcher
        """
        _check_valid_batch_dims(batch_dims)
        assert GeluPrimitive.outer_primitive is not None
        inputs, = batched_args
        inputs_bdim, = batch_dims

        out_bdims = inputs_bdim
        return GeluPrimitive.outer_primitive.bind(inputs), out_bdims

    @staticmethod
    def infer_sharding_from_operands(mesh, arg_infos, result_infos):
        """
        gated_gelu infer_sharding_from_operands
        """
        del result_infos    # Unused.
        x_spec = get_padded_spec(arg_infos[0])
        out_sharding = NamedSharding(mesh, PartitionSpec(*x_spec))
        return out_sharding

    @staticmethod
    def partition(mesh, arg_infos, result_infos):
        """
        gated_gelu partitioning
        """
        del result_infos
        x_spec = get_padded_spec(arg_infos[0])
        arg_shardings = tuple(arg_i.sharding for arg_i in arg_infos)
        out_sharding = NamedSharding(mesh, PartitionSpec(*x_spec))
        impl = GeluPrimitive.impl
        return mesh, impl, out_sharding, arg_shardings


register_primitive(GeluPrimitive)


def gelu(inputs: jnp.ndarray) -> jnp.ndarray:
    """
    gelu wrapper
    Return geglu(inputs)
    Assume inputs has two dimensions shape and the memory layout is (N..., H)
    """
    return GeluPrimitive.outer_primitive.bind(inputs)


class DGeluPrimitive(BasePrimitive):
    """
    Dgated Gelu Primitive
    """
    name = "te_dgelu"
    multiple_results = False
    inner_primitive = None
    outer_primitive = None
    impl_static_args = ()

    @staticmethod
    def abstract(dz_aval, x_aval):
        """
        dgelu abstract
        """
        dtype = dtypes.canonicalize_dtype(dz_aval.dtype)
        assert dtype in [jnp.float32, jnp.float16, jnp.bfloat16]
        assert x_aval.dtype == dtype
        assert dz_aval.shape == x_aval.shape

        out_aval = core.raise_to_shaped(x_aval)
        return out_aval

    @staticmethod
    def lowering(ctx, dz, x):
        """
        dgelu lowering rules
        """
        in_aval, gi_aval = ctx.avals_in
        assert in_aval.dtype in [jnp.float32, jnp.float16, jnp.bfloat16]
        assert gi_aval.dtype == in_aval.dtype
        ir_in_type = ir.RankedTensorType(dz.type)
        ir_in_shape = ir_in_type.shape
        gi_type = ir.RankedTensorType(x.type)
        gi_shape = gi_type.shape
        assert ir_in_shape == gi_shape

        ir_batch_size = reduce(operator.mul, ir_in_shape[:-1])
        i_hidden_size = ir_in_shape[-1]
        out_dtype = ir_in_type.element_type
        out_shape = gi_shape

        out_types = [
            ir.RankedTensorType.get(out_shape, out_dtype),
        ]
        operands = [dz, x]
        operand_shapes = [ir_in_shape, gi_shape]
        args = CustomCallArgsWrapper(out_types, operands, operand_shapes)

        in_dtype = jax_dtype_to_te_dtype(in_aval.dtype)
        opaque = transformer_engine_jax.pack_common_descriptor((ir_batch_size, i_hidden_size),
                                                               in_dtype, in_dtype)

        out = custom_caller(DGeluPrimitive.name, args, opaque, False)

        return [out]

    @staticmethod
    def impl(dz, x):
        """
        dgelu implementation
        """
        assert DGeluPrimitive.inner_primitive is not None
        dx = DGeluPrimitive.inner_primitive.bind(dz, x)
        return dx

    @staticmethod
    def batcher(batched_args, batch_dims):
        """
        dgelu batcher
        """
        _check_valid_batch_dims(batch_dims)
        assert DGeluPrimitive.outer_primitive is not None
        dz, x = batched_args
        _, x_bdim = batch_dims

        out_bdims = x_bdim
        return DGeluPrimitive.outer_primitive.bind(dz, x), out_bdims

    @staticmethod
    def infer_sharding_from_operands(mesh, arg_infos, result_infos):
        """
        dgelu infer_sharding_from_operands
        """
        del result_infos    # Unused.
        gelu_out_spec = get_padded_spec(arg_infos[1])
        dx_sharding = NamedSharding(mesh, PartitionSpec(*gelu_out_spec))
        return dx_sharding

    @staticmethod
    def partition(mesh, arg_infos, result_infos):
        """
        dgelu partition
        """
        del result_infos
        dx_sharding = NamedSharding(mesh, PartitionSpec(*get_padded_spec(arg_infos[1])))
        arg_shardings = tuple(arg_i.sharding for arg_i in arg_infos)
        out_shardings = dx_sharding
        impl = DGeluPrimitive.impl
        return mesh, impl, out_shardings, arg_shardings


register_primitive(DGeluPrimitive)


def dgelu(inputs: jnp.ndarray, gelu_inputs: jnp.ndarray) -> jnp.ndarray:
    """
    dgelu fusion wrapper
    Return dgeglu(inputs)
    """
    return DGeluPrimitive.outer_primitive.bind(inputs, gelu_inputs)


3445
class GatedGeluPrimitive(BasePrimitive):
3446
    """
3447
    Gated Gelu Froward Primitive
3448
    """
3449
    name = "te_gated_gelu"
3450
    multiple_results = False
3451
3452
3453
    inner_primitive = None
    outer_primitive = None
    impl_static_args = ()
3454
3455

    @staticmethod
3456
    def abstract(x_aval):
3457
        """
3458
        gated_gelu abstract
3459
        """
3460
3461
3462
3463
3464
3465
3466
3467
3468
3469
        dtype = dtypes.canonicalize_dtype(x_aval.dtype)
        assert dtype in [jnp.float32, jnp.float16, jnp.bfloat16]
        x_shape = x_aval.shape
        assert x_shape[-2] == 2    # Assume x in (....., 2, hidden)
        hidden_size = x_shape[-1]
        batch_shapes = x_shape[:-2]
        x_shape = x_aval.shape
        out_aval = core.raise_to_shaped(x_aval)
        out_shape = (batch_shapes) + (hidden_size,)
        out_aval = out_aval.update(shape=out_shape, dtype=dtype)
3470

3471
        return out_aval
3472
3473

    @staticmethod
3474
    def lowering(ctx, x):
3475
        """
3476
        gated_gelu lowering rules
3477
        """
3478
3479
3480
3481
3482
        (x_aval,) = ctx.avals_in
        assert x_aval.dtype in [jnp.float32, jnp.float16, jnp.bfloat16]
        ir_x_type = ir.RankedTensorType(x.type)
        ir_x_shape = ir_x_type.shape
        out_shape = ir_x_shape[:-2] + [ir_x_shape[-1]]
3483

3484
3485
3486
3487
3488
3489
        out_types = [
            ir.RankedTensorType.get(out_shape, ir_x_type.element_type),
        ]
        operands = [x]
        operand_shapes = [ir_x_shape]
        args = CustomCallArgsWrapper(out_types, operands, operand_shapes)
3490

3491
3492
3493
3494
3495
        hidden_size = ir_x_shape[-1]
        batch_size = reduce(operator.mul, ir_x_shape[:-2])
        in_dtype = jax_dtype_to_te_dtype(x_aval.dtype)
        opaque = transformer_engine_jax.pack_common_descriptor((batch_size, hidden_size), in_dtype,
                                                               in_dtype)
3496

3497
        out = custom_caller(GatedGeluPrimitive.name, args, opaque, False)
3498

3499
        return [out]
3500

3501
3502
3503
3504
3505
    @staticmethod
    def impl(x):
        assert GatedGeluPrimitive.inner_primitive is not None
        out = GatedGeluPrimitive.inner_primitive.bind(x)
        return out
3506

3507
3508
3509
3510
3511
3512
3513
3514
3515
    @staticmethod
    def batcher(batched_args, batch_dims):
        """
        gated_gelu batcher
        """
        _check_valid_batch_dims(batch_dims)
        assert GatedGeluPrimitive.outer_primitive is not None
        inputs, = batched_args
        inputs_bdim, = batch_dims
3516

3517
3518
        out_bdims = inputs_bdim
        return GatedGeluPrimitive.outer_primitive.bind(inputs), out_bdims
3519

3520
3521
3522
3523
3524
3525
3526
3527
3528
    @staticmethod
    def infer_sharding_from_operands(mesh, arg_infos, result_infos):
        """
        gated_gelu infer_sharding_from_operands
        """
        del result_infos    # Unused.
        x_spec = get_padded_spec(arg_infos[0])
        out_sharding = NamedSharding(mesh, PartitionSpec(*x_spec[:-2], x_spec[-1]))
        return out_sharding
3529

3530
3531
3532
3533
3534
3535
3536
3537
3538
3539
3540
    @staticmethod
    def partition(mesh, arg_infos, result_infos):
        """
        gated_gelu partitioning
        """
        del result_infos
        x_spec = get_padded_spec(arg_infos[0])
        arg_shardings = tuple(arg_i.sharding for arg_i in arg_infos)
        out_sharding = NamedSharding(mesh, PartitionSpec(*x_spec[:-2], x_spec[-1]))
        impl = GatedGeluPrimitive.impl
        return mesh, impl, out_sharding, arg_shardings
3541
3542


3543
register_primitive(GatedGeluPrimitive)
3544
3545


3546
def gated_gelu(inputs: jnp.ndarray) -> jnp.ndarray:
3547
    """
3548
3549
3550
    gated gelu wrapper
    Return FP8(geglu(inputs))
    Assume inputs has two dimensions shape and the memory layout is (N, 2, H)
3551
    """
3552
    return GatedGeluPrimitive.outer_primitive.bind(inputs)
3553
3554


3555
class DgatedGeluPrimitive(BasePrimitive):
3556
    """
3557
    Dgated Gelu Primitive
3558
    """
3559
3560
3561
3562
3563
    name = "te_dgated_gelu"
    multiple_results = False
    inner_primitive = None
    outer_primitive = None
    impl_static_args = ()
3564
3565

    @staticmethod
3566
    def abstract(dz_aval, x_aval):
3567
        """
3568
        dgated_gelu abstract
3569
        """
3570
3571
3572
3573
3574
        dtype = dtypes.canonicalize_dtype(dz_aval.dtype)
        assert dtype in [jnp.float32, jnp.float16, jnp.bfloat16]
        assert x_aval.dtype == dtype
        for axis in range(len(dz_aval.shape) - 1):
            assert dz_aval.shape[axis] == x_aval.shape[axis]
3575

3576
        assert x_aval.shape[-2] == 2    # Assume x in (....., 2, hidden)
3577

3578
3579
3580
3581
3582
        i_hidden_size = dz_aval.shape[-1]
        g_hidden_size = x_aval.shape[-1]
        assert i_hidden_size == g_hidden_size
        out_aval = core.raise_to_shaped(x_aval)
        return out_aval
3583
3584

    @staticmethod
3585
    def lowering(ctx, dz, x):
3586
        """
3587
        dgated_gelu lowering rules
3588
        """
3589
3590
3591
3592
3593
3594
3595
3596
3597
        in_aval, gi_aval = ctx.avals_in
        assert in_aval.dtype in [jnp.float32, jnp.float16, jnp.bfloat16]
        assert gi_aval.dtype == in_aval.dtype
        ir_in_type = ir.RankedTensorType(dz.type)
        ir_in_shape = ir_in_type.shape
        gi_type = ir.RankedTensorType(x.type)
        gi_shape = gi_type.shape
        for axis in range(len(ir_in_shape) - 1):
            assert ir_in_shape[axis] == gi_shape[axis]
3598

3599
3600
3601
3602
3603
3604
        ir_batch_size = reduce(operator.mul, ir_in_shape[:-1])
        i_hidden_size = ir_in_shape[-1]
        g_hidden_size = gi_shape[-1]
        assert i_hidden_size == g_hidden_size
        out_dtype = ir_in_type.element_type
        out_shape = gi_shape
3605
3606

        out_types = [
3607
            ir.RankedTensorType.get(out_shape, out_dtype),
3608
        ]
3609
3610
        operands = [dz, x]
        operand_shapes = [ir_in_shape, gi_shape]
3611
3612
        args = CustomCallArgsWrapper(out_types, operands, operand_shapes)

3613
3614
3615
        in_dtype = jax_dtype_to_te_dtype(in_aval.dtype)
        opaque = transformer_engine_jax.pack_common_descriptor((ir_batch_size, i_hidden_size),
                                                               in_dtype, in_dtype)
3616

3617
        out = custom_caller(DgatedGeluPrimitive.name, args, opaque, False)
3618
3619
3620
3621

        return [out]

    @staticmethod
3622
3623
3624
3625
3626
3627
3628
    def impl(dz, x):
        """
        dgated_gelu implementation
        """
        assert DgatedGeluPrimitive.inner_primitive is not None
        dx = DgatedGeluPrimitive.inner_primitive.bind(dz, x)
        return dx
3629
3630

    @staticmethod
3631
    def batcher(batched_args, batch_dims):
3632
        """
3633
        dgated_gelu batcher
3634
        """
3635
3636
3637
3638
        _check_valid_batch_dims(batch_dims)
        assert DgatedGeluPrimitive.outer_primitive is not None
        dz, x = batched_args
        _, x_bdim = batch_dims
3639

3640
3641
        out_bdims = x_bdim
        return DgatedGeluPrimitive.outer_primitive.bind(dz, x), out_bdims
3642
3643

    @staticmethod
3644
    def infer_sharding_from_operands(mesh, arg_infos, result_infos):
3645
        """
3646
        dgated_gelu infer_sharding_from_operands
3647
        """
3648
3649
3650
3651
        del result_infos    # Unused.
        gelu_out_spec = get_padded_spec(arg_infos[1])
        dx_sharding = NamedSharding(mesh, PartitionSpec(*gelu_out_spec))
        return dx_sharding
3652

3653
3654
3655
3656
3657
3658
3659
3660
3661
3662
3663
    @staticmethod
    def partition(mesh, arg_infos, result_infos):
        """
        dgated_gelu partition
        """
        del result_infos
        dx_sharding = NamedSharding(mesh, PartitionSpec(*get_padded_spec(arg_infos[1])))
        arg_shardings = tuple(arg_i.sharding for arg_i in arg_infos)
        out_shardings = dx_sharding
        impl = DgatedGeluPrimitive.impl
        return mesh, impl, out_shardings, arg_shardings
3664
3665


3666
register_primitive(DgatedGeluPrimitive)
3667
3668


3669
3670
3671
3672
3673
3674
def dgated_gelu(inputs: jnp.ndarray, gelu_inputs: jnp.ndarray) -> jnp.ndarray:
    """
    dgated_gelu fusion wrapper
    Return dgeglu(inputs)
    """
    return DgatedGeluPrimitive.outer_primitive.bind(inputs, gelu_inputs)
3675
3676


3677
3678
def _normalize_axis_boundary(axis, ndim):
    return axis if axis >= 0 else ndim + axis
3679
3680


3681
def _multidim_transpose(shape, static_axis_boundary, transpose_axis_boundary):
3682
    """
3683
3684
3685
3686
3687
3688
3689
3690
3691
3692
3693
3694
3695
3696
3697
3698
3699
3700
    te_cast_transpose_p multi-dims transpose

    static_axis_boundary: int, Indicate those axes <= static_axis_boundary would not be
        involved into transpose, -1 means all axes involve into transpose.
    transpose_axis_boundary: int, Indicate how to split multi-dimensions tensors to 2D matrix for
        transpose. Note, transpose_axis_boundary should be greater than static_axis_boundary

    examples:
        X in shape (dim0, dim1, dim2, dim3, dim4)

        static_axis_boundary == -1, transpose_axis_boundary == 2
            Xt = (dim2, dim3, dim4, dim0, dim1)

        static_axis_boundary == 0, transpose_axis_boundary == 2
            Xt = (dim0, dim2, dim3, dim4, dim1)

        static_axis_boundary == 0, transpose_axis_boundary == 3
            Xt = (dim0, dim3, dim4, dim1. dim2)
3701
    """
3702
3703
3704
3705
3706
3707
3708
3709
    if static_axis_boundary < 0:
        static_axis_boundary = -1    # means no static axes
    assert static_axis_boundary < len(shape) - 2    # at least 2 remaining for transpose.
    transpose_start_idx = static_axis_boundary + 1
    transpose_axis_boundary = _normalize_axis_boundary(transpose_axis_boundary, len(shape))
    assert transpose_start_idx < transpose_axis_boundary
    return (*shape[:transpose_start_idx], *shape[transpose_axis_boundary:],
            *shape[transpose_start_idx:transpose_axis_boundary])
3710
3711


3712
class CastTransposePrimitive(BasePrimitive):
3713
    """
3714
    Cast Transpose Primitive
3715
    """
3716
3717
3718
3719
3720
    name = "te_cast_transpose"
    multiple_results = True
    impl_static_args = (4, 5, 6)
    inner_primitive = None
    outer_primitive = None
3721
3722

    @staticmethod
3723
3724
3725
3726
3727
3728
3729
3730
3731
3732
3733
3734
3735
3736
3737
3738
3739
3740
3741
    def abstract(x_aval, amax_aval, scale_aval, scale_inv_aval, *, out_dtype, static_axis_boundary,
                 transpose_axis_boundary):
        """
        te_cast_transpose_p abstract
        """
        dtype = dtypes.canonicalize_dtype(x_aval.dtype)
        assert dtype in [jnp.float32, jnp.float16, jnp.bfloat16]
        assert amax_aval.dtype == jnp.float32
        assert scale_aval.dtype == jnp.float32
        assert scale_inv_aval.dtype == jnp.float32

        transposed_x_shape = _multidim_transpose(x_aval.shape, static_axis_boundary,
                                                 transpose_axis_boundary)

        casted_x_aval = x_aval.update(shape=x_aval.shape, dtype=out_dtype)
        casted_xt_aval = x_aval.update(shape=transposed_x_shape, dtype=out_dtype)
        updated_amax_aval = amax_aval.update(shape=amax_aval.shape, dtype=amax_aval.dtype)

        return casted_x_aval, casted_xt_aval, updated_amax_aval
3742
3743

    @staticmethod
3744
3745
    def lowering(ctx, x, amax, scale, scale_inv, *, out_dtype, static_axis_boundary,
                 transpose_axis_boundary):
3746
        """
3747
        te_cast_transpose_p lowering rules
3748
        """
3749
3750
3751
3752
3753
3754
3755
3756
3757
3758
3759
3760
3761
3762
3763
3764
3765
3766
3767
3768
3769
3770
3771
3772
3773
3774
3775
3776
3777
3778
3779
3780
3781
3782
3783
3784
3785
3786
3787
3788
3789
3790
        x_aval, amax_aval, scale_aval, scale_inv_aval = ctx.avals_in
        assert x_aval.dtype in [jnp.float32, jnp.float16, jnp.bfloat16]
        assert amax_aval.dtype == jnp.float32
        assert scale_aval.dtype == jnp.float32
        assert scale_inv_aval.dtype == jnp.float32
        ir_x_type = ir.RankedTensorType(x.type)
        ir_x_shape = ir_x_type.shape
        if static_axis_boundary >= 0:
            for i in range(static_axis_boundary + 1):
                assert ir_x_shape[i] == 1
        ir_out_dtype = jax_dtype_to_ir_dtype(out_dtype)
        ir_amax_type = ir.RankedTensorType(amax.type)
        ir_amax_dtype = ir_amax_type.element_type
        ir_amax_shape = ir_amax_type.shape
        ir_scale_shape = ir_amax_shape
        ir_scale_inv_shape = ir_amax_shape

        transposed_x_shape = _multidim_transpose(ir_x_shape, static_axis_boundary,
                                                 transpose_axis_boundary)

        out_types = [
            ir.RankedTensorType.get(ir_x_shape, ir_out_dtype),
            ir.RankedTensorType.get(transposed_x_shape, ir_out_dtype),
            ir.RankedTensorType.get(ir_amax_shape, ir_amax_dtype),
        ]
        operands = [x, amax, scale, scale_inv]
        operand_shapes = [ir_x_shape, ir_amax_shape, ir_scale_shape, ir_scale_inv_shape]
        args = CustomCallArgsWrapper(out_types, operands, operand_shapes)

        contracted_x_shape = (reduce(operator.mul, ir_x_shape[:transpose_axis_boundary]),
                              reduce(operator.mul, ir_x_shape[transpose_axis_boundary:]))
        opaque = transformer_engine_jax.pack_common_descriptor(contracted_x_shape,
                                                               jax_dtype_to_te_dtype(x_aval.dtype),
                                                               jax_dtype_to_te_dtype(out_dtype))

        out = custom_caller(CastTransposePrimitive.name,
                            args,
                            opaque,
                            False,
                            operand_output_aliases={1: 2})

        return out
3791
3792

    @staticmethod
3793
    def impl(x, amax, scale, scale_inv, out_dtype, static_axis_boundary, transpose_axis_boundary):
3794
        """
3795
        te_cast_transpose implementation
3796
        """
3797
3798
3799
3800
3801
3802
3803
        assert CastTransposePrimitive.inner_primitive is not None
        casted_x, casted_transposed_x, updated_amax = \
            CastTransposePrimitive.inner_primitive.bind(
                x, amax, scale, scale_inv, out_dtype=out_dtype,
                static_axis_boundary=static_axis_boundary,
                transpose_axis_boundary=transpose_axis_boundary)
        return casted_x, casted_transposed_x, updated_amax
3804

3805
3806
3807
3808
3809
3810
    @staticmethod
    def batcher(batched_args, batch_dims, *, out_dtype, static_axis_boundary,
                transpose_axis_boundary):
        _check_valid_batch_dims(batch_dims)
        assert CastTransposePrimitive.outer_primitive is not None
        assert static_axis_boundary < 0
3811

3812
3813
        x, amax, scale, scale_inv = batched_args
        x_bdim, amax_bdim, *_ = batch_dims
3814

3815
3816
3817
        # Minus batch dim.
        transpose_axis_boundary = _normalize_axis_boundary(transpose_axis_boundary, x.ndim - 1)
        transpose_axis_boundary += 1    # Plus batch dim
3818

3819
3820
3821
3822
3823
3824
3825
3826
3827
        out_bdims = x_bdim, x_bdim, amax_bdim
        return CastTransposePrimitive.outer_primitive.bind(
            x,
            amax,
            scale,
            scale_inv,
            out_dtype=out_dtype,
            static_axis_boundary=x_bdim,
            transpose_axis_boundary=transpose_axis_boundary), out_bdims
3828

3829
3830
3831
3832
3833
3834
3835
3836
3837
3838
3839
3840
3841
3842
3843
3844
3845
3846
3847
3848
3849
3850
3851
3852
3853
3854
3855
3856
3857
3858
3859
3860
3861
3862
3863
3864
3865
3866
3867
3868
3869
3870
    @staticmethod
    def infer_sharding_from_operands(out_dtype, static_axis_boundary, transpose_axis_boundary, mesh,
                                     arg_infos, result_infos):
        del out_dtype, result_infos
        x_spec = get_padded_spec(arg_infos[0])
        casted_x_sharding = NamedSharding(mesh, PartitionSpec(*x_spec))
        xt_spec = _multidim_transpose(x_spec, static_axis_boundary, transpose_axis_boundary)
        casted_transposed_x_sharding = NamedSharding(mesh, PartitionSpec(*xt_spec))
        amax_sharding = NamedSharding(mesh, PartitionSpec(*get_padded_spec(arg_infos[1])))
        return (casted_x_sharding, casted_transposed_x_sharding, amax_sharding)

    @staticmethod
    def partition(out_dtype, static_axis_boundary, transpose_axis_boundary, mesh, arg_infos,
                  result_infos):
        del result_infos
        x_spec = get_padded_spec(arg_infos[0])
        casted_x_sharding = NamedSharding(mesh, PartitionSpec(*x_spec))
        xt_spec = _multidim_transpose(x_spec, static_axis_boundary, transpose_axis_boundary)
        casted_transposed_x_sharding = NamedSharding(mesh, PartitionSpec(*xt_spec))
        amax_sharding = NamedSharding(mesh, PartitionSpec(*get_padded_spec(arg_infos[1])))
        arg_shardings = tuple(arg_i.sharding for arg_i in arg_infos)
        out_shardings = (casted_x_sharding, casted_transposed_x_sharding, amax_sharding)

        def sharded_impl(x, amax, scale, scale_inv):
            local_cx, local_cxt, local_updated_amax = \
                CastTransposePrimitive.impl(x, amax, scale, scale_inv,
                     out_dtype=out_dtype,
                       static_axis_boundary=static_axis_boundary,
                       transpose_axis_boundary=transpose_axis_boundary)
            global_updated_amax = all_reduce_max_along_all_axes_except_PP(local_updated_amax)

            return local_cx, local_cxt, global_updated_amax

        return mesh, sharded_impl, out_shardings, arg_shardings


register_primitive(CastTransposePrimitive)


def cast_transpose(x: jnp.ndarray, amax: jnp.ndarray, scale: jnp.ndarray, scale_inv: jnp.ndarray,
                   out_dtype: jnp.dtype, static_axis_boundary: int,
                   transpose_axis_boundary: int) -> Tuple[jnp.ndarray, jnp.ndarray, jnp.ndarray]:
3871
    """
3872
3873
    cast transpose wrapper
    Return two tensors, FP8(inputs) and FP8(inputs.T), which are scaled by `scale`
3874
    """
3875
3876
3877
3878
3879
3880
3881
3882
    return CastTransposePrimitive.outer_primitive.bind(
        x,
        amax,
        scale,
        scale_inv,
        out_dtype=out_dtype,
        static_axis_boundary=static_axis_boundary,
        transpose_axis_boundary=transpose_axis_boundary)
3883
3884


3885
3886
3887
3888
3889
3890
3891
3892
3893
3894
3895
3896
3897
3898
3899
3900
3901
3902
3903
3904
3905
3906
3907
3908
3909
3910
3911
3912
3913
3914
3915
3916
3917
3918
3919
3920
3921
3922
3923
3924
3925
3926
3927
3928
3929
3930
3931
3932
3933
3934
3935
3936
3937
3938
3939
3940
3941
3942
3943
3944
3945
3946
3947
3948
3949
3950
3951
3952
3953
3954
3955
3956
3957
3958
3959
3960
3961
3962
3963
3964
3965
3966
3967
3968
class CastFP8Primitive(BasePrimitive):
    """
    Cast Primitive
    """
    name = "te_quantize"
    multiple_results = True
    impl_static_args = (4,)
    inner_primitive = None
    outer_primitive = None

    @staticmethod
    def abstract(x_aval, amax_aval, scale_aval, scale_inv_aval, *, out_dtype):
        """
        te_cast abstract
        """
        dtype = dtypes.canonicalize_dtype(x_aval.dtype)
        assert dtype in [jnp.float32, jnp.float16, jnp.bfloat16]
        assert amax_aval.dtype == jnp.float32
        assert scale_aval.dtype == jnp.float32
        assert scale_inv_aval.dtype == jnp.float32

        casted_x_aval = x_aval.update(shape=x_aval.shape, dtype=out_dtype)
        updated_amax_aval = amax_aval.update(shape=amax_aval.shape, dtype=amax_aval.dtype)

        return casted_x_aval, updated_amax_aval

    @staticmethod
    def lowering(ctx, x, amax, scale, scale_inv, *, out_dtype):
        """
        te_cast lowering rules
        """
        x_aval, amax_aval, scale_aval, scale_inv_aval = ctx.avals_in
        assert x_aval.dtype in [jnp.float32, jnp.float16, jnp.bfloat16]
        assert amax_aval.dtype == jnp.float32
        assert scale_aval.dtype == jnp.float32
        assert scale_inv_aval.dtype == jnp.float32
        ir_x_type = ir.RankedTensorType(x.type)
        ir_x_shape = ir_x_type.shape
        ir_out_dtype = jax_dtype_to_ir_dtype(out_dtype)
        ir_amax_type = ir.RankedTensorType(amax.type)
        ir_amax_dtype = ir_amax_type.element_type
        ir_amax_shape = ir_amax_type.shape
        ir_scale_shape = ir_amax_shape
        ir_scale_inv_shape = ir_amax_shape

        out_types = [
            ir.RankedTensorType.get(ir_x_shape, ir_out_dtype),
            ir.RankedTensorType.get(ir_amax_shape, ir_amax_dtype),
        ]
        operands = [x, amax, scale, scale_inv]
        operand_shapes = [ir_x_shape, ir_amax_shape, ir_scale_shape, ir_scale_inv_shape]
        args = CustomCallArgsWrapper(out_types, operands, operand_shapes)

        opaque = transformer_engine_jax.pack_common_descriptor(ir_x_shape,
                                                               jax_dtype_to_te_dtype(x_aval.dtype),
                                                               jax_dtype_to_te_dtype(out_dtype))

        out = custom_caller(CastFP8Primitive.name,
                            args,
                            opaque,
                            False,
                            operand_output_aliases={1: 1})

        return out

    @staticmethod
    def impl(x, amax, scale, scale_inv, out_dtype):
        """
        te_cast implementation
        """
        assert CastFP8Primitive.inner_primitive is not None
        casted_x, updated_amax = \
            CastFP8Primitive.inner_primitive.bind(
                x, amax, scale, scale_inv, out_dtype=out_dtype)
        return casted_x, updated_amax

    @staticmethod
    def batcher(batched_args, batch_dims, *, out_dtype):
        _check_valid_batch_dims(batch_dims)
        assert CastFP8Primitive.outer_primitive is not None

        x, amax, scale, scale_inv = batched_args
        x_bdim, amax_bdim, *_ = batch_dims

3969
        out_bdims = x_bdim, amax_bdim
3970
3971
3972
3973
3974
3975
3976
3977
3978
3979
3980
3981
3982
3983
3984
3985
3986
3987
3988
3989
3990
3991
3992
3993
3994
3995
3996
3997
3998
3999
4000
4001
4002
4003
4004
4005
4006
4007
4008
4009
4010
4011
        return CastFP8Primitive.outer_primitive.bind(x, amax, scale, scale_inv,
                                                     out_dtype=out_dtype), out_bdims

    @staticmethod
    def infer_sharding_from_operands(out_dtype, mesh, arg_infos, result_infos):
        del out_dtype, result_infos
        x_spec = get_padded_spec(arg_infos[0])
        casted_x_sharding = NamedSharding(mesh, PartitionSpec(*x_spec))
        amax_sharding = NamedSharding(mesh, PartitionSpec(*get_padded_spec(arg_infos[1])))
        return (casted_x_sharding, amax_sharding)

    @staticmethod
    def partition(out_dtype, mesh, arg_infos, result_infos):
        del result_infos
        x_spec = get_padded_spec(arg_infos[0])
        casted_x_sharding = NamedSharding(mesh, PartitionSpec(*x_spec))
        amax_sharding = NamedSharding(mesh, PartitionSpec(*get_padded_spec(arg_infos[1])))
        arg_shardings = tuple(arg_i.sharding for arg_i in arg_infos)
        out_shardings = (casted_x_sharding, amax_sharding)

        def sharded_impl(x, amax, scale, scale_inv):
            local_cx, local_updated_amax = \
                CastFP8Primitive.impl(x, amax, scale, scale_inv, out_dtype=out_dtype)
            global_updated_amax = all_reduce_max_along_all_axes_except_PP(local_updated_amax)

            return local_cx, global_updated_amax

        return mesh, sharded_impl, out_shardings, arg_shardings


register_primitive(CastFP8Primitive)


def cast_fp8(x: jnp.ndarray, amax: jnp.ndarray, scale: jnp.ndarray, scale_inv: jnp.ndarray,
             out_dtype: TEDType) -> Tuple[jnp.ndarray, jnp.ndarray]:
    """
    Cast wrapper
    Return FP8 tensor
    """
    return CastFP8Primitive.outer_primitive.bind(x, amax, scale, scale_inv, out_dtype=out_dtype)


4012
class TransposePrimitive(BasePrimitive):
4013
    """
4014
    Transpose Primitive
4015
    """
4016
    name = "te_transpose"
4017
    multiple_results = False
4018
4019
4020
    impl_static_args = (1, 2)
    inner_primitive = None
    outer_primitive = None
4021
4022

    @staticmethod
4023
    def abstract(x_aval, *, static_axis_boundary, transpose_axis_boundary):
4024
        """
4025
        _transpose abstract
4026
        """
4027
4028
4029
        transposed_x_shape = _multidim_transpose(x_aval.shape, static_axis_boundary,
                                                 transpose_axis_boundary)
        xt_aval = x_aval.update(shape=transposed_x_shape, dtype=x_aval.dtype)
4030

4031
        return xt_aval
4032
4033

    @staticmethod
4034
    def lowering(ctx, x, *, static_axis_boundary, transpose_axis_boundary):
4035
        """
4036
        _transpose cuda lowering
4037
4038
        """

4039
4040
4041
4042
        x_aval = ctx.avals_in[0]
        assert x_aval.dtype in [
            jnp.float32, jnp.float16, jnp.bfloat16, jnp.float8_e4m3fn, jnp.float8_e5m2
        ]
4043

4044
4045
4046
4047
4048
4049
        ir_x_type = ir.RankedTensorType(x.type)
        ir_x_shape = ir_x_type.shape
        ir_out_dtype = jax_dtype_to_ir_dtype(x_aval.dtype)
        if static_axis_boundary >= 0:
            for i in range(static_axis_boundary + 1):
                assert ir_x_shape[i] == 1
4050

4051
4052
4053
4054
4055
4056
        transposed_x_shape = _multidim_transpose(ir_x_shape, static_axis_boundary,
                                                 transpose_axis_boundary)

        out_types = [ir.RankedTensorType.get(transposed_x_shape, ir_out_dtype)]
        operands = [x]
        operand_shapes = [ir_x_shape]
4057
4058
        args = CustomCallArgsWrapper(out_types, operands, operand_shapes)

4059
4060
4061
4062
4063
        te_dtype = jax_dtype_to_te_dtype(x_aval.dtype)
        contracted_x_shape = (reduce(operator.mul, ir_x_shape[:transpose_axis_boundary]),
                              reduce(operator.mul, ir_x_shape[transpose_axis_boundary:]))
        opaque = transformer_engine_jax.pack_common_descriptor(contracted_x_shape, te_dtype,
                                                               te_dtype)
4064

4065
        out = custom_caller(TransposePrimitive.name, args, opaque, False)
4066
4067
4068

        return [out]

4069
4070
4071
4072
4073
4074
4075
4076
4077
4078
4079
    @staticmethod
    def impl(x, static_axis_boundary, transpose_axis_boundary):
        """
        tcast_transpose implementation
        """
        assert TransposePrimitive.inner_primitive is not None
        transposed_x = \
            TransposePrimitive.inner_primitive.bind(x,
                                                    static_axis_boundary=static_axis_boundary,
                                                    transpose_axis_boundary=transpose_axis_boundary)
        return transposed_x
4080

4081
4082
4083
4084
4085
    @staticmethod
    def batcher(batched_args, batch_dims, *, static_axis_boundary, transpose_axis_boundary):
        _check_valid_batch_dims(batch_dims)
        assert TransposePrimitive.outer_primitive is not None
        assert static_axis_boundary < 0
4086

4087
4088
        x, = batched_args
        x_bdim, = batch_dims
4089

4090
4091
4092
        # Minus batch dim.
        transpose_axis_boundary = _normalize_axis_boundary(transpose_axis_boundary, x.ndim - 1)
        transpose_axis_boundary += 1    # Plus batch dim
4093

4094
4095
4096
4097
        out_bdims = x_bdim
        return TransposePrimitive.outer_primitive.bind(
            x, static_axis_boundary=x_bdim,
            transpose_axis_boundary=transpose_axis_boundary), out_bdims
4098
4099

    @staticmethod
4100
4101
4102
4103
4104
4105
4106
    def infer_sharding_from_operands(static_axis_boundary, transpose_axis_boundary, mesh, arg_infos,
                                     result_infos):
        del result_infos
        x_spec = get_padded_spec(arg_infos[0])
        xt_spec = _multidim_transpose(x_spec, static_axis_boundary, transpose_axis_boundary)
        transposed_x_sharding = NamedSharding(mesh, PartitionSpec(*xt_spec))
        return transposed_x_sharding
4107
4108

    @staticmethod
4109
4110
4111
4112
4113
4114
4115
    def partition(static_axis_boundary, transpose_axis_boundary, mesh, arg_infos, result_infos):
        del result_infos
        x_spec = get_padded_spec(arg_infos[0])
        xt_spec = _multidim_transpose(x_spec, static_axis_boundary, transpose_axis_boundary)
        transposed_x_sharding = NamedSharding(mesh, PartitionSpec(*xt_spec))
        arg_shardings = tuple(arg_i.sharding for arg_i in arg_infos)
        out_shardings = transposed_x_sharding
4116

4117
4118
4119
        impl = partial(TransposePrimitive.impl,
                       static_axis_boundary=static_axis_boundary,
                       transpose_axis_boundary=transpose_axis_boundary)
4120

4121
        return mesh, impl, out_shardings, arg_shardings
4122
4123


4124
register_primitive(TransposePrimitive)
4125
4126


4127
4128
def transpose(x: jnp.ndarray, static_axis_boundary: int,
              transpose_axis_boundary: int) -> jnp.ndarray:
4129
    """
4130
    transpose wrapper
4131
    """
4132
4133
4134
    return TransposePrimitive.outer_primitive.bind(x,
                                                   static_axis_boundary=static_axis_boundary,
                                                   transpose_axis_boundary=transpose_axis_boundary)
4135
4136


4137
class LayerNormFwdFp8Primitive(BasePrimitive):
4138
    """
4139
    Layer Normalization Forward FP8 Primitive
4140
    """
4141
4142
4143
4144
4145
    name = "te_layernorm_forward_fp8"
    multiple_results = True
    impl_static_args = (6, 7, 8)    # out_type, zero_centered_gamma, epsilon
    inner_primitive = None
    outer_primitive = None
4146
4147

    @staticmethod
4148
4149
    def abstract(x_aval, gamma_aval, beta_aval, amax_aval, scale_aval, scale_inv_aval, *, out_dtype,
                 zero_centered_gamma, epsilon):
4150
        """
4151
        LayerNorm fwd (fp8 out) inner primitive abstract
4152
        """
4153
        x_dtype = dtypes.canonicalize_dtype(x_aval.dtype)
4154

4155
4156
4157
4158
        assert x_dtype in [jnp.float32, jnp.float16, jnp.bfloat16]
        assert amax_aval.dtype == jnp.float32
        assert scale_aval.dtype == jnp.float32
        assert scale_inv_aval.dtype == jnp.float32
4159

4160
4161
4162
4163
        mu_rsigama_dtype = jnp.float32

        assert gamma_aval.size == beta_aval.size

4164
        wkspace_info, barrier_info = transformer_engine_jax.get_layernorm_fwd_workspace_sizes(
4165
4166
4167
4168
            x_aval.size // gamma_aval.size,    # batch size
            gamma_aval.size,    # hidden size
            jax_dtype_to_te_dtype(x_aval.dtype),    # in type
            jax_dtype_to_te_dtype(gamma_aval.dtype),    # weight type
4169
            jax_dtype_to_te_dtype(out_dtype),
4170
4171
4172
            True,
            zero_centered_gamma,
            epsilon)
4173

4174
4175
4176
        out_aval = x_aval.update(shape=x_aval.shape, dtype=out_dtype)
        mu_aval = rsigma_aval = out_aval.update(shape=out_aval.shape[:-1], dtype=mu_rsigama_dtype)
        updated_amax_aval = amax_aval.update(shape=amax_aval.shape, dtype=amax_aval.dtype)
4177
4178
4179
4180
4181
4182
        wkspace_aval = x_aval.update(shape=wkspace_info[0],
                                     dtype=te_dtype_to_jax_dtype(wkspace_info[1]))
        barrier_aval = x_aval.update(shape=barrier_info[0],
                                     dtype=te_dtype_to_jax_dtype(barrier_info[1]))

        return out_aval, mu_aval, rsigma_aval, updated_amax_aval, wkspace_aval, barrier_aval
4183

4184
4185
4186
4187
4188
4189
4190
    @staticmethod
    def outer_abstract(*args, **kwargs):
        """
        LayerNorm fwd (fp8 out) outer primitive abstract
        """
        out_aval, mu_aval, rsigma_aval, updated_amax_aval, _, _ = \
            LayerNormFwdFp8Primitive.abstract(*args, **kwargs)
4191
        return out_aval, mu_aval, rsigma_aval, updated_amax_aval
4192
4193

    @staticmethod
4194
4195
    def lowering(ctx, x, gamma, beta, amax, scale, scale_inv, *, out_dtype, zero_centered_gamma,
                 epsilon):
4196
        """
4197
        LayerNorm fwd (fp8 out) lowering rules
4198
        """
4199
        x_aval, gamma_aval, beta_aval, amax_aval, scale_aval, scale_inv_aval = ctx.avals_in
4200

4201
4202
        # Currently only support casting to E4M3 only in C side.
        assert out_dtype == jnp.float8_e4m3fn
4203

4204
4205
4206
4207
4208
        assert x_aval.dtype in [jnp.float32, jnp.float16, jnp.bfloat16]
        assert gamma_aval.dtype == beta_aval.dtype
        assert amax_aval.dtype == jnp.float32
        assert scale_aval.dtype == jnp.float32
        assert scale_inv_aval.dtype == jnp.float32
4209

4210
4211
4212
4213
4214
4215
        x_type = ir.RankedTensorType(x.type)
        x_shape = x_type.shape
        g_type = ir.RankedTensorType(gamma.type)
        g_shape = g_type.shape
        b_type = ir.RankedTensorType(beta.type)
        b_shape = b_type.shape
4216

4217
4218
        assert g_type == b_type
        assert g_shape == b_shape
4219

4220
4221
4222
4223
4224
4225
4226
4227
        ir_out_dtype = jax_dtype_to_ir_dtype(out_dtype)
        ir_mu_dtype = ir.F32Type.get()
        ir_rsigma_dtype = ir.F32Type.get()
        ir_amax_type = ir.RankedTensorType(amax.type)
        ir_amax_dtype = ir_amax_type.element_type
        ir_amax_shape = ir_amax_type.shape
        ir_scale_shape = ir_amax_shape
        ir_scale_inv_shape = ir_amax_shape
4228

4229
4230
4231
4232
        out_shape = x_shape
        hidden_size = reduce(operator.mul, g_shape)
        batch_shape = out_shape[:-1]
        batch_size = reduce(operator.mul, x_shape) // hidden_size
4233

4234
4235
        wkspace_aval, barrier_aval = ctx.avals_out[-2:]

4236
4237
4238
4239
4240
        out_types = [
            ir.RankedTensorType.get(out_shape, ir_out_dtype),
            ir.RankedTensorType.get(batch_shape, ir_mu_dtype),
            ir.RankedTensorType.get(batch_shape, ir_rsigma_dtype),
            ir.RankedTensorType.get(ir_amax_shape, ir_amax_dtype),
4241
4242
            ir.RankedTensorType.get(wkspace_aval.shape, jax_dtype_to_ir_dtype(wkspace_aval.dtype)),
            ir.RankedTensorType.get(barrier_aval.shape, jax_dtype_to_ir_dtype(barrier_aval.dtype))
4243
4244
4245
4246
4247
4248
        ]
        operands = [x, gamma, beta, amax, scale, scale_inv]
        operand_shapes = [
            x_shape, g_shape, b_shape, ir_amax_shape, ir_scale_shape, ir_scale_inv_shape
        ]
        args = CustomCallArgsWrapper(out_types, operands, operand_shapes)
4249

4250
4251
        sm_margin = int(os.getenv("NVTE_FWD_LAYERNORM_SM_MARGIN", "0"))

4252
4253
4254
        opaque = transformer_engine_jax.pack_norm_descriptor(
            batch_size,
            hidden_size,
4255
4256
            wkspace_aval.size,
            barrier_aval.size,
4257
4258
            0,    # no dgamma_part in FWD pass
            0,    # no dbeta_part in BWD pass
4259
4260
            jax_dtype_to_te_dtype(x_aval.dtype),
            jax_dtype_to_te_dtype(gamma_aval.dtype),
4261
4262
            jax_dtype_to_te_dtype(wkspace_aval.dtype),
            jax_dtype_to_te_dtype(barrier_aval.dtype),
4263
4264
            TEDType.kByte,    # dummy dgamma_part te_dtype
            TEDType.kByte,    # dummy dbeta_part te_dtype
4265
4266
            zero_centered_gamma,
            epsilon,
4267
            sm_margin,
4268
        )
4269

4270
4271
4272
4273
4274
        out = custom_caller(LayerNormFwdFp8Primitive.name,
                            args,
                            opaque,
                            False,
                            operand_output_aliases={3: 3})
4275

4276
        return out
4277
4278

    @staticmethod
4279
    def impl(x, gamma, beta, amax, scale, scale_inv, out_dtype, zero_centered_gamma, epsilon):
4280
        """
4281
        to describe implementation
4282
        """
4283
        assert LayerNormFwdFp8Primitive.inner_primitive is not None
4284
        out, mu, rsigma, updated_amax, _, _ = LayerNormFwdFp8Primitive.inner_primitive.bind(
4285
4286
4287
4288
4289
4290
4291
4292
4293
4294
            x,
            gamma,
            beta,
            amax,
            scale,
            scale_inv,
            out_dtype=out_dtype,
            zero_centered_gamma=zero_centered_gamma,
            epsilon=epsilon)
        return out, mu, rsigma, updated_amax
4295
4296

    @staticmethod
4297
    def batcher(batched_args, batch_dims, *, out_dtype, zero_centered_gamma, epsilon):
4298
        """
4299
        to describe batch rules for vmap
4300
        """
4301
4302
4303
4304
4305
4306
4307
4308
4309
4310
4311
4312
4313
4314
4315
4316
4317
4318
4319
4320
4321
4322
4323
4324
4325
4326
4327
4328
4329
4330
4331
4332
4333
4334
4335
4336
4337
4338
4339
4340
4341
4342
4343
4344
4345
4346
4347
4348
4349
4350
4351
4352
4353
        _check_valid_batch_dims(batch_dims)
        assert LayerNormFwdFp8Primitive.outer_primitive is not None
        x, gamma, beta, amax, scale, scale_inv = batched_args
        x_bdim, _, _, amax_bdim, _, _ = batch_dims

        out_bdims = x_bdim, x_bdim, x_bdim, amax_bdim
        return LayerNormFwdFp8Primitive.outer_primitive.bind(
            x,
            gamma,
            beta,
            amax,
            scale,
            scale_inv,
            out_dtype=out_dtype,
            zero_centered_gamma=zero_centered_gamma,
            epsilon=epsilon), out_bdims

    @staticmethod
    def infer_sharding_from_operands(out_dtype, zero_centered_gamma, epsilon, mesh, arg_infos,
                                     result_infos):
        del out_dtype, zero_centered_gamma, epsilon, result_infos
        x_spec = get_padded_spec(arg_infos[0])
        if x_spec[-1] is not None:
            warnings.warn(
                f"Does not support to shard hidden dim in {LayerNormFwdPrimitive.name}! " \
                f"Force to not shard the hidden dim, which might introduce extra collective ops, " \
                f"and hurt performance.")

        out_sharding = NamedSharding(mesh, PartitionSpec(*x_spec[:-1], None))
        mu_sharding = rsigma_sharding = NamedSharding(mesh, PartitionSpec(*x_spec[:-1]))
        amax_sharding = NamedSharding(mesh, PartitionSpec(*get_padded_spec(arg_infos[3])))
        return (out_sharding, mu_sharding, rsigma_sharding, amax_sharding)

    @staticmethod
    def partition(out_dtype, zero_centered_gamma, epsilon, mesh, arg_infos, result_infos):
        del result_infos
        x_spec = get_padded_spec(arg_infos[0])
        if x_spec[-1] is not None:
            warnings.warn(
                f"Does not support to shard hidden dim in {LayerNormFwdPrimitive.name}! " \
                f"Force to not shard the hidden dim, which might introduce extra collective ops, " \
                f"and hurt performance."
            )
        x_sharding = NamedSharding(mesh, PartitionSpec(*x_spec[:-1], None))
        g_sharding = NamedSharding(mesh, PartitionSpec(*get_padded_spec(arg_infos[1])))
        b_sharding = NamedSharding(mesh, PartitionSpec(*get_padded_spec(arg_infos[2])))
        out_sharding = x_sharding
        mu_sharding = rsigma_sharding = NamedSharding(
            mesh, PartitionSpec(*get_padded_spec(arg_infos[0])[:-1]))
        amax_sharding = NamedSharding(mesh, PartitionSpec(*get_padded_spec(arg_infos[3])))
        fp8_meta_sharding = amax_sharding
        arg_shardings = (x_sharding, g_sharding, b_sharding) + (fp8_meta_sharding,) * 3
        out_shardings = (out_sharding, mu_sharding, rsigma_sharding, amax_sharding)
4354

4355
4356
4357
4358
4359
4360
4361
        def sharded_impl(x, gamma, beta, amax, scale, scale_inv):
            local_x, local_mu, local_rsigma, local_amax = \
                LayerNormFwdFp8Primitive.impl(x, gamma, beta, amax, scale, scale_inv,
                                            out_dtype=out_dtype,
                                            zero_centered_gamma=zero_centered_gamma,
                                            epsilon=epsilon)
            global_updated_amax = all_reduce_max_along_all_axes_except_PP(local_amax)
4362

4363
            return local_x, local_mu, local_rsigma, global_updated_amax
4364

4365
        return mesh, sharded_impl, out_shardings, arg_shardings
4366

4367
4368
4369
4370
4371
4372
4373

register_primitive(LayerNormFwdFp8Primitive)


def layernorm_fwd_fp8(x: jnp.ndarray, gamma: jnp.ndarray, beta: jnp.ndarray, amax: jnp.ndarray,
                      scale: jnp.ndarray, scale_inv: jnp.ndarray, out_dtype: jnp.dtype,
                      zero_centered_gamma: bool, epsilon: float):
4374
    """
4375
    Wrapper for TE layernorm fwd (fp8 out)
4376
    """
4377
4378
4379
4380
4381
4382
4383
4384
4385
    return LayerNormFwdFp8Primitive.outer_primitive.bind(x,
                                                         gamma,
                                                         beta,
                                                         amax,
                                                         scale,
                                                         scale_inv,
                                                         out_dtype=out_dtype,
                                                         zero_centered_gamma=zero_centered_gamma,
                                                         epsilon=epsilon)
4386
4387


4388
class RmsNormFwdFp8Primitive(BasePrimitive):
4389
    """
4390
    RMS Normalization Forward FP8 Primitive
4391
    """
4392
4393
4394
4395
4396
    name = "te_rmsnorm_forward_fp8"
    multiple_results = True
    impl_static_args = (5, 6)    # out_dtype, epsilon
    inner_primitive = None
    outer_primitive = None
4397

4398
4399
    @staticmethod
    def abstract(x_aval, gamma_aval, amax_aval, scale_aval, scale_inv_aval, out_dtype, epsilon):
4400
        """
4401
        RMSNorm fwd (fp8 out) inner primitive abstract
4402
        """
4403
        x_dtype = dtypes.canonicalize_dtype(x_aval.dtype)
4404

4405
4406
4407
4408
        assert x_dtype in [jnp.float32, jnp.float16, jnp.bfloat16]
        assert amax_aval.dtype == jnp.float32
        assert scale_aval.dtype == jnp.float32
        assert scale_inv_aval.dtype == jnp.float32
4409

4410
4411
        hidden_size = gamma_aval.size
        assert x_aval.size % hidden_size == 0
4412

4413
        rsigama_dtype = jnp.float32
4414

4415
        wkspace_info, barrier_info = transformer_engine_jax.get_layernorm_fwd_workspace_sizes(
4416
            x_aval.size // hidden_size,    # batch_size
4417
            hidden_size,
4418
4419
4420
4421
4422
4423
            jax_dtype_to_te_dtype(x_aval.dtype),    # in te_dtype
            jax_dtype_to_te_dtype(gamma_aval.dtype),    # weight te_dtype
            jax_dtype_to_te_dtype(out_dtype),    # out te_dtype
            False,
            False,
            epsilon)
4424

4425
4426
4427
        out_aval = x_aval.update(shape=x_aval.shape, dtype=out_dtype)
        rsigma_aval = out_aval.update(shape=out_aval.shape[:-1], dtype=rsigama_dtype)
        amax_aval = out_aval.update(shape=amax_aval.shape, dtype=amax_aval.dtype)
4428
4429
4430
4431
4432
4433
        wkspace_aval = x_aval.update(shape=wkspace_info[0],
                                     dtype=te_dtype_to_jax_dtype(wkspace_info[1]))
        barrier_aval = x_aval.update(shape=barrier_info[0],
                                     dtype=te_dtype_to_jax_dtype(barrier_info[1]))

        return out_aval, rsigma_aval, amax_aval, wkspace_aval, barrier_aval
4434

4435
4436
4437
4438
4439
4440
    @staticmethod
    def outer_abstract(*args, **kwargs):
        """
        RMSNorm fwd (fp8 out) outer primitive abstract
        """
        out_aval, rsigma_aval, amax_aval, _, _ = RmsNormFwdFp8Primitive.abstract(*args, **kwargs)
4441
        return out_aval, rsigma_aval, amax_aval
4442
4443

    @staticmethod
4444
    def lowering(ctx, x, gamma, amax, scale, scale_inv, *, out_dtype, epsilon):
4445
        """
4446
        RMSNorm fwd (fp8 out) lowering rules
4447
4448
        """

4449
4450
        # Currently only support casting to E4M3 only in C side.
        assert out_dtype == jnp.float8_e4m3fn
4451

4452
        x_aval, gamma_aval, amax_aval, scale_aval, scale_inv_aval = ctx.avals_in
4453

4454
4455
4456
4457
        assert x_aval.dtype in [jnp.float32, jnp.float16, jnp.bfloat16]
        assert amax_aval.dtype == jnp.float32
        assert scale_aval.dtype == jnp.float32
        assert scale_inv_aval.dtype == jnp.float32
4458

4459
4460
4461
4462
4463
4464
4465
4466
4467
4468
4469
4470
4471
4472
4473
4474
4475
        x_type = ir.RankedTensorType(x.type)
        x_shape = x_type.shape
        g_type = ir.RankedTensorType(gamma.type)
        g_shape = g_type.shape

        ir_out_dtype = jax_dtype_to_ir_dtype(out_dtype)
        ir_rsigma_dtype = ir.F32Type.get()
        ir_amax_type = ir.RankedTensorType(amax.type)
        ir_amax_dtype = ir_amax_type.element_type
        ir_amax_shape = ir_amax_type.shape
        ir_scale_shape = ir_amax_shape
        ir_scale_inv_shape = ir_amax_shape

        out_shape = x_shape
        hidden_size = reduce(operator.mul, g_shape)
        batch_shape = out_shape[:-1]
        batch_size = reduce(operator.mul, x_shape) // hidden_size
4476

4477
4478
        wkspace_aval, barrier_aval = ctx.avals_out[-2:]

4479
4480
4481
4482
        out_types = [
            ir.RankedTensorType.get(out_shape, ir_out_dtype),
            ir.RankedTensorType.get(batch_shape, ir_rsigma_dtype),
            ir.RankedTensorType.get(ir_amax_shape, ir_amax_dtype),
4483
4484
            ir.RankedTensorType.get(wkspace_aval.shape, jax_dtype_to_ir_dtype(wkspace_aval.dtype)),
            ir.RankedTensorType.get(barrier_aval.shape, jax_dtype_to_ir_dtype(barrier_aval.dtype))
4485
4486
4487
4488
4489
        ]
        operands = [x, gamma, amax, scale, scale_inv]
        operand_shapes = [x_shape, g_shape, ir_amax_shape, ir_scale_shape, ir_scale_inv_shape]
        args = CustomCallArgsWrapper(out_types, operands, operand_shapes)

4490
4491
        sm_margin = int(os.getenv("NVTE_FWD_LAYERNORM_SM_MARGIN", "0"))

4492
4493
4494
        opaque = transformer_engine_jax.pack_norm_descriptor(
            batch_size,
            hidden_size,
4495
4496
            wkspace_aval.size,
            barrier_aval.size,
4497
4498
            0,    # no dgamma_part in FWD pass
            0,    # no dbeta_part in BWD pass
4499
4500
            jax_dtype_to_te_dtype(x_aval.dtype),
            jax_dtype_to_te_dtype(gamma_aval.dtype),
4501
4502
            jax_dtype_to_te_dtype(wkspace_aval.dtype),
            jax_dtype_to_te_dtype(barrier_aval.dtype),
4503
4504
            TEDType.kByte,    # dummy dgamma_part te_dtype
            TEDType.kByte,    # dummy dbeta_part te_dtype
4505
4506
            False,    # RMSNorm doesn't support zero_centered_gamma
            epsilon,
4507
            sm_margin,
4508
4509
        )

4510
4511
4512
4513
4514
4515
4516
4517
        out = custom_caller(RmsNormFwdFp8Primitive.name,
                            args,
                            opaque,
                            False,
                            operand_output_aliases={2: 2})

        return out

4518
    @staticmethod
4519
    def impl(x, gamma, amax, scale, scale_inv, out_dtype, epsilon):
4520
        """
4521
        to describe implementation
4522
        """
4523
        assert RmsNormFwdFp8Primitive.inner_primitive is not None
4524
4525
4526
4527
4528
4529
4530
        out, rsigma, amax, _, _ = RmsNormFwdFp8Primitive.inner_primitive.bind(x,
                                                                              gamma,
                                                                              amax,
                                                                              scale,
                                                                              scale_inv,
                                                                              out_dtype=out_dtype,
                                                                              epsilon=epsilon)
4531
        return out, rsigma, amax
4532

4533
4534
4535
4536
4537
4538
4539
4540
4541
4542
4543
4544
4545
4546
4547
4548
4549
    @staticmethod
    def batcher(batched_args, batch_dims, *, out_dtype, epsilon):
        """
        to describe batch rules for vmap
        """
        _check_valid_batch_dims(batch_dims)
        assert RmsNormFwdFp8Primitive.outer_primitive is not None
        x, gamma, amax, scale, scale_inv = batched_args
        x_bdim, _, amax_bdim, _, _ = batch_dims
        out_bdims = x_bdim, x_bdim, amax_bdim
        return RmsNormFwdFp8Primitive.outer_primitive.bind(x,
                                                           gamma,
                                                           amax,
                                                           scale,
                                                           scale_inv,
                                                           out_dtype=out_dtype,
                                                           epsilon=epsilon), out_bdims
4550

4551
4552
4553
4554
4555
4556
4557
4558
4559
4560
4561
4562
4563
4564
    @staticmethod
    def infer_sharding_from_operands(out_dtype, epsilon, mesh, arg_infos, result_infos):
        del out_dtype, epsilon, result_infos
        x_spec = get_padded_spec(arg_infos[0])
        if x_spec[-1] is not None:
            warnings.warn(
                f"Does not support to shard hidden dim in {RmsNormFwdFp8Primitive.name}! " \
                f"Force to not shard the hidden dim, which might introduce extra collective ops, " \
                f"and hurt performance."
            )
        out_sharding = NamedSharding(mesh, PartitionSpec(*x_spec[:-1], None))
        rsigma_sharding = NamedSharding(mesh, PartitionSpec(*x_spec[:-1]))
        amax_sharding = NamedSharding(mesh, PartitionSpec(*get_padded_spec(arg_infos[2])))
        return (out_sharding, rsigma_sharding, amax_sharding)
4565

4566
4567
4568
4569
4570
4571
4572
4573
4574
4575
4576
4577
4578
4579
4580
4581
4582
4583
    @staticmethod
    def partition(out_dtype, epsilon, mesh, arg_infos, result_infos):
        del result_infos
        x_spec = get_padded_spec(arg_infos[0])
        if x_spec[-1] is not None:
            warnings.warn(
                f"Does not support to shard hidden dim in {RmsNormFwdFp8Primitive.name}! " \
                f"Force to not shard the hidden dim, which might introduce extra collective ops, " \
                f"and hurt performance."
            )
        x_sharding = NamedSharding(mesh, PartitionSpec(*x_spec[:-1], None))
        g_sharding = NamedSharding(mesh, PartitionSpec(*get_padded_spec(arg_infos[1])))
        out_sharding = x_sharding
        rsigma_sharding = NamedSharding(mesh, PartitionSpec(*get_padded_spec(arg_infos[0])[:-1]))
        amax_sharding = NamedSharding(mesh, PartitionSpec(*get_padded_spec(arg_infos[2])))
        fp8_meta_sharding = amax_sharding
        arg_shardings = (x_sharding, g_sharding) + (fp8_meta_sharding,) * 3
        out_shardings = (out_sharding, rsigma_sharding, amax_sharding)
4584

4585
4586
4587
4588
4589
        def sharded_impl(x, gamma, amax, scale, scale_inv):
            local_x, local_rsigma, local_amax= \
                RmsNormFwdFp8Primitive.impl(x, gamma, amax, scale, scale_inv,
                                            out_dtype=out_dtype, epsilon=epsilon)
            global_updated_amax = all_reduce_max_along_all_axes_except_PP(local_amax)
4590

4591
            return local_x, local_rsigma, global_updated_amax
4592

4593
        return mesh, sharded_impl, out_shardings, arg_shardings
4594
4595


4596
register_primitive(RmsNormFwdFp8Primitive)
4597

4598
4599
4600

def rmsnorm_fwd_fp8(x: jnp.ndarray, gamma: jnp.ndarray, amax: jnp.ndarray, scale: jnp.ndarray,
                    scale_inv: jnp.ndarray, out_dtype: jnp.dtype, epsilon: float):
4601
    """
4602
    Wrapper for TE rmsnorm fwd (fp8 out)
4603
    """
4604
4605
4606
4607
4608
4609
4610
    return RmsNormFwdFp8Primitive.outer_primitive.bind(x,
                                                       gamma,
                                                       amax,
                                                       scale,
                                                       scale_inv,
                                                       out_dtype=out_dtype,
                                                       epsilon=epsilon)
4611
4612


4613
4614
4615
4616
4617
4618
4619
4620
4621
4622
4623
4624
4625
4626
4627
4628
4629
4630
4631
4632
4633
4634
4635
4636
4637
4638
4639
4640
4641
4642
4643
4644
4645
4646
4647
4648
4649
4650
4651
4652
4653
4654
4655
4656
4657
4658
4659
4660
4661
4662
4663
4664
4665
4666
4667
4668
4669
4670
4671
4672
4673
4674
4675
4676
4677
4678
4679
4680
4681
4682
4683
4684
4685
4686
4687
4688
4689
4690
4691
4692
4693
4694
4695
4696
4697
4698
4699
4700
4701
4702
4703
4704
4705
4706
4707
4708
4709
4710
4711
4712
4713
4714
4715
4716
4717
4718
4719
4720
4721
4722
4723
4724
4725
4726
4727
4728
4729
4730
4731
4732
4733
4734
4735
4736
4737
4738
4739
4740
4741
4742
4743
4744
4745
4746
4747
4748
4749
4750
4751
4752
4753
4754
4755
4756
4757
4758
4759
4760
4761
4762
4763
4764
4765
4766
4767
4768
4769
4770
4771
4772
4773
4774
4775
4776
4777
4778
4779
4780
4781
4782
4783
4784
4785
4786
4787
4788
4789
4790
4791
4792
4793
4794
4795
4796
4797
4798
4799
4800
4801
4802
4803
4804
4805
4806
4807
4808
4809
4810
4811
4812
4813
4814
4815
4816
4817
4818
4819
4820
4821
4822
4823
4824
4825
4826
4827
4828
4829
4830
4831
4832
4833
4834
4835
4836
4837
4838
4839
4840
4841
4842
4843
4844
4845
4846
4847
4848
4849
4850
4851
4852
4853
4854
4855
4856
4857
4858
4859
4860
4861
4862
4863
4864
4865
4866
4867
4868
4869
4870
4871
4872
4873
4874
4875
4876
4877
4878
4879
4880
4881
4882
4883
4884
4885
4886
4887
4888
4889
4890
4891
4892
4893
4894
4895
4896
4897
4898
4899
4900
4901
4902
4903
4904
4905
4906
4907
4908
4909
4910
4911
4912
4913
4914
4915
4916
4917
4918
4919
4920
4921
4922
4923
4924
4925
4926
4927
4928
4929
4930
4931
4932
4933
4934
4935
4936
4937
4938
4939
4940
4941
4942
4943
4944
4945
4946
4947
4948
4949
4950
4951
4952
4953
4954
4955
4956
4957
4958
4959
4960
4961
4962
4963
4964
4965
4966
4967
4968
4969
4970
4971
4972
4973
4974
4975
4976
4977
4978
4979
4980
4981
4982
4983
4984
4985
class GeluFp8Primitive(BasePrimitive):
    """
    Gelu FP8 Primitive
    """
    name = "te_gelu_fp8"
    multiple_results = True
    impl_static_args = (4,)    #out_dtype
    inner_primitive = None
    outer_primitive = None

    @staticmethod
    def abstract(x_aval, amax_aval, scale_aval, scale_inv_aval, *, out_dtype):
        """
        te_gelu_p abstract
        """
        dtype = dtypes.canonicalize_dtype(x_aval.dtype)
        # Currently only support casting to E4M3 only in C side.
        assert out_dtype == jnp.float8_e4m3fn
        assert dtype in [jnp.float32, jnp.float16, jnp.bfloat16]
        assert amax_aval.dtype == jnp.float32
        assert scale_aval.dtype == jnp.float32
        assert scale_inv_aval.dtype == jnp.float32

        out_aval = x_aval.update(shape=x_aval.shape, dtype=out_dtype)
        updated_amax_aval = amax_aval.update(shape=amax_aval.shape, dtype=amax_aval.dtype)

        return out_aval, updated_amax_aval

    @staticmethod
    def lowering(ctx, x, amax, scale, scale_inv, *, out_dtype):
        """
        te_gated_gelu_p lowering rules
        """
        x_aval, amax_aval, scale_aval, scale_inv_aval = ctx.avals_in
        assert x_aval.dtype in [jnp.float32, jnp.float16, jnp.bfloat16]
        assert amax_aval.dtype == jnp.float32
        assert scale_aval.dtype == jnp.float32
        assert scale_inv_aval.dtype == jnp.float32
        ir_x_type = ir.RankedTensorType(x.type)
        ir_x_shape = ir_x_type.shape
        ir_out_dtype = jax_dtype_to_ir_dtype(out_dtype)
        ir_amax_type = ir.RankedTensorType(amax.type)
        ir_amax_dtype = ir_amax_type.element_type
        ir_amax_shape = ir_amax_type.shape
        ir_scale_shape = ir_amax_shape
        ir_scale_inv_shape = ir_amax_shape

        hidden_size = ir_x_shape[-1]
        batch_size = reduce(operator.mul, ir_x_shape[:-1])
        out_shape = ir_x_shape
        out_types = [
            ir.RankedTensorType.get(out_shape, ir_out_dtype),
            ir.RankedTensorType.get(ir_amax_shape, ir_amax_dtype),
        ]
        operands = [x, amax, scale, scale_inv]
        operand_shapes = [ir_x_shape, ir_amax_shape, ir_scale_shape, ir_scale_inv_shape]
        args = CustomCallArgsWrapper(out_types, operands, operand_shapes)

        opaque = transformer_engine_jax.pack_common_descriptor((batch_size, hidden_size),
                                                               jax_dtype_to_te_dtype(x_aval.dtype),
                                                               jax_dtype_to_te_dtype(out_dtype))

        out = custom_caller(GeluFp8Primitive.name,
                            args,
                            opaque,
                            False,
                            operand_output_aliases={1: 1})

        return out

    @staticmethod
    def impl(x, amax, scale, scale_inv, out_dtype):
        """
        to describe implementation
        """
        assert GeluFp8Primitive.inner_primitive is not None
        out, updated_amax = GeluFp8Primitive.inner_primitive.bind(x,
                                                                  amax,
                                                                  scale,
                                                                  scale_inv,
                                                                  out_dtype=out_dtype)
        return out, updated_amax

    @staticmethod
    def batcher(batched_args, batch_dims, *, out_dtype):
        """
        to describe batch rules for vmap
        """
        _check_valid_batch_dims(batch_dims)
        assert GeluFp8Primitive.outer_primitive is not None
        x, amax, scale, scale_inv = batched_args
        x_bdim, amax_bdim, _, _ = batch_dims

        out_bdims = x_bdim, amax_bdim
        return GeluFp8Primitive.outer_primitive.bind(x, amax, scale, scale_inv,
                                                     out_dtype=out_dtype), out_bdims

    @staticmethod
    def infer_sharding_from_operands(out_dtype, mesh, arg_infos, result_infos):
        del out_dtype, result_infos
        x_spec = get_padded_spec(arg_infos[0])
        out_sharding = NamedSharding(mesh, PartitionSpec(*x_spec))
        amax_sharding = NamedSharding(mesh, PartitionSpec(*get_padded_spec(arg_infos[1])))
        return (out_sharding, amax_sharding)

    @staticmethod
    def partition(out_dtype, mesh, arg_infos, result_infos):
        del result_infos
        x_spec = get_padded_spec(arg_infos[0])
        out_sharding = NamedSharding(mesh, PartitionSpec(*x_spec))
        amax_sharding = NamedSharding(mesh, PartitionSpec(*get_padded_spec(arg_infos[1])))
        arg_shardings = tuple(arg_i.sharding for arg_i in arg_infos)
        out_shardings = (out_sharding, amax_sharding)

        def sharded_impl(x, amax, scale, scale_inv):
            local_x, local_amax = GeluFp8Primitive.impl(x,
                                                        amax,
                                                        scale,
                                                        scale_inv,
                                                        out_dtype=out_dtype)
            global_updated_amax = all_reduce_max_along_all_axes_except_PP(local_amax)

            return local_x, global_updated_amax

        return mesh, sharded_impl, out_shardings, arg_shardings


register_primitive(GeluFp8Primitive)


def gelu_fp8(x: jnp.ndarray, amax: jnp.ndarray, scale: jnp.ndarray, scale_inv: jnp.ndarray,
             out_dtype: jnp.dtype) -> Tuple[jnp.ndarray, jnp.ndarray, jnp.ndarray]:
    """
    gated gelu wrapper
    Return FP8(geglu(x))
    """
    return GeluFp8Primitive.outer_primitive.bind(x, amax, scale, scale_inv, out_dtype=out_dtype)


class DGeluDBiasCastTransposePrimitive(BasePrimitive):
    """
    DGelu DBias Cast Transpose Primitive
    """
    name = "te_dgelu_dbias_cast_transpose"
    multiple_results = True
    # out_dtype, static_axis_boundary, transpose_axis_boundary
    impl_static_args = (5, 6, 7)
    inner_primitive = None
    outer_primitive = None

    @staticmethod
    def abstract(dz_aval, x_aval, amax_aval, scale_aval, scale_inv_aval, *, out_dtype,
                 static_axis_boundary, transpose_axis_boundary):
        """
        te_dgelu_dbais_cast_transpose_p abstract
        """
        dtype = dtypes.canonicalize_dtype(dz_aval.dtype)
        assert dtype in [jnp.float32, jnp.float16, jnp.bfloat16]
        assert x_aval.dtype == dtype
        assert amax_aval.dtype == jnp.float32
        assert scale_aval.dtype == jnp.float32
        assert scale_inv_aval.dtype == jnp.float32
        ir_hidden_szie = dz_aval.shape[-1]
        gi_hidden_size = x_aval.shape[-1]
        assert ir_hidden_szie == gi_hidden_size
        t_shape = _multidim_transpose(x_aval.shape, static_axis_boundary, transpose_axis_boundary)
        out = dz_aval.update(shape=x_aval.shape, dtype=out_dtype)
        t_out = dz_aval.update(shape=t_shape, dtype=out_dtype)

        dbias_shape = (*x_aval.shape[:static_axis_boundary + 1], gi_hidden_size)
        dbias = dz_aval.update(shape=dbias_shape, dtype=dtype)

        updated_amax_aval = amax_aval.update(shape=amax_aval.shape, dtype=amax_aval.dtype)

        wkspace_info, = transformer_engine_jax.get_dgelu_dbias_ct_workspace_sizes(
            x_aval.size // gi_hidden_size,
            gi_hidden_size,
            jax_dtype_to_te_dtype(x_aval.dtype),
            jax_dtype_to_te_dtype(out_dtype),
        )
        wkspace_aval = x_aval.update(shape=wkspace_info[0],
                                     dtype=te_dtype_to_jax_dtype(wkspace_info[1]))

        return out, t_out, dbias, updated_amax_aval, wkspace_aval

    @staticmethod
    def outer_abstract(*args, **kwargs):
        """
        te_dgelu_dbais_cast_transpose_p outer abstract
        """

        out, t_out, dbias, updated_amax_aval, _ = \
            DGeluDBiasCastTransposePrimitive.abstract(*args, **kwargs)
        return out, t_out, dbias, updated_amax_aval

    @staticmethod
    def lowering(ctx, dz, x, amax, scale, scale_inv, *, out_dtype, static_axis_boundary,
                 transpose_axis_boundary):
        """
        te_dgated_gelu_cast_transpose_p lowering rules
        """
        dz_aval, x_aval, amax_aval, scale_aval, scale_inv_aval = ctx.avals_in
        assert dz_aval.dtype in [jnp.float32, jnp.float16, jnp.bfloat16]
        assert x_aval.dtype == dz_aval.dtype
        assert amax_aval.dtype == jnp.float32
        assert scale_aval.dtype == jnp.float32
        assert scale_inv_aval.dtype == jnp.float32
        ir_dz_type = ir.RankedTensorType(dz.type)
        ir_dz_shape = ir_dz_type.shape
        x_type = ir.RankedTensorType(x.type)
        x_shape = x_type.shape
        assert ir_dz_shape == x_shape

        batch_szie = reduce(operator.mul, ir_dz_shape[:-1])
        ir_hidden_szie = ir_dz_shape[-1]
        contracted_x_shape = (batch_szie, ir_hidden_szie)

        ir_out_dtype = jax_dtype_to_ir_dtype(out_dtype)
        ir_amax_type = ir.RankedTensorType(amax.type)
        ir_amax_dtype = ir_amax_type.element_type
        ir_amax_shape = ir_amax_type.shape
        ir_scale_shape = ir_amax_shape
        ir_scale_inv_shape = ir_amax_shape
        transposed_x_shape = _multidim_transpose(x_shape, static_axis_boundary,
                                                 transpose_axis_boundary)
        dbias_shape = (*x_shape[:static_axis_boundary + 1], ir_hidden_szie)

        wkspace_aval = ctx.avals_out[-1]

        out_types = [
            ir.RankedTensorType.get(x_shape, ir_out_dtype),
            ir.RankedTensorType.get(transposed_x_shape, ir_out_dtype),
            ir.RankedTensorType.get(dbias_shape, ir_dz_type.element_type),
            ir.RankedTensorType.get(ir_amax_shape, ir_amax_dtype),
            ir.RankedTensorType.get(wkspace_aval.shape, jax_dtype_to_ir_dtype(wkspace_aval.dtype)),
        ]
        operands = [dz, x, amax, scale, scale_inv]
        operand_shapes = [ir_dz_shape, x_shape, ir_amax_shape, ir_scale_shape, ir_scale_inv_shape]
        args = CustomCallArgsWrapper(out_types, operands, operand_shapes)
        opaque = transformer_engine_jax.pack_common_wk_descriptor(
            contracted_x_shape, wkspace_aval.shape, jax_dtype_to_te_dtype(dz_aval.dtype),
            jax_dtype_to_te_dtype(out_dtype), jax_dtype_to_te_dtype(wkspace_aval.dtype))

        out = custom_caller(DGeluDBiasCastTransposePrimitive.name,
                            args,
                            opaque,
                            False,
                            operand_output_aliases={2: 3})

        return out

    @staticmethod
    def impl(dz, x, amax, scale, scale_inv, out_dtype, static_axis_boundary,
             transpose_axis_boundary):
        """
        to describe implementation
        """
        assert DGeluDBiasCastTransposePrimitive.inner_primitive is not None
        out, t_out, dbias, updated_amax, _ = DGeluDBiasCastTransposePrimitive.inner_primitive.bind(
            dz,
            x,
            amax,
            scale,
            scale_inv,
            out_dtype=out_dtype,
            static_axis_boundary=static_axis_boundary,
            transpose_axis_boundary=transpose_axis_boundary)
        return out, t_out, dbias, updated_amax

    @staticmethod
    def batcher(batched_args, batch_dims, *, out_dtype, static_axis_boundary,
                transpose_axis_boundary):
        """
        to describe batch rules for vmap
        """
        del static_axis_boundary
        _check_valid_batch_dims(batch_dims)
        assert DGeluDBiasCastTransposePrimitive.outer_primitive is not None
        dz, x, amax, scale, scale_inv = batched_args
        x_bdim, _, amax_bdim, _, _ = batch_dims

        # Minus batch dim.
        transpose_axis_boundary = _normalize_axis_boundary(transpose_axis_boundary, x.ndim - 1)
        transpose_axis_boundary += 1    # Plus batch dim

        out_bdims = x_bdim, x_bdim, x_bdim, amax_bdim
        return DGeluDBiasCastTransposePrimitive.outer_primitive.bind(
            dz,
            x,
            amax,
            scale,
            scale_inv,
            out_dtype=out_dtype,
            static_axis_boundary=x_bdim,
            transpose_axis_boundary=transpose_axis_boundary), out_bdims

    @staticmethod
    def infer_sharding_from_operands(out_dtype, static_axis_boundary, transpose_axis_boundary, mesh,
                                     arg_infos, result_infos):
        del out_dtype, result_infos
        x_spec = get_padded_spec(arg_infos[1])
        out_sharding = NamedSharding(mesh, PartitionSpec(*x_spec))
        xt_spec = _multidim_transpose(x_spec, static_axis_boundary, transpose_axis_boundary)
        tranposed_out_sharding = NamedSharding(mesh, PartitionSpec(*xt_spec))
        dbias_shaprding = NamedSharding(
            mesh, PartitionSpec(*x_spec[:static_axis_boundary + 1], x_spec[-1]))
        amax_sharding = NamedSharding(mesh, PartitionSpec(*get_padded_spec(arg_infos[2])))
        return (out_sharding, tranposed_out_sharding, dbias_shaprding, amax_sharding)

    @staticmethod
    def partition(out_dtype, static_axis_boundary, transpose_axis_boundary, mesh, arg_infos,
                  result_infos):
        del result_infos
        x_spec = get_padded_spec(arg_infos[1])
        casted_x_sharding = NamedSharding(mesh, PartitionSpec(*x_spec))
        xt_spec = _multidim_transpose(x_spec, static_axis_boundary, transpose_axis_boundary)
        casted_transposed_x_sharding = NamedSharding(mesh, PartitionSpec(*xt_spec))

        dbias_shaprding = NamedSharding(
            mesh, PartitionSpec(*x_spec[:static_axis_boundary + 1], x_spec[-1]))

        amax_sharding = NamedSharding(mesh, PartitionSpec(*get_padded_spec(arg_infos[2])))
        arg_shardings = tuple(arg_i.sharding for arg_i in arg_infos)
        out_shardings = (casted_x_sharding, casted_transposed_x_sharding, dbias_shaprding,
                         amax_sharding)

        def sharded_impl(dz, x, amax, scale, scale_inv):
            local_out, local_t_out, local_dbias, local_amax = DGeluDBiasCastTransposePrimitive.impl(
                dz,
                x,
                amax,
                scale,
                scale_inv,
                out_dtype=out_dtype,
                static_axis_boundary=static_axis_boundary,
                transpose_axis_boundary=transpose_axis_boundary)
            global_dbias = all_reduce_sum_along_dp_fsdp(local_dbias)
            global_updated_amax = all_reduce_max_along_all_axes_except_PP(local_amax)
            return local_out, local_t_out, global_dbias, global_updated_amax

        return mesh, sharded_impl, out_shardings, arg_shardings


register_primitive(DGeluDBiasCastTransposePrimitive)


def dgelu_dbias_cast_transpose(
        dz: jnp.ndarray,
        x: jnp.ndarray,
        amax: jnp.ndarray,
        scale: jnp.ndarray,
        scale_inv: jnp.ndarray,
        out_dtype: TEDType,
        static_axis_boundary: int,
        transpose_axis_boundary: int = -1) -> Tuple[jnp.ndarray, jnp.ndarray, jnp.ndarray]:
    """
    cast transpose dgelu and dbias fusion wrapper
    Return FP8(dgeglu(inputs)), dbias
    """
    if static_axis_boundary < 0:
        static_axis_boundary = -1    # means no static axes

    return DGeluDBiasCastTransposePrimitive.outer_primitive.bind(
        dz,
        x,
        amax,
        scale,
        scale_inv,
        out_dtype=out_dtype,
        static_axis_boundary=static_axis_boundary,
        transpose_axis_boundary=transpose_axis_boundary)


4986
class GatedGeluFp8Primitive(BasePrimitive):
4987
    """
4988
    Gated Gelu FP8 Primitive
4989
    """
4990
    name = "te_gated_gelu_fp8"
4991
    multiple_results = True
4992
4993
4994
    impl_static_args = (4,)    #out_dtype
    inner_primitive = None
    outer_primitive = None
4995
4996

    @staticmethod
4997
    def abstract(x_aval, amax_aval, scale_aval, scale_inv_aval, *, out_dtype):
4998
        """
4999
        te_gated_gelu_p abstract
5000
        """
5001
5002
5003
5004
5005
5006
5007
        dtype = dtypes.canonicalize_dtype(x_aval.dtype)
        # Currently only support casting to E4M3 only in C side.
        assert out_dtype == jnp.float8_e4m3fn
        assert dtype in [jnp.float32, jnp.float16, jnp.bfloat16]
        assert amax_aval.dtype == jnp.float32
        assert scale_aval.dtype == jnp.float32
        assert scale_inv_aval.dtype == jnp.float32
5008

5009
5010
5011
5012
5013
5014
        assert x_aval.shape[-2] == 2    # Assume x in (....., 2, hidden)
        hidden_size = x_aval.shape[-1]
        batch_shape = x_aval.shape[:-2]
        out_shape = (batch_shape) + (hidden_size,)
        out_aval = x_aval.update(shape=out_shape, dtype=out_dtype)
        updated_amax_aval = amax_aval.update(shape=amax_aval.shape, dtype=amax_aval.dtype)
5015

5016
        return out_aval, updated_amax_aval
5017
5018

    @staticmethod
5019
    def lowering(ctx, x, amax, scale, scale_inv, *, out_dtype):
5020
        """
5021
        te_gated_gelu_p lowering rules
5022
        """
5023
5024
5025
5026
5027
5028
5029
5030
5031
5032
5033
5034
5035
        x_aval, amax_aval, scale_aval, scale_inv_aval = ctx.avals_in
        assert x_aval.dtype in [jnp.float32, jnp.float16, jnp.bfloat16]
        assert amax_aval.dtype == jnp.float32
        assert scale_aval.dtype == jnp.float32
        assert scale_inv_aval.dtype == jnp.float32
        ir_x_type = ir.RankedTensorType(x.type)
        ir_x_shape = ir_x_type.shape
        ir_out_dtype = jax_dtype_to_ir_dtype(out_dtype)
        ir_amax_type = ir.RankedTensorType(amax.type)
        ir_amax_dtype = ir_amax_type.element_type
        ir_amax_shape = ir_amax_type.shape
        ir_scale_shape = ir_amax_shape
        ir_scale_inv_shape = ir_amax_shape
5036

5037
5038
5039
5040
        hidden_size = ir_x_shape[-1]
        batch_shape = ir_x_shape[:-2]
        batch_size = reduce(operator.mul, batch_shape)
        out_shape = batch_shape + [hidden_size]
5041
        out_types = [
5042
5043
            ir.RankedTensorType.get(out_shape, ir_out_dtype),
            ir.RankedTensorType.get(ir_amax_shape, ir_amax_dtype),
5044
        ]
5045
5046
        operands = [x, amax, scale, scale_inv]
        operand_shapes = [ir_x_shape, ir_amax_shape, ir_scale_shape, ir_scale_inv_shape]
5047
        args = CustomCallArgsWrapper(out_types, operands, operand_shapes)
5048

5049
5050
5051
        opaque = transformer_engine_jax.pack_common_descriptor((batch_size, out_shape[-1]),
                                                               jax_dtype_to_te_dtype(x_aval.dtype),
                                                               jax_dtype_to_te_dtype(out_dtype))
5052

5053
5054
5055
5056
5057
        out = custom_caller(GatedGeluFp8Primitive.name,
                            args,
                            opaque,
                            False,
                            operand_output_aliases={1: 1})
5058
5059
5060
5061

        return out

    @staticmethod
5062
    def impl(x, amax, scale, scale_inv, out_dtype):
5063
        """
5064
        to describe implementation
5065
        """
5066
5067
5068
5069
5070
5071
5072
        assert GatedGeluFp8Primitive.inner_primitive is not None
        out, updated_amax = GatedGeluFp8Primitive.inner_primitive.bind(x,
                                                                       amax,
                                                                       scale,
                                                                       scale_inv,
                                                                       out_dtype=out_dtype)
        return out, updated_amax
5073
5074

    @staticmethod
5075
    def batcher(batched_args, batch_dims, *, out_dtype):
5076
        """
5077
        to describe batch rules for vmap
5078
        """
5079
5080
5081
5082
        _check_valid_batch_dims(batch_dims)
        assert GatedGeluFp8Primitive.outer_primitive is not None
        x, amax, scale, scale_inv = batched_args
        x_bdim, amax_bdim, _, _ = batch_dims
5083

5084
5085
5086
5087
5088
5089
        out_bdims = x_bdim, amax_bdim
        return GatedGeluFp8Primitive.outer_primitive.bind(x,
                                                          amax,
                                                          scale,
                                                          scale_inv,
                                                          out_dtype=out_dtype), out_bdims
5090

5091
5092
5093
5094
5095
5096
5097
    @staticmethod
    def infer_sharding_from_operands(out_dtype, mesh, arg_infos, result_infos):
        del out_dtype, result_infos
        x_spec = get_padded_spec(arg_infos[0])
        out_sharding = NamedSharding(mesh, PartitionSpec(*x_spec[:-2], x_spec[-1]))
        amax_sharding = NamedSharding(mesh, PartitionSpec(*get_padded_spec(arg_infos[1])))
        return (out_sharding, amax_sharding)
5098

5099
5100
5101
5102
5103
5104
5105
5106
    @staticmethod
    def partition(out_dtype, mesh, arg_infos, result_infos):
        del result_infos
        x_spec = get_padded_spec(arg_infos[0])
        out_sharding = NamedSharding(mesh, PartitionSpec(*x_spec[:-2], x_spec[-1]))
        amax_sharding = NamedSharding(mesh, PartitionSpec(*get_padded_spec(arg_infos[1])))
        arg_shardings = tuple(arg_i.sharding for arg_i in arg_infos)
        out_shardings = (out_sharding, amax_sharding)
5107

5108
5109
5110
5111
5112
5113
5114
        def sharded_impl(x, amax, scale, scale_inv):
            local_x, local_amax = GatedGeluFp8Primitive.impl(x,
                                                             amax,
                                                             scale,
                                                             scale_inv,
                                                             out_dtype=out_dtype)
            global_updated_amax = all_reduce_max_along_all_axes_except_PP(local_amax)
5115

5116
            return local_x, global_updated_amax
5117

5118
        return mesh, sharded_impl, out_shardings, arg_shardings
5119
5120


5121
register_primitive(GatedGeluFp8Primitive)
5122

5123
5124
5125

def gated_gelu_fp8(x: jnp.ndarray, amax: jnp.ndarray, scale: jnp.ndarray, scale_inv: jnp.ndarray,
                   out_dtype: jnp.dtype) -> Tuple[jnp.ndarray, jnp.ndarray, jnp.ndarray]:
5126
    """
5127
5128
    gated gelu wrapper
    Return FP8(geglu(x))
5129
    """
5130
5131
5132
5133
5134
    return GatedGeluFp8Primitive.outer_primitive.bind(x,
                                                      amax,
                                                      scale,
                                                      scale_inv,
                                                      out_dtype=out_dtype)
5135
5136


5137
class DgatedGeluCastTransposePrimitive(BasePrimitive):
5138
    """
5139
    Dgated Gelu Cast Transpose Primitive
5140
    """
5141
    name = "te_dgated_gelu_cast_transpose"
5142
    multiple_results = True
5143
5144
5145
    impl_static_args = (5, 6)    # out_dtype, static_axis_boundary
    inner_primitive = None
    outer_primitive = None
5146
5147

    @staticmethod
5148
5149
    def abstract(dz_aval, x_aval, amax_aval, scale_aval, scale_inv_aval, *, out_dtype,
                 static_axis_boundary):
5150
        """
5151
        te_dgated_gelu_cast_transpose_p abstract
5152
        """
5153
5154
5155
5156
5157
5158
5159
5160
5161
5162
5163
5164
5165
5166
5167
        dtype = dtypes.canonicalize_dtype(dz_aval.dtype)
        assert dtype in [jnp.float32, jnp.float16, jnp.bfloat16]
        assert x_aval.dtype == dtype
        assert x_aval.shape[-2] == 2    # Linear + GeLU
        assert amax_aval.dtype == jnp.float32
        assert scale_aval.dtype == jnp.float32
        assert scale_inv_aval.dtype == jnp.float32
        ir_hidden_szie = dz_aval.shape[-1]
        gi_hidden_size = x_aval.shape[-1]
        assert ir_hidden_szie == gi_hidden_size
        t_shape = _multidim_transpose(x_aval.shape, static_axis_boundary, -2)
        out = dz_aval.update(shape=x_aval.shape, dtype=out_dtype)
        t_out = dz_aval.update(shape=t_shape, dtype=out_dtype)
        updated_amax_aval = amax_aval.update(shape=amax_aval.shape, dtype=amax_aval.dtype)
        return out, t_out, updated_amax_aval
5168

5169
5170
5171
5172
5173
5174
5175
5176
5177
5178
5179
5180
5181
5182
5183
5184
5185
5186
5187
5188
5189
5190
5191
5192
5193
5194
5195
5196
5197
5198
5199
5200
5201
5202
5203
5204
5205
5206
5207
5208
5209
5210
5211
5212
5213
5214
5215
5216
5217
    @staticmethod
    def lowering(ctx, dz, x, amax, scale, scale_inv, *, out_dtype, static_axis_boundary):
        """
        te_dgated_gelu_cast_transpose_p lowering rules
        """
        dz_aval, x_aval, amax_aval, scale_aval, scale_inv_aval = ctx.avals_in
        assert dz_aval.dtype in [jnp.float32, jnp.float16, jnp.bfloat16]
        assert x_aval.dtype == dz_aval.dtype
        assert amax_aval.dtype == jnp.float32
        assert scale_aval.dtype == jnp.float32
        assert scale_inv_aval.dtype == jnp.float32
        ir_dz_type = ir.RankedTensorType(dz.type)
        ir_dz_shape = ir_dz_type.shape
        x_type = ir.RankedTensorType(x.type)
        x_shape = x_type.shape
        dz_batch_szie = reduce(operator.mul, ir_dz_shape[:-1])
        x_batch_size = reduce(operator.mul, x_shape[:-2])
        assert dz_batch_szie == x_batch_size
        assert x_shape[-2] == 2    # Linear + GeLU
        ir_hidden_szie = ir_dz_shape[-1]
        gi_hidden_size = x_shape[-1]
        assert ir_hidden_szie == gi_hidden_size
        ir_out_dtype = jax_dtype_to_ir_dtype(out_dtype)
        ir_amax_type = ir.RankedTensorType(amax.type)
        ir_amax_dtype = ir_amax_type.element_type
        ir_amax_shape = ir_amax_type.shape
        ir_scale_shape = ir_amax_shape
        ir_scale_inv_shape = ir_amax_shape
        transposed_x_shape = _multidim_transpose(x_shape, static_axis_boundary, -2)
        out_types = [
            ir.RankedTensorType.get(x_shape, ir_out_dtype),
            ir.RankedTensorType.get(transposed_x_shape, ir_out_dtype),
            ir.RankedTensorType.get(ir_amax_shape, ir_amax_dtype),
        ]
        operands = [dz, x, amax, scale, scale_inv]
        operand_shapes = [ir_dz_shape, x_shape, ir_amax_shape, ir_scale_shape, ir_scale_inv_shape]
        args = CustomCallArgsWrapper(out_types, operands, operand_shapes)
        contracted_x_shape = (x_batch_size, x_shape[-1])
        opaque = transformer_engine_jax.pack_common_descriptor(contracted_x_shape,
                                                               jax_dtype_to_te_dtype(dz_aval.dtype),
                                                               jax_dtype_to_te_dtype(out_dtype))

        out = custom_caller(DgatedGeluCastTransposePrimitive.name,
                            args,
                            opaque,
                            False,
                            operand_output_aliases={2: 2})

        return out
5218
5219

    @staticmethod
5220
    def impl(dz, x, amax, scale, scale_inv, out_dtype, static_axis_boundary):
5221
        """
5222
        to describe implementation
5223
        """
5224
5225
5226
5227
5228
5229
5230
5231
5232
5233
        assert DgatedGeluCastTransposePrimitive.inner_primitive is not None
        out, t_out, updated_amax = DgatedGeluCastTransposePrimitive.inner_primitive.bind(
            dz,
            x,
            amax,
            scale,
            scale_inv,
            out_dtype=out_dtype,
            static_axis_boundary=static_axis_boundary)
        return out, t_out, updated_amax
5234

5235
5236
5237
5238
5239
5240
5241
5242
5243
5244
    @staticmethod
    def batcher(batched_args, batch_dims, *, out_dtype, static_axis_boundary):
        """
        to describe batch rules for vmap
        """
        del static_axis_boundary
        _check_valid_batch_dims(batch_dims)
        assert DgatedGeluCastTransposePrimitive.outer_primitive is not None
        dz, x, amax, scale, scale_inv = batched_args
        x_bdim, _, amax_bdim, _, _ = batch_dims
5245

5246
5247
5248
5249
        out_bdims = x_bdim, x_bdim, amax_bdim
        return DgatedGeluCastTransposePrimitive.outer_primitive.bind(
            dz, x, amax, scale, scale_inv, out_dtype=out_dtype,
            static_axis_boundary=x_bdim), out_bdims
5250

5251
5252
5253
5254
5255
5256
5257
5258
5259
5260
    @staticmethod
    def infer_sharding_from_operands(out_dtype, static_axis_boundary, mesh, arg_infos,
                                     result_infos):
        del out_dtype, result_infos
        x_spec = get_padded_spec(arg_infos[1])
        out_sharding = NamedSharding(mesh, PartitionSpec(*x_spec))
        xt_spec = _multidim_transpose(x_spec, static_axis_boundary, -2)
        tranposed_out_sharding = NamedSharding(mesh, PartitionSpec(*xt_spec))
        amax_sharding = NamedSharding(mesh, PartitionSpec(*get_padded_spec(arg_infos[2])))
        return (out_sharding, tranposed_out_sharding, amax_sharding)
5261

5262
5263
5264
5265
5266
5267
5268
    @staticmethod
    def partition(out_dtype, static_axis_boundary, mesh, arg_infos, result_infos):
        del result_infos
        x_spec = get_padded_spec(arg_infos[1])
        casted_x_sharding = NamedSharding(mesh, PartitionSpec(*x_spec))
        xt_spec = _multidim_transpose(x_spec, static_axis_boundary, -2)
        casted_transposed_x_sharding = NamedSharding(mesh, PartitionSpec(*xt_spec))
5269

5270
5271
5272
        amax_sharding = NamedSharding(mesh, PartitionSpec(*get_padded_spec(arg_infos[2])))
        arg_shardings = tuple(arg_i.sharding for arg_i in arg_infos)
        out_shardings = (casted_x_sharding, casted_transposed_x_sharding, amax_sharding)
5273

5274
5275
5276
5277
5278
5279
5280
5281
5282
5283
5284
        def sharded_impl(dz, x, amax, scale, scale_inv):
            local_out, local_t_out, local_amax = DgatedGeluCastTransposePrimitive.impl(
                dz,
                x,
                amax,
                scale,
                scale_inv,
                out_dtype=out_dtype,
                static_axis_boundary=static_axis_boundary)
            global_updated_amax = all_reduce_max_along_all_axes_except_PP(local_amax)
            return local_out, local_t_out, global_updated_amax
5285

5286
        return mesh, sharded_impl, out_shardings, arg_shardings
5287
5288


5289
register_primitive(DgatedGeluCastTransposePrimitive)
5290

5291
5292
5293
5294
5295

def dgated_gelu_cast_transpose(
        dz: jnp.ndarray, x: jnp.ndarray, amax: jnp.ndarray, scale: jnp.ndarray,
        scale_inv: jnp.ndarray, out_dtype: TEDType,
        static_axis_boundary: int) -> Tuple[jnp.ndarray, jnp.ndarray, jnp.ndarray]:
5296
    """
5297
5298
    cast transpose d_gated_gelu fusion wrapper
    Return FP8(dgeglu(inputs))
5299
    """
5300
5301
5302
5303
5304
5305
5306
5307
    return DgatedGeluCastTransposePrimitive.outer_primitive.bind(
        dz,
        x,
        amax,
        scale,
        scale_inv,
        out_dtype=out_dtype,
        static_axis_boundary=static_axis_boundary)