linear.py 64.7 KB
Newer Older
1
# Copyright (c) 2022-2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
2
3
4
5
#
# See LICENSE for license information.

"""Linear API"""
6
from typing import Callable, Dict, Optional, Tuple, Union, List
7
8
from functools import reduce
from operator import mul as multiply_op
9
import warnings
10
11
12

import torch

13
import transformer_engine_torch as tex
14

15
from transformer_engine.common.recipe import Recipe
16
from transformer_engine.pytorch import torch_version
17

18
from .base import (
19
20
    fill_userbuffers_buffer_for_all_gather,
    get_dummy_wgrad,
21
    get_ub,
22
    get_workspace,
23
24
25
26
27
    TransformerEngineBaseModule,
    _2X_ACC_FPROP,
    _2X_ACC_DGRAD,
    _2X_ACC_WGRAD,
)
28
from ._common import noop_cat, WeightGradStore
29
from ..fp8 import FP8GlobalStateManager
30
31
from ..utils import (
    cast_if_needed,
32
    clear_tensor_data,
33
    divide,
34
    init_method_constant,
35
36
    requires_grad,
    needs_quantized_gemm,
37
    assert_dim_for_fp8_exec,
38
39
    nvtx_range_pop,
    nvtx_range_push,
40
41
42
43
44
)
from ..distributed import (
    set_tensor_model_parallel_attributes,
    get_distributed_world_size,
    allreduce,
45
    symmetric_all_reduce,
46
47
    reduce_scatter_along_first_dim,
    gather_along_first_dim,
48
    is_fp8_activation_recompute_enabled,
49
    in_fp8_activation_recompute_phase,
50
51
    _fsdp_scatter_tensors,
    _fsdp_gather_tensors,
52
53
)
from ..cpp_extensions import (
54
    general_gemm,
55
)
56
from ..constants import GemmParallelModes, dist_group_type
57
from ..jit import no_torch_dynamo
58
from ..graph import is_graph_capturing
59
60
from ..tensor.quantized_tensor import (
    QuantizedTensor,
61
    QuantizedTensorBase,
62
63
64
65
    Quantizer,
    prepare_for_saving,
    restore_from_saved,
)
66
from ..tensor.float8_tensor import Float8CurrentScalingQuantizer, Float8Quantizer
67
from ..tensor.mxfp8_tensor import MXFP8Quantizer
68
from ..tensor._internal.mxfp8_tensor_base import MXFP8TensorBase
69
from ..tensor._internal.float8_blockwise_tensor_base import Float8BlockwiseQTensorBase
70
from ..cpu_offload import is_cpu_offload_enabled, mark_activation_offload
71
72
from ...debug.pytorch.debug_state import TEDebugState
from ...debug.pytorch.utils import any_feature_enabled
73

74
75
76
77
78
79
80
81
82
83
84
__all__ = ["Linear"]


class _Linear(torch.autograd.Function):
    """Linear semi-top level module
    Calls custom cuda extensions.
    """

    @staticmethod
    def forward(
        ctx,
85
        weight: torch.Tensor,
86
        inp: torch.Tensor,
87
        bias: Optional[torch.Tensor],
88
89
90
        is_first_microbatch: Union[bool, None],
        fp8: bool,
        fp8_calibration: bool,
91
        wgrad_store: WeightGradStore,
92
93
94
95
        input_quantizer: Optional[Quantizer],
        weight_quantizer: Optional[Quantizer],
        output_quantizer: Optional[Quantizer],
        grad_input_quantizer: Optional[Quantizer],
96
97
        grad_weight_quantizer: Optional[Quantizer],
        grad_output_quantizer: Optional[Quantizer],
98
        fuse_wgrad_accumulation: bool,
99
        cpu_offloading: bool,
100
101
102
103
104
105
106
        tp_group: Union[dist_group_type, None],
        tp_size: int,
        sequence_parallel: bool,
        tensor_parallel: bool,
        activation_dtype: torch.dtype,
        parallel_mode: Union[str, None],
        is_grad_enabled: bool,
107
108
109
110
111
112
        ub_overlap_rs_fprop: bool,
        ub_overlap_ag_dgrad: bool,
        ub_overlap_ag_fprop: bool,
        ub_overlap_rs_dgrad: bool,
        ub_bulk_dgrad: bool,
        ub_bulk_wgrad: bool,
113
        ub_name: str,
114
        fp8_output: bool,  # pylint: disable=unused-argument
115
        fsdp_group: Union[dist_group_type, None],
116
117
        module: torch.nn.Module,
        skip_fp8_weight_update: bool,
118
        symmetric_ar_type: str,
119
        debug: Optional[bool] = False,
120
    ) -> torch.Tensor:
121
        # pylint: disable=missing-function-docstring
122

123
124
125
126
127
        # NVTX label for profiling
        nvtx_label = "transformer_engine._Linear.forward"
        if ub_name is not None:
            nvtx_label = f"{nvtx_label}.{ub_name}"

128
        # Make sure input dimensions are compatible
129
        out_features, in_features = weight.shape
130
        assert inp.shape[-1] == in_features, "GEMM not possible"
131

132
        # Configure tensor-parallel communication
133
        tp_world_size = get_distributed_world_size(tp_group)
134
135
136
137
        backward_needs_input = is_grad_enabled and weight.requires_grad
        with_input_all_gather_nccl = (
            parallel_mode == "column" and sequence_parallel and not ub_overlap_ag_fprop
        )
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156

        # Configure Userbuffers communication (comm+GEMM overlap)
        ub_obj = None
        ub_type = None
        if ub_overlap_rs_fprop:
            ub_obj = get_ub(ub_name + "_fprop")
            ub_type = tex.CommOverlapType.RS
        elif ub_overlap_ag_fprop:
            ub_obj = get_ub(ub_name + "_fprop")
            ub_type = tex.CommOverlapType.AG

        # ------------------------------------------------------
        # Prepare input tensor
        # Note: Cast to expected dtype and perform tensor-parallel communication
        # ------------------------------------------------------
        nvtx_range_push(f"{nvtx_label}.input_cast_comm")
        inputmat = inp  # Input tensor to save for backward (maybe sharded)
        inputmat_total = None  # Input tensor to pass to GEMM (gathered)
        own_quantized_input = False
157
        if fp8:
158
            assert_dim_for_fp8_exec(inputmat, weight)
159
160
161
162
163
164
        if with_input_all_gather_nccl or ub_overlap_ag_fprop:  # All-gather input tensor

            # Cast local input tensor if needed
            if fp8 or debug:
                if input_quantizer is None:
                    raise ValueError("Missing quantizer for input tensor")
165
                if not isinstance(inputmat, QuantizedTensorBase):
166
167
168
169
170
171
172
173
                    input_quantizer.set_usage(rowwise=True, columnwise=backward_needs_input)
                    if isinstance(
                        input_quantizer, (Float8Quantizer, Float8CurrentScalingQuantizer)
                    ):
                        # All-gather is not supported with FP8 column-wise data
                        input_quantizer.set_usage(columnwise=False)
                    inputmat = input_quantizer(inputmat)
                    own_quantized_input = True
174
            else:
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
                inputmat = cast_if_needed(inp, activation_dtype)  # Cast for AMP

            # Initialize gathered input tensor
            quantizer = None
            if fp8 or debug:
                quantizer = input_quantizer
                quantizer.set_usage(rowwise=True, columnwise=False)
            if with_input_all_gather_nccl:  # Perform NCCL all-gather
                inputmat_total, _ = gather_along_first_dim(
                    inputmat,
                    tp_group,
                    quantizer=quantizer,
                )
            elif ub_overlap_ag_fprop:  # Initialize Userbuffers all-gather
                inputmat_total, _ = fill_userbuffers_buffer_for_all_gather(
                    ub_obj,
                    inputmat,
                    quantizer,
                    tp_group,
                )

        else:  # Do not all-gather input tensor
            if fp8 or debug:
                if isinstance(inputmat, QuantizedTensorBase):
                    inputmat.update_usage(rowwise_usage=True)
200
                else:
201
202
203
                    if input_quantizer is None:
                        raise ValueError("Missing quantizer for input tensor")
                    input_quantizer.set_usage(rowwise=True, columnwise=backward_needs_input)
204
                    inputmat = input_quantizer(inputmat)
205
                    own_quantized_input = True
206
            else:
207
208
                inputmat = cast_if_needed(inp, activation_dtype)  # Cast for AMP
            inputmat_total = inputmat
209
        nvtx_range_pop(f"{nvtx_label}.input_cast_comm")
210
211
212
        # ------------------------------------------------------
        # Input tensor is ready for GEMM...
        # ------------------------------------------------------
213

214
215
216
        # ------------------------------------------------------
        # Prepare weight tensor
        # ------------------------------------------------------
217
218
        weightmat = weight
        if fp8 or debug:
219
220
221
222
223
224
225
226
227
            # Configure quantizer
            if weight_quantizer is not None:
                columnwise_usage = is_grad_enabled and inp.requires_grad
                if not columnwise_usage:
                    columnwise_usage = (
                        is_fp8_activation_recompute_enabled()
                        and not in_fp8_activation_recompute_phase()
                    )
                weight_quantizer.set_usage(rowwise=True, columnwise=columnwise_usage)
228
229

            # Get quantized weight
230
231
232
233
234
235
236
237
            update_workspace = is_first_microbatch is None or is_first_microbatch
            weightmat = module.get_weight_workspace(
                tensor=weight,
                quantizer=weight_quantizer,
                cache_name=(None if is_first_microbatch is None else "weight"),
                update_workspace=update_workspace,
                skip_update_flag=skip_fp8_weight_update,
                fsdp_group=fsdp_group,
238
                workspace_dtype=activation_dtype,
239
            )
240
241
            weightmat.update_usage(rowwise_usage=True)

242
        else:
243
244
245
246
            weightmat = cast_if_needed(weightmat, activation_dtype)  # Cast for AMP
        # ------------------------------------------------------
        # Weight tensor is ready for GEMM...
        # ------------------------------------------------------
247
248
249

        # Cast bias to expected dtype
        bias_dtype = activation_dtype
250
        if needs_quantized_gemm(inputmat_total) and activation_dtype == torch.float32:
251
            # cuBLAS does not support FP8 GEMM with FP32 bias, so we cast to BF16
252
253
254
255
256
257
258
259
260
261
            bias_dtype = torch.bfloat16
        bias = cast_if_needed(bias, bias_dtype) if bias is not None else bias

        # Calibrate quantizers if needed
        if not fp8 and fp8_calibration:
            if input_quantizer is not None:
                input_quantizer.calibrate(inputmat_total)
            if weight_quantizer is not None:
                weight_quantizer.calibrate(weight)

262
263
        # Choose whether to use GEMM kernel with split accumulator
        use_split_accumulator = _2X_ACC_FPROP
264
265
266
        if fp8:
            recipe = FP8GlobalStateManager.get_fp8_recipe()
            if hasattr(recipe, "fp8_gemm_fprop"):
267
268
269
270
271
                use_split_accumulator = recipe.fp8_gemm_fprop.use_split_accumulator

        # Configure output quantizer
        if output_quantizer is not None:
            output_quantizer.set_usage(rowwise=True, columnwise=False)
272

273
274
275
276
277
278
279
280
281
282
283
284
285
286
        # Output buffer for Userbuffers reduce-scatter
        reduce_scatter_out = None
        if ub_overlap_rs_fprop:
            out_shape = list(inp.shape)
            out_shape[0] //= tp_world_size
            out_shape[-1] = out_features
            reduce_scatter_out = torch.empty(out_shape, dtype=activation_dtype, device=inp.device)

        # ------------------------------------------------------
        # Forward GEMM
        # Note: y = x * w^T
        # ------------------------------------------------------
        nvtx_range_push(f"{nvtx_label}.gemm")
        gemm_out, *_, reduce_scatter_out = general_gemm(
287
288
289
290
            weightmat,
            inputmat_total,
            get_workspace(),
            quantization_params=output_quantizer,
291
            out_dtype=activation_dtype,
292
            bias=bias,
293
            use_split_accumulator=use_split_accumulator,
294
295
            ub=ub_obj,
            ub_type=ub_type,
296
            extra_output=reduce_scatter_out,
297
        )
298
        nvtx_range_pop(f"{nvtx_label}.gemm")
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
        # ------------------------------------------------------
        # Finished forward GEMM...
        # ------------------------------------------------------

        # ------------------------------------------------------
        # Prepare output tensor
        # Note: Perform tensor-parallel communication
        # ------------------------------------------------------
        out = None
        if ub_overlap_rs_fprop:
            out = reduce_scatter_out
        elif parallel_mode == "row" and tp_size > 1:
            nvtx_range_push(f"{nvtx_label}.row_parallel_comm")
            out = gemm_out
            if sequence_parallel:
                out, _ = reduce_scatter_along_first_dim(out, tp_group)
            elif tensor_parallel:
                if symmetric_ar_type is not None:
                    out, _ = symmetric_all_reduce(out, tp_group, all_reduce_type=symmetric_ar_type)
                else:
                    out, _ = allreduce(out, tp_group)
            nvtx_range_pop(f"{nvtx_label}.row_parallel_comm")
        else:
            out = gemm_out
        # ------------------------------------------------------
        # Output tensor is ready to return...
        # ------------------------------------------------------

        # ------------------------------------------------------
        # Cache state for backward pass
        # ------------------------------------------------------
330
331

        if is_grad_enabled:
332
            ctx.weight_quantizer = weight_quantizer
333
            saved_inputmat = None
334
335
336
337
338

            ctx.backward_input_needs_gather = (
                weight.requires_grad and parallel_mode == "column" and sequence_parallel
            )

339
            if backward_needs_input:
340
                if own_quantized_input and isinstance(inputmat, QuantizedTensorBase):
341
342
343
                    # For sequence parallel in vanilla FP8, rowwise data is
                    # to gather the input. For MXFP8, columnwise only data
                    # can be allgathered.
344
345
346
347
                    if (
                        isinstance(inputmat, (MXFP8TensorBase, Float8BlockwiseQTensorBase))
                        or not ctx.backward_input_needs_gather
                    ):
348
                        inputmat.update_usage(rowwise_usage=False, columnwise_usage=True)
349
                saved_inputmat = inputmat
350

351
352
            # Weight with column-wise usage is needed for dgrad GEMM.
            if inp.requires_grad:
353
                if isinstance(weightmat, QuantizedTensorBase):
354
355
                    weightmat.update_usage(columnwise_usage=True)

356
357
            if cpu_offloading and saved_inputmat is not None:
                mark_activation_offload(saved_inputmat)
358

359
360
            # Scatter intermediate/activation tensors saved for the backward pass
            # NOTE: FSDP sharding is not valid for models initialized with primary Fp8 weights
361
            nvtx_range_push(f"{nvtx_label}.fsdp_scatter")
362
363
364
            ctx.fsdp_group = fsdp_group
            ctx.fsdp_shapes = _fsdp_scatter_tensors(
                fsdp_group,
365
                saved_inputmat,
366
                weightmat if fp8 and not isinstance(weight, QuantizedTensorBase) else None,
367
            )
368
            nvtx_range_pop(f"{nvtx_label}.fsdp_scatter")
369

370
371
372
373
374
375
376
377
378
379
380
            if cpu_offloading:
                ctx.grad_added_to_main_grad = hasattr(weight, "grad_added_to_main_grad")

                if ctx.grad_added_to_main_grad:
                    # If you are passing torch.nn.Parameter through the Torch hooks, you will
                    # get back torch.Tensor. Torch rips off the Parameter wrapper.
                    # You need to preserve the weight object to have all the attributes user
                    # sets for the weights. Because of this, it is not recommended to offload
                    # weights if weights are externally touched outside this module
                    ctx.weight_object = weight

381
382
            # TODO(ksivamani): Check memory usage
            tensors_to_save, tensor_objects = prepare_for_saving(
383
                saved_inputmat,
384
                weightmat,
385
                weight,
386
                bias,
387
            )
388
389
            ctx.save_for_backward(*tensors_to_save)
            ctx.tensor_objects = tensor_objects
390

391
392
            ctx.activation_dtype = activation_dtype
            ctx.fp8 = fp8
393
            ctx.fp8_recipe = FP8GlobalStateManager.get_fp8_recipe() if fp8 else None
394
395
            ctx.input_quantizer = input_quantizer
            ctx.grad_input_quantizer = grad_input_quantizer
396
397
            ctx.grad_weight_quantizer = grad_weight_quantizer
            ctx.grad_output_quantizer = grad_output_quantizer
398
            ctx.fuse_wgrad_accumulation = fuse_wgrad_accumulation
399
400
401
            if fuse_wgrad_accumulation and weight.requires_grad:
                ctx.main_grad = weight.main_grad

402
            ctx.debug = debug
403
            ctx.cpu_offloading = cpu_offloading
404
            ctx.is_first_microbatch = is_first_microbatch
405
            ctx.use_bias = bias is not None
406
407
            ctx.sequence_parallel = sequence_parallel
            ctx.tensor_parallel = tensor_parallel
408
            ctx.inp_shape = inp.shape
409
410
            ctx.parallel_mode = parallel_mode
            ctx.tp_group = tp_group
411
412
413
414
            ctx.ub_overlap_ag = ub_overlap_ag_dgrad
            ctx.ub_overlap_rs_dgrad = ub_overlap_rs_dgrad
            ctx.ub_bulk_dgrad = ub_bulk_dgrad
            ctx.ub_bulk_wgrad = ub_bulk_wgrad
415
            ctx.ub_name = ub_name
416
417
            ctx.tp_size = tp_size
            ctx.requires_dgrad = inp.requires_grad
418
            ctx.requires_wgrad = weight.requires_grad
419
            ctx.reduce_and_update_bwd_fp8_tensors = False
420

421
            ctx.owns_input = saved_inputmat is not inp
422
            if ctx.fp8 and requires_grad(inp, weight, bias):
423
424
425
426
                _first_fp8_module = FP8GlobalStateManager.IS_FIRST_FP8_MODULE
                ctx.reduce_and_update_bwd_fp8_tensors = FP8GlobalStateManager.is_first_fp8_module()
                if in_fp8_activation_recompute_phase():
                    FP8GlobalStateManager.IS_FIRST_FP8_MODULE = _first_fp8_module
427
            ctx.wgrad_store = wgrad_store
428

429
430
431
        # ------------------------------------------------------
        # Cached state for backward pass is ready...
        # ------------------------------------------------------
432

433
        return out
434
435

    @staticmethod
436
    def backward(ctx, grad_output: torch.Tensor) -> Tuple[Union[torch.Tensor, None], ...]:
437
        # pylint: disable=missing-function-docstring
438

439
440
441
442
443
        # NVTX label for profiling
        nvtx_label = "transformer_engine._Linear.backward"
        if ctx.ub_name is not None:
            nvtx_label = f"{nvtx_label}.{ctx.ub_name}"

444
        with torch.cuda.nvtx.range("_Linear_backward"):
445
446
447
448
            saved_tensors = ctx.saved_tensors
            inputmat, weight_fp8, weight, bias = (  # pylint: disable=unbalanced-tuple-unpacking
                restore_from_saved(ctx.tensor_objects, saved_tensors)
            )
449

450
451
452
            # Delete the references to tensor objects once they've been consumed
            # by the `restore_from_saved` method to construct back the actual tensors.
            ctx.tensor_objects = None
453
454
455
456
457
458
459
460

            # Since main_grad can be modified inplace, it should not be a part of saved_tensors
            main_grad = (
                ctx.main_grad
                if weight is not None and ctx.fuse_wgrad_accumulation and ctx.requires_wgrad
                else None
            )

461
462
463
464
465
            if ctx.cpu_offloading:
                if ctx.grad_added_to_main_grad:
                    weight = ctx.weight_object
                if ctx.requires_wgrad and ctx.fuse_wgrad_accumulation:
                    weight.main_grad = main_grad
466

467
468
469
            # Gather intermediate/activation tensors if needed
            # NOTE: weight_fp8 = weight when ctx.fp8 == False and torch.disttributed.FSDP already
            #       shards/unshards the base weights so we don't do it ourselves
470
            nvtx_range_push(f"{nvtx_label}.fsdp_gather")
471
472
473
474
            _fsdp_gather_tensors(
                ctx.fsdp_group,
                ctx.fsdp_shapes,
                inputmat,
475
                weight_fp8,
476
            )
477
            nvtx_range_pop(f"{nvtx_label}.fsdp_gather")
478

479
            # Configure Userbuffers communication (comm+GEMM overlap)
480
            ctx.ub_obj_gradout = None
481
            ub_obj_dgrad = None
482
            ub_obj_wgrad = None
483
484
            ub_type_dgrad = None
            ub_type_wgrad = None
485
            dgrad_shape = [reduce(multiply_op, ctx.inp_shape[:-1]), ctx.inp_shape[-1]]
486
            if ctx.ub_overlap_ag:
487
                # Overlap grad_output all-gather with dgrad compute
488
                ctx.ub_obj_gradout = get_ub(ctx.ub_name + "_dgrad")
489
490
                ub_obj_dgrad = ctx.ub_obj_gradout
                ub_type_dgrad = tex.CommOverlapType.AG
491
492
493
            elif ctx.ub_overlap_rs_dgrad:
                # Overlap dgrad reduce-scatter with dgrad compute
                ctx.ub_obj_gradout = get_ub(ctx.ub_name + "_dgrad")
494
495
                ub_obj_dgrad = ctx.ub_obj_gradout
                ub_type_dgrad = tex.CommOverlapType.RS
496
497
498
499
            else:
                if ctx.ub_bulk_dgrad:
                    # Overlap inputmat all-gather with dgrad compute
                    ctx.ub_obj_gradout = get_ub(ctx.ub_name + "_dgrad")
500
501
                    ub_obj_dgrad = ctx.ub_obj_gradout
                    ub_type_dgrad = tex.CommOverlapType.AG
502
503
504
                if ctx.ub_bulk_wgrad:
                    # Overlap dgrad reduce-scatter with wgrad compute
                    ub_obj_wgrad = get_ub(ctx.ub_name + "_wgrad")
505
                    ub_type_wgrad = tex.CommOverlapType.RS
506
507
508
509
510
511
512
513

            # --------------------------------------------------
            # Prepare grad output tensor
            # Note: Cast to expected dtype and perform tensor-parallel communication
            # --------------------------------------------------

            # Unmodified grad output tensor
            grad_output_arg = grad_output
514

515
516
517
518
            # Configure quantizer for grad output tensor
            # Note: dgrad GEMM requires row-wise usage, wgrad GEMM
            # requires column-wise usage
            if ctx.grad_output_quantizer is not None:
519
520
521
522
523
524
525
                quantizer = ctx.grad_output_quantizer
                quantizer.set_usage(rowwise=True, columnwise=True)
                if ctx.ub_overlap_ag:
                    # Userbuffers only supports communication for one
                    # tensor usage at a time. Configure quantizer with
                    # usage for only dgrad GEMM.
                    quantizer.set_usage(columnwise=False)
526

527
            # Prepare grad output tensor
528
            nvtx_range_push(f"{nvtx_label}.grad_output_preprocess")
529
530
531
532
            (
                grad_output,
                grad_bias,
            ) = TransformerEngineBaseModule.grad_output_preprocess(
533
534
535
536
                ctx,
                grad_output,
                ctx.parallel_mode == "row",
                ctx.grad_output_quantizer,
537
            )
538
            nvtx_range_pop(f"{nvtx_label}.grad_output_preprocess")
539

540
541
542
543
544
545
546
547
548
549
            # --------------------------------------------------
            # Grad output tensor is ready for computing grad input...
            # --------------------------------------------------

            # --------------------------------------------------
            # Prepare input tensor
            # Note: Input tensor is needed for wgrad GEMM.
            # Tensor-parallel communication is overlapped with dgrad
            # GEMM.
            # --------------------------------------------------
550
            inputmat_total = None
551
            inputmat_total_work = None
552
            if ctx.backward_input_needs_gather:
553
                quantizer = None
554
                if ctx.fp8 or ctx.debug:
555
                    quantizer = ctx.input_quantizer
556
557
558
559
560
561
                    if isinstance(quantizer, (Float8Quantizer, Float8CurrentScalingQuantizer)):
                        # If data is in FP8, we compute FP8 transposes manually
                        quantizer.set_usage(rowwise=True, columnwise=False)
                    else:
                        # wgrad GEMM requires input with column-wise usage
                        quantizer.set_usage(rowwise=False, columnwise=True)
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
                if ctx.ub_bulk_dgrad:
                    inputmat_total, _ = fill_userbuffers_buffer_for_all_gather(
                        ub_obj_dgrad,
                        inputmat,
                        quantizer,
                        ctx.tp_group,
                    )
                else:
                    nvtx_range_push(f"{nvtx_label}.column_parallel_comm_input")
                    inputmat_total, inputmat_total_work = gather_along_first_dim(
                        inputmat,
                        ctx.tp_group,
                        async_op=True,
                        quantizer=quantizer,
                    )
                    nvtx_range_pop(f"{nvtx_label}.column_parallel_comm_input")
578
579
            else:
                inputmat_total = inputmat
580
581
582
            # --------------------------------------------------
            # Input tensor is ready for computing grad weight...
            # --------------------------------------------------
583

584
            # --------------------------------------------------
585
            # Compute grad input tensor
586
587
            # --------------------------------------------------

588
589
            dgrad = None
            dgrad_work = None
590
            if ctx.requires_dgrad:
591

592
593
594
595
596
597
598
599
                # Make sure required data is available
                if isinstance(grad_output, QuantizedTensorBase):
                    grad_output.update_usage(rowwise_usage=True)
                if ctx.weight_quantizer is not None and isinstance(weight_fp8, QuantizedTensorBase):
                    weight_fp8.update_usage(columnwise_usage=True)

                # Choose whether to use GEMM kernel with split accumulator
                use_split_accumulator = _2X_ACC_DGRAD
600
601
602
                if ctx.fp8:
                    recipe = ctx.fp8_recipe
                    if hasattr(recipe, "fp8_gemm_dgrad"):
603
                        use_split_accumulator = recipe.fp8_gemm_dgrad.use_split_accumulator
604

605
606
607
608
609
610
611
612
613
614
                # Update grad input quantizer
                if ctx.grad_input_quantizer is not None:
                    ctx.grad_input_quantizer.set_usage(rowwise=True, columnwise=False)

                # Output buffers for Userbuffers reduce-scatter
                gemm_out = None
                reduce_scatter_out = None
                if ctx.ub_overlap_rs_dgrad:
                    reduce_scatter_out = torch.empty(
                        dgrad_shape, dtype=ctx.activation_dtype, device=grad_output_arg.device
615
                    )
616
617
                elif ctx.ub_bulk_wgrad:
                    gemm_out = ub_obj_wgrad.get_buffer(local_chunk=False)
618

619
620
621
622
                # dgrad GEMM
                # Note: dx = dy * w
                nvtx_range_push(f"{nvtx_label}.dgrad_gemm")
                gemm_out, *_, reduce_scatter_out = general_gemm(
623
624
625
626
627
628
                    weight_fp8,
                    grad_output,
                    get_workspace(),
                    layout="NN",
                    grad=True,
                    quantization_params=ctx.grad_input_quantizer,
629
                    out=gemm_out,
630
                    out_dtype=ctx.activation_dtype,
631
                    use_split_accumulator=use_split_accumulator,
632
633
                    ub=ub_obj_dgrad,
                    ub_type=ub_type_dgrad,
634
                    extra_output=reduce_scatter_out,
635
636
                    bulk_overlap=ctx.ub_bulk_dgrad,
                )
637
                nvtx_range_pop(f"{nvtx_label}.dgrad_gemm")
638

639
640
                # Prepare grad input tensor
                # Note: Perform tensor-parallel communication
641
                if ctx.ub_overlap_rs_dgrad:
642
643
644
645
                    dgrad = reduce_scatter_out
                elif ctx.ub_bulk_wgrad:
                    dgrad = ub_obj_wgrad.get_buffer(local_chunk=True)
                elif ctx.parallel_mode == "column" and ctx.tp_size > 1:
646
                    nvtx_range_push(f"{nvtx_label}.column_parallel_comm_dgrad")
647
                    dgrad = gemm_out
648
649
650
651
652
                    if ctx.sequence_parallel:
                        dgrad, dgrad_work = reduce_scatter_along_first_dim(
                            dgrad,
                            ctx.tp_group,
                            async_op=True,
653
                        )
654
                    else:
655
                        dgrad, dgrad_work = allreduce(dgrad, ctx.tp_group, async_op=True)
656
                    nvtx_range_pop(f"{nvtx_label}.column_parallel_comm_dgrad")
657
658
659
660
661
662
663
664
665
666
                else:
                    dgrad = gemm_out

            # --------------------------------------------------
            # Grad input tensor has been computed...
            # --------------------------------------------------

            # --------------------------------------------------
            # Compute grad weight
            # --------------------------------------------------
667

668
669
            wgrad = None
            if ctx.requires_wgrad:
670

671
672
673
                # Prepare input tensor
                # Note: Synchronize tensor-parallel communication and
                # make sure required data is available
674
675
676
                if inputmat_total_work is not None:
                    inputmat_total_work.wait()
                    inputmat_total_work = None
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
                if ctx.fp8 or ctx.debug:
                    if isinstance(inputmat_total, QuantizedTensorBase):
                        inputmat_total.update_usage(columnwise_usage=True)
                    else:
                        ctx.input_quantizer.set_usage(rowwise=False, columnwise=True)
                        inputmat_total = ctx.input_quantizer(inputmat_total)

                # Prepare grad output tensor
                # Note: Synchronize tensor-parallel communication and
                # make sure required data is available
                if ctx.ub_overlap_ag and isinstance(ctx.grad_output_quantizer, MXFP8Quantizer):
                    # UB does not support overlapping grad output
                    # all-gather with wgrad GEMM. Also, we can't
                    # convert row-scaled MXFP8 to column-scaled, so we
                    # can't reuse the grad output that was gathered
692
693
                    # for the dgrad GEMM. We work around by explicitly
                    # overlapping the NCCL operation with the dgrad GEMM.
694
                    ctx.grad_output_quantizer.set_usage(rowwise=False, columnwise=True)
695
696
697
698
699
700
701
702
703
704
705
706
707
708
                    # Get the communication stream from the dgrad GEMM and set it as the current torch stream
                    dgrad_comm_stream = ub_obj_dgrad.get_communication_stream()
                    with torch.cuda.stream(dgrad_comm_stream):
                        # Syncs with the current stream (dgrad_comm_stream) before starting the all-gather
                        # This ensures that we don't start until all communication for the dgrad GEMM is complete
                        grad_output, grad_output_work = gather_along_first_dim(
                            grad_output_arg,
                            ctx.tp_group,
                            async_op=True,
                            quantizer=ctx.grad_output_quantizer,
                        )
                    # Synchronize with the main stream
                    grad_output_work.wait()

709
710
711
712
713
714
                if ctx.fp8 or ctx.debug:
                    if isinstance(grad_output, QuantizedTensorBase):
                        grad_output.update_usage(columnwise_usage=True)
                    else:
                        ctx.grad_output_quantizer.set_usage(rowwise=False, columnwise=True)
                        grad_output = ctx.grad_output_quantizer(grad_output)
715

716
717
718
719
720
721
722
                # Figure out whether to use split accumulator
                use_split_accumulator = _2X_ACC_WGRAD
                if ctx.fp8:
                    recipe = ctx.fp8_recipe
                    if hasattr(recipe, "fp8_gemm_wgrad"):
                        use_split_accumulator = recipe.fp8_gemm_wgrad.use_split_accumulator

723
724
725
726
727
728
729
730
731
                # Figure out whether to output wgrad GEMM directly into main grad
                if ctx.is_first_microbatch is not None:
                    accumulate_wgrad_into_param_main_grad = (
                        ctx.fuse_wgrad_accumulation and not ctx.is_first_microbatch
                    )
                else:
                    accumulate_wgrad_into_param_main_grad = ctx.fuse_wgrad_accumulation

                # Output buffer for overlapping FP8 grad input
732
                # reduce-scatter with wgrad GEMM
733
                reduce_scatter_out = None
734
                if ctx.ub_bulk_wgrad and ub_obj_wgrad.is_fp8_ubuf():
735
736
                    reduce_scatter_out = torch.empty(
                        dgrad_shape, dtype=ctx.activation_dtype, device=grad_output_arg.device
737
738
                    )

739
740
741
742
                # Arguments to include in wgrad GEMM closure
                wgrad_gemm_kwargs = {
                    "workspace": get_workspace(),
                    "out_dtype": (
743
744
                        main_grad.dtype if ctx.fuse_wgrad_accumulation else ctx.activation_dtype
                    ),
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
                    "quantization_params": ctx.grad_weight_quantizer,
                    "accumulate": accumulate_wgrad_into_param_main_grad,
                    "layout": "NT",
                    "out": main_grad if ctx.fuse_wgrad_accumulation else None,
                    "bias": (bias if (grad_bias is None and not ctx.fp8) else None),
                    "use_split_accumulator": use_split_accumulator,
                    "grad": True,
                    "ub": ub_obj_wgrad,
                    "ub_type": ub_type_wgrad,
                    "extra_output": reduce_scatter_out,
                    "bulk_overlap": ctx.ub_bulk_wgrad,
                }

                def wgrad_gemm(
                    x: torch.Tensor,
                    dy: torch.Tensor,
                ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
                    """Perform wgrad GEMM: dw = dy^T * x

                    May be fused with bgrad computation.

                    May be called outside of this function to enable
                    some advanced communication/compute overlapping.

                    """
                    nvtx_range_push(f"{nvtx_label}.wgrad_gemm")
                    dw, db, *_ = general_gemm(x, dy, **wgrad_gemm_kwargs)
                    nvtx_range_pop(f"{nvtx_label}.wgrad_gemm")
                    return dw, db

                # Choose whether to call wgrad GEMM now or delay
776
                if ctx.wgrad_store is not None and ctx.wgrad_store.delay_wgrad_compute():
777
778
779
780
781
782
783
784
785
786
787
                    if (
                        wgrad_gemm_kwargs["ub"] is not None
                        or wgrad_gemm_kwargs["ub_type"] is not None
                        or wgrad_gemm_kwargs["extra_output"] is not None
                        or wgrad_gemm_kwargs["bulk_overlap"]
                    ):
                        raise NotImplementedError(
                            "Delayed weight grad computation is not supported "
                            "with Userbuffers (tensor-parallel communication overlapping)"
                        )
                    ctx.wgrad_store.put([inputmat_total, grad_output], wgrad_gemm)
788
789
                else:

790
791
792
793
                    # Call wgrad GEMM now
                    wgrad, grad_bias_ = wgrad_gemm(inputmat_total, grad_output)

                    # Update grad bias if needed
794
795
796
797
                    if grad_bias is None:
                        grad_bias = grad_bias_
                    del grad_bias_

798
                    # Deallocate input tensor if permitted
799
800
                    if ctx.owns_input:
                        clear_tensor_data(inputmat_total)
801

802
                # Update grad input if overlapping reduce-scatter with wgrad GEMM
803
804
                if ctx.ub_bulk_wgrad:
                    if ub_obj_wgrad.is_fp8_ubuf():
805
                        dgrad = reduce_scatter_out
806
                    else:
807
808
809
810
811
                        dgrad = ub_obj_wgrad.get_buffer(local_chunk=True).clone()

            # --------------------------------------------------
            # Grad weight has been computed...
            # --------------------------------------------------
812

813
            # Don't return grad bias if not needed
814
815
816
            if not ctx.use_bias:
                grad_bias = None

817
            # Make sure all tensor-parallel communication is finished
818
819
820
821
822
823
824
825
            if inputmat_total_work is not None:
                inputmat_total_work.wait()
                inputmat_total_work = None
            if dgrad_work is not None:
                dgrad_work.wait()
                dgrad_work = None

        if ctx.requires_wgrad:
826
            # Handle custom DDP from mcore.
827
828
829
830
831
            if (
                ctx.fuse_wgrad_accumulation
                and weight is not None
                and hasattr(weight, "grad_added_to_main_grad")
            ):
832
                weight.grad_added_to_main_grad = True
833
                if getattr(weight, "zero_out_wgrad", False):
834
835
836
837
                    wgrad = get_dummy_wgrad(
                        list(weight.main_grad.shape),
                        weight.dtype,
                        zero=True,
838
                    )
839
                else:
840
841
842
                    wgrad = get_dummy_wgrad(
                        list(weight.main_grad.shape),
                        weight.dtype,
843
                    )
844
845
846
847
            elif ctx.fuse_wgrad_accumulation:
                wgrad = None
        else:
            wgrad = None
848

849
        # Update FP8 scaling factors if needed
850
        if ctx.reduce_and_update_bwd_fp8_tensors and not is_graph_capturing():
851
            nvtx_range_push(f"{nvtx_label}.reduce_and_update_fp8_tensors")
852
            FP8GlobalStateManager.reduce_and_update_fp8_tensors(forward=False)
853
            nvtx_range_pop(f"{nvtx_label}.reduce_and_update_fp8_tensors")
854

855
        # Scatter fp8 weight buffers
856
        if ctx.fp8 and not isinstance(weight, QuantizedTensorBase):
857
            _fsdp_scatter_tensors(ctx.fsdp_group, weight_fp8)
858
        return (
859
            wgrad,
860
861
            dgrad.view(ctx.inp_shape) if ctx.requires_dgrad else None,
            grad_bias,
862
863
864
            None,  # is_first_microbatch
            None,  # fp8
            None,  # fp8_calibration
865
            None,  # wgrad_store
866
867
868
869
            None,  # input_quantizer
            None,  # weight_quantizer
            None,  # output_quantizer
            None,  # grad_input_quantizer
870
871
            None,  # grad_weight_quantizer
            None,  # grad_output_quantizer
872
873
874
875
876
877
878
879
880
            None,  # fuse_wgrad_accumulation
            None,  # cpu_offloading
            None,  # tp_group
            None,  # tp_size
            None,  # sequence_parallel
            None,  # tensor_parallel
            None,  # activation_dtype
            None,  # parallel_mode
            None,  # is_grad_enabled
881
882
883
884
885
886
            None,  # ub_overlap_rs_fprop
            None,  # ub_overlap_ag_dgrad
            None,  # ub_overlap_ag_fprop
            None,  # ub_overlap_rs_dgrad
            None,  # ub_bulk_dgrad
            None,  # ub_bulk_wgrad
887
            None,  # ub_name
888
            None,  # fp8_output
889
            None,  # fsdp_group
890
891
            None,  # module
            None,  # skip_fp8_weight_update
892
            None,  # symmetric_ar_type
893
            None,  # debug
894
895
896
897
        )


class Linear(TransformerEngineBaseModule):
898
    """Applies a linear transformation to the incoming data :math:`y = xA^T + b`
899
900
901
902
903
904
905
906
907
908
909
910
911
912

    On NVIDIA GPUs it is a drop-in replacement for `torch.nn.Linear`.

    Parameters
    ----------
    in_features : int
                 size of each input sample.
    out_features : int
                  size of each output sample.
    bias : bool, default = `True`
          if set to `False`, the layer will not learn an additive bias.
    init_method : Callable, default = `None`
                 used for initializing weights in the following way: `init_method(weight)`.
                 When set to `None`, defaults to `torch.nn.init.normal_(mean=0.0, std=0.023)`.
913
    get_rng_state_tracker : Callable, default = `None`
914
                 used to get the random number generator state tracker for initializing weights.
915
916
    rng_tracker_name : str, default = `None`
                 the param passed to get_rng_state_tracker to get the specific rng tracker.
cyanguwa's avatar
cyanguwa committed
917
    parameters_split : Optional[Union[Tuple[str, ...], Dict[str, int]]], default = None
918
919
920
921
922
923
924
                      Configuration for splitting the weight and bias tensors along dim 0 into
                      multiple PyTorch parameters. If a list or tuple of strings is provided,
                      they are used to make the names of equally-sized parameters. If a dict
                      (preferably an OrderedDict) is provided, the keys are used as names and
                      values as split sizes along dim 0. The resulting parameters will have
                      names that end in `_weight` or `_bias`, so trailing underscores are
                      stripped from any provided names.
925
    device : Union[torch.device, str], default = "cuda"
926
          The device on which the parameters of the model will be allocated. It is the user's
927
928
          responsibility to ensure all parameters are moved to the GPU before running the
          forward pass.
929
930
    name: str, default = `None`
        name of the module, currently used for debugging purposes.
931
932
933
934
935
936
937
938
939
940
941
942
943

    Parallelism parameters
    ----------------------
    sequence_parallel : bool, default = `False`
                       if set to `True`, uses sequence parallelism.
    tp_group : ProcessGroup, default = `None`
              tensor parallel process group.
    tp_size : int, default = 1
             used as TP (tensor parallel) world size when TP groups are not formed during
             initialization. In this case, users must call the
             `set_tensor_parallel_group(tp_group)` method on the initialized module before the
             forward pass to supply the tensor parallel group needed for tensor and sequence
             parallel collectives.
944
    parallel_mode : {None, 'column', 'row'}, default = `None`
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
                   used to decide whether this Linear layer is Column Parallel Linear or Row
                   Parallel Linear as described `here <https://arxiv.org/pdf/1909.08053.pdf>`_.
                   When set to `None`, no communication is performed.

    Optimization parameters
    -----------------------
    fuse_wgrad_accumulation : bool, default = 'False'
                             if set to `True`, enables fusing of creation and accumulation of
                             the weight gradient. When enabled, it is assumed that the weights
                             have an additional `main_grad` attribute (used instead of the
                             regular `grad`) which is a pre-allocated buffer of the correct
                             size to accumulate gradients in.
    return_bias : bool, default = `False`
                 when set to `True`, this module will not apply the additive bias itself, but
                 instead return the bias value during the forward pass together with the
                 output of the linear transformation :math:`y = xA^T`. This is useful when
                 the bias addition can be fused to subsequent operations.
962
    params_dtype : torch.dtype, default = `torch.get_default_dtype()`
963
964
965
                  it controls the type used to allocate the initial parameters. Useful when
                  the model is trained with lower precision and the original FP32 parameters
                  would not fit in GPU memory.
966
967
968
969
    delay_wgrad_compute : bool, default = `False`
                         Whether or not to delay weight gradient computation. If set to `True`,
                         it's the user's responsibility to call `module.backward_dw` to compute
                         weight gradients.
970
971
972
973
974
    symmetric_ar_type : {None, 'multimem_all_reduce', 'two_shot', 'one_shot'}, default = None
                   Type of symmetric memory all-reduce to use during the forward pass.
                   This can help in latency bound communication situations.
                   Requires PyTorch version 2.7.0 or higher. When set to None, standard all-reduce
                   is used.
975
976
977
978
979
980
981
982
983
984
985
    """

    def __init__(
        self,
        in_features: int,
        out_features: int,
        sequence_parallel: bool = False,
        fuse_wgrad_accumulation: bool = False,
        tp_group: Optional[dist_group_type] = None,
        tp_size: int = 1,
        get_rng_state_tracker: Optional[Callable] = None,
986
        rng_tracker_name: Optional[str] = None,
987
988
989
        init_method: Optional[Callable] = None,
        bias: bool = True,
        return_bias: bool = False,
990
        params_dtype: Optional[torch.dtype] = None,
991
        parallel_mode: Optional[str] = None,
cyanguwa's avatar
cyanguwa committed
992
        parameters_split: Optional[Union[Tuple[str, ...], Dict[str, int]]] = None,
993
        device: Union[torch.device, str] = "cuda",
994
        ub_overlap_ag: bool = False,
995
        ub_overlap_rs: bool = False,
996
        ub_overlap_rs_dgrad: bool = False,
997
998
        ub_bulk_dgrad: bool = False,
        ub_bulk_wgrad: bool = False,
999
        ub_name: Optional[str] = None,
1000
        delay_wgrad_compute: bool = False,
1001
        symmetric_ar_type: Optional[str] = None,
1002
        name: Optional[str] = None,
1003
1004
    ) -> None:
        super().__init__()
1005
1006

        params_dtype = torch.get_default_dtype() if params_dtype is None else params_dtype
1007
1008
1009
1010
1011
1012
        self.in_features = in_features
        self.out_features = out_features
        self.fuse_wgrad_accumulation = fuse_wgrad_accumulation
        self.use_bias = bias
        self.return_bias = return_bias
        self.apply_bias = bias and not return_bias
1013
1014
        self.get_rng_state_tracker = get_rng_state_tracker
        self.rng_tracker_name = rng_tracker_name
1015
        self.symmetric_ar_type = symmetric_ar_type
1016
1017
1018
1019
        self.name = name

        if TEDebugState.debug_enabled:
            self._turn_off_unsupported_features_in_debug()  # turn off userbuffers
1020

1021
1022
        self.wgrad_store = WeightGradStore(delay_wgrad_compute, ub_bulk_wgrad)

1023
1024
        if device == "meta":
            assert parameters_split is None, "Cannot split module parameters on 'meta' device."
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
        if tp_group is None:
            self.tp_size = tp_size
            if tp_size == 1:
                self.set_tensor_parallel_group(tp_group)
        else:
            self.tp_size = get_distributed_world_size(tp_group)
            self.set_tensor_parallel_group(tp_group)
        self.set_nccl_overlap_warning_if_tp()

        self.parallel_mode = parallel_mode
        assert (
            self.parallel_mode in GemmParallelModes
        ), f"parallel_mode {parallel_mode} not supported"

        if self.parallel_mode == "column":
            self.out_features = divide(self.out_features, self.tp_size)
        elif self.parallel_mode == "row":
            self.in_features = divide(self.in_features, self.tp_size)

        self.sequence_parallel = (self.tp_size > 1) and sequence_parallel

1046
        # Column parallel TP overlap options
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
        self.ub_overlap_ag_fprop = (
            self.parallel_mode == "column" and self.sequence_parallel and ub_overlap_ag
        )
        self.ub_overlap_rs_dgrad = (
            self.parallel_mode == "column" and self.sequence_parallel and ub_overlap_rs_dgrad
        )
        self.ub_bulk_dgrad = (
            self.parallel_mode == "column"
            and self.sequence_parallel
            and ub_bulk_dgrad
            and not self.ub_overlap_rs_dgrad
        )
        self.ub_bulk_wgrad = (
            self.parallel_mode == "column"
            and self.sequence_parallel
            and ub_bulk_wgrad
            and not self.ub_overlap_rs_dgrad
        )
1065
1066

        # Row parallel TP overlap options
1067
1068
1069
1070
1071
1072
        self.ub_overlap_rs_fprop = (
            self.parallel_mode == "row" and self.sequence_parallel and ub_overlap_rs
        )
        self.ub_overlap_ag_dgrad = (
            self.parallel_mode == "row" and self.sequence_parallel and ub_overlap_ag
        )
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086

        if any(
            [
                self.ub_overlap_rs_fprop,
                self.ub_overlap_ag_dgrad,
                self.ub_overlap_ag_fprop,
                self.ub_overlap_rs_dgrad,
                self.ub_bulk_dgrad,
                self.ub_bulk_wgrad,
            ]
        ):
            assert ub_name is not None, f"Comm+GEMM overlap layer '{ub_name}' is not initialized."
        self.ub_name = ub_name

1087
1088
1089
1090
1091
1092
1093
        if self.symmetric_ar_type is not None:
            assert torch_version() >= (
                2,
                7,
                0,
            ), "Torch version must be at least 2.7 to use symmetric memory"

1094
1095
1096
        # Initialize params in FP8
        with_fp8_params = FP8GlobalStateManager.with_fp8_parameters()

1097
1098
1099
1100
1101
1102
1103
1104
        # Contiguous buffers for params
        weight_tensor = torch.empty(
            self.out_features,
            self.in_features,
            device=device,
            dtype=params_dtype,
        )
        bias_tensor = None
1105
        if self.use_bias:
1106
1107
1108
1109
1110
            bias_tensor = torch.empty(
                self.out_features,
                device=device,
                dtype=params_dtype,
            )
1111

1112
1113
1114
1115
        # Configure parameter splits
        self.weight_names = []
        self.bias_names = []
        self.parameter_split_sizes = []
1116
        if parameters_split is None:
1117
1118
1119
1120
1121
1122
            # Split into a single parameter by default
            self.weight_names = ["weight"]
            self.bias_names = ["bias"]
            self.parameter_split_sizes = [out_features]
        elif not parameters_split:
            raise ValueError("Cannot split weight buffer into 0 parameters")
cyanguwa's avatar
cyanguwa committed
1123
        elif isinstance(parameters_split, dict):
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
            # Split parameters with provided sizes
            for name, split_size in parameters_split.items():
                self.weight_names.append(f"{name.rstrip('_')}_weight")
                self.bias_names.append(f"{name.rstrip('_')}_bias")
                self.parameter_split_sizes.append(split_size)
        elif all(isinstance(name, str) for name in parameters_split):
            # Split parameters evenly
            split_size = out_features // len(parameters_split)
            for name in parameters_split:
                self.weight_names.append(f"{name.rstrip('_')}_weight")
                self.bias_names.append(f"{name.rstrip('_')}_bias")
                self.parameter_split_sizes.append(split_size)
cyanguwa's avatar
cyanguwa committed
1136
        else:
1137
            raise TypeError("Invalid configuration for parameters split")
1138

1139
1140
1141
1142
1143
1144
        # Make sure parameter splits are valid
        if sum(self.parameter_split_sizes) != out_features:
            raise ValueError(
                f"Trying to split weight buffer ({out_features=}) "
                f"with split sizes {self.parameter_split_sizes}"
            )
1145

1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
        # Adjust parameter splits for tensor-parallel distribution
        if self.parallel_mode == "column":
            for i, size in enumerate(self.parameter_split_sizes):
                if size % self.tp_size != 0:
                    raise RuntimeError(
                        f"Attempting to distribute a parameter with out_features={size} "
                        f"between {self.tp_size} tensor-parallel processes"
                    )
                self.parameter_split_sizes[i] = size // self.tp_size

1156
1157
1158
1159
1160
        # Construct weight parameters
        # Note: Register weights together so that they are adjacent to
        # each other in Linear.parameters(). This makes it more likely
        # that they will stay contiguous if the weights are
        # manipulated externally, e.g. by FSDP.
1161
1162
1163
1164
1165
1166
1167
1168
        offset = 0
        for i, split_size in enumerate(self.parameter_split_sizes):
            split_start = offset
            offset += split_size
            split_end = offset

            # Check if parameters are subviews of buffers
            is_subview = (split_start, split_end) != (0, self.out_features)
1169
            if is_subview and with_fp8_params:
1170
1171
1172
                raise RuntimeError(
                    "Splitting QuantizedTensor into multiple params is not supported"
                )
1173

1174
            # Construct weight parameter
1175
1176
1177
1178
1179
1180
1181
            self.register_parameter(
                self.weight_names[i],
                torch.nn.Parameter(weight_tensor[split_start:split_end]),
                init_fn=init_method,
                get_rng_state_tracker=get_rng_state_tracker,
                fp8_meta_index=tex.FP8FwdTensors.GEMM1_WEIGHT,
            )
1182

1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
        # Construct bias parameters if needed
        if self.use_bias:
            offset = 0
            for i, split_size in enumerate(self.parameter_split_sizes):
                split_start = offset
                offset += split_size
                split_end = offset
                self.register_parameter(
                    self.bias_names[i],
                    torch.nn.Parameter(bias_tensor[split_start:split_end]),
                    init_fn=init_method_constant(0.0),
                )
        else:
            for name in self.bias_names:
                bias = torch.Tensor().to(dtype=params_dtype, device=device)
                setattr(self, name, bias)
cyanguwa's avatar
cyanguwa committed
1199

1200
        if with_fp8_params:
1201
1202
            self.init_fp8_metadata()

1203
        self.reset_parameters(defer_init=device == "meta")
1204

1205
1206
1207
1208
1209
1210
1211
        # For RPL, bias has to be added after TP collectives
        # So it cannot be fused with the GEMM
        if self.parallel_mode == "row" and self.apply_bias:
            self.gemm_bias_unfused_add = True
        else:
            self.gemm_bias_unfused_add = False

1212
1213
1214
1215
1216
1217
1218
1219
    def set_meta_tensor(self, fwd: bool, recipe: Recipe) -> None:
        """Init scales and amaxes for fwd | bwd."""
        super().set_meta_tensor(fwd, recipe)

        # customize quantizers based on each recipe & layer configs
        recipe = FP8GlobalStateManager.get_fp8_recipe()
        if recipe.float8_current_scaling():
            self._customize_quantizers_float8_current_scaling(fwd, recipe)
1220
1221
        elif recipe.float8_block_scaling():
            self._customize_quantizers_float8_blockwise_scaling(fwd, recipe)
1222
1223
        # elif for other recipes (mxfp8, etc.)

1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
    def reset_parameters(self, defer_init=False):
        super().reset_parameters(defer_init=defer_init)

        if not defer_init:
            # Set parallelism attributes for linear weights
            for weight in self.weight_names:
                set_tensor_model_parallel_attributes(
                    tensor=getattr(self, weight),
                    is_parallel=True,
                    dim=1 if self.parallel_mode == "row" else 0,
                    stride=1,
                )

            # Set parallelism attributes for linear biases
            if self.use_bias:
                for bias in self.bias_names:
                    if self.parallel_mode == "row":
                        setattr(getattr(self, bias), "sequence_parallel", self.sequence_parallel)
                    elif self.parallel_mode == "column":
                        set_tensor_model_parallel_attributes(getattr(self, bias), True, 0, 1)

1245
    @no_torch_dynamo()
1246
1247
1248
1249
    def forward(
        self,
        inp: torch.Tensor,
        is_first_microbatch: Optional[bool] = None,
1250
        fp8_output: Optional[bool] = False,
1251
        fp8_grad: Optional[bool] = False,
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
    ) -> Union[torch.Tensor, Tuple[torch.Tensor, ...]]:
        """
        Apply the linear transformation to the input.

        Parameters
        ----------
        inp : torch.Tensor
             Input tensor.
        is_first_microbatch : {True, False, None}, default = None
                             During training using either gradient accumulation or
                             pipeline parallelism a minibatch of data is further split
                             into microbatches. Between the microbatches of the same minibatch
                             the model weights are not updated. Setting this parameter indicates
                             whether the current microbatch is the first in a minibatch or not.
                             When set, this parameter enables additional optimizations:

                             * during FP8 training, it allows caching of the FP8 versions of
                               the weights
                             * it also allows skipping gradient accumulation during the
                               first microbatch (since it is the first gradient being
                               produced)
        """
1274
1275
1276
1277
        debug = TEDebugState.debug_enabled
        if debug:
            self._validate_name()

1278
1279
1280
1281
        if FP8GlobalStateManager.fp8_graph_capturing():
            skip_fp8_weight_update = FP8GlobalStateManager.get_skip_fp8_weight_update_tensor()
        else:
            skip_fp8_weight_update = None
1282
1283
1284
        if skip_fp8_weight_update is not None:
            is_first_microbatch = False

1285
1286
1287
1288
1289
1290
1291
        if self.ub_overlap_rs_fprop:
            if get_ub(self.ub_name + "_fprop").is_fp8_ubuf():
                fp8_output = True
        if self.ub_overlap_rs_dgrad:
            if get_ub(self.ub_name + "_dgrad").is_fp8_ubuf():
                fp8_grad = True

1292
1293
        with self.prepare_forward(
            inp,
1294
            allow_non_contiguous=isinstance(inp, QuantizedTensor),
1295
        ) as inp:
1296
1297

            # Get concatenated weight and bias tensors
1298
            unfused_weights = self._get_weight_tensors()
1299
            weight_tensor = noop_cat(unfused_weights)
1300
            if self.use_bias:
1301
                bias_tensor = noop_cat([getattr(self, name) for name in self.bias_names])
1302
            else:
1303
1304
                bias_tensor = None

1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
            quantizers = (
                self._get_quantizers(fp8_output, fp8_grad)
                if not debug
                else self._get_debug_quantizers(fp8_output, fp8_grad)
            )
            if debug:
                if not any_feature_enabled(quantizers):
                    # If no feature is used, then run faster implementation with debug = False.
                    quantizers = self._get_quantizers(fp8_output, fp8_grad)
                    debug = False

                if isinstance(weight_tensor, QuantizedTensor):
                    raise RuntimeError("FP8 weights are not supported in debug mode.")

1319
1320
1321
1322
1323
            (
                input_quantizer,
                weight_quantizer,
                output_quantizer,
                grad_input_quantizer,
1324
1325
1326
                grad_weight_quantizer,
                grad_output_quantizer,
            ) = quantizers
1327

1328
1329
1330
1331
1332
1333
1334
1335
1336
            if torch.is_grad_enabled():
                linear_fn = _Linear.apply
                args = []
            else:
                linear_fn = _Linear.forward
                args = [None]
            args += (
                weight_tensor,
                inp,
1337
                bias_tensor if (self.apply_bias and not self.gemm_bias_unfused_add) else None,
1338
1339
1340
                is_first_microbatch,
                self.fp8,
                self.fp8_calibration,
1341
                self.wgrad_store,
1342
1343
1344
1345
                input_quantizer,
                weight_quantizer,
                output_quantizer,
                grad_input_quantizer,
1346
1347
                grad_weight_quantizer,
                grad_output_quantizer,
1348
                self.fuse_wgrad_accumulation,
1349
                is_cpu_offload_enabled(),
1350
1351
1352
1353
1354
1355
1356
                self.tp_group,
                self.tp_size,
                self.sequence_parallel,
                self.tp_size > 1,
                self.activation_dtype,
                self.parallel_mode,
                torch.is_grad_enabled(),
1357
1358
1359
1360
1361
1362
                self.ub_overlap_rs_fprop,
                self.ub_overlap_ag_dgrad,
                self.ub_overlap_ag_fprop,
                self.ub_overlap_rs_dgrad,
                self.ub_bulk_dgrad,
                self.ub_bulk_wgrad,
1363
                self.ub_name,
1364
                fp8_output,
1365
                self.fsdp_group,
1366
1367
                self,
                skip_fp8_weight_update,
1368
                self.symmetric_ar_type,
1369
                debug,
1370
1371
1372
1373
1374
1375
1376
1377
            )
            out = linear_fn(*args)
        if self.gemm_bias_unfused_add:
            out = out + cast_if_needed(bias_tensor, self.activation_dtype)

        if self.return_bias:
            return out, cast_if_needed(bias_tensor, self.activation_dtype)
        return out
1378
1379
1380

    def _get_quantizers(self, fp8_output, fp8_grad):
        if not self.fp8:
1381
            return [None] * 6
1382
        grad_input_quantizer = None
1383
        grad_weight_quantizer = None
1384
1385
1386
        grad_output_quantizer = None
        output_quantizer = None
        input_quantizer = self.quantizers["scaling_fwd"][tex.FP8FwdTensors.GEMM1_INPUT]
1387
        input_quantizer.internal = True
1388
        (weight_quantizer,) = self._get_weight_quantizers()
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
        if fp8_output:
            output_quantizer = self.quantizers["scaling_fwd"][tex.FP8FwdTensors.GEMM1_OUTPUT]
        if torch.is_grad_enabled():
            grad_output_quantizer = self.quantizers["scaling_bwd"][tex.FP8BwdTensors.GRAD_OUTPUT1]
            grad_output_quantizer.internal = True
            if fp8_grad:
                grad_input_quantizer = self.quantizers["scaling_bwd"][tex.FP8BwdTensors.GRAD_INPUT1]
        return (
            input_quantizer,
            weight_quantizer,
            output_quantizer,
            grad_input_quantizer,
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
            grad_weight_quantizer,
            grad_output_quantizer,
        )

    def _get_debug_quantizers(self, fp8_output, fp8_grad):
        original_quantizers = self._get_quantizers(fp8_output, fp8_grad)
        assert TEDebugState.debug_enabled
        from ...debug.pytorch.debug_quantization import DebugQuantizer

        names = ["activation", "weight", "output", "dgrad", "wgrad", "gradient"]
        return tuple(
            DebugQuantizer(self.name, name, q, self.tp_group)
            for name, q in zip(names, original_quantizers)
1414
        )
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461

    def _customize_quantizers_float8_current_scaling(self, fwd: bool, recipe: Recipe) -> None:
        """Customize quantizers based on current scaling recipe + linear."""
        assert (
            recipe.float8_current_scaling()
        ), "current scaling recipe quantizer customization here"
        if fwd:
            # set configs about amax epsilon and power_2_scale
            self.quantizers["scaling_fwd"][
                tex.FP8FwdTensors.GEMM1_INPUT
            ].force_pow_2_scales = recipe.fp8_quant_fwd_inp.power_2_scale
            self.quantizers["scaling_fwd"][
                tex.FP8FwdTensors.GEMM1_INPUT
            ].amax_epsilon = recipe.fp8_quant_fwd_inp.amax_epsilon
            # also set weight quantizer with same amax_epsilon & power_2_scale
            self.quantizers["scaling_fwd"][
                tex.FP8FwdTensors.GEMM1_WEIGHT
            ].force_pow_2_scales = recipe.fp8_quant_fwd_weight.power_2_scale
            self.quantizers["scaling_fwd"][
                tex.FP8FwdTensors.GEMM1_WEIGHT
            ].amax_epsilon = recipe.fp8_quant_fwd_weight.amax_epsilon
            # paralle related
            if self.sequence_parallel and self.parallel_mode == "column":
                # customize input_quantizer with amax reduction TP group
                self.quantizers["scaling_fwd"][
                    tex.FP8FwdTensors.GEMM1_INPUT
                ].with_amax_reduction = True
                self.quantizers["scaling_fwd"][
                    tex.FP8FwdTensors.GEMM1_INPUT
                ].amax_reduction_group = self.tp_group
        else:
            # set grad_output_quantizer with amax epsilon and power_2_scale
            self.quantizers["scaling_bwd"][
                tex.FP8BwdTensors.GRAD_OUTPUT1
            ].force_pow_2_scales = recipe.fp8_quant_bwd_grad.power_2_scale
            self.quantizers["scaling_bwd"][
                tex.FP8BwdTensors.GRAD_OUTPUT1
            ].amax_epsilon = recipe.fp8_quant_bwd_grad.amax_epsilon
            # parallel related
            if self.sequence_parallel and self.parallel_mode == "row":
                # customize grad_output_quantizer with amax reduction TP group
                self.quantizers["scaling_bwd"][
                    tex.FP8BwdTensors.GRAD_OUTPUT1
                ].with_amax_reduction = True
                self.quantizers["scaling_bwd"][
                    tex.FP8BwdTensors.GRAD_OUTPUT1
                ].amax_reduction_group = self.tp_group
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486

    def _get_weight_tensors(self) -> List[Union[torch.Tensor, QuantizedTensorBase]]:
        """Get the weight tensors of the module."""
        unfused_weights = [getattr(self, name) for name in self.weight_names]
        if any(isinstance(w, QuantizedTensor) for w in unfused_weights):
            if self.fp8:
                if len(unfused_weights) != 1:
                    raise RuntimeError(
                        "Splitting QuantizedTensor into multiple params is not supported"
                    )
            else:
                warnings.warn(
                    "You are using quantized weights without quantized compute. "
                    "Please make sure this is intentional."
                )
                unfused_weights = [w.dequantize() for w in unfused_weights]
        return unfused_weights

    def _get_weight_quantizers(self) -> List[Quantizer]:
        """Get the weight quantizers of the module."""
        if not self.fp8:
            return [None]
        weight_quantizer = self.quantizers["scaling_fwd"][tex.FP8FwdTensors.GEMM1_WEIGHT]
        weight_quantizer.internal = True
        return [weight_quantizer]
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505

    def _customize_quantizers_float8_blockwise_scaling(self, fwd: bool, recipe: Recipe) -> None:
        """Customize quantizers based on blockwise scaling recipe + linear."""
        assert (
            recipe.float8_block_scaling()
        ), "blockwise scaling recipe quantizer customization here"

        if fwd:
            if self.sequence_parallel and self.parallel_mode == "column":
                # set compact for inp tensor X
                self.quantizers["scaling_fwd"][
                    tex.FP8FwdTensors.GEMM1_INPUT
                ].all_gather_usage = True
        else:
            if self.sequence_parallel and self.parallel_mode == "row":
                # set compact for grad_output tensor dY
                self.quantizers["scaling_bwd"][
                    tex.FP8BwdTensors.GRAD_OUTPUT1
                ].all_gather_usage = True