linear.py 74.6 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
import os
11
12
13

import torch

14
import transformer_engine_torch as tex
15

16
from transformer_engine.common.recipe import Recipe
17
from transformer_engine.pytorch import torch_version
18

19
from .base import (
20
21
    fill_userbuffers_buffer_for_all_gather,
    get_dummy_wgrad,
22
    get_ub,
23
    get_workspace,
24
25
26
27
28
    TransformerEngineBaseModule,
    _2X_ACC_FPROP,
    _2X_ACC_DGRAD,
    _2X_ACC_WGRAD,
)
29
from ._common import noop_cat, WeightGradStore
30
from ..fp8 import FP8GlobalStateManager
31
32
from ..utils import (
    cast_if_needed,
33
    clear_tensor_data,
34
    divide,
35
    init_method_constant,
36
37
    requires_grad,
    needs_quantized_gemm,
38
    assert_dim_for_fp8_exec,
39
40
    nvtx_range_pop,
    nvtx_range_push,
dongcl's avatar
dongcl committed
41
    get_activation_offloading,
42
43
44
45
46
)
from ..distributed import (
    set_tensor_model_parallel_attributes,
    get_distributed_world_size,
    allreduce,
47
    symmetric_all_reduce,
48
49
    reduce_scatter_along_first_dim,
    gather_along_first_dim,
50
    is_fp8_activation_recompute_enabled,
51
    in_fp8_activation_recompute_phase,
52
53
    _fsdp_scatter_tensors,
    _fsdp_gather_tensors,
54
55
)
from ..cpp_extensions import (
56
    general_gemm,
57
)
58
from ..constants import GemmParallelModes, dist_group_type
59
from ..jit import no_torch_dynamo
60
from ..graph import is_graph_capturing
61
62
from ..tensor.quantized_tensor import (
    QuantizedTensor,
63
    QuantizedTensorBase,
64
65
66
67
    Quantizer,
    prepare_for_saving,
    restore_from_saved,
)
68
from ..tensor.float8_tensor import Float8CurrentScalingQuantizer, Float8Quantizer
69
from ..tensor.mxfp8_tensor import MXFP8Quantizer
70
from ..export import is_in_onnx_export_mode, assert_warmed_up
71
from ..cpu_offload import is_cpu_offload_enabled, mark_activation_offload
72
from ...debug.pytorch.debug_state import TEDebugState
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
        fine_grained_activation_offloading: bool,
115
        fp8_output: bool,  # pylint: disable=unused-argument
116
        fsdp_group: Union[dist_group_type, None],
117
118
        module: torch.nn.Module,
        skip_fp8_weight_update: bool,
119
        symmetric_ar_type: str,
120
        save_original_input: bool = False,
121
        debug: Optional[bool] = False,
122
    ) -> torch.Tensor:
123
        # pylint: disable=missing-function-docstring
124

125
126
127
128
129
        # NVTX label for profiling
        nvtx_label = "transformer_engine._Linear.forward"
        if ub_name is not None:
            nvtx_label = f"{nvtx_label}.{ub_name}"

130
        # Make sure input dimensions are compatible
131
        out_features, in_features = weight.shape
132
        assert inp.shape[-1] == in_features, "GEMM not possible"
133

134
        # Configure tensor-parallel communication
135
        tp_world_size = get_distributed_world_size(tp_group)
136
137
138
139
        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
        )
140
141

        # Configure Userbuffers communication (comm+GEMM overlap)
142
143
144
145
146
147
        if debug:  # turn off userbuffers in debug mode
            ub_overlap_rs_fprop = False
            ub_overlap_ag_fprop = False
            ub_overlap_rs_dgrad = False
            ub_bulk_wgrad = False
            ub_bulk_dgrad = False
148
149
150
        ub_obj = None
        ub_type = None
        if ub_overlap_rs_fprop:
151
            ub_obj = get_ub(ub_name + "_fprop", fp8)
152
153
            ub_type = tex.CommOverlapType.RS
        elif ub_overlap_ag_fprop:
154
            ub_obj = get_ub(ub_name + "_fprop", fp8)
155
156
157
158
159
160
161
162
163
164
            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
165
        if fp8:
166
            assert_dim_for_fp8_exec(inputmat, weight)
167
168
169
170
171
            if save_original_input:
                assert not isinstance(
                    input_quantizer, Float8Quantizer
                ), "DelayedScaling recipe is not supported with save_original_input"

172
173
174
175
176
177
        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")
178
                if not isinstance(inputmat, QuantizedTensorBase):
179
180
                    own_quantized_input = True
                    input_quantizer.set_usage(rowwise=True, columnwise=backward_needs_input)
181
182
183
184
185
                    if isinstance(
                        input_quantizer, (Float8Quantizer, Float8CurrentScalingQuantizer)
                    ):
                        # All-gather is not supported with FP8 column-wise data
                        input_quantizer.set_usage(columnwise=False)
186
187
188
189
190
                    if save_original_input:
                        # No need for column-wise data since this
                        # tensor will not be cached for backward pass
                        input_quantizer.set_usage(columnwise=False)
                        own_quantized_input = False
191
                    inputmat = input_quantizer(inputmat)
192
            else:
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
                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)
218
                else:
219
220
                    if input_quantizer is None:
                        raise ValueError("Missing quantizer for input tensor")
221
222
223
                    input_quantizer.set_usage(
                        rowwise=True, columnwise=backward_needs_input and not save_original_input
                    )
224
                    inputmat = input_quantizer(inputmat)
225
                    own_quantized_input = True
226
            else:
227
228
                inputmat = cast_if_needed(inp, activation_dtype)  # Cast for AMP
            inputmat_total = inputmat
229
        nvtx_range_pop(f"{nvtx_label}.input_cast_comm")
230
231
232
        # ------------------------------------------------------
        # Input tensor is ready for GEMM...
        # ------------------------------------------------------
233

234
235
236
        # ------------------------------------------------------
        # Prepare weight tensor
        # ------------------------------------------------------
237
238
        weightmat = weight
        if fp8 or debug:
239
240
241
242
243
244
245
246
247
            # 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)
248
249

            # Get quantized weight
250
251
252
253
254
255
256
257
            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,
258
                workspace_dtype=activation_dtype,
259
            )
260
261
            weightmat.update_usage(rowwise_usage=True)

262
        else:
263
264
265
266
            weightmat = cast_if_needed(weightmat, activation_dtype)  # Cast for AMP
        # ------------------------------------------------------
        # Weight tensor is ready for GEMM...
        # ------------------------------------------------------
267
268
269

        # Cast bias to expected dtype
        bias_dtype = activation_dtype
270
        if needs_quantized_gemm(inputmat_total) and activation_dtype == torch.float32:
271
            # cuBLAS does not support FP8 GEMM with FP32 bias, so we cast to BF16
272
273
274
275
276
277
278
279
280
281
            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)

282
283
        # Choose whether to use GEMM kernel with split accumulator
        use_split_accumulator = _2X_ACC_FPROP
284
285
286
        if fp8:
            recipe = FP8GlobalStateManager.get_fp8_recipe()
            if hasattr(recipe, "fp8_gemm_fprop"):
287
288
289
290
291
                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)
292

293
294
295
296
297
298
299
300
301
302
303
304
305
306
        # 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(
307
308
309
310
            weightmat,
            inputmat_total,
            get_workspace(),
            quantization_params=output_quantizer,
311
            out_dtype=activation_dtype,
312
            bias=bias,
313
            use_split_accumulator=use_split_accumulator,
314
315
            ub=ub_obj,
            ub_type=ub_type,
316
            extra_output=reduce_scatter_out,
317
        )
318
        nvtx_range_pop(f"{nvtx_label}.gemm")
319
320
321
322
        # ------------------------------------------------------
        # Finished forward GEMM...
        # ------------------------------------------------------

323
324
325
326
327
328
329
        # Deallocate GEMM input tensor if no longer needed
        # TODO(yuzhongw, tmoon): Figure out why inputmat_total is not automatically
        # deallocated by GC. Manually deallocating is a temporary hack.
        if with_input_all_gather_nccl:
            clear_tensor_data(inputmat_total)
            inputmat_total = None

330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
        # ------------------------------------------------------
        # 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
        # ------------------------------------------------------
357
358

        if is_grad_enabled:
359
360
361
            if save_original_input:
                inputmat = inp

362
            ctx.weight_quantizer = weight_quantizer
363
364
365
366
367

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

368
369
370
371
372
373
            # Discard unneeded data in input tensor
            if (
                backward_needs_input
                and own_quantized_input
                and isinstance(inputmat, QuantizedTensorBase)
            ):
374
375
376
                if (
                    ctx.backward_input_needs_gather
                    and weight_quantizer.supports_only_rowwise_all_gather()
377
378
379
380
381
382
383
384
385
                ):
                    # All-gather is not supported with FP8 column-wise data
                    inputmat.update_usage(rowwise_usage=True, columnwise_usage=False)
                else:
                    # Discard row-wise data since it is not needed in backward pass
                    inputmat.update_usage(rowwise_usage=False, columnwise_usage=True)

            # Cached input tensor
            saved_inputmat = None
386
387
            if backward_needs_input:
                saved_inputmat = inputmat
388

389
390
            # Weight with column-wise usage is needed for dgrad GEMM.
            if inp.requires_grad:
391
                if isinstance(weightmat, QuantizedTensorBase):
392
393
                    weightmat.update_usage(columnwise_usage=True)

394
395
            if cpu_offloading and saved_inputmat is not None:
                mark_activation_offload(saved_inputmat)
396

397
398
            # Scatter intermediate/activation tensors saved for the backward pass
            # NOTE: FSDP sharding is not valid for models initialized with primary Fp8 weights
399
            nvtx_range_push(f"{nvtx_label}.fsdp_scatter")
400
401
402
            ctx.fsdp_group = fsdp_group
            ctx.fsdp_shapes = _fsdp_scatter_tensors(
                fsdp_group,
403
                saved_inputmat,
404
                weightmat if fp8 and not isinstance(weight, QuantizedTensorBase) else None,
405
            )
406
            nvtx_range_pop(f"{nvtx_label}.fsdp_scatter")
407

408
            ctx.fine_grained_activation_offloading = fine_grained_activation_offloading
dongcl's avatar
dongcl committed
409

410
            if fine_grained_activation_offloading and cpu_offloading:
dongcl's avatar
dongcl committed
411
                raise ValueError(
412
                    f"Do not use fine_grained_activation_offloading and cpu_offloading at the same time."
dongcl's avatar
dongcl committed
413
414
                )

415
416
417
418
419
420
421
            if (
                fine_grained_activation_offloading
                and weight.requires_grad
                and fuse_wgrad_accumulation
            ):
                if hasattr(weight, "grad_added_to_main_grad"):
                    ctx.has_grad_added_to_main_grad = True
dongcl's avatar
dongcl committed
422
423
                    ctx.grad_added_to_main_grad = weight.grad_added_to_main_grad
                    weight.grad_added_to_main_grad = True
424
425
426
                    ctx.weight_object = weight
                else:
                    ctx.has_grad_added_to_main_grad = False
dongcl's avatar
dongcl committed
427

evt_fugx1's avatar
evt_fugx1 committed
428
            if cpu_offloading or int(os.getenv("NVTE_SWAP_OVERLAP_GRAD", "0")):
dongcl's avatar
dongcl committed
429
                ctx.has_grad_added_to_main_grad = hasattr(weight, "grad_added_to_main_grad")
430

dongcl's avatar
dongcl committed
431
                if ctx.has_grad_added_to_main_grad:
432
433
434
435
436
437
438
                    # 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

439
440
441
442
443
444
            # Do not offload weights and biases
            weight.offloading_activation = False
            weightmat.offloading_activation = False
            if bias is not None:
                bias.offloading_activation = False

445
446
            # TODO(ksivamani): Check memory usage
            tensors_to_save, tensor_objects = prepare_for_saving(
447
                saved_inputmat,
448
                weightmat,
449
                weight,
450
                bias,
451
            )
452
453
            ctx.save_for_backward(*tensors_to_save)
            ctx.tensor_objects = tensor_objects
454

455
456
            ctx.activation_dtype = activation_dtype
            ctx.fp8 = fp8
457
            ctx.fp8_recipe = FP8GlobalStateManager.get_fp8_recipe() if fp8 else None
458
459
            ctx.input_quantizer = input_quantizer
            ctx.grad_input_quantizer = grad_input_quantizer
460
461
            ctx.grad_weight_quantizer = grad_weight_quantizer
            ctx.grad_output_quantizer = grad_output_quantizer
462
            ctx.fuse_wgrad_accumulation = fuse_wgrad_accumulation
463
            if fuse_wgrad_accumulation and weight.requires_grad:
464
465
466
467
468
469
470
471
                # This check is needed to ensure that main_grad is not created
                # during the forward pass when using MCore FSDP as it creates
                # the main_grad buffer lazily before backprop
                if hasattr(weight, "__fsdp_param__"):
                    # MCore FSDP creates main_grad lazily before backward
                    ctx.main_grad_func = weight.get_main_grad
                else:
                    ctx.main_grad_func = lambda: weight.main_grad
472

473
            ctx.debug = debug
474
            ctx.cpu_offloading = cpu_offloading
475
            ctx.is_first_microbatch = is_first_microbatch
476
            ctx.use_bias = bias is not None
477
478
            ctx.sequence_parallel = sequence_parallel
            ctx.tensor_parallel = tensor_parallel
479
            ctx.inp_shape = inp.shape
480
481
            ctx.parallel_mode = parallel_mode
            ctx.tp_group = tp_group
482
483
484
485
            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
486
            ctx.ub_name = ub_name
487
488
            ctx.tp_size = tp_size
            ctx.requires_dgrad = inp.requires_grad
489
            ctx.requires_wgrad = weight.requires_grad
490
            ctx.reduce_and_update_bwd_fp8_tensors = False
491

492
            ctx.owns_input = saved_inputmat is not inp
493
            if ctx.fp8 and requires_grad(inp, weight, bias):
494
495
496
497
                _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
498
            ctx.wgrad_store = wgrad_store
499

500
501
502
        # ------------------------------------------------------
        # Cached state for backward pass is ready...
        # ------------------------------------------------------
503

504
        return out
505
506

    @staticmethod
507
    def backward(ctx, grad_output: torch.Tensor) -> Tuple[Union[torch.Tensor, None], ...]:
508
        # pylint: disable=missing-function-docstring
509

510
511
512
513
514
        # 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}"

515
        with torch.cuda.nvtx.range("_Linear_backward"):
516
517
518
519
            saved_tensors = ctx.saved_tensors
            inputmat, weight_fp8, weight, bias = (  # pylint: disable=unbalanced-tuple-unpacking
                restore_from_saved(ctx.tensor_objects, saved_tensors)
            )
520

521
522
523
            # 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
524
525
526

            # Since main_grad can be modified inplace, it should not be a part of saved_tensors
            main_grad = (
527
                ctx.main_grad_func()
528
529
530
531
                if weight is not None and ctx.fuse_wgrad_accumulation and ctx.requires_wgrad
                else None
            )

532
            if ctx.cpu_offloading or ctx.fine_grained_activation_offloading or int(os.getenv("NVTE_SWAP_OVERLAP_GRAD", "0")):
dongcl's avatar
dongcl committed
533
                if ctx.has_grad_added_to_main_grad:
534
                    weight = ctx.weight_object
535
                    if ctx.fine_grained_activation_offloading:
dongcl's avatar
dongcl committed
536
                        weight.grad_added_to_main_grad = ctx.grad_added_to_main_grad
537
538
                if ctx.requires_wgrad and ctx.fuse_wgrad_accumulation:
                    weight.main_grad = main_grad
539

540
541
542
            # 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
543
            nvtx_range_push(f"{nvtx_label}.fsdp_gather")
544
545
546
547
            _fsdp_gather_tensors(
                ctx.fsdp_group,
                ctx.fsdp_shapes,
                inputmat,
548
                weight_fp8,
549
            )
550
            nvtx_range_pop(f"{nvtx_label}.fsdp_gather")
551

552
            # Configure Userbuffers communication (comm+GEMM overlap)
553
            ctx.ub_obj_gradout = None
554
            ub_obj_dgrad = None
555
            ub_obj_wgrad = None
556
557
            ub_type_dgrad = None
            ub_type_wgrad = None
558
            dgrad_shape = [reduce(multiply_op, ctx.inp_shape[:-1]), ctx.inp_shape[-1]]
559
            if ctx.ub_overlap_ag:
560
                # Overlap grad_output all-gather with dgrad compute
561
                ctx.ub_obj_gradout = get_ub(ctx.ub_name + "_dgrad", ctx.fp8)
562
563
                ub_obj_dgrad = ctx.ub_obj_gradout
                ub_type_dgrad = tex.CommOverlapType.AG
564
565
            elif ctx.ub_overlap_rs_dgrad:
                # Overlap dgrad reduce-scatter with dgrad compute
566
                ctx.ub_obj_gradout = get_ub(ctx.ub_name + "_dgrad", ctx.fp8)
567
568
                ub_obj_dgrad = ctx.ub_obj_gradout
                ub_type_dgrad = tex.CommOverlapType.RS
569
570
571
            else:
                if ctx.ub_bulk_dgrad:
                    # Overlap inputmat all-gather with dgrad compute
572
                    ctx.ub_obj_gradout = get_ub(ctx.ub_name + "_dgrad", ctx.fp8)
573
574
                    ub_obj_dgrad = ctx.ub_obj_gradout
                    ub_type_dgrad = tex.CommOverlapType.AG
575
576
                if ctx.ub_bulk_wgrad:
                    # Overlap dgrad reduce-scatter with wgrad compute
577
                    ub_obj_wgrad = get_ub(ctx.ub_name + "_wgrad", ctx.fp8)
578
                    ub_type_wgrad = tex.CommOverlapType.RS
579
580
581
582
583
584
585
586

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

            # Unmodified grad output tensor
            grad_output_arg = grad_output
587

588
589
590
591
            # 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:
592
593
594
595
596
597
598
                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)
599

600
601
602
603
604
605
606
607
608
609
610
611
612
            # Adjust the quantization direction approach depending
            # on whether wgrad calculations will be performed.
            # NOTE: If requires_dgrad is False, disabling `rowwise` quantization and keeping `columnwise` quantization
            #       results in `Assertion failed: output_tensor->has_data(). Quantizing in only the columnwise direction not supported yet!`
            # NOTE: For `ctx.bias is True`, selected quantize kernel errors with
            #       `cast_kernels.cuh:1322 in function fp8_quantize_arch_l_100: Not implemented scaling mode or fusion: NVTE_DELAYED_TENSOR_SCALING or IS_DBIAS=true on GPU with compute capability < 10.0.`
            if (
                not ctx.use_bias
                and not ctx.requires_wgrad
                and ctx.grad_output_quantizer is not None
            ):
                ctx.grad_output_quantizer.set_usage(columnwise=False)

613
            # Prepare grad output tensor
614
            nvtx_range_push(f"{nvtx_label}.grad_output_preprocess")
615
616
617
618
            (
                grad_output,
                grad_bias,
            ) = TransformerEngineBaseModule.grad_output_preprocess(
619
620
621
622
                ctx,
                grad_output,
                ctx.parallel_mode == "row",
                ctx.grad_output_quantizer,
623
            )
624
            nvtx_range_pop(f"{nvtx_label}.grad_output_preprocess")
625

626
627
628
629
630
631
632
633
634
635
            # --------------------------------------------------
            # 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.
            # --------------------------------------------------
636
            inputmat_total = None
637
            inputmat_total_work = None
638
639
            if ctx.requires_wgrad:
                if ctx.fp8 or ctx.debug:
640
641
642
643
644
645
646
647
                    if isinstance(inputmat, QuantizedTensorBase):
                        # Input tensor is already quantized
                        pass
                    elif ctx.debug:
                        # Debug quantizer will be applied immediately before wgrad GEMM
                        pass
                    else:
                        # Quantize input tensor
648
                        quantizer = ctx.input_quantizer
649
                        if quantizer.supports_only_rowwise_all_gather():
650
                            # All-gather is not supported with FP8 column-wise data
651
652
653
654
655
656
657
658
                            quantizer.set_usage(
                                rowwise=True,
                                columnwise=not ctx.backward_input_needs_gather,
                            )
                        else:
                            quantizer.set_usage(rowwise=False, columnwise=True)
                        inputmat = quantizer(inputmat)
                else:
659
                    if isinstance(inputmat, QuantizedTensorBase):
660
661
662
                        inputmat = inputmat.dequantize(dtype=ctx.activation_dtype)
                    else:
                        inputmat = cast_if_needed(inputmat, ctx.activation_dtype)
663
            if ctx.backward_input_needs_gather:
664
                quantizer = None
665
                if ctx.fp8 or ctx.debug:
666
                    quantizer = ctx.input_quantizer
667
                    if quantizer.supports_only_rowwise_all_gather():
668
669
670
671
672
                        # 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)
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
                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")
689
690
            else:
                inputmat_total = inputmat
691
692
693
            # --------------------------------------------------
            # Input tensor is ready for computing grad weight...
            # --------------------------------------------------
694

695
            # --------------------------------------------------
696
            # Compute grad input tensor
697
698
            # --------------------------------------------------

699
700
            dgrad = None
            dgrad_work = None
701
            if ctx.requires_dgrad:
702

703
704
705
706
707
708
709
710
                # 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
711
712
713
                if ctx.fp8:
                    recipe = ctx.fp8_recipe
                    if hasattr(recipe, "fp8_gemm_dgrad"):
714
                        use_split_accumulator = recipe.fp8_gemm_dgrad.use_split_accumulator
715

716
717
718
719
720
721
722
723
724
725
                # 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
726
                    )
727
728
                elif ctx.ub_bulk_wgrad:
                    gemm_out = ub_obj_wgrad.get_buffer(local_chunk=False)
729

730
731
732
733
                # dgrad GEMM
                # Note: dx = dy * w
                nvtx_range_push(f"{nvtx_label}.dgrad_gemm")
                gemm_out, *_, reduce_scatter_out = general_gemm(
734
735
736
737
738
739
                    weight_fp8,
                    grad_output,
                    get_workspace(),
                    layout="NN",
                    grad=True,
                    quantization_params=ctx.grad_input_quantizer,
740
                    out=gemm_out,
741
                    out_dtype=ctx.activation_dtype,
742
                    use_split_accumulator=use_split_accumulator,
743
744
                    ub=ub_obj_dgrad,
                    ub_type=ub_type_dgrad,
745
                    extra_output=reduce_scatter_out,
746
747
                    bulk_overlap=ctx.ub_bulk_dgrad,
                )
748
                nvtx_range_pop(f"{nvtx_label}.dgrad_gemm")
749

750
751
                # Prepare grad input tensor
                # Note: Perform tensor-parallel communication
752
                if ctx.ub_overlap_rs_dgrad:
753
754
755
756
                    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:
757
                    nvtx_range_push(f"{nvtx_label}.column_parallel_comm_dgrad")
758
                    dgrad = gemm_out
759
760
761
762
763
                    if ctx.sequence_parallel:
                        dgrad, dgrad_work = reduce_scatter_along_first_dim(
                            dgrad,
                            ctx.tp_group,
                            async_op=True,
764
                        )
765
                    else:
766
                        dgrad, dgrad_work = allreduce(dgrad, ctx.tp_group, async_op=True)
767
                    nvtx_range_pop(f"{nvtx_label}.column_parallel_comm_dgrad")
768
769
770
771
772
773
774
775
776
777
                else:
                    dgrad = gemm_out

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

            # --------------------------------------------------
            # Compute grad weight
            # --------------------------------------------------
778

779
780
            wgrad = None
            if ctx.requires_wgrad:
781

782
783
784
                # Prepare input tensor
                # Note: Synchronize tensor-parallel communication and
                # make sure required data is available
785
786
787
                if inputmat_total_work is not None:
                    inputmat_total_work.wait()
                    inputmat_total_work = None
788
789
790
791
792
793
794
795
796
797
798
                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):
799
                    # UB does not support pipelined overlapping grad output
800
801
802
                    # 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
803
                    # for the dgrad GEMM. We work around by explicitly
804
805
806
807
808
809
                    # overlapping the AG operation with the dgrad GEMM.

                    # Get the communication stream from the dgrad GEMM to use for the AG
                    dgrad_send_stream, dgrad_recv_stream = ub_obj_dgrad.get_communication_stream()

                    # This object is separate from the ub_obj_wgrad object which is passed to the GEMM
810
                    ub_obj_overlap_wgrad = get_ub(ctx.ub_name + "_wgrad", ctx.fp8)
811

812
                    ctx.grad_output_quantizer.set_usage(rowwise=False, columnwise=True)
813
814
815
816
817
818
819

                    # We use the send stream to copy into the userbuffers.
                    # This is the same stream that we will use to access the data in the AG,
                    # so we dont need to add any syncs yet.
                    with torch.cuda.stream(dgrad_send_stream):
                        grad_output, _ = fill_userbuffers_buffer_for_all_gather(
                            ub_obj_overlap_wgrad,
820
                            grad_output_arg,
821
                            ctx.grad_output_quantizer,
822
823
                            ctx.tp_group,
                        )
824
825
826
827
828

                    # Allgather grad_outputs[0] using the dgrad streams so we can overlap with the fc2_dgrad gemm
                    tex.bulk_overlap_ag_with_external_gemm(
                        ub_obj_overlap_wgrad, dgrad_send_stream, dgrad_recv_stream
                    )
829

830
831
832
833
834
835
                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)
836

837
838
839
840
841
842
843
                # 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

844
845
846
847
848
849
850
851
852
                # 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
853
                # reduce-scatter with wgrad GEMM
854
                reduce_scatter_out = None
855
                if ctx.ub_bulk_wgrad and ub_obj_wgrad.is_fp8_ubuf():
856
857
                    reduce_scatter_out = torch.empty(
                        dgrad_shape, dtype=ctx.activation_dtype, device=grad_output_arg.device
858
859
                    )

860
861
862
863
                # Arguments to include in wgrad GEMM closure
                wgrad_gemm_kwargs = {
                    "workspace": get_workspace(),
                    "out_dtype": (
864
865
                        main_grad.dtype if ctx.fuse_wgrad_accumulation else ctx.activation_dtype
                    ),
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
                    "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
897
                if ctx.wgrad_store is not None and ctx.wgrad_store.delay_wgrad_compute():
898
899
900
901
902
903
904
905
906
907
908
                    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)
909
910
911
                    # overlap_grad_reduce, dongcl
                    if int(os.getenv("NVTE_OVERLAP_GRAD_REDUCE", "0")):
                        wgrad = torch.empty(weight.size(), dtype=ctx.activation_dtype, device=weight.device)
912
913
                else:

914
915
916
917
                    # Call wgrad GEMM now
                    wgrad, grad_bias_ = wgrad_gemm(inputmat_total, grad_output)

                    # Update grad bias if needed
918
919
920
921
                    if grad_bias is None:
                        grad_bias = grad_bias_
                    del grad_bias_

922
                    # Deallocate tensors if permitted
923
                    if ctx.owns_input:
924
                        # Input tensor is internal
925
                        clear_tensor_data(inputmat_total)
926
927
                    elif ctx.backward_input_needs_gather:
                        # Gathered input tensor is internal
928
                        clear_tensor_data(inputmat_total)
929
930
931
                    if ctx.parallel_mode == "row" and ctx.sequence_parallel:
                        # Gathered grad output tensor is internal
                        clear_tensor_data(grad_output)
932

933
                # Update grad input if overlapping reduce-scatter with wgrad GEMM
934
935
                if ctx.ub_bulk_wgrad:
                    if ub_obj_wgrad.is_fp8_ubuf():
936
                        dgrad = reduce_scatter_out
937
                    else:
938
939
940
941
942
                        dgrad = ub_obj_wgrad.get_buffer(local_chunk=True).clone()

            # --------------------------------------------------
            # Grad weight has been computed...
            # --------------------------------------------------
943

944
            # Don't return grad bias if not needed
945
946
947
            if not ctx.use_bias:
                grad_bias = None

948
            # Make sure all tensor-parallel communication is finished
949
950
951
952
953
954
955
956
            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:
957
            # Handle custom DDP from mcore.
958
959
960
961
962
            if (
                ctx.fuse_wgrad_accumulation
                and weight is not None
                and hasattr(weight, "grad_added_to_main_grad")
            ):
963
                weight.grad_added_to_main_grad = True
964
                if getattr(weight, "zero_out_wgrad", False):
965
966
967
968
                    wgrad = get_dummy_wgrad(
                        list(weight.main_grad.shape),
                        weight.dtype,
                        zero=True,
969
                    )
970
                else:
971
972
973
                    wgrad = get_dummy_wgrad(
                        list(weight.main_grad.shape),
                        weight.dtype,
974
                    )
975
976
977
978
            elif ctx.fuse_wgrad_accumulation:
                wgrad = None
        else:
            wgrad = None
979

980
        # Update FP8 scaling factors if needed
981
        if ctx.reduce_and_update_bwd_fp8_tensors and not is_graph_capturing():
982
            nvtx_range_push(f"{nvtx_label}.reduce_and_update_fp8_tensors")
983
            FP8GlobalStateManager.reduce_and_update_fp8_tensors(forward=False)
984
            nvtx_range_pop(f"{nvtx_label}.reduce_and_update_fp8_tensors")
985

986
        # Scatter fp8 weight buffers
987
        if ctx.fp8 and not isinstance(weight, QuantizedTensorBase):
988
            _fsdp_scatter_tensors(ctx.fsdp_group, weight_fp8)
989
        return (
990
            wgrad,
991
992
            dgrad.view(ctx.inp_shape) if ctx.requires_dgrad else None,
            grad_bias,
993
994
995
            None,  # is_first_microbatch
            None,  # fp8
            None,  # fp8_calibration
996
            None,  # wgrad_store
997
998
999
1000
            None,  # input_quantizer
            None,  # weight_quantizer
            None,  # output_quantizer
            None,  # grad_input_quantizer
1001
1002
            None,  # grad_weight_quantizer
            None,  # grad_output_quantizer
1003
1004
1005
1006
1007
1008
1009
1010
1011
            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
1012
1013
1014
1015
1016
1017
            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
1018
            None,  # ub_name
1019
            None,  # fine_grained_activation_offloading
1020
            None,  # fp8_output
1021
            None,  # fsdp_group
1022
1023
            None,  # module
            None,  # skip_fp8_weight_update
1024
            None,  # symmetric_ar_type
1025
            None,  # save_original_input
1026
            None,  # debug
1027
1028
1029
1030
        )


class Linear(TransformerEngineBaseModule):
1031
    """Applies a linear transformation to the incoming data :math:`y = xA^T + b`
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045

    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)`.
1046
    get_rng_state_tracker : Callable, default = `None`
1047
                 used to get the random number generator state tracker for initializing weights.
1048
1049
    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
1050
    parameters_split : Optional[Union[Tuple[str, ...], Dict[str, int]]], default = None
1051
1052
1053
1054
1055
1056
1057
                      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.
1058
    device : Union[torch.device, str], default = "cuda"
1059
          The device on which the parameters of the model will be allocated. It is the user's
1060
1061
          responsibility to ensure all parameters are moved to the GPU before running the
          forward pass.
1062
1063
    name: str, default = `None`
        name of the module, currently used for debugging purposes.
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076

    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.
1077
    parallel_mode : {None, 'column', 'row'}, default = `None`
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
                   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.
1095
    params_dtype : torch.dtype, default = `torch.get_default_dtype()`
1096
1097
1098
                  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.
1099
1100
1101
1102
    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.
1103
1104
1105
1106
1107
    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.
1108
1109
1110
1111
1112
    save_original_input : bool, default = `False`
                       If set to `True`, always saves the original input tensor rather than the
                       cast tensor. In some scenarios, the input tensor is used by multiple modules,
                       and saving the original input tensor may reduce the memory usage.
                       Cannot work with FP8 DelayedScaling recipe.
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
    """

    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,
1124
        rng_tracker_name: Optional[str] = None,
1125
1126
1127
        init_method: Optional[Callable] = None,
        bias: bool = True,
        return_bias: bool = False,
1128
        params_dtype: Optional[torch.dtype] = None,
1129
        parallel_mode: Optional[str] = None,
cyanguwa's avatar
cyanguwa committed
1130
        parameters_split: Optional[Union[Tuple[str, ...], Dict[str, int]]] = None,
1131
        device: Union[torch.device, str] = "cuda",
1132
        ub_overlap_ag: bool = False,
1133
        ub_overlap_rs: bool = False,
1134
        ub_overlap_rs_dgrad: bool = False,
1135
1136
        ub_bulk_dgrad: bool = False,
        ub_bulk_wgrad: bool = False,
1137
        ub_name: Optional[str] = None,
1138
        delay_wgrad_compute: bool = False,
1139
        symmetric_ar_type: Optional[str] = None,
1140
        save_original_input: bool = False,
1141
        name: Optional[str] = None,
1142
        fine_grained_activation_offloading: bool = False,
1143
1144
    ) -> None:
        super().__init__()
1145
1146

        params_dtype = torch.get_default_dtype() if params_dtype is None else params_dtype
1147
1148
1149
1150
1151
1152
        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
1153
1154
        self.get_rng_state_tracker = get_rng_state_tracker
        self.rng_tracker_name = rng_tracker_name
1155
        self.symmetric_ar_type = symmetric_ar_type
1156
        self.save_original_input = save_original_input
1157
        self.name = name
1158
        self.fine_grained_activation_offloading = fine_grained_activation_offloading
1159

1160
1161
        self.wgrad_store = WeightGradStore(delay_wgrad_compute, ub_bulk_wgrad)

1162
1163
        if device == "meta":
            assert parameters_split is None, "Cannot split module parameters on 'meta' device."
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
        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

1185
        # Column parallel TP overlap options
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
        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
        )
1204
1205

        # Row parallel TP overlap options
1206
1207
1208
1209
1210
1211
        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
        )
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225

        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

1226
1227
1228
1229
1230
1231
1232
        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"

1233
1234
1235
        # Initialize params in FP8
        with_fp8_params = FP8GlobalStateManager.with_fp8_parameters()

1236
1237
1238
1239
1240
1241
1242
1243
        # Contiguous buffers for params
        weight_tensor = torch.empty(
            self.out_features,
            self.in_features,
            device=device,
            dtype=params_dtype,
        )
        bias_tensor = None
1244
        if self.use_bias:
1245
1246
1247
1248
1249
            bias_tensor = torch.empty(
                self.out_features,
                device=device,
                dtype=params_dtype,
            )
1250

1251
1252
1253
1254
        # Configure parameter splits
        self.weight_names = []
        self.bias_names = []
        self.parameter_split_sizes = []
1255
        if parameters_split is None:
1256
1257
1258
1259
1260
1261
            # 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
1262
        elif isinstance(parameters_split, dict):
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
            # 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
1275
        else:
1276
            raise TypeError("Invalid configuration for parameters split")
1277

1278
1279
1280
1281
1282
1283
        # 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}"
            )
1284

1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
        # 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

1295
1296
1297
1298
1299
        # 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.
1300
1301
1302
1303
1304
1305
1306
1307
        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)
1308
            if is_subview and with_fp8_params:
1309
1310
1311
                raise RuntimeError(
                    "Splitting QuantizedTensor into multiple params is not supported"
                )
1312

1313
            # Construct weight parameter
1314
1315
1316
1317
1318
1319
1320
            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,
            )
1321

1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
        # 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
1338

1339
        if with_fp8_params:
1340
1341
            self.init_fp8_metadata()

1342
        self.reset_parameters(defer_init=device == "meta")
1343

1344
1345
1346
1347
1348
1349
1350
        # 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

1351
1352
1353
1354
1355
        if self.wgrad_store.delay_wgrad_compute():
            for name, param in self.named_parameters():
                if name in self.weight_names or name in self.bias_names:
                    param.skip_backward_post_hook = True

1356
1357
1358
1359
1360
1361
1362
1363
    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)
1364
1365
        elif recipe.float8_block_scaling():
            self._customize_quantizers_float8_blockwise_scaling(fwd, recipe)
1366
1367
        # elif for other recipes (mxfp8, etc.)

1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
    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)

1389
    @no_torch_dynamo()
1390
1391
1392
1393
    def forward(
        self,
        inp: torch.Tensor,
        is_first_microbatch: Optional[bool] = None,
1394
        fp8_output: Optional[bool] = False,
1395
        fp8_grad: Optional[bool] = False,
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
    ) -> 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)
        """
1418
1419
1420
        if is_in_onnx_export_mode():
            return self.onnx_forward(inp, fp8_output)

1421
        debug = self.is_debug_iter()
1422

1423
1424
1425
1426
        if FP8GlobalStateManager.fp8_graph_capturing():
            skip_fp8_weight_update = FP8GlobalStateManager.get_skip_fp8_weight_update_tensor()
        else:
            skip_fp8_weight_update = None
1427
1428
1429
        if skip_fp8_weight_update is not None:
            is_first_microbatch = False

1430
        if self.ub_overlap_rs_fprop:
1431
1432
1433
            if get_ub(
                self.ub_name + "_fprop", FP8GlobalStateManager.is_fp8_enabled()
            ).is_fp8_ubuf():
1434
1435
                fp8_output = True
        if self.ub_overlap_rs_dgrad:
1436
1437
1438
            if get_ub(
                self.ub_name + "_dgrad", FP8GlobalStateManager.is_fp8_enabled()
            ).is_fp8_ubuf():
1439
1440
                fp8_grad = True

1441
1442
1443
        with torch.cuda.device(
            getattr(self, list(self.named_parameters())[0][0]).device
        ), self.prepare_forward(
1444
            inp,
1445
            allow_non_contiguous=isinstance(inp, QuantizedTensor),
1446
        ) as inp:
1447

1448
            weight_tensor, bias_tensor = self._get_weight_and_bias_tensors()
1449

1450
1451
1452
1453
1454
            quantizers = (
                self._get_quantizers(fp8_output, fp8_grad)
                if not debug
                else self._get_debug_quantizers(fp8_output, fp8_grad)
            )
1455

1456
            if debug:
1457
                if self.no_debug_features_active(quantizers):
1458
                    debug = False
1459
                    quantizers = self._get_quantizers(fp8_output, fp8_grad)
1460

1461
1462
1463
1464
1465
            (
                input_quantizer,
                weight_quantizer,
                output_quantizer,
                grad_input_quantizer,
1466
1467
1468
                grad_weight_quantizer,
                grad_output_quantizer,
            ) = quantizers
1469

1470
1471
1472
1473
1474
1475
1476
1477
1478
            if torch.is_grad_enabled():
                linear_fn = _Linear.apply
                args = []
            else:
                linear_fn = _Linear.forward
                args = [None]
            args += (
                weight_tensor,
                inp,
1479
                bias_tensor if (self.apply_bias and not self.gemm_bias_unfused_add) else None,
1480
1481
1482
                is_first_microbatch,
                self.fp8,
                self.fp8_calibration,
1483
                self.wgrad_store,
1484
1485
1486
1487
                input_quantizer,
                weight_quantizer,
                output_quantizer,
                grad_input_quantizer,
1488
1489
                grad_weight_quantizer,
                grad_output_quantizer,
1490
                self.fuse_wgrad_accumulation,
1491
                is_cpu_offload_enabled(),
1492
1493
1494
1495
1496
1497
1498
                self.tp_group,
                self.tp_size,
                self.sequence_parallel,
                self.tp_size > 1,
                self.activation_dtype,
                self.parallel_mode,
                torch.is_grad_enabled(),
1499
1500
1501
1502
1503
1504
                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,
1505
                self.ub_name,
1506
                self.fine_grained_activation_offloading,
1507
                fp8_output,
1508
                self.fsdp_group,
1509
1510
                self,
                skip_fp8_weight_update,
1511
                self.symmetric_ar_type,
1512
                self.save_original_input,
1513
                debug,
1514
1515
1516
1517
1518
1519
1520
1521
            )
            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
1522
1523
1524

    def _get_quantizers(self, fp8_output, fp8_grad):
        if not self.fp8:
1525
            return [None] * 6
1526
        grad_input_quantizer = None
1527
        grad_weight_quantizer = None
1528
1529
1530
        grad_output_quantizer = None
        output_quantizer = None
        input_quantizer = self.quantizers["scaling_fwd"][tex.FP8FwdTensors.GEMM1_INPUT]
1531
        input_quantizer.internal = True
1532
        (weight_quantizer,) = self._get_weight_quantizers()
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
        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,
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
            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)
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
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
    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_and_bias_tensors(self) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
        # Get concatenated weight and bias tensors
        unfused_weights = self._get_weight_tensors()
        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]

        weight_tensor = noop_cat(unfused_weights)
        if self.use_bias:
            bias_tensor = noop_cat([getattr(self, name) for name in self.bias_names])
        else:
            bias_tensor = None

        return weight_tensor, bias_tensor

    def onnx_forward(
        self,
        inp: torch.Tensor,
        fp8_output: bool,
    ) -> torch.Tensor:
        """
        ONNX-compatible version of the forward function that provides numerical equivalence
        while only using operations that have defined ONNX symbolic translations.
        This simplified implementation is designed specifically for inference scenarios.
        """
        from ..export import onnx_gemm

        assert_warmed_up(self)
        assert not TEDebugState.debug_enabled, "Debug mode is not supported in ONNX export."
        weight_tensor, bias_tensor = self._get_weight_and_bias_tensors()
        (
            input_quantizer,
            weight_quantizer,
            output_quantizer,
            *_,
        ) = self._get_quantizers(fp8_output, False)
        inp_dtype = inp.dtype

        if input_quantizer is not None:
            inp_q = input_quantizer.onnx_quantize(inp)
            inp = input_quantizer.onnx_dequantize(inp_q)
            inp = inp.to(inp_dtype)

        if weight_quantizer is not None:
            weight_q = weight_quantizer.onnx_quantize(weight_tensor)
            weight_tensor = weight_quantizer.onnx_dequantize(weight_q)
        if bias_tensor is not None:
            bias_tensor = bias_tensor.to(inp_dtype)
        weight_tensor = weight_tensor.to(inp_dtype)

        if self.apply_bias:
            output = onnx_gemm(weight_tensor, inp, bias_tensor)
        else:
            output = onnx_gemm(weight_tensor, inp, None)

        if output_quantizer is not None:
            raise NotImplementedError("ONNX export of quantized output is not supported")

        if self.return_bias:
            return output, bias_tensor

        return output

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
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
    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
1695
1696
1697

    def _get_weight_quantizers(self) -> List[Quantizer]:
        """Get the weight quantizers of the module."""
1698
        if not self.fp8 and not self.fp8_calibration:
1699
1700
1701
1702
            return [None]
        weight_quantizer = self.quantizers["scaling_fwd"][tex.FP8FwdTensors.GEMM1_WEIGHT]
        weight_quantizer.internal = True
        return [weight_quantizer]
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721

    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