"...wan2.1-t2v-14b.git" did not exist on "873911fa18763d4e3813a9ac7386958498932d82"
base.py 69.6 KB
Newer Older
1
# Copyright (c) 2022-2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
2
3
4
5
#
# See LICENSE for license information.

"""Base modules and utilities for TransformerEngine PyTorch API"""
6
import io
7
import math
8
9
10
import os
import pickle
import warnings
11
from enum import Enum
12
from abc import ABC, abstractmethod
13
from typing import Any, Dict, Generator, List, Optional, Set, Tuple, Union
14
from contextlib import contextmanager
15
import logging
16
from types import MethodType
17
18
19

import torch
import torch.nn.functional as F
20
from torch.distributed.tensor import DTensor
21

22
import transformer_engine_torch as tex
23

24
from ._common import _ParameterInitMeta, noop_cat
25
from ..quantization import (
26
27
    MXFP8BlockScalingRecipeState,
    DelayedScalingRecipeState,
28
    Float8CurrentScalingRecipeState,
29
    Float8BlockScalingRecipeState,
30
    NVFP4BlockScalingRecipeState,
31
    FP8GlobalStateManager,
32
    RecipeState,
33
34
35
36
37
)
from ..distributed import (
    gather_along_first_dim,
    is_fp8_activation_recompute_enabled,
    in_fp8_activation_recompute_phase,
38
    _fsdp_gather_tensors,
39
40
)
from ..constants import dist_group_type
41
from ..cpp_extensions.gemm import _NUM_MAX_UB_STREAMS
42
from ..quantized_tensor import QuantizedTensor, QuantizedTensorStorage, Quantizer
43
44
from ..tensor.float8_tensor import Float8Quantizer, Float8CurrentScalingQuantizer
from ..tensor.mxfp8_tensor import MXFP8Quantizer
45
from ..tensor.float8_blockwise_tensor import Float8BlockQuantizer
46
47
from ..tensor.storage.float8_tensor_storage import Float8TensorStorage
from ..tensor.storage.mxfp8_tensor_storage import MXFP8TensorStorage
48
49
50
51
52
from ..utils import (
    is_non_tn_fp8_gemm_supported,
    torch_get_autocast_gpu_dtype,
    get_nvtx_range_context,
)
53
from ..tensor.storage.float8_blockwise_tensor_storage import Float8BlockwiseQTensorStorage
54
from ...common.recipe import DelayedScaling, Recipe
55
56
from ...debug.pytorch.debug_state import TEDebugState
from ...debug.pytorch.debug_quantization import DebugQuantizer, DebugQuantizedTensor
57
from ...debug.pytorch.utils import next_iter_when_debug_should_be_run, any_feature_enabled
58

59
__all__ = ["initialize_ub", "destroy_ub", "UserBufferQuantizationMode"]
60

61
62
63
_2X_ACC_FPROP = False
_2X_ACC_DGRAD = True
_2X_ACC_WGRAD = True
64
_dummy_wgrads = {}
65
_ub_communicators = None
66
_MIN_STREAM_PRIORITY, _MAX_STREAM_PRIORITY = None, None
67
layers_atomic_ring_exchange = []
68
69


70
71
72
73
74
75
76
77
78
class UserBufferQuantizationMode(Enum):
    """
    UserBufferQuantizationMode is an enum that represents the quantization mode of the UserBuffer.
    """

    NONE = "none"
    FP8 = "fp8"


79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
def get_dummy_wgrad(shape: list, dtype: torch.dtype, zero=False) -> torch.Tensor:
    """Returns a dummy tensor of given shape."""
    assert len(shape) == 2
    global _dummy_wgrads
    if (shape[0], shape[1], dtype) not in _dummy_wgrads:
        _dummy_wgrads[(shape[0], shape[1], dtype)] = torch.empty(
            shape,
            dtype=dtype,
            device="cuda",
            requires_grad=False,
        )
    if zero:
        _dummy_wgrads[(shape[0], shape[1], dtype)].fill_(0)
    return _dummy_wgrads[(shape[0], shape[1], dtype)].detach()


95
96
def initialize_ub(
    shape: list,
97
    tp_size: int,
98
    use_fp8: bool = False,
99
    quantization_modes: List[UserBufferQuantizationMode] = None,
100
    dtype: torch.dtype = torch.bfloat16,
101
    ub_cfgs: Optional[Union[dict, List[dict]]] = None,
102
    bootstrap_backend: Union[str, torch.distributed.Backend] = None,
103
) -> None:
104
105
    r"""
    Initialize the Userbuffers communicator for overlapping tensor-parallel communications with
Paweł Gadziński's avatar
Paweł Gadziński committed
106
    GEMM compute in ``te.Linear``, ``te.LayerNormLinear`` and ``te.LayerNormMLP`` modules.
107
108
109
110
111

    Parameters
    ----------
    shape : list
            shape of the communication buffer, typically set to be the same as the global shape of
Paweł Gadziński's avatar
Paweł Gadziński committed
112
113
            the input tensor to a ``te.TransformerLayer`` forward pass, with the sequence and batch
            dimensions collapsed together -- i.e.: ``(sequence_length * batch_size, hidden_size)``
114
115
116
    tp_size : int
              number of GPUs in the tensor-parallel process group
    use_fp8 : bool = False
117
              allocate the communication buffer for FP8 GEMM inputs/outputs.
Paweł Gadziński's avatar
Paweł Gadziński committed
118
              DEPRECATED: Please use ``quantization_modes`` instead.
119
120
    quantization_modes : List[UserBufferQuantizationMode] = None
              if a list of UserBufferQuantizationMode is provided, a UB communicator is created for each quantization setting in the list.
Paweł Gadziński's avatar
Paweł Gadziński committed
121
              falls back to the legacy ``use_fp8`` parameter if ``None`` is provided.
122
    dtype : torch.dtype = torch.bfloat16
Paweł Gadziński's avatar
Paweł Gadziński committed
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
            non-FP8 data type of the communication buffer when ``use_fp8 = False``
    ub_cfgs : dict = None
             Configuration dictionary with the structure::

                 {
                    <gemm_name> : {
                        "method": <"ring_exchange" or "pipeline">,
                        "is_reduce_scatter": bool,
                        "num_sm": int,
                        "cga_size": int,
                        "set_sm_margin": bool,
                        "num_splits": int,
                        "aggregate": bool,
                        "atomic_gemm": bool,
                        "use_ce": bool,
                        "fp8_buf": bool,
                    }
                 }

             for ``te.TransformerLayer`` GEMM layers in ``["qkv_fprop", "qkv_dgrad", "qkv_wgrad",
143
             "proj_fprop", "proj_dgrad", "proj_wgrad", "fc1_fprop", "fc1_dgrad", "fc2_dgrad",
Paweł Gadziński's avatar
Paweł Gadziński committed
144
145
             "fc2_fprop", "fc2_wgrad"]``.
             a list may be provided to specify different overlap configurations for different the quantization settings in ``quantization_modes``
146
    bootstrap_backend : str = None
Paweł Gadziński's avatar
Paweł Gadziński committed
147
                        ``torch.distributed`` communication backend for the all-gather, broadcast and
148
149
150
151
                        barrier collectives during Userbuffers initialization. Not all backends are
                        valid for every cluster configuration and distributed launch method even if
                        they are available in PyTorch. When left unset, the initialization prefers
                        to use the MPI backend, falling back first on Gloo and then NCCL if MPI is
Paweł Gadziński's avatar
Paweł Gadziński committed
152
                        not available. Setting ``NVTE_UB_WITH_MPI=1`` when building TE overrides this
153
                        option and always initializes Userbuffers with direct MPI calls in C++,
Paweł Gadziński's avatar
Paweł Gadziński committed
154
                        which also requires ``MPI_HOME=/path/to/mpi/root`` to be set at compile time.
155
    """
156
    if not tex.device_supports_multicast():
157
        assert bool(int(os.getenv("UB_SKIPMC", "0"))), (
158
159
160
161
            "CUDA device, driver and/or toolkit version does not support comm+GEMM overlap with "
            + "CUDA Multicast. Launch app with UB_SKIPMC=1 to try CUDA IPC instead."
        )

162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
    if not quantization_modes:
        warnings.warn(
            "Initializing Userbuffers with use_fp8 is deprecated. Please use quantization_modes"
            " instead.",
            DeprecationWarning,
        )
        quantization_modes = [
            UserBufferQuantizationMode.FP8 if use_fp8 else UserBufferQuantizationMode.NONE
        ]
    else:
        assert isinstance(quantization_modes, list), "quantization_modes must be a list"
        assert all(
            isinstance(mode, UserBufferQuantizationMode) for mode in quantization_modes
        ), "quantization_modes must be a list of UserBufferQuantizationMode"

    if isinstance(ub_cfgs, dict) or ub_cfgs is None:
        ub_cfgs = [ub_cfgs] * len(quantization_modes)
    else:
        assert len(ub_cfgs) == len(
            quantization_modes
        ), "Number of ub_cfgs settings must match number of quantization configurations"

184
185
186
    global _ub_communicators
    assert _ub_communicators is None, "UB communicators are already initialized."
    _ub_communicators = {}
187
188

    if tex.ubuf_built_with_mpi():
189
190
        # We're bootstrapping with direct calls to MPI in Userbuffers code so we need to force
        # an MPI_Init() here by creating a new MPI process group...
191
        assert torch.distributed.is_mpi_available()
192
193
        _ = torch.distributed.new_group(backend="mpi")
        helper = tex.CommOverlapHelper()
194
    else:
195
196
        # Bootstrapping with torch.distributed API, so check backend and construct
        # intra/inter-node process groups...
197
198
199
200
201
        assert (
            torch.distributed.is_initialized()
        ), "torch.distributed must be initialized before Userbuffers"
        if bootstrap_backend is None:
            bootstrap_backend = "nccl"
202
            if torch.distributed.is_mpi_available():
203
                bootstrap_backend = "mpi"
204
205
            elif torch.distributed.is_gloo_available():
                bootstrap_backend = "gloo"
206
        else:
207
208
209
210
211
212
213
214
215
            assert bootstrap_backend in [
                "gloo",
                "mpi",
                "nccl",
            ], "Invalid torch.distributed backend for bootstrapping Userbuffers!"
            assert torch.distributed.is_backend_available(bootstrap_backend), (
                f"PyTorch must be compiled with '{bootstrap_backend}' support in order to "
                f"bootstrap Userbuffers with '{bootstrap_backend}' collectives."
            )
216
217
218
219
220

        world_group = torch.distributed.new_group(backend=bootstrap_backend)
        world_rank = torch.distributed.get_rank(world_group)
        world_size = torch.distributed.get_world_size(world_group)

221
222
        num_domains = world_size // tp_size
        mydomain_idx = world_rank // tp_size
223
        if num_domains > 1:
224
225
226
227
            ranks_per_domain_list = [
                [i * tp_size + t for t in range(tp_size)] for i in range(num_domains)
            ]
            tp_domain_group, _ = torch.distributed.new_subgroups_by_enumeration(
228
229
                ranks_per_domain_list, backend=bootstrap_backend
            )
230
231
            local_rank = torch.distributed.get_rank(tp_domain_group)
            tp_domain_ranks = torch.distributed.get_process_group_ranks(tp_domain_group)
232

233
            helper = tex.CommOverlapHelper(world_group, tp_domain_group)
234
        else:
235
236
            # TP model on single NVLink domain, no replication, no data-parallelism
            mydomain_idx = 0
237
            local_rank = world_rank
238
            tp_domain_ranks = list(range(world_size))
239
240

            helper = tex.CommOverlapHelper(world_group)
241

242
        if world_rank == 0:
243
            print(f"!!! [UB] Number of TP domains: {num_domains}\n", end="", flush=True)
244
245
        if local_rank == 0:
            print(
246
                f"!!! [UB] Global ranks on TP domain {mydomain_idx}: {tp_domain_ranks}\n",
247
248
249
250
                end="",
                flush=True,
            )

251
    # Default buffer precision: AllGather buffers use fp8 when using fp8 recipe
252
    layers_all_gather_overlap = [
253
254
255
        "qkv_fprop",
        "qkv_dgrad",
        "proj_dgrad",
256
        "proj_wgrad",
257
258
259
        "fc1_fprop",
        "fc1_dgrad",
        "fc2_dgrad",
260
        "fc2_wgrad",
261
    ]
262
    layers_reduce_scatter_overlap = ["proj_fprop", "fc2_fprop", "qkv_wgrad", "fc1_wgrad"]
Jaemin Choi's avatar
Jaemin Choi committed
263
    dgrad_reduce_scatter_overlap = ["qkv_dgrad", "fc1_dgrad"]
264
265
    # Default overlap methods for layers
    methods = {
266
267
268
269
270
271
        "ring_exchange": [
            "qkv_fprop",
            "fc1_fprop",
            "proj_dgrad",
            "fc2_dgrad",
        ],
272
273
        "pipeline": ["proj_fprop", "fc2_fprop"],
        "bulk": ["qkv_dgrad", "qkv_wgrad", "fc1_dgrad", "fc1_wgrad"],
274
        "external": ["proj_wgrad", "fc2_wgrad"],
275
276
    }

277
    # AG-RS overlap pairs of layers forming a tensor-parallel block
278
279
    ag_rs_pairs = {"qkv_fprop": "proj_fprop", "fc1_fprop": "fc2_fprop"}
    rs_ag_pairs = {v: k for k, v in ag_rs_pairs.items()}
280
    external_gemm_to_overlap = {"proj_wgrad": "proj_dgrad", "fc2_wgrad": "fc2_dgrad"}
281
282
283
    global layers_atomic_ring_exchange
    layers_atomic_ring_exchange = []

284
285
286
287
288
289
    def get_method(name):
        for method, names in methods.items():
            if name in names:
                return method
        raise KeyError(f"Given layer name {name} does not exist.")

290
    def get_default_config(name):
291
        global _MIN_STREAM_PRIORITY, _MAX_STREAM_PRIORITY
292
293
        method = get_method(name)
        is_reduce_scatter = name in layers_reduce_scatter_overlap
294
295
        if _MIN_STREAM_PRIORITY is None or _MAX_STREAM_PRIORITY is None:
            _MIN_STREAM_PRIORITY, _MAX_STREAM_PRIORITY = tex.get_stream_priority_range()
296
297
298
299
300
        default_cfg = {
            "method": method,
            "is_reduce_scatter": is_reduce_scatter,
            "num_sm": 1 if method == "ring_exchange" else 16,
            "cga_size": 1 if method == "ring_exchange" else 2,
301
302
            "set_sm_margin": not method == "ring_exchange",
            "num_splits": tp_size if method == "ring_exchange" else 4,
303
304
305
306
            "aggregate": False,
            "atomic_gemm": False,
            "use_ce": True,
            "fp8_buf": name in layers_all_gather_overlap,
307
308
309
            "comm_priority": _MAX_STREAM_PRIORITY,
            "gemm_priority": _MIN_STREAM_PRIORITY,
            "pipeline_rs_overlap_first_gemm": False,
310
311
312
        }
        return default_cfg

313
314
    def add_ub(
        name: str,
315
        quantization_mode: UserBufferQuantizationMode,
316
        method: str,
317
        is_reduce_scatter: bool,
318
319
        num_sm: int = 16,
        cga_size: int = 2,
320
        set_sm_margin: bool = False,
321
        num_splits: int = 0,
322
323
        aggregate: bool = False,
        atomic_gemm: bool = False,
324
        use_ce: bool = True,
325
        fp8_buf: bool = False,
326
327
328
        comm_priority: int = 0,
        gemm_priority: int = 0,
        pipeline_rs_overlap_first_gemm: bool = False,
329
    ) -> None:
330
331
332
333
        if atomic_gemm:
            warnings.warn(
                "Atomic GEMM uses a beta API from cublas and is not tested for all use cases."
            )
334
335
336
            assert (
                quantization_mode == UserBufferQuantizationMode.FP8
            ), "Atomic GEMM overlap supported only for FP8 GEMM."
337
            if method in ("bulk", "external"):
338
                warnings.warn(
339
                    f"At {name}, atoimic GEMM not is supported for a bulk overlap."
340
341
342
                    "Defaulting to `atomic_gemm=False`."
                )
                atomic_gemm = 0
343
        if not is_reduce_scatter and method == "pipeline":
344
            raise ValueError(
345
                f"At {name}, `pipeline` overlap method is not supported for AllGather."
346
            )
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
        # Check if both AG and RS overlaps use `atomic GEMM`` + `p2p ring-exchange`.
        # Using atomic GEMM + p2p ring-exchange in only one of the pair breaks functionality.
        global layers_atomic_ring_exchange
        if atomic_gemm and method == "ring_exchange" and name in ag_rs_pairs:
            layers_atomic_ring_exchange += [name, ag_rs_pairs[name]]
        if name in rs_ag_pairs:
            assert_message = (
                f"At {name}, atomic AG-GEMM overlap with `ring_exchange` shuffles GEMM chunk "
                "outputs, and  RS-GEMM overlap un-suffle them. When one of the GEMM-AG and "
                "GEMM-RS overlaps forming a TP block (e.g., qkv_fprop and proj_fprop) uses "
                "`atomic gemm` and `ring_exhcnage`, its pair must use the same overlap config "
                "for functionality."
            )
            if name in layers_atomic_ring_exchange:
                assert atomic_gemm and method == "ring_exchange", assert_message
            else:
                if atomic_gemm and method == "ring_exchange":
                    assert rs_ag_pairs[name] in layers_atomic_ring_exchange, assert_message

366
367
368
369
370
371
372
373
374
375
        if name in external_gemm_to_overlap:
            assert method == "external", (
                f"At {name}, `external` overlap method is specified, but the selected method is"
                f" {method}"
            )
            assert external_gemm_to_overlap[name] in methods["ring_exchange"], (
                f"At {name}, `external` overlap method is specified, but the external gemm"
                f" {external_gemm_to_overlap[name]} is not using `ring_exchange` overlap method"
            )

376
377
378
379
380
        buffer_dtype = (
            torch.uint8
            if (quantization_mode == UserBufferQuantizationMode.FP8 and fp8_buf)
            else dtype
        )
381
        if method == "ring_exchange":
382
383
384
385
            ub_obj = tex.CommOverlapP2P(
                shape,  # Communication buffer shape
                buffer_dtype,  # Communication buffer data type
                helper,  # Helper for torch.distributed callbacks during bootstrapping
386
                tp_size,  # Tensor-parallel group size (may be different than local_size)
387
388
389
390
391
392
393
394
                tex.CommOverlapType.RS if is_reduce_scatter else tex.CommOverlapType.AG,
                num_max_streams=_NUM_MAX_UB_STREAMS,
                comm_cga_size=cga_size,
                num_comm_sm=num_sm,
                set_sm_margin=set_sm_margin,
                atomic_gemm=atomic_gemm,
                use_ce=use_ce,
                aggregate=aggregate,
395
396
                gemm_priority=gemm_priority,
                comm_priority=comm_priority,
397
            )
398
        else:
399
400
401
402
            ub_obj = tex.CommOverlap(
                shape,  # Communication buffer shape
                buffer_dtype,  # Communication buffer data type
                helper,  # Helper for torch.distributed callbacks during bootstrapping
403
                tp_size,  # Tensor-parallel group size (may be different than local_size)
404
405
406
407
408
409
                num_splits=num_splits,
                num_max_streams=_NUM_MAX_UB_STREAMS,
                comm_cga_size=cga_size,
                num_comm_sm=num_sm,
                set_sm_margin=set_sm_margin,
                atomic_gemm=atomic_gemm,
410
411
412
                gemm_priority=gemm_priority,
                comm_priority=comm_priority,
                rs_overlap_first_gemm=pipeline_rs_overlap_first_gemm,
413
            )
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
        _ub_communicators[(name, quantization_mode)] = ub_obj

    for quantization_mode, user_ub_cfg in zip(quantization_modes, ub_cfgs):
        if user_ub_cfg is not None:
            for name in dgrad_reduce_scatter_overlap:
                if (
                    name in user_ub_cfg
                    and "method" in user_ub_cfg[name]
                    and user_ub_cfg[name]["method"] != "bulk"
                ):
                    wgrad_name = name.replace("dgrad", "wgrad")
                    assert wgrad_name not in user_ub_cfg
                    layers_reduce_scatter_overlap.remove(wgrad_name)
                    layers_all_gather_overlap.remove(name)
                    layers_reduce_scatter_overlap.append(name)
                    methods["bulk"].remove(name)
                    new_method = user_ub_cfg[name]["method"]
                    methods[new_method].append(name)

        for name in (
            methods["ring_exchange"] + methods["pipeline"] + methods["bulk"] + methods["external"]
        ):
            ub_cfg = get_default_config(name)
            if user_ub_cfg is not None and name in user_ub_cfg:
                fp8_buf = (name in layers_all_gather_overlap) or (
                    user_ub_cfg[name].get("fp8_buf", False) and name in methods["pipeline"]
                )
441
                ub_cfg.update(user_ub_cfg[name])
442
443
                ub_cfg["fp8_buf"] = fp8_buf
            add_ub(name, quantization_mode, **ub_cfg)
444
445


446
def get_ub(name: str, use_fp8: bool):
447
    """Get userbuffer communicator corresponding to give key."""
448
449
450
451
    # For now use `use_fp8` boolean input as it matches the current design in the modules
    # So favour simplicity until the correct design becomes clear.
    # This is mainly an internal API so we don't need to worry about future changes
    key = (name, UserBufferQuantizationMode.FP8 if use_fp8 else UserBufferQuantizationMode.NONE)
452
    assert _ub_communicators is not None, "UB manager is not initialized."
453
454
    assert key in _ub_communicators, f"UB for {name} with use_fp8={use_fp8} is not registered."
    return _ub_communicators[key]
455

456

457
458
459
460
461
462
463
def destroy_ub():
    """Destroy all allocated userbuffer communicators."""
    global _ub_communicators
    _ub_communicators = None
    global layers_atomic_ring_exchange
    layers_atomic_ring_exchange = []

464

465
466
467
468
469
def fill_userbuffers_buffer_for_all_gather(
    comm,
    local_tensor: torch.Tensor,
    quantizer: Optional[Quantizer],
    process_group,
470
) -> tuple[torch.Tensor | QuantizedTensorStorage, torch.Tensor | QuantizedTensorStorage]:
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
    """Fill local shard of Userbuffers buffer with data for all-gather

    Returns the full tensor and the local shard, both using the
    Userbuffers buffer as their underlying data. These tensors should
    be used carefully (e.g. only immediately before and after a
    Userbuffers operation) since the underlying data may be
    overwritten by other Userbuffers operations.

    May perform blocking communication if needed for the gathered
    tensor's metadata, e.g. scaling factors.

    """

    # Tensor dimensions
    local_shape = local_tensor.size()
    if not local_shape:
        raise ValueError(f"Invalid local tensor (shape={tuple(local_shape)})")
    process_group_size = torch.distributed.get_world_size(process_group)
    global_shape = list(local_shape)
    global_shape[0] *= process_group_size

    # Unquantized data
    if quantizer is None:
494
        if isinstance(local_tensor, QuantizedTensorStorage):
495
496
497
498
499
500
501
502
503
504
505
506
            local_tensor = local_tensor.dequantize()
        if comm.is_fp8_ubuf():
            raise RuntimeError(
                "Attempting to all-gather unquantized tensor, "
                "but Userbuffers is initialized with FP8 buffers"
            )
        comm.copy_into_buffer(local_tensor, local_chunk=True)
        global_tensor = comm.get_buffer(shape=global_shape)
        return global_tensor, local_tensor

    # FP8 data
    if isinstance(quantizer, (Float8Quantizer, Float8CurrentScalingQuantizer)):
507
508
        if not isinstance(local_tensor, Float8TensorStorage):
            if isinstance(local_tensor, QuantizedTensorStorage):
509
510
511
512
513
514
515
516
517
518
                local_tensor.dequantize()
            quantizer.set_usage(rowwise=True, columnwise=False)
            local_tensor = quantizer(local_tensor)
        if not comm.is_fp8_ubuf():
            raise RuntimeError(
                "Attempting to all-gather FP8 tensor, "
                "but Userbuffers is not initialized with FP8 buffers"
            )
        comm.copy_into_buffer(local_tensor._data, local_chunk=True)
        global_tensor_data = comm.get_buffer(shape=global_shape)
519
        global_tensor = Float8TensorStorage(
520
521
522
523
524
525
526
527
528
529
530
            data=global_tensor_data,
            fp8_scale_inv=local_tensor._scale_inv,
            fp8_dtype=local_tensor._fp8_dtype,
            quantizer=quantizer,
        )
        return global_tensor, local_tensor

    # MXFP8 data
    if isinstance(quantizer, MXFP8Quantizer):

        # Cast to MXFP8 if needed
531
532
        if not isinstance(local_tensor, MXFP8TensorStorage):
            if isinstance(local_tensor, QuantizedTensorStorage):
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
                local_tensor.dequantize()
            local_tensor = quantizer(local_tensor)
        if not comm.is_fp8_ubuf():
            raise RuntimeError(
                "Attempting to all-gather MXFP8 tensor, "
                "but Userbuffers is not initialized with FP8 buffers"
            )

        # Check which MXFP8 buffer to communicate
        if quantizer.rowwise_usage == quantizer.columnwise_usage:
            raise ValueError(
                "Userbuffers can only communicate one MXFP8 buffer at a time, "
                f"but quantizer has rowwise_usage={quantizer.rowwise_usage}, "
                f"columnwise_usage={quantizer.columnwise_usage}"
            )
        with_rowwise_data = quantizer.rowwise_usage

        # Copy MXFP8 data to local chunk of Userbuffers buffer
        local_data = (
            local_tensor._rowwise_data if with_rowwise_data else local_tensor._columnwise_data
        )
        comm.copy_into_buffer(local_data, local_chunk=True)

        # Gather scaling-inverses
        if math.prod(local_shape[:-1]) % 128 != 0:
            raise ValueError(
                "Userbuffers requires MXFP8 tensor dims that are divisible by 128, "
                f"but got MXFP8 tensor with shape={tuple(local_shape)}"
            )
        local_scale_inv = (
            local_tensor._rowwise_scale_inv
            if with_rowwise_data
            else local_tensor._columnwise_scale_inv
        )
        local_scale_inv_size = list(local_scale_inv.size())
        global_scale_inv = torch.empty(
            [process_group_size * local_scale_inv_size[0]] + local_scale_inv_size[1:],
            dtype=local_scale_inv.dtype,
            device=local_scale_inv.device,
        )
        torch.distributed.all_gather_into_tensor(
            global_scale_inv,
            local_scale_inv,
            group=process_group,
        )

        # Construct MXFP8 tensor with Userbuffers buffer
        rowwise_data, rowwise_scale_inv = None, None
        columnwise_data, columnwise_scale_inv = None, None
        global_data = comm.get_buffer(shape=global_shape)
        if with_rowwise_data:
            rowwise_data, rowwise_scale_inv = global_data, global_scale_inv
        else:
            columnwise_data, columnwise_scale_inv = global_data, global_scale_inv
587
        global_tensor = MXFP8TensorStorage(
588
589
590
591
592
593
594
595
596
597
598
599
600
            rowwise_data=rowwise_data,
            rowwise_scale_inv=rowwise_scale_inv,
            columnwise_data=columnwise_data,
            columnwise_scale_inv=columnwise_scale_inv,
            fp8_dtype=local_tensor._fp8_dtype,
            quantizer=quantizer,
        )
        return global_tensor, local_tensor

    # Unsupported data format
    raise ValueError(f"Unsupported quantizer for Userbuffers ({quantizer})")


601
602
603
604
605
606
class TransformerEngineBaseModule(torch.nn.Module, ABC):
    """Base TE module."""

    def __init__(self) -> None:
        super().__init__()
        assert torch.cuda.is_available(), "TransformerEngine needs CUDA."
607
        self.name = None
608
        self.next_iter_when_debug_should_be_run = 0
609
610
611
612
        self.fp8_initialized = False
        self.fp8 = False
        self.fp8_calibration = False
        self.fp8_meta = {}
613
        self.fp8_meta["fp8_checkpoint"] = False
614
615
        self.fp8_meta["fp8_group"] = None
        self.fp8_meta_tensors_initialized = False
616
        self.quantizers = {"scaling_fwd": {}, "scaling_bwd": {}}
617
618
619
        self.tp_group = None
        self.tp_size = 1
        self.sequence_parallel = False
620
621
        self.param_init_meta = {}
        self.primary_weights_in_fp8 = FP8GlobalStateManager.with_fp8_parameters()
622
        self.preserve_high_precision_init_val = FP8GlobalStateManager.with_high_precision_init_val()
623
624
        self.fsdp_wrapped = False
        self.fsdp_group = None
625
        self._fp8_workspaces: Dict[str, QuantizedTensor] = {}
626
        self.activation_dtype: Optional[torch.dtype] = None
627
        self.wgrad_accumulation_and_reduce_hooks = []
628
        self.wgrad_store = None
629

630
631
632
        if not TEDebugState.debug_enabled:
            TEDebugState.initialize()

633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
    # Names of attributes that can be set quickly (see __setattr__
    # method)
    _fast_setattr_names: Set[str] = {
        "activation_dtype",
        "fp8",
        "fp8_initialized",
        "fp8_calibration",
        "fp8_parameters",
    }

    def __setattr__(self, name: str, value: Any) -> None:
        if name in TransformerEngineBaseModule._fast_setattr_names:
            # torch.nn.Module has a custom __setattr__ that handles
            # modules, parameters, and buffers. This is unnecessary
            # overhead when setting plain attrs.
            self.__dict__[name] = value
        else:
            # Default case
            super().__setattr__(name, value)
652

653
    def adjust_amax_history_length(self, length: int, fwd: Optional[bool] = None) -> None:
654
655
656
657
        """
        Delayed scaling only.

        Increase or decrease size of amax history based on given `length`.
658
659
660
661
662
663
664
665
666
667

        .. warning::
            This changes the underlying amax memory location.
        """
        if fwd is None:
            fp8_meta_tensor_keys = ("scaling_fwd", "scaling_bwd")
        else:
            fp8_meta_tensor_keys = ("scaling_fwd" if fwd else "scaling_bwd",)

        for meta_key in fp8_meta_tensor_keys:
668
669
670
            if meta_key not in self.fp8_meta:
                # Handles non-parameter FP8 modules, e.g. DPA.
                continue
671
672
673
674
675
            curr_len = self.fp8_meta[meta_key].amax_history.shape[0]
            if length == curr_len:
                continue
            if length < curr_len:
                self.fp8_meta[meta_key].amax_history = (
676
677
                    self.fp8_meta[meta_key].amax_history[:length].clone()
                )
678
679
680
681
682
683
            elif length > curr_len:
                extra_rows = length - curr_len
                self.fp8_meta[meta_key].amax_history = F.pad(
                    self.fp8_meta[meta_key].amax_history, pad=(0, 0, 0, extra_rows)
                )

684
685
            # Update quantizers with new amax pointers.
            self.quantizers[meta_key] = self.fp8_meta[meta_key].make_quantizers()
686
687
            # Make sure weight tensors has correct quantizers
            self._update_weight_quantizers()
688

689
690
            # Update the global buffers with new amax and history pointers.
            if FP8GlobalStateManager.get_buffer_info() in self.fp8_meta:
691
692
693
                fwd_pos, fwd_key, bwd_pos, bwd_key = self.fp8_meta[
                    FP8GlobalStateManager.get_buffer_info()
                ]
694
695
696
697
698
                for pos, buffer_key in zip((fwd_pos, bwd_pos), (fwd_key, bwd_key)):
                    if buffer_key in FP8GlobalStateManager.global_amax_buffer:
                        assert (
                            buffer_key in FP8GlobalStateManager.global_amax_history_buffer
                        ), "TE internal error during amax history change."
699
700
701
                        FP8GlobalStateManager.global_amax_buffer[buffer_key][pos] = self.fp8_meta[
                            meta_key
                        ].amax_history[0]
702
                        FP8GlobalStateManager.global_amax_history_buffer[buffer_key][pos] = (
703
704
                            self.fp8_meta[meta_key].amax_history
                        )
705

706
    def set_meta_tensor(self, fwd: bool, recipe: Recipe) -> None:
707
708
709
        """Init scales and amaxes for fwd | bwd."""
        fp8_meta_tensor_key = "scaling_fwd" if fwd else "scaling_bwd"

710
        # Return early if recipe state matches recipe
711
        if self.fp8_meta_tensors_initialized:
712
713
714
715
716
717
            recipe_state = self.fp8_meta[fp8_meta_tensor_key]
            if recipe.delayed() and isinstance(recipe_state, DelayedScalingRecipeState):
                self.adjust_amax_history_length(recipe.amax_history_len, fwd=fwd)
                return
            if recipe.mxfp8() and isinstance(recipe_state, MXFP8BlockScalingRecipeState):
                return
718
719
720
721
            if recipe.float8_current_scaling() and isinstance(
                recipe_state, Float8CurrentScalingRecipeState
            ):
                return
722
723
724
725
            if recipe.float8_block_scaling() and isinstance(
                recipe_state, Float8BlockScalingRecipeState
            ):
                return
726
727
            if recipe.nvfp4() and isinstance(recipe_state, NVFP4BlockScalingRecipeState):
                return
728
729
730

        # Max. number of fp8 tensors per GEMM = 3 (input, weight, output) for fwd and
        # 2 (grad_output and grad_input) for bwd
731
        num_fp8_tensors = self.fp8_meta["num_gemms"] * 3 if fwd else self.fp8_meta["num_gemms"] * 2
732

733
734
735
736
737
        # Initialize recipe state and quantizers
        recipe_state = RecipeState.create(
            recipe,
            mode=("forward" if fwd else "backward"),
            num_quantizers=num_fp8_tensors,
738
739
        )

740
741
742
        self.fp8_meta[fp8_meta_tensor_key] = recipe_state
        self.quantizers[fp8_meta_tensor_key] = recipe_state.make_quantizers()

743
744
745
746
747
748
749
750
751
    def _update_weight_quantizers(self) -> None:
        """Update the quantizers for the weight tensors."""
        weight_tensors = self._get_weight_tensors()
        weight_quantizers = self._get_weight_quantizers()
        assert len(weight_tensors) == len(weight_quantizers), (
            f"Number of weight tensors ({len(weight_tensors)}) and quantizers "
            f"({len(weight_quantizers)}) must match"
        )
        for weight, quantizer in zip(weight_tensors, weight_quantizers):
752
            if quantizer is not None and isinstance(weight, QuantizedTensorStorage):
753
754
                weight.update_quantizer(quantizer)

755
    def _get_weight_tensors(self) -> List[Union[torch.Tensor, QuantizedTensorStorage]]:
756
757
758
759
760
761
762
763
764
765
766
        """Get the weight tensors of the module."""
        raise NotImplementedError(
            f"{self.__class__.__name__} class does not implement _get_weight_tensors function"
        )

    def _get_weight_quantizers(self) -> List[Quantizer]:
        """Get the weight quantizers of the module."""
        raise NotImplementedError(
            f"{self.__class__.__name__} class does not implement _get_weight_quantizers function"
        )

767
    def init_fp8_meta_tensors(self, recipe: Recipe) -> None:
768
        """Init scales and amaxes."""
769
770
771
        self.set_meta_tensor(True, recipe)
        self.set_meta_tensor(False, recipe)

772
773
        self.fp8_meta_tensors_initialized = True

774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
    def get_fp8_meta_tensors(self) -> None:
        """Get scales and amaxes."""
        fwd_key, bwd_key = "scaling_fwd", "scaling_bwd"
        if fwd_key not in self.fp8_meta or bwd_key not in self.fp8_meta:
            return None

        fp8_meta_tensors = {fwd_key: [], bwd_key: []}
        with torch.no_grad():
            for key in (fwd_key, bwd_key):
                fp8_meta_tensors[key].append(self.fp8_meta[key].scale.clone())
                fp8_meta_tensors[key].append(self.fp8_meta[key].amax_history.clone())
        return fp8_meta_tensors

    def reset_fp8_meta_tensors(self, fp8_meta_tensors=None) -> None:
        """Reset scales and amaxes."""
789

790
791
792
793
794
        def reset(key):
            if key in self.fp8_meta:
                if fp8_meta_tensors is None:
                    self.fp8_meta[key].scale.copy_(torch.ones_like(self.fp8_meta[key].scale))
                    self.fp8_meta[key].amax_history.copy_(
795
796
                        torch.zeros_like(self.fp8_meta[key].amax_history)
                    )
797
798
799
                else:
                    assert key in fp8_meta_tensors, "Cannot reset fp8 tensors."
                    self.fp8_meta[key].scale.copy_(fp8_meta_tensors[key][0])
800
                    self.fp8_meta[key].amax_history.copy_(fp8_meta_tensors[key][1])
801

802
803
804
805
        with torch.no_grad():
            reset("scaling_fwd")
            reset("scaling_bwd")

806
    def get_extra_state(self) -> torch.Tensor:
807
        """Save before checkpointing."""
808

809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
        # This implementation is working around a few issues:
        #
        # (1) PyTorch's "extra state" infrastructure might be able to
        #     support any picklable type, but they make no guarantees.
        #     We have experienced problems (e.g. in ONNX export) with
        #     non-tensor extra state.
        # (2) PyTorch's checkpointing infrastructure does not remap
        #     devices for "extra state" like it does for "state dict".
        #     Thus, we want to avoid putting extra state on the GPU
        #     since it may be loaded on the wrong device.
        # (3) The extra state consists of many small tensors. If we
        #     want to copy them all to CPU, then we need to avoid the
        #     overhead of many GPU-CPU memory transfers.
        #
        # See: https://github.com/NVIDIA/TransformerEngine/pull/351
        # See: https://github.com/NVIDIA/TransformerEngine/pull/363

        def to_cpu(src: torch.Tensor) -> torch.Tensor:
            """Helper function to make CPU copy of tensor

            Memory transfer is asynchronous w.r.t. host, so GPU should
            be synchronized before using result.

            """
            dst = torch.empty_like(src, device="cpu")
            dst.copy_(src, non_blocking=True)
            return dst

        # Store FP8 state if needed
        state = None
839
        fp8_checkpoint = self.fp8_meta["fp8_checkpoint"] or self.fp8 or self.fp8_calibration
840
        if not fp8_checkpoint:
841
            return torch.empty(0, dtype=torch.uint8)
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859

        # Copy tensors to CPU and store
        state = {}
        state["recipe"] = self.fp8_meta["recipe"]
        if state["recipe"].delayed():
            state["scale_fwd"] = to_cpu(self.fp8_meta["scaling_fwd"].scale)
            state["amax_history_fwd"] = to_cpu(self.fp8_meta["scaling_fwd"].amax_history)
            state["scale_bwd"] = to_cpu(self.fp8_meta["scaling_bwd"].scale)
            state["amax_history_bwd"] = to_cpu(self.fp8_meta["scaling_bwd"].amax_history)

        # Store other pickelable values
        extra = {}
        for k, v in self.fp8_meta.items():
            if k != "buffer_index_and_autocast_key" and isinstance(
                v, (bool, int, float, str, tuple, list)
            ):
                extra[k] = v
        state["extra_fp8_variables"] = extra
860

861
862
863
864
        # Serialize state into byte tensor
        torch.cuda.synchronize()
        state_serialized = bytearray(pickle.dumps(state))
        state_serialized = torch.frombuffer(state_serialized, dtype=torch.uint8)
865
        return state_serialized
866

867
    def set_extra_state(self, state: torch.Tensor) -> None:
868
        """Load previous state."""
869
870
871
872
873

        # Maintain backwards compatibility with older checkpoints.
        if state is None:
            return

874
        # Load state
875
        if isinstance(state, torch.Tensor):
876
877
878
            # No FP8 is indicated by an empty tensor we don't need to unpickle.
            if state.numel() == 0:
                return
879
            # Default format: byte tensor with pickled data
880
            state = pickle.loads(state.detach().cpu().numpy().tobytes())
881
        elif isinstance(state, io.BytesIO):
882
            # Deprecated format with io.BytesIO
883
            state.seek(0)
884
            state = torch.load(state, map_location="cuda")
885
886
        else:
            raise RuntimeError("Unsupported checkpoint format.")
887
888
889

        if state is None:
            return
890

891
892
893
894
895
896
897
898
        # TE 1.x checkpoint compatibility: add DelayedScaling recipe if missing
        if "recipe" not in state:
            # TE 1.x only supported delayed scaling, which was the default recipe
            state["recipe"] = DelayedScaling()
            # TE 1.x also saved scale_inv, which is not needed with Recipe object
            state.pop("scale_inv_fwd", None)
            state.pop("scale_inv_bwd", None)

899
        # Load extra items
900
        self.fp8_meta.update(state["extra_fp8_variables"])
901
        self.fp8_meta["recipe"] = state["recipe"]
902
903
904
        if "global_fp8_buffer_pos_fwd_recompute" in self.fp8_meta:
            del self.fp8_meta["global_fp8_buffer_pos_fwd_recompute"]

905
        # Initialize before loading
906
        self.init_fp8_meta_tensors(self.fp8_meta["recipe"])
907
908
909
910
911
912
913
914
915
916
917

        def copy_tensor(src: torch.Tensor, dst: torch.Tensor) -> None:
            """Helper function to copy tensor from CPU

            Memory transfer is asynchronous w.r.t. host, so GPU should
            be synchronized before using result.

            """
            dst.copy_(src, non_blocking=True)

        # Load tensors
918
919
920
921
922
        if self.fp8_meta["recipe"].delayed():
            copy_tensor(state["scale_fwd"], self.fp8_meta["scaling_fwd"].scale)
            copy_tensor(state["amax_history_fwd"], self.fp8_meta["scaling_fwd"].amax_history)
            copy_tensor(state["scale_bwd"], self.fp8_meta["scaling_bwd"].scale)
            copy_tensor(state["amax_history_bwd"], self.fp8_meta["scaling_bwd"].amax_history)
923
        torch.cuda.synchronize()
924
925
926
927
928

    def set_activation_dtype(self, inp: torch.Tensor) -> None:
        """Get activation data type for AMP."""
        # Native AMP (`torch.autocast`) gets highest priority
        if torch.is_autocast_enabled():
929
            self.activation_dtype = torch_get_autocast_gpu_dtype()
930
931
932
            return

        # All checks after this have already been performed once, thus skip
933
        if self.activation_dtype == inp.dtype:
934
935
            return

936
        dtype = inp.dtype
937
938
939
940
941
942
943
        if not self.allow_different_data_and_param_types:
            for name, param in self.named_parameters():
                if param is not None:
                    assert dtype == param.dtype, (
                        "Data types for parameters must match when outside of autocasted region. "
                        f" Found input dtype: {dtype} and {name!r} dtype: {param.dtype}"
                    )
944
        self.activation_dtype = dtype
945
946

    def set_tensor_parallel_group(self, tp_group: Union[dist_group_type, None]) -> None:
947
948
949
950
951
952
        """
        Set the tensor parallel group for the given
        module before executing the forward pass.

        Parameters
        ----------
Paweł Gadziński's avatar
Paweł Gadziński committed
953
        tp_group : ProcessGroup, default = None
954
955
                  tensor parallel process group.
        """
956
957
958
        self.tp_group = tp_group
        self.tp_group_initialized = True

959
960
961
    def _get_fp8_params(self) -> Union[List[torch.Tensor], None]:
        """returns the FP8 weights."""
        fp8_params = []
962
        for param in self.parameters(recurse=False):
963
            if isinstance(param, QuantizedTensor) and param.requires_grad:
964
965
966
967
968
                fp8_params.append(param)
        if len(fp8_params) == 0:
            return None
        return fp8_params

969
970
    # This routine is shared across FP8 and FP8_calibration paths so should not actually
    # assume FP8 execution.
971
    def init_fp8_metadata(self, num_gemms: int = 1) -> None:
972
        """Initialize fp8 related metadata and tensors during fprop."""
973
974
        _original_recipe = self.fp8_meta.get("recipe", None)

975
        self.fp8_parameters = FP8GlobalStateManager.with_fp8_parameters()
976
977
        self.fp8 = FP8GlobalStateManager.is_fp8_enabled()
        self.fp8_calibration = FP8GlobalStateManager.is_fp8_calibration()
978
        fp8_enabled = self.fp8 or self.fp8_calibration
979
        self.fp8_meta["fp8_checkpoint"] = self.fp8 or self.fp8_calibration
980

981
        if self.fp8_parameters or fp8_enabled:
982
983
984
985
            if (
                self.fp8_initialized
                and FP8GlobalStateManager.get_fp8_recipe() == self.fp8_meta["recipe"]
            ):
986
                # FP8 init has already been run and recipe is the same, don't do anything.
987
                return
988
            self.fp8_meta["recipe"] = FP8GlobalStateManager.get_fp8_recipe()
989
990
991
992
993
994
995
996
997
998
999
        else:
            # If fp8 isn't enabled, turn off and return.
            self.fp8_initialized = False
            return

        if self.fp8_parameters and not self.fp8_initialized:
            self.fp8_meta["num_gemms"] = num_gemms
            self.init_fp8_meta_tensors(self.fp8_meta["recipe"])

        if fp8_enabled:
            # Set FP8 and other FP8 metadata
1000
            self.fp8_meta["num_gemms"] = num_gemms
1001
            self.fp8_meta["fp8_group"] = FP8GlobalStateManager.get_fp8_group()
1002
1003

            # Set FP8_MAX per tensor according to recipe
1004
1005
1006
            if hasattr(self.fp8_meta["recipe"], "fp8_format"):
                self.fp8_meta["fp8_max_fwd"] = self.fp8_meta["recipe"].fp8_format.value.max_fwd
                self.fp8_meta["fp8_max_bwd"] = self.fp8_meta["recipe"].fp8_format.value.max_bwd
1007
1008

            # Allocate scales and amaxes
1009
            self.init_fp8_meta_tensors(self.fp8_meta["recipe"])
1010
            self.fp8_initialized = True
1011
1012

            self.fp8_meta["recipe"] = FP8GlobalStateManager.get_fp8_recipe()
1013

1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
        _current_recipe = self.fp8_meta["recipe"]
        if _original_recipe is not None and not (
            issubclass(_current_recipe.__class__, _original_recipe.__class__)
            or issubclass(_original_recipe.__class__, _current_recipe.__class__)
        ):
            warnings.warn(
                f"Recipe type changed from {_original_recipe.__class__.__name__} "
                f"to {_current_recipe.__class__.__name__}. "
                "This may affect model behavior."
            )
            # Clear cached workspaces as they were created with the old recipe/quantizer type
            self._fp8_workspaces.clear()

1027
1028
1029
1030
1031
    @contextmanager
    def prepare_forward(
        self,
        inp: torch.Tensor,
        num_gemms: int = 1,
1032
        allow_non_contiguous: bool = False,
1033
        allow_different_data_and_param_types: bool = False,
Jan Bielak's avatar
Jan Bielak committed
1034
    ) -> Generator[torch.Tensor, None, None]:
1035
1036
1037
1038
1039
1040
        """Checks and prep for FWD.
        The context manager is needed because there isn't a way for a module to know
        if it's the last FP8 module in the forward autocast. It is useful
        to setup the forward aggregated amax reduction for every module
        just in case. The autocast exit will pick up the most recent one.
        """
1041
        self.allow_different_data_and_param_types = allow_different_data_and_param_types
1042
        self.forwarded_at_least_once = True
1043

1044
1045
        # Activation recomputation is used and this is the second forward phase.
        if self.fp8 and in_fp8_activation_recompute_phase():
1046
            delayed_scaling_recipe = self.fp8_meta["recipe"].delayed()
1047
            FP8GlobalStateManager.get_old_fp8_meta_tensors_for_recompute(self.fp8_meta)
1048
1049
1050
1051
1052
1053
1054
        else:
            assert inp.is_cuda, "TransformerEngine needs CUDA."

            if self.tp_size > 1:
                assert self.tp_group_initialized, "TP group not initialized."

            self.set_activation_dtype(inp)
1055
            self.init_fp8_metadata(num_gemms=num_gemms)
1056
            self._check_weight_tensor_recipe_correspondence()
1057

1058
1059
1060
1061
1062
1063
1064
            delayed_scaling_recipe = self.fp8 and self.fp8_meta["recipe"].delayed()
            if delayed_scaling_recipe:
                if self.sequence_parallel:
                    assert self.fp8_meta["recipe"].reduce_amax, (
                        "Amax reduction across tensor parallel group is "
                        "necessary when using sequence parallelism with FP8."
                    )
1065

1066
1067
                if not FP8GlobalStateManager.fp8_graph_capturing():
                    FP8GlobalStateManager.add_fp8_tensors_to_global_buffer(self.fp8_meta)
1068

1069
1070
1071
                # Activation recomputation is used and this is the first forward phase.
                if self.training and is_fp8_activation_recompute_enabled():
                    FP8GlobalStateManager.copy_forward_fp8_meta_tensors_for_recompute(self.fp8_meta)
1072

1073
        with get_nvtx_range_context(self.__class__.__name__ + " forward"):
1074
1075
1076
            if not allow_non_contiguous and not inp.is_contiguous():
                inp = inp.contiguous()
            yield inp
1077

1078
        if delayed_scaling_recipe and self.fp8 and in_fp8_activation_recompute_phase():
1079
            FP8GlobalStateManager.restore_fp8_meta_tensors(self.fp8_meta)
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099

    def set_nccl_overlap_warning_if_tp(self) -> None:
        """When using TP, the NCCL communication needs to be scheduled
        before the GEMM for there to be a guaranteed overlap. From the
        host side in TE, the comm calls are always launched first, but
        to ensure that the GEMM isn't scheduled first, the environment
        variable `CUDA_DEVICE_MAX_CONNECTIONS` needs to be set to 1 to
        force a single channel.
        """
        if self.tp_size == 1:
            return
        num_cuda_work_queues = int(os.getenv("CUDA_DEVICE_MAX_CONNECTIONS", "0"))
        if num_cuda_work_queues != 1:
            warnings.warn(
                "To guarantee overlapping TP and SP collectives with the backward"
                "GEMMs, set environment variable CUDA_DEVICE_MAX_CONNECTIONS = 1"
            )

    @staticmethod
    def grad_output_preprocess(
1100
1101
1102
1103
        ctx,
        grad_output: torch.Tensor,
        row_parallel_mode: bool,
        quantizer: Optional[Quantizer],
1104
1105
1106
    ) -> Tuple[Union[torch.Tensor, None], ...]:
        """Utility function for backward.
        Returns tuple in order (all optional/None based on training precion/recipe):
1107
1108
            R1: gathered `grad_output`.
            R2: bias gradient on R1.
1109
1110

        """
1111
1112
        grad_output = grad_output.reshape((-1, grad_output.shape[-1]))
        grad_output = grad_output.contiguous()
1113
1114
        gather_grad_output = row_parallel_mode and ctx.sequence_parallel

1115
        # Non-FP8 case: bgrad is fused with wgrad for this case.
1116
        if not ctx.fp8 and not ctx.debug:
1117
            if gather_grad_output:
1118
                if not ctx.ub_overlap_ag:  # Perform NCCL all-gather
1119
                    grad_output, _ = gather_along_first_dim(grad_output, ctx.tp_group)
1120
1121
1122
1123
1124
1125
1126
                else:  # Initialize Userbuffers all-gather
                    grad_output, _ = fill_userbuffers_buffer_for_all_gather(
                        ctx.ub_obj_gradout,
                        grad_output,
                        None,
                        ctx.tp_group,
                    )
1127
1128
1129
            return grad_output, None

        # FP8 with all-gather: unfused bgrad, fused cast + transpose
1130
        # Also supports debug quantization, which is handled inside gather_along_first_dim.
1131
1132
        if gather_grad_output:
            grad_bias = None
1133
            if ctx.use_bias:
1134
                grad_bias = grad_output.view(-1, grad_output.shape[-1]).sum(dim=0)
1135
            if ctx.ub_overlap_ag:
1136
1137
                # Quantize the gradient if needed
                if not isinstance(
1138
1139
1140
                    grad_output,
                    (
                        QuantizedTensor,
1141
1142
1143
                        Float8TensorStorage,
                        MXFP8TensorStorage,
                        Float8BlockwiseQTensorStorage,
1144
                    ),
1145
1146
1147
1148
                ):
                    grad_output = quantizer(grad_output)

                # Copy into communication buffer, and replace original gradient with it
1149
1150
1151
1152
1153
1154
                grad_output, _ = fill_userbuffers_buffer_for_all_gather(
                    ctx.ub_obj_gradout,
                    grad_output,
                    quantizer,
                    ctx.tp_group,
                )
1155
            else:
1156
1157
1158
1159
                grad_output, _ = gather_along_first_dim(
                    grad_output,
                    ctx.tp_group,
                    quantizer=quantizer,
1160
                )
1161
            return grad_output, grad_bias
1162

1163
1164
1165
1166
        # Debug without all-gather: unfused cast and bgrad
        # bgrad only if wgrad is in FP8, otherwise it is fused with wgrad and we return None
        if ctx.debug:
            grad_output_ = quantizer(grad_output)
1167
            if ctx.use_bias:
1168
1169
1170
1171
1172
1173
                grad_bias = grad_output.view(-1, grad_output.shape[-1]).sum(dim=0)
            else:
                grad_bias = None
            grad_output = grad_output_
            return grad_output, grad_bias

1174
1175
        # FP8 without all-gather: fused bgrad + cast + transpose
        grad_bias = None
1176
        if ctx.use_bias:
1177
1178
            if isinstance(
                grad_output,
1179
1180
1181
1182
1183
1184
                (
                    QuantizedTensor,
                    Float8TensorStorage,
                    MXFP8TensorStorage,
                    Float8BlockwiseQTensorStorage,
                ),
1185
            ):
1186
                grad_bias = grad_output.dequantize().view(-1, grad_output.shape[-1]).sum(dim=0)
1187
            else:
1188
                if isinstance(quantizer, Float8BlockQuantizer):
1189
1190
1191
1192
                    # unfuse bgrad for now until cast_transpose + dgrad calculation is ready for Float8BlockQuantizer.
                    grad_bias = grad_output.view(-1, grad_output.shape[-1]).sum(dim=0)
                else:
                    grad_bias, grad_output = tex.bgrad_quantize(grad_output, quantizer)
1193
        if not isinstance(grad_output, QuantizedTensorStorage):
1194
1195
            grad_output = quantizer(grad_output)
        return grad_output, grad_bias
1196

1197
1198
1199
1200
1201
1202
    def register_parameter(self, name, param, **kwargs):
        """
        Thin wrapper around PyTorch parameter registration to stash additional parameter
        metedata used in deferred initialization.
        """
        super().register_parameter(name, param)
1203
1204
1205
1206
1207
1208
        # Initialize param_init_meta exactly once during the init. FSDP2 can call
        # register parameter again to change parameters to DTensors. And it calls
        # it without custom fp8 specific kwargs that we need. And so we dont want
        # to reset/loose our fp8 init attributes.
        if hasattr(self, "param_init_meta") and name not in self.param_init_meta:
            self.param_init_meta[name] = _ParameterInitMeta(**kwargs)
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219

    def reset_parameters(self, defer_init: Optional[bool] = False) -> None:
        """
        Reset all module parameters to initial values. Unless deferred initialization
        is specified, all parameters on a 'meta' device are also materialized on a real cuda
        device before the values are reset to initial.
        """
        if defer_init:
            return

        for name, param in self.named_parameters(recurse=False):
1220
1221
1222
1223
1224
            # Check if parameter is a DTensor (FSDP2) or regular tensor
            is_dtensor = isinstance(param, DTensor)
            dtensor_param = param if is_dtensor else None
            # Need to update/quantize local tensor in case of DTensor
            param = param._local_tensor if is_dtensor else param
1225
            # Ensure parameter is on a real device
1226
1227
            if param.device == torch.device("meta"):
                param = torch.empty_like(param, device="cuda")
1228
1229
1230
1231
1232
1233
            # Initialize the parameter values on device
            init_fn = self.param_init_meta[name].init_fn
            get_rng_state_tracker = self.param_init_meta[name].get_rng_state_tracker
            if get_rng_state_tracker is None:
                init_fn(param)
            else:
1234
1235
1236
1237
1238
1239
                if hasattr(self, "rng_tracker_name") and self.rng_tracker_name:
                    with get_rng_state_tracker().fork(self.rng_tracker_name):
                        init_fn(param)
                else:
                    with get_rng_state_tracker().fork():
                        init_fn(param)
1240

1241
            # Wrap parameters in QuantizedTensor if needed
1242
            fp8_meta_index = self.param_init_meta[name].fp8_meta_index
1243
            high_precision_init_val = None
1244
            if self.primary_weights_in_fp8 and fp8_meta_index is not None:
1245
1246

                # Keep high-precision values on CPU if needed
1247
1248
1249
                if self.preserve_high_precision_init_val:
                    high_precision_init_val = param.detach().cpu()

1250
                # Configure quantizer
1251
                quantizer = self.quantizers["scaling_fwd"][fp8_meta_index]
1252
1253
1254
                if quantizer is None:
                    raise RuntimeError("Weight quantizer has not been initialized")
                quantizer.set_usage(rowwise=True, columnwise=torch.is_grad_enabled())
1255
                quantizer.internal = False
1256
1257
1258
1259
1260
1261
1262
1263
1264
                if is_dtensor and isinstance(quantizer, Float8CurrentScalingQuantizer):
                    device_mesh = dtensor_param.device_mesh
                    amax_reduction_group = (
                        device_mesh.get_group(mesh_dim="shard")
                        if device_mesh.ndim > 1
                        else device_mesh.get_group()
                    )
                    quantizer.amax_reduction_group = amax_reduction_group
                    quantizer.with_amax_reduction = True
1265
                # Quantize parameter
1266
                param = quantizer(param)
1267
1268
1269
1270
1271

            # Redo parameter wrap in case we broke it above
            # NOTE: Currently this can only be broken when primary weights are in Fp8 but
            #       re-applying the nn.Parameter() wrap is a no-op when the input is already
            #       a parameter so we always re-apply it just for extra safety.
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
            if is_dtensor:
                # recreate the DTensor from the parameter.
                dtensor_param = DTensor.from_local(
                    param,
                    device_mesh=dtensor_param.device_mesh,
                    placements=dtensor_param.placements,
                    shape=dtensor_param.size(),
                    stride=dtensor_param.stride(),
                )
                dtensor_param = torch.nn.Parameter(dtensor_param)
            else:
                param = torch.nn.Parameter(param)
1284
1285

            # Keep high-precision values on CPU if needed
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
            if high_precision_init_val is not None:

                # - Master weights are initialized from model weights, if we use fp8 primary
                #   weights to initialize master weights, the numerical values of master weights
                #   are not consistent with the numerical values when we initialize them from
                #   bf16/fp16 weights.
                # - So we add a `_high_precision_init_val` attribute to each model weight to store
                #   the original bf16/fp16 weight on cpu before casting it to fp8. And users can
                #   use `get_high_precision_init_val` to get this cpu tensor.
                # - This cpu tensor is not needed once the master weight is initialized, so users
                #   should call `clear_high_precision_init_val` to remove it after master weight
                #   is initialized.

                def get(self):
                    if hasattr(self, "_high_precision_init_val"):
                        return self._high_precision_init_val
                    return None

                def clear(self):
                    if hasattr(self, "_high_precision_init_val"):
                        del self._high_precision_init_val

                param._high_precision_init_val = high_precision_init_val
                param.get_high_precision_init_val = MethodType(get, param)
                param.clear_high_precision_init_val = MethodType(clear, param)
1311
                # Update the parameter based on its type
1312

1313
1314
1315
1316
            if not is_dtensor:
                setattr(self, name, param)
            else:
                setattr(self, name, dtensor_param)
1317

1318
1319
1320
    @abstractmethod
    def forward(self):
        """Needs override."""
1321

1322
    def get_weight_workspace(
1323
        self,
1324
1325
        *,
        tensor: Optional[torch.Tensor] = None,
1326
        quantizer: Optional[Quantizer] = None,
1327
1328
1329
        cache_name: Optional[str] = None,
        update_workspace: bool = True,
        skip_update_flag: Optional[torch.Tensor] = None,
1330
        fsdp_group: Optional[dist_group_type] = None,
1331
        workspace_dtype: Optional[torch.dtype] = None,
1332
    ) -> QuantizedTensor:
1333
        """Get workspace buffer for weights and maybe update its values
1334
1335
1336
1337
1338
1339
1340
1341

        The workspace buffer may be cached for future function calls.

        Parameters
        ----------
        tensor : torch.Tensor, optional
            Values to copy into workspace. Required if the workspace
            is being constructed or updated.
1342
1343
1344
        quantizer: Quantizer, optional
            Quantizer used to cast the weights. Required if the
            workspace is being constructed or updated.
1345
1346
        cache_name: str, optional
            Key for caching.
Paweł Gadziński's avatar
Paweł Gadziński committed
1347
        update_workspace: bool, default = True
1348
1349
1350
1351
            Update workspace with values from `tensor`.
        skip_update_flag: torch.Tensor, optional
            GPU flag to skip updating the workspace. Take precedence
            over `update_workspace` if provided.
1352
1353
        fsdp_group: bool, default = None
            FSDP process group that the weights are distributed over.
1354
1355
1356
        workspace_dtype: torch.dtype, default = None
            If weight workspace contains high-precision tensor - for example
            for debug quantization, this is dtype of the tensor.
1357
1358
        """

1359
1360
1361
        # Handle case where weights are already quantized
        # Note: Make sure weights have required usages, but do not
        # destroy unnecessary usages since they may be used later.
1362
        if isinstance(tensor, QuantizedTensor):
1363
1364
1365
1366
1367
1368
            update_rowwise_usage = True if quantizer.rowwise_usage else None
            update_columnwise_usage = True if quantizer.columnwise_usage else None
            tensor.update_usage(
                rowwise_usage=update_rowwise_usage,
                columnwise_usage=update_columnwise_usage,
            )
1369
1370
            return tensor

1371
        # Try getting workspace from cache
1372
1373
1374
        out = None
        if cache_name is not None:
            out = self._fp8_workspaces.get(cache_name, None)
1375
1376
1377
1378

        # Reset cache if workspace is invalid
        if out is not None and quantizer is not None:
            reset_cache = False
1379
            if isinstance(out, Float8TensorStorage):
1380
1381
1382
1383
1384
1385
                if (
                    not is_non_tn_fp8_gemm_supported()
                    and quantizer.columnwise_usage
                    and out._transpose is None
                ):
                    reset_cache = True
1386
            elif isinstance(out, MXFP8TensorStorage):
1387
                if quantizer.rowwise_usage and out._rowwise_data is None:
1388
                    reset_cache = True
1389
                elif quantizer.columnwise_usage and out._columnwise_data is None:
1390
1391
1392
1393
                    reset_cache = True
            if isinstance(out, DebugQuantizedTensor) != isinstance(quantizer, DebugQuantizer):
                reset_cache = True
            if reset_cache:
1394
                out = None
1395
                del self._fp8_workspaces[cache_name]
1396

1397
1398
1399
1400
1401
        # Gather cached Fp8 workspace if it's distributed
        # NOTE: FSDP sharding is supported only for Fp8 buffers and will not work
        #       for models initialized with Fp8 primary weights.
        if (
            out is not None
1402
            and tensor is not None
1403
            and fsdp_group is not None
1404
            and out.data.shape != tensor.data.shape
1405
1406
1407
1408
        ):
            _fsdp_gather_tensors(fsdp_group, [tensor.data.shape], out)

        # Construct workspace if needed
1409
        if out is None:
1410
            if tensor is None or quantizer is None:
1411
                raise ValueError(
1412
                    "tensor and quantizer kwargs must be provided to construct FP8 workspace"
1413
                )
1414
1415
1416
1417
1418
1419
1420

            if cache_name is not None:
                # Ensure the tensor in the cache is an instance of torch.Tensor,
                # as it persists beyond a single forward pass.
                # Setting internal=True would cause the data to be removed in prepare_for_saving(...).
                quantizer_internal = quantizer.internal
                quantizer.internal = False
1421
            out = quantizer.quantize(tensor, dtype=workspace_dtype)
1422
1423
            if cache_name is not None:
                quantizer.internal = quantizer_internal
1424
1425

            # Update cache
1426
1427
            if cache_name is not None:
                self._fp8_workspaces[cache_name] = out
1428
            return out
1429
1430
1431
1432
1433
1434

        # Update workspace if needed
        if skip_update_flag is not None:
            update_workspace = True
        if update_workspace:
            if tensor is None:
1435
                raise ValueError("tensor kwarg must be provided to update FP8 workspace")
1436
            if hasattr(out, "quantize_"):
1437
                out.quantize_(tensor, noop_flag=skip_update_flag)
1438
1439
            else:
                tex.quantize(tensor, quantizer, out, skip_update_flag)
1440
        return out
1441

1442
1443
1444
    def _load_from_state_dict(
        self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs
    ):
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
        """
        This function loads tensors and extra state including fp8 metadata.
        This metadata is essential for copying fp8 tensors, as the copy_ function
        uses the scale_inv parameter from fp8_meta to set the correct scaling factor
        for the new tensor.
        Hence, this extra state must be loaded before the tensor copying process,
        not after, as is typically done in _load_from_state_dict.
        Tensors are copied into fp8 tensors only when self.primary_weights_in_fp8=True,
        otherwise, this behavior is not required.
        """
        if self.primary_weights_in_fp8:
            extra_state_key = prefix + torch.nn.modules.module._EXTRA_STATE_KEY_SUFFIX
            if extra_state_key in state_dict:
                self.set_extra_state(state_dict[extra_state_key])
1459
1460
1461
        super()._load_from_state_dict(
            state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs
        )
1462

1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
    def register_wgrad_accumulation_and_reduce_hooks(self, wgrad_accumulation_and_reduce_hook):
        """
        This method is used to manually control the weight gradient accumulation and reduce.
        This method should be called before the backward() method.
        Set the skip_wgrad_accumulation_and_reduce to True to skip the weight gradient accumulation
        and reduce in backward();
        And register the wgrad_accumulation_and_reduce_func to be called in backward_dw() method.
        """
        self.wgrad_accumulation_and_reduce_hooks.append(wgrad_accumulation_and_reduce_hook)

1473
1474
1475
1476
1477
1478
1479
1480
1481
    def need_backward_dw(self):
        """
        Check if this module needs to execute the delayed weight gradient computation.
        This method should be used at the beginning of self.backward_dw() to determine if it
        should actually be executed or just return without doing anything.
        User can also manually call this method to check that before calling into backward_dw().
        """
        return self.wgrad_store is not None and self.wgrad_store.delay_wgrad_compute()

1482
1483
1484
1485
1486
    def backward_dw(self):
        """
        Execute the delayed weight gradient computation.
        This method is called after the main backward pass to compute weight gradients.
        """
1487
        if not self.need_backward_dw():
1488
            return
1489
        with get_nvtx_range_context(f"_{self.__class__.__name__}_wgrad"):
1490
            (wgrad, bgrad), _ = self.wgrad_store.pop()
1491
            if not self.fuse_wgrad_accumulation:
1492
                weight_tensor = noop_cat(self._get_weight_tensors())
1493
                weight_tensor.grad = wgrad.to(weight_tensor.dtype)
1494
1495
1496
            if self.use_bias:
                bias_tensor = noop_cat([getattr(self, name) for name in self.bias_names])
                if bias_tensor.grad is None:
1497
                    bias_tensor.grad = bgrad.to(bias_tensor.dtype)
1498
1499
1500
1501
            del wgrad
            del bgrad
            for wgrad_accumulation_and_reduce_hook in self.wgrad_accumulation_and_reduce_hooks:
                wgrad_accumulation_and_reduce_hook()
1502

1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
    def is_debug_iter(self) -> bool:
        """
        This function checks if the debug should be enabled for this layer.
        """
        debug = TEDebugState.debug_enabled
        if not debug:
            return False
        self._validate_name()

        # If layer is run first time in new iteration,
        # we need to check if the debug should be enabled for this layer -
        # maybe in previous iterations debug features returned information
        # that no feature will be active for this layer for multiple next iterations.
        started_new_iteration = TEDebugState.get_iteration() != getattr(
            self, "debug_last_iteration", None
        )
        if started_new_iteration:
            if self.next_iter_when_debug_should_be_run is None:
                debug = False
            else:
                debug = TEDebugState.get_iteration() >= self.next_iter_when_debug_should_be_run
1524
1525
1526
1527
1528
1529
1530
            self.debug_last_iteration = TEDebugState.get_iteration()
            self.debug_enabled_in_this_iteration = debug
        else:
            # If this is the same iteration as previous invocation of the module,
            # we use the debug value from the first invocation in the iteration.
            debug = self.debug_enabled_in_this_iteration

1531
1532
1533
1534
1535
1536
        self.debug_last_iteration = TEDebugState.get_iteration()

        if self.wgrad_store is not None:
            if debug and self.wgrad_store.delay_wgrad_compute():
                raise RuntimeError("Delayed wgrad compute is not supported in debug mode.")

1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
        return debug

    def no_debug_features_active(self, quantizers):
        """
        Checks if any debug feature is active for this layer.
        """
        run_current = any_feature_enabled(quantizers)

        # Sometimes features inform that they will not be enabled for particular layer
        # for multiple next iterations.
        self.next_iter_when_debug_should_be_run = next_iter_when_debug_should_be_run(quantizers)

        if not run_current:
            return True

        if self.primary_weights_in_fp8:
            raise RuntimeError("FP8 weights are not supported in debug mode.")
        return False

1556
1557
1558
1559
1560
1561
    def _validate_name(self):
        """
        Validate name passed to the module.
        This is invoked in the forward() method as module names are assigned after Model is initialized in Megatron-LM.
        If no name is assigned, it creates a default name with layer count as the variable.
        """
1562
1563
        if self.name is not None:
            return
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
        assert TEDebugState.debug_enabled
        import nvdlfw_inspect.api as debug_api

        if self.name is None:
            debug_api.log_message(
                "Names are not provided to debug modules. ",
                "Creating and using generic names. Pass names to debug modules for better"
                " insight. ",
                level=logging.WARNING,
            )
            self.name = f"Layer_{TEDebugState.get_layer_count()}"

1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
    def _check_weight_tensor_recipe_correspondence(self) -> None:
        """
        Verify that the weight tensor types match their corresponding recipe type.
        This is invoked in the forward().

        This establishes a 1:1 correspondence between recipe types and tensor types:
        - DelayedScaling → Float8Tensor
        - Float8CurrentScaling → Float8Tensor
        - MXFP8BlockScaling → MXFP8Tensor
        - Float8BlockScaling → Float8BlockTensor

1587
1588
        Example case to check: recipe is DelayedScaling (DelayedScaling is set in autocast()),
        but the weight tensor is MXFP8Tensor (MXFP8BlockScaling is set in quantized_model_init()).
1589
1590
1591
        """
        if not self.fp8 and not self.fp8_calibration:
            return
1592
1593
        if not self.primary_weights_in_fp8:
            return
1594
1595
1596
1597
1598
1599
        if not hasattr(self, "weight_names") or not self.weight_names:
            return

        recipe = self.fp8_meta["recipe"]
        weight_tensors = [getattr(self, name) for name in self.weight_names]
        for i, tensor in enumerate(weight_tensors):
1600
            if isinstance(tensor, QuantizedTensorStorage):
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
                quantizer = tensor._get_quantizer()
                if quantizer is None:
                    continue
                compatible_recipe_class = quantizer._get_compatible_recipe()
                if compatible_recipe_class is None:
                    continue
                if not isinstance(recipe, compatible_recipe_class):
                    raise RuntimeError(
                        f"Recipe mismatch for '{self.weight_names[i]}': tensor supports recipe"
                        f" {compatible_recipe_class.__name__}, but got {recipe.__class__.__name__}."
1611
1612
                        " Please check the recipes assigned during quantized_model_init() and"
                        " autocast() calls."
1613
                    )