test_numerics.py 91.4 KB
Newer Older
1
# Copyright (c) 2022-2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
2
3
4
#
# See LICENSE for license information.

5
import math
6
import os
7
from typing import Dict, List, Tuple, Optional
8
import pytest
9
import random
10
11
12
13

import torch
import torch.nn as nn
from torch.nn import Parameter
yuguo's avatar
yuguo committed
14
from torch.utils.cpp_extension import IS_HIP_EXTENSION
15

16
17
18
19
20
from transformer_engine.pytorch.fp8 import (
    FP8GlobalStateManager,
    fp8_autocast,
    fp8_model_init,
)
21
22
23
from transformer_engine.pytorch.utils import (
    init_method_normal,
    scaled_init_method_normal,
24
    attention_mask_func,
25
    is_bf16_compatible,
26
27
)
from transformer_engine.pytorch import (
28
29
30
31
    DotProductAttention,
    LayerNormLinear,
    LayerNormMLP,
    Linear,
32
    GroupedLinear,
33
34
35
36
    MultiheadAttention,
    RMSNorm,
    TransformerLayer,
    LayerNorm,
37
38
    Fp8Padding,
    Fp8Unpadding,
39
)
40
from transformer_engine.pytorch import torch_version
41
from transformer_engine.pytorch.distributed import checkpoint as te_checkpoint
42
from transformer_engine.pytorch.cpp_extensions import general_gemm, general_grouped_gemm
43
from transformer_engine.pytorch.cpp_extensions.fused_attn import FusedAttnBackend
44
45
46
47
48
from transformer_engine.pytorch.tensor.float8_tensor import (
    Float8Quantizer,
    Float8CurrentScalingQuantizer,
)
from transformer_engine.pytorch.tensor.mxfp8_tensor import MXFP8Quantizer
49
from transformer_engine.pytorch.module.base import get_multi_stream_cublas_workspace, get_workspace
50
from transformer_engine.pytorch.utils import get_device_compute_capability
51
from transformer_engine.common import recipe
52
import transformer_engine_torch as tex
53
from utils import ModelConfig, reset_rng_states, get_available_attention_backends
54

55

56
# Only run FP8 tests on supported devices.
57
fp8_available, reason_for_no_fp8 = FP8GlobalStateManager.is_fp8_available()
58
mxfp8_available, reason_for_no_mxfp8 = FP8GlobalStateManager.is_mxfp8_available()
59
fp8_block_scaling_available, _ = FP8GlobalStateManager.is_fp8_block_scaling_available()
60

61
sm_80plus = get_device_compute_capability() >= (8, 0)
62

63
seed = 1234
64
65
# Reset RNG states.
reset_rng_states()
66

67
68
69
70
if torch_version() >= (2, 7, 0):
    torch._dynamo.config.recompile_limit = 16
else:
    torch._dynamo.config.cache_size_limit = 16
71
72
73


model_configs = {
74
75
    "small": ModelConfig(1, 128, 8, 16, num_layers=4),
    "126m": ModelConfig(1, 2048, 12, 64, num_layers=12),
76
}
77
model_configs_inference = {
78
    "126m": ModelConfig(1, 256, 12, 64, num_layers=12),
79
}
80
backends_inference = ["FlashAttention", "UnfusedAttention", "FusedAttention"]
81
82
83
module_inference = ["TransformerLayer", "MultiheadAttention"]
input_formats_inference = ["sbhd", "bshd"]

84
param_types = [torch.float32, torch.float16]
85
if is_bf16_compatible():  # bf16 requires sm_80 or higher
86
87
88
89
90
91
    param_types.append(torch.bfloat16)

batch_sizes = [1, 2]

all_boolean = [True, False]

92
93
94
95
96
97
98
99
100
101
102
103
all_activations = [
    "gelu",
    "geglu",
    "qgelu",
    "qgeglu",
    "relu",
    "reglu",
    "srelu",
    "sreglu",
    "silu",
    "swiglu",
]
104

105
106
all_normalizations = ["LayerNorm", "RMSNorm"]

107
108
mask_types = ["causal", "no_mask"]

109
NVTE_TEST_NVINSPECT_ENABLED = int(os.environ.get("NVTE_TEST_NVINSPECT_ENABLED", "0"))
110
111
112
113
114
115
116
117
118
119
120
121
122

if NVTE_TEST_NVINSPECT_ENABLED:
    # The numerics of all the layers should work the same,
    # when debug=True. I fed them with dummy feature
    # to prevent switching off debug, which can happen if
    # no feature is active.
    import nvdlfw_inspect.api as debug_api

    debug_api.initialize(
        os.environ["NVTE_TEST_NVINSPECT_CONFIG_FILE"],
        feature_dirs=os.environ["NVTE_TEST_NVINSPECT_FEATURE_DIRS"],
    )

123
124
125
126
127
128
129
130
131

fp8_recipes = []
if mxfp8_available:
    fp8_recipes.append(recipe.MXFP8BlockScaling())
if fp8_block_scaling_available:
    fp8_recipes.append(recipe.Float8BlockScaling())
if fp8_available:
    fp8_recipes.append(recipe.Float8CurrentScaling())
    fp8_recipes.append(recipe.DelayedScaling())
132

133
134
135
136
137
use_cutlass_grouped_gemm = [False]
# Only enable cutlass grouped gemm on Hopper
if torch.cuda.get_device_capability() == (9, 0):
    use_cutlass_grouped_gemm.append(True)

138

139
def is_fused_attn_available(
140
141
142
143
144
    config: ModelConfig,
    dtype: torch.dtype,
    qkv_layout="bshd_bshd_bshd",
    is_training=True,
    deterministic=False,
145
):
146
    _, _, fused_attn_backends = get_available_attention_backends(
147
148
149
150
        config,
        qkv_dtype=dtype,
        qkv_layout=qkv_layout,
        is_training=is_training,
151
        deterministic=deterministic,
152
153
    )
    return FusedAttnBackend["F16_arbitrary_seqlen"] in fused_attn_backends
154

155

156
157
158
159
def get_causal_attn_mask(sq: int) -> torch.Tensor:
    return torch.triu(torch.ones(sq, sq, device="cuda"), diagonal=1).bool()


160
161
def dtype_tols(dtype: torch.dtype) -> Dict[str, float]:
    """Estimated numerical error for a datatype
162

163
    Based on tolerances for torch.testing.assert_close.
164

165
166
167
168
169
170
171
172
173
174
175
    """
    if dtype == torch.float32:
        return dict(rtol=1.3e-6, atol=1e-5)
    if dtype == torch.float16:
        return dict(rtol=1e-3, atol=1e-5)
    if dtype == torch.bfloat16:
        return dict(rtol=1.6e-2, atol=1e-5)
    raise ValueError(f"Unsuppored dtype ({dtype})")


def assert_allclose(
176
    l1: List[torch.Tensor], l2: List[torch.Tensor], atol: float = None, rtol: float = None
177
) -> bool:
178
179
    """Ensures two lists are equal."""
    assert len(l1) == len(l2), "Unequal number of outputs."
180
    for i, (t1, t2) in enumerate(zip(l1, l2)):
181
        tols = dtype_tols(t2.dtype)
182
183
        if rtol is not None:
            tols["rtol"] = rtol
184
185
        if atol is not None:
            tols["atol"] = atol
186
        result = torch.allclose(t1, t2, **tols)
187
        if not result:
188
            diff = torch.abs(t1 - t2)
189
            tol = tols["atol"] + (tols["rtol"] * torch.abs(t2))
190
191
192
193
194
195
196
197
198
199
200
201
            exceed_mask = diff > tol
            if exceed_mask.any():
                indices = torch.nonzero(exceed_mask, as_tuple=True)
                max_diff = diff[exceed_mask].max()
                max_idx = (diff[exceed_mask] == max_diff).nonzero(as_tuple=True)[0][0]
                max_location = [idx[max_idx].item() for idx in indices]
                msg = (
                    f"Outputs not close enough in tensor at idx={i}. "
                    f"Maximum difference at location {max_location} "
                    f"with {t1[exceed_mask][max_idx].item()} vs {t2[exceed_mask][max_idx].item()} "
                    f"(diff {max_diff.item()})."
                )
202
            raise AssertionError(msg)
203
204


205
206
207
208
@pytest.fixture(autouse=True)
def reset_global_fp8_state():
    yield
    FP8GlobalStateManager.reset()
209
210


211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
class TorchScaledMaskedSoftmax(nn.Module):
    def __init__(self) -> None:
        super().__init__()

    def forward(
        self, inp: torch.Tensor, mask: torch.Tensor, scale: Optional[float] = None
    ) -> torch.Tensor:
        dtype = inp.dtype
        inp = inp.float()

        if scale is not None:
            inp = inp * scale
        mask_output = attention_mask_func(inp, mask) if mask is not None else inp

        probs = torch.nn.Softmax(dim=-1)(mask_output)
        probs = probs.to(dtype)
        return probs


class TorchDotProductAttention(torch.nn.Module):
    def __init__(
        self,
        kv_channels: int,
        attention_dropout: float = 0.0,
    ) -> None:
        super().__init__()

        self.norm_factor = math.sqrt(kv_channels)
        self.scale_mask_softmax = TorchScaledMaskedSoftmax()
        self.attention_dropout = torch.nn.Dropout(attention_dropout)

    def forward(
        self,
        query_layer: torch.Tensor,
        key_layer: torch.Tensor,
        value_layer: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        batch_size, seqlen = query_layer.shape[1], query_layer.shape[0]

        # [b, np, sq, sk]
        output_size = (
            query_layer.size(1),
            query_layer.size(2),
            query_layer.size(0),
            key_layer.size(0),
        )

        # [sq, b, np, hn] -> [sq, b * np, hn]
260
        query_layer = query_layer.reshape(output_size[2], output_size[0] * output_size[1], -1)
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
        # [sk, b, np, hn] -> [sk, b * np, hn]
        key_layer = key_layer.reshape(output_size[3], output_size[0] * output_size[1], -1)

        # preallocting result tensor: [b * np, sq, sk]
        matmul_result = torch.empty(
            output_size[0] * output_size[1],
            output_size[2],
            output_size[3],
            dtype=query_layer.dtype,
            device=torch.cuda.current_device(),
        )

        # Raw attention scores. [b * np, sq, sk]
        matmul_result = torch.baddbmm(
            matmul_result,
            query_layer.transpose(0, 1),  # [b * np, sq, hn]
            key_layer.transpose(0, 1).transpose(1, 2),  # [b * np, hn, sk]
            beta=0.0,
            alpha=(1.0 / self.norm_factor),
        )

        # change view to [b, np, sq, sk]
        attention_scores = matmul_result.view(*output_size)

        # attention scores and attention mask [b, np, sq, sk]
        attention_probs = self.scale_mask_softmax(attention_scores, attention_mask)
        attention_probs = self.attention_dropout(attention_probs)

        # value_layer -> context layer.
        # [sk, b, np, hn] --> [b, np, sq, hn]
        output_size = (
            value_layer.size(1),
            value_layer.size(2),
            query_layer.size(0),
            value_layer.size(3),
        )

        # change view [sk, b * np, hn]
299
        value_layer = value_layer.reshape(value_layer.size(0), output_size[0] * output_size[1], -1)
300
301

        # change view [b * np, sq, sk]
302
        attention_probs = attention_probs.view(output_size[0] * output_size[1], output_size[2], -1)
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317

        # matmul: [b * np, sq, hn]
        context_layer = torch.bmm(attention_probs, value_layer.transpose(0, 1))

        # change view [b, np, sq, hn]
        context_layer = context_layer.view(*output_size)

        # [b, np, sq, hn] --> [sq, b, np, hn]
        context_layer = context_layer.permute(2, 0, 1, 3).contiguous()

        # [sq, b, np, hn] --> [sq, b, hp]
        context_layer = context_layer.view(seqlen, batch_size, -1)

        return context_layer

318

319
class TorchLayerNorm(nn.Module):
320
    def __init__(self, in_features: int, eps: float, zero_centered_gamma: bool):
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
        super().__init__()
        self.eps = eps
        self.in_features = in_features
        self.zero_centered_gamma = zero_centered_gamma

        initial_value = torch.ones(in_features) if zero_centered_gamma else torch.zeros(in_features)
        self.weight = nn.Parameter(initial_value)
        self.bias = nn.Parameter(torch.zeros(in_features))
        self.register_parameter("weight", self.weight)
        self.register_parameter("bias", self.bias)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        w = self.weight if not self.zero_centered_gamma else 1 + self.weight
        w = w.to(torch.float32)
        b = self.bias.to(torch.float32)
        inp = x.to(torch.float32)
337
338
339
        out = torch.nn.functional.layer_norm(
            inp, (self.in_features,), weight=w, bias=b, eps=self.eps
        )
340
341
        return out.to(x.dtype)

342

343
344
# Adapted from https://github.com/bzhangGo/rmsnorm/blob/c6691f20ec0af4128c8159c903071f7575404295/rmsnorm_torch.py
class TorchRMSNorm(nn.Module):
345
    def __init__(self, in_features, zero_centered_gamma, eps=1e-5):
346
347
348
349
        super().__init__()

        self.eps = eps
        self.in_features = in_features
350
        self.zero_centered_gamma = zero_centered_gamma
351

352
353
        initial_value = torch.ones(in_features) if zero_centered_gamma else torch.zeros(in_features)
        self.weight = nn.Parameter(initial_value)
354
355
356
        self.register_parameter("weight", self.weight)

    def forward(self, x):
357
        norm_x2 = torch.sum(x.float() ** 2, dim=-1, keepdim=True)
358
359
        d_x = self.in_features

360
        rms_x2 = norm_x2 / d_x + self.eps
361
        r_rms_x = rms_x2 ** (-1.0 / 2)
362
        x_normed = x * r_rms_x
363

364
365
366
367
        w = self.weight.float()
        if self.zero_centered_gamma:
            w = 1 + w
        return (w * x_normed).to(x.dtype)
368

369

370
class TorchLayerNormLinear(nn.Module):
371
372
373
374
375
376
377
    def __init__(
        self,
        in_features: int,
        out_features: int,
        eps: float,
        normalization: str = "LayerNorm",
        zero_centered_gamma: bool = False,
378
        bias: bool = True,
379
    ):
380
        super().__init__()
381
        if normalization == "LayerNorm":
382
383
384
            self.layernorm = TorchLayerNorm(
                in_features, eps=eps, zero_centered_gamma=zero_centered_gamma
            )
385
        elif normalization == "RMSNorm":
386
387
388
            self.layernorm = TorchRMSNorm(
                in_features, eps=eps, zero_centered_gamma=zero_centered_gamma
            )
389
390
391
        else:
            raise RuntimeError("Unsupported normalization")

392
        self.linear = nn.Linear(in_features, out_features, bias=bias)
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return self.linear(self.layernorm(x))


class TorchMHA(nn.Module):
    def __init__(self, hidden_size: int, num_attention_heads: int):
        super().__init__()
        self.mhsa = nn.MultiheadAttention(
            embed_dim=hidden_size,
            num_heads=num_attention_heads,
            dropout=0.1,
            bias=True,
            batch_first=False,
        )

409
410
    def forward(self, x, attention_mask=None):
        output = self.mhsa(x, x, x, attn_mask=attention_mask, need_weights=False)
411
412
413
414
        if isinstance(output, tuple):
            output = output[0]
        return output

415

416
417
418
class TorchQuickGELU(nn.Module):
    def forward(self, input: torch.Tensor) -> torch.Tensor:
        return input * torch.sigmoid(1.702 * input)
419

420

421
422
423
424
class TorchSquaredRELU(nn.Module):
    def forward(self, input: torch.Tensor) -> torch.Tensor:
        return (input > 0) * input * input

425

426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
class TorchGroupedLinearWithPadding(nn.Module):

    def __init__(
        self, num_gemms, in_features, out_features, bias, params_dtype, parallel_mode, fp8
    ) -> None:
        super().__init__()

        self.padding = Fp8Padding(num_gemms)
        self.linear_fn = GroupedLinear(
            num_gemms,
            in_features,
            out_features,
            bias=bias,
            params_dtype=params_dtype,
            parallel_mode=parallel_mode,
            device="cuda",
        )
        self.unpadding = Fp8Unpadding(num_gemms)

        self.fp8 = fp8

    def forward(self, inp: torch.Tensor, m_splits: List[int]) -> torch.Tensor:
        if self.fp8:
            orig_m_splits = m_splits
            inp, m_splits = self.padding(inp, m_splits)

        out = self.linear_fn(inp, m_splits)

        if self.fp8:
            out = self.unpadding(out, orig_m_splits)

        return out


460
461
_supported_act = {
    "gelu": nn.GELU(approximate="tanh"),
462
    "geglu": nn.GELU(approximate="tanh"),
463
    "qgelu": TorchQuickGELU(),
464
465
466
    "qgeglu": TorchQuickGELU(),
    "relu": nn.ReLU(),
    "reglu": nn.ReLU(),
467
    "srelu": TorchSquaredRELU(),
468
469
470
    "sreglu": TorchSquaredRELU(),
    "silu": nn.SiLU(),
    "swiglu": nn.SiLU(),
471
}
472

473

474
475
476
477
478
479
480
class TorchGLU(nn.Module):
    def __init__(self, activation: str):
        super().__init__()
        self.act = _supported_act[activation]

    def forward(self, x):
        shape = x.size(-1)
481
482
        a = x[..., : shape // 2]
        b = x[..., (shape // 2) :]
483
484
        a = self.act(a)
        return a * b
485

486

487
class TorchLayerNormMLP(nn.Module):
488
489
490
491
492
493
494
    def __init__(
        self,
        hidden_size: int,
        ffn_hidden_size: int,
        eps: float = 1e-5,
        activation="gelu",
        normalization: str = "LayerNorm",
495
        bias: bool = True,
496
    ):
497
        super().__init__()
498
        if normalization == "LayerNorm":
499
            self.ln = TorchLayerNorm(hidden_size, eps=eps, zero_centered_gamma=False)
500
        elif normalization == "RMSNorm":
501
            self.ln = TorchRMSNorm(hidden_size, eps=eps, zero_centered_gamma=False)
502
503
        else:
            raise RuntimeError("Unsupported normalization")
504
        if "glu" in activation:
505
506
507
508
509
510
            fc1_output_features = 2 * ffn_hidden_size
            self.gelu = TorchGLU(activation)
        else:
            fc1_output_features = ffn_hidden_size
            self.gelu = _supported_act[activation]

511
512
        self.fc1 = nn.Linear(hidden_size, fc1_output_features, bias=bias)
        self.fc2 = nn.Linear(ffn_hidden_size, hidden_size, bias=bias)
513
514

    def forward(self, x):
515
516
        t = self.gelu(self.fc1(self.ln(x)))
        return self.fc2(t)
517
518
519


class TorchGPT(nn.Module):
520
521
522
    def __init__(
        self, hidden_size: int, eps: float, num_attention_heads: int, parallel_attention_mlp: bool
    ):
523
        super().__init__()
524
        self.ln = nn.LayerNorm(hidden_size, eps=eps)
525
        self.causal_attn = TorchMHA(hidden_size, num_attention_heads)
526
        self.ln_mlp = TorchLayerNormMLP(hidden_size, 4 * hidden_size, eps)
527
        self.parallel_attention_mlp = parallel_attention_mlp
528
529
530
531

    def forward(
        self,
        x: torch.Tensor,
532
        attention_mask: Optional[torch.Tensor] = None,
533
    ) -> torch.Tensor:
534
        a = self.ln(x)
535
        b = self.causal_attn(a, attention_mask)
536
537
538
539
540
541
542
        if self.parallel_attention_mlp:
            n = self.ln_mlp(x)
            x = x + nn.functional.dropout(b + n, p=0.1, training=self.training)
        else:
            x = x + nn.functional.dropout(b, p=0.1, training=self.training)
            n = self.ln_mlp(x)
            x = x + nn.functional.dropout(n, p=0.1, training=self.training)
543
544
545
        return x


546
547
548
def _test_e2e_selective_recompute(
    bs, dtype, config, fp8, recipe, fp8_model_params=False, recompute=False
):
549
    reset_rng_states()
550
    FP8GlobalStateManager.reset()
551
552
553
554
555

    sigma = 0.023
    init_method = init_method_normal(sigma)
    output_layer_init_method = scaled_init_method_normal(sigma, config.num_layers)

556
    with fp8_model_init(enabled=fp8 and fp8_model_params, recipe=recipe):
557
558
559
        block = TransformerLayer(
            config.hidden_size,
            4 * config.hidden_size,
560
            config.num_heads,
561
562
563
564
565
            layernorm_epsilon=config.eps,
            init_method=init_method,
            output_layer_init_method=output_layer_init_method,
            hidden_dropout=0.1,
            attention_dropout=0.1,
566
            kv_channels=config.kv_channels,
567
568
569
570
571
            apply_residual_connection_post_layernorm=False,
            output_layernorm=False,
            params_dtype=dtype,
            fuse_qkv_params=True,
            device="cuda",
572
573
574
        )

    te_inp_hidden_states = torch.randn(
575
        (config.max_seqlen_q, bs, config.hidden_size),
576
577
578
579
        dtype=dtype,
        device="cuda",
        requires_grad=True,
    )
580
    te_inp_hidden_states.retain_grad()
581
    te_inp_attn_mask = get_causal_attn_mask(config.max_seqlen_q)
582

583
    with fp8_autocast(enabled=fp8, fp8_recipe=recipe):
584
585
        te_out = block(
            te_inp_hidden_states,
586
            attention_mask=te_inp_attn_mask,
587
            checkpoint_core_attention=recompute,
588
589
590
591
592
593
594
595
596
597
598
599
600
601
        )
    loss = te_out.sum()
    loss.backward()
    torch.cuda.synchronize()

    outputs = [te_out, te_inp_hidden_states.grad]
    for p in block.parameters():
        if p.requires_grad:
            outputs.append(p.grad)
    return outputs


@pytest.mark.parametrize("dtype", param_types)
@pytest.mark.parametrize("bs", batch_sizes)
602
@pytest.mark.parametrize("model", ["126m"])
603
@pytest.mark.parametrize("fp8", all_boolean)
604
@pytest.mark.parametrize("recipe", fp8_recipes)
605
@pytest.mark.parametrize("fp8_model_params", all_boolean)
606
def test_gpt_selective_activation_recompute(dtype, bs, model, fp8, recipe, fp8_model_params):
607
608
    if fp8_model_params and NVTE_TEST_NVINSPECT_ENABLED:
        pytest.skip("FP8 parameters are not supported in debug mode.")
609

610
611
    config = model_configs[model]

612
    outputs = _test_e2e_selective_recompute(
613
        bs, dtype, config, fp8, recipe, fp8_model_params, recompute=False
614
615
    )
    outputs_recompute = _test_e2e_selective_recompute(
616
        bs, dtype, config, fp8, recipe, fp8_model_params, recompute=True
617
    )
618
619
620
621
622
623
624

    # Check that results match
    tols = dtype_tols(dtype)
    if dtype in (torch.float16, torch.bfloat16):
        tols["atol"] = 1e-4
    if fp8 or fp8_model_params:
        tols.update(dict(rtol=0.125, atol=0.0675))
625

626
627
628
629
630
631
632
    for i, (ref, test) in enumerate(zip(outputs, outputs_recompute)):
        torch.testing.assert_close(
            test,
            ref,
            msg=f"Mismatch in tensor {i}",
            **tols,
        )
633
634


635
def _test_e2e_full_recompute(
636
    bs, dtype, config, fp8, recipe, fp8_model_params=False, recompute=False, use_reentrant=True
637
):
638
639
640
    reset_rng_states()
    FP8GlobalStateManager.reset()

641
642
643
644
    sigma = 0.023
    init_method = init_method_normal(sigma)
    output_layer_init_method = scaled_init_method_normal(sigma, config.num_layers)

645
    with fp8_model_init(enabled=fp8 and fp8_model_params, recipe=recipe):
646
        block = TransformerLayer(
647
648
            config.hidden_size,
            4 * config.hidden_size,
649
            config.num_heads,
650
651
652
653
654
            layernorm_epsilon=config.eps,
            init_method=init_method,
            output_layer_init_method=output_layer_init_method,
            hidden_dropout=0.1,
            attention_dropout=0.1,
655
            kv_channels=config.kv_channels,
656
657
658
            apply_residual_connection_post_layernorm=False,
            output_layernorm=False,
            params_dtype=dtype,
659
            fuse_qkv_params=True,
660
            device="cuda",
661
        )
662

663
    te_inp_hidden_states = torch.randn(
664
        (config.max_seqlen_q, bs, config.hidden_size),
665
666
667
668
        dtype=dtype,
        device="cuda",
        requires_grad=use_reentrant,
    )
669
670
    if use_reentrant:
        te_inp_hidden_states.retain_grad()
671
    te_inp_attn_mask = get_causal_attn_mask(config.max_seqlen_q)
672

673
    with fp8_autocast(enabled=fp8, fp8_recipe=recipe):
674
675
676
677
678
679
        if recompute:
            te_out = te_checkpoint(
                block,
                te_inp_hidden_states,
                attention_mask=te_inp_attn_mask,
                checkpoint_core_attention=False,
680
681
682
                distribute_saved_activations=False,
                tp_group=None,
                use_reentrant=use_reentrant,
683
684
685
686
687
688
689
690
691
692
693
            )
        else:
            te_out = block(
                te_inp_hidden_states,
                attention_mask=te_inp_attn_mask,
                checkpoint_core_attention=False,
            )
    loss = te_out.sum()
    loss.backward()
    torch.cuda.synchronize()

694
695
696
697
698
699
    outputs = [te_out]
    names = ["output"]
    if use_reentrant:
        outputs.append(te_inp_hidden_states.grad)
        names.append("input")
    for name, p in block.named_parameters():
700
701
        if p.requires_grad:
            outputs.append(p.grad)
702
703
704
            names.append(name)

    return outputs, names
705
706
707
708


@pytest.mark.parametrize("dtype", param_types)
@pytest.mark.parametrize("bs", batch_sizes)
709
@pytest.mark.parametrize("model", ["126m"])
710
@pytest.mark.parametrize("fp8", all_boolean)
711
@pytest.mark.parametrize("recipe", fp8_recipes)
712
@pytest.mark.parametrize("fp8_model_params", all_boolean)
713
@pytest.mark.parametrize("use_reentrant", all_boolean)
714
715
716
def test_gpt_full_activation_recompute(
    dtype, bs, model, fp8, recipe, fp8_model_params, use_reentrant
):
717
718
    if fp8_model_params and NVTE_TEST_NVINSPECT_ENABLED:
        pytest.skip("FP8 parameters are not supported in debug mode.")
719
720
721

    config = model_configs[model]

722
723
724
725
    if not use_reentrant:
        # Non-reentrant checkpoint becomes non-deterministic with bias+GELU fusion
        os.environ["NVTE_BIAS_GELU_NVFUSION"] = "0"

726
    outputs, names = _test_e2e_full_recompute(
727
728
729
730
731
732
733
734
        bs,
        dtype,
        config,
        fp8,
        recipe,
        fp8_model_params,
        recompute=False,
        use_reentrant=use_reentrant,
735
736
    )
    outputs_recompute, _ = _test_e2e_full_recompute(
737
738
739
740
741
742
743
744
        bs,
        dtype,
        config,
        fp8,
        recipe,
        fp8_model_params,
        recompute=True,
        use_reentrant=use_reentrant,
745
    )
746
747
748
749
750

    if not use_reentrant:
        # Reset bias+GELU fusion flag to avoid contaminating other tests
        del os.environ["NVTE_BIAS_GELU_NVFUSION"]

751
752
753
754
755
756
757
758
759
760
761
762
763
    # Check that results match
    tols = dtype_tols(dtype)
    if dtype in (torch.float16, torch.bfloat16):
        tols["atol"] = 1e-3
    if fp8 or fp8_model_params:
        tols.update(dict(rtol=0.125, atol=0.0675))
    for i, (ref, test) in enumerate(zip(outputs, outputs_recompute)):
        torch.testing.assert_close(
            test,
            ref,
            msg=f"Mismatch in tensor {i}",
            **tols,
        )
764
765
766
767
768
769


def _test_e2e_checkpointing_get_model(config, dtype):
    sigma = 0.023
    init_method = init_method_normal(sigma)
    output_layer_init_method = scaled_init_method_normal(sigma, config.num_layers)
770

771
772
773
    return TransformerLayer(
        config.hidden_size,
        4 * config.hidden_size,
774
        config.num_heads,
775
776
777
778
779
        layernorm_epsilon=config.eps,
        init_method=init_method,
        output_layer_init_method=output_layer_init_method,
        hidden_dropout=0.1,
        attention_dropout=0.1,
780
        kv_channels=config.kv_channels,
781
782
783
784
        apply_residual_connection_post_layernorm=False,
        output_layernorm=False,
        params_dtype=dtype,
        device="cuda",
785
786
787
788
789
790
791
    )


def _test_e2e_checkpointing(bs, dtype, config, checkpoint=False, steps=10, path="checkpoint.pt"):
    reset_rng_states()

    te_inp_hidden_states = torch.randn(
792
        (config.max_seqlen_q, bs, config.hidden_size),
793
794
795
796
        dtype=dtype,
        device="cuda",
        requires_grad=True,
    )
797
798
799
800
801
802
803
    te_inp_hidden_states.retain_grad()

    block = _test_e2e_checkpointing_get_model(config, dtype)

    for _ in range(steps // 2):
        te_out = block(
            te_inp_hidden_states,
804
            None,
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
        )
        loss = te_out.sum()
        loss.backward()

    if checkpoint:
        # This process is necessary so that we can start afresh with
        # a new model while erasing all internal state to ensure that
        # loading from a checkpoint gives bitwise identical results.
        # Since gradients are being accumulated, it is important to
        # restore them post loading the checkpoint.
        torch.save(block.state_dict(), path)

        param_grads = []
        for p in block.parameters():
            if p.requires_grad:
                param_grads.append(p.grad.clone())

822
823
824
        _cpu_rng_state = torch.get_rng_state()
        _cuda_rng_state = torch.cuda.get_rng_state()

825
826
        del block
        block = _test_e2e_checkpointing_get_model(config, dtype)
827
        block.load_state_dict(torch.load(path, weights_only=False))
828
829
        torch.set_rng_state(_cpu_rng_state)
        torch.cuda.set_rng_state(_cuda_rng_state)
830
831
832
833
834
835
836
837
838
839

        for p in block.parameters():
            if p.requires_grad:
                p.grad = param_grads.pop(0)

        assert not param_grads, "Oops!"

    for _ in range(steps // 2):
        te_out = block(
            te_inp_hidden_states,
840
            None,
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
        )
        loss = te_out.sum()
        loss.backward()

    torch.cuda.synchronize()

    if os.path.exists(path):
        os.remove(path)

    outputs = [te_out, te_inp_hidden_states.grad]
    for p in block.parameters():
        if p.requires_grad:
            outputs.append(p.grad)
    return outputs


@pytest.mark.parametrize("dtype", param_types)
@pytest.mark.parametrize("bs", batch_sizes)
859
@pytest.mark.parametrize("model", ["126m"])
860
861
def test_gpt_checkpointing(dtype, bs, model):
    config = model_configs[model]
862
    if not is_fused_attn_available(config, dtype, deterministic=True):
863
        pytest.skip("No attention backend available.")
864
    outputs = _test_e2e_checkpointing(bs, dtype, config, checkpoint=False)
865
    outputs_checkpoint = _test_e2e_checkpointing(bs, dtype, config, checkpoint=True)
866
867
868
869
870
871
872
873
874
875
876
877

    # Check that results match
    tols = dtype_tols(dtype)
    if dtype in (torch.float16, torch.bfloat16):
        tols.update(dict(rtol=2e-2, atol=2e-3))
    for i, (ref, test) in enumerate(zip(outputs, outputs_checkpoint)):
        torch.testing.assert_close(
            test,
            ref,
            msg=f"Mismatch in tensor {i}",
            **tols,
        )
878
879
880
881
882
883


def _test_e2e_gpt_accuracy(block, bs, dtype, config):
    reset_rng_states()

    inp_hidden_states = torch.randn(
884
        (config.max_seqlen_q, bs, config.hidden_size),
885
886
887
888
        dtype=dtype,
        device="cuda",
        requires_grad=True,
    )
889
    inp_hidden_states.retain_grad()
890
    inp_attn_mask = get_causal_attn_mask(config.max_seqlen_q)
891

892
    out = block(inp_hidden_states, attention_mask=inp_attn_mask)
893
894
895
896
897
898
899
900
901
902
903
904
905
    loss = out.sum()
    loss.backward()

    torch.cuda.synchronize()
    outputs = [out, inp_hidden_states.grad]
    for p in block.parameters():
        if p.requires_grad:
            outputs.append(p.grad)
    return outputs


@pytest.mark.parametrize("dtype", param_types)
@pytest.mark.parametrize("bs", batch_sizes)
906
@pytest.mark.parametrize("model", ["small"])
907
908
@pytest.mark.parametrize("parallel_attention_mlp", all_boolean)
def test_gpt_accuracy(dtype, bs, model, parallel_attention_mlp):
909
    config = model_configs[model]
910
911
912
    if not is_fused_attn_available(
        config, dtype, qkv_layout="sb3hd", is_training=True, deterministic=True
    ):
913
        pytest.skip("No attention backend available.")
914

915
916
917
    te_gpt = TransformerLayer(
        hidden_size=config.hidden_size,
        ffn_hidden_size=4 * config.hidden_size,
918
        num_attention_heads=config.num_heads,
919
920
921
922
923
924
925
926
927
        layernorm_epsilon=config.eps,
        attention_dropout=0.1,
        hidden_dropout=0.1,
        params_dtype=dtype,
        fuse_qkv_params=True,
        qkv_weight_interleaved=False,
        parallel_attention_mlp=parallel_attention_mlp,
        device="cuda",
    ).eval()
928
929
930
931
932

    torch_gpt = (
        TorchGPT(
            config.hidden_size,
            config.eps,
933
            config.num_heads,
934
            parallel_attention_mlp=parallel_attention_mlp,
935
936
937
938
939
940
941
942
        )
        .to(dtype=dtype)
        .cuda()
        .eval()
    )

    # Share params
    with torch.no_grad():
943
        torch_gpt.ln.weight = Parameter(
944
945
            te_gpt.self_attention.layernorm_qkv.layer_norm_weight.clone()
        )
946
        torch_gpt.ln.bias = Parameter(te_gpt.self_attention.layernorm_qkv.layer_norm_bias.clone())
947
948
949
950
951
952
953
954
955
956
957
958
        torch_gpt.causal_attn.mhsa.in_proj_weight = Parameter(
            te_gpt.self_attention.layernorm_qkv.weight.clone()
        )
        torch_gpt.causal_attn.mhsa.in_proj_bias = Parameter(
            te_gpt.self_attention.layernorm_qkv.bias.clone()
        )
        torch_gpt.causal_attn.mhsa.out_proj.weight = Parameter(
            te_gpt.self_attention.proj.weight.clone()
        )
        torch_gpt.causal_attn.mhsa.out_proj.bias = Parameter(
            te_gpt.self_attention.proj.bias.clone()
        )
959
960
961
962
963
964
        torch_gpt.ln_mlp.ln.weight = Parameter(te_gpt.layernorm_mlp.layer_norm_weight.clone())
        torch_gpt.ln_mlp.ln.bias = Parameter(te_gpt.layernorm_mlp.layer_norm_bias.clone())
        torch_gpt.ln_mlp.fc1.weight = Parameter(te_gpt.layernorm_mlp.fc1_weight.clone())
        torch_gpt.ln_mlp.fc1.bias = Parameter(te_gpt.layernorm_mlp.fc1_bias.clone())
        torch_gpt.ln_mlp.fc2.weight = Parameter(te_gpt.layernorm_mlp.fc2_weight.clone())
        torch_gpt.ln_mlp.fc2.bias = Parameter(te_gpt.layernorm_mlp.fc2_bias.clone())
965
966
967
968

    te_outputs = _test_e2e_gpt_accuracy(te_gpt, bs, dtype, config)
    torch_outputs = _test_e2e_gpt_accuracy(torch_gpt, bs, dtype, config)

969
970
971
972
973
974
    atol = {
        torch.float32: 5e-3,
        torch.half: 5e-2,
        torch.bfloat16: 1e-1,
    }

975
    # Check output.
976
977
978
979
980
981
982
983
984
985
986
987
    assert_allclose(te_outputs[0], torch_outputs[0], atol[dtype])

    # Check gradients, only for small model
    if model == "small":
        atol[torch.float32] = 5e-2
        rtol = {
            torch.float32: 1e-2,
            torch.half: 1e-2,
            torch.bfloat16: 1e-2,
        }
        for te_output, torch_output in zip(te_outputs[1:], torch_outputs[1:]):
            assert_allclose(te_output, torch_output, atol[dtype], rtol[dtype])
988
989


990
def _test_mha_accuracy(block, bs, dtype, config, mask_type, te=True):
991
992
993
    reset_rng_states()

    inp_hidden_states = torch.randn(
994
        (config.max_seqlen_q, bs, config.hidden_size),
995
996
997
998
        dtype=dtype,
        device="cuda",
        requires_grad=True,
    )
999
    inp_hidden_states.retain_grad()
1000
    inp_attn_mask = get_causal_attn_mask(config.max_seqlen_q) if mask_type == "causal" else None
1001

1002
1003
1004
1005
1006
1007
    forward_kwargs = {}
    if te:
        forward_kwargs["attn_mask_type"] = mask_type
    forward_kwargs["attention_mask"] = inp_attn_mask

    out = block(inp_hidden_states, **forward_kwargs)
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
    loss = out.sum()
    loss.backward()

    torch.cuda.synchronize()
    outputs = [out, inp_hidden_states.grad]
    for p in block.parameters():
        if p.requires_grad:
            outputs.append(p.grad)
    return outputs


@pytest.mark.parametrize("dtype", param_types)
@pytest.mark.parametrize("bs", batch_sizes)
1021
@pytest.mark.parametrize("model", ["small"])
1022
1023
1024
@pytest.mark.parametrize("mask_type", mask_types)
def test_mha_accuracy(dtype, bs, model, mask_type):
    config = model_configs[model]
1025
1026
1027
    if not is_fused_attn_available(
        config, dtype, qkv_layout="sb3hd", is_training=True, deterministic=True
    ):
1028
        pytest.skip("No attention backend available.")
1029

1030
1031
    te_mha = MultiheadAttention(
        config.hidden_size,
1032
        config.num_heads,
1033
1034
1035
1036
1037
1038
        fuse_qkv_params=True,
        params_dtype=dtype,
        qkv_weight_interleaved=False,
        input_layernorm=False,
        device="cuda",
    ).eval()
1039
1040
1041
1042

    torch_mha = (
        TorchMHA(
            config.hidden_size,
1043
            config.num_heads,
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
        )
        .to(dtype=dtype)
        .cuda()
        .eval()
    )

    # Share params
    with torch.no_grad():
        torch_mha.mhsa.in_proj_weight = Parameter(te_mha.qkv.weight.clone())
        torch_mha.mhsa.in_proj_bias = Parameter(te_mha.qkv.bias.clone())
        torch_mha.mhsa.out_proj.weight = Parameter(te_mha.proj.weight.clone())
        torch_mha.mhsa.out_proj.bias = Parameter(te_mha.proj.bias.clone())

1057
1058
    te_outputs = _test_mha_accuracy(te_mha, bs, dtype, config, mask_type, te=True)
    torch_outputs = _test_mha_accuracy(torch_mha, bs, dtype, config, mask_type, te=False)
1059
1060
1061
1062
1063
1064
1065

    # Check output.
    if dtype == torch.float32:
        assert_allclose(te_outputs[0], torch_outputs[0], 5e-3)
    else:
        assert_allclose(te_outputs[0], torch_outputs[0], 5e-2)

1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
    # Check gradients, only for small model
    if model == "small":
        atol = {
            torch.float32: 5e-2,
            torch.half: 5e-2,
            torch.bfloat16: 5e-2,
        }
        rtol = {
            torch.float32: 1e-2,
            torch.half: 1e-2,
            torch.bfloat16: 1e-2,
        }
        for te_output, torch_output in zip(te_outputs[1:], torch_outputs[1:]):
            assert_allclose(te_output, torch_output, atol[dtype], rtol[dtype])

1081

1082
def _test_granular_accuracy(block, bs, dtype, config, delay_wgrad_compute=False, recipe=None):
1083
    reset_rng_states()
1084
1085
1086
    fp8 = recipe is not None
    if fp8:
        FP8GlobalStateManager.reset()
1087
1088

    inp_hidden_states = torch.randn(
1089
        (config.max_seqlen_q, bs, config.hidden_size),
1090
1091
1092
1093
        dtype=dtype,
        device="cuda",
        requires_grad=True,
    )
1094
1095
    inp_hidden_states.retain_grad()

1096
1097
1098
1099
    with fp8_autocast(enabled=fp8, fp8_recipe=recipe):
        out = block(inp_hidden_states)
        if isinstance(out, (List, Tuple)):
            out = out[0]
1100
1101
    loss = out.sum()
    loss.backward()
1102
1103
    if delay_wgrad_compute:
        block.backward_dw()
1104
1105
1106
1107
1108

    torch.cuda.synchronize()
    outputs = [out, inp_hidden_states.grad]
    for p in block.parameters():
        if p.requires_grad:
1109
1110
1111
1112
1113
            if getattr(p, "main_grad", None) is not None:
                outputs.append(p.main_grad)
                assert p.grad is None  # grad should be None if fuse_wgrad_accumulation is True
            else:
                outputs.append(p.grad)
1114
1115
1116
    return outputs


1117
1118
1119
def _test_dpa_accuracy(block, bs, dtype, config):
    reset_rng_states()

1120
    mask = torch.triu(
1121
1122
        torch.ones(config.max_seqlen_q, config.max_seqlen_kv, dtype=torch.bool, device="cuda"),
        diagonal=1,
1123
    )
1124
    query, key, value = [
1125
        torch.randn(
1126
            (config.max_seqlen_q, bs, config.num_heads, config.kv_channels),
1127
1128
1129
1130
1131
1132
            dtype=dtype,
            device="cuda",
            requires_grad=True,
        )
        for _ in range(3)
    ]
1133
1134
1135
1136
1137

    query.retain_grad()
    key.retain_grad()
    value.retain_grad()

1138
    out = block(query, key, value, attention_mask=mask)
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
    loss = out.sum()
    loss.backward()

    torch.cuda.synchronize()

    return [out, query.grad, key.grad, value.grad]


@pytest.mark.parametrize("dtype", param_types)
@pytest.mark.parametrize("bs", batch_sizes)
1149
@pytest.mark.parametrize("model", ["126m"])
1150
1151
1152
1153
1154
def test_dpa_accuracy(dtype, bs, model):
    config = model_configs[model]

    te_dpa = (
        DotProductAttention(
1155
1156
            config.num_heads,
            config.kv_channels,
1157
            attention_dropout=0.0,  # disable dropout, FU uses rng differently
1158
1159
1160
1161
1162
1163
1164
        )
        .to(dtype=dtype)
        .cuda()
    )

    torch_dpa = (
        TorchDotProductAttention(
1165
            config.kv_channels,
1166
            0.0,  # dropout
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
        )
        .to(dtype=dtype)
        .cuda()
    )

    te_outputs = _test_dpa_accuracy(te_dpa, bs, dtype, config)
    torch_outputs = _test_dpa_accuracy(torch_dpa, bs, dtype, config)

    # Check output.
    if dtype == torch.float32:
        assert_allclose(te_outputs[0], torch_outputs[0], 5e-3)
    else:
        assert_allclose(te_outputs[0], torch_outputs[0], 5e-2)

1181
1182
1183
    for te_output, torch_output in zip(te_outputs[1:], torch_outputs[1:]):
        assert_allclose(te_output, torch_output, atol=5e-2, rtol=1e-2)

1184

1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
class TestReturnBiasModule(nn.Module):
    def __init__(self, mod, **kwargs):
        super().__init__()
        self.te_module = mod(**kwargs)
        self.return_bias = kwargs["return_bias"]
        self.bias = kwargs["bias"]

    def forward(self, x):
        if self.return_bias:
            out, bias = self.te_module(x)
            if self.bias:
                out = out + bias
            return out
        return self.te_module(x)


1201
1202
@pytest.mark.parametrize("dtype", param_types)
@pytest.mark.parametrize("bs", batch_sizes)
1203
@pytest.mark.parametrize("model", ["small"])
1204
1205
1206
@pytest.mark.parametrize("return_bias", all_boolean)
@pytest.mark.parametrize("bias", all_boolean)
def test_linear_accuracy(dtype, bs, model, return_bias, bias):
1207
1208
    config = model_configs[model]

1209
1210
1211
1212
    te_linear = TestReturnBiasModule(
        Linear,
        in_features=config.hidden_size,
        out_features=4 * config.hidden_size,
1213
        params_dtype=dtype,
1214
1215
        return_bias=return_bias,
        bias=bias,
1216
        device="cuda",
1217
    )
1218

1219
1220
1221
    torch_linear = torch.nn.Linear(
        config.hidden_size,
        4 * config.hidden_size,
1222
        bias=bias,
1223
1224
        device="cuda",
        dtype=dtype,
1225
    )
1226
1227
1228

    # Share params
    with torch.no_grad():
1229
1230
1231
        torch_linear.weight = Parameter(te_linear.te_module.weight.clone())
        if bias:
            torch_linear.bias = Parameter(te_linear.te_module.bias.clone())
1232
1233
1234
1235
1236

    te_outputs = _test_granular_accuracy(te_linear, bs, dtype, config)
    torch_outputs = _test_granular_accuracy(torch_linear, bs, dtype, config)

    # Check output.
1237
1238
1239
1240
1241
1242
1243
1244
1245
    if model == "small":
        tolerance = 5e-3 if dtype == torch.float32 else 5e-2
        rtol = {
            torch.float32: 1.3e-6,
            torch.half: 1e-2,
            torch.bfloat16: 2e-2,
        }
        for te_output, torch_output in zip(te_outputs, torch_outputs):
            assert_allclose(te_output, torch_output, tolerance, rtol[dtype])
1246

1247

1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
@pytest.mark.parametrize("dtype", param_types)
@pytest.mark.parametrize("bs", batch_sizes)
@pytest.mark.parametrize("model", ["small"])
@pytest.mark.parametrize("bias", all_boolean)
@pytest.mark.parametrize("fuse_wgrad_accumulation", all_boolean)
def test_linear_accuracy_delay_wgrad_compute(dtype, bs, model, bias, fuse_wgrad_accumulation):
    config = model_configs[model]

    te_linear_ref = Linear(
        config.hidden_size,
        4 * config.hidden_size,
        bias=bias,
        params_dtype=dtype,
        device="cuda",
        delay_wgrad_compute=False,
        fuse_wgrad_accumulation=fuse_wgrad_accumulation,
    ).eval()

    te_linear = Linear(
        config.hidden_size,
        4 * config.hidden_size,
        bias=bias,
        params_dtype=dtype,
        device="cuda",
        delay_wgrad_compute=True,
        fuse_wgrad_accumulation=fuse_wgrad_accumulation,
    ).eval()

    # Share params
    with torch.no_grad():
        te_linear_ref.weight = Parameter(te_linear.weight.clone())
        if bias:
            te_linear_ref.bias = Parameter(te_linear.bias.clone())
        if fuse_wgrad_accumulation:
            weight = getattr(te_linear, f"weight")
            weight.main_grad = torch.rand_like(weight, dtype=torch.float32)
            te_linear_ref.weight.main_grad = weight.main_grad.clone()

    te_outputs = _test_granular_accuracy(te_linear, bs, dtype, config, delay_wgrad_compute=True)
    te_outputs_ref = _test_granular_accuracy(
        te_linear_ref, bs, dtype, config, delay_wgrad_compute=False
    )

1291
1292
    # Should be bit-wise match
    for _, (o, o_ref) in enumerate(zip(te_outputs, te_outputs_ref)):
1293
1294
1295
        torch.testing.assert_close(o, o_ref, rtol=0, atol=0)


1296
1297
1298
1299
1300
1301
1302
1303
@pytest.mark.parametrize("dtype", param_types)
@pytest.mark.parametrize("model", ["small"])
@pytest.mark.parametrize("recipe", fp8_recipes + [None])
def test_linear_accuracy_save_original_input(dtype, model, recipe):
    bs = 1
    fuse_wgrad_accumulation = True
    fp8_model_params = False
    fp8 = recipe is not None
1304

1305
1306
1307
1308
    if fp8 and recipe.delayed():
        pytest.skip("DelayedScaling recipe is not supported with save_original_input")

    config = model_configs[model]
1309
    if config.max_seqlen_q % 16 != 0 and fp8:
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
        pytest.skip("FP8 requires sequence length to be divisible by 16.")

    with fp8_model_init(enabled=fp8 and fp8_model_params, recipe=recipe):
        te_linear_ref = Linear(
            config.hidden_size,
            4 * config.hidden_size,
            bias=False,
            params_dtype=dtype,
            device="cuda",
            fuse_wgrad_accumulation=fuse_wgrad_accumulation,
            save_original_input=False,
        ).eval()

        te_linear = Linear(
            config.hidden_size,
            4 * config.hidden_size,
            bias=False,
            params_dtype=dtype,
            device="cuda",
            fuse_wgrad_accumulation=fuse_wgrad_accumulation,
            save_original_input=True,
        ).eval()

    # Share params
    with torch.no_grad():
        te_linear_ref.weight = Parameter(te_linear.weight.clone())
        if fuse_wgrad_accumulation:
            weight = getattr(te_linear, f"weight")
            weight.main_grad = torch.rand_like(weight, dtype=torch.float32)
            te_linear_ref.weight.main_grad = weight.main_grad.clone()

    te_outputs = _test_granular_accuracy(te_linear, bs, dtype, config, recipe=recipe)
    te_outputs_ref = _test_granular_accuracy(te_linear_ref, bs, dtype, config, recipe=recipe)

    # Shoule be bit-wise match
    for i, (o, o_ref) in enumerate(zip(te_outputs, te_outputs_ref)):
        torch.testing.assert_close(o, o_ref, rtol=0, atol=0)


1349
1350
@pytest.mark.parametrize("dtype", param_types)
@pytest.mark.parametrize("bs", batch_sizes)
1351
@pytest.mark.parametrize("model", ["126m"])
1352
@pytest.mark.parametrize("eps", [1e-1, 1e-3, 1e-5, 1e-7])
1353
1354
@pytest.mark.parametrize("zero_centered_gamma", all_boolean)
def test_rmsnorm_accuracy(dtype, bs, model, eps, zero_centered_gamma):
1355
1356
    config = model_configs[model]

1357
1358
1359
1360
1361
1362
1363
    te_rmsnorm = RMSNorm(
        config.hidden_size,
        eps=eps,
        params_dtype=dtype,
        zero_centered_gamma=zero_centered_gamma,
        device="cuda",
    ).eval()
1364
1365

    torch_rmsnorm = (
1366
        TorchRMSNorm(config.hidden_size, eps=eps, zero_centered_gamma=zero_centered_gamma)
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
        .to(dtype=dtype)
        .cuda()
        .eval()
    )

    # Share params
    with torch.no_grad():
        torch_rmsnorm.weight = Parameter(te_rmsnorm.weight.clone())

    te_outputs = _test_granular_accuracy(te_rmsnorm, bs, dtype, config)
    torch_outputs = _test_granular_accuracy(torch_rmsnorm, bs, dtype, config)

1379
1380
1381
1382
    atol = {
        torch.float32: 1e-7,
        torch.half: 2e-3,
        torch.bfloat16: 2e-2,
1383
    }
1384
1385

    # Check output.
1386
1387
    assert_allclose(te_outputs[0], torch_outputs[0], atol[dtype])

1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
    atol[torch.float32] = 2e-3
    rtol = {
        torch.float32: 1.3e-6,
        torch.half: 1e-3,
        torch.bfloat16: 1.6e-2,
    }
    # Check gradients
    for te_output, torch_output in zip(te_outputs[1:], torch_outputs[1:]):
        assert_allclose(te_output, torch_output, atol[dtype], rtol[dtype])

1398

1399
1400
@pytest.mark.parametrize("dtype", param_types)
@pytest.mark.parametrize("bs", batch_sizes)
1401
@pytest.mark.parametrize("model", ["126m"])
1402
1403
1404
1405
1406
@pytest.mark.parametrize("eps", [1e-1, 1e-3, 1e-5, 1e-7])
@pytest.mark.parametrize("zero_centered_gamma", all_boolean)
def test_layernorm_accuracy(dtype, bs, model, eps, zero_centered_gamma):
    config = model_configs[model]

1407
1408
1409
1410
1411
1412
1413
    te_layernorm = LayerNorm(
        config.hidden_size,
        eps=eps,
        params_dtype=dtype,
        zero_centered_gamma=zero_centered_gamma,
        device="cuda",
    ).eval()
1414
1415

    torch_layernorm = (
1416
        TorchLayerNorm(config.hidden_size, eps=eps, zero_centered_gamma=zero_centered_gamma)
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
        .to(dtype=dtype)
        .cuda()
        .eval()
    )

    # Share params
    with torch.no_grad():
        torch_layernorm.weight = Parameter(te_layernorm.weight.clone())
        torch_layernorm.bias = Parameter(te_layernorm.bias.clone())

    te_outputs = _test_granular_accuracy(te_layernorm, bs, dtype, config)
    torch_outputs = _test_granular_accuracy(torch_layernorm, bs, dtype, config)

1430
1431
1432
1433
    atol = {
        torch.float32: 1e-7,
        torch.half: 2e-3,
        torch.bfloat16: 2e-2,
1434
    }
1435
1436

    # Check output.
1437
    assert_allclose(te_outputs[0], torch_outputs[0], atol[dtype])
1438

1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
    rtol = {
        torch.float32: 1.3e-6,
        torch.half: 1e-3,
        torch.bfloat16: 1.6e-2,
    }
    atol[torch.float32] = 1e-4
    # Check gradients
    for te_output, torch_output in zip(te_outputs[1:], torch_outputs[1:]):
        assert_allclose(te_output, torch_output, atol[dtype], rtol[dtype])

1449

1450
1451
@pytest.mark.parametrize("dtype", param_types)
@pytest.mark.parametrize("bs", batch_sizes)
1452
@pytest.mark.parametrize("model", ["small"])
1453
@pytest.mark.parametrize("normalization", all_normalizations)
1454
@pytest.mark.parametrize("zero_centered_gamma", all_boolean)
1455
1456
1457
1458
1459
@pytest.mark.parametrize("return_bias", all_boolean)
@pytest.mark.parametrize("bias", all_boolean)
def test_layernorm_linear_accuracy(
    dtype, bs, model, normalization, zero_centered_gamma, return_bias, bias
):
1460
1461
    config = model_configs[model]

1462
1463
1464
1465
1466
    te_ln_linear = TestReturnBiasModule(
        LayerNormLinear,
        in_features=config.hidden_size,
        out_features=4 * config.hidden_size,
        eps=config.eps,
1467
1468
1469
        normalization=normalization,
        params_dtype=dtype,
        zero_centered_gamma=zero_centered_gamma,
1470
1471
        return_bias=return_bias,
        bias=bias,
1472
        device="cuda",
1473
    )
1474
1475
1476
1477
1478
1479

    torch_ln_linear = (
        TorchLayerNormLinear(
            config.hidden_size,
            4 * config.hidden_size,
            config.eps,
1480
            normalization=normalization,
1481
            zero_centered_gamma=zero_centered_gamma,
1482
            bias=bias,
1483
1484
1485
1486
1487
1488
1489
        )
        .to(dtype=dtype)
        .cuda()
    )

    # Share params
    with torch.no_grad():
1490
1491
1492
        torch_ln_linear.layernorm.weight = Parameter(
            te_ln_linear.te_module.layer_norm_weight.clone()
        )
1493
        if normalization != "RMSNorm":
1494
1495
1496
1497
1498
1499
            torch_ln_linear.layernorm.bias = Parameter(
                te_ln_linear.te_module.layer_norm_bias.clone()
            )
        torch_ln_linear.linear.weight = Parameter(te_ln_linear.te_module.weight.clone())
        if bias:
            torch_ln_linear.linear.bias = Parameter(te_ln_linear.te_module.bias.clone())
1500
1501
1502
1503

    te_outputs = _test_granular_accuracy(te_ln_linear, bs, dtype, config)
    torch_outputs = _test_granular_accuracy(torch_ln_linear, bs, dtype, config)

1504
1505
1506
1507
    atol = {
        torch.float32: 2.5e-4,
        torch.half: 2e-3,
        torch.bfloat16: 2e-2,
1508
    }
1509
1510
1511
1512
1513
    rtol = {
        torch.float32: 1e-3,
        torch.half: 4e-2,
        torch.bfloat16: 4e-2,
    }
1514
1515

    # Check output.
1516
    assert_allclose(te_outputs[0], torch_outputs[0], atol[dtype], rtol[dtype])
1517

1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
    if model == "small":
        atol = {
            torch.float32: 1e-3,
            torch.half: 5e-2,
            torch.bfloat16: 5e-2,
        }
        rtol = {
            torch.float32: 1e-3,
            torch.half: 4e-2,
            torch.bfloat16: 4e-2,
        }
        # Check gradients
        for te_output, torch_output in zip(te_outputs[1:], torch_outputs[1:]):
            assert_allclose(te_output, torch_output, atol[dtype], rtol[dtype])

1533

1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
@pytest.mark.parametrize("dtype", param_types)
@pytest.mark.parametrize("bs", batch_sizes)
@pytest.mark.parametrize("model", ["small"])
@pytest.mark.parametrize("normalization", all_normalizations)
@pytest.mark.parametrize("zero_centered_gamma", all_boolean)
@pytest.mark.parametrize("bias", all_boolean)
@pytest.mark.parametrize("fuse_wgrad_accumulation", all_boolean)
def test_layernorm_linear_accuracy_delay_wgrad_compute(
    dtype, bs, model, normalization, zero_centered_gamma, bias, fuse_wgrad_accumulation
):
    config = model_configs[model]

    ln_linear_ref = LayerNormLinear(
        config.hidden_size,
        4 * config.hidden_size,
        config.eps,
        bias=bias,
        normalization=normalization,
        params_dtype=dtype,
        zero_centered_gamma=zero_centered_gamma,
        device="cuda",
        delay_wgrad_compute=False,
        fuse_wgrad_accumulation=fuse_wgrad_accumulation,
    ).eval()

    ln_linear = LayerNormLinear(
        config.hidden_size,
        4 * config.hidden_size,
        config.eps,
        bias=bias,
        normalization=normalization,
        params_dtype=dtype,
        zero_centered_gamma=zero_centered_gamma,
        device="cuda",
        delay_wgrad_compute=True,
        fuse_wgrad_accumulation=fuse_wgrad_accumulation,
    ).eval()

    # Share params
    with torch.no_grad():
        ln_linear_ref.layer_norm_weight = Parameter(ln_linear.layer_norm_weight.clone())
        if normalization != "RMSNorm":
            ln_linear_ref.layer_norm_bias = Parameter(ln_linear.layer_norm_bias.clone())
        ln_linear_ref.weight = Parameter(ln_linear.weight.clone())
        if bias:
            ln_linear_ref.bias = Parameter(ln_linear.bias.clone())
        if fuse_wgrad_accumulation:
            weight = getattr(ln_linear, f"weight")
            weight.main_grad = torch.rand_like(weight, dtype=torch.float32)
            ln_linear_ref.weight.main_grad = weight.main_grad.clone()

    te_outputs = _test_granular_accuracy(ln_linear, bs, dtype, config, delay_wgrad_compute=True)
    te_outputs_ref = _test_granular_accuracy(
        ln_linear_ref, bs, dtype, config, delay_wgrad_compute=False
    )

    # Shoule be bit-wise match
    for i, (o, o_ref) in enumerate(zip(te_outputs, te_outputs_ref)):
        torch.testing.assert_close(o, o_ref, rtol=0, atol=0)


1595
1596
@pytest.mark.parametrize("dtype", param_types)
@pytest.mark.parametrize("bs", batch_sizes)
1597
@pytest.mark.parametrize("model", ["small"])
1598
@pytest.mark.parametrize("activation", all_activations)
1599
@pytest.mark.parametrize("normalization", all_normalizations)
1600
1601
1602
@pytest.mark.parametrize("return_bias", all_boolean)
@pytest.mark.parametrize("bias", all_boolean)
def test_layernorm_mlp_accuracy(dtype, bs, model, activation, normalization, return_bias, bias):
1603
1604
    config = model_configs[model]

1605
1606
1607
1608
    te_ln_mlp = TestReturnBiasModule(
        LayerNormMLP,
        hidden_size=config.hidden_size,
        ffn_hidden_size=4 * config.hidden_size,
1609
1610
1611
        activation=activation,
        normalization=normalization,
        params_dtype=dtype,
1612
1613
        return_bias=return_bias,
        bias=bias,
1614
        device="cuda",
1615
    )
1616
1617
1618
1619
1620

    torch_ln_mlp = (
        TorchLayerNormMLP(
            config.hidden_size,
            4 * config.hidden_size,
1621
            activation=activation,
1622
            normalization=normalization,
1623
            bias=bias,
1624
1625
1626
1627
1628
1629
1630
        )
        .to(dtype=dtype)
        .cuda()
    )

    # Share params
    with torch.no_grad():
1631
        torch_ln_mlp.ln.weight = Parameter(te_ln_mlp.te_module.layer_norm_weight.clone())
1632
        if normalization != "RMSNorm":
1633
1634
1635
1636
1637
1638
            torch_ln_mlp.ln.bias = Parameter(te_ln_mlp.te_module.layer_norm_bias.clone())
        torch_ln_mlp.fc1.weight = Parameter(te_ln_mlp.te_module.fc1_weight.clone())
        torch_ln_mlp.fc2.weight = Parameter(te_ln_mlp.te_module.fc2_weight.clone())
        if bias:
            torch_ln_mlp.fc1.bias = Parameter(te_ln_mlp.te_module.fc1_bias.clone())
            torch_ln_mlp.fc2.bias = Parameter(te_ln_mlp.te_module.fc2_bias.clone())
1639
1640
1641
1642

    te_outputs = _test_granular_accuracy(te_ln_mlp, bs, dtype, config)
    torch_outputs = _test_granular_accuracy(torch_ln_mlp, bs, dtype, config)

1643
1644
1645
1646
1647
1648
    atol = {
        torch.float32: 2e-2,
        torch.half: 5e-2,
        torch.bfloat16: 5e-2,
    }

1649
1650
1651
1652
1653
1654
    rtol = {
        torch.float32: 1e-3,
        torch.half: 4e-2,
        torch.bfloat16: 4e-2,
    }

1655
    # Check output.
1656
    assert_allclose(te_outputs[0], torch_outputs[0], atol[dtype], rtol[dtype])
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668

    # Check gradients, only for small model
    rtol = {
        torch.float32: 1e-3,
        torch.half: 1e-2,
        torch.bfloat16: 4e-2,
    }
    atol[torch.half] = 2e-1
    atol[torch.bfloat16] = 2e-1
    if model == "small":
        for te_output, torch_output in zip(te_outputs[1:], torch_outputs[1:]):
            assert_allclose(te_output, torch_output, atol[dtype], rtol[dtype])
1669
1670


1671
@pytest.mark.parametrize("dtype", param_types)
1672
@pytest.mark.parametrize("bs", [2])
1673
1674
1675
1676
@pytest.mark.parametrize("model", ["small"])
@pytest.mark.parametrize("bias", all_boolean)
@pytest.mark.parametrize("fuse_wgrad_accumulation", all_boolean)
def test_layernorm_mlp_accuracy_delay_wgrad_compute(
1677
    dtype, bs, model, bias, fuse_wgrad_accumulation
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
):
    config = model_configs[model]

    ln_mlp = LayerNormMLP(
        hidden_size=config.hidden_size,
        ffn_hidden_size=4 * config.hidden_size,
        eps=config.eps,
        bias=bias,
        params_dtype=dtype,
        device="cuda",
        delay_wgrad_compute=True,
        fuse_wgrad_accumulation=fuse_wgrad_accumulation,
    ).eval()

    ln_mlp_ref = LayerNormMLP(
        hidden_size=config.hidden_size,
        ffn_hidden_size=4 * config.hidden_size,
        eps=config.eps,
        bias=bias,
        params_dtype=dtype,
        device="cuda",
        delay_wgrad_compute=False,
        fuse_wgrad_accumulation=fuse_wgrad_accumulation,
    ).eval()

    # Share params
    with torch.no_grad():
        ln_mlp_ref.layer_norm_weight = Parameter(ln_mlp.layer_norm_weight.clone())
1706
        ln_mlp_ref.layer_norm_bias = Parameter(ln_mlp.layer_norm_bias.clone())
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
        ln_mlp_ref.fc1_weight = Parameter(ln_mlp.fc1_weight.clone())
        ln_mlp_ref.fc2_weight = Parameter(ln_mlp.fc2_weight.clone())
        if bias:
            ln_mlp_ref.fc1_bias = Parameter(ln_mlp.fc1_bias.clone())
            ln_mlp_ref.fc2_bias = Parameter(ln_mlp.fc2_bias.clone())
        if fuse_wgrad_accumulation:
            ln_mlp.fc1_weight.main_grad = torch.rand_like(ln_mlp.fc1_weight, dtype=torch.float32)
            ln_mlp_ref.fc1_weight.main_grad = ln_mlp.fc1_weight.main_grad.clone()
            ln_mlp.fc2_weight.main_grad = torch.rand_like(ln_mlp.fc2_weight, dtype=torch.float32)
            ln_mlp_ref.fc2_weight.main_grad = ln_mlp.fc2_weight.main_grad.clone()

    te_outputs = _test_granular_accuracy(ln_mlp, bs, dtype, config, delay_wgrad_compute=True)
    te_outputs_ref = _test_granular_accuracy(
        ln_mlp_ref, bs, dtype, config, delay_wgrad_compute=False
    )

    # Shoule be bit-wise match
    for i, (o, o_ref) in enumerate(zip(te_outputs, te_outputs_ref)):
        torch.testing.assert_close(o, o_ref, rtol=0, atol=0)


1728
def _test_grouped_linear_accuracy(
1729
1730
1731
1732
1733
1734
1735
1736
1737
    block,
    num_gemms,
    bs,
    dtype,
    config,
    recipe,
    fp8,
    fuse_wgrad_accumulation,
    delay_wgrad_compute=False,
1738
):
1739
1740
1741
1742
1743
    reset_rng_states()
    if fp8:
        FP8GlobalStateManager.reset()

    inp_hidden_states = torch.randn(
1744
        (config.max_seqlen_q, bs, config.hidden_size),
1745
1746
1747
1748
1749
1750
        dtype=dtype,
        device="cuda",
        requires_grad=True,
    )
    inp_hidden_states.retain_grad()

1751
    if num_gemms > 1:
1752
1753
        split_size = 1
        if fp8:
1754
            split_size = 16
1755
1756
            if recipe.mxfp8():
                split_size = 128
1757
        m = config.max_seqlen_q // split_size
1758
1759
1760
        dist = torch.sort(torch.randint(0, m, (num_gemms - 2,))).values.tolist()
        dist.append(dist[-1])  # Manually add a zero
        m_splits = torch.tensor(dist + [m]) - torch.tensor([0] + dist)
1761
        m_splits = m_splits * split_size
1762
        assert m_splits.sum() == config.max_seqlen_q and len(m_splits) == num_gemms
1763
    else:
1764
        m_splits = torch.tensor([config.max_seqlen_q])
1765

1766
    with fp8_autocast(enabled=fp8, fp8_recipe=recipe):
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
        if isinstance(block, GroupedLinear):
            m_splits = m_splits * bs
            out = block(inp_hidden_states, m_splits.tolist())
        else:
            out = torch.cat(
                [
                    block[i](inp)
                    for i, inp in enumerate(torch.split(inp_hidden_states, m_splits.tolist()))
                ]
            )
    loss = out.sum()
    loss.backward()
1779
1780
1781
1782
1783
1784
    if delay_wgrad_compute:
        if isinstance(block, GroupedLinear):
            block.backward_dw()
        else:
            for i in range(num_gemms):
                block[i].backward_dw()
1785
1786
1787
1788
1789

    torch.cuda.synchronize()
    outputs = [out, inp_hidden_states.grad]
    for p in block.parameters():
        if p.requires_grad:
1790
1791
1792
1793
1794
            if getattr(p, "main_grad", None) is not None:
                outputs.append(p.main_grad)
                assert p.grad is None  # grad should be None if fuse_wgrad_accumulation is True
            else:
                outputs.append(p.grad)
1795
1796
1797
    return outputs


1798
@pytest.mark.parametrize("dtype", param_types, ids=str)
1799
1800
@pytest.mark.parametrize("num_gemms", [3, 6])
@pytest.mark.parametrize("bs", batch_sizes)
1801
@pytest.mark.parametrize("model", ["126m"])
1802
@pytest.mark.parametrize("recipe", fp8_recipes + [None])
1803
@pytest.mark.parametrize("fp8_model_params", all_boolean)
1804
@pytest.mark.parametrize("fuse_wgrad_accumulation", all_boolean)
1805
1806
@pytest.mark.parametrize("bias", all_boolean)
@pytest.mark.parametrize("delay_wgrad_compute", all_boolean)
1807
def test_grouped_linear_accuracy(
1808
1809
1810
1811
1812
1813
1814
    dtype,
    num_gemms,
    bs,
    model,
    recipe,
    fp8_model_params,
    fuse_wgrad_accumulation,
1815
1816
    bias,
    delay_wgrad_compute,
1817
    parallel_mode=None,
1818
    use_cutlass=False,
1819
):
1820
    fp8 = recipe is not None
1821
    if fp8 and fp8_model_params and NVTE_TEST_NVINSPECT_ENABLED:
1822
        pytest.skip("FP8 parameters are not supported in debug mode.")
1823
1824

    config = model_configs[model]
1825
    if config.max_seqlen_q % 16 != 0 and fp8:
1826
1827
        pytest.skip("FP8 requires sequence length to be divisible by 16.")

1828
    with fp8_model_init(enabled=fp8 and fp8_model_params, recipe=recipe):
1829
1830
1831
1832
        grouped_linear = GroupedLinear(
            num_gemms,
            config.hidden_size,
            4 * config.hidden_size,
1833
            bias=bias,
1834
            params_dtype=dtype,
1835
            parallel_mode=parallel_mode,
1836
            device="cuda",
1837
            fuse_wgrad_accumulation=fuse_wgrad_accumulation,
1838
            delay_wgrad_compute=delay_wgrad_compute,
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
            save_original_input=False,
        ).eval()
        sequential_linear = torch.nn.ModuleList(
            [
                Linear(
                    config.hidden_size,
                    4 * config.hidden_size,
                    bias=bias,
                    params_dtype=dtype,
                    parallel_mode=parallel_mode,
                    device="cuda",
                    fuse_wgrad_accumulation=fuse_wgrad_accumulation,
                ).eval()
                for _ in range(num_gemms)
            ]
        )

    # Share params
    with torch.no_grad():
        for i in range(num_gemms):
            sequential_linear[i].weight = Parameter(getattr(grouped_linear, f"weight{i}").clone())
            if bias:
                sequential_linear[i].bias = Parameter(getattr(grouped_linear, f"bias{i}").clone())
            if fuse_wgrad_accumulation:
                weight_i = getattr(grouped_linear, f"weight{i}")
                weight_i.main_grad = torch.rand_like(weight_i, dtype=torch.float32)
                sequential_linear[i].weight.main_grad = weight_i.main_grad.clone()

    outputs_ref = _test_grouped_linear_accuracy(
        sequential_linear,
        num_gemms,
        bs,
        dtype,
        config,
        recipe,
        fp8,
        fuse_wgrad_accumulation,
        delay_wgrad_compute,
    )
    outputs = _test_grouped_linear_accuracy(
        grouped_linear,
        num_gemms,
        bs,
        dtype,
        config,
        recipe,
        fp8,
        fuse_wgrad_accumulation,
        delay_wgrad_compute,
    )

1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
    for o, o_ref in zip(outputs, outputs_ref):
        if use_cutlass:
            torch.testing.assert_close(o, o_ref, rtol=1e-3, atol=1e-3)
        else:
            # cuBLAS implementation should be bit-wise match
            torch.testing.assert_close(o, o_ref, rtol=0, atol=0)


@pytest.mark.skipif(
    torch.cuda.get_device_capability() != (9, 0),
    reason="Only enable CUTLASS grouped gemm on Hopper",
)
@pytest.mark.parametrize("dtype", param_types, ids=str)
@pytest.mark.parametrize("num_gemms", [3, 6])
@pytest.mark.parametrize("bs", batch_sizes)
@pytest.mark.parametrize("model", ["126m"])
@pytest.mark.parametrize("fuse_wgrad_accumulation", all_boolean)
@pytest.mark.parametrize("delay_wgrad_compute", all_boolean)
def test_grouped_linear_accuracy_cutlass(
    dtype,
    num_gemms,
    bs,
    model,
    fuse_wgrad_accumulation,
    delay_wgrad_compute,
):
    os.environ["NVTE_USE_CUTLASS_GROUPED_GEMM"] = "1"
    test_grouped_linear_accuracy(
        dtype,
        num_gemms,
        bs,
        model,
        None,
        False,
        fuse_wgrad_accumulation,
        False,
        delay_wgrad_compute,
        None,
        use_cutlass=True,
    )
    os.environ.pop("NVTE_USE_CUTLASS_GROUPED_GEMM", None)
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954


@pytest.mark.parametrize("dtype", param_types, ids=str)
@pytest.mark.parametrize("num_gemms", [3])
@pytest.mark.parametrize("bs", [1])
@pytest.mark.parametrize("model", ["126m"])
@pytest.mark.parametrize("recipe", fp8_recipes + [None])
@pytest.mark.parametrize("fp8_model_params", [False])
@pytest.mark.parametrize("fuse_wgrad_accumulation", [True])
@pytest.mark.parametrize("bias", [False])
@pytest.mark.parametrize("delay_wgrad_compute", [True])
def test_grouped_linear_accuracy_save_original_input(
    dtype,
    num_gemms,
    bs,
    model,
    recipe,
    fp8_model_params,
    fuse_wgrad_accumulation,
    bias,
    delay_wgrad_compute,
    parallel_mode=None,
):
    fp8 = recipe is not None
1955
    if fp8 and fp8_model_params and NVTE_TEST_NVINSPECT_ENABLED:
1956
1957
1958
1959
1960
        pytest.skip("FP8 parameters are not supported in debug mode.")
    if fp8 and recipe.delayed():
        pytest.skip("DelayedScaling recipe is not supported with save_original_input")

    config = model_configs[model]
1961
    if config.max_seqlen_q % 16 != 0 and fp8:
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
        pytest.skip("FP8 requires sequence length to be divisible by 16.")

    with fp8_model_init(enabled=fp8 and fp8_model_params, recipe=recipe):
        grouped_linear = GroupedLinear(
            num_gemms,
            config.hidden_size,
            4 * config.hidden_size,
            bias=bias,
            params_dtype=dtype,
            parallel_mode=parallel_mode,
            device="cuda",
            fuse_wgrad_accumulation=fuse_wgrad_accumulation,
            delay_wgrad_compute=delay_wgrad_compute,
            save_original_input=True,
1976
1977
1978
1979
1980
1981
        ).eval()
        sequential_linear = torch.nn.ModuleList(
            [
                Linear(
                    config.hidden_size,
                    4 * config.hidden_size,
1982
                    bias=bias,
1983
                    params_dtype=dtype,
1984
                    parallel_mode=parallel_mode,
1985
                    device="cuda",
1986
                    fuse_wgrad_accumulation=fuse_wgrad_accumulation,
1987
1988
1989
1990
1991
1992
1993
1994
1995
                ).eval()
                for _ in range(num_gemms)
            ]
        )

    # Share params
    with torch.no_grad():
        for i in range(num_gemms):
            sequential_linear[i].weight = Parameter(getattr(grouped_linear, f"weight{i}").clone())
1996
1997
            if bias:
                sequential_linear[i].bias = Parameter(getattr(grouped_linear, f"bias{i}").clone())
1998
1999
2000
2001
            if fuse_wgrad_accumulation:
                weight_i = getattr(grouped_linear, f"weight{i}")
                weight_i.main_grad = torch.rand_like(weight_i, dtype=torch.float32)
                sequential_linear[i].weight.main_grad = weight_i.main_grad.clone()
2002

2003
2004
2005
2006
    # Force the sequential_linear and grouped_linear to use hipblaslt rather than hipblas
    if IS_HIP_EXTENSION:
        os.environ["NVTE_FORCE_ROCM_GEMM"] = "1"
    
2007
    outputs_ref = _test_grouped_linear_accuracy(
2008
2009
2010
2011
2012
2013
2014
2015
2016
        sequential_linear,
        num_gemms,
        bs,
        dtype,
        config,
        recipe,
        fp8,
        fuse_wgrad_accumulation,
        delay_wgrad_compute,
2017
2018
    )
    outputs = _test_grouped_linear_accuracy(
2019
2020
2021
2022
2023
2024
2025
2026
2027
        grouped_linear,
        num_gemms,
        bs,
        dtype,
        config,
        recipe,
        fp8,
        fuse_wgrad_accumulation,
        delay_wgrad_compute,
2028
    )
2029
2030
    if IS_HIP_EXTENSION:
        os.environ["NVTE_FORCE_ROCM_GEMM"] = "0"
2031
2032
2033
2034
2035
2036

    # Shoule be bit-wise match
    for i, (o, o_ref) in enumerate(zip(outputs, outputs_ref)):
        torch.testing.assert_close(o, o_ref, rtol=0, atol=0)


2037
@pytest.mark.parametrize("recipe", fp8_recipes + [None])
2038
def test_grouped_linear_accuracy_single_gemm(recipe):
2039
2040
2041
2042
2043
    """Split the tests to save CI time"""
    test_grouped_linear_accuracy(
        dtype=torch.float32,
        num_gemms=1,
        bs=2,
2044
        model="126m",
2045
        recipe=recipe,
2046
        fp8_model_params=True,
2047
        fuse_wgrad_accumulation=True,
2048
2049
        bias=True,
        delay_wgrad_compute=False,
2050
2051
2052
    )


2053
def _test_padding_grouped_linear_accuracy(block, num_gemms, bs, dtype, config, recipe, fp8=False):
2054
2055

    def _pad_tensor_for_fp8(hidden_states, tokens_per_expert):
2056
2057
2058
        align_size = 16
        if recipe.mxfp8():
            align_size = 32
2059
        padded_tokens_per_expert = [
2060
2061
            (num_tokens + align_size - 1) // align_size * align_size
            for num_tokens in tokens_per_expert
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
        ]
        hidden_states = torch.split(hidden_states, tokens_per_expert)
        padded_hidden_states = []
        for hidden_state, actual_num_tokens, padded_num_tokens in zip(
            hidden_states, tokens_per_expert, padded_tokens_per_expert
        ):
            padded_hidden_states.append(hidden_state)
            if padded_num_tokens > actual_num_tokens:
                pad_tensor = torch.zeros(
                    padded_num_tokens - actual_num_tokens,
                    hidden_state.shape[1],
                    dtype=hidden_state.dtype,
                    device=hidden_state.device,
                )
                padded_hidden_states.append(pad_tensor)
        padded_hidden_states = torch.cat(padded_hidden_states, dim=0)
        return padded_hidden_states, padded_tokens_per_expert

    def _unpad_tensor_for_fp8(padded_hidden_states, actual_tokens_per_expert, tokens_per_expert):
        inputmats = torch.split(
            padded_hidden_states.view(-1, padded_hidden_states.shape[-1]), tokens_per_expert
        )
        hidden_states = torch.cat(
            [
                grad_output_mat[: actual_tokens_per_expert[i]]
                for i, grad_output_mat in enumerate(inputmats)
            ],
            dim=0,
        )

        return hidden_states

    def _generate_random_numbers(n, total_sum):
        if n <= 0:
            return []

        # reset seed
        random.seed(seed)

        breaks = sorted(random.sample(range(1, total_sum), n - 1))
        random_numbers = (
            [breaks[0]]
            + [breaks[i] - breaks[i - 1] for i in range(1, n - 1)]
            + [total_sum - breaks[-1]]
        )

        return random_numbers

    reset_rng_states()
    if fp8:
        FP8GlobalStateManager.reset()

    inp_hidden_states = torch.randn(
2115
        (config.max_seqlen_q * bs, config.hidden_size),
2116
2117
2118
2119
2120
2121
        dtype=dtype,
        device="cuda",
        requires_grad=True,
    )
    inp_hidden_states.retain_grad()

2122
    m_splits = _generate_random_numbers(num_gemms, config.max_seqlen_q * bs)
2123

2124
    with fp8_autocast(enabled=fp8, fp8_recipe=recipe):
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
        if isinstance(block, TorchGroupedLinearWithPadding):
            out = block(inp_hidden_states, m_splits)
        else:
            if fp8:
                padded_inp_hidden_states, padding_m_splits = _pad_tensor_for_fp8(
                    inp_hidden_states, m_splits
                )
                padded_inp_hidden_states = block(padded_inp_hidden_states, padding_m_splits)
                out = _unpad_tensor_for_fp8(padded_inp_hidden_states, m_splits, padding_m_splits)
            else:
                out = block(inp_hidden_states, m_splits)

    loss = out.sum()
    loss.backward()

    torch.cuda.synchronize()
    outputs = [out, inp_hidden_states.grad]
    for p in block.parameters():
        if p.requires_grad:
            outputs.append(p.grad)
    return outputs


@pytest.mark.parametrize("dtype", param_types)
@pytest.mark.parametrize("num_gemms", [3, 6])
@pytest.mark.parametrize("bs", batch_sizes)
2151
@pytest.mark.parametrize("model", ["126m"])
2152
@pytest.mark.parametrize("fp8", [True])
2153
@pytest.mark.parametrize("recipe", fp8_recipes)
2154
2155
@pytest.mark.parametrize("fp8_model_params", all_boolean)
def test_padding_grouped_linear_accuracy(
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
    dtype,
    num_gemms,
    bs,
    model,
    fp8,
    recipe,
    fp8_model_params,
    parallel_mode=None,
):
    if fp8_model_params and NVTE_TEST_NVINSPECT_ENABLED:
        pytest.skip("FP8 parameters are not supported in debug mode.")

    config = model_configs[model]
2169
    if config.max_seqlen_q % 16 != 0 and fp8:
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
        pytest.skip("FP8 requires sequence length to be divisible by 16.")

    with fp8_model_init(enabled=fp8 and fp8_model_params, recipe=recipe):
        grouped_linear = TorchGroupedLinearWithPadding(
            num_gemms,
            config.hidden_size,
            4 * config.hidden_size,
            bias=False,
            params_dtype=dtype,
            parallel_mode=parallel_mode,
            fp8=fp8,
        ).eval()

    with fp8_model_init(enabled=fp8 and fp8_model_params, recipe=recipe):
        ref_grouped_linear = GroupedLinear(
            num_gemms,
            config.hidden_size,
            4 * config.hidden_size,
            bias=False,
            params_dtype=dtype,
            parallel_mode=parallel_mode,
            device="cuda",
            save_original_input=False,
        ).eval()

    # Share params
    with torch.no_grad():
        inner_grouped_linear = grouped_linear.linear_fn
        for i in range(num_gemms):
            setattr(
                ref_grouped_linear,
                f"weight{i}",
                Parameter(getattr(inner_grouped_linear, f"weight{i}").clone()),
            )

    outputs = _test_padding_grouped_linear_accuracy(
        grouped_linear, num_gemms, bs, dtype, config, recipe, fp8
    )
    outputs_ref = _test_padding_grouped_linear_accuracy(
        ref_grouped_linear, num_gemms, bs, dtype, config, recipe, fp8
    )

    # Shoule be bit-wise match
    for i, (o, o_ref) in enumerate(zip(outputs, outputs_ref)):
        torch.testing.assert_close(o, o_ref, rtol=0, atol=0)


@pytest.mark.parametrize("dtype", param_types)
@pytest.mark.parametrize("num_gemms", [3])
@pytest.mark.parametrize("bs", [1])
@pytest.mark.parametrize("model", ["126m"])
@pytest.mark.parametrize("fp8", [True])
@pytest.mark.parametrize("recipe", fp8_recipes)
@pytest.mark.parametrize("fp8_model_params", [False])
def test_padding_grouped_linear_accuracy_save_original_input(
    dtype,
    num_gemms,
    bs,
    model,
    fp8,
    recipe,
    fp8_model_params,
    parallel_mode=None,
2233
):
2234
2235
    if fp8_model_params and NVTE_TEST_NVINSPECT_ENABLED:
        pytest.skip("FP8 parameters are not supported in debug mode.")
2236
2237
    if fp8 and recipe.delayed():
        pytest.skip("DelayedScaling recipe is not supported with save_original_input")
2238
2239

    config = model_configs[model]
2240
    if config.max_seqlen_q % 16 != 0 and fp8:
2241
2242
        pytest.skip("FP8 requires sequence length to be divisible by 16.")

2243
    with fp8_model_init(enabled=fp8 and fp8_model_params, recipe=recipe):
2244
2245
2246
2247
2248
2249
2250
2251
2252
2253
        grouped_linear = TorchGroupedLinearWithPadding(
            num_gemms,
            config.hidden_size,
            4 * config.hidden_size,
            bias=False,
            params_dtype=dtype,
            parallel_mode=parallel_mode,
            fp8=fp8,
        ).eval()

2254
    with fp8_model_init(enabled=fp8 and fp8_model_params, recipe=recipe):
2255
2256
2257
2258
2259
2260
2261
2262
        ref_grouped_linear = GroupedLinear(
            num_gemms,
            config.hidden_size,
            4 * config.hidden_size,
            bias=False,
            params_dtype=dtype,
            parallel_mode=parallel_mode,
            device="cuda",
2263
            save_original_input=True,
2264
2265
2266
2267
2268
2269
2270
2271
2272
2273
2274
2275
2276
        ).eval()

    # Share params
    with torch.no_grad():
        inner_grouped_linear = grouped_linear.linear_fn
        for i in range(num_gemms):
            setattr(
                ref_grouped_linear,
                f"weight{i}",
                Parameter(getattr(inner_grouped_linear, f"weight{i}").clone()),
            )

    outputs = _test_padding_grouped_linear_accuracy(
2277
        grouped_linear, num_gemms, bs, dtype, config, recipe, fp8
2278
2279
    )
    outputs_ref = _test_padding_grouped_linear_accuracy(
2280
        ref_grouped_linear, num_gemms, bs, dtype, config, recipe, fp8
2281
2282
2283
2284
2285
2286
2287
    )

    # Shoule be bit-wise match
    for i, (o, o_ref) in enumerate(zip(outputs, outputs_ref)):
        torch.testing.assert_close(o, o_ref, rtol=0, atol=0)


2288
2289
2290
2291
2292
2293
2294
def _test_gpt_e2e_cuda_graph(block, bs, dtype, config, graph):
    reset_rng_states()

    # Initialize loss function and optimizer.
    loss_fn = torch.nn.MSELoss()
    optimizer = torch.optim.SGD(block.parameters(), lr=0.1)

2295
    # Placeholders used for graph capture.
2296
    static_input = torch.randn(
2297
2298
2299
2300
        config.max_seqlen_q, bs, config.hidden_size, device="cuda", dtype=dtype, requires_grad=True
    )
    static_target = torch.randn(
        config.max_seqlen_q, bs, config.hidden_size, device="cuda", dtype=dtype
2301
    )
2302
2303
2304
2305

    real_input = torch.rand_like(static_input)
    real_target = torch.rand_like(static_target)

2306
2307
2308
2309
2310
2311
2312
2313
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
    # Basic training loop.
    def train_step():
        optimizer.zero_grad(set_to_none=False)
        out = block(static_input)
        loss = loss_fn(out, static_target)
        loss.backward()
        optimizer.step()
        return out

    # Warmup steps in a separate stream.
    s = torch.cuda.Stream()
    s.wait_stream(torch.cuda.current_stream())
    with torch.cuda.stream(s):
        for _ in range(3):
            train_step()
    torch.cuda.current_stream().wait_stream(s)

    # Capture graph.
    g = None
    static_output = None
2326
2327
2328
    if graph:
        g = torch.cuda.CUDAGraph()
        with torch.cuda.graph(g):
2329
2330
2331
2332
2333
2334
2335
            static_output = train_step()

    # Run with new data.
    with torch.no_grad():
        static_input.copy_(real_input)
        static_target.copy_(real_target)
    if graph:
2336
2337
        g.replay()
    else:
2338
        static_output = train_step()
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351

    grads = [static_input.grad]
    for p in block.parameters():
        if p.requires_grad:
            grads.append(p.grad)

    with torch.no_grad():
        output = static_output.clone()
    return output, grads


@pytest.mark.parametrize("dtype", param_types)
@pytest.mark.parametrize("bs", batch_sizes)
2352
@pytest.mark.parametrize("model", ["126m"])
2353
def test_gpt_cuda_graph(dtype, bs, model):
2354
2355
    if NVTE_TEST_NVINSPECT_ENABLED:
        pytest.skip("Cuda Graphs are not supported in debug mode.")
2356
2357
2358
2359
2360
2361
    config = model_configs[model]

    sigma = 0.023
    init_method = init_method_normal(sigma)
    output_layer_init_method = scaled_init_method_normal(sigma, config.num_layers)

2362
    block_args = (
2363
2364
        config.hidden_size,
        4 * config.hidden_size,
2365
        config.num_heads,
2366
2367
    )
    block_kwargs = dict(
2368
2369
2370
2371
2372
        layernorm_epsilon=config.eps,
        init_method=init_method,
        output_layer_init_method=output_layer_init_method,
        hidden_dropout=0.1,
        attention_dropout=0.1,
2373
        kv_channels=config.kv_channels,
2374
2375
2376
2377
        params_dtype=dtype,
        apply_residual_connection_post_layernorm=False,
        output_layernorm=False,
        device="cuda",
2378
    )
2379
2380
2381
2382
2383
    block = TransformerLayer(*block_args, **block_kwargs)
    graphed_block = TransformerLayer(*block_args, **block_kwargs)
    with torch.no_grad():
        for param1, param2 in zip(block.parameters(), graphed_block.parameters()):
            param2.copy_(param1)
2384

2385
2386
2387
2388
    out, grads = _test_gpt_e2e_cuda_graph(block, bs, dtype, config, False)
    graphed_out, graphed_grads = _test_gpt_e2e_cuda_graph(graphed_block, bs, dtype, config, True)
    params = list(block.parameters())
    graphed_params = list(graphed_block.parameters())
2389

2390
2391
2392
2393
    # Check that results match
    assert_allclose(out, graphed_out, 1e-3)
    assert_allclose(params, graphed_params, 1e-3)
    assert_allclose(grads, graphed_grads, 1e-3)
2394
2395


2396
def _test_gpt_fp8_parameters(bs, dtype, config, fp8_model_params, recipe):
2397
2398
2399
2400
2401
2402
2403
    reset_rng_states()
    FP8GlobalStateManager.reset()

    sigma = 0.023
    init_method = init_method_normal(sigma)
    output_layer_init_method = scaled_init_method_normal(sigma, config.num_layers)

2404
    with fp8_model_init(enabled=fp8_model_params, recipe=recipe):
2405
2406
2407
        block = TransformerLayer(
            config.hidden_size,
            4 * config.hidden_size,
2408
            config.num_heads,
2409
2410
2411
2412
2413
            layernorm_epsilon=config.eps,
            init_method=init_method,
            output_layer_init_method=output_layer_init_method,
            hidden_dropout=0.1,
            attention_dropout=0.1,
2414
            kv_channels=config.kv_channels,
2415
2416
2417
2418
2419
            apply_residual_connection_post_layernorm=False,
            output_layernorm=False,
            params_dtype=dtype,
            fuse_qkv_params=True,
            device="cuda",
2420
2421
2422
        )

    te_inp_hidden_states = torch.randn(
2423
        (config.max_seqlen_q, bs, config.hidden_size),
2424
2425
2426
2427
        dtype=dtype,
        device="cuda",
        requires_grad=True,
    )
2428
    te_inp_hidden_states.retain_grad()
2429
    te_inp_attn_mask = get_causal_attn_mask(config.max_seqlen_q)
2430

2431
    with fp8_autocast(enabled=True, fp8_recipe=recipe):
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
2443
2444
2445
        te_out = block(te_inp_hidden_states, attention_mask=te_inp_attn_mask)
    loss = te_out.sum()
    loss.backward()
    torch.cuda.synchronize()

    outputs = [te_out, te_inp_hidden_states.grad]
    for p in block.parameters():
        if p.requires_grad:
            outputs.append(p.grad)
    return outputs


@pytest.mark.parametrize("dtype", param_types)
@pytest.mark.parametrize("bs", batch_sizes)
2446
@pytest.mark.parametrize("model", ["126m"])
2447
2448
@pytest.mark.parametrize("recipe", fp8_recipes)
def test_gpt_fp8_parameters(dtype, bs, model, recipe):
2449
2450
    if NVTE_TEST_NVINSPECT_ENABLED:
        pytest.skip("FP8 parameters are not supported in debug mode.")
2451
2452
2453

    config = model_configs[model]

2454
2455
    outputs = _test_gpt_fp8_parameters(bs, dtype, config, False, recipe)
    outputs_fp8_params = _test_gpt_fp8_parameters(bs, dtype, config, True, recipe)
2456
2457
2458
2459
2460
2461
2462
2463
2464
2465
2466
2467

    # Check that results match
    tols = dict(rtol=0.125, atol=0.0675)
    for i, (ref, test) in enumerate(zip(outputs, outputs_fp8_params)):
        torch.testing.assert_close(
            test,
            ref,
            msg=f"Mismatch in tensor {i}",
            rtol=0.125,
            atol=0.0675,
        )

2468
2469
2470

@pytest.mark.parametrize("dtype", param_types)
@pytest.mark.parametrize("bs", batch_sizes)
2471
@pytest.mark.parametrize("model", ["126m"])
2472
2473
2474
2475
2476
2477
2478
2479
2480
2481
2482
def test_transformer_layer_hidden_states_format(dtype, bs, model):
    config = model_configs[model]

    sigma = 0.023
    init_method = init_method_normal(sigma)
    output_layer_init_method = scaled_init_method_normal(sigma, config.num_layers)

    # Set `torch.manual_seed` to make sure the weights are identical to the
    # other layer. Set `*dropout` values to 0 to make sure the forward pass
    # is identical to the other layer.
    torch.manual_seed(0)
2483
2484
2485
    block_sbhd = TransformerLayer(
        config.hidden_size,
        4 * config.hidden_size,
2486
        config.num_heads,
2487
2488
2489
2490
2491
        layernorm_epsilon=config.eps,
        init_method=init_method,
        output_layer_init_method=output_layer_init_method,
        hidden_dropout=0,
        attention_dropout=0,
2492
        kv_channels=config.kv_channels,
2493
2494
2495
2496
2497
        params_dtype=dtype,
        apply_residual_connection_post_layernorm=False,
        output_layernorm=False,
        device="cuda",
        attn_input_format="sbhd",
2498
2499
2500
2501
2502
2503
    )

    # Set `torch.manual_seed` to make sure the weights are identical to the
    # other layer. Set `*dropout` values to 0 to make sure the forward pass
    # is identical to the other layer.
    torch.manual_seed(0)
2504
2505
2506
    block_bshd = TransformerLayer(
        config.hidden_size,
        4 * config.hidden_size,
2507
        config.num_heads,
2508
2509
2510
2511
2512
        layernorm_epsilon=config.eps,
        init_method=init_method,
        output_layer_init_method=output_layer_init_method,
        hidden_dropout=0,
        attention_dropout=0,
2513
        kv_channels=config.kv_channels,
2514
2515
2516
2517
2518
        params_dtype=dtype,
        apply_residual_connection_post_layernorm=False,
        output_layernorm=False,
        device="cuda",
        attn_input_format="bshd",
2519
2520
    )

2521
2522
2523
2524
    torch.manual_seed(0)
    block_thd = TransformerLayer(
        config.hidden_size,
        4 * config.hidden_size,
2525
        config.num_heads,
2526
2527
2528
2529
2530
        layernorm_epsilon=config.eps,
        init_method=init_method,
        output_layer_init_method=output_layer_init_method,
        hidden_dropout=0,
        attention_dropout=0,
2531
        kv_channels=config.kv_channels,
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
2543
        params_dtype=dtype,
        apply_residual_connection_post_layernorm=False,
        output_layernorm=False,
        device="cuda",
        attn_input_format="thd",
        self_attn_mask_type="padding_causal",
    )

    for (n1, p1), (n2, p2), (n3, p3) in zip(
        block_bshd.named_parameters(), block_sbhd.named_parameters(), block_thd.named_parameters()
    ):
        assert torch.all(torch.eq(p1, p2) & torch.eq(p1, p3)), f"{n1}, {n2} and {n3} not identical"
2544
2545

    x_sbhd = torch.randn(
2546
        (config.max_seqlen_q, bs, config.hidden_size),
2547
2548
2549
2550
        dtype=dtype,
        device="cuda",
        requires_grad=True,
    )
2551

2552
    x_bshd = x_sbhd.transpose(0, 1).contiguous()
2553
2554
    x_thd = x_bshd.reshape(bs * config.max_seqlen_q, config.hidden_size).contiguous()
    x_thd_cumsum = torch.arange(bs + 1, device="cuda", dtype=torch.int32) * config.max_seqlen_q
2555
2556
2557
2558
2559
2560
2561
2562
2563
2564
2565

    # To make sure forward is also identical (just in case some module decides
    # to act fancy)
    torch.manual_seed(0)
    y_sbhd = block_sbhd(x_sbhd)

    # To make sure forward is also identical (just in case some module decides
    # to act fancy)
    torch.manual_seed(0)
    y_bshd = block_bshd(x_bshd)

2566
2567
2568
    # Check that results match
    torch.testing.assert_close(
        y_bshd,
2569
        y_sbhd.transpose(0, 1).contiguous(),
2570
    )
2571

2572
2573
2574
2575
2576
2577
2578
2579
2580
    # THD is not supported in float32 and on GPUs older than Ampere, skip the test here
    if dtype != torch.float32 and sm_80plus:
        # To make sure forward is also identical (just in case some module decides
        # to act fancy)
        torch.manual_seed(0)
        y_thd = block_thd(
            x_thd,
            cu_seqlens_q=x_thd_cumsum,
            cu_seqlens_kv=x_thd_cumsum,
2581
2582
            max_seqlen_q=config.max_seqlen_q,
            max_seqlen_kv=config.max_seqlen_kv,
2583
2584
2585
2586
        )

        torch.testing.assert_close(
            y_bshd,
2587
            y_thd.reshape(bs, config.max_seqlen_q, config.hidden_size).contiguous(),
2588
        )
2589

2590
2591
2592
2593
2594
2595
2596
2597
2598
2599

@pytest.mark.parametrize(
    "shape",
    [
        (1, 127, 128, 512),
        (8, 15, 128, 512),
        (8, 1027, 128, 512),
        (16, 10027, 128, 512),
    ],
)
2600
@pytest.mark.parametrize("dtype", param_types, ids=str)
2601
2602
@pytest.mark.parametrize("layout", ["TN", "NN", "NT"])
@pytest.mark.parametrize("accumulate", [False, True])
2603
2604
@pytest.mark.parametrize("use_cutlass", use_cutlass_grouped_gemm)
def test_grouped_gemm(shape, dtype, layout, accumulate, use_cutlass):
2605
2606
2607
2608
2609
2610
2611
2612
2613
2614
    torch.manual_seed(0)
    z, m, k, n = shape

    dist = torch.sort(torch.randint(0, m, (z - 1,))).values.tolist()
    m_splits = torch.tensor(dist + [m]) - torch.tensor([0] + dist)
    assert m_splits.sum() == m and len(m_splits) == z
    m_splits = m_splits.tolist()

    if layout == "TN":
        A = [torch.randn(n, k, dtype=dtype, device="cuda") for _ in range(z)]  # weight
2615
2616
2617
        B = list(torch.split(torch.randn(m, k, dtype=dtype, device="cuda"), m_splits))  # input
        out = [torch.randn(m, n, dtype=dtype, device="cuda")]  # output
        out_ref = [o.clone() for o in torch.split(out[0], m_splits)]
2618
        grad = False
2619
        single_output = True
2620
2621
    elif layout == "NN":
        A = [torch.randn(n, k, dtype=dtype, device="cuda") for _ in range(z)]  # weight
2622
2623
2624
2625
2626
        B = list(
            torch.split(torch.randn(m, n, dtype=dtype, device="cuda"), m_splits)
        )  # grad_output
        out = [torch.randn(m, k, dtype=dtype, device="cuda")]  # dgrad
        out_ref = [o.clone() for o in torch.split(out[0], m_splits)]
2627
        grad = True
2628
        single_output = True
2629
    else:  # layout == "NT"
2630
2631
2632
2633
        A = list(torch.split(torch.randn(m, k, dtype=dtype, device="cuda"), m_splits))  # input
        B = list(
            torch.split(torch.randn(m, n, dtype=dtype, device="cuda"), m_splits)
        )  # grad_output
2634
        out = [torch.randn(n, k, dtype=dtype, device="cuda") for _ in range(z)]  # wgrad
2635
        out_ref = [o.clone() for o in out]
2636
        grad = True
2637
        single_output = False
2638

2639
2640
2641
    if use_cutlass:
        os.environ["NVTE_USE_CUTLASS_GROUPED_GEMM"] = "1"

2642
2643
    # Force the sequential_linear and grouped_linear to use hipblaslt rather than hipblas
    if IS_HIP_EXTENSION:
2644
        ori_force_rocm_gemm = os.environ.get("NVTE_FORCE_ROCM_GEMM", None)
2645
2646
        os.environ["NVTE_FORCE_ROCM_GEMM"] = "1"

2647
    for i in range(z):
2648
        general_gemm(
2649
2650
2651
            A[i],
            B[i],
            get_workspace(),
2652
            dtype,
2653
2654
2655
2656
2657
            grad=grad,
            accumulate=accumulate,
            layout=layout,
            out=out_ref[i],
        )
2658
2659
    if single_output:
        out_ref = [torch.cat(out_ref)]
2660

2661
    general_grouped_gemm(
2662
        A,
2663
2664
        B,
        out,
2665
2666
        dtype,
        get_multi_stream_cublas_workspace(),
2667
        m_splits=m_splits,
2668
2669
2670
        grad=grad,
        accumulate=accumulate,
        layout=layout,
2671
        single_output=single_output,
2672
    )
2673
    if IS_HIP_EXTENSION:
2674
2675
2676
2677
        if ori_force_rocm_gemm is not None:
            os.environ["NVTE_FORCE_ROCM_GEMM"] = ori_force_rocm_gemm
        else:
            del os.environ["NVTE_FORCE_ROCM_GEMM"]
2678
2679

    for o, o_ref in zip(out, out_ref):
2680
2681
2682
2683
2684
2685
2686
2687
        if not use_cutlass:
            # cublas implementation should be bit-wise match
            torch.testing.assert_close(o, o_ref, rtol=0, atol=0)
        else:
            torch.testing.assert_close(o, o_ref, rtol=1.5e-2, atol=1.5e-2)

    if use_cutlass:
        os.environ.pop("NVTE_USE_CUTLASS_GROUPED_GEMM", None)
2688
2689


2690
2691
2692
2693
2694
2695
2696
2697
2698
2699
2700
2701
2702
2703
2704
2705
2706
2707
2708
2709
2710
2711
2712
2713
2714
2715
2716
2717
2718
2719
2720
2721
2722
2723
2724
2725
2726
2727
2728
2729
2730
2731
2732
2733
2734
2735
2736
2737
2738
2739
2740
2741
2742
2743
2744
2745
2746
2747
2748
2749
2750
2751
2752
2753
2754
@pytest.mark.parametrize("N", [32])
@pytest.mark.parametrize("datatype", [torch.float16, torch.bfloat16])
@pytest.mark.parametrize(
    "input_quantizer",
    [
        Float8CurrentScalingQuantizer(fp8_dtype=tex.DType.kFloat8E4M3, device="cuda"),
        MXFP8Quantizer(fp8_dtype=tex.DType.kFloat8E4M3),
    ],
)
@pytest.mark.parametrize(
    "out_quantizer",
    [
        Float8CurrentScalingQuantizer(fp8_dtype=tex.DType.kFloat8E4M3, device="cuda"),
        MXFP8Quantizer(fp8_dtype=tex.DType.kFloat8E4M3),
        Float8Quantizer(
            torch.ones(1).cuda().squeeze(), torch.ones(1).cuda().squeeze(), tex.DType.kFloat8E4M3
        ),
    ],
)
def test_fp8gemm_with_unfused_quantization(N, datatype, input_quantizer, out_quantizer):
    # For MXFP8 and CurrentScaling, below unfused quantization should happen
    # FP8 input --> cublas GEMM --> BF16 output --> Quantize to FP8 --> fp8 Output
    # Skip invalid configurations
    is_mxfp8_needed = isinstance(input_quantizer, MXFP8Quantizer) or isinstance(
        out_quantizer, MXFP8Quantizer
    )
    if not fp8_available:
        pytest.skip(reason_for_no_fp8)
    if is_mxfp8_needed and not mxfp8_available:
        pytest.skip(reason_for_no_mxfp8)
    inp_fp8 = input_quantizer(torch.randn(N, N, device="cuda", dtype=datatype))
    weight_fp8 = input_quantizer(torch.randn(N, N, device="cuda", dtype=datatype))
    outp_type = torch.float32
    quantized_out, *_ = general_gemm(
        weight_fp8,
        inp_fp8,
        get_workspace(),
        outp_type,
        quantization_params=out_quantizer,
        bias=None,
        use_split_accumulator=False,
    )

    out, *_ = general_gemm(
        weight_fp8,
        inp_fp8,
        get_workspace(),
        outp_type,
        quantization_params=None,
        bias=None,
        use_split_accumulator=False,
    )
    expected_quantized_out = out_quantizer(out)

    # Match results again Pytorch GEMM and allow for quantization tolerance
    pytorch_out = torch.matmul(
        inp_fp8.dequantize().to(torch.float64),
        torch.transpose(weight_fp8.dequantize().to(torch.float64), 0, 1),
    )
    fp8_tols = dict(rtol=0.125, atol=0.0675)
    torch.testing.assert_close(
        pytorch_out.to(outp_type), expected_quantized_out.dequantize(), **fp8_tols
    )
    # Match results between quantization happening inside vs outside general_gemm
    torch.testing.assert_close(expected_quantized_out.dequantize(), quantized_out.dequantize())
2755
2756
2757
2758
2759
2760
2761
2762
2763
2764
2765


@pytest.mark.parametrize(
    "shape",
    [
        (1, 128, 128, 512),
        (8, 1024, 128, 512),
        (16, 4096, 128, 512),
    ],
)
@pytest.mark.parametrize("accumulate", [False, True])
2766
def test_fp8_grouped_gemm(shape, accumulate):
2767
2768
2769
2770
    if not fp8_available:
        pytest.skip(reason_for_no_fp8)

    z, m, k, n = shape
2771
    m_splits = [m // z] * z
2772
2773
2774
2775
2776
2777
2778
2779

    dtype = torch.bfloat16
    A = [torch.randn(n, k, dtype=dtype, device="cuda") for _ in range(z)]  # weight
    B = torch.split(torch.randn(m, k, dtype=dtype, device="cuda"), m_splits)  # input
    out = torch.split(torch.randn(m, n, dtype=dtype, device="cuda"), m_splits)  # output
    out_ref = [o.clone() for o in out]

    # fp8 should be robust enough to this fake scale
2780
2781
    scale = 1 + torch.rand(1, dtype=torch.float32, device="cuda").squeeze()
    amax = torch.zeros(1, 1, dtype=torch.float32, device="cuda")
2782

2783
2784
2785
2786
    a_quantizers = [
        Float8Quantizer(
            scale.clone(),
            amax.clone(),
2787
2788
            tex.DType.kFloat8E4M3,
        )
2789
        for _ in range(z)
2790
    ]
2791
2792
2793
2794
2795
    b_quantizers = [
        Float8Quantizer(
            scale.clone(),
            amax.clone(),
            tex.DType.kFloat8E4M3,
2796
        )
2797
        for _ in range(z)
2798
2799
    ]

2800
2801
2802
2803
2804
2805
    A_fp8 = []
    B_fp8 = []

    for i in range(z):
        A_fp8.append(a_quantizers[i](A[i]))
        B_fp8.append(b_quantizers[i](B[i]))
2806
2807
2808

    # baseline
    for i in range(z):
2809
        general_gemm(
2810
2811
2812
            A_fp8[i],
            B_fp8[i],
            get_workspace(),
2813
            dtype,
2814
2815
2816
            out=out_ref[i],
            accumulate=accumulate,
        )
2817
2818
2819
2820
2821
2822
    general_grouped_gemm(
        A_fp8,
        B_fp8,
        out,
        dtype,
        get_multi_stream_cublas_workspace(),
2823
        m_splits=m_splits,
2824
2825
        accumulate=accumulate,
    )
2826
2827
2828
2829

    # should be bit-wise match
    for o, o_ref in zip(out, out_ref):
        torch.testing.assert_close(o, o_ref, rtol=0, atol=0)
2830
2831
2832
2833
2834
2835
2836
2837
2838
2839
2840
2841
2842
2843
2844
2845
2846
2847
2848
2849
2850
2851
2852
2853
2854
2855
2856
2857
2858
2859
2860
2861
2862
2863
2864
2865
2866
2867
2868
2869
2870
2871
2872
2873
2874
2875
2876
2877
2878


def test_noncontiguous():
    def _create2modules(m, params):
        mod1 = m(*params)
        mod2 = m(*params)
        for p1, p2 in zip(mod1.parameters(), mod2.parameters()):
            p2.data = p1.data.clone()

        return mod1, mod2

    def _run_module(m, inp):
        out = m(inp)
        out.sum().backward()
        ret = [out]
        if inp.grad is not None:
            ret.append(inp.grad)

        for p in m.parameters():
            if p.requires_grad:
                ret.append(p.grad)
        return ret

    a = torch.randn((128, 256), device="cuda", requires_grad=True)
    a = a.T
    assert not a.is_contiguous(), "The test is supposed to test noncontiguous input."

    b = a.contiguous()

    # LayerNorm
    ln1, ln2 = _create2modules(LayerNorm, [128])
    outT = _run_module(ln1, a)
    out = _run_module(ln2, b)

    assert_allclose(out, outT, 1e-7)

    # RMSNorm
    ln1, ln2 = _create2modules(RMSNorm, [128])
    outT = _run_module(ln1, a)
    out = _run_module(ln2, b)

    assert_allclose(out, outT, 1e-7)

    # GEMM
    g1, g2 = _create2modules(Linear, [128, 128])
    outT = _run_module(g1, a)
    out = _run_module(g2, b)

    assert_allclose(out, outT, 1e-7)