benchmark_cutlass_moe_nvfp4.py 16.1 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
4
5
6
7
8
"""
Benchmark the performance of the cutlass_moe_fp4 kernel vs the triton_moe
kernel. The cutlass_moe_fp4 kernel takes in fp4 quantized weights and 16-bit
activations. The triton_moe kernel takes in fp8 weights(tensor scaled to fp8)
and 16-bit activations.
"""
9

10
11
12
13
import nvtx
import torch
import torch.utils.benchmark as benchmark

14
import vllm.model_executor.layers.fused_moe.modular_kernel as mk
15
from tests.kernels.moe.utils import make_dummy_moe_config
16
17
from vllm import _custom_ops as ops
from vllm.config import ParallelConfig, VllmConfig, set_current_vllm_config
18
19
20
from vllm.model_executor.layers.fused_moe.all2all_utils import (
    maybe_make_prepare_finalize,
)
21
22
23
24
from vllm.model_executor.layers.fused_moe.config import (
    fp8_w8a8_moe_quant_config,
    nvfp4_moe_quant_config,
)
25
26
27
from vllm.model_executor.layers.fused_moe.cutlass_moe import (
    CutlassExpertsFp4,
)
28
from vllm.model_executor.layers.fused_moe.fused_moe import fused_experts, fused_topk
29
from vllm.scalar_type import scalar_types
30
from vllm.utils.argparse_utils import FlexibleArgumentParser
31
from vllm.v1.worker.workspace import init_workspace_manager
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53

WEIGHT_SHAPES_MOE = {
    "nvidia/DeepSeek-R1-FP4": [
        [256, 8, 2048, 7168],
    ],
}

DEFAULT_MODELS = [
    "nvidia/DeepSeek-R1-FP4",
]

DEFAULT_BATCH_SIZES = [4, 8, 16, 32, 64, 128, 256, 512, 1024, 2048]
DEFAULT_TP_SIZES = [1]

PER_ACT_TOKEN_OPTS = [False]
PER_OUT_CH_OPTS = [False]
FLOAT4_E2M1_MAX = scalar_types.float4_e2m1f.max()
FLOAT8_E4M3_MAX = torch.finfo(torch.float8_e4m3fn).max


def to_fp8(tensor: torch.Tensor):
    finfo = torch.finfo(torch.float8_e4m3fn)
54
55
56
    return torch.round(tensor.clamp(min=finfo.min, max=finfo.max)).to(
        dtype=torch.float8_e4m3fn
    )
57
58


59
60
61
62
63
64
65
66
67
def bench_run(
    results: list[benchmark.Measurement],
    model: str,
    num_experts: int,
    topk: int,
    per_act_token: bool,
    per_out_ch: bool,
    mkn: tuple[int, int, int],
):
68
69
70
    label = "NVFP4 Blockscaled CUTLASS MOE vs FP8 Tensor Scaled Triton"

    sub_label = (
71
72
73
74
        "{}, num_experts={}, topk={}, per_act_token={} per_out_ch={}, MKN=({})".format(
            model, num_experts, topk, per_act_token, per_out_ch, mkn
        )
    )
75
76
77
78
79
80
81
82
83
84
85
86
87

    print(f"Testing: {sub_label}")

    (m, k, n) = mkn

    dtype = torch.half
    device = "cuda"
    a = torch.randn((m, k), device=device, dtype=dtype) / 10
    w1 = torch.randn((num_experts, 2 * n, k), device=device, dtype=dtype) / 10
    w2 = torch.randn((num_experts, k, n), device=device, dtype=dtype) / 10

    _, a_fp8_scale = ops.scaled_fp8_quant(a)

88
89
90
91
92
93
    w1_fp8q = torch.empty(
        (num_experts, 2 * n, k), device=device, dtype=torch.float8_e4m3fn
    )
    w2_fp8q = torch.empty((num_experts, k, n), device=device, dtype=torch.float8_e4m3fn)
    w1_fp8scale = torch.empty((num_experts, 1, 1), device=device, dtype=torch.float32)
    w2_fp8scale = torch.empty((num_experts, 1, 1), device=device, dtype=torch.float32)
94
95
96
97
98
99
100
101
102
103
104
105

    for expert in range(num_experts):
        w1_fp8q[expert], w1_fp8scale[expert] = ops.scaled_fp8_quant(w1[expert])
        w2_fp8q[expert], w2_fp8scale[expert] = ops.scaled_fp8_quant(w2[expert])

    w1_fp8q_notransp = w1_fp8q.clone()
    w2_fp8q_notransp = w2_fp8q.clone()
    w1_fp8q = w1_fp8q.transpose(1, 2)
    w2_fp8q = w2_fp8q.transpose(1, 2)

    score = torch.randn((m, num_experts), device=device, dtype=dtype)

106
    topk_weights, topk_ids, _ = fused_topk(a, score, topk, renormalize=False)
107
108

    quant_blocksize = 16
109
110
111
112
113
114
115
116
    w1_blockscale = torch.empty(
        (num_experts, 2 * n, k // quant_blocksize),
        device=device,
        dtype=torch.float8_e4m3fn,
    )
    w2_blockscale = torch.empty(
        (num_experts, k, n // quant_blocksize), device=device, dtype=torch.float8_e4m3fn
    )
117
118
119

    # n_b_scales = 2 * n if per_out_ch else 1
    # k_b_scales = k if per_out_ch else 1
120
121
122
123
124
125
126
    w1_fp4 = torch.empty((num_experts, 2 * n, k // 2), device=device, dtype=torch.uint8)
    w2_fp4 = torch.empty((num_experts, k, n // 2), device=device, dtype=torch.uint8)

    w1_gs = torch.empty((num_experts,), device=device, dtype=torch.float32)
    w2_gs = torch.empty((num_experts,), device=device, dtype=torch.float32)
    a1_gs = torch.ones((num_experts,), device=device, dtype=torch.float32)
    a2_gs = torch.ones((num_experts,), device=device, dtype=torch.float32)
127
128
129
130
131
132
133
134
135
136

    for expert in range(num_experts):
        w1_e = w1[expert]
        w2_e = w2[expert]
        w1_amax = torch.abs(w1_e).max().to(torch.float32)
        w2_amax = torch.abs(w2_e).max().to(torch.float32)
        w1_gs[expert] = FLOAT8_E4M3_MAX * FLOAT4_E2M1_MAX / w1_amax
        w2_gs[expert] = FLOAT8_E4M3_MAX * FLOAT4_E2M1_MAX / w2_amax

        w1_fp4[expert], w1_blockscale[expert] = ops.scaled_fp4_quant(
137
138
            w1_e, w1_gs[expert]
        )
139
140

        w2_fp4[expert], w2_blockscale[expert] = ops.scaled_fp4_quant(
141
142
143
144
145
146
147
148
149
150
151
152
153
154
            w2_e, w2_gs[expert]
        )

    def run_triton_moe(
        a: torch.Tensor,
        w1: torch.Tensor,
        w2: torch.Tensor,
        topk_weights: torch.Tensor,
        topk_ids: torch.Tensor,
        w1_scale: torch.Tensor,
        w2_scale: torch.Tensor,
        a_fp8_scale: torch.Tensor,
        num_repeats: int,
    ):
155
156
157
158
159
160
        quant_config = fp8_w8a8_moe_quant_config(
            w1_scale=w1_scale,
            w2_scale=w2_scale,
            a1_scale=a_fp8_scale,
        )

161
        for _ in range(num_repeats):
162
163
164
165
166
167
            fused_experts(
                a,
                w1,
                w2,
                topk_weights,
                topk_ids,
168
                quant_config=quant_config,
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
            )

    def run_cutlass_moe_fp4(
        a: torch.Tensor,
        w1_fp4: torch.Tensor,
        w2_fp4: torch.Tensor,
        w1_blockscale: torch.Tensor,
        w2_blockscale: torch.Tensor,
        w1_gs: torch.Tensor,
        w2_gs: torch.Tensor,
        a1_gs: torch.Tensor,
        a2_gs: torch.Tensor,
        topk_weights: torch.Tensor,
        topk_ids: torch.Tensor,
        m: int,
        n: int,
        k: int,
        e: int,
        device: torch.device,
        num_repeats: int,
    ):
190
191
192
193
194
195
196
197
        quant_config = nvfp4_moe_quant_config(
            a1_gscale=a1_gs,
            a2_gscale=a2_gs,
            w1_scale=w1_blockscale,
            w2_scale=w2_blockscale,
            g1_alphas=w1_gs,
            g2_alphas=w2_gs,
        )
198

199
200
201
202
203
204
205
206
207
208
209
210
211
        moe_config = make_dummy_moe_config(
            num_experts=num_experts,
            hidden_dim=k,
            intermediate_size_per_partition=n,
            in_dtype=a.dtype,
        )
        kernel = mk.FusedMoEKernel(
            maybe_make_prepare_finalize(
                moe=moe_config,
                quant_config=quant_config,
                allow_new_interface=True,
                use_monolithic=False,
            ),
212
            CutlassExpertsFp4(
213
                moe_config=moe_config,
214
215
216
217
                quant_config=quant_config,
            ),
        )

218
219
        for _ in range(num_repeats):
            with nvtx.annotate("cutlass_moe_fp4", color="green"):
220
221
222
223
                kernel(
                    hidden_states=a,
                    w1=w1_fp4,
                    w2=w2_fp4,
224
225
226
                    topk_weights=topk_weights,
                    topk_ids=topk_ids,
                )
227
228

    def run_cutlass_from_graph(
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
        a: torch.Tensor,
        a1_gscale: torch.Tensor,
        w1_fp4: torch.Tensor,
        w1_blockscale: torch.Tensor,
        w1_alphas: torch.Tensor,
        a2_gscale: torch.Tensor,
        w2_fp4: torch.Tensor,
        w2_blockscale: torch.Tensor,
        w2_alphas: torch.Tensor,
        topk_weights: torch.Tensor,
        topk_ids: torch.Tensor,
        m: int,
        n: int,
        k: int,
        e: int,
        device: torch.device,
    ):
246
247
248
249
250
251
252
253
        quant_config = nvfp4_moe_quant_config(
            a1_gscale=a1_gs,
            a2_gscale=a2_gs,
            w1_scale=w1_blockscale,
            w2_scale=w2_blockscale,
            g1_alphas=w1_gs,
            g2_alphas=w2_gs,
        )
254
        moe_config = make_dummy_moe_config()
255

256
257
258
259
260
261
262
        kernel = mk.FusedMoEKernel(
            maybe_make_prepare_finalize(
                moe=moe_config,
                quant_config=quant_config,
                allow_new_interface=True,
                use_monolithic=False,
            ),
263
            CutlassExpertsFp4(
264
                moe_config=moe_config,
265
266
267
268
                quant_config=quant_config,
            ),
        )

269
        with set_current_vllm_config(
270
271
            VllmConfig(parallel_config=ParallelConfig(pipeline_parallel_size=1))
        ):
272
273
274
275
            return kernel(
                hidden_states=a,
                w1=w1_fp4,
                w2=w2_fp4,
276
277
278
279
280
281
282
283
284
285
286
287
288
289
                topk_weights=topk_weights,
                topk_ids=topk_ids,
            )

    def run_triton_from_graph(
        a: torch.Tensor,
        w1: torch.Tensor,
        w2: torch.Tensor,
        topk_weights: torch.Tensor,
        topk_ids: torch.Tensor,
        w1_scale: torch.Tensor,
        w2_scale: torch.Tensor,
        a_fp8_scale: torch.Tensor,
    ):
290
        with set_current_vllm_config(
291
292
            VllmConfig(parallel_config=ParallelConfig(pipeline_parallel_size=1))
        ):
293
294
295
296
297
            quant_config = fp8_w8a8_moe_quant_config(
                w1_scale=w1_scale,
                w2_scale=w2_scale,
                a1_scale=a_fp8_scale,
            )
298
299
300
301
302
303
            return fused_experts(
                a,
                w1,
                w2,
                topk_weights,
                topk_ids,
304
                quant_config=quant_config,
305
            )
306
307
308
309

    def replay_graph(graph, num_repeats):
        for _ in range(num_repeats):
            graph.replay()
310
        torch.accelerator.synchronize()
311
312
313
314

    cutlass_stream = torch.cuda.Stream()
    cutlass_graph = torch.cuda.CUDAGraph()
    with torch.cuda.graph(cutlass_graph, stream=cutlass_stream):
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
        run_cutlass_from_graph(
            a=a,
            a1_gscale=a1_gs,
            w1_fp4=w1_fp4,
            w1_blockscale=w1_blockscale,
            w1_alphas=w1_gs,
            a2_gscale=a2_gs,
            w2_fp4=w2_fp4,
            w2_blockscale=w2_blockscale,
            w2_alphas=w2_gs,
            topk_weights=topk_weights,
            topk_ids=topk_ids,
            m=m,
            n=n,
            k=k,
            e=num_experts,
            device=device,
        )
333
    torch.accelerator.synchronize()
334
335
336
337

    triton_stream = torch.cuda.Stream()
    triton_graph = torch.cuda.CUDAGraph()
    with torch.cuda.graph(triton_graph, stream=triton_stream):
338
339
340
341
342
343
344
345
346
347
        run_triton_from_graph(
            a,
            w1_fp8q_notransp,
            w2_fp8q_notransp,
            topk_weights,
            topk_ids,
            w1_fp8scale,
            w2_fp8scale,
            a_fp8_scale,
        )
348
    torch.accelerator.synchronize()
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393

    min_run_time = 5
    num_warmup = 5
    num_runs = 25

    globals = {
        # Baseline params
        "w1": w1,
        "w2": w2,
        "score": score,
        "topk": topk,
        "w1_fp8q_notransp": w1_fp8q_notransp,
        "w2_fp8q_notransp": w2_fp8q_notransp,
        "w1_fp8scale": w1_fp8scale,
        "w2_fp8scale": w2_fp8scale,
        "a_fp8_scale": a_fp8_scale,
        # Cutlass params
        "a": a,
        "a1_gscale": a1_gs,
        "w1_fp4": w1_fp4,
        "w1_blockscale": w1_blockscale,
        "w1_alphas": w1_gs,
        "a2_gscale": a2_gs,
        "w2_fp4": w2_fp4,
        "w2_blockscale": w2_blockscale,
        "w2_alphas": w2_gs,
        "topk_weights": topk_weights,
        "topk_ids": topk_ids,
        "m": m,
        "n": n,
        "k": k,
        "e": num_experts,
        "device": device,
        # cuda graph params
        "cutlass_graph": cutlass_graph,
        "triton_graph": triton_graph,
        # Gen params
        "num_runs": num_runs,
        # Kernels
        "run_triton_moe": run_triton_moe,
        "run_cutlass_moe_fp4": run_cutlass_moe_fp4,
        "replay_graph": replay_graph,
    }

    # Warmup
394
395
396
397
398
399
400
401
402
403
404
    run_triton_moe(
        a,
        w1_fp8q_notransp,
        w2_fp8q_notransp,
        topk_weights,
        topk_ids,
        w1_fp8scale,
        w2_fp8scale,
        a_fp8_scale,
        num_warmup,
    )
405
406
407

    results.append(
        benchmark.Timer(
408
            stmt="run_triton_moe(a, w1_fp8q_notransp, w2_fp8q_notransp, topk_weights, topk_ids, w1_fp8scale, w2_fp8scale, a_fp8_scale, num_runs)",  # noqa: E501
409
410
411
412
            globals=globals,
            label=label,
            sub_label=sub_label,
            description="triton_moe",
413
414
        ).blocked_autorange(min_run_time=min_run_time)
    )
415
416
417
418
419
420
421
422
423
424
425

    # Warmup
    replay_graph(triton_graph, num_warmup)

    results.append(
        benchmark.Timer(
            stmt="replay_graph(triton_graph, num_runs)",
            globals=globals,
            label=label,
            sub_label=sub_label,
            description="triton_moe_cuda_graphs",
426
427
        ).blocked_autorange(min_run_time=min_run_time)
    )
428
429
430

    # Warmup

431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
    run_cutlass_moe_fp4(
        a,
        w1_fp4,
        w2_fp4,
        w1_blockscale,
        w2_blockscale,
        w1_gs,
        w2_gs,
        a1_gs,
        a2_gs,
        topk_weights,
        topk_ids,
        m,
        n,
        k,
        num_experts,
        device,
        num_warmup,
    )
450
451
452

    results.append(
        benchmark.Timer(
453
            stmt="run_cutlass_moe_fp4(a, w1_fp4, w2_fp4, w1_blockscale, w2_blockscale, w1_alphas, w2_alphas, a1_gscale, a2_gscale, topk_weights, topk_ids, m, n, k, e, device, num_runs)",  # noqa: E501
454
455
456
457
            globals=globals,
            label=label,
            sub_label=sub_label,
            description="cutlass_moe_fp4",
458
459
        ).blocked_autorange(min_run_time=min_run_time)
    )
460
461
462
463
464
465
466
467
468
469
470

    # Warmup
    replay_graph(cutlass_graph, num_warmup)

    results.append(
        benchmark.Timer(
            stmt="replay_graph(cutlass_graph, num_runs)",
            globals=globals,
            label=label,
            sub_label=sub_label,
            description="cutlass_moe_fp4_cuda_graphs",
471
472
        ).blocked_autorange(min_run_time=min_run_time)
    )
473
474
475


def main(args):
476
477
478
479
    # Initialize workspace manager (required for CUTLASS MoE kernels)
    device = torch.device("cuda:0")
    init_workspace_manager(device)

480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
    print("Benchmarking models:")
    for i, model in enumerate(args.models):
        print(f"[{i}]  {model}")

    results: list[benchmark.Measurement] = []

    for model in args.models:
        for tp in args.tp_sizes:
            for layer in WEIGHT_SHAPES_MOE[model]:
                num_experts = layer[0]
                topk = layer[1]
                size_k = layer[2]
                size_n = layer[3] // tp

                if len(args.limit_k) > 0 and size_k not in args.limit_k:
                    continue

                if len(args.limit_n) > 0 and size_n not in args.limit_n:
                    continue

                for per_act_token in PER_ACT_TOKEN_OPTS:
                    for per_out_ch in PER_OUT_CH_OPTS:
                        for size_m in args.batch_sizes:
                            mkn = (size_m, size_k, size_n)
504
505
506
507
508
509
510
511
512
                            bench_run(
                                results,
                                model,
                                num_experts,
                                topk,
                                per_act_token,
                                per_out_ch,
                                mkn,
                            )
513
514
515
516
517
518
519

    compare = benchmark.Compare(results)
    compare.print()


if __name__ == "__main__":
    parser = FlexibleArgumentParser(
520
521
        description="Benchmark NVFP4 CUTLASS MOE across specified models/shapes/batches"
    )
522
523
524
525
526
527
528
    parser.add_argument(
        "--models",
        nargs="+",
        type=str,
        default=DEFAULT_MODELS,
        choices=WEIGHT_SHAPES_MOE.keys(),
    )
529
530
531
532
    parser.add_argument("--tp-sizes", nargs="+", type=int, default=DEFAULT_TP_SIZES)
    parser.add_argument(
        "--batch-sizes", nargs="+", type=int, default=DEFAULT_BATCH_SIZES
    )
533
534
535
    parser.add_argument("--limit-k", nargs="+", type=int, default=[])
    parser.add_argument("--limit-n", nargs="+", type=int, default=[])
    parser.add_argument("--limit-num-groups", nargs="+", type=int, default=[])
536
    parser.add_argument("--limit-per-act-token", nargs="+", type=int, default=[])
537
538
539
540
    parser.add_argument("--limit-per-out-ch", nargs="+", type=int, default=[])

    args = parser.parse_args()
    main(args)