benchmark_moe.py 29 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
import argparse
5
import json
6
import os
7
import time
8
from contextlib import nullcontext
9
from datetime import datetime
10
from itertools import product
11
from typing import Any, TypedDict, Optional
12
13
14
15
16

import ray
import torch
from ray.experimental.tqdm_ray import tqdm

17
18
19
20
from vllm.model_executor.layers.fused_moe.config import (
    FusedMoEQuantConfig,
    _get_config_dtype_str,
)
21
from vllm.model_executor.layers.fused_moe.fused_moe import *
22
from vllm.transformers_utils.config import get_config
23
from vllm.triton_utils import triton
24
from vllm.utils import FlexibleArgumentParser
25

26
27
# 移除全局的 current_platform 导入,改为在需要时局部导入
# FP8_DTYPE = current_platform.fp8_dtype()
28
29


30
def ensure_divisibility(numerator, denominator, text):
31
    """Ensure that numerator is divisible by the denominator."""
32
33
    assert numerator % denominator == 0, "{} {} is not divisible by tp {}.".format(
        text, numerator, denominator
34
35
36
    )


37
38
39
40
41
42
43
class BenchmarkConfig(TypedDict):
    BLOCK_SIZE_M: int
    BLOCK_SIZE_N: int
    BLOCK_SIZE_K: int
    GROUP_SIZE_M: int
    num_warps: int
    num_stages: int
zhuwenwen's avatar
zhuwenwen committed
44
    num_ldmatrixes: Optional[int]
45
46


47
48
49
50
51
52
53
54
55
56
57
def benchmark_config(
    config: BenchmarkConfig,
    num_tokens: int,
    num_experts: int,
    shard_intermediate_size: int,
    hidden_size: int,
    topk: int,
    dtype: torch.dtype,
    use_fp8_w8a8: bool,
    use_int8_w8a16: bool,
    num_iters: int = 100,
58
    block_quant_shape: list[int] = None,
59
    use_deep_gemm: bool = False,
zhuwenwen's avatar
zhuwenwen committed
60
    nn_moe: Optional[bool] = False
61
) -> float:
62
63
64
    from vllm.platforms import current_platform
    device = torch.cuda.current_device()

65
    init_dtype = torch.float16 if use_fp8_w8a8 else dtype
66
    x = torch.randn(num_tokens, hidden_size, dtype=dtype, device=device)
67
    if use_int8_w8a16:
zhuwenwen's avatar
zhuwenwen committed
68
        if not nn_moe:
zhuwenwen's avatar
zhuwenwen committed
69
70
71
72
73
74
75
76
77
            w1 = torch.randint(
                -127,
                127,
                (
                    num_experts,
                    shard_intermediate_size,
                    hidden_size,
                ),
                dtype=torch.int8,
78
                device=device,
zhuwenwen's avatar
zhuwenwen committed
79
80
81
82
83
84
85
86
87
88
            )
            w2 = torch.randint(
                -127,
                127,
                (
                    num_experts,
                    hidden_size,
                    shard_intermediate_size // 2,
                ),
                dtype=torch.int8,
89
                device=device,
zhuwenwen's avatar
zhuwenwen committed
90
            )
zhuwenwen's avatar
zhuwenwen committed
91
        else:
zhuwenwen's avatar
zhuwenwen committed
92
93
94
95
96
97
98
99
100
            w1 = torch.randint(
                -127,
                127,
                (
                    num_experts,
                    hidden_size,
                    shard_intermediate_size,
                ),
                dtype=torch.int8,
101
                device=device,
zhuwenwen's avatar
zhuwenwen committed
102
103
104
105
106
107
108
109
110
111
            )
            w2 = torch.randint(
                -127,
                127,
                (
                    num_experts,
                    shard_intermediate_size // 2,
                    hidden_size,
                ),
                dtype=torch.int8,
112
                device=device,
zhuwenwen's avatar
zhuwenwen committed
113
            )
114
    else:
zhuwenwen's avatar
zhuwenwen committed
115
        if not nn_moe:
zhuwenwen's avatar
zhuwenwen committed
116
            w1 = torch.randn(
117
                num_experts, shard_intermediate_size, hidden_size, dtype=init_dtype, device=device
zhuwenwen's avatar
zhuwenwen committed
118
119
            )
            w2 = torch.randn(
120
                num_experts, hidden_size, shard_intermediate_size // 2, dtype=init_dtype, device=device
zhuwenwen's avatar
zhuwenwen committed
121
            )
zhuwenwen's avatar
zhuwenwen committed
122
        else:
zhuwenwen's avatar
zhuwenwen committed
123
            w1 = torch.randn(
124
                num_experts, hidden_size, shard_intermediate_size, dtype=init_dtype, device=device
zhuwenwen's avatar
zhuwenwen committed
125
126
            )
            w2 = torch.randn(
127
                num_experts, shard_intermediate_size // 2, hidden_size, dtype=init_dtype, device=device
zhuwenwen's avatar
zhuwenwen committed
128
            )
129
    gating_output = torch.randn(num_iters, num_tokens, num_experts, dtype=torch.float32, device=device)
130
131
132
133
134

    w1_scale = None
    w2_scale = None
    a1_scale = None
    a2_scale = None
135
    if use_int8_w8a16:
136
        w1_scale = torch.randn(
137
            (num_experts, 2 * shard_intermediate_size), dtype=torch.float32, device=device
138
        )
139
        w2_scale = torch.randn((hidden_size, num_experts), dtype=torch.float32, device=device)
140
141
142
    if use_deep_gemm:
        # we use the default block shape for deepgemm
        block_quant_shape = [128, 128]
143
    if use_fp8_w8a8:
144
145
146
147
148
149
150
151
152
153
        if block_quant_shape:
            block_n, block_k = block_quant_shape[0], block_quant_shape[1]
            E = num_experts
            N = shard_intermediate_size // 2
            K = hidden_size
            factor_for_scale = 1e-2
            n_tiles_w1 = (2 * N + block_n - 1) // block_n
            n_tiles_w2 = (K + block_n - 1) // block_n
            k_tiles_w1 = (K + block_k - 1) // block_k
            k_tiles_w2 = (N + block_k - 1) // block_k
154
            w1_scale = (
155
                torch.rand((E, n_tiles_w1, k_tiles_w1), dtype=torch.float32, device=device)
156
157
158
                * factor_for_scale
            )
            w2_scale = (
159
                torch.rand((E, n_tiles_w2, k_tiles_w2), dtype=torch.float32, device=device)
160
161
                * factor_for_scale
            )
162
        else:
163
164
            w1_scale = torch.randn(num_experts, dtype=torch.float32, device=device)
            w2_scale = torch.randn(num_experts, dtype=torch.float32, device=device)
165

166
167
        a1_scale = torch.randn(1, dtype=torch.float32, device=device)
        a2_scale = torch.randn(1, dtype=torch.float32, device=device)
168

169
170
        # 获取 FP8_DTYPE
        FP8_DTYPE = current_platform.fp8_dtype()
171
172
        w1 = w1.to(FP8_DTYPE)
        w2 = w2.to(FP8_DTYPE)
173

174
    input_gating = torch.empty(num_tokens, num_experts, dtype=torch.float32, device=device)
175
176
177
178
179

    def prepare(i: int):
        input_gating.copy_(gating_output[i])

    def run():
180
        from vllm.model_executor.layers.fused_moe import override_config
181

182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
        if use_fp8_w8a8:
            quant_dtype = torch.float8_e4m3fn
        elif use_int8_w8a16:
            quant_dtype = torch.int8
        else:
            quant_dtype = None

        quant_config = FusedMoEQuantConfig.make(
            quant_dtype=quant_dtype,
            w1_scale=w1_scale,
            w2_scale=w2_scale,
            a1_scale=a1_scale,
            a2_scale=a2_scale,
            block_shape=block_quant_shape,
        )

198
        with override_config(config):
199
200
201
202
203
204
205
206
207
208
209
210
            topk_weights, topk_ids, token_expert_indices = fused_topk(
                x, input_gating, topk, renormalize=not use_deep_gemm
            )
            return fused_experts(
                x,
                w1,
                w2,
                topk_weights,
                topk_ids,
                inplace=True,
                quant_config=quant_config,
                allow_deep_gemm=use_deep_gemm,
zhuwenwen's avatar
zhuwenwen committed
211
                use_nn_moe=nn_moe,
212
            )
213
214
215
216
217
218

    # JIT compilation & warmup
    run()
    torch.cuda.synchronize()

    # Capture 10 invocations with CUDA graph
219
220
221
222
223
    graph = torch.cuda.CUDAGraph()
    with torch.cuda.graph(graph):
        for _ in range(10):
            run()
    torch.cuda.synchronize()
224
225
226

    # Warmup
    for _ in range(5):
227
228
        graph.replay()
        # run()
229
230
231
232
233
    torch.cuda.synchronize()

    start_event = torch.cuda.Event(enable_timing=True)
    end_event = torch.cuda.Event(enable_timing=True)

234
    latencies: list[float] = []
235
236
237
238
239
    for i in range(num_iters):
        prepare(i)
        torch.cuda.synchronize()

        start_event.record()
240
241
        graph.replay()
        # run()
242
243
244
245
        end_event.record()
        end_event.synchronize()
        latencies.append(start_event.elapsed_time(end_event))
    avg = sum(latencies) / (num_iters * 10) * 1000  # us
246
    graph.reset()
247
248
249
    return avg


zhuwenwen's avatar
zhuwenwen committed
250
def get_rocm_tuning_space(use_fp16, nn_moe: Optional[bool] = False):
251
252
    block_m_range = [16, 32, 64, 128, 256]
    block_n_range = [32, 64, 128, 256]
253
    block_k_range = [32, 64, 128, 256]
254
255
    if not use_fp16:
        block_k_range.remove(16)  # BLOCK_K=16 not supported for fp8
256
257
258
259
260
261
    num_warps_range = [2, 4, 8]
    group_m_range = [1, 16, 32, 64]
    num_stage_range = [2, 3, 4, 5]
    # waves_per_eu_range = [0]
    # matrix_instr_nonkdim_range = [16, 32] if use_fp16 else []
    # kpack_range = [1, 2] if use_fp16 else []
262
263

    param_ranges = {
264
265
        "BLOCK_SIZE_M": block_m_range,
        "BLOCK_SIZE_N": block_n_range,
266
267
268
269
        "BLOCK_SIZE_K": block_k_range,
        "GROUP_SIZE_M": group_m_range,
        "num_warps": num_warps_range,
        "num_stages": num_stage_range,
270
        # "waves_per_eu": waves_per_eu_range,
271
    }
zhuwenwen's avatar
zhuwenwen committed
272
    if nn_moe:
273
274
275
276
277
278
        param_ranges["num_ldmatrixes"] = [1]
    
    # DCU currently does not support the following parameters
    # if use_fp16:
    #     param_ranges["matrix_instr_nonkdim"] = matrix_instr_nonkdim_range
    #     param_ranges["kpack"] = kpack_range
279
280
281
282

    return param_ranges


zhuwenwen's avatar
zhuwenwen committed
283
def get_configs_compute_bound(use_fp16, block_quant_shape, nn_moe: Optional[bool] = False) -> list[dict[str, int]]:
284
    configs: list[BenchmarkConfig] = []
285
286
287
    
    # 局部导入 current_platform
    from vllm.platforms import current_platform
288
289

    if current_platform.is_rocm():
zhuwenwen's avatar
zhuwenwen committed
290
        param_ranges = get_rocm_tuning_space(use_fp16, nn_moe)
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
    else:
        # Reduced search space for faster tuning.
        # TODO(woosuk): Increase the search space and use a performance model to
        # prune the search space.
        block_m_range = [16, 32, 64, 128, 256]
        block_n_range = [32, 64, 128, 256]
        block_k_range = [64, 128, 256]
        num_warps_range = [4, 8]
        group_m_range = [1, 16, 32, 64]
        num_stage_range = [2, 3, 4, 5]

        param_ranges = {
            "BLOCK_SIZE_M": block_m_range,
            "BLOCK_SIZE_N": block_n_range,
            "BLOCK_SIZE_K": block_k_range,
            "GROUP_SIZE_M": group_m_range,
            "num_warps": num_warps_range,
            "num_stages": num_stage_range,
        }

    keys, values = zip(*param_ranges.items())
    for config_values in product(*values):
        config = dict(zip(keys, config_values))
        configs.append(config)
315
316
317
318
319
320
321

    # Remove configs that are not compatible with fp8 block quantization
    # BLOCK_SIZE_K must be a multiple of block_k
    # BLOCK_SIZE_N must be a multiple of block_n
    if block_quant_shape is not None and not use_fp16:
        block_n, block_k = block_quant_shape[0], block_quant_shape[1]
        for config in configs[:]:
322
323
324
325
            if (
                config["BLOCK_SIZE_K"] % block_k != 0
                or config["BLOCK_SIZE_N"] % block_n != 0
            ):
326
                configs.remove(config)
327
328
329
    return configs


330
331
332
def prune_rocm_search_space(
    num_tokens, shard_intermediate_size, hidden_size, search_space, is_fp16, topk
):
333
334
    N1, K1 = shard_intermediate_size, hidden_size
    N2, K2 = hidden_size, shard_intermediate_size // 2
335
336
337
338
339
340
    pruned_space_1 = prune_rocm_configs(
        num_tokens * topk, N1, K1, search_space, is_fp16
    )
    pruned_space_2 = prune_rocm_configs(
        num_tokens * topk, N2, K2, search_space, is_fp16
    )
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
    search_space = merge_unique_dicts(pruned_space_1, pruned_space_2)
    return search_space


# The following code is inspired by ROCm/Triton GEMM tuning script:
# https://github.com/ROCm/triton/blob/triton-mlir/scripts/amd/gemm/tune_gemm.py#L89
def prune_rocm_configs(M, N, K, configs, is_fp16=True):
    pruned_configs = []
    elemBytes_a = 2 if is_fp16 else 1
    elemBytes_b = 2 if is_fp16 else 1

    mfma = 16 if M < 32 or N < 32 else 32

    # TODO (zhanglx): figure out the boundary between large and small gemms
    large_gemm = False
    if M >= 2048 and N >= 2048:
        large_gemm = True

    for config in configs:
        BLOCK_SIZE_M = config.get("BLOCK_SIZE_M")
        BLOCK_SIZE_N = config.get("BLOCK_SIZE_N")
        BLOCK_SIZE_K = config.get("BLOCK_SIZE_K")
        num_warps = config.get("num_warps")

365
366
367
368
369
        # DCU currently does not support matrix_instr_nonkdim param
        # if is_fp16:
        #     matrix_instr_nonkdim = config.get("matrix_instr_nonkdim")
        #     if matrix_instr_nonkdim > mfma:
        #         continue
370
371
372
373
374
375
376
377
        if mfma == 4 and BLOCK_SIZE_K < 64:
            continue
        # some layouts could not work properly in case
        # number elements per thread is less 1
        if BLOCK_SIZE_M * BLOCK_SIZE_N < 64:
            continue
        SPLIT_K = config.get("SPLIT_K", 1)
        GROUP_M = config.get("GROUP_SIZE_M")
378
379
380

        # DCU currently does not support matrix_instr_nonkdim param
        # if is_fp16:
zhuwenwen's avatar
zhuwenwen committed
381
382
383
384
        #     if (
        #         matrix_instr_nonkdim > BLOCK_SIZE_M
        #         or matrix_instr_nonkdim > BLOCK_SIZE_N
        #     ):
385
        #         continue
zhuwenwen's avatar
zhuwenwen committed
386
        #     if matrix_instr_nonkdim >= M and matrix_instr_nonkdim != BLOCK_SIZE_M:
387
        #         continue
zhuwenwen's avatar
zhuwenwen committed
388
        #     if matrix_instr_nonkdim >= N and matrix_instr_nonkdim != BLOCK_SIZE_N:
389
        #         continue
zhuwenwen's avatar
zhuwenwen committed
390
        
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
        # Skip BLOCK_SIZE that is too large compare to M/N
        # unless BLOCK_SIZE is already small enough
        if M * 2 < BLOCK_SIZE_M and BLOCK_SIZE_M != 16:
            continue
        if N * 2 < BLOCK_SIZE_N and BLOCK_SIZE_N != 16:
            continue
        # skip large split_k when not necessary
        if SPLIT_K != 1 and not need_split_k(M, N, K):
            continue
        # skip split_k that leads to EVEN_K = false
        leap = SPLIT_K * BLOCK_SIZE_K
        modv = K % leap
        if modv != 0:
            continue
        # skip large GROUP_M
        if GROUP_M * BLOCK_SIZE_M > M and GROUP_M != 1:
            continue
        # out of shared memory resource
        # TODO (zhanglx): This does not consider the LDS usage in the epilogue
410
411
412
413
        LDS = (
            BLOCK_SIZE_K * BLOCK_SIZE_M * elemBytes_a
            + BLOCK_SIZE_K * BLOCK_SIZE_N * elemBytes_b
        )
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
        if LDS > 65536:
            continue
        # Skip small block sizes and num_warps for large gemm
        # For fp16 and f8, we want to only use BLOCK_SIZE >= 64
        if large_gemm:
            if BLOCK_SIZE_M < 64 or BLOCK_SIZE_N < 64:
                continue
            if BLOCK_SIZE_K < 64:
                continue
            if num_warps < 4:
                continue

        pruned_configs.append(config)

    return pruned_configs


def need_split_k(SIZE_M, SIZE_N, SIZE_K):
    return (SIZE_M < 64 or SIZE_N < 64) and SIZE_K > 1024


def merge_unique_dicts(list1, list2):
    result = []
    combined_list = list1.copy()
    combined_list.extend(list2)
    for dictionary in combined_list:
        if dictionary not in result:
            result.append(dictionary)
    return result


445
446
@ray.remote(num_gpus=1)
class BenchmarkWorker:
王敏's avatar
王敏 committed
447
    def __init__(self, seed: int, device_id: int) -> None:
448
449
450
451
452
453
454
455
456
        from vllm.platforms import current_platform
        import os
        
        if current_platform.is_rocm():
            # In ROCm environment with Ray, let Ray handle device assignment
            # Don't manually set default device as it may conflict with Ray's device mapping
            pass
        else:
            torch.set_default_device("cuda:"+ str(device_id))
457
        current_platform.seed_everything(seed)
458
        self.seed = seed
459
        # Store the logical device ID for Ray
王敏's avatar
王敏 committed
460
        self.device_id = device_id
461
462
463
464
465
466
467
468
469

    def benchmark(
        self,
        num_tokens: int,
        num_experts: int,
        shard_intermediate_size: int,
        hidden_size: int,
        topk: int,
        dtype: torch.dtype,
470
471
        use_fp8_w8a8: bool,
        use_int8_w8a16: bool,
472
        block_quant_shape: list[int] = None,
473
        use_deep_gemm: bool = False,
474
        nn_moe: Optional[bool] = False,
475
    ) -> tuple[dict[str, int], float]:
476
477
        # 局部导入 current_platform
        from vllm.platforms import current_platform
478
        current_platform.seed_everything(self.seed)
479
        dtype_str = _get_config_dtype_str(
480
481
            dtype, use_int8_w8a16=use_int8_w8a16, use_fp8_w8a8=use_fp8_w8a8
        )
482
483
        # NOTE(woosuk): The current naming convention uses w2.shape[2], which
        # is the intermediate size after silu_and_mul.
484
485
        block_n = block_quant_shape[0] if block_quant_shape else None
        block_k = block_quant_shape[1] if block_quant_shape else None
486
        op_config = get_moe_configs(
487
            num_experts, shard_intermediate_size // 2, dtype_str, block_n, block_k, use_nn_moe=nn_moe
488
        )
489
        if op_config is None:
490
491
492
493
494
495
496
            config = get_default_config(
                num_tokens,
                num_experts,
                shard_intermediate_size,
                hidden_size,
                topk,
                dtype_str,
497
                block_quant_shape,
498
                use_nn_moe=nn_moe,
499
            )
500
        else:
501
502
503
504
505
506
507
508
509
510
511
512
513
514
            config = op_config[min(op_config.keys(), key=lambda x: abs(x - num_tokens))]
        kernel_time = benchmark_config(
            config,
            num_tokens,
            num_experts,
            shard_intermediate_size,
            hidden_size,
            topk,
            dtype,
            use_fp8_w8a8,
            use_int8_w8a16,
            num_iters=100,
            block_quant_shape=block_quant_shape,
            use_deep_gemm=use_deep_gemm,
zhuwenwen's avatar
zhuwenwen committed
515
            use_nn_moe=nn_moe
516
        )
517
518
519
520
521
522
523
524
525
526
        return config, kernel_time

    def tune(
        self,
        num_tokens: int,
        num_experts: int,
        shard_intermediate_size: int,
        hidden_size: int,
        topk: int,
        dtype: torch.dtype,
527
528
        use_fp8_w8a8: bool,
        use_int8_w8a16: bool,
529
        search_space: list[dict[str, int]],
530
        block_quant_shape: list[int],
531
        use_deep_gemm: bool,
王敏's avatar
王敏 committed
532
        nn_moe: Optional[bool] = False,
533
    ) -> dict[str, int]:
534
535
536
        from vllm.platforms import current_platform
        import os

537
538
        best_config = None
        best_time = float("inf")
539
540
        if current_platform.is_rocm():
            is_fp16 = not (use_fp8_w8a8 or use_int8_w8a16)
541
542
543
544
545
546
547
548
            search_space = prune_rocm_search_space(
                num_tokens,
                shard_intermediate_size,
                hidden_size,
                search_space,
                is_fp16,
                topk,
            )
549

550
551
        # In ROCm environments with Ray, device context is already handled by Ray
        # Using torch.cuda.device() may cause device ordinal conflicts
552
553
        need_device_guard = False
        if current_platform.is_rocm():
554
555
556
557
558
559
            # For ROCm with Ray, skip additional device context management
            need_device_guard = False
        else:
            # For other platforms, use device guard if needed
            visible_devices = os.environ.get("CUDA_VISIBLE_DEVICES", None)
            if visible_devices is not None and len(visible_devices.split(',')) > 1:
560
                need_device_guard = True
561

562
        with torch.cuda.device(self.device_id) if need_device_guard else nullcontext():
563
564
            for config in tqdm(search_space):
                try:
565
566
567
568
569
570
571
572
573
574
575
                    kernel_time = benchmark_config(
                        config,
                        num_tokens,
                        num_experts,
                        shard_intermediate_size,
                        hidden_size,
                        topk,
                        dtype,
                        use_fp8_w8a8,
                        use_int8_w8a16,
                        num_iters=20,
zhuwenwen's avatar
zhuwenwen committed
576
                        block_quant_shape=block_quant_shape,
zhuwenwen's avatar
zhuwenwen committed
577
578
                        use_deep_gemm=use_deep_gemm,
                        nn_moe=nn_moe)
579
580
581
582
583
584
585
                except triton.runtime.autotuner.OutOfResources:
                    # Some configurations may be invalid and fail to compile.
                    continue

                if kernel_time < best_time:
                    best_time = kernel_time
                    best_config = config
586
587
        now = datetime.now()
        print(f"{now.ctime()}] Completed tuning for batch_size={num_tokens}")
588
        assert best_config is not None
589
590
591
        return best_config


592
def sort_config(config: BenchmarkConfig) -> BenchmarkConfig:
593
594

    return {
595
596
597
598
599
600
        "BLOCK_SIZE_M": config["BLOCK_SIZE_M"],
        "BLOCK_SIZE_N": config["BLOCK_SIZE_N"],
        "BLOCK_SIZE_K": config["BLOCK_SIZE_K"],
        "GROUP_SIZE_M": config["GROUP_SIZE_M"],
        "num_warps": config["num_warps"],
        "num_stages": config["num_stages"],
zhuwenwen's avatar
zhuwenwen committed
601
602
603
        **(
            {"num_ldmatrixes": config["num_ldmatrixes"]} if "num_ldmatrixes" in config else {}
        ),
604
605
606
607
608
609
610
611
612
        **(
            {"waves_per_eu": config["waves_per_eu"]} if "waves_per_eu" in config else {}
        ),
        **(
            {"matrix_instr_nonkdim": config["matrix_instr_nonkdim"]}
            if "matrix_instr_nonkdim" in config
            else {}
        ),
        **({"kpack": config["kpack"]} if "kpack" in config else {}),
613
614
615
    }


616
617
618
619
620
621
622
623
624
def save_configs(
    configs: dict[int, BenchmarkConfig],
    num_experts: int,
    shard_intermediate_size: int,
    hidden_size: int,
    topk: int,
    dtype: torch.dtype,
    use_fp8_w8a8: bool,
    use_int8_w8a16: bool,
625
    block_quant_shape: list[int],
626
    save_dir: str,
zhuwenwen's avatar
zhuwenwen committed
627
    use_nn_moe: Optional[bool] = False,
628
) -> None:
629
    dtype_str = _get_config_dtype_str(
630
631
        dtype, use_int8_w8a16=use_int8_w8a16, use_fp8_w8a8=use_fp8_w8a8
    )
632

633
634
    # NOTE(woosuk): The current naming convention uses w2.shape[2], which
    # is the intermediate size after silu_and_mul.
635
    filename = get_config_file_name(
zhuwenwen's avatar
zhuwenwen committed
636
        num_experts, shard_intermediate_size // 2, dtype_str, block_quant_shape, use_nn_moe=use_nn_moe
637
    )
638
639
    os.makedirs(save_dir, exist_ok=True)
    filename = os.path.join(save_dir, filename)
640
641
    print(f"Writing best config to {filename}...")
    with open(filename, "w") as f:
642
        json.dump({"triton_version": triton.__version__, **configs}, f, indent=4)
643
644
645
        f.write("\n")


646
def get_weight_block_size_safety(config, default_value=None):
647
    quantization_config = getattr(config, "quantization_config", {})
648
    if isinstance(quantization_config, dict):
649
        return quantization_config.get("weight_block_size", default_value)
650
651
652
    return default_value


653
def main(args: argparse.Namespace):
654
655
    import os
    import logging
656

657
658
659
660
    from vllm.platforms import current_platform
    
    logger = logging.getLogger(__name__)

661
    print(args)
zhuwenwen's avatar
zhuwenwen committed
662
    
王敏's avatar
王敏 committed
663
    tp_size = args.tp_size
664
    config = get_config(model=args.model, trust_remote_code=args.trust_remote_code)
665
666
    if args.model_prefix:
        config = getattr(config, args.model_prefix)
王敏's avatar
王敏 committed
667

668
669
670
671
    if config.architectures[0] == "DbrxForCausalLM":
        E = config.ffn_config.moe_num_experts
        topk = config.ffn_config.moe_top_k
        intermediate_size = config.ffn_config.ffn_hidden_size
672
673
674
675
    elif config.architectures[0] == "JambaForCausalLM":
        E = config.num_experts
        topk = config.num_experts_per_tok
        intermediate_size = config.intermediate_size
Yuxuan Zhang's avatar
Yuxuan Zhang committed
676
677
    elif config.architectures[0] in (
        "DeepseekV2ForCausalLM",
678
679
        "DeepseekV3ForCausalLM",
        "DeepseekV32ForCausalLM",
Yuxuan Zhang's avatar
Yuxuan Zhang committed
680
681
        "Glm4MoeForCausalLM",
    ):
682
        E = config.n_routed_experts
王敏's avatar
王敏 committed
683
684
        topk = config.num_experts_per_tok
        intermediate_size = config.moe_intermediate_size
685
686
687
688
689
    elif config.architectures[0] in (
        "Qwen2MoeForCausalLM",
        "Qwen3MoeForCausalLM",
        "Qwen3NextForCausalLM",
    ):
王敏's avatar
王敏 committed
690
        E = config.num_experts
691
692
        topk = config.num_experts_per_tok
        intermediate_size = config.moe_intermediate_size
693
694
695
696
    elif config.architectures[0] in ("HunYuanMoEV1ForCausalLM"):
        E = config.num_experts
        topk = config.moe_topk[0]
        intermediate_size = config.moe_intermediate_size[0]
zhuwenwen's avatar
zhuwenwen committed
697
698
699
700
    elif config.architectures[0] in ("Step3VLForConditionalGeneration"):
        E = config.text_config.moe_num_experts
        topk = config.text_config.moe_top_k
        intermediate_size = config.text_config.moe_intermediate_size
701
    else:
702
703
        # Support for llama4
        config = config.get_text_config()
704
705
706
707
        # Default: Mixtral.
        E = config.num_local_experts
        topk = config.num_experts_per_tok
        intermediate_size = config.intermediate_size
708
709
    enable_ep = bool(args.enable_expert_parallel)
    if enable_ep:
zhuwenwen's avatar
zhuwenwen committed
710
711
        ensure_divisibility(E, tp_size, "Number of experts")
        E = E // tp_size
712
713
        shard_intermediate_size = 2 * intermediate_size
    else:
zhuwenwen's avatar
zhuwenwen committed
714
715
        ensure_divisibility(intermediate_size, tp_size, "intermediate_size")
        shard_intermediate_size = 2 * intermediate_size // tp_size
716
    hidden_size = config.hidden_size
717
    dtype = torch.float16 if current_platform.is_rocm() else config.torch_dtype
718
719
    use_fp8_w8a8 = args.dtype == "fp8_w8a8"
    use_int8_w8a16 = args.dtype == "int8_w8a16"
720
    block_quant_shape = get_weight_block_size_safety(config)
721
722

    if args.batch_size is None:
723
        batch_sizes = [
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
            1,
            2,
            4,
            8,
            16,
            24,
            32,
            48,
            64,
            96,
            128,
            256,
            512,
            1024,
            1536,
            2048,
            3072,
            4096,
742
        ]
743
    else:
744
        batch_sizes = args.batch_size
745

746
747
    use_deep_gemm = bool(args.use_deep_gemm)

748
749
750
751
    if current_platform.is_rocm() and "HIP_VISIBLE_DEVICES" in os.environ:
        # Ray will set ROCR_VISIBLE_DEVICES for device visibility
        logger.warning(
            "Ray uses ROCR_VISIBLE_DEVICES to control device accessibility."
752
753
            "Replacing HIP_VISIBLE_DEVICES with ROCR_VISIBLE_DEVICES."
        )
754
755
756
757
        val = os.environ["HIP_VISIBLE_DEVICES"]
        os.environ["ROCR_VISIBLE_DEVICES"] = val
        del os.environ["HIP_VISIBLE_DEVICES"]

zhuwenwen's avatar
zhuwenwen committed
758
    ray.init(address=None, ignore_reinit_error=True, num_gpus=args.num_gpus)
759
    num_gpus = int(ray.available_resources()["GPU"])
王敏's avatar
王敏 committed
760
    workers = [BenchmarkWorker.remote(args.seed, i) for i in range(num_gpus)]
761

762
    def _distribute(method: str, inputs: list[Any]) -> list[Any]:
763
764
765
766
767
768
769
770
771
772
773
        outputs = []
        worker_idx = 0
        for input_args in inputs:
            worker = workers[worker_idx]
            worker_method = getattr(worker, method)
            output = worker_method.remote(*input_args)
            outputs.append(output)
            worker_idx = (worker_idx + 1) % num_gpus
        return ray.get(outputs)

    if args.tune:
774
        is_fp16 = not (use_fp8_w8a8 or use_int8_w8a16)
zhuwenwen's avatar
zhuwenwen committed
775
        search_space = get_configs_compute_bound(is_fp16, block_quant_shape, args.nn_moe)
776
        print(f"Start tuning over {len(search_space)} configurations...")
777
778
779
780
781
        if use_deep_gemm:
            raise ValueError(
                "Tuning with --use-deep-gemm is not supported as it only tunes Triton "
                "kernels. Please remove the flag."
            )
782
783
        start = time.time()
        configs = _distribute(
784
785
786
787
788
789
790
791
792
793
794
795
796
797
            "tune",
            [
                (
                    batch_size,
                    E,
                    shard_intermediate_size,
                    hidden_size,
                    topk,
                    dtype,
                    use_fp8_w8a8,
                    use_int8_w8a16,
                    search_space,
                    block_quant_shape,
                    use_deep_gemm,
zhuwenwen's avatar
zhuwenwen committed
798
                    args.nn_moe,
799
800
801
802
                )
                for batch_size in batch_sizes
            ],
        )
803
        best_configs = {
804
            M: sort_config(config) for M, config in zip(batch_sizes, configs)
805
        }
806
807
808
809
810
811
812
813
814
815
        save_configs(
            best_configs,
            E,
            shard_intermediate_size,
            hidden_size,
            topk,
            dtype,
            use_fp8_w8a8,
            use_int8_w8a16,
            block_quant_shape,
816
            args.save_dir,
zhuwenwen's avatar
zhuwenwen committed
817
            use_nn_moe=args.nn_moe,
818
        )
819
820
821
        end = time.time()
        print(f"Tuning took {end - start:.2f} seconds")
    else:
822
        outputs = _distribute(
823
            "benchmark",
824
825
826
827
828
829
830
831
832
833
834
835
            [
                (
                    batch_size,
                    E,
                    shard_intermediate_size,
                    hidden_size,
                    topk,
                    dtype,
                    use_fp8_w8a8,
                    use_int8_w8a16,
                    block_quant_shape,
                    use_deep_gemm,
zhuwenwen's avatar
zhuwenwen committed
836
                    args.nn_moe,
837
838
839
840
                )
                for batch_size in batch_sizes
            ],
        )
841
842
843
844
845
846
847

        for batch_size, (config, kernel_time) in zip(batch_sizes, outputs):
            print(f"Batch size: {batch_size}, config: {config}")
            print(f"Kernel time: {kernel_time:.2f} us")


if __name__ == "__main__":
848
    parser = FlexibleArgumentParser()
849
850
851
852
853
854
    parser.add_argument(
        "--model", type=str, default="mistralai/Mixtral-8x7B-Instruct-v0.1"
    )
    parser.add_argument(
        "--tp-size", "-tp", "--tensor-parallel-size", type=int, default=2
    )
855
    parser.add_argument("--enable-expert-parallel", "-enable-ep", action="store_true")
856
857
858
    parser.add_argument(
        "--dtype", type=str, choices=["auto", "fp8_w8a8", "int8_w8a16"], default="auto"
    )
859
    parser.add_argument("--use-deep-gemm", action="store_true")
860
861
862
    parser.add_argument(
        "--save-dir", type=str, default="./", help="Directory to save tuned results"
    )
863
    parser.add_argument("--seed", type=int, default=0)
864
    parser.add_argument("--batch-size", type=int, nargs="+", required=False)
865
    parser.add_argument("--tune", action="store_true")
王敏's avatar
王敏 committed
866
    parser.add_argument("--nn-moe", action='store_true', default=False)
867
    parser.add_argument("--trust-remote-code", action="store_true")
868
    parser.add_argument("--model-prefix", type=str, required=False)
王敏's avatar
王敏 committed
869
    parser.add_argument("--num-gpus", type=int, default=1)
870
871
872
    args = parser.parse_args()

    main(args)