benchmark_moe.py 24.9 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
12
13
14
15
16
17

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

from vllm.model_executor.layers.fused_moe.fused_moe import *
18
from vllm.platforms import current_platform
19
from vllm.transformers_utils.config import get_config
20
from vllm.triton_utils import triton
21
from vllm.utils import FlexibleArgumentParser
22

23
FP8_DTYPE = current_platform.fp8_dtype()
24
25


26
def ensure_divisibility(numerator, denominator, text):
27
    """Ensure that numerator is divisible by the denominator."""
28
29
    assert numerator % denominator == 0, "{} {} is not divisible by tp {}.".format(
        text, numerator, denominator
30
31
32
    )


33
34
35
36
37
38
39
40
41
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


42
43
44
45
46
47
48
49
50
51
52
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,
53
    block_quant_shape: list[int] = None,
54
55
    use_deep_gemm: bool = False,
) -> float:
56
    init_dtype = torch.float16 if use_fp8_w8a8 else dtype
57
    x = torch.randn(num_tokens, hidden_size, dtype=dtype)
58
    if use_int8_w8a16:
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
        w1 = torch.randint(
            -127,
            127,
            (
                num_experts,
                shard_intermediate_size,
                hidden_size,
            ),
            dtype=torch.int8,
        )
        w2 = torch.randint(
            -127,
            127,
            (
                num_experts,
                hidden_size,
                shard_intermediate_size // 2,
            ),
            dtype=torch.int8,
        )
79
    else:
80
81
82
83
84
85
86
        w1 = torch.randn(
            num_experts, shard_intermediate_size, hidden_size, dtype=init_dtype
        )
        w2 = torch.randn(
            num_experts, hidden_size, shard_intermediate_size // 2, dtype=init_dtype
        )
    gating_output = torch.randn(num_iters, num_tokens, num_experts, dtype=torch.float32)
87
88
89
90
91

    w1_scale = None
    w2_scale = None
    a1_scale = None
    a2_scale = None
92
    if use_int8_w8a16:
93
94
95
        w1_scale = torch.randn(
            (num_experts, 2 * shard_intermediate_size), dtype=torch.float32
        )
96
        w2_scale = torch.randn((hidden_size, num_experts), dtype=torch.float32)
97
98
99
    if use_deep_gemm:
        # we use the default block shape for deepgemm
        block_quant_shape = [128, 128]
100
    if use_fp8_w8a8:
101
102
103
104
105
106
107
108
109
110
        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
111
112
113
114
115
116
117
118
            w1_scale = (
                torch.rand((E, n_tiles_w1, k_tiles_w1), dtype=torch.float32)
                * factor_for_scale
            )
            w2_scale = (
                torch.rand((E, n_tiles_w2, k_tiles_w2), dtype=torch.float32)
                * factor_for_scale
            )
119
120
121
122
        else:
            w1_scale = torch.randn(num_experts, dtype=torch.float32)
            w2_scale = torch.randn(num_experts, dtype=torch.float32)

123
124
125
        a1_scale = torch.randn(1, dtype=torch.float32)
        a2_scale = torch.randn(1, dtype=torch.float32)

126
127
        w1 = w1.to(FP8_DTYPE)
        w2 = w2.to(FP8_DTYPE)
128
129
130
131
132
133
134

    input_gating = torch.empty(num_tokens, num_experts, dtype=torch.float32)

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

    def run():
135
        from vllm.model_executor.layers.fused_moe import override_config
136

137
        with override_config(config):
138
            if use_deep_gemm:
139
                topk_weights, topk_ids, token_expert_indices = fused_topk(
140
141
                    x, input_gating, topk, False
                )
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
                return fused_experts(
                    x,
                    w1,
                    w2,
                    topk_weights,
                    topk_ids,
                    inplace=True,
                    use_fp8_w8a8=use_fp8_w8a8,
                    w1_scale=w1_scale,
                    w2_scale=w2_scale,
                    a1_scale=a1_scale,
                    a2_scale=a2_scale,
                    block_shape=block_quant_shape,
                    allow_deep_gemm=True,
                )
            else:
                fused_moe(
                    x,
                    w1,
                    w2,
                    input_gating,
                    topk,
                    renormalize=True,
                    inplace=True,
                    use_fp8_w8a8=use_fp8_w8a8,
                    use_int8_w8a16=use_int8_w8a16,
                    w1_scale=w1_scale,
                    w2_scale=w2_scale,
                    a1_scale=a1_scale,
                    a2_scale=a2_scale,
                    block_shape=block_quant_shape,
                )
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193

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

    # Capture 10 invocations with CUDA graph
    graph = torch.cuda.CUDAGraph()
    with torch.cuda.graph(graph):
        for _ in range(10):
            run()
    torch.cuda.synchronize()

    # Warmup
    for _ in range(5):
        graph.replay()
    torch.cuda.synchronize()

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

194
    latencies: list[float] = []
195
196
197
198
199
200
201
202
203
204
205
206
207
208
    for i in range(num_iters):
        prepare(i)
        torch.cuda.synchronize()

        start_event.record()
        graph.replay()
        end_event.record()
        end_event.synchronize()
        latencies.append(start_event.elapsed_time(end_event))
    avg = sum(latencies) / (num_iters * 10) * 1000  # us
    graph.reset()
    return avg


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
def get_rocm_tuning_space(use_fp16):
    block_mn_range = [16, 32, 64, 128, 256]
    block_k_range = [16, 32, 64, 128, 256]
    if not use_fp16:
        block_k_range.remove(16)  # BLOCK_K=16 not supported for fp8
    num_warps_range = [1, 2, 4, 8]
    group_m_range = [1, 4, 8, 16, 32]
    num_stage_range = [2]
    waves_per_eu_range = [0]
    matrix_instr_nonkdim_range = [16, 32] if use_fp16 else []
    kpack_range = [1, 2] if use_fp16 else []

    param_ranges = {
        "BLOCK_SIZE_M": block_mn_range,
        "BLOCK_SIZE_N": block_mn_range,
        "BLOCK_SIZE_K": block_k_range,
        "GROUP_SIZE_M": group_m_range,
        "num_warps": num_warps_range,
        "num_stages": num_stage_range,
        "waves_per_eu": waves_per_eu_range,
    }
    if use_fp16:
        param_ranges["matrix_instr_nonkdim"] = matrix_instr_nonkdim_range
        param_ranges["kpack"] = kpack_range

    return param_ranges


237
def get_configs_compute_bound(use_fp16, block_quant_shape) -> list[dict[str, int]]:
238
    configs: list[BenchmarkConfig] = []
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265

    if current_platform.is_rocm():
        param_ranges = get_rocm_tuning_space(use_fp16)
    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)
266
267
268
269
270
271
272

    # 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[:]:
273
274
275
276
            if (
                config["BLOCK_SIZE_K"] % block_k != 0
                or config["BLOCK_SIZE_N"] % block_n != 0
            ):
277
                configs.remove(config)
278
279
280
    return configs


281
282
283
def prune_rocm_search_space(
    num_tokens, shard_intermediate_size, hidden_size, search_space, is_fp16, topk
):
284
285
    N1, K1 = shard_intermediate_size, hidden_size
    N2, K2 = hidden_size, shard_intermediate_size // 2
286
287
288
289
290
291
    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
    )
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
    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")

        if is_fp16:
            matrix_instr_nonkdim = config.get("matrix_instr_nonkdim")
            if matrix_instr_nonkdim > mfma:
                continue
        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")
        if is_fp16:
329
330
331
332
            if (
                matrix_instr_nonkdim > BLOCK_SIZE_M
                or matrix_instr_nonkdim > BLOCK_SIZE_N
            ):
333
                continue
334
            if matrix_instr_nonkdim >= M and matrix_instr_nonkdim != BLOCK_SIZE_M:
335
                continue
336
            if matrix_instr_nonkdim >= N and matrix_instr_nonkdim != BLOCK_SIZE_N:
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
                continue
        # 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
357
358
359
360
        LDS = (
            BLOCK_SIZE_K * BLOCK_SIZE_M * elemBytes_a
            + BLOCK_SIZE_K * BLOCK_SIZE_N * elemBytes_b
        )
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
        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


392
393
394
395
@ray.remote(num_gpus=1)
class BenchmarkWorker:
    def __init__(self, seed: int) -> None:
        torch.set_default_device("cuda")
396
        current_platform.seed_everything(seed)
397
        self.seed = seed
398
399
400
401
        # Get the device ID to allocate tensors and kernels
        # on the respective GPU. This is required for Ray to work
        # correctly with multi-GPU tuning on the ROCm platform.
        self.device_id = int(ray.get_gpu_ids()[0])
402
403
404
405
406
407
408
409
410

    def benchmark(
        self,
        num_tokens: int,
        num_experts: int,
        shard_intermediate_size: int,
        hidden_size: int,
        topk: int,
        dtype: torch.dtype,
411
412
        use_fp8_w8a8: bool,
        use_int8_w8a16: bool,
413
        block_quant_shape: list[int] = None,
414
        use_deep_gemm: bool = False,
415
    ) -> tuple[dict[str, int], float]:
416
        current_platform.seed_everything(self.seed)
417
418
419
        dtype_str = get_config_dtype_str(
            dtype, use_int8_w8a16=use_int8_w8a16, use_fp8_w8a8=use_fp8_w8a8
        )
420
421
        # NOTE(woosuk): The current naming convention uses w2.shape[2], which
        # is the intermediate size after silu_and_mul.
422
423
424
        op_config = get_moe_configs(
            num_experts, shard_intermediate_size // 2, dtype_str
        )
425
        if op_config is None:
426
427
428
429
430
431
432
433
            config = get_default_config(
                num_tokens,
                num_experts,
                shard_intermediate_size,
                hidden_size,
                topk,
                dtype_str,
            )
434
        else:
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
            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,
        )
450
451
452
453
454
455
456
457
458
459
        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,
460
461
        use_fp8_w8a8: bool,
        use_int8_w8a16: bool,
462
        search_space: list[dict[str, int]],
463
        block_quant_shape: list[int],
464
        use_deep_gemm: bool,
465
    ) -> dict[str, int]:
466
467
        best_config = None
        best_time = float("inf")
468
469
        if current_platform.is_rocm():
            is_fp16 = not (use_fp8_w8a8 or use_int8_w8a16)
470
471
472
473
474
475
476
477
            search_space = prune_rocm_search_space(
                num_tokens,
                shard_intermediate_size,
                hidden_size,
                search_space,
                is_fp16,
                topk,
            )
478

479
480
481
482
483
484
        need_device_guard = False
        if current_platform.is_rocm():
            visible_device = os.environ.get("ROCR_VISIBLE_DEVICES", None)
            if visible_device != f"{self.device_id}":
                need_device_guard = True

485
        with torch.cuda.device(self.device_id) if need_device_guard else nullcontext():
486
487
            for config in tqdm(search_space):
                try:
488
489
490
491
492
493
494
495
496
497
498
                    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,
499
                        block_quant_shape=block_quant_shape,
500
501
                        use_deep_gemm=use_deep_gemm,
                    )
502
503
504
505
506
507
508
                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
509
510
        now = datetime.now()
        print(f"{now.ctime()}] Completed tuning for batch_size={num_tokens}")
511
        assert best_config is not None
512
513
514
        return best_config


515
def sort_config(config: BenchmarkConfig) -> BenchmarkConfig:
516
    return {
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
        "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"],
        **(
            {"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 {}),
532
533
534
    }


535
536
537
538
539
540
541
542
543
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,
544
    block_quant_shape: list[int],
545
    save_dir: str,
546
547
548
549
) -> None:
    dtype_str = get_config_dtype_str(
        dtype, use_int8_w8a16=use_int8_w8a16, use_fp8_w8a8=use_fp8_w8a8
    )
550

551
552
    # NOTE(woosuk): The current naming convention uses w2.shape[2], which
    # is the intermediate size after silu_and_mul.
553
554
555
    filename = get_config_file_name(
        num_experts, shard_intermediate_size // 2, dtype_str, block_quant_shape
    )
556
557
    os.makedirs(save_dir, exist_ok=True)
    filename = os.path.join(save_dir, filename)
558
559
560
561
562
563
    print(f"Writing best config to {filename}...")
    with open(filename, "w") as f:
        json.dump(configs, f, indent=4)
        f.write("\n")


564
def get_weight_block_size_safety(config, default_value=None):
565
    quantization_config = getattr(config, "quantization_config", {})
566
    if isinstance(quantization_config, dict):
567
        return quantization_config.get("weight_block_size", default_value)
568
569
570
    return default_value


571
572
def main(args: argparse.Namespace):
    print(args)
573

574
    config = get_config(model=args.model, trust_remote_code=args.trust_remote_code)
575
576
577
    if args.model_prefix:
        config = getattr(config, args.model_prefix)

578
579
580
581
    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
582
583
584
585
    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
586
587
588
589
590
    elif config.architectures[0] in (
        "DeepseekV3ForCausalLM",
        "DeepseekV2ForCausalLM",
        "Glm4MoeForCausalLM",
    ):
591
592
593
        E = config.n_routed_experts
        topk = config.num_experts_per_tok
        intermediate_size = config.moe_intermediate_size
594
    elif config.architectures[0] in ("Qwen2MoeForCausalLM", "Qwen3MoeForCausalLM"):
595
596
597
        E = config.num_experts
        topk = config.num_experts_per_tok
        intermediate_size = config.moe_intermediate_size
598
599
600
601
    elif config.architectures[0] in ("HunYuanMoEV1ForCausalLM"):
        E = config.num_experts
        topk = config.moe_topk[0]
        intermediate_size = config.moe_intermediate_size[0]
602
    else:
603
604
        # Support for llama4
        config = config.get_text_config()
605
606
607
608
        # Default: Mixtral.
        E = config.num_local_experts
        topk = config.num_experts_per_tok
        intermediate_size = config.intermediate_size
609
610
611
612
613
614
615
    enable_ep = bool(args.enable_expert_parallel)
    if enable_ep:
        ensure_divisibility(E, args.tp_size, "Number of experts")
        E = E // args.tp_size
        shard_intermediate_size = 2 * intermediate_size
    else:
        ensure_divisibility(intermediate_size, args.tp_size, "intermediate_size")
616
617
        shard_intermediate_size = 2 * intermediate_size // args.tp_size
    hidden_size = config.hidden_size
618
    dtype = torch.float16 if current_platform.is_rocm() else config.torch_dtype
619
620
    use_fp8_w8a8 = args.dtype == "fp8_w8a8"
    use_int8_w8a16 = args.dtype == "int8_w8a16"
621
    block_quant_shape = get_weight_block_size_safety(config)
622
623

    if args.batch_size is None:
624
        batch_sizes = [
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
            1,
            2,
            4,
            8,
            16,
            24,
            32,
            48,
            64,
            96,
            128,
            256,
            512,
            1024,
            1536,
            2048,
            3072,
            4096,
643
        ]
644
    else:
645
        batch_sizes = args.batch_size
646

647
648
    use_deep_gemm = bool(args.use_deep_gemm)

649
650
651
652
    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."
653
654
            "Replacing HIP_VISIBLE_DEVICES with ROCR_VISIBLE_DEVICES."
        )
655
656
657
658
        val = os.environ["HIP_VISIBLE_DEVICES"]
        os.environ["ROCR_VISIBLE_DEVICES"] = val
        del os.environ["HIP_VISIBLE_DEVICES"]

659
660
661
662
    ray.init()
    num_gpus = int(ray.available_resources()["GPU"])
    workers = [BenchmarkWorker.remote(args.seed) for _ in range(num_gpus)]

663
    def _distribute(method: str, inputs: list[Any]) -> list[Any]:
664
665
666
667
668
669
670
671
672
673
674
        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:
675
        is_fp16 = not (use_fp8_w8a8 or use_int8_w8a16)
676
        search_space = get_configs_compute_bound(is_fp16, block_quant_shape)
677
678
679
680
        print(f"Start tuning over {len(search_space)} configurations...")

        start = time.time()
        configs = _distribute(
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
            "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,
                )
                for batch_size in batch_sizes
            ],
        )
699
        best_configs = {
700
            M: sort_config(config) for M, config in zip(batch_sizes, configs)
701
        }
702
703
704
705
706
707
708
709
710
711
        save_configs(
            best_configs,
            E,
            shard_intermediate_size,
            hidden_size,
            topk,
            dtype,
            use_fp8_w8a8,
            use_int8_w8a16,
            block_quant_shape,
712
            args.save_dir,
713
        )
714
715
716
        end = time.time()
        print(f"Tuning took {end - start:.2f} seconds")
    else:
717
        outputs = _distribute(
718
            "benchmark",
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
            [
                (
                    batch_size,
                    E,
                    shard_intermediate_size,
                    hidden_size,
                    topk,
                    dtype,
                    use_fp8_w8a8,
                    use_int8_w8a16,
                    block_quant_shape,
                    use_deep_gemm,
                )
                for batch_size in batch_sizes
            ],
        )
735
736
737
738
739
740
741

        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__":
742
    parser = FlexibleArgumentParser()
743
744
745
746
747
748
    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
    )
749
    parser.add_argument("--enable-expert-parallel", "-enable-ep", action="store_true")
750
751
752
    parser.add_argument(
        "--dtype", type=str, choices=["auto", "fp8_w8a8", "int8_w8a16"], default="auto"
    )
753
    parser.add_argument("--use-deep-gemm", action="store_true")
754
755
756
    parser.add_argument(
        "--save-dir", type=str, default="./", help="Directory to save tuned results"
    )
757
    parser.add_argument("--seed", type=int, default=0)
758
    parser.add_argument("--batch-size", type=int, nargs="+", required=False)
759
    parser.add_argument("--tune", action="store_true")
760
    parser.add_argument("--trust-remote-code", action="store_true")
761
    parser.add_argument("--model-prefix", type=str, required=False)
762
763
764
    args = parser.parse_args()

    main(args)