benchmark_moe.py 34.6 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 gc
6
import json
7
import os
8
import time
9
from contextlib import nullcontext
10
from datetime import datetime
11
from itertools import product
12
from typing import Any, TypedDict
13
14
15
16
17

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

18
from vllm.model_executor.layers.fused_moe import fused_topk
19
from vllm.model_executor.layers.fused_moe.activation import MoEActivation
20
from vllm.model_executor.layers.fused_moe.config import (
21
22
    FusedMoEConfig,
    FusedMoEParallelConfig,
23
    FusedMoEQuantConfig,
24
    RoutingMethodType,
25
26
    _get_config_dtype_str,
)
27
from vllm.model_executor.layers.fused_moe.fused_moe import *
28
29
30
from vllm.model_executor.layers.fused_moe.triton_deep_gemm_moe import (
    TritonOrDeepGemmExperts,
)
31
from vllm.transformers_utils.config import get_config
32
from vllm.triton_utils import triton
33
from vllm.utils.argparse_utils import FlexibleArgumentParser
34
from vllm.utils.torch_utils import set_random_seed
35

36
FP8_DTYPE = current_platform.fp8_dtype()
37

38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
# Default interval for clearing Triton JIT cache during tuning
# Set to 0 to disable automatic cache clearing
_CACHE_CLEAR_INTERVAL_ENV = "VLLM_MOE_TUNE_CACHE_CLEAR_INTERVAL"
TRITON_CACHE_CLEAR_INTERVAL = int(os.environ.get(_CACHE_CLEAR_INTERVAL_ENV, "50"))


def clear_triton_cache():
    """Clear Triton JIT compilation cache and Python/CUDA memory.

    This helps prevent OOM during tuning with large models (many experts).
    """
    # Force Python garbage collection
    gc.collect()

    # Clear CUDA memory cache
    if torch.cuda.is_available():
        torch.cuda.empty_cache()

    # Try to clear Triton's runtime cache
    try:
        if (
            hasattr(triton, "runtime")
            and hasattr(triton.runtime, "cache")
            and hasattr(triton.runtime.cache, "clear")
        ):
            triton.runtime.cache.clear()
    except ImportError:
        # Triton not installed, skip cache clearing
        pass
    except AttributeError:
        # Triton version doesn't have expected cache API
        pass
    except Exception as e:
        print(f"Warning: Failed to clear Triton cache: {e}")

    # Additional garbage collection after clearing caches
    gc.collect()

76

77
def ensure_divisibility(numerator, denominator, text):
78
    """Ensure that numerator is divisible by the denominator."""
79
80
    assert numerator % denominator == 0, "{} {} is not divisible by tp {}.".format(
        text, numerator, denominator
81
82
83
    )


84
85
86
87
88
89
90
91
92
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


93
94
95
96
97
98
99
100
101
102
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,
103
    use_int4_w4a16: bool = False,
104
    num_iters: int = 100,
105
    block_quant_shape: list[int] = None,
106
107
    use_deep_gemm: bool = False,
) -> float:
108
    init_dtype = torch.float16 if use_fp8_w8a8 else dtype
109
    x = torch.randn(num_tokens, hidden_size, dtype=dtype)
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
    if use_int4_w4a16:
        # Int4 packed weights: 2 int4 values per uint8 byte
        # K dimension is packed (halved)
        intermediate_size = shard_intermediate_size // 2  # after silu_and_mul
        w1 = torch.randint(
            0,
            255,
            (
                num_experts,
                shard_intermediate_size,
                hidden_size // 2,  # int4 packing
            ),
            dtype=torch.uint8,
        )
        w2 = torch.randint(
            0,
            255,
            (
                num_experts,
                hidden_size,
                intermediate_size // 2,  # int4 packing
            ),
            dtype=torch.uint8,
        )
    elif use_int8_w8a16:
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
        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,
        )
155
    else:
156
157
158
159
160
161
162
        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)
163
164
165
166
167

    w1_scale = None
    w2_scale = None
    a1_scale = None
    a2_scale = None
168
169
170
171
172
173
174
175
176
177
178
179
180
181
    if use_int4_w4a16:
        if block_quant_shape is None:
            raise ValueError("block_quant_shape is required for int4_w4a16")
        group_size = block_quant_shape[1]
        # Scales shape: (E, N, K // group_size) in fp16
        w1_scale = torch.rand(
            (num_experts, shard_intermediate_size, hidden_size // group_size),
            dtype=dtype,
        )
        w2_scale = torch.rand(
            (num_experts, hidden_size, intermediate_size // group_size),
            dtype=dtype,
        )
    elif use_int8_w8a16:
182
183
184
        w1_scale = torch.randn(
            (num_experts, 2 * shard_intermediate_size), dtype=torch.float32
        )
185
        w2_scale = torch.randn((hidden_size, num_experts), dtype=torch.float32)
186
187
188
    if use_deep_gemm:
        # we use the default block shape for deepgemm
        block_quant_shape = [128, 128]
189
    if use_fp8_w8a8:
190
191
192
193
194
195
196
197
198
199
        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
200
201
202
203
204
205
206
207
            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
            )
208
209
210
211
        else:
            w1_scale = torch.randn(num_experts, dtype=torch.float32)
            w2_scale = torch.randn(num_experts, dtype=torch.float32)

212
213
214
        a1_scale = torch.randn(1, dtype=torch.float32)
        a2_scale = torch.randn(1, dtype=torch.float32)

215
216
        w1 = w1.to(FP8_DTYPE)
        w2 = w2.to(FP8_DTYPE)
217
218
219
220
221
222
223

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

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

    def run():
224
        from vllm.model_executor.layers.fused_moe import override_config
225

226
227
228
229
230
231
232
233
234
235
236
237
238
239
        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,
240
            weight_dtype="int4" if use_int4_w4a16 else None,
241
242
        )

243
244
245
246
247
248
249
250
251
252
253
        deep_gemm_experts = None
        if use_deep_gemm:
            deep_gemm_experts = mk.FusedMoEModularKernel(
                prepare_finalize=MoEPrepareAndFinalizeNoEP(),
                fused_experts=TritonOrDeepGemmExperts(
                    moe_config=FusedMoEConfig(
                        num_experts=num_experts,
                        experts_per_token=topk,
                        hidden_dim=hidden_size,
                        intermediate_size_per_partition=shard_intermediate_size,
                        num_local_experts=num_experts,
254
255
                        num_logical_experts=num_experts,
                        activation=MoEActivation.SILU,
256
257
258
259
260
261
                        moe_parallel_config=FusedMoEParallelConfig.make_no_parallel(),
                        in_dtype=init_dtype,
                        routing_method=RoutingMethodType.TopK,
                        device="cuda",
                    ),
                    quant_config=quant_config,
262
                ),
263
            )
264

265
        with override_config(config):
266
267
268
            topk_weights, topk_ids, token_expert_indices = fused_topk(
                x, input_gating, topk, renormalize=not use_deep_gemm
            )
269

270
            inplace = not disable_inplace()
271
272
            if use_deep_gemm:
                return deep_gemm_experts(
273
                    x, w1, w2, topk_weights, topk_ids, inplace=inplace
274
                )
275
276
277
278
279
280
            return fused_experts(
                x,
                w1,
                w2,
                topk_weights,
                topk_ids,
281
                inplace=inplace,
282
283
                quant_config=quant_config,
            )
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300

    # 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()

301
302
    start_event = torch.Event(enable_timing=True)
    end_event = torch.Event(enable_timing=True)
303

304
    latencies: list[float] = []
305
306
307
308
309
310
311
312
313
314
315
316
317
318
    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


319
320
321
322
323
324
325
326
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]
327
    waves_per_eu_range = [0, 1, 2, 4]
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
    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


347
def get_configs_compute_bound(use_fp16, block_quant_shape) -> list[dict[str, int]]:
348
    configs: list[BenchmarkConfig] = []
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

    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)
376
377
378
379
380
381
382

    # 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[:]:
383
384
385
386
            if (
                config["BLOCK_SIZE_K"] % block_k != 0
                or config["BLOCK_SIZE_N"] % block_n != 0
            ):
387
                configs.remove(config)
388
389
390
    return configs


391
392
393
def prune_rocm_search_space(
    num_tokens, shard_intermediate_size, hidden_size, search_space, is_fp16, topk
):
394
395
    N1, K1 = shard_intermediate_size, hidden_size
    N2, K2 = hidden_size, shard_intermediate_size // 2
396
397
398
399
400
401
    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
    )
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
    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:
439
440
441
442
            if (
                matrix_instr_nonkdim > BLOCK_SIZE_M
                or matrix_instr_nonkdim > BLOCK_SIZE_N
            ):
443
                continue
444
            if matrix_instr_nonkdim >= M and matrix_instr_nonkdim != BLOCK_SIZE_M:
445
                continue
446
            if matrix_instr_nonkdim >= N and matrix_instr_nonkdim != BLOCK_SIZE_N:
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
                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
467
468
469
470
        LDS = (
            BLOCK_SIZE_K * BLOCK_SIZE_M * elemBytes_a
            + BLOCK_SIZE_K * BLOCK_SIZE_N * elemBytes_b
        )
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
        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


502
503
504
505
@ray.remote(num_gpus=1)
class BenchmarkWorker:
    def __init__(self, seed: int) -> None:
        torch.set_default_device("cuda")
506
        set_random_seed(seed)
507
        self.seed = seed
508
509
510
511
        # 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])
512
513
514
515
516
517
518
519
520

    def benchmark(
        self,
        num_tokens: int,
        num_experts: int,
        shard_intermediate_size: int,
        hidden_size: int,
        topk: int,
        dtype: torch.dtype,
521
522
        use_fp8_w8a8: bool,
        use_int8_w8a16: bool,
523
        use_int4_w4a16: bool = False,
524
        block_quant_shape: list[int] = None,
525
        use_deep_gemm: bool = False,
526
    ) -> tuple[dict[str, int], float]:
527
528
        # local import to allow serialization by ray

529
        set_random_seed(self.seed)
530
        dtype_str = _get_config_dtype_str(
531
532
533
534
            dtype,
            use_int8_w8a16=use_int8_w8a16,
            use_fp8_w8a8=use_fp8_w8a8,
            use_int4_w4a16=use_int4_w4a16,
535
        )
536
537
        # NOTE(woosuk): The current naming convention uses w2.shape[2], which
        # is the intermediate size after silu_and_mul.
538
539
        block_n = block_quant_shape[0] if block_quant_shape else None
        block_k = block_quant_shape[1] if block_quant_shape else None
540
        op_config = get_moe_configs(
541
            num_experts, shard_intermediate_size // 2, dtype_str, block_n, block_k
542
        )
543
        if op_config is None:
544
545
546
547
548
549
550
            config = get_default_config(
                num_tokens,
                num_experts,
                shard_intermediate_size,
                hidden_size,
                topk,
                dtype_str,
551
                block_quant_shape,
552
            )
553
        else:
554
555
556
557
558
559
560
561
562
563
564
            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,
565
            use_int4_w4a16=use_int4_w4a16,
566
567
568
569
            num_iters=100,
            block_quant_shape=block_quant_shape,
            use_deep_gemm=use_deep_gemm,
        )
570
571
572
573
574
575
576
577
578
579
        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,
580
581
        use_fp8_w8a8: bool,
        use_int8_w8a16: bool,
582
        use_int4_w4a16: bool,
583
        search_space: list[dict[str, int]],
584
        block_quant_shape: list[int],
585
        use_deep_gemm: bool,
586
    ) -> dict[str, int]:
587
588
589
        # local import to allow serialization by ray
        from vllm.platforms import current_platform

590
591
        best_config = None
        best_time = float("inf")
592
        if current_platform.is_rocm():
593
            is_fp16 = not (use_fp8_w8a8 or use_int8_w8a16 or use_int4_w4a16)
594
595
596
597
598
599
600
601
            search_space = prune_rocm_search_space(
                num_tokens,
                shard_intermediate_size,
                hidden_size,
                search_space,
                is_fp16,
                topk,
            )
602

603
604
605
606
607
608
        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

609
        with torch.cuda.device(self.device_id) if need_device_guard else nullcontext():
610
            for idx, config in enumerate(tqdm(search_space)):
611
                try:
612
613
614
615
616
617
618
619
620
621
                    kernel_time = benchmark_config(
                        config,
                        num_tokens,
                        num_experts,
                        shard_intermediate_size,
                        hidden_size,
                        topk,
                        dtype,
                        use_fp8_w8a8,
                        use_int8_w8a16,
622
                        use_int4_w4a16,
623
                        num_iters=20,
624
                        block_quant_shape=block_quant_shape,
625
626
                        use_deep_gemm=use_deep_gemm,
                    )
627
628
629
630
631
632
633
                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
634
635
636
637
638
639
640
641
642
643
644
645
646

                # Periodically clear Triton JIT cache to prevent OOM
                # This is especially important for large models with many experts
                if (
                    TRITON_CACHE_CLEAR_INTERVAL > 0
                    and idx > 0
                    and idx % TRITON_CACHE_CLEAR_INTERVAL == 0
                ):
                    clear_triton_cache()

        # Final cleanup after tuning completes
        clear_triton_cache()

647
648
        now = datetime.now()
        print(f"{now.ctime()}] Completed tuning for batch_size={num_tokens}")
649
        assert best_config is not None
650
651
652
        return best_config


653
def sort_config(config: BenchmarkConfig) -> BenchmarkConfig:
654
    return {
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
        "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 {}),
670
        **({"SPLIT_K": config["SPLIT_K"]} if "SPLIT_K" in config else {}),
671
672
673
    }


674
675
676
677
678
679
680
681
682
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,
683
    use_int4_w4a16: bool,
684
    block_quant_shape: list[int],
685
    save_dir: str,
686
) -> None:
687
    dtype_str = _get_config_dtype_str(
688
689
690
691
        dtype,
        use_int8_w8a16=use_int8_w8a16,
        use_fp8_w8a8=use_fp8_w8a8,
        use_int4_w4a16=use_int4_w4a16,
692
    )
693

694
695
    # NOTE(woosuk): The current naming convention uses w2.shape[2], which
    # is the intermediate size after silu_and_mul.
696
697
698
    filename = get_config_file_name(
        num_experts, shard_intermediate_size // 2, dtype_str, block_quant_shape
    )
699
700
    os.makedirs(save_dir, exist_ok=True)
    filename = os.path.join(save_dir, filename)
701
702
    print(f"Writing best config to {filename}...")
    with open(filename, "w") as f:
703
        json.dump({"triton_version": triton.__version__, **configs}, f, indent=4)
704
705
706
        f.write("\n")


707
708
709
710
711
712
713
714
715
716
def get_compressed_tensors_block_structure(config, default_value=None):
    config_groups = config.get("config_groups", {})
    if len(config_groups) != 1:
        return default_value
    group = next(iter(config_groups.values()))
    weights = group.get("weights", {})
    block_structure = weights.get("block_structure", default_value)
    return block_structure


717
def get_weight_block_size_safety(config, default_value=None):
718
    quantization_config = getattr(config, "quantization_config", {})
719
    if isinstance(quantization_config, dict):
720
721
722
723
724
        if "weight_block_size" in quantization_config:
            return quantization_config["weight_block_size"]
        return get_compressed_tensors_block_structure(
            quantization_config, default_value
        )
725
726
727
    return default_value


728
def get_model_params(config):
729
730
731
732
    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
733
        hidden_size = config.hidden_size
734
735
736
737
    elif config.architectures[0] == "JambaForCausalLM":
        E = config.num_experts
        topk = config.num_experts_per_tok
        intermediate_size = config.intermediate_size
738
        hidden_size = config.hidden_size
Yuxuan Zhang's avatar
Yuxuan Zhang committed
739
740
    elif config.architectures[0] in (
        "DeepseekV2ForCausalLM",
741
742
        "DeepseekV3ForCausalLM",
        "DeepseekV32ForCausalLM",
Jee Jee Li's avatar
Jee Jee Li committed
743
        "GlmMoeDsaForCausalLM",
Yuxuan Zhang's avatar
Yuxuan Zhang committed
744
        "Glm4MoeForCausalLM",
745
        "Glm4MoeLiteForCausalLM",
746
        "NemotronHForCausalLM",
747
        "MistralLarge3ForCausalLM",
Yuxuan Zhang's avatar
Yuxuan Zhang committed
748
    ):
749
750
751
        E = config.n_routed_experts
        topk = config.num_experts_per_tok
        intermediate_size = config.moe_intermediate_size
752
        hidden_size = config.hidden_size
753
754
755
756
757
    elif config.architectures[0] in (
        "Qwen2MoeForCausalLM",
        "Qwen3MoeForCausalLM",
        "Qwen3NextForCausalLM",
    ):
758
759
760
        E = config.num_experts
        topk = config.num_experts_per_tok
        intermediate_size = config.moe_intermediate_size
761
762
763
764
765
766
767
        hidden_size = config.hidden_size
    elif config.architectures[0] == "Qwen3VLMoeForConditionalGeneration":
        text_config = config.get_text_config()
        E = text_config.num_experts
        topk = text_config.num_experts_per_tok
        intermediate_size = text_config.moe_intermediate_size
        hidden_size = text_config.hidden_size
768
    elif config.architectures[0] == "HunYuanMoEV1ForCausalLM":
769
770
771
        E = config.num_experts
        topk = config.moe_topk[0]
        intermediate_size = config.moe_intermediate_size[0]
772
        hidden_size = config.hidden_size
773
    elif config.architectures[0] == "Qwen3OmniMoeForConditionalGeneration":
774
775
776
777
        E = config.thinker_config.text_config.num_experts
        topk = config.thinker_config.text_config.num_experts_per_tok
        intermediate_size = config.thinker_config.text_config.moe_intermediate_size
        hidden_size = config.thinker_config.text_config.hidden_size
778
779
780
781
    elif config.architectures[0] == "PixtralForConditionalGeneration":
        # Pixtral can contain different LLM architectures,
        # recurse to get their parameters
        return get_model_params(config.get_text_config())
782
    else:
783
784
        # Support for llama4
        config = config.get_text_config()
785
786
787
788
        # Default: Mixtral.
        E = config.num_local_experts
        topk = config.num_experts_per_tok
        intermediate_size = config.intermediate_size
789
        hidden_size = config.hidden_size
790
791
792
    return E, topk, intermediate_size, hidden_size


793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
def get_quantization_group_size(config) -> int | None:
    """Extract the quantization group size from the HF model config.

    This reads directly from the HuggingFace config object (as returned by
    ``get_config()``), not from vLLM's quantization config classes.

    Supports AWQ/GPTQ-style configs (direct 'group_size' key) and
    compressed-tensors configs (nested inside 'config_groups').
    """
    quantization_config = getattr(config, "quantization_config", {})
    if not isinstance(quantization_config, dict):
        return None
    # AWQ / GPTQ style: group_size is a top-level key
    gs = quantization_config.get("group_size")
    if gs is not None:
        return gs
    # compressed-tensors style: group_size is nested in config_groups
    config_groups = quantization_config.get("config_groups", {})
    if not isinstance(config_groups, dict):
        return None
    for group_cfg in config_groups.values():
        if not isinstance(group_cfg, dict):
            continue
        weights = group_cfg.get("weights", {})
        if not isinstance(weights, dict):
            continue
        gs = weights.get("group_size")
        if gs is not None:
            return gs
    return None


825
826
827
828
829
830
831
def main(args: argparse.Namespace):
    print(args)

    config = get_config(model=args.model, trust_remote_code=args.trust_remote_code)
    if args.model_prefix:
        config = getattr(config, args.model_prefix)
    E, topk, intermediate_size, hidden_size = get_model_params(config)
832
833
834
835
836
837
838
    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")
839
        shard_intermediate_size = 2 * intermediate_size // args.tp_size
840
    dtype = torch.float16 if current_platform.is_rocm() else config.dtype
841
842
    use_fp8_w8a8 = args.dtype == "fp8_w8a8"
    use_int8_w8a16 = args.dtype == "int8_w8a16"
843
    use_int4_w4a16 = args.dtype == "int4_w4a16"
844
    block_quant_shape = get_weight_block_size_safety(config)
845
846
847
848
849
850
851
852
853
854
855
856
    if use_int4_w4a16:
        group_size = get_quantization_group_size(config)
        if group_size is None:
            raise ValueError(
                "Could not determine group_size from model config. "
                "The model's quantization_config must contain a 'group_size' "
                "field (AWQ/GPTQ) or 'config_groups.*.weights.group_size' "
                "(compressed-tensors)."
            )
        # For int4_w4a16, block_shape = [0, group_size]
        # block_shape[0]=0 means no block quantization on N dimension
        block_quant_shape = [0, group_size]
857
858

    if args.batch_size is None:
859
        batch_sizes = [
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
            1,
            2,
            4,
            8,
            16,
            24,
            32,
            48,
            64,
            96,
            128,
            256,
            512,
            1024,
            1536,
            2048,
            3072,
            4096,
878
        ]
879
    else:
880
        batch_sizes = args.batch_size
881

882
883
    use_deep_gemm = bool(args.use_deep_gemm)

884
885
886
887
    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."
888
889
            "Replacing HIP_VISIBLE_DEVICES with ROCR_VISIBLE_DEVICES."
        )
890
891
892
893
        val = os.environ["HIP_VISIBLE_DEVICES"]
        os.environ["ROCR_VISIBLE_DEVICES"] = val
        del os.environ["HIP_VISIBLE_DEVICES"]

894
895
896
897
    ray.init()
    num_gpus = int(ray.available_resources()["GPU"])
    workers = [BenchmarkWorker.remote(args.seed) for _ in range(num_gpus)]

898
    def _distribute(method: str, inputs: list[Any]) -> list[Any]:
899
900
901
902
903
904
905
906
907
908
909
        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:
910
911
912
913
914
915
916
917
918
919
920
921
922
923
        # int4_w4a16 weights are uint8-packed, not fp16; treat like fp8 for
        # search space generation (no matrix_instr_nonkdim/kpack exploration).
        is_fp16 = not (use_fp8_w8a8 or use_int8_w8a16 or use_int4_w4a16)
        # For int4_w4a16, the group_size constraint on BLOCK_SIZE_K does not
        # apply: the gptq_awq kernel handles arbitrary BLOCK_SIZE_K regardless
        # of group_size. Skip block_quant_shape filtering to keep the full
        # search space (e.g. BLOCK_SIZE_K=64 with group_size=128).
        tune_block_quant_shape = None if use_int4_w4a16 else block_quant_shape
        search_space = get_configs_compute_bound(is_fp16, tune_block_quant_shape)
        if use_int4_w4a16:
            # SPLIT_K is a required kernel constexpr for gptq_awq kernel;
            # only SPLIT_K=1 is used at runtime, so fix it during tuning.
            for cfg in search_space:
                cfg["SPLIT_K"] = 1
924
        print(f"Start tuning over {len(search_space)} configurations...")
925
926
927
928
929
        if use_deep_gemm:
            raise ValueError(
                "Tuning with --use-deep-gemm is not supported as it only tunes Triton "
                "kernels. Please remove the flag."
            )
930
931
        start = time.time()
        configs = _distribute(
932
933
934
935
936
937
938
939
940
941
942
            "tune",
            [
                (
                    batch_size,
                    E,
                    shard_intermediate_size,
                    hidden_size,
                    topk,
                    dtype,
                    use_fp8_w8a8,
                    use_int8_w8a16,
943
                    use_int4_w4a16,
944
945
946
947
948
949
950
                    search_space,
                    block_quant_shape,
                    use_deep_gemm,
                )
                for batch_size in batch_sizes
            ],
        )
951
        best_configs = {
952
            M: sort_config(config) for M, config in zip(batch_sizes, configs)
953
        }
954
955
956
957
958
959
960
961
962
        save_configs(
            best_configs,
            E,
            shard_intermediate_size,
            hidden_size,
            topk,
            dtype,
            use_fp8_w8a8,
            use_int8_w8a16,
963
            use_int4_w4a16,
964
            block_quant_shape,
965
            args.save_dir,
966
        )
967
968
969
        end = time.time()
        print(f"Tuning took {end - start:.2f} seconds")
    else:
970
        outputs = _distribute(
971
            "benchmark",
972
973
974
975
976
977
978
979
980
981
            [
                (
                    batch_size,
                    E,
                    shard_intermediate_size,
                    hidden_size,
                    topk,
                    dtype,
                    use_fp8_w8a8,
                    use_int8_w8a16,
982
                    use_int4_w4a16,
983
984
985
986
987
988
                    block_quant_shape,
                    use_deep_gemm,
                )
                for batch_size in batch_sizes
            ],
        )
989
990
991
992
993
994
995

        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__":
996
    parser = FlexibleArgumentParser()
997
998
999
1000
1001
1002
    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
    )
1003
    parser.add_argument("--enable-expert-parallel", "-enable-ep", action="store_true")
1004
    parser.add_argument(
1005
1006
1007
1008
        "--dtype",
        type=str,
        choices=["auto", "fp8_w8a8", "int8_w8a16", "int4_w4a16"],
        default="auto",
1009
    )
1010
    parser.add_argument("--use-deep-gemm", action="store_true")
1011
1012
1013
    parser.add_argument(
        "--save-dir", type=str, default="./", help="Directory to save tuned results"
    )
1014
    parser.add_argument("--seed", type=int, default=0)
1015
    parser.add_argument("--batch-size", type=int, nargs="+", required=False)
1016
    parser.add_argument("--tune", action="store_true")
1017
    parser.add_argument("--trust-remote-code", action="store_true")
1018
    parser.add_argument("--model-prefix", type=str, required=False)
1019
1020
1021
    args = parser.parse_args()

    main(args)