common.py 21.1 KB
Newer Older
1
2
3
4
5
6
7
8
9
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from dataclasses import dataclass
from typing import Any, Optional, Union

import torch

import vllm._custom_ops as ops
import vllm.model_executor.layers.fused_moe.modular_kernel as mk
10
from tests.kernels.moe.utils import make_test_weights, per_token_cast_to_fp8
11
12
13
14
15
from tests.kernels.quantization.nvfp4_utils import (
    FLOAT4_E2M1_MAX,
    FLOAT8_E4M3_MAX,
    dequantize_nvfp4_to_dtype,
)
16
17
18
from tests.kernels.utils import torch_experts
from vllm.config import VllmConfig
from vllm.distributed import get_dp_group, get_tensor_model_parallel_world_size
19
from vllm.forward_context import set_forward_context
20
from vllm.model_executor.layers.fused_moe.config import (
21
22
23
24
    FusedMoEConfig,
    FusedMoEParallelConfig,
    FusedMoEQuantConfig,
)
25
26
27
from vllm.model_executor.layers.fused_moe.fused_moe import fused_topk
from vllm.utils import has_deep_ep, has_deep_gemm, has_pplx

28
29
30
31
32
33
34
from .mk_objects import (
    TestMoEQuantConfig,
    expert_info,
    make_fused_experts,
    make_prepare_finalize,
    prepare_finalize_info,
)
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
from .parallel_utils import ProcessGroupInfo


def _describe_tensor(t: Optional[torch.Tensor], name: str) -> str:
    if t is None:
        return f"{name} : None"
    else:
        return f"{name} : {t.shape} {t.dtype} {t.device}"


@dataclass
class Config:
    Ms: Union[list[int], int]
    K: int
    N: int
    E: int
    topks: Union[list[int], int]
    dtype: torch.dtype
53
    quant_config: Optional[TestMoEQuantConfig]
54
55
56
57
58
59
60
61
62

    prepare_finalize_type: mk.FusedMoEPrepareAndFinalize
    fused_experts_type: mk.FusedMoEPermuteExpertsUnpermute

    fused_moe_chunk_size: Optional[int]
    world_size: int

    torch_trace_dir_path: Optional[str] = None

63
64
    def __post_init__(self):
        if self.quant_config is None:
65
            self.quant_config = TestMoEQuantConfig(None, False, False, None)
66

67
68
    def describe(self) -> str:
        s = ""
69
70
71
72
73
74
75
76
77
78
79
80
        s += "== Config:\n"
        s += f" world_size={self.world_size}\n"
        s += f" PF={self.prepare_finalize_type.__name__}\n"
        s += f" FE={self.fused_experts_type.__name__}\n"
        s += f" E={self.E}\n"
        s += f" Ms={self.Ms}\n"
        s += f" N={self.N}\n"
        s += f" K={self.K}\n"
        s += f" topk={self.topks}\n"
        s += f" dtype={self.dtype}\n"
        s += f" fused_moe_chunk_size={self.fused_moe_chunk_size}\n"
        s += " Quant:\n"
81
        if self.quant_config is not None:
82
83
84
85
            s += f"     q_dtype={self.quant_dtype}\n"
            s += f"     q_block_shape={self.quant_block_shape}\n"
            s += f"     q_per_out_ch_quant={self.is_per_out_ch_quant}\n"
            s += f"     q_per_act_token={self.is_per_act_token_quant}\n"
86
        else:
87
            s += "     quant=None\n"
88
89
90
91
92
93
94
95
        return s

    @property
    def M(self) -> int:
        assert isinstance(self.Ms, int)
        return self.Ms

    @property
96
97
    def quant_dtype(self) -> Union[torch.dtype, str, None]:
        assert self.quant_config is not None
98
99
100
101
        return self.quant_config.quant_dtype

    @property
    def is_per_act_token_quant(self) -> bool:
102
        assert self.quant_config is not None
103
104
105
106
        return self.quant_config.per_act_token_quant

    @property
    def is_per_tensor_act_quant(self) -> bool:
107
        return not self.is_per_act_token_quant and self.quant_block_shape is None
108
109
110

    @property
    def is_per_out_ch_quant(self) -> bool:
111
        assert self.quant_config is not None
112
113
114
115
        return self.quant_config.per_out_ch_quant

    @property
    def quant_block_shape(self) -> Optional[list[int]]:
116
        assert self.quant_config is not None
117
118
119
120
121
122
123
124
125
126
127
128
129
        return self.quant_config.block_shape

    @property
    def topk(self) -> int:
        assert isinstance(self.topks, int)
        return self.topks

    @property
    def num_local_experts(self) -> int:
        return self.E // self.world_size

    def make_env_data(self) -> tuple[VllmConfig, dict[Any, Any]]:
        """
130
        make env data for vllm launch.
131
132
133
134
135
136
137
138
        """
        vllm_config = VllmConfig()
        vllm_config.parallel_config.data_parallel_size = self.world_size
        vllm_config.parallel_config.enable_expert_parallel = True

        env_dict = {
            "VLLM_USE_DEEP_GEMM": str(int(self.needs_deep_gemm())),
        }
139
140
141
142
143

        backend = self.all2all_backend()
        if backend is not None:
            env_dict.update({"VLLM_ALL2ALL_BACKEND": backend})

144
145
        if self.fused_moe_chunk_size is not None:
            env_dict.update(
146
147
                {"VLLM_FUSED_MOE_CHUNK_SIZE": str(self.fused_moe_chunk_size)}
            )
148

149
150
151
        return vllm_config, env_dict

    def is_fp8_block_quantized(self):
152
153
154
155
        return (
            self.quant_dtype == torch.float8_e4m3fn
            and self.quant_block_shape is not None
        )
156
157

    def is_batched_prepare_finalize(self):
158
        info = prepare_finalize_info(self.prepare_finalize_type)
159
        return mk.FusedMoEActivationFormat.BatchedExperts == info.activation_format
160
161

    def is_batched_fused_experts(self):
162
        info = expert_info(self.fused_experts_type)
163
        return mk.FusedMoEActivationFormat.BatchedExperts == info.activation_format
164
165

    def is_standard_fused_experts(self):
166
167
168
169
170
171
172
173
174
175
176
177
178
179
        info = expert_info(self.fused_experts_type)
        return mk.FusedMoEActivationFormat.Standard == info.activation_format

    def fe_supported_types(self):
        info = expert_info(self.fused_experts_type)
        return info.supported_dtypes

    def pf_supported_types(self):
        info = prepare_finalize_info(self.prepare_finalize_type)
        return info.supported_dtypes

    def is_block_quant_supported(self):
        info = expert_info(self.fused_experts_type)
        return info.blocked_quantization_support
180
181

    def is_fe_supports_chunking(self):
182
183
184
185
186
187
188
189
190
191
        info = expert_info(self.fused_experts_type)
        return info.supports_chunking

    def supports_expert_map(self):
        info = expert_info(self.fused_experts_type)
        return info.supports_expert_map

    def supports_apply_weight_on_input(self):
        info = prepare_finalize_info(self.prepare_finalize_type)
        return info.supports_apply_weight_on_input
192
193

    def needs_deep_gemm(self):
194
195
        info = expert_info(self.fused_experts_type)
        return info.needs_deep_gemm
196
197

    def needs_pplx(self):
198
199
        info = prepare_finalize_info(self.prepare_finalize_type)
        return info.backend == "pplx"
200
201

    def needs_deep_ep(self):
202
        info = prepare_finalize_info(self.prepare_finalize_type)
203
204
205
206
        return (
            info.backend == "deepep_high_throughput"
            or info.backend == "deepep_low_latency"
        )
207
208

    def all2all_backend(self):
209
210
        info = prepare_finalize_info(self.prepare_finalize_type)
        return info.backend
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225

    def is_valid(self):
        # Check prepare-finalize and fused-experts compatibility
        if self.is_batched_prepare_finalize():
            if not self.is_batched_fused_experts():
                return False
        else:
            if not self.is_standard_fused_experts():
                return False

        use_chunking = self.fused_moe_chunk_size is not None
        if use_chunking and not self.is_fe_supports_chunking():
            return False

        # Check quantization sanity
226
227
228
229
230
        if (
            int(self.is_per_act_token_quant)
            + int(self.is_per_tensor_act_quant)
            + int(self.quant_block_shape is not None)
        ) > 1:
231
232
233
            # invalid quant config
            return False

234
235
        # check type support
        if self.quant_dtype is None:
236
237
238
239
            if (
                self.dtype not in self.pf_supported_types()
                or self.dtype not in self.fe_supported_types()
            ):
240
241
                return False
        else:
242
243
244
245
            if (
                self.quant_dtype not in self.pf_supported_types()
                or self.quant_dtype not in self.fe_supported_types()
            ):
246
                return False
247

248
        # Check block quanization support
249
        is_block_quatized = self.quant_block_shape is not None
250
        if is_block_quatized and self.quant_dtype is None:
251
            return False
252
        if is_block_quatized and not self.is_block_quant_supported():
253
254
255
256
257
258
            return False

        # deep_gemm only works with block-quantized
        if self.needs_deep_gemm() and not is_block_quatized:
            return False

259
        # Check dependencies (turn into asserts?)
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
        if self.needs_deep_ep() and not has_deep_ep():
            return False
        if self.needs_deep_gemm() and not has_deep_gemm():
            return False
        if self.needs_pplx() and not has_pplx():  # noqa: SIM103
            return False

        return True


@dataclass
class WeightTensors:
    w1: torch.Tensor
    w2: torch.Tensor
    w1_scale: Optional[torch.Tensor]
    w2_scale: Optional[torch.Tensor]
276
277
    w1_gs: Optional[torch.Tensor] = None
    w2_gs: Optional[torch.Tensor] = None
278
279
280
281

    def describe(self):
        s = ""
        s += "== Weight Tensors: \n"
282
283
284
285
286
287
        s += f" - {_describe_tensor(self.w1, 'w1')} \n"
        s += f" - {_describe_tensor(self.w2, 'w2')} \n"
        s += f" - {_describe_tensor(self.w1_scale, 'w1_scale')} \n"
        s += f" - {_describe_tensor(self.w2_scale, 'w2_scale')} \n"
        s += f" - {_describe_tensor(self.w1_gs, 'w1_gs')} \n"
        s += f" - {_describe_tensor(self.w2_gs, 'w2_gs')} \n"
288
289
        return s

290
291
    def is_quantized(self) -> bool:
        # or w1_scale is not None?
292
293
294
295
296
        return (
            self.w1.dtype == torch.float8_e4m3fn
            or self.w1.dtype == torch.uint8
            or self.w1.dtype == torch.int8
        )
297

298
    def to_current_device(self):
299
300
301
        device = torch.cuda.current_device()
        self.w1 = self.w1.to(device=device)
        self.w2 = self.w2.to(device=device)
302

303
304
305
306
        if self.w1_scale is not None:
            self.w1_scale = self.w1_scale.to(device=device)
        if self.w2_scale is not None:
            self.w2_scale = self.w2_scale.to(device=device)
307

308
        if self.w1_gs is not None:
309
310
311
            self.w1_gs = self.w1_gs.to(device=device)
        if self.w2_gs is not None:
            self.w2_gs = self.w2_gs.to(device=device)
312

313
    def slice_weights(self, rank: int, num_local_experts: int) -> "WeightTensors":
314
315
316
317
        s = rank * num_local_experts
        e = s + num_local_experts
        w1 = self.w1[s:e, :, :]
        w2 = self.w2[s:e, :, :]
318
319
        w1_scale = self.w1_scale[s:e, :, :] if self.w1_scale is not None else None
        w2_scale = self.w2_scale[s:e, :, :] if self.w2_scale is not None else None
320
321
        w1_gs = self.w1_gs[s:e] if self.w1_gs is not None else None
        w2_gs = self.w2_gs[s:e] if self.w2_gs is not None else None
322

323
        return WeightTensors(w1, w2, w1_scale, w2_scale, w1_gs, w2_gs)
324

325
326
327
    @staticmethod
    def make(config: Config) -> "WeightTensors":
        (_, w1, w1_scale, w1_gs), (_, w2, w2_scale, w2_gs) = make_test_weights(
328
329
330
            e=config.E,
            n=config.N,
            k=config.K,
331
332
333
            in_dtype=config.dtype,
            quant_dtype=config.quant_dtype,
            block_shape=config.quant_block_shape,
334
335
            # or config.is_per_out_ch_quant
            per_out_ch_quant=config.is_per_act_token_quant,
336
337
338
        )
        return WeightTensors(
            w1=w1, w2=w2, w1_scale=w1_scale, w2_scale=w2_scale, w1_gs=w1_gs, w2_gs=w2_gs
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
        )


@dataclass
class RankTensors:
    hidden_states: torch.Tensor
    hidden_states_scale: Optional[torch.Tensor]

    topk_weights: torch.Tensor
    topk_ids: torch.Tensor
    expert_map: Optional[torch.Tensor]

    def describe(self):
        s = ""
        s += "== Rank Tensors: \n"
354
355
356
357
358
        s += f" - {_describe_tensor(self.hidden_states, 'HS')} \n"
        s += f" - {_describe_tensor(self.hidden_states_scale, 'HS_scale')} \n"
        s += f" - {_describe_tensor(self.topk_weights, 'topk_weights')} \n"
        s += f" - {_describe_tensor(self.topk_ids, 'topk_ids')} \n"
        s += f" - {_describe_tensor(self.expert_map, 'expert_map')} \n"
359
360
361
362
        return s

    @staticmethod
    def make_hidden_states(
363
364
        config: Config,
    ) -> tuple[torch.Tensor, Optional[torch.Tensor]]:
365
366
367
368
        """
        Return hidden_states
        """
        m, k, dtype = (config.M, config.K, config.dtype)
369
        a = torch.randn((m, k), device=torch.cuda.current_device(), dtype=dtype) / 15.0
370
371
372
373
374
375
376

        if config.quant_dtype is None:
            return a, None

        # We dequant and use that as hidden_states so the tests are stable.
        # quantizing and dequantizing yield slightly different results
        # depending on the hardware. Here we, quantize and dequantize
377
        # first - so further quantize and dequantize will yield the same
378
379
        # values.
        if config.is_per_tensor_act_quant:
380
            a_q, a_scales = ops.scaled_fp8_quant(a, use_per_token_if_dynamic=False)
381
382
383
            return a_q.float().mul(a_scales).to(dtype), a_scales

        if config.is_per_act_token_quant:
384
            a_q, a_scales = ops.scaled_fp8_quant(a, use_per_token_if_dynamic=True)
385
386
387
388
389
            return a_q.float().mul(a_scales).to(dtype), None

        assert config.quant_block_shape is not None
        block_k = config.quant_block_shape[1]
        a_q, a_scales = per_token_cast_to_fp8(a, block_size=block_k)
390
391
392
        return a_q.float().view((-1, block_k)).mul(a_scales.view(-1, 1)).view(m, k).to(
            dtype
        ), None
393
394
395
396
397

    @staticmethod
    def make(config: Config, pgi: ProcessGroupInfo):
        dtype = config.dtype
        topk, m, _ = (config.topk, config.M, config.K)
398
        hidden_states, hidden_states_scale = RankTensors.make_hidden_states(config)
399

400
401
402
        num_local_experts, global_num_experts = (config.num_local_experts, config.E)
        score = torch.randn((m, global_num_experts), device="cuda", dtype=dtype)
        topk_weights, topk_ids, _ = fused_topk(hidden_states, score, topk, False)
403
404
405
406
407
408
409

        # distribute topk_ids evenly
        for mi in range(m):
            topk_ids[mi] = torch.randperm(config.E)[:topk]
        topk_ids = topk_ids.to(device=torch.cuda.current_device())

        expert_map = None
410
        if config.world_size > 1 and config.supports_expert_map():
411
412
413
            expert_map = torch.full(
                (global_num_experts,), fill_value=-1, dtype=torch.int32
            )
414
415
416
            s = pgi.rank * num_local_experts
            e = s + num_local_experts
            expert_map[s:e] = torch.tensor(list(range(num_local_experts)))
417
418
419
            expert_map = expert_map.to(
                device=torch.cuda.current_device(), dtype=torch.int32
            )
420
421
422
423
424
425
426
427
428
429

        return RankTensors(
            hidden_states=hidden_states,
            hidden_states_scale=hidden_states_scale,
            topk_weights=topk_weights,
            topk_ids=topk_ids,
            expert_map=expert_map,
        )


430
431
432
def reference_moe_impl(
    config: Config, weights: WeightTensors, rank_tensors: RankTensors
) -> torch.Tensor:
433
434
435
436
437
438
439
440
441
442
443
444
    if config.quant_dtype == "nvfp4":
        quant_blocksize = 16
        dtype = config.dtype

        w1_q = weights.w1
        w1_blockscale = weights.w1_scale
        w1_gs = weights.w1_gs

        w2_q = weights.w2
        w2_blockscale = weights.w2_scale
        w2_gs = weights.w2_gs

445
446
447
448
        a_global_scale = (
            (FLOAT8_E4M3_MAX * FLOAT4_E2M1_MAX)
            / torch.amax(rank_tensors.hidden_states.flatten(), dim=-1)
        ).to(torch.float32)
449
450
451
452
453
454
455
456
457
458
459
460

        assert w1_gs is not None
        assert w2_gs is not None
        assert w1_blockscale is not None
        assert w2_blockscale is not None

        assert w1_blockscale.shape[1] % 128 == 0
        assert w1_blockscale.shape[2] % 4 == 0
        assert w2_blockscale.shape[1] % 128 == 0
        assert w2_blockscale.shape[2] % 4 == 0

        a_fp4, a_scale_interleaved = ops.scaled_fp4_quant(
461
462
            rank_tensors.hidden_states, a_global_scale
        )
463

464
465
466
467
468
469
470
471
        a = dequantize_nvfp4_to_dtype(
            a_fp4,
            a_scale_interleaved,
            a_global_scale,
            dtype=dtype,
            device=a_fp4.device,
            block_size=quant_blocksize,
        )
472
473
474
475
476
477
478
479
480

        e = w1_q.shape[0]
        n = w1_q.shape[1] // 2
        k = w2_q.shape[1]

        w1 = torch.zeros((e, 2 * n, k), device="cuda", dtype=dtype)
        w2 = torch.zeros((e, k, n), device="cuda", dtype=dtype)

        for idx in range(0, e):
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
            w1[idx] = dequantize_nvfp4_to_dtype(
                w1_q[idx],
                w1_blockscale[idx],
                w1_gs[idx],
                dtype=dtype,
                device=w1_q.device,
                block_size=quant_blocksize,
            )
            w2[idx] = dequantize_nvfp4_to_dtype(
                w2_q[idx],
                w2_blockscale[idx],
                w2_gs[idx],
                dtype=dtype,
                device=w2_q.device,
                block_size=quant_blocksize,
            )
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
        a_scale = None
        w1_scale = None
        w2_scale = None
        quant_dtype = None
        per_act_token_quant = False
        block_shape = None
    else:
        a = rank_tensors.hidden_states
        a_scale = rank_tensors.hidden_states_scale
        w1 = weights.w1
        w1_scale = weights.w1_scale
        w2 = weights.w2
        w2_scale = weights.w2_scale
        quant_dtype = config.quant_dtype
        per_act_token_quant = config.is_per_act_token_quant
        block_shape = config.quant_block_shape

514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
    return torch_experts(
        a=a,
        w1=w1,
        w2=w2,
        topk_weight=rank_tensors.topk_weights,
        topk_ids=rank_tensors.topk_ids,
        global_num_experts=config.E,
        expert_map=None,
        w1_scale=w1_scale,
        w2_scale=w2_scale,
        a1_scale=a_scale,
        quant_dtype=quant_dtype,
        per_act_token_quant=per_act_token_quant,
        block_shape=block_shape,
        apply_router_weights_on_input=config.topk == 1
        and config.supports_apply_weight_on_input(),
    )
531
532


533
def _make_gscale(num_experts: int) -> torch.Tensor:
534
535
536
    return torch.ones(
        (num_experts,), device=torch.cuda.current_device(), dtype=torch.float32
    )
537
538


539
540
541
def make_modular_kernel(
    config: Config,
    vllm_config: VllmConfig,
542
    quant_config: FusedMoEQuantConfig,
543
) -> mk.FusedMoEModularKernel:
544
545
    def next_power_of_2(x):
        import math
546

547
548
        if x == 0:
            return 1
549
        return 2 ** math.ceil(math.log2(x))
550
551
552
553
554
555
556

    # make moe config
    moe_parallel_config: FusedMoEParallelConfig = FusedMoEParallelConfig.make(
        tp_size_=get_tensor_model_parallel_world_size(),
        dp_size_=get_dp_group().world_size,
        vllm_parallel_config=vllm_config.parallel_config,
    )
557

558
559
560
561
562
563
564
565
566
567
568
    moe = FusedMoEConfig(
        num_experts=config.E,
        experts_per_token=config.topk,
        hidden_dim=config.K,
        num_local_experts=config.num_local_experts,
        moe_parallel_config=moe_parallel_config,
        in_dtype=config.dtype,
        max_num_tokens=next_power_of_2(config.M),
    )

    # make modular kernel
569
570
571
    prepare_finalize = make_prepare_finalize(
        config.prepare_finalize_type, config.all2all_backend(), moe, quant_config
    )
572
573
574
575

    fused_experts = make_fused_experts(
        config.fused_experts_type,
        moe,
576
        quant_config,
577
        prepare_finalize.num_dispatchers(),
578
        config.N,
579
    )
580
581

    modular_kernel = mk.FusedMoEModularKernel(
582
583
        prepare_finalize=prepare_finalize, fused_experts=fused_experts
    )
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600

    return modular_kernel


def run_modular_kernel(
    pgi: ProcessGroupInfo,
    vllm_config: VllmConfig,
    config: Config,
    weights: WeightTensors,
    rank_tensors: RankTensors,
) -> torch.Tensor:
    assert isinstance(config.Ms, int)
    assert isinstance(config.topks, int)

    # weights for rank
    rank_weights = weights.slice_weights(pgi.rank, config.num_local_experts)

601
602
603
604
605
606
607
608
609
610
    if config.quant_dtype == "nvfp4":
        gscale = _make_gscale(config.num_local_experts)
    else:
        gscale = None

    quant_config = FusedMoEQuantConfig.make(
        config.quant_dtype,
        w1_scale=rank_weights.w1_scale,
        w2_scale=rank_weights.w2_scale,
        a1_scale=rank_tensors.hidden_states_scale,
611
612
        g1_alphas=(1 / rank_weights.w1_gs) if rank_weights.w1_gs is not None else None,
        g2_alphas=(1 / rank_weights.w2_gs) if rank_weights.w2_gs is not None else None,
613
614
615
616
617
618
619
620
621
622
623
624
        a1_gscale=gscale,
        a2_gscale=gscale,
        block_shape=config.quant_block_shape,
        per_act_token_quant=config.is_per_act_token_quant,
        per_out_ch_quant=config.is_per_out_ch_quant,
    )

    mk = make_modular_kernel(config, vllm_config, quant_config)

    # impls might update the tensor in place
    hidden_states = rank_tensors.hidden_states.clone()

625
    topk_ids = rank_tensors.topk_ids.to(mk.prepare_finalize.topk_indices_dtype())
626
627

    mk_kwargs = {
628
629
630
631
632
633
634
635
636
        "hidden_states": hidden_states,
        "w1": rank_weights.w1,
        "w2": rank_weights.w2,
        "topk_weights": rank_tensors.topk_weights,
        "topk_ids": topk_ids,
        "expert_map": rank_tensors.expert_map,
        "global_num_experts": config.E,
        "apply_router_weight_on_input": config.topk == 1
        and config.supports_apply_weight_on_input(),
637
    }
638
639

    num_tokens = rank_tensors.hidden_states.shape[0]
640
641
642
    num_tokens_across_dp = torch.tensor(
        [num_tokens] * config.world_size, device="cuda", dtype=torch.int
    )
643
644

    with set_forward_context(
645
646
647
648
        None,
        vllm_config,
        num_tokens=num_tokens,
        num_tokens_across_dp=num_tokens_across_dp,
649
650
    ):
        out = mk.forward(**mk_kwargs)
651
652

    return out