_custom_ops.py 99.8 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
from typing import TYPE_CHECKING, Literal
5
6

import torch
gaoqiong's avatar
gaoqiong committed
7

8
import vllm.envs as envs
9
from vllm.logger import init_logger
10
from vllm.platforms import current_platform
11
from vllm.scalar_type import ScalarType
12

13
from vllm.utils.torch_utils import direct_register_custom_op
14

15
try:
gaoqiong's avatar
gaoqiong committed
16
    from lmslim import quant_ops 
gaoqiong's avatar
gaoqiong committed
17
    from lmslim import quant_tools 
18
except Exception:
gaoqiong's avatar
gaoqiong committed
19
    print("INFO: Please install lmslim if you want to infer gptq or awq  or w8a8 model.\n") 
yangql's avatar
yangql committed
20
try:
21
    import lightop
yangql's avatar
yangql committed
22
except Exception:
23
    print("INFO: Please install lightop if you want to infer awq of marlin.\n") 
24

25
26
logger = init_logger(__name__)

27
current_platform.import_kernels()
28

29
if TYPE_CHECKING:
30
31
32
33
34
35
36
37
38

    def register_fake(fn):
        return lambda name: fn
else:
    try:
        from torch.library import register_fake
    except ImportError:
        from torch.library import impl_abstract as register_fake

39
40
41

# page attention ops
def paged_attention_v1(
42
43
44
45
46
47
48
49
50
51
    out: torch.Tensor,
    query: torch.Tensor,
    key_cache: torch.Tensor,
    value_cache: torch.Tensor,
    num_kv_heads: int,
    scale: float,
    block_tables: torch.Tensor,
    seq_lens: torch.Tensor,
    block_size: int,
    max_seq_len: int,
52
    alibi_slopes: torch.Tensor | None,
53
    kv_cache_dtype: str,
54
55
    k_scale: torch.Tensor,
    v_scale: torch.Tensor,
56
57
58
59
    tp_rank: int = 0,
    blocksparse_local_blocks: int = 0,
    blocksparse_vert_stride: int = 0,
    blocksparse_block_size: int = 64,
60
    blocksparse_head_sliding_step: int = 0,
61
62
) -> None:
    torch.ops._C.paged_attention_v1(
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
        out,
        query,
        key_cache,
        value_cache,
        num_kv_heads,
        scale,
        block_tables,
        seq_lens,
        block_size,
        max_seq_len,
        alibi_slopes,
        kv_cache_dtype,
        k_scale,
        v_scale,
        tp_rank,
        blocksparse_local_blocks,
        blocksparse_vert_stride,
        blocksparse_block_size,
        blocksparse_head_sliding_step,
    )
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98


def paged_attention_v2(
    out: torch.Tensor,
    exp_sum: torch.Tensor,
    max_logits: torch.Tensor,
    tmp_out: torch.Tensor,
    query: torch.Tensor,
    key_cache: torch.Tensor,
    value_cache: torch.Tensor,
    num_kv_heads: int,
    scale: float,
    block_tables: torch.Tensor,
    seq_lens: torch.Tensor,
    block_size: int,
    max_seq_len: int,
99
    alibi_slopes: torch.Tensor | None,
100
    kv_cache_dtype: str,
101
102
    k_scale: torch.Tensor,
    v_scale: torch.Tensor,
103
104
105
106
    tp_rank: int = 0,
    blocksparse_local_blocks: int = 0,
    blocksparse_vert_stride: int = 0,
    blocksparse_block_size: int = 64,
107
    blocksparse_head_sliding_step: int = 0,
108
109
) -> None:
    torch.ops._C.paged_attention_v2(
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
        out,
        exp_sum,
        max_logits,
        tmp_out,
        query,
        key_cache,
        value_cache,
        num_kv_heads,
        scale,
        block_tables,
        seq_lens,
        block_size,
        max_seq_len,
        alibi_slopes,
        kv_cache_dtype,
        k_scale,
        v_scale,
        tp_rank,
        blocksparse_local_blocks,
        blocksparse_vert_stride,
        blocksparse_block_size,
        blocksparse_head_sliding_step,
    )
133
134


135
136
137
138
139
140
141
142
143
144
145
146
# def paged_attention_rocm(
#     out: torch.Tensor,
#     exp_sum: torch.Tensor,
#     max_logits: torch.Tensor,
#     tmp_out: torch.Tensor,
#     query: torch.Tensor,
#     key_cache: torch.Tensor,
#     value_cache: torch.Tensor,
#     num_kv_heads: int,
#     scale: float,
#     block_tables: torch.Tensor,
#     seq_lens: torch.Tensor,
147
#     query_start_loc: torch.Tensor | None,
148
149
#     block_size: int,
#     max_seq_len: int,
150
#     alibi_slopes: torch.Tensor | None,
151
152
153
#     kv_cache_dtype: str,
#     k_scale: torch.Tensor,
#     v_scale: torch.Tensor,
154
#     fp8_out_scale: torch.Tensor | None = None,
155
#     mfma_type: str = "fp8" if envs.VLLM_ROCM_FP8_MFMA_PAGE_ATTN else "f16",
156
# ) -> None:
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
#     torch.ops._rocm_C.paged_attention(
#         out,
#         exp_sum,
#         max_logits,
#         tmp_out,
#         query,
#         key_cache,
#         value_cache,
#         num_kv_heads,
#         scale,
#         block_tables,
#         seq_lens,
#         query_start_loc,
#         block_size,
#         max_seq_len,
#         alibi_slopes,
#         kv_cache_dtype,
#         k_scale,
#         v_scale,
#         fp8_out_scale,
#         mfma_type,
#     )
179

Thien Tran's avatar
Thien Tran committed
180
181
182
183
184
185
186
187
def mla_decode_kvcache_cpu(
    out: torch.Tensor,
    query: torch.Tensor,
    kv_cache: torch.Tensor,
    scale: float,
    block_tables: torch.Tensor,
    seq_lens: torch.Tensor,
) -> None:
188
189
190
    torch.ops._C_cpu.mla_decode_kvcache(
        out, query, kv_cache, scale, block_tables, seq_lens
    )
191
192


193
# merge attn states ops
194
195
196
197
198
199
def merge_attn_states(
    output: torch.Tensor,
    prefix_output: torch.Tensor,
    prefix_lse: torch.Tensor,
    suffix_output: torch.Tensor,
    suffix_lse: torch.Tensor,
200
    output_lse: torch.Tensor | None = None,
201
202
203
204
) -> None:
    torch.ops._C.merge_attn_states(
        output, output_lse, prefix_output, prefix_lse, suffix_output, suffix_lse
    )
205
206


207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
def convert_vertical_slash_indexes(
    q_seqlens: torch.Tensor,  # [BATCH, ]
    kv_seqlens: torch.Tensor,  # [BATCH, ]
    vertical_indexes: torch.Tensor,  # [BATCH, N_HEADS, NNZ_V]
    slash_indexes: torch.Tensor,  # [BATCH, N_HEADS, NNZ_S]
    context_size: int,
    block_size_M: int,
    block_size_N: int,
    causal: bool = True,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
    batch_size = slash_indexes.size(0)
    num_heads = slash_indexes.size(1)
    nnz_slash = slash_indexes.size(2)
    nnz_vertical = vertical_indexes.size(2)
    num_rows = (context_size + block_size_M - 1) // block_size_M

223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
    block_count = torch.zeros(
        batch_size, num_heads, num_rows, dtype=q_seqlens.dtype, device=q_seqlens.device
    )
    block_offset = torch.zeros(
        batch_size,
        num_heads,
        num_rows,
        nnz_slash,
        dtype=q_seqlens.dtype,
        device=q_seqlens.device,
    )
    column_count = torch.zeros(
        batch_size, num_heads, num_rows, dtype=q_seqlens.dtype, device=q_seqlens.device
    )
    column_index = torch.zeros(
        batch_size,
        num_heads,
        num_rows,
        nnz_vertical,
        dtype=q_seqlens.dtype,
        device=q_seqlens.device,
    )
245
246

    torch.ops._C.convert_vertical_slash_indexes(
247
248
249
250
251
252
253
254
255
256
257
258
259
        block_count,
        block_offset,
        column_count,
        column_index,
        q_seqlens,
        kv_seqlens,
        vertical_indexes,
        slash_indexes,
        context_size,
        block_size_M,
        block_size_N,
        causal,
    )
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
    return block_count, block_offset, column_count, column_index


def convert_vertical_slash_indexes_mergehead(
    q_seqlens: torch.Tensor,  # [BATCH, ]
    kv_seqlens: torch.Tensor,  # [BATCH, ]
    vertical_indexes: torch.Tensor,  # [BATCH, N_HEADS, NNZ_V]
    slash_indexes: torch.Tensor,  # [BATCH, N_HEADS, NNZ_S]
    # [N_HEADS] : different head use different number of indices
    vertical_indices_count: torch.Tensor,
    slash_indices_count: torch.Tensor,
    context_size: int,
    block_size_M: int,
    block_size_N: int,
    causal: bool = True,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
    batch_size = slash_indexes.size(0)
    num_heads = slash_indexes.size(1)
    nnz_slash = slash_indexes.size(2)
    nnz_vertical = vertical_indexes.size(2)
    num_rows = (context_size + block_size_M - 1) // block_size_M

282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
    block_count = torch.empty(
        batch_size, num_heads, num_rows, dtype=q_seqlens.dtype, device=q_seqlens.device
    )
    block_offset = torch.empty(
        batch_size,
        num_heads,
        num_rows,
        nnz_slash,
        dtype=q_seqlens.dtype,
        device=q_seqlens.device,
    )
    column_count = torch.empty(
        batch_size, num_heads, num_rows, dtype=q_seqlens.dtype, device=q_seqlens.device
    )
    column_index = torch.empty(
        batch_size,
        num_heads,
        num_rows,
        nnz_vertical,
        dtype=q_seqlens.dtype,
        device=q_seqlens.device,
    )
304
305

    torch.ops._C.convert_vertical_slash_indexes_mergehead(
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
        block_count,
        block_offset,
        column_count,
        column_index,
        q_seqlens,
        kv_seqlens,
        vertical_indexes,
        slash_indexes,
        vertical_indices_count,
        slash_indices_count,
        context_size,
        block_size_M,
        block_size_N,
        causal,
    )
321
322
323
    return block_count, block_offset, column_count, column_index


324
325
326
327
# pos encoding ops
def rotary_embedding(
    positions: torch.Tensor,
    query: torch.Tensor,
328
    key: torch.Tensor | None,
329
330
331
332
    head_size: int,
    cos_sin_cache: torch.Tensor,
    is_neox: bool,
) -> None:
333
334
335
    torch.ops._C.rotary_embedding(
        positions, query, key, head_size, cos_sin_cache, is_neox
    )
336
337
338


# layer norm ops
339
340
341
def rms_norm(
    out: torch.Tensor, input: torch.Tensor, weight: torch.Tensor, epsilon: float
) -> None:
342
    torch.ops._C.rms_norm(out, input, weight, epsilon)
343
344


345
346
347
def fused_add_rms_norm(
    input: torch.Tensor, residual: torch.Tensor, weight: torch.Tensor, epsilon: float
) -> None:
348
    torch.ops._C.fused_add_rms_norm(input, residual, weight, epsilon)
zhuwenwen's avatar
zhuwenwen committed
349
350
351
352
353
354
355
356
357
358
359
    

# layer norm ops (opt)
def rms_norm_opt(out: torch.Tensor, input: torch.Tensor, weight: torch.Tensor,
             epsilon: float) -> None:
    torch.ops._C.rms_norm_opt(out, input, weight, epsilon)


def fused_add_rms_norm_opt(input: torch.Tensor, residual: torch.Tensor,
                       weight: torch.Tensor, epsilon: float) -> None:
    torch.ops._C.fused_add_rms_norm_opt(input, residual, weight, epsilon)
360
361


362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
def fused_qk_norm_rope(
    qkv: torch.Tensor,
    num_heads_q: int,
    num_heads_k: int,
    num_heads_v: int,
    head_dim: int,
    eps: float,
    q_weight: torch.Tensor,
    k_weight: torch.Tensor,
    cos_sin_cache: torch.Tensor,
    is_neox: bool,
    position_ids: torch.Tensor,
) -> None:
    torch.ops._C.fused_qk_norm_rope(
        qkv,
        num_heads_q,
        num_heads_k,
        num_heads_v,
        head_dim,
        eps,
        q_weight,
        k_weight,
        cos_sin_cache,
        is_neox,
        position_ids,
    )
388
389


390
def apply_repetition_penalties_torch(
391
392
393
394
395
    logits: torch.Tensor,
    prompt_mask: torch.Tensor,
    output_mask: torch.Tensor,
    repetition_penalties: torch.Tensor,
) -> None:
396
    repetition_penalties = repetition_penalties.unsqueeze(dim=1).repeat(
397
398
        1, logits.size(1)
    )
399
    # If token appears in prompt or output, apply, otherwise use 1.0 for no-op.
400
    penalties = torch.where(prompt_mask | output_mask, repetition_penalties, 1.0)
401
402
403
404
405
406
    # If logits are positive, divide by penalty, otherwise multiply by penalty.
    scaling = torch.where(logits > 0, 1.0 / penalties, penalties)
    logits *= scaling


def apply_repetition_penalties_cuda(
407
408
409
410
411
412
413
414
    logits: torch.Tensor,
    prompt_mask: torch.Tensor,
    output_mask: torch.Tensor,
    repetition_penalties: torch.Tensor,
) -> None:
    torch.ops._C.apply_repetition_penalties_(
        logits, prompt_mask, output_mask, repetition_penalties
    )
415
416


417
418
419
420
421
422
def apply_repetition_penalties(
    logits: torch.Tensor,
    prompt_mask: torch.Tensor,
    output_mask: torch.Tensor,
    repetition_penalties: torch.Tensor,
) -> None:
423
424
425
426
427
428
429
430
    """Apply repetition penalties to logits in-place.

    Args:
        logits: The logits tensor of shape [num_seqs, vocab_size].
        prompt_mask: A boolean tensor indicating which tokens appear in the prompt.
        output_mask: A boolean tensor indicating which tokens appear in the output.
        repetition_penalties: The repetition penalties of shape (num_seqs, ).
    """
431
    if logits.is_cuda and logits.is_contiguous():
432
433
434
        apply_repetition_penalties_cuda(
            logits, prompt_mask, output_mask, repetition_penalties
        )
435
    else:
436
437
438
        apply_repetition_penalties_torch(
            logits, prompt_mask, output_mask, repetition_penalties
        )
439
440


zhuwenwen's avatar
zhuwenwen committed
441
442
443
444
445
# trans_w16
def trans_w16_gemm(dst: torch.Tensor, src: torch.Tensor,
                row:int, col:int) -> None :
    torch.ops._C.trans_w16_gemm(dst,src,row,col)
    
446

447
448
449
450
451
452
# fused quant layer norm ops
def rms_norm_dynamic_per_token_quant(
    input: torch.Tensor,
    weight: torch.Tensor,
    epsilon: float,
    quant_dtype: torch.dtype,
453
454
    scale_ub: torch.Tensor | None = None,
    residual: torch.Tensor | None = None,
455
) -> tuple[torch.Tensor, torch.Tensor]:
456
    output = torch.empty_like(input, dtype=quant_dtype)
457
458
459
    scales = torch.empty(
        (input.numel() // input.shape[-1], 1), device=input.device, dtype=torch.float32
    )
460

461
462
463
    torch.ops._C.rms_norm_dynamic_per_token_quant(
        output, input, weight, scales, epsilon, scale_ub, residual
    )
464
465
466
    return output, scales


467
468
469
470
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
502
503
504
505
506
# fused quant layer norm ops blocked
def rms_norm_per_block_quant(
    input: torch.Tensor,
    weight: torch.Tensor,
    epsilon: float,
    quant_dtype: torch.dtype,
    group_size: list[int],
    scale_ub: torch.Tensor | None = None,
    residual: torch.Tensor | None = None,
    is_scale_transposed: bool = False,
) -> tuple[torch.Tensor, torch.Tensor]:
    assert len(group_size) == 2
    output = torch.empty_like(input, dtype=quant_dtype)
    if is_scale_transposed:
        scales = torch.empty(
            (input.shape[-1] // group_size[1], input.numel() // input.shape[-1]),
            device=input.device,
            dtype=torch.float32,
        ).transpose(0, 1)
    else:
        scales = torch.empty(
            (input.numel() // input.shape[-1], input.shape[-1] // group_size[1]),
            device=input.device,
            dtype=torch.float32,
        )

    torch.ops._C.rms_norm_per_block_quant(
        output,
        input,
        weight,
        scales,
        epsilon,
        scale_ub,
        residual,
        group_size[1],
        is_scale_transposed,
    )
    return output, scales


507
508
# quantization ops
# awq
zhuwenwen's avatar
zhuwenwen committed
509
510
511
512
513
514
def GetAWQShareWorkspaceSize()->int:
    return quant_ops.GetAWQShareWorkspaceSize()

def GetAWQShareWorkspace()->torch.Tensor:
    return quant_ops.GetAWQShareWorkspace()

515
516
517
518
519
520
521
522
def awq_dequantize(
    qweight: torch.Tensor,
    scales: torch.Tensor,
    zeros: torch.Tensor,
    split_k_iters: int,
    thx: int,
    thy: int,
) -> torch.Tensor:
523
524
    if envs.VLLM_USE_TRITON_AWQ:
        from vllm.model_executor.layers.quantization.awq_triton import (
525
526
527
            awq_dequantize_triton,
        )

528
        return awq_dequantize_triton(qweight, scales, zeros)
529
    return torch.ops._C.awq_dequantize(qweight, scales, zeros, split_k_iters, thx, thy)
530

531
532
533
534
# def awq_gemm(
#     input: torch.Tensor,
#     qweight: torch.Tensor,
#     scales: torch.Tensor,
535
#     qzeros: torch.Tensor,
536
537
#     split_k_iters: int,
# ) -> torch.Tensor:
zhuwenwen's avatar
zhuwenwen committed
538
#     if envs.VLLM_USE_TRITON_AWQ:
539
#         from vllm.model_executor.layers.quantization.awq_triton import awq_gemm_triton
540

541
542
#         return awq_gemm_triton(input, qweight, scales, qzeros, split_k_iters)
#     return torch.ops._C.awq_gemm(input, qweight, scales, qzeros, split_k_iters)
zhuwenwen's avatar
zhuwenwen committed
543

544
545
546
547
548
549
550
551
552
553
def awq_gemm(
    input: torch.Tensor, 
    weight: torch.Tensor,
    zeros_and_scales:torch.Tensor,
    m:int,n:int,k:int,
    group_size:int,
    padding_group:int,
    splikspace:torch.Tensor,
    splikspacesize:int
    ) -> torch.Tensor:
gaoqiong's avatar
gaoqiong committed
554
555
556
557
558
559
560
561
562
563
564
    return quant_ops.awq_gemm(input,
                              weight,
                              zeros_and_scales,
                              m,
                              n,
                              k,
                              group_size,
                              padding_group,
                              splikspace,
                              splikspacesize)

565
566
567
568
569
570
571
572
573
574
575
576
def awq_gemm_fake(
    input: torch.Tensor, 
    weight: torch.Tensor,
    zeros_and_scales:torch.Tensor,
    m:int,
    n:int,
    k:int,
    group_size:int,
    padding_group:int,
    splikspace:torch.Tensor,
    splikspacesize:int,
    ) -> torch.Tensor:
577
578
    return torch.empty((m, n), dtype=input.dtype, device=input.device)

gaoqiong's avatar
gaoqiong committed
579
580
581
582
583
584
585
def convert_s4(qw: torch.Tensor, qz: torch.Tensor, s: torch.Tensor,
               group_size: int):
    return quant_ops.convert_s4(qw,qz,s,group_size)

def sz_permute(sz:torch.Tensor)-> torch.Tensor:
    return quant_ops.sz_permute(sz)

586
587
588
589
590
591
592
def dequant_w4_gemm_colmajor(
    qweight:torch.Tensor,
    zeros_and_scale:torch.Tensor,
    k:int,
    n:int,
    group_size:int
    )->torch.Tensor:
gaoqiong's avatar
gaoqiong committed
593
    return quant_ops.dequant_w4_gemm_colmajor(qweight,zeros_and_scale,k,n,group_size)
594
595
596


# gptq
597
598
599
600
601
602
603
def gptq_gemm(
    a: torch.Tensor,
    b_q_weight: torch.Tensor,
    b_gptq_qzeros: torch.Tensor,
    b_gptq_scales: torch.Tensor,
    b_g_idx: torch.Tensor,
    use_exllama: bool,
604
    # use_v2_format: bool,
605
606
    bit: int,
) -> torch.Tensor:
607
608
609
610
611
612
613
614
615
616
617
618
619
620
    # return torch.ops._C.gptq_gemm(
    #     a,
    #     b_q_weight,
    #     b_gptq_qzeros,
    #     b_gptq_scales,
    #     b_g_idx,
    #     use_exllama,
    #     use_v2_format,
    #     bit,
    # )
    return quant_ops.gptq_gemm(
        a, 
        b_q_weight, 
        b_gptq_qzeros, 
621
        b_gptq_scales,
622
623
624
        b_g_idx, 
        use_exllama, 
        bit)
625
626


627
if hasattr(torch.ops._C, "gptq_gemm"):
628

629
    @register_fake("_C::gptq_gemm")
630
631
632
633
634
635
636
    def _gptq_gemm_fake(
        a: torch.Tensor,
        b_q_weight: torch.Tensor,
        b_gptq_qzeros: torch.Tensor,
        b_gptq_scales: torch.Tensor,
        b_g_idx: torch.Tensor,
        use_exllama: bool,
637
        use_v2_format: bool,
638
639
640
641
642
        bit: int,
    ) -> torch.Tensor:
        return torch.empty(
            (a.size(0), b_q_weight.size(1)), dtype=a.dtype, device=a.device
        )
643
644


645
def gptq_shuffle(q_weight: torch.Tensor, q_perm: torch.Tensor, bit: int) -> None:
646
    # torch.ops._C.gptq_shuffle(q_weight, q_perm, bit)
647
    quant_ops.gptq_shuffle(q_weight, q_perm, bit)
648
649


650
# marlin_24
651
652
653
654
655
656
657
658
659
660
661
662
663
664
def gptq_marlin_24_gemm(
    a: torch.Tensor,
    b_q_weight: torch.Tensor,
    b_meta: torch.Tensor,
    b_scales: torch.Tensor,
    workspace: torch.Tensor,
    b_q_type: ScalarType,
    size_m: int,
    size_n: int,
    size_k: int,
) -> torch.Tensor:
    return torch.ops._C.gptq_marlin_24_gemm(
        a, b_q_weight, b_meta, b_scales, workspace, b_q_type.id, size_m, size_n, size_k
    )
665
666


zhuwenwen's avatar
zhuwenwen committed
667
668
669
# if hasattr(torch.ops._C, "gptq_marlin_24_gemm"):

#     @register_fake("_C::gptq_marlin_24_gemm")
670
671
672
673
674
675
676
677
678
679
680
#     def _gptq_marlin_24_gemm_fake(
#         a: torch.Tensor,
#         b_q_weight: torch.Tensor,
#         b_meta: torch.Tensor,
#         b_scales: torch.Tensor,
#         workspace: torch.Tensor,
#         b_q_type: ScalarType,
#         size_m: torch.SymInt,
#         size_n: torch.SymInt,
#         size_k: torch.SymInt,
#     ) -> torch.Tensor:
zhuwenwen's avatar
zhuwenwen committed
681
682
683
#         return torch.empty((size_m, size_n), device=a.device, dtype=a.dtype)

#     @register_fake("_C::gptq_marlin_gemm")
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
#     def _gptq_marlin_gemm_fake(
#         a: torch.Tensor,
#         c: torch.Tensor | None,
#         b_q_weight: torch.Tensor,
#         b_bias: torch.Tensor | None,
#         b_scales: torch.Tensor,
#         a_scales: torch.Tensor | None,
#         global_scale: torch.Tensor | None,
#         b_zeros: torch.Tensor | None,
#         g_idx: torch.Tensor | None,
#         perm: torch.Tensor | None,
#         workspace: torch.Tensor,
#         b_q_type_id: int,
#         size_m: torch.SymInt,
#         size_n: torch.SymInt,
#         size_k: torch.SymInt,
#         is_k_full: bool = True,
#         use_atomic_add: bool = False,
#         use_fp32_reduce: bool = False,
#         is_zp_float: bool = False,
#     ) -> torch.Tensor:
#         dtype = a.dtype
#         if dtype not in [torch.half, torch.bfloat16]:
#             dtype = b_scales.dtype
#         return torch.empty((size_m, size_n), device=a.device, dtype=dtype)
zhuwenwen's avatar
zhuwenwen committed
709
710

#     @register_fake("_C::awq_dequantize")
711
712
713
714
715
716
717
718
#     def _awq_dequantize_fake(
#         qweight: torch.Tensor,
#         scales: torch.Tensor,
#         zeros: torch.Tensor,
#         split_k_iters: torch.SymInt,
#         thx: int,
#         thy: int,
#     ) -> torch.Tensor:
zhuwenwen's avatar
zhuwenwen committed
719
720
721
#         in_c = qweight.size(0)
#         qout_c = qweight.size(1)
#         out_c = qout_c * 8
722
#         return torch.empty((in_c, out_c), dtype=scales.dtype, device=scales.device)
zhuwenwen's avatar
zhuwenwen committed
723
724

#     @register_fake("_C::awq_gemm")
725
726
727
728
#     def _awq_gemm_fake(
#         input: torch.Tensor,
#         qweight: torch.Tensor,
#         scales: torch.Tensor,
729
#         qzeros: torch.Tensor,
730
731
#         split_k_iters: torch.SymInt,
#     ) -> torch.Tensor:
zhuwenwen's avatar
zhuwenwen committed
732
#         num_in_feats = input.size(0)
733
734
735
736
737
#         return torch.empty(
#             (split_k_iters, num_in_feats, qweight.size(1) * 8),
#             dtype=input.dtype,
#             device=input.device,
#         ).sum(0)
zhuwenwen's avatar
zhuwenwen committed
738
739
740
741
742
743
744

#     @register_fake("_C::machete_mm")
#     def machete_mm_fake(
#         a: torch.Tensor,
#         # b_q Should be the tensor returned by machete_prepack_B
#         b_q: torch.Tensor,
#         b_type: ScalarType,
745
746
747
748
749
750
751
#         out_type: torch.dtype | None = None,
#         b_group_scales: torch.Tensor | None = None,
#         b_group_zeros: torch.Tensor | None = None,
#         b_group_size: int | None = None,
#         b_channel_scales: torch.Tensor | None = None,
#         a_token_scales: torch.Tensor | None = None,
#         schedule: str | None = None,
zhuwenwen's avatar
zhuwenwen committed
752
753
754
755
#     ) -> torch.Tensor:
#         m = a.size(0)
#         n = b_q.size(1)
#         return torch.empty((m, n), device=a.device, dtype=a.dtype)
756

zhuwenwen's avatar
zhuwenwen committed
757
758
#     @register_fake("_C::machete_prepack_B")
#     def machete_prepack_B_fake(
759
760
761
762
763
764
#         b_q_weight: torch.Tensor,
#         a_type: torch.dtype,
#         b_type: ScalarType,
#         group_scales_type: torch.dtype | None,
#     ) -> torch.Tensor:
#         return torch.empty_like(b_q_weight, memory_format=torch.contiguous_format)
765

766
767
#     @register_fake("_C::cutlass_w4a8_mm")
#     def cutlass_w4a8_mm_fake(
768
769
770
771
772
773
774
775
776
777
#         a: torch.Tensor,
#         # b_q Should be the tensor returned by cutlass_encode_and_reorder_int4b
#         b_q: torch.Tensor,
#         b_group_scales: torch.Tensor,
#         b_group_size: int,
#         b_channel_scales: torch.Tensor,
#         a_token_scales: torch.Tensor,
#         out_type: torch.dtype | None = None,
#         maybe_schedule: str | None = None,
#     ) -> torch.Tensor:
778
779
780
781
782
783
784
785
786
787
788
789
790
#         m = a.size(0)
#         n = b_q.size(1)
#         out_dtype = out_type if out_type is not None else torch.bfloat16
#         return torch.empty((m, n), device=a.device, dtype=out_dtype)

#     @register_fake("_C::cutlass_pack_scale_fp8")
#     def cutlass_pack_scale_fp8_fake(scales: torch.Tensor) -> torch.Tensor:
#         return torch.empty_like(scales, memory_format=torch.contiguous_format)

#     @register_fake("_C::cutlass_encode_and_reorder_int4b")
#     def cutlass_encode_and_reorder_int4b_fake(b: torch.Tensor) -> torch.Tensor:
#         return torch.empty_like(b, memory_format=torch.contiguous_format)

791
792
793
794
#     @register_fake("_C::cutlass_encode_and_reorder_int4b_grouped")
#     def cutlass_encode_and_reorder_int4b_grouped_fake(b: torch.Tensor) -> torch.Tensor:
#         return torch.empty_like(b, memory_format=torch.contiguous_format)

795

zhuwenwen's avatar
zhuwenwen committed
796
# if hasattr(torch.ops._C, "allspark_w8a16_gemm"):
797

zhuwenwen's avatar
zhuwenwen committed
798
#     @register_fake("_C::allspark_w8a16_gemm")
799
800
801
802
803
804
805
806
807
808
809
810
811
#     def _allspark_w8a16_gemm_fake(
#         a: torch.Tensor,
#         b_qweight: torch.Tensor,
#         b_scales: torch.Tensor,
#         b_qzeros: torch.Tensor | None,
#         n: torch.SymInt,
#         group_size: torch.SymInt,
#         sm_count: torch.SymInt,
#         sm_version: torch.SymInt,
#         CUBLAS_M_THRESHOLD: torch.SymInt,
#         has_zp: bool,
#         n32k16_reorder: bool,
#     ) -> torch.Tensor:
zhuwenwen's avatar
zhuwenwen committed
812
813
#         m = a.size(0)
#         return torch.empty((m, n), device=a.device, dtype=a.dtype)
814

815

816
if hasattr(torch.ops._C, "ggml_dequantize"):
zhuwenwen's avatar
zhuwenwen committed
817

818
    @register_fake("_C::ggml_dequantize")
819
    def _ggml_dequantize_fake(
820
821
822
823
        W: torch.Tensor,
        quant_type: int,
        m: torch.SymInt,
        n: torch.SymInt,
824
        dtype: torch.dtype | None = None,
825
    ) -> torch.Tensor:
826
827
828
829
830
831
832
833
834
        return torch.empty((m, n), dtype=torch.float16, device=W.device)

    @register_fake("_C::ggml_mul_mat_vec_a8")
    def _ggml_mul_mat_vec_a8_fake(
        W: torch.Tensor,
        X: torch.Tensor,
        quant_type: int,
        row: torch.SymInt,
    ) -> torch.Tensor:
835
        return torch.empty((X.shape[0], row), dtype=X.dtype, device=W.device)
836
837
838
839
840
841
842
843
844

    @register_fake("_C::ggml_mul_mat_a8")
    def _ggml_mul_mat_a8_fake(
        W: torch.Tensor,
        X: torch.Tensor,
        quant_type: int,
        row: torch.SymInt,
    ) -> torch.Tensor:
        batch = X.size(0)
845
        return torch.empty((batch, row), dtype=X.dtype, device=W.device)
846

847
848
849
850
851
852
853
854
855
856
857
858
859
    @register_fake("_C::ggml_moe_a8")
    def _ggml_moe_a8_fake(
        X: torch.Tensor,
        W: torch.Tensor,
        sorted_token_ids: torch.Tensor,
        expert_ids: torch.Tensor,
        num_tokens_post_padded: torch.Tensor,
        quant_type: int,
        row: torch.SymInt,
        top_k: torch.SymInt,
        tokens: torch.SymInt,
    ) -> torch.Tensor:
        tokens = X.size(0)
860
        return torch.empty((tokens * top_k, row), dtype=torch.float16, device=W.device)
861
862


863
864
865
866
867
868
869
870
871
872
873
874
875
if hasattr(torch.ops._C, "ggml_moe_a8_vec"):

    @register_fake("_C::ggml_moe_a8_vec")
    def _ggml_moe_a8_vec_fake(
        X: torch.Tensor,
        W: torch.Tensor,
        topk_ids: torch.Tensor,
        top_k: int,
        quant_type: int,
        row: torch.SymInt,
        tokens: torch.SymInt,
    ) -> torch.Tensor:
        tokens = X.size(0)
876
        return torch.empty((tokens * top_k, row), dtype=X.dtype, device=W.device)
877
878


879
# cutlass
880
881
882
883
def cutlass_scaled_mm_supports_fp4(cuda_device_capability: int) -> bool:
    return torch.ops._C.cutlass_scaled_mm_supports_fp4(cuda_device_capability)


884
885
886
887
888
889
890
891
892
def cutlass_blockwise_scaled_grouped_mm(
    output: torch.Tensor,
    a: torch.Tensor,
    b: torch.Tensor,
    scales_a: torch.Tensor,
    scales_b: torch.Tensor,
    problem_sizes: torch.Tensor,
    expert_offsets: torch.Tensor,
):
893
894
895
    torch.ops._C.cutlass_blockwise_scaled_grouped_mm(
        output, a, b, scales_a, scales_b, problem_sizes, expert_offsets
    )
896
897


898
899
900
901
902
903
904
905
def cutlass_scaled_fp4_mm(
    a: torch.Tensor,
    b: torch.Tensor,
    block_scale_a: torch.Tensor,
    block_scale_b: torch.Tensor,
    alpha: torch.Tensor,
    out_dtype: torch.dtype,
) -> torch.Tensor:
906
907
908
    assert a.ndim == 2 and b.ndim == 2
    m, n = a.shape[0], b.shape[0]
    out = torch.empty((m, n), dtype=out_dtype, device=a.device)
909
    torch.ops._C.cutlass_scaled_fp4_mm(out, a, b, block_scale_a, block_scale_b, alpha)
910
911
912
    return out


913
914
915
916
def cutlass_scaled_mm_supports_fp8(cuda_device_capability: int) -> bool:
    return torch.ops._C.cutlass_scaled_mm_supports_fp8(cuda_device_capability)


917
def cutlass_scaled_mm_supports_block_fp8(cuda_device_capability: int) -> bool:
918
    return torch.ops._C.cutlass_scaled_mm_supports_block_fp8(cuda_device_capability)
919
920


921
922
923
924
925
926
def cutlass_scaled_mm(
    a: torch.Tensor,
    b: torch.Tensor,
    scale_a: torch.Tensor,
    scale_b: torch.Tensor,
    out_dtype: torch.dtype,
927
    bias: torch.Tensor | None = None,
928
) -> torch.Tensor:
929
    """
930
    `cutlass_scaled_mm` implements a fused version of
931
        `output = torch.mm((scale_a * a), (scale_b * b)).to(out_dtype)`
932
933
934
935
936
937
938
939
    where scale_a * a and scale_b * b are implemented using numpy-style
    broadcasting.

    In order to support blockwise scaling like found in DeepSeek V3 we also
    support extended "group" broadcast rules. We extend the numpy-style
    broadcasting rules with the following rule:
        "if the extent of a dimension in the source shape is between 1 and
        corresponding extent in the target shape we repeat each element along
940
941
942
943
944
945
946
947
948
949
950
        that dimension  src_shape[dim] // target_shape[dim] times consecutively"
    example if we have:
          a = [[1, 2], and target_shape = (2, 4)
               [3, 4]]
    then we would expand a to:
          a = [[1, 1, 2, 2],
               [3, 3, 4, 4]]
    currently we only support the case:
        scale_a.shape * [1, 128] == a.shape
        scale_b.shape * [128, 128] == b.shape
    """
951
952
    assert out_dtype is torch.bfloat16 or out_dtype is torch.float16
    assert bias is None or bias.numel() == b.shape[1] and bias.dtype == out_dtype
953

954
    # Massage the input to be 2D
955
956
    # target_shape = (*a.shape[:-1], b.shape[1])
    # a = a.view(-1, a.shape[-1])
957

zhuwenwen's avatar
zhuwenwen committed
958
959
    # cutlass_compatible_b = (b.shape[0] % 16 == 0 and b.shape[1] % 16 == 0)
    # if current_platform.is_rocm() or not cutlass_compatible_b:
zhuwenwen's avatar
zhuwenwen committed
960
961
    #     from vllm.model_executor.layers.quantization.compressed_tensors.triton_scaled_mm import (  # noqa
    #         triton_scaled_mm)
962
963
964
965
966
967
968
969
    #     out = triton_scaled_mm(a, b, scale_a, scale_b, out_dtype, bias)
    # else:
    #     out = torch.empty((a.shape[0], b.shape[1]),
    #                       dtype=out_dtype,
    #                       device=a.device)
    #     torch.ops._C.cutlass_scaled_mm(out, a, b, scale_a, scale_b, bias)

    # return out.view(*target_shape)
gaoqiong's avatar
gaoqiong committed
970
    return quant_ops.rocblas_scaled_mm_nn(a, b, scale_a, scale_b, out_dtype, bias)
971

972
973
974
975
976
def rocblas_scaled_mm(a: torch.Tensor,
                      b: torch.Tensor,
                      scale_a: torch.Tensor,
                      scale_b: torch.Tensor,
                      out_dtype: torch.dtype,
977
                      bias: torch.Tensor | None = None) -> torch.Tensor:
978

979
980
981
982
983
984
985
986
987
988
    # cutlass_compatible_b = b.shape[0] % 16 == 0 and b.shape[1] % 16 == 0
    # if current_platform.is_rocm() or not cutlass_compatible_b:
    #     from vllm.model_executor.layers.quantization.compressed_tensors.triton_scaled_mm import (  # noqa
    #         triton_scaled_mm,
    #     )

    #     out = triton_scaled_mm(a, b, scale_a, scale_b, out_dtype, bias)
    # else:
    #     out = torch.empty((a.shape[0], b.shape[1]), dtype=out_dtype, device=a.device)
    #     torch.ops._C.cutlass_scaled_mm(out, a, b, scale_a, scale_b, bias)
989
990
    return quant_ops.rocblas_scaled_mm_nn(a, b, scale_a, scale_b, out_dtype, bias)

zhuwenwen's avatar
zhuwenwen committed
991
992
993
994
995
def blaslt_scaled_mm(a: torch.Tensor,
                      b: torch.Tensor,
                      scale_a: torch.Tensor,
                      scale_b: torch.Tensor,
                      out_dtype: torch.dtype,
996
                      bias: torch.Tensor | None = None) -> torch.Tensor:
zhuwenwen's avatar
zhuwenwen committed
997
998
999
1000
    m = a.shape[0]
    n = b.shape[0]
    k = a.shape[1]
    _, out = quant_ops.hipblaslt_w8a8_gemm(a, b, scale_a, scale_b, m, n, k, 'NT', out_dtype)
zhuwenwen's avatar
zhuwenwen committed
1001
1002
    if bias is not None:
        out += bias
zhuwenwen's avatar
zhuwenwen committed
1003
1004
    return out

1005
1006
1007
1008
1009
def triton_scaled_mm(a: torch.Tensor,
                      b: torch.Tensor,
                      scale_a: torch.Tensor,
                      scale_b: torch.Tensor,
                      out_dtype: torch.dtype,
1010
1011
                      bias: torch.Tensor | None = None,
                      best_config: list | None = None) -> torch.Tensor:
1012

gaoqiong's avatar
gaoqiong committed
1013
    return quant_ops.triton_scaled_mm(a, b,scale_a,scale_b,out_dtype,bias,best_config)
1014

gaoqiong's avatar
gaoqiong committed
1015
1016
1017
1018
1019
1020
def triton_int8_gemm_helper(m: int,
                             n: int,
                             k: int,
                             per_token_act_quant: bool,
                             per_out_channel_weight_quant: bool,
                             use_bias: bool,
zhuwenwen's avatar
zhuwenwen committed
1021
                             out_dtype: type[torch.dtype] = torch.float16,
1022
                             device: str = "cuda:0",
1023
1024
                             best_config: list | None = None,
                             repeat: int | None = 2):
1025
1026
1027
1028
1029
1030
1031
1032
1033
    return quant_tools.triton_int8_gemm_helper(m,n,k,per_token_act_quant,per_out_channel_weight_quant,use_bias,out_dtype,device,best_config,repeat)

def triton_blockint8_gemm_helper(m: int,
                                n: int,
                                k: int,
                                block_size:list=[128,128],
                                use_bias: bool=False,
                                out_dtype: type[torch.dtype] = torch.bfloat16,
                                device: str = "cuda:0",
1034
1035
                                best_config: dict | None = None,
                                repeat: int | None = 2):
1036
1037

    return quant_tools.triton_blockint8_gemm_helper(m,n,k,block_size,use_bias,out_dtype,device,best_config,repeat)
gaoqiong's avatar
gaoqiong committed
1038
1039


1040
1041
1042
1043
1044
1045
1046
def cutlass_scaled_mm_azp(
    a: torch.Tensor,
    b: torch.Tensor,
    scale_a: torch.Tensor,
    scale_b: torch.Tensor,
    out_dtype: torch.dtype,
    azp_adj: torch.Tensor,
1047
1048
    azp: torch.Tensor | None = None,
    bias: torch.Tensor | None = None,
1049
) -> torch.Tensor:
1050
1051
1052
1053
1054
    """
    :param azp_adj: In the per-tensor case, this should include the azp.
    Always per-channel.
    :param azp: Only set in the per-token case. Per-token if set.
    """
1055
1056
1057
    assert b.shape[0] % 16 == 0 and b.shape[1] % 16 == 0
    assert out_dtype is torch.bfloat16 or out_dtype is torch.float16
    assert bias is None or bias.numel() == b.shape[1] and bias.dtype == out_dtype
1058

1059
1060
1061
1062
    # Massage the input to be 2D
    target_shape = (*a.shape[:-1], b.shape[1])
    a = a.view(-1, a.shape[-1])
    assert azp is None or azp.numel() == a.shape[0]
1063

1064
1065
    out = torch.empty((a.shape[0], b.shape[1]), dtype=out_dtype, device=a.device)
    torch.ops._C.cutlass_scaled_mm_azp(out, a, b, scale_a, scale_b, azp_adj, azp, bias)
1066
    return out.view(*target_shape)
1067
1068


1069
def cutlass_sparse_scaled_mm_supported(cuda_device_capability: int) -> bool:
1070
    return torch.ops._C.cutlass_sparse_scaled_mm_supported(cuda_device_capability)
1071
1072


1073
def cutlass_group_gemm_supported(cuda_device_capability: int) -> bool:
1074
1075
1076
1077
1078
    try:
        return torch.ops._C.cutlass_group_gemm_supported(cuda_device_capability)
    except AttributeError:
        # Return False on non-CUDA platforms where it is not available
        return False
1079

1080

1081
def cutlass_sparse_compress(a: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
1082
1083
1084
1085
1086
1087
1088
1089
    """
    Compresses a sparse matrix for use with Cutlass sparse operations.

    This function takes a dense tensor and compresses it into two components:
    non-zero elements and metadata. The compressed representation is compatible
    with Cutlass sparse kernels.

    Args:
1090
        a (torch.Tensor):
1091
1092
1093
1094
1095
1096
1097
            The input tensor to be compressed. Must have one of the following data types:
            - `torch.int8`
            - `torch.float8_e4m3fn`
            - `torch.bfloat16`
            - `torch.float16`

    Returns:
1098
        tuple[torch.Tensor, torch.Tensor]:
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
            A tuple containing:
            - `a_nzs` (torch.Tensor): A tensor containing non-zero elements of `a`.
            - `a_meta` (torch.Tensor): A tensor containing metadata for the sparse representation.

    Raises:
        ValueError: If the compression operation fails.

    Notes:
        - The `a_meta` tensor has a data type of `torch.uint8`.
        - Each metadata element encodes the sparsity of 4 non-zero elements (i.e., `elemsPerMetaElem = 4`).
        - The shape of `a_nzs` is `(m, k // 2)`, where `m` and `k` are the dimensions of the input tensor.
        - The shape of `a_meta` is `(m, k // 2 // elemsPerMetaElem)`.
    """
1112
1113
    assert a.dtype in [torch.int8, torch.float8_e4m3fn, torch.bfloat16, torch.float16]
    assert a.is_contiguous()
1114
1115
1116

    # a_meta.dtype: torch.uint8 so elemsPerMetaElem = 8b / 2b_per_nz = 4
    elemsPerMetaElem = 4
1117
    assert a.shape[1] % (2 * elemsPerMetaElem) == 0
1118

1119
    return torch.ops._C.cutlass_sparse_compress(a)
1120
1121
1122


def cutlass_scaled_sparse_mm(
1123
1124
1125
1126
1127
1128
    a: torch.Tensor,
    bt_nzs: torch.Tensor,
    bt_meta: torch.Tensor,
    scale_a: torch.Tensor,
    scale_b: torch.Tensor,
    out_dtype: torch.dtype,
1129
    bias: torch.Tensor | None = None,
1130
) -> torch.Tensor:
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
    """
    Performs a scaled sparse matrix multiplication using Cutlass.

    Steps:
    1. Create a dense matrix `a` of shape (m, k) on the CUDA device:
    `a = torch.randn((m, k), device='cuda')`.

    2. Create a dense matrix `b` of shape (k, n) on the CUDA device:
    `b = torch.randn((k, n), device='cuda')`.

    3. Prune matrix `b` to 2:4 sparsity along the specified dimension:
    `b = prune_to_2_4(b, dim=0)`.

    4. Compress the transposed sparse matrix `b.t()`:
    `bt_nzs, bt_meta = cutlass_sparse_compress(b.t())`.

    5. Perform sparse matrix multiplication using the compressed matrix,
    applying scaling factors for `a` and `b`, and the output data type:
    `out = cutlass_scaled_sparse_mm(a, bt_nzs, bt_meta, scale_a, scale_b, out_dtype)`.

    Returns:
    - The result of the scaled sparse matrix multiplication.
    """
1154
1155
1156
    assert bt_nzs.shape[0] % 16 == 0 and bt_nzs.shape[1] % 16 == 0
    assert out_dtype is torch.bfloat16 or out_dtype is torch.float16
    assert bias is None or bias.shape[0] == bt_nzs.shape[0] and bias.dtype == out_dtype
1157
1158
1159
1160
1161

    m = a.shape[0]
    n = bt_nzs.shape[0]
    out = torch.empty((m, n), dtype=out_dtype, device=a.device)

1162
1163
1164
    torch.ops._C.cutlass_scaled_sparse_mm(
        out, a, bt_nzs, bt_meta, scale_a, scale_b, bias
    )
1165
1166
1167
1168

    return out


1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
def get_cutlass_moe_mm_data(
    topk_ids: torch.Tensor,
    expert_offsets: torch.Tensor,
    problem_sizes1: torch.Tensor,
    problem_sizes2: torch.Tensor,
    input_permutation: torch.Tensor,
    output_permutation: torch.Tensor,
    num_experts: int,
    n: int,
    k: int,
1179
    blockscale_offsets: torch.Tensor | None = None,
1180
):
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
    """
    Prepare data necessary to perform CUTLASS grouped matrix multiplications
    used in CUTLASS-based fused MoE.

    The function takes in topk_ids (token-expert mapping) and uses it to
    compute:
    - expert_offsets: Indices that mark at which token index each expert begins
                      its computation after the input is sorted with
                      input_permutation. The number of tokens computed with
                      expert E is expert_offsets[E + 1] - expert_offsets[E]
    - problem_sizes1, problem_sizes2: MxNxK sizes of each expert's
                                      multiplication in two grouped MMs used in
                                      the fused MoE operation.
    - input_permutation: Permutation that must be used to shuffle the input
                         before executing the MMs.
    - output_permutation: Permutation that must be used to shuffle the output
                          after executing the MMs.
1198
1199
1200
1201
1202
    - blockscale_offsets: Optional argument passed for fp4 moe. Indices that
                          mark at which block scale index each expert begins
                          its computation. The number of block scale rows
                          computed with expert E is blockscale_offsets[E + 1] -
                          blockscale_offsets[E]
1203
    """
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
    return torch.ops._C.get_cutlass_moe_mm_data(
        topk_ids,
        expert_offsets,
        problem_sizes1,
        problem_sizes2,
        input_permutation,
        output_permutation,
        num_experts,
        n,
        k,
        blockscale_offsets,
    )
1216
1217


1218
def get_cutlass_moe_mm_problem_sizes(
1219
1220
1221
1222
1223
1224
    topk_ids: torch.Tensor,
    problem_sizes1: torch.Tensor,
    problem_sizes2: torch.Tensor,
    num_experts: int,
    n: int,
    k: int,
1225
    blockscale_offsets: torch.Tensor | None = None,
1226
    force_swap_ab: bool | None = None,
1227
):
1228
1229
1230
1231
1232
1233
1234
1235
    """
    Compute only the per-expert problem sizes needed by the two grouped matrix
    multiplications used in CUTLASS-based fused MoE.

    The function takes in topk_ids (token→expert mapping) and computes:
    - problem_sizes1, problem_sizes2: M×N×K sizes of each expert's
                                    multiplication for the two grouped MMs
                                    used in the fused MoE operation.
1236
1237
1238
1239
    Optional:
    - force_swap_ab: If set to True or False, explicitly enable or disable the
                     A/B input swap optimization. If None (default), the swap
                     is selected automatically based on tensor sizes.
1240
1241
    """
    return torch.ops._C.get_cutlass_moe_mm_problem_sizes(
1242
1243
1244
1245
1246
1247
1248
1249
        topk_ids,
        problem_sizes1,
        problem_sizes2,
        num_experts,
        n,
        k,
        blockscale_offsets,
        force_swap_ab,
1250
    )
1251
1252


1253
1254
1255
1256
1257
1258
def shuffle_rows(input_tensor: torch.Tensor, dst2src_map: torch.Tensor):
    """
    Shuffle and expand the input tensor according to the dst2src_map and store the result in output_tensor.
    This is used in MoE to permute the input tensor before performing grouped matrix multiplications.
    """
    num_tokens_permuted = dst2src_map.shape[0]
1259
1260
1261
1262
1263
    output_tensor = torch.empty(
        (num_tokens_permuted, input_tensor.shape[1]),
        device=input_tensor.device,
        dtype=input_tensor.dtype,
    )
1264
1265
    torch.ops._moe_C.shuffle_rows(input_tensor, dst2src_map, output_tensor)
    return output_tensor
1266
1267


1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
def get_cutlass_pplx_moe_mm_data(
    expert_offsets: torch.Tensor,
    problem_sizes1: torch.Tensor,
    problem_sizes2: torch.Tensor,
    expert_num_tokens: torch.Tensor,
    num_local_experts: int,
    padded_m: int,
    n: int,
    k: int,
):
1278
1279
1280
1281
1282
    """
    Prepare data necessary to perform CUTLASS grouped matrix multiplications
    used in CUTLASS-based fused MoE.

    The function takes in expert_num_tokens (token count per expert) and
1283
    non_zero_expert_idxs (consecutive indices of experts with non-zero token
1284
1285
1286
1287
1288
1289
1290
1291
    counts) and uses them to compute:
    - expert_offsets: Indices that mark at which token index each expert begins
                      its computation.
    - problem_sizes1, problem_sizes2: MxNxK sizes of each expert's
                                      multiplication in two grouped MMs used in
                                      the fused MoE operation.
    """
    return torch.ops._C.get_cutlass_pplx_moe_mm_data(
1292
1293
1294
1295
1296
1297
1298
1299
1300
        expert_offsets,
        problem_sizes1,
        problem_sizes2,
        expert_num_tokens,
        num_local_experts,
        padded_m,
        n,
        k,
    )
1301
1302


1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
def cutlass_moe_mm(
    out_tensors: torch.Tensor,
    a_tensors: torch.Tensor,
    b_tensors: torch.Tensor,
    a_scales: torch.Tensor,
    b_scales: torch.Tensor,
    expert_offsets: torch.Tensor,
    problem_sizes: torch.Tensor,
    a_strides: torch.Tensor,
    b_strides: torch.Tensor,
    c_strides: torch.Tensor,
    per_act_token: bool,
    per_out_ch: bool,
):
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
    """
    A single grouped matrix multiplication used in CUTLASS-based fused MoE.
    The function executes fp8-quantized OUT = AB matrix multiplication.

    - expert_offsets: Indices that mark at which token index each expert begins
                      its computation. The number of tokens computed with
                      expert E is expert_offsets[E + 1] - expert_offsets[E]
    - problem_sizes: MxNxK sizes of each expert's multiplication in two grouped
                     MMs used in the fused MoE operation.
    - a/b/c_strides: The data strides passed to grouped matrix multiplication.
    """
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
    return torch.ops._C.cutlass_moe_mm(
        out_tensors,
        a_tensors,
        b_tensors,
        a_scales,
        b_scales,
        expert_offsets,
        problem_sizes,
        a_strides,
        b_strides,
        c_strides,
        per_act_token,
        per_out_ch,
    )
1342
1343


1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
def cutlass_fp4_moe_mm(
    out_tensors: torch.Tensor,
    a_tensors: torch.Tensor,
    b_tensors: torch.Tensor,
    a_scales: torch.Tensor,
    b_scales: torch.Tensor,
    alphas: torch.Tensor,
    problem_sizes: torch.Tensor,
    expert_offsets: torch.Tensor,
    sf_offsets: torch.Tensor,
):
1355
    """
1356
    An FP4 Blockscaled Group Gemm that takes in  a_tensors, b_tensors and runs
1357
1358
1359
1360
1361
1362
    the gemms for each combination based on the specified problem sizes.

    This is used as the MoE gemm during NVFP4 Quantized FusedMoE forward.
    - a/b_tensors: the NVFP4 a_ptrs and b_ptrs tensors which are quantized
                     input and expert weights.
    - a_/b_scales: The blockscales in FP8-E4M3 precision
1363
1364
1365
1366
    - expert_offsets/sf_offsets: Indices that mark at which token index
                    each expert begins its computation. The number of tokens
                    computed with expert E is expert_offsets[E + 1] -
                    expert_offsets[E] And the sf_size per expert is
1367
1368
1369
1370
                    sf_offset[E+1] - sf_offset[E]
    - problem_sizes: MxNxK sizes of each expert's multiplication in two grouped
                     MMs used in the fused MoE operation.
    """
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
    return torch.ops._C.cutlass_fp4_group_mm(
        out_tensors,
        a_tensors,
        b_tensors,
        a_scales,
        b_scales,
        alphas,
        problem_sizes,
        expert_offsets,
        sf_offsets,
    )
1382
1383


1384
# gptq_marlin
1385
1386
1387
1388
1389
1390
def gptq_marlin_repack(
    b_q_weight: torch.Tensor,
    perm: torch.Tensor,
    size_k: int,
    size_n: int,
    num_bits: int,
1391
    is_a_8bit: bool = False,
1392
) -> torch.Tensor:
1393
1394
1395
    return torch.ops._C.gptq_marlin_repack(
        b_q_weight, perm, size_k, size_n, num_bits, is_a_8bit
    )
1396
1397


1398
1399
1400
1401
1402
1403
1404
1405
1406
if hasattr(torch.ops._C, "gptq_marlin_repack"):

    @register_fake("_C::gptq_marlin_repack")
    def _gptq_marlin_repack_fake(
        b_q_weight: torch.Tensor,
        perm: torch.Tensor,
        size_k: torch.SymInt,
        size_n: torch.SymInt,
        num_bits: int,
1407
        is_a_8bit: bool = False,
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
    ) -> torch.Tensor:
        pack_factor = 32 // num_bits
        marlin_tile_size = 16
        return torch.empty(
            (size_k // marlin_tile_size, size_n * marlin_tile_size // pack_factor),
            dtype=b_q_weight.dtype,
            device=b_q_weight.device,
        )


# awq_marlin
1419
def awq_marlin_repack(
1420
1421
1422
1423
1424
    b_q_weight: torch.Tensor,
    size_k: int,
    size_n: int,
    num_bits: int,
    is_a_8bit: bool = False,
1425
) -> torch.Tensor:
1426
1427
1428
    return torch.ops._C.awq_marlin_repack(
        b_q_weight, size_k, size_n, num_bits, is_a_8bit
    )
1429
1430


1431
if hasattr(torch.ops._C, "awq_marlin_repack"):
1432

1433
1434
1435
1436
1437
1438
    @register_fake("_C::awq_marlin_repack")
    def _awq_marlin_repack_fake(
        b_q_weight: torch.Tensor,
        size_k: torch.SymInt,
        size_n: torch.SymInt,
        num_bits: int,
1439
        is_a_8bit: bool = False,
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
    ) -> torch.Tensor:
        pack_factor = 32 // num_bits
        marlin_tile_size = 16
        return torch.empty(
            (size_k // marlin_tile_size, size_n * marlin_tile_size // pack_factor),
            dtype=b_q_weight.dtype,
            device=b_q_weight.device,
        )


1450
1451
1452
1453
1454
1455
def gptq_marlin_moe_repack(
    b_q_weight: torch.Tensor,
    perm: torch.Tensor,
    size_k: int,
    size_n: int,
    num_bits: int,
1456
    is_a_8bit: bool = False,
1457
) -> torch.Tensor:
1458
1459
    num_experts = b_q_weight.shape[0]
    assert size_k % 16 == 0
1460
1461
1462
1463
1464
    output = torch.empty(
        (num_experts, size_k // 16, size_n * (num_bits // 2)),
        device=b_q_weight.device,
        dtype=b_q_weight.dtype,
    )
1465
    for e in range(num_experts):
1466
        output[e] = torch.ops._C.gptq_marlin_repack(
1467
            b_q_weight[e], perm[e], size_k, size_n, num_bits, is_a_8bit
1468
        )
1469
1470
1471
    return output


1472
1473
1474
1475
1476
1477
def awq_marlin_moe_repack(
    b_q_weight: torch.Tensor,
    perm: torch.Tensor,
    size_k: int,
    size_n: int,
    num_bits: int,
1478
    is_a_8bit: bool = False,
1479
) -> torch.Tensor:
1480
1481
    num_experts = b_q_weight.shape[0]
    assert size_k % 16 == 0
1482
1483
1484
1485
1486
    output = torch.empty(
        (num_experts, size_k // 16, size_n * (num_bits // 2)),
        device=b_q_weight.device,
        dtype=b_q_weight.dtype,
    )
1487
    for e in range(num_experts):
1488
        output[e] = torch.ops._C.awq_marlin_repack(
1489
            b_q_weight[e], size_k, size_n, num_bits, is_a_8bit
1490
        )
1491
1492
1493
    return output


1494
1495
1496
1497
1498
1499
1500
1501
def marlin_int4_fp8_preprocess(
    qweight: torch.Tensor,
    qzeros_or_none: torch.Tensor | None = None,
    inplace: bool = False,
):
    return torch.ops._C.marlin_int4_fp8_preprocess(qweight, qzeros_or_none, inplace)


1502
1503
def gptq_marlin_gemm(
    a: torch.Tensor,
1504
    c: torch.Tensor | None,
1505
    b_q_weight: torch.Tensor,
1506
    b_bias: torch.Tensor | None,
1507
    b_scales: torch.Tensor,
1508
    a_scales: torch.Tensor | None,
1509
1510
1511
1512
    global_scale: torch.Tensor | None,
    b_zeros: torch.Tensor | None,
    g_idx: torch.Tensor | None,
    perm: torch.Tensor | None,
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
    workspace: torch.Tensor,
    b_q_type: ScalarType,
    size_m: int,
    size_n: int,
    size_k: int,
    is_k_full: bool = True,
    use_atomic_add: bool = False,
    use_fp32_reduce: bool = False,
    is_zp_float: bool = False,
) -> torch.Tensor:
    return torch.ops._C.gptq_marlin_gemm(
        a,
        c,
        b_q_weight,
        b_bias,
        b_scales,
1529
        a_scales,
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
        global_scale,
        b_zeros,
        g_idx,
        perm,
        workspace,
        b_q_type.id,
        size_m,
        size_n,
        size_k,
        is_k_full,
        use_atomic_add,
        use_fp32_reduce,
        is_zp_float,
    )
1544
1545


1546
# machete
1547
def machete_supported_schedules(
1548
1549
    a_type: torch.dtype,
    b_type: ScalarType,
1550
1551
1552
1553
1554
    group_scales_type: torch.dtype | None,
    group_zeros_type: torch.dtype | None = None,
    channel_scales_type: torch.dtype | None = None,
    token_scales_type: torch.dtype | None = None,
    out_type: torch.dtype | None = None,
1555
) -> list[str]:
1556
    return torch.ops._C.machete_supported_schedules(
1557
1558
1559
1560
1561
1562
1563
1564
        a_type,
        b_type.id,
        group_scales_type,
        group_zeros_type,
        channel_scales_type,
        token_scales_type,
        out_type,
    )
1565
1566


1567
def machete_mm(
1568
1569
1570
1571
    a: torch.Tensor,
    # b_q Should be the tensor returned by machete_prepack_B
    b_q: torch.Tensor,
    b_type: ScalarType,
1572
1573
1574
1575
1576
1577
1578
    out_type: torch.dtype | None = None,
    b_group_scales: torch.Tensor | None = None,
    b_group_zeros: torch.Tensor | None = None,
    b_group_size: int | None = None,
    b_channel_scales: torch.Tensor | None = None,
    a_token_scales: torch.Tensor | None = None,
    schedule: str | None = None,
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
) -> torch.Tensor:
    return torch.ops._C.machete_mm(
        a,
        b_q,
        b_type.id,
        out_type,
        b_group_scales,
        b_group_zeros,
        b_group_size,
        b_channel_scales,
        a_token_scales,
        schedule,
    )
1592
1593
1594


def machete_prepack_B(
1595
1596
1597
    b_q_weight: torch.Tensor,
    a_type: torch.dtype,
    b_type: ScalarType,
1598
    group_scales_type: torch.dtype | None,
1599
1600
1601
1602
) -> torch.Tensor:
    return torch.ops._C.machete_prepack_B(
        b_q_weight, a_type, b_type.id, group_scales_type
    )
1603
1604


1605
1606
# CUTLASS W4A8
def cutlass_w4a8_mm(
1607
1608
1609
1610
1611
1612
1613
    a: torch.Tensor,
    # b_q Should be the tensor returned by cutlass_encode_and_reorder_int4b
    b_q: torch.Tensor,
    b_group_scales: torch.Tensor,
    b_group_size: int,
    b_channel_scales: torch.Tensor,
    a_token_scales: torch.Tensor,
1614
1615
    out_type: torch.dtype | None = None,
    maybe_schedule: str | None = None,
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
) -> torch.Tensor:
    return torch.ops._C.cutlass_w4a8_mm(
        a,
        b_q,
        b_group_scales,
        b_group_size,
        b_channel_scales,
        a_token_scales,
        out_type,
        maybe_schedule,
    )
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636


def cutlass_pack_scale_fp8(scales: torch.Tensor) -> torch.Tensor:
    return torch.ops._C.cutlass_pack_scale_fp8(scales)


def cutlass_encode_and_reorder_int4b(b: torch.Tensor) -> torch.Tensor:
    return torch.ops._C.cutlass_encode_and_reorder_int4b(b)


1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
def cutlass_w4a8_moe_mm(
    out_tensors: torch.Tensor,
    a_tensors: torch.Tensor,
    b_tensors: torch.Tensor,
    a_scales: torch.Tensor,
    b_scales: torch.Tensor,
    b_group_scales: torch.Tensor,
    b_group_size: int,
    expert_offsets: torch.Tensor,
    problem_sizes: torch.Tensor,
    a_strides: torch.Tensor,
    b_strides: torch.Tensor,
    c_strides: torch.Tensor,
    group_scale_strides: torch.Tensor,
    maybe_schedule: str | None = None,
):
    """
    Executes the CUTLASS-based fused-MoE grouped matrix multiplication for the
    W4A8 quantization scheme. Uses group-wise quantization (INT4 -> FP8)
    and both per-channel + per-token scaling in the epilogue.

    Args:
        out_tensors:
            Output buffer for all experts (updated in-place).
        a_tensors:
            FP8 (E4M3FN) activations for all experts.
        b_tensors:
            INT4-packed weight matrix for all experts, packed to INT32
        a_scales:
            Per-token FP8 activation scales, applied in the epilogue.
        b_scales:
            Per-channel FP8 weight scales for each expert, applied in the epilogue.
        b_group_scales:
            FP8 scale values for group-wise INT4 weight blocks.
        b_group_size:
            Number of elements grouped under each entry of b_group_scales.
        expert_offsets:
            Cumulative token offsets
        problem_sizes:
            Per-expert (M, N, K) GEMM sizes used by the grouped GEMM launcher.
        a/b/c/group_scale_strides:
            Strides describing the memory layout of the input tensors.
        maybe_schedule:
            Optional override to choose a specific kernel or epilogue schedule.

    Returns:
        out_tensors updated in-place with the dequantized INT4xFP8 grouped GEMM result.
    """
    return torch.ops._C.cutlass_w4a8_moe_mm(
        out_tensors,
        a_tensors,
        b_tensors,
        a_scales,
        b_scales,
        b_group_scales,
        b_group_size,
        expert_offsets,
        problem_sizes,
        a_strides,
        b_strides,
        c_strides,
        group_scale_strides,
        maybe_schedule,
    )


def cutlass_encode_and_reorder_int4b_grouped(
    b_tensors: torch.Tensor,
) -> tuple[torch.Tensor, torch.Tensor]:
    return torch.ops._C.cutlass_encode_and_reorder_int4b_grouped(b_tensors)


1709
if hasattr(torch.ops._C, "permute_cols"):
1710

1711
    @register_fake("_C::permute_cols")
1712
    def _permute_cols_fake(a: torch.Tensor, perm: torch.Tensor) -> torch.Tensor:
1713
1714
1715
1716
1717
1718
1719
        return torch.empty_like(a)


def permute_cols(a: torch.Tensor, perm: torch.Tensor) -> torch.Tensor:
    return torch.ops._C.permute_cols(a, perm)


1720
1721
# fp4
def scaled_fp4_quant(
1722
1723
    input: torch.Tensor, input_global_scale: torch.Tensor
) -> tuple[torch.Tensor, torch.Tensor]:
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
    """
    Quantize input tensor to FP4 and return quantized tensor and scale.

    This function quantizes the last dimension of the given tensor `input`. For
    every 16 consecutive elements, a single dynamically computed scaling factor
    is shared. This scaling factor is quantized using the `input_global_scale`
    and is stored in a swizzled layout (see
    https://docs.nvidia.com/cuda/parallel-thread-execution/#tcgen05-mma-scale-factor-b-layout-4x).

    Args:
        input: The input tensor to be quantized to FP4
        input_global_scale: A scalar scaling factor for the entire tensor.

    Returns:
1738
        tuple[torch.Tensor, torch.Tensor]: The output tensor in FP4 but every
1739
1740
1741
            two values are packed into a uint8 and float8_e4m3 scaling factors
            in the sizzled layout.
    """
1742
    assert not current_platform.is_rocm()
1743
    assert input.ndim >= 1, f"input.ndim needs to be >= 1, but got {input.ndim}."
1744
1745
1746
1747
1748
1749
    other_dims = 1 if input.ndim == 1 else -1
    input = input.reshape(other_dims, input.shape[-1])
    m, n = input.shape
    block_size = 16
    device = input.device

1750
    assert n % block_size == 0, f"last dim has to be multiple of 16, but got {n}."
1751
    assert input.dtype in (torch.float16, torch.bfloat16), (
1752
1753
        f"input.dtype needs to be fp16 or bf16 but got {input.dtype}."
    )
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766

    # Two fp4 values will be packed into an uint8.
    output = torch.empty((m, n // 2), device=device, dtype=torch.uint8)

    # We use the rounded values to store the swizzled values. Due to the
    # requirement of the Tensor Core, the minimum tile is 128x4 for the scales.
    # So, we first pad the scales to multiples of 128 and 4. Then, the scales
    # (in float8_e4m3fn) are packed into an int32 for every 4 values. More:
    # https://docs.nvidia.com/cuda/parallel-thread-execution/#tcgen05-mma-scale-factor-b-layout-4x
    round_up = lambda x, y: (x + y - 1) // y * y
    rounded_m = round_up(m, 128)
    scale_n = n // block_size
    rounded_n = round_up(scale_n, 4)
1767
    output_scale = torch.empty(
1768
1769
        (rounded_m, rounded_n // 4), device=device, dtype=torch.int32
    )
1770

1771
    torch.ops._C.scaled_fp4_quant(output, input, output_scale, input_global_scale)
1772
1773
1774
1775
    output_scale = output_scale.view(torch.float8_e4m3fn)
    return output, output_scale


1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
def scaled_fp4_experts_quant(
    input_tensor: torch.Tensor,
    input_global_scale: torch.Tensor,
    expert_offsets: torch.Tensor,
    blockscale_offsets: torch.Tensor,
    topk: int,
) -> tuple[torch.Tensor, torch.Tensor]:
    """
    Quantize input tensor to FP4 and return quantized tensor and scale, for
    packed MoE Inputs.
    Args:
1787
        input_tensor: The input tensor to be quantized to FP4
1788
1789
1790
1791
1792
1793
1794
1795
1796
        input_global_scale: A scalar scaling factor for the entire tensor.
        expert_offsets: The expert offsets tensor
        blockscale_offsets: The blockscale offsets tensor
    Outputs:
        output: The quantized tensor in FP4
        output_scales: The blockscale tensor in FP8-E4M3
    """
    assert not current_platform.is_rocm()
    assert input_tensor.ndim == 2, (
1797
1798
        f"input.ndim needs to be == 2, but got {input_tensor.ndim}."
    )
1799

1800
1801
1802
1803
1804
    # Control the maximum number of tokens per expert supported by the
    # NVFP4 MoE Expert Quantization. This is used to prevent the kernel
    # from running out of memory. This value can also be increased to support
    # larger models.
    MAX_TOKENS_PER_EXPERT = envs.VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE
1805
1806
    m_numtopk, k = input_tensor.shape

1807
    assert m_numtopk <= MAX_TOKENS_PER_EXPERT * topk, (
1808
1809
1810
        f"m_numtopk must be less than MAX_TOKENS_PER_EXPERT("
        f"{MAX_TOKENS_PER_EXPERT})"
        f" for cutlass_moe_fp4, observed m_numtopk = {m_numtopk}. Use"
1811
1812
        f" VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE to set this value."
    )
1813
1814
1815
1816
    scales_k = k // 16
    padded_k = (scales_k + (4 - 1)) // 4

    # output is uint8 and packed fp4 values
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
    output = torch.empty(
        m_numtopk, k // 2, device=input_tensor.device, dtype=torch.uint8
    )
    output_scales = torch.empty(
        MAX_TOKENS_PER_EXPERT * topk,
        padded_k,
        dtype=torch.int32,
        device=input_tensor.device,
    )
    torch.ops._C.scaled_fp4_experts_quant(
        output,
        output_scales,
        input_tensor,
        input_global_scale,
        expert_offsets,
        blockscale_offsets,
    )
1834
1835
1836
1837
    output_scales = output_scales.view(torch.float8_e4m3fn)
    return output, output_scales


1838
# fp8
zhuwenwen's avatar
zhuwenwen committed
1839
1840
# def scaled_fp8_quant(
#     input: torch.Tensor,
1841
1842
1843
#     scale: torch.Tensor | None = None,
#     num_token_padding: int | None = None,
#     scale_ub: torch.Tensor | None = None,
1844
#     use_per_token_if_dynamic: bool = False,
1845
#     output: torch.Tensor | None = None,
zhuwenwen's avatar
zhuwenwen committed
1846
# ) -> tuple[torch.Tensor, torch.Tensor]:
zhuwenwen's avatar
zhuwenwen committed
1847
1848
1849
1850
1851
1852
#     """
#     Quantize input tensor to FP8 and return quantized tensor and scale.

#     This function supports both static and dynamic quantization: If you
#     provide the scale, it will use static scaling and if you omit it,
#     the scale will be determined dynamically. The function also allows
1853
#     optional padding of the output tensors for downstream kernels that
zhuwenwen's avatar
zhuwenwen committed
1854
1855
1856
1857
1858
#     will benefit from padding.

#     Args:
#         input: The input tensor to be quantized to FP8
#         scale: Optional scaling factor for the FP8 quantization
zhuwenwen's avatar
zhuwenwen committed
1859
#         scale_ub: Optional upper bound for scaling factor in dynamic
1860
#             per token case
1861
#         num_token_padding: If specified, pad the first dimension
zhuwenwen's avatar
zhuwenwen committed
1862
#             of the output to at least this value.
zhuwenwen's avatar
zhuwenwen committed
1863
#         use_per_token_if_dynamic: Whether to do per_tensor or per_token
1864
#             in the dynamic quantization case.
zhuwenwen's avatar
zhuwenwen committed
1865
1866

#     Returns:
zhuwenwen's avatar
zhuwenwen committed
1867
#         tuple[torch.Tensor, torch.Tensor]: The output tensor in FP8 and
zhuwenwen's avatar
zhuwenwen committed
1868
1869
#             scaling factor.
#     """
1870
#     # This code assumes batch_dim and num_tokens are flattened
1871
1872
#     assert input.ndim == 2
#     shape: tuple[int, int] | torch.Size = input.shape
zhuwenwen's avatar
zhuwenwen committed
1873
1874
#     # For ROCm on MI300, the output fp8 dtype is torch.float_e3m3fnuz
#     out_dtype: torch.dtype = current_platform.fp8_dtype()
1875
1876
#     if num_token_padding:
#         shape = (max(num_token_padding, input.shape[0]), shape[1])
zhuwenwen's avatar
zhuwenwen committed
1877
1878
1879
#     if output is None:
#         output = torch.empty(shape, device=input.device, dtype=out_dtype)
#     else:
1880
#         assert num_token_padding is None, "padding not supported if output passed in"
zhuwenwen's avatar
zhuwenwen committed
1881
#         assert output.dtype == out_dtype
1882

zhuwenwen's avatar
zhuwenwen committed
1883
#     if scale is None:
1884
#         if use_per_token_if_dynamic:
1885
#             scale = torch.empty((shape[0], 1), device=input.device, dtype=torch.float32)
1886
#             torch.ops._C.dynamic_per_token_scaled_fp8_quant(
1887
1888
#                 output, input, scale, scale_ub
#             )
1889
#         else:
1890
#             scale = torch.empty(1, device=input.device, dtype=torch.float32)
1891
#             torch.ops._C.dynamic_scaled_fp8_quant(output, input, scale)
zhuwenwen's avatar
zhuwenwen committed
1892
#     else:
zhuwenwen's avatar
zhuwenwen committed
1893
#         assert scale.numel() == 1, f"{scale.shape}"
zhuwenwen's avatar
zhuwenwen committed
1894
#         torch.ops._C.static_scaled_fp8_quant(output, input, scale)
1895

zhuwenwen's avatar
zhuwenwen committed
1896
#     return output, scale
1897
1898


1899
1900
# gptq allspark
def allspark_repack_weight(
1901
1902
    qweight: torch.Tensor,
    scale: torch.Tensor,
1903
    zero_point: torch.Tensor | None = None,
1904
    has_zp: bool = False,
1905
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
1906
    """
1907
    Rearrange qweight, scale, and zero_point(if asymmetric) to n32k16 format
1908
1909
1910
1911
1912
1913
1914
1915
    for Ampere W8A16 Fused Gemm kernel

    Args:
        qweight: uint8 weight tensor, original k x n format.
        scale: fp16/bf16 weight scale tensor, 1 x n format.
        zero_point: fp16/bf16 weight zero_point tensor, 1 x n format.
            Must be provided for asymmetric quantization.
        has_zp: if use symmetric quantization, has_zp = False.
1916
1917
            if use asymmetric quantization, has_zp = True.

1918
    Returns:
1919
        tuple[torch.Tensor, torch.Tensor, torch.Tensor | None] :
1920
1921
1922
1923
1924
1925
            rearranged weight, scale, and optionally zero_point.
    """
    K = qweight.shape[0]
    N = qweight.shape[1]
    N_32align = (N + 32 - 1) // 32 * 32

1926
1927
1928
1929
    qweight_reorder = torch.empty(
        (N_32align, K), device=qweight.device, dtype=qweight.dtype
    )
    scale_reorder = torch.empty((1, N_32align), device=scale.device, dtype=scale.dtype)
1930
1931
1932
    zero_point_reorder = None
    if has_zp:
        assert zero_point is not None, (
1933
1934
1935
1936
1937
            "zero_point must be provided for asymmetric quantization."
        )
        zero_point_reorder = torch.empty(
            (1, N_32align), device=zero_point.device, dtype=zero_point.dtype
        )
1938
1939

    torch.ops._C.rearrange_kn_weight_as_n32k16_order(
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
        qweight,
        scale,
        zero_point,
        has_zp,
        qweight_reorder,
        scale_reorder,
        zero_point_reorder,
        K,
        N,
        N_32align,
    )
1951
1952
1953
1954

    return qweight_reorder, scale_reorder, zero_point_reorder


1955
1956
1957
1958
def allspark_w8a16_gemm(
    a: torch.Tensor,
    b_qweight: torch.Tensor,
    b_scales: torch.Tensor,
1959
    b_qzeros: torch.Tensor | None,
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
    n: int,
    group_size: int,
    sm_count: int,
    sm_version: int,
    CUBLAS_M_THRESHOLD: int,
    has_zp: bool,
    n32k16_reorder: bool,
) -> torch.Tensor:
    return torch.ops._C.allspark_w8a16_gemm(
        a,
        b_qweight,
        b_scales,
        b_qzeros,
        n,
        group_size,
        sm_count,
        sm_version,
        CUBLAS_M_THRESHOLD,
        has_zp,
        n32k16_reorder,
    )
1981
1982


1983
# int8
1984
def scaled_int8_quant(
1985
    input: torch.Tensor,
1986
1987
    scale: torch.Tensor | None = None,
    azp: torch.Tensor | None = None,
1988
    symmetric: bool = True,
1989
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor | None]:
zhuwenwen's avatar
zhuwenwen committed
1990
    """
1991
    Quantize the input tensor to int8 and return the quantized tensor and scale, and maybe azp.
zhuwenwen's avatar
zhuwenwen committed
1992
1993
1994

    Args:
        input: The input tensor to be quantized to int8.
1995
1996
        scale: Optional scaling factor for the int8 quantization.
            When not provided, we invoke dynamic-per-token quantization.
1997
1998
1999
        azp: Optional zero-point for the int8 quantization.
            Must be provided for asymmetric quantization if `scale` is provided.
        symmetric: Whether to use symmetric quantization (scale only, azp ignored).
zhuwenwen's avatar
zhuwenwen committed
2000
2001

    Returns:
2002
      tuple[torch.Tensor, torch.Tensor, torch.Tensor | None] : Output int8 tensor, scales, and optionally azp.
zhuwenwen's avatar
zhuwenwen committed
2003
    """
2004
2005
2006
    output = torch.empty_like(input, dtype=torch.int8)
    if scale is not None:
        # static-per-tensor quantization.
2007
2008
2009
        assert symmetric == (azp is None), (
            "azp must only be provided for asymmetric quantization."
        )
2010
        torch.ops._C.static_scaled_int8_quant(output, input, scale, azp)
2011
        return output, scale, azp
2012
2013

    # dynamic-per-token quantization.
2014
2015
2016
2017
2018
2019
2020
    input_scales = torch.empty(
        (input.numel() // input.shape[-1], 1), device=input.device, dtype=torch.float32
    )
    input_azp = None if symmetric else torch.empty_like(input_scales, dtype=torch.int32)
    torch.ops._C.dynamic_scaled_int8_quant(
        output, input.contiguous(), input_scales, input_azp
    )
2021
    return output, input_scales, input_azp
2022
2023


2024
# gguf
2025
def ggml_dequantize(
2026
    W: torch.Tensor, quant_type: int, m: int, n: int, dtype: torch.dtype | None
2027
) -> torch.Tensor:
2028
    return torch.ops._C.ggml_dequantize(W, quant_type, m, n, dtype)
2029
2030
2031
2032
2033
2034
2035


def ggml_mul_mat_vec_a8(
    W: torch.Tensor,
    X: torch.Tensor,
    quant_type: int,
    row: int,
2036
) -> torch.Tensor:
2037
2038
2039
2040
2041
2042
2043
2044
    return torch.ops._C.ggml_mul_mat_vec_a8(W, X, quant_type, row)


def ggml_mul_mat_a8(
    W: torch.Tensor,
    X: torch.Tensor,
    quant_type: int,
    row: int,
2045
) -> torch.Tensor:
2046
2047
2048
    return torch.ops._C.ggml_mul_mat_a8(W, X, quant_type, row)


2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
def ggml_moe_a8(
    X: torch.Tensor,
    W: torch.Tensor,
    sorted_token_ids: torch.Tensor,
    expert_ids: torch.Tensor,
    num_tokens_post_padded: torch.Tensor,
    quant_type: int,
    row: int,
    top_k: int,
    tokens: int,
) -> torch.Tensor:
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
    return torch.ops._C.ggml_moe_a8(
        X,
        W,
        sorted_token_ids,
        expert_ids,
        num_tokens_post_padded,
        quant_type,
        row,
        top_k,
        tokens,
    )
2071
2072


2073
2074
2075
2076
2077
2078
2079
2080
2081
def ggml_moe_a8_vec(
    X: torch.Tensor,
    W: torch.Tensor,
    topk_ids: torch.Tensor,
    top_k: int,
    quant_type: int,
    row: torch.SymInt,
    tokens: torch.SymInt,
) -> torch.Tensor:
2082
    return torch.ops._C.ggml_moe_a8_vec(X, W, topk_ids, top_k, quant_type, row, tokens)
2083
2084


2085
2086
2087
2088
def ggml_moe_get_block_size(quant_type: int) -> int:
    return torch.ops._C.ggml_moe_get_block_size(quant_type)


2089
# mamba
2090
2091
2092
2093
2094
2095
def selective_scan_fwd(
    u: torch.Tensor,
    delta: torch.Tensor,
    A: torch.Tensor,
    B: torch.Tensor,
    C: torch.Tensor,
2096
2097
2098
    D_: torch.Tensor | None,
    z_: torch.Tensor | None,
    delta_bias_: torch.Tensor | None,
2099
    delta_softplus: bool,
2100
2101
2102
    query_start_loc: torch.Tensor | None,
    cache_indices: torch.Tensor | None,
    has_initial_state: torch.Tensor | None,
2103
2104
    ssm_states: torch.Tensor,
    pad_slot_id: int,
2105
2106
2107
2108
    block_size: int = 1024,
    block_idx_first_scheduled_token: torch.Tensor | None = None,
    block_idx_last_scheduled_token: torch.Tensor | None = None,
    initial_state_idx: torch.Tensor | None = None,
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
):
    torch.ops._C.selective_scan_fwd(
        u,
        delta,
        A,
        B,
        C,
        D_,
        z_,
        delta_bias_,
        delta_softplus,
        query_start_loc,
        cache_indices,
        has_initial_state,
        ssm_states,
        pad_slot_id,
2125
2126
2127
2128
        block_size,
        block_idx_first_scheduled_token,
        block_idx_last_scheduled_token,
        initial_state_idx,
2129
    )
2130
2131


2132
# ROCm skinny gemms
2133
def LLMM1(a: torch.Tensor, b: torch.Tensor, rows_per_block: int) -> torch.Tensor:
2134
2135
2136
    return torch.ops._rocm_C.LLMM1(a, b, rows_per_block)


2137
2138
2139
def wvSplitK(
    a: torch.Tensor, b: torch.Tensor, cu_count: int, bias: torch.Tensor = None
) -> torch.Tensor:
2140
    return torch.ops._rocm_C.wvSplitK(a, b, bias, cu_count)
2141
2142


2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
def wvSplitKQ(
    a: torch.Tensor,
    b: torch.Tensor,
    out_dtype: torch.dtype,
    scale_a: torch.Tensor,
    scale_b: torch.Tensor,
    cu_count: int,
    bias: torch.Tensor = None,
) -> torch.Tensor:
    out = torch.empty((b.shape[0], a.shape[0]), dtype=out_dtype, device=b.device)
2153
    torch.ops._rocm_C.wvSplitKQ(a, b, bias, out, scale_a, scale_b, cu_count)
2154
2155
2156
    return out


2157
# moe
2158
2159
def moe_sum(input: torch.Tensor, output: torch.Tensor):
    torch.ops._moe_C.moe_sum(input, output)
zhuwenwen's avatar
zhuwenwen committed
2160
2161
2162
    
def moe_sum_opt1(input: torch.Tensor, output: torch.Tensor):
    torch.ops._moe_C.moe_sum_opt1(input, output)
2163
2164


2165
2166
2167
2168
2169
2170
2171
def moe_align_block_size(
    topk_ids: torch.Tensor,
    num_experts: int,
    block_size: int,
    sorted_token_ids: torch.Tensor,
    experts_ids: torch.Tensor,
    num_tokens_post_pad: torch.Tensor,
2172
    expert_map: torch.Tensor | None = None,
2173
2174
2175
2176
2177
2178
2179
2180
) -> None:
    torch.ops._moe_C.moe_align_block_size(
        topk_ids,
        num_experts,
        block_size,
        sorted_token_ids,
        experts_ids,
        num_tokens_post_pad,
2181
        expert_map,
2182
    )
2183
2184


2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
def batched_moe_align_block_size(
    max_tokens_per_batch: int,
    block_size: int,
    expert_num_tokens: torch.Tensor,
    sorted_ids: torch.Tensor,
    expert_ids: torch.Tensor,
    num_tokens_post_pad: torch.Tensor,
) -> None:
    torch.ops._moe_C.batched_moe_align_block_size(
        max_tokens_per_batch,
        block_size,
        expert_num_tokens,
        sorted_ids,
        expert_ids,
        num_tokens_post_pad,
    )


2203
2204
2205
2206
2207
2208
def moe_lora_align_block_size(
    topk_ids: torch.Tensor,
    token_lora_mapping: torch.Tensor,
    num_experts: int,
    block_size: int,
    max_loras: int,
2209
2210
    max_num_tokens_padded: int,
    max_num_m_blocks: int,
2211
2212
2213
    sorted_token_ids: torch.Tensor,
    experts_ids: torch.Tensor,
    num_tokens_post_pad: torch.Tensor,
2214
2215
    adapter_enabled: torch.Tensor,
    lora_ids: torch.Tensor,
gnovack's avatar
gnovack committed
2216
    expert_map: torch.Tensor | None = None,
2217
2218
2219
2220
2221
2222
2223
) -> None:
    torch.ops._moe_C.moe_lora_align_block_size(
        topk_ids,
        token_lora_mapping,
        num_experts,
        block_size,
        max_loras,
2224
2225
        max_num_tokens_padded,
        max_num_m_blocks,
2226
2227
2228
        sorted_token_ids,
        experts_ids,
        num_tokens_post_pad,
2229
2230
        adapter_enabled,
        lora_ids,
gnovack's avatar
gnovack committed
2231
        expert_map,
2232
    )
2233
2234


2235
2236
2237
2238
2239
def moe_wna16_gemm(
    input: torch.Tensor,
    output: torch.Tensor,
    b_qweight: torch.Tensor,
    b_scales: torch.Tensor,
2240
2241
    b_qzeros: torch.Tensor | None,
    topk_weights: torch.Tensor | None,
2242
2243
2244
2245
2246
2247
2248
2249
2250
    sorted_token_ids: torch.Tensor,
    experts_ids: torch.Tensor,
    num_tokens_post_pad: torch.Tensor,
    top_k: int,
    BLOCK_SIZE_M: int,
    BLOCK_SIZE_N: int,
    BLOCK_SIZE_K: int,
    bit: int,
) -> torch.Tensor:
2251
2252
    if not current_platform.is_cuda():
        raise NotImplementedError(
2253
2254
2255
2256
2257
2258
2259
2260
2261
2262
2263
2264
2265
2266
2267
2268
2269
2270
            "The optimized moe_wna16_gemm kernel is only available on CUDA platforms"
        )
    torch.ops._moe_C.moe_wna16_gemm(
        input,
        output,
        b_qweight,
        b_scales,
        b_qzeros,
        topk_weights,
        sorted_token_ids,
        experts_ids,
        num_tokens_post_pad,
        top_k,
        BLOCK_SIZE_M,
        BLOCK_SIZE_N,
        BLOCK_SIZE_K,
        bit,
    )
2271
2272


2273
2274
2275
2276
2277
def topk_softmax(
    topk_weights: torch.Tensor,
    topk_ids: torch.Tensor,
    token_expert_indices: torch.Tensor,
    gating_output: torch.Tensor,
2278
    renormalize: bool = False,
2279
2280
) -> None:
    torch.ops._moe_C.topk_softmax(
2281
        topk_weights, topk_ids, token_expert_indices, gating_output, renormalize
2282
    )
2283

2284

2285
2286
2287
2288
2289
2290
2291
def grouped_topk(
    scores: torch.Tensor,
    num_expert_group: int,
    topk_group: int,
    topk: int,
    renormalize: bool,
    routed_scaling_factor: float,
2292
2293
    bias: torch.Tensor,
    scoring_func: int = 0,
2294
):
2295
2296
    """
    Perform grouped top-k routing for mixture of experts.
2297

2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
    Args:
        scores: Raw inputs (logits if scoring_func=1, scores if scoring_func=0)
        num_expert_group: Number of expert groups
        topk_group: Number of groups to select
        topk: Number of experts to select per token
        renormalize: Whether to renormalize the output weights
        routed_scaling_factor: Scaling factor for routing weights
        bias: Bias tensor (e_score_correction_bias). Always fused in kernel.
        scoring_func: 0=none (no activation), 1=sigmoid
    """
2308
    if not current_platform.is_cuda():
2309
2310
2311
2312
2313
2314
2315
2316
2317
2318
        raise NotImplementedError(
            "The fused grouped_topk kernel is only available on CUDA platforms"
        )
    return torch.ops._moe_C.grouped_topk(
        scores,
        num_expert_group,
        topk_group,
        topk,
        renormalize,
        routed_scaling_factor,
2319
2320
        bias,
        scoring_func,
2321
2322
2323
2324
2325
    )


def moe_wna16_marlin_gemm(
    input: torch.Tensor,
2326
    output: torch.Tensor | None,
2327
    b_qweight: torch.Tensor,
2328
    b_bias: torch.Tensor | None,
2329
    b_scales: torch.Tensor,
2330
    a_scales: torch.Tensor | None,
2331
2332
2333
2334
    global_scale: torch.Tensor | None,
    b_qzeros: torch.Tensor | None,
    g_idx: torch.Tensor | None,
    perm: torch.Tensor | None,
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351
    workspace: torch.Tensor,
    sorted_token_ids: torch.Tensor,
    expert_ids: torch.Tensor,
    num_tokens_past_padded: torch.Tensor,
    topk_weights: torch.Tensor,
    moe_block_size: int,
    top_k: int,
    mul_topk_weights: bool,
    is_ep: bool,
    b_q_type: ScalarType,
    size_m: int,
    size_n: int,
    size_k: int,
    is_k_full: bool,
    use_atomic_add: bool,
    use_fp32_reduce: bool,
    is_zp_float: bool,
2352
2353
2354
    thread_k: int = -1,
    thread_n: int = -1,
    blocks_per_sm: int = -1,
2355
) -> torch.Tensor:
2356
    return torch.ops._moe_C.moe_wna16_marlin_gemm(
2357
2358
2359
2360
2361
        input,
        output,
        b_qweight,
        b_bias,
        b_scales,
2362
        a_scales,
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
        global_scale,
        b_qzeros,
        g_idx,
        perm,
        workspace,
        sorted_token_ids,
        expert_ids,
        num_tokens_past_padded,
        topk_weights,
        moe_block_size,
        top_k,
        mul_topk_weights,
        is_ep,
        b_q_type.id,
        size_m,
        size_n,
        size_k,
        is_k_full,
        use_atomic_add,
        use_fp32_reduce,
        is_zp_float,
2384
2385
2386
        thread_k,
        thread_n,
        blocks_per_sm,
2387
    )
2388
2389


2390
if hasattr(torch.ops, "_moe_C") and hasattr(torch.ops._moe_C, "marlin_gemm_moe"):
2391

2392
    @register_fake("_moe_C::marlin_gemm_moe")
2393
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403
2404
2405
2406
2407
2408
2409
2410
2411
2412
2413
2414
2415
    def marlin_gemm_moe_fake(
        a: torch.Tensor,
        b_q_weights: torch.Tensor,
        sorted_ids: torch.Tensor,
        topk_weights: torch.Tensor,
        topk_ids: torch.Tensor,
        b_scales: torch.Tensor,
        b_zero_points: torch.Tensor,
        g_idx: torch.Tensor,
        perm: torch.Tensor,
        workspace: torch.Tensor,
        b_q_type: ScalarType,
        size_m: torch.SymInt,
        size_n: torch.SymInt,
        size_k: torch.SymInt,
        is_k_full: bool,
        num_experts: int,
        topk: int,
        moe_block_size: int,
        replicate_input: bool,
        apply_weights: bool,
    ) -> torch.Tensor:
        return torch.empty((size_m, topk, size_n), dtype=a.dtype, device=a.device)
2416

2417
    @register_fake("_moe_C::moe_wna16_marlin_gemm")
2418
2419
    def moe_wna16_marlin_gemm_fake(
        input: torch.Tensor,
2420
        output: torch.Tensor | None,
2421
        b_qweight: torch.Tensor,
2422
        b_bias: torch.Tensor | None,
2423
        b_scales: torch.Tensor,
2424
2425
        a_scales: torch.Tensor | None,
        global_scale: torch.Tensor | None,
2426
2427
2428
        b_qzeros: torch.Tensor | None,
        g_idx: torch.Tensor | None,
        perm: torch.Tensor | None,
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
2443
2444
2445
        workspace: torch.Tensor,
        sorted_token_ids: torch.Tensor,
        expert_ids: torch.Tensor,
        num_tokens_past_padded: torch.Tensor,
        topk_weights: torch.Tensor,
        moe_block_size: int,
        top_k: int,
        mul_topk_weights: bool,
        is_ep: bool,
        b_q_type: ScalarType,
        size_m: int,
        size_n: int,
        size_k: int,
        is_k_full: bool,
        use_atomic_add: bool,
        use_fp32_reduce: bool,
        is_zp_float: bool,
2446
    ):
2447
2448
2449
        return torch.empty(
            (size_m * top_k, size_n), dtype=input.dtype, device=input.device
        )
2450

2451

2452
2453
2454
2455
2456
2457
2458
def reshape_and_cache(
    key: torch.Tensor,
    value: torch.Tensor,
    key_cache: torch.Tensor,
    value_cache: torch.Tensor,
    slot_mapping: torch.Tensor,
    kv_cache_dtype: str,
2459
2460
    k_scale: torch.Tensor,
    v_scale: torch.Tensor,
2461
) -> None:
2462
2463
2464
2465
2466
2467
2468
2469
2470
2471
    torch.ops._C_cache_ops.reshape_and_cache(
        key,
        value,
        key_cache,
        value_cache,
        slot_mapping,
        kv_cache_dtype,
        k_scale,
        v_scale,
    )
2472
2473


zhuwenwen's avatar
zhuwenwen committed
2474
2475
2476
2477
2478
2479
2480
2481
2482
2483
2484
2485
2486
2487
2488
def reshape_and_cache_cuda(
    key: torch.Tensor,
    value: torch.Tensor,
    key_cache: torch.Tensor,
    value_cache: torch.Tensor,
    slot_mapping: torch.Tensor,
    kv_cache_dtype: str,
    k_scale: torch.Tensor,
    v_scale: torch.Tensor,
) -> None:
    torch.ops._C_cache_ops.reshape_and_cache_cuda(key, value, key_cache, 
                                                  value_cache, slot_mapping, 
                                                  kv_cache_dtype, k_scale, v_scale)


2489
2490
2491
2492
2493
2494
2495
def reshape_and_cache_flash(
    key: torch.Tensor,
    value: torch.Tensor,
    key_cache: torch.Tensor,
    value_cache: torch.Tensor,
    slot_mapping: torch.Tensor,
    kv_cache_dtype: str,
2496
2497
    k_scale: torch.Tensor,
    v_scale: torch.Tensor,
2498
) -> None:
2499
2500
2501
2502
2503
2504
2505
2506
2507
2508
    torch.ops._C_cache_ops.reshape_and_cache_flash(
        key,
        value,
        key_cache,
        value_cache,
        slot_mapping,
        kv_cache_dtype,
        k_scale,
        v_scale,
    )
2509
2510


2511
2512
2513
2514
2515
2516
2517
2518
def concat_and_cache_mla(
    kv_c: torch.Tensor,
    k_pe: torch.Tensor,
    kv_cache: torch.Tensor,
    slot_mapping: torch.Tensor,
    kv_cache_dtype: str,
    scale: torch.Tensor,
) -> None:
2519
2520
2521
    torch.ops._C_cache_ops.concat_and_cache_mla(
        kv_c, k_pe, kv_cache, slot_mapping, kv_cache_dtype, scale
    )
2522
2523


2524
2525
2526
2527
2528
def copy_blocks(
    key_caches: list[torch.Tensor],
    value_caches: list[torch.Tensor],
    block_mapping: torch.Tensor,
) -> None:
2529
    torch.ops._C_cache_ops.copy_blocks(key_caches, value_caches, block_mapping)
2530
2531


2532
def copy_blocks_mla(kv_caches: list[torch.Tensor], block_mapping: torch.Tensor) -> None:
2533
2534
2535
    torch.ops._C_cache_ops.copy_blocks_mla(kv_caches, block_mapping)


2536
2537
2538
def swap_blocks(
    src: torch.Tensor, dst: torch.Tensor, block_mapping: torch.Tensor
) -> None:
2539
    torch.ops._C_cache_ops.swap_blocks(src, dst, block_mapping)
2540
2541


2542
2543
2544
def convert_fp8(
    output: torch.Tensor, input: torch.Tensor, scale: float = 1.0, kv_dtype: str = "fp8"
) -> None:
2545
2546
2547
    torch.ops._C_cache_ops.convert_fp8(output, input, scale, kv_dtype)


2548
def gather_and_maybe_dequant_cache(
2549
2550
2551
2552
    src_cache: torch.Tensor,
    dst: torch.Tensor,
    block_table: torch.Tensor,
    cu_seq_lens: torch.Tensor,
2553
2554
    token_to_seq: torch.Tensor,
    num_tokens: int,
2555
2556
    kv_cache_dtype: str,
    scale: torch.Tensor,
2557
    seq_starts: torch.Tensor | None = None,
2558
) -> None:
2559
    torch.ops._C_cache_ops.gather_and_maybe_dequant_cache(
2560
2561
2562
2563
        src_cache,
        dst,
        block_table,
        cu_seq_lens,
2564
2565
        token_to_seq,
        num_tokens,
2566
2567
2568
2569
        kv_cache_dtype,
        scale,
        seq_starts,
    )
2570
2571


2572
2573
2574
2575
2576
2577
def cp_gather_cache(
    src_cache: torch.Tensor,
    dst: torch.Tensor,
    block_table: torch.Tensor,
    cu_seq_lens: torch.Tensor,
    batch_size: int,
2578
    seq_starts: torch.Tensor | None = None,
2579
2580
2581
2582
) -> None:
    torch.ops._C_cache_ops.cp_gather_cache(
        src_cache, dst, block_table, cu_seq_lens, batch_size, seq_starts
    )
2583
2584


2585
2586
2587
2588
2589
2590
2591
2592
2593
2594
2595
2596
2597
2598
2599
2600
2601
2602
2603
2604
2605
2606
2607
def cp_gather_and_upconvert_fp8_kv_cache(
    src_cache: torch.Tensor,
    dst: torch.Tensor,
    block_table: torch.Tensor,
    seq_lens: torch.Tensor,
    workspace_starts: torch.Tensor,
    batch_size: int,
) -> None:
    """Gather and upconvert FP8 KV cache to BF16 workspace.

    Args:
        src_cache: FP8 KV cache [num_blocks, block_size, 656]
        dst: BF16 output workspace [total_tokens, 576]
        block_table: Block indices [num_reqs, max_blocks]
        seq_lens: Sequence lengths [num_reqs]
        workspace_starts: Workspace start offsets [num_reqs]
        batch_size: Number of requests
    """
    torch.ops._C_cache_ops.cp_gather_and_upconvert_fp8_kv_cache(
        src_cache, dst, block_table, seq_lens, workspace_starts, batch_size
    )


2608
2609
2610
2611
2612
2613
2614
2615
2616
2617
def indexer_k_quant_and_cache(
    k: torch.Tensor,
    kv_cache: torch.Tensor,
    slot_mapping: torch.Tensor,
    quant_block_size: int,
    kv_cache_dtype: str,
) -> None:
    torch.ops._C_cache_ops.indexer_k_quant_and_cache(
        k, kv_cache, slot_mapping, quant_block_size, kv_cache_dtype
    )
zhuwenwen's avatar
zhuwenwen committed
2618
2619
2620
2621
2622
2623
2624
    
def indexer_k_cache(k: torch.Tensor, kv_cache: torch.Tensor,
                    slot_mapping: torch.Tensor,
                    kv_cache_dtype: str) -> None:
    torch.ops._C_cache_ops.indexer_k_cache(
        k, kv_cache, slot_mapping, kv_cache_dtype
    )
2625
2626


2627
2628
2629
2630
2631
2632
2633
2634
2635
2636
def cp_gather_indexer_k_quant_cache(
    kv_cache: torch.Tensor,
    dst_k: torch.Tensor,
    dst_scale: torch.Tensor,
    block_table: torch.Tensor,
    cu_seq_lens: torch.Tensor,
) -> None:
    torch.ops._C_cache_ops.cp_gather_indexer_k_quant_cache(
        kv_cache, dst_k, dst_scale, block_table, cu_seq_lens
    )
2637
2638


2639
2640
2641
2642
2643
2644
2645
2646
2647
def indexer_k_quant_and_cache(k: torch.Tensor, kv_cache: torch.Tensor,
                              slot_mapping: torch.Tensor,
                              quant_block_size: int,
                              kv_cache_dtype: str) -> None:
    torch.ops._C_cache_ops.indexer_k_quant_and_cache(k, kv_cache, slot_mapping,
                                                     quant_block_size,
                                                     kv_cache_dtype)


2648
2649
2650
2651
2652
2653
2654
def get_device_attribute(attribute: int, device: int) -> int:
    return torch.ops._C_cuda_utils.get_device_attribute(attribute, device)


def get_max_shared_memory_per_block_device_attribute(device: int) -> int:
    # ruff: noqa: E501
    return torch.ops._C_cuda_utils.get_max_shared_memory_per_block_device_attribute(
2655
2656
        device
    )
2657
2658
2659


# custom ar
2660
2661
2662
2663
2664
2665
2666
2667
2668
def init_custom_ar(
    ipc_tensors: list[torch.Tensor],
    rank_data: torch.Tensor,
    rank: int,
    fully_connected: bool,
) -> int:
    return torch.ops._C_custom_ar.init_custom_ar(
        ipc_tensors, rank_data, rank, fully_connected
    )
2669

2670

2671
2672
2673
2674
2675
2676
2677
2678
def all_reduce(
    fa: int,
    inp: torch.Tensor,
    out: torch.Tensor,
    reg_buffer: int,
    reg_buffer_sz_bytes: int,
) -> None:
    torch.ops._C_custom_ar.all_reduce(fa, inp, out, reg_buffer, reg_buffer_sz_bytes)
2679

2680
2681
2682
2683
2684
2685
2686
2687
2688

def dispose(fa: int) -> None:
    torch.ops._C_custom_ar.dispose(fa)


def meta_size() -> int:
    return torch.ops._C_custom_ar.meta_size()


2689
def register_buffer(fa: int, ipc_tensors: list[int]) -> None:
2690
    return torch.ops._C_custom_ar.register_buffer(fa, ipc_tensors)
2691
2692


2693
def get_graph_buffer_ipc_meta(fa: int) -> tuple[list[int], list[int]]:
2694
2695
2696
    return torch.ops._C_custom_ar.get_graph_buffer_ipc_meta(fa)


2697
2698
2699
def register_graph_buffers(
    fa: int, handles: list[list[int]], offsets: list[list[int]]
) -> None:
2700
2701
    torch.ops._C_custom_ar.register_graph_buffers(fa, handles, offsets)

2702

zhuwenwen's avatar
zhuwenwen committed
2703
2704
2705
2706
2707
2708
2709
2710
2711
2712
2713
2714
def allocate_shared_buffer_and_handle(size: int) -> tuple[int, torch.Tensor]:
    return torch.ops._C_custom_ar.allocate_shared_buffer_and_handle(size)


def open_mem_handle(mem_handle: torch.Tensor):
    return torch.ops._C_custom_ar.open_mem_handle(mem_handle)


def free_shared_buffer(ptr: int) -> None:
    torch.ops._C_custom_ar.free_shared_buffer(ptr)


2715
2716
2717
def read_cache(
        keys: torch.Tensor,
        values: torch.Tensor,
zhuwenwen's avatar
zhuwenwen committed
2718
2719
        key_caches: list[torch.Tensor],
        value_caches: list[torch.Tensor],
2720
2721
2722
2723
2724
2725
2726
2727
2728
2729
        slot_mapping: torch.Tensor,
        kv_cache_dtype: str
) -> None:
    torch.ops._C_cache_ops.read_cache(keys, values, key_caches,
                                      value_caches, slot_mapping,
                                      kv_cache_dtype)

def write_cache_multi_layers(
        keys: torch.Tensor,
        values: torch.Tensor,
zhuwenwen's avatar
zhuwenwen committed
2730
2731
        key_caches: list[torch.Tensor],
        value_caches: list[torch.Tensor],
2732
2733
2734
2735
2736
2737
        slot_mapping: torch.Tensor,
        kv_cache_dtype: str
) -> None:
    torch.ops._C_cache_ops.write_cache_multi_layers(keys, values, key_caches,
                                                    value_caches, slot_mapping,
                                                    kv_cache_dtype)
zhuwenwen's avatar
zhuwenwen committed
2738

2739
# quick all reduce
2740
def init_custom_qr(rank: int, world_size: int, qr_max_size: int | None = None) -> int:
2741
2742
2743
2744
2745
2746
2747
    return torch.ops._C_custom_ar.init_custom_qr(rank, world_size, qr_max_size)


def qr_destroy(fa: int) -> None:
    torch.ops._C_custom_ar.qr_destroy(fa)


2748
2749
2750
2751
2752
2753
2754
2755
def qr_all_reduce(
    fa: int,
    inp: torch.Tensor,
    out: torch.Tensor,
    quant_level: int,
    cast_bf2half: bool = False,
) -> None:
    torch.ops._C_custom_ar.qr_all_reduce(fa, inp, out, quant_level, cast_bf2half)
2756
2757
2758
2759
2760
2761
2762
2763
2764
2765
2766
2767
2768


def qr_get_handle(fa: int) -> torch.Tensor:
    return torch.ops._C_custom_ar.qr_get_handle(fa)


def qr_open_handles(fa: int, handles: list[torch.Tensor]) -> None:
    return torch.ops._C_custom_ar.qr_open_handles(fa, handles)


def qr_max_size() -> int:
    return torch.ops._C_custom_ar.qr_max_size()

zhuwenwen's avatar
zhuwenwen committed
2769

2770
2771
2772
2773
def get_flash_mla_metadata(
    cache_seqlens: torch.Tensor,
    num_heads_per_head_k: int,
    num_heads_k: int,
2774
) -> tuple[torch.Tensor, torch.Tensor]:
2775
2776
2777
2778
2779
2780
2781
2782
2783
2784
    """
    Arguments:
        cache_seqlens: (batch_size), dtype torch.int32.
        num_heads_per_head_k: Equals to seq_len_q * num_heads_q // num_heads_k.
        num_heads_k: num_heads_k.

    Return:
        tile_scheduler_metadata: (num_sm_parts, TileSchedulerMetaDataSize), dtype torch.int32.
        num_splits: (batch_size + 1), dtype torch.int32.
    """
2785
2786
2787
    return torch.ops._C.get_flash_mla_metadata(
        cache_seqlens, num_heads_per_head_k, num_heads_k
    )
2788
2789
2790
2791
2792
2793
2794
2795
2796
2797


def flash_mla_with_kvcache(
    q: torch.Tensor,
    k_cache: torch.Tensor,
    block_table: torch.Tensor,
    cache_seqlens: torch.Tensor,
    head_dim_v: int,
    tile_scheduler_metadata: torch.Tensor,
    num_splits: torch.Tensor,
2798
    softmax_scale: float | None = None,
2799
    causal: bool = False,
2800
) -> tuple[torch.Tensor, torch.Tensor]:
2801
2802
2803
2804
2805
2806
2807
2808
2809
2810
2811
2812
2813
2814
2815
2816
2817
    """
    Arguments:
        q: (batch_size, seq_len_q, num_heads_q, head_dim).
        k_cache: (num_blocks, page_block_size, num_heads_k, head_dim).
        block_table: (batch_size, max_num_blocks_per_seq), torch.int32.
        cache_seqlens: (batch_size), torch.int32.
        head_dim_v: Head_dim of v.
        tile_scheduler_metadata: (num_sm_parts, TileSchedulerMetaDataSize), torch.int32, return by get_mla_metadata.
        num_splits: (batch_size + 1), torch.int32, return by get_mla_metadata.
        softmax_scale: float. The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim).
        causal: bool. Whether to apply causal attention mask.

    Return:
        out: (batch_size, seq_len_q, num_heads_q, head_dim_v).
        softmax_lse: (batch_size, num_heads_q, seq_len_q), torch.float32.
    """
    if softmax_scale is None:
2818
        softmax_scale = q.shape[-1] ** (-0.5)
2819
2820
2821
2822
2823
2824
2825
2826
2827
2828
2829
2830
2831
    out, softmax_lse = torch.ops._C.flash_mla_fwd_kvcache(
        q,
        k_cache,
        None,
        head_dim_v,
        cache_seqlens,
        block_table,
        softmax_scale,
        causal,
        tile_scheduler_metadata,
        num_splits,
    )
    return out, softmax_lse
2832
2833


2834
2835
2836
2837
2838
2839
2840
2841
2842
2843
2844
2845
2846
2847
2848
2849
2850
2851
2852
2853
2854
2855
2856
2857
2858
2859
2860
2861
2862
2863
2864
2865
2866
2867
2868
2869
2870
2871
2872
2873
2874
2875
2876
2877
2878
def moe_fused_gate(
    input_tensor,
    bias,
    num_expert_group,
    topk_group,
    topk,
    n_share_experts_fusion=0,
    routed_scaling_factor=0,
):
    # This fused kernel function is used to select topk expert in a hierarchical 2-layer fashion
    # it split group of expert into num_expert_group, and use top2 expert weight sum in each group
    # as the group weight to select exerpt groups and then select topk experts within the selected groups
    # the #experts is decided by the input tensor shape and we currently only support power of 2 #experts
    # and #experts should be divisible by num_expert_group. #expert/num_expert_group <= 32 is limitted for now.
    # for non-supported case, we suggestion to use the biased_grouped_topk func in sglang.srt.layers.moe.topk
    # n_share_experts_fusion: if > 0, the last expert will be replaced with a round-robin shared expert
    # routed_scaling_factor: if > 0, the last expert will be scaled by this factor
    return torch.ops._moe_C.moe_fused_gate(
        input_tensor,
        bias,
        num_expert_group,
        topk_group,
        topk,
        n_share_experts_fusion,
        routed_scaling_factor,
    )

if hasattr(torch.ops._moe_C, "moe_fused_gate"):

    @register_fake("_moe_C::moe_fused_gate")
    def moe_fused_gate_fake(
        input_tensor: torch.Tensor,
        bias: torch.Tensor,
        num_expert_group: int,
        topk_group: int,
        topk: int,
        n_share_experts_fusion: int,
        routed_scaling_factor: int,
    ):
        return torch.empty((input_tensor.size(0), topk),
                           dtype=input_tensor.dtype,
                           device=input_tensor.device), \
                    torch.empty((input_tensor.size(0), topk),
                           dtype=input_tensor.dtype,
                           device=input_tensor.device)
王敏's avatar
王敏 committed
2879

2880

2881
2882
2883
2884
2885
2886
2887
2888
2889
2890
2891
2892
2893
2894
2895
2896
2897
2898
2899
2900
2901
2902
2903
2904
def sm100_cutlass_mla_decode(
    out: torch.Tensor,
    lse: torch.Tensor,
    q_nope: torch.Tensor,
    q_pe: torch.Tensor,
    kv_c_and_k_pe_cache: torch.Tensor,
    seq_lens: torch.Tensor,
    page_table: torch.Tensor,
    workspace: torch.Tensor,
    scale: float,
    num_kv_splits: int,
) -> torch.Tensor:
    torch.ops._C.sm100_cutlass_mla_decode(
        out,
        lse,
        q_nope,
        q_pe,
        kv_c_and_k_pe_cache,
        seq_lens,
        page_table,
        workspace,
        scale,
        num_kv_splits,
    )
2905
2906
2907
    return out


2908
2909
2910
def sm100_cutlass_mla_get_workspace_size(
    max_seq_len: int, num_batches: int, sm_count: int, num_kv_splits: int
) -> int:
2911
    return torch.ops._C.sm100_cutlass_mla_get_workspace_size(
2912
2913
        max_seq_len, num_batches, sm_count, num_kv_splits
    )
2914
2915


2916
2917
2918
if hasattr(torch.ops._C, "weight_packed_linear"):

    @register_fake("_C::weight_packed_linear")
2919
2920
2921
    def weight_packed_linear_fake(
        mat1: torch.Tensor,
        mat2: torch.Tensor,
2922
        bias: torch.Tensor | None,
2923
2924
2925
2926
2927
        is_vnni: bool,
    ) -> torch.Tensor:
        return torch.empty(
            (mat1.size(0), mat2.size(0)), dtype=mat1.dtype, device=mat2.device
        )
2928
2929
2930
2931
2932
2933
2934
2935
2936
2937
2938
2939
2940
2941


if hasattr(torch.ops._C, "fused_experts_cpu"):

    @register_fake("_C::fused_experts_cpu")
    def fused_experts_cpu_fake(
        hidden_states: torch.Tensor,
        w1: torch.Tensor,
        w2: torch.Tensor,
        topk_weights: torch.Tensor,
        topk_ids: torch.Tensor,
        inplace: bool,
        use_int8_w8a8: bool,
        use_fp8_w8a16: bool,
2942
2943
2944
2945
2946
        w1_scale: torch.Tensor | None,
        w2_scale: torch.Tensor | None,
        block_size: list[int] | None,
        a1_scale: torch.Tensor | None,
        a2_scale: torch.Tensor | None,
2947
2948
2949
2950
2951
2952
2953
2954
2955
2956
2957
2958
        is_vnni: bool,
    ) -> torch.Tensor:
        return torch.empty_like(hidden_states)


if hasattr(torch.ops._C, "int8_scaled_mm_with_quant"):

    @register_fake("_C::int8_scaled_mm_with_quant")
    def int8_scaled_mm_with_quant_fake(
        mat1: torch.Tensor,
        mat2: torch.Tensor,
        scales2: torch.Tensor,
2959
        bias: torch.Tensor | None,
2960
2961
2962
2963
2964
2965
        out_dtype: torch.dtype,
        is_vnni: bool,
    ) -> torch.Tensor:
        M = mat1.size(0)
        N = mat2.size(0)
        return torch.empty((M, N), dtype=out_dtype)
2966
    
2967
2968
2969

class CPUDNNLGEMMHandler:
    def __init__(self) -> None:
2970
        self.handler: int | None = None
2971
2972
2973
2974
2975
2976
2977
2978
        self.n = -1
        self.k = -1

    def __del__(self):
        if self.handler is not None:
            torch.ops._C.release_dnnl_matmul_handler(self.handler)


2979
_supports_onednn = bool(hasattr(torch.ops._C, "create_onednn_mm_handler"))
2980
2981


2982
2983
def is_onednn_acl_supported():
    return torch.ops._C.is_onednn_acl_supported()
2984
2985
2986
2987
2988
2989
2990
2991
2992


def create_onednn_mm(
    weight: torch.Tensor,  # [K, N]
    primitive_cache_size: int = 128,
) -> CPUDNNLGEMMHandler:
    handler = CPUDNNLGEMMHandler()
    handler.k, handler.n = weight.size()
    handler.handler = torch.ops._C.create_onednn_mm_handler(
2993
2994
        weight, primitive_cache_size
    )
2995
2996
2997
2998
2999
3000
    return handler


def onednn_mm(
    dnnl_handler: CPUDNNLGEMMHandler,
    x: torch.Tensor,
3001
    bias: torch.Tensor | None,
3002
3003
) -> torch.Tensor:
    output = torch.empty((*x.shape[0:-1], dnnl_handler.n), dtype=x.dtype)
3004
3005
3006
    torch.ops._C.onednn_mm(
        output, x.reshape(-1, dnnl_handler.k), bias, dnnl_handler.handler
    )
3007
3008
3009
3010

    return output


3011
3012
3013
3014
3015
3016
3017
3018
3019
3020
3021
def create_onednn_scaled_mm(
    weight: torch.Tensor,  # [K, N]
    weight_scales: torch.Tensor,
    output_type: torch.dtype,
    dynamic_quant: bool,
    use_azp: bool,
    primitive_cache_size: int = 128,
) -> CPUDNNLGEMMHandler:
    handler = CPUDNNLGEMMHandler()
    handler.k, handler.n = weight.size()
    handler.handler = torch.ops._C.create_onednn_scaled_mm_handler(
3022
3023
        weight, weight_scales, output_type, dynamic_quant, use_azp, primitive_cache_size
    )
3024
3025
3026
    return handler


3027
3028
def onednn_scaled_int8_quant(
    input: torch.Tensor,
3029
3030
    scale: torch.Tensor | None = None,
    azp: torch.Tensor | None = None,
3031
3032
    symmetric: bool = True,
):
3033
3034
3035
3036
3037
3038
3039
3040
3041
3042
3043
3044
    """
    Quantize the input tensor to int8 and return the quantized tensor and scale, and maybe azp.

    Args:
        input: The input tensor to be quantized to int8.
        scale: Optional scaling factor for the int8 quantization.
            When not provided, we invoke dynamic-per-token quantization.
        azp: Optional zero-point for the int8 quantization.
            Must be provided for asymmetric quantization if `scale` is provided.
        symmetric: Whether to use symmetric quantization (scale only, azp ignored).

    Returns:
3045
      tuple[torch.Tensor, torch.Tensor, torch.Tensor | None] : Output int8 tensor, scales, and optionally azp.
3046
3047
3048
3049
3050
3051
    """
    output = torch.empty_like(input, dtype=torch.int8)
    token_num = input.numel() // input.shape[-1]
    input = input.view((token_num, input.shape[-1]))
    if scale is not None:
        # static-per-tensor quantization.
3052
3053
3054
        assert symmetric == (azp is None), (
            "azp must only be provided for asymmetric quantization."
        )
3055
3056
3057
3058
        torch.ops._C.static_scaled_int8_quant(output, input, scale, azp)
        return output, scale, azp

    # dynamic-per-token quantization.
3059
3060
3061
    input_scales = torch.empty((token_num, 1), device=input.device, dtype=torch.float32)
    input_azp = None if symmetric else torch.empty_like(input_scales, dtype=torch.int32)
    torch.ops._C.dynamic_scaled_int8_quant(output, input, input_scales, input_azp)
3062
3063
3064
3065
3066
3067
3068
    return output, input_scales, input_azp


def onednn_scaled_mm(
    dnnl_handler: CPUDNNLGEMMHandler,
    x: torch.Tensor,
    output: torch.Tensor,
3069
3070
3071
3072
    input_scale: torch.Tensor | None,
    input_zp: torch.Tensor | None,
    input_zp_adj: torch.Tensor | None,
    bias: torch.Tensor | None,
3073
) -> torch.Tensor:
3074
3075
3076
    torch.ops._C.onednn_scaled_mm(
        output, x, input_scale, input_zp, input_zp_adj, bias, dnnl_handler.handler
    )
3077
3078

    return output
3079
3080


3081
3082
3083
3084
3085
3086
3087
3088
3089
3090
3091
3092
def cpu_attn_get_scheduler_metadata(
    num_reqs: int,
    num_heads: int,
    num_kv_heads: int,
    head_dim: int,
    seq_lens: torch.Tensor,
    dtype: torch.dtype,
    query_start_loc: torch.Tensor,
    causal: bool,
    sliding_window_size: int,
    isa: str,
    enable_kv_split: bool,
3093
) -> torch.Tensor:
3094
3095
3096
3097
3098
3099
3100
3101
3102
3103
3104
3105
3106
3107
3108
3109
3110
3111
3112
3113
3114
3115
3116
3117
3118
3119
3120
3121
3122
3123
3124
3125
3126
3127
3128
3129
3130
3131
3132
3133
3134
3135
3136
3137
3138
3139
3140
3141
3142
3143
3144
3145
3146
3147
3148
3149
3150
3151
3152
3153
3154
3155
3156
3157
3158
3159
3160
3161
    sheduler_metadata = torch.ops._C.get_scheduler_metadata(
        num_reqs,
        num_heads,
        num_kv_heads,
        head_dim,
        seq_lens,
        dtype,
        query_start_loc,
        causal,
        sliding_window_size,
        isa,
        enable_kv_split,
    )
    return sheduler_metadata


def cpu_attn_reshape_and_cache(
    key: torch.Tensor,
    value: torch.Tensor,
    key_cache: torch.Tensor,
    value_cache: torch.Tensor,
    slot_mapping: torch.Tensor,
    isa: str,
) -> None:
    torch.ops._C.cpu_attn_reshape_and_cache(
        key,
        value,
        key_cache,
        value_cache,
        slot_mapping,
        isa,
    )


def cpu_attention_with_kv_cache(
    query: torch.Tensor,
    key_cache: torch.Tensor,
    value_cache: torch.Tensor,
    output: torch.Tensor,
    query_start_loc: torch.Tensor,
    seq_lens: torch.Tensor,
    scale: float,
    causal: bool,
    alibi_slopes: torch.Tensor | None,
    sliding_window: tuple[int, int],
    block_table: torch.Tensor,
    softcap: float,
    scheduler_metadata: torch.Tensor,
    s_aux: torch.Tensor | None,
) -> None:
    torch.ops._C.cpu_attention_with_kv_cache(
        query,
        key_cache,
        value_cache,
        output,
        query_start_loc,
        seq_lens,
        scale,
        causal,
        alibi_slopes,
        sliding_window[0],
        sliding_window[1],
        block_table,
        softcap,
        scheduler_metadata,
        s_aux,
    )

3162

Li, Jiang's avatar
Li, Jiang committed
3163
3164
3165
3166
3167
3168
3169
3170
3171
3172
3173
3174
3175
3176
3177
3178
3179
3180
3181
3182
3183
3184
def cpu_gemm_wna16(
    input: torch.Tensor,
    q_weight: torch.Tensor,
    scales: torch.Tensor,
    zeros: torch.Tensor | None,
    g_idx: torch.Tensor | None,
    bias: torch.Tensor | None,
    pack_factor: int,
    isa_hint: str,
) -> torch.Tensor:
    output = torch.empty((input.size(0), scales.size(1)), dtype=input.dtype)
    torch.ops._C.cpu_gemm_wna16(
        input,
        q_weight,
        output,
        scales,
        zeros,
        g_idx,
        bias,
        pack_factor,
        isa_hint,
    )
3185
    return output
3186

3187

3188
3189
3190
3191
3192
3193
3194
3195
3196
3197
3198
3199
3200
3201
3202
3203
3204
3205
3206
3207
3208
3209
3210
3211
3212
3213
3214
3215
3216
3217
3218
3219
3220
3221
3222
3223
3224
3225
3226
3227
3228
3229
3230
3231
3232
3233
3234
3235
3236
3237
3238
3239
3240
3241
3242
3243
3244
3245
3246
3247
3248
3249
3250
3251
3252
3253
3254
3255
3256
3257
3258
3259
3260
3261
3262
3263
3264
3265
3266
3267
3268
3269
3270
3271
3272
3273
3274
3275
3276
3277
3278
3279
3280
3281
3282
3283
3284
3285
3286
3287
3288
3289
3290
3291
3292
3293
3294
3295
3296
3297
3298
3299
3300
3301
3302
3303
3304
3305
3306
3307
3308
3309
3310
3311
3312
3313
3314
3315
3316
3317
3318
3319
3320
3321
3322
3323
3324
3325
if hasattr(torch.ops._qutlass_C, "matmul_mxf4_bf16_tn"):

    @register_fake("_qutlass_C::matmul_mxf4_bf16_tn")
    def _fake_matmul_mxf4_bf16_tn(
        a: torch.Tensor,
        b: torch.Tensor,
        a_sf: torch.Tensor,
        b_sf: torch.Tensor,
        alpha: torch.Tensor,
    ):
        return a.new_empty(*a.shape[:-1], b.shape[0], dtype=torch.bfloat16)


def matmul_mxf4_bf16_tn(
    a: torch.Tensor,
    b: torch.Tensor,
    a_sf: torch.Tensor,
    b_sf: torch.Tensor,
    alpha: torch.Tensor,
) -> torch.Tensor:
    return torch.ops._qutlass_C.matmul_mxf4_bf16_tn(a, b, a_sf, b_sf, alpha)


if hasattr(torch.ops._qutlass_C, "matmul_ada_mxf4_bf16_tn"):

    @register_fake("_qutlass_C::matmul_ada_mxf4_bf16_tn")
    def _fake_matmul_ada_mxf4_bf16_tn(
        a: torch.Tensor,
        b: torch.Tensor,
        a_sf: torch.Tensor,
        b_sf: torch.Tensor,
        alpha: torch.Tensor,
    ):
        return a.new_empty(*a.shape[:-1], b.shape[0], dtype=torch.bfloat16)


def matmul_ada_mxf4_bf16_tn(
    a: torch.Tensor,
    b: torch.Tensor,
    a_sf: torch.Tensor,
    b_sf: torch.Tensor,
    alpha: torch.Tensor,
) -> torch.Tensor:
    return torch.ops._qutlass_C.matmul_ada_mxf4_bf16_tn(a, b, a_sf, b_sf, alpha)


def ceil_div(a, b):
    return (a + b - 1) // b


if hasattr(torch.ops._qutlass_C, "fusedQuantizeMxQuest"):

    @register_fake("_qutlass_C::fusedQuantizeMxQuest")
    def _fake_fused_quantize_mx_quest(
        a: torch.Tensor, b: torch.Tensor, xh_e2m1: torch.Tensor, xh_e8m0: torch.Tensor
    ):
        return xh_e2m1, xh_e8m0


if hasattr(torch.ops._qutlass_C, "fusedQuantizeMxAbsMax"):

    @register_fake("_qutlass_C::fusedQuantizeMxAbsMax")
    def _fake_fused_quantize_mx_absmax(
        a: torch.Tensor, b: torch.Tensor, xh_e2m1: torch.Tensor, xh_e8m0: torch.Tensor
    ):
        return xh_e2m1, xh_e8m0


def fusedQuantizeMx(
    a: torch.Tensor, b: torch.Tensor, *, method: Literal["quest", "abs_max"] = "quest"
) -> tuple[torch.Tensor, torch.Tensor]:
    if a.dim() == 0:
        raise ValueError("`a` must have at least 1 dimension.")
    if a.size(-1) % 32 != 0:
        raise ValueError(f"last dim of `a` must be divisible by 32, got {a.size(-1)}.")
    if b.device != a.device:
        raise ValueError("`a` and `b` must be on the same device.")

    xh_e2m1 = torch.empty(
        *a.shape[:-1], a.size(-1) // 2, dtype=torch.uint8, device=a.device
    )

    rows, cols = a.numel() // a.size(-1), a.size(-1) // 32
    n_row_blocks = ceil_div(rows, 128)
    n_col_blocks = ceil_div(cols, 4)
    padded_rows = n_row_blocks * 128
    padded_cols = n_col_blocks * 4

    xh_e8m0 = torch.empty(
        padded_rows, padded_cols, dtype=torch.float8_e8m0fnu, device=a.device
    )

    if not hasattr(torch.ops, "_qutlass_C"):
        raise RuntimeError(
            "The `_qutlass_C` extension is not loaded. "
            "Make sure your custom op library is imported before calling fusedQuantizeMx."
        )

    if method == "quest":
        return torch.ops._qutlass_C.fusedQuantizeMxQuest(a, b, xh_e2m1, xh_e8m0)
    elif method == "abs_max":
        return torch.ops._qutlass_C.fusedQuantizeMxAbsMax(a, b, xh_e2m1, xh_e8m0)
    else:
        raise ValueError(f"invalid method {method!r}, must be 'quest' or 'abs_max'")


if hasattr(torch.ops._qutlass_C, "fusedQuantizeNv"):

    @register_fake("_qutlass_C::fusedQuantizeNv")
    def _fake_fused_quantize_nv(
        a: torch.Tensor,
        b: torch.Tensor,
        xh_e2m1: torch.Tensor,
        xh_e4m3: torch.Tensor,
        global_scale: torch.Tensor,
    ):
        return xh_e2m1, xh_e4m3


def fusedQuantizeNv(
    a: torch.Tensor, b: torch.Tensor, global_scale: torch.Tensor
) -> tuple[torch.Tensor, torch.Tensor]:
    xh_e2m1 = torch.empty(
        *a.shape[:-1], a.size(-1) // 2, dtype=torch.uint8, device=a.device
    )

    rows, cols = a.numel() // a.size(-1), a.size(-1) // 16
    n_row_blocks = ceil_div(rows, 128)
    n_col_blocks = ceil_div(cols, 4)
    padded_rows = n_row_blocks * 128
    padded_cols = n_col_blocks * 4
    xh_e4m3 = torch.empty(
        padded_rows, padded_cols, dtype=torch.float8_e4m3fn, device=a.device
    )

    return torch.ops._qutlass_C.fusedQuantizeNv(a, b, xh_e2m1, xh_e4m3, global_scale)


3326
3327
3328
3329
3330
3331
3332
3333
def hadacore_transform(x: torch.Tensor, inplace: bool = True) -> torch.Tensor:
    """
    Perform Hadamard transforms using [Hadacore](https://arxiv.org/abs/2412.08832)
    kernels. Note that these kernels exploit the recursive properties of
    Sylvester Hadamards, and therefore do not require transform weight data

    Note that sylvester hadamard transforms are also symmetric, which means that
    this function is also applies the (transpose <=> inverse) transform.
3334

3335
3336
3337
3338
3339
3340
3341
3342
3343
3344
    :param x: value to be transformed inplace
    :param inplace: modify value in place
    :return: value after transformation
    """
    return torch.ops._C.hadacore_transform(x, inplace)


if hasattr(torch.ops._C, "hadacore_transform"):

    @register_fake("_C::hadacore_transform")
3345
    def _hadacore_transform_fake(x: torch.Tensor, inplace: bool) -> torch.Tensor:
3346
        return torch.empty_like(x) if not inplace else x
3347
3348


3349
3350
3351
3352
3353
direct_register_custom_op(
    op_name="awq_gemm",
    op_func=awq_gemm,
    mutates_args=[],
    fake_impl=awq_gemm_fake,
3354
)