_custom_ops.py 83.4 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
7

import torch

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
14

logger = init_logger(__name__)

15
current_platform.import_kernels()
16

17
if TYPE_CHECKING:
18
19
20
21
22
23
24
25
26

    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

27
28
29
30
31
32
33
34
35
36

# page attention ops
def paged_attention_v1(
    out: torch.Tensor,
    query: torch.Tensor,
    key_cache: torch.Tensor,
    value_cache: torch.Tensor,
    num_kv_heads: int,
    scale: float,
    block_tables: torch.Tensor,
37
    seq_lens: torch.Tensor,
38
    block_size: int,
39
    max_seq_len: int,
40
    alibi_slopes: torch.Tensor | None,
41
    kv_cache_dtype: str,
42
43
    k_scale: torch.Tensor,
    v_scale: torch.Tensor,
44
45
46
47
48
    tp_rank: int = 0,
    blocksparse_local_blocks: int = 0,
    blocksparse_vert_stride: int = 0,
    blocksparse_block_size: int = 64,
    blocksparse_head_sliding_step: int = 0,
49
) -> None:
50
    torch.ops._C.paged_attention_v1(
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
        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,
    )
71
72
73
74
75
76
77
78
79
80
81
82
83


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,
84
    seq_lens: torch.Tensor,
85
    block_size: int,
86
    max_seq_len: int,
87
    alibi_slopes: torch.Tensor | None,
88
    kv_cache_dtype: str,
89
90
    k_scale: torch.Tensor,
    v_scale: torch.Tensor,
91
92
93
94
95
    tp_rank: int = 0,
    blocksparse_local_blocks: int = 0,
    blocksparse_vert_stride: int = 0,
    blocksparse_block_size: int = 64,
    blocksparse_head_sliding_step: int = 0,
96
) -> None:
97
    torch.ops._C.paged_attention_v2(
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
        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,
    )
121
122


123
124
125
126
127
128
129
130
131
132
133
134
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,
135
    query_start_loc: torch.Tensor | None,
136
137
    block_size: int,
    max_seq_len: int,
138
    alibi_slopes: torch.Tensor | None,
139
    kv_cache_dtype: str,
140
141
    k_scale: torch.Tensor,
    v_scale: torch.Tensor,
142
    fp8_out_scale: torch.Tensor | None = None,
xiao-llm's avatar
xiao-llm committed
143
    mfma_type: str = "fp8" if envs.VLLM_ROCM_FP8_MFMA_PAGE_ATTN else "f16",
144
) -> None:
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
    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,
    )
167
168


Thien Tran's avatar
Thien Tran committed
169
170
171
172
173
174
175
176
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:
177
178
179
    torch.ops._C_cpu.mla_decode_kvcache(
        out, query, kv_cache, scale, block_tables, seq_lens
    )
Thien Tran's avatar
Thien Tran committed
180
181


182
# merge attn states ops
183
184
185
186
187
188
def merge_attn_states(
    output: torch.Tensor,
    prefix_output: torch.Tensor,
    prefix_lse: torch.Tensor,
    suffix_output: torch.Tensor,
    suffix_lse: torch.Tensor,
189
    output_lse: torch.Tensor | None = None,
190
191
192
193
) -> None:
    torch.ops._C.merge_attn_states(
        output, output_lse, prefix_output, prefix_lse, suffix_output, suffix_lse
    )
194
195


196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
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

212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
    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,
    )
234
235

    torch.ops._C.convert_vertical_slash_indexes(
236
237
238
239
240
241
242
243
244
245
246
247
248
        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,
    )
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
    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

271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
    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,
    )
293
294

    torch.ops._C.convert_vertical_slash_indexes_mergehead(
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
        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,
    )
310
311
312
    return block_count, block_offset, column_count, column_index


313
314
315
316
# pos encoding ops
def rotary_embedding(
    positions: torch.Tensor,
    query: torch.Tensor,
317
    key: torch.Tensor | None,
318
319
320
321
    head_size: int,
    cos_sin_cache: torch.Tensor,
    is_neox: bool,
) -> None:
322
323
324
    torch.ops._C.rotary_embedding(
        positions, query, key, head_size, cos_sin_cache, is_neox
    )
325
326
327


# layer norm ops
328
329
330
def rms_norm(
    out: torch.Tensor, input: torch.Tensor, weight: torch.Tensor, epsilon: float
) -> None:
331
    torch.ops._C.rms_norm(out, input, weight, epsilon)
332
333


334
335
336
def fused_add_rms_norm(
    input: torch.Tensor, residual: torch.Tensor, weight: torch.Tensor, epsilon: float
) -> None:
337
    torch.ops._C.fused_add_rms_norm(input, residual, weight, epsilon)
338
339


340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
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,
    )


368
def apply_repetition_penalties_torch(
369
370
371
372
373
    logits: torch.Tensor,
    prompt_mask: torch.Tensor,
    output_mask: torch.Tensor,
    repetition_penalties: torch.Tensor,
) -> None:
374
    repetition_penalties = repetition_penalties.unsqueeze(dim=1).repeat(
375
376
        1, logits.size(1)
    )
377
    # If token appears in prompt or output, apply, otherwise use 1.0 for no-op.
378
    penalties = torch.where(prompt_mask | output_mask, repetition_penalties, 1.0)
379
380
381
382
383
384
    # 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(
385
386
387
388
389
390
391
392
    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
    )
393
394


395
396
397
398
399
400
def apply_repetition_penalties(
    logits: torch.Tensor,
    prompt_mask: torch.Tensor,
    output_mask: torch.Tensor,
    repetition_penalties: torch.Tensor,
) -> None:
401
402
403
404
405
406
407
408
    """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, ).
    """
409
    if logits.is_cuda and logits.is_contiguous():
410
411
412
        apply_repetition_penalties_cuda(
            logits, prompt_mask, output_mask, repetition_penalties
        )
413
    else:
414
415
416
        apply_repetition_penalties_torch(
            logits, prompt_mask, output_mask, repetition_penalties
        )
417
418


419
420
421
422
423
424
# fused quant layer norm ops
def rms_norm_dynamic_per_token_quant(
    input: torch.Tensor,
    weight: torch.Tensor,
    epsilon: float,
    quant_dtype: torch.dtype,
425
426
    scale_ub: torch.Tensor | None = None,
    residual: torch.Tensor | None = None,
427
) -> tuple[torch.Tensor, torch.Tensor]:
428
    output = torch.empty_like(input, dtype=quant_dtype)
429
430
431
    scales = torch.empty(
        (input.numel() // input.shape[-1], 1), device=input.device, dtype=torch.float32
    )
432

433
434
435
    torch.ops._C.rms_norm_dynamic_per_token_quant(
        output, input, weight, scales, epsilon, scale_ub, residual
    )
436
437
438
    return output, scales


439
440
# quantization ops
# awq
441
442
443
444
445
446
447
448
def awq_dequantize(
    qweight: torch.Tensor,
    scales: torch.Tensor,
    zeros: torch.Tensor,
    split_k_iters: int,
    thx: int,
    thy: int,
) -> torch.Tensor:
449
450
    if envs.VLLM_USE_TRITON_AWQ:
        from vllm.model_executor.layers.quantization.awq_triton import (
451
452
453
            awq_dequantize_triton,
        )

454
        return awq_dequantize_triton(qweight, scales, zeros)
455
    return torch.ops._C.awq_dequantize(qweight, scales, zeros, split_k_iters, thx, thy)
456
457


458
459
460
461
462
463
464
def awq_gemm(
    input: torch.Tensor,
    qweight: torch.Tensor,
    qzeros: torch.Tensor,
    scales: torch.Tensor,
    split_k_iters: int,
) -> torch.Tensor:
465
    if envs.VLLM_USE_TRITON_AWQ:
466
467
        from vllm.model_executor.layers.quantization.awq_triton import awq_gemm_triton

468
        return awq_gemm_triton(input, qweight, qzeros, scales, split_k_iters)
469
    return torch.ops._C.awq_gemm(input, qweight, qzeros, scales, split_k_iters)
470
471
472


# gptq
473
474
475
476
477
478
479
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,
480
    use_v2_format: bool,
481
482
483
    bit: int,
) -> torch.Tensor:
    return torch.ops._C.gptq_gemm(
484
485
486
487
488
489
490
491
        a,
        b_q_weight,
        b_gptq_qzeros,
        b_gptq_scales,
        b_g_idx,
        use_exllama,
        use_v2_format,
        bit,
492
    )
493
494


495
if hasattr(torch.ops._C, "gptq_gemm"):
496

497
    @register_fake("_C::gptq_gemm")
498
499
500
501
502
503
504
    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,
505
        use_v2_format: bool,
506
507
508
509
510
        bit: int,
    ) -> torch.Tensor:
        return torch.empty(
            (a.size(0), b_q_weight.size(1)), dtype=a.dtype, device=a.device
        )
511
512


513
def gptq_shuffle(q_weight: torch.Tensor, q_perm: torch.Tensor, bit: int) -> None:
514
    torch.ops._C.gptq_shuffle(q_weight, q_perm, bit)
515
516


517
# marlin_24
518
519
520
521
522
523
524
525
526
527
528
529
530
531
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
    )
532
533


534
if hasattr(torch.ops._C, "gptq_marlin_24_gemm"):
535

536
    @register_fake("_C::gptq_marlin_24_gemm")
537
538
539
540
541
542
543
544
545
546
547
    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:
548
549
        return torch.empty((size_m, size_n), device=a.device, dtype=a.dtype)

550
    @register_fake("_C::gptq_marlin_gemm")
551
552
    def _gptq_marlin_gemm_fake(
        a: torch.Tensor,
553
        c: torch.Tensor | None,
554
        b_q_weight: torch.Tensor,
555
        b_bias: torch.Tensor | None,
556
        b_scales: torch.Tensor,
557
558
559
560
        global_scale: torch.Tensor | None,
        b_zeros: torch.Tensor | None,
        g_idx: torch.Tensor | None,
        perm: torch.Tensor | None,
561
562
563
564
565
566
567
568
569
570
        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:
571
572
        return torch.empty((size_m, size_n), device=a.device, dtype=a.dtype)

573
    @register_fake("_C::awq_dequantize")
574
575
576
577
578
579
580
581
    def _awq_dequantize_fake(
        qweight: torch.Tensor,
        scales: torch.Tensor,
        zeros: torch.Tensor,
        split_k_iters: torch.SymInt,
        thx: int,
        thy: int,
    ) -> torch.Tensor:
582
583
584
        in_c = qweight.size(0)
        qout_c = qweight.size(1)
        out_c = qout_c * 8
585
        return torch.empty((in_c, out_c), dtype=scales.dtype, device=scales.device)
586

587
    @register_fake("_C::awq_gemm")
588
589
590
591
592
593
594
    def _awq_gemm_fake(
        input: torch.Tensor,
        qweight: torch.Tensor,
        qzeros: torch.Tensor,
        scales: torch.Tensor,
        split_k_iters: torch.SymInt,
    ) -> torch.Tensor:
595
        num_in_feats = input.size(0)
596
597
598
599
600
        return torch.empty(
            (split_k_iters, num_in_feats, qweight.size(1) * 8),
            dtype=input.dtype,
            device=input.device,
        ).sum(0)
601

602
603
    @register_fake("_C::machete_mm")
    def machete_mm_fake(
604
        a: torch.Tensor,
605
        # b_q Should be the tensor returned by machete_prepack_B
606
        b_q: torch.Tensor,
607
        b_type: ScalarType,
608
609
610
611
612
613
614
        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,
615
616
617
618
619
    ) -> torch.Tensor:
        m = a.size(0)
        n = b_q.size(1)
        return torch.empty((m, n), device=a.device, dtype=a.dtype)

620
    @register_fake("_C::machete_prepack_B")
621
    def machete_prepack_B_fake(
622
623
624
        b_q_weight: torch.Tensor,
        a_type: torch.dtype,
        b_type: ScalarType,
625
        group_scales_type: torch.dtype | None,
626
627
    ) -> torch.Tensor:
        return torch.empty_like(b_q_weight, memory_format=torch.contiguous_format)
628

629
630
    @register_fake("_C::cutlass_w4a8_mm")
    def cutlass_w4a8_mm_fake(
631
632
633
634
635
636
637
        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,
638
639
        out_type: torch.dtype | None = None,
        maybe_schedule: str | None = None,
640
    ) -> torch.Tensor:
641
642
643
644
645
646
647
648
649
650
651
652
653
        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)

654

655
656
657
if hasattr(torch.ops._C, "allspark_w8a16_gemm"):

    @register_fake("_C::allspark_w8a16_gemm")
658
659
660
661
    def _allspark_w8a16_gemm_fake(
        a: torch.Tensor,
        b_qweight: torch.Tensor,
        b_scales: torch.Tensor,
662
        b_qzeros: torch.Tensor | None,
663
664
665
666
667
668
669
670
        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:
671
672
673
674
        m = a.size(0)
        return torch.empty((m, n), device=a.device, dtype=a.dtype)


675
676
677
if hasattr(torch.ops._C, "ggml_dequantize"):

    @register_fake("_C::ggml_dequantize")
678
    def _ggml_dequantize_fake(
679
680
681
682
        W: torch.Tensor,
        quant_type: int,
        m: torch.SymInt,
        n: torch.SymInt,
683
        dtype: torch.dtype | None = None,
684
    ) -> torch.Tensor:
685
686
687
688
689
690
691
692
693
        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:
694
        return torch.empty((X.shape[0], row), dtype=X.dtype, device=W.device)
695
696
697
698
699
700
701
702
703

    @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)
704
        return torch.empty((batch, row), dtype=X.dtype, device=W.device)
705

706
707
708
709
710
711
712
713
714
715
716
717
718
    @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)
719
        return torch.empty((tokens * top_k, row), dtype=torch.float16, device=W.device)
720

721

722
723
724
725
726
727
728
729
730
731
732
733
734
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)
735
        return torch.empty((tokens * top_k, row), dtype=X.dtype, device=W.device)
736
737


738
# cutlass
739
740
741
742
def cutlass_scaled_mm_supports_fp4(cuda_device_capability: int) -> bool:
    return torch.ops._C.cutlass_scaled_mm_supports_fp4(cuda_device_capability)


743
744
745
746
747
748
749
750
751
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,
):
752
753
754
    torch.ops._C.cutlass_blockwise_scaled_grouped_mm(
        output, a, b, scales_a, scales_b, problem_sizes, expert_offsets
    )
755
756


757
758
759
760
761
762
763
764
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:
765
766
767
    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)
768
    torch.ops._C.cutlass_scaled_fp4_mm(out, a, b, block_scale_a, block_scale_b, alpha)
769
770
771
    return out


772
773
774
775
def cutlass_scaled_mm_supports_fp8(cuda_device_capability: int) -> bool:
    return torch.ops._C.cutlass_scaled_mm_supports_fp8(cuda_device_capability)


776
def cutlass_scaled_mm_supports_block_fp8(cuda_device_capability: int) -> bool:
777
    return torch.ops._C.cutlass_scaled_mm_supports_block_fp8(cuda_device_capability)
778
779


780
781
782
783
784
785
def cutlass_scaled_mm(
    a: torch.Tensor,
    b: torch.Tensor,
    scale_a: torch.Tensor,
    scale_b: torch.Tensor,
    out_dtype: torch.dtype,
786
    bias: torch.Tensor | None = None,
787
) -> torch.Tensor:
788
    """
789
    `cutlass_scaled_mm` implements a fused version of
790
        `output = torch.mm((scale_a * a), (scale_b * b)).to(out_dtype)`
791
792
793
794
795
796
797
798
    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
799
800
801
802
803
804
805
806
807
808
809
        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
    """
810
811
    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
812

813
814
815
    # Massage the input to be 2D
    target_shape = (*a.shape[:-1], b.shape[1])
    a = a.view(-1, a.shape[-1])
816

817
    cutlass_compatible_b = b.shape[0] % 16 == 0 and b.shape[1] % 16 == 0
818
    if current_platform.is_rocm() or not cutlass_compatible_b:
819
        from vllm.model_executor.layers.quantization.compressed_tensors.triton_scaled_mm import (  # noqa
820
821
822
            triton_scaled_mm,
        )

823
824
        out = triton_scaled_mm(a, b, scale_a, scale_b, out_dtype, bias)
    else:
825
        out = torch.empty((a.shape[0], b.shape[1]), dtype=out_dtype, device=a.device)
826
        torch.ops._C.cutlass_scaled_mm(out, a, b, scale_a, scale_b, bias)
827

828
    return out.view(*target_shape)
829
830


831
832
833
834
835
836
837
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,
838
839
    azp: torch.Tensor | None = None,
    bias: torch.Tensor | None = None,
840
) -> torch.Tensor:
841
842
843
844
845
    """
    :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.
    """
846
847
848
    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
849

850
851
852
853
    # 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]
854

855
856
    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)
857
    return out.view(*target_shape)
858
859


860
def cutlass_sparse_scaled_mm_supported(cuda_device_capability: int) -> bool:
861
    return torch.ops._C.cutlass_sparse_scaled_mm_supported(cuda_device_capability)
862
863


864
def cutlass_group_gemm_supported(cuda_device_capability: int) -> bool:
865
866
867
868
869
    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
870

871

872
def cutlass_sparse_compress(a: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
873
874
875
876
877
878
879
880
    """
    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:
881
        a (torch.Tensor):
882
883
884
885
886
887
888
            The input tensor to be compressed. Must have one of the following data types:
            - `torch.int8`
            - `torch.float8_e4m3fn`
            - `torch.bfloat16`
            - `torch.float16`

    Returns:
889
        tuple[torch.Tensor, torch.Tensor]:
890
891
892
893
894
895
896
897
898
899
900
901
902
            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)`.
    """
903
904
    assert a.dtype in [torch.int8, torch.float8_e4m3fn, torch.bfloat16, torch.float16]
    assert a.is_contiguous()
905
906
907

    # a_meta.dtype: torch.uint8 so elemsPerMetaElem = 8b / 2b_per_nz = 4
    elemsPerMetaElem = 4
908
    assert a.shape[1] % (2 * elemsPerMetaElem) == 0
909

910
    return torch.ops._C.cutlass_sparse_compress(a)
911
912
913


def cutlass_scaled_sparse_mm(
914
915
916
917
918
919
    a: torch.Tensor,
    bt_nzs: torch.Tensor,
    bt_meta: torch.Tensor,
    scale_a: torch.Tensor,
    scale_b: torch.Tensor,
    out_dtype: torch.dtype,
920
    bias: torch.Tensor | None = None,
921
) -> torch.Tensor:
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
    """
    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.
    """
945
946
947
    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
948
949
950
951
952

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

953
954
955
    torch.ops._C.cutlass_scaled_sparse_mm(
        out, a, bt_nzs, bt_meta, scale_a, scale_b, bias
    )
956
957
958
959

    return out


960
961
962
963
964
965
966
967
968
969
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,
970
    blockscale_offsets: torch.Tensor | None = None,
971
):
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
    """
    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.
989
990
991
992
993
    - 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]
994
    """
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
    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,
    )
1007
1008


1009
def get_cutlass_moe_mm_problem_sizes(
1010
1011
1012
1013
1014
1015
    topk_ids: torch.Tensor,
    problem_sizes1: torch.Tensor,
    problem_sizes2: torch.Tensor,
    num_experts: int,
    n: int,
    k: int,
1016
    blockscale_offsets: torch.Tensor | None = None,
1017
):
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
    """
    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.
    """
    return torch.ops._C.get_cutlass_moe_mm_problem_sizes(
1028
1029
        topk_ids, problem_sizes1, problem_sizes2, num_experts, n, k, blockscale_offsets
    )
1030
1031


1032
1033
1034
1035
1036
1037
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]
1038
1039
1040
1041
1042
    output_tensor = torch.empty(
        (num_tokens_permuted, input_tensor.shape[1]),
        device=input_tensor.device,
        dtype=input_tensor.dtype,
    )
1043
1044
    torch.ops._moe_C.shuffle_rows(input_tensor, dst2src_map, output_tensor)
    return output_tensor
1045
1046


1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
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,
):
1057
1058
1059
1060
1061
    """
    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
1062
    non_zero_expert_idxs (consecutive indices of experts with non-zero token
1063
1064
1065
1066
1067
1068
1069
1070
    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(
1071
1072
1073
1074
1075
1076
1077
1078
1079
        expert_offsets,
        problem_sizes1,
        problem_sizes2,
        expert_num_tokens,
        num_local_experts,
        padded_m,
        n,
        k,
    )
1080
1081


1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
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,
):
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
    """
    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.
    """
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
    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,
    )
1121
1122


1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
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,
):
1134
    """
1135
    An FP4 Blockscaled Group Gemm that takes in  a_tensors, b_tensors and runs
1136
1137
1138
1139
1140
1141
    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
1142
1143
1144
1145
    - 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
1146
1147
1148
1149
                    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.
    """
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
    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,
    )
1161
1162


1163
# gptq_marlin
1164
1165
1166
1167
1168
1169
1170
1171
def gptq_marlin_repack(
    b_q_weight: torch.Tensor,
    perm: torch.Tensor,
    size_k: int,
    size_n: int,
    num_bits: int,
) -> torch.Tensor:
    return torch.ops._C.gptq_marlin_repack(b_q_weight, perm, size_k, size_n, num_bits)
1172
1173


1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
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,
    ) -> 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
1194
1195
1196
def awq_marlin_repack(
    b_q_weight: torch.Tensor, size_k: int, size_n: int, num_bits: int
) -> torch.Tensor:
1197
1198
1199
    return torch.ops._C.awq_marlin_repack(b_q_weight, size_k, size_n, num_bits)


1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
if hasattr(torch.ops._C, "awq_marlin_repack"):

    @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,
    ) -> 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,
        )


1218
1219
1220
1221
1222
1223
1224
def gptq_marlin_moe_repack(
    b_q_weight: torch.Tensor,
    perm: torch.Tensor,
    size_k: int,
    size_n: int,
    num_bits: int,
) -> torch.Tensor:
1225
1226
    num_experts = b_q_weight.shape[0]
    assert size_k % 16 == 0
1227
1228
1229
1230
1231
    output = torch.empty(
        (num_experts, size_k // 16, size_n * (num_bits // 2)),
        device=b_q_weight.device,
        dtype=b_q_weight.dtype,
    )
1232
    for e in range(num_experts):
1233
1234
1235
        output[e] = torch.ops._C.gptq_marlin_repack(
            b_q_weight[e], perm[e], size_k, size_n, num_bits
        )
1236
1237
1238
    return output


1239
1240
1241
1242
1243
1244
1245
def awq_marlin_moe_repack(
    b_q_weight: torch.Tensor,
    perm: torch.Tensor,
    size_k: int,
    size_n: int,
    num_bits: int,
) -> torch.Tensor:
1246
1247
    num_experts = b_q_weight.shape[0]
    assert size_k % 16 == 0
1248
1249
1250
1251
1252
    output = torch.empty(
        (num_experts, size_k // 16, size_n * (num_bits // 2)),
        device=b_q_weight.device,
        dtype=b_q_weight.dtype,
    )
1253
    for e in range(num_experts):
1254
1255
1256
        output[e] = torch.ops._C.awq_marlin_repack(
            b_q_weight[e], size_k, size_n, num_bits
        )
1257
1258
1259
    return output


1260
1261
def gptq_marlin_gemm(
    a: torch.Tensor,
1262
    c: torch.Tensor | None,
1263
    b_q_weight: torch.Tensor,
1264
    b_bias: torch.Tensor | None,
1265
    b_scales: torch.Tensor,
1266
1267
1268
1269
    global_scale: torch.Tensor | None,
    b_zeros: torch.Tensor | None,
    g_idx: torch.Tensor | None,
    perm: torch.Tensor | None,
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
    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,
        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,
    )
1300
1301


1302
# machete
1303
def machete_supported_schedules(
1304
1305
    a_type: torch.dtype,
    b_type: ScalarType,
1306
1307
1308
1309
1310
    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,
1311
) -> list[str]:
1312
    return torch.ops._C.machete_supported_schedules(
1313
1314
1315
1316
1317
1318
1319
1320
        a_type,
        b_type.id,
        group_scales_type,
        group_zeros_type,
        channel_scales_type,
        token_scales_type,
        out_type,
    )
1321
1322
1323


def machete_mm(
1324
1325
1326
1327
    a: torch.Tensor,
    # b_q Should be the tensor returned by machete_prepack_B
    b_q: torch.Tensor,
    b_type: ScalarType,
1328
1329
1330
1331
1332
1333
1334
    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,
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
) -> 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,
    )
1348
1349
1350


def machete_prepack_B(
1351
1352
1353
    b_q_weight: torch.Tensor,
    a_type: torch.dtype,
    b_type: ScalarType,
1354
    group_scales_type: torch.dtype | None,
1355
1356
1357
1358
) -> torch.Tensor:
    return torch.ops._C.machete_prepack_B(
        b_q_weight, a_type, b_type.id, group_scales_type
    )
1359
1360


1361
1362
# CUTLASS W4A8
def cutlass_w4a8_mm(
1363
1364
1365
1366
1367
1368
1369
    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,
1370
1371
    out_type: torch.dtype | None = None,
    maybe_schedule: str | None = None,
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
) -> 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,
    )
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392


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)


1393
if hasattr(torch.ops._C, "permute_cols"):
1394

1395
    @register_fake("_C::permute_cols")
1396
    def _permute_cols_fake(a: torch.Tensor, perm: torch.Tensor) -> torch.Tensor:
1397
1398
1399
1400
1401
1402
1403
        return torch.empty_like(a)


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


1404
1405
# fp4
def scaled_fp4_quant(
1406
1407
    input: torch.Tensor, input_global_scale: torch.Tensor
) -> tuple[torch.Tensor, torch.Tensor]:
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
    """
    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:
1422
        tuple[torch.Tensor, torch.Tensor]: The output tensor in FP4 but every
1423
1424
1425
            two values are packed into a uint8 and float8_e4m3 scaling factors
            in the sizzled layout.
    """
1426
    assert not current_platform.is_rocm()
1427
    assert input.ndim >= 1, f"input.ndim needs to be >= 1, but got {input.ndim}."
1428
1429
1430
1431
1432
1433
    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

1434
    assert n % block_size == 0, f"last dim has to be multiple of 16, but got {n}."
1435
    assert input.dtype in (torch.float16, torch.bfloat16), (
1436
1437
        f"input.dtype needs to be fp16 or bf16 but got {input.dtype}."
    )
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450

    # 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)
1451
    output_scale = torch.empty(
1452
1453
        (rounded_m, rounded_n // 4), device=device, dtype=torch.int32
    )
1454

1455
    torch.ops._C.scaled_fp4_quant(output, input, output_scale, input_global_scale)
1456
1457
1458
1459
    output_scale = output_scale.view(torch.float8_e4m3fn)
    return output, output_scale


1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
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:
1471
        input_tensor: The input tensor to be quantized to FP4
1472
1473
1474
1475
1476
1477
1478
1479
1480
        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, (
1481
1482
        f"input.ndim needs to be == 2, but got {input_tensor.ndim}."
    )
1483

1484
1485
1486
1487
1488
    # 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
1489
1490
    m_numtopk, k = input_tensor.shape

1491
    assert m_numtopk <= MAX_TOKENS_PER_EXPERT * topk, (
1492
1493
1494
        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"
1495
1496
        f" VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE to set this value."
    )
1497
1498
1499
1500
    scales_k = k // 16
    padded_k = (scales_k + (4 - 1)) // 4

    # output is uint8 and packed fp4 values
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
    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,
    )
1518
1519
1520
1521
    output_scales = output_scales.view(torch.float8_e4m3fn)
    return output, output_scales


1522
# fp8
1523
1524
def scaled_fp8_quant(
    input: torch.Tensor,
1525
1526
1527
    scale: torch.Tensor | None = None,
    num_token_padding: int | None = None,
    scale_ub: torch.Tensor | None = None,
1528
    use_per_token_if_dynamic: bool = False,
1529
    output: torch.Tensor | None = None,
1530
) -> tuple[torch.Tensor, torch.Tensor]:
1531
1532
1533
1534
1535
1536
    """
    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
1537
    optional padding of the output tensors for downstream kernels that
1538
1539
1540
1541
1542
    will benefit from padding.

    Args:
        input: The input tensor to be quantized to FP8
        scale: Optional scaling factor for the FP8 quantization
1543
        scale_ub: Optional upper bound for scaling factor in dynamic
1544
            per token case
1545
        num_token_padding: If specified, pad the first dimension
1546
            of the output to at least this value.
1547
        use_per_token_if_dynamic: Whether to do per_tensor or per_token
1548
            in the dynamic quantization case.
1549
1550

    Returns:
1551
        tuple[torch.Tensor, torch.Tensor]: The output tensor in FP8 and
1552
1553
            scaling factor.
    """
1554
    # This code assumes batch_dim and num_tokens are flattened
1555
    assert input.ndim == 2
1556
    shape: tuple[int, int] | torch.Size = input.shape
1557
1558
    # For ROCm on MI300, the output fp8 dtype is torch.float_e3m3fnuz
    out_dtype: torch.dtype = current_platform.fp8_dtype()
1559
1560
    if num_token_padding:
        shape = (max(num_token_padding, input.shape[0]), shape[1])
1561
1562
1563
    if output is None:
        output = torch.empty(shape, device=input.device, dtype=out_dtype)
    else:
1564
        assert num_token_padding is None, "padding not supported if output passed in"
1565
        assert output.dtype == out_dtype
1566

1567
    if scale is None:
1568
        if use_per_token_if_dynamic:
1569
            scale = torch.empty((shape[0], 1), device=input.device, dtype=torch.float32)
1570
            torch.ops._C.dynamic_per_token_scaled_fp8_quant(
1571
1572
                output, input, scale, scale_ub
            )
1573
        else:
1574
            scale = torch.empty((1, 1), device=input.device, dtype=torch.float32)
1575
            torch.ops._C.dynamic_scaled_fp8_quant(output, input, scale)
1576
    else:
bnellnm's avatar
bnellnm committed
1577
        assert scale.numel() == 1, f"{scale.shape}"
1578
        torch.ops._C.static_scaled_fp8_quant(output, input, scale)
1579

1580
    return output, scale
1581
1582


1583
1584
# gptq allspark
def allspark_repack_weight(
1585
1586
    qweight: torch.Tensor,
    scale: torch.Tensor,
1587
    zero_point: torch.Tensor | None = None,
1588
    has_zp: bool = False,
1589
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
1590
    """
1591
    Rearrange qweight, scale, and zero_point(if asymmetric) to n32k16 format
1592
1593
1594
1595
1596
1597
1598
1599
    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.
1600
1601
            if use asymmetric quantization, has_zp = True.

1602
    Returns:
1603
        tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]] :
1604
1605
1606
1607
1608
1609
            rearranged weight, scale, and optionally zero_point.
    """
    K = qweight.shape[0]
    N = qweight.shape[1]
    N_32align = (N + 32 - 1) // 32 * 32

1610
1611
1612
1613
    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)
1614
1615
1616
    zero_point_reorder = None
    if has_zp:
        assert zero_point is not None, (
1617
1618
1619
1620
1621
            "zero_point must be provided for asymmetric quantization."
        )
        zero_point_reorder = torch.empty(
            (1, N_32align), device=zero_point.device, dtype=zero_point.dtype
        )
1622
1623

    torch.ops._C.rearrange_kn_weight_as_n32k16_order(
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
        qweight,
        scale,
        zero_point,
        has_zp,
        qweight_reorder,
        scale_reorder,
        zero_point_reorder,
        K,
        N,
        N_32align,
    )
1635
1636
1637
1638

    return qweight_reorder, scale_reorder, zero_point_reorder


1639
1640
1641
1642
def allspark_w8a16_gemm(
    a: torch.Tensor,
    b_qweight: torch.Tensor,
    b_scales: torch.Tensor,
1643
    b_qzeros: torch.Tensor | None,
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
    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,
    )
1665
1666


1667
# int8
1668
def scaled_int8_quant(
1669
    input: torch.Tensor,
1670
1671
    scale: torch.Tensor | None = None,
    azp: torch.Tensor | None = None,
1672
    symmetric: bool = True,
1673
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor | None]:
1674
    """
1675
    Quantize the input tensor to int8 and return the quantized tensor and scale, and maybe azp.
1676
1677
1678

    Args:
        input: The input tensor to be quantized to int8.
1679
1680
        scale: Optional scaling factor for the int8 quantization.
            When not provided, we invoke dynamic-per-token quantization.
1681
1682
1683
        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).
1684
1685

    Returns:
1686
      tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]] : Output int8 tensor, scales, and optionally azp.
1687
    """
1688
1689
1690
    output = torch.empty_like(input, dtype=torch.int8)
    if scale is not None:
        # static-per-tensor quantization.
1691
1692
1693
        assert symmetric == (azp is None), (
            "azp must only be provided for asymmetric quantization."
        )
1694
        torch.ops._C.static_scaled_int8_quant(output, input, scale, azp)
1695
        return output, scale, azp
1696
1697

    # dynamic-per-token quantization.
1698
1699
1700
1701
1702
1703
1704
    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
    )
1705
    return output, input_scales, input_azp
1706
1707


1708
# gguf
1709
def ggml_dequantize(
1710
    W: torch.Tensor, quant_type: int, m: int, n: int, dtype: torch.dtype | None
1711
) -> torch.Tensor:
1712
    return torch.ops._C.ggml_dequantize(W, quant_type, m, n, dtype)
1713
1714
1715
1716
1717
1718
1719


def ggml_mul_mat_vec_a8(
    W: torch.Tensor,
    X: torch.Tensor,
    quant_type: int,
    row: int,
1720
) -> torch.Tensor:
1721
1722
1723
1724
1725
1726
1727
1728
    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,
1729
) -> torch.Tensor:
1730
1731
1732
    return torch.ops._C.ggml_mul_mat_a8(W, X, quant_type, row)


1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
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:
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
    return torch.ops._C.ggml_moe_a8(
        X,
        W,
        sorted_token_ids,
        expert_ids,
        num_tokens_post_padded,
        quant_type,
        row,
        top_k,
        tokens,
    )
1755
1756


1757
1758
1759
1760
1761
1762
1763
1764
1765
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:
1766
    return torch.ops._C.ggml_moe_a8_vec(X, W, topk_ids, top_k, quant_type, row, tokens)
1767
1768


1769
1770
1771
1772
def ggml_moe_get_block_size(quant_type: int) -> int:
    return torch.ops._C.ggml_moe_get_block_size(quant_type)


1773
# mamba
1774
1775
1776
1777
1778
1779
def selective_scan_fwd(
    u: torch.Tensor,
    delta: torch.Tensor,
    A: torch.Tensor,
    B: torch.Tensor,
    C: torch.Tensor,
1780
1781
1782
    D_: torch.Tensor | None,
    z_: torch.Tensor | None,
    delta_bias_: torch.Tensor | None,
1783
    delta_softplus: bool,
1784
1785
1786
    query_start_loc: torch.Tensor | None,
    cache_indices: torch.Tensor | None,
    has_initial_state: torch.Tensor | None,
1787
1788
    ssm_states: torch.Tensor,
    pad_slot_id: int,
1789
1790
1791
1792
    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,
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
):
    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,
1809
1810
1811
1812
        block_size,
        block_idx_first_scheduled_token,
        block_idx_last_scheduled_token,
        initial_state_idx,
1813
    )
1814
1815


1816
# ROCm skinny gemms
1817
def LLMM1(a: torch.Tensor, b: torch.Tensor, rows_per_block: int) -> torch.Tensor:
1818
1819
1820
    return torch.ops._rocm_C.LLMM1(a, b, rows_per_block)


1821
1822
1823
def wvSplitK(
    a: torch.Tensor, b: torch.Tensor, cu_count: int, bias: torch.Tensor = None
) -> torch.Tensor:
1824
1825
1826
    return torch.ops._rocm_C.wvSplitK(a, b, bias, cu_count)


1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
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)
1837
    torch.ops._rocm_C.wvSplitKQ(a, b, bias, out, scale_a, scale_b, cu_count)
1838
1839
1840
    return out


1841
# moe
1842
1843
1844
1845
def moe_sum(input: torch.Tensor, output: torch.Tensor):
    torch.ops._moe_C.moe_sum(input, output)


1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
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,
) -> None:
    torch.ops._moe_C.moe_align_block_size(
        topk_ids,
        num_experts,
        block_size,
        sorted_token_ids,
        experts_ids,
        num_tokens_post_pad,
    )
1862
1863


1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
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,
    )


1882
1883
1884
1885
1886
1887
def moe_lora_align_block_size(
    topk_ids: torch.Tensor,
    token_lora_mapping: torch.Tensor,
    num_experts: int,
    block_size: int,
    max_loras: int,
1888
1889
    max_num_tokens_padded: int,
    max_num_m_blocks: int,
1890
1891
1892
    sorted_token_ids: torch.Tensor,
    experts_ids: torch.Tensor,
    num_tokens_post_pad: torch.Tensor,
1893
1894
    adapter_enabled: torch.Tensor,
    lora_ids: torch.Tensor,
1895
1896
1897
1898
1899
1900
1901
) -> None:
    torch.ops._moe_C.moe_lora_align_block_size(
        topk_ids,
        token_lora_mapping,
        num_experts,
        block_size,
        max_loras,
1902
1903
        max_num_tokens_padded,
        max_num_m_blocks,
1904
1905
1906
        sorted_token_ids,
        experts_ids,
        num_tokens_post_pad,
1907
1908
        adapter_enabled,
        lora_ids,
1909
1910
1911
    )


1912
1913
1914
1915
1916
def moe_wna16_gemm(
    input: torch.Tensor,
    output: torch.Tensor,
    b_qweight: torch.Tensor,
    b_scales: torch.Tensor,
1917
1918
    b_qzeros: torch.Tensor | None,
    topk_weights: torch.Tensor | None,
1919
1920
1921
1922
1923
1924
1925
1926
1927
    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:
1928
1929
    if not current_platform.is_cuda():
        raise NotImplementedError(
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
            "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,
    )
1948
1949


1950
1951
1952
1953
1954
def topk_softmax(
    topk_weights: torch.Tensor,
    topk_ids: torch.Tensor,
    token_expert_indices: torch.Tensor,
    gating_output: torch.Tensor,
1955
    renormalize: bool = False,
1956
1957
) -> None:
    torch.ops._moe_C.topk_softmax(
1958
        topk_weights, topk_ids, token_expert_indices, gating_output, renormalize
1959
    )
1960
1961


1962
1963
1964
1965
1966
1967
1968
def grouped_topk(
    scores: torch.Tensor,
    num_expert_group: int,
    topk_group: int,
    topk: int,
    renormalize: bool,
    routed_scaling_factor: float,
1969
1970
    bias: torch.Tensor,
    scoring_func: int = 0,
1971
):
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
    """
    Perform grouped top-k routing for mixture of experts.

    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
    """
1985
    if not current_platform.is_cuda():
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
        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,
1996
1997
        bias,
        scoring_func,
1998
1999
2000
2001
2002
    )


def moe_wna16_marlin_gemm(
    input: torch.Tensor,
2003
    output: torch.Tensor | None,
2004
    b_qweight: torch.Tensor,
2005
    b_bias: torch.Tensor | None,
2006
    b_scales: torch.Tensor,
2007
2008
2009
2010
    global_scale: torch.Tensor | None,
    b_qzeros: torch.Tensor | None,
    g_idx: torch.Tensor | None,
    perm: torch.Tensor | None,
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
    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,
) -> torch.Tensor:
2029
    return torch.ops._moe_C.moe_wna16_marlin_gemm(
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
        input,
        output,
        b_qweight,
        b_bias,
        b_scales,
        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,
    )
2057
2058


2059
if hasattr(torch.ops, "_moe_C") and hasattr(torch.ops._moe_C, "marlin_gemm_moe"):
2060

2061
    @register_fake("_moe_C::marlin_gemm_moe")
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
    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)
2085

2086
    @register_fake("_moe_C::moe_wna16_marlin_gemm")
2087
2088
    def moe_wna16_marlin_gemm_fake(
        input: torch.Tensor,
2089
        output: torch.Tensor | None,
2090
2091
        b_qweight: torch.Tensor,
        b_scales: torch.Tensor,
2092
2093
2094
        b_qzeros: torch.Tensor | None,
        g_idx: torch.Tensor | None,
        perm: torch.Tensor | None,
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
        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,
    ) -> torch.Tensor:
        return torch.empty(
            (size_m * top_k, size_n), dtype=input.dtype, device=input.device
        )
2116

2117

2118
2119
2120
2121
2122
2123
2124
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,
2125
2126
    k_scale: torch.Tensor,
    v_scale: torch.Tensor,
2127
) -> None:
2128
2129
2130
2131
2132
2133
2134
2135
2136
2137
    torch.ops._C_cache_ops.reshape_and_cache(
        key,
        value,
        key_cache,
        value_cache,
        slot_mapping,
        kv_cache_dtype,
        k_scale,
        v_scale,
    )
2138
2139


2140
2141
2142
2143
2144
2145
2146
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,
2147
2148
    k_scale: torch.Tensor,
    v_scale: torch.Tensor,
2149
) -> None:
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
    torch.ops._C_cache_ops.reshape_and_cache_flash(
        key,
        value,
        key_cache,
        value_cache,
        slot_mapping,
        kv_cache_dtype,
        k_scale,
        v_scale,
    )
2160
2161


2162
2163
2164
2165
2166
2167
2168
2169
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:
2170
2171
2172
    torch.ops._C_cache_ops.concat_and_cache_mla(
        kv_c, k_pe, kv_cache, slot_mapping, kv_cache_dtype, scale
    )
2173
2174


2175
2176
2177
2178
2179
def copy_blocks(
    key_caches: list[torch.Tensor],
    value_caches: list[torch.Tensor],
    block_mapping: torch.Tensor,
) -> None:
2180
    torch.ops._C_cache_ops.copy_blocks(key_caches, value_caches, block_mapping)
2181
2182


2183
def copy_blocks_mla(kv_caches: list[torch.Tensor], block_mapping: torch.Tensor) -> None:
2184
2185
2186
    torch.ops._C_cache_ops.copy_blocks_mla(kv_caches, block_mapping)


2187
2188
2189
def swap_blocks(
    src: torch.Tensor, dst: torch.Tensor, block_mapping: torch.Tensor
) -> None:
2190
    torch.ops._C_cache_ops.swap_blocks(src, dst, block_mapping)
2191
2192


2193
2194
2195
def convert_fp8(
    output: torch.Tensor, input: torch.Tensor, scale: float = 1.0, kv_dtype: str = "fp8"
) -> None:
2196
2197
2198
    torch.ops._C_cache_ops.convert_fp8(output, input, scale, kv_dtype)


2199
def gather_and_maybe_dequant_cache(
2200
2201
2202
2203
2204
2205
2206
    src_cache: torch.Tensor,
    dst: torch.Tensor,
    block_table: torch.Tensor,
    cu_seq_lens: torch.Tensor,
    batch_size: int,
    kv_cache_dtype: str,
    scale: torch.Tensor,
2207
    seq_starts: torch.Tensor | None = None,
2208
) -> None:
2209
    torch.ops._C_cache_ops.gather_and_maybe_dequant_cache(
2210
2211
2212
2213
2214
2215
2216
2217
2218
        src_cache,
        dst,
        block_table,
        cu_seq_lens,
        batch_size,
        kv_cache_dtype,
        scale,
        seq_starts,
    )
2219
2220


2221
2222
2223
2224
2225
2226
def cp_gather_cache(
    src_cache: torch.Tensor,
    dst: torch.Tensor,
    block_table: torch.Tensor,
    cu_seq_lens: torch.Tensor,
    batch_size: int,
2227
    seq_starts: torch.Tensor | None = None,
2228
2229
2230
2231
) -> None:
    torch.ops._C_cache_ops.cp_gather_cache(
        src_cache, dst, block_table, cu_seq_lens, batch_size, seq_starts
    )
2232
2233


2234
2235
2236
2237
2238
2239
2240
2241
2242
2243
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
    )
2244
2245


2246
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
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
    )


2258
2259
2260
2261
2262
2263
2264
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(
2265
2266
        device
    )
2267
2268
2269


# custom ar
2270
2271
2272
2273
2274
2275
2276
2277
2278
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
    )
2279
2280


2281
2282
2283
2284
2285
2286
2287
2288
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)
2289

2290
2291
2292
2293
2294
2295
2296
2297
2298

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


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


2299
def register_buffer(fa: int, ipc_tensors: list[int]) -> None:
2300
    return torch.ops._C_custom_ar.register_buffer(fa, ipc_tensors)
2301
2302


2303
def get_graph_buffer_ipc_meta(fa: int) -> tuple[list[int], list[int]]:
2304
2305
2306
    return torch.ops._C_custom_ar.get_graph_buffer_ipc_meta(fa)


2307
2308
2309
def register_graph_buffers(
    fa: int, handles: list[list[int]], offsets: list[list[int]]
) -> None:
2310
    torch.ops._C_custom_ar.register_graph_buffers(fa, handles, offsets)
2311
2312


2313
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
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)


2325
# quick all reduce
2326
def init_custom_qr(rank: int, world_size: int, qr_max_size: int | None = None) -> int:
2327
2328
2329
2330
2331
2332
2333
    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)


2334
2335
2336
2337
2338
2339
2340
2341
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)
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351
2352
2353
2354
2355


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()


2356
2357
2358
2359
def get_flash_mla_metadata(
    cache_seqlens: torch.Tensor,
    num_heads_per_head_k: int,
    num_heads_k: int,
2360
) -> tuple[torch.Tensor, torch.Tensor]:
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
    """
    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.
    """
2371
2372
2373
    return torch.ops._C.get_flash_mla_metadata(
        cache_seqlens, num_heads_per_head_k, num_heads_k
    )
2374
2375
2376
2377
2378
2379
2380
2381
2382
2383


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,
2384
    softmax_scale: float | None = None,
2385
    causal: bool = False,
2386
) -> tuple[torch.Tensor, torch.Tensor]:
2387
2388
2389
2390
2391
2392
2393
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403
    """
    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:
2404
        softmax_scale = q.shape[-1] ** (-0.5)
2405
2406
2407
2408
2409
2410
2411
2412
2413
2414
2415
2416
2417
    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
2418
2419


2420
2421
2422
2423
2424
2425
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
2443
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,
    )
2444
2445
2446
    return out


2447
2448
2449
def sm100_cutlass_mla_get_workspace_size(
    max_seq_len: int, num_batches: int, sm_count: int, num_kv_splits: int
) -> int:
2450
    return torch.ops._C.sm100_cutlass_mla_get_workspace_size(
2451
2452
        max_seq_len, num_batches, sm_count, num_kv_splits
    )
2453
2454


2455
2456
2457
if hasattr(torch.ops._C, "weight_packed_linear"):

    @register_fake("_C::weight_packed_linear")
2458
2459
2460
    def weight_packed_linear_fake(
        mat1: torch.Tensor,
        mat2: torch.Tensor,
2461
        bias: torch.Tensor | None,
2462
2463
2464
2465
2466
        is_vnni: bool,
    ) -> torch.Tensor:
        return torch.empty(
            (mat1.size(0), mat2.size(0)), dtype=mat1.dtype, device=mat2.device
        )
2467
2468
2469
2470
2471
2472
2473
2474
2475
2476
2477
2478
2479
2480


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,
2481
2482
2483
2484
2485
        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,
2486
2487
2488
2489
2490
2491
2492
2493
2494
2495
2496
2497
        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,
2498
        bias: torch.Tensor | None,
2499
2500
2501
2502
2503
2504
        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)
2505
2506
2507
2508


class CPUDNNLGEMMHandler:
    def __init__(self) -> None:
2509
        self.handler: int | None = None
2510
2511
2512
2513
2514
2515
2516
2517
        self.n = -1
        self.k = -1

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


2518
_supports_onednn = bool(hasattr(torch.ops._C, "create_onednn_mm_handler"))
2519
2520


2521
2522
2523
2524
def is_onednn_acl_supported():
    return torch.ops._C.is_onednn_acl_supported()


2525
2526
2527
2528
2529
2530
2531
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(
2532
2533
        weight, primitive_cache_size
    )
2534
2535
2536
2537
2538
2539
    return handler


def onednn_mm(
    dnnl_handler: CPUDNNLGEMMHandler,
    x: torch.Tensor,
2540
    bias: torch.Tensor | None,
2541
2542
) -> torch.Tensor:
    output = torch.empty((*x.shape[0:-1], dnnl_handler.n), dtype=x.dtype)
2543
2544
2545
    torch.ops._C.onednn_mm(
        output, x.reshape(-1, dnnl_handler.k), bias, dnnl_handler.handler
    )
2546
2547
2548
2549

    return output


2550
2551
2552
2553
2554
2555
2556
2557
2558
2559
2560
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(
2561
2562
        weight, weight_scales, output_type, dynamic_quant, use_azp, primitive_cache_size
    )
2563
2564
2565
    return handler


2566
2567
def onednn_scaled_int8_quant(
    input: torch.Tensor,
2568
2569
    scale: torch.Tensor | None = None,
    azp: torch.Tensor | None = None,
2570
2571
    symmetric: bool = True,
):
2572
2573
2574
2575
2576
2577
2578
2579
2580
2581
2582
2583
2584
2585
2586
2587
2588
2589
2590
    """
    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:
      tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]] : Output int8 tensor, scales, and optionally azp.
    """
    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.
2591
2592
2593
        assert symmetric == (azp is None), (
            "azp must only be provided for asymmetric quantization."
        )
2594
2595
2596
2597
        torch.ops._C.static_scaled_int8_quant(output, input, scale, azp)
        return output, scale, azp

    # dynamic-per-token quantization.
2598
2599
2600
    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)
2601
2602
2603
2604
2605
2606
2607
    return output, input_scales, input_azp


def onednn_scaled_mm(
    dnnl_handler: CPUDNNLGEMMHandler,
    x: torch.Tensor,
    output: torch.Tensor,
2608
2609
2610
2611
    input_scale: torch.Tensor | None,
    input_zp: torch.Tensor | None,
    input_zp_adj: torch.Tensor | None,
    bias: torch.Tensor | None,
2612
) -> torch.Tensor:
2613
2614
2615
    torch.ops._C.onednn_scaled_mm(
        output, x, input_scale, input_zp, input_zp_adj, bias, dnnl_handler.handler
    )
2616
2617

    return output
2618
2619


2620
2621
2622
2623
2624
2625
2626
2627
2628
2629
2630
2631
2632
2633
2634
2635
2636
2637
2638
2639
2640
2641
2642
2643
2644
2645
2646
2647
2648
2649
2650
2651
2652
2653
2654
2655
2656
2657
2658
2659
2660
2661
2662
2663
2664
2665
2666
2667
2668
2669
2670
2671
2672
2673
2674
2675
2676
2677
2678
2679
2680
2681
2682
2683
2684
2685
2686
2687
2688
2689
2690
2691
2692
2693
2694
2695
2696
2697
2698
2699
2700
2701
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,
) -> torch.Tensor:
    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,
    )


Li, Jiang's avatar
Li, Jiang committed
2702
2703
2704
2705
2706
2707
2708
2709
2710
2711
2712
2713
2714
2715
2716
2717
2718
2719
2720
2721
2722
2723
2724
2725
2726
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,
    )
    return output


2727
2728
2729
2730
2731
2732
2733
2734
2735
2736
2737
2738
2739
2740
2741
2742
2743
2744
2745
2746
2747
2748
2749
2750
2751
2752
2753
2754
2755
2756
2757
2758
2759
2760
2761
2762
2763
2764
2765
2766
2767
2768
2769
2770
2771
2772
2773
2774
2775
2776
2777
2778
2779
2780
2781
2782
2783
2784
2785
2786
2787
2788
2789
2790
2791
2792
2793
2794
2795
2796
2797
2798
2799
2800
2801
2802
2803
2804
2805
2806
2807
2808
2809
2810
2811
2812
2813
2814
2815
2816
2817
2818
2819
2820
2821
2822
2823
2824
2825
2826
2827
2828
2829
2830
2831
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
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)


2865
2866
2867
2868
2869
2870
2871
2872
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.
2873

2874
2875
2876
2877
2878
2879
2880
2881
2882
2883
    :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")
2884
    def _hadacore_transform_fake(x: torch.Tensor, inplace: bool) -> torch.Tensor:
2885
        return torch.empty_like(x) if not inplace else x