_custom_ops.py 41.5 KB
Newer Older
1
import contextlib
2
import importlib
3
from typing import TYPE_CHECKING, List, Optional, Tuple, Union
4
5

import torch
6
import torch.library
7

8
import vllm.envs as envs
9
from vllm.logger import init_logger
10
from vllm.platforms import current_platform
11
from vllm.scalar_type import ScalarType
12
13
14

logger = init_logger(__name__)

15
if not current_platform.is_tpu() and not current_platform.is_hpu():
16
17
18
19
    try:
        import vllm._C
    except ImportError as e:
        logger.warning("Failed to import from vllm._C with %r", e)
20

21
supports_moe_ops = False
22
with contextlib.suppress(ImportError):
23
    import vllm._moe_C  # noqa: F401
24
    supports_moe_ops = True
25

26
if TYPE_CHECKING:
27
28
29
30
31
32
33
34
35

    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

36
37
38

# activation ops
def silu_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
39
    torch.ops._C.silu_and_mul(out, x)
40
41
42


def gelu_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
43
    torch.ops._C.gelu_and_mul(out, x)
44
45
46


def gelu_tanh_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
47
    torch.ops._C.gelu_tanh_and_mul(out, x)
48
49


50
51
52
53
54
55
def fatrelu_and_mul(out: torch.Tensor,
                    x: torch.Tensor,
                    threshold: float = 0.0) -> None:
    torch.ops._C.fatrelu_and_mul(out, x, threshold)


56
def gelu_fast(out: torch.Tensor, x: torch.Tensor) -> None:
57
    torch.ops._C.gelu_fast(out, x)
58
59
60


def gelu_new(out: torch.Tensor, x: torch.Tensor) -> None:
61
    torch.ops._C.gelu_new(out, x)
62
63


64
65
66
67
def gelu_quick(out: torch.Tensor, x: torch.Tensor) -> None:
    torch.ops._C.gelu_quick(out, x)


68
69
70
71
72
73
74
75
76
# 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,
77
    seq_lens: torch.Tensor,
78
    block_size: int,
79
    max_seq_len: int,
80
81
    alibi_slopes: Optional[torch.Tensor],
    kv_cache_dtype: str,
82
83
    k_scale: float,
    v_scale: float,
84
85
86
87
88
    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,
89
) -> None:
90
    torch.ops._C.paged_attention_v1(
91
92
        out, query, key_cache, value_cache, num_kv_heads, scale, block_tables,
        seq_lens, block_size, max_seq_len, alibi_slopes, kv_cache_dtype,
93
94
95
        k_scale, v_scale, tp_rank, blocksparse_local_blocks,
        blocksparse_vert_stride, blocksparse_block_size,
        blocksparse_head_sliding_step)
96
97
98
99
100
101
102
103
104
105
106
107
108


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,
109
    seq_lens: torch.Tensor,
110
    block_size: int,
111
    max_seq_len: int,
112
113
    alibi_slopes: Optional[torch.Tensor],
    kv_cache_dtype: str,
114
115
    k_scale: float,
    v_scale: float,
116
117
118
119
120
    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,
121
) -> None:
122
    torch.ops._C.paged_attention_v2(
123
124
        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,
125
        alibi_slopes, kv_cache_dtype, k_scale, v_scale, tp_rank,
126
127
        blocksparse_local_blocks, blocksparse_vert_stride,
        blocksparse_block_size, blocksparse_head_sliding_step)
128
129


130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
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,
    block_size: int,
    max_seq_len: int,
    alibi_slopes: Optional[torch.Tensor],
    kv_cache_dtype: str,
146
147
    k_scale: float,
    v_scale: float,
148
149
150
151
152
) -> None:
    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,
                                      block_size, max_seq_len, alibi_slopes,
153
                                      kv_cache_dtype, k_scale, v_scale)
154
155


156
157
158
159
160
161
162
163
164
# pos encoding ops
def rotary_embedding(
    positions: torch.Tensor,
    query: torch.Tensor,
    key: torch.Tensor,
    head_size: int,
    cos_sin_cache: torch.Tensor,
    is_neox: bool,
) -> None:
165
166
    torch.ops._C.rotary_embedding(positions, query, key, head_size,
                                  cos_sin_cache, is_neox)
167
168
169
170
171
172
173


def batched_rotary_embedding(positions: torch.Tensor, query: torch.Tensor,
                             key: torch.Tensor, head_size: int,
                             cos_sin_cache: torch.Tensor, is_neox: bool,
                             rot_dim: int,
                             cos_sin_cache_offsets: torch.Tensor) -> None:
174
175
176
    torch.ops._C.batched_rotary_embedding(positions, query, key, head_size,
                                          cos_sin_cache, is_neox, rot_dim,
                                          cos_sin_cache_offsets)
177
178
179
180
181


# layer norm ops
def rms_norm(out: torch.Tensor, input: torch.Tensor, weight: torch.Tensor,
             epsilon: float) -> None:
182
    torch.ops._C.rms_norm(out, input, weight, epsilon)
183
184
185
186


def fused_add_rms_norm(input: torch.Tensor, residual: torch.Tensor,
                       weight: torch.Tensor, epsilon: float) -> None:
187
    torch.ops._C.fused_add_rms_norm(input, residual, weight, epsilon)
188
189


190
191
192
193
194
195
def advance_step_flashattn(num_seqs: int, num_queries: int, block_size: int,
                           input_tokens: torch.Tensor,
                           sampled_token_ids: torch.Tensor,
                           input_positions: torch.Tensor,
                           seq_lens: torch.Tensor, slot_mapping: torch.Tensor,
                           block_tables: torch.Tensor) -> None:
196
    """Advance a step on GPU for existing inputs for a multi-step runner"""
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
    return torch.ops._C.advance_step_flashattn(num_seqs, num_queries,
                                               block_size, input_tokens,
                                               sampled_token_ids,
                                               input_positions, seq_lens,
                                               slot_mapping, block_tables)


def advance_step_flashinfer(num_seqs: int, num_queries: int, block_size: int,
                            input_tokens: torch.Tensor,
                            sampled_token_ids: torch.Tensor,
                            input_positions: torch.Tensor,
                            seq_lens: torch.Tensor, slot_mapping: torch.Tensor,
                            block_tables: torch.Tensor,
                            paged_kv_indices: torch.Tensor,
                            paged_kv_indptr: torch.Tensor,
                            paged_kv_last_page_len: torch.Tensor,
                            block_table_bound: torch.Tensor) -> None:

    return torch.ops._C.advance_step_flashinfer(
        num_seqs, num_queries, block_size, input_tokens, sampled_token_ids,
        input_positions, seq_lens, slot_mapping, block_tables,
        paged_kv_indices, paged_kv_indptr, paged_kv_last_page_len,
        block_table_bound)
220
221


222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
# fused quant layer norm ops
def rms_norm_dynamic_per_token_quant(
    input: torch.Tensor,
    weight: torch.Tensor,
    epsilon: float,
    quant_dtype: torch.dtype,
    scale_ub: Optional[torch.Tensor] = None,
    residual: Optional[torch.Tensor] = None
) -> Tuple[torch.Tensor, torch.Tensor]:
    output = torch.empty_like(input, dtype=quant_dtype)
    scales = torch.empty((input.numel() // input.shape[-1], 1),
                         device=input.device,
                         dtype=torch.float32)

    torch.ops._C.rms_norm_dynamic_per_token_quant(output, input, weight,
                                                  scales, epsilon, scale_ub,
                                                  residual)
    return output, scales


242
243
244
245
246
# quantization ops
# awq
def awq_dequantize(qweight: torch.Tensor, scales: torch.Tensor,
                   zeros: torch.Tensor, split_k_iters: int, thx: int,
                   thy: int) -> torch.Tensor:
247
248
249
250
    if envs.VLLM_USE_TRITON_AWQ:
        from vllm.model_executor.layers.quantization.awq_triton import (
            awq_dequantize_triton)
        return awq_dequantize_triton(qweight, scales, zeros)
251
252
    return torch.ops._C.awq_dequantize(qweight, scales, zeros, split_k_iters,
                                       thx, thy)
253
254
255
256


def awq_gemm(input: torch.Tensor, qweight: torch.Tensor, qzeros: torch.Tensor,
             scales: torch.Tensor, split_k_iters: int) -> torch.Tensor:
257
258
259
260
    if envs.VLLM_USE_TRITON_AWQ:
        from vllm.model_executor.layers.quantization.awq_triton import (
            awq_gemm_triton)
        return awq_gemm_triton(input, qweight, qzeros, scales, split_k_iters)
261
    return torch.ops._C.awq_gemm(input, qweight, qzeros, scales, split_k_iters)
262
263
264
265
266
267
268


# gptq
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,
              bit: int) -> torch.Tensor:
269
270
    return torch.ops._C.gptq_gemm(a, b_q_weight, b_gptq_qzeros, b_gptq_scales,
                                  b_g_idx, use_exllama, bit)
271
272


273
if hasattr(torch.ops._C, "gptq_gemm"):
274

275
    @register_fake("_C::gptq_gemm")
276
277
278
279
280
281
282
283
284
    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, bit: int) -> torch.Tensor:
        return torch.empty((a.size(0), b_q_weight.size(1)),
                           dtype=a.dtype,
                           device=a.device)


285
286
def gptq_shuffle(q_weight: torch.Tensor, q_perm: torch.Tensor,
                 bit: int) -> None:
287
    torch.ops._C.gptq_shuffle(q_weight, q_perm, bit)
288
289
290
291
292
293


# marlin
def marlin_gemm(a: torch.Tensor, b_q_weight: torch.Tensor,
                b_scales: torch.Tensor, workspace: torch.Tensor, size_m: int,
                size_n: int, size_k: int) -> torch.Tensor:
294
295
    return torch.ops._C.marlin_gemm(a, b_q_weight, b_scales, workspace, size_m,
                                    size_n, size_k)
296
297


298
299
300
# marlin_24
def gptq_marlin_24_gemm(a: torch.Tensor, b_q_weight: torch.Tensor,
                        b_meta: torch.Tensor, b_scales: torch.Tensor,
301
302
                        workspace: torch.Tensor, b_q_type: ScalarType,
                        size_m: int, size_n: int, size_k: int) -> torch.Tensor:
303
    return torch.ops._C.gptq_marlin_24_gemm(a, b_q_weight, b_meta, b_scales,
304
                                            workspace, b_q_type.id, size_m,
305
                                            size_n, size_k)
306
307


308
if hasattr(torch.ops._C, "gptq_marlin_24_gemm"):
309

310
    @register_fake("_C::gptq_marlin_24_gemm")
311
312
313
    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,
314
315
316
                                  b_q_type: ScalarType, size_m: torch.SymInt,
                                  size_n: torch.SymInt,
                                  size_k: torch.SymInt) -> torch.Tensor:
317
318
        return torch.empty((size_m, size_n), device=a.device, dtype=a.dtype)

319
    @register_fake("_C::gptq_marlin_gemm")
320
321
322
323
324
325
326
327
    def _gptq_marlin_gemm_fake(a: torch.Tensor,
                               b_q_weight: torch.Tensor,
                               b_scales: torch.Tensor,
                               b_zeros: torch.Tensor,
                               g_idx: torch.Tensor,
                               perm: torch.Tensor,
                               workspace: torch.Tensor,
                               b_q_type: ScalarType,
328
329
330
                               size_m: torch.SymInt,
                               size_n: torch.SymInt,
                               size_k: torch.SymInt,
331
332
                               is_k_full: bool,
                               has_zp: bool = False,
333
334
                               use_fp32_reduce: bool = False,
                               is_zp_float: bool = False) -> torch.Tensor:
335
336
        return torch.empty((size_m, size_n), device=a.device, dtype=a.dtype)

337
    @register_fake("_C::marlin_qqq_gemm")
338
339
340
    def _marlin_qqq_gemm_fake(a: torch.Tensor, b_q_weight: torch.Tensor,
                              s_tok: torch.Tensor, s_ch: torch.Tensor,
                              s_group: torch.Tensor, workspace: torch.Tensor,
341
342
                              size_m: torch.SymInt, size_n: torch.SymInt,
                              size_k: torch.SymInt) -> torch.Tensor:
343
344
345
346
        return torch.empty((size_m, size_n),
                           dtype=torch.float16,
                           device=a.device)

347
    @register_fake("_C::marlin_gemm")
348
349
    def _marlin_gemm_fake(a: torch.Tensor, b_q_weight: torch.Tensor,
                          b_scales: torch.Tensor, workspace: torch.Tensor,
350
351
                          size_m: torch.SymInt, size_n: torch.SymInt,
                          size_k: torch.SymInt) -> torch.Tensor:
352
353
354
355
        return torch.empty((size_m, size_n),
                           dtype=torch.float16,
                           device=a.device)

356
    @register_fake("_C::awq_dequantize")
357
    def _awq_dequantize_fake(qweight: torch.Tensor, scales: torch.Tensor,
358
359
                             zeros: torch.Tensor, split_k_iters: torch.SymInt,
                             thx: int, thy: int) -> torch.Tensor:
360
361
362
363
364
365
366
        in_c = qweight.size(0)
        qout_c = qweight.size(1)
        out_c = qout_c * 8
        return torch.empty((in_c, out_c),
                           dtype=scales.dtype,
                           device=scales.device)

367
    @register_fake("_C::awq_gemm")
368
369
    def _awq_gemm_fake(input: torch.Tensor, qweight: torch.Tensor,
                       qzeros: torch.Tensor, scales: torch.Tensor,
370
                       split_k_iters: torch.SymInt) -> torch.Tensor:
371
372
373
374
375
        num_in_feats = input.size(0)
        return torch.empty((split_k_iters, num_in_feats, qweight.size(1) * 8),
                           dtype=input.dtype,
                           device=input.device).sum(0)

376
    @register_fake("_C::aqlm_gemm")
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
    def _aqlm_gemm_fake(input: torch.Tensor, codes: torch.Tensor,
                        codebooks: torch.Tensor, scales: torch.Tensor,
                        codebook_partition_sizes: List[int],
                        bias: Optional[torch.Tensor]) -> torch.Tensor:
        out_features = codes.size(0) * codebooks.size(2)
        flat_input = input.reshape((-1, input.size(-1)))
        flat_output = torch.empty((flat_input.size(0), out_features),
                                  dtype=input.dtype,
                                  device=input.device)

        output_sizes = list(input.shape)
        output_sizes.pop()
        output_sizes.append(-1)
        return flat_output.reshape(tuple(output_sizes))

392
    @register_fake("_C::aqlm_dequant")
393
394
395
396
397
398
399
400
401
    def _aqlm_dequant_fake(
            codes: torch.Tensor, codebooks: torch.Tensor,
            codebook_partition_sizes: List[int]) -> torch.Tensor:
        in_features = codes.size(1) * 8
        out_features = codes.size(0)
        return torch.empty((out_features, in_features),
                           dtype=codebooks.dtype,
                           device=codebooks.device)

402
    @register_fake("_C::fp8_marlin_gemm")
403
404
    def _fp8_marlin_gemm_fake(a: torch.Tensor, b_q_weight: torch.Tensor,
                              b_scales: torch.Tensor, workspace: torch.Tensor,
405
406
407
                              num_bits: int, size_m: torch.SymInt,
                              size_n: torch.SymInt,
                              size_k: torch.SymInt) -> torch.Tensor:
408
409
        return torch.empty((size_m, size_n), dtype=a.dtype, device=a.device)

410
411
    @register_fake("_C::machete_mm")
    def machete_mm_fake(
412
        a: torch.Tensor,
413
        # b_q Should be the tensor returned by machete_prepack_B
414
        b_q: torch.Tensor,
415
        b_type: ScalarType,
416
417
418
        out_type: Optional[torch.dtype] = None,
        b_group_scales: Optional[torch.Tensor] = None,
        b_group_zeros: Optional[torch.Tensor] = None,
419
        b_group_size: Optional[int] = None,
420
421
        b_channel_scales: Optional[torch.Tensor] = None,
        a_token_scales: Optional[torch.Tensor] = None,
422
423
424
425
426
427
        schedule: Optional[str] = None,
    ) -> torch.Tensor:
        m = a.size(0)
        n = b_q.size(1)
        return torch.empty((m, n), device=a.device, dtype=a.dtype)

428
    @register_fake("_C::machete_prepack_B")
429
430
431
    def machete_prepack_B_fake(
            b_q_weight: torch.Tensor, a_type: torch.dtype, b_type: ScalarType,
            group_scales_type: Optional[torch.dtype]) -> torch.Tensor:
432
433
        return torch.empty_like(b_q_weight,
                                memory_format=torch.contiguous_format)
434
435


436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
if hasattr(torch.ops._C, "ggml_dequantize"):

    @register_fake("_C::ggml_dequantize")
    def _ggml_dequantize_fake(W: torch.Tensor, quant_type: int,
                              m: torch.SymInt,
                              n: torch.SymInt) -> torch.Tensor:
        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:
        return torch.empty((1, row), dtype=torch.float16, device=W.device)

    @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)
        return torch.empty((batch, row), dtype=torch.float16, device=W.device)


464
# cutlass
465
466
467
468
def cutlass_scaled_mm_supports_fp8(cuda_device_capability: int) -> bool:
    return torch.ops._C.cutlass_scaled_mm_supports_fp8(cuda_device_capability)


469
470
471
def cutlass_scaled_mm(a: torch.Tensor,
                      b: torch.Tensor,
                      scale_a: torch.Tensor,
472
                      scale_b: torch.Tensor,
473
                      out_dtype: torch.dtype,
474
                      bias: Optional[torch.Tensor] = None) -> torch.Tensor:
475
476
    assert (b.shape[0] % 16 == 0 and b.shape[1] % 16 == 0)
    assert (out_dtype is torch.bfloat16 or out_dtype is torch.float16)
477
478
    assert bias is None or bias.shape[0] == b.shape[
        1] and bias.dtype == out_dtype
479
480
481

    m = a.shape[0]
    n = b.shape[1]
482
483
484
485
486
487
488
489

    if current_platform.is_rocm():
        triton_scaled_mm_module = importlib.import_module(
            "vllm.model_executor.layers.quantization.compressed_tensors."
            "triton_scaled_mm")
        triton_scaled_mm = triton_scaled_mm_module.triton_scaled_mm
        return triton_scaled_mm(a, b, scale_a, scale_b, out_dtype, bias)

490
491
    out = torch.empty((m, n), dtype=out_dtype, device=a.device)

492
493
    torch.ops._C.cutlass_scaled_mm(out, a, b, scale_a, scale_b, bias)

494
495
496
    return out


497
498
499
500
501
502
503
504
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,
                          azp: Optional[torch.Tensor] = None,
                          bias: Optional[torch.Tensor] = None) -> torch.Tensor:
505
506
507
508
509
    """
    :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.
    """
510
511
512
513
    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
514
    assert azp is None or azp.numel() == a.shape[0]
515
516
517
518
519
520
521
522
523
524

    m = a.shape[0]
    n = b.shape[1]
    out = torch.empty((m, n), dtype=out_dtype, device=a.device)

    torch.ops._C.cutlass_scaled_mm_azp(out, a, b, scale_a, scale_b, azp_adj,
                                       azp, bias)
    return out


525
526
527
528
529
def cutlass_sparse_scaled_mm_supported(cuda_device_capability: int) -> bool:
    return torch.ops._C.cutlass_sparse_scaled_mm_supported(
        cuda_device_capability)


530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
def cutlass_sparse_compress(a: torch.Tensor) \
    -> Tuple[torch.Tensor, torch.Tensor]:
    """
    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:
        a (torch.Tensor): 
            The input tensor to be compressed. Must have one of the following data types:
            - `torch.int8`
            - `torch.float8_e4m3fn`
            - `torch.bfloat16`
            - `torch.float16`

    Returns:
        Tuple[torch.Tensor, torch.Tensor]: 
            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)`.
    """
    assert (a.dtype in [
        torch.int8, torch.float8_e4m3fn, torch.bfloat16, torch.float16
    ])
    assert (a.is_contiguous())

    # a_meta.dtype: torch.uint8 so elemsPerMetaElem = 8b / 2b_per_nz = 4
    elemsPerMetaElem = 4

    m = a.shape[0]
    k = a.shape[1]
    assert (k % 2 == 0)
    a_nzs = torch.empty((m, k // 2), dtype=a.dtype, device=a.device)
    a_meta = torch.empty((m, k // 2 // elemsPerMetaElem),
                         dtype=torch.uint8,
                         device=a.device)

    if not (torch.ops._C.cutlass_sparse_compress_entry(a_nzs, a_meta, a)):
        raise ValueError

    assert (a_nzs.is_contiguous())
    assert (a_meta.is_contiguous())

    return a_nzs, a_meta


def cutlass_scaled_sparse_mm(
        a: torch.Tensor,
        bt_nzs: torch.Tensor,
        bt_meta: torch.Tensor,
        scale_a: torch.Tensor,
        scale_b: torch.Tensor,
        out_dtype: torch.dtype,
        bias: Optional[torch.Tensor] = None) -> torch.Tensor:
    """
    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.
    """
    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

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

    torch.ops._C.cutlass_scaled_sparse_mm(out, a, bt_nzs, bt_meta, scale_a,
                                          scale_b, bias)

    return out


633
634
635
# aqlm
def aqlm_gemm(input: torch.Tensor, codes: torch.Tensor,
              codebooks: torch.Tensor, scales: torch.Tensor,
636
              codebook_partition_sizes: List[int],
637
              bias: Optional[torch.Tensor]) -> torch.Tensor:
638
639
    return torch.ops._C.aqlm_gemm(input, codes, codebooks, scales,
                                  codebook_partition_sizes, bias)
640
641
642


def aqlm_dequant(codes: torch.Tensor, codebooks: torch.Tensor,
643
                 codebook_partition_sizes: List[int]) -> torch.Tensor:
644
645
    return torch.ops._C.aqlm_dequant(codes, codebooks,
                                     codebook_partition_sizes)
646
647


648
649
# gptq_marlin
def gptq_marlin_repack(b_q_weight: torch.Tensor, perm: torch.Tensor,
650
651
                       size_k: int, size_n: int,
                       num_bits: int) -> torch.Tensor:
652
653
    return torch.ops._C.gptq_marlin_repack(b_q_weight, perm, size_k, size_n,
                                           num_bits)
654
655


656
657
658
659
660
661
# gptq_marlin
def awq_marlin_repack(b_q_weight: torch.Tensor, size_k: int, size_n: int,
                      num_bits: int) -> torch.Tensor:
    return torch.ops._C.awq_marlin_repack(b_q_weight, size_k, size_n, num_bits)


662
663
664
665
666
def gptq_marlin_moe_repack(b_q_weight: torch.Tensor, perm: torch.Tensor,
                           size_k: int, size_n: int,
                           num_bits: int) -> torch.Tensor:
    num_experts = b_q_weight.shape[0]
    assert size_k % 16 == 0
667
    output = torch.empty((num_experts, size_k // 16, size_n * (num_bits // 2)),
668
669
670
671
672
673
674
675
                         device=b_q_weight.device,
                         dtype=b_q_weight.dtype)
    for e in range(num_experts):
        output[e] = torch.ops._C.gptq_marlin_repack(b_q_weight[e], perm[e],
                                                    size_k, size_n, num_bits)
    return output


676
677
678
679
680
681
682
683
684
685
686
687
688
689
def awq_marlin_moe_repack(b_q_weight: torch.Tensor, perm: torch.Tensor,
                          size_k: int, size_n: int,
                          num_bits: int) -> torch.Tensor:
    num_experts = b_q_weight.shape[0]
    assert size_k % 16 == 0
    output = torch.empty((num_experts, size_k // 16, size_n * (num_bits // 2)),
                         device=b_q_weight.device,
                         dtype=b_q_weight.dtype)
    for e in range(num_experts):
        output[e] = torch.ops._C.awq_marlin_repack(b_q_weight[e], size_k,
                                                   size_n, num_bits)
    return output


690
691
692
693
694
695
696
697
698
699
700
701
702
def gptq_marlin_gemm(a: torch.Tensor,
                     b_q_weight: torch.Tensor,
                     b_scales: torch.Tensor,
                     b_zeros: torch.Tensor,
                     g_idx: torch.Tensor,
                     perm: torch.Tensor,
                     workspace: torch.Tensor,
                     b_q_type: ScalarType,
                     size_m: int,
                     size_n: int,
                     size_k: int,
                     is_k_full: bool,
                     has_zp: bool = False,
703
704
                     use_fp32_reduce: bool = False,
                     is_zp_float: bool = False) -> torch.Tensor:
705
    return torch.ops._C.gptq_marlin_gemm(a, b_q_weight, b_scales, b_zeros,
706
                                         g_idx, perm, workspace, b_q_type.id,
707
                                         size_m, size_n, size_k, is_k_full,
708
                                         has_zp, use_fp32_reduce, is_zp_float)
709
710


711
712
713
714
715
716
717
718
719
# fp8 marlin
def fp8_marlin_gemm(a: torch.Tensor, b_q_weight: torch.Tensor,
                    b_scales: torch.Tensor, workspace: torch.Tensor,
                    num_bits: int, size_m: int, size_n: int,
                    size_k: int) -> torch.Tensor:
    return torch.ops._C.fp8_marlin_gemm(a, b_q_weight, b_scales, workspace,
                                        num_bits, size_m, size_n, size_k)


720
# machete
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
def machete_supported_schedules(
        a_type: torch.dtype,
        b_type: ScalarType,
        group_scales_type: Optional[torch.dtype],
        group_zeros_type: Optional[torch.dtype] = None,
        channel_scales_type: Optional[torch.dtype] = None,
        token_scales_type: Optional[torch.dtype] = None,
        out_type: Optional[torch.dtype] = None) -> List[str]:
    return torch.ops._C.machete_supported_schedules(
        a_type, b_type.id, group_scales_type, group_zeros_type,
        channel_scales_type, token_scales_type, out_type)


def machete_mm(
        a: torch.Tensor,
        # b_q Should be the tensor returned by machete_prepack_B
        b_q: torch.Tensor,
        b_type: ScalarType,
        out_type: Optional[torch.dtype] = None,
        b_group_scales: Optional[torch.Tensor] = None,
        b_group_zeros: Optional[torch.Tensor] = None,
        b_group_size: Optional[int] = None,
        b_channel_scales: Optional[torch.Tensor] = None,
        a_token_scales: Optional[torch.Tensor] = None,
        schedule: Optional[str] = None) -> 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)


def machete_prepack_B(
        b_q_weight: torch.Tensor, a_type: torch.dtype, b_type: ScalarType,
        group_scales_type: Optional[torch.dtype]) -> torch.Tensor:
    return torch.ops._C.machete_prepack_B(b_q_weight, a_type, b_type.id,
                                          group_scales_type)
756
757


758
if hasattr(torch.ops._C, "permute_cols"):
759

760
    @register_fake("_C::permute_cols")
761
762
763
764
765
766
767
768
769
    def _permute_cols_fake(a: torch.Tensor,
                           perm: torch.Tensor) -> torch.Tensor:
        return torch.empty_like(a)


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


770
# fp8
771
772
773
def scaled_fp8_quant(
    input: torch.Tensor,
    scale: Optional[torch.Tensor] = None,
774
    num_token_padding: Optional[int] = None,
775
    scale_ub: Optional[torch.Tensor] = None,
776
    use_per_token_if_dynamic: bool = False,
777
) -> Tuple[torch.Tensor, torch.Tensor]:
778
779
780
781
782
783
    """
    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
784
    optional padding of the output tensors for downstream kernels that
785
786
787
788
789
    will benefit from padding.

    Args:
        input: The input tensor to be quantized to FP8
        scale: Optional scaling factor for the FP8 quantization
790
        scale_ub: Optional upper bound for scaling factor in dynamic
791
            per token case
792
        num_token_padding: If specified, pad the first dimension
793
            of the output to at least this value.
794
        use_per_token_if_dynamic: Whether to do per_tensor or per_token
795
            in the dynamic quantization case.
796
797
798
799
800

    Returns:
        Tuple[torch.Tensor, torch.Tensor]: The output tensor in FP8 and
            scaling factor.
    """
801
802
    # This code assumes batch_dim and num_tokens are flattened
    assert (input.ndim == 2)
803
    shape: Union[Tuple[int, int], torch.Size] = input.shape
804
    # For rocm, the output fp8 dtype is torch.float_e3m3fnuz
805
806
    out_dtype: torch.dtype = torch.float8_e4m3fnuz \
            if current_platform.is_rocm() else torch.float8_e4m3fn
807
808
    if num_token_padding:
        shape = (max(num_token_padding, input.shape[0]), shape[1])
809
    output = torch.empty(shape, device=input.device, dtype=out_dtype)
810

811
    if scale is None:
812
        if use_per_token_if_dynamic:
813
            scale = torch.empty((shape[0], 1),
814
815
816
                                device=input.device,
                                dtype=torch.float32)
            torch.ops._C.dynamic_per_token_scaled_fp8_quant(
817
                output, input, scale, scale_ub)
818
819
820
        else:
            scale = torch.zeros(1, device=input.device, dtype=torch.float32)
            torch.ops._C.dynamic_scaled_fp8_quant(output, input, scale)
821
    else:
822
823
        # num_token_padding not implemented for this case
        assert (scale.numel() == 1 or num_token_padding is None)
824
        torch.ops._C.static_scaled_fp8_quant(output, input, scale)
825

826
    return output, scale
827
828


829
# int8
830
def scaled_int8_quant(
831
832
833
834
835
    input: torch.Tensor,
    scale: Optional[torch.Tensor] = None,
    azp: Optional[torch.Tensor] = None,
    symmetric: bool = True
) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
836
    """
837
    Quantize the input tensor to int8 and return the quantized tensor and scale, and maybe azp.
838
839
840

    Args:
        input: The input tensor to be quantized to int8.
841
842
        scale: Optional scaling factor for the int8 quantization.
            When not provided, we invoke dynamic-per-token quantization.
843
844
845
        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).
846
847

    Returns:
848
      Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]] : Output int8 tensor, scales, and optionally azp.
849
    """
850
851
852
    output = torch.empty_like(input, dtype=torch.int8)
    if scale is not None:
        # static-per-tensor quantization.
853
854
855
856
        assert symmetric == (
            azp is
            None), "azp must only be provided for asymmetric quantization."
        torch.ops._C.static_scaled_int8_quant(output, input, scale, azp)
857
        return output, scale, azp
858
859
860
861
862

    # dynamic-per-token quantization.
    input_scales = torch.empty((input.numel() // input.shape[-1], 1),
                               device=input.device,
                               dtype=torch.float32)
863
864
865
866
867
    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)
    return output, input_scales, input_azp
868
869


870
871
872
873
874
875
876
877
878
# qqq ops
def marlin_qqq_gemm(a: torch.Tensor, b_q_weight: torch.Tensor,
                    s_tok: torch.Tensor, s_ch: torch.Tensor,
                    s_group: torch.Tensor, workspace: torch.Tensor,
                    size_m: int, size_n: int, size_k: int) -> torch.Tensor:
    return torch.ops._C.marlin_qqq_gemm(a, b_q_weight, s_tok, s_ch, s_group,
                                        workspace, size_m, size_n, size_k)


879
# gguf
880
881
def ggml_dequantize(W: torch.Tensor, quant_type: int, m: int,
                    n: int) -> torch.Tensor:
882
883
884
885
886
887
888
889
    return torch.ops._C.ggml_dequantize(W, quant_type, m, n)


def ggml_mul_mat_vec_a8(
    W: torch.Tensor,
    X: torch.Tensor,
    quant_type: int,
    row: int,
890
) -> torch.Tensor:
891
892
893
894
895
896
897
898
    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,
899
) -> torch.Tensor:
900
901
902
    return torch.ops._C.ggml_mul_mat_a8(W, X, quant_type, row)


903
904
905
# mamba
def causal_conv1d_fwd(x: torch.Tensor, weight: torch.Tensor,
                      bias_: Optional[torch.Tensor],
906
907
908
909
                      conv_states: Optional[torch.Tensor],
                      query_start_loc: Optional[torch.Tensor],
                      cache_indices: Optional[torch.Tensor],
                      has_initial_state: Optional[torch.Tensor],
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
                      silu_activation: bool, pad_slot_id: int):
    torch.ops._C.causal_conv1d_fwd(x, weight, bias_, conv_states,
                                   query_start_loc, cache_indices,
                                   has_initial_state, silu_activation,
                                   pad_slot_id)


def causal_conv1d_update(x: torch.Tensor, conv_state: torch.Tensor,
                         weight: torch.Tensor, bias_: Optional[torch.Tensor],
                         silu_activation: bool,
                         cache_seqlens: Optional[torch.Tensor],
                         conv_state_indices: Optional[torch.Tensor],
                         pad_slot_id: int):
    torch.ops._C.causal_conv1d_update(x, conv_state, weight, bias_,
                                      silu_activation, cache_seqlens,
                                      conv_state_indices, pad_slot_id)


def selective_scan_fwd(u: torch.Tensor, delta: torch.Tensor, A: torch.Tensor,
                       B: torch.Tensor, C: torch.Tensor,
                       D_: Optional[torch.Tensor], z_: Optional[torch.Tensor],
                       delta_bias_: Optional[torch.Tensor],
                       delta_softplus: bool,
                       query_start_loc: Optional[torch.Tensor],
                       cache_indices: Optional[torch.Tensor],
                       has_initial_state: Optional[torch.Tensor],
                       ssm_states: torch.Tensor, pad_slot_id: int):
937
938
939
    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,
940
                                    ssm_states, pad_slot_id)
941
942


943
# moe
944
945
946
947
def moe_sum(input: torch.Tensor, output: torch.Tensor):
    torch.ops._moe_C.moe_sum(input, output)


948
949
950
951
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:
952
953
954
    torch.ops._moe_C.moe_align_block_size(topk_ids, num_experts, block_size,
                                          sorted_token_ids, experts_ids,
                                          num_tokens_post_pad)
955
956
957
958
959
960
961


def topk_softmax(topk_weights: torch.Tensor, topk_ids: torch.Tensor,
                 token_expert_indicies: torch.Tensor,
                 gating_output: float) -> None:
    torch.ops._moe_C.topk_softmax(topk_weights, topk_ids,
                                  token_expert_indicies, gating_output)
962
963


964
965
if supports_moe_ops and hasattr(torch.ops._moe_C, "marlin_gemm_moe"):

966
    @register_fake("_moe_C::marlin_gemm_moe")
967
968
969
970
    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,
971
972
                             b_zero_points: torch.Tensor, g_idx: torch.Tensor,
                             perm: torch.Tensor, workspace: torch.Tensor,
973
974
975
976
                             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,
977
978
979
980
981
982
                             apply_weights: bool) -> torch.Tensor:
        return torch.empty((size_m, topk, size_n),
                           dtype=a.dtype,
                           device=a.device)


983
984
985
986
987
988
989
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,
990
991
    k_scale: float,
    v_scale: float,
992
) -> None:
993
994
    torch.ops._C_cache_ops.reshape_and_cache(key, value, key_cache,
                                             value_cache, slot_mapping,
995
                                             kv_cache_dtype, k_scale, v_scale)
996
997


998
999
1000
1001
1002
1003
1004
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,
1005
1006
    k_scale: float,
    v_scale: float,
1007
) -> None:
1008
1009
    torch.ops._C_cache_ops.reshape_and_cache_flash(key, value, key_cache,
                                                   value_cache, slot_mapping,
1010
1011
                                                   kv_cache_dtype, k_scale,
                                                   v_scale)
1012
1013


1014
1015
def copy_blocks(key_caches: List[torch.Tensor],
                value_caches: List[torch.Tensor],
1016
                block_mapping: torch.Tensor) -> None:
1017
    torch.ops._C_cache_ops.copy_blocks(key_caches, value_caches, block_mapping)
1018
1019
1020


def swap_blocks(src: torch.Tensor, dst: torch.Tensor,
1021
                block_mapping: torch.Tensor) -> None:
1022
    torch.ops._C_cache_ops.swap_blocks(src, dst, block_mapping)
1023
1024


1025
1026
1027
1028
def convert_fp8(output: torch.Tensor,
                input: torch.Tensor,
                scale: float = 1.0,
                kv_dtype: str = "fp8") -> None:
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
    torch.ops._C_cache_ops.convert_fp8(output, input, scale, kv_dtype)


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(
        device)


# custom ar
1043
1044
1045
1046
def init_custom_ar(ipc_tensors: List[torch.Tensor], rank_data: torch.Tensor,
                   rank: int, full_nvlink: bool) -> int:
    return torch.ops._C_custom_ar.init_custom_ar(ipc_tensors, rank_data, rank,
                                                 full_nvlink)
1047
1048


1049
1050
1051
1052
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)
1053

1054
1055
1056
1057
1058
1059
1060
1061
1062

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


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


1063
1064
def register_buffer(fa: int, ipc_tensors: List[int]) -> None:
    return torch.ops._C_custom_ar.register_buffer(fa, ipc_tensors)
1065
1066


1067
def get_graph_buffer_ipc_meta(fa: int) -> Tuple[List[int], List[int]]:
1068
1069
1070
    return torch.ops._C_custom_ar.get_graph_buffer_ipc_meta(fa)


1071
def register_graph_buffers(fa: int, handles: List[List[int]],
1072
1073
                           offsets: List[List[int]]) -> None:
    torch.ops._C_custom_ar.register_graph_buffers(fa, handles, offsets)