_custom_ops.py 57.6 KB
Newer Older
1
2
# SPDX-License-Identifier: Apache-2.0

3
import contextlib
4
import importlib
5
from typing import TYPE_CHECKING, Optional, Union
6
7

import torch
8
import torch.library
9

10
import vllm.envs as envs
11
from vllm.logger import init_logger
12
from vllm.platforms import current_platform
13
from vllm.scalar_type import ScalarType
14
15
16

logger = init_logger(__name__)

17
if not current_platform.is_tpu() and not current_platform.is_hpu():
18
19
20
21
    try:
        import vllm._C
    except ImportError as e:
        logger.warning("Failed to import from vllm._C with %r", e)
22

23
supports_moe_ops = False
24
with contextlib.suppress(ImportError):
25
    import vllm._moe_C  # noqa: F401
26
    supports_moe_ops = True
27

28
if TYPE_CHECKING:
29
30
31
32
33
34
35
36
37

    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

38
39
40
41
42
43
44
45
46
47

# 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,
48
    seq_lens: torch.Tensor,
49
    block_size: int,
50
    max_seq_len: int,
51
52
    alibi_slopes: Optional[torch.Tensor],
    kv_cache_dtype: str,
53
54
    k_scale: torch.Tensor,
    v_scale: torch.Tensor,
55
56
57
58
59
    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,
60
) -> None:
61
    torch.ops._C.paged_attention_v1(
62
63
        out, query, key_cache, value_cache, num_kv_heads, scale, block_tables,
        seq_lens, block_size, max_seq_len, alibi_slopes, kv_cache_dtype,
64
65
66
        k_scale, v_scale, tp_rank, blocksparse_local_blocks,
        blocksparse_vert_stride, blocksparse_block_size,
        blocksparse_head_sliding_step)
67
68
69
70
71
72
73
74
75
76
77
78
79


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,
80
    seq_lens: torch.Tensor,
81
    block_size: int,
82
    max_seq_len: int,
83
84
    alibi_slopes: Optional[torch.Tensor],
    kv_cache_dtype: str,
85
86
    k_scale: torch.Tensor,
    v_scale: torch.Tensor,
87
88
89
90
91
    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,
92
) -> None:
93
    torch.ops._C.paged_attention_v2(
94
95
        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,
96
        alibi_slopes, kv_cache_dtype, k_scale, v_scale, tp_rank,
97
98
        blocksparse_local_blocks, blocksparse_vert_stride,
        blocksparse_block_size, blocksparse_head_sliding_step)
99
100


101
102
103
104
105
106
107
108
109
110
111
112
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,
113
    query_start_loc: Optional[torch.Tensor],
114
115
116
117
    block_size: int,
    max_seq_len: int,
    alibi_slopes: Optional[torch.Tensor],
    kv_cache_dtype: str,
118
119
    k_scale: torch.Tensor,
    v_scale: torch.Tensor,
120
121
122
123
) -> 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,
124
125
126
                                      query_start_loc, block_size, max_seq_len,
                                      alibi_slopes, kv_cache_dtype, k_scale,
                                      v_scale)
127
128


Thien Tran's avatar
Thien Tran committed
129
130
131
132
133
134
135
136
137
138
139
140
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:
    torch.ops._C_cpu.mla_decode_kvcache(out, query, kv_cache, scale,
                                        block_tables, seq_lens)


141
142
143
144
145
146
147
148
149
# 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:
150
151
    torch.ops._C.rotary_embedding(positions, query, key, head_size,
                                  cos_sin_cache, is_neox)
152
153
154
155
156
157
158


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:
159
160
161
    torch.ops._C.batched_rotary_embedding(positions, query, key, head_size,
                                          cos_sin_cache, is_neox, rot_dim,
                                          cos_sin_cache_offsets)
162
163
164
165
166


# layer norm ops
def rms_norm(out: torch.Tensor, input: torch.Tensor, weight: torch.Tensor,
             epsilon: float) -> None:
167
    torch.ops._C.rms_norm(out, input, weight, epsilon)
168
169
170
171


def fused_add_rms_norm(input: torch.Tensor, residual: torch.Tensor,
                       weight: torch.Tensor, epsilon: float) -> None:
172
    torch.ops._C.fused_add_rms_norm(input, residual, weight, epsilon)
173
174


175
176
177
178
179
180
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:
181
    """Advance a step on GPU for existing inputs for a multi-step runner"""
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
    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)
205
206


207
208
209
210
211
212
213
214
# 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
215
) -> tuple[torch.Tensor, torch.Tensor]:
216
217
218
219
220
221
222
223
224
225
226
    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


227
228
229
230
231
# 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:
232
233
234
235
    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)
236
237
    return torch.ops._C.awq_dequantize(qweight, scales, zeros, split_k_iters,
                                       thx, thy)
238
239
240
241


def awq_gemm(input: torch.Tensor, qweight: torch.Tensor, qzeros: torch.Tensor,
             scales: torch.Tensor, split_k_iters: int) -> torch.Tensor:
242
243
244
245
    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)
246
    return torch.ops._C.awq_gemm(input, qweight, qzeros, scales, split_k_iters)
247
248
249
250
251
252
253


# 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:
254
255
    return torch.ops._C.gptq_gemm(a, b_q_weight, b_gptq_qzeros, b_gptq_scales,
                                  b_g_idx, use_exllama, bit)
256
257


258
if hasattr(torch.ops._C, "gptq_gemm"):
259

260
    @register_fake("_C::gptq_gemm")
261
262
263
264
265
266
267
268
269
    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)


270
271
def gptq_shuffle(q_weight: torch.Tensor, q_perm: torch.Tensor,
                 bit: int) -> None:
272
    torch.ops._C.gptq_shuffle(q_weight, q_perm, bit)
273
274
275
276
277
278


# 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:
279
280
    return torch.ops._C.marlin_gemm(a, b_q_weight, b_scales, workspace, size_m,
                                    size_n, size_k)
281
282


283
284
285
# marlin_24
def gptq_marlin_24_gemm(a: torch.Tensor, b_q_weight: torch.Tensor,
                        b_meta: torch.Tensor, b_scales: torch.Tensor,
286
287
                        workspace: torch.Tensor, b_q_type: ScalarType,
                        size_m: int, size_n: int, size_k: int) -> torch.Tensor:
288
    return torch.ops._C.gptq_marlin_24_gemm(a, b_q_weight, b_meta, b_scales,
289
                                            workspace, b_q_type.id, size_m,
290
                                            size_n, size_k)
291
292


293
if hasattr(torch.ops._C, "gptq_marlin_24_gemm"):
294

295
    @register_fake("_C::gptq_marlin_24_gemm")
296
297
298
    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,
299
300
301
                                  b_q_type: ScalarType, size_m: torch.SymInt,
                                  size_n: torch.SymInt,
                                  size_k: torch.SymInt) -> torch.Tensor:
302
303
        return torch.empty((size_m, size_n), device=a.device, dtype=a.dtype)

304
    @register_fake("_C::gptq_marlin_gemm")
305
306
307
308
309
310
311
312
    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,
313
314
315
                               size_m: torch.SymInt,
                               size_n: torch.SymInt,
                               size_k: torch.SymInt,
316
317
                               is_k_full: bool,
                               has_zp: bool = False,
318
                               use_atomic_add: bool = False,
319
320
                               use_fp32_reduce: bool = False,
                               is_zp_float: bool = False) -> torch.Tensor:
321
322
        return torch.empty((size_m, size_n), device=a.device, dtype=a.dtype)

323
    @register_fake("_C::marlin_qqq_gemm")
324
325
326
    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,
327
328
                              size_m: torch.SymInt, size_n: torch.SymInt,
                              size_k: torch.SymInt) -> torch.Tensor:
329
330
331
332
        return torch.empty((size_m, size_n),
                           dtype=torch.float16,
                           device=a.device)

333
    @register_fake("_C::marlin_gemm")
334
335
    def _marlin_gemm_fake(a: torch.Tensor, b_q_weight: torch.Tensor,
                          b_scales: torch.Tensor, workspace: torch.Tensor,
336
337
                          size_m: torch.SymInt, size_n: torch.SymInt,
                          size_k: torch.SymInt) -> torch.Tensor:
338
339
340
341
        return torch.empty((size_m, size_n),
                           dtype=torch.float16,
                           device=a.device)

342
    @register_fake("_C::awq_dequantize")
343
    def _awq_dequantize_fake(qweight: torch.Tensor, scales: torch.Tensor,
344
345
                             zeros: torch.Tensor, split_k_iters: torch.SymInt,
                             thx: int, thy: int) -> torch.Tensor:
346
347
348
349
350
351
352
        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)

353
    @register_fake("_C::awq_gemm")
354
355
    def _awq_gemm_fake(input: torch.Tensor, qweight: torch.Tensor,
                       qzeros: torch.Tensor, scales: torch.Tensor,
356
                       split_k_iters: torch.SymInt) -> torch.Tensor:
357
358
359
360
361
        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)

362
    @register_fake("_C::aqlm_gemm")
363
364
    def _aqlm_gemm_fake(input: torch.Tensor, codes: torch.Tensor,
                        codebooks: torch.Tensor, scales: torch.Tensor,
365
                        codebook_partition_sizes: list[int],
366
367
368
369
370
371
372
373
374
375
376
377
                        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))

378
    @register_fake("_C::aqlm_dequant")
379
380
    def _aqlm_dequant_fake(
            codes: torch.Tensor, codebooks: torch.Tensor,
381
            codebook_partition_sizes: list[int]) -> torch.Tensor:
382
383
384
385
386
387
        in_features = codes.size(1) * 8
        out_features = codes.size(0)
        return torch.empty((out_features, in_features),
                           dtype=codebooks.dtype,
                           device=codebooks.device)

388
    @register_fake("_C::fp8_marlin_gemm")
389
390
    def _fp8_marlin_gemm_fake(a: torch.Tensor, b_q_weight: torch.Tensor,
                              b_scales: torch.Tensor, workspace: torch.Tensor,
391
392
393
                              num_bits: int, size_m: torch.SymInt,
                              size_n: torch.SymInt,
                              size_k: torch.SymInt) -> torch.Tensor:
394
395
        return torch.empty((size_m, size_n), dtype=a.dtype, device=a.device)

396
397
    @register_fake("_C::machete_mm")
    def machete_mm_fake(
398
        a: torch.Tensor,
399
        # b_q Should be the tensor returned by machete_prepack_B
400
        b_q: torch.Tensor,
401
        b_type: ScalarType,
402
403
404
        out_type: Optional[torch.dtype] = None,
        b_group_scales: Optional[torch.Tensor] = None,
        b_group_zeros: Optional[torch.Tensor] = None,
405
        b_group_size: Optional[int] = None,
406
407
        b_channel_scales: Optional[torch.Tensor] = None,
        a_token_scales: Optional[torch.Tensor] = None,
408
409
410
411
412
413
        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)

414
    @register_fake("_C::machete_prepack_B")
415
416
417
    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:
418
419
        return torch.empty_like(b_q_weight,
                                memory_format=torch.contiguous_format)
420
421


422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
if hasattr(torch.ops._C, "allspark_w8a16_gemm"):

    @register_fake("_C::allspark_w8a16_gemm")
    def _allspark_w8a16_gemm_fake(a: torch.Tensor, b_qweight: torch.Tensor,
                                  b_scales: torch.Tensor,
                                  b_qzeros: Optional[torch.Tensor],
                                  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:
        m = a.size(0)
        return torch.empty((m, n), device=a.device, dtype=a.dtype)


438
439
440
if hasattr(torch.ops._C, "ggml_dequantize"):

    @register_fake("_C::ggml_dequantize")
441
442
443
444
445
446
    def _ggml_dequantize_fake(
            W: torch.Tensor,
            quant_type: int,
            m: torch.SymInt,
            n: torch.SymInt,
            dtype: Optional[torch.dtype] = None) -> torch.Tensor:
447
448
449
450
451
452
453
454
455
        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:
456
        return torch.empty((1, row), dtype=X.dtype, device=W.device)
457
458
459
460
461
462
463
464
465

    @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)
466
        return torch.empty((batch, row), dtype=X.dtype, device=W.device)
467

468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
    @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)
        return torch.empty((tokens * top_k, row),
                           dtype=torch.float16,
                           device=W.device)

485

486
# cutlass
487
488
489
490
def cutlass_scaled_mm_supports_fp4(cuda_device_capability: int) -> bool:
    return torch.ops._C.cutlass_scaled_mm_supports_fp4(cuda_device_capability)


491
492
493
494
495
496
497
498
499
500
501
502
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:
    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)
    torch.ops._C.cutlass_scaled_fp4_mm(out, a, b, block_scale_a, block_scale_b,
                                       alpha)
    return out


503
504
505
506
def cutlass_scaled_mm_supports_fp8(cuda_device_capability: int) -> bool:
    return torch.ops._C.cutlass_scaled_mm_supports_fp8(cuda_device_capability)


507
508
509
510
511
def cutlass_scaled_mm_supports_block_fp8(cuda_device_capability: int) -> bool:
    return torch.ops._C.cutlass_scaled_mm_supports_block_fp8(
        cuda_device_capability)


512
513
514
def cutlass_scaled_mm(a: torch.Tensor,
                      b: torch.Tensor,
                      scale_a: torch.Tensor,
515
                      scale_b: torch.Tensor,
516
                      out_dtype: torch.dtype,
517
                      bias: Optional[torch.Tensor] = None) -> torch.Tensor:
518
    """
519
    `cutlass_scaled_mm` implements a fused version of
520
        `output = torch.mm((scale_a * a), (scale_b * b)).to(out_dtype)`
521
522
523
524
525
526
527
528
    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
529
530
531
532
533
534
535
536
537
538
539
        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
    """
540
541
    assert (b.shape[0] % 16 == 0 and b.shape[1] % 16 == 0)
    assert (out_dtype is torch.bfloat16 or out_dtype is torch.float16)
542
543
    assert bias is None or bias.shape[0] == b.shape[
        1] and bias.dtype == out_dtype
544
545
546

    m = a.shape[0]
    n = b.shape[1]
547
548
549
550
551
552
553
554

    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)

555
556
    out = torch.empty((m, n), dtype=out_dtype, device=a.device)

557
558
    torch.ops._C.cutlass_scaled_mm(out, a, b, scale_a, scale_b, bias)

559
560
561
    return out


562
563
564
565
566
567
568
569
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:
570
571
572
573
574
    """
    :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.
    """
575
576
577
578
    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
579
    assert azp is None or azp.numel() == a.shape[0]
580
581
582
583
584
585
586
587
588
589

    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


590
591
592
593
594
def cutlass_sparse_scaled_mm_supported(cuda_device_capability: int) -> bool:
    return torch.ops._C.cutlass_sparse_scaled_mm_supported(
        cuda_device_capability)


595
596
597
def cutlass_group_gemm_supported(cuda_device_capability: int) -> bool:
    return torch.ops._C.cutlass_group_gemm_supported(cuda_device_capability)

598
def cutlass_sparse_compress(a: torch.Tensor) \
599
    -> tuple[torch.Tensor, torch.Tensor]:
600
601
602
603
604
605
606
607
    """
    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:
608
        a (torch.Tensor):
609
610
611
612
613
614
615
            The input tensor to be compressed. Must have one of the following data types:
            - `torch.int8`
            - `torch.float8_e4m3fn`
            - `torch.bfloat16`
            - `torch.float16`

    Returns:
616
        tuple[torch.Tensor, torch.Tensor]:
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
            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
637
    assert (a.shape[1] % (2 * elemsPerMetaElem) == 0)
638

639
    return torch.ops._C.cutlass_sparse_compress(a)
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687


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


688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
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):
    """
    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.
    """
    torch.ops._C.get_cutlass_moe_mm_data(topk_ids, expert_offsets,
                                         problem_sizes1, problem_sizes2,
                                         input_permutation, output_permutation,
                                         num_experts, n, k)


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):
    """
    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.
    """
    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)


738
739
740
# aqlm
def aqlm_gemm(input: torch.Tensor, codes: torch.Tensor,
              codebooks: torch.Tensor, scales: torch.Tensor,
741
              codebook_partition_sizes: list[int],
742
              bias: Optional[torch.Tensor]) -> torch.Tensor:
743
744
    return torch.ops._C.aqlm_gemm(input, codes, codebooks, scales,
                                  codebook_partition_sizes, bias)
745
746
747


def aqlm_dequant(codes: torch.Tensor, codebooks: torch.Tensor,
748
                 codebook_partition_sizes: list[int]) -> torch.Tensor:
749
750
    return torch.ops._C.aqlm_dequant(codes, codebooks,
                                     codebook_partition_sizes)
751
752


753
754
# gptq_marlin
def gptq_marlin_repack(b_q_weight: torch.Tensor, perm: torch.Tensor,
755
756
                       size_k: int, size_n: int,
                       num_bits: int) -> torch.Tensor:
757
758
    return torch.ops._C.gptq_marlin_repack(b_q_weight, perm, size_k, size_n,
                                           num_bits)
759
760


761
762
763
764
765
766
# 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)


767
768
769
770
771
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
772
    output = torch.empty((num_experts, size_k // 16, size_n * (num_bits // 2)),
773
774
775
776
777
778
779
780
                         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


781
782
783
784
785
786
787
788
789
790
791
792
793
794
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


795
796
797
798
799
800
801
802
803
804
805
806
807
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,
808
                     use_atomic_add: bool = False,
809
810
                     use_fp32_reduce: bool = False,
                     is_zp_float: bool = False) -> torch.Tensor:
811
    return torch.ops._C.gptq_marlin_gemm(a, b_q_weight, b_scales, b_zeros,
812
                                         g_idx, perm, workspace, b_q_type.id,
813
                                         size_m, size_n, size_k, is_k_full,
814
815
                                         has_zp, use_atomic_add,
                                         use_fp32_reduce, is_zp_float)
816
817


818
819
820
821
822
823
824
825
826
# 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)


827
# machete
828
829
830
831
832
833
834
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,
835
        out_type: Optional[torch.dtype] = None) -> list[str]:
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
    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)
863
864


865
if hasattr(torch.ops._C, "permute_cols"):
866

867
    @register_fake("_C::permute_cols")
868
869
870
871
872
873
874
875
876
    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)


877
878
879
# fp4
def scaled_fp4_quant(
        input: torch.Tensor,
880
        input_global_scale: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
881
882
883
884
885
886
887
888
889
890
891
892
893
894
    """
    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:
895
        tuple[torch.Tensor, torch.Tensor]: The output tensor in FP4 but every
896
897
898
            two values are packed into a uint8 and float8_e4m3 scaling factors
            in the sizzled layout.
    """
899
    assert not current_platform.is_rocm()
900
901
902
903
904
905
906
907
908
909
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
    assert input.ndim >= 1, (
        f'input.ndim needs to be >= 1, but got {input.ndim}.')
    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

    assert n % block_size == 0, (
        f'last dim has to be multiple of 16, but got {n}.')
    assert input.dtype in (torch.float16, torch.bfloat16), (
        f'input.dtype needs to be fp16 or bf16 but got {input.dtype}.')

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

    torch.ops._C.scaled_fp4_quant(output, input, output_scale,
                                  input_global_scale)
    output_scale = output_scale.view(torch.float8_e4m3fn)
    return output, output_scale


935
# fp8
936
937
938
def scaled_fp8_quant(
    input: torch.Tensor,
    scale: Optional[torch.Tensor] = None,
939
    num_token_padding: Optional[int] = None,
940
    scale_ub: Optional[torch.Tensor] = None,
941
    use_per_token_if_dynamic: bool = False,
942
) -> tuple[torch.Tensor, torch.Tensor]:
943
944
945
946
947
948
    """
    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
949
    optional padding of the output tensors for downstream kernels that
950
951
952
953
954
    will benefit from padding.

    Args:
        input: The input tensor to be quantized to FP8
        scale: Optional scaling factor for the FP8 quantization
955
        scale_ub: Optional upper bound for scaling factor in dynamic
956
            per token case
957
        num_token_padding: If specified, pad the first dimension
958
            of the output to at least this value.
959
        use_per_token_if_dynamic: Whether to do per_tensor or per_token
960
            in the dynamic quantization case.
961
962

    Returns:
963
        tuple[torch.Tensor, torch.Tensor]: The output tensor in FP8 and
964
965
            scaling factor.
    """
966
967
    # This code assumes batch_dim and num_tokens are flattened
    assert (input.ndim == 2)
968
    shape: Union[tuple[int, int], torch.Size] = input.shape
969
970
    # For ROCm on MI300, the output fp8 dtype is torch.float_e3m3fnuz
    out_dtype: torch.dtype = current_platform.fp8_dtype()
971
972
    if num_token_padding:
        shape = (max(num_token_padding, input.shape[0]), shape[1])
973
    output = torch.empty(shape, device=input.device, dtype=out_dtype)
974

975
    if scale is None:
976
        if use_per_token_if_dynamic:
977
            scale = torch.empty((shape[0], 1),
978
979
980
                                device=input.device,
                                dtype=torch.float32)
            torch.ops._C.dynamic_per_token_scaled_fp8_quant(
981
                output, input, scale, scale_ub)
982
983
984
        else:
            scale = torch.zeros(1, device=input.device, dtype=torch.float32)
            torch.ops._C.dynamic_scaled_fp8_quant(output, input, scale)
985
    else:
986
987
        # num_token_padding not implemented for this case
        assert (scale.numel() == 1 or num_token_padding is None)
988
        torch.ops._C.static_scaled_fp8_quant(output, input, scale)
989

990
    return output, scale
991
992


993
994
995
996
997
998
# gptq allspark
def allspark_repack_weight(
        qweight: torch.Tensor,
        scale: torch.Tensor,
        zero_point: Optional[torch.Tensor] = None,
        has_zp: bool = False
999
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
1000
    """
1001
    Rearrange qweight, scale, and zero_point(if asymmetric) to n32k16 format
1002
1003
1004
1005
1006
1007
1008
1009
    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.
1010
1011
            if use asymmetric quantization, has_zp = True.

1012
    Returns:
1013
        tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]] :
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
            rearranged weight, scale, and optionally zero_point.
    """
    K = qweight.shape[0]
    N = qweight.shape[1]
    N_32align = (N + 32 - 1) // 32 * 32

    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)
    zero_point_reorder = None
    if has_zp:
        assert zero_point is not None, (
            "zero_point must be provided for asymmetric quantization.")
        zero_point_reorder = torch.empty((1, N_32align),
                                         device=zero_point.device,
                                         dtype=zero_point.dtype)

    torch.ops._C.rearrange_kn_weight_as_n32k16_order(
        qweight, scale, zero_point, has_zp, qweight_reorder, scale_reorder,
        zero_point_reorder, K, N, N_32align)

    return qweight_reorder, scale_reorder, zero_point_reorder


def allspark_w8a16_gemm(a: torch.Tensor, b_qweight: torch.Tensor,
                        b_scales: torch.Tensor,
                        b_qzeros: Optional[torch.Tensor], 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)


1054
# int8
1055
def scaled_int8_quant(
1056
1057
1058
1059
    input: torch.Tensor,
    scale: Optional[torch.Tensor] = None,
    azp: Optional[torch.Tensor] = None,
    symmetric: bool = True
1060
) -> tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
1061
    """
1062
    Quantize the input tensor to int8 and return the quantized tensor and scale, and maybe azp.
1063
1064
1065

    Args:
        input: The input tensor to be quantized to int8.
1066
1067
        scale: Optional scaling factor for the int8 quantization.
            When not provided, we invoke dynamic-per-token quantization.
1068
1069
1070
        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).
1071
1072

    Returns:
1073
      tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]] : Output int8 tensor, scales, and optionally azp.
1074
    """
1075
1076
1077
    output = torch.empty_like(input, dtype=torch.int8)
    if scale is not None:
        # static-per-tensor quantization.
1078
        assert symmetric == (
1079
1080
            azp
            is None), "azp must only be provided for asymmetric quantization."
1081
        torch.ops._C.static_scaled_int8_quant(output, input, scale, azp)
1082
        return output, scale, azp
1083
1084
1085
1086
1087

    # dynamic-per-token quantization.
    input_scales = torch.empty((input.numel() // input.shape[-1], 1),
                               device=input.device,
                               dtype=torch.float32)
1088
1089
1090
1091
1092
    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
1093
1094


1095
1096
1097
1098
1099
1100
1101
1102
1103
# 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)


1104
# gguf
1105
1106
1107
def ggml_dequantize(W: torch.Tensor, quant_type: int, m: int, n: int,
                    dtype: Optional[torch.dtype]) -> torch.Tensor:
    return torch.ops._C.ggml_dequantize(W, quant_type, m, n, dtype)
1108
1109
1110
1111
1112
1113
1114


def ggml_mul_mat_vec_a8(
    W: torch.Tensor,
    X: torch.Tensor,
    quant_type: int,
    row: int,
1115
) -> torch.Tensor:
1116
1117
1118
1119
1120
1121
1122
1123
    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,
1124
) -> torch.Tensor:
1125
1126
1127
    return torch.ops._C.ggml_mul_mat_a8(W, X, quant_type, row)


1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
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:
    return torch.ops._C.ggml_moe_a8(X, W, sorted_token_ids, expert_ids,
                                    num_tokens_post_padded, quant_type, row,
                                    top_k, tokens)


def ggml_moe_get_block_size(quant_type: int) -> int:
    return torch.ops._C.ggml_moe_get_block_size(quant_type)


1148
1149
1150
# mamba
def causal_conv1d_fwd(x: torch.Tensor, weight: torch.Tensor,
                      bias_: Optional[torch.Tensor],
1151
1152
1153
1154
                      conv_states: Optional[torch.Tensor],
                      query_start_loc: Optional[torch.Tensor],
                      cache_indices: Optional[torch.Tensor],
                      has_initial_state: Optional[torch.Tensor],
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
                      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):
1182
1183
1184
    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,
1185
                                    ssm_states, pad_slot_id)
1186
1187


1188
# moe
1189
1190
1191
1192
def moe_sum(input: torch.Tensor, output: torch.Tensor):
    torch.ops._moe_C.moe_sum(input, output)


1193
1194
1195
1196
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:
1197
1198
1199
    torch.ops._moe_C.moe_align_block_size(topk_ids, num_experts, block_size,
                                          sorted_token_ids, experts_ids,
                                          num_tokens_post_pad)
1200
1201


1202
1203
1204
1205
1206
1207
1208
1209
1210
def sgl_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.sgl_moe_align_block_size(topk_ids, num_experts,
                                              block_size, sorted_token_ids,
                                              experts_ids, num_tokens_post_pad)


1211
1212
1213
1214
1215
1216
1217
1218
def moe_wna16_gemm(input: torch.Tensor, output: torch.Tensor,
                   b_qweight: torch.Tensor, b_scales: torch.Tensor,
                   b_qzeros: Optional[torch.Tensor],
                   topk_weights: Optional[torch.Tensor],
                   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:
1219
1220
1221
1222
    if not current_platform.is_cuda():
        raise NotImplementedError(
            "The optimized moe_wna16_gemm kernel is only "
            "available on CUDA platforms")
1223
1224
1225
1226
1227
1228
1229
    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)


1230
1231
def topk_softmax(topk_weights: torch.Tensor, topk_ids: torch.Tensor,
                 token_expert_indicies: torch.Tensor,
1232
                 gating_output: torch.Tensor) -> None:
1233
1234
    torch.ops._moe_C.topk_softmax(topk_weights, topk_ids,
                                  token_expert_indicies, gating_output)
1235
1236


1237
1238
if supports_moe_ops and hasattr(torch.ops._moe_C, "marlin_gemm_moe"):

1239
    @register_fake("_moe_C::marlin_gemm_moe")
1240
1241
1242
1243
    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,
1244
1245
                             b_zero_points: torch.Tensor, g_idx: torch.Tensor,
                             perm: torch.Tensor, workspace: torch.Tensor,
1246
1247
1248
1249
                             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,
1250
1251
1252
1253
1254
1255
                             apply_weights: bool) -> torch.Tensor:
        return torch.empty((size_m, topk, size_n),
                           dtype=a.dtype,
                           device=a.device)


1256
1257
1258
1259
1260
1261
1262
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,
1263
1264
    k_scale: torch.Tensor,
    v_scale: torch.Tensor,
1265
) -> None:
1266
1267
    torch.ops._C_cache_ops.reshape_and_cache(key, value, key_cache,
                                             value_cache, slot_mapping,
1268
                                             kv_cache_dtype, k_scale, v_scale)
1269
1270


1271
1272
1273
1274
1275
1276
1277
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,
1278
1279
    k_scale: torch.Tensor,
    v_scale: torch.Tensor,
1280
) -> None:
1281
1282
    torch.ops._C_cache_ops.reshape_and_cache_flash(key, value, key_cache,
                                                   value_cache, slot_mapping,
1283
1284
                                                   kv_cache_dtype, k_scale,
                                                   v_scale)
1285
1286


1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
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:
    torch.ops._C_cache_ops.concat_and_cache_mla(kv_c, k_pe, kv_cache,
                                                slot_mapping, kv_cache_dtype,
                                                scale)


1300
1301
def copy_blocks(key_caches: list[torch.Tensor],
                value_caches: list[torch.Tensor],
1302
                block_mapping: torch.Tensor) -> None:
1303
    torch.ops._C_cache_ops.copy_blocks(key_caches, value_caches, block_mapping)
1304
1305


1306
def copy_blocks_mla(kv_caches: list[torch.Tensor],
1307
1308
1309
1310
                    block_mapping: torch.Tensor) -> None:
    torch.ops._C_cache_ops.copy_blocks_mla(kv_caches, block_mapping)


1311
def swap_blocks(src: torch.Tensor, dst: torch.Tensor,
1312
                block_mapping: torch.Tensor) -> None:
1313
    torch.ops._C_cache_ops.swap_blocks(src, dst, block_mapping)
1314
1315


1316
1317
1318
1319
def convert_fp8(output: torch.Tensor,
                input: torch.Tensor,
                scale: float = 1.0,
                kv_dtype: str = "fp8") -> None:
1320
1321
1322
    torch.ops._C_cache_ops.convert_fp8(output, input, scale, kv_dtype)


1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
def gather_cache(src_cache: torch.Tensor,
                 dst: torch.Tensor,
                 block_table: torch.Tensor,
                 cu_seq_lens: torch.Tensor,
                 batch_size: int,
                 seq_starts: Optional[torch.Tensor] = None) -> None:
    torch.ops._C_cache_ops.gather_cache(src_cache, dst, block_table,
                                        cu_seq_lens, batch_size, seq_starts)


1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
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
1344
def init_custom_ar(ipc_tensors: list[torch.Tensor], rank_data: torch.Tensor,
1345
                   rank: int, fully_connected: bool) -> int:
1346
    return torch.ops._C_custom_ar.init_custom_ar(ipc_tensors, rank_data, rank,
1347
                                                 fully_connected)
1348
1349


1350
1351
1352
1353
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)
1354

1355
1356
1357
1358
1359
1360
1361
1362
1363

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


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


1364
def register_buffer(fa: int, ipc_tensors: list[int]) -> None:
1365
    return torch.ops._C_custom_ar.register_buffer(fa, ipc_tensors)
1366
1367


1368
def get_graph_buffer_ipc_meta(fa: int) -> tuple[list[int], list[int]]:
1369
1370
1371
    return torch.ops._C_custom_ar.get_graph_buffer_ipc_meta(fa)


1372
1373
def register_graph_buffers(fa: int, handles: list[list[int]],
                           offsets: list[list[int]]) -> None:
1374
    torch.ops._C_custom_ar.register_graph_buffers(fa, handles, offsets)
1375
1376


1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
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)


1389
1390
1391
1392
def get_flash_mla_metadata(
    cache_seqlens: torch.Tensor,
    num_heads_per_head_k: int,
    num_heads_k: int,
1393
) -> tuple[torch.Tensor, torch.Tensor]:
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
    """
    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.
    """
    return torch.ops._C.get_flash_mla_metadata(cache_seqlens,
                                               num_heads_per_head_k,
                                               num_heads_k)


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,
    softmax_scale: Optional[float] = None,
    causal: bool = False,
1419
) -> tuple[torch.Tensor, torch.Tensor]:
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
    """
    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:
        softmax_scale = q.shape[-1]**(-0.5)
    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