_custom_ops.py 80.8 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
from typing import TYPE_CHECKING, Optional, Union
5
6
7

import torch

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

logger = init_logger(__name__)

15
16
current_platform.import_core_kernels()
supports_moe_ops = current_platform.try_import_moe_kernels()
17

18
if TYPE_CHECKING:
19
20
21
22
23
24
25
26
27

    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

28
29
30
31
32
33
34
35
36
37

# 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,
38
    seq_lens: torch.Tensor,
39
    block_size: int,
40
    max_seq_len: int,
41
42
    alibi_slopes: Optional[torch.Tensor],
    kv_cache_dtype: str,
43
44
    k_scale: torch.Tensor,
    v_scale: torch.Tensor,
45
46
47
48
49
    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,
50
) -> None:
51
    torch.ops._C.paged_attention_v1(
52
53
        out, query, key_cache, value_cache, num_kv_heads, scale, block_tables,
        seq_lens, block_size, max_seq_len, alibi_slopes, kv_cache_dtype,
54
55
56
        k_scale, v_scale, tp_rank, blocksparse_local_blocks,
        blocksparse_vert_stride, blocksparse_block_size,
        blocksparse_head_sliding_step)
57
58
59
60
61
62
63
64
65
66
67
68
69


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,
70
    seq_lens: torch.Tensor,
71
    block_size: int,
72
    max_seq_len: int,
73
74
    alibi_slopes: Optional[torch.Tensor],
    kv_cache_dtype: str,
75
76
    k_scale: torch.Tensor,
    v_scale: torch.Tensor,
77
78
79
80
81
    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,
82
) -> None:
83
    torch.ops._C.paged_attention_v2(
84
85
        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,
86
        alibi_slopes, kv_cache_dtype, k_scale, v_scale, tp_rank,
87
88
        blocksparse_local_blocks, blocksparse_vert_stride,
        blocksparse_block_size, blocksparse_head_sliding_step)
89
90


91
92
93
94
95
96
97
98
99
100
101
102
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,
103
    query_start_loc: Optional[torch.Tensor],
104
105
106
107
    block_size: int,
    max_seq_len: int,
    alibi_slopes: Optional[torch.Tensor],
    kv_cache_dtype: str,
108
109
    k_scale: torch.Tensor,
    v_scale: torch.Tensor,
110
    fp8_out_scale: Optional[torch.Tensor] = None,
xiao-llm's avatar
xiao-llm committed
111
    mfma_type: str = "fp8" if envs.VLLM_ROCM_FP8_MFMA_PAGE_ATTN else "f16",
112
113
114
115
) -> 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,
116
117
                                      query_start_loc, block_size, max_seq_len,
                                      alibi_slopes, kv_cache_dtype, k_scale,
xiao-llm's avatar
xiao-llm committed
118
                                      v_scale, fp8_out_scale, mfma_type)
119
120


Thien Tran's avatar
Thien Tran committed
121
122
123
124
125
126
127
128
129
130
131
132
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)


133
134
135
136
137
138
139
140
141
142
143
# merge attn states ops
def merge_attn_states(output: torch.Tensor,
                      prefix_output: torch.Tensor,
                      prefix_lse: torch.Tensor,
                      suffix_output: torch.Tensor,
                      suffix_lse: torch.Tensor,
                      output_lse: Optional[torch.Tensor] = None) -> None:
    torch.ops._C.merge_attn_states(output, output_lse, prefix_output,
                                   prefix_lse, suffix_output, suffix_lse)


144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
def convert_vertical_slash_indexes(
    q_seqlens: torch.Tensor,  # [BATCH, ]
    kv_seqlens: torch.Tensor,  # [BATCH, ]
    vertical_indexes: torch.Tensor,  # [BATCH, N_HEADS, NNZ_V]
    slash_indexes: torch.Tensor,  # [BATCH, N_HEADS, NNZ_S]
    context_size: int,
    block_size_M: int,
    block_size_N: int,
    causal: bool = True,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
    batch_size = slash_indexes.size(0)
    num_heads = slash_indexes.size(1)
    nnz_slash = slash_indexes.size(2)
    nnz_vertical = vertical_indexes.size(2)
    num_rows = (context_size + block_size_M - 1) // block_size_M

    block_count = torch.zeros(batch_size,
                              num_heads,
                              num_rows,
                              dtype=q_seqlens.dtype,
                              device=q_seqlens.device)
    block_offset = torch.zeros(batch_size,
                               num_heads,
                               num_rows,
                               nnz_slash,
                               dtype=q_seqlens.dtype,
                               device=q_seqlens.device)
    column_count = torch.zeros(batch_size,
                               num_heads,
                               num_rows,
                               dtype=q_seqlens.dtype,
                               device=q_seqlens.device)
    column_index = torch.zeros(batch_size,
                               num_heads,
                               num_rows,
                               nnz_vertical,
                               dtype=q_seqlens.dtype,
                               device=q_seqlens.device)

    torch.ops._C.convert_vertical_slash_indexes(
        block_count, block_offset, column_count, column_index, q_seqlens,
        kv_seqlens, vertical_indexes, slash_indexes, context_size,
        block_size_M, block_size_N, causal)
    return block_count, block_offset, column_count, column_index


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

    block_count = torch.empty(batch_size,
                              num_heads,
                              num_rows,
                              dtype=q_seqlens.dtype,
                              device=q_seqlens.device)
    block_offset = torch.empty(batch_size,
                               num_heads,
                               num_rows,
                               nnz_slash,
                               dtype=q_seqlens.dtype,
                               device=q_seqlens.device)
    column_count = torch.empty(batch_size,
                               num_heads,
                               num_rows,
                               dtype=q_seqlens.dtype,
                               device=q_seqlens.device)
    column_index = torch.empty(batch_size,
                               num_heads,
                               num_rows,
                               nnz_vertical,
                               dtype=q_seqlens.dtype,
                               device=q_seqlens.device)

    torch.ops._C.convert_vertical_slash_indexes_mergehead(
        block_count, block_offset, column_count, column_index, q_seqlens,
        kv_seqlens, vertical_indexes, slash_indexes, vertical_indices_count,
        slash_indices_count, context_size, block_size_M, block_size_N, causal)
    return block_count, block_offset, column_count, column_index


239
240
241
242
# pos encoding ops
def rotary_embedding(
    positions: torch.Tensor,
    query: torch.Tensor,
243
    key: Optional[torch.Tensor],
244
245
246
247
    head_size: int,
    cos_sin_cache: torch.Tensor,
    is_neox: bool,
) -> None:
248
249
    torch.ops._C.rotary_embedding(positions, query, key, head_size,
                                  cos_sin_cache, is_neox)
250
251
252
253
254


# layer norm ops
def rms_norm(out: torch.Tensor, input: torch.Tensor, weight: torch.Tensor,
             epsilon: float) -> None:
255
256
257
    # TODO: Remove this contiguous call when the kernel is updated to support non-contiguous input
    input_contiguous = input.contiguous()
    torch.ops._C.rms_norm(out, input_contiguous, weight, epsilon)
258
259
260
261


def fused_add_rms_norm(input: torch.Tensor, residual: torch.Tensor,
                       weight: torch.Tensor, epsilon: float) -> None:
262
    torch.ops._C.fused_add_rms_norm(input, residual, weight, epsilon)
263
264


265
266
267
268
269
270
271
def poly_norm(out: torch.Tensor, input: torch.Tensor, weight: torch.Tensor,
              bias: torch.Tensor, epsilon: float) -> None:
    # TODO: Remove this contiguous call when the kernel is updated to support non-contiguous input
    input_contiguous = input.contiguous()
    torch.ops._C.poly_norm(out, input_contiguous, weight, bias, epsilon)


272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
def apply_repetition_penalties_torch(
        logits: torch.Tensor, prompt_mask: torch.Tensor,
        output_mask: torch.Tensor, repetition_penalties: torch.Tensor) -> None:
    repetition_penalties = repetition_penalties.unsqueeze(dim=1).repeat(
        1, logits.size(1))
    # If token appears in prompt or output, apply, otherwise use 1.0 for no-op.
    penalties = torch.where(prompt_mask | output_mask, repetition_penalties,
                            1.0)
    # If logits are positive, divide by penalty, otherwise multiply by penalty.
    scaling = torch.where(logits > 0, 1.0 / penalties, penalties)
    logits *= scaling


def apply_repetition_penalties_cuda(
        logits: torch.Tensor, prompt_mask: torch.Tensor,
        output_mask: torch.Tensor, repetition_penalties: torch.Tensor) -> None:
    torch.ops._C.apply_repetition_penalties_(logits, prompt_mask, output_mask,
                                             repetition_penalties)


def apply_repetition_penalties(logits: torch.Tensor, prompt_mask: torch.Tensor,
                               output_mask: torch.Tensor,
                               repetition_penalties: torch.Tensor) -> None:
    """Apply repetition penalties to logits in-place.

    Args:
        logits: The logits tensor of shape [num_seqs, vocab_size].
        prompt_mask: A boolean tensor indicating which tokens appear in the prompt.
        output_mask: A boolean tensor indicating which tokens appear in the output.
        repetition_penalties: The repetition penalties of shape (num_seqs, ).
    """
303
    if logits.is_cuda and logits.is_contiguous():
304
305
306
307
308
309
310
        apply_repetition_penalties_cuda(logits, prompt_mask, output_mask,
                                        repetition_penalties)
    else:
        apply_repetition_penalties_torch(logits, prompt_mask, output_mask,
                                         repetition_penalties)


311
312
313
314
315
316
317
318
# 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
319
) -> tuple[torch.Tensor, torch.Tensor]:
320
321
322
323
324
325
326
327
328
329
330
    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


331
332
333
334
335
# 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:
336
337
338
339
    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)
340
341
    return torch.ops._C.awq_dequantize(qweight, scales, zeros, split_k_iters,
                                       thx, thy)
342
343
344
345


def awq_gemm(input: torch.Tensor, qweight: torch.Tensor, qzeros: torch.Tensor,
             scales: torch.Tensor, split_k_iters: int) -> torch.Tensor:
346
347
348
349
    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)
350
    return torch.ops._C.awq_gemm(input, qweight, qzeros, scales, split_k_iters)
351
352
353
354
355
356
357


# 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:
358
359
    return torch.ops._C.gptq_gemm(a, b_q_weight, b_gptq_qzeros, b_gptq_scales,
                                  b_g_idx, use_exllama, bit)
360
361


362
if hasattr(torch.ops._C, "gptq_gemm"):
363

364
    @register_fake("_C::gptq_gemm")
365
366
367
368
369
370
371
372
373
    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)


374
375
def gptq_shuffle(q_weight: torch.Tensor, q_perm: torch.Tensor,
                 bit: int) -> None:
376
    torch.ops._C.gptq_shuffle(q_weight, q_perm, bit)
377
378


379
380
381
# marlin_24
def gptq_marlin_24_gemm(a: torch.Tensor, b_q_weight: torch.Tensor,
                        b_meta: torch.Tensor, b_scales: torch.Tensor,
382
383
                        workspace: torch.Tensor, b_q_type: ScalarType,
                        size_m: int, size_n: int, size_k: int) -> torch.Tensor:
384
    return torch.ops._C.gptq_marlin_24_gemm(a, b_q_weight, b_meta, b_scales,
385
                                            workspace, b_q_type.id, size_m,
386
                                            size_n, size_k)
387
388


389
if hasattr(torch.ops._C, "gptq_marlin_24_gemm"):
390

391
    @register_fake("_C::gptq_marlin_24_gemm")
392
393
394
    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,
395
396
397
                                  b_q_type: ScalarType, size_m: torch.SymInt,
                                  size_n: torch.SymInt,
                                  size_k: torch.SymInt) -> torch.Tensor:
398
399
        return torch.empty((size_m, size_n), device=a.device, dtype=a.dtype)

400
    @register_fake("_C::gptq_marlin_gemm")
401
    def _gptq_marlin_gemm_fake(a: torch.Tensor,
402
                               c: Optional[torch.Tensor],
403
                               b_q_weight: torch.Tensor,
404
                               b_bias: Optional[torch.Tensor],
405
                               b_scales: torch.Tensor,
406
                               global_scale: Optional[torch.Tensor],
407
408
409
                               b_zeros: Optional[torch.Tensor],
                               g_idx: Optional[torch.Tensor],
                               perm: Optional[torch.Tensor],
410
                               workspace: torch.Tensor,
411
                               b_q_type_id: int,
412
413
414
                               size_m: torch.SymInt,
                               size_n: torch.SymInt,
                               size_k: torch.SymInt,
415
                               is_k_full: bool = True,
416
                               use_atomic_add: bool = False,
417
418
                               use_fp32_reduce: bool = False,
                               is_zp_float: bool = False) -> torch.Tensor:
419
420
        return torch.empty((size_m, size_n), device=a.device, dtype=a.dtype)

421
    @register_fake("_C::awq_dequantize")
422
    def _awq_dequantize_fake(qweight: torch.Tensor, scales: torch.Tensor,
423
424
                             zeros: torch.Tensor, split_k_iters: torch.SymInt,
                             thx: int, thy: int) -> torch.Tensor:
425
426
427
428
429
430
431
        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)

432
    @register_fake("_C::awq_gemm")
433
434
    def _awq_gemm_fake(input: torch.Tensor, qweight: torch.Tensor,
                       qzeros: torch.Tensor, scales: torch.Tensor,
435
                       split_k_iters: torch.SymInt) -> torch.Tensor:
436
437
438
439
440
        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)

441
442
    @register_fake("_C::machete_mm")
    def machete_mm_fake(
443
        a: torch.Tensor,
444
        # b_q Should be the tensor returned by machete_prepack_B
445
        b_q: torch.Tensor,
446
        b_type: ScalarType,
447
448
449
        out_type: Optional[torch.dtype] = None,
        b_group_scales: Optional[torch.Tensor] = None,
        b_group_zeros: Optional[torch.Tensor] = None,
450
        b_group_size: Optional[int] = None,
451
452
        b_channel_scales: Optional[torch.Tensor] = None,
        a_token_scales: Optional[torch.Tensor] = None,
453
454
455
456
457
458
        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)

459
    @register_fake("_C::machete_prepack_B")
460
461
462
    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:
463
464
        return torch.empty_like(b_q_weight,
                                memory_format=torch.contiguous_format)
465

466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
    @register_fake("_C::cutlass_w4a8_mm")
    def cutlass_w4a8_mm_fake(
            a: torch.Tensor,
            # b_q Should be the tensor returned by cutlass_encode_and_reorder_int4b
            b_q: torch.Tensor,
            b_group_scales: torch.Tensor,
            b_group_size: int,
            b_channel_scales: torch.Tensor,
            a_token_scales: torch.Tensor,
            out_type: Optional[torch.dtype] = None,
            maybe_schedule: Optional[str] = None) -> torch.Tensor:
        m = a.size(0)
        n = b_q.size(1)
        out_dtype = out_type if out_type is not None else torch.bfloat16
        return torch.empty((m, n), device=a.device, dtype=out_dtype)

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

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

490

491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
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)


507
508
509
if hasattr(torch.ops._C, "ggml_dequantize"):

    @register_fake("_C::ggml_dequantize")
510
511
512
513
514
515
    def _ggml_dequantize_fake(
            W: torch.Tensor,
            quant_type: int,
            m: torch.SymInt,
            n: torch.SymInt,
            dtype: Optional[torch.dtype] = None) -> torch.Tensor:
516
517
518
519
520
521
522
523
524
        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:
525
        return torch.empty((X.shape[0], row), dtype=X.dtype, device=W.device)
526
527
528
529
530
531
532
533
534

    @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)
535
        return torch.empty((batch, row), dtype=X.dtype, device=W.device)
536

537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
    @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)

554

555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
if hasattr(torch.ops._C, "ggml_moe_a8_vec"):

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


573
# cutlass
574
575
576
577
def cutlass_scaled_mm_supports_fp4(cuda_device_capability: int) -> bool:
    return torch.ops._C.cutlass_scaled_mm_supports_fp4(cuda_device_capability)


578
579
580
581
582
583
584
585
586
587
588
589
590
591
def cutlass_blockwise_scaled_grouped_mm(
    output: torch.Tensor,
    a: torch.Tensor,
    b: torch.Tensor,
    scales_a: torch.Tensor,
    scales_b: torch.Tensor,
    problem_sizes: torch.Tensor,
    expert_offsets: torch.Tensor,
):
    torch.ops._C.cutlass_blockwise_scaled_grouped_mm(output, a, b, scales_a,
                                                     scales_b, problem_sizes,
                                                     expert_offsets)


592
593
594
595
596
597
598
599
600
601
602
603
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


604
605
606
607
def cutlass_scaled_mm_supports_fp8(cuda_device_capability: int) -> bool:
    return torch.ops._C.cutlass_scaled_mm_supports_fp8(cuda_device_capability)


608
609
610
611
612
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)


613
614
615
def cutlass_scaled_mm(a: torch.Tensor,
                      b: torch.Tensor,
                      scale_a: torch.Tensor,
616
                      scale_b: torch.Tensor,
617
                      out_dtype: torch.dtype,
618
                      bias: Optional[torch.Tensor] = None) -> torch.Tensor:
619
    """
620
    `cutlass_scaled_mm` implements a fused version of
621
        `output = torch.mm((scale_a * a), (scale_b * b)).to(out_dtype)`
622
623
624
625
626
627
628
629
    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
630
631
632
633
634
635
636
637
638
639
640
        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
    """
641
    assert (out_dtype is torch.bfloat16 or out_dtype is torch.float16)
642
643
    assert bias is None or bias.numel(
    ) == b.shape[1] and bias.dtype == out_dtype
644

645
646
647
    # Massage the input to be 2D
    target_shape = (*a.shape[:-1], b.shape[1])
    a = a.view(-1, a.shape[-1])
648

649
650
    cutlass_compatible_b = (b.shape[0] % 16 == 0 and b.shape[1] % 16 == 0)
    if current_platform.is_rocm() or not cutlass_compatible_b:
651
652
        from vllm.model_executor.layers.quantization.compressed_tensors.triton_scaled_mm import (  # noqa
            triton_scaled_mm)
653
654
655
656
657
658
        out = triton_scaled_mm(a, b, scale_a, scale_b, out_dtype, bias)
    else:
        out = torch.empty((a.shape[0], b.shape[1]),
                          dtype=out_dtype,
                          device=a.device)
        torch.ops._C.cutlass_scaled_mm(out, a, b, scale_a, scale_b, bias)
659

660
    return out.view(*target_shape)
661
662


663
664
665
666
667
668
669
670
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:
671
672
673
674
675
    """
    :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.
    """
676
677
678
679
680
    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

681
682
683
684
    # Massage the input to be 2D
    target_shape = (*a.shape[:-1], b.shape[1])
    a = a.view(-1, a.shape[-1])
    assert azp is None or azp.numel() == a.shape[0]
685

686
687
688
    out = torch.empty((a.shape[0], b.shape[1]),
                      dtype=out_dtype,
                      device=a.device)
689
690
    torch.ops._C.cutlass_scaled_mm_azp(out, a, b, scale_a, scale_b, azp_adj,
                                       azp, bias)
691
    return out.view(*target_shape)
692
693


694
695
696
697
698
def cutlass_sparse_scaled_mm_supported(cuda_device_capability: int) -> bool:
    return torch.ops._C.cutlass_sparse_scaled_mm_supported(
        cuda_device_capability)


699
700
701
def cutlass_group_gemm_supported(cuda_device_capability: int) -> bool:
    return torch.ops._C.cutlass_group_gemm_supported(cuda_device_capability)

702

703
def cutlass_sparse_compress(a: torch.Tensor) \
704
    -> tuple[torch.Tensor, torch.Tensor]:
705
706
707
708
709
710
711
712
    """
    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:
713
        a (torch.Tensor):
714
715
716
717
718
719
720
            The input tensor to be compressed. Must have one of the following data types:
            - `torch.int8`
            - `torch.float8_e4m3fn`
            - `torch.bfloat16`
            - `torch.float16`

    Returns:
721
        tuple[torch.Tensor, torch.Tensor]:
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
            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
742
    assert (a.shape[1] % (2 * elemsPerMetaElem) == 0)
743

744
    return torch.ops._C.cutlass_sparse_compress(a)
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792


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


793
794
795
796
797
798
799
800
801
802
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,
                            blockscale_offsets: Optional[torch.Tensor] = None):
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
    """
    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.
820
821
822
823
824
    - blockscale_offsets: Optional argument passed for fp4 moe. Indices that
                          mark at which block scale index each expert begins
                          its computation. The number of block scale rows
                          computed with expert E is blockscale_offsets[E + 1] -
                          blockscale_offsets[E]
825
    """
826
827
828
829
    return torch.ops._C.get_cutlass_moe_mm_data(topk_ids, expert_offsets,
                                                problem_sizes1, problem_sizes2,
                                                input_permutation,
                                                output_permutation,
830
831
832
833
                                                num_experts, n, k,
                                                blockscale_offsets)


834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
def get_cutlass_moe_mm_problem_sizes(
        topk_ids: torch.Tensor,
        problem_sizes1: torch.Tensor,
        problem_sizes2: torch.Tensor,
        num_experts: int,
        n: int,
        k: int,
        blockscale_offsets: Optional[torch.Tensor] = None):
    """
    Compute only the per-expert problem sizes needed by the two grouped matrix
    multiplications used in CUTLASS-based fused MoE.

    The function takes in topk_ids (token→expert mapping) and computes:
    - problem_sizes1, problem_sizes2: M×N×K sizes of each expert's
                                    multiplication for the two grouped MMs
                                    used in the fused MoE operation.
    """
    return torch.ops._C.get_cutlass_moe_mm_problem_sizes(
        topk_ids, problem_sizes1, problem_sizes2, num_experts, n, k,
        blockscale_offsets)


856
857
858
859
860
861
862
863
864
865
866
def shuffle_rows(input_tensor: torch.Tensor, dst2src_map: torch.Tensor):
    """
    Shuffle and expand the input tensor according to the dst2src_map and store the result in output_tensor.
    This is used in MoE to permute the input tensor before performing grouped matrix multiplications.
    """
    num_tokens_permuted = dst2src_map.shape[0]
    output_tensor = torch.empty((num_tokens_permuted, input_tensor.shape[1]),
                                device=input_tensor.device,
                                dtype=input_tensor.dtype)
    torch.ops._moe_C.shuffle_rows(input_tensor, dst2src_map, output_tensor)
    return output_tensor
867
868


869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
def get_cutlass_pplx_moe_mm_data(expert_offsets: torch.Tensor,
                                 problem_sizes1: torch.Tensor,
                                 problem_sizes2: torch.Tensor,
                                 expert_num_tokens: torch.Tensor,
                                 num_local_experts: int, padded_m: int, n: int,
                                 k: int):
    """
    Prepare data necessary to perform CUTLASS grouped matrix multiplications
    used in CUTLASS-based fused MoE.

    The function takes in expert_num_tokens (token count per expert) and
    non_zero_expert_idxs (consecutive indices of experts with non-zero token 
    counts) and uses them to compute:
    - expert_offsets: Indices that mark at which token index each expert begins
                      its computation.
    - problem_sizes1, problem_sizes2: MxNxK sizes of each expert's
                                      multiplication in two grouped MMs used in
                                      the fused MoE operation.
    """
    return torch.ops._C.get_cutlass_pplx_moe_mm_data(
        expert_offsets, problem_sizes1, problem_sizes2, expert_num_tokens,
        num_local_experts, padded_m, n, k)


893
894
895
896
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,
897
898
                   b_strides: torch.Tensor, c_strides: torch.Tensor,
                   per_act_token: bool, per_out_ch: bool):
899
900
901
902
903
904
905
906
907
908
909
    """
    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.
    """
910
911
912
    return torch.ops._C.cutlass_moe_mm(out_tensors, a_tensors, b_tensors,
                                       a_scales, b_scales, expert_offsets,
                                       problem_sizes, a_strides, b_strides,
913
                                       c_strides, per_act_token, per_out_ch)
914
915


916
917
918
919
920
def cutlass_fp4_moe_mm(out_tensors: torch.Tensor, a_tensors: torch.Tensor,
                       b_tensors: torch.Tensor, a_scales: torch.Tensor,
                       b_scales: torch.Tensor, alphas: torch.Tensor,
                       problem_sizes: torch.Tensor,
                       expert_offsets: torch.Tensor, sf_offsets: torch.Tensor):
921
    """
922
    An FP4 Blockscaled Group Gemm that takes in  a_tensors, b_tensors and runs
923
924
925
926
927
928
    the gemms for each combination based on the specified problem sizes.

    This is used as the MoE gemm during NVFP4 Quantized FusedMoE forward.
    - a/b_tensors: the NVFP4 a_ptrs and b_ptrs tensors which are quantized
                     input and expert weights.
    - a_/b_scales: The blockscales in FP8-E4M3 precision
929
930
931
932
    - expert_offsets/sf_offsets: Indices that mark at which token index
                    each expert begins its computation. The number of tokens
                    computed with expert E is expert_offsets[E + 1] -
                    expert_offsets[E] And the sf_size per expert is
933
934
935
936
                    sf_offset[E+1] - sf_offset[E]
    - problem_sizes: MxNxK sizes of each expert's multiplication in two grouped
                     MMs used in the fused MoE operation.
    """
937
938
939
940
    return torch.ops._C.cutlass_fp4_group_mm(out_tensors, a_tensors, b_tensors,
                                             a_scales, b_scales, alphas,
                                             problem_sizes, expert_offsets,
                                             sf_offsets)
941
942


943
944
# gptq_marlin
def gptq_marlin_repack(b_q_weight: torch.Tensor, perm: torch.Tensor,
945
946
                       size_k: int, size_n: int,
                       num_bits: int) -> torch.Tensor:
947
948
    return torch.ops._C.gptq_marlin_repack(b_q_weight, perm, size_k, size_n,
                                           num_bits)
949
950


951
952
953
954
955
956
# 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)


957
958
959
960
961
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
962
    output = torch.empty((num_experts, size_k // 16, size_n * (num_bits // 2)),
963
964
965
966
967
968
969
970
                         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


971
972
973
974
975
976
977
978
979
980
981
982
983
984
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


985
def gptq_marlin_gemm(a: torch.Tensor,
986
                     c: Optional[torch.Tensor],
987
                     b_q_weight: torch.Tensor,
988
                     b_bias: Optional[torch.Tensor],
989
                     b_scales: torch.Tensor,
990
                     global_scale: Optional[torch.Tensor],
991
992
993
                     b_zeros: Optional[torch.Tensor],
                     g_idx: Optional[torch.Tensor],
                     perm: Optional[torch.Tensor],
994
995
996
997
998
                     workspace: torch.Tensor,
                     b_q_type: ScalarType,
                     size_m: int,
                     size_n: int,
                     size_k: int,
999
                     is_k_full: bool = True,
1000
                     use_atomic_add: bool = False,
1001
1002
                     use_fp32_reduce: bool = False,
                     is_zp_float: bool = False) -> torch.Tensor:
1003
    return torch.ops._C.gptq_marlin_gemm(a, c, b_q_weight, b_bias, b_scales,
1004
1005
1006
                                         global_scale, b_zeros, g_idx, perm,
                                         workspace, b_q_type.id, size_m,
                                         size_n, size_k, is_k_full,
1007
1008
                                         use_atomic_add, use_fp32_reduce,
                                         is_zp_float)
1009
1010


1011
# machete
1012
1013
1014
1015
1016
1017
1018
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,
1019
        out_type: Optional[torch.dtype] = None) -> list[str]:
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
    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)
1047
1048


1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
# CUTLASS W4A8
def cutlass_w4a8_mm(
        a: torch.Tensor,
        # b_q Should be the tensor returned by cutlass_encode_and_reorder_int4b
        b_q: torch.Tensor,
        b_group_scales: torch.Tensor,
        b_group_size: int,
        b_channel_scales: torch.Tensor,
        a_token_scales: torch.Tensor,
        out_type: Optional[torch.dtype] = None,
        maybe_schedule: Optional[str] = None) -> torch.Tensor:
    return torch.ops._C.cutlass_w4a8_mm(a, b_q, b_group_scales, b_group_size,
                                        b_channel_scales, a_token_scales,
                                        out_type, maybe_schedule)


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


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


1073
if hasattr(torch.ops._C, "permute_cols"):
1074

1075
    @register_fake("_C::permute_cols")
1076
1077
1078
1079
1080
1081
1082
1083
1084
    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)


1085
1086
1087
# fp4
def scaled_fp4_quant(
        input: torch.Tensor,
1088
        input_global_scale: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
    """
    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:
1103
        tuple[torch.Tensor, torch.Tensor]: The output tensor in FP4 but every
1104
1105
1106
            two values are packed into a uint8 and float8_e4m3 scaling factors
            in the sizzled layout.
    """
1107
    assert not current_platform.is_rocm()
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
    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


1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
def scaled_fp4_experts_quant(
    input_tensor: torch.Tensor,
    input_global_scale: torch.Tensor,
    expert_offsets: torch.Tensor,
    blockscale_offsets: torch.Tensor,
    topk: int,
) -> tuple[torch.Tensor, torch.Tensor]:
    """
    Quantize input tensor to FP4 and return quantized tensor and scale, for
    packed MoE Inputs.
    Args:
1154
        input_tensor: The input tensor to be quantized to FP4
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
        input_global_scale: A scalar scaling factor for the entire tensor.
        expert_offsets: The expert offsets tensor
        blockscale_offsets: The blockscale offsets tensor
    Outputs:
        output: The quantized tensor in FP4
        output_scales: The blockscale tensor in FP8-E4M3
    """
    assert not current_platform.is_rocm()
    assert input_tensor.ndim == 2, (
        f'input.ndim needs to be == 2, but got {input_tensor.ndim}.')

1166
1167
1168
1169
1170
    # Control the maximum number of tokens per expert supported by the
    # NVFP4 MoE Expert Quantization. This is used to prevent the kernel
    # from running out of memory. This value can also be increased to support
    # larger models.
    MAX_TOKENS_PER_EXPERT = envs.VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE
1171
1172
    m_numtopk, k = input_tensor.shape

1173
    assert (m_numtopk <= MAX_TOKENS_PER_EXPERT * topk), (
1174
1175
1176
1177
        f"m_numtopk must be less than MAX_TOKENS_PER_EXPERT("
        f"{MAX_TOKENS_PER_EXPERT})"
        f" for cutlass_moe_fp4, observed m_numtopk = {m_numtopk}. Use"
        f" VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE to set this value.")
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
    scales_k = k // 16
    padded_k = (scales_k + (4 - 1)) // 4

    # output is uint8 and packed fp4 values
    output = torch.empty(m_numtopk,
                         k // 2,
                         device=input_tensor.device,
                         dtype=torch.uint8)
    output_scales = torch.empty(MAX_TOKENS_PER_EXPERT * topk,
                                padded_k,
                                dtype=torch.int32,
                                device=input_tensor.device)
    torch.ops._C.scaled_fp4_experts_quant(output, output_scales, input_tensor,
                                          input_global_scale, expert_offsets,
                                          blockscale_offsets)
    output_scales = output_scales.view(torch.float8_e4m3fn)
    return output, output_scales


1197
# fp8
1198
1199
1200
def scaled_fp8_quant(
    input: torch.Tensor,
    scale: Optional[torch.Tensor] = None,
1201
    num_token_padding: Optional[int] = None,
1202
    scale_ub: Optional[torch.Tensor] = None,
1203
    use_per_token_if_dynamic: bool = False,
1204
    output: Optional[torch.Tensor] = None,
1205
) -> tuple[torch.Tensor, torch.Tensor]:
1206
1207
1208
1209
1210
1211
    """
    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
1212
    optional padding of the output tensors for downstream kernels that
1213
1214
1215
1216
1217
    will benefit from padding.

    Args:
        input: The input tensor to be quantized to FP8
        scale: Optional scaling factor for the FP8 quantization
1218
        scale_ub: Optional upper bound for scaling factor in dynamic
1219
            per token case
1220
        num_token_padding: If specified, pad the first dimension
1221
            of the output to at least this value.
1222
        use_per_token_if_dynamic: Whether to do per_tensor or per_token
1223
            in the dynamic quantization case.
1224
1225

    Returns:
1226
        tuple[torch.Tensor, torch.Tensor]: The output tensor in FP8 and
1227
1228
            scaling factor.
    """
1229
1230
    # This code assumes batch_dim and num_tokens are flattened
    assert (input.ndim == 2)
1231
    shape: Union[tuple[int, int], torch.Size] = input.shape
1232
1233
    # For ROCm on MI300, the output fp8 dtype is torch.float_e3m3fnuz
    out_dtype: torch.dtype = current_platform.fp8_dtype()
1234
1235
    if num_token_padding:
        shape = (max(num_token_padding, input.shape[0]), shape[1])
1236
1237
1238
1239
1240
1241
    if output is None:
        output = torch.empty(shape, device=input.device, dtype=out_dtype)
    else:
        assert num_token_padding is None, \
            "padding not supported if output passed in"
        assert output.dtype == out_dtype
1242

1243
    if scale is None:
1244
        if use_per_token_if_dynamic:
1245
            scale = torch.empty((shape[0], 1),
1246
1247
1248
                                device=input.device,
                                dtype=torch.float32)
            torch.ops._C.dynamic_per_token_scaled_fp8_quant(
1249
                output, input, scale, scale_ub)
1250
        else:
1251
1252
            scale = torch.empty(1, device=input.device, dtype=torch.float32)
            torch.ops._C.dynamic_scaled_fp8_quant(output, input, scale)
1253
    else:
bnellnm's avatar
bnellnm committed
1254
        assert scale.numel() == 1, f"{scale.shape}"
1255
        torch.ops._C.static_scaled_fp8_quant(output, input, scale)
1256

1257
    return output, scale
1258
1259


1260
1261
1262
1263
1264
1265
# gptq allspark
def allspark_repack_weight(
        qweight: torch.Tensor,
        scale: torch.Tensor,
        zero_point: Optional[torch.Tensor] = None,
        has_zp: bool = False
1266
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
1267
    """
1268
    Rearrange qweight, scale, and zero_point(if asymmetric) to n32k16 format
1269
1270
1271
1272
1273
1274
1275
1276
    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.
1277
1278
            if use asymmetric quantization, has_zp = True.

1279
    Returns:
1280
        tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]] :
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
            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)


1321
# int8
1322
def scaled_int8_quant(
1323
1324
1325
1326
    input: torch.Tensor,
    scale: Optional[torch.Tensor] = None,
    azp: Optional[torch.Tensor] = None,
    symmetric: bool = True
1327
) -> tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
1328
    """
1329
    Quantize the input tensor to int8 and return the quantized tensor and scale, and maybe azp.
1330
1331
1332

    Args:
        input: The input tensor to be quantized to int8.
1333
1334
        scale: Optional scaling factor for the int8 quantization.
            When not provided, we invoke dynamic-per-token quantization.
1335
1336
1337
        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).
1338
1339

    Returns:
1340
      tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]] : Output int8 tensor, scales, and optionally azp.
1341
    """
1342
1343
1344
    output = torch.empty_like(input, dtype=torch.int8)
    if scale is not None:
        # static-per-tensor quantization.
1345
        assert symmetric == (
1346
1347
            azp
            is None), "azp must only be provided for asymmetric quantization."
1348
        torch.ops._C.static_scaled_int8_quant(output, input, scale, azp)
1349
        return output, scale, azp
1350
1351
1352
1353
1354

    # dynamic-per-token quantization.
    input_scales = torch.empty((input.numel() // input.shape[-1], 1),
                               device=input.device,
                               dtype=torch.float32)
1355
1356
    input_azp = None if symmetric else torch.empty_like(input_scales,
                                                        dtype=torch.int32)
1357
1358
    torch.ops._C.dynamic_scaled_int8_quant(output, input.contiguous(),
                                           input_scales, input_azp)
1359
    return output, input_scales, input_azp
1360
1361


1362
# gguf
1363
1364
1365
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)
1366
1367
1368
1369
1370
1371
1372


def ggml_mul_mat_vec_a8(
    W: torch.Tensor,
    X: torch.Tensor,
    quant_type: int,
    row: int,
1373
) -> torch.Tensor:
1374
1375
1376
1377
1378
1379
1380
1381
    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,
1382
) -> torch.Tensor:
1383
1384
1385
    return torch.ops._C.ggml_mul_mat_a8(W, X, quant_type, row)


1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
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)


1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
def ggml_moe_a8_vec(
    X: torch.Tensor,
    W: torch.Tensor,
    topk_ids: torch.Tensor,
    top_k: int,
    quant_type: int,
    row: torch.SymInt,
    tokens: torch.SymInt,
) -> torch.Tensor:
    return torch.ops._C.ggml_moe_a8_vec(X, W, topk_ids, top_k, quant_type, row,
                                        tokens)


1415
1416
1417
1418
def ggml_moe_get_block_size(quant_type: int) -> int:
    return torch.ops._C.ggml_moe_get_block_size(quant_type)


1419
# mamba
1420
1421
1422
1423
1424
1425
1426
1427
1428
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):
1429
1430
1431
    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,
1432
                                    ssm_states, pad_slot_id)
1433
1434


1435
1436
1437
1438
1439
1440
# ROCm skinny gemms
def LLMM1(a: torch.Tensor, b: torch.Tensor,
          rows_per_block: int) -> torch.Tensor:
    return torch.ops._rocm_C.LLMM1(a, b, rows_per_block)


1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
def wvSplitK(a: torch.Tensor,
             b: torch.Tensor,
             cu_count: int,
             bias: torch.Tensor = None) -> torch.Tensor:
    return torch.ops._rocm_C.wvSplitK(a, b, bias, cu_count)


def wvSplitKQ(a: torch.Tensor,
              b: torch.Tensor,
              out_dtype: torch.dtype,
              scale_a: torch.Tensor,
              scale_b: torch.Tensor,
              cu_count: int,
              bias: torch.Tensor = None) -> torch.Tensor:
1455
1456
1457
    out = torch.empty((b.shape[0], a.shape[0]),
                      dtype=out_dtype,
                      device=b.device)
1458
    torch.ops._rocm_C.wvSplitKQ(a, b, bias, out, scale_a, scale_b, cu_count)
1459
1460
1461
    return out


1462
# moe
1463
1464
1465
1466
def moe_sum(input: torch.Tensor, output: torch.Tensor):
    torch.ops._moe_C.moe_sum(input, output)


1467
1468
1469
1470
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:
1471
1472
1473
    torch.ops._moe_C.moe_align_block_size(topk_ids, num_experts, block_size,
                                          sorted_token_ids, experts_ids,
                                          num_tokens_post_pad)
1474
1475


1476
1477
1478
1479
1480
1481
1482
1483
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:
1484
1485
1486
1487
    if not current_platform.is_cuda():
        raise NotImplementedError(
            "The optimized moe_wna16_gemm kernel is only "
            "available on CUDA platforms")
1488
1489
1490
1491
1492
1493
1494
    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)


1495
def topk_softmax(topk_weights: torch.Tensor, topk_ids: torch.Tensor,
1496
                 token_expert_indices: torch.Tensor,
1497
                 gating_output: torch.Tensor) -> None:
1498
1499
    torch.ops._moe_C.topk_softmax(topk_weights, topk_ids, token_expert_indices,
                                  gating_output)
1500
1501


1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
def grouped_topk(scores: torch.Tensor, scores_with_bias: torch.Tensor,
                 num_expert_group: int, topk_group: int, topk: int,
                 renormalize: bool, routed_scaling_factor: float):
    if not current_platform.is_cuda():
        raise NotImplementedError("The fused grouped_topk kernel is only "
                                  "available on CUDA platforms")
    return torch.ops._moe_C.grouped_topk(scores, scores_with_bias,
                                         num_expert_group, topk_group, topk,
                                         renormalize, routed_scaling_factor)


1513
def moe_wna16_marlin_gemm(input: torch.Tensor, output: Optional[torch.Tensor],
1514
1515
1516
                          b_qweight: torch.Tensor,
                          b_bias: Optional[torch.Tensor],
                          b_scales: torch.Tensor,
1517
                          global_scale: Optional[torch.Tensor],
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
                          b_qzeros: Optional[torch.Tensor],
                          g_idx: Optional[torch.Tensor],
                          perm: Optional[torch.Tensor],
                          workspace: torch.Tensor,
                          sorted_token_ids: torch.Tensor,
                          expert_ids: torch.Tensor,
                          num_tokens_past_padded: torch.Tensor,
                          topk_weights: torch.Tensor, moe_block_size: int,
                          top_k: int, mul_topk_weights: bool, is_ep: bool,
                          b_q_type: ScalarType, size_m: int, size_n: int,
                          size_k: int, is_k_full: bool, use_atomic_add: bool,
                          use_fp32_reduce: bool,
                          is_zp_float: bool) -> torch.Tensor:
    return torch.ops._moe_C.moe_wna16_marlin_gemm(
1532
1533
1534
1535
1536
        input, output, b_qweight, b_bias, b_scales, global_scale, b_qzeros,
        g_idx, perm, workspace, sorted_token_ids, expert_ids,
        num_tokens_past_padded, topk_weights, moe_block_size, top_k,
        mul_topk_weights, is_ep, b_q_type.id, size_m, size_n, size_k,
        is_k_full, use_atomic_add, use_fp32_reduce, is_zp_float)
1537
1538


1539
1540
if supports_moe_ops and hasattr(torch.ops._moe_C, "marlin_gemm_moe"):

1541
    @register_fake("_moe_C::marlin_gemm_moe")
1542
1543
1544
1545
    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,
1546
1547
                             b_zero_points: torch.Tensor, g_idx: torch.Tensor,
                             perm: torch.Tensor, workspace: torch.Tensor,
1548
1549
1550
1551
                             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,
1552
1553
1554
1555
1556
                             apply_weights: bool) -> torch.Tensor:
        return torch.empty((size_m, topk, size_n),
                           dtype=a.dtype,
                           device=a.device)

1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
    @register_fake("_moe_C::moe_wna16_marlin_gemm")
    def moe_wna16_marlin_gemm_fake(input: torch.Tensor,
                                   output: Optional[torch.Tensor],
                                   b_qweight: torch.Tensor,
                                   b_scales: torch.Tensor,
                                   b_qzeros: Optional[torch.Tensor],
                                   g_idx: Optional[torch.Tensor],
                                   perm: Optional[torch.Tensor],
                                   workspace: torch.Tensor,
                                   sorted_token_ids: torch.Tensor,
                                   expert_ids: torch.Tensor,
                                   num_tokens_past_padded: torch.Tensor,
                                   topk_weights: torch.Tensor,
                                   moe_block_size: int, top_k: int,
                                   mul_topk_weights: bool, is_ep: bool,
                                   b_q_type: ScalarType, size_m: int,
                                   size_n: int, size_k: int, is_k_full: bool,
                                   use_atomic_add: bool, use_fp32_reduce: bool,
                                   is_zp_float: bool) -> torch.Tensor:
        return torch.empty((size_m * top_k, size_n),
                           dtype=input.dtype,
                           device=input.device)

1580

1581
1582
1583
1584
1585
1586
1587
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,
1588
1589
    k_scale: torch.Tensor,
    v_scale: torch.Tensor,
1590
) -> None:
1591
1592
    torch.ops._C_cache_ops.reshape_and_cache(key, value, key_cache,
                                             value_cache, slot_mapping,
1593
                                             kv_cache_dtype, k_scale, v_scale)
1594
1595


1596
1597
1598
1599
1600
1601
1602
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,
1603
1604
    k_scale: torch.Tensor,
    v_scale: torch.Tensor,
1605
) -> None:
1606
1607
    torch.ops._C_cache_ops.reshape_and_cache_flash(key, value, key_cache,
                                                   value_cache, slot_mapping,
1608
1609
                                                   kv_cache_dtype, k_scale,
                                                   v_scale)
1610
1611


1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
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)


1625
1626
def copy_blocks(key_caches: list[torch.Tensor],
                value_caches: list[torch.Tensor],
1627
                block_mapping: torch.Tensor) -> None:
1628
    torch.ops._C_cache_ops.copy_blocks(key_caches, value_caches, block_mapping)
1629
1630


1631
def copy_blocks_mla(kv_caches: list[torch.Tensor],
1632
1633
1634
1635
                    block_mapping: torch.Tensor) -> None:
    torch.ops._C_cache_ops.copy_blocks_mla(kv_caches, block_mapping)


1636
def swap_blocks(src: torch.Tensor, dst: torch.Tensor,
1637
                block_mapping: torch.Tensor) -> None:
1638
    torch.ops._C_cache_ops.swap_blocks(src, dst, block_mapping)
1639
1640


1641
1642
1643
1644
def convert_fp8(output: torch.Tensor,
                input: torch.Tensor,
                scale: float = 1.0,
                kv_dtype: str = "fp8") -> None:
1645
1646
1647
    torch.ops._C_cache_ops.convert_fp8(output, input, scale, kv_dtype)


1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
def gather_and_maybe_dequant_cache(
        src_cache: torch.Tensor,
        dst: torch.Tensor,
        block_table: torch.Tensor,
        cu_seq_lens: torch.Tensor,
        batch_size: int,
        kv_cache_dtype: str,
        scale: torch.Tensor,
        seq_starts: Optional[torch.Tensor] = None) -> None:
    torch.ops._C_cache_ops.gather_and_maybe_dequant_cache(
        src_cache, dst, block_table, cu_seq_lens, batch_size, kv_cache_dtype,
        scale, seq_starts)
1660
1661


1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
def cp_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.cp_gather_cache(src_cache, dst, block_table,
                                           cu_seq_lens, batch_size, seq_starts)


1672
1673
1674
1675
1676
1677
1678
1679
1680
def indexer_k_quant_and_cache(k: torch.Tensor, kv_cache: torch.Tensor,
                              slot_mapping: torch.Tensor,
                              quant_block_size: int,
                              kv_cache_dtype: str) -> None:
    torch.ops._C_cache_ops.indexer_k_quant_and_cache(k, kv_cache, slot_mapping,
                                                     quant_block_size,
                                                     kv_cache_dtype)


1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
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
1692
def init_custom_ar(ipc_tensors: list[torch.Tensor], rank_data: torch.Tensor,
1693
                   rank: int, fully_connected: bool) -> int:
1694
    return torch.ops._C_custom_ar.init_custom_ar(ipc_tensors, rank_data, rank,
1695
                                                 fully_connected)
1696
1697


1698
1699
1700
1701
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)
1702

1703
1704
1705
1706
1707
1708
1709
1710
1711

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


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


1712
def register_buffer(fa: int, ipc_tensors: list[int]) -> None:
1713
    return torch.ops._C_custom_ar.register_buffer(fa, ipc_tensors)
1714
1715


1716
def get_graph_buffer_ipc_meta(fa: int) -> tuple[list[int], list[int]]:
1717
1718
1719
    return torch.ops._C_custom_ar.get_graph_buffer_ipc_meta(fa)


1720
1721
def register_graph_buffers(fa: int, handles: list[list[int]],
                           offsets: list[list[int]]) -> None:
1722
    torch.ops._C_custom_ar.register_graph_buffers(fa, handles, offsets)
1723
1724


1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
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)


1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
# quick all reduce
def init_custom_qr(rank: int,
                   world_size: int,
                   qr_max_size: Optional[int] = None) -> int:
    return torch.ops._C_custom_ar.init_custom_qr(rank, world_size, qr_max_size)


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


def qr_all_reduce(fa: int,
                  inp: torch.Tensor,
                  out: torch.Tensor,
                  quant_level: int,
                  cast_bf2half: bool = False) -> None:
    torch.ops._C_custom_ar.qr_all_reduce(fa, inp, out, quant_level,
                                         cast_bf2half)


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


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


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


1769
1770
1771
1772
def get_flash_mla_metadata(
    cache_seqlens: torch.Tensor,
    num_heads_per_head_k: int,
    num_heads_k: int,
1773
) -> tuple[torch.Tensor, torch.Tensor]:
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
    """
    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,
1799
) -> tuple[torch.Tensor, torch.Tensor]:
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
    """
    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
1831
1832


1833
1834
def sm100_cutlass_mla_decode(out: torch.Tensor, lse: torch.Tensor,
                             q_nope: torch.Tensor, q_pe: torch.Tensor,
1835
1836
1837
1838
                             kv_c_and_k_pe_cache: torch.Tensor,
                             seq_lens: torch.Tensor, page_table: torch.Tensor,
                             workspace: torch.Tensor, scale: float,
                             num_kv_splits: int) -> torch.Tensor:
1839
    torch.ops._C.sm100_cutlass_mla_decode(out, lse, q_nope, q_pe,
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
                                          kv_c_and_k_pe_cache, seq_lens,
                                          page_table, workspace, scale,
                                          num_kv_splits)
    return out


def sm100_cutlass_mla_get_workspace_size(max_seq_len: int, num_batches: int,
                                         sm_count: int,
                                         num_kv_splits: int) -> int:
    return torch.ops._C.sm100_cutlass_mla_get_workspace_size(
        max_seq_len, num_batches, sm_count, num_kv_splits)


1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
if hasattr(torch.ops._C, "weight_packed_linear"):

    @register_fake("_C::weight_packed_linear")
    def weight_packed_linear_fake(mat1: torch.Tensor, mat2: torch.Tensor,
                                  bias: Optional[torch.Tensor],
                                  is_vnni: bool) -> torch.Tensor:
        return torch.empty((mat1.size(0), mat2.size(0)),
                           dtype=mat1.dtype,
                           device=mat2.device)


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

    @register_fake("_C::fused_experts_cpu")
    def fused_experts_cpu_fake(
        hidden_states: torch.Tensor,
        w1: torch.Tensor,
        w2: torch.Tensor,
        topk_weights: torch.Tensor,
        topk_ids: torch.Tensor,
        inplace: bool,
        use_int8_w8a8: bool,
        use_fp8_w8a16: bool,
        w1_scale: Optional[torch.Tensor],
        w2_scale: Optional[torch.Tensor],
        block_size: Optional[list[int]],
        a1_scale: Optional[torch.Tensor],
        a2_scale: Optional[torch.Tensor],
        is_vnni: bool,
    ) -> torch.Tensor:
        return torch.empty_like(hidden_states)


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

    @register_fake("_C::int8_scaled_mm_with_quant")
    def int8_scaled_mm_with_quant_fake(
        mat1: torch.Tensor,
        mat2: torch.Tensor,
        scales2: torch.Tensor,
        bias: Optional[torch.Tensor],
        out_dtype: torch.dtype,
        is_vnni: bool,
    ) -> torch.Tensor:
        M = mat1.size(0)
        N = mat2.size(0)
        return torch.empty((M, N), dtype=out_dtype)
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913


class CPUDNNLGEMMHandler:

    def __init__(self) -> None:
        self.handler: Optional[int] = None
        self.n = -1
        self.k = -1

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


1914
1915
1916
1917
1918
1919
if hasattr(torch.ops._C, "create_onednn_mm_handler"):
    _supports_onednn = True
else:
    _supports_onednn = False


1920
1921
1922
1923
def is_onednn_acl_supported():
    return torch.ops._C.is_onednn_acl_supported()


1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
def create_onednn_mm(
    weight: torch.Tensor,  # [K, N]
    primitive_cache_size: int = 128,
) -> CPUDNNLGEMMHandler:
    handler = CPUDNNLGEMMHandler()
    handler.k, handler.n = weight.size()
    handler.handler = torch.ops._C.create_onednn_mm_handler(
        weight, primitive_cache_size)
    return handler


def onednn_mm(
    dnnl_handler: CPUDNNLGEMMHandler,
    x: torch.Tensor,
    bias: Optional[torch.Tensor],
) -> torch.Tensor:
    output = torch.empty((*x.shape[0:-1], dnnl_handler.n), dtype=x.dtype)
    torch.ops._C.onednn_mm(output, x.reshape(-1, dnnl_handler.k), bias,
                           dnnl_handler.handler)

    return output


1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
def create_onednn_scaled_mm(
    weight: torch.Tensor,  # [K, N]
    weight_scales: torch.Tensor,
    output_type: torch.dtype,
    dynamic_quant: bool,
    use_azp: bool,
    primitive_cache_size: int = 128,
) -> CPUDNNLGEMMHandler:
    handler = CPUDNNLGEMMHandler()
    handler.k, handler.n = weight.size()
    handler.handler = torch.ops._C.create_onednn_scaled_mm_handler(
        weight, weight_scales, output_type, dynamic_quant, use_azp,
        primitive_cache_size)
    return handler


def onednn_scaled_int8_quant(input: torch.Tensor,
                             scale: Optional[torch.Tensor] = None,
                             azp: Optional[torch.Tensor] = None,
                             symmetric: bool = True):
    """
    Quantize the input tensor to int8 and return the quantized tensor and scale, and maybe azp.

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

    Returns:
      tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]] : Output int8 tensor, scales, and optionally azp.
    """
    output = torch.empty_like(input, dtype=torch.int8)
    token_num = input.numel() // input.shape[-1]
    input = input.view((token_num, input.shape[-1]))
    if scale is not None:
        # static-per-tensor quantization.
        assert symmetric == (
            azp
            is None), "azp must only be provided for asymmetric quantization."
        torch.ops._C.static_scaled_int8_quant(output, input, scale, azp)
        return output, scale, azp

    # dynamic-per-token quantization.
    input_scales = torch.empty((token_num, 1),
                               device=input.device,
                               dtype=torch.float32)
    input_azp = None if symmetric else torch.empty_like(input_scales,
                                                        dtype=torch.int32)
    torch.ops._C.dynamic_scaled_int8_quant(output, input, input_scales,
                                           input_azp)
    return output, input_scales, input_azp


def onednn_scaled_mm(
    dnnl_handler: CPUDNNLGEMMHandler,
    x: torch.Tensor,
    output: torch.Tensor,
    input_scale: Optional[torch.Tensor],
    input_zp: Optional[torch.Tensor],
    input_zp_adj: Optional[torch.Tensor],
    bias: Optional[torch.Tensor],
) -> torch.Tensor:
    torch.ops._C.onednn_scaled_mm(output, x, input_scale, input_zp,
                                  input_zp_adj, bias, dnnl_handler.handler)

    return output
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039


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

    Note that sylvester hadamard transforms are also symmetric, which means that
    this function is also applies the (transpose <=> inverse) transform.
    
    :param x: value to be transformed inplace
    :param inplace: modify value in place
    :return: value after transformation
    """
    return torch.ops._C.hadacore_transform(x, inplace)


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

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