_custom_ops.py 66.8 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
    fp8_out_scale: Optional[torch.Tensor] = None,
121
122
123
124
) -> 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,
125
126
                                      query_start_loc, block_size, max_seq_len,
                                      alibi_slopes, kv_cache_dtype, k_scale,
127
                                      v_scale, fp8_out_scale)
128
129


Thien Tran's avatar
Thien Tran committed
130
131
132
133
134
135
136
137
138
139
140
141
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)


142
143
144
145
146
147
148
149
150
151
152
# 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)


153
154
155
156
# pos encoding ops
def rotary_embedding(
    positions: torch.Tensor,
    query: torch.Tensor,
157
    key: Optional[torch.Tensor],
158
159
160
161
    head_size: int,
    cos_sin_cache: torch.Tensor,
    is_neox: bool,
) -> None:
162
163
    # TODO: Remove this contiguous call when the kernel is updated to support tensor slices
    query_contiguous = query.contiguous()
164
    key_contiguous = key.contiguous() if key is not None else None
165
166
167
    torch.ops._C.rotary_embedding(positions, query_contiguous, key_contiguous,
                                  head_size, cos_sin_cache, is_neox)
    query.copy_(query_contiguous)
168
169
    if key is not None:
        key.copy_(key_contiguous)
170
171
172


def batched_rotary_embedding(positions: torch.Tensor, query: torch.Tensor,
173
                             key: Optional[torch.Tensor], head_size: int,
174
175
176
                             cos_sin_cache: torch.Tensor, is_neox: bool,
                             rot_dim: int,
                             cos_sin_cache_offsets: torch.Tensor) -> None:
177
178
    # TODO: Remove this contiguous call when the kernel is updated to support tensor slices
    query_contiguous = query.contiguous()
179
    key_contiguous = key.contiguous() if key is not None else None
180
181
    torch.ops._C.batched_rotary_embedding(positions, query_contiguous,
                                          key_contiguous, head_size,
182
183
                                          cos_sin_cache, is_neox, rot_dim,
                                          cos_sin_cache_offsets)
184
    query.copy_(query_contiguous)
185
186
    if key is not None:
        key.copy_(key_contiguous)
187
188
189
190
191


# layer norm ops
def rms_norm(out: torch.Tensor, input: torch.Tensor, weight: torch.Tensor,
             epsilon: float) -> None:
192
193
194
    # 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)
195
196
197
198


def fused_add_rms_norm(input: torch.Tensor, residual: torch.Tensor,
                       weight: torch.Tensor, epsilon: float) -> None:
199
    torch.ops._C.fused_add_rms_norm(input, residual, weight, epsilon)
200
201


202
203
204
205
206
207
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:
208
    """Advance a step on GPU for existing inputs for a multi-step runner"""
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
    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)
232
233


234
235
236
237
238
239
240
241
# fused quant layer norm ops
def rms_norm_dynamic_per_token_quant(
    input: torch.Tensor,
    weight: torch.Tensor,
    epsilon: float,
    quant_dtype: torch.dtype,
    scale_ub: Optional[torch.Tensor] = None,
    residual: Optional[torch.Tensor] = None
242
) -> tuple[torch.Tensor, torch.Tensor]:
243
244
245
246
247
248
249
250
251
252
253
    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


254
255
256
257
258
# 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:
259
260
261
262
    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)
263
264
    return torch.ops._C.awq_dequantize(qweight, scales, zeros, split_k_iters,
                                       thx, thy)
265
266
267
268


def awq_gemm(input: torch.Tensor, qweight: torch.Tensor, qzeros: torch.Tensor,
             scales: torch.Tensor, split_k_iters: int) -> torch.Tensor:
269
270
271
272
    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)
273
    return torch.ops._C.awq_gemm(input, qweight, qzeros, scales, split_k_iters)
274
275
276
277
278
279
280


# 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:
281
282
    return torch.ops._C.gptq_gemm(a, b_q_weight, b_gptq_qzeros, b_gptq_scales,
                                  b_g_idx, use_exllama, bit)
283
284


285
if hasattr(torch.ops._C, "gptq_gemm"):
286

287
    @register_fake("_C::gptq_gemm")
288
289
290
291
292
293
294
295
296
    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)


297
298
def gptq_shuffle(q_weight: torch.Tensor, q_perm: torch.Tensor,
                 bit: int) -> None:
299
    torch.ops._C.gptq_shuffle(q_weight, q_perm, bit)
300
301
302
303
304
305


# 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:
306
307
    return torch.ops._C.marlin_gemm(a, b_q_weight, b_scales, workspace, size_m,
                                    size_n, size_k)
308
309


310
311
312
# marlin_24
def gptq_marlin_24_gemm(a: torch.Tensor, b_q_weight: torch.Tensor,
                        b_meta: torch.Tensor, b_scales: torch.Tensor,
313
314
                        workspace: torch.Tensor, b_q_type: ScalarType,
                        size_m: int, size_n: int, size_k: int) -> torch.Tensor:
315
    return torch.ops._C.gptq_marlin_24_gemm(a, b_q_weight, b_meta, b_scales,
316
                                            workspace, b_q_type.id, size_m,
317
                                            size_n, size_k)
318
319


320
if hasattr(torch.ops._C, "gptq_marlin_24_gemm"):
321

322
    @register_fake("_C::gptq_marlin_24_gemm")
323
324
325
    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,
326
327
328
                                  b_q_type: ScalarType, size_m: torch.SymInt,
                                  size_n: torch.SymInt,
                                  size_k: torch.SymInt) -> torch.Tensor:
329
330
        return torch.empty((size_m, size_n), device=a.device, dtype=a.dtype)

331
    @register_fake("_C::gptq_marlin_gemm")
332
    def _gptq_marlin_gemm_fake(a: torch.Tensor,
333
                               c: Optional[torch.Tensor],
334
335
                               b_q_weight: torch.Tensor,
                               b_scales: torch.Tensor,
336
                               global_scale: Optional[torch.Tensor],
337
338
339
                               b_zeros: Optional[torch.Tensor],
                               g_idx: Optional[torch.Tensor],
                               perm: Optional[torch.Tensor],
340
                               workspace: torch.Tensor,
341
                               b_q_type_id: int,
342
343
344
                               size_m: torch.SymInt,
                               size_n: torch.SymInt,
                               size_k: torch.SymInt,
345
                               is_k_full: bool = True,
346
                               use_atomic_add: bool = False,
347
348
                               use_fp32_reduce: bool = False,
                               is_zp_float: bool = False) -> torch.Tensor:
349
350
        return torch.empty((size_m, size_n), device=a.device, dtype=a.dtype)

351
    @register_fake("_C::marlin_qqq_gemm")
352
353
354
    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,
355
356
                              size_m: torch.SymInt, size_n: torch.SymInt,
                              size_k: torch.SymInt) -> torch.Tensor:
357
358
359
360
        return torch.empty((size_m, size_n),
                           dtype=torch.float16,
                           device=a.device)

361
    @register_fake("_C::marlin_gemm")
362
363
    def _marlin_gemm_fake(a: torch.Tensor, b_q_weight: torch.Tensor,
                          b_scales: torch.Tensor, workspace: torch.Tensor,
364
365
                          size_m: torch.SymInt, size_n: torch.SymInt,
                          size_k: torch.SymInt) -> torch.Tensor:
366
367
368
369
        return torch.empty((size_m, size_n),
                           dtype=torch.float16,
                           device=a.device)

370
    @register_fake("_C::awq_dequantize")
371
    def _awq_dequantize_fake(qweight: torch.Tensor, scales: torch.Tensor,
372
373
                             zeros: torch.Tensor, split_k_iters: torch.SymInt,
                             thx: int, thy: int) -> torch.Tensor:
374
375
376
377
378
379
380
        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)

381
    @register_fake("_C::awq_gemm")
382
383
    def _awq_gemm_fake(input: torch.Tensor, qweight: torch.Tensor,
                       qzeros: torch.Tensor, scales: torch.Tensor,
384
                       split_k_iters: torch.SymInt) -> torch.Tensor:
385
386
387
388
389
        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)

390
    @register_fake("_C::aqlm_gemm")
391
392
    def _aqlm_gemm_fake(input: torch.Tensor, codes: torch.Tensor,
                        codebooks: torch.Tensor, scales: torch.Tensor,
393
                        codebook_partition_sizes: list[int],
394
395
396
397
398
399
400
401
402
403
404
405
                        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))

406
    @register_fake("_C::aqlm_dequant")
407
408
    def _aqlm_dequant_fake(
            codes: torch.Tensor, codebooks: torch.Tensor,
409
            codebook_partition_sizes: list[int]) -> torch.Tensor:
410
411
412
413
414
415
        in_features = codes.size(1) * 8
        out_features = codes.size(0)
        return torch.empty((out_features, in_features),
                           dtype=codebooks.dtype,
                           device=codebooks.device)

416
417
    @register_fake("_C::machete_mm")
    def machete_mm_fake(
418
        a: torch.Tensor,
419
        # b_q Should be the tensor returned by machete_prepack_B
420
        b_q: torch.Tensor,
421
        b_type: ScalarType,
422
423
424
        out_type: Optional[torch.dtype] = None,
        b_group_scales: Optional[torch.Tensor] = None,
        b_group_zeros: Optional[torch.Tensor] = None,
425
        b_group_size: Optional[int] = None,
426
427
        b_channel_scales: Optional[torch.Tensor] = None,
        a_token_scales: Optional[torch.Tensor] = None,
428
429
430
431
432
433
        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)

434
    @register_fake("_C::machete_prepack_B")
435
436
437
    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:
438
439
        return torch.empty_like(b_q_weight,
                                memory_format=torch.contiguous_format)
440
441


442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
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)


458
459
460
if hasattr(torch.ops._C, "ggml_dequantize"):

    @register_fake("_C::ggml_dequantize")
461
462
463
464
465
466
    def _ggml_dequantize_fake(
            W: torch.Tensor,
            quant_type: int,
            m: torch.SymInt,
            n: torch.SymInt,
            dtype: Optional[torch.dtype] = None) -> torch.Tensor:
467
468
469
470
471
472
473
474
475
        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:
476
        return torch.empty((1, row), dtype=X.dtype, device=W.device)
477
478
479
480
481
482
483
484
485

    @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)
486
        return torch.empty((batch, row), dtype=X.dtype, device=W.device)
487

488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
    @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)

505

506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
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)


524
# cutlass
525
526
527
528
def cutlass_scaled_mm_supports_fp4(cuda_device_capability: int) -> bool:
    return torch.ops._C.cutlass_scaled_mm_supports_fp4(cuda_device_capability)


529
530
531
532
533
534
535
536
537
538
539
540
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


541
542
543
544
def cutlass_scaled_mm_supports_fp8(cuda_device_capability: int) -> bool:
    return torch.ops._C.cutlass_scaled_mm_supports_fp8(cuda_device_capability)


545
546
547
548
549
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)


550
551
552
def cutlass_scaled_mm(a: torch.Tensor,
                      b: torch.Tensor,
                      scale_a: torch.Tensor,
553
                      scale_b: torch.Tensor,
554
                      out_dtype: torch.dtype,
555
                      bias: Optional[torch.Tensor] = None) -> torch.Tensor:
556
    """
557
    `cutlass_scaled_mm` implements a fused version of
558
        `output = torch.mm((scale_a * a), (scale_b * b)).to(out_dtype)`
559
560
561
562
563
564
565
566
    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
567
568
569
570
571
572
573
574
575
576
577
        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
    """
578
    assert (out_dtype is torch.bfloat16 or out_dtype is torch.float16)
579
580
    assert bias is None or bias.shape[0] == b.shape[
        1] and bias.dtype == out_dtype
581
582
583

    m = a.shape[0]
    n = b.shape[1]
584

585
586
    cutlass_compatible_b = (b.shape[0] % 16 == 0 and b.shape[1] % 16 == 0)
    if current_platform.is_rocm() or not cutlass_compatible_b:
587
588
589
590
591
592
        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)

593
594
    out = torch.empty((m, n), dtype=out_dtype, device=a.device)

595
596
    torch.ops._C.cutlass_scaled_mm(out, a, b, scale_a, scale_b, bias)

597
598
599
    return out


600
601
602
603
604
605
606
607
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:
608
609
610
611
612
    """
    :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.
    """
613
614
615
616
    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
617
    assert azp is None or azp.numel() == a.shape[0]
618
619
620
621
622
623
624
625
626
627

    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


628
629
630
631
632
def cutlass_sparse_scaled_mm_supported(cuda_device_capability: int) -> bool:
    return torch.ops._C.cutlass_sparse_scaled_mm_supported(
        cuda_device_capability)


633
634
635
def cutlass_group_gemm_supported(cuda_device_capability: int) -> bool:
    return torch.ops._C.cutlass_group_gemm_supported(cuda_device_capability)

636
def cutlass_sparse_compress(a: torch.Tensor) \
637
    -> tuple[torch.Tensor, torch.Tensor]:
638
639
640
641
642
643
644
645
    """
    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:
646
        a (torch.Tensor):
647
648
649
650
651
652
653
            The input tensor to be compressed. Must have one of the following data types:
            - `torch.int8`
            - `torch.float8_e4m3fn`
            - `torch.bfloat16`
            - `torch.float16`

    Returns:
654
        tuple[torch.Tensor, torch.Tensor]:
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
            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
675
    assert (a.shape[1] % (2 * elemsPerMetaElem) == 0)
676

677
    return torch.ops._C.cutlass_sparse_compress(a)
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725


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


726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
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.
    """
749
750
751
752
753
    return torch.ops._C.get_cutlass_moe_mm_data(topk_ids, expert_offsets,
                                                problem_sizes1, problem_sizes2,
                                                input_permutation,
                                                output_permutation,
                                                num_experts, n, k)
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771


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.
    """
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
    return torch.ops._C.cutlass_moe_mm(out_tensors, a_tensors, b_tensors,
                                       a_scales, b_scales, expert_offsets,
                                       problem_sizes, a_strides, b_strides,
                                       c_strides)


def cutlass_fp4_moe_mm(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,
                       out_dtype: torch.dtype, device: torch.device):
    """
    An FP4 Blockscaled Group Gemm that takes in  a_tensors, b_tensors and runs 
    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
    - 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 
                    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.
    """
    m_topk = a_tensors.shape[0]
    n = b_tensors.shape[1]
    c_shape = (m_topk, n)
    c = torch.empty(c_shape, device=device, dtype=out_dtype)
    torch.ops._C.cutlass_fp4_group_mm(c, a_tensors, b_tensors, a_scales,
                                      b_scales, alphas, problem_sizes,
                                      expert_offsets, sf_offsets)
    return c.to(out_dtype)
807
808


809
810
811
# aqlm
def aqlm_gemm(input: torch.Tensor, codes: torch.Tensor,
              codebooks: torch.Tensor, scales: torch.Tensor,
812
              codebook_partition_sizes: list[int],
813
              bias: Optional[torch.Tensor]) -> torch.Tensor:
814
815
    return torch.ops._C.aqlm_gemm(input, codes, codebooks, scales,
                                  codebook_partition_sizes, bias)
816
817
818


def aqlm_dequant(codes: torch.Tensor, codebooks: torch.Tensor,
819
                 codebook_partition_sizes: list[int]) -> torch.Tensor:
820
821
    return torch.ops._C.aqlm_dequant(codes, codebooks,
                                     codebook_partition_sizes)
822
823


824
825
# gptq_marlin
def gptq_marlin_repack(b_q_weight: torch.Tensor, perm: torch.Tensor,
826
827
                       size_k: int, size_n: int,
                       num_bits: int) -> torch.Tensor:
828
829
    return torch.ops._C.gptq_marlin_repack(b_q_weight, perm, size_k, size_n,
                                           num_bits)
830
831


832
833
834
835
836
837
# 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)


838
839
840
841
842
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
843
    output = torch.empty((num_experts, size_k // 16, size_n * (num_bits // 2)),
844
845
846
847
848
849
850
851
                         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


852
853
854
855
856
857
858
859
860
861
862
863
864
865
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


866
def gptq_marlin_gemm(a: torch.Tensor,
867
                     c: Optional[torch.Tensor],
868
869
                     b_q_weight: torch.Tensor,
                     b_scales: torch.Tensor,
870
                     global_scale: Optional[torch.Tensor],
871
872
873
                     b_zeros: Optional[torch.Tensor],
                     g_idx: Optional[torch.Tensor],
                     perm: Optional[torch.Tensor],
874
875
876
877
878
                     workspace: torch.Tensor,
                     b_q_type: ScalarType,
                     size_m: int,
                     size_n: int,
                     size_k: int,
879
                     is_k_full: bool = True,
880
                     use_atomic_add: bool = False,
881
882
                     use_fp32_reduce: bool = False,
                     is_zp_float: bool = False) -> torch.Tensor:
883
884
885
886
    return torch.ops._C.gptq_marlin_gemm(a, c, b_q_weight, b_scales,
                                         global_scale, b_zeros, g_idx, perm,
                                         workspace, b_q_type.id, size_m,
                                         size_n, size_k, is_k_full,
887
888
                                         use_atomic_add, use_fp32_reduce,
                                         is_zp_float)
889
890


891
# machete
892
893
894
895
896
897
898
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,
899
        out_type: Optional[torch.dtype] = None) -> list[str]:
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
    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)
927
928


929
if hasattr(torch.ops._C, "permute_cols"):
930

931
    @register_fake("_C::permute_cols")
932
933
934
935
936
937
938
939
940
    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)


941
942
943
# fp4
def scaled_fp4_quant(
        input: torch.Tensor,
944
        input_global_scale: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
945
946
947
948
949
950
951
952
953
954
955
956
957
958
    """
    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:
959
        tuple[torch.Tensor, torch.Tensor]: The output tensor in FP4 but every
960
961
962
            two values are packed into a uint8 and float8_e4m3 scaling factors
            in the sizzled layout.
    """
963
    assert not current_platform.is_rocm()
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
    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


999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
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
def scaled_fp4_experts_quant(
    input_tensor: torch.Tensor,
    input_global_scale: torch.Tensor,
    expert_offsets: torch.Tensor,
    blockscale_offsets: torch.Tensor,
    topk: int,
    expert_map: Optional[torch.Tensor] = None,
    MAX_TOKENS_PER_EXPERT: int = 163840,
) -> tuple[torch.Tensor, torch.Tensor]:
    """
    Quantize input tensor to FP4 and return quantized tensor and scale, for
    packed MoE Inputs.
    Args:
        input: The input tensor to be quantized to FP4
        expert_map: The expert map tensor
        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}.')

    input_tensor = input_tensor[
        expert_map] if expert_map is not None else input_tensor
    m_numtopk, k = input_tensor.shape
    assert (m_numtopk <= MAX_TOKENS_PER_EXPERT * topk), (
        f"m_numtopk must be less than MAX_TOKENS_PER_EXPERT * topk for"
        f" scaled_fp4_experts_quant kernel, observed m_numtopk = {m_numtopk}")
    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


1050
# fp8
1051
1052
1053
def scaled_fp8_quant(
    input: torch.Tensor,
    scale: Optional[torch.Tensor] = None,
1054
    num_token_padding: Optional[int] = None,
1055
    scale_ub: Optional[torch.Tensor] = None,
1056
    use_per_token_if_dynamic: bool = False,
1057
) -> tuple[torch.Tensor, torch.Tensor]:
1058
1059
1060
1061
1062
1063
    """
    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
1064
    optional padding of the output tensors for downstream kernels that
1065
1066
1067
1068
1069
    will benefit from padding.

    Args:
        input: The input tensor to be quantized to FP8
        scale: Optional scaling factor for the FP8 quantization
1070
        scale_ub: Optional upper bound for scaling factor in dynamic
1071
            per token case
1072
        num_token_padding: If specified, pad the first dimension
1073
            of the output to at least this value.
1074
        use_per_token_if_dynamic: Whether to do per_tensor or per_token
1075
            in the dynamic quantization case.
1076
1077

    Returns:
1078
        tuple[torch.Tensor, torch.Tensor]: The output tensor in FP8 and
1079
1080
            scaling factor.
    """
1081
1082
    # This code assumes batch_dim and num_tokens are flattened
    assert (input.ndim == 2)
1083
    shape: Union[tuple[int, int], torch.Size] = input.shape
1084
1085
    # For ROCm on MI300, the output fp8 dtype is torch.float_e3m3fnuz
    out_dtype: torch.dtype = current_platform.fp8_dtype()
1086
1087
    if num_token_padding:
        shape = (max(num_token_padding, input.shape[0]), shape[1])
1088
    output = torch.empty(shape, device=input.device, dtype=out_dtype)
1089

1090
    if scale is None:
1091
        if use_per_token_if_dynamic:
1092
            scale = torch.empty((shape[0], 1),
1093
1094
1095
                                device=input.device,
                                dtype=torch.float32)
            torch.ops._C.dynamic_per_token_scaled_fp8_quant(
1096
                output, input, scale, scale_ub)
1097
1098
1099
        else:
            scale = torch.zeros(1, device=input.device, dtype=torch.float32)
            torch.ops._C.dynamic_scaled_fp8_quant(output, input, scale)
1100
    else:
1101
1102
        # num_token_padding not implemented for this case
        assert (scale.numel() == 1 or num_token_padding is None)
1103
        torch.ops._C.static_scaled_fp8_quant(output, input, scale)
1104

1105
    return output, scale
1106
1107


1108
1109
1110
1111
1112
1113
# gptq allspark
def allspark_repack_weight(
        qweight: torch.Tensor,
        scale: torch.Tensor,
        zero_point: Optional[torch.Tensor] = None,
        has_zp: bool = False
1114
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
1115
    """
1116
    Rearrange qweight, scale, and zero_point(if asymmetric) to n32k16 format
1117
1118
1119
1120
1121
1122
1123
1124
    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.
1125
1126
            if use asymmetric quantization, has_zp = True.

1127
    Returns:
1128
        tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]] :
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
            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)


1169
# int8
1170
def scaled_int8_quant(
1171
1172
1173
1174
    input: torch.Tensor,
    scale: Optional[torch.Tensor] = None,
    azp: Optional[torch.Tensor] = None,
    symmetric: bool = True
1175
) -> tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
1176
    """
1177
    Quantize the input tensor to int8 and return the quantized tensor and scale, and maybe azp.
1178
1179
1180

    Args:
        input: The input tensor to be quantized to int8.
1181
1182
        scale: Optional scaling factor for the int8 quantization.
            When not provided, we invoke dynamic-per-token quantization.
1183
1184
1185
        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).
1186
1187

    Returns:
1188
      tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]] : Output int8 tensor, scales, and optionally azp.
1189
    """
1190
1191
1192
    output = torch.empty_like(input, dtype=torch.int8)
    if scale is not None:
        # static-per-tensor quantization.
1193
        assert symmetric == (
1194
1195
            azp
            is None), "azp must only be provided for asymmetric quantization."
1196
        torch.ops._C.static_scaled_int8_quant(output, input, scale, azp)
1197
        return output, scale, azp
1198
1199
1200
1201
1202

    # dynamic-per-token quantization.
    input_scales = torch.empty((input.numel() // input.shape[-1], 1),
                               device=input.device,
                               dtype=torch.float32)
1203
1204
1205
1206
1207
    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
1208
1209


1210
1211
1212
1213
1214
1215
1216
1217
1218
# 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)


1219
# gguf
1220
1221
1222
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)
1223
1224
1225
1226
1227
1228
1229


def ggml_mul_mat_vec_a8(
    W: torch.Tensor,
    X: torch.Tensor,
    quant_type: int,
    row: int,
1230
) -> torch.Tensor:
1231
1232
1233
1234
1235
1236
1237
1238
    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,
1239
) -> torch.Tensor:
1240
1241
1242
    return torch.ops._C.ggml_mul_mat_a8(W, X, quant_type, row)


1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
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)


1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
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)


1272
1273
1274
1275
def ggml_moe_get_block_size(quant_type: int) -> int:
    return torch.ops._C.ggml_moe_get_block_size(quant_type)


1276
1277
1278
# mamba
def causal_conv1d_fwd(x: torch.Tensor, weight: torch.Tensor,
                      bias_: Optional[torch.Tensor],
1279
1280
1281
1282
                      conv_states: Optional[torch.Tensor],
                      query_start_loc: Optional[torch.Tensor],
                      cache_indices: Optional[torch.Tensor],
                      has_initial_state: Optional[torch.Tensor],
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
                      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):
1310
1311
1312
    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,
1313
                                    ssm_states, pad_slot_id)
1314
1315


1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
# 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)


def wvSplitK(a: torch.Tensor, b: torch.Tensor, cu_count: int) -> torch.Tensor:
    return torch.ops._rocm_C.wvSplitK(a, b, 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) -> torch.Tensor:
    out = torch.empty((b.shape[0], a.shape[0]),
                      dtype=out_dtype,
                      device=b.device)
    torch.ops._rocm_C.wvSplitKQ(a, b, out, scale_a, scale_b, cu_count)
    return out


1336
# moe
1337
1338
1339
1340
def moe_sum(input: torch.Tensor, output: torch.Tensor):
    torch.ops._moe_C.moe_sum(input, output)


1341
1342
1343
1344
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:
1345
1346
1347
    torch.ops._moe_C.moe_align_block_size(topk_ids, num_experts, block_size,
                                          sorted_token_ids, experts_ids,
                                          num_tokens_post_pad)
1348
1349


1350
1351
1352
1353
1354
1355
1356
1357
1358
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)


1359
1360
1361
1362
1363
1364
1365
1366
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:
1367
1368
1369
1370
    if not current_platform.is_cuda():
        raise NotImplementedError(
            "The optimized moe_wna16_gemm kernel is only "
            "available on CUDA platforms")
1371
1372
1373
1374
1375
1376
1377
    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)


1378
1379
def topk_softmax(topk_weights: torch.Tensor, topk_ids: torch.Tensor,
                 token_expert_indicies: torch.Tensor,
1380
                 gating_output: torch.Tensor) -> None:
1381
1382
    torch.ops._moe_C.topk_softmax(topk_weights, topk_ids,
                                  token_expert_indicies, gating_output)
1383
1384


1385
1386
def moe_wna16_marlin_gemm(input: torch.Tensor, output: Optional[torch.Tensor],
                          b_qweight: torch.Tensor, b_scales: torch.Tensor,
1387
                          global_scale: Optional[torch.Tensor],
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
                          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(
1402
1403
1404
1405
1406
        input, output, b_qweight, 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)
1407
1408


1409
1410
if supports_moe_ops and hasattr(torch.ops._moe_C, "marlin_gemm_moe"):

1411
    @register_fake("_moe_C::marlin_gemm_moe")
1412
1413
1414
1415
    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,
1416
1417
                             b_zero_points: torch.Tensor, g_idx: torch.Tensor,
                             perm: torch.Tensor, workspace: torch.Tensor,
1418
1419
1420
1421
                             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,
1422
1423
1424
1425
1426
                             apply_weights: bool) -> torch.Tensor:
        return torch.empty((size_m, topk, size_n),
                           dtype=a.dtype,
                           device=a.device)

1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
    @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)

1450

1451
1452
1453
1454
1455
1456
1457
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,
1458
1459
    k_scale: torch.Tensor,
    v_scale: torch.Tensor,
1460
) -> None:
1461
1462
    torch.ops._C_cache_ops.reshape_and_cache(key, value, key_cache,
                                             value_cache, slot_mapping,
1463
                                             kv_cache_dtype, k_scale, v_scale)
1464
1465


1466
1467
1468
1469
1470
1471
1472
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,
1473
1474
    k_scale: torch.Tensor,
    v_scale: torch.Tensor,
1475
) -> None:
1476
1477
    torch.ops._C_cache_ops.reshape_and_cache_flash(key, value, key_cache,
                                                   value_cache, slot_mapping,
1478
1479
                                                   kv_cache_dtype, k_scale,
                                                   v_scale)
1480
1481


1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
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)


1495
1496
def copy_blocks(key_caches: list[torch.Tensor],
                value_caches: list[torch.Tensor],
1497
                block_mapping: torch.Tensor) -> None:
1498
    torch.ops._C_cache_ops.copy_blocks(key_caches, value_caches, block_mapping)
1499
1500


1501
def copy_blocks_mla(kv_caches: list[torch.Tensor],
1502
1503
1504
1505
                    block_mapping: torch.Tensor) -> None:
    torch.ops._C_cache_ops.copy_blocks_mla(kv_caches, block_mapping)


1506
def swap_blocks(src: torch.Tensor, dst: torch.Tensor,
1507
                block_mapping: torch.Tensor) -> None:
1508
    torch.ops._C_cache_ops.swap_blocks(src, dst, block_mapping)
1509
1510


1511
1512
1513
1514
def convert_fp8(output: torch.Tensor,
                input: torch.Tensor,
                scale: float = 1.0,
                kv_dtype: str = "fp8") -> None:
1515
1516
1517
    torch.ops._C_cache_ops.convert_fp8(output, input, scale, kv_dtype)


1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
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)


1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
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
1539
def init_custom_ar(ipc_tensors: list[torch.Tensor], rank_data: torch.Tensor,
1540
                   rank: int, fully_connected: bool) -> int:
1541
    return torch.ops._C_custom_ar.init_custom_ar(ipc_tensors, rank_data, rank,
1542
                                                 fully_connected)
1543
1544


1545
1546
1547
1548
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)
1549

1550
1551
1552
1553
1554
1555
1556
1557
1558

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


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


1559
def register_buffer(fa: int, ipc_tensors: list[int]) -> None:
1560
    return torch.ops._C_custom_ar.register_buffer(fa, ipc_tensors)
1561
1562


1563
def get_graph_buffer_ipc_meta(fa: int) -> tuple[list[int], list[int]]:
1564
1565
1566
    return torch.ops._C_custom_ar.get_graph_buffer_ipc_meta(fa)


1567
1568
def register_graph_buffers(fa: int, handles: list[list[int]],
                           offsets: list[list[int]]) -> None:
1569
    torch.ops._C_custom_ar.register_graph_buffers(fa, handles, offsets)
1570
1571


1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
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)


1584
1585
1586
1587
def get_flash_mla_metadata(
    cache_seqlens: torch.Tensor,
    num_heads_per_head_k: int,
    num_heads_k: int,
1588
) -> tuple[torch.Tensor, torch.Tensor]:
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
    """
    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,
1614
) -> tuple[torch.Tensor, torch.Tensor]:
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
    """
    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
1646
1647
1648
1649
1650
1651
1652
1653
1654


def cutlass_mla_decode(out: torch.Tensor, q_nope: torch.Tensor,
                       q_pe: torch.Tensor, kv_c_and_k_pe_cache: torch.Tensor,
                       seq_lens: torch.Tensor, page_table: torch.Tensor,
                       scale: float) -> torch.Tensor:
    torch.ops._C.cutlass_mla_decode(out, q_nope, q_pe, kv_c_and_k_pe_cache,
                                    seq_lens, page_table, scale)
    return out