_custom_ops.py 61.4 KB
Newer Older
1
2
# SPDX-License-Identifier: Apache-2.0

3
import contextlib
4
import importlib
zhuwenwen's avatar
zhuwenwen committed
5
from typing import TYPE_CHECKING, List, Optional, Tuple, Union, Type
6
7

import torch
8
import torch.library
gaoqiong's avatar
gaoqiong committed
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
try:
gaoqiong's avatar
gaoqiong committed
16
    from lmslim import quant_ops 
gaoqiong's avatar
gaoqiong committed
17
    from lmslim import quant_tools 
18
except Exception:
gaoqiong's avatar
gaoqiong committed
19
    print("INFO: Please install lmslim if you want to infer gptq or awq  or w8a8 model.\n") 
20

21
22
logger = init_logger(__name__)

23
if not current_platform.is_tpu() and not current_platform.is_hpu():
24
25
26
27
    try:
        import vllm._C
    except ImportError as e:
        logger.warning("Failed to import from vllm._C with %r", e)
28

29

30
supports_moe_ops = False
31
with contextlib.suppress(ImportError):
32
    import vllm._moe_C  # noqa: F401
33
    supports_moe_ops = True
34

35
if TYPE_CHECKING:
36
37
38
39
40
41
42
43
44

    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

45
46
47

# page attention ops
def paged_attention_v1(
48
49
50
51
52
53
54
55
56
57
58
59
    out: torch.Tensor,
    query: torch.Tensor,
    key_cache: torch.Tensor,
    value_cache: torch.Tensor,
    num_kv_heads: int,
    scale: float,
    block_tables: torch.Tensor,
    seq_lens: torch.Tensor,
    block_size: int,
    max_seq_len: int,
    alibi_slopes: Optional[torch.Tensor],
    kv_cache_dtype: str,
60
61
    k_scale: torch.Tensor,
    v_scale: torch.Tensor,
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
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
) -> None:
    torch.ops._C.paged_attention_v1(
        out, query, key_cache, value_cache, num_kv_heads, scale, block_tables,
        seq_lens, block_size, max_seq_len, alibi_slopes, kv_cache_dtype,
        k_scale, v_scale, tp_rank, blocksparse_local_blocks,
        blocksparse_vert_stride, blocksparse_block_size,
        blocksparse_head_sliding_step)


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,
    seq_lens: torch.Tensor,
    block_size: int,
    max_seq_len: int,
    alibi_slopes: Optional[torch.Tensor],
    kv_cache_dtype: str,
92
93
    k_scale: torch.Tensor,
    v_scale: torch.Tensor,
94
95
96
97
98
99
100
101
102
103
104
105
106
107
    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
) -> None:
    torch.ops._C.paged_attention_v2(
        out, exp_sum, max_logits, tmp_out, query, key_cache, value_cache,
        num_kv_heads, scale, block_tables, seq_lens, block_size, max_seq_len,
        alibi_slopes, kv_cache_dtype, k_scale, v_scale, tp_rank,
        blocksparse_local_blocks, blocksparse_vert_stride,
        blocksparse_block_size, blocksparse_head_sliding_step)
    
def paged_attention_v1_with_mask(
108
109
110
111
112
113
114
    out: torch.Tensor,
    query: torch.Tensor,
    key_cache: torch.Tensor,
    value_cache: torch.Tensor,
    num_kv_heads: int,
    scale: float,
    block_tables: torch.Tensor,
115
    seq_lens: torch.Tensor,
116
    block_size: int,
117
    max_seq_len: int,
118
119
    alibi_slopes: Optional[torch.Tensor],
    kv_cache_dtype: str,
120
121
    k_scale: torch.Tensor,
    v_scale: torch.Tensor,
122
123
124
125
126
    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,
127
128
    attn_masks: Optional[torch.Tensor] = None,
    attn_masks_stride: int = 0,
129
) -> None:
130
    torch.ops._C.paged_attention_v1_with_mask(
131
132
        out, query, key_cache, value_cache, num_kv_heads, scale, block_tables,
        seq_lens, block_size, max_seq_len, alibi_slopes, kv_cache_dtype,
133
134
        k_scale, v_scale, tp_rank, blocksparse_local_blocks,
        blocksparse_vert_stride, blocksparse_block_size,
135
136
        blocksparse_head_sliding_step,attn_masks, 
        attn_masks_stride)
137
138


139
def paged_attention_v2_with_mask(
140
141
142
143
144
145
146
147
148
149
    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,
150
    seq_lens: torch.Tensor,
151
    block_size: int,
152
    max_seq_len: int,
153
154
    alibi_slopes: Optional[torch.Tensor],
    kv_cache_dtype: str,
155
156
    k_scale: torch.Tensor,
    v_scale: torch.Tensor,
157
158
159
160
161
    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,
162
163
    attn_masks: Optional[torch.Tensor] = None,
    attn_masks_stride: int = 0,
164
) -> None:
165
    torch.ops._C.paged_attention_v2_with_mask(
166
167
        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,
168
        alibi_slopes, kv_cache_dtype, k_scale, v_scale, tp_rank,
169
        blocksparse_local_blocks, blocksparse_vert_stride,
170
171
        blocksparse_block_size, blocksparse_head_sliding_step,
        attn_masks, attn_masks_stride)
172
173


zhuwenwen's avatar
zhuwenwen committed
174
175
# page attention ops (opt)
def paged_attention_v1_opt(
176
177
178
179
180
181
182
183
184
185
186
187
    out: torch.Tensor,
    query: torch.Tensor,
    key_cache: torch.Tensor,
    value_cache: torch.Tensor,
    num_kv_heads: int,
    scale: float,
    block_tables: torch.Tensor,
    seq_lens: torch.Tensor,
    block_size: int,
    max_seq_len: int,
    alibi_slopes: Optional[torch.Tensor],
    kv_cache_dtype: str,
188
189
    k_scale: torch.Tensor,
    v_scale: torch.Tensor,
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
    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
) -> None:
    torch.ops._C.paged_attention_v1_opt(
        out, query, key_cache, value_cache, num_kv_heads, scale, block_tables,
        seq_lens, block_size, max_seq_len, alibi_slopes, kv_cache_dtype,
        k_scale, v_scale, tp_rank, blocksparse_local_blocks,
        blocksparse_vert_stride, blocksparse_block_size,
        blocksparse_head_sliding_step)


def paged_attention_v2_opt(
    out: torch.Tensor,
    exp_sum: torch.Tensor,
    max_logits: torch.Tensor,
    tmp_out: torch.Tensor,
    query: torch.Tensor,
    key_cache: torch.Tensor,
    value_cache: torch.Tensor,
    num_kv_heads: int,
    scale: float,
    block_tables: torch.Tensor,
    seq_lens: torch.Tensor,
    block_size: int,
    max_seq_len: int,
    alibi_slopes: Optional[torch.Tensor],
    kv_cache_dtype: str,
220
221
    k_scale: torch.Tensor,
    v_scale: torch.Tensor,
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
    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
) -> None:
    torch.ops._C.paged_attention_v2_opt(
        out, exp_sum, max_logits, tmp_out, query, key_cache, value_cache,
        num_kv_heads, scale, block_tables, seq_lens, block_size, max_seq_len,
        alibi_slopes, kv_cache_dtype, k_scale, v_scale, tp_rank,
        blocksparse_local_blocks, blocksparse_vert_stride,
        blocksparse_block_size, blocksparse_head_sliding_step)


def paged_attention_v1_opt_with_mask(
zhuwenwen's avatar
zhuwenwen committed
237
238
239
240
241
242
243
244
245
246
247
248
    out: torch.Tensor,
    query: torch.Tensor,
    key_cache: torch.Tensor,
    value_cache: torch.Tensor,
    num_kv_heads: int,
    scale: float,
    block_tables: torch.Tensor,
    seq_lens: torch.Tensor,
    block_size: int,
    max_seq_len: int,
    alibi_slopes: Optional[torch.Tensor],
    kv_cache_dtype: str,
249
250
    k_scale: torch.Tensor,
    v_scale: torch.Tensor,
zhuwenwen's avatar
zhuwenwen committed
251
252
253
254
255
    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,
256
257
    attn_masks: Optional[torch.Tensor] = None,
    attn_masks_stride: int = 0,
zhuwenwen's avatar
zhuwenwen committed
258
) -> None:
259
    torch.ops._C.paged_attention_v1_opt_with_mask(
zhuwenwen's avatar
zhuwenwen committed
260
261
262
263
        out, query, key_cache, value_cache, num_kv_heads, scale, block_tables,
        seq_lens, block_size, max_seq_len, alibi_slopes, kv_cache_dtype,
        k_scale, v_scale, tp_rank, blocksparse_local_blocks,
        blocksparse_vert_stride, blocksparse_block_size,
264
265
        blocksparse_head_sliding_step, attn_masks,
        attn_masks_stride)
zhuwenwen's avatar
zhuwenwen committed
266
267


268
def paged_attention_v2_opt_with_mask(
zhuwenwen's avatar
zhuwenwen committed
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
    out: torch.Tensor,
    exp_sum: torch.Tensor,
    max_logits: torch.Tensor,
    tmp_out: torch.Tensor,
    query: torch.Tensor,
    key_cache: torch.Tensor,
    value_cache: torch.Tensor,
    num_kv_heads: int,
    scale: float,
    block_tables: torch.Tensor,
    seq_lens: torch.Tensor,
    block_size: int,
    max_seq_len: int,
    alibi_slopes: Optional[torch.Tensor],
    kv_cache_dtype: str,
284
285
    k_scale: torch.Tensor,
    v_scale: torch.Tensor,
zhuwenwen's avatar
zhuwenwen committed
286
287
288
289
290
    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,
291
292
    attn_masks: Optional[torch.Tensor] = None,
    attn_masks_stride: int = 0,
zhuwenwen's avatar
zhuwenwen committed
293
) -> None:
294
    torch.ops._C.paged_attention_v2_opt_with_mask(
zhuwenwen's avatar
zhuwenwen committed
295
296
297
298
        out, exp_sum, max_logits, tmp_out, query, key_cache, value_cache,
        num_kv_heads, scale, block_tables, seq_lens, block_size, max_seq_len,
        alibi_slopes, kv_cache_dtype, k_scale, v_scale, tp_rank,
        blocksparse_local_blocks, blocksparse_vert_stride,
299
300
        blocksparse_block_size, blocksparse_head_sliding_step,
        attn_masks, attn_masks_stride)
zhuwenwen's avatar
zhuwenwen committed
301
302


303
304
# page attention ops (opt)
def paged_attention_v1_opt_tc(
305
306
307
308
309
310
311
312
313
314
315
316
    out: torch.Tensor,
    query: torch.Tensor,
    key_cache: torch.Tensor,
    value_cache: torch.Tensor,
    num_kv_heads: int,
    scale: float,
    block_tables: torch.Tensor,
    seq_lens: torch.Tensor,
    block_size: int,
    max_seq_len: int,
    alibi_slopes: Optional[torch.Tensor],
    kv_cache_dtype: str,
317
318
    k_scale: torch.Tensor,
    v_scale: torch.Tensor,
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
    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
) -> None:
    torch.ops._C.paged_attention_v1_opt_tc(
        out, query, key_cache, value_cache, num_kv_heads, scale, block_tables,
        seq_lens, block_size, max_seq_len, alibi_slopes, kv_cache_dtype,
        k_scale, v_scale, tp_rank, blocksparse_local_blocks, blocksparse_vert_stride,
        blocksparse_block_size, blocksparse_head_sliding_step)


def paged_attention_v2_opt_tc(
    out: torch.Tensor,
    exp_sum: torch.Tensor,
    max_logits: torch.Tensor,
    tmp_out: torch.Tensor,
    query: torch.Tensor,
    key_cache: torch.Tensor,
    value_cache: torch.Tensor,
    num_kv_heads: int,
    scale: float,
    block_tables: torch.Tensor,
    seq_lens: torch.Tensor,
    block_size: int,
    max_seq_len: int,
    alibi_slopes: Optional[torch.Tensor],
    kv_cache_dtype: str,
348
349
    k_scale: torch.Tensor,
    v_scale: torch.Tensor,
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
    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
) -> None:
    torch.ops._C.paged_attention_v2_opt_tc(
        out, exp_sum, max_logits, tmp_out, query, key_cache, value_cache,
        num_kv_heads, scale, block_tables, seq_lens, block_size, max_seq_len,
        alibi_slopes, kv_cache_dtype, k_scale, v_scale, tp_rank,
        blocksparse_local_blocks, blocksparse_vert_stride,
        blocksparse_block_size, blocksparse_head_sliding_step)
    

# page attention ops (opt)
def paged_attention_v1_opt_tc_with_mask(
366
367
368
369
370
371
372
373
374
375
376
377
    out: torch.Tensor,
    query: torch.Tensor,
    key_cache: torch.Tensor,
    value_cache: torch.Tensor,
    num_kv_heads: int,
    scale: float,
    block_tables: torch.Tensor,
    seq_lens: torch.Tensor,
    block_size: int,
    max_seq_len: int,
    alibi_slopes: Optional[torch.Tensor],
    kv_cache_dtype: str,
378
379
    k_scale: torch.Tensor,
    v_scale: torch.Tensor,
380
381
382
383
384
    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,
385
386
    attn_masks: Optional[torch.Tensor] = None,
    attn_masks_stride: int = 0,
387
) -> None:
388
    torch.ops._C.paged_attention_v1_opt_tc_with_mask(
389
390
        out, query, key_cache, value_cache, num_kv_heads, scale, block_tables,
        seq_lens, block_size, max_seq_len, alibi_slopes, kv_cache_dtype,
zhuwenwen's avatar
zhuwenwen committed
391
        k_scale, v_scale, tp_rank, blocksparse_local_blocks, blocksparse_vert_stride,
392
393
        blocksparse_block_size, blocksparse_head_sliding_step,
        attn_masks, attn_masks_stride)
394
395


396
def paged_attention_v2_opt_tc_with_mask(
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
    out: torch.Tensor,
    exp_sum: torch.Tensor,
    max_logits: torch.Tensor,
    tmp_out: torch.Tensor,
    query: torch.Tensor,
    key_cache: torch.Tensor,
    value_cache: torch.Tensor,
    num_kv_heads: int,
    scale: float,
    block_tables: torch.Tensor,
    seq_lens: torch.Tensor,
    block_size: int,
    max_seq_len: int,
    alibi_slopes: Optional[torch.Tensor],
    kv_cache_dtype: str,
412
413
    k_scale: torch.Tensor,
    v_scale: torch.Tensor,
414
415
416
417
418
    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,
419
420
    attn_masks: Optional[torch.Tensor] = None,
    attn_masks_stride: int = 0,
421
) -> None:
422
    torch.ops._C.paged_attention_v2_opt_tc_with_mask(
423
424
        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,
zhuwenwen's avatar
zhuwenwen committed
425
        alibi_slopes, kv_cache_dtype, k_scale, v_scale, tp_rank,
426
        blocksparse_local_blocks, blocksparse_vert_stride,
427
428
        blocksparse_block_size, blocksparse_head_sliding_step,
        attn_masks, attn_masks_stride)
429
430


431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
def paged_attention_rocm(
    out: torch.Tensor,
    exp_sum: torch.Tensor,
    max_logits: torch.Tensor,
    tmp_out: torch.Tensor,
    query: torch.Tensor,
    key_cache: torch.Tensor,
    value_cache: torch.Tensor,
    num_kv_heads: int,
    scale: float,
    block_tables: torch.Tensor,
    seq_lens: torch.Tensor,
    block_size: int,
    max_seq_len: int,
    alibi_slopes: Optional[torch.Tensor],
    kv_cache_dtype: str,
447
448
    k_scale: torch.Tensor,
    v_scale: torch.Tensor,
449
450
451
452
453
) -> None:
    torch.ops._rocm_C.paged_attention(out, exp_sum, max_logits, tmp_out, query,
                                      key_cache, value_cache, num_kv_heads,
                                      scale, block_tables, seq_lens,
                                      block_size, max_seq_len, alibi_slopes,
454
                                      kv_cache_dtype, k_scale, v_scale)
455
456


457
458
459
460
461
462
463
464
465
# pos encoding ops
def rotary_embedding(
    positions: torch.Tensor,
    query: torch.Tensor,
    key: torch.Tensor,
    head_size: int,
    cos_sin_cache: torch.Tensor,
    is_neox: bool,
) -> None:
466
467
    torch.ops._C.rotary_embedding(positions, query, key, head_size,
                                  cos_sin_cache, is_neox)
468
469
470
471
472
473
474


def batched_rotary_embedding(positions: torch.Tensor, query: torch.Tensor,
                             key: torch.Tensor, head_size: int,
                             cos_sin_cache: torch.Tensor, is_neox: bool,
                             rot_dim: int,
                             cos_sin_cache_offsets: torch.Tensor) -> None:
475
476
477
    torch.ops._C.batched_rotary_embedding(positions, query, key, head_size,
                                          cos_sin_cache, is_neox, rot_dim,
                                          cos_sin_cache_offsets)
478
479
480
481
482


# layer norm ops
def rms_norm(out: torch.Tensor, input: torch.Tensor, weight: torch.Tensor,
             epsilon: float) -> None:
483
    torch.ops._C.rms_norm(out, input, weight, epsilon)
484
485
486
487


def fused_add_rms_norm(input: torch.Tensor, residual: torch.Tensor,
                       weight: torch.Tensor, epsilon: float) -> None:
488
    torch.ops._C.fused_add_rms_norm(input, residual, weight, epsilon)
zhuwenwen's avatar
zhuwenwen committed
489
490
491
492
493
494
495
496
497
498
499
    

# layer norm ops (opt)
def rms_norm_opt(out: torch.Tensor, input: torch.Tensor, weight: torch.Tensor,
             epsilon: float) -> None:
    torch.ops._C.rms_norm_opt(out, input, weight, epsilon)


def fused_add_rms_norm_opt(input: torch.Tensor, residual: torch.Tensor,
                       weight: torch.Tensor, epsilon: float) -> None:
    torch.ops._C.fused_add_rms_norm_opt(input, residual, weight, epsilon)
500
501


502
503
504
505
506
507
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:
508
    """Advance a step on GPU for existing inputs for a multi-step runner"""
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
    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)
532

zhuwenwen's avatar
zhuwenwen committed
533
534
535
536
537
# trans_w16
def trans_w16_gemm(dst: torch.Tensor, src: torch.Tensor,
                row:int, col:int) -> None :
    torch.ops._C.trans_w16_gemm(dst,src,row,col)
    
538

539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
# fused quant layer norm ops
def rms_norm_dynamic_per_token_quant(
    input: torch.Tensor,
    weight: torch.Tensor,
    epsilon: float,
    quant_dtype: torch.dtype,
    scale_ub: Optional[torch.Tensor] = None,
    residual: Optional[torch.Tensor] = None
) -> Tuple[torch.Tensor, torch.Tensor]:
    output = torch.empty_like(input, dtype=quant_dtype)
    scales = torch.empty((input.numel() // input.shape[-1], 1),
                         device=input.device,
                         dtype=torch.float32)

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


559
560
# quantization ops
# awq
zhuwenwen's avatar
zhuwenwen committed
561
562
563
564
565
566
def GetAWQShareWorkspaceSize()->int:
    return quant_ops.GetAWQShareWorkspaceSize()

def GetAWQShareWorkspace()->torch.Tensor:
    return quant_ops.GetAWQShareWorkspace()

567
568
569
def awq_dequantize(qweight: torch.Tensor, scales: torch.Tensor,
                   zeros: torch.Tensor, split_k_iters: int, thx: int,
                   thy: int) -> torch.Tensor:
570
571
572
573
    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)
574
575
    return torch.ops._C.awq_dequantize(qweight, scales, zeros, split_k_iters,
                                       thx, thy)
576
577


gaoqiong's avatar
gaoqiong committed
578
579
# def awq_gemm(input: torch.Tensor, qweight: torch.Tensor, qzeros: torch.Tensor,
#              scales: torch.Tensor, split_k_iters: int) -> torch.Tensor:
zhuwenwen's avatar
zhuwenwen committed
580
581
582
583
584
585
#     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)
#     return torch.ops._C.awq_gemm(input, qweight, qzeros, scales, split_k_iters)

gaoqiong's avatar
gaoqiong committed
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616

def awq_gemm(input: torch.Tensor, weight: torch.Tensor,
             zeros_and_scales:torch.Tensor,
             m:int,n:int,k:int,
             group_size:int,padding_group:int,splikspace:torch.Tensor,
            splikspacesize:int) -> torch.Tensor:
    return quant_ops.awq_gemm(input,
                              weight,
                              zeros_and_scales,
                              m,
                              n,
                              k,
                              group_size,
                              padding_group,
                              splikspace,
                              splikspacesize)

def convert_s4(qw: torch.Tensor, qz: torch.Tensor, s: torch.Tensor,
               group_size: int):
    return quant_ops.convert_s4(qw,qz,s,group_size)

def sz_permute(sz:torch.Tensor)-> torch.Tensor:
    return quant_ops.sz_permute(sz)

def dequant_w4_gemm_colmajor(qweight:torch.Tensor,
                                zeros_and_scale:torch.Tensor,
                                k:int,
                                n:int,
                                group_size:int
                             )->torch.Tensor:
    return quant_ops.dequant_w4_gemm_colmajor(qweight,zeros_and_scale,k,n,group_size)
617
618
619
620
621
622
623


# 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:
gaoqiong's avatar
gaoqiong committed
624
    return quant_ops.gptq_gemm(a, b_q_weight, b_gptq_qzeros, b_gptq_scales,
625
                                  b_g_idx, use_exllama, bit)
626
627
    # return torch.ops._C.gptq_gemm(a, b_q_weight, b_gptq_qzeros, b_gptq_scales,
    #                               b_g_idx, use_exllama, bit)
628
629


630
if hasattr(torch.ops._C, "gptq_gemm"):
631

632
    @register_fake("_C::gptq_gemm")
633
634
635
636
637
638
639
640
641
    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)


642
643
def gptq_shuffle(q_weight: torch.Tensor, q_perm: torch.Tensor,
                 bit: int) -> None:
gaoqiong's avatar
gaoqiong committed
644
    quant_ops.gptq_shuffle(q_weight, q_perm, bit)
645
    # torch.ops._C.gptq_shuffle(q_weight, q_perm, bit)
646
647
648
649
650
651


# 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:
652
653
    return torch.ops._C.marlin_gemm(a, b_q_weight, b_scales, workspace, size_m,
                                    size_n, size_k)
654
655


656
657
658
# marlin_24
def gptq_marlin_24_gemm(a: torch.Tensor, b_q_weight: torch.Tensor,
                        b_meta: torch.Tensor, b_scales: torch.Tensor,
659
660
                        workspace: torch.Tensor, b_q_type: ScalarType,
                        size_m: int, size_n: int, size_k: int) -> torch.Tensor:
661
    return torch.ops._C.gptq_marlin_24_gemm(a, b_q_weight, b_meta, b_scales,
662
                                            workspace, b_q_type.id, size_m,
663
                                            size_n, size_k)
664
665


666
if hasattr(torch.ops._C, "gptq_marlin_24_gemm"):
667

668
    @register_fake("_C::gptq_marlin_24_gemm")
669
670
671
    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,
672
673
674
                                  b_q_type: ScalarType, size_m: torch.SymInt,
                                  size_n: torch.SymInt,
                                  size_k: torch.SymInt) -> torch.Tensor:
675
676
        return torch.empty((size_m, size_n), device=a.device, dtype=a.dtype)

677
    @register_fake("_C::gptq_marlin_gemm")
678
679
680
681
682
683
684
685
    def _gptq_marlin_gemm_fake(a: torch.Tensor,
                               b_q_weight: torch.Tensor,
                               b_scales: torch.Tensor,
                               b_zeros: torch.Tensor,
                               g_idx: torch.Tensor,
                               perm: torch.Tensor,
                               workspace: torch.Tensor,
                               b_q_type: ScalarType,
686
687
688
                               size_m: torch.SymInt,
                               size_n: torch.SymInt,
                               size_k: torch.SymInt,
689
690
                               is_k_full: bool,
                               has_zp: bool = False,
691
692
                               use_fp32_reduce: bool = False,
                               is_zp_float: bool = False) -> torch.Tensor:
693
694
        return torch.empty((size_m, size_n), device=a.device, dtype=a.dtype)

695
    @register_fake("_C::marlin_qqq_gemm")
696
697
698
    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,
699
700
                              size_m: torch.SymInt, size_n: torch.SymInt,
                              size_k: torch.SymInt) -> torch.Tensor:
701
702
703
704
        return torch.empty((size_m, size_n),
                           dtype=torch.float16,
                           device=a.device)

705
    @register_fake("_C::marlin_gemm")
706
707
    def _marlin_gemm_fake(a: torch.Tensor, b_q_weight: torch.Tensor,
                          b_scales: torch.Tensor, workspace: torch.Tensor,
708
709
                          size_m: torch.SymInt, size_n: torch.SymInt,
                          size_k: torch.SymInt) -> torch.Tensor:
710
711
712
713
        return torch.empty((size_m, size_n),
                           dtype=torch.float16,
                           device=a.device)

714
    @register_fake("_C::awq_dequantize")
715
    def _awq_dequantize_fake(qweight: torch.Tensor, scales: torch.Tensor,
716
717
                             zeros: torch.Tensor, split_k_iters: torch.SymInt,
                             thx: int, thy: int) -> torch.Tensor:
718
719
720
721
722
723
724
        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)

725
    @register_fake("_C::awq_gemm")
726
727
    def _awq_gemm_fake(input: torch.Tensor, qweight: torch.Tensor,
                       qzeros: torch.Tensor, scales: torch.Tensor,
728
                       split_k_iters: torch.SymInt) -> torch.Tensor:
729
730
731
732
733
        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)

734
    @register_fake("_C::aqlm_gemm")
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
    def _aqlm_gemm_fake(input: torch.Tensor, codes: torch.Tensor,
                        codebooks: torch.Tensor, scales: torch.Tensor,
                        codebook_partition_sizes: List[int],
                        bias: Optional[torch.Tensor]) -> torch.Tensor:
        out_features = codes.size(0) * codebooks.size(2)
        flat_input = input.reshape((-1, input.size(-1)))
        flat_output = torch.empty((flat_input.size(0), out_features),
                                  dtype=input.dtype,
                                  device=input.device)

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

750
    @register_fake("_C::aqlm_dequant")
751
752
753
754
755
756
757
758
759
    def _aqlm_dequant_fake(
            codes: torch.Tensor, codebooks: torch.Tensor,
            codebook_partition_sizes: List[int]) -> torch.Tensor:
        in_features = codes.size(1) * 8
        out_features = codes.size(0)
        return torch.empty((out_features, in_features),
                           dtype=codebooks.dtype,
                           device=codebooks.device)

760
    @register_fake("_C::fp8_marlin_gemm")
761
762
    def _fp8_marlin_gemm_fake(a: torch.Tensor, b_q_weight: torch.Tensor,
                              b_scales: torch.Tensor, workspace: torch.Tensor,
763
764
765
                              num_bits: int, size_m: torch.SymInt,
                              size_n: torch.SymInt,
                              size_k: torch.SymInt) -> torch.Tensor:
766
767
        return torch.empty((size_m, size_n), dtype=a.dtype, device=a.device)

768
769
    @register_fake("_C::machete_mm")
    def machete_mm_fake(
770
        a: torch.Tensor,
771
        # b_q Should be the tensor returned by machete_prepack_B
772
        b_q: torch.Tensor,
773
        b_type: ScalarType,
774
775
776
        out_type: Optional[torch.dtype] = None,
        b_group_scales: Optional[torch.Tensor] = None,
        b_group_zeros: Optional[torch.Tensor] = None,
777
        b_group_size: Optional[int] = None,
778
779
        b_channel_scales: Optional[torch.Tensor] = None,
        a_token_scales: Optional[torch.Tensor] = None,
780
781
782
783
784
785
        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)

786
    @register_fake("_C::machete_prepack_B")
787
788
789
    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:
790
791
        return torch.empty_like(b_q_weight,
                                memory_format=torch.contiguous_format)
792
793


zhuwenwen's avatar
zhuwenwen committed
794
# if hasattr(torch.ops._C, "ggml_dequantize"):
795

zhuwenwen's avatar
zhuwenwen committed
796
797
798
799
800
#     @register_fake("_C::ggml_dequantize")
#     def _ggml_dequantize_fake(W: torch.Tensor, quant_type: int,
#                               m: torch.SymInt,
#                               n: torch.SymInt) -> torch.Tensor:
#         return torch.empty((m, n), dtype=torch.float16, device=W.device)
801

zhuwenwen's avatar
zhuwenwen committed
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
#     @register_fake("_C::ggml_mul_mat_vec_a8")
#     def _ggml_mul_mat_vec_a8_fake(
#         W: torch.Tensor,
#         X: torch.Tensor,
#         quant_type: int,
#         row: torch.SymInt,
#     ) -> torch.Tensor:
#         return torch.empty((1, row), dtype=torch.float16, device=W.device)

#     @register_fake("_C::ggml_mul_mat_a8")
#     def _ggml_mul_mat_a8_fake(
#         W: torch.Tensor,
#         X: torch.Tensor,
#         quant_type: int,
#         row: torch.SymInt,
#     ) -> torch.Tensor:
#         batch = X.size(0)
#         return torch.empty((batch, row), dtype=torch.float16, device=W.device)
820
821


822
# cutlass
823
824
825
826
def cutlass_scaled_mm_supports_fp8(cuda_device_capability: int) -> bool:
    return torch.ops._C.cutlass_scaled_mm_supports_fp8(cuda_device_capability)


827
828
829
830
831
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)


832
833
834
def cutlass_scaled_mm(a: torch.Tensor,
                      b: torch.Tensor,
                      scale_a: torch.Tensor,
835
                      scale_b: torch.Tensor,
836
                      out_dtype: torch.dtype,
837
                      bias: Optional[torch.Tensor] = None) -> torch.Tensor:
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
    """
    `cutlass_scaled_mm` implements a fused version of 
        `output = torch.mm((scale_a * a), (scale_b * b)).to(out_dtype)`
    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 
        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
    """
860
861
    assert (b.shape[0] % 16 == 0 and b.shape[1] % 16 == 0)
    assert (out_dtype is torch.bfloat16 or out_dtype is torch.float16)
862
863
    assert bias is None or bias.shape[0] == b.shape[
        1] and bias.dtype == out_dtype
864

zhuwenwen's avatar
zhuwenwen committed
865
866
    # m = a.shape[0]
    # n = b.shape[1]
867

zhuwenwen's avatar
zhuwenwen committed
868
869
870
871
872
873
    # if current_platform.is_rocm():
    #     triton_scaled_mm_module = importlib.import_module(
    #         "vllm.model_executor.layers.quantization.compressed_tensors."
    #         "triton_scaled_mm")
    #     triton_scaled_mm = triton_scaled_mm_module.triton_scaled_mm
    #     return triton_scaled_mm(a, b, scale_a, scale_b, out_dtype, bias)
874

zhuwenwen's avatar
zhuwenwen committed
875
    # out = torch.empty((m, n), dtype=out_dtype, device=a.device)
876

zhuwenwen's avatar
zhuwenwen committed
877
    # torch.ops._C.cutlass_scaled_mm(out, a, b, scale_a, scale_b, bias)
878

zhuwenwen's avatar
zhuwenwen committed
879
    # return out
gaoqiong's avatar
gaoqiong committed
880
881
    #return quant_ops.cutlass_scaled_mm(a, b, scale_a, scale_b, out_dtype, bias)
    return quant_ops.rocblas_scaled_mm_nn(a, b, scale_a, scale_b, out_dtype, bias)
882

883
884
885
886
887
888
889
890
891
892
893
894
895
896
def rocblas_scaled_mm(a: torch.Tensor,
                      b: torch.Tensor,
                      scale_a: torch.Tensor,
                      scale_b: torch.Tensor,
                      out_dtype: torch.dtype,
                      bias: Optional[torch.Tensor] = None) -> torch.Tensor:

    return quant_ops.rocblas_scaled_mm_nn(a, b, scale_a, scale_b, out_dtype, bias)

def triton_scaled_mm(a: torch.Tensor,
                      b: torch.Tensor,
                      scale_a: torch.Tensor,
                      scale_b: torch.Tensor,
                      out_dtype: torch.dtype,
gaoqiong's avatar
gaoqiong committed
897
898
                      bias: Optional[torch.Tensor] = None,
                      best_config:Optional[list] = None) -> torch.Tensor:
899

gaoqiong's avatar
gaoqiong committed
900
    return quant_ops.triton_scaled_mm(a, b,scale_a,scale_b,out_dtype,bias,best_config)
901

gaoqiong's avatar
gaoqiong committed
902
903
904
905
906
907
908
909
910
def triton_int8_gemm_helper(m: int,
                             n: int,
                             k: int,
                             per_token_act_quant: bool,
                             per_out_channel_weight_quant: bool,
                             use_bias: bool,
                             out_dtype: Type[torch.dtype] = torch.float16,
                             device: str = "cuda",
                             best_config:Optional[list] = None):
zhuwenwen's avatar
zhuwenwen committed
911
    return quant_tools.triton_int8_gemm_helper(m,n,k,per_token_act_quant,per_out_channel_weight_quant,use_bias,out_dtype,device,best_config)
gaoqiong's avatar
gaoqiong committed
912
913


914
915
916
917
918
919
920
921
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:
922
923
924
925
926
    """
    :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.
    """
927
928
929
930
    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
931
    assert azp is None or azp.numel() == a.shape[0]
932
933
934
935
936
937
938
939
940
941

    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


942
943
944
945
946
def cutlass_sparse_scaled_mm_supported(cuda_device_capability: int) -> bool:
    return torch.ops._C.cutlass_sparse_scaled_mm_supported(
        cuda_device_capability)


947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
def cutlass_sparse_compress(a: torch.Tensor) \
    -> Tuple[torch.Tensor, torch.Tensor]:
    """
    Compresses a sparse matrix for use with Cutlass sparse operations.

    This function takes a dense tensor and compresses it into two components:
    non-zero elements and metadata. The compressed representation is compatible
    with Cutlass sparse kernels.

    Args:
        a (torch.Tensor): 
            The input tensor to be compressed. Must have one of the following data types:
            - `torch.int8`
            - `torch.float8_e4m3fn`
            - `torch.bfloat16`
            - `torch.float16`

    Returns:
        Tuple[torch.Tensor, torch.Tensor]: 
            A tuple containing:
            - `a_nzs` (torch.Tensor): A tensor containing non-zero elements of `a`.
            - `a_meta` (torch.Tensor): A tensor containing metadata for the sparse representation.

    Raises:
        ValueError: If the compression operation fails.

    Notes:
        - The `a_meta` tensor has a data type of `torch.uint8`.
        - Each metadata element encodes the sparsity of 4 non-zero elements (i.e., `elemsPerMetaElem = 4`).
        - The shape of `a_nzs` is `(m, k // 2)`, where `m` and `k` are the dimensions of the input tensor.
        - The shape of `a_meta` is `(m, k // 2 // elemsPerMetaElem)`.
    """
    assert (a.dtype in [
        torch.int8, torch.float8_e4m3fn, torch.bfloat16, torch.float16
    ])
    assert (a.is_contiguous())

    # a_meta.dtype: torch.uint8 so elemsPerMetaElem = 8b / 2b_per_nz = 4
    elemsPerMetaElem = 4
986
    assert (a.shape[1] % (2 * elemsPerMetaElem) == 0)
987

988
    return torch.ops._C.cutlass_sparse_compress(a)
989
990
991
992
993
994
995
996
997
998
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


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


1037
1038
1039
# aqlm
def aqlm_gemm(input: torch.Tensor, codes: torch.Tensor,
              codebooks: torch.Tensor, scales: torch.Tensor,
1040
              codebook_partition_sizes: List[int],
1041
              bias: Optional[torch.Tensor]) -> torch.Tensor:
1042
1043
    return torch.ops._C.aqlm_gemm(input, codes, codebooks, scales,
                                  codebook_partition_sizes, bias)
1044
1045
1046


def aqlm_dequant(codes: torch.Tensor, codebooks: torch.Tensor,
1047
                 codebook_partition_sizes: List[int]) -> torch.Tensor:
1048
1049
    return torch.ops._C.aqlm_dequant(codes, codebooks,
                                     codebook_partition_sizes)
1050
1051


1052
1053
# gptq_marlin
def gptq_marlin_repack(b_q_weight: torch.Tensor, perm: torch.Tensor,
1054
1055
                       size_k: int, size_n: int,
                       num_bits: int) -> torch.Tensor:
1056
1057
    return torch.ops._C.gptq_marlin_repack(b_q_weight, perm, size_k, size_n,
                                           num_bits)
1058
1059


1060
1061
1062
1063
1064
1065
# 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)


1066
1067
1068
1069
1070
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
1071
    output = torch.empty((num_experts, size_k // 16, size_n * (num_bits // 2)),
1072
1073
1074
1075
1076
1077
1078
1079
                         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


1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
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


1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
def gptq_marlin_gemm(a: torch.Tensor,
                     b_q_weight: torch.Tensor,
                     b_scales: torch.Tensor,
                     b_zeros: torch.Tensor,
                     g_idx: torch.Tensor,
                     perm: torch.Tensor,
                     workspace: torch.Tensor,
                     b_q_type: ScalarType,
                     size_m: int,
                     size_n: int,
                     size_k: int,
                     is_k_full: bool,
                     has_zp: bool = False,
1107
1108
                     use_fp32_reduce: bool = False,
                     is_zp_float: bool = False) -> torch.Tensor:
1109
    return torch.ops._C.gptq_marlin_gemm(a, b_q_weight, b_scales, b_zeros,
1110
                                         g_idx, perm, workspace, b_q_type.id,
1111
                                         size_m, size_n, size_k, is_k_full,
1112
                                         has_zp, use_fp32_reduce, is_zp_float)
1113
1114


1115
1116
1117
1118
1119
1120
1121
1122
1123
# fp8 marlin
def fp8_marlin_gemm(a: torch.Tensor, b_q_weight: torch.Tensor,
                    b_scales: torch.Tensor, workspace: torch.Tensor,
                    num_bits: int, size_m: int, size_n: int,
                    size_k: int) -> torch.Tensor:
    return torch.ops._C.fp8_marlin_gemm(a, b_q_weight, b_scales, workspace,
                                        num_bits, size_m, size_n, size_k)


1124
# machete
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
def machete_supported_schedules(
        a_type: torch.dtype,
        b_type: ScalarType,
        group_scales_type: Optional[torch.dtype],
        group_zeros_type: Optional[torch.dtype] = None,
        channel_scales_type: Optional[torch.dtype] = None,
        token_scales_type: Optional[torch.dtype] = None,
        out_type: Optional[torch.dtype] = None) -> List[str]:
    return torch.ops._C.machete_supported_schedules(
        a_type, b_type.id, group_scales_type, group_zeros_type,
        channel_scales_type, token_scales_type, out_type)
1136
1137


1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
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)
1160
1161


1162
if hasattr(torch.ops._C, "permute_cols"):
1163

1164
    @register_fake("_C::permute_cols")
1165
1166
1167
1168
1169
1170
1171
1172
1173
    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)


1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
# fp4
def scaled_fp4_quant(
        input: torch.Tensor,
        input_global_scale: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
    """
    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:
        Tuple[torch.Tensor, torch.Tensor]: The output tensor in FP4 but every
            two values are packed into a uint8 and float8_e4m3 scaling factors
            in the sizzled layout.
    """
1196
    assert not current_platform.is_rocm()
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
    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


1232
# fp8
zhuwenwen's avatar
zhuwenwen committed
1233
1234
1235
# def scaled_fp8_quant(
#     input: torch.Tensor,
#     scale: Optional[torch.Tensor] = None,
1236
#     num_token_padding: Optional[int] = None,
1237
1238
#     scale_ub: Optional[torch.Tensor] = None,
#     use_per_token_if_dynamic: bool = False,
zhuwenwen's avatar
zhuwenwen committed
1239
# ) -> Tuple[torch.Tensor, torch.Tensor]:
zhuwenwen's avatar
zhuwenwen committed
1240
1241
1242
1243
1244
1245
#     """
#     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
1246
#     optional padding of the output tensors for downstream kernels that
zhuwenwen's avatar
zhuwenwen committed
1247
1248
1249
1250
1251
#     will benefit from padding.

#     Args:
#         input: The input tensor to be quantized to FP8
#         scale: Optional scaling factor for the FP8 quantization
zhuwenwen's avatar
zhuwenwen committed
1252
#         scale_ub: Optional upper bound for scaling factor in dynamic
1253
#             per token case
1254
#         num_token_padding: If specified, pad the first dimension
zhuwenwen's avatar
zhuwenwen committed
1255
#             of the output to at least this value.
zhuwenwen's avatar
zhuwenwen committed
1256
#         use_per_token_if_dynamic: Whether to do per_tensor or per_token
1257
#             in the dynamic quantization case.
zhuwenwen's avatar
zhuwenwen committed
1258
1259
1260
1261
1262

#     Returns:
#         Tuple[torch.Tensor, torch.Tensor]: The output tensor in FP8 and
#             scaling factor.
#     """
1263
1264
1265
#     # This code assumes batch_dim and num_tokens are flattened
#     assert (input.ndim == 2)
#     shape: Union[Tuple[int, int], torch.Size] = input.shape
1266
#     # For rocm, the output fp8 dtype is torch.float_e3m3fnuz
zhuwenwen's avatar
zhuwenwen committed
1267
1268
#     out_dtype: torch.dtype = torch.float8_e4m3fnuz \
#             if current_platform.is_rocm() else torch.float8_e4m3fn
1269
1270
#     if num_token_padding:
#         shape = (max(num_token_padding, input.shape[0]), shape[1])
1271
#     output = torch.empty(shape, device=input.device, dtype=out_dtype)
1272

zhuwenwen's avatar
zhuwenwen committed
1273
#     if scale is None:
1274
#         if use_per_token_if_dynamic:
1275
#             scale = torch.empty((shape[0], 1),
1276
1277
1278
1279
1280
1281
1282
#                                 device=input.device,
#                                 dtype=torch.float32)
#             torch.ops._C.dynamic_per_token_scaled_fp8_quant(
#                 output, input, scale, scale_ub)
#         else:
#             scale = torch.zeros(1, device=input.device, dtype=torch.float32)
#             torch.ops._C.dynamic_scaled_fp8_quant(output, input, scale)
zhuwenwen's avatar
zhuwenwen committed
1283
#     else:
1284
1285
#         # num_token_padding not implemented for this case
#         assert (scale.numel() == 1 or num_token_padding is None)
zhuwenwen's avatar
zhuwenwen committed
1286
#         torch.ops._C.static_scaled_fp8_quant(output, input, scale)
1287

zhuwenwen's avatar
zhuwenwen committed
1288
#     return output, scale
1289
1290


1291
# int8
1292
def scaled_int8_quant(
1293
1294
1295
1296
1297
    input: torch.Tensor,
    scale: Optional[torch.Tensor] = None,
    azp: Optional[torch.Tensor] = None,
    symmetric: bool = True
) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
zhuwenwen's avatar
zhuwenwen committed
1298
    """
1299
    Quantize the input tensor to int8 and return the quantized tensor and scale, and maybe azp.
zhuwenwen's avatar
zhuwenwen committed
1300
1301
1302

    Args:
        input: The input tensor to be quantized to int8.
1303
1304
        scale: Optional scaling factor for the int8 quantization.
            When not provided, we invoke dynamic-per-token quantization.
1305
1306
1307
        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).
zhuwenwen's avatar
zhuwenwen committed
1308
1309

    Returns:
1310
      Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]] : Output int8 tensor, scales, and optionally azp.
zhuwenwen's avatar
zhuwenwen committed
1311
    """
1312
1313
1314
    output = torch.empty_like(input, dtype=torch.int8)
    if scale is not None:
        # static-per-tensor quantization.
1315
        assert symmetric == (
1316
1317
            azp
            is None), "azp must only be provided for asymmetric quantization."
1318
        torch.ops._C.static_scaled_int8_quant(output, input, scale, azp)
1319
        return output, scale, azp
1320
1321
1322
1323
1324

    # dynamic-per-token quantization.
    input_scales = torch.empty((input.numel() // input.shape[-1], 1),
                               device=input.device,
                               dtype=torch.float32)
1325
1326
1327
1328
1329
    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
1330
1331


1332
1333
1334
1335
1336
1337
1338
1339
1340
# 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)


1341
# gguf
zhuwenwen's avatar
zhuwenwen committed
1342
1343
1344
# def ggml_dequantize(W: torch.Tensor, quant_type: int, m: int,
#                     n: int) -> torch.Tensor:
#     return torch.ops._C.ggml_dequantize(W, quant_type, m, n)
1345
1346
1347
1348
1349
1350
1351


def ggml_mul_mat_vec_a8(
    W: torch.Tensor,
    X: torch.Tensor,
    quant_type: int,
    row: int,
1352
) -> torch.Tensor:
1353
1354
1355
1356
1357
1358
1359
1360
    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,
1361
) -> torch.Tensor:
1362
1363
1364
    return torch.ops._C.ggml_mul_mat_a8(W, X, quant_type, row)


1365
1366
1367
# mamba
def causal_conv1d_fwd(x: torch.Tensor, weight: torch.Tensor,
                      bias_: Optional[torch.Tensor],
1368
1369
1370
1371
                      conv_states: Optional[torch.Tensor],
                      query_start_loc: Optional[torch.Tensor],
                      cache_indices: Optional[torch.Tensor],
                      has_initial_state: Optional[torch.Tensor],
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
                      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)
1388
1389
1390
1391
1392
1393


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],
1394
1395
1396
1397
1398
                       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):
1399
1400
1401
    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,
1402
                                    ssm_states, pad_slot_id)
1403
1404


1405
# moe
1406
1407
1408
1409
def moe_sum(input: torch.Tensor, output: torch.Tensor):
    torch.ops._moe_C.moe_sum(input, output)


1410
1411
1412
1413
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:
1414
1415
1416
    torch.ops._moe_C.moe_align_block_size(topk_ids, num_experts, block_size,
                                          sorted_token_ids, experts_ids,
                                          num_tokens_post_pad)
1417
1418


1419
1420
1421
1422
1423
1424
1425
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)
王敏's avatar
王敏 committed
1426
1427
1428
1429
1430
1431
    
def ep_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,
                         start_expert, end_expert) -> None:
王敏's avatar
王敏 committed
1432
    torch.ops._moe_C.ep_moe_align_block_size(topk_ids, num_experts, block_size,
王敏's avatar
王敏 committed
1433
1434
1435
                                      sorted_token_ids, experts_ids,
                                      num_tokens_post_pad, start_expert, 
                                      end_expert)
1436
1437


1438
1439
1440
1441
1442
def topk_softmax(topk_weights: torch.Tensor, topk_ids: torch.Tensor,
                 token_expert_indicies: torch.Tensor,
                 gating_output: float) -> None:
    torch.ops._moe_C.topk_softmax(topk_weights, topk_ids,
                                  token_expert_indicies, gating_output)
1443
1444


1445
1446
if supports_moe_ops and hasattr(torch.ops._moe_C, "marlin_gemm_moe"):

1447
    @register_fake("_moe_C::marlin_gemm_moe")
1448
1449
1450
1451
    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,
1452
1453
                             b_zero_points: torch.Tensor, g_idx: torch.Tensor,
                             perm: torch.Tensor, workspace: torch.Tensor,
1454
1455
                             b_q_type: ScalarType, size_m: torch.SymInt,
                             size_n: torch.SymInt, size_k: torch.SymInt,
1456
1457
1458
1459
1460
1461
1462
1463
                             is_k_full: bool, num_experts: int, topk: int,
                             moe_block_size: int, replicate_input: bool,
                             apply_weights: bool) -> torch.Tensor:
        return torch.empty((size_m, topk, size_n),
                           dtype=a.dtype,
                           device=a.device)


1464
1465
1466
1467
1468
1469
1470
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,
1471
1472
    k_scale: torch.Tensor,
    v_scale: torch.Tensor,
1473
) -> None:
1474
1475
    torch.ops._C_cache_ops.reshape_and_cache(key, value, key_cache,
                                             value_cache, slot_mapping,
1476
                                             kv_cache_dtype, k_scale, v_scale)
1477
1478


1479
1480
1481
1482
1483
1484
1485
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,
1486
1487
    k_scale: torch.Tensor,
    v_scale: torch.Tensor,
1488
) -> None:
1489
1490
    torch.ops._C_cache_ops.reshape_and_cache_flash(key, value, key_cache,
                                                   value_cache, slot_mapping,
1491
1492
                                                   kv_cache_dtype, k_scale,
                                                   v_scale)
1493
1494


1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
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)


1508
1509
def copy_blocks(key_caches: List[torch.Tensor],
                value_caches: List[torch.Tensor],
1510
                block_mapping: torch.Tensor) -> None:
1511
    torch.ops._C_cache_ops.copy_blocks(key_caches, value_caches, block_mapping)
1512
1513


1514
1515
1516
1517
1518
def copy_blocks_mla(kv_caches: List[torch.Tensor],
                    block_mapping: torch.Tensor) -> None:
    torch.ops._C_cache_ops.copy_blocks_mla(kv_caches, block_mapping)


1519
def swap_blocks(src: torch.Tensor, dst: torch.Tensor,
1520
                block_mapping: torch.Tensor) -> None:
1521
    torch.ops._C_cache_ops.swap_blocks(src, dst, block_mapping)
1522
1523


1524
1525
1526
1527
def convert_fp8(output: torch.Tensor,
                input: torch.Tensor,
                scale: float = 1.0,
                kv_dtype: str = "fp8") -> None:
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
    torch.ops._C_cache_ops.convert_fp8(output, input, scale, kv_dtype)


def get_device_attribute(attribute: int, device: int) -> int:
    return torch.ops._C_cuda_utils.get_device_attribute(attribute, device)


def get_max_shared_memory_per_block_device_attribute(device: int) -> int:
    # ruff: noqa: E501
    return torch.ops._C_cuda_utils.get_max_shared_memory_per_block_device_attribute(
        device)


# custom ar
1542
1543
1544
1545
def init_custom_ar(ipc_tensors: List[torch.Tensor], rank_data: torch.Tensor,
                   rank: int, full_nvlink: bool) -> int:
    return torch.ops._C_custom_ar.init_custom_ar(ipc_tensors, rank_data, rank,
                                                 full_nvlink)
1546

1547

1548
1549
1550
1551
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)
1552

1553
1554
1555
1556
1557
1558
1559
1560
1561

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


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


1562
1563
def register_buffer(fa: int, ipc_tensors: List[int]) -> None:
    return torch.ops._C_custom_ar.register_buffer(fa, ipc_tensors)
1564
1565


1566
def get_graph_buffer_ipc_meta(fa: int) -> Tuple[List[int], List[int]]:
1567
1568
1569
    return torch.ops._C_custom_ar.get_graph_buffer_ipc_meta(fa)


1570
def register_graph_buffers(fa: int, handles: List[List[int]],
1571
1572
1573
                           offsets: List[List[int]]) -> None:
    torch.ops._C_custom_ar.register_graph_buffers(fa, handles, offsets)

1574

1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
def read_cache(
        keys: torch.Tensor,
        values: torch.Tensor,
        key_caches: List[torch.Tensor],
        value_caches: List[torch.Tensor],
        slot_mapping: torch.Tensor,
        kv_cache_dtype: str
) -> None:
    torch.ops._C_cache_ops.read_cache(keys, values, key_caches,
                                      value_caches, slot_mapping,
                                      kv_cache_dtype)

def write_cache_multi_layers(
        keys: torch.Tensor,
        values: torch.Tensor,
        key_caches: List[torch.Tensor],
        value_caches: List[torch.Tensor],
        slot_mapping: torch.Tensor,
        kv_cache_dtype: str
) -> None:
    torch.ops._C_cache_ops.write_cache_multi_layers(keys, values, key_caches,
                                                    value_caches, slot_mapping,
                                                    kv_cache_dtype)