_custom_ops.py 51.9 KB
Newer Older
1
import contextlib
2
import functools
3
from typing import List, Optional, Tuple, Union
4
5

import torch
gaoqiong's avatar
gaoqiong committed
6

7
import vllm.envs as envs
8
from vllm._core_ext import ScalarType
9
from vllm.logger import init_logger
10
from vllm.platforms import current_platform
11

12
try:
gaoqiong's avatar
gaoqiong committed
13
    from lmslim import quant_ops 
14
except Exception:
gaoqiong's avatar
gaoqiong committed
15
    print("INFO: Please install lmslim if you want to infer gptq or awq  or w8a8 model.\n") 
16

17
18
logger = init_logger(__name__)

19
20
21
22
23
if not current_platform.is_tpu():
    try:
        import vllm._C
    except ImportError as e:
        logger.warning("Failed to import from vllm._C with %r", e)
24

25
26
# if current_platform.is_rocm():
#     import vllm._rocm_C  # noqa: F401
27

28
supports_moe_ops = False
29
with contextlib.suppress(ImportError):
30
    import vllm._moe_C  # noqa: F401
31
    supports_moe_ops = True
32
33


34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
def hint_on_error(fn):

    @functools.wraps(fn)
    def wrapper(*args, **kwargs):
        try:
            return fn(*args, **kwargs)
        except AttributeError as e:
            msg = (
                "Error in calling custom op %s: %s\n"
                "Possibly you have built or installed an obsolete version of vllm.\n"
                "Please try a clean build and install of vllm,"
                "or remove old built files such as vllm/*cpython*.so and build/ ."
            )
            logger.error(msg, fn.__name__, e)
            raise e

    return wrapper


53
54
# activation ops
def silu_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
55
    torch.ops._C.silu_and_mul(out, x)
56
57
58


def gelu_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
59
    torch.ops._C.gelu_and_mul(out, x)
60
61
62


def gelu_tanh_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
63
    torch.ops._C.gelu_tanh_and_mul(out, x)
zhuwenwen's avatar
zhuwenwen committed
64
65
66
67
68
69
70
71
72
73
74
75
    
    
def silu_and_mul_opt(out: torch.Tensor, x: torch.Tensor) -> None:
    torch.ops._C.silu_and_mul_opt(out, x)


def gelu_and_mul_opt(out: torch.Tensor, x: torch.Tensor) -> None:
    torch.ops._C.gelu_and_mul_opt(out, x)


def gelu_tanh_and_mul_opt(out: torch.Tensor, x: torch.Tensor) -> None:
    torch.ops._C.gelu_tanh_and_mul_opt(out, x)
76
77
78


def gelu_fast(out: torch.Tensor, x: torch.Tensor) -> None:
79
    torch.ops._C.gelu_fast(out, x)
80
81
82


def gelu_new(out: torch.Tensor, x: torch.Tensor) -> None:
83
    torch.ops._C.gelu_new(out, x)
84
85


86
87
88
89
def gelu_quick(out: torch.Tensor, x: torch.Tensor) -> None:
    torch.ops._C.gelu_quick(out, x)


90
91
# page attention ops
def paged_attention_v1(
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
    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,
    k_scale: float,
    v_scale: float,
    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,
    k_scale: float,
    v_scale: float,
    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(
152
153
154
155
156
157
158
    out: torch.Tensor,
    query: torch.Tensor,
    key_cache: torch.Tensor,
    value_cache: torch.Tensor,
    num_kv_heads: int,
    scale: float,
    block_tables: torch.Tensor,
159
    seq_lens: torch.Tensor,
160
    block_size: int,
161
    max_seq_len: int,
162
163
    alibi_slopes: Optional[torch.Tensor],
    kv_cache_dtype: str,
164
165
    k_scale: float,
    v_scale: float,
166
167
168
169
170
    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,
171
172
    attn_masks: Optional[torch.Tensor] = None,
    attn_masks_stride: int = 0,
173
) -> None:
174
    torch.ops._C.paged_attention_v1_with_mask(
175
176
        out, query, key_cache, value_cache, num_kv_heads, scale, block_tables,
        seq_lens, block_size, max_seq_len, alibi_slopes, kv_cache_dtype,
177
178
        k_scale, v_scale, tp_rank, blocksparse_local_blocks,
        blocksparse_vert_stride, blocksparse_block_size,
179
180
        blocksparse_head_sliding_step,attn_masks, 
        attn_masks_stride)
181
182


183
def paged_attention_v2_with_mask(
184
185
186
187
188
189
190
191
192
193
    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,
194
    seq_lens: torch.Tensor,
195
    block_size: int,
196
    max_seq_len: int,
197
198
    alibi_slopes: Optional[torch.Tensor],
    kv_cache_dtype: str,
199
200
    k_scale: float,
    v_scale: float,
201
202
203
204
205
    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,
206
207
    attn_masks: Optional[torch.Tensor] = None,
    attn_masks_stride: int = 0,
208
) -> None:
209
    torch.ops._C.paged_attention_v2_with_mask(
210
211
        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,
212
        alibi_slopes, kv_cache_dtype, k_scale, v_scale, tp_rank,
213
        blocksparse_local_blocks, blocksparse_vert_stride,
214
215
        blocksparse_block_size, blocksparse_head_sliding_step,
        attn_masks, attn_masks_stride)
216
217


zhuwenwen's avatar
zhuwenwen committed
218
219
# page attention ops (opt)
def paged_attention_v1_opt(
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
    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,
    k_scale: float,
    v_scale: float,
    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,
    k_scale: float,
    v_scale: float,
    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
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
    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,
    k_scale: float,
    v_scale: float,
    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,
300
301
    attn_masks: Optional[torch.Tensor] = None,
    attn_masks_stride: int = 0,
zhuwenwen's avatar
zhuwenwen committed
302
) -> None:
303
    torch.ops._C.paged_attention_v1_opt_with_mask(
zhuwenwen's avatar
zhuwenwen committed
304
305
306
307
        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,
308
309
        blocksparse_head_sliding_step, attn_masks,
        attn_masks_stride)
zhuwenwen's avatar
zhuwenwen committed
310
311


312
def paged_attention_v2_opt_with_mask(
zhuwenwen's avatar
zhuwenwen committed
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
    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,
    k_scale: float,
    v_scale: float,
    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,
335
336
    attn_masks: Optional[torch.Tensor] = None,
    attn_masks_stride: int = 0,
zhuwenwen's avatar
zhuwenwen committed
337
) -> None:
338
    torch.ops._C.paged_attention_v2_opt_with_mask(
zhuwenwen's avatar
zhuwenwen committed
339
340
341
342
        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,
343
344
        blocksparse_block_size, blocksparse_head_sliding_step,
        attn_masks, attn_masks_stride)
zhuwenwen's avatar
zhuwenwen committed
345
346


347
348
# page attention ops (opt)
def paged_attention_v1_opt_tc(
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
    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,
    k_scale: float,
    v_scale: float,
    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,
    k_scale: float,
    v_scale: float,
    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(
410
411
412
413
414
415
416
417
418
419
420
421
    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,
zhuwenwen's avatar
zhuwenwen committed
422
423
    k_scale: float,
    v_scale: float,
424
425
426
427
428
    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,
429
430
    attn_masks: Optional[torch.Tensor] = None,
    attn_masks_stride: int = 0,
431
) -> None:
432
    torch.ops._C.paged_attention_v1_opt_tc_with_mask(
433
434
        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
435
        k_scale, v_scale, tp_rank, blocksparse_local_blocks, blocksparse_vert_stride,
436
437
        blocksparse_block_size, blocksparse_head_sliding_step,
        attn_masks, attn_masks_stride)
438
439


440
def paged_attention_v2_opt_tc_with_mask(
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
    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,
zhuwenwen's avatar
zhuwenwen committed
456
457
    k_scale: float,
    v_scale: float,
458
459
460
461
462
    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,
463
464
    attn_masks: Optional[torch.Tensor] = None,
    attn_masks_stride: int = 0,
465
) -> None:
466
    torch.ops._C.paged_attention_v2_opt_tc_with_mask(
467
468
        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
469
        alibi_slopes, kv_cache_dtype, k_scale, v_scale, tp_rank,
470
        blocksparse_local_blocks, blocksparse_vert_stride,
471
472
        blocksparse_block_size, blocksparse_head_sliding_step,
        attn_masks, attn_masks_stride)
473
474


475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
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,
491
492
    k_scale: float,
    v_scale: float,
493
494
495
496
497
) -> 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,
498
                                      kv_cache_dtype, k_scale, v_scale)
499
500


501
502
503
504
505
506
507
508
509
# 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:
510
511
    torch.ops._C.rotary_embedding(positions, query, key, head_size,
                                  cos_sin_cache, is_neox)
512
513
514
515
516
517
518


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:
519
520
521
    torch.ops._C.batched_rotary_embedding(positions, query, key, head_size,
                                          cos_sin_cache, is_neox, rot_dim,
                                          cos_sin_cache_offsets)
522
523
524
525
526


# layer norm ops
def rms_norm(out: torch.Tensor, input: torch.Tensor, weight: torch.Tensor,
             epsilon: float) -> None:
527
    torch.ops._C.rms_norm(out, input, weight, epsilon)
528
529
530
531


def fused_add_rms_norm(input: torch.Tensor, residual: torch.Tensor,
                       weight: torch.Tensor, epsilon: float) -> None:
532
    torch.ops._C.fused_add_rms_norm(input, residual, weight, epsilon)
zhuwenwen's avatar
zhuwenwen committed
533
534
535
536
537
538
539
540
541
542
543
    

# 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)
544
545


546
547
548
549
550
551
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:
552
    """Advance a step on GPU for existing inputs for a multi-step runner"""
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
    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)
576

zhuwenwen's avatar
zhuwenwen committed
577
578
579
580
581
# 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)
    
582

583
584
# quantization ops
# awq
zhuwenwen's avatar
zhuwenwen committed
585
586
587
588
589
590
def GetAWQShareWorkspaceSize()->int:
    return quant_ops.GetAWQShareWorkspaceSize()

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

591
592
593
def awq_dequantize(qweight: torch.Tensor, scales: torch.Tensor,
                   zeros: torch.Tensor, split_k_iters: int, thx: int,
                   thy: int) -> torch.Tensor:
594
595
596
597
    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)
598
599
    return torch.ops._C.awq_dequantize(qweight, scales, zeros, split_k_iters,
                                       thx, thy)
600
601


gaoqiong's avatar
gaoqiong committed
602
603
# 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
604
605
606
607
608
609
#     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
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640

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)
641
642
643
644
645
646
647


# 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
648
    return quant_ops.gptq_gemm(a, b_q_weight, b_gptq_qzeros, b_gptq_scales,
649
                                  b_g_idx, use_exllama, bit)
650
651
    # return torch.ops._C.gptq_gemm(a, b_q_weight, b_gptq_qzeros, b_gptq_scales,
    #                               b_g_idx, use_exllama, bit)
652
653


654
if hasattr(torch.ops._C, "gptq_gemm"):
655
656
657
658
659
660
661
662
663
664
665

    @torch.library.register_fake("_C::gptq_gemm")
    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)


666
667
def gptq_shuffle(q_weight: torch.Tensor, q_perm: torch.Tensor,
                 bit: int) -> None:
gaoqiong's avatar
gaoqiong committed
668
    quant_ops.gptq_shuffle(q_weight, q_perm, bit)
669
    # torch.ops._C.gptq_shuffle(q_weight, q_perm, bit)
670
671
672
673
674
675


# 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:
676
677
    return torch.ops._C.marlin_gemm(a, b_q_weight, b_scales, workspace, size_m,
                                    size_n, size_k)
678
679


680
681
682
# marlin_24
def gptq_marlin_24_gemm(a: torch.Tensor, b_q_weight: torch.Tensor,
                        b_meta: torch.Tensor, b_scales: torch.Tensor,
683
684
                        workspace: torch.Tensor, b_q_type: ScalarType,
                        size_m: int, size_n: int, size_k: int) -> torch.Tensor:
685
    return torch.ops._C.gptq_marlin_24_gemm(a, b_q_weight, b_meta, b_scales,
686
                                            workspace, b_q_type, size_m,
687
                                            size_n, size_k)
688
689


690
if hasattr(torch.ops._C, "gptq_marlin_24_gemm"):
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815

    @torch.library.register_fake("_C::gptq_marlin_24_gemm")
    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,
                                  b_q_type: ScalarType, size_m: int,
                                  size_n: int, size_k: int) -> torch.Tensor:
        return torch.empty((size_m, size_n), device=a.device, dtype=a.dtype)

    @torch.library.register_fake("_C::gptq_marlin_gemm")
    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,
                               size_m: int,
                               size_n: int,
                               size_k: int,
                               is_k_full: bool,
                               has_zp: bool = False,
                               use_fp32_reduce: bool = False) -> torch.Tensor:
        return torch.empty((size_m, size_n), device=a.device, dtype=a.dtype)

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

    @torch.library.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: int,
    ) -> torch.Tensor:
        return torch.empty((1, row), dtype=torch.float16, device=W.device)

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

    @torch.library.register_fake("_C::marlin_qqq_gemm")
    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,
                              size_m: int, size_n: int,
                              size_k: int) -> torch.Tensor:
        return torch.empty((size_m, size_n),
                           dtype=torch.float16,
                           device=a.device)

    @torch.library.register_fake("_C::marlin_gemm")
    def _marlin_gemm_fake(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:
        return torch.empty((size_m, size_n),
                           dtype=torch.float16,
                           device=a.device)

    @torch.library.register_fake("_C::awq_dequantize")
    def _awq_dequantize_fake(qweight: torch.Tensor, scales: torch.Tensor,
                             zeros: torch.Tensor, split_k_iters: int, thx: int,
                             thy: int) -> torch.Tensor:
        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)

    @torch.library.register_fake("_C::awq_gemm")
    def _awq_gemm_fake(input: torch.Tensor, qweight: torch.Tensor,
                       qzeros: torch.Tensor, scales: torch.Tensor,
                       split_k_iters: int) -> torch.Tensor:
        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)

    @torch.library.register_fake("_C::aqlm_gemm")
    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))

    @torch.library.register_fake("_C::aqlm_dequant")
    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)

    @torch.library.register_fake("_C::fp8_marlin_gemm")
    def _fp8_marlin_gemm_fake(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.empty((size_m, size_n), dtype=a.dtype, device=a.device)

    @torch.library.register_fake("_C::machete_gemm")
    def machete_gemm_fake(
        a: torch.Tensor,
816
817
        # Should be the tensor returned by machete_prepack_B
        b_q: torch.Tensor,
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
        b_type: ScalarType,
        b_scales: Optional[torch.Tensor] = None,
        b_zeros: Optional[torch.Tensor] = None,
        b_group_size: Optional[int] = None,
        c: Optional[torch.Tensor] = None,
        alpha: Optional[float] = None,
        beta: Optional[float] = None,
        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)

    @torch.library.register_fake("_C::machete_prepack_B")
    def machete_prepack_B_fake(b_q_weight: torch.Tensor,
                               b_type: ScalarType) -> torch.Tensor:
834
835
        return torch.empty_like(b_q_weight,
                                memory_format=torch.contiguous_format)
836
837
838
839
840
841
842
843
844
845
846

    @torch.library.register_fake("_C::causal_conv1d_fwd")
    def causal_conv1d_fwd_fake(x: torch.Tensor, weight: torch.Tensor,
                               bias_: Optional[torch.Tensor],
                               seq_idx_: Optional[torch.Tensor],
                               initial_states_: Optional[torch.Tensor],
                               final_states_out_: Optional[torch.Tensor],
                               silu_activation: bool) -> torch.Tensor:
        return torch.empty_like(x)

    @torch.library.register_fake("_C::causal_conv1d_update")
847
848
849
850
    def causal_conv1d_update_fake(
            x: torch.Tensor, conv_state: torch.Tensor, weight: torch.Tensor,
            bias_: Optional[torch.Tensor], silu_activation: bool,
            conv_state_indices: Optional[torch.Tensor]) -> torch.Tensor:
851
852
853
854
855
856
857
858
859
860
861
862
        return torch.empty_like(x)

    @torch.library.register_fake("_C::selective_scan_fwd")
    def selective_scan_fwd_fake(
            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, index_: Optional[torch.Tensor],
            x: Optional[torch.Tensor]) -> List[torch.Tensor]:
        a = torch.empty_like(u)
        if z_ is not None:
            c = torch.empty_like(z_)
863
            return [a, c]
864
        else:
865
            return [a]
866
867


868
# cutlass
869
870
871
872
def cutlass_scaled_mm_supports_fp8(cuda_device_capability: int) -> bool:
    return torch.ops._C.cutlass_scaled_mm_supports_fp8(cuda_device_capability)


873
874
875
def cutlass_scaled_mm(a: torch.Tensor,
                      b: torch.Tensor,
                      scale_a: torch.Tensor,
876
                      scale_b: torch.Tensor,
877
                      out_dtype: torch.dtype,
878
                      bias: Optional[torch.Tensor] = None) -> torch.Tensor:
zhuwenwen's avatar
zhuwenwen committed
879
880
881
882
    # 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.shape[0] == b.shape[
    #     1] and bias.dtype == out_dtype
883

zhuwenwen's avatar
zhuwenwen committed
884
885
886
    # m = a.shape[0]
    # n = b.shape[1]
    # out = torch.empty((m, n), dtype=out_dtype, device=a.device)
887

zhuwenwen's avatar
zhuwenwen committed
888
    # torch.ops._C.cutlass_scaled_mm(out, a, b, scale_a, scale_b, bias)
889

zhuwenwen's avatar
zhuwenwen committed
890
    # return out
gaoqiong's avatar
gaoqiong committed
891
892
    #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)
893

894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
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,
                      bias: Optional[torch.Tensor] = None) -> torch.Tensor:

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

912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
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:
    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

    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


934
935
936
# aqlm
def aqlm_gemm(input: torch.Tensor, codes: torch.Tensor,
              codebooks: torch.Tensor, scales: torch.Tensor,
937
              codebook_partition_sizes: List[int],
938
              bias: Optional[torch.Tensor]) -> torch.Tensor:
939
940
    return torch.ops._C.aqlm_gemm(input, codes, codebooks, scales,
                                  codebook_partition_sizes, bias)
941
942
943


def aqlm_dequant(codes: torch.Tensor, codebooks: torch.Tensor,
944
                 codebook_partition_sizes: List[int]) -> torch.Tensor:
945
946
    return torch.ops._C.aqlm_dequant(codes, codebooks,
                                     codebook_partition_sizes)
947
948


949
950
# gptq_marlin
def gptq_marlin_repack(b_q_weight: torch.Tensor, perm: torch.Tensor,
951
952
                       size_k: int, size_n: int,
                       num_bits: int) -> torch.Tensor:
953
954
    return torch.ops._C.gptq_marlin_repack(b_q_weight, perm, size_k, size_n,
                                           num_bits)
955
956


957
958
959
960
961
962
# 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)


963
964
965
966
967
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
968
    output = torch.empty((num_experts, size_k // 16, size_n * (num_bits // 2)),
969
970
971
972
973
974
975
976
                         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


977
978
979
980
981
982
983
984
985
986
987
988
989
990
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,
                     use_fp32_reduce: bool = False) -> torch.Tensor:
991
    return torch.ops._C.gptq_marlin_gemm(a, b_q_weight, b_scales, b_zeros,
992
                                         g_idx, perm, workspace, b_q_type,
993
                                         size_m, size_n, size_k, is_k_full,
994
                                         has_zp, use_fp32_reduce)
995
996


997
998
999
1000
1001
1002
1003
1004
1005
# 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)


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
# machete
def machete_supported_schedules(b_type: ScalarType) -> List[str]:
    return torch.ops._C.machete_supported_schedules(b_type)


def machete_gemm(
    a: torch.Tensor,
    b_q: torch.Tensor,  # Should be the tensor returned by machete_prepack_B
    b_type: ScalarType,
    b_scales: Optional[torch.Tensor] = None,
    b_zeros: Optional[torch.Tensor] = None,
    b_group_size: Optional[int] = None,
    c: Optional[torch.Tensor] = None,
    alpha: Optional[float] = None,
    beta: Optional[float] = None,
    schedule: Optional[str] = None,
) -> torch.Tensor:
    return torch.ops._C.machete_gemm(a, b_q, b_type, b_scales, b_zeros,
                                     b_group_size, c, alpha, beta, schedule)


def machete_prepack_B(b_q_weight: torch.Tensor,
                      b_type: ScalarType) -> torch.Tensor:
    return torch.ops._C.machete_prepack_B(b_q_weight, b_type)


1032
if hasattr(torch.ops._C, "permute_cols"):
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043

    @torch.library.register_fake("_C::permute_cols")
    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)


1044
# fp8
zhuwenwen's avatar
zhuwenwen committed
1045
1046
1047
# def scaled_fp8_quant(
#     input: torch.Tensor,
#     scale: Optional[torch.Tensor] = None,
1048
#     num_token_padding: Optional[int] = None,
1049
1050
#     scale_ub: Optional[torch.Tensor] = None,
#     use_per_token_if_dynamic: bool = False,
zhuwenwen's avatar
zhuwenwen committed
1051
# ) -> Tuple[torch.Tensor, torch.Tensor]:
zhuwenwen's avatar
zhuwenwen committed
1052
1053
1054
1055
1056
1057
#     """
#     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
1058
#     optional padding of the output tensors for downstream kernels that
zhuwenwen's avatar
zhuwenwen committed
1059
1060
1061
1062
1063
#     will benefit from padding.

#     Args:
#         input: The input tensor to be quantized to FP8
#         scale: Optional scaling factor for the FP8 quantization
1064
1065
#         scale_ub: Optional upper bound for scaling factor in dynamic 
#             per token case
1066
#         num_token_padding: If specified, pad the first dimension
zhuwenwen's avatar
zhuwenwen committed
1067
#             of the output to at least this value.
1068
1069
#         use_per_token_if_dynamic: Whether to do per_tensor or per_token 
#             in the dynamic quantization case.
zhuwenwen's avatar
zhuwenwen committed
1070
1071
1072
1073
1074

#     Returns:
#         Tuple[torch.Tensor, torch.Tensor]: The output tensor in FP8 and
#             scaling factor.
#     """
1075
1076
1077
#     # This code assumes batch_dim and num_tokens are flattened
#     assert (input.ndim == 2)
#     shape: Union[Tuple[int, int], torch.Size] = input.shape
1078
1079
1080
#     # For rocm, the output fp8 dtype is torch.float_e3m3fnuz
#     out_dtype: torch.dtype = torch.float8_e4m3fnuz if vllm.utils.is_hip() \
#         else torch.float8_e4m3fn
1081
1082
#     if num_token_padding:
#         shape = (max(num_token_padding, input.shape[0]), shape[1])
1083
#     output = torch.empty(shape, device=input.device, dtype=out_dtype)
1084

zhuwenwen's avatar
zhuwenwen committed
1085
#     if scale is None:
1086
#         if use_per_token_if_dynamic:
1087
#             scale = torch.empty((shape[0], 1),
1088
1089
1090
1091
1092
1093
1094
#                                 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
1095
#     else:
1096
1097
#         # num_token_padding not implemented for this case
#         assert (scale.numel() == 1 or num_token_padding is None)
zhuwenwen's avatar
zhuwenwen committed
1098
#         torch.ops._C.static_scaled_fp8_quant(output, input, scale)
1099

zhuwenwen's avatar
zhuwenwen committed
1100
#     return output, scale
1101
1102


1103
# int8
1104
def scaled_int8_quant(
1105
1106
1107
1108
1109
    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
1110
    """
1111
    Quantize the input tensor to int8 and return the quantized tensor and scale, and maybe azp.
zhuwenwen's avatar
zhuwenwen committed
1112
1113
1114

    Args:
        input: The input tensor to be quantized to int8.
1115
1116
        scale: Optional scaling factor for the int8 quantization.
            When not provided, we invoke dynamic-per-token quantization.
1117
1118
1119
        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
1120
1121

    Returns:
1122
      Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]] : Output int8 tensor, scales, and optionally azp.
zhuwenwen's avatar
zhuwenwen committed
1123
    """
1124
1125
1126
    output = torch.empty_like(input, dtype=torch.int8)
    if scale is not None:
        # static-per-tensor quantization.
1127
1128
1129
1130
1131
        assert symmetric == (
            azp is
            None), "azp must only be provided for asymmetric quantization."
        torch.ops._C.static_scaled_int8_quant(output, input, scale, azp)
        return output, scale, None
1132
1133
1134
1135
1136

    # dynamic-per-token quantization.
    input_scales = torch.empty((input.numel() // input.shape[-1], 1),
                               device=input.device,
                               dtype=torch.float32)
1137
1138
1139
1140
1141
    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
1142
1143


1144
1145
1146
1147
1148
1149
1150
1151
1152
# 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)


1153
# gguf
1154
1155
def ggml_dequantize(W: torch.Tensor, quant_type: int, m: int,
                    n: int) -> torch.Tensor:
1156
1157
1158
1159
1160
1161
1162
1163
    return torch.ops._C.ggml_dequantize(W, quant_type, m, n)


def ggml_mul_mat_vec_a8(
    W: torch.Tensor,
    X: torch.Tensor,
    quant_type: int,
    row: int,
1164
) -> torch.Tensor:
1165
1166
1167
1168
1169
1170
1171
1172
    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,
1173
) -> torch.Tensor:
1174
1175
1176
    return torch.ops._C.ggml_mul_mat_a8(W, X, quant_type, row)


1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
# mamba
def causal_conv1d_fwd(x: torch.Tensor, weight: torch.Tensor,
                      bias_: Optional[torch.Tensor],
                      seq_idx_: Optional[torch.Tensor],
                      initial_states_: Optional[torch.Tensor],
                      final_states_out_: Optional[torch.Tensor],
                      silu_activation: bool) -> torch.Tensor:
    return torch.ops._C.causal_conv1d_fwd(x, weight, bias_, seq_idx_,
                                          initial_states_, final_states_out_,
                                          silu_activation)


1189
1190
1191
1192
1193
1194
1195
1196
def causal_conv1d_update(
    x: torch.Tensor,
    conv_state: torch.Tensor,
    weight: torch.Tensor,
    bias_: Optional[torch.Tensor],
    silu_activation: bool,
    conv_state_indices: Optional[torch.Tensor],
) -> torch.Tensor:
1197
    return torch.ops._C.causal_conv1d_update(x, conv_state, weight, bias_,
1198
1199
                                             silu_activation,
                                             conv_state_indices)
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212


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, index_: Optional[torch.Tensor],
                       x: Optional[torch.Tensor]) -> List[torch.Tensor]:
    return torch.ops._C.selective_scan_fwd(u, delta, A, B, C, D_, z_,
                                           delta_bias_, delta_softplus, index_,
                                           x)


1213
1214
1215
1216
1217
# moe
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:
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
    torch.ops._C.moe_align_block_size(topk_ids, num_experts, block_size,
                                      sorted_token_ids, experts_ids,
                                      num_tokens_post_pad)


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)
1228
1229


1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
if supports_moe_ops and hasattr(torch.ops._moe_C, "marlin_gemm_moe"):

    @torch.library.register_fake("_moe_C::marlin_gemm_moe")
    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,
                             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, 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)


1248
1249
1250
1251
1252
1253
1254
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,
1255
1256
    k_scale: float,
    v_scale: float,
1257
) -> None:
1258
1259
    torch.ops._C_cache_ops.reshape_and_cache(key, value, key_cache,
                                             value_cache, slot_mapping,
1260
                                             kv_cache_dtype, k_scale, v_scale)
1261
1262


1263
1264
1265
1266
1267
1268
1269
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,
1270
1271
    k_scale: float,
    v_scale: float,
1272
) -> None:
1273
1274
    torch.ops._C_cache_ops.reshape_and_cache_flash(key, value, key_cache,
                                                   value_cache, slot_mapping,
1275
1276
                                                   kv_cache_dtype, k_scale,
                                                   v_scale)
1277
1278


1279
1280
def copy_blocks(key_caches: List[torch.Tensor],
                value_caches: List[torch.Tensor],
1281
                block_mapping: torch.Tensor) -> None:
1282
    torch.ops._C_cache_ops.copy_blocks(key_caches, value_caches, block_mapping)
1283
1284
1285


def swap_blocks(src: torch.Tensor, dst: torch.Tensor,
1286
                block_mapping: torch.Tensor) -> None:
1287
    torch.ops._C_cache_ops.swap_blocks(src, dst, block_mapping)
1288
1289


1290
1291
1292
1293
def convert_fp8(output: torch.Tensor,
                input: torch.Tensor,
                scale: float = 1.0,
                kv_dtype: str = "fp8") -> None:
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
    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
def init_custom_ar(meta: torch.Tensor, rank_data: torch.Tensor,
                   handles: List[str], offsets: List[int], rank: int,
                   full_nvlink: bool) -> int:
    return torch.ops._C_custom_ar.init_custom_ar(meta, rank_data, handles,
                                                 offsets, rank, full_nvlink)


def all_reduce_reg(fa: int, inp: torch.Tensor, out: torch.Tensor) -> None:
    torch.ops._C_custom_ar.all_reduce_reg(fa, inp, out)

1318

1319
1320
1321
def all_reduce_unreg(fa: int, inp: torch.Tensor, reg_buffer: torch.Tensor,
                     out: torch.Tensor) -> None:
    torch.ops._C_custom_ar.all_reduce_unreg(fa, inp, reg_buffer, out)
1322

1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344

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


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


def register_buffer(fa: int, t: torch.Tensor, handles: List[str],
                    offsets: List[int]) -> None:
    return torch.ops._C_custom_ar.register_buffer(fa, t, handles, offsets)


def get_graph_buffer_ipc_meta(fa: int) -> Tuple[List[str], List[int]]:
    return torch.ops._C_custom_ar.get_graph_buffer_ipc_meta(fa)


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

1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
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)


1370

1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
# temporary fix for https://github.com/vllm-project/vllm/issues/5456
# TODO: remove this in v0.6.0
names_and_values = globals()
names_and_values_to_update = {}
# prepare variables to avoid dict size change during iteration
k, v, arg = None, None, None
fn_type = type(lambda x: x)
for k, v in names_and_values.items():
    # find functions that are defined in this file and have torch.Tensor
    # in their annotations. `arg == "torch.Tensor"` is used to handle
    # the case when users use `import __annotations__` to turn type
    # hints into strings.
    if isinstance(v, fn_type) \
        and v.__code__.co_filename == __file__ \
        and any(arg is torch.Tensor or arg == "torch.Tensor"
1386
                for arg in v.__annotations__.values()):
1387
1388
1389
1390
        names_and_values_to_update[k] = hint_on_error(v)

names_and_values.update(names_and_values_to_update)
del names_and_values_to_update, names_and_values, v, k, fn_type