_custom_ops.py 55.6 KB
Newer Older
1
import contextlib
2
import functools
zhuwenwen's avatar
zhuwenwen committed
3

4
import importlib
5
from typing import TYPE_CHECKING, List, Optional, Tuple, Union
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
36
# neuron has torch version that doesn't even have impl_abstract
if TYPE_CHECKING or current_platform.is_neuron():
37
38
39
40
41
42
43
44
45

    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

46

47
48
49
50
51
52
def hint_on_error(fn):

    @functools.wraps(fn)
    def wrapper(*args, **kwargs):
        try:
            return fn(*args, **kwargs)
53
54
55
56
57
58
59
60
61

        except NotImplementedError as e:
            msg = (
                "Error in calling custom op %s: %s\n"
                "Not implemented or built, mostly likely because the current current device "
                "does not support this kernel (less likely TORCH_CUDA_ARCH_LIST was set "
                "incorrectly while building)")
            logger.error(msg, fn.__name__, e)
            raise NotImplementedError(msg % (fn.__name__, e)) from e
62
63
64
65
66
67
68
69
70
71
72
73
74
        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


75
76
# activation ops
def silu_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
77
    torch.ops._C.silu_and_mul(out, x)
78
79
80


def gelu_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
81
    torch.ops._C.gelu_and_mul(out, x)
82
83
84


def gelu_tanh_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
85
    torch.ops._C.gelu_tanh_and_mul(out, x)
zhuwenwen's avatar
zhuwenwen committed
86
87
88
89
90
91
92
93
94
95
96
97
    
    
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)
98
99


100
101
102
103
104
105
def fatrelu_and_mul(out: torch.Tensor,
                    x: torch.Tensor,
                    threshold: float = 0.0) -> None:
    torch.ops._C.fatrelu_and_mul(out, x, threshold)


106
def gelu_fast(out: torch.Tensor, x: torch.Tensor) -> None:
107
    torch.ops._C.gelu_fast(out, x)
108
109
110


def gelu_new(out: torch.Tensor, x: torch.Tensor) -> None:
111
    torch.ops._C.gelu_new(out, x)
112
113


114
115
116
117
def gelu_quick(out: torch.Tensor, x: torch.Tensor) -> None:
    torch.ops._C.gelu_quick(out, x)


118
119
# page attention ops
def paged_attention_v1(
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
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
    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(
180
181
182
183
184
185
186
    out: torch.Tensor,
    query: torch.Tensor,
    key_cache: torch.Tensor,
    value_cache: torch.Tensor,
    num_kv_heads: int,
    scale: float,
    block_tables: torch.Tensor,
187
    seq_lens: torch.Tensor,
188
    block_size: int,
189
    max_seq_len: int,
190
191
    alibi_slopes: Optional[torch.Tensor],
    kv_cache_dtype: str,
192
193
    k_scale: float,
    v_scale: float,
194
195
196
197
198
    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,
199
200
    attn_masks: Optional[torch.Tensor] = None,
    attn_masks_stride: int = 0,
201
) -> None:
202
    torch.ops._C.paged_attention_v1_with_mask(
203
204
        out, query, key_cache, value_cache, num_kv_heads, scale, block_tables,
        seq_lens, block_size, max_seq_len, alibi_slopes, kv_cache_dtype,
205
206
        k_scale, v_scale, tp_rank, blocksparse_local_blocks,
        blocksparse_vert_stride, blocksparse_block_size,
207
208
        blocksparse_head_sliding_step,attn_masks, 
        attn_masks_stride)
209
210


211
def paged_attention_v2_with_mask(
212
213
214
215
216
217
218
219
220
221
    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,
222
    seq_lens: torch.Tensor,
223
    block_size: int,
224
    max_seq_len: int,
225
226
    alibi_slopes: Optional[torch.Tensor],
    kv_cache_dtype: str,
227
228
    k_scale: float,
    v_scale: float,
229
230
231
232
233
    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,
234
235
    attn_masks: Optional[torch.Tensor] = None,
    attn_masks_stride: int = 0,
236
) -> None:
237
    torch.ops._C.paged_attention_v2_with_mask(
238
239
        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,
240
        alibi_slopes, kv_cache_dtype, k_scale, v_scale, tp_rank,
241
        blocksparse_local_blocks, blocksparse_vert_stride,
242
243
        blocksparse_block_size, blocksparse_head_sliding_step,
        attn_masks, attn_masks_stride)
244
245


zhuwenwen's avatar
zhuwenwen committed
246
247
# page attention ops (opt)
def paged_attention_v1_opt(
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
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
    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
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
    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,
328
329
    attn_masks: Optional[torch.Tensor] = None,
    attn_masks_stride: int = 0,
zhuwenwen's avatar
zhuwenwen committed
330
) -> None:
331
    torch.ops._C.paged_attention_v1_opt_with_mask(
zhuwenwen's avatar
zhuwenwen committed
332
333
334
335
        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,
336
337
        blocksparse_head_sliding_step, attn_masks,
        attn_masks_stride)
zhuwenwen's avatar
zhuwenwen committed
338
339


340
def paged_attention_v2_opt_with_mask(
zhuwenwen's avatar
zhuwenwen committed
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
    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,
363
364
    attn_masks: Optional[torch.Tensor] = None,
    attn_masks_stride: int = 0,
zhuwenwen's avatar
zhuwenwen committed
365
) -> None:
366
    torch.ops._C.paged_attention_v2_opt_with_mask(
zhuwenwen's avatar
zhuwenwen committed
367
368
369
370
        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,
371
372
        blocksparse_block_size, blocksparse_head_sliding_step,
        attn_masks, attn_masks_stride)
zhuwenwen's avatar
zhuwenwen committed
373
374


375
376
# page attention ops (opt)
def paged_attention_v1_opt_tc(
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
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
    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(
438
439
440
441
442
443
444
445
446
447
448
449
    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
450
451
    k_scale: float,
    v_scale: float,
452
453
454
455
456
    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,
457
458
    attn_masks: Optional[torch.Tensor] = None,
    attn_masks_stride: int = 0,
459
) -> None:
460
    torch.ops._C.paged_attention_v1_opt_tc_with_mask(
461
462
        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
463
        k_scale, v_scale, tp_rank, blocksparse_local_blocks, blocksparse_vert_stride,
464
465
        blocksparse_block_size, blocksparse_head_sliding_step,
        attn_masks, attn_masks_stride)
466
467


468
def paged_attention_v2_opt_tc_with_mask(
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
    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
484
485
    k_scale: float,
    v_scale: float,
486
487
488
489
490
    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,
491
492
    attn_masks: Optional[torch.Tensor] = None,
    attn_masks_stride: int = 0,
493
) -> None:
494
    torch.ops._C.paged_attention_v2_opt_tc_with_mask(
495
496
        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
497
        alibi_slopes, kv_cache_dtype, k_scale, v_scale, tp_rank,
498
        blocksparse_local_blocks, blocksparse_vert_stride,
499
500
        blocksparse_block_size, blocksparse_head_sliding_step,
        attn_masks, attn_masks_stride)
501
502


503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
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,
519
520
    k_scale: float,
    v_scale: float,
521
522
523
524
525
) -> 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,
526
                                      kv_cache_dtype, k_scale, v_scale)
527
528


529
530
531
532
533
534
535
536
537
# 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:
538
539
    torch.ops._C.rotary_embedding(positions, query, key, head_size,
                                  cos_sin_cache, is_neox)
540
541
542
543
544
545
546


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:
547
548
549
    torch.ops._C.batched_rotary_embedding(positions, query, key, head_size,
                                          cos_sin_cache, is_neox, rot_dim,
                                          cos_sin_cache_offsets)
550
551
552
553
554


# layer norm ops
def rms_norm(out: torch.Tensor, input: torch.Tensor, weight: torch.Tensor,
             epsilon: float) -> None:
555
    torch.ops._C.rms_norm(out, input, weight, epsilon)
556
557
558
559


def fused_add_rms_norm(input: torch.Tensor, residual: torch.Tensor,
                       weight: torch.Tensor, epsilon: float) -> None:
560
    torch.ops._C.fused_add_rms_norm(input, residual, weight, epsilon)
zhuwenwen's avatar
zhuwenwen committed
561
562
563
564
565
566
567
568
569
570
571
    

# 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)
572
573


574
575
576
577
578
579
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:
580
    """Advance a step on GPU for existing inputs for a multi-step runner"""
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
    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)
604

zhuwenwen's avatar
zhuwenwen committed
605
606
607
608
609
# 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)
    
610

611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
# 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


631
632
# quantization ops
# awq
zhuwenwen's avatar
zhuwenwen committed
633
634
635
636
637
638
def GetAWQShareWorkspaceSize()->int:
    return quant_ops.GetAWQShareWorkspaceSize()

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

639
640
641
def awq_dequantize(qweight: torch.Tensor, scales: torch.Tensor,
                   zeros: torch.Tensor, split_k_iters: int, thx: int,
                   thy: int) -> torch.Tensor:
642
643
644
645
    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)
646
647
    return torch.ops._C.awq_dequantize(qweight, scales, zeros, split_k_iters,
                                       thx, thy)
648
649


gaoqiong's avatar
gaoqiong committed
650
651
# 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
652
653
654
655
656
657
#     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
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688

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)
689
690
691
692
693
694
695


# 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
696
    return quant_ops.gptq_gemm(a, b_q_weight, b_gptq_qzeros, b_gptq_scales,
697
                                  b_g_idx, use_exllama, bit)
698
699
    # return torch.ops._C.gptq_gemm(a, b_q_weight, b_gptq_qzeros, b_gptq_scales,
    #                               b_g_idx, use_exllama, bit)
700
701


702
if hasattr(torch.ops._C, "gptq_gemm"):
703

704
    @register_fake("_C::gptq_gemm")
705
706
707
708
709
710
711
712
713
    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)


714
715
def gptq_shuffle(q_weight: torch.Tensor, q_perm: torch.Tensor,
                 bit: int) -> None:
gaoqiong's avatar
gaoqiong committed
716
    quant_ops.gptq_shuffle(q_weight, q_perm, bit)
717
    # torch.ops._C.gptq_shuffle(q_weight, q_perm, bit)
718
719
720
721
722
723


# 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:
724
725
    return torch.ops._C.marlin_gemm(a, b_q_weight, b_scales, workspace, size_m,
                                    size_n, size_k)
726
727


728
729
730
# marlin_24
def gptq_marlin_24_gemm(a: torch.Tensor, b_q_weight: torch.Tensor,
                        b_meta: torch.Tensor, b_scales: torch.Tensor,
731
732
                        workspace: torch.Tensor, b_q_type: ScalarType,
                        size_m: int, size_n: int, size_k: int) -> torch.Tensor:
733
    return torch.ops._C.gptq_marlin_24_gemm(a, b_q_weight, b_meta, b_scales,
734
                                            workspace, b_q_type.id, size_m,
735
                                            size_n, size_k)
736
737


738
if hasattr(torch.ops._C, "gptq_marlin_24_gemm"):
739

740
    @register_fake("_C::gptq_marlin_24_gemm")
741
742
743
    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,
744
745
746
                                  b_q_type: ScalarType, size_m: torch.SymInt,
                                  size_n: torch.SymInt,
                                  size_k: torch.SymInt) -> torch.Tensor:
747
748
        return torch.empty((size_m, size_n), device=a.device, dtype=a.dtype)

749
    @register_fake("_C::gptq_marlin_gemm")
750
751
752
753
754
755
756
757
    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,
758
759
760
                               size_m: torch.SymInt,
                               size_n: torch.SymInt,
                               size_k: torch.SymInt,
761
762
                               is_k_full: bool,
                               has_zp: bool = False,
763
764
                               use_fp32_reduce: bool = False,
                               is_zp_float: bool = False) -> torch.Tensor:
765
766
        return torch.empty((size_m, size_n), device=a.device, dtype=a.dtype)

767
    @register_fake("_C::marlin_qqq_gemm")
768
769
770
    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,
771
772
                              size_m: torch.SymInt, size_n: torch.SymInt,
                              size_k: torch.SymInt) -> torch.Tensor:
773
774
775
776
        return torch.empty((size_m, size_n),
                           dtype=torch.float16,
                           device=a.device)

777
    @register_fake("_C::marlin_gemm")
778
779
    def _marlin_gemm_fake(a: torch.Tensor, b_q_weight: torch.Tensor,
                          b_scales: torch.Tensor, workspace: torch.Tensor,
780
781
                          size_m: torch.SymInt, size_n: torch.SymInt,
                          size_k: torch.SymInt) -> torch.Tensor:
782
783
784
785
        return torch.empty((size_m, size_n),
                           dtype=torch.float16,
                           device=a.device)

786
    @register_fake("_C::awq_dequantize")
787
    def _awq_dequantize_fake(qweight: torch.Tensor, scales: torch.Tensor,
788
789
                             zeros: torch.Tensor, split_k_iters: torch.SymInt,
                             thx: int, thy: int) -> torch.Tensor:
790
791
792
793
794
795
796
        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)

797
    @register_fake("_C::awq_gemm")
798
799
    def _awq_gemm_fake(input: torch.Tensor, qweight: torch.Tensor,
                       qzeros: torch.Tensor, scales: torch.Tensor,
800
                       split_k_iters: torch.SymInt) -> torch.Tensor:
801
802
803
804
805
        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)

806
    @register_fake("_C::aqlm_gemm")
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
    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))

822
    @register_fake("_C::aqlm_dequant")
823
824
825
826
827
828
829
830
831
    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)

832
    @register_fake("_C::fp8_marlin_gemm")
833
834
    def _fp8_marlin_gemm_fake(a: torch.Tensor, b_q_weight: torch.Tensor,
                              b_scales: torch.Tensor, workspace: torch.Tensor,
835
836
837
                              num_bits: int, size_m: torch.SymInt,
                              size_n: torch.SymInt,
                              size_k: torch.SymInt) -> torch.Tensor:
838
839
        return torch.empty((size_m, size_n), dtype=a.dtype, device=a.device)

840
841
    @register_fake("_C::machete_mm")
    def machete_mm_fake(
842
        a: torch.Tensor,
843
        # b_q Should be the tensor returned by machete_prepack_B
844
        b_q: torch.Tensor,
845
        b_type: ScalarType,
846
847
848
        out_type: Optional[torch.dtype] = None,
        b_group_scales: Optional[torch.Tensor] = None,
        b_group_zeros: Optional[torch.Tensor] = None,
849
        b_group_size: Optional[int] = None,
850
851
        b_channel_scales: Optional[torch.Tensor] = None,
        a_token_scales: Optional[torch.Tensor] = None,
852
853
854
855
856
857
        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)

858
    @register_fake("_C::machete_prepack_B")
859
860
861
    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:
862
863
        return torch.empty_like(b_q_weight,
                                memory_format=torch.contiguous_format)
864
865


866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
if hasattr(torch.ops._C, "ggml_dequantize"):

    @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)

    @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)
892
893


894
# cutlass
895
896
897
898
def cutlass_scaled_mm_supports_fp8(cuda_device_capability: int) -> bool:
    return torch.ops._C.cutlass_scaled_mm_supports_fp8(cuda_device_capability)


899
900
901
def cutlass_scaled_mm(a: torch.Tensor,
                      b: torch.Tensor,
                      scale_a: torch.Tensor,
902
                      scale_b: torch.Tensor,
903
                      out_dtype: torch.dtype,
904
                      bias: Optional[torch.Tensor] = None) -> torch.Tensor:
zhuwenwen's avatar
zhuwenwen committed
905
906
907
908
    # 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
909

zhuwenwen's avatar
zhuwenwen committed
910
911
    # m = a.shape[0]
    # n = b.shape[1]
912

zhuwenwen's avatar
zhuwenwen committed
913
914
915
916
917
918
    # 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)
919

zhuwenwen's avatar
zhuwenwen committed
920
    # out = torch.empty((m, n), dtype=out_dtype, device=a.device)
921

zhuwenwen's avatar
zhuwenwen committed
922
    # torch.ops._C.cutlass_scaled_mm(out, a, b, scale_a, scale_b, bias)
923

zhuwenwen's avatar
zhuwenwen committed
924
    # return out
gaoqiong's avatar
gaoqiong committed
925
926
    #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)
927

928
929
930
931
932
933
934
935
936
937
938
939
940
941
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
942
943
                      bias: Optional[torch.Tensor] = None,
                      best_config:Optional[list] = None) -> torch.Tensor:
944

gaoqiong's avatar
gaoqiong committed
945
    return quant_ops.triton_scaled_mm(a, b,scale_a,scale_b,out_dtype,bias,best_config)
946

gaoqiong's avatar
gaoqiong committed
947
948
949
950
951
952
953
954
955
956
957
958
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):
    return quant_tools.triton_int8_gemm_helper(m,n,k,per_token_act_quant,per_out_channel_weight_quant,use_bias,out_dtype,device,est_config)


959
960
961
962
963
964
965
966
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:
967
968
969
970
971
    """
    :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.
    """
972
973
974
975
    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
976
    assert azp is None or azp.numel() == a.shape[0]
977
978
979
980
981
982
983
984
985
986

    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


987
988
989
# aqlm
def aqlm_gemm(input: torch.Tensor, codes: torch.Tensor,
              codebooks: torch.Tensor, scales: torch.Tensor,
990
              codebook_partition_sizes: List[int],
991
              bias: Optional[torch.Tensor]) -> torch.Tensor:
992
993
    return torch.ops._C.aqlm_gemm(input, codes, codebooks, scales,
                                  codebook_partition_sizes, bias)
994
995
996


def aqlm_dequant(codes: torch.Tensor, codebooks: torch.Tensor,
997
                 codebook_partition_sizes: List[int]) -> torch.Tensor:
998
999
    return torch.ops._C.aqlm_dequant(codes, codebooks,
                                     codebook_partition_sizes)
1000
1001


1002
1003
# gptq_marlin
def gptq_marlin_repack(b_q_weight: torch.Tensor, perm: torch.Tensor,
1004
1005
                       size_k: int, size_n: int,
                       num_bits: int) -> torch.Tensor:
1006
1007
    return torch.ops._C.gptq_marlin_repack(b_q_weight, perm, size_k, size_n,
                                           num_bits)
1008
1009


1010
1011
1012
1013
1014
1015
# 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)


1016
1017
1018
1019
1020
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
1021
    output = torch.empty((num_experts, size_k // 16, size_n * (num_bits // 2)),
1022
1023
1024
1025
1026
1027
1028
1029
                         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


1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
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


1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
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,
1057
1058
                     use_fp32_reduce: bool = False,
                     is_zp_float: bool = False) -> torch.Tensor:
1059
    return torch.ops._C.gptq_marlin_gemm(a, b_q_weight, b_scales, b_zeros,
1060
                                         g_idx, perm, workspace, b_q_type.id,
1061
                                         size_m, size_n, size_k, is_k_full,
1062
                                         has_zp, use_fp32_reduce, is_zp_float)
1063
1064


1065
1066
1067
1068
1069
1070
1071
1072
1073
# 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)


1074
# machete
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
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)
1086
1087


1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
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)
1110
1111


1112
if hasattr(torch.ops._C, "permute_cols"):
1113

1114
    @register_fake("_C::permute_cols")
1115
1116
1117
1118
1119
1120
1121
1122
1123
    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)


1124
# fp8
zhuwenwen's avatar
zhuwenwen committed
1125
1126
1127
# def scaled_fp8_quant(
#     input: torch.Tensor,
#     scale: Optional[torch.Tensor] = None,
1128
#     num_token_padding: Optional[int] = None,
1129
1130
#     scale_ub: Optional[torch.Tensor] = None,
#     use_per_token_if_dynamic: bool = False,
zhuwenwen's avatar
zhuwenwen committed
1131
# ) -> Tuple[torch.Tensor, torch.Tensor]:
zhuwenwen's avatar
zhuwenwen committed
1132
1133
1134
1135
1136
1137
#     """
#     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
1138
#     optional padding of the output tensors for downstream kernels that
zhuwenwen's avatar
zhuwenwen committed
1139
1140
1141
1142
1143
#     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
1144
#         scale_ub: Optional upper bound for scaling factor in dynamic
1145
#             per token case
1146
#         num_token_padding: If specified, pad the first dimension
zhuwenwen's avatar
zhuwenwen committed
1147
#             of the output to at least this value.
zhuwenwen's avatar
zhuwenwen committed
1148
#         use_per_token_if_dynamic: Whether to do per_tensor or per_token
1149
#             in the dynamic quantization case.
zhuwenwen's avatar
zhuwenwen committed
1150
1151
1152
1153
1154

#     Returns:
#         Tuple[torch.Tensor, torch.Tensor]: The output tensor in FP8 and
#             scaling factor.
#     """
1155
1156
1157
#     # This code assumes batch_dim and num_tokens are flattened
#     assert (input.ndim == 2)
#     shape: Union[Tuple[int, int], torch.Size] = input.shape
1158
#     # For rocm, the output fp8 dtype is torch.float_e3m3fnuz
zhuwenwen's avatar
zhuwenwen committed
1159
1160
#     out_dtype: torch.dtype = torch.float8_e4m3fnuz \
#             if current_platform.is_rocm() else torch.float8_e4m3fn
1161
1162
#     if num_token_padding:
#         shape = (max(num_token_padding, input.shape[0]), shape[1])
1163
#     output = torch.empty(shape, device=input.device, dtype=out_dtype)
1164

zhuwenwen's avatar
zhuwenwen committed
1165
#     if scale is None:
1166
#         if use_per_token_if_dynamic:
1167
#             scale = torch.empty((shape[0], 1),
1168
1169
1170
1171
1172
1173
1174
#                                 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
1175
#     else:
1176
1177
#         # num_token_padding not implemented for this case
#         assert (scale.numel() == 1 or num_token_padding is None)
zhuwenwen's avatar
zhuwenwen committed
1178
#         torch.ops._C.static_scaled_fp8_quant(output, input, scale)
1179

zhuwenwen's avatar
zhuwenwen committed
1180
#     return output, scale
1181
1182


1183
# int8
1184
def scaled_int8_quant(
1185
1186
1187
1188
1189
    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
1190
    """
1191
    Quantize the input tensor to int8 and return the quantized tensor and scale, and maybe azp.
zhuwenwen's avatar
zhuwenwen committed
1192
1193
1194

    Args:
        input: The input tensor to be quantized to int8.
1195
1196
        scale: Optional scaling factor for the int8 quantization.
            When not provided, we invoke dynamic-per-token quantization.
1197
1198
1199
        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
1200
1201

    Returns:
1202
      Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]] : Output int8 tensor, scales, and optionally azp.
zhuwenwen's avatar
zhuwenwen committed
1203
    """
1204
1205
1206
    output = torch.empty_like(input, dtype=torch.int8)
    if scale is not None:
        # static-per-tensor quantization.
1207
1208
1209
1210
        assert symmetric == (
            azp is
            None), "azp must only be provided for asymmetric quantization."
        torch.ops._C.static_scaled_int8_quant(output, input, scale, azp)
1211
        return output, scale, azp
1212
1213
1214
1215
1216

    # dynamic-per-token quantization.
    input_scales = torch.empty((input.numel() // input.shape[-1], 1),
                               device=input.device,
                               dtype=torch.float32)
1217
1218
1219
1220
1221
    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
1222
1223


1224
1225
1226
1227
1228
1229
1230
1231
1232
# 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)


1233
# gguf
1234
1235
def ggml_dequantize(W: torch.Tensor, quant_type: int, m: int,
                    n: int) -> torch.Tensor:
1236
1237
1238
1239
1240
1241
1242
1243
    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,
1244
) -> torch.Tensor:
1245
1246
1247
1248
1249
1250
1251
1252
    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,
1253
) -> torch.Tensor:
1254
1255
1256
    return torch.ops._C.ggml_mul_mat_a8(W, X, quant_type, row)


1257
1258
1259
# mamba
def causal_conv1d_fwd(x: torch.Tensor, weight: torch.Tensor,
                      bias_: Optional[torch.Tensor],
1260
1261
1262
1263
                      conv_states: Optional[torch.Tensor],
                      query_start_loc: Optional[torch.Tensor],
                      cache_indices: Optional[torch.Tensor],
                      has_initial_state: Optional[torch.Tensor],
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
                      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)
1280
1281
1282
1283
1284
1285


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],
1286
1287
1288
1289
1290
                       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):
1291
1292
1293
    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,
1294
                                    ssm_states, pad_slot_id)
1295
1296


1297
# moe
1298
1299
1300
1301
def moe_sum(input: torch.Tensor, output: torch.Tensor):
    torch.ops._moe_C.moe_sum(input, output)


1302
1303
1304
1305
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:
1306
1307
1308
    torch.ops._moe_C.moe_align_block_size(topk_ids, num_experts, block_size,
                                          sorted_token_ids, experts_ids,
                                          num_tokens_post_pad)
1309
1310
1311
1312
1313
1314
1315


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)
1316
1317


1318
1319
if supports_moe_ops and hasattr(torch.ops._moe_C, "marlin_gemm_moe"):

1320
    @register_fake("_moe_C::marlin_gemm_moe")
1321
1322
1323
1324
    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,
1325
1326
                             b_zero_points: torch.Tensor, g_idx: torch.Tensor,
                             perm: torch.Tensor, workspace: torch.Tensor,
1327
1328
                             b_q_type: ScalarType, size_m: torch.SymInt,
                             size_n: torch.SymInt, size_k: torch.SymInt,
1329
1330
1331
1332
1333
1334
1335
1336
                             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)


1337
1338
1339
1340
1341
1342
1343
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,
1344
1345
    k_scale: float,
    v_scale: float,
1346
) -> None:
1347
1348
    torch.ops._C_cache_ops.reshape_and_cache(key, value, key_cache,
                                             value_cache, slot_mapping,
1349
                                             kv_cache_dtype, k_scale, v_scale)
1350
1351


1352
1353
1354
1355
1356
1357
1358
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,
1359
1360
    k_scale: float,
    v_scale: float,
1361
) -> None:
1362
1363
    torch.ops._C_cache_ops.reshape_and_cache_flash(key, value, key_cache,
                                                   value_cache, slot_mapping,
1364
1365
                                                   kv_cache_dtype, k_scale,
                                                   v_scale)
1366
1367


1368
1369
def copy_blocks(key_caches: List[torch.Tensor],
                value_caches: List[torch.Tensor],
1370
                block_mapping: torch.Tensor) -> None:
1371
    torch.ops._C_cache_ops.copy_blocks(key_caches, value_caches, block_mapping)
1372
1373
1374


def swap_blocks(src: torch.Tensor, dst: torch.Tensor,
1375
                block_mapping: torch.Tensor) -> None:
1376
    torch.ops._C_cache_ops.swap_blocks(src, dst, block_mapping)
1377
1378


1379
1380
1381
1382
def convert_fp8(output: torch.Tensor,
                input: torch.Tensor,
                scale: float = 1.0,
                kv_dtype: str = "fp8") -> None:
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
    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
1397
1398
1399
1400
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)
1401

1402

1403
1404
1405
1406
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)
1407

1408
1409
1410
1411
1412
1413
1414
1415
1416

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


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


1417
1418
def register_buffer(fa: int, ipc_tensors: List[int]) -> None:
    return torch.ops._C_custom_ar.register_buffer(fa, ipc_tensors)
1419
1420


1421
def get_graph_buffer_ipc_meta(fa: int) -> Tuple[List[int], List[int]]:
1422
1423
1424
    return torch.ops._C_custom_ar.get_graph_buffer_ipc_meta(fa)


1425
def register_graph_buffers(fa: int, handles: List[List[int]],
1426
1427
1428
                           offsets: List[List[int]]) -> None:
    torch.ops._C_custom_ar.register_graph_buffers(fa, handles, offsets)

1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
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)


1454

1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
# 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"
1470
                for arg in v.__annotations__.values()):
1471
1472
1473
1474
        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