_custom_ops.py 38.1 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 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
with contextlib.suppress(ImportError):
26
    import vllm._moe_C  # noqa: F401
27
28


29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
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


48
49
# activation ops
def silu_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
50
    torch.ops._C.silu_and_mul(out, x)
51
52
53


def gelu_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
54
    torch.ops._C.gelu_and_mul(out, x)
55
56
57


def gelu_tanh_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
58
    torch.ops._C.gelu_tanh_and_mul(out, x)
zhuwenwen's avatar
zhuwenwen committed
59
60
61
62
63
64
65
66
67
68
69
70
    
    
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)
71
72
73


def gelu_fast(out: torch.Tensor, x: torch.Tensor) -> None:
74
    torch.ops._C.gelu_fast(out, x)
75
76
77


def gelu_new(out: torch.Tensor, x: torch.Tensor) -> None:
78
    torch.ops._C.gelu_new(out, x)
79
80


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


85
86
87
88
89
90
91
92
93
# page attention ops
def paged_attention_v1(
    out: torch.Tensor,
    query: torch.Tensor,
    key_cache: torch.Tensor,
    value_cache: torch.Tensor,
    num_kv_heads: int,
    scale: float,
    block_tables: torch.Tensor,
94
    seq_lens: torch.Tensor,
95
    block_size: int,
96
    max_seq_len: int,
97
98
    alibi_slopes: Optional[torch.Tensor],
    kv_cache_dtype: str,
99
100
    k_scale: float,
    v_scale: float,
101
102
103
104
105
    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,
106
) -> None:
107
    torch.ops._C.paged_attention_v1(
108
109
        out, query, key_cache, value_cache, num_kv_heads, scale, block_tables,
        seq_lens, block_size, max_seq_len, alibi_slopes, kv_cache_dtype,
110
111
112
        k_scale, v_scale, tp_rank, blocksparse_local_blocks,
        blocksparse_vert_stride, blocksparse_block_size,
        blocksparse_head_sliding_step)
113
114
115
116
117
118
119
120
121
122
123
124
125


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,
126
    seq_lens: torch.Tensor,
127
    block_size: int,
128
    max_seq_len: int,
129
130
    alibi_slopes: Optional[torch.Tensor],
    kv_cache_dtype: str,
131
132
    k_scale: float,
    v_scale: float,
133
134
135
136
137
    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,
138
) -> None:
139
    torch.ops._C.paged_attention_v2(
140
141
        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,
142
        alibi_slopes, kv_cache_dtype, k_scale, v_scale, tp_rank,
143
144
        blocksparse_local_blocks, blocksparse_vert_stride,
        blocksparse_block_size, blocksparse_head_sliding_step)
145
146


zhuwenwen's avatar
zhuwenwen committed
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
# page attention ops (opt)
def paged_attention_v1_opt(
    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)
    
    
209
210
211
212
213
214
215
216
217
# 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:
218
219
    torch.ops._C.rotary_embedding(positions, query, key, head_size,
                                  cos_sin_cache, is_neox)
220
221
222
223
224
225
226


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:
227
228
229
    torch.ops._C.batched_rotary_embedding(positions, query, key, head_size,
                                          cos_sin_cache, is_neox, rot_dim,
                                          cos_sin_cache_offsets)
230
231
232
233
234


# layer norm ops
def rms_norm(out: torch.Tensor, input: torch.Tensor, weight: torch.Tensor,
             epsilon: float) -> None:
235
    torch.ops._C.rms_norm(out, input, weight, epsilon)
236
237
238
239


def fused_add_rms_norm(input: torch.Tensor, residual: torch.Tensor,
                       weight: torch.Tensor, epsilon: float) -> None:
240
    torch.ops._C.fused_add_rms_norm(input, residual, weight, epsilon)
zhuwenwen's avatar
zhuwenwen committed
241
242
243
244
245
246
247
248
249
250
251
    

# 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)
252
253


254
255
256
257
258
259
260
261
262
263
264
def advance_step(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:
    """Advance a step on GPU for existing inputs for a multi-step runner"""
    return torch.ops._C.advance_step(num_seqs, num_queries, block_size,
                                     input_tokens, sampled_token_ids,
                                     input_positions, seq_lens, slot_mapping,
                                     block_tables)

zhuwenwen's avatar
zhuwenwen committed
265
266
267
268
269
# 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)
    
270

271
272
# quantization ops
# awq
zhuwenwen's avatar
zhuwenwen committed
273
274
275
276
277
278
def GetAWQShareWorkspaceSize()->int:
    return quant_ops.GetAWQShareWorkspaceSize()

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

279
280
281
def awq_dequantize(qweight: torch.Tensor, scales: torch.Tensor,
                   zeros: torch.Tensor, split_k_iters: int, thx: int,
                   thy: int) -> torch.Tensor:
282
283
284
285
    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)
286
287
    return torch.ops._C.awq_dequantize(qweight, scales, zeros, split_k_iters,
                                       thx, thy)
288
289


gaoqiong's avatar
gaoqiong committed
290
291
# 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
292
293
294
295
296
297
#     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
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328

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)
329
330
331
332
333
334
335


# 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
336
    return quant_ops.gptq_gemm(a, b_q_weight, b_gptq_qzeros, b_gptq_scales,
337
                                  b_g_idx, use_exllama, bit)
338
339
    # return torch.ops._C.gptq_gemm(a, b_q_weight, b_gptq_qzeros, b_gptq_scales,
    #                               b_g_idx, use_exllama, bit)
340
341


342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
# TODO: has to be a better way to do this
try:
    torch.ops._C.gptq_gemm  # noqa B018

    @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)
except Exception:
    pass


358
359
def gptq_shuffle(q_weight: torch.Tensor, q_perm: torch.Tensor,
                 bit: int) -> None:
gaoqiong's avatar
gaoqiong committed
360
    quant_ops.gptq_shuffle(q_weight, q_perm, bit)
361
    # torch.ops._C.gptq_shuffle(q_weight, q_perm, bit)
362
363
364
365
366
367


# 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:
368
369
    return torch.ops._C.marlin_gemm(a, b_q_weight, b_scales, workspace, size_m,
                                    size_n, size_k)
370
371


372
373
374
# marlin_24
def gptq_marlin_24_gemm(a: torch.Tensor, b_q_weight: torch.Tensor,
                        b_meta: torch.Tensor, b_scales: torch.Tensor,
375
376
                        workspace: torch.Tensor, b_q_type: ScalarType,
                        size_m: int, size_n: int, size_k: int) -> torch.Tensor:
377
    return torch.ops._C.gptq_marlin_24_gemm(a, b_q_weight, b_meta, b_scales,
378
                                            workspace, b_q_type, size_m,
379
                                            size_n, size_k)
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
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
# TODO: has to be a better way to do this
try:
    torch.ops._C.gptq_marlin_24_gemm  # noqa B018

    @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,
        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:
        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:
        return torch.empty_like(b_q_weight)

    @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")
    def causal_conv1d_update_fake(x: torch.Tensor, conv_state: torch.Tensor,
                                  weight: torch.Tensor,
                                  bias_: Optional[torch.Tensor],
                                  silu_activation: bool) -> torch.Tensor:
        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 x is not None:
            b = x
        else:
            b = torch.empty((u.size(0), u.size(1), A.size(1)),
                            dtype=u.dtype,
                            device=u.device)
        if z_ is not None:
            c = torch.empty_like(z_)
            return [a, b, c]
        else:
            return [a, b]

except Exception:
    pass


570
# cutlass
571
572
573
574
def cutlass_scaled_mm_supports_fp8(cuda_device_capability: int) -> bool:
    return torch.ops._C.cutlass_scaled_mm_supports_fp8(cuda_device_capability)


575
576
577
def cutlass_scaled_mm(a: torch.Tensor,
                      b: torch.Tensor,
                      scale_a: torch.Tensor,
578
                      scale_b: torch.Tensor,
579
                      out_dtype: torch.dtype,
580
                      bias: Optional[torch.Tensor] = None) -> torch.Tensor:
581
582
    assert (b.shape[0] % 16 == 0 and b.shape[1] % 16 == 0)
    assert (out_dtype is torch.bfloat16 or out_dtype is torch.float16)
583
584
    assert bias is None or bias.shape[0] == b.shape[
        1] and bias.dtype == out_dtype
585
586
587
588
589

    m = a.shape[0]
    n = b.shape[1]
    out = torch.empty((m, n), dtype=out_dtype, device=a.device)

590
    torch.ops._C.cutlass_scaled_mm(out, a, b, scale_a, scale_b, bias)
591
592
593
594

    return out


595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
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


617
618
619
# aqlm
def aqlm_gemm(input: torch.Tensor, codes: torch.Tensor,
              codebooks: torch.Tensor, scales: torch.Tensor,
620
              codebook_partition_sizes: List[int],
621
              bias: Optional[torch.Tensor]) -> torch.Tensor:
622
623
    return torch.ops._C.aqlm_gemm(input, codes, codebooks, scales,
                                  codebook_partition_sizes, bias)
624
625
626


def aqlm_dequant(codes: torch.Tensor, codebooks: torch.Tensor,
627
                 codebook_partition_sizes: List[int]) -> torch.Tensor:
628
629
    return torch.ops._C.aqlm_dequant(codes, codebooks,
                                     codebook_partition_sizes)
630
631


632
633
# gptq_marlin
def gptq_marlin_repack(b_q_weight: torch.Tensor, perm: torch.Tensor,
634
635
                       size_k: int, size_n: int,
                       num_bits: int) -> torch.Tensor:
636
637
    return torch.ops._C.gptq_marlin_repack(b_q_weight, perm, size_k, size_n,
                                           num_bits)
638
639


640
641
642
643
644
645
# 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)


646
647
648
649
650
651
652
653
654
655
656
657
658
659
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
    output = torch.empty((num_experts, size_k // 16, size_n * 2),
                         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


660
661
662
663
664
665
666
667
668
669
670
671
672
673
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:
674
    return torch.ops._C.gptq_marlin_gemm(a, b_q_weight, b_scales, b_zeros,
675
                                         g_idx, perm, workspace, b_q_type,
676
                                         size_m, size_n, size_k, is_k_full,
677
                                         has_zp, use_fp32_reduce)
678
679


680
681
682
683
684
685
686
687
688
# 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)


689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
# 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)


715
# fp8
zhuwenwen's avatar
zhuwenwen committed
716
717
718
# def scaled_fp8_quant(
#     input: torch.Tensor,
#     scale: Optional[torch.Tensor] = None,
719
#     num_token_padding: Optional[int] = None,
720
721
#     scale_ub: Optional[torch.Tensor] = None,
#     use_per_token_if_dynamic: bool = False,
zhuwenwen's avatar
zhuwenwen committed
722
# ) -> Tuple[torch.Tensor, torch.Tensor]:
zhuwenwen's avatar
zhuwenwen committed
723
724
725
726
727
728
#     """
#     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
729
#     optional padding of the output tensors for downstream kernels that
zhuwenwen's avatar
zhuwenwen committed
730
731
732
733
734
#     will benefit from padding.

#     Args:
#         input: The input tensor to be quantized to FP8
#         scale: Optional scaling factor for the FP8 quantization
735
736
#         scale_ub: Optional upper bound for scaling factor in dynamic 
#             per token case
737
#         num_token_padding: If specified, pad the first dimension
zhuwenwen's avatar
zhuwenwen committed
738
#             of the output to at least this value.
739
740
#         use_per_token_if_dynamic: Whether to do per_tensor or per_token 
#             in the dynamic quantization case.
zhuwenwen's avatar
zhuwenwen committed
741
742
743
744
745

#     Returns:
#         Tuple[torch.Tensor, torch.Tensor]: The output tensor in FP8 and
#             scaling factor.
#     """
746
747
748
#     # This code assumes batch_dim and num_tokens are flattened
#     assert (input.ndim == 2)
#     shape: Union[Tuple[int, int], torch.Size] = input.shape
749
750
751
#     # 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
752
753
#     if num_token_padding:
#         shape = (max(num_token_padding, input.shape[0]), shape[1])
754
#     output = torch.empty(shape, device=input.device, dtype=out_dtype)
755

zhuwenwen's avatar
zhuwenwen committed
756
#     if scale is None:
757
#         if use_per_token_if_dynamic:
758
#             scale = torch.empty((shape[0], 1),
759
760
761
762
763
764
765
#                                 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
766
#     else:
767
768
#         # num_token_padding not implemented for this case
#         assert (scale.numel() == 1 or num_token_padding is None)
zhuwenwen's avatar
zhuwenwen committed
769
#         torch.ops._C.static_scaled_fp8_quant(output, input, scale)
770

zhuwenwen's avatar
zhuwenwen committed
771
#     return output, scale
772
773


774
# int8
775
776
777
778
def scaled_int8_quant(
        input: torch.Tensor,
        scale: Optional[torch.Tensor] = None
) -> Tuple[torch.Tensor, torch.Tensor]:
zhuwenwen's avatar
zhuwenwen committed
779
    """
780
    Quantize the input tensor to int8 and return the quantized tensor and scale.
zhuwenwen's avatar
zhuwenwen committed
781
782
783

    Args:
        input: The input tensor to be quantized to int8.
784
785
        scale: Optional scaling factor for the int8 quantization.
            When not provided, we invoke dynamic-per-token quantization.
zhuwenwen's avatar
zhuwenwen committed
786
787

    Returns:
788
      Tuple[Torch.Tensor, Torch.Tensor] : Output int8 tensor and scales.
zhuwenwen's avatar
zhuwenwen committed
789
    """
790
791
792
    output = torch.empty_like(input, dtype=torch.int8)
    if scale is not None:
        # static-per-tensor quantization.
793
        torch.ops._C.static_scaled_int8_quant(output, input, scale)
794
795
796
797
798
799
        return output, scale

    # dynamic-per-token quantization.
    input_scales = torch.empty((input.numel() // input.shape[-1], 1),
                               device=input.device,
                               dtype=torch.float32)
800
    torch.ops._C.dynamic_scaled_int8_quant(output, input, input_scales)
801
    return output, input_scales
802
803


804
805
806
807
808
809
810
811
812
# 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)


813
# gguf
814
815
def ggml_dequantize(W: torch.Tensor, quant_type: int, m: int,
                    n: int) -> torch.Tensor:
816
817
818
819
820
821
822
823
    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,
824
) -> torch.Tensor:
825
826
827
828
829
830
831
832
    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,
833
) -> torch.Tensor:
834
835
836
    return torch.ops._C.ggml_mul_mat_a8(W, X, quant_type, row)


837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
# 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)


def causal_conv1d_update(x: torch.Tensor, conv_state: torch.Tensor,
                         weight: torch.Tensor, bias_: Optional[torch.Tensor],
                         silu_activation: bool) -> torch.Tensor:
    return torch.ops._C.causal_conv1d_update(x, conv_state, weight, bias_,
                                             silu_activation)


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)


867
868
869
870
871
# 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:
872
873
874
875
876
877
878
879
880
881
    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)
882
883
884
885
886
887
888
889
890


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,
891
892
    k_scale: float,
    v_scale: float,
893
) -> None:
894
895
    torch.ops._C_cache_ops.reshape_and_cache(key, value, key_cache,
                                             value_cache, slot_mapping,
896
                                             kv_cache_dtype, k_scale, v_scale)
897
898


899
900
901
902
903
904
905
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,
906
907
    k_scale: float,
    v_scale: float,
908
) -> None:
909
910
    torch.ops._C_cache_ops.reshape_and_cache_flash(key, value, key_cache,
                                                   value_cache, slot_mapping,
911
912
                                                   kv_cache_dtype, k_scale,
                                                   v_scale)
913
914


915
916
def copy_blocks(key_caches: List[torch.Tensor],
                value_caches: List[torch.Tensor],
917
                block_mapping: torch.Tensor) -> None:
918
    torch.ops._C_cache_ops.copy_blocks(key_caches, value_caches, block_mapping)
919
920
921


def swap_blocks(src: torch.Tensor, dst: torch.Tensor,
922
                block_mapping: torch.Tensor) -> None:
923
    torch.ops._C_cache_ops.swap_blocks(src, dst, block_mapping)
924
925


926
927
928
929
def convert_fp8(output: torch.Tensor,
                input: torch.Tensor,
                scale: float = 1.0,
                kv_dtype: str = "fp8") -> None:
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
    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 should_custom_ar(inp: torch.Tensor, max_size: int, world_size: int,
                     full_nvlink: bool) -> bool:
    return torch.ops._C_custom_ar.should_custom_ar(inp, max_size, world_size,
                                                   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)

960

961
962
963
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)
964

965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987

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)


988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
# 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"
1003
                for arg in v.__annotations__.values()):
1004
1005
1006
1007
        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