_custom_ops.py 45.7 KB
Newer Older
1
import contextlib
2
import functools
3
from typing import TYPE_CHECKING, List, Optional, Tuple, Union
4
5

import torch
6
import torch.library
gaoqiong's avatar
gaoqiong committed
7

8
import vllm.envs as envs
9
from vllm._core_ext import ScalarType
10
from vllm.logger import init_logger
11
from vllm.platforms import current_platform
12

13
try:
gaoqiong's avatar
gaoqiong committed
14
    from lmslim import quant_ops 
15
except Exception:
gaoqiong's avatar
gaoqiong committed
16
    print("INFO: Please install lmslim if you want to infer gptq or awq model.\n") 
17

18
19
logger = init_logger(__name__)

20
21
22
23
24
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)
25

26
27
# if current_platform.is_rocm():
#     import vllm._rocm_C  # noqa: F401
28

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

34
35
36
37
38
39
40
41
42
43
if TYPE_CHECKING:

    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

44

45
46
47
48
49
50
def hint_on_error(fn):

    @functools.wraps(fn)
    def wrapper(*args, **kwargs):
        try:
            return fn(*args, **kwargs)
51
52
53
54
55
56
57
58
59

        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
60
61
62
63
64
65
66
67
68
69
70
71
72
        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


73
74
# activation ops
def silu_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
75
    torch.ops._C.silu_and_mul(out, x)
76
77
78


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


def gelu_tanh_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
83
    torch.ops._C.gelu_tanh_and_mul(out, x)
zhuwenwen's avatar
zhuwenwen committed
84
85
86
87
88
89
90
91
92
93
94
95
    
    
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)
96
97
98


def gelu_fast(out: torch.Tensor, x: torch.Tensor) -> None:
99
    torch.ops._C.gelu_fast(out, x)
100
101
102


def gelu_new(out: torch.Tensor, x: torch.Tensor) -> None:
103
    torch.ops._C.gelu_new(out, x)
104
105


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


110
111
112
113
114
115
116
117
118
# 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,
119
    seq_lens: torch.Tensor,
120
    block_size: int,
121
    max_seq_len: int,
122
123
    alibi_slopes: Optional[torch.Tensor],
    kv_cache_dtype: str,
124
125
    k_scale: float,
    v_scale: float,
126
127
128
129
130
    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,
131
) -> None:
132
    torch.ops._C.paged_attention_v1(
133
134
        out, query, key_cache, value_cache, num_kv_heads, scale, block_tables,
        seq_lens, block_size, max_seq_len, alibi_slopes, kv_cache_dtype,
135
136
137
        k_scale, v_scale, tp_rank, blocksparse_local_blocks,
        blocksparse_vert_stride, blocksparse_block_size,
        blocksparse_head_sliding_step)
138
139
140
141
142
143
144
145
146
147
148
149
150


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,
151
    seq_lens: torch.Tensor,
152
    block_size: int,
153
    max_seq_len: int,
154
155
    alibi_slopes: Optional[torch.Tensor],
    kv_cache_dtype: str,
156
157
    k_scale: float,
    v_scale: float,
158
159
160
161
162
    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,
163
) -> None:
164
    torch.ops._C.paged_attention_v2(
165
166
        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,
167
        alibi_slopes, kv_cache_dtype, k_scale, v_scale, tp_rank,
168
169
        blocksparse_local_blocks, blocksparse_vert_stride,
        blocksparse_block_size, blocksparse_head_sliding_step)
170
171


zhuwenwen's avatar
zhuwenwen committed
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
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
# 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)
zhuwenwen's avatar
zhuwenwen committed
232
233


234
# page attention ops (opt_tc)
235
236
237
238
239
240
241
242
243
244
245
246
247
def paged_attention_v1_opt_tc(
    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
248
249
    k_scale: float,
    v_scale: float,
250
251
252
253
254
255
256
257
258
    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,
zhuwenwen's avatar
zhuwenwen committed
259
        k_scale, v_scale, tp_rank, blocksparse_local_blocks, blocksparse_vert_stride,
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
        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,
zhuwenwen's avatar
zhuwenwen committed
279
280
    k_scale: float,
    v_scale: float,
281
282
283
284
285
286
287
288
289
    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,
zhuwenwen's avatar
zhuwenwen committed
290
        alibi_slopes, kv_cache_dtype, k_scale, v_scale, tp_rank,
291
292
293
294
        blocksparse_local_blocks, blocksparse_vert_stride,
        blocksparse_block_size, blocksparse_head_sliding_step)


295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
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,
311
312
    k_scale: float,
    v_scale: float,
313
314
315
316
317
) -> 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,
318
                                      kv_cache_dtype, k_scale, v_scale)
319
320


321
322
323
324
325
326
327
328
329
# 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:
330
331
    torch.ops._C.rotary_embedding(positions, query, key, head_size,
                                  cos_sin_cache, is_neox)
332
333
334
335
336
337
338


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:
339
340
341
    torch.ops._C.batched_rotary_embedding(positions, query, key, head_size,
                                          cos_sin_cache, is_neox, rot_dim,
                                          cos_sin_cache_offsets)
342
343
344
345
346


# layer norm ops
def rms_norm(out: torch.Tensor, input: torch.Tensor, weight: torch.Tensor,
             epsilon: float) -> None:
347
    torch.ops._C.rms_norm(out, input, weight, epsilon)
348
349
350
351


def fused_add_rms_norm(input: torch.Tensor, residual: torch.Tensor,
                       weight: torch.Tensor, epsilon: float) -> None:
352
    torch.ops._C.fused_add_rms_norm(input, residual, weight, epsilon)
zhuwenwen's avatar
zhuwenwen committed
353
354
355
356
357
358
359
360
361
362
363
    

# 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)
364
365


366
367
368
369
370
371
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:
372
    """Advance a step on GPU for existing inputs for a multi-step runner"""
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
    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)
396

zhuwenwen's avatar
zhuwenwen committed
397
398
399
400
401
# 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)
    
402

403
404
# quantization ops
# awq
zhuwenwen's avatar
zhuwenwen committed
405
406
407
408
409
410
def GetAWQShareWorkspaceSize()->int:
    return quant_ops.GetAWQShareWorkspaceSize()

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

411
412
413
def awq_dequantize(qweight: torch.Tensor, scales: torch.Tensor,
                   zeros: torch.Tensor, split_k_iters: int, thx: int,
                   thy: int) -> torch.Tensor:
414
415
416
417
    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)
418
419
    return torch.ops._C.awq_dequantize(qweight, scales, zeros, split_k_iters,
                                       thx, thy)
420
421


gaoqiong's avatar
gaoqiong committed
422
423
# 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
424
425
426
427
428
429
#     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
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

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)
461
462
463
464
465
466
467


# 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
468
    return quant_ops.gptq_gemm(a, b_q_weight, b_gptq_qzeros, b_gptq_scales,
469
                                  b_g_idx, use_exllama, bit)
470
471
    # return torch.ops._C.gptq_gemm(a, b_q_weight, b_gptq_qzeros, b_gptq_scales,
    #                               b_g_idx, use_exllama, bit)
472
473


474
if hasattr(torch.ops._C, "gptq_gemm"):
475

476
    @register_fake("_C::gptq_gemm")
477
478
479
480
481
482
483
484
485
    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)


486
487
def gptq_shuffle(q_weight: torch.Tensor, q_perm: torch.Tensor,
                 bit: int) -> None:
gaoqiong's avatar
gaoqiong committed
488
    quant_ops.gptq_shuffle(q_weight, q_perm, bit)
489
    # torch.ops._C.gptq_shuffle(q_weight, q_perm, bit)
490
491
492
493
494
495


# 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:
496
497
    return torch.ops._C.marlin_gemm(a, b_q_weight, b_scales, workspace, size_m,
                                    size_n, size_k)
498
499


500
501
502
# marlin_24
def gptq_marlin_24_gemm(a: torch.Tensor, b_q_weight: torch.Tensor,
                        b_meta: torch.Tensor, b_scales: torch.Tensor,
503
504
                        workspace: torch.Tensor, b_q_type: ScalarType,
                        size_m: int, size_n: int, size_k: int) -> torch.Tensor:
505
    return torch.ops._C.gptq_marlin_24_gemm(a, b_q_weight, b_meta, b_scales,
506
                                            workspace, b_q_type, size_m,
507
                                            size_n, size_k)
508
509


510
if hasattr(torch.ops._C, "gptq_marlin_24_gemm"):
511

512
    @register_fake("_C::gptq_marlin_24_gemm")
513
514
515
516
517
518
519
    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)

520
    @register_fake("_C::gptq_marlin_gemm")
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
    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)

537
    @register_fake("_C::ggml_dequantize")
538
539
540
541
    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)

542
    @register_fake("_C::ggml_mul_mat_vec_a8")
543
544
545
546
547
548
549
550
    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)

551
    @register_fake("_C::ggml_mul_mat_a8")
552
553
554
555
556
557
558
559
560
    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)

561
    @register_fake("_C::marlin_qqq_gemm")
562
563
564
565
566
567
568
569
570
    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)

571
    @register_fake("_C::marlin_gemm")
572
573
574
575
576
577
578
579
    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)

580
    @register_fake("_C::awq_dequantize")
581
582
583
584
585
586
587
588
589
590
    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)

591
    @register_fake("_C::awq_gemm")
592
593
594
595
596
597
598
599
    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)

600
    @register_fake("_C::aqlm_gemm")
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
    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))

616
    @register_fake("_C::aqlm_dequant")
617
618
619
620
621
622
623
624
625
    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)

626
    @register_fake("_C::fp8_marlin_gemm")
627
628
629
630
631
632
    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)

633
    @register_fake("_C::machete_gemm")
634
635
    def machete_gemm_fake(
        a: torch.Tensor,
636
637
        # Should be the tensor returned by machete_prepack_B
        b_q: torch.Tensor,
638
639
640
641
642
643
644
645
646
647
648
649
650
        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)

651
    @register_fake("_C::machete_prepack_B")
652
653
    def machete_prepack_B_fake(b_q_weight: torch.Tensor,
                               b_type: ScalarType) -> torch.Tensor:
654
655
        return torch.empty_like(b_q_weight,
                                memory_format=torch.contiguous_format)
656

657
    @register_fake("_C::causal_conv1d_fwd")
658
659
    def causal_conv1d_fwd_fake(x: torch.Tensor, weight: torch.Tensor,
                               bias_: Optional[torch.Tensor],
660
661
662
663
                               conv_states: Optional[torch.Tensor],
                               cu_seq_len: Optional[torch.Tensor],
                               cache_indices: Optional[torch.Tensor],
                               has_initial_state: Optional[torch.Tensor],
664
665
                               silu_activation: bool, pad_slot_id: int):
        return None
666

667
    @register_fake("_C::causal_conv1d_update")
668
669
670
671
672
673
674
675
    def causal_conv1d_update_fake(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) -> None:
        return None
676

677
    @register_fake("_C::selective_scan_fwd")
678
679
680
681
682
683
684
685
686
    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,
                                cu_seq_len: Optional[torch.Tensor],
                                cache_indices: Optional[torch.Tensor],
                                has_initial_state: Optional[torch.Tensor],
687
688
                                ssm_states: Optional[torch.Tensor],
                                pad_slot_id: int) -> None:
689
        return None
690
691


692
# cutlass
693
694
695
696
def cutlass_scaled_mm_supports_fp8(cuda_device_capability: int) -> bool:
    return torch.ops._C.cutlass_scaled_mm_supports_fp8(cuda_device_capability)


697
698
699
def cutlass_scaled_mm(a: torch.Tensor,
                      b: torch.Tensor,
                      scale_a: torch.Tensor,
700
                      scale_b: torch.Tensor,
701
                      out_dtype: torch.dtype,
702
                      bias: Optional[torch.Tensor] = None) -> torch.Tensor:
zhuwenwen's avatar
zhuwenwen committed
703
704
705
706
    # 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
707

zhuwenwen's avatar
zhuwenwen committed
708
709
710
    # m = a.shape[0]
    # n = b.shape[1]
    # out = torch.empty((m, n), dtype=out_dtype, device=a.device)
711

zhuwenwen's avatar
zhuwenwen committed
712
    # torch.ops._C.cutlass_scaled_mm(out, a, b, scale_a, scale_b, bias)
713

zhuwenwen's avatar
zhuwenwen committed
714
715
    # return out
    return quant_ops.rocblas_scaled_mm_nn(a, b, scale_a, scale_b, out_dtype, bias)
716
717


718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
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


740
741
742
# aqlm
def aqlm_gemm(input: torch.Tensor, codes: torch.Tensor,
              codebooks: torch.Tensor, scales: torch.Tensor,
743
              codebook_partition_sizes: List[int],
744
              bias: Optional[torch.Tensor]) -> torch.Tensor:
745
746
    return torch.ops._C.aqlm_gemm(input, codes, codebooks, scales,
                                  codebook_partition_sizes, bias)
747
748
749


def aqlm_dequant(codes: torch.Tensor, codebooks: torch.Tensor,
750
                 codebook_partition_sizes: List[int]) -> torch.Tensor:
751
752
    return torch.ops._C.aqlm_dequant(codes, codebooks,
                                     codebook_partition_sizes)
753
754


755
756
# gptq_marlin
def gptq_marlin_repack(b_q_weight: torch.Tensor, perm: torch.Tensor,
757
758
                       size_k: int, size_n: int,
                       num_bits: int) -> torch.Tensor:
759
760
    return torch.ops._C.gptq_marlin_repack(b_q_weight, perm, size_k, size_n,
                                           num_bits)
761
762


763
764
765
766
767
768
# 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)


769
770
771
772
773
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
774
    output = torch.empty((num_experts, size_k // 16, size_n * (num_bits // 2)),
775
776
777
778
779
780
781
782
                         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


783
784
785
786
787
788
789
790
791
792
793
794
795
796
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


797
798
799
800
801
802
803
804
805
806
807
808
809
810
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:
811
    return torch.ops._C.gptq_marlin_gemm(a, b_q_weight, b_scales, b_zeros,
812
                                         g_idx, perm, workspace, b_q_type,
813
                                         size_m, size_n, size_k, is_k_full,
814
                                         has_zp, use_fp32_reduce)
815
816


817
818
819
820
821
822
823
824
825
# 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)


826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
# 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)


852
if hasattr(torch.ops._C, "permute_cols"):
853

854
    @register_fake("_C::permute_cols")
855
856
857
858
859
860
861
862
863
    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)


864
# fp8
zhuwenwen's avatar
zhuwenwen committed
865
866
867
# def scaled_fp8_quant(
#     input: torch.Tensor,
#     scale: Optional[torch.Tensor] = None,
868
#     num_token_padding: Optional[int] = None,
869
870
#     scale_ub: Optional[torch.Tensor] = None,
#     use_per_token_if_dynamic: bool = False,
zhuwenwen's avatar
zhuwenwen committed
871
# ) -> Tuple[torch.Tensor, torch.Tensor]:
zhuwenwen's avatar
zhuwenwen committed
872
873
874
875
876
877
#     """
#     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
878
#     optional padding of the output tensors for downstream kernels that
zhuwenwen's avatar
zhuwenwen committed
879
880
881
882
883
#     will benefit from padding.

#     Args:
#         input: The input tensor to be quantized to FP8
#         scale: Optional scaling factor for the FP8 quantization
884
885
#         scale_ub: Optional upper bound for scaling factor in dynamic 
#             per token case
886
#         num_token_padding: If specified, pad the first dimension
zhuwenwen's avatar
zhuwenwen committed
887
#             of the output to at least this value.
888
889
#         use_per_token_if_dynamic: Whether to do per_tensor or per_token 
#             in the dynamic quantization case.
zhuwenwen's avatar
zhuwenwen committed
890
891
892
893
894

#     Returns:
#         Tuple[torch.Tensor, torch.Tensor]: The output tensor in FP8 and
#             scaling factor.
#     """
895
896
897
#     # This code assumes batch_dim and num_tokens are flattened
#     assert (input.ndim == 2)
#     shape: Union[Tuple[int, int], torch.Size] = input.shape
898
899
900
#     # 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
901
902
#     if num_token_padding:
#         shape = (max(num_token_padding, input.shape[0]), shape[1])
903
#     output = torch.empty(shape, device=input.device, dtype=out_dtype)
904

zhuwenwen's avatar
zhuwenwen committed
905
#     if scale is None:
906
#         if use_per_token_if_dynamic:
907
#             scale = torch.empty((shape[0], 1),
908
909
910
911
912
913
914
#                                 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
915
#     else:
916
917
#         # num_token_padding not implemented for this case
#         assert (scale.numel() == 1 or num_token_padding is None)
zhuwenwen's avatar
zhuwenwen committed
918
#         torch.ops._C.static_scaled_fp8_quant(output, input, scale)
919

zhuwenwen's avatar
zhuwenwen committed
920
#     return output, scale
921
922


923
# int8
924
def scaled_int8_quant(
925
926
927
928
929
    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
930
    """
931
    Quantize the input tensor to int8 and return the quantized tensor and scale, and maybe azp.
zhuwenwen's avatar
zhuwenwen committed
932
933
934

    Args:
        input: The input tensor to be quantized to int8.
935
936
        scale: Optional scaling factor for the int8 quantization.
            When not provided, we invoke dynamic-per-token quantization.
937
938
939
        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
940
941

    Returns:
942
      Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]] : Output int8 tensor, scales, and optionally azp.
zhuwenwen's avatar
zhuwenwen committed
943
    """
944
945
946
    output = torch.empty_like(input, dtype=torch.int8)
    if scale is not None:
        # static-per-tensor quantization.
947
948
949
950
951
        assert symmetric == (
            azp is
            None), "azp must only be provided for asymmetric quantization."
        torch.ops._C.static_scaled_int8_quant(output, input, scale, azp)
        return output, scale, None
952
953
954
955
956

    # dynamic-per-token quantization.
    input_scales = torch.empty((input.numel() // input.shape[-1], 1),
                               device=input.device,
                               dtype=torch.float32)
957
958
959
960
961
    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
962
963


964
965
966
967
968
969
970
971
972
# 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)


973
# gguf
974
975
def ggml_dequantize(W: torch.Tensor, quant_type: int, m: int,
                    n: int) -> torch.Tensor:
976
977
978
979
980
981
982
983
    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,
984
) -> torch.Tensor:
985
986
987
988
989
990
991
992
    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,
993
) -> torch.Tensor:
994
995
996
    return torch.ops._C.ggml_mul_mat_a8(W, X, quant_type, row)


997
998
999
# mamba
def causal_conv1d_fwd(x: torch.Tensor, weight: torch.Tensor,
                      bias_: Optional[torch.Tensor],
1000
1001
1002
1003
                      conv_states: Optional[torch.Tensor],
                      query_start_loc: Optional[torch.Tensor],
                      cache_indices: Optional[torch.Tensor],
                      has_initial_state: Optional[torch.Tensor],
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
                      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)
1020
1021
1022
1023
1024
1025


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],
1026
1027
1028
1029
1030
                       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):
1031
1032
1033
    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,
1034
                                    ssm_states, pad_slot_id)
1035
1036


1037
1038
1039
1040
1041
# 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:
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
    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)
1052
1053


1054
1055
if supports_moe_ops and hasattr(torch.ops._moe_C, "marlin_gemm_moe"):

1056
    @register_fake("_moe_C::marlin_gemm_moe")
1057
1058
1059
1060
    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,
1061
1062
1063
1064
1065
1066
                             b_zero_points: torch.Tensor, g_idx: torch.Tensor,
                             perm: torch.Tensor, workspace: torch.Tensor,
                             b_q_type: ScalarType, size_m: int, size_n: int,
                             size_k: int, is_k_full: bool, num_experts: int,
                             topk: int, moe_block_size: int,
                             replicate_input: bool,
1067
1068
1069
1070
1071
1072
                             apply_weights: bool) -> torch.Tensor:
        return torch.empty((size_m, topk, size_n),
                           dtype=a.dtype,
                           device=a.device)


1073
1074
1075
1076
1077
1078
1079
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,
1080
1081
    k_scale: float,
    v_scale: float,
1082
) -> None:
1083
1084
    torch.ops._C_cache_ops.reshape_and_cache(key, value, key_cache,
                                             value_cache, slot_mapping,
1085
                                             kv_cache_dtype, k_scale, v_scale)
1086
1087


1088
1089
1090
1091
1092
1093
1094
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,
1095
1096
    k_scale: float,
    v_scale: float,
1097
) -> None:
1098
1099
    torch.ops._C_cache_ops.reshape_and_cache_flash(key, value, key_cache,
                                                   value_cache, slot_mapping,
1100
1101
                                                   kv_cache_dtype, k_scale,
                                                   v_scale)
1102
1103


1104
1105
def copy_blocks(key_caches: List[torch.Tensor],
                value_caches: List[torch.Tensor],
1106
                block_mapping: torch.Tensor) -> None:
1107
    torch.ops._C_cache_ops.copy_blocks(key_caches, value_caches, block_mapping)
1108
1109
1110


def swap_blocks(src: torch.Tensor, dst: torch.Tensor,
1111
                block_mapping: torch.Tensor) -> None:
1112
    torch.ops._C_cache_ops.swap_blocks(src, dst, block_mapping)
1113
1114


1115
1116
1117
1118
def convert_fp8(output: torch.Tensor,
                input: torch.Tensor,
                scale: float = 1.0,
                kv_dtype: str = "fp8") -> None:
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
    torch.ops._C_cache_ops.convert_fp8(output, input, scale, kv_dtype)


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


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


# custom ar
def init_custom_ar(meta: torch.Tensor, rank_data: torch.Tensor,
                   handles: List[str], offsets: List[int], rank: int,
                   full_nvlink: bool) -> int:
    return torch.ops._C_custom_ar.init_custom_ar(meta, rank_data, handles,
                                                 offsets, rank, full_nvlink)


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

1143

1144
1145
1146
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)
1147

1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170

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)


1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
# 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"
1186
                for arg in v.__annotations__.values()):
1187
1188
1189
1190
        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