Repository navigation
Expand file tree
/
Copy path_xpu_ops.py
<
8000
svg data-component="Octicon" aria-hidden="true" focusable="false" class="octicon octicon-search" viewBox="0 0 16 16" width="16" height="16" fill="currentColor" display="inline-block" overflow="visible" style="vertical-align:text-bottom">
More file actions
624 lines (540 loc) · 21.9 KB
/
Copy path_xpu_ops.py
File metadata and controls
624 lines (540 loc) · 21.9 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
8000
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
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
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
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
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from typing import TYPE_CHECKING
import torch
from vllm_xpu_kernels.flash_attn_interface import flash_attn_varlen_func
from vllm.logger import init_logger
from vllm.platforms import current_platform
from vllm.utils.torch_utils import direct_register_custom_op
logger = init_logger(__name__)
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
if hasattr(torch.ops._xpu_C, "fp8_gemm"):
@register_fake("_xpu_C::fp8_gemm")
def _fp8_gemm_fake(
q_input: torch.Tensor,
q_weight: torch.Tensor,
out_dtype: torch.dtype,
input_scales: torch.Tensor,
weight_scale: torch.Tensor,
bias: torch.Tensor | None = None,
) -> torch.Tensor:
input_2d = q_input.view(-1, q_input.shape[-1])
M = input_2d.size(0)
N = q_weight.size(1)
return torch.empty((M, N), dtype=out_dtype, device=q_input.device)
if hasattr(torch.ops._xpu_C, "fp8_gemm_w8a16"):
@register_fake("_xpu_C::fp8_gemm_w8a16")
def _fp8_gemm_w8a16_fake(
input: torch.Tensor,
q_weight: torch.Tensor,
weight_scale: torch.Tensor,
bias: torch.Tensor | None = None,
) -> torch.Tensor:
input_2d = input.view(-1, input.shape[-1])
M = input_2d.size(0)
N = q_weight.size(1)
return torch.empty((M, N), dtype=input.dtype, device=input.device)
if hasattr(torch.ops._xpu_C, "int4_gemm_w4a8"):
@register_fake("_xpu_C::int4_gemm_w4a8")
def _int4_gemm_w4a8_fake(
input: torch.Tensor,
input_scales: torch.Tensor,
input_zero_points: torch.Tensor,
q_weight: torch.Tensor,
weight_scale: torch.Tensor,
weight_zp: torch.Tensor,
group_size: int,
g_idx: torch.Tensor | None = None,
bias: torch.Tensor | None = None,
) -> torch.Tensor:
input_2d = input.view(-1, input.shape[-1])
M = input_2d.size(0)
N = q_weight.size(1)
return torch.empty((M, N), dtype=torch.float16, device=input.device)
if hasattr(torch.ops._xpu_C, "int4_gemm_w4a16"):
@register_fake("_xpu_C::int4_gemm_w4a16")
def _int4_gemm_w4a16_fake(
input: torch.Tensor,
q_weight: torch.Tensor,
bias: torch.Tensor | None,
weight_scale: torch.Tensor,
qzeros: torch.Tensor,
group_size: int,
group_idx: torch.Tensor | None = None,
) -> torch.Tensor:
input_2d = input.view(-1, input.shape[-1])
M = input_2d.size(0)
N = q_weight.size(1)
return torch.empty((M, N), dtype=input.dtype, device=input.device)
def _xpu_ops_deepseek_scaling_rope_impl(
positions: torch.Tensor,
query: torch.Tensor,
key: torch.Tensor | None,
offsets: torch.Tensor | None,
cos_sin_cache: torch.Tensor | None,
rotary_dim: int,
is_neox_style: bool,
) -> tuple[torch.Tensor, torch.Tensor]:
assert key is not None
return torch.ops._xpu_C.deepseek_scaling_rope(
positions, query, key, offsets, cos_sin_cache, rotary_dim, is_neox_style
)
def _xpu_ops_deepseek_scaling_rope_fake(
positions: torch.Tensor,
query: torch.Tensor,
key: torch.Tensor | None,
offsets: torch.Tensor | None,
cos_sin_cache: torch.Tensor | None,
rotary_dim: int,
is_neox_style: bool,
) -> tuple[torch.Tensor, torch.Tensor]:
return query, key
def _xpu_mxfp8_quantize_impl(
x: torch.Tensor, dtype: torch.dtype | None = None
) -> tuple[torch.Tensor, torch.Tensor]:
MXFP8_BLOCK_SIZE = 32
assert x.shape[-1] % MXFP8_BLOCK_SIZE == 0
if dtype is not None:
assert dtype in (torch.float8_e4m3fn, torch.float8_e5m2), (
f"Unsupported dtype for xpu_mxfp8_quantize: {dtype}. "
f"Expected torch.float8_e4m3fn or torch.float8_e5m2."
)
else:
dtype = current_platform.fp8_dtype()
finfo = torch.finfo(dtype)
fp8_min = finfo.min
fp8_max = finfo.max
eps = 1e-10
x_q = torch.empty_like(x, device=x.device, dtype=dtype)
shape = x.shape[:-1] + (x.shape[-1] // MXFP8_BLOCK_SIZE,)
x_s = torch.empty(shape, device=x.device, dtype=torch.float32)
torch.ops._C.per_token_group_fp8_quant(
x, x_q, x_s, MXFP8_BLOCK_SIZE, eps, fp8_min, fp8_max, True
)
x_s = x_s.to(torch.float8_e8m0fnu)
return x_q, x_s
def _xpu_mxfp8_quantize_fake(
x: torch.Tensor, dtype: torch.dtype | None = None
) -> tuple[torch.Tensor, torch.Tensor]:
if dtype is None:
dtype = current_platform.fp8_dtype()
MXFP8_BLOCK_SIZE = 32
shape = x.shape[:-1] + (x.shape[-1] // MXFP8_BLOCK_SIZE,)
x_s = torch.zeros(shape, device=x.device, dtype=torch.float32)
return x.to(dtype), x_s.to(torch.float8_e8m0fnu)
def _xpu_mxfp4_quantize_impl(
x: torch.Tensor,
) -> tuple[torch.Tensor, torch.Tensor]:
MXFP4_BLOCK_SIZE = 32
eps = 1e-10
assert x.ndim == 2, "input must be 2-D"
assert x.shape[-1] % MXFP4_BLOCK_SIZE == 0, (
f"last dimension {x.shape[-1]} must be divisible by group_size "
f"{MXFP4_BLOCK_SIZE}"
)
assert x.is_contiguous(), "input groups must be contiguous"
M, N = x.shape
# Packed FP4 output: two nibbles per byte
x_q = torch.empty(M, N // 2, device=x.device, dtype=torch.uint8)
x_s = torch.empty(M, N // MXFP4_BLOCK_SIZE, device=x.device, dtype=torch.float32)
torch.ops._C.per_token_group_quant_mxfp4(x, x_q, x_s, MXFP4_BLOCK_SIZE, eps)
x_q = x_q.view(torch.float4_e2m1fn_x2)
x_s = x_s.to(dtype=torch.float8_e8m0fnu, memory_format=torch.preserve_format)
return x_q, x_s
def _xpu_mxfp4_quantize_fake(
x: torch.Tensor,
) -> tuple[torch.Tensor, torch.Tensor]:
MXFP4_BLOCK_SIZE = 32
M, N = x.shape
# Packed FP4 output: two nibbles per byte
x_q = torch.empty(M, N // 2, device=x.device, dtype=torch.uint8)
x_s = torch.empty(M, N // MXFP4_BLOCK_SIZE, device=x.device, dtype=torch.float32)
x_q = x_q.view(torch.float4_e2m1fn_x2)
x_s = x_s.to(dtype=torch.float8_e8m0fnu, memory_format=torch.preserve_format)
return x_q, x_s
# Global flag to ensure ops are registered only once
_OPS_REGISTERED = False
class xpu_ops:
@staticmethod
@torch.compile
def dynamic_per_token_int8_quant_ref(
input: torch.Tensor, use_sym_quant: bool, bits: int
):
original_sizes = input.size()
# view is not safe in torch.compile if input is not contiguous
input = input.reshape(
-1, original_sizes[-1]
) # Flatten except for the last dimension
qmin = -(2 ** (bits - 1)) if use_sym_quant else 0
qmax = 2 ** (bits - 1) - 1 if use_sym_quant else 2**bits - 1
min_val = torch.min(input, dim=-1)[0].to(dtype=torch.float32).unsqueeze(-1)
max_val = torch.max(input, dim=-1)[0].to(dtype=torch.float32).unsqueeze(-1)
if use_sym_quant:
scale = (
torch.maximum(torch.abs(min_val), torch.abs(max_val)) / qmax
).clamp(min=1e-5)
zero_point = torch.zeros_like(scale).to(dtype=torch.int32)
else:
scale = ((max_val - min_val) / qmax).clamp(min=1e-5)
zero_point = -1 * torch.round(min_val / scale).to(dtype=torch.int32)
scale = scale.to(dtype=input.dtype)
quantized = torch.clamp(
torch.round(input / scale.to(dtype=torch.float32) + zero_point),
qmin,
qmax,
).to(dtype=torch.int8 if use_sym_quant else torch.uint8)
return (
quantized.view(original_sizes),
scale.view(original_sizes[:-1] + (1,)),
zero_point.view(original_sizes[:-1] + (1,)),
)
@staticmethod
def flash_attn_varlen_func(
q: torch.Tensor,
k: torch.Tensor,
v: torch.Tensor,
cu_seqlens_q: torch.Tensor,
max_seqlen_q: int,
max_seqlen_k: int,
softmax_scale: float | None = None,
causal: bool = False,
out: torch.Tensor | None = None,
block_table: torch.Tensor | None = None,
alibi_slopes: torch.Tensor | None = None,
window_size: list[int] | None = None,
softcap: float | None = 0.0,
seqused_k: torch.Tensor | None = None,
cu_seqlens_k: torch.Tensor | None = None,
# passed in qwen vl
dropout_p: float = 0.0,
# The following parameters are not used in xpu kernel currently,
# we keep API compatible to CUDA's.
scheduler_metadata=None,
fa_version: int = 2,
q_descale=None,
k_descale=None,
v_descale=None,
num_splits=0,
return_softmax_lse: bool | None = False,
s_aux: torch.Tensor | None = None,
return_attn_probs: bool | None = False,
):
assert cu_seqlens_k is not None or seqused_k is not None, (
"cu_seqlens_k or seqused_k must be provided"
)
assert cu_seqlens_k is None or seqused_k is None, (
"cu_seqlens_k and seqused_k cannot be provided at the same time"
)
assert block_table is None or seqused_k is not None, (
"when enable block_table, seqused_k is needed"
)
assert block_table is not None or cu_seqlens_k is not None, (
"when block_table is disabled, cu_seqlens_k is needed"
)
if out is None:
out = torch.empty(q.shape, dtype=q.dtype, device=q.device)
real_window_size: tuple[int, int]
if window_size is None:
real_window_size = (-1, -1)
else:
assert len(window_size) == 2
real_window_size = (window_size[0], window_size[1]) # noqa: F841
# In encode attention, k and v maybe not contiguous and current
# kernel can't handle it
if block_table is None:
k = k.contiguous()
v = v.contiguous()
return flash_attn_varlen_func(
q=q.contiguous(),
k=k,
v=v,
cu_seqlens_q=cu_seqlens_q,
cu_seqlens_k=cu_seqlens_k,
seqused_k=seqused_k,
max_seqlen_q=max_seqlen_q,
max_seqlen_k=max_seqlen_k,
softmax_scale=softmax_scale,
causal=causal,
block_table=block_table,
s_aux=s_aux,
window_size=real_window_size,
# alibi_slopes = alibi_slopes,
# softcap=softcap,
return_softmax_lse=return_softmax_lse,
q_descale=q_descale,
k_descale=k_descale,
v_descale=v_descale,
)
@staticmethod
def get_scheduler_metadata(
batch_size,
max_seqlen_q,
max_seqlen_k,
num_heads_q,
num_heads_kv,
headdim,
cache_seqlens: torch.Tensor,
qkv_dtype=torch.bfloat16,
headdim_v=None,
cu_seqlens_q: torch.Tensor | None = None,
cu_seqlens_k_new: torch.Tensor | None = None,
cache_leftpad: torch.Tensor | None = None,
page_size: int | None = None,
max_seqlen_k_new=0,
causal=False,
window_size=(-1, -1), # -1 means infinite context window
has_softcap=False,
num_splits=0, # Can be tuned for speed
pack_gqa=None, # Can be tuned for speed
sm_margin=0, # Can be tuned if some SMs are used for communication
) -> None:
logger.warning_once(
"get_scheduler_metadata is not implemented for xpu_ops, returning None."
)
return None
@staticmethod
def indexer_k_quant_and_cache(
k: torch.Tensor,
kv_cache: torch.Tensor,
slot_mapping: torch.Tensor,
quant_block_size: int,
scale_fmt: str | None,
) -> None:
head_dim = k.shape[-1]
k = k.view(-1, head_dim) # [total_tokens, head_dim]
def group_quant_torch(
x: torch.Tensor,
group_size: int,
eps: float = 1e-10,
dtype: torch.dtype | None = None,
column_major_scales: bool = False,
out_q: torch.Tensor | None = None,
use_ue8m0: bool | None = None,
) -> tuple[torch.Tensor, torch.Tensor]:
if use_ue8m0 is None:
# Default fallback - could import is_deep_gemm_e8m0_used if needed
use_ue8m0 = False
if dtype is None:
dtype = current_platform.fp8_dtype()
# Validate inputs
assert x.shape[-1] % group_size == 0, (
f"Last dimension {x.shape[-1]} must be divisible by "
f"group_size {group_size}"
)
assert x.stride(-1) == 1, "Input tensor groups must be contiguous"
# Prepare output tensor
if out_q is None:
x_q = torch.empty_like(x, dtype=dtype)
else:
assert out_q.shape == x.shape
x_q = out_q
# Reshape input for group processing
# Original shape: (..., last_dim)
# Target shape: (..., num_groups, group_size)
original_shape = x.shape
num_groups = original_shape[-1] // group_size
# Reshape to separate groups
group_shape = original_shape[:-1] + (num_groups, group_size)
x_grouped = x.view(group_shape)
# Compute per-group absolute maximum values
# Shape: (..., num_groups)
abs_max = torch.amax(torch.abs(x_grouped), dim=-1, keepdim=False)
abs_max = torch.maximum(
abs_max, torch.tensor(eps, device=x.device, dtype=x.dtype)
)
# Compute scales
FP8_MAX = torch.finfo(dtype).max
FP8_MIN = torch.finfo(dtype).min
scale_raw = abs_max / FP8_MAX
if use_ue8m0:
# For UE8M0 format, scales must be powers of 2
scales = torch.pow(2.0, torch.ceil(torch.log2(scale_raw)))
else:
scales = scale_raw
# Expand scales for broadcasting with grouped data
# Shape: (..., num_groups, 1)
scales_expanded = scales.unsqueeze(-1)
# Quantize the grouped data
x_scaled = x_grouped / scales_expanded
x_clamped = torch.clamp(x_scaled, FP8_MIN, FP8_MAX)
x_quantized = x_clamped.to(dtype)
# Reshape back to original shape
x_q.copy_(x_quantized.view(original_shape))
# Prepare scales tensor in requested format
if column_major_scales:
# Column-major: (num_groups,) + batch_dims
# Transpose the scales to put group dimension first
scales_shape = (num_groups,) + original_shape[:-1]
x_s = scales.permute(-1, *range(len(original_shape) - 1))
x_s = x_s.contiguous().view(scales_shape)
else:
# Row-major: batch_dims + (num_groups,)
x_s = scales.contiguous()
# Ensure scales are float32
return x_q, x_s.float()
k_fp8, k_scale = group_quant_torch(
k,
group_size=quant_block_size,
column_major_scales=False,
use_ue8m0=(scale_fmt == "ue8m0"),
)
k_fp8_bytes = k_fp8.view(-1, head_dim).view(torch.uint8)
scale_bytes = k_scale.view(torch.uint8).view(-1, 4)
k = torch.cat(
[k_fp8_bytes, scale_bytes], dim=-1
) # [total_tokens, head_dim + 4]
slot_mapping = slot_mapping.flatten()
# kv_cache: [num_block, block_size, head_dim + 4]
kv_cache.view(-1, kv_cache.shape[-1]).index_copy_(0, slot_mapping, k)
@staticmethod
def cp_gather_indexer_k_quant_cache(
kv_cache: torch.Tensor,
dst_k: torch.Tensor,
dst_scale: torch.Tensor,
block_table: torch.Tensor,
cu_seq_lens: torch.Tensor,
) -> None:
"""
Args:
kv_cache: [num_blocks, block_size, cache_stride] - quantized KV cache
Layout per block: [k_values, scale_values]
- k_values: [block_size * head_dim]
- scale_values: [block_size * head_dim * 4 / quant_block_size]
dst_k: [num_tokens, head_dim] - output tensor for K values
dst_scale: [num_tokens, head_dim / quant_block_size * 4]
- output tensor for scale values
block_table: [batch_size, num_blocks] - block table for indexing
cu_seq_lens: [batch_size + 1] - cumulative sequence lengths
"""
batch_size = block_table.size(0)
num_tokens = dst_k.size(0)
head_dim = dst_k.size(1)
cache_block_size = kv_cache.size(1)
quant_block_size = head_dim * 4 // dst_scale.size(1)
# For each token, find which batch it belongs to using searchsorted
token_indices = torch.arange(num_tokens, device=dst_k.device) + 1
# cu_seq_lens is [batch_size + 1], we need to find which interval each
# token belongs to
batch_indices = torch.searchsorted(cu_seq_lens, token_indices) - 1
batch_indices = torch.clamp(batch_indices, 0, batch_size - 1)
# Calculate the in-batch sequence index for each token
inbatch_seq_indices = token_indices - cu_seq_lens[batch_indices]
# Find which block each token belongs to
block_indices_in_table = inbatch_seq_indices // cache_block_size
physical_block_indices = block_table[batch_indices, block_indices_in_table]
# Calculate the offset within each block
inblock_offsets = (inbatch_seq_indices - 1) % cache_block_size
# Calculate strides
block_stride = kv_cache.stride(0) # stride for each block
# Flatten kv_cache for easier indexing
kv_cache_flat = kv_cache.view(-1)
# Calculate source offset for K values for all tokens (vectorized)
src_block_offsets = physical_block_indices * block_stride
src_k_offsets = src_block_offsets + inblock_offsets * head_dim
# Gather K values using advanced indexing
# Create indices for all elements we need to gather
k_indices = src_k_offsets.unsqueeze(1) + torch.arange(
head_dim, device=dst_k.device
)
dst_k[:] = kv_cache_flat[k_indices]
# Calculate source offset for scale values (vectorized)
# Scales are stored after all K values for each block
scale_size = head_dim * 4 // quant_block_size
src_scale_offsets = src_block_offsets + head_dim + inblock_offsets * scale_size
# Gather scale values
scale_indices = src_scale_offsets.unsqueeze(1) + torch.arange(
scale_size, device=dst_scale.device
)
dst_scale[:] = kv_cache_flat[scale_indices]
@staticmethod
def top_k_per_row_prefill(
logits: torch.Tensor,
cu_seqlen_ks: torch.Tensor,
cu_seqlen_ke: torch.Tensor,
raw_topk_indices: torch.Tensor,
num_rows: int,
stride0: int,
strdide1: int,
topk_tokens: int,
) -> torch.Tensor:
real_topk = min(topk_tokens, logits.shape[-1])
topk_indices = logits.topk(real_topk, dim=-1)[1].to(torch.int32)
topk_indices -= cu_seqlen_ks[:, None]
mask_lo = topk_indices >= 0
mask_hi = topk_indices - (cu_seqlen_ke - cu_seqlen_ks)[:, None] < 0
mask = torch.full_like(
topk_indices, False, dtype=torch.bool, device=topk_indices.device
)
mask = mask_lo & mask_hi
topk_indices.masked_fill_(~mask, -1)
raw_topk_indices[: topk_indices.shape[0], : topk_indices.shape[1]] = (
topk_indices
)
@staticmethod
def top_k_per_row_decode(
logits: torch.Tensor,
next_n: int,
seq_lens: torch.Tensor,
raw_topk_indices: torch.Tensor,
num_rows: int,
stride0: int,
stride1: int,
topk_tokens: int,
) -> torch.Tensor:
device = logits.device
batch_size = seq_lens.size(0)
# padded query len
padded_num_tokens = batch_size * next_n
positions = (
torch.arange(logits.shape[-1], device=device)
.unsqueeze(0)
.expand(batch_size * next_n, -1)
)
row_indices = torch.arange(padded_num_tokens, device=device) // next_n
next_n_offset = torch.arange(padded_num_tokens, device=device) % next_n
index_end_pos = (seq_lens[row_indices] - next_n + next_n_offset).unsqueeze(1)
# index_end_pos: [B * N, 1]
mask = positions <= index_end_pos
# mask: [B * N, L]
logits = logits.masked_fill(~mask, float("-inf"))
real_topk = min(topk_tokens, logits.shape[-1])
topk_indices = logits.topk(real_topk, dim=-1)[1].to(torch.int32) # [B * N, K]
# ensure we don't set indices for the top k
# that is out of range(masked already)
# this will happen if context length is shorter than K
topk_indices[topk_indices > index_end_pos] = -1
raw_topk_indices[: topk_indices.shape[0], : topk_indices.shape[1]] = (
topk_indices
)
@staticmethod
def register_ops_once() -> None:
global _OPS_REGISTERED
if not _OPS_REGISTERED:
# register all the custom ops here
direct_register_custom_op(
op_name="xpu_ops_deepseek_scaling_rope",
op_func=_xpu_ops_deepseek_scaling_rope_impl,
mutates_args=[],
fake_impl=_xpu_ops_deepseek_scaling_rope_fake,
dispatch_key=current_platform.dispatch_key,
)
direct_register_custom_op(
op_name="xpu_mxfp8_quantize",
op_func=_xpu_mxfp8_quantize_impl,
fake_impl=_xpu_mxfp8_quantize_fake,
)
direct_register_custom_op(
op_name="xpu_mxfp4_quantize",
op_func=_xpu_mxfp4_quantize_impl,
fake_impl=_xpu_mxfp4_quantize_fake,
)
_OPS_REGISTERED = True
xpu_ops.register_ops_once()