Coverage for src/flag_gems/ops/__init__.py: 100%

1from flag_gems.ops._euclidean_dist import _euclidean_dist 

2from flag_gems.ops._functional_sym_constrain_range_for_size import ( 

3 _functional_sym_constrain_range_for_size, 

4) 

5from flag_gems.ops._is_all_true import _is_all_true 

6from flag_gems.ops._safe_softmax import _safe_softmax 

7from flag_gems.ops._upsample_nearest_exact1d import _upsample_nearest_exact1d 

8from flag_gems.ops.abs import abs, abs_ 

9from flag_gems.ops.absolute import absolute 

10from flag_gems.ops.acos import acos 

11from flag_gems.ops.act_quant import act_quant_triton 

12from flag_gems.ops.add import add, add_ 

13from flag_gems.ops.addcdiv import addcdiv, addcdiv_out 

14from flag_gems.ops.addcmul import addcmul, addcmul_out 

15from flag_gems.ops.addmm import addmm, addmm_dtype, addmm_dtype_out, addmm_out 

16from flag_gems.ops.addmv import addmv, addmv_out 

17from flag_gems.ops.addr import addr 

18from flag_gems.ops.affine_grid_generator import affine_grid_generator 

19from flag_gems.ops.alias_copy import alias_copy, alias_copy_out 

20from flag_gems.ops.all import all, all_dim, all_dims 

21from flag_gems.ops.amax import amax 

22from flag_gems.ops.aminmax import aminmax 

23from flag_gems.ops.angle import angle 

24from flag_gems.ops.any import any, any_dim, any_dims 

25from flag_gems.ops.arange import arange, arange_start 

26from flag_gems.ops.arcsinh import arcsinh, arcsinh_out 

27from flag_gems.ops.arcsinh_ import arcsinh_ 

28from flag_gems.ops.arctanh_ import arctanh_ 

29from flag_gems.ops.argmax import argmax 

30from flag_gems.ops.argmin import argmin 

31from flag_gems.ops.asinh import asinh, asinh_out 

32from flag_gems.ops.asinh_ import asinh_ 

33from flag_gems.ops.assert_async import _assert_async 

34from flag_gems.ops.atan import atan, atan_ 

35from flag_gems.ops.atan2 import atan2, atan2_out 

36from flag_gems.ops.attention import ( 

37 ScaleDotProductAttention, 

38 flash_attention_forward, 

39 flash_attn_varlen_func, 

40 flash_attn_varlen_opt_func, 

41 scaled_dot_product_attention, 

42 scaled_dot_product_attention_backward, 

43 scaled_dot_product_attention_forward, 

44) 

45from flag_gems.ops.avg_pool2d import avg_pool2d, avg_pool2d_backward 

46from flag_gems.ops.avg_pool3d import avg_pool3d, avg_pool3d_backward 

47from flag_gems.ops.baddbmm import baddbmm, baddbmm_out 

48from flag_gems.ops.batch_norm import batch_norm, batch_norm_backward 

49from flag_gems.ops.bernoulli_ import bernoulli_ 

50from flag_gems.ops.bincount import bincount 

51from flag_gems.ops.bitwise_and import ( 

52 bitwise_and_scalar, 

53 bitwise_and_scalar_, 

54 bitwise_and_scalar_tensor, 

55 bitwise_and_tensor, 

56 bitwise_and_tensor_, 

57) 

58from flag_gems.ops.bitwise_left_shift import bitwise_left_shift 

59from flag_gems.ops.bitwise_not import bitwise_not, bitwise_not_ 

60from flag_gems.ops.bitwise_or import ( 

61 bitwise_or_scalar, 

62 bitwise_or_scalar_, 

63 bitwise_or_scalar_tensor, 

64 bitwise_or_tensor, 

65 bitwise_or_tensor_, 

66) 

67from flag_gems.ops.bitwise_right_shift import bitwise_right_shift 

68from flag_gems.ops.bmm import bmm, bmm_out 

69from flag_gems.ops.cat import cat, cat_out 

70from flag_gems.ops.cauchy import cauchy, cauchy_ 

71from flag_gems.ops.ceil import ceil, ceil_, ceil_out 

72from flag_gems.ops.celu import celu, celu_ 

73from flag_gems.ops.clamp import ( 

74 clamp, 

75 clamp_, 

76 clamp_min, 

77 clamp_min_, 

78 clamp_tensor, 

79 clamp_tensor_, 

80) 

81from flag_gems.ops.clip import clip, clip_ 

82from flag_gems.ops.col2im import col2im 

83from flag_gems.ops.concatenate import concatenate 

84from flag_gems.ops.conj_physical import conj_physical 

85from flag_gems.ops.contiguous import contiguous 

86from flag_gems.ops.conv1d import conv1d 

87from flag_gems.ops.conv2d import conv2d 

88from flag_gems.ops.conv3d import conv3d 

89from flag_gems.ops.conv_depthwise2d import _conv_depthwise2d 

90from flag_gems.ops.conv_transpose1d import conv_transpose1d 

91from flag_gems.ops.conv_transpose2d import conv_transpose2d 

92from flag_gems.ops.copy import copy, copy_ 

93from flag_gems.ops.copysign import copysign, copysign_out 

94from flag_gems.ops.cos import cos, cos_ 

95from flag_gems.ops.cosh import cosh, cosh_, cosh_out 

96from flag_gems.ops.count_nonzero import count_nonzero 

97from flag_gems.ops.ctc_loss import ctc_loss 

98from flag_gems.ops.cudnn_convolution import cudnn_convolution 

99from flag_gems.ops.cummax import cummax 

100from flag_gems.ops.cummin import cummin 

101from flag_gems.ops.cumprod import cumprod, cumprod_ 

102from flag_gems.ops.cumsum import cumsum, cumsum_out, normed_cumsum 

103from flag_gems.ops.diag import diag 

104from flag_gems.ops.diag_embed import diag_embed 

105from flag_gems.ops.diagonal import diagonal_backward 

106from flag_gems.ops.diff import diff 

107from flag_gems.ops.digamma_ import digamma_ 

108from flag_gems.ops.div import ( 

109 div_mode, 

110 div_mode_, 

111 floor_divide, 

112 floor_divide_, 

113 remainder, 

114 remainder_, 

115 true_divide, 

116 true_divide_, 

117 true_divide_out, 

118) 

119from flag_gems.ops.dot import dot 

120from flag_gems.ops.dropout import dropout, dropout_backward 

121from flag_gems.ops.elu import elu, elu_, elu_backward 

122from flag_gems.ops.embedding import embedding, embedding_backward 

123from flag_gems.ops.embedding_dense_backward import embedding_dense_backward 

124from flag_gems.ops.eq import eq, eq_scalar, equal 

125from flag_gems.ops.erf import erf, erf_ 

126from flag_gems.ops.exp import exp, exp_, exp_out 

127from flag_gems.ops.exp2 import exp2, exp2_ 

128from flag_gems.ops.expm1 import expm1, expm1_, expm1_out 

129from flag_gems.ops.exponential_ import exponential_ 

130from flag_gems.ops.eye import eye 

131from flag_gems.ops.eye_m import eye_m 

132from flag_gems.ops.feature_dropout import feature_dropout, feature_dropout_ 

133from flag_gems.ops.fft import fft 

134from flag_gems.ops.fill import ( 

135 fill_scalar, 

136 fill_scalar_, 

137 fill_scalar_out, 

138 fill_tensor, 

139 fill_tensor_, 

140 fill_tensor_out, 

141) 

142from flag_gems.ops.flip import flip 

143from flag_gems.ops.floor_ import floor_ 

144from flag_gems.ops.fmin import fmin, fmin_out 

145from flag_gems.ops.fmod import fmod_scalar, fmod_scalar_, fmod_tensor, fmod_tensor_ 

146from flag_gems.ops.fp8_matmul import fp8_matmul 

147from flag_gems.ops.fp8_mqa_logits import fp8_mqa_logits 

148from flag_gems.ops.full import full 

149from flag_gems.ops.full_like import full_like 

150from flag_gems.ops.gather import gather, gather_backward 

151from flag_gems.ops.gcd import gcd, gcd_out 

152from flag_gems.ops.ge import ge, ge_scalar 

153from flag_gems.ops.gelu import gelu, gelu_, gelu_backward 

154from flag_gems.ops.get_paged_mqa_logits_metadata import get_paged_mqa_logits_metadata 

155from flag_gems.ops.get_scheduler_metadata import get_scheduler_metadata 

156from flag_gems.ops.glu import glu, glu_backward 

157from flag_gems.ops.greater import ( 

158 greater, 

159 greater_out, 

160 greater_scalar, 

161 greater_scalar_out, 

162) 

163from flag_gems.ops.grid_sample import grid_sample 

164from flag_gems.ops.group_gemm import group_mm 

165from flag_gems.ops.groupnorm import group_norm, group_norm_backward 

166from flag_gems.ops.gt import gt, gt_scalar 

167from flag_gems.ops.hadamard_transform import ( 

168 hadamard_transform, 

169 hadamard_transform_12N, 

170 hadamard_transform_20N, 

171 hadamard_transform_28N, 

172 hadamard_transform_40N, 

173) 

174from flag_gems.ops.hardsigmoid import hardsigmoid, hardsigmoid_out 

175from flag_gems.ops.hardswish_ import hardswish_ 

176from flag_gems.ops.histc import histc 

177from flag_gems.ops.hstack import hstack 

178from flag_gems.ops.hypot import hypot, hypot_out 

179from flag_gems.ops.i0 import i0, i0_out 

180from flag_gems.ops.i0_ import i0_ 

181from flag_gems.ops.index import index 

182from flag_gems.ops.index_add import index_add, index_add_ 

183from flag_gems.ops.index_copy_ import index_copy, index_copy_ 

184from flag_gems.ops.index_put import _index_put_impl_, index_put, index_put_ 

185from flag_gems.ops.index_select import index_select 

186from flag_gems.ops.isclose import allclose, isclose 

187from flag_gems.ops.isfinite import isfinite 

188from flag_gems.ops.isin import isin 

189from flag_gems.ops.isinf import isinf 

190from flag_gems.ops.isnan import isnan 

191from flag_gems.ops.isneginf import isneginf, isneginf_out 

192from flag_gems.ops.kron import kron 

193from flag_gems.ops.layernorm import layer_norm, layer_norm_backward 

194from flag_gems.ops.le import le, le_scalar 

195from flag_gems.ops.leaky_relu import leaky_relu, leaky_relu_, leaky_relu_out 

196from flag_gems.ops.lerp import lerp_scalar, lerp_scalar_, lerp_tensor, lerp_tensor_ 

197from flag_gems.ops.lift_fresh_copy import lift_fresh_copy, lift_fresh_copy_out 

198from flag_gems.ops.linspace import linspace 

199from flag_gems.ops.log import log 

200from flag_gems.ops.log1p_ import log1p_ 

201from flag_gems.ops.log10 import log10, log10_, log10_out 

202from flag_gems.ops.log_sigmoid import log_sigmoid 

203from flag_gems.ops.log_softmax import ( 

204 log_softmax, 

205 log_softmax_backward, 

206 log_softmax_backward_out, 

207 log_softmax_out, 

208) 

209from flag_gems.ops.logaddexp import logaddexp, logaddexp_out 

210from flag_gems.ops.logical_and import logical_and, logical_and_ 

211from flag_gems.ops.logical_not import logical_not 

212from flag_gems.ops.logical_or import logical_or, logical_or_ 

213from flag_gems.ops.logical_xor import logical_xor 

214from flag_gems.ops.logit import logit, logit_out 

215from flag_gems.ops.logit_ import logit_ 

216from flag_gems.ops.logspace import logspace 

217from flag_gems.ops.logsumexp import logsumexp 

218from flag_gems.ops.lt import lt, lt_scalar 

219from flag_gems.ops.margin_ranking_loss import margin_ranking_loss 

220from flag_gems.ops.masked_fill import masked_fill, masked_fill_ 

221from flag_gems.ops.masked_scatter import masked_scatter, masked_scatter_ 

222from flag_gems.ops.masked_select import masked_select 

223from flag_gems.ops.max import max, max_dim 

224from flag_gems.ops.max_pool2d_with_indices import ( 

225 max_pool2d_backward, 

226 max_pool2d_with_indices, 

227) 

228from flag_gems.ops.max_pool3d_with_indices import ( 

229 max_pool3d_backward, 

230 max_pool3d_with_indices, 

231) 

232from flag_gems.ops.maximum import maximum 

233from flag_gems.ops.mean import mean, mean_dim 

234from flag_gems.ops.median import median, median_dim, median_dim_values, median_out 

235from flag_gems.ops.min import min, min_dim 

236from flag_gems.ops.minimum import minimum 

237from flag_gems.ops.mm import mm, mm_out, router_gemm 

238from flag_gems.ops.mse_loss import mse_loss 

239from flag_gems.ops.mul import mul, mul_ 

240from flag_gems.ops.multinomial import multinomial 

241from flag_gems.ops.mv import mv 

242from flag_gems.ops.nan_to_num import nan_to_num 

243from flag_gems.ops.ne import ne, ne_scalar 

244from flag_gems.ops.neg import neg, neg_ 

245from flag_gems.ops.new_full import new_full 

246from flag_gems.ops.nll_loss_nd import nll_loss_nd_backward, nll_loss_nd_forward 

247from flag_gems.ops.nllloss import ( 

248 nll_loss2d_backward, 

249 nll_loss2d_forward, 

250 nll_loss_backward, 

251 nll_loss_forward, 

252) 

253from flag_gems.ops.nonzero import nonzero 

254from flag_gems.ops.nonzero_numpy import nonzero_numpy 

255from flag_gems.ops.normal import ( 

256 normal_, 

257 normal_float_tensor, 

258 normal_tensor_float, 

259 normal_tensor_tensor, 

260) 

261from flag_gems.ops.one_hot import one_hot 

262from flag_gems.ops.ones import ones 

263from flag_gems.ops.ones_like import ones_like 

264from flag_gems.ops.pad import constant_pad_nd, pad 

265from flag_gems.ops.per_token_group_quant_fp8 import ( 

266 SUPPORTED_FP8_DTYPE, 

267 per_token_group_quant_fp8, 

268) 

269from flag_gems.ops.pixel_shuffle import pixel_shuffle 

270from flag_gems.ops.pixel_unshuffle import pixel_unshuffle, pixel_unshuffle_out 

271from flag_gems.ops.poisson import poisson 

272from flag_gems.ops.polar import polar 

273from flag_gems.ops.pow import ( 

274 pow_scalar, 

275 pow_tensor_scalar, 

276 pow_tensor_scalar_, 

277 pow_tensor_tensor, 

278 pow_tensor_tensor_, 

279) 

280from flag_gems.ops.prelu import prelu 

281from flag_gems.ops.prod import prod, prod_dim 

282from flag_gems.ops.quantile import quantile 

283from flag_gems.ops.rad2deg import rad2deg, rad2deg_ 

284from flag_gems.ops.rand import rand 

285from flag_gems.ops.rand_like import rand_like 

286from flag_gems.ops.randint import randint 

287from flag_gems.ops.randn import randn 

288from flag_gems.ops.randn_like import randn_like 

289from flag_gems.ops.randperm import randperm 

290from flag_gems.ops.reciprocal import reciprocal, reciprocal_ 

291from flag_gems.ops.reflection_pad1d import reflection_pad1d, reflection_pad1d_out 

292from flag_gems.ops.reflection_pad2d import reflection_pad2d, reflection_pad2d_out 

293from flag_gems.ops.relu import relu, relu_ 

294from flag_gems.ops.relu6 import relu6 

295from flag_gems.ops.repeat import repeat 

296from flag_gems.ops.repeat_interleave import ( 

297 repeat_interleave_self_int, 

298 repeat_interleave_self_tensor, 

299 repeat_interleave_tensor, 

300) 

301from flag_gems.ops.replication_pad1d import replication_pad1d, replication_pad1d_out 

302from flag_gems.ops.replication_pad3d import replication_pad3d 

303from flag_gems.ops.resolve_conj import resolve_conj 

304from flag_gems.ops.resolve_neg import resolve_neg 

305from flag_gems.ops.rms_norm import rms_norm, rms_norm_backward, rms_norm_forward 

306from flag_gems.ops.roll import roll 

307from flag_gems.ops.round import round, round_, round_out 

308from flag_gems.ops.rrelu_with_noise_backward import rrelu_with_noise_backward 

309from flag_gems.ops.rsqrt import rsqrt, rsqrt_ 

310from flag_gems.ops.rsub import rsub_scalar, rsub_tensor 

311from flag_gems.ops.scaled_softmax import scaled_softmax_backward, scaled_softmax_forward 

312from flag_gems.ops.scatter import scatter, scatter_ 

313from flag_gems.ops.scatter_add_ import scatter_add_ 

314from flag_gems.ops.scatter_reduce import ( 

315 scatter_reduce, 

316 scatter_reduce_, 

317 scatter_reduce_out, 

318) 

319from flag_gems.ops.select_backward import select_backward 

320from flag_gems.ops.select_scatter import select_scatter 

321from flag_gems.ops.selu import selu 

322from flag_gems.ops.selu_ import selu_ 

323from flag_gems.ops.sgn_ import sgn_ 

324from flag_gems.ops.sigmoid import sigmoid, sigmoid_, sigmoid_backward 

325from flag_gems.ops.signbit import signbit, signbit_out 

326from flag_gems.ops.silu import silu, silu_, silu_backward 

327from flag_gems.ops.sin import sin, sin_ 

328from flag_gems.ops.sinh_ import sinh_ 

329from flag_gems.ops.slice_backward import slice_backward 

330from flag_gems.ops.slice_scatter import slice_scatter 

331from flag_gems.ops.smooth_l1_loss import ( 

332 smooth_l1_loss, 

333 smooth_l1_loss_backward, 

334 smooth_l1_loss_out, 

335) 

336from flag_gems.ops.soft_margin_loss import soft_margin_loss, soft_margin_loss_out 

337from flag_gems.ops.softmax import ( 

338 softmax, 

339 softmax_backward, 

340 softmax_backward_out, 

341 softmax_out, 

342) 

343from flag_gems.ops.softplus import softplus 

344from flag_gems.ops.softshrink import softshrink, softshrink_out 

345from flag_gems.ops.sort import sort, sort_stable 

346from flag_gems.ops.special_i0e import special_i0e, special_i0e_out 

347from flag_gems.ops.special_i1 import special_i1, special_i1_out 

348from flag_gems.ops.sqrt import sqrt, sqrt_ 

349from flag_gems.ops.square import square, square_, square_out 

350from flag_gems.ops.stack import stack 

351from flag_gems.ops.std import std 

352from flag_gems.ops.sub import sub, sub_ 

353from flag_gems.ops.sum import sum, sum_dim, sum_dim_out, sum_out 

354from flag_gems.ops.svd import svd 

355from flag_gems.ops.t_copy import t_copy, t_copy_out 

356from flag_gems.ops.tan import tan, tan_ 

357from flag_gems.ops.tanh import tanh, tanh_, tanh_backward 

358from flag_gems.ops.threshold import threshold, threshold_backward 

359from flag_gems.ops.tile import tile 

360from flag_gems.ops.to import to_copy 

361from flag_gems.ops.topk import topk 

362from flag_gems.ops.trace import trace 

363from flag_gems.ops.tril import tril, tril_, tril_out 

364from flag_gems.ops.triu import triu, triu_ 

365from flag_gems.ops.unfold_backward import unfold_backward 

366from flag_gems.ops.uniform import uniform_ 

367from flag_gems.ops.unique import _unique2 

368from flag_gems.ops.unique_consecutive import unique_consecutive 

369from flag_gems.ops.upsample_bicubic2d import upsample_bicubic2d 

370from flag_gems.ops.upsample_bicubic2d_aa import _upsample_bicubic2d_aa 

371from flag_gems.ops.upsample_bicubic2d_aa_backward import _upsample_bicubic2d_aa_backward 

372from flag_gems.ops.upsample_linear1d import upsample_linear1d 

373from flag_gems.ops.upsample_nearest1d import upsample_nearest1d 

374from flag_gems.ops.upsample_nearest2d import upsample_nearest2d 

375from flag_gems.ops.upsample_nearest3d import upsample_nearest3d 

376from flag_gems.ops.var import var, var_correction, var_dim 

377from flag_gems.ops.var_mean import var_mean 

378from flag_gems.ops.vdot import vdot 

379from flag_gems.ops.vector_norm import vector_norm 

380from flag_gems.ops.vstack import vstack 

381from flag_gems.ops.w8a8_block_fp8_matmul import w8a8_block_fp8_matmul 

382from flag_gems.ops.weightnorm import ( 

383 weight_norm_interface, 

384 weight_norm_interface_backward, 

385) 

386from flag_gems.ops.where import ( 

387 where_scalar_other, 

388 where_scalar_self, 

389 where_self, 

390 where_self_out, 

391) 

392from flag_gems.ops.zero import zero, zero_out 

393from flag_gems.ops.zeros import zero_, zeros 

394from flag_gems.ops.zeros_like import zeros_like 

395 

396__all__ = [ 

397 "_assert_async", 

398 "_conv_depthwise2d", 

399 "_euclidean_dist", 

400 "_functional_sym_constrain_range_for_size", 

401 "_index_put_impl_", 

402 "_is_all_true", 

403 "_safe_softmax", 

404 "_unique2", 

405 "_upsample_bicubic2d_aa", 

406 "_upsample_bicubic2d_aa_backward", 

407 "_upsample_nearest_exact1d", 

408 "abs", 

409 "abs_", 

410 "absolute", 

411 "acos", 

412 "act_quant_triton", 

413 "add", 

414 "add_", 

415 "addcdiv", 

416 "addcdiv_out", 

417 "addcmul", 

418 "addcmul_out", 

419 "addmm", 

420 "addmm_dtype", 

421 "addmm_dtype_out", 

422 "addmm_out", 

423 "addmv", 

424 "addmv_out", 

425 "addr", 

426 "affine_grid_generator", 

427 "alias_copy", 

428 "alias_copy_out", 

429 "all", 

430 "all_dim", 

431 "all_dims", 

432 "allclose", 

433 "amax", 

434 "aminmax", 

435 "angle", 

436 "any", 

437 "any_dim", 

438 "any_dims", 

439 "arange", 

440 "arange_start", 

441 "arcsinh", 

442 "arcsinh_", 

443 "arcsinh_out", 

444 "arctanh_", 

445 "argmax", 

446 "argmin", 

447 "asinh", 

448 "asinh_", 

449 "asinh_out", 

450 "atan", 

451 "atan2", 

452 "atan2_out", 

453 "atan_", 

454 "avg_pool2d", 

455 "avg_pool2d_backward", 

456 "avg_pool3d", 

457 "avg_pool3d_backward", 

458 "baddbmm", 

459 "baddbmm_out", 

460 "batch_norm", 

461 "batch_norm_backward", 

462 "bernoulli_", 

463 "bincount", 

464 "bitwise_and_scalar", 

465 "bitwise_and_scalar_", 

466 "bitwise_and_scalar_tensor", 

467 "bitwise_and_tensor", 

468 "bitwise_and_tensor_", 

469 "bitwise_left_shift", 

470 "bitwise_not", 

471 "bitwise_not_", 

472 "bitwise_or_scalar", 

473 "bitwise_or_scalar_", 

474 "bitwise_or_scalar_tensor", 

475 "bitwise_or_tensor", 

476 "bitwise_or_tensor_", 

477 "bitwise_right_shift", 

478 "bmm", 

479 "bmm_out", 

480 "cat", 

481 "cat_out", 

482 "cauchy", 

483 "cauchy_", 

484 "ceil", 

485 "ceil_", 

486 "ceil_out", 

487 "celu", 

488 "celu_", 

489 "clamp", 

490 "clamp_", 

491 "clamp_min", 

492 "clamp_min_", 

493 "clamp_tensor", 

494 "clamp_tensor_", 

495 "clip", 

496 "clip_", 

497 "col2im", 

498 "concatenate", 

499 "conj_physical", 

500 "constant_pad_nd", 

501 "contiguous", 

502 "conv1d", 

503 "conv2d", 

504 "conv3d", 

505 "conv_transpose1d", 

506 "conv_transpose2d", 

507 "copy", 

508 "copy_", 

509 "copysign", 

510 "copysign_out", 

511 "cos", 

512 "cos_", 

513 "cosh", 

514 "cosh_", 

515 "cosh_out", 

516 "count_nonzero", 

517 "ctc_loss", 

518 "cudnn_convolution", 

519 "cummax", 

520 "cummin", 

521 "cumprod", 

522 "cumprod_", 

523 "cumsum", 

524 "cumsum_out", 

525 "diag", 

526 "diag_embed", 

527 "diagonal_backward", 

528 "diff", 

529 "digamma_", 

530 "div_mode", 

531 "div_mode_", 

532 "dot", 

533 "dropout", 

534 "dropout_backward", 

535 "elu", 

536 "elu_", 

537 "elu_backward", 

538 "embedding", 

539 "embedding_backward", 

540 "embedding_dense_backward", 

541 "eq", 

542 "eq_scalar", 

543 "equal", 

544 "erf", 

545 "erf_", 

546 "exp", 

547 "exp2", 

548 "exp2_", 

549 "exp_", 

550 "exp_out", 

551 "expm1", 

552 "expm1_", 

553 "expm1_out", 

554 "exponential_", 

555 "eye", 

556 "eye_m", 

557 "feature_dropout", 

558 "feature_dropout_", 

559 "fft", 

560 "fill_scalar", 

561 "fill_scalar_", 

562 "fill_scalar_out", 

563 "fill_tensor", 

564 "fill_tensor_", 

565 "fill_tensor_out", 

566 "flash_attention_forward", 

567 "flash_attn_varlen_func", 

568 "flash_attn_varlen_opt_func", 

569 "flip", 

570 "floor_", 

571 "floor_divide", 

572 "floor_divide_", 

573 "fmin", 

574 "fmin_out", 

575 "fmod_scalar", 

576 "fmod_scalar_", 

577 "fmod_tensor", 

578 "fmod_tensor_", 

579 "fp8_matmul", 

580 "fp8_mqa_logits", 

581 "full", 

582 "full_like", 

583 "gather", 

584 "gather_backward", 

585 "gcd", 

586 "gcd_out", 

587 "ge", 

588 "ge_scalar", 

589 "gelu", 

590 "gelu_", 

591 "gelu_backward", 

592 "get_paged_mqa_logits_metadata", 

593 "get_scheduler_metadata", 

594 "glu", 

595 "glu_backward", 

596 "greater", 

597 "greater_out", 

598 "greater_scalar", 

599 "greater_scalar_out", 

600 "grid_sample", 

601 "group_mm", 

602 "group_norm", 

603 "group_norm_backward", 

604 "gt", 

605 "gt_scalar", 

606 "hadamard_transform", 

607 "hadamard_transform_12N", 

608 "hadamard_transform_20N", 

609 "hadamard_transform_28N", 

610 "hadamard_transform_40N", 

611 "hardsigmoid", 

612 "hardsigmoid_out", 

613 "hardswish_", 

614 "histc", 

615 "hstack", 

616 "hypot", 

617 "hypot_out", 

618 "i0", 

619 "i0_", 

620 "i0_out", 

621 "index", 

622 "index_add", 

623 "index_add_", 

624 "index_copy", 

625 "index_copy_", 

626 "index_put", 

627 "index_put_", 

628 "index_select", 

629 "isclose", 

630 "isfinite", 

631 "isin", 

632 "isinf", 

633 "isnan", 

634 "isneginf", 

635 "isneginf_out", 

636 "kron", 

637 "layer_norm", 

638 "layer_norm_backward", 

639 "le", 

640 "le_scalar", 

641 "leaky_relu", 

642 "leaky_relu_", 

643 "leaky_relu_out", 

644 "lerp_scalar", 

645 "lerp_scalar_", 

646 "lerp_tensor", 

647 "lerp_tensor_", 

648 "lift_fresh_copy", 

649 "lift_fresh_copy_out", 

650 "linspace", 

651 "log", 

652 "log10", 

653 "log10_", 

654 "log10_out", 

655 "log1p_", 

656 "log_sigmoid", 

657 "log_softmax", 

658 "log_softmax_backward", 

659 "log_softmax_backward_out", 

660 "log_softmax_out", 

661 "logaddexp", 

662 "logaddexp_out", 

663 "logical_and", 

664 "logical_and_", 

665 "logical_not", 

666 "logical_or", 

667 "logical_or_", 

668 "logical_xor", 

669 "logit", 

670 "logit_", 

671 "logit_out", 

672 "logspace", 

673 "logsumexp", 

674 "lt", 

675 "lt_scalar", 

676 "margin_ranking_loss", 

677 "masked_fill", 

678 "masked_fill_", 

679 "masked_scatter", 

680 "masked_scatter_", 

681 "masked_select", 

682 "max", 

683 "max_dim", 

684 "max_pool2d_backward", 

685 "max_pool2d_with_indices", 

686 "max_pool3d_backward", 

687 "max_pool3d_with_indices", 

688 "maximum", 

689 "mean", 

690 "mean_dim", 

691 "median", 

692 "median_dim", 

693 "median_dim_values", 

694 "median_out", 

695 "min", 

696 "min_dim", 

697 "minimum", 

698 "mm", 

699 "mm_out", 

700 "mse_loss", 

701 "mul", 

702 "mul_", 

703 "multinomial", 

704 "mv", 

705 "nan_to_num", 

706 "ne", 

707 "ne_scalar", 

708 "neg", 

709 "neg_", 

710 "new_full", 

711 "nll_loss2d_backward", 

712 "nll_loss2d_forward", 

713 "nll_loss_backward", 

714 "nll_loss_forward", 

715 "nll_loss_nd_backward", 

716 "nll_loss_nd_forward", 

717 "nonzero", 

718 "nonzero_numpy", 

719 "normal_", 

720 "normal_float_tensor", 

721 "normal_tensor_float", 

722 "normal_tensor_tensor", 

723 "normed_cumsum", 

724 "one_hot", 

725 "ones", 

726 "ones_like", 

727 "pad", 

728 "per_token_group_quant_fp8", 

729 "pixel_shuffle", 

730 "pixel_unshuffle", 

731 "pixel_unshuffle_out", 

732 "poisson", 

733 "polar", 

734 "pow_scalar", 

735 "pow_tensor_scalar", 

736 "pow_tensor_scalar_", 

737 "pow_tensor_tensor", 

738 "pow_tensor_tensor_", 

739 "prelu", 

740 "prod", 

741 "prod_dim", 

742 "quantile", 

743 "rad2deg", 

744 "rad2deg_", 

745 "rand", 

746 "rand_like", 

747 "randint", 

748 "randn", 

749 "randn_like", 

750 "randperm", 

751 "reciprocal", 

752 "reciprocal_", 

753 "reflection_pad1d", 

754 "reflection_pad1d_out", 

755 "reflection_pad2d", 

756 "reflection_pad2d_out", 

757 "relu", 

758 "relu6", 

759 "relu_", 

760 "remainder", 

761 "remainder_", 

762 "repeat", 

763 "repeat_interleave_self_int", 

764 "repeat_interleave_self_tensor", 

765 "repeat_interleave_tensor", 

766 "replication_pad1d", 

767 "replication_pad1d_out", 

768 "replication_pad3d", 

769 "resolve_conj", 

770 "resolve_neg", 

771 "rms_norm", 

772 "rms_norm_backward", 

773 "rms_norm_forward", 

774 "roll", 

775 "round", 

776 "round_", 

777 "round_out", 

778 "router_gemm", 

779 "rrelu_with_noise_backward", 

780 "rsqrt", 

781 "rsqrt_", 

782 "rsub_scalar", 

783 "rsub_tensor", 

784 "scaled_dot_product_attention", 

785 "scaled_dot_product_attention_backward", 

786 "scaled_dot_product_attention_forward", 

787 "scaled_softmax_backward", 

788 "scaled_softmax_forward", 

789 "ScaleDotProductAttention", 

790 "scatter", 

791 "scatter_", 

792 "scatter_add_", 

793 "scatter_reduce", 

794 "scatter_reduce_", 

795 "scatter_reduce_out", 

796 "select_backward", 

797 "select_scatter", 

798 "selu", 

799 "selu_", 

800 "sgn_", 

801 "sigmoid", 

802 "sigmoid_", 

803 "sigmoid_backward", 

804 "signbit", 

805 "signbit_out", 

806 "silu", 

807 "silu_", 

808 "silu_backward", 

809 "sin", 

810 "sin_", 

811 "sinh_", 

812 "slice_backward", 

813 "slice_scatter", 

814 "smooth_l1_loss", 

815 "smooth_l1_loss_backward", 

816 "smooth_l1_loss_out", 

817 "soft_margin_loss", 

818 "soft_margin_loss_out", 

819 "softmax", 

820 "softmax_backward", 

821 "softmax_backward_out", 

822 "softmax_out", 

823 "softplus", 

824 "softshrink", 

825 "softshrink_out", 

826 "sort", 

827 "sort_stable", 

828 "special_i0e", 

829 "special_i0e_out", 

830 "special_i1", 

831 "special_i1_out", 

832 "sqrt", 

833 "sqrt_", 

834 "square", 

835 "square_", 

836 "square_out", 

837 "stack", 

838 "std", 

839 "sub", 

840 "sub_", 

841 "sum", 

842 "sum_dim", 

843 "sum_dim_out", 

844 "sum_out", 

845 "SUPPORTED_FP8_DTYPE", 

846 "svd", 

847 "t_copy", 

848 "t_copy_out", 

849 "tan", 

850 "tan_", 

851 "tanh", 

852 "tanh_", 

853 "tanh_backward", 

854 "threshold", 

855 "threshold_backward", 

856 "tile", 

857 "to_copy", 

858 "topk", 

859 "trace", 

860 "tril", 

861 "tril_", 

862 "tril_out",