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.ceil import ceil, ceil_, ceil_out 

71from flag_gems.ops.celu import celu, celu_ 

72from flag_gems.ops.clamp import ( 

73 clamp, 

74 clamp_, 

75 clamp_min, 

76 clamp_min_, 

77 clamp_tensor, 

78 clamp_tensor_, 

79) 

80from flag_gems.ops.clip import clip, clip_ 

81from flag_gems.ops.col2im import col2im 

82from flag_gems.ops.concatenate import concatenate 

83from flag_gems.ops.conj_physical import conj_physical 

84from flag_gems.ops.contiguous import contiguous 

85from flag_gems.ops.conv1d import conv1d 

86from flag_gems.ops.conv2d import conv2d 

87from flag_gems.ops.conv3d import conv3d 

88from flag_gems.ops.conv_depthwise2d import _conv_depthwise2d 

89from flag_gems.ops.conv_transpose1d import conv_transpose1d 

90from flag_gems.ops.conv_transpose2d import conv_transpose2d 

91from flag_gems.ops.copy import copy, copy_ 

92from flag_gems.ops.copysign import copysign, copysign_out 

93from flag_gems.ops.cos import cos, cos_ 

94from flag_gems.ops.cosh import cosh, cosh_, cosh_out 

95from flag_gems.ops.count_nonzero import count_nonzero 

96from flag_gems.ops.ctc_loss import ctc_loss 

97from flag_gems.ops.cudnn_convolution import cudnn_convolution 

98from flag_gems.ops.cummax import cummax 

99from flag_gems.ops.cummin import cummin 

100from flag_gems.ops.cumprod import cumprod, cumprod_ 

101from flag_gems.ops.cumsum import cumsum, cumsum_out, normed_cumsum 

102from flag_gems.ops.diag import diag 

103from flag_gems.ops.diag_embed import diag_embed 

104from flag_gems.ops.diagonal import diagonal_backward 

105from flag_gems.ops.diff import diff 

106from flag_gems.ops.digamma_ import digamma_ 

107from flag_gems.ops.div import ( 

108 div_mode, 

109 div_mode_, 

110 floor_divide, 

111 floor_divide_, 

112 remainder, 

113 remainder_, 

114 true_divide, 

115 true_divide_, 

116 true_divide_out, 

117) 

118from flag_gems.ops.dot import dot 

119from flag_gems.ops.dropout import dropout, dropout_backward 

120from flag_gems.ops.elu import elu, elu_, elu_backward 

121from flag_gems.ops.embedding import embedding, embedding_backward 

122from flag_gems.ops.embedding_dense_backward import embedding_dense_backward 

123from flag_gems.ops.eq import eq, eq_scalar, equal 

124from flag_gems.ops.erf import erf, erf_ 

125from flag_gems.ops.exp import exp, exp_, exp_out 

126from flag_gems.ops.exp2 import exp2, exp2_ 

127from flag_gems.ops.expm1 import expm1, expm1_, expm1_out 

128from flag_gems.ops.exponential_ import exponential_ 

129from flag_gems.ops.eye import eye 

130from flag_gems.ops.eye_m import eye_m 

131from flag_gems.ops.feature_dropout import feature_dropout, feature_dropout_ 

132from flag_gems.ops.fft import fft 

133from flag_gems.ops.fill import ( 

134 fill_scalar, 

135 fill_scalar_, 

136 fill_scalar_out, 

137 fill_tensor, 

138 fill_tensor_, 

139 fill_tensor_out, 

140) 

141from flag_gems.ops.flip import flip 

142from flag_gems.ops.floor_ import floor_ 

143from flag_gems.ops.fmin import fmin, fmin_out 

144from flag_gems.ops.fmod import fmod_scalar, fmod_scalar_, fmod_tensor, fmod_tensor_ 

145from flag_gems.ops.fp8_matmul import fp8_matmul 

146from flag_gems.ops.fp8_mqa_logits import fp8_mqa_logits 

147from flag_gems.ops.full import full 

148from flag_gems.ops.full_like import full_like 

149from flag_gems.ops.gather import gather, gather_backward 

150from flag_gems.ops.gcd import gcd, gcd_out 

151from flag_gems.ops.ge import ge, ge_scalar 

152from flag_gems.ops.gelu import gelu, gelu_, gelu_backward 

153from flag_gems.ops.get_paged_mqa_logits_metadata import get_paged_mqa_logits_metadata 

154from flag_gems.ops.get_scheduler_metadata import get_scheduler_metadata 

155from flag_gems.ops.glu import glu, glu_backward 

156from flag_gems.ops.greater import ( 

157 greater, 

158 greater_out, 

159 greater_scalar, 

160 greater_scalar_out, 

161) 

162from flag_gems.ops.grid_sample import grid_sample 

163from flag_gems.ops.group_gemm import group_mm 

164from flag_gems.ops.groupnorm import group_norm, group_norm_backward 

165from flag_gems.ops.gt import gt, gt_scalar 

166from flag_gems.ops.hadamard_transform import ( 

167 hadamard_transform, 

168 hadamard_transform_12N, 

169 hadamard_transform_20N, 

170 hadamard_transform_28N, 

171 hadamard_transform_40N, 

172) 

173from flag_gems.ops.hardsigmoid import hardsigmoid, hardsigmoid_out 

174from flag_gems.ops.hardswish_ import hardswish_ 

175from flag_gems.ops.histc import histc 

176from flag_gems.ops.hstack import hstack 

177from flag_gems.ops.hypot import hypot, hypot_out 

178from flag_gems.ops.i0 import i0, i0_out 

179from flag_gems.ops.i0_ import i0_ 

180from flag_gems.ops.index import index 

181from flag_gems.ops.index_add import index_add, index_add_ 

182from flag_gems.ops.index_copy_ import index_copy, index_copy_ 

183from flag_gems.ops.index_put import _index_put_impl_, index_put, index_put_ 

184from flag_gems.ops.index_select import index_select 

185from flag_gems.ops.isclose import allclose, isclose 

186from flag_gems.ops.isfinite import isfinite 

187from flag_gems.ops.isin import isin 

188from flag_gems.ops.isinf import isinf 

189from flag_gems.ops.isnan import isnan 

190from flag_gems.ops.isneginf import isneginf, isneginf_out 

191from flag_gems.ops.kron import kron 

192from flag_gems.ops.layernorm import layer_norm, layer_norm_backward 

193from flag_gems.ops.le import le, le_scalar 

194from flag_gems.ops.leaky_relu import leaky_relu, leaky_relu_, leaky_relu_out 

195from flag_gems.ops.lerp import lerp_scalar, lerp_scalar_, lerp_tensor, lerp_tensor_ 

196from flag_gems.ops.lift_fresh_copy import lift_fresh_copy, lift_fresh_copy_out 

197from flag_gems.ops.linspace import linspace 

198from flag_gems.ops.log import log 

199from flag_gems.ops.log1p_ import log1p_ 

200from flag_gems.ops.log10 import log10, log10_, log10_out 

201from flag_gems.ops.log_sigmoid import log_sigmoid 

202from flag_gems.ops.log_softmax import ( 

203 log_softmax, 

204 log_softmax_backward, 

205 log_softmax_backward_out, 

206 log_softmax_out, 

207) 

208from flag_gems.ops.logaddexp import logaddexp, logaddexp_out 

209from flag_gems.ops.logical_and import logical_and, logical_and_ 

210from flag_gems.ops.logical_not import logical_not 

211from flag_gems.ops.logical_or import logical_or, logical_or_ 

212from flag_gems.ops.logical_xor import logical_xor 

213from flag_gems.ops.logit import logit, logit_out 

214from flag_gems.ops.logit_ import logit_ 

215from flag_gems.ops.logspace import logspace 

216from flag_gems.ops.logsumexp import logsumexp 

217from flag_gems.ops.lt import lt, lt_scalar 

218from flag_gems.ops.margin_ranking_loss import margin_ranking_loss 

219from flag_gems.ops.masked_fill import masked_fill, masked_fill_ 

220from flag_gems.ops.masked_scatter import masked_scatter, masked_scatter_ 

221from flag_gems.ops.masked_select import masked_select 

222from flag_gems.ops.max import max, max_dim 

223from flag_gems.ops.max_pool2d_with_indices import ( 

224 max_pool2d_backward, 

225 max_pool2d_with_indices, 

226) 

227from flag_gems.ops.max_pool3d_with_indices import ( 

228 max_pool3d_backward, 

229 max_pool3d_with_indices, 

230) 

231from flag_gems.ops.maximum import maximum 

232from flag_gems.ops.mean import mean, mean_dim 

233from flag_gems.ops.median import median, median_dim, median_dim_values, median_out 

234from flag_gems.ops.min import min, min_dim 

235from flag_gems.ops.minimum import minimum 

236from flag_gems.ops.mm import mm, mm_out, router_gemm 

237from flag_gems.ops.mse_loss import mse_loss 

238from flag_gems.ops.mul import mul, mul_ 

239from flag_gems.ops.multinomial import multinomial 

240from flag_gems.ops.mv import mv 

241from flag_gems.ops.nan_to_num import nan_to_num 

242from flag_gems.ops.ne import ne, ne_scalar 

243from flag_gems.ops.neg import neg, neg_ 

244from flag_gems.ops.new_full import new_full 

245from flag_gems.ops.nll_loss_nd import nll_loss_nd_backward, nll_loss_nd_forward 

246from flag_gems.ops.nllloss import ( 

247 nll_loss2d_backward, 

248 nll_loss2d_forward, 

249 nll_loss_backward, 

250 nll_loss_forward, 

251) 

252from flag_gems.ops.nonzero import nonzero 

253from flag_gems.ops.nonzero_numpy import nonzero_numpy 

254from flag_gems.ops.normal import ( 

255 normal_, 

256 normal_float_tensor, 

257 normal_tensor_float, 

258 normal_tensor_tensor, 

259) 

260from flag_gems.ops.one_hot import one_hot 

261from flag_gems.ops.ones import ones 

262from flag_gems.ops.ones_like import ones_like 

263from flag_gems.ops.pad import constant_pad_nd, pad 

264from flag_gems.ops.per_token_group_quant_fp8 import ( 

265 SUPPORTED_FP8_DTYPE, 

266 per_token_group_quant_fp8, 

267) 

268from flag_gems.ops.pixel_shuffle import pixel_shuffle 

269from flag_gems.ops.pixel_unshuffle import pixel_unshuffle, pixel_unshuffle_out 

270from flag_gems.ops.poisson import poisson 

271from flag_gems.ops.polar import polar 

272from flag_gems.ops.pow import ( 

273 pow_scalar, 

274 pow_tensor_scalar, 

275 pow_tensor_scalar_, 

276 pow_tensor_tensor, 

277 pow_tensor_tensor_, 

278) 

279from flag_gems.ops.prelu import prelu 

280from flag_gems.ops.prod import prod, prod_dim 

281from flag_gems.ops.quantile import quantile 

282from flag_gems.ops.rand import rand 

283from flag_gems.ops.rand_like import rand_like 

284from flag_gems.ops.randn import randn 

285from flag_gems.ops.randn_like import randn_like 

286from flag_gems.ops.randperm import randperm 

287from flag_gems.ops.reciprocal import reciprocal, reciprocal_ 

288from flag_gems.ops.reflection_pad1d import reflection_pad1d, reflection_pad1d_out 

289from flag_gems.ops.reflection_pad2d import reflection_pad2d, reflection_pad2d_out 

290from flag_gems.ops.relu import relu, relu_ 

291from flag_gems.ops.relu6 import relu6 

292from flag_gems.ops.repeat import repeat 

293from flag_gems.ops.repeat_interleave import ( 

294 repeat_interleave_self_int, 

295 repeat_interleave_self_tensor, 

296 repeat_interleave_tensor, 

297) 

298from flag_gems.ops.replication_pad1d import replication_pad1d, replication_pad1d_out 

299from flag_gems.ops.replication_pad3d import replication_pad3d 

300from flag_gems.ops.resolve_conj import resolve_conj 

301from flag_gems.ops.resolve_neg import resolve_neg 

302from flag_gems.ops.rms_norm import rms_norm, rms_norm_backward, rms_norm_forward 

303from flag_gems.ops.roll import roll 

304from flag_gems.ops.round import round, round_, round_out 

305from flag_gems.ops.rrelu_with_noise_backward import rrelu_with_noise_backward 

306from flag_gems.ops.rsqrt import rsqrt, rsqrt_ 

307from flag_gems.ops.rsub import rsub_scalar, rsub_tensor 

308from flag_gems.ops.scaled_softmax import scaled_softmax_backward, scaled_softmax_forward 

309from flag_gems.ops.scatter import scatter, scatter_ 

310from flag_gems.ops.scatter_add_ import scatter_add_ 

311from flag_gems.ops.scatter_reduce import ( 

312 scatter_reduce, 

313 scatter_reduce_, 

314 scatter_reduce_out, 

315) 

316from flag_gems.ops.select_backward import select_backward 

317from flag_gems.ops.select_scatter import select_scatter 

318from flag_gems.ops.selu import selu 

319from flag_gems.ops.selu_ import selu_ 

320from flag_gems.ops.sgn_ import sgn_ 

321from flag_gems.ops.sigmoid import sigmoid, sigmoid_, sigmoid_backward 

322from flag_gems.ops.signbit import signbit, signbit_out 

323from flag_gems.ops.silu import silu, silu_, silu_backward 

324from flag_gems.ops.sin import sin, sin_ 

325from flag_gems.ops.sinh_ import sinh_ 

326from flag_gems.ops.slice_backward import slice_backward 

327from flag_gems.ops.slice_scatter import slice_scatter 

328from flag_gems.ops.smooth_l1_loss import ( 

329 smooth_l1_loss, 

330 smooth_l1_loss_backward, 

331 smooth_l1_loss_out, 

332) 

333from flag_gems.ops.soft_margin_loss import soft_margin_loss, soft_margin_loss_out 

334from flag_gems.ops.softmax import ( 

335 softmax, 

336 softmax_backward, 

337 softmax_backward_out, 

338 softmax_out, 

339) 

340from flag_gems.ops.softplus import softplus 

341from flag_gems.ops.softshrink import softshrink, softshrink_out 

342from flag_gems.ops.sort import sort, sort_stable 

343from flag_gems.ops.special_i0e import special_i0e, special_i0e_out 

344from flag_gems.ops.special_i1 import special_i1, special_i1_out 

345from flag_gems.ops.sqrt import sqrt, sqrt_ 

346from flag_gems.ops.square import square, square_, square_out 

347from flag_gems.ops.stack import stack 

348from flag_gems.ops.std import std 

349from flag_gems.ops.sub import sub, sub_ 

350from flag_gems.ops.sum import sum, sum_dim, sum_dim_out, sum_out 

351from flag_gems.ops.svd import svd 

352from flag_gems.ops.t_copy import t_copy, t_copy_out 

353from flag_gems.ops.tan import tan, tan_ 

354from flag_gems.ops.tanh import tanh, tanh_, tanh_backward 

355from flag_gems.ops.threshold import threshold, threshold_backward 

356from flag_gems.ops.tile import tile 

357from flag_gems.ops.to import to_copy 

358from flag_gems.ops.topk import topk 

359from flag_gems.ops.trace import trace 

360from flag_gems.ops.tril import tril, tril_, tril_out 

361from flag_gems.ops.triu import triu, triu_ 

362from flag_gems.ops.unfold_backward import unfold_backward 

363from flag_gems.ops.uniform import uniform_ 

364from flag_gems.ops.unique import _unique2 

365from flag_gems.ops.unique_consecutive import unique_consecutive 

366from flag_gems.ops.upsample_bicubic2d import upsample_bicubic2d 

367from flag_gems.ops.upsample_bicubic2d_aa import _upsample_bicubic2d_aa 

368from flag_gems.ops.upsample_bicubic2d_aa_backward import _upsample_bicubic2d_aa_backward 

369from flag_gems.ops.upsample_linear1d import upsample_linear1d 

370from flag_gems.ops.upsample_nearest1d import upsample_nearest1d 

371from flag_gems.ops.upsample_nearest2d import upsample_nearest2d 

372from flag_gems.ops.upsample_nearest3d import upsample_nearest3d 

373from flag_gems.ops.var import var, var_correction, var_dim 

374from flag_gems.ops.var_mean import var_mean 

375from flag_gems.ops.vdot import vdot 

376from flag_gems.ops.vector_norm import vector_norm 

377from flag_gems.ops.vstack import vstack 

378from flag_gems.ops.w8a8_block_fp8_matmul import w8a8_block_fp8_matmul 

379from flag_gems.ops.weightnorm import ( 

380 weight_norm_interface, 

381 weight_norm_interface_backward, 

382) 

383from flag_gems.ops.where import ( 

384 where_scalar_other, 

385 where_scalar_self, 

386 where_self, 

387 where_self_out, 

388) 

389from flag_gems.ops.zero import zero, zero_out 

390from flag_gems.ops.zeros import zero_, zeros 

391from flag_gems.ops.zeros_like import zeros_like 

392 

393__all__ = [ 

394 "_assert_async", 

395 "_conv_depthwise2d", 

396 "_euclidean_dist", 

397 "_functional_sym_constrain_range_for_size", 

398 "_index_put_impl_", 

399 "_is_all_true", 

400 "_safe_softmax", 

401 "_unique2", 

402 "_upsample_bicubic2d_aa", 

403 "_upsample_bicubic2d_aa_backward", 

404 "_upsample_nearest_exact1d", 

405 "abs", 

406 "abs_", 

407 "absolute", 

408 "act_quant_triton", 

409 "acos", 

410 "add", 

411 "add_", 

412 "addcdiv", 

413 "addcdiv_out", 

414 "addcmul", 

415 "addcmul_out", 

416 "addmm", 

417 "addmm_dtype", 

418 "addmm_dtype_out", 

419 "addmm_out", 

420 "addmv", 

421 "addmv_out", 

422 "addr", 

423 "affine_grid_generator", 

424 "alias_copy", 

425 "alias_copy_out", 

426 "all", 

427 "all_dim", 

428 "all_dims", 

429 "allclose", 

430 "amax", 

431 "aminmax", 

432 "angle", 

433 "any", 

434 "any_dim", 

435 "any_dims", 

436 "arange", 

437 "arange_start", 

438 "arcsinh", 

439 "arcsinh_out", 

440 "arctanh_", 

441 "arcsinh_", 

442 "argmax", 

443 "argmin", 

444 "asinh", 

445 "asinh_", 

446 "asinh_out", 

447 "atan", 

448 "atan_", 

449 "atan2", 

450 "atan2_out", 

451 "avg_pool2d", 

452 "avg_pool2d_backward", 

453 "avg_pool3d", 

454 "avg_pool3d_backward", 

455 "baddbmm", 

456 "baddbmm_out", 

457 "batch_norm", 

458 "batch_norm_backward", 

459 "bernoulli_", 

460 "bincount", 

461 "bitwise_and_scalar", 

462 "bitwise_and_scalar_", 

463 "bitwise_and_scalar_tensor", 

464 "bitwise_and_tensor", 

465 "bitwise_and_tensor_", 

466 "bitwise_left_shift", 

467 "bitwise_not", 

468 "bitwise_not_", 

469 "bitwise_or_scalar", 

470 "bitwise_or_scalar_", 

471 "bitwise_or_scalar_tensor", 

472 "bitwise_or_tensor", 

473 "bitwise_or_tensor_", 

474 "bitwise_right_shift", 

475 "bmm", 

476 "bmm_out", 

477 "cat", 

478 "cat_out", 

479 "ceil", 

480 "ceil_", 

481 "ceil_out", 

482 "celu", 

483 "celu_", 

484 "col2im", 

485 "clamp", 

486 "clamp_", 

487 "clamp_min", 

488 "clamp_min_", 

489 "clamp_tensor", 

490 "clamp_tensor_", 

491 "clip", 

492 "clip_", 

493 "concatenate", 

494 "constant_pad_nd", 

495 "contiguous", 

496 "conv1d", 

497 "conv2d", 

498 "conv3d", 

499 "conv_transpose1d", 

500 "conv_transpose2d", 

501 "copy", 

502 "copy_", 

503 "copysign", 

504 "copysign_out", 

505 "cos", 

506 "cos_", 

507 "cosh", 

508 "cosh_", 

509 "cosh_out", 

510 "count_nonzero", 

511 "ctc_loss", 

512 "cudnn_convolution", 

513 "cummax", 

514 "cummin", 

515 "cumprod", 

516 "cumprod_", 

517 "cumsum", 

518 "cumsum_out", 

519 "conj_physical", 

520 "diag", 

521 "diag_embed", 

522 "diagonal_backward", 

523 "diff", 

524 "digamma_", 

525 "div_mode", 

526 "div_mode_", 

527 "dot", 

528 "dropout", 

529 "dropout_backward", 

530 "elu", 

531 "elu_", 

532 "elu_backward", 

533 "embedding", 

534 "embedding_backward", 

535 "embedding_dense_backward", 

536 "eq", 

537 "eq_scalar", 

538 "equal", 

539 "erf", 

540 "erf_", 

541 "exp", 

542 "exp_", 

543 "exp_out", 

544 "exp2", 

545 "exp2_", 

546 "expm1", 

547 "expm1_", 

548 "expm1_out", 

549 "exponential_", 

550 "eye", 

551 "eye_m", 

552 "feature_dropout", 

553 "feature_dropout_", 

554 "fill_scalar", 

555 "fill_scalar_", 

556 "fill_scalar_out", 

557 "fill_tensor", 

558 "fill_tensor_", 

559 "fill_tensor_out", 

560 "flash_attention_forward", 

561 "flash_attn_varlen_func", 

562 "flash_attn_varlen_opt_func", 

563 "flip", 

564 "fft", 

565 "floor_", 

566 "floor_divide", 

567 "floor_divide_", 

568 "fmin", 

569 "fmin_out", 

570 "fmod_scalar", 

571 "fmod_scalar_", 

572 "fmod_tensor", 

573 "fmod_tensor_", 

574 "fp8_mqa_logits", 

575 "full", 

576 "full_like", 

577 "gather", 

578 "gather_backward", 

579 "gcd", 

580 "gcd_out", 

581 "ge", 

582 "ge_scalar", 

583 "gelu", 

584 "gelu_", 

585 "gelu_backward", 

586 "get_paged_mqa_logits_metadata", 

587 "get_scheduler_metadata", 

588 "glu", 

589 "glu_backward", 

590 "grid_sample", 

591 "greater", 

592 "greater_out", 

593 "greater_scalar", 

594 "greater_scalar_out", 

595 "group_mm", 

596 "group_norm", 

597 "group_norm_backward", 

598 "gt", 

599 "gt_scalar", 

600 "hadamard_transform", 

601 "hadamard_transform_12N", 

602 "hadamard_transform_20N", 

603 "hadamard_transform_28N", 

604 "hadamard_transform_40N", 

605 "hardsigmoid", 

606 "hardsigmoid_out", 

607 "hardswish_", 

608 "histc", 

609 "hstack", 

610 "hypot", 

611 "hypot_out", 

612 "i0", 

613 "i0_out", 

614 "i0_", 

615 "index", 

616 "index_add", 

617 "index_add_", 

618 "index_copy", 

619 "index_copy_", 

620 "index_put", 

621 "index_put_", 

622 "index_select", 

623 "isclose", 

624 "isfinite", 

625 "isin", 

626 "isinf", 

627 "isnan", 

628 "isneginf", 

629 "isneginf_out", 

630 "kron", 

631 "layer_norm", 

632 "layer_norm_backward", 

633 "leaky_relu", 

634 "leaky_relu_", 

635 "leaky_relu_out", 

636 "le", 

637 "le_scalar", 

638 "lerp_scalar", 

639 "lerp_scalar_", 

640 "lerp_tensor", 

641 "lerp_tensor_", 

642 "lift_fresh_copy", 

643 "lift_fresh_copy_out", 

644 "linspace", 

645 "log", 

646 "log10", 

647 "log10_", 

648 "log10_out", 

649 "log_sigmoid", 

650 "log_softmax", 

651 "log_softmax_backward", 

652 "log_softmax_backward_out", 

653 "log_softmax_out", 

654 "log1p_", 

655 "logaddexp", 

656 "logaddexp_out", 

657 "logical_and", 

658 "logical_and_", 

659 "logical_not", 

660 "logical_or", 

661 "logical_or_", 

662 "logical_xor", 

663 "logit", 

664 "logit_out", 

665 "logit_", 

666 "logspace", 

667 "logsumexp", 

668 "lt", 

669 "lt_scalar", 

670 "margin_ranking_loss", 

671 "masked_fill", 

672 "masked_fill_", 

673 "masked_scatter", 

674 "masked_scatter_", 

675 "masked_select", 

676 "max", 

677 "max_dim", 

678 "max_pool2d_with_indices", 

679 "max_pool2d_backward", 

680 "max_pool3d_with_indices", 

681 "max_pool3d_backward", 

682 "maximum", 

683 "mean", 

684 "mean_dim", 

685 "median", 

686 "median_dim", 

687 "median_dim_values", 

688 "median_out", 

689 "min", 

690 "min_dim", 

691 "minimum", 

692 "mm", 

693 "mm_out", 

694 "mse_loss", 

695 "mul", 

696 "mul_", 

697 "multinomial", 

698 "mv", 

699 "nan_to_num", 

700 "ne", 

701 "ne_scalar", 

702 "neg", 

703 "neg_", 

704 "new_full", 

705 "nll_loss_backward", 

706 "nll_loss_forward", 

707 "nll_loss2d_backward", 

708 "nll_loss2d_forward", 

709 "nll_loss_nd_forward", 

710 "nll_loss_nd_backward", 

711 "nonzero", 

712 "nonzero_numpy", 

713 "normal_float_tensor", 

714 "normal_tensor_float", 

715 "normal_tensor_tensor", 

716 "normal_", 

717 "normed_cumsum", 

718 "ones", 

719 "ones_like", 

720 "one_hot", 

721 "pad", 

722 "per_token_group_quant_fp8", 

723 "pixel_shuffle", 

724 "pixel_unshuffle", 

725 "pixel_unshuffle_out", 

726 "poisson", 

727 "polar", 

728 "pow_scalar", 

729 "pow_tensor_scalar", 

730 "pow_tensor_scalar_", 

731 "pow_tensor_tensor", 

732 "pow_tensor_tensor_", 

733 "prelu", 

734 "prod", 

735 "prod_dim", 

736 "quantile", 

737 "rand", 

738 "rand_like", 

739 "randn", 

740 "randn_like", 

741 "randperm", 

742 "reciprocal", 

743 "reciprocal_", 

744 "reflection_pad2d", 

745 "reflection_pad2d_out", 

746 "reflection_pad1d", 

747 "reflection_pad1d_out", 

748 "relu", 

749 "relu_", 

750 "relu6", 

751 "remainder", 

752 "remainder_", 

753 "repeat", 

754 "repeat_interleave_self_int", 

755 "repeat_interleave_self_tensor", 

756 "repeat_interleave_tensor", 

757 "replication_pad1d", 

758 "replication_pad1d_out", 

759 "replication_pad3d", 

760 "resolve_conj", 

761 "resolve_neg", 

762 "rms_norm", 

763 "rms_norm_backward", 

764 "rms_norm_forward", 

765 "roll", 

766 "round", 

767 "round_", 

768 "round_out", 

769 "rrelu_with_noise_backward", 

770 "rsqrt", 

771 "rsqrt_", 

772 "rsub_scalar", 

773 "rsub_tensor", 

774 "scaled_dot_product_attention", 

775 "scaled_dot_product_attention_backward", 

776 "scaled_dot_product_attention_forward", 

777 "scaled_softmax_backward", 

778 "scaled_softmax_forward", 

779 "scatter", 

780 "scatter_", 

781 "scatter_add_", 

782 "scatter_reduce", 

783 "scatter_reduce_", 

784 "scatter_reduce_out", 

785 "select_backward", 

786 "select_scatter", 

787 "selu", 

788 "selu_", 

789 "sgn_", 

790 "sigmoid", 

791 "sigmoid_", 

792 "sigmoid_backward", 

793 "signbit", 

794 "signbit_out", 

795 "silu", 

796 "silu_", 

797 "silu_backward", 

798 "sin", 

799 "sin_", 

800 "sinh_", 

801 "slice_backward", 

802 "slice_scatter", 

803 "smooth_l1_loss", 

804 "smooth_l1_loss_backward", 

805 "smooth_l1_loss_out", 

806 "soft_margin_loss", 

807 "soft_margin_loss_out", 

808 "softmax", 

809 "softmax_backward", 

810 "softmax_backward_out", 

811 "softmax_out", 

812 "softplus", 

813 "softshrink", 

814 "softshrink_out", 

815 "sort", 

816 "sort_stable", 

817 "special_i1", 

818 "special_i1_out", 

819 "special_i0e", 

820 "special_i0e_out", 

821 "sqrt", 

822 "sqrt_", 

823 "square", 

824 "square_", 

825 "square_out", 

826 "stack", 

827 "std", 

828 "sub", 

829 "sub_", 

830 "sum", 

831 "sum_dim", 

832 "sum_dim_out", 

833 "sum_out", 

834 "svd", 

835 "ScaleDotProductAttention", 

836 "SUPPORTED_FP8_DTYPE", 

837 "t_copy", 

838 "t_copy_out", 

839 "tan", 

840 "tan_", 

841 "tanh", 

842 "tanh_", 

843 "tanh_backward", 

844 "threshold", 

845 "threshold_backward", 

846 "tile", 

847 "to_copy", 

848 "topk", 

849 "trace", 

850 "tril", 

851 "tril_", 

852 "tril_out", 

853 "triu", 

854 "triu_", 

855 "true_divide", 

856 "true_divide_", 

857 "true_divide_out", 

858 "unfold_backward", 

859 "uniform_", 

860 "unique_consecutive", 

861 "upsample_bicubic2d", 

862 "upsample_linear1d", 

863 "upsample_nearest1d", 

864 "upsample_nearest2d", 

865 "upsample_nearest3d", 

866 "var_mean",