gistfile1.txt

import torch
from torch import tensor, device
import torch.fx as fx
from torch._dynamo.testing import rand_strided
from math import inf
import torch._inductor.inductor_prims

import torch._dynamo.config
import torch._inductor.config
import torch._functorch.config
torch._dynamo.config.cache_size_limit = 128
torch._dynamo.config.suppress_errors = True




isolate_fails_code_str = None



# torch version: 2.1.2+cu121
# torch cuda version: 12.1
# torch git version: a8e7c98cb95ff97bb30a728c6b2a1ce6bff946eb


# CUDA Info:
# nvcc not found
# GPU Hardware Info:
# NVIDIA GeForce RTX 4090 : 1


from torch.nn import *
class Repro(torch.nn.Module):
    def __init__(self):
        super().__init__()



    def forward(self, primals_410, sym_size, sym_size_1, sym_size_14, primals_116, primals_122, primals_128, primals_134, primals_140, primals_146, primals_152, primals_158, primals_164, primals_170, primals_176, primals_182, primals_188, primals_194, primals_200, primals_206, primals_212, primals_218, primals_224, primals_230, primals_236, primals_242, primals_248, primals_254, primals_260, primals_266, primals_272, primals_278, primals_284, primals_290, primals_296, primals_302, primals_308, primals_314, primals_320, primals_326, primals_332, primals_338, primals_344, primals_350, primals_356, primals_362, primals_368, primals_374, primals_380, primals_386, primals_392, primals_398, primals_404, primals_415, view, sin, cos, convert_element_type_2, addmm, convert_element_type_4, addmm_1, convert_element_type_8, addmm_2, convert_element_type_12, iota_2, lt_1, view_292, mm_72, view_294, addmm_3, convert_element_type_205, convert_element_type_208, convert_element_type_209, view_296, clone_99, cat_4, getitem_1, rsqrt_25, view_298, permute_149, cat_6, cat_8, getitem_6, getitem_7, getitem_8, getitem_9, getitem_12, getitem_13, view_310, add_75, getitem_16, rsqrt_26, view_312, view_314, permute_155, permute_156, permute_157, getitem_17, getitem_18, getitem_19, getitem_20, getitem_23, getitem_24, view_322, add_78, getitem_27, rsqrt_27, view_324, getitem_28, getitem_29, view_326, add_81, getitem_31, rsqrt_28, view_328, permute_167, cat_10, cat_12, getitem_36, getitem_37, getitem_38, getitem_39, getitem_42, getitem_43, view_340, add_86, getitem_46, rsqrt_29, view_342, permute_173, permute_174, permute_175, getitem_47, getitem_48, getitem_49, getitem_50, getitem_53, getitem_54, view_352, add_89, getitem_57, rsqrt_30, view_354, getitem_58, getitem_59, view_356, add_92, getitem_61, rsqrt_31, view_358, permute_185, cat_14, cat_16, getitem_66, getitem_67, getitem_68, getitem_69, getitem_72, getitem_73, view_370, add_97, getitem_76, rsqrt_32, view_372, permute_191, permute_192, permute_193, getitem_77, getitem_78, getitem_79, getitem_80, getitem_83, getitem_84, view_382, add_100, getitem_87, rsqrt_33, view_384, getitem_88, getitem_89, view_386, add_103, getitem_91, rsqrt_34, view_388, permute_203, cat_18, cat_20, getitem_96, getitem_97, getitem_98, getitem_99, getitem_102, getitem_103, view_400, add_108, getitem_106, rsqrt_35, view_402, permute_209, permute_210, permute_211, getitem_107, getitem_108, getitem_109, getitem_110, getitem_113, getitem_114, view_412, add_111, getitem_117, rsqrt_36, view_414, getitem_118, getitem_119, view_416, add_114, getitem_121, rsqrt_37, view_418, permute_221, cat_22, cat_24, getitem_126, getitem_127, getitem_128, getitem_129, getitem_132, getitem_133, view_430, add_119, getitem_136, rsqrt_38, view_432, permute_227, permute_228, permute_229, getitem_137, getitem_138, getitem_139, getitem_140, getitem_143, getitem_144, view_442, add_122, getitem_147, rsqrt_39, view_444, getitem_148, getitem_149, view_446, add_125, getitem_151, rsqrt_40, view_448, permute_239, cat_26, cat_28, getitem_156, getitem_157, getitem_158, getitem_159, getitem_162, getitem_163, view_460, add_130, getitem_166, rsqrt_41, view_462, permute_245, permute_246, permute_247, getitem_167, getitem_168, getitem_169, getitem_170, getitem_173, getitem_174, view_472, add_133, getitem_177, rsqrt_42, view_474, getitem_178, getitem_179, view_476, add_136, getitem_181, rsqrt_43, view_478, permute_257, cat_30, cat_32, getitem_186, getitem_187, getitem_188, getitem_189, getitem_192, getitem_193, view_490, add_141, getitem_196, rsqrt_44, view_492, permute_263, permute_264, permute_265, getitem_197, getitem_198, getitem_199, getitem_200, getitem_203, getitem_204, view_502, add_144, getitem_207, rsqrt_45, view_504, getitem_208, getitem_209, view_506, add_147, getitem_211, rsqrt_46, view_508, permute_275, cat_34, cat_36, getitem_216, getitem_217, getitem_218, getitem_219, getitem_222, getitem_223, view_520, add_152, getitem_226, rsqrt_47, view_522, permute_281, permute_282, permute_283, getitem_227, getitem_228, getitem_229, getitem_230, getitem_233, getitem_234, view_532, add_155, getitem_237, rsqrt_48, view_534, getitem_238, getitem_239, view_536, add_158, getitem_241, rsqrt_49, view_538, permute_293, cat_38, cat_40, getitem_246, getitem_247, getitem_248, getitem_249, getitem_252, getitem_253, view_550, add_163, getitem_256, rsqrt_50, view_552, permute_299, permute_300, permute_301, getitem_257, getitem_258, getitem_259, getitem_260, getitem_263, getitem_264, view_562, add_166, getitem_267, rsqrt_51, view_564, getitem_268, getitem_269, view_566, add_169, getitem_271, rsqrt_52, view_568, permute_311, cat_42, cat_44, getitem_276, getitem_277, getitem_278, getitem_279, getitem_282, getitem_283, view_580, add_174, getitem_286, rsqrt_53, view_582, permute_317, permute_318, permute_319, getitem_287, getitem_288, getitem_289, getitem_290, getitem_293, getitem_294, view_592, add_177, getitem_297, rsqrt_54, view_594, getitem_298, getitem_299, view_596, add_180, getitem_301, rsqrt_55, view_598, permute_329, cat_46, cat_48, getitem_306, getitem_307, getitem_308, getitem_309, getitem_312, getitem_313, view_610, add_185, getitem_316, rsqrt_56, view_612, permute_335, permute_336, permute_337, getitem_317, getitem_318, getitem_319, getitem_320, getitem_323, getitem_324, view_622, add_188, getitem_327, rsqrt_57, view_624, getitem_328, getitem_329, view_626, add_191, getitem_331, rsqrt_58, view_628, permute_347, cat_50, cat_52, getitem_336, getitem_337, getitem_338, getitem_339, getitem_342, getitem_343, view_640, add_196, getitem_346, rsqrt_59, view_642, permute_353, permute_354, permute_355, getitem_347, getitem_348, getitem_349, getitem_350, getitem_353, getitem_354, view_652, add_199, getitem_357, rsqrt_60, view_654, getitem_358, getitem_359, view_656, add_202, getitem_361, rsqrt_61, view_658, permute_365, cat_54, cat_56, getitem_366, getitem_367, getitem_368, getitem_369, getitem_372, getitem_373, view_670, add_207, getitem_376, rsqrt_62, view_672, permute_371, permute_372, permute_373, getitem_377, getitem_378, getitem_379, getitem_380, getitem_383, getitem_384, view_682, add_210, getitem_387, rsqrt_63, view_684, getitem_388, getitem_389, view_686, add_213, getitem_391, rsqrt_64, view_688, permute_383, cat_58, cat_60, getitem_396, getitem_397, getitem_398, getitem_399, getitem_402, getitem_403, view_700, add_218, getitem_406, rsqrt_65, view_702, permute_389, permute_390, permute_391, getitem_407, getitem_408, getitem_409, getitem_410, getitem_413, getitem_414, view_712, add_221, getitem_417, rsqrt_66, view_714, getitem_418, getitem_419, view_716, add_224, getitem_421, rsqrt_67, view_718, permute_401, cat_62, cat_64, getitem_426, getitem_427, getitem_428, getitem_429, getitem_432, getitem_433, view_730, add_229, getitem_436, rsqrt_68, view_732, permute_407, permute_408, permute_409, getitem_437, getitem_438, getitem_439, getitem_440, getitem_443, getitem_444, view_742, add_232, getitem_447, rsqrt_69, view_744, getitem_448, getitem_449, view_746, add_235, getitem_451, rsqrt_70, view_748, permute_419, cat_66, cat_68, getitem_456, getitem_457, getitem_458, getitem_459, getitem_462, getitem_463, view_760, add_240, getitem_466, rsqrt_71, view_762, permute_425, permute_426, permute_427, getitem_467, getitem_468, getitem_469, getitem_470, getitem_473, getitem_474, view_772, add_243, getitem_477, rsqrt_72, view_774, getitem_478, getitem_479, view_776, add_246, getitem_481, rsqrt_73, view_778, slice_263, convert_element_type_572, permute_435, permute_439, permute_444, permute_450, permute_458, permute_462, permute_466, permute_470, permute_478, permute_482, permute_486, permute_488, permute_493, permute_499, permute_507, permute_511, permute_515, permute_519, permute_527, permute_531, permute_535, permute_537, permute_542, permute_548, permute_556, permute_560, permute_564, permute_568, permute_576, permute_580, permute_584, permute_586, permute_591, permute_597, permute_605, permute_609, permute_613, permute_617, permute_625, permute_629, permute_633, permute_635, permute_640, permute_646, permute_654, permute_658, permute_662, permute_666, permute_674, permute_678, permute_682, permute_684, permute_689, permute_695, permute_703, permute_707, permute_711, permute_715, permute_723, permute_727, permute_731, permute_733, permute_738, permute_744, permute_752, permute_756, permute_760, permute_764, permute_772, permute_776, permute_780, permute_782, permute_787, permute_793, permute_801, permute_805, permute_809, permute_813, permute_821, permute_825, permute_829, permute_831, permute_836, permute_842, permute_850, permute_854, permute_858, permute_862, permute_870, permute_874, permute_878, permute_880, permute_885, permute_891, permute_899, permute_903, permute_907, permute_911, permute_919, permute_923, permute_927, permute_929, permute_934, permute_940, permute_948, permute_952, permute_956, permute_960, permute_968, permute_972, permute_976, permute_978, permute_983, permute_989, permute_997, permute_1001, permute_1005, permute_1009, permute_1017, permute_1021, permute_1025, permute_1027, permute_1032, permute_1038, permute_1046, permute_1050, permute_1054, permute_1058, permute_1066, permute_1070, permute_1074, permute_1076, permute_1081, permute_1087, permute_1095, permute_1099, permute_1103, permute_1107, permute_1115, permute_1119, permute_1123, permute_1125, permute_1130, permute_1136, permute_1144, permute_1148, permute_1152, permute_1156, permute_1164, permute_1168, permute_1172, permute_1174, permute_1179, permute_1185, permute_1193, permute_1197, permute_1201, permute_1205, permute_1213, permute_1217, permute_1221, permute_1225, permute_1228, permute_1233, permute_1239, permute_1244, permute_1246, permute_1254, tangents_1, tangents_2, tangents_3, tangents_4, tangents_5, tangents_6, tangents_7, tangents_8, tangents_9, tangents_10, tangents_11, tangents_12, tangents_13, tangents_14, tangents_15, tangents_16, tangents_17, tangents_18, tangents_19, tangents_20, tangents_21, tangents_22, tangents_23, tangents_24, tangents_25, tangents_26, tangents_27, tangents_28, tangents_29, tangents_30, tangents_31, tangents_32, tangents_33, tangents_34, tangents_35, tangents_36, tangents_37, tangents_38, tangents_39, tangents_40, tangents_41, tangents_42, tangents_43, tangents_44, tangents_45, tangents_46, tangents_47, tangents_48, tangents_49, tangents_50, tangents_51, tangents_52, tangents_53, tangents_54, tangents_55, tangents_56, tangents_57, tangents_58, tangents_59, tangents_60, tangents_61, tangents_62, tangents_63, tangents_64, tangents_65):
        view_2 = torch.ops.aten.view.default(addmm, [sym_size, 768]);  addmm = None
        convert_element_type_3 = torch.ops.prims.convert_element_type.default(view_2, torch.float32);  view_2 = None
        mul_4 = torch.ops.aten.mul.Tensor(convert_element_type_3, 0.7071067811865476)
        erf = torch.ops.aten.erf.default(mul_4);  mul_4 = None
        add = torch.ops.aten.add.Tensor(erf, 1);  erf = None
        convert_element_type_7 = torch.ops.prims.convert_element_type.default(addmm_1, torch.float32);  addmm_1 = None
        mul_7 = torch.ops.aten.mul.Tensor(convert_element_type_7, 0.7071067811865476)
        erf_1 = torch.ops.aten.erf.default(mul_7);  mul_7 = None
        add_1 = torch.ops.aten.add.Tensor(erf_1, 1);  erf_1 = None
        convert_element_type_11 = torch.ops.prims.convert_element_type.default(addmm_2, torch.float32);  addmm_2 = None
        mul_10 = torch.ops.aten.mul.Tensor(convert_element_type_11, 0.7071067811865476)
        erf_2 = torch.ops.aten.erf.default(mul_10);  mul_10 = None
        add_2 = torch.ops.aten.add.Tensor(erf_2, 1);  erf_2 = None
        view_293 = torch.ops.aten.view.default(mm_72, [sym_size_1, 128, 1024]);  mm_72 = None
        convert_element_type_216 = torch.ops.prims.convert_element_type.default(clone_99, torch.float32);  clone_99 = None
        sub_14 = torch.ops.aten.sub.Tensor(convert_element_type_216, getitem_1);  convert_element_type_216 = getitem_1 = None
        mul_168 = torch.ops.aten.mul.Tensor(sub_14, rsqrt_25);  sub_14 = None
        slice_9 = torch.ops.aten.slice.Tensor(cat_4, 0, -513, 9223372036854775807);  cat_4 = None
        slice_10 = torch.ops.aten.slice.Tensor(slice_9, 1, 0, 9223372036854775807);  slice_9 = None
        cos_1 = torch.ops.aten.cos.default(slice_10)
        sin_1 = torch.ops.aten.sin.default(slice_10);  slice_10 = None
        convert_element_type_224 = torch.ops.prims.convert_element_type.default(add_75, torch.float32);  add_75 = None
        sub_15 = torch.ops.aten.sub.Tensor(convert_element_type_224, getitem_16);  convert_element_type_224 = getitem_16 = None
        mul_182 = torch.ops.aten.mul.Tensor(sub_15, rsqrt_26);  sub_15 = None
        convert_element_type_230 = torch.ops.prims.convert_element_type.default(add_78, torch.float32);  add_78 = None
        sub_16 = torch.ops.aten.sub.Tensor(convert_element_type_230, getitem_27);  convert_element_type_230 = getitem_27 = None
        mul_188 = torch.ops.aten.mul.Tensor(sub_16, rsqrt_27);  sub_16 = None
        convert_element_type_234 = torch.ops.prims.convert_element_type.default(getitem_29, torch.float32)
        sigmoid_2 = torch.ops.aten.sigmoid.default(convert_element_type_234)
        mul_191 = torch.ops.aten.mul.Tensor(convert_element_type_234, sigmoid_2);  convert_element_type_234 = sigmoid_2 = None
        convert_element_type_235 = torch.ops.prims.convert_element_type.default(mul_191, torch.float16);  mul_191 = None
        convert_element_type_238 = torch.ops.prims.convert_element_type.default(add_81, torch.float32);  add_81 = None
        sub_17 = torch.ops.aten.sub.Tensor(convert_element_type_238, getitem_31);  convert_element_type_238 = getitem_31 = None
        mul_195 = torch.ops.aten.mul.Tensor(sub_17, rsqrt_28);  sub_17 = None
        convert_element_type_246 = torch.ops.prims.convert_element_type.default(add_86, torch.float32);  add_86 = None
        sub_18 = torch.ops.aten.sub.Tensor(convert_element_type_246, getitem_46);  convert_element_type_246 = getitem_46 = None
        mul_209 = torch.ops.aten.mul.Tensor(sub_18, rsqrt_29);  sub_18 = None
        convert_element_type_252 = torch.ops.prims.convert_element_type.default(add_89, torch.float32);  add_89 = None
        sub_19 = torch.ops.aten.sub.Tensor(convert_element_type_252, getitem_57);  convert_element_type_252 = getitem_57 = None
        mul_215 = torch.ops.aten.mul.Tensor(sub_19, rsqrt_30);  sub_19 = None
        convert_element_type_256 = torch.ops.prims.convert_element_type.default(getitem_59, torch.float32)
        sigmoid_3 = torch.ops.aten.sigmoid.default(convert_element_type_256)
        mul_218 = torch.ops.aten.mul.Tensor(convert_element_type_256, sigmoid_3);  convert_element_type_256 = sigmoid_3 = None
        convert_element_type_257 = torch.ops.prims.convert_element_type.default(mul_218, torch.float16);  mul_218 = None
        convert_element_type_260 = torch.ops.prims.convert_element_type.default(add_92, torch.float32);  add_92 = None
        sub_20 = torch.ops.aten.sub.Tensor(convert_element_type_260, getitem_61);  convert_element_type_260 = getitem_61 = None
        mul_222 = torch.ops.aten.mul.Tensor(sub_20, rsqrt_31);  sub_20 = None
        convert_element_type_268 = torch.ops.prims.convert_element_type.default(add_97, torch.float32);  add_97 = None
        sub_21 = torch.ops.aten.sub.Tensor(convert_element_type_268, getitem_76);  convert_element_type_268 = getitem_76 = None
        mul_236 = torch.ops.aten.mul.Tensor(sub_21, rsqrt_32);  sub_21 = None
        convert_element_type_274 = torch.ops.prims.convert_element_type.default(add_100, torch.float32);  add_100 = None
        sub_22 = torch.ops.aten.sub.Tensor(convert_element_type_274, getitem_87);  convert_element_type_274 = getitem_87 = None
        mul_242 = torch.ops.aten.mul.Tensor(sub_22, rsqrt_33);  sub_22 = None
        convert_element_type_278 = torch.ops.prims.convert_element_type.default(getitem_89, torch.float32)
        sigmoid_4 = torch.ops.aten.sigmoid.default(convert_element_type_278)
        mul_245 = torch.ops.aten.mul.Tensor(convert_element_type_278, sigmoid_4);  convert_element_type_278 = sigmoid_4 = None
        convert_element_type_279 = torch.ops.prims.convert_element_type.default(mul_245, torch.float16);  mul_245 = None
        convert_element_type_282 = torch.ops.prims.convert_element_type.default(add_103, torch.float32);  add_103 = None
        sub_23 = torch.ops.aten.sub.Tensor(convert_element_type_282, getitem_91);  convert_element_type_282 = getitem_91 = None
        mul_249 = torch.ops.aten.mul.Tensor(sub_23, rsqrt_34);  sub_23 = None
        convert_element_type_290 = torch.ops.prims.convert_element_type.default(add_108, torch.float32);  add_108 = None
        sub_24 = torch.ops.aten.sub.Tensor(convert_element_type_290, getitem_106);  convert_element_type_290 = getitem_106 = None
        mul_263 = torch.ops.aten.mul.Tensor(sub_24, rsqrt_35);  sub_24 = None
        convert_element_type_296 = torch.ops.prims.convert_element_type.default(add_111, torch.float32);  add_111 = None
        sub_25 = torch.ops.aten.sub.Tensor(convert_element_type_296, getitem_117);  convert_element_type_296 = getitem_117 = None
        mul_269 = torch.ops.aten.mul.Tensor(sub_25, rsqrt_36);  sub_25 = None
        convert_element_type_300 = torch.ops.prims.convert_element_type.default(getitem_119, torch.float32)
        sigmoid_5 = torch.ops.aten.sigmoid.default(convert_element_type_300)
        mul_272 = torch.ops.aten.mul.Tensor(convert_element_type_300, sigmoid_5);  convert_element_type_300 = sigmoid_5 = None
        convert_element_type_301 = torch.ops.prims.convert_element_type.default(mul_272, torch.float16);  mul_272 = None
        convert_element_type_304 = torch.ops.prims.convert_element_type.default(add_114, torch.float32);  add_114 = None
        sub_26 = torch.ops.aten.sub.Tensor(convert_element_type_304, getitem_121);  convert_element_type_304 = getitem_121 = None
        mul_276 = torch.ops.aten.mul.Tensor(sub_26, rsqrt_37);  sub_26 = None
        convert_element_type_312 = torch.ops.prims.convert_element_type.default(add_119, torch.float32);  add_119 = None
        sub_27 = torch.ops.aten.sub.Tensor(convert_element_type_312, getitem_136);  convert_element_type_312 = getitem_136 = None
        mul_290 = torch.ops.aten.mul.Tensor(sub_27, rsqrt_38);  sub_27 = None
        convert_element_type_318 = torch.ops.prims.convert_element_type.default(add_122, torch.float32);  add_122 = None
        sub_28 = torch.ops.aten.sub.Tensor(convert_element_type_318, getitem_147);  convert_element_type_318 = getitem_147 = None
        mul_296 = torch.ops.aten.mul.Tensor(sub_28, rsqrt_39);  sub_28 = None
        convert_element_type_322 = torch.ops.prims.convert_element_type.default(getitem_149, torch.float32)
        sigmoid_6 = torch.ops.aten.sigmoid.default(convert_element_type_322)
        mul_299 = torch.ops.aten.mul.Tensor(convert_element_type_322, sigmoid_6);  convert_element_type_322 = sigmoid_6 = None
        convert_element_type_323 = torch.ops.prims.convert_element_type.default(mul_299, torch.float16);  mul_299 = None
        convert_element_type_326 = torch.ops.prims.convert_element_type.default(add_125, torch.float32);  add_125 = None
        sub_29 = torch.ops.aten.sub.Tensor(convert_element_type_326, getitem_151);  convert_element_type_326 = getitem_151 = None
        mul_303 = torch.ops.aten.mul.Tensor(sub_29, rsqrt_40);  sub_29 = None
        convert_element_type_334 = torch.ops.prims.convert_element_type.default(add_130, torch.float32);  add_130 = None
        sub_30 = torch.ops.aten.sub.Tensor(convert_element_type_334, getitem_166);  convert_element_type_334 = getitem_166 = None
        mul_317 = torch.ops.aten.mul.Tensor(sub_30, rsqrt_41);  sub_30 = None
        convert_element_type_340 = torch.ops.prims.convert_element_type.default(add_133, torch.float32);  add_133 = None
        sub_31 = torch.ops.aten.sub.Tensor(convert_element_type_340, getitem_177);  convert_element_type_340 = getitem_177 = None
        mul_323 = torch.ops.aten.mul.Tensor(sub_31, rsqrt_42);  sub_31 = None
        convert_element_type_344 = torch.ops.prims.convert_element_type.default(getitem_179, torch.float32)
        sigmoid_7 = torch.ops.aten.sigmoid.default(convert_element_type_344)
        mul_326 = torch.ops.aten.mul.Tensor(convert_element_type_344, sigmoid_7);  convert_element_type_344 = sigmoid_7 = None
        convert_element_type_345 = torch.ops.prims.convert_element_type.default(mul_326, torch.float16);  mul_326 = None
        convert_element_type_348 = torch.ops.prims.convert_element_type.default(add_136, torch.float32);  add_136 = None
        sub_32 = torch.ops.aten.sub.Tensor(convert_element_type_348, getitem_181);  convert_element_type_348 = getitem_181 = None
        mul_330 = torch.ops.aten.mul.Tensor(sub_32, rsqrt_43);  sub_32 = None
        convert_element_type_356 = torch.ops.prims.convert_element_type.default(add_141, torch.float32);  add_141 = None
        sub_33 = torch.ops.aten.sub.Tensor(convert_element_type_356, getitem_196);  convert_element_type_356 = getitem_196 = None
        mul_344 = torch.ops.aten.mul.Tensor(sub_33, rsqrt_44);  sub_33 = None
        convert_element_type_362 = torch.ops.prims.convert_element_type.default(add_144, torch.float32);  add_144 = None
        sub_34 = torch.ops.aten.sub.Tensor(convert_element_type_362, getitem_207);  convert_element_type_362 = getitem_207 = None
        mul_350 = torch.ops.aten.mul.Tensor(sub_34, rsqrt_45);  sub_34 = None
        convert_element_type_366 = torch.ops.prims.convert_element_type.default(getitem_209, torch.float32)
        sigmoid_8 = torch.ops.aten.sigmoid.default(convert_element_type_366)
        mul_353 = torch.ops.aten.mul.Tensor(convert_element_type_366, sigmoid_8);  convert_element_type_366 = sigmoid_8 = None
        convert_element_type_367 = torch.ops.prims.convert_element_type.default(mul_353, torch.float16);  mul_353 = None
        convert_element_type_370 = torch.ops.prims.convert_element_type.default(add_147, torch.float32);  add_147 = None
        sub_35 = torch.ops.aten.sub.Tensor(convert_element_type_370, getitem_211);  convert_element_type_370 = getitem_211 = None
        mul_357 = torch.ops.aten.mul.Tensor(sub_35, rsqrt_46);  sub_35 = None
        convert_element_type_378 = torch.ops.prims.convert_element_type.default(add_152, torch.float32);  add_152 = None
        sub_36 = torch.ops.aten.sub.Tensor(convert_element_type_378, getitem_226);  convert_element_type_378 = getitem_226 = None
        mul_371 = torch.ops.aten.mul.Tensor(sub_36, rsqrt_47);  sub_36 = None
        convert_element_type_384 = torch.ops.prims.convert_element_type.default(add_155, torch.float32);  add_155 = None
        sub_37 = torch.ops.aten.sub.Tensor(convert_element_type_384, getitem_237);  convert_element_type_384 = getitem_237 = None
        mul_377 = torch.ops.aten.mul.Tensor(sub_37, rsqrt_48);  sub_37 = None
        convert_element_type_388 = torch.ops.prims.convert_element_type.default(getitem_239, torch.float32)
        sigmoid_9 = torch.ops.aten.sigmoid.default(convert_element_type_388)
        mul_380 = torch.ops.aten.mul.Tensor(convert_element_type_388, sigmoid_9);  convert_element_type_388 = sigmoid_9 = None
        convert_element_type_389 = torch.ops.prims.convert_element_type.default(mul_380, torch.float16);  mul_380 = None
        convert_element_type_392 = torch.ops.prims.convert_element_type.default(add_158, torch.float32);  add_158 = None
        sub_38 = torch.ops.aten.sub.Tensor(convert_element_type_392, getitem_241);  convert_element_type_392 = getitem_241 = None
        mul_384 = torch.ops.aten.mul.Tensor(sub_38, rsqrt_49);  sub_38 = None
        convert_element_type_400 = torch.ops.prims.convert_element_type.default(add_163, torch.float32);  add_163 = None
        sub_39 = torch.ops.aten.sub.Tensor(convert_element_type_400, getitem_256);  convert_element_type_400 = getitem_256 = None
        mul_398 = torch.ops.aten.mul.Tensor(sub_39, rsqrt_50);  sub_39 = None
        convert_element_type_406 = torch.ops.prims.convert_element_type.default(add_166, torch.float32);  add_166 = None
        sub_40 = torch.ops.aten.sub.Tensor(convert_element_type_406, getitem_267);  convert_element_type_406 = getitem_267 = None
        mul_404 = torch.ops.aten.mul.Tensor(sub_40, rsqrt_51);  sub_40 = None
        convert_element_type_410 = torch.ops.prims.convert_element_type.default(getitem_269, torch.float32)
        sigmoid_10 = torch.ops.aten.sigmoid.default(convert_element_type_410)
        mul_407 = torch.ops.aten.mul.Tensor(convert_element_type_410, sigmoid_10);  convert_element_type_410 = sigmoid_10 = None
        convert_element_type_411 = torch.ops.prims.convert_element_type.default(mul_407, torch.float16);  mul_407 = None
        convert_element_type_414 = torch.ops.prims.convert_element_type.default(add_169, torch.float32);  add_169 = None
        sub_41 = torch.ops.aten.sub.Tensor(convert_element_type_414, getitem_271);  convert_element_type_414 = getitem_271 = None
        mul_411 = torch.ops.aten.mul.Tensor(sub_41, rsqrt_52);  sub_41 = None
        convert_element_type_422 = torch.ops.prims.convert_element_type.default(add_174, torch.float32);  add_174 = None
        sub_42 = torch.ops.aten.sub.Tensor(convert_element_type_422, getitem_286);  convert_element_type_422 = getitem_286 = None
        mul_425 = torch.ops.aten.mul.Tensor(sub_42, rsqrt_53);  sub_42 = None
        convert_element_type_428 = torch.ops.prims.convert_element_type.default(add_177, torch.float32);  add_177 = None
        sub_43 = torch.ops.aten.sub.Tensor(convert_element_type_428, getitem_297);  convert_element_type_428 = getitem_297 = None
        mul_431 = torch.ops.aten.mul.Tensor(sub_43, rsqrt_54);  sub_43 = None
        convert_element_type_432 = torch.ops.prims.convert_element_type.default(getitem_299, torch.float32)
        sigmoid_11 = torch.ops.aten.sigmoid.default(convert_element_type_432)
        mul_434 = torch.ops.aten.mul.Tensor(convert_element_type_432, sigmoid_11);  convert_element_type_432 = sigmoid_11 = None
        convert_element_type_433 = torch.ops.prims.convert_element_type.default(mul_434, torch.float16);  mul_434 = None
        convert_element_type_436 = torch.ops.prims.convert_element_type.default(add_180, torch.float32);  add_180 = None
        sub_44 = torch.ops.aten.sub.Tensor(convert_element_type_436, getitem_301);  convert_element_type_436 = getitem_301 = None
        mul_438 = torch.ops.aten.mul.Tensor(sub_44, rsqrt_55);  sub_44 = None
        convert_element_type_444 = torch.ops.prims.convert_element_type.default(add_185, torch.float32);  add_185 = None
        sub_45 = torch.ops.aten.sub.Tensor(convert_element_type_444, getitem_316);  convert_element_type_444 = getitem_316 = None
        mul_452 = torch.ops.aten.mul.Tensor(sub_45, rsqrt_56);  sub_45 = None
        convert_element_type_450 = torch.ops.prims.convert_element_type.default(add_188, torch.float32);  add_188 = None
        sub_46 = torch.ops.aten.sub.Tensor(convert_element_type_450, getitem_327);  convert_element_type_450 = getitem_327 = None
        mul_458 = torch.ops.aten.mul.Tensor(sub_46, rsqrt_57);  sub_46 = None
        convert_element_type_454 = torch.ops.prims.convert_element_type.default(getitem_329, torch.float32)
        sigmoid_12 = torch.ops.aten.sigmoid.default(convert_element_type_454)
        mul_461 = torch.ops.aten.mul.Tensor(convert_element_type_454, sigmoid_12);  convert_element_type_454 = sigmoid_12 = None
        convert_element_type_455 = torch.ops.prims.convert_element_type.default(mul_461, torch.float16);  mul_461 = None
        convert_element_type_458 = torch.ops.prims.convert_element_type.default(add_191, torch.float32);  add_191 = None
        sub_47 = torch.ops.aten.sub.Tensor(convert_element_type_458, getitem_331);  convert_element_type_458 = getitem_331 = None
        mul_465 = torch.ops.aten.mul.Tensor(sub_47, rsqrt_58);  sub_47 = None
        convert_element_type_466 = torch.ops.prims.convert_element_type.default(add_196, torch.float32);  add_196 = None
        sub_48 = torch.ops.aten.sub.Tensor(convert_element_type_466, getitem_346);  convert_element_type_466 = getitem_346 = None
        mul_479 = torch.ops.aten.mul.Tensor(sub_48, rsqrt_59);  sub_48 = None
        convert_element_type_472 = torch.ops.prims.convert_element_type.default(add_199, torch.float32);  add_199 = None
        sub_49 = torch.ops.aten.sub.Tensor(convert_element_type_472, getitem_357);  convert_element_type_472 = getitem_357 = None
        mul_485 = torch.ops.aten.mul.Tensor(sub_49, rsqrt_60);  sub_49 = None
        convert_element_type_476 = torch.ops.prims.convert_element_type.default(getitem_359, torch.float32)
        sigmoid_13 = torch.ops.aten.sigmoid.default(convert_element_type_476)
        mul_488 = torch.ops.aten.mul.Tensor(convert_element_type_476, sigmoid_13);  convert_element_type_476 = sigmoid_13 = None
        convert_element_type_477 = torch.ops.prims.convert_element_type.default(mul_488, torch.float16);  mul_488 = None
        convert_element_type_480 = torch.ops.prims.convert_element_type.default(add_202, torch.float32);  add_202 = None
        sub_50 = torch.ops.aten.sub.Tensor(convert_element_type_480, getitem_361);  convert_element_type_480 = getitem_361 = None
        mul_492 = torch.ops.aten.mul.Tensor(sub_50, rsqrt_61);  sub_50 = None
        convert_element_type_488 = torch.ops.prims.convert_element_type.default(add_207, torch.float32);  add_207 = None
        sub_51 = torch.ops.aten.sub.Tensor(convert_element_type_488, getitem_376);  convert_element_type_488 = getitem_376 = None
        mul_506 = torch.ops.aten.mul.Tensor(sub_51, rsqrt_62);  sub_51 = None
        convert_element_type_494 = torch.ops.prims.convert_element_type.default(add_210, torch.float32);  add_210 = None
        sub_52 = torch.ops.aten.sub.Tensor(convert_element_type_494, getitem_387);  convert_element_type_494 = getitem_387 = None
        mul_512 = torch.ops.aten.mul.Tensor(sub_52, rsqrt_63);  sub_52 = None
        convert_element_type_498 = torch.ops.prims.convert_element_type.default(getitem_389, torch.float32)
        sigmoid_14 = torch.ops.aten.sigmoid.default(convert_element_type_498)
        mul_515 = torch.ops.aten.mul.Tensor(convert_element_type_498, sigmoid_14);  convert_element_type_498 = sigmoid_14 = None
        convert_element_type_499 = torch.ops.prims.convert_element_type.default(mul_515, torch.float16);  mul_515 = None
        convert_element_type_502 = torch.ops.prims.convert_element_type.default(add_213, torch.float32);  add_213 = None
        sub_53 = torch.ops.aten.sub.Tensor(convert_element_type_502, getitem_391);  convert_element_type_502 = getitem_391 = None
        mul_519 = torch.ops.aten.mul.Tensor(sub_53, rsqrt_64);  sub_53 = None
        convert_element_type_510 = torch.ops.prims.convert_element_type.default(add_218, torch.float32);  add_218 = None
        sub_54 = torch.ops.aten.sub.Tensor(convert_element_type_510, getitem_406);  convert_element_type_510 = getitem_406 = None
        mul_533 = torch.ops.aten.mul.Tensor(sub_54, rsqrt_65);  sub_54 = None
        convert_element_type_516 = torch.ops.prims.convert_element_type.default(add_221, torch.float32);  add_221 = None
        sub_55 = torch.ops.aten.sub.Tensor(convert_element_type_516, getitem_417);  convert_element_type_516 = getitem_417 = None
        mul_539 = torch.ops.aten.mul.Tensor(sub_55, rsqrt_66);  sub_55 = None
        convert_element_type_520 = torch.ops.prims.convert_element_type.default(getitem_419, torch.float32)
        sigmoid_15 = torch.ops.aten.sigmoid.default(convert_element_type_520)
        mul_542 = torch.ops.aten.mul.Tensor(convert_element_type_520, sigmoid_15);  convert_element_type_520 = sigmoid_15 = None
        convert_element_type_521 = torch.ops.prims.convert_element_type.default(mul_542, torch.float16);  mul_542 = None
        convert_element_type_524 = torch.ops.prims.convert_element_type.default(add_224, torch.float32);  add_224 = None
        sub_56 = torch.ops.aten.sub.Tensor(convert_element_type_524, getitem_421);  convert_element_type_524 = getitem_421 = None
        mul_546 = torch.ops.aten.mul.Tensor(sub_56, rsqrt_67);  sub_56 = None
        convert_element_type_532 = torch.ops.prims.convert_element_type.default(add_229, torch.float32);  add_229 = None
        sub_57 = torch.ops.aten.sub.Tensor(convert_element_type_532, getitem_436);  convert_element_type_532 = getitem_436 = None
        mul_560 = torch.ops.aten.mul.Tensor(sub_57, rsqrt_68);  sub_57 = None
        convert_element_type_538 = torch.ops.prims.convert_element_type.default(add_232, torch.float32);  add_232 = None
        sub_58 = torch.ops.aten.sub.Tensor(convert_element_type_538, getitem_447);  convert_element_type_538 = getitem_447 = None
        mul_566 = torch.ops.aten.mul.Tensor(sub_58, rsqrt_69);  sub_58 = None
        convert_element_type_542 = torch.ops.prims.convert_element_type.default(getitem_449, torch.float32)
        sigmoid_16 = torch.ops.aten.sigmoid.default(convert_element_type_542)
        mul_569 = torch.ops.aten.mul.Tensor(convert_element_type_542, sigmoid_16);  convert_element_type_542 = sigmoid_16 = None
        convert_element_type_543 = torch.ops.prims.convert_element_type.default(mul_569, torch.float16);  mul_569 = None
        convert_element_type_546 = torch.ops.prims.convert_element_type.default(add_235, torch.float32);  add_235 = None
        sub_59 = torch.ops.aten.sub.Tensor(convert_element_type_546, getitem_451);  convert_element_type_546 = getitem_451 = None
        mul_573 = torch.ops.aten.mul.Tensor(sub_59, rsqrt_70);  sub_59 = None
        convert_element_type_554 = torch.ops.prims.convert_element_type.default(add_240, torch.float32);  add_240 = None
        sub_60 = torch.ops.aten.sub.Tensor(convert_element_type_554, getitem_466);  convert_element_type_554 = getitem_466 = None
        mul_587 = torch.ops.aten.mul.Tensor(sub_60, rsqrt_71);  sub_60 = None
        convert_element_type_560 = torch.ops.prims.convert_element_type.default(add_243, torch.float32);  add_243 = None
        sub_61 = torch.ops.aten.sub.Tensor(convert_element_type_560, getitem_477);  convert_element_type_560 = getitem_477 = None
        mul_593 = torch.ops.aten.mul.Tensor(sub_61, rsqrt_72);  sub_61 = None
        convert_element_type_564 = torch.ops.prims.convert_element_type.default(getitem_479, torch.float32)
        sigmoid_17 = torch.ops.aten.sigmoid.default(convert_element_type_564)
        mul_596 = torch.ops.aten.mul.Tensor(convert_element_type_564, sigmoid_17);  convert_element_type_564 = sigmoid_17 = None
        convert_element_type_565 = torch.ops.prims.convert_element_type.default(mul_596, torch.float16);  mul_596 = None
        convert_element_type_568 = torch.ops.prims.convert_element_type.default(add_246, torch.float32);  add_246 = None
        sub_62 = torch.ops.aten.sub.Tensor(convert_element_type_568, getitem_481);  convert_element_type_568 = getitem_481 = None
        mul_600 = torch.ops.aten.mul.Tensor(sub_62, rsqrt_73);  sub_62 = None
        convolution_backward = torch.ops.aten.convolution_backward.default(tangents_1, slice_263, convert_element_type_572, [0], [1], [0], [1], False, [0], 1, [True, True, False]);  slice_263 = convert_element_type_572 = None
        getitem_482 = convolution_backward[0]
        getitem_483 = convolution_backward[1];  convolution_backward = None
        add_250 = torch.ops.aten.add.Tensor(tangents_1, getitem_482);  tangents_1 = getitem_482 = None
        convert_element_type_573 = torch.ops.prims.convert_element_type.default(getitem_483, torch.float32);  getitem_483 = None
        full_3 = torch.ops.aten.full.default([sym_size_14, 32, 513], 0, dtype = torch.float16, layout = torch.strided, device = device(type='cuda', index=0), pin_memory = False);  sym_size_14 = None
        slice_scatter = torch.ops.aten.slice_scatter.default(full_3, add_250, 2, 1, 9223372036854775807);  add_250 = None
        slice_scatter_1 = torch.ops.aten.slice_scatter.default(full_3, slice_scatter, 1, 0, 9223372036854775807);  full_3 = slice_scatter = None
        full_5 = torch.ops.aten.full.default([sym_size, 32, 513], 0, dtype = torch.float16, layout = torch.strided, device = device(type='cuda', index=0), pin_memory = False)
        slice_scatter_2 = torch.ops.aten.slice_scatter.default(full_5, slice_scatter_1, 0, 0, 9223372036854775807);  full_5 = slice_scatter_1 = None
        permute_434 = torch.ops.aten.permute.default(slice_scatter_2, [0, 2, 1]);  slice_scatter_2 = None
        clone_116 = torch.ops.aten.clone.default(permute_434, memory_format = torch.contiguous_format);  permute_434 = None
        sym_size_15 = torch.ops.aten.sym_size(view_778, 0)
        view_780 = torch.ops.aten.view.default(clone_116, [sym_size_15, 32]);  clone_116 = sym_size_15 = None
        mm_203 = torch.ops.aten.mm.default(view_780, permute_435);  permute_435 = None
        permute_436 = torch.ops.aten.permute.default(view_780, [1, 0])
        mm_204 = torch.ops.aten.mm.default(permute_436, view_778);  permute_436 = view_778 = None
        permute_437 = torch.ops.aten.permute.default(mm_204, [1, 0]);  mm_204 = None
        sum_13 = torch.ops.aten.sum.dim_IntList(view_780, [0], True, dtype = torch.float32);  view_780 = None
        view_781 = torch.ops.aten.view.default(sum_13, [32]);  sum_13 = None
        permute_438 = torch.ops.aten.permute.default(permute_437, [1, 0]);  permute_437 = None
        view_782 = torch.ops.aten.view.default(mm_203, [sym_size, 513, 1024]);  mm_203 = None
        convert_element_type_575 = torch.ops.prims.convert_element_type.default(view_782, torch.float32);  view_782 = None
        convert_element_type_576 = torch.ops.prims.convert_element_type.default(permute_438, torch.float32);  permute_438 = None
        convert_element_type_default_38 = torch.ops.prims.convert_element_type.default(view_781, torch.float32);  view_781 = None
        mul_604 = torch.ops.aten.mul.Tensor(convert_element_type_575, primals_404);  primals_404 = None
        mul_605 = torch.ops.aten.mul.Tensor(mul_604, 1024)
        sum_14 = torch.ops.aten.sum.dim_IntList(mul_604, [2], True)
        mul_606 = torch.ops.aten.mul.Tensor(mul_604, mul_600);  mul_604 = None
        sum_15 = torch.ops.aten.sum.dim_IntList(mul_606, [2], True);  mul_606 = None
        mul_607 = torch.ops.aten.mul.Tensor(mul_600, sum_15);  sum_15 = None
        sub_64 = torch.ops.aten.sub.Tensor(mul_605, sum_14);  mul_605 = sum_14 = None
        sub_65 = torch.ops.aten.sub.Tensor(sub_64, mul_607);  sub_64 = mul_607 = None
        div_15 = torch.ops.aten.div.Tensor(rsqrt_73, 1024);  rsqrt_73 = None
        mul_608 = torch.ops.aten.mul.Tensor(div_15, sub_65);  div_15 = sub_65 = None
        mul_609 = torch.ops.aten.mul.Tensor(convert_element_type_575, mul_600);  mul_600 = None
        sum_16 = torch.ops.aten.sum.dim_IntList(mul_609, [0, 1]);  mul_609 = None
        sum_17 = torch.ops.aten.sum.dim_IntList(convert_element_type_575, [0, 1]);  convert_element_type_575 = None
        convert_element_type_578 = torch.ops.prims.convert_element_type.default(mul_608, torch.float16);  mul_608 = None
        sym_size_16 = torch.ops.aten.sym_size(view_776, 0)
        view_783 = torch.ops.aten.view.default(convert_element_type_578, [sym_size_16, 1024]);  sym_size_16 = None
        mm_205 = torch.ops.aten.mm.default(view_783, permute_439);  permute_439 = None
        permute_440 = torch.ops.aten.permute.default(view_783, [1, 0])
        mm_206 = torch.ops.aten.mm.default(permute_440, view_776);  permute_440 = view_776 = None
        permute_441 = torch.ops.aten.permute.default(mm_206, [1, 0]);  mm_206 = None
        sum_18 = torch.ops.aten.sum.dim_IntList(view_783, [0], True, dtype = torch.float32);  view_783 = None
        view_784 = torch.ops.aten.view.default(sum_18, [1024]);  sum_18 = None
        permute_442 = torch.ops.aten.permute.default(permute_441, [1, 0]);  permute_441 = None
        view_785 = torch.ops.aten.view.default(mm_205, [sym_size, 513, 4096]);  mm_205 = None
        convert_element_type_580 = torch.ops.prims.convert_element_type.default(permute_442, torch.float32);  permute_442 = None
        convert_element_type_default_37 = torch.ops.prims.convert_element_type.default(view_784, torch.float32);  view_784 = None
        mul_610 = torch.ops.aten.mul.Tensor(view_785, 1.0);  view_785 = None
        mul_611 = torch.ops.aten.mul.Tensor(mul_610, getitem_478);  getitem_478 = None
        mul_612 = torch.ops.aten.mul.Tensor(mul_610, convert_element_type_565);  mul_610 = convert_element_type_565 = None
        sigmoid_18 = torch.ops.aten.sigmoid.default(getitem_479)
        full_6 = torch.ops.aten.full.default([sym_size, 513, 4096], 1, dtype = torch.float16, layout = torch.strided, device = device(type='cuda', index=0), pin_memory = False)
        sub_66 = torch.ops.aten.sub.Tensor(full_6, sigmoid_18)
        mul_613 = torch.ops.aten.mul.Tensor(getitem_479, sub_66);  getitem_479 = sub_66 = None
        add_251 = torch.ops.aten.add.Scalar(mul_613, 1);  mul_613 = None
        mul_614 = torch.ops.aten.mul.Tensor(sigmoid_18, add_251);  sigmoid_18 = add_251 = None
        mul_615 = torch.ops.aten.mul.Tensor(mul_611, mul_614);  mul_611 = mul_614 = None
        cat_69 = torch.ops.aten.cat.default([mul_612, mul_615], 2);  mul_612 = mul_615 = None
        sym_size_17 = torch.ops.aten.sym_size(view_774, 0)
        view_786 = torch.ops.aten.view.default(cat_69, [sym_size_17, 8192]);  cat_69 = sym_size_17 = None
        print("any inf: " + str(torch.any(torch.isinf(view_786))))
        print("max value: " + str(torch.max(view_786)))
        mm_207 = torch.ops.aten.mm.default(view_786, permute_444);  permute_444 = None
        print("any inf: " + str(torch.any(torch.isinf(mm_207))))
        return [mm_207]

def load_args(reader):
    reader.symint(2)  # primals_410
    reader.symint(2)  # sym_size
    reader.symint(2)  # sym_size_1
    reader.symint(2)  # sym_size_14
    buf0 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf0, (1024,), requires_grad=True, is_leaf=True)  # primals_116
    buf1 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf1, (1024,), requires_grad=True, is_leaf=True)  # primals_122
    buf2 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf2, (1024,), requires_grad=True, is_leaf=True)  # primals_128
    buf3 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf3, (1024,), requires_grad=True, is_leaf=True)  # primals_134
    buf4 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf4, (1024,), requires_grad=True, is_leaf=True)  # primals_140
    buf5 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf5, (1024,), requires_grad=True, is_leaf=True)  # primals_146
    buf6 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf6, (1024,), requires_grad=True, is_leaf=True)  # primals_152
    buf7 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf7, (1024,), requires_grad=True, is_leaf=True)  # primals_158
    buf8 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf8, (1024,), requires_grad=True, is_leaf=True)  # primals_164
    buf9 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf9, (1024,), requires_grad=True, is_leaf=True)  # primals_170
    buf10 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf10, (1024,), requires_grad=True, is_leaf=True)  # primals_176
    buf11 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf11, (1024,), requires_grad=True, is_leaf=True)  # primals_182
    buf12 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf12, (1024,), requires_grad=True, is_leaf=True)  # primals_188
    buf13 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf13, (1024,), requires_grad=True, is_leaf=True)  # primals_194
    buf14 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf14, (1024,), requires_grad=True, is_leaf=True)  # primals_200
    buf15 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf15, (1024,), requires_grad=True, is_leaf=True)  # primals_206
    buf16 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf16, (1024,), requires_grad=True, is_leaf=True)  # primals_212
    buf17 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf17, (1024,), requires_grad=True, is_leaf=True)  # primals_218
    buf18 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf18, (1024,), requires_grad=True, is_leaf=True)  # primals_224
    buf19 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf19, (1024,), requires_grad=True, is_leaf=True)  # primals_230
    buf20 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf20, (1024,), requires_grad=True, is_leaf=True)  # primals_236
    buf21 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf21, (1024,), requires_grad=True, is_leaf=True)  # primals_242
    buf22 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf22, (1024,), requires_grad=True, is_leaf=True)  # primals_248
    buf23 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf23, (1024,), requires_grad=True, is_leaf=True)  # primals_254
    buf24 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf24, (1024,), requires_grad=True, is_leaf=True)  # primals_260
    buf25 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf25, (1024,), requires_grad=True, is_leaf=True)  # primals_266
    buf26 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf26, (1024,), requires_grad=True, is_leaf=True)  # primals_272
    buf27 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf27, (1024,), requires_grad=True, is_leaf=True)  # primals_278
    buf28 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf28, (1024,), requires_grad=True, is_leaf=True)  # primals_284
    buf29 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf29, (1024,), requires_grad=True, is_leaf=True)  # primals_290
    buf30 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf30, (1024,), requires_grad=True, is_leaf=True)  # primals_296
    buf31 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf31, (1024,), requires_grad=True, is_leaf=True)  # primals_302
    buf32 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf32, (1024,), requires_grad=True, is_leaf=True)  # primals_308
    buf33 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf33, (1024,), requires_grad=True, is_leaf=True)  # primals_314
    buf34 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf34, (1024,), requires_grad=True, is_leaf=True)  # primals_320
    buf35 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf35, (1024,), requires_grad=True, is_leaf=True)  # primals_326
    buf36 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf36, (1024,), requires_grad=True, is_leaf=True)  # primals_332
    buf37 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf37, (1024,), requires_grad=True, is_leaf=True)  # primals_338
    buf38 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf38, (1024,), requires_grad=True, is_leaf=True)  # primals_344
    buf39 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf39, (1024,), requires_grad=True, is_leaf=True)  # primals_350
    buf40 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf40, (1024,), requires_grad=True, is_leaf=True)  # primals_356
    buf41 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf41, (1024,), requires_grad=True, is_leaf=True)  # primals_362
    buf42 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf42, (1024,), requires_grad=True, is_leaf=True)  # primals_368
    buf43 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf43, (1024,), requires_grad=True, is_leaf=True)  # primals_374
    buf44 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf44, (1024,), requires_grad=True, is_leaf=True)  # primals_380
    buf45 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf45, (1024,), requires_grad=True, is_leaf=True)  # primals_386
    buf46 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf46, (1024,), requires_grad=True, is_leaf=True)  # primals_392
    buf47 = reader.storage('00902d755c30625138a13e9a1a34aed94ec0283d', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf47, (1024,), requires_grad=True, is_leaf=True)  # primals_398
    buf48 = reader.storage('25e419eaf593b0fdafe540b8b6d38a44d7d19442', 4096, device=device(type='cuda', index=0))
    reader.tensor(buf48, (1024,), requires_grad=True, is_leaf=True)  # primals_404
    buf49 = reader.storage('dc8b59b130ea2921436fd4079b10294f9062e8dc', 1024, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf49, (2, 64), dtype=torch.int64, is_leaf=True)  # primals_415
    buf50 = reader.storage('55e8d0e974b197405e6eefc5a30e14b3f77f5b67', 8, device=device(type='cuda', index=0))
    reader.tensor(buf50, (2, 1), is_leaf=True)  # view
    buf51 = reader.storage('9c43329cf6f78bed26784ab49685a903e6987492', 1024, device=device(type='cuda', index=0))
    reader.tensor(buf51, (2, 128), is_leaf=True)  # sin
    buf52 = reader.storage('854a89f08725f03446987f1a68a6abe130f4ed85', 1024, device=device(type='cuda', index=0))
    reader.tensor(buf52, (2, 128), is_leaf=True)  # cos
    buf53 = reader.storage('ad92e1929c8965a97672aafb3a0f434a29f2542b', 1028, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf53, (2, 257), dtype=torch.float16, is_leaf=True)  # convert_element_type_2
    buf54 = reader.storage('4c793b8ceaa49379e4f30213eb5ed0a4a88e48dc', 3072, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf54, (2, 768), dtype=torch.float16, is_leaf=True)  # addmm
    buf55 = reader.storage('3d2f28ccca4e52c699d4c112ee29a60605e36f58', 3072, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf55, (2, 768), dtype=torch.float16, is_leaf=True)  # convert_element_type_4
    buf56 = reader.storage('37ca8120d383629bf58bff85bcc3b7bb1039a934', 3072, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf56, (2, 768), dtype=torch.float16, is_leaf=True)  # addmm_1
    buf57 = reader.storage('fee32d84bc17f0f41a264eb8f2951b5c1a044d7d', 3072, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf57, (2, 768), dtype=torch.float16, is_leaf=True)  # convert_element_type_8
    buf58 = reader.storage('6307b3c0b75e9399d10f76bdc62e387f1eecb518', 3072, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf58, (2, 768), dtype=torch.float16, is_leaf=True)  # addmm_2
    buf59 = reader.storage('a52d97e6e57bbefb7edabadffdeb7cddff2b33be', 3072, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf59, (2, 768), dtype=torch.float16, is_leaf=True)  # convert_element_type_12
    buf60 = reader.storage('9b333cd7f2c5fc90bb73853268f7b0b3a1bcd8dc', 1024, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf60, (128,), dtype=torch.int64, is_leaf=True)  # iota_2
    buf61 = reader.storage('66373706417e24633a31ac4e06738d099d50b450', 2, device=device(type='cuda', index=0), dtype_hint=torch.bool)
    reader.tensor(buf61, (2, 1, 1), dtype=torch.bool, is_leaf=True)  # lt_1
    buf62 = reader.storage('2202682558f96a882239e49db25925415622060c', 393216, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf62, (256, 768), dtype=torch.float16, is_leaf=True)  # view_292
    buf63 = reader.storage('160aea8193b046d72a73f9c8214474f9dbf4cc87', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf63, (256, 1024), dtype=torch.float16, is_leaf=True)  # mm_72
    buf64 = reader.storage('2fb8558ea8370080ea4a008398965239a2985a4f', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf64, (256, 1024), dtype=torch.float16, is_leaf=True)  # view_294
    buf65 = reader.storage('fd6a8fd234b312b48eb48db35d08db783ee8d4ff', 4096, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf65, (2, 1024), dtype=torch.float16, is_leaf=True)  # addmm_3
    buf66 = reader.storage('d488bd3781c0d5d29c46564c77cca6f8feca99e0', 4096, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf66, (2, 1024), dtype=torch.float16, is_leaf=True)  # convert_element_type_205
    buf67 = reader.storage('ed9d62b445ecb1e32399a090ffc887a2e134200b', 2048, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf67, (32, 32, 1), dtype=torch.float16, is_leaf=True)  # convert_element_type_208
    buf68 = reader.storage('327852b0949133965266dcf7a1d294ede1703095', 65536, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf68, (2, 32, 512), dtype=torch.float16, is_leaf=True)  # convert_element_type_209
    buf69 = reader.storage('a64d5ded7052b8e3a8a26817faf1602fa1493721', 65536, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf69, (1024, 32), dtype=torch.float16, is_leaf=True)  # view_296
    buf70 = reader.storage('4359caa64437e0803fcf9a5750788f5cb8ed204c', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf70, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # clone_99
    buf71 = reader.storage('30736c9cff503fffd0758c5b9d188617406ccd22', 32832, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf71, (513, 32), dtype=torch.float16, is_leaf=True)  # cat_4
    buf72 = reader.storage('937d253d3536014be1847ddd65288a9b17ef6a4b', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf72, (2, 513, 1), is_leaf=True)  # getitem_1
    buf73 = reader.storage('f2cefecc90a070c074159182f946596fb7b6eb29', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf73, (2, 513, 1), is_leaf=True)  # rsqrt_25
    buf74 = reader.storage('968f812a635aebd24312ddccda26fcf3ddffab9a', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf74, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_298
    buf75 = reader.storage('f3c8870d91b1eec8cd9dc0ff275ce140b039d1a7', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf75, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_149
    buf76 = reader.storage('4fd46df994ad3bb0babd689b79e295bd127a652b', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf76, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # cat_6
    buf77 = reader.storage('ceb1aea3f278a97275c7f44946d5c17f330097d5', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf77, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # cat_8
    buf78 = reader.storage('8d0adec2ba11fb3c179c0ab5df39376936ea941a', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf78, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # getitem_6
    buf79 = reader.storage('a87da01de756062e2c1c6dc8ea39f37f4df6287e', 67584, device=device(type='cuda', index=0))
    reader.tensor(buf79, (2, 16, 528), is_leaf=True)  # getitem_7
    buf80 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf80, (3,), dtype=torch.int32, is_leaf=True)  # getitem_8
    buf81 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf81, (3,), dtype=torch.int32, is_leaf=True)  # getitem_9
    buf82 = reader.storage('6daffc41d3ab8fe90bb0a7747066dbdd476dc2fe', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf82, (), dtype=torch.int64, is_leaf=True)  # getitem_12
    buf83 = reader.storage('7650095f0f29a31e731dd85f663178c17665cbf1', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf83, (), dtype=torch.int64, is_leaf=True)  # getitem_13
    reader.tensor(buf78, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_310
    buf84 = reader.storage('33b0b89f6176bf7154be0aeca0d54fd2ec6fbd17', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf84, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_75
    buf85 = reader.storage('6eaa673fd5452d8a907172144e9adf0ccb218d3a', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf85, (2, 513, 1), is_leaf=True)  # getitem_16
    buf86 = reader.storage('0f0997333c41627d0e92f5f46d9175f8a0a33893', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf86, (2, 513, 1), is_leaf=True)  # rsqrt_26
    buf87 = reader.storage('719588c17c81dc9fcabfb7ea7d651168eeaca035', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf87, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_312
    buf88 = reader.storage('a1a91eca264bb5336cf129df9501f47290ee7f73', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf88, (256, 1024), dtype=torch.float16, is_leaf=True)  # view_314
    buf89 = reader.storage('72acefd78c49cf3f402e05391499ec352a5d1728', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf89, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_155
    buf90 = reader.storage('2f7ee829d1f116aab436df4109f8602879c8b87f', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf90, (2, 16, 128, 64), (131072, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_156
    buf91 = reader.storage('ec9e3d206912149cb06581e5f92bf68fc573e28e', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf91, (2, 16, 128, 64), (131072, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_157
    buf92 = reader.storage('8f8c4db44b1e521d1f599da235a4c213426d20cb', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf92, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # getitem_17
    buf93 = reader.storage('84af1b7bbcbd3df9043f159cffaae637ca9ffde9', 67584, device=device(type='cuda', index=0))
    reader.tensor(buf93, (2, 16, 528), is_leaf=True)  # getitem_18
    buf94 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf94, (3,), dtype=torch.int32, is_leaf=True)  # getitem_19
    buf95 = reader.storage('1cbaf6f990ca04e947eaf166e1fbcbd05ce76bf1', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf95, (3,), dtype=torch.int32, is_leaf=True)  # getitem_20
    buf96 = reader.storage('4c7dba836b8213f91f5f23d0c4553137590fa16e', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf96, (), dtype=torch.int64, is_leaf=True)  # getitem_23
    buf97 = reader.storage('86ea2f52438c0026d22bc25b58ee7d54e72d57a1', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf97, (), dtype=torch.int64, is_leaf=True)  # getitem_24
    reader.tensor(buf92, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_322
    buf98 = reader.storage('aea28e2503901e5913aaf701776021ea5109b89a', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf98, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_78
    buf99 = reader.storage('f3c71eac9010220deacaf8451a6626527b435526', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf99, (2, 513, 1), is_leaf=True)  # getitem_27
    buf100 = reader.storage('c8134010703c3b2f84fd3a241dcac8ade619b49c', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf100, (2, 513, 1), is_leaf=True)  # rsqrt_27
    buf101 = reader.storage('b3882e3e9299cd47697dee9f612f4ea4b0a39552', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf101, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_324
    buf102 = reader.storage('d5979e782596d8dbaa6b2302664973d5b85c6e62', 16809984, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf102, (2, 513, 4096), (4202496, 8192, 1), dtype=torch.float16, is_leaf=True)  # getitem_28
    reader.tensor(buf102, (2, 513, 4096), (4202496, 8192, 1), dtype=torch.float16, storage_offset=4096, is_leaf=True)  # getitem_29
    buf103 = reader.storage('3984ddccdf300382cf9e645820685bc1e52ab123', 8404992, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf103, (1026, 4096), dtype=torch.float16, is_leaf=True)  # view_326
    buf104 = reader.storage('cd9ec9336b2e5589d417b7e101056b050aed6c28', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf104, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_81
    buf105 = reader.storage('efbe436200eabb73e1f28758bb885bf3e57f1fb9', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf105, (2, 513, 1), is_leaf=True)  # getitem_31
    buf106 = reader.storage('62cc9dbd7ffd52252167c458fbc6c8ca9fbd11de', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf106, (2, 513, 1), is_leaf=True)  # rsqrt_28
    buf107 = reader.storage('038dfd3594e25271dca0de69563aeae6acaa3be2', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf107, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_328
    buf108 = reader.storage('dd2292013123206d7963f4c5fda2ac4a6dfb1e44', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf108, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_167
    buf109 = reader.storage('b35c6a7adb2ee47aad5a6906e6c8d35c05d02ebb', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf109, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # cat_10
    buf110 = reader.storage('15f34c81e355f4429797ddb5c49fed52629527f4', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf110, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # cat_12
    buf111 = reader.storage('dca1ffd62839137878bbb8ced3ec48d0b3512edb', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf111, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # getitem_36
    buf112 = reader.storage('7b5776a803b499b080eb14213dfef6544a6d1165', 67584, device=device(type='cuda', index=0))
    reader.tensor(buf112, (2, 16, 528), is_leaf=True)  # getitem_37
    buf113 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf113, (3,), dtype=torch.int32, is_leaf=True)  # getitem_38
    buf114 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf114, (3,), dtype=torch.int32, is_leaf=True)  # getitem_39
    buf115 = reader.storage('a0b5ddd4089c82efe9a583f6c19274e0a10ec258', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf115, (), dtype=torch.int64, is_leaf=True)  # getitem_42
    buf116 = reader.storage('e6b0c031ae8ba78da6633870adf1729513c6f512', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf116, (), dtype=torch.int64, is_leaf=True)  # getitem_43
    reader.tensor(buf111, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_340
    buf117 = reader.storage('6d83358efc76695d5f5c347de20240d15c28b78c', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf117, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_86
    buf118 = reader.storage('99a75bde705dba4ff2e070815978be461689eb6a', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf118, (2, 513, 1), is_leaf=True)  # getitem_46
    buf119 = reader.storage('f5915ff27fbe1209cf390fc69b70a1cbfb0e4dd9', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf119, (2, 513, 1), is_leaf=True)  # rsqrt_29
    buf120 = reader.storage('85a80be4619fdadf9bd1b7747efa9e1f0a938e89', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf120, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_342
    buf121 = reader.storage('7e862050f3969dc638d8cb7d57becd51096fcd71', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf121, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_173
    buf122 = reader.storage('6d7d9127a70d06d7fa6ec759cbc2c6441c88f3f4', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf122, (2, 16, 128, 64), (131072, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_174
    buf123 = reader.storage('b98083e7b4cfc318d96b7aa1fc680cd4891a9325', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf123, (2, 16, 128, 64), (131072, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_175
    buf124 = reader.storage('dfe0a181774b0a5ec8432dc7bf761b557fc560b4', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf124, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # getitem_47
    buf125 = reader.storage('9d27f74e0c646080278d994581660b66563c7756', 67584, device=device(type='cuda', index=0))
    reader.tensor(buf125, (2, 16, 528), is_leaf=True)  # getitem_48
    buf126 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf126, (3,), dtype=torch.int32, is_leaf=True)  # getitem_49
    buf127 = reader.storage('1cbaf6f990ca04e947eaf166e1fbcbd05ce76bf1', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf127, (3,), dtype=torch.int32, is_leaf=True)  # getitem_50
    buf128 = reader.storage('45cafbdc5e2b8caffb227196433e3d88c6768168', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf128, (), dtype=torch.int64, is_leaf=True)  # getitem_53
    buf129 = reader.storage('c7b03f4b0af87d4a2e3a9d83447b6ff5ce3994fd', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf129, (), dtype=torch.int64, is_leaf=True)  # getitem_54
    reader.tensor(buf124, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_352
    buf130 = reader.storage('c9b5801873ebd7092a1f80c56ec4e899b8616e8a', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf130, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_89
    buf131 = reader.storage('c93abdbae19088bce70b761870390d2b5e306489', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf131, (2, 513, 1), is_leaf=True)  # getitem_57
    buf132 = reader.storage('68acc2a82bfe3ae52642e2273957e4a54f95b043', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf132, (2, 513, 1), is_leaf=True)  # rsqrt_30
    buf133 = reader.storage('76cda6ae69c39f049898589f4b29ed101ed7ba4f', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf133, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_354
    buf134 = reader.storage('3b936981cb40725ae6a2eec7a8def7070578a4f2', 16809984, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf134, (2, 513, 4096), (4202496, 8192, 1), dtype=torch.float16, is_leaf=True)  # getitem_58
    reader.tensor(buf134, (2, 513, 4096), (4202496, 8192, 1), dtype=torch.float16, storage_offset=4096, is_leaf=True)  # getitem_59
    buf135 = reader.storage('2c266f05b82644782cfcb7c28b5104a2f69930c6', 8404992, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf135, (1026, 4096), dtype=torch.float16, is_leaf=True)  # view_356
    buf136 = reader.storage('054d4cb849d23ddc28d930eca287144240dd355b', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf136, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_92
    buf137 = reader.storage('a7189cf372984ef31307c4a7e01a884fad32e938', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf137, (2, 513, 1), is_leaf=True)  # getitem_61
    buf138 = reader.storage('b7284c82a2ed47cd1dcc77732ca0f10a74b385be', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf138, (2, 513, 1), is_leaf=True)  # rsqrt_31
    buf139 = reader.storage('1abecae4eacec322ccf09f6eaf894ffdee086d77', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf139, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_358
    buf140 = reader.storage('b1518e4f025b519cc011b4d487dd3273a939cfb8', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf140, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_185
    buf141 = reader.storage('0aa8db6b3212a5f052244ef1e4760d2baa33abda', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf141, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # cat_14
    buf142 = reader.storage('1784c8e04d3b7ffb6eb66a505195cc01581764f2', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf142, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # cat_16
    buf143 = reader.storage('bd8f1aa2489fd10a2c13abdce3ec2e4d48a76c90', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf143, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # getitem_66
    buf144 = reader.storage('c4456b33d3595c0d345db47ce33d465d28499130', 67584, device=device(type='cuda', index=0))
    reader.tensor(buf144, (2, 16, 528), is_leaf=True)  # getitem_67
    buf145 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf145, (3,), dtype=torch.int32, is_leaf=True)  # getitem_68
    buf146 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf146, (3,), dtype=torch.int32, is_leaf=True)  # getitem_69
    buf147 = reader.storage('a97801e75ae38f449002eda5544431cd8d61314b', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf147, (), dtype=torch.int64, is_leaf=True)  # getitem_72
    buf148 = reader.storage('0bcdf15c62a17d7e14489b572fff10b6ae107859', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf148, (), dtype=torch.int64, is_leaf=True)  # getitem_73
    reader.tensor(buf143, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_370
    buf149 = reader.storage('455c043ba9c671c023fdb67ab0cb60768d1f7e63', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf149, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_97
    buf150 = reader.storage('f623318150899994128d82ff5e1b877e2d571802', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf150, (2, 513, 1), is_leaf=True)  # getitem_76
    buf151 = reader.storage('19b4e65f94fc1b36b86423ae2d4fc90c5416bab9', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf151, (2, 513, 1), is_leaf=True)  # rsqrt_32
    buf152 = reader.storage('8af8818551cf5b223bb6a073f64d80980aa3ae47', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf152, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_372
    buf153 = reader.storage('f3f17f8fe369e46831266e26a472bbc8fca8e971', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf153, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_191
    buf154 = reader.storage('70269f08edca8bf8914da999cf6aa747466b1a0f', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf154, (2, 16, 128, 64), (131072, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_192
    buf155 = reader.storage('ac1ef677f46ec7593bba97719ad062cb3bee46f1', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf155, (2, 16, 128, 64), (131072, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_193
    buf156 = reader.storage('dd19b064756099a7564cc97e15f46445a1c85f91', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf156, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # getitem_77
    buf157 = reader.storage('f188b41c2e890a23b42196ed0a7ed05459f1f190', 67584, device=device(type='cuda', index=0))
    reader.tensor(buf157, (2, 16, 528), is_leaf=True)  # getitem_78
    buf158 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf158, (3,), dtype=torch.int32, is_leaf=True)  # getitem_79
    buf159 = reader.storage('1cbaf6f990ca04e947eaf166e1fbcbd05ce76bf1', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf159, (3,), dtype=torch.int32, is_leaf=True)  # getitem_80
    buf160 = reader.storage('54eb2ebcd2dd22f73fdc8b3e680b3df8190fc900', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf160, (), dtype=torch.int64, is_leaf=True)  # getitem_83
    buf161 = reader.storage('6de79f4d451761873a72c46a249b835d43d5f5e8', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf161, (), dtype=torch.int64, is_leaf=True)  # getitem_84
    reader.tensor(buf156, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_382
    buf162 = reader.storage('2ecb1f25264afa30f6ca05edf963d0a86d5dc58f', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf162, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_100
    buf163 = reader.storage('b64a8b8e4565c71a23517b25098394173ec61866', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf163, (2, 513, 1), is_leaf=True)  # getitem_87
    buf164 = reader.storage('b8261ecee764738a6c906a8e23141343fb2a761a', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf164, (2, 513, 1), is_leaf=True)  # rsqrt_33
    buf165 = reader.storage('d3656df567b2bb646e2e21d01e25562ae4767bf5', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf165, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_384
    buf166 = reader.storage('814357cf92e5ec22ed52a291fec4395ee5946785', 16809984, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf166, (2, 513, 4096), (4202496, 8192, 1), dtype=torch.float16, is_leaf=True)  # getitem_88
    reader.tensor(buf166, (2, 513, 4096), (4202496, 8192, 1), dtype=torch.float16, storage_offset=4096, is_leaf=True)  # getitem_89
    buf167 = reader.storage('a639dcbc50033923dd7b12770053fe644a85bb97', 8404992, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf167, (1026, 4096), dtype=torch.float16, is_leaf=True)  # view_386
    buf168 = reader.storage('7fa71242c7b7c2432235a6f98fa657f15a8418c5', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf168, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_103
    buf169 = reader.storage('830d513b6791168c38703a9728a4aff5720e7284', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf169, (2, 513, 1), is_leaf=True)  # getitem_91
    buf170 = reader.storage('8c04cd53c92d1f047b551512766e6faf98096a60', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf170, (2, 513, 1), is_leaf=True)  # rsqrt_34
    buf171 = reader.storage('cade8540c236fd95959f7318c4fd5bf7097cb1d3', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf171, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_388
    buf172 = reader.storage('4f164ad9219a884821940b086d34607ecc2faff2', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf172, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_203
    buf173 = reader.storage('27847540aac65f94bcf9f75afec5dcea4de1b3a4', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf173, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # cat_18
    buf174 = reader.storage('b252f0bce0fc198ae4315832bb95b194cc425b10', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf174, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # cat_20
    buf175 = reader.storage('f83e1f4988818e4088cfa76db38026fe046ab89b', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf175, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # getitem_96
    buf176 = reader.storage('87155a05e7c6b19cbe43b16a0b5da6296c513be8', 67584, device=device(type='cuda', index=0))
    reader.tensor(buf176, (2, 16, 528), is_leaf=True)  # getitem_97
    buf177 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf177, (3,), dtype=torch.int32, is_leaf=True)  # getitem_98
    buf178 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf178, (3,), dtype=torch.int32, is_leaf=True)  # getitem_99
    buf179 = reader.storage('f215d79cda670c4da1a314240c0a1c1c523db83a', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf179, (), dtype=torch.int64, is_leaf=True)  # getitem_102
    buf180 = reader.storage('df95b346d98027f00d56979d3a15995472b0256f', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf180, (), dtype=torch.int64, is_leaf=True)  # getitem_103
    reader.tensor(buf175, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_400
    buf181 = reader.storage('e759d62d191779f50301d0a21f8613648e7fa356', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf181, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_108
    buf182 = reader.storage('236ed2260b8dd296d0730e010dc968e8d006f546', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf182, (2, 513, 1), is_leaf=True)  # getitem_106
    buf183 = reader.storage('af938dbf90724b0e45fa64e635107ec28ab1c491', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf183, (2, 513, 1), is_leaf=True)  # rsqrt_35
    buf184 = reader.storage('83d01df04804b06be1b2126dd96fa9d3decd83a3', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf184, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_402
    buf185 = reader.storage('5ad60cbeac8b45216a758b6a9af66df4c12c4eae', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf185, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_209
    buf186 = reader.storage('ce89ac00b6ac3af622e5675ef143180a3e0d6bde', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf186, (2, 16, 128, 64), (131072, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_210
    buf187 = reader.storage('ae381657fbc3625a734415e378a0a077afab31cf', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf187, (2, 16, 128, 64), (131072, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_211
    buf188 = reader.storage('1139367cbf3c10a00c4ef829abf6cf4796917fcc', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf188, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # getitem_107
    buf189 = reader.storage('00eff4ee8678fc1a46318f807b21b4638d7a3dfc', 67584, device=device(type='cuda', index=0))
    reader.tensor(buf189, (2, 16, 528), is_leaf=True)  # getitem_108
    buf190 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf190, (3,), dtype=torch.int32, is_leaf=True)  # getitem_109
    buf191 = reader.storage('1cbaf6f990ca04e947eaf166e1fbcbd05ce76bf1', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf191, (3,), dtype=torch.int32, is_leaf=True)  # getitem_110
    buf192 = reader.storage('6de79f4d451761873a72c46a249b835d43d5f5e8', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf192, (), dtype=torch.int64, is_leaf=True)  # getitem_113
    buf193 = reader.storage('6de79f4d451761873a72c46a249b835d43d5f5e8', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf193, (), dtype=torch.int64, is_leaf=True)  # getitem_114
    reader.tensor(buf188, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_412
    buf194 = reader.storage('5aaca475271d4f65e1867fd4854ba6feb8f8781b', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf194, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_111
    buf195 = reader.storage('01045a55ddfa788cd9c713163a2df09e8611d799', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf195, (2, 513, 1), is_leaf=True)  # getitem_117
    buf196 = reader.storage('de9fbf12d199a5255b12951ec356069a6f2d01b1', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf196, (2, 513, 1), is_leaf=True)  # rsqrt_36
    buf197 = reader.storage('caf0bc4cc4483a8dc9d7a91d2dfdecd6c2485826', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf197, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_414
    buf198 = reader.storage('7009c1300a3d5779108564874d1530ba818c862e', 16809984, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf198, (2, 513, 4096), (4202496, 8192, 1), dtype=torch.float16, is_leaf=True)  # getitem_118
    reader.tensor(buf198, (2, 513, 4096), (4202496, 8192, 1), dtype=torch.float16, storage_offset=4096, is_leaf=True)  # getitem_119
    buf199 = reader.storage('b5d21345d2e5f2ee0db7fe0ad611f77a537e97c6', 8404992, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf199, (1026, 4096), dtype=torch.float16, is_leaf=True)  # view_416
    buf200 = reader.storage('2173ef7c18fbdf258643936052a6a2bce18d4958', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf200, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_114
    buf201 = reader.storage('86fb9271414b303487ec275e2e4aebe3940b84aa', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf201, (2, 513, 1), is_leaf=True)  # getitem_121
    buf202 = reader.storage('aa96a71e07ea2d8cc9eca70c67571f356cc0b861', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf202, (2, 513, 1), is_leaf=True)  # rsqrt_37
    buf203 = reader.storage('bd9113b5c16f0472118339049d05298e6eb3808d', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf203, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_418
    buf204 = reader.storage('a23e0c322ab63efa18b9d46a8c435013463d082e', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf204, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_221
    buf205 = reader.storage('d94c8c6cf6fbba1deacf7f23b7a73bc500818be1', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf205, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # cat_22
    buf206 = reader.storage('70999cc633435f755b2182fbe45b3993fa3377ed', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf206, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # cat_24
    buf207 = reader.storage('53858fd4d6b1bb6ce3e8a2f3bb7c059a95e0154e', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf207, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # getitem_126
    buf208 = reader.storage('1ec92a4b750d4ccb67bfc7ab70e7e0a1e8cd713b', 67584, device=device(type='cuda', index=0))
    reader.tensor(buf208, (2, 16, 528), is_leaf=True)  # getitem_127
    buf209 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf209, (3,), dtype=torch.int32, is_leaf=True)  # getitem_128
    buf210 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf210, (3,), dtype=torch.int32, is_leaf=True)  # getitem_129
    buf211 = reader.storage('6de79f4d451761873a72c46a249b835d43d5f5e8', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf211, (), dtype=torch.int64, is_leaf=True)  # getitem_132
    buf212 = reader.storage('6de79f4d451761873a72c46a249b835d43d5f5e8', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf212, (), dtype=torch.int64, is_leaf=True)  # getitem_133
    reader.tensor(buf207, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_430
    buf213 = reader.storage('defb2f33a04cc05b710e526b6b781dd6b2b8ca25', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf213, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_119
    buf214 = reader.storage('3b3704044bad4b659ee8fe5431a84a43ab08dbc1', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf214, (2, 513, 1), is_leaf=True)  # getitem_136
    buf215 = reader.storage('cd4eeb6e628b042fefb031ad0278330101169e1e', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf215, (2, 513, 1), is_leaf=True)  # rsqrt_38
    buf216 = reader.storage('a1a9a42389dbd5bdaf1676b1214f9fc7ebd7b342', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf216, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_432
    buf217 = reader.storage('04750490d7d9eea64224ac817be075e53724e893', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf217, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_227
    buf218 = reader.storage('1b28bd45bdcf39a9c520538aa30115e28db2ebc5', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf218, (2, 16, 128, 64), (131072, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_228
    buf219 = reader.storage('2e3c159748d23c6d34c03aa76d4ac52597919889', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf219, (2, 16, 128, 64), (131072, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_229
    buf220 = reader.storage('7694ffed70b1df298ff419da44779b340141b2ac', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf220, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # getitem_137
    buf221 = reader.storage('f4ab6c928be0566450d290e182bc8bdd8d180f7f', 67584, device=device(type='cuda', index=0))
    reader.tensor(buf221, (2, 16, 528), is_leaf=True)  # getitem_138
    buf222 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf222, (3,), dtype=torch.int32, is_leaf=True)  # getitem_139
    buf223 = reader.storage('1cbaf6f990ca04e947eaf166e1fbcbd05ce76bf1', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf223, (3,), dtype=torch.int32, is_leaf=True)  # getitem_140
    buf224 = reader.storage('ec14db113a385189886bd194d1b024cd05b2b1de', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf224, (), dtype=torch.int64, is_leaf=True)  # getitem_143
    buf225 = reader.storage('8d5f0d89261c7e07a1d1cb0a2a93e4993987c298', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf225, (), dtype=torch.int64, is_leaf=True)  # getitem_144
    reader.tensor(buf220, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_442
    buf226 = reader.storage('836b001dc9d410b418322f5fdbcfd6909df84952', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf226, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_122
    buf227 = reader.storage('2d401cb29b1327f3de9a7901eede8a900ba679a1', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf227, (2, 513, 1), is_leaf=True)  # getitem_147
    buf228 = reader.storage('a4322c8915e3962f918be7cabcf13a15b0fea69f', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf228, (2, 513, 1), is_leaf=True)  # rsqrt_39
    buf229 = reader.storage('680f0b5c4a4cb8cf8176bc47ca11a6705a9c4a95', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf229, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_444
    buf230 = reader.storage('6559a8aedc626d0cca64e26d18a4d987fc1a96c8', 16809984, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf230, (2, 513, 4096), (4202496, 8192, 1), dtype=torch.float16, is_leaf=True)  # getitem_148
    reader.tensor(buf230, (2, 513, 4096), (4202496, 8192, 1), dtype=torch.float16, storage_offset=4096, is_leaf=True)  # getitem_149
    buf231 = reader.storage('2632a9a035c4376c2559d5105d9640488fa481d2', 8404992, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf231, (1026, 4096), dtype=torch.float16, is_leaf=True)  # view_446
    buf232 = reader.storage('41e02523688fb4baad1f45ceb3fbc2a0f391a0d5', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf232, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_125
    buf233 = reader.storage('53e27b1b3c252071fcc61b14f839cc575b0a7761', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf233, (2, 513, 1), is_leaf=True)  # getitem_151
    buf234 = reader.storage('c2416c30666e31e17b58f82495e4648447b2fb26', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf234, (2, 513, 1), is_leaf=True)  # rsqrt_40
    buf235 = reader.storage('687f06db3f06d3cdc11caa14def0f20c9b3000f9', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf235, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_448
    buf236 = reader.storage('2284543a63c770b8e97b070fe82e628b369ee46e', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf236, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_239
    buf237 = reader.storage('10cc242f01309e40da218acf461d65b0215d48e0', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf237, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # cat_26
    buf238 = reader.storage('2622c0213c05c2b74e3f7018e90a9d9bd2807174', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf238, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # cat_28
    buf239 = reader.storage('ded88bf4fc85afbaf4595bffb8e5856c6bcfcd69', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf239, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # getitem_156
    buf240 = reader.storage('29656400b77975d7785bb9e955bc42a13b3b7ec9', 67584, device=device(type='cuda', index=0))
    reader.tensor(buf240, (2, 16, 528), is_leaf=True)  # getitem_157
    buf241 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf241, (3,), dtype=torch.int32, is_leaf=True)  # getitem_158
    buf242 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf242, (3,), dtype=torch.int32, is_leaf=True)  # getitem_159
    buf243 = reader.storage('ec14db113a385189886bd194d1b024cd05b2b1de', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf243, (), dtype=torch.int64, is_leaf=True)  # getitem_162
    buf244 = reader.storage('8d5f0d89261c7e07a1d1cb0a2a93e4993987c298', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf244, (), dtype=torch.int64, is_leaf=True)  # getitem_163
    reader.tensor(buf239, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_460
    buf245 = reader.storage('4f55e56b54667db21a77a897a9a384a7c1e21049', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf245, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_130
    buf246 = reader.storage('45e99cd31d1bfeb2f697db900f4c699efdc0ec3b', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf246, (2, 513, 1), is_leaf=True)  # getitem_166
    buf247 = reader.storage('3daec914bee5c059b357429d52487f190f803c07', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf247, (2, 513, 1), is_leaf=True)  # rsqrt_41
    buf248 = reader.storage('156c09a13e1b3551c05517dec66de1b013c26df3', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf248, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_462
    buf249 = reader.storage('cf71af2fd5cad122b5240e8c04d98e4d6ef5f86e', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf249, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_245
    buf250 = reader.storage('e9836661df4939d0d1ebed3994d8d832a6dbb6d6', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf250, (2, 16, 128, 64), (131072, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_246
    buf251 = reader.storage('537f4e9acded57cd12267f6500e1ec6724857fbe', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf251, (2, 16, 128, 64), (131072, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_247
    buf252 = reader.storage('009c6509d18e488c51bb4d78036d2efb209d5c06', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf252, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # getitem_167
    buf253 = reader.storage('1a2f032f48d6acb501ce9c80c86629819f31deb5', 67584, device=device(type='cuda', index=0))
    reader.tensor(buf253, (2, 16, 528), is_leaf=True)  # getitem_168
    buf254 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf254, (3,), dtype=torch.int32, is_leaf=True)  # getitem_169
    buf255 = reader.storage('1cbaf6f990ca04e947eaf166e1fbcbd05ce76bf1', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf255, (3,), dtype=torch.int32, is_leaf=True)  # getitem_170
    buf256 = reader.storage('6de79f4d451761873a72c46a249b835d43d5f5e8', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf256, (), dtype=torch.int64, is_leaf=True)  # getitem_173
    buf257 = reader.storage('6de79f4d451761873a72c46a249b835d43d5f5e8', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf257, (), dtype=torch.int64, is_leaf=True)  # getitem_174
    reader.tensor(buf252, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_472
    buf258 = reader.storage('5f32eccf907d61d6ad75a6b2a075467565a3b26c', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf258, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_133
    buf259 = reader.storage('ce526cd2039bb739d4394f5b59f8eb49664ab15e', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf259, (2, 513, 1), is_leaf=True)  # getitem_177
    buf260 = reader.storage('d49e991cf1368a7c26a2a2a4705f3640b5f5a8fa', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf260, (2, 513, 1), is_leaf=True)  # rsqrt_42
    buf261 = reader.storage('1d1cb1f9a4f62f2059565ffea1ed9d0b32b3a7a8', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf261, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_474
    buf262 = reader.storage('2550a8765b80daf180b0f257a05f20872c9ba6d2', 16809984, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf262, (2, 513, 4096), (4202496, 8192, 1), dtype=torch.float16, is_leaf=True)  # getitem_178
    reader.tensor(buf262, (2, 513, 4096), (4202496, 8192, 1), dtype=torch.float16, storage_offset=4096, is_leaf=True)  # getitem_179
    buf263 = reader.storage('e8295c05509f93077074f726b1fd879e36da3ece', 8404992, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf263, (1026, 4096), dtype=torch.float16, is_leaf=True)  # view_476
    buf264 = reader.storage('f8e4023e3d7c3a2756522698e741055e6aa90677', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf264, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_136
    buf265 = reader.storage('003aec91a38ac0485e24a7fd779c62d686310cec', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf265, (2, 513, 1), is_leaf=True)  # getitem_181
    buf266 = reader.storage('9833ac45320a927798072c197122bf3d0401eb55', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf266, (2, 513, 1), is_leaf=True)  # rsqrt_43
    buf267 = reader.storage('8ed3c7fd5dfd64ca3674720b1e17f6e0482f5e64', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf267, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_478
    buf268 = reader.storage('7e6a471d70f411bf5468430ec7f70330ca41f2cf', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf268, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_257
    buf269 = reader.storage('161374b5f2a276169f48d7930d986b87088ed80a', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf269, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # cat_30
    buf270 = reader.storage('71de6c3535deadb9c298fc74b8ae7f3407488c47', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf270, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # cat_32
    buf271 = reader.storage('67f57e6056652831ba3f9b3a4e7d0bceb6874759', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf271, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # getitem_186
    buf272 = reader.storage('86fd2619705ece53dabd674d49e128b2ab208c51', 67584, device=device(type='cuda', index=0))
    reader.tensor(buf272, (2, 16, 528), is_leaf=True)  # getitem_187
    buf273 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf273, (3,), dtype=torch.int32, is_leaf=True)  # getitem_188
    buf274 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf274, (3,), dtype=torch.int32, is_leaf=True)  # getitem_189
    buf275 = reader.storage('6de79f4d451761873a72c46a249b835d43d5f5e8', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf275, (), dtype=torch.int64, is_leaf=True)  # getitem_192
    buf276 = reader.storage('6de79f4d451761873a72c46a249b835d43d5f5e8', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf276, (), dtype=torch.int64, is_leaf=True)  # getitem_193
    reader.tensor(buf271, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_490
    buf277 = reader.storage('51ca7dafd72c030f765efb08c0c4a1e83e37cb5a', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf277, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_141
    buf278 = reader.storage('0a7ad4cb15c63f946d05c72d584d0d486163fb80', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf278, (2, 513, 1), is_leaf=True)  # getitem_196
    buf279 = reader.storage('bcd8a618545c0d9235df34a8234591d2699ce47c', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf279, (2, 513, 1), is_leaf=True)  # rsqrt_44
    buf280 = reader.storage('727028b9a5c8be5a34c7d33f7b334751e7e2352e', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf280, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_492
    buf281 = reader.storage('1fa3b30efe38e30ef55f05e14f2432f8653cd93a', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf281, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_263
    buf282 = reader.storage('e4d8d7941bb1b0d3255ce5f38ad990444061e456', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf282, (2, 16, 128, 64), (131072, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_264
    buf283 = reader.storage('aca939c78aa29e40b7c63e4aa7948f330150b88d', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf283, (2, 16, 128, 64), (131072, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_265
    buf284 = reader.storage('ed7107ac9fb6a43b3a8ae22c6f11af8e4ce14600', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf284, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # getitem_197
    buf285 = reader.storage('96bddc81495d33a2ba9d2e33f8ec533f007ba11d', 67584, device=device(type='cuda', index=0))
    reader.tensor(buf285, (2, 16, 528), is_leaf=True)  # getitem_198
    buf286 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf286, (3,), dtype=torch.int32, is_leaf=True)  # getitem_199
    buf287 = reader.storage('1cbaf6f990ca04e947eaf166e1fbcbd05ce76bf1', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf287, (3,), dtype=torch.int32, is_leaf=True)  # getitem_200
    buf288 = reader.storage('6de79f4d451761873a72c46a249b835d43d5f5e8', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf288, (), dtype=torch.int64, is_leaf=True)  # getitem_203
    buf289 = reader.storage('6de79f4d451761873a72c46a249b835d43d5f5e8', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf289, (), dtype=torch.int64, is_leaf=True)  # getitem_204
    reader.tensor(buf284, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_502
    buf290 = reader.storage('e07321cf8d9bdcd43b45b9fc84b520bec66bac33', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf290, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_144
    buf291 = reader.storage('e5825a29eea6cfc05c0db239a0efeff14f91ca52', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf291, (2, 513, 1), is_leaf=True)  # getitem_207
    buf292 = reader.storage('fee953e76b2345c91e24d7bb5a80cd3256d8767e', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf292, (2, 513, 1), is_leaf=True)  # rsqrt_45
    buf293 = reader.storage('e53080c4561fd38273213bf4ef7f8be9f06f4bf4', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf293, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_504
    buf294 = reader.storage('581220d8e8094a539ee0abcac2c11b4d10c02209', 16809984, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf294, (2, 513, 4096), (4202496, 8192, 1), dtype=torch.float16, is_leaf=True)  # getitem_208
    reader.tensor(buf294, (2, 513, 4096), (4202496, 8192, 1), dtype=torch.float16, storage_offset=4096, is_leaf=True)  # getitem_209
    buf295 = reader.storage('701fa4cafd6a781dde33ddcafb84fc192fac08ba', 8404992, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf295, (1026, 4096), dtype=torch.float16, is_leaf=True)  # view_506
    buf296 = reader.storage('60f17b8d1dc191acab7305dd76a3cfbf647a44b2', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf296, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_147
    buf297 = reader.storage('568dc1406721082a26caf44dbd8d38c6705c8cce', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf297, (2, 513, 1), is_leaf=True)  # getitem_211
    buf298 = reader.storage('d68aefce66690a63e18921cb0a766105d027e318', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf298, (2, 513, 1), is_leaf=True)  # rsqrt_46
    buf299 = reader.storage('2448eed6ce166832b76559018add17695b7239a1', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf299, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_508
    buf300 = reader.storage('26e169c83e264351d736aed78e68a4f7286382ca', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf300, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_275
    buf301 = reader.storage('7a4183ccd047c4c2aea95afb0f72f5918f46bb40', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf301, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # cat_34
    buf302 = reader.storage('67383be2fcf3094659219dac0677ccd4dc782147', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf302, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # cat_36
    buf303 = reader.storage('49260b3487205151a93e9cde9ea4a96a983643ef', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf303, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # getitem_216
    buf304 = reader.storage('ae3e65ded6ceb75a2dfcb6851ce8df22c9eaf4e4', 67584, device=device(type='cuda', index=0))
    reader.tensor(buf304, (2, 16, 528), is_leaf=True)  # getitem_217
    buf305 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf305, (3,), dtype=torch.int32, is_leaf=True)  # getitem_218
    buf306 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf306, (3,), dtype=torch.int32, is_leaf=True)  # getitem_219
    buf307 = reader.storage('6de79f4d451761873a72c46a249b835d43d5f5e8', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf307, (), dtype=torch.int64, is_leaf=True)  # getitem_222
    buf308 = reader.storage('6de79f4d451761873a72c46a249b835d43d5f5e8', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf308, (), dtype=torch.int64, is_leaf=True)  # getitem_223
    reader.tensor(buf303, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_520
    buf309 = reader.storage('58fc1eb4a674f231ea6a13bf856217e8a7a56e64', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf309, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_152
    buf310 = reader.storage('3b019cc0d7c5f1e8c3ec221ac2c1dc49ab8c302e', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf310, (2, 513, 1), is_leaf=True)  # getitem_226
    buf311 = reader.storage('8a81a12e41144bb5515d3cff6d218c582cd819b0', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf311, (2, 513, 1), is_leaf=True)  # rsqrt_47
    buf312 = reader.storage('ef7aa30d276bacc7e9c14b5db1b8037040b9d751', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf312, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_522
    buf313 = reader.storage('30ee0f5fe93c90d6ebba868d8da967c71f07591b', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf313, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_281
    buf314 = reader.storage('5c3f7bfdd4da75333a3ecc6046205ada88b52aed', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf314, (2, 16, 128, 64), (131072, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_282
    buf315 = reader.storage('dc287c63238b45074430b491ba2d1e20b9d92cda', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf315, (2, 16, 128, 64), (131072, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_283
    buf316 = reader.storage('d1873d240aa09b776ff30b916eab6f5d16b1e0d7', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf316, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # getitem_227
    buf317 = reader.storage('a5aa69529d4817e744712494ffc9cc80a4dee38c', 67584, device=device(type='cuda', index=0))
    reader.tensor(buf317, (2, 16, 528), is_leaf=True)  # getitem_228
    buf318 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf318, (3,), dtype=torch.int32, is_leaf=True)  # getitem_229
    buf319 = reader.storage('1cbaf6f990ca04e947eaf166e1fbcbd05ce76bf1', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf319, (3,), dtype=torch.int32, is_leaf=True)  # getitem_230
    buf320 = reader.storage('b2788bbff402a39a83c1429b89604f69dc7db49d', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf320, (), dtype=torch.int64, is_leaf=True)  # getitem_233
    buf321 = reader.storage('6b311d8f751ee568a15fd0a9b7556eedaf040ebf', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf321, (), dtype=torch.int64, is_leaf=True)  # getitem_234
    reader.tensor(buf316, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_532
    buf322 = reader.storage('c7002aac5739e6f60730005f389a1e16ac03852d', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf322, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_155
    buf323 = reader.storage('0d57f3d4f64dcf7b87885b1eacd3937e9f81f28b', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf323, (2, 513, 1), is_leaf=True)  # getitem_237
    buf324 = reader.storage('57dc177bdc9a8c69a021137ba6fb769abffa5d2c', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf324, (2, 513, 1), is_leaf=True)  # rsqrt_48
    buf325 = reader.storage('582a1fddf9ca378385b8a19b49d90c221734ad61', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf325, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_534
    buf326 = reader.storage('b843fdd24ccd6ef4f0ab231314ad2c21685175b6', 16809984, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf326, (2, 513, 4096), (4202496, 8192, 1), dtype=torch.float16, is_leaf=True)  # getitem_238
    reader.tensor(buf326, (2, 513, 4096), (4202496, 8192, 1), dtype=torch.float16, storage_offset=4096, is_leaf=True)  # getitem_239
    buf327 = reader.storage('c6b0b4f06e7d565d9f637ea782db2c3828c0a6b5', 8404992, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf327, (1026, 4096), dtype=torch.float16, is_leaf=True)  # view_536
    buf328 = reader.storage('762b377aa6bcf85c7c5c71474ec8f74069540c1f', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf328, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_158
    buf329 = reader.storage('b6f1d917f8acbc83ea5214e3883a1770debf705d', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf329, (2, 513, 1), is_leaf=True)  # getitem_241
    buf330 = reader.storage('11dd6812e2a66b88dbfec9e92e05328ff294f629', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf330, (2, 513, 1), is_leaf=True)  # rsqrt_49
    buf331 = reader.storage('ba383edf702cb07619529ab3d1e9d67612b7640b', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf331, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_538
    buf332 = reader.storage('dea3e18b4d022b0dd4ae25ce231e42883c2030a7', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf332, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_293
    buf333 = reader.storage('108ef3c2b30c55f0fa73f5aada1bd7e340f769e5', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf333, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # cat_38
    buf334 = reader.storage('20cfd5e5847739cc4061100d8d4de2096ae323dd', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf334, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # cat_40
    buf335 = reader.storage('015e9544d7bae90fbb83801517b4978138aff40d', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf335, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # getitem_246
    buf336 = reader.storage('4795a4cd656280d953cf7a7f7beb6621472a1604', 67584, device=device(type='cuda', index=0))
    reader.tensor(buf336, (2, 16, 528), is_leaf=True)  # getitem_247
    buf337 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf337, (3,), dtype=torch.int32, is_leaf=True)  # getitem_248
    buf338 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf338, (3,), dtype=torch.int32, is_leaf=True)  # getitem_249
    buf339 = reader.storage('4a388065c77f56b617fc3f13aedf905339c42149', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf339, (), dtype=torch.int64, is_leaf=True)  # getitem_252
    buf340 = reader.storage('8500910c33e929f6eb08cb7f14ad3c765a35fbb1', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf340, (), dtype=torch.int64, is_leaf=True)  # getitem_253
    reader.tensor(buf335, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_550
    buf341 = reader.storage('291e9a0d7f2ffa68d37a4780b1675f823b59bd08', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf341, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_163
    buf342 = reader.storage('93d2d1099ff871795722fa0f2849de4117ea4297', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf342, (2, 513, 1), is_leaf=True)  # getitem_256
    buf343 = reader.storage('246cd51570c2f754f674d2aaed582b0061420b26', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf343, (2, 513, 1), is_leaf=True)  # rsqrt_50
    buf344 = reader.storage('fce1956028e7a6ce58dac9326e73014555d84e5c', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf344, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_552
    buf345 = reader.storage('0ac34765222cde92a8afa061c9b6e8c290355f28', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf345, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_299
    buf346 = reader.storage('6fefed45b68a1faee5d6f2f95099eccef7aac221', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf346, (2, 16, 128, 64), (131072, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_300
    buf347 = reader.storage('7b5e8c5037e2de4ef2cae4dca4ee449c5d2596dc', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf347, (2, 16, 128, 64), (131072, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_301
    buf348 = reader.storage('defcfaec78f398652294a1d7b426d431cb72fa4b', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf348, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # getitem_257
    buf349 = reader.storage('cab38b84c41394a69d96ba49dc0dd0e72c0986eb', 67584, device=device(type='cuda', index=0))
    reader.tensor(buf349, (2, 16, 528), is_leaf=True)  # getitem_258
    buf350 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf350, (3,), dtype=torch.int32, is_leaf=True)  # getitem_259
    buf351 = reader.storage('1cbaf6f990ca04e947eaf166e1fbcbd05ce76bf1', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf351, (3,), dtype=torch.int32, is_leaf=True)  # getitem_260
    buf352 = reader.storage('fed7c85c48d8b47fe4776446262a1c68a7d0df98', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf352, (), dtype=torch.int64, is_leaf=True)  # getitem_263
    buf353 = reader.storage('f3dddfae9c0e25f15690ade4ef8551b2edd0768e', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf353, (), dtype=torch.int64, is_leaf=True)  # getitem_264
    reader.tensor(buf348, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_562
    buf354 = reader.storage('b90fecadc7e4614a9955722b04cdcb7d46b80c86', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf354, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_166
    buf355 = reader.storage('13a15b816d3ade90a5e4dd1664df50a749328f9e', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf355, (2, 513, 1), is_leaf=True)  # getitem_267
    buf356 = reader.storage('4e4163bac653981f5a67505fef9bc604725e9850', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf356, (2, 513, 1), is_leaf=True)  # rsqrt_51
    buf357 = reader.storage('67c1a1ddb1826c32626ad4cb4c439e8baf25698c', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf357, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_564
    buf358 = reader.storage('d9a28758bedfa018262d2d41a908e45963743f1d', 16809984, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf358, (2, 513, 4096), (4202496, 8192, 1), dtype=torch.float16, is_leaf=True)  # getitem_268
    reader.tensor(buf358, (2, 513, 4096), (4202496, 8192, 1), dtype=torch.float16, storage_offset=4096, is_leaf=True)  # getitem_269
    buf359 = reader.storage('ac78f3f88f6113fe678b6d5b1dd7801cbfca75c1', 8404992, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf359, (1026, 4096), dtype=torch.float16, is_leaf=True)  # view_566
    buf360 = reader.storage('a927d54acd5f423ad3abf0da14fa1a40a20232b0', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf360, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_169
    buf361 = reader.storage('1c29d40926f0a78fb6ee42b767262b29ae3ce380', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf361, (2, 513, 1), is_leaf=True)  # getitem_271
    buf362 = reader.storage('42e804b6d5f7e3f56fc36564547f4f45f33d4ae1', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf362, (2, 513, 1), is_leaf=True)  # rsqrt_52
    buf363 = reader.storage('a9622a85721c116c568f6bafbe9b2ebbdb5b3e53', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf363, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_568
    buf364 = reader.storage('48a297b7b6ef02137ad8f362b64ae11c6796defb', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf364, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_311
    buf365 = reader.storage('6d31d963b4bb395ab7957ce9dbcda37281a6f387', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf365, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # cat_42
    buf366 = reader.storage('fdc6f3a3bcc5d66b09a9731a0d8e767efd7d3f0e', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf366, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # cat_44
    buf367 = reader.storage('344a53a68e3d9fb2c5483b19411dd2858ca372ee', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf367, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # getitem_276
    buf368 = reader.storage('c1c05c2ca2db880532c2119b3b168569081836f6', 67584, device=device(type='cuda', index=0))
    reader.tensor(buf368, (2, 16, 528), is_leaf=True)  # getitem_277
    buf369 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf369, (3,), dtype=torch.int32, is_leaf=True)  # getitem_278
    buf370 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf370, (3,), dtype=torch.int32, is_leaf=True)  # getitem_279
    buf371 = reader.storage('0e15b2da1dcacef7123f1cb21cb55802437921e0', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf371, (), dtype=torch.int64, is_leaf=True)  # getitem_282
    buf372 = reader.storage('5f973e7f7eb1c13257980a439e01ed5926bb5725', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf372, (), dtype=torch.int64, is_leaf=True)  # getitem_283
    reader.tensor(buf367, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_580
    buf373 = reader.storage('a9beaf5c6c5c61d10a78a03a5cf18cf55bb3588e', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf373, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_174
    buf374 = reader.storage('67cff5c04856d995356ed53715ffa092a1aaac44', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf374, (2, 513, 1), is_leaf=True)  # getitem_286
    buf375 = reader.storage('030597246ab53004ad1f7f98acf61da00a61688d', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf375, (2, 513, 1), is_leaf=True)  # rsqrt_53
    buf376 = reader.storage('d70bd086d16b7549d2e967d28bae1a8f15ed195a', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf376, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_582
    buf377 = reader.storage('49996cb27a99b7b4d1b93634c1f4b8ff780487c1', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf377, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_317
    buf378 = reader.storage('9e259f7e992dae4e339c1030b981e16da67a8275', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf378, (2, 16, 128, 64), (131072, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_318
    buf379 = reader.storage('89c77e351ef45a638e220e7448d8bcc5e5d810e7', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf379, (2, 16, 128, 64), (131072, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_319
    buf380 = reader.storage('4f914f2608fcfa9dd525fd2da4e23ffbfc7041f4', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf380, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # getitem_287
    buf381 = reader.storage('b81246a6563b6f0ef06b4ff6b521335cf3df22e5', 67584, device=device(type='cuda', index=0))
    reader.tensor(buf381, (2, 16, 528), is_leaf=True)  # getitem_288
    buf382 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf382, (3,), dtype=torch.int32, is_leaf=True)  # getitem_289
    buf383 = reader.storage('1cbaf6f990ca04e947eaf166e1fbcbd05ce76bf1', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf383, (3,), dtype=torch.int32, is_leaf=True)  # getitem_290
    buf384 = reader.storage('98cc91c934e7fafa601f084fd7bca327a541a4ed', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf384, (), dtype=torch.int64, is_leaf=True)  # getitem_293
    buf385 = reader.storage('9f639ff633ffa31c5c0be8d11ea102ecc82d07cb', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf385, (), dtype=torch.int64, is_leaf=True)  # getitem_294
    reader.tensor(buf380, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_592
    buf386 = reader.storage('0c65598218d8f62595f5e1b83c233549228b7479', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf386, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_177
    buf387 = reader.storage('95ac37443bb68ce4f82ca5515b96b16608045246', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf387, (2, 513, 1), is_leaf=True)  # getitem_297
    buf388 = reader.storage('7546466b613ded7ac3e57bd4b85ed070373a4a69', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf388, (2, 513, 1), is_leaf=True)  # rsqrt_54
    buf389 = reader.storage('dc2665feae3ddd87a2860eef2f320c994681d770', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf389, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_594
    buf390 = reader.storage('b3a0e1fcdbd061d8dcfdf151b67a85c7bddc1755', 16809984, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf390, (2, 513, 4096), (4202496, 8192, 1), dtype=torch.float16, is_leaf=True)  # getitem_298
    reader.tensor(buf390, (2, 513, 4096), (4202496, 8192, 1), dtype=torch.float16, storage_offset=4096, is_leaf=True)  # getitem_299
    buf391 = reader.storage('65c6e4e2da42213b261dd0ad28139cf2269b968c', 8404992, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf391, (1026, 4096), dtype=torch.float16, is_leaf=True)  # view_596
    buf392 = reader.storage('7e50b55efd31324bfd8791f61f429c5142a9c2cb', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf392, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_180
    buf393 = reader.storage('1d26ffa99bfb62787c064bb4a87d022c76bfc72e', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf393, (2, 513, 1), is_leaf=True)  # getitem_301
    buf394 = reader.storage('dec479fe9d1023ea693b6a16da99cc10926ee1b1', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf394, (2, 513, 1), is_leaf=True)  # rsqrt_55
    buf395 = reader.storage('b702087d3991d5f10807222a69d395454c9c51e3', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf395, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_598
    buf396 = reader.storage('9689c815d2f14fc54003384c05a26d4706437917', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf396, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_329
    buf397 = reader.storage('e74572909992ad2abc06fac5beb4b126c5745f5b', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf397, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # cat_46
    buf398 = reader.storage('75ced79e52052354ce5f424f996181d176d5171d', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf398, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # cat_48
    buf399 = reader.storage('c97feb264811f1baaf9084d2a9244cd765db8a65', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf399, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # getitem_306
    buf400 = reader.storage('75d4020fe107f9f78cddf2fdbaf6e51452c8143d', 67584, device=device(type='cuda', index=0))
    reader.tensor(buf400, (2, 16, 528), is_leaf=True)  # getitem_307
    buf401 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf401, (3,), dtype=torch.int32, is_leaf=True)  # getitem_308
    buf402 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf402, (3,), dtype=torch.int32, is_leaf=True)  # getitem_309
    buf403 = reader.storage('bf30973635248386f554f81b8f4b1498f98813b1', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf403, (), dtype=torch.int64, is_leaf=True)  # getitem_312
    buf404 = reader.storage('271f4046509352569abd6a0b54a40388dea74501', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf404, (), dtype=torch.int64, is_leaf=True)  # getitem_313
    reader.tensor(buf399, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_610
    buf405 = reader.storage('93bcf812c0f3e2588b5ddb37761440064b930b0c', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf405, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_185
    buf406 = reader.storage('05d27abe8766daa50537ebcd82f3bf4a9250e420', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf406, (2, 513, 1), is_leaf=True)  # getitem_316
    buf407 = reader.storage('cb777d34cf11d3030e391370f9299d61a0ee7baa', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf407, (2, 513, 1), is_leaf=True)  # rsqrt_56
    buf408 = reader.storage('3477041f6af436d8386bdb6d61bd750c9505f350', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf408, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_612
    buf409 = reader.storage('37a20032539b1647300d2af2d1ec04028302db25', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf409, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_335
    buf410 = reader.storage('059130ba24ce7b69cd51de6f1f4fd0cb01f998dc', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf410, (2, 16, 128, 64), (131072, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_336
    buf411 = reader.storage('4211f673c173eaa1a14f744d5597dce7e8558d38', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf411, (2, 16, 128, 64), (131072, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_337
    buf412 = reader.storage('5202f78cd88ea7837e6b5bf532b53c32730c5ea6', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf412, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # getitem_317
    buf413 = reader.storage('9b03d68712519c27041589909815c8c84bbd2df3', 67584, device=device(type='cuda', index=0))
    reader.tensor(buf413, (2, 16, 528), is_leaf=True)  # getitem_318
    buf414 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf414, (3,), dtype=torch.int32, is_leaf=True)  # getitem_319
    buf415 = reader.storage('1cbaf6f990ca04e947eaf166e1fbcbd05ce76bf1', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf415, (3,), dtype=torch.int32, is_leaf=True)  # getitem_320
    buf416 = reader.storage('6de79f4d451761873a72c46a249b835d43d5f5e8', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf416, (), dtype=torch.int64, is_leaf=True)  # getitem_323
    buf417 = reader.storage('6de79f4d451761873a72c46a249b835d43d5f5e8', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf417, (), dtype=torch.int64, is_leaf=True)  # getitem_324
    reader.tensor(buf412, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_622
    buf418 = reader.storage('fa3bdd26a7ec808dd01f645c2cf380cff040de87', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf418, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_188
    buf419 = reader.storage('6b8b214fd3ad9efb2754226e44a9a37850bdbeb9', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf419, (2, 513, 1), is_leaf=True)  # getitem_327
    buf420 = reader.storage('b35f64d52c620e9e3cec8c3094f675a5dafea3b4', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf420, (2, 513, 1), is_leaf=True)  # rsqrt_57
    buf421 = reader.storage('dd7373797cb07134b5763a2dc4e5f38fa853c1f2', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf421, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_624
    buf422 = reader.storage('ca73ffa4456ac78d6568baddfc1e4cfb35221154', 16809984, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf422, (2, 513, 4096), (4202496, 8192, 1), dtype=torch.float16, is_leaf=True)  # getitem_328
    reader.tensor(buf422, (2, 513, 4096), (4202496, 8192, 1), dtype=torch.float16, storage_offset=4096, is_leaf=True)  # getitem_329
    buf423 = reader.storage('2f0a2e3afc1616e9643ebf412432ee6c008bcabb', 8404992, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf423, (1026, 4096), dtype=torch.float16, is_leaf=True)  # view_626
    buf424 = reader.storage('a436bd8fa6b30959e8b0e37fa25c63a625b76267', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf424, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_191
    buf425 = reader.storage('5d9579609a307016acc14ae626c574dc89578d50', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf425, (2, 513, 1), is_leaf=True)  # getitem_331
    buf426 = reader.storage('b75fade2c662deea3aececdb62ce83c5e88932f3', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf426, (2, 513, 1), is_leaf=True)  # rsqrt_58
    buf427 = reader.storage('b80830992545a3a08234b0acf476258eae6b5e13', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf427, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_628
    buf428 = reader.storage('a482a233f9de7eacbeeb3a907ba71e8c1fb7857f', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf428, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_347
    buf429 = reader.storage('a6c2b491fde9026382fa0037a3a8337d21acfe6e', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf429, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # cat_50
    buf430 = reader.storage('17763bd611fa51819d7e76cf82d54cd683db3ee6', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf430, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # cat_52
    buf431 = reader.storage('f45a358c6393c1c518cc67c02aa06efce723395f', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf431, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # getitem_336
    buf432 = reader.storage('6e16cb23d58be3ea49812ff26a68d44b3c0b9be9', 67584, device=device(type='cuda', index=0))
    reader.tensor(buf432, (2, 16, 528), is_leaf=True)  # getitem_337
    buf433 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf433, (3,), dtype=torch.int32, is_leaf=True)  # getitem_338
    buf434 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf434, (3,), dtype=torch.int32, is_leaf=True)  # getitem_339
    buf435 = reader.storage('6de79f4d451761873a72c46a249b835d43d5f5e8', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf435, (), dtype=torch.int64, is_leaf=True)  # getitem_342
    buf436 = reader.storage('6de79f4d451761873a72c46a249b835d43d5f5e8', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf436, (), dtype=torch.int64, is_leaf=True)  # getitem_343
    reader.tensor(buf431, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_640
    buf437 = reader.storage('6ab1d8b12a450be3280fa1ef7538b10fbbf74dd4', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf437, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_196
    buf438 = reader.storage('2036a0fdef82202858433bad8421ca3b5ae04cae', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf438, (2, 513, 1), is_leaf=True)  # getitem_346
    buf439 = reader.storage('5d956aa5ab15bbe754ae9bfa1a2c403268253633', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf439, (2, 513, 1), is_leaf=True)  # rsqrt_59
    buf440 = reader.storage('cf1fdd05b7f518968e7128b65847595e5224028d', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf440, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_642
    buf441 = reader.storage('f20eb0723d2c2af8f03b0e7129789becf74c1c30', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf441, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_353
    buf442 = reader.storage('5c5f033518ba4f526e4dae37bd25a86c31b0083c', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf442, (2, 16, 128, 64), (131072, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_354
    buf443 = reader.storage('cd5ee8f201ece3967943970d7602b28d1977765a', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf443, (2, 16, 128, 64), (131072, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_355
    buf444 = reader.storage('eb63ef8d2121e9bb460b3fa654d2edcd32ae9950', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf444, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # getitem_347
    buf445 = reader.storage('4a1ce4741ac7c461920ac23b4033799bbff14b2d', 67584, device=device(type='cuda', index=0))
    reader.tensor(buf445, (2, 16, 528), is_leaf=True)  # getitem_348
    buf446 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf446, (3,), dtype=torch.int32, is_leaf=True)  # getitem_349
    buf447 = reader.storage('1cbaf6f990ca04e947eaf166e1fbcbd05ce76bf1', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf447, (3,), dtype=torch.int32, is_leaf=True)  # getitem_350
    buf448 = reader.storage('6de79f4d451761873a72c46a249b835d43d5f5e8', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf448, (), dtype=torch.int64, is_leaf=True)  # getitem_353
    buf449 = reader.storage('6de79f4d451761873a72c46a249b835d43d5f5e8', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf449, (), dtype=torch.int64, is_leaf=True)  # getitem_354
    reader.tensor(buf444, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_652
    buf450 = reader.storage('eb1969d1bdeb3c9fb3956f877263818dd9f17024', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf450, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_199
    buf451 = reader.storage('77ac59b8d7fe426b079e85975cec533b43d89801', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf451, (2, 513, 1), is_leaf=True)  # getitem_357
    buf452 = reader.storage('496d8bc9d59cbdfd4ec86baac4bfe47afc0d2c56', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf452, (2, 513, 1), is_leaf=True)  # rsqrt_60
    buf453 = reader.storage('d30b48ae46555da57bc871e15a3ea1f8574941a4', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf453, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_654
    buf454 = reader.storage('bb563e7bba3acbc9ac6d68720815edc473105332', 16809984, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf454, (2, 513, 4096), (4202496, 8192, 1), dtype=torch.float16, is_leaf=True)  # getitem_358
    reader.tensor(buf454, (2, 513, 4096), (4202496, 8192, 1), dtype=torch.float16, storage_offset=4096, is_leaf=True)  # getitem_359
    buf455 = reader.storage('98497e2e71ef6bd3f0ba2d402f0b700ea249aded', 8404992, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf455, (1026, 4096), dtype=torch.float16, is_leaf=True)  # view_656
    buf456 = reader.storage('8e1170fee73ac1bc0ef4726e28dc4f8b4088f78d', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf456, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_202
    buf457 = reader.storage('f9cb2098ba5de0b44d91bc0372d92470fa2ab1f5', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf457, (2, 513, 1), is_leaf=True)  # getitem_361
    buf458 = reader.storage('27ce41836960801bdcb1c2b3d9b098368c408322', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf458, (2, 513, 1), is_leaf=True)  # rsqrt_61
    buf459 = reader.storage('4359fa72077e44b4b4766266e7b8d9bebceb46e5', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf459, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_658
    buf460 = reader.storage('68e9b2b0fadfffe0b30e9659138aecf1ba5cdfbd', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf460, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_365
    buf461 = reader.storage('d7578674d760d6e86633b20b9fd1dce99e926742', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf461, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # cat_54
    buf462 = reader.storage('0965ca6a3985c4041505310e1d933453ffce66be', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf462, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # cat_56
    buf463 = reader.storage('7d9c92199b61c892e1ed8d378f3928cbe6bef313', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf463, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # getitem_366
    buf464 = reader.storage('9ead307501e2d49b095874eca41f00f8f0d853f7', 67584, device=device(type='cuda', index=0))
    reader.tensor(buf464, (2, 16, 528), is_leaf=True)  # getitem_367
    buf465 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf465, (3,), dtype=torch.int32, is_leaf=True)  # getitem_368
    buf466 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf466, (3,), dtype=torch.int32, is_leaf=True)  # getitem_369
    buf467 = reader.storage('6de79f4d451761873a72c46a249b835d43d5f5e8', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf467, (), dtype=torch.int64, is_leaf=True)  # getitem_372
    buf468 = reader.storage('6de79f4d451761873a72c46a249b835d43d5f5e8', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf468, (), dtype=torch.int64, is_leaf=True)  # getitem_373
    reader.tensor(buf463, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_670
    buf469 = reader.storage('ec999fb408966b0c5ce735b57e31e55c0b515488', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf469, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_207
    buf470 = reader.storage('63b5f226d7a98b14675635dc0574bf96c9b33beb', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf470, (2, 513, 1), is_leaf=True)  # getitem_376
    buf471 = reader.storage('f976593c9a3f433003e403ec73e6dac249dc4fa4', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf471, (2, 513, 1), is_leaf=True)  # rsqrt_62
    buf472 = reader.storage('d62d548142ee9e6553c5dea2666f9569d3a102e0', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf472, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_672
    buf473 = reader.storage('ba8bb7856c245cc886a8c771adc28ba2581e14e8', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf473, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_371
    buf474 = reader.storage('07bf5831ec489ac44be50a8389d31336ce55a731', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf474, (2, 16, 128, 64), (131072, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_372
    buf475 = reader.storage('6e8fdd740ffa3ebc6da7e458fd740d1098aa13e9', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf475, (2, 16, 128, 64), (131072, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_373
    buf476 = reader.storage('7214042864522ac44956fb5d447c2cb78a6c9e9e', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf476, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # getitem_377
    buf477 = reader.storage('07bc8fc97f1c39ba1223ea598ba47c71e4b2cd68', 67584, device=device(type='cuda', index=0))
    reader.tensor(buf477, (2, 16, 528), is_leaf=True)  # getitem_378
    buf478 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf478, (3,), dtype=torch.int32, is_leaf=True)  # getitem_379
    buf479 = reader.storage('1cbaf6f990ca04e947eaf166e1fbcbd05ce76bf1', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf479, (3,), dtype=torch.int32, is_leaf=True)  # getitem_380
    buf480 = reader.storage('6de79f4d451761873a72c46a249b835d43d5f5e8', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf480, (), dtype=torch.int64, is_leaf=True)  # getitem_383
    buf481 = reader.storage('6de79f4d451761873a72c46a249b835d43d5f5e8', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf481, (), dtype=torch.int64, is_leaf=True)  # getitem_384
    reader.tensor(buf476, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_682
    buf482 = reader.storage('988e460f00a2ad4505f68678d9e417892f214507', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf482, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_210
    buf483 = reader.storage('6b740c5613fa8ef9b2331b924589d03d31e9d30c', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf483, (2, 513, 1), is_leaf=True)  # getitem_387
    buf484 = reader.storage('953b74a7d1242f0974628f9d9c502920b0bb2513', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf484, (2, 513, 1), is_leaf=True)  # rsqrt_63
    buf485 = reader.storage('7d39e1d51f063ea95a82c5b8c6120ff09d7cca15', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf485, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_684
    buf486 = reader.storage('7b77595c868a412d848d4c913d14057628fe28b7', 16809984, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf486, (2, 513, 4096), (4202496, 8192, 1), dtype=torch.float16, is_leaf=True)  # getitem_388
    reader.tensor(buf486, (2, 513, 4096), (4202496, 8192, 1), dtype=torch.float16, storage_offset=4096, is_leaf=True)  # getitem_389
    buf487 = reader.storage('8f07094d29feed2cbeac79c62e6b23c36e1845e1', 8404992, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf487, (1026, 4096), dtype=torch.float16, is_leaf=True)  # view_686
    buf488 = reader.storage('d10f8a85c3fb50b9148bcbf94920ce89882df6de', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf488, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_213
    buf489 = reader.storage('6638b7f7b893a4af535b36e25480b904511477cf', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf489, (2, 513, 1), is_leaf=True)  # getitem_391
    buf490 = reader.storage('db1b8d9e11140d8ccd89fdc7d6ad15408a6262b0', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf490, (2, 513, 1), is_leaf=True)  # rsqrt_64
    buf491 = reader.storage('047bd6dc4503c0d2da17b491e22ce5a287d91d23', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf491, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_688
    buf492 = reader.storage('f778ced9d714f293fd7a37fb4c3920a6cf94cc44', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf492, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_383
    buf493 = reader.storage('b07a45615b26ae084abd3fe4b05d01246664053a', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf493, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # cat_58
    buf494 = reader.storage('472550fcec132ce212f6287f047f9e30a3c2a5cc', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf494, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # cat_60
    buf495 = reader.storage('f2b5682c451b5789dade8bf68a1619233760aad8', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf495, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # getitem_396
    buf496 = reader.storage('7ae3b3b9562db7ec93aa744bcd0f2198e5694255', 67584, device=device(type='cuda', index=0))
    reader.tensor(buf496, (2, 16, 528), is_leaf=True)  # getitem_397
    buf497 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf497, (3,), dtype=torch.int32, is_leaf=True)  # getitem_398
    buf498 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf498, (3,), dtype=torch.int32, is_leaf=True)  # getitem_399
    buf499 = reader.storage('6de79f4d451761873a72c46a249b835d43d5f5e8', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf499, (), dtype=torch.int64, is_leaf=True)  # getitem_402
    buf500 = reader.storage('6de79f4d451761873a72c46a249b835d43d5f5e8', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf500, (), dtype=torch.int64, is_leaf=True)  # getitem_403
    reader.tensor(buf495, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_700
    buf501 = reader.storage('8b44e37ba7d5f42e74cb8b403f549a6ed27b3020', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf501, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_218
    buf502 = reader.storage('5159f1d69f08d6762bcef6b3ae0867314a92d5b9', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf502, (2, 513, 1), is_leaf=True)  # getitem_406
    buf503 = reader.storage('2089ac78a549c0684e409507417e3edc0fba999e', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf503, (2, 513, 1), is_leaf=True)  # rsqrt_65
    buf504 = reader.storage('180160fc471c72c15fa2eb6cb787178f765a9529', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf504, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_702
    buf505 = reader.storage('337394d234d6d06f87ae991645b46a6095a19236', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf505, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_389
    buf506 = reader.storage('493ae69bba25616f01c7b538cdcc112b3d143130', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf506, (2, 16, 128, 64), (131072, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_390
    buf507 = reader.storage('7452eaa5b795b7b632eba34fd44672f6b9c56eb6', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf507, (2, 16, 128, 64), (131072, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_391
    buf508 = reader.storage('aa280e85000e562e9fe7c148b60d023f9c4c7bea', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf508, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # getitem_407
    buf509 = reader.storage('d8f4c1c69984649396c81538787fc827f371775e', 67584, device=device(type='cuda', index=0))
    reader.tensor(buf509, (2, 16, 528), is_leaf=True)  # getitem_408
    buf510 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf510, (3,), dtype=torch.int32, is_leaf=True)  # getitem_409
    buf511 = reader.storage('1cbaf6f990ca04e947eaf166e1fbcbd05ce76bf1', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf511, (3,), dtype=torch.int32, is_leaf=True)  # getitem_410
    buf512 = reader.storage('6de79f4d451761873a72c46a249b835d43d5f5e8', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf512, (), dtype=torch.int64, is_leaf=True)  # getitem_413
    buf513 = reader.storage('6de79f4d451761873a72c46a249b835d43d5f5e8', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf513, (), dtype=torch.int64, is_leaf=True)  # getitem_414
    reader.tensor(buf508, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_712
    buf514 = reader.storage('f20171f83ab5bb814ce210fe54bdc652a3603fe6', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf514, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_221
    buf515 = reader.storage('0e91436368e5caaff63d12710df1bd08c3873c33', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf515, (2, 513, 1), is_leaf=True)  # getitem_417
    buf516 = reader.storage('cd09e1ac5037a0a29fc3218f717e648fc0304996', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf516, (2, 513, 1), is_leaf=True)  # rsqrt_66
    buf517 = reader.storage('bf677b63c0186d5b31d3f5b05fb432ba7c9430d9', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf517, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_714
    buf518 = reader.storage('ecf0d109d47ce1bfaedc767da49897749bdb2bf1', 16809984, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf518, (2, 513, 4096), (4202496, 8192, 1), dtype=torch.float16, is_leaf=True)  # getitem_418
    reader.tensor(buf518, (2, 513, 4096), (4202496, 8192, 1), dtype=torch.float16, storage_offset=4096, is_leaf=True)  # getitem_419
    buf519 = reader.storage('668a686782c92634c9849114fd6907ceef4fb263', 8404992, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf519, (1026, 4096), dtype=torch.float16, is_leaf=True)  # view_716
    buf520 = reader.storage('b8b99b1e201bca8ba5542b31f8e866a6b4817d24', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf520, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_224
    buf521 = reader.storage('334bb1dee7c6839e6ffb242083a98567b090d087', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf521, (2, 513, 1), is_leaf=True)  # getitem_421
    buf522 = reader.storage('ed086e9c94142a3e497c6ad1204e69991561a6f1', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf522, (2, 513, 1), is_leaf=True)  # rsqrt_67
    buf523 = reader.storage('9c711f9ac99aa6f995d29874522d609c75a3c0d3', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf523, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_718
    buf524 = reader.storage('ca7872adbb6d2a5875b7653186fe4083bf70b03f', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf524, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_401
    buf525 = reader.storage('c0afdd35e03b7707157cbf01e82c3ffe3cbffac9', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf525, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # cat_62
    buf526 = reader.storage('2544e1b417471f2abfc6aada1b76dd5606d7a068', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf526, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # cat_64
    buf527 = reader.storage('c2eff0125b53b9158ac54481ec47dae641ec66b0', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf527, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # getitem_426
    buf528 = reader.storage('138b507dccb5447d5dbe24fab4bb989053a61662', 67584, device=device(type='cuda', index=0))
    reader.tensor(buf528, (2, 16, 528), is_leaf=True)  # getitem_427
    buf529 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf529, (3,), dtype=torch.int32, is_leaf=True)  # getitem_428
    buf530 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf530, (3,), dtype=torch.int32, is_leaf=True)  # getitem_429
    buf531 = reader.storage('6de79f4d451761873a72c46a249b835d43d5f5e8', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf531, (), dtype=torch.int64, is_leaf=True)  # getitem_432
    buf532 = reader.storage('6de79f4d451761873a72c46a249b835d43d5f5e8', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf532, (), dtype=torch.int64, is_leaf=True)  # getitem_433
    reader.tensor(buf527, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_730
    buf533 = reader.storage('9a6b4983d3d7bb378befca126bfc624be425b68a', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf533, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_229
    buf534 = reader.storage('7ae140137149b0e9d9d76a7b8b954fec234bd185', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf534, (2, 513, 1), is_leaf=True)  # getitem_436
    buf535 = reader.storage('c768f191f94e1edae0e1d210a4877b9bc077f731', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf535, (2, 513, 1), is_leaf=True)  # rsqrt_68
    buf536 = reader.storage('4e8ab7e87e0af8c88698a89f53932cee3c386a98', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf536, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_732
    buf537 = reader.storage('f0f809faefffccabbb9d4340bc089c681036e496', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf537, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_407
    buf538 = reader.storage('ced148c33c7506fef3176cebd1062569cedcc531', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf538, (2, 16, 128, 64), (131072, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_408
    buf539 = reader.storage('76510fb7233022f9a77caf6efb9c61d9aa3f7a5a', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf539, (2, 16, 128, 64), (131072, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_409
    buf540 = reader.storage('935326073eea2a3b6f78506fe1999db2f2ed63de', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf540, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # getitem_437
    buf541 = reader.storage('0dda49fcb60547c8098354ad87c5e796ca25424d', 67584, device=device(type='cuda', index=0))
    reader.tensor(buf541, (2, 16, 528), is_leaf=True)  # getitem_438
    buf542 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf542, (3,), dtype=torch.int32, is_leaf=True)  # getitem_439
    buf543 = reader.storage('1cbaf6f990ca04e947eaf166e1fbcbd05ce76bf1', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf543, (3,), dtype=torch.int32, is_leaf=True)  # getitem_440
    buf544 = reader.storage('6de79f4d451761873a72c46a249b835d43d5f5e8', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf544, (), dtype=torch.int64, is_leaf=True)  # getitem_443
    buf545 = reader.storage('6de79f4d451761873a72c46a249b835d43d5f5e8', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf545, (), dtype=torch.int64, is_leaf=True)  # getitem_444
    reader.tensor(buf540, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_742
    buf546 = reader.storage('081773acf360e0fa52e7bfc0a38565f317151f4c', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf546, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_232
    buf547 = reader.storage('9ab33f00ebae605cba5d5d964c11f1ab9cd8e20b', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf547, (2, 513, 1), is_leaf=True)  # getitem_447
    buf548 = reader.storage('0be2d23e7960bcdc2b0324135eb7de854343d149', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf548, (2, 513, 1), is_leaf=True)  # rsqrt_69
    buf549 = reader.storage('aacb705d7d98c307d3562379990cfeabc11ede29', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf549, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_744
    buf550 = reader.storage('e224c34b90bdcee2f3dc305ceee7a198470b7a08', 16809984, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf550, (2, 513, 4096), (4202496, 8192, 1), dtype=torch.float16, is_leaf=True)  # getitem_448
    reader.tensor(buf550, (2, 513, 4096), (4202496, 8192, 1), dtype=torch.float16, storage_offset=4096, is_leaf=True)  # getitem_449
    buf551 = reader.storage('60eb1870f7e3b95d75e8671daa9157c8f9308c0e', 8404992, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf551, (1026, 4096), dtype=torch.float16, is_leaf=True)  # view_746
    buf552 = reader.storage('10d3f9282474d898bd4e2d03e81513a665eceed5', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf552, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_235
    buf553 = reader.storage('513305cd52415c4fab4d666c874885b448bf135d', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf553, (2, 513, 1), is_leaf=True)  # getitem_451
    buf554 = reader.storage('0665e0e5f6020c6f4831c5aac4ebdc4cfa792c88', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf554, (2, 513, 1), is_leaf=True)  # rsqrt_70
    buf555 = reader.storage('356ba8ca08c1c7f03d1a2176384bf83dcce23162', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf555, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_748
    buf556 = reader.storage('2cd0a13fd6a8c62e8152871f05e664b1d46929ac', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf556, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_419
    buf557 = reader.storage('68b07d4148b169d6df245e4b395cb41a5a163a24', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf557, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # cat_66
    buf558 = reader.storage('7a12ed05db6fdfbba16ab4ccbb4dd27992e883f5', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf558, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # cat_68
    buf559 = reader.storage('dcb6bface8198c89b7f409f2db7f7f073b53fa08', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf559, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # getitem_456
    buf560 = reader.storage('e6c9b2da0f9273d40352b225b43dbae7b9e936ed', 67584, device=device(type='cuda', index=0))
    reader.tensor(buf560, (2, 16, 528), is_leaf=True)  # getitem_457
    buf561 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf561, (3,), dtype=torch.int32, is_leaf=True)  # getitem_458
    buf562 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf562, (3,), dtype=torch.int32, is_leaf=True)  # getitem_459
    buf563 = reader.storage('6de79f4d451761873a72c46a249b835d43d5f5e8', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf563, (), dtype=torch.int64, is_leaf=True)  # getitem_462
    buf564 = reader.storage('6de79f4d451761873a72c46a249b835d43d5f5e8', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf564, (), dtype=torch.int64, is_leaf=True)  # getitem_463
    reader.tensor(buf559, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_760
    buf565 = reader.storage('dedcf11df504e1d134bfe053306842b4bd61c1dc', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf565, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_240
    buf566 = reader.storage('439b5095b123291dc458c7002d4b2d78d9cfe8ed', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf566, (2, 513, 1), is_leaf=True)  # getitem_466
    buf567 = reader.storage('0d776bbf6c52c2b6272d56c9a915d77eb234984d', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf567, (2, 513, 1), is_leaf=True)  # rsqrt_71
    buf568 = reader.storage('c3661df7f395d60088e91fb6098f9f3189d408cc', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf568, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_762
    buf569 = reader.storage('12414172f77955d471b4fc3d469fb9507dc2b508', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf569, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_425
    buf570 = reader.storage('7262286da8a64e72f70ed5a7ba24cb5b56de950a', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf570, (2, 16, 128, 64), (131072, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_426
    buf571 = reader.storage('d2d287531d3b9edbf0a50a421b6c2a88b99634e4', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf571, (2, 16, 128, 64), (131072, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # permute_427
    buf572 = reader.storage('5ff5006198a30d3dd6c47884957117913fd85aa4', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf572, (2, 16, 513, 64), (525312, 64, 1024, 1), dtype=torch.float16, is_leaf=True)  # getitem_467
    buf573 = reader.storage('c97feb5aa2bdcacae027d544bb0a7a61707c6e58', 67584, device=device(type='cuda', index=0))
    reader.tensor(buf573, (2, 16, 528), is_leaf=True)  # getitem_468
    buf574 = reader.storage('a7f9e74f029ee5e84909ee8161ed548679df3358', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf574, (3,), dtype=torch.int32, is_leaf=True)  # getitem_469
    buf575 = reader.storage('1cbaf6f990ca04e947eaf166e1fbcbd05ce76bf1', 12, device=device(type='cuda', index=0), dtype_hint=torch.int32)
    reader.tensor(buf575, (3,), dtype=torch.int32, is_leaf=True)  # getitem_470
    buf576 = reader.storage('6de79f4d451761873a72c46a249b835d43d5f5e8', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf576, (), dtype=torch.int64, is_leaf=True)  # getitem_473
    buf577 = reader.storage('6de79f4d451761873a72c46a249b835d43d5f5e8', 8, device=device(type='cuda', index=0), dtype_hint=torch.int64)
    reader.tensor(buf577, (), dtype=torch.int64, is_leaf=True)  # getitem_474
    reader.tensor(buf572, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_772
    buf578 = reader.storage('4cb4b84aa592aed9fc76fd404df5e6ca8af54012', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf578, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_243
    buf579 = reader.storage('8e730b8fb9fb4490e976c8adc5b55d1fdca0436c', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf579, (2, 513, 1), is_leaf=True)  # getitem_477
    buf580 = reader.storage('308da1033f17c04ac565d82b37c9fa9fc87e6b71', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf580, (2, 513, 1), is_leaf=True)  # rsqrt_72
    buf581 = reader.storage('7515463a015125c74cbfae0993a2431e4f2fe244', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf581, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_774
    buf582 = reader.storage('fc654ba6a616d000e993c0dde9fed77d8cce1c56', 16809984, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf582, (2, 513, 4096), (4202496, 8192, 1), dtype=torch.float16, is_leaf=True)  # getitem_478
    reader.tensor(buf582, (2, 513, 4096), (4202496, 8192, 1), dtype=torch.float16, storage_offset=4096, is_leaf=True)  # getitem_479
    buf583 = reader.storage('2fe40076563f24451623ce7678642435b85a3560', 8404992, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf583, (1026, 4096), dtype=torch.float16, is_leaf=True)  # view_776
    buf584 = reader.storage('0fdb851b8e813dad72a8f767db0c5b852804082c', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf584, (2, 513, 1024), dtype=torch.float16, is_leaf=True)  # add_246
    buf585 = reader.storage('fe581b81c6f43f3edcb76dd482fd9ef120c9088a', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf585, (2, 513, 1), is_leaf=True)  # getitem_481
    buf586 = reader.storage('179711a4f740ca6dbac1901315e3953b4d0a8345', 4104, device=device(type='cuda', index=0))
    reader.tensor(buf586, (2, 513, 1), is_leaf=True)  # rsqrt_73
    buf587 = reader.storage('0d0a03c3ded769df9ef3ae6c32f50bfd04113562', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf587, (1026, 1024), dtype=torch.float16, is_leaf=True)  # view_778
    buf588 = reader.storage('a142c19b02ff61b74d45011fbe491c9defb721b6', 65664, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf588, (2, 32, 512), (16416, 1, 32), dtype=torch.float16, storage_offset=32, is_leaf=True)  # slice_263
    buf589 = reader.storage('2d88809871846a7c71ef536440284e93671a5a82', 2048, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf589, (32, 32, 1), dtype=torch.float16, is_leaf=True)  # convert_element_type_572
    buf590 = reader.storage('1333668f175a7bb146c455e95b79dd9da6d6f5d8', 65536, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf590, (32, 1024), dtype=torch.float16, is_leaf=True)  # permute_435
    buf591 = reader.storage('6ce69334bc0a732555bc5bfc82b4952e6a7fa6ee', 8388608, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf591, (1024, 4096), dtype=torch.float16, is_leaf=True)  # permute_439
    buf592 = reader.storage('14ef142d8af1ab1b1068b64ae0c9421d4f1705a3', 16777216, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf592, (8192, 1024), dtype=torch.float16, is_leaf=True)  # permute_444
    buf593 = reader.storage('91992d6c0701da87f8fbf6351b30752879a86255', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf593, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_450
    buf594 = reader.storage('fc11945a8b543e820ef515368e8ac8c5c8a756b9', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf594, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_458
    buf595 = reader.storage('718f44c18e37339eaeeecd6a442e630038a9d402', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf595, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_462
    buf596 = reader.storage('5249fcabbaa3262131d9b8e0efe9176b825d5ab2', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf596, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_466
    buf597 = reader.storage('877424955ae87bcf06fde737f54d47d8a7b8c441', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf597, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_470
    buf598 = reader.storage('d6edc2c2cfe84a6155a6569ed439d02711ab5ca4', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf598, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_478
    buf599 = reader.storage('5988de566208ecf6c98736ea3a2a3a5e71d7b2f6', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf599, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_482
    buf600 = reader.storage('7225c85d371ae5eb95730776af3807138a41b7f1', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf600, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_486
    buf601 = reader.storage('f081ba6dba4248961f2c216d683d176c17b0493e', 8388608, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf601, (1024, 4096), dtype=torch.float16, is_leaf=True)  # permute_488
    buf602 = reader.storage('db14806de2c8d25b26bb5736e5be190327fca7ad', 16777216, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf602, (8192, 1024), dtype=torch.float16, is_leaf=True)  # permute_493
    buf603 = reader.storage('15034227a568374ee6c642f57170fccfc8422a6c', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf603, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_499
    buf604 = reader.storage('ba937955f31a9f7b1ebfd567673d9070fce319a1', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf604, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_507
    buf605 = reader.storage('e5b512b807af4e580057976c42254fd7ce40c3b7', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf605, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_511
    buf606 = reader.storage('6258f8fa20f0c4aede26ffd7ceba720a2b3b13f6', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf606, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_515
    buf607 = reader.storage('f997e7d86f68835cd14f4b7426e06f9e3528c1a9', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf607, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_519
    buf608 = reader.storage('504906192eb2e579a4dcf511b79f8e7d181c790c', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf608, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_527
    buf609 = reader.storage('7bd46da3bc26264a7f888e1b8da8ac7b3c76fdaf', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf609, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_531
    buf610 = reader.storage('0fd3e5f4fd37735761ea492cda3121671ee87719', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf610, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_535
    buf611 = reader.storage('98e6b2a0369463af0046a725bf1b06a795fae4ac', 8388608, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf611, (1024, 4096), dtype=torch.float16, is_leaf=True)  # permute_537
    buf612 = reader.storage('4b1899e59a5679beb7af46b51e0d2fe2dbec416f', 16777216, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf612, (8192, 1024), dtype=torch.float16, is_leaf=True)  # permute_542
    buf613 = reader.storage('3941e9a57ee2a107f3a21522249bac87b7df807c', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf613, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_548
    buf614 = reader.storage('d39916a38db2d0a90f9c2bb94c8139b970b34618', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf614, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_556
    buf615 = reader.storage('b04f87bfa5e058864b3f0c494593335600a1c733', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf615, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_560
    buf616 = reader.storage('c26bccd9b1af281b2d19a1f04bbf235c4d2ef8f9', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf616, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_564
    buf617 = reader.storage('526d15f47861d8a6ec8becf14e7ad63f02ce7d73', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf617, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_568
    buf618 = reader.storage('378260ccaecbc579347c7cf8a7cc746b2b307fbd', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf618, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_576
    buf619 = reader.storage('c9328871cee8bd64257c494e0b653fe3591a892c', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf619, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_580
    buf620 = reader.storage('26a5c1101e680a6fecffa71833bbf7d5920f7f0f', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf620, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_584
    buf621 = reader.storage('3f77f53e5a755fa37465ee4fb39eede3704643c2', 8388608, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf621, (1024, 4096), dtype=torch.float16, is_leaf=True)  # permute_586
    buf622 = reader.storage('2eb82291df259d931c8cf6bfbcba12f25d5c2a2d', 16777216, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf622, (8192, 1024), dtype=torch.float16, is_leaf=True)  # permute_591
    buf623 = reader.storage('8a4d3fa646ab7e8708ca09c82f9585a9982610be', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf623, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_597
    buf624 = reader.storage('3b01750dd00d52e7e23137232d2831bfff71315d', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf624, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_605
    buf625 = reader.storage('9998744f81cac3335e7dec64531ffa5b4a192619', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf625, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_609
    buf626 = reader.storage('2b062b7243e19ee99095178cc280cc1b99b5e8d7', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf626, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_613
    buf627 = reader.storage('fa976cb1160fb03a8872df388d5f83557de3532e', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf627, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_617
    buf628 = reader.storage('b733180e94235951b4774e497e38e62b4f0c1c14', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf628, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_625
    buf629 = reader.storage('8a1cd0771e59be79d7f08c2a66aaf919729cf327', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf629, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_629
    buf630 = reader.storage('7dd5530d257d690948a663b36a3eb56d4607dda0', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf630, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_633
    buf631 = reader.storage('c333fe3d339fb03996adcd600ed9e86a2324e49a', 8388608, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf631, (1024, 4096), dtype=torch.float16, is_leaf=True)  # permute_635
    buf632 = reader.storage('04a405a7ea8f96a79ecf6893d1cb7e90648f78eb', 16777216, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf632, (8192, 1024), dtype=torch.float16, is_leaf=True)  # permute_640
    buf633 = reader.storage('a3a8f9a9558d9d58ef7a1e6e968b9c38da368686', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf633, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_646
    buf634 = reader.storage('40737a55374b48a0adf2a91feb40cf9bf32cc3ae', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf634, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_654
    buf635 = reader.storage('5e72bd88cee73828282210548ae2889377c0f88c', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf635, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_658
    buf636 = reader.storage('8b7ae9589f4dd47965a3e27c6d44361a87061530', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf636, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_662
    buf637 = reader.storage('4f861d135b97510eae6fbe373dbc12acd3dd4d94', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf637, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_666
    buf638 = reader.storage('cb0cc0c4294132e6c45e1579ec43905a13ce1ae3', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf638, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_674
    buf639 = reader.storage('1490f6b6130ddd9ffae3fa932acf976232551bd3', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf639, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_678
    buf640 = reader.storage('0ce3dcf13ba78b1b56e76c635d43662ffe1be798', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf640, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_682
    buf641 = reader.storage('687267c7419e4aeb478793ca34f17a8bbfe06369', 8388608, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf641, (1024, 4096), dtype=torch.float16, is_leaf=True)  # permute_684
    buf642 = reader.storage('0f734569ff4dbfa66b5b3ab9fabd1ef52b50e563', 16777216, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf642, (8192, 1024), dtype=torch.float16, is_leaf=True)  # permute_689
    buf643 = reader.storage('6721e54d0540ccbdb66edb7d5204fb0f78a31385', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf643, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_695
    buf644 = reader.storage('9ee9091e8c3867a74d58276098a7b6e9d5e2d9b4', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf644, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_703
    buf645 = reader.storage('d34eea7249dbd6a34d4f2adb8e11e613bfb65091', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf645, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_707
    buf646 = reader.storage('4fa4358414688432fa25533a4aa424a01afa4ac1', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf646, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_711
    buf647 = reader.storage('8248ba6f1938f3299701514c267f464462e1a2fb', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf647, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_715
    buf648 = reader.storage('9d58b0ccfd8d12fe92d4406431bafcc0b208261f', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf648, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_723
    buf649 = reader.storage('0762ac514156d0c4661d629524ee6ec5d7ac3ca1', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf649, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_727
    buf650 = reader.storage('acb10c9d0e8c0658b7bea39e3b594afb9154a50b', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf650, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_731
    buf651 = reader.storage('e4694b5abf73a4b0bee6c36617dbde4b3582c6de', 8388608, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf651, (1024, 4096), dtype=torch.float16, is_leaf=True)  # permute_733
    buf652 = reader.storage('03cd047e52cd0d4e73005a9e777ca06e985da1dd', 16777216, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf652, (8192, 1024), dtype=torch.float16, is_leaf=True)  # permute_738
    buf653 = reader.storage('bb06a5d3cc5d7e154b5c1a8c0070716ddc98c6b4', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf653, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_744
    buf654 = reader.storage('f85adac30328be04d33a4c2d639422481922a5b8', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf654, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_752
    buf655 = reader.storage('9f6111bb3729325f7edb4d525d48e6940c20f333', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf655, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_756
    buf656 = reader.storage('d8964a4a8dd6e78ce9cb9b7c7becb854a9ffe02f', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf656, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_760
    buf657 = reader.storage('7f2de893fc16dce068e276a1c7f983f01ca8ea14', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf657, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_764
    buf658 = reader.storage('343d605e3d1881b34afd7f3c15e0dad3d589113e', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf658, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_772
    buf659 = reader.storage('3fd106d06a707178b67375ebeb0e9ddc92e6f931', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf659, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_776
    buf660 = reader.storage('13a3c49c040627b92a205e4dfce0d08a1ac3dde4', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf660, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_780
    buf661 = reader.storage('56ece6e360dccdd6b6d3f7ad3b42368b42ed8613', 8388608, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf661, (1024, 4096), dtype=torch.float16, is_leaf=True)  # permute_782
    buf662 = reader.storage('db8aef2831d5fa4e2908dc45f40e5297c1bcaa1e', 16777216, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf662, (8192, 1024), dtype=torch.float16, is_leaf=True)  # permute_787
    buf663 = reader.storage('ce6350862c4506aa2ba973f78181be435f14bb20', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf663, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_793
    buf664 = reader.storage('e85f13d3c4b900d86f0c4c194ad5753dd54d2a75', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf664, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_801
    buf665 = reader.storage('0533cba002871ea54f3556e0f178105530bbd50b', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf665, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_805
    buf666 = reader.storage('e38960ed2c369c31c82025b365b4be34146e82f7', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf666, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_809
    buf667 = reader.storage('446ab9fba0b4eeea53815fcdef3df7fe78a2b0e0', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf667, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_813
    buf668 = reader.storage('d6a28a0e44daee8f6115379d142ff051f09d071c', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf668, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_821
    buf669 = reader.storage('b5ad9b8ac321571f7c07695f5d55bec12fd64061', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf669, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_825
    buf670 = reader.storage('261a9ee7bb85244935eabebf55e9642e15e667f3', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf670, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_829
    buf671 = reader.storage('350914f682871f3557332e8e87a70456136d045a', 8388608, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf671, (1024, 4096), dtype=torch.float16, is_leaf=True)  # permute_831
    buf672 = reader.storage('2f69cdb3087ffb6757241ebc0d763b551b60a90b', 16777216, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf672, (8192, 1024), dtype=torch.float16, is_leaf=True)  # permute_836
    buf673 = reader.storage('ecc3905c12fcb221f24263e9e0be0ed21f5fd81e', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf673, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_842
    buf674 = reader.storage('6dc585275d6e55016ea4df23a2d2ba785921cfb0', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf674, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_850
    buf675 = reader.storage('58a10925550fe6e7cd218e105ab9a4f461462943', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf675, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_854
    buf676 = reader.storage('2fc85e6b8e9b1d027494bfe75cb34583439f0207', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf676, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_858
    buf677 = reader.storage('51d3dcc45fffbb73a8013577db914f533b35b906', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf677, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_862
    buf678 = reader.storage('ec2f772e6af0f580d2243613ad3a6f3e70aac418', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf678, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_870
    buf679 = reader.storage('9bae67c42f0363280b6faa5877155c29b7a972a3', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf679, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_874
    buf680 = reader.storage('cda066ce6d35ddbfe3764692a787fe22aed3bf90', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf680, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_878
    buf681 = reader.storage('c544c0cbf1f44e104d6132dd70e94b30aff758d5', 8388608, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf681, (1024, 4096), dtype=torch.float16, is_leaf=True)  # permute_880
    buf682 = reader.storage('b0e17e12868dca1460dc8351373ad091a076ae75', 16777216, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf682, (8192, 1024), dtype=torch.float16, is_leaf=True)  # permute_885
    buf683 = reader.storage('3165c63df5e36bc2937a068a61d9d22367a6d8b9', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf683, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_891
    buf684 = reader.storage('2722374fae04237f7ddf0fc1746352c7e60ac74e', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf684, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_899
    buf685 = reader.storage('19bd416e8c1ce78dd173484e9ede03aa84b23c71', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf685, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_903
    buf686 = reader.storage('7f2114268a449b964e28983f94173100600b16fe', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf686, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_907
    buf687 = reader.storage('b931d4073b988bb404eb39ba418ceff9fd3d9bfd', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf687, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_911
    buf688 = reader.storage('a51680c956ec0629fd1bca1ae24da808e241e1bb', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf688, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_919
    buf689 = reader.storage('e8a1221ebd148f3140119464004f488b7d748b86', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf689, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_923
    buf690 = reader.storage('be65bc1c078e885ea9317f66bc6a7c1e613043ee', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf690, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_927
    buf691 = reader.storage('29b565f087eb4969e061bbc2fb81aef56763510d', 8388608, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf691, (1024, 4096), dtype=torch.float16, is_leaf=True)  # permute_929
    buf692 = reader.storage('afc24915c9d82ea954a1613d8e9e112aa871c5d4', 16777216, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf692, (8192, 1024), dtype=torch.float16, is_leaf=True)  # permute_934
    buf693 = reader.storage('7e0a7de49cc85177a352ffbb2d84fc7658722c0c', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf693, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_940
    buf694 = reader.storage('470028030989db305128a20b5fd30eee4aa41cb6', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf694, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_948
    buf695 = reader.storage('1b7c962c911c86e98b6e2a1ad4926460da7cf6e8', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf695, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_952
    buf696 = reader.storage('7488cc730ba4d599f9d12740b4a788f00d717037', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf696, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_956
    buf697 = reader.storage('441625057bd81cca6ba4c28a0ff6edc1b85e9beb', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf697, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_960
    buf698 = reader.storage('d83e9fbf85ffcc5b8a3455c53224a63d0e917238', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf698, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_968
    buf699 = reader.storage('4f56265a7a1f985a4d2ee472956a6a24d7223749', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf699, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_972
    buf700 = reader.storage('dd859bdfbb27dbf0e4a62c4b2cbcf457527234f8', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf700, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_976
    buf701 = reader.storage('f3070bbc76be83b8c4f66120ca4347521d3b6cb0', 8388608, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf701, (1024, 4096), dtype=torch.float16, is_leaf=True)  # permute_978
    buf702 = reader.storage('b02e8c5c5dd861a0f06af1419d6a6780c89502b4', 16777216, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf702, (8192, 1024), dtype=torch.float16, is_leaf=True)  # permute_983
    buf703 = reader.storage('5661b6f8403e4800cd4cb30cb484245740b8841a', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf703, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_989
    buf704 = reader.storage('d6fd2745dc7b4b20acb925ad3f5424207af1a818', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf704, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_997
    buf705 = reader.storage('7a352b0b202b334bfc2d2d947202644d6ca72411', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf705, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_1001
    buf706 = reader.storage('62bbab8106712c1651e33f486460d40d8641feed', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf706, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_1005
    buf707 = reader.storage('865f6ba09d9b3ce18176c4a1378b75dd49e98534', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf707, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_1009
    buf708 = reader.storage('02d404ed00cac0673b15b82edd73e926c78cec0f', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf708, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_1017
    buf709 = reader.storage('5bb3ac14a50ff0641a052356e8ceaac175aeeaf6', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf709, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_1021
    buf710 = reader.storage('0e641fbf26dd6ab0701066868845e437bb21aef3', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf710, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_1025
    buf711 = reader.storage('928b9fb0413270d29b8e848dfc6248a1312754c8', 8388608, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf711, (1024, 4096), dtype=torch.float16, is_leaf=True)  # permute_1027
    buf712 = reader.storage('57ccc9ad7f827eed6a2e06787ea30e194d987ca7', 16777216, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf712, (8192, 1024), dtype=torch.float16, is_leaf=True)  # permute_1032
    buf713 = reader.storage('89c76bbfc9355b4282e86a7e4663b26d56f39240', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf713, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_1038
    buf714 = reader.storage('782bc04fbe6e866733762fda4667b518772d4fc3', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf714, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_1046
    buf715 = reader.storage('1b64ca01215ff5148406ba209af42999f60879a1', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf715, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_1050
    buf716 = reader.storage('096f564e131f43a9ac8408841140eca0180a9394', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf716, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_1054
    buf717 = reader.storage('b02db646f2f58c8ed912345df8fee95404a40ac0', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf717, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_1058
    buf718 = reader.storage('3dcffc3695fdbbc2083e326c8308d9a8bcd86b46', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf718, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_1066
    buf719 = reader.storage('bef8948839015a62d6e3c91fe1d2045db1b779f7', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf719, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_1070
    buf720 = reader.storage('1b86889d121c03b0d653250c4106809b076d4b58', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf720, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_1074
    buf721 = reader.storage('1e166b8022e3af058716858c709dc39e911b80e7', 8388608, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf721, (1024, 4096), dtype=torch.float16, is_leaf=True)  # permute_1076
    buf722 = reader.storage('e4674f31379c60481c0129106c1f93acd340a2f6', 16777216, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf722, (8192, 1024), dtype=torch.float16, is_leaf=True)  # permute_1081
    buf723 = reader.storage('d71ec7fd4f37c5745432c5559736cebb2ed5abd9', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf723, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_1087
    buf724 = reader.storage('30debd51b05a1a07e78f07a377d8a3b28d8ffca5', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf724, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_1095
    buf725 = reader.storage('3613394e7da6711c145712a8e59938ef25ddf627', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf725, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_1099
    buf726 = reader.storage('b895d3395abaafe4c8b4cc468a1d0c4a470f2733', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf726, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_1103
    buf727 = reader.storage('c1502dcf91009508ce2bd7993733d1a1c6953384', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf727, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_1107
    buf728 = reader.storage('d7c7271936e37836a4f051f7eebe1c70f7ad28ae', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf728, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_1115
    buf729 = reader.storage('84e31d9e4cb26b4149b05443f48b70b36f50bead', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf729, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_1119
    buf730 = reader.storage('ce7825f2c1d7281f2a33c97f034b85a3096a12b6', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf730, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_1123
    buf731 = reader.storage('ab8578318795eabfe244bf829fe6f13adf564b9b', 8388608, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf731, (1024, 4096), dtype=torch.float16, is_leaf=True)  # permute_1125
    buf732 = reader.storage('e03489b7fd58406f7d9fff76b6a3c5cd934de12c', 16777216, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf732, (8192, 1024), dtype=torch.float16, is_leaf=True)  # permute_1130
    buf733 = reader.storage('61000eeb7a4c945b228b39fa45220bfdd99929a4', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf733, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_1136
    buf734 = reader.storage('2ae81c0dc5b08ed636d929d2e6e7432f7f0e39c4', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf734, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_1144
    buf735 = reader.storage('c23047cd0b73ccee87b4855caad9ab088d5296eb', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf735, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_1148
    buf736 = reader.storage('1565aecab26e2add1a9e0c7324c982b2872841b3', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf736, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_1152
    buf737 = reader.storage('129697e170bec65c174ea37e47a1e08f0b718afa', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf737, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_1156
    buf738 = reader.storage('28cf6a90bb44b88bb719809ecb908a2954475e43', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf738, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_1164
    buf739 = reader.storage('e4dbbb412bf8adf4bdc70cef57bef1f63c5d75e0', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf739, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_1168
    buf740 = reader.storage('7ab89e58d5e1e30577abd5effbd86bbf38036a5d', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf740, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_1172
    buf741 = reader.storage('43c52c21bee33bea13e199f35d8973d6edad718a', 8388608, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf741, (1024, 4096), dtype=torch.float16, is_leaf=True)  # permute_1174
    buf742 = reader.storage('564bc7c993c6c2277846d92711d45c48de7b1877', 16777216, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf742, (8192, 1024), dtype=torch.float16, is_leaf=True)  # permute_1179
    buf743 = reader.storage('62102b5cd235d76e28e0111d67961bf81047a9f7', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf743, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_1185
    buf744 = reader.storage('bafe8152acf97ffd48f64684e00192920b964f9a', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf744, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_1193
    buf745 = reader.storage('3343d09c697b9c3ae4284b958364b0d4bfbbb49f', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf745, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_1197
    buf746 = reader.storage('b2ee5caf652b10735267b51341a4c8da9c3a2ab4', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf746, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_1201
    buf747 = reader.storage('f4a72b1cc34d5efcb2797c73cb49913a57e60f23', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf747, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_1205
    buf748 = reader.storage('03c798ad6c3c181046136d3e8855019597a00018', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf748, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_1213
    buf749 = reader.storage('693d1fdc42e5fa1e4343782773433cf1e201c83c', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf749, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_1217
    buf750 = reader.storage('e24af27041541b3b2a5cb2c84fff78b69159734f', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf750, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_1221
    buf751 = reader.storage('dbfaf9e558dff76ed47db41be816c33edd0df4b9', 65536, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf751, (1024, 32), dtype=torch.float16, is_leaf=True)  # permute_1225
    buf752 = reader.storage('17d59617dba308084591d40c0e863a969367e323', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf752, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_1228
    buf753 = reader.storage('302cbe1ad4053f732abd70ff2c25f048bdac2431', 1572864, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf753, (1024, 768), dtype=torch.float16, is_leaf=True)  # permute_1233
    buf754 = reader.storage('ccc2e1fca7f8294b30a5afcab58a501309dfa933', 2097152, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf754, (1024, 1024), dtype=torch.float16, is_leaf=True)  # permute_1239
    buf755 = reader.storage('b2dbdeb8f1377fc8175e303b745728e810c4279c', 1572864, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf755, (1024, 768), dtype=torch.float16, is_leaf=True)  # permute_1244
    buf756 = reader.storage('49cf10c0b12563ca2b027a7a915d223fa19e8cfd', 1179648, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf756, (768, 768), dtype=torch.float16, is_leaf=True)  # permute_1246
    buf757 = reader.storage('e562d91623bbf637fb07a57f978202ee19e68c40', 394752, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf757, (768, 257), dtype=torch.float16, is_leaf=True)  # permute_1254
    buf758 = reader.storage('2ca18a43909aba68e389d7f3f820398e0783d556', 65536, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf758, (2, 32, 512), dtype=torch.float16, is_leaf=True)  # tangents_1
    buf759 = reader.storage('1b2b7e957b3a212a7fea15ed5dab5b646e0a494d', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf759, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # tangents_2
    buf760 = reader.storage('1b2b7e957b3a212a7fea15ed5dab5b646e0a494d', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf760, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # tangents_3
    buf761 = reader.storage('65beb0b1081345a9374cd1e755797afb287e5bf4', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf761, (2, 16, 128, 64), dtype=torch.float16, is_leaf=True)  # tangents_4
    buf762 = reader.storage('65beb0b1081345a9374cd1e755797afb287e5bf4', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf762, (2, 16, 128, 64), dtype=torch.float16, is_leaf=True)  # tangents_5
    buf763 = reader.storage('1b2b7e957b3a212a7fea15ed5dab5b646e0a494d', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf763, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # tangents_6
    buf764 = reader.storage('1b2b7e957b3a212a7fea15ed5dab5b646e0a494d', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf764, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # tangents_7
    buf765 = reader.storage('65beb0b1081345a9374cd1e755797afb287e5bf4', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf765, (2, 16, 128, 64), dtype=torch.float16, is_leaf=True)  # tangents_8
    buf766 = reader.storage('65beb0b1081345a9374cd1e755797afb287e5bf4', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf766, (2, 16, 128, 64), dtype=torch.float16, is_leaf=True)  # tangents_9
    buf767 = reader.storage('1b2b7e957b3a212a7fea15ed5dab5b646e0a494d', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf767, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # tangents_10
    buf768 = reader.storage('1b2b7e957b3a212a7fea15ed5dab5b646e0a494d', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf768, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # tangents_11
    buf769 = reader.storage('65beb0b1081345a9374cd1e755797afb287e5bf4', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf769, (2, 16, 128, 64), dtype=torch.float16, is_leaf=True)  # tangents_12
    buf770 = reader.storage('65beb0b1081345a9374cd1e755797afb287e5bf4', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf770, (2, 16, 128, 64), dtype=torch.float16, is_leaf=True)  # tangents_13
    buf771 = reader.storage('1b2b7e957b3a212a7fea15ed5dab5b646e0a494d', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf771, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # tangents_14
    buf772 = reader.storage('1b2b7e957b3a212a7fea15ed5dab5b646e0a494d', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf772, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # tangents_15
    buf773 = reader.storage('65beb0b1081345a9374cd1e755797afb287e5bf4', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf773, (2, 16, 128, 64), dtype=torch.float16, is_leaf=True)  # tangents_16
    buf774 = reader.storage('65beb0b1081345a9374cd1e755797afb287e5bf4', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf774, (2, 16, 128, 64), dtype=torch.float16, is_leaf=True)  # tangents_17
    buf775 = reader.storage('1b2b7e957b3a212a7fea15ed5dab5b646e0a494d', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf775, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # tangents_18
    buf776 = reader.storage('1b2b7e957b3a212a7fea15ed5dab5b646e0a494d', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf776, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # tangents_19
    buf777 = reader.storage('65beb0b1081345a9374cd1e755797afb287e5bf4', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf777, (2, 16, 128, 64), dtype=torch.float16, is_leaf=True)  # tangents_20
    buf778 = reader.storage('65beb0b1081345a9374cd1e755797afb287e5bf4', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf778, (2, 16, 128, 64), dtype=torch.float16, is_leaf=True)  # tangents_21
    buf779 = reader.storage('1b2b7e957b3a212a7fea15ed5dab5b646e0a494d', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf779, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # tangents_22
    buf780 = reader.storage('1b2b7e957b3a212a7fea15ed5dab5b646e0a494d', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf780, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # tangents_23
    buf781 = reader.storage('65beb0b1081345a9374cd1e755797afb287e5bf4', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf781, (2, 16, 128, 64), dtype=torch.float16, is_leaf=True)  # tangents_24
    buf782 = reader.storage('65beb0b1081345a9374cd1e755797afb287e5bf4', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf782, (2, 16, 128, 64), dtype=torch.float16, is_leaf=True)  # tangents_25
    buf783 = reader.storage('1b2b7e957b3a212a7fea15ed5dab5b646e0a494d', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf783, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # tangents_26
    buf784 = reader.storage('1b2b7e957b3a212a7fea15ed5dab5b646e0a494d', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf784, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # tangents_27
    buf785 = reader.storage('65beb0b1081345a9374cd1e755797afb287e5bf4', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf785, (2, 16, 128, 64), dtype=torch.float16, is_leaf=True)  # tangents_28
    buf786 = reader.storage('65beb0b1081345a9374cd1e755797afb287e5bf4', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf786, (2, 16, 128, 64), dtype=torch.float16, is_leaf=True)  # tangents_29
    buf787 = reader.storage('1b2b7e957b3a212a7fea15ed5dab5b646e0a494d', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf787, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # tangents_30
    buf788 = reader.storage('1b2b7e957b3a212a7fea15ed5dab5b646e0a494d', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf788, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # tangents_31
    buf789 = reader.storage('65beb0b1081345a9374cd1e755797afb287e5bf4', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf789, (2, 16, 128, 64), dtype=torch.float16, is_leaf=True)  # tangents_32
    buf790 = reader.storage('65beb0b1081345a9374cd1e755797afb287e5bf4', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf790, (2, 16, 128, 64), dtype=torch.float16, is_leaf=True)  # tangents_33
    buf791 = reader.storage('1b2b7e957b3a212a7fea15ed5dab5b646e0a494d', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf791, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # tangents_34
    buf792 = reader.storage('1b2b7e957b3a212a7fea15ed5dab5b646e0a494d', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf792, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # tangents_35
    buf793 = reader.storage('65beb0b1081345a9374cd1e755797afb287e5bf4', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf793, (2, 16, 128, 64), dtype=torch.float16, is_leaf=True)  # tangents_36
    buf794 = reader.storage('65beb0b1081345a9374cd1e755797afb287e5bf4', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf794, (2, 16, 128, 64), dtype=torch.float16, is_leaf=True)  # tangents_37
    buf795 = reader.storage('1b2b7e957b3a212a7fea15ed5dab5b646e0a494d', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf795, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # tangents_38
    buf796 = reader.storage('1b2b7e957b3a212a7fea15ed5dab5b646e0a494d', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf796, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # tangents_39
    buf797 = reader.storage('65beb0b1081345a9374cd1e755797afb287e5bf4', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf797, (2, 16, 128, 64), dtype=torch.float16, is_leaf=True)  # tangents_40
    buf798 = reader.storage('65beb0b1081345a9374cd1e755797afb287e5bf4', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf798, (2, 16, 128, 64), dtype=torch.float16, is_leaf=True)  # tangents_41
    buf799 = reader.storage('1b2b7e957b3a212a7fea15ed5dab5b646e0a494d', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf799, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # tangents_42
    buf800 = reader.storage('1b2b7e957b3a212a7fea15ed5dab5b646e0a494d', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf800, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # tangents_43
    buf801 = reader.storage('65beb0b1081345a9374cd1e755797afb287e5bf4', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf801, (2, 16, 128, 64), dtype=torch.float16, is_leaf=True)  # tangents_44
    buf802 = reader.storage('65beb0b1081345a9374cd1e755797afb287e5bf4', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf802, (2, 16, 128, 64), dtype=torch.float16, is_leaf=True)  # tangents_45
    buf803 = reader.storage('1b2b7e957b3a212a7fea15ed5dab5b646e0a494d', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf803, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # tangents_46
    buf804 = reader.storage('1b2b7e957b3a212a7fea15ed5dab5b646e0a494d', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf804, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # tangents_47
    buf805 = reader.storage('65beb0b1081345a9374cd1e755797afb287e5bf4', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf805, (2, 16, 128, 64), dtype=torch.float16, is_leaf=True)  # tangents_48
    buf806 = reader.storage('65beb0b1081345a9374cd1e755797afb287e5bf4', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf806, (2, 16, 128, 64), dtype=torch.float16, is_leaf=True)  # tangents_49
    buf807 = reader.storage('1b2b7e957b3a212a7fea15ed5dab5b646e0a494d', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf807, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # tangents_50
    buf808 = reader.storage('1b2b7e957b3a212a7fea15ed5dab5b646e0a494d', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf808, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # tangents_51
    buf809 = reader.storage('65beb0b1081345a9374cd1e755797afb287e5bf4', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf809, (2, 16, 128, 64), dtype=torch.float16, is_leaf=True)  # tangents_52
    buf810 = reader.storage('65beb0b1081345a9374cd1e755797afb287e5bf4', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf810, (2, 16, 128, 64), dtype=torch.float16, is_leaf=True)  # tangents_53
    buf811 = reader.storage('1b2b7e957b3a212a7fea15ed5dab5b646e0a494d', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf811, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # tangents_54
    buf812 = reader.storage('1b2b7e957b3a212a7fea15ed5dab5b646e0a494d', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf812, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # tangents_55
    buf813 = reader.storage('65beb0b1081345a9374cd1e755797afb287e5bf4', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf813, (2, 16, 128, 64), dtype=torch.float16, is_leaf=True)  # tangents_56
    buf814 = reader.storage('65beb0b1081345a9374cd1e755797afb287e5bf4', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf814, (2, 16, 128, 64), dtype=torch.float16, is_leaf=True)  # tangents_57
    buf815 = reader.storage('1b2b7e957b3a212a7fea15ed5dab5b646e0a494d', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf815, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # tangents_58
    buf816 = reader.storage('1b2b7e957b3a212a7fea15ed5dab5b646e0a494d', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf816, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # tangents_59
    buf817 = reader.storage('65beb0b1081345a9374cd1e755797afb287e5bf4', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf817, (2, 16, 128, 64), dtype=torch.float16, is_leaf=True)  # tangents_60
    buf818 = reader.storage('65beb0b1081345a9374cd1e755797afb287e5bf4', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf818, (2, 16, 128, 64), dtype=torch.float16, is_leaf=True)  # tangents_61
    buf819 = reader.storage('1b2b7e957b3a212a7fea15ed5dab5b646e0a494d', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf819, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # tangents_62
    buf820 = reader.storage('1b2b7e957b3a212a7fea15ed5dab5b646e0a494d', 2101248, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf820, (2, 16, 513, 64), dtype=torch.float16, is_leaf=True)  # tangents_63
    buf821 = reader.storage('65beb0b1081345a9374cd1e755797afb287e5bf4', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf821, (2, 16, 128, 64), dtype=torch.float16, is_leaf=True)  # tangents_64
    buf822 = reader.storage('65beb0b1081345a9374cd1e755797afb287e5bf4', 524288, device=device(type='cuda', index=0), dtype_hint=torch.float16)
    reader.tensor(buf822, (2, 16, 128, 64), dtype=torch.float16, is_leaf=True)  # tangents_65
load_args._version = 0
mod = Repro()
if __name__ == '__main__':
    from torch._dynamo.repro.after_aot import run_repro
    run_repro(mod, load_args, accuracy=True, command='run', save_dir='the-model/torch_compile_debug/run_2024_01_06_18_15_47_002351-pid_2494068/minifier/checkpoints', tracing_mode='real', check_str=None)