aufklarer/Qwen3-ASR-CoreML
4751
1program(1.3)2[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}})]3{4 func main<ios18>(tensor<fp32, [1, 1, 128, 1024]> attention_mask, tensor<fp32, [1, 128, 1024]> input_embeds, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_0, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_1, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_10, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_11, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_12, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_13, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_2, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_3, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_4, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_5, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_6, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_7, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_8, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_9, tensor<int32, [128]> positions, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_0, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_1, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_10, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_11, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_12, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_13, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_2, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_3, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_4, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_5, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_6, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_7, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_8, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_9) {5 int32 var_68_one_hot_vector_size_0 = const()[name = string("op_68_one_hot_vector_size_0"), val = int32(1024)];6 int32 var_68_axis_0 = const()[name = string("op_68_axis_0"), val = int32(-1)];7 int32 var_68_on_value_0 = const()[name = string("op_68_on_value_0"), val = int32(1)];8 int32 var_68_off_value_0 = const()[name = string("op_68_off_value_0"), val = int32(0)];9 tensor<int32, [128, 1024]> var_68 = one_hot(axis = var_68_axis_0, indices = positions, off_value = var_68_off_value_0, on_value = var_68_on_value_0, one_hot_vector_size = var_68_one_hot_vector_size_0)[name = string("op_68")];10 tensor<int32, [1]> var_78_axes_0 = const()[name = string("op_78_axes_0"), val = tensor<int32, [1]>([0])];11 bool var_78_keep_dims_0 = const()[name = string("op_78_keep_dims_0"), val = bool(false)];12 string cast_1_to_fp16_dtype_0 = const()[name = string("cast_1_to_fp16_dtype_0"), val = string("fp16")];13 tensor<fp16, [128, 1024]> var_68_to_fp16 = cast(dtype = cast_1_to_fp16_dtype_0, x = var_68)[name = string("cast_3")];14 tensor<fp16, [1024]> var_78_cast_fp16 = reduce_sum(axes = var_78_axes_0, keep_dims = var_78_keep_dims_0, x = var_68_to_fp16)[name = string("op_78_cast_fp16")];15 tensor<int32, [4]> var_83 = const()[name = string("op_83"), val = tensor<int32, [4]>([1, 1, 1024, 1])];16 tensor<fp16, [1, 1, 1024, 1]> var_84_cast_fp16 = reshape(shape = var_83, x = var_78_cast_fp16)[name = string("op_84_cast_fp16")];17 int32 var_99 = const()[name = string("op_99"), val = int32(-1)];18 string input_embeds_to_fp16_dtype_0 = const()[name = string("input_embeds_to_fp16_dtype_0"), val = string("fp16")];19 fp16 var_98_promoted_to_fp16 = const()[name = string("op_98_promoted_to_fp16"), val = fp16(0x1p+1)];20 tensor<fp16, [1, 128, 1024]> input_embeds_to_fp16 = cast(dtype = input_embeds_to_fp16_dtype_0, x = input_embeds)[name = string("cast_2")];21 tensor<fp16, [1, 128, 1024]> var_108_cast_fp16 = pow(x = input_embeds_to_fp16, y = var_98_promoted_to_fp16)[name = string("op_108_cast_fp16")];22 tensor<int32, [1]> var_110_axes_0 = const()[name = string("op_110_axes_0"), val = tensor<int32, [1]>([-1])];23 bool var_110_keep_dims_0 = const()[name = string("op_110_keep_dims_0"), val = bool(true)];24 tensor<fp16, [1, 128, 1]> var_110_cast_fp16 = reduce_mean(axes = var_110_axes_0, keep_dims = var_110_keep_dims_0, x = var_108_cast_fp16)[name = string("op_110_cast_fp16")];25 fp16 var_111_to_fp16 = const()[name = string("op_111_to_fp16"), val = fp16(0x1.1p-20)];26 tensor<fp16, [1, 128, 1]> var_112_cast_fp16 = add(x = var_110_cast_fp16, y = var_111_to_fp16)[name = string("op_112_cast_fp16")];27 fp32 norm_1_epsilon_0 = const()[name = string("norm_1_epsilon_0"), val = fp32(0x1.197998p-40)];28 tensor<fp16, [1, 128, 1]> norm_1_cast_fp16 = rsqrt(epsilon = norm_1_epsilon_0, x = var_112_cast_fp16)[name = string("norm_1_cast_fp16")];29 tensor<fp16, [1, 128, 1024]> var_114_cast_fp16 = mul(x = input_embeds_to_fp16, y = norm_1_cast_fp16)[name = string("op_114_cast_fp16")];30 tensor<fp16, [1024]> layers_0_input_layernorm_weight_to_fp16 = const()[name = string("layers_0_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];31 tensor<fp16, [1, 128, 1024]> var_115_cast_fp16 = mul(x = var_114_cast_fp16, y = layers_0_input_layernorm_weight_to_fp16)[name = string("op_115_cast_fp16")];32 tensor<fp16, [2048, 1024]> layers_0_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2176))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2099392))))[name = string("layers_0_self_attn_q_proj_weight_to_fp16_palettized")];33 tensor<fp16, [2048]> linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2099968)))];34 tensor<fp16, [1, 128, 2048]> linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_0_self_attn_q_proj_weight_to_fp16_palettized, x = var_115_cast_fp16)[name = string("linear_0_cast_fp16")];35 tensor<int32, [4]> var_131 = const()[name = string("op_131"), val = tensor<int32, [4]>([1, 128, 16, 128])];36 tensor<fp16, [1, 128, 16, 128]> var_132_cast_fp16 = reshape(shape = var_131, x = linear_0_cast_fp16)[name = string("op_132_cast_fp16")];37 tensor<int32, [4]> x_5_perm_0 = const()[name = string("x_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];38 tensor<fp16, [1024, 1024]> layers_0_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2104128))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3152768))))[name = string("layers_0_self_attn_k_proj_weight_to_fp16_palettized")];39 tensor<fp16, [1024]> linear_1_bias_0_to_fp16 = const()[name = string("linear_1_bias_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3153344)))];40 tensor<fp16, [1, 128, 1024]> linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_0_self_attn_k_proj_weight_to_fp16_palettized, x = var_115_cast_fp16)[name = string("linear_1_cast_fp16")];41 tensor<int32, [4]> var_136 = const()[name = string("op_136"), val = tensor<int32, [4]>([1, 128, 8, 128])];42 tensor<fp16, [1, 128, 8, 128]> var_137_cast_fp16 = reshape(shape = var_136, x = linear_1_cast_fp16)[name = string("op_137_cast_fp16")];43 tensor<int32, [4]> x_9_perm_0 = const()[name = string("x_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];44 tensor<fp16, [1024, 1024]> layers_0_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3155456))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4204096))))[name = string("layers_0_self_attn_v_proj_weight_to_fp16_palettized")];45 tensor<fp16, [1, 128, 1024]> linear_2_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_0_self_attn_v_proj_weight_to_fp16_palettized, x = var_115_cast_fp16)[name = string("linear_2_cast_fp16")];46 tensor<int32, [4]> var_141 = const()[name = string("op_141"), val = tensor<int32, [4]>([1, 128, 8, 128])];47 tensor<fp16, [1, 128, 8, 128]> var_142_cast_fp16 = reshape(shape = var_141, x = linear_2_cast_fp16)[name = string("op_142_cast_fp16")];48 tensor<int32, [4]> transpose_56_perm_0 = const()[name = string("transpose_56_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];49 fp16 var_98_promoted_1_to_fp16 = const()[name = string("op_98_promoted_1_to_fp16"), val = fp16(0x1p+1)];50 tensor<fp16, [1, 16, 128, 128]> x_5_cast_fp16 = transpose(perm = x_5_perm_0, x = var_132_cast_fp16)[name = string("transpose_97")];51 tensor<fp16, [1, 16, 128, 128]> var_146_cast_fp16 = pow(x = x_5_cast_fp16, y = var_98_promoted_1_to_fp16)[name = string("op_146_cast_fp16")];52 tensor<int32, [1]> var_148_axes_0 = const()[name = string("op_148_axes_0"), val = tensor<int32, [1]>([-1])];53 bool var_148_keep_dims_0 = const()[name = string("op_148_keep_dims_0"), val = bool(true)];54 tensor<fp16, [1, 16, 128, 1]> var_148_cast_fp16 = reduce_mean(axes = var_148_axes_0, keep_dims = var_148_keep_dims_0, x = var_146_cast_fp16)[name = string("op_148_cast_fp16")];55 fp16 var_149_to_fp16 = const()[name = string("op_149_to_fp16"), val = fp16(0x1.1p-20)];56 tensor<fp16, [1, 16, 128, 1]> var_150_cast_fp16 = add(x = var_148_cast_fp16, y = var_149_to_fp16)[name = string("op_150_cast_fp16")];57 fp32 norm_3_epsilon_0 = const()[name = string("norm_3_epsilon_0"), val = fp32(0x1.197998p-40)];58 tensor<fp16, [1, 16, 128, 1]> norm_3_cast_fp16 = rsqrt(epsilon = norm_3_epsilon_0, x = var_150_cast_fp16)[name = string("norm_3_cast_fp16")];59 tensor<fp16, [1, 16, 128, 128]> var_152_cast_fp16 = mul(x = x_5_cast_fp16, y = norm_3_cast_fp16)[name = string("op_152_cast_fp16")];60 tensor<fp16, [128]> layers_0_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_0_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4204672)))];61 tensor<fp16, [1, 16, 128, 128]> var_153_cast_fp16 = mul(x = var_152_cast_fp16, y = layers_0_self_attn_q_norm_weight_to_fp16)[name = string("op_153_cast_fp16")];62 fp16 var_98_promoted_2_to_fp16 = const()[name = string("op_98_promoted_2_to_fp16"), val = fp16(0x1p+1)];63 tensor<fp16, [1, 8, 128, 128]> x_9_cast_fp16 = transpose(perm = x_9_perm_0, x = var_137_cast_fp16)[name = string("transpose_96")];64 tensor<fp16, [1, 8, 128, 128]> var_157_cast_fp16 = pow(x = x_9_cast_fp16, y = var_98_promoted_2_to_fp16)[name = string("op_157_cast_fp16")];65 tensor<int32, [1]> var_159_axes_0 = const()[name = string("op_159_axes_0"), val = tensor<int32, [1]>([-1])];66 bool var_159_keep_dims_0 = const()[name = string("op_159_keep_dims_0"), val = bool(true)];67 tensor<fp16, [1, 8, 128, 1]> var_159_cast_fp16 = reduce_mean(axes = var_159_axes_0, keep_dims = var_159_keep_dims_0, x = var_157_cast_fp16)[name = string("op_159_cast_fp16")];68 fp16 var_160_to_fp16 = const()[name = string("op_160_to_fp16"), val = fp16(0x1.1p-20)];69 tensor<fp16, [1, 8, 128, 1]> var_161_cast_fp16 = add(x = var_159_cast_fp16, y = var_160_to_fp16)[name = string("op_161_cast_fp16")];70 fp32 norm_5_epsilon_0 = const()[name = string("norm_5_epsilon_0"), val = fp32(0x1.197998p-40)];71 tensor<fp16, [1, 8, 128, 1]> norm_5_cast_fp16 = rsqrt(epsilon = norm_5_epsilon_0, x = var_161_cast_fp16)[name = string("norm_5_cast_fp16")];72 tensor<fp16, [1, 8, 128, 128]> var_163_cast_fp16 = mul(x = x_9_cast_fp16, y = norm_5_cast_fp16)[name = string("op_163_cast_fp16")];73 tensor<fp16, [128]> layers_0_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_0_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4204992)))];74 tensor<fp16, [1, 8, 128, 128]> var_164_cast_fp16 = mul(x = var_163_cast_fp16, y = layers_0_self_attn_k_norm_weight_to_fp16)[name = string("op_164_cast_fp16")];75 tensor<int32, [1]> var_168_axes_0 = const()[name = string("op_168_axes_0"), val = tensor<int32, [1]>([-1])];76 string cast_12_to_fp16_dtype_0 = const()[name = string("cast_12_to_fp16_dtype_0"), val = string("fp16")];77 tensor<fp16, [128]> positions_to_fp16 = cast(dtype = cast_12_to_fp16_dtype_0, x = positions)[name = string("cast_1")];78 tensor<fp16, [128, 1]> var_168_cast_fp16 = expand_dims(axes = var_168_axes_0, x = positions_to_fp16)[name = string("op_168_cast_fp16")];79 tensor<fp16, [64]> layers_0_self_attn_rope_inv_freq_to_fp16 = const()[name = string("layers_0_self_attn_rope_inv_freq_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4205312)))];80 tensor<fp16, [128, 64]> freqs_1_cast_fp16 = mul(x = var_168_cast_fp16, y = layers_0_self_attn_rope_inv_freq_to_fp16)[name = string("freqs_1_cast_fp16")];81 tensor<fp16, [128, 64]> var_170_cast_fp16 = cos(x = freqs_1_cast_fp16)[name = string("op_170_cast_fp16")];82 tensor<int32, [4]> var_172 = const()[name = string("op_172"), val = tensor<int32, [4]>([1, 1, -1, 64])];83 tensor<fp16, [1, 1, 128, 64]> cos_val_1_cast_fp16 = reshape(shape = var_172, x = var_170_cast_fp16)[name = string("cos_val_1_cast_fp16")];84 tensor<fp16, [128, 64]> var_174_cast_fp16 = sin(x = freqs_1_cast_fp16)[name = string("op_174_cast_fp16")];85 tensor<int32, [4]> var_176 = const()[name = string("op_176"), val = tensor<int32, [4]>([1, 1, -1, 64])];86 tensor<fp16, [1, 1, 128, 64]> sin_val_1_cast_fp16 = reshape(shape = var_176, x = var_174_cast_fp16)[name = string("sin_val_1_cast_fp16")];87 tensor<int32, [4]> x1_1_begin_0 = const()[name = string("x1_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];88 tensor<int32, [4]> x1_1_end_0 = const()[name = string("x1_1_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];89 tensor<bool, [4]> x1_1_end_mask_0 = const()[name = string("x1_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];90 tensor<fp16, [1, 16, 128, 64]> x1_1_cast_fp16 = slice_by_index(begin = x1_1_begin_0, end = x1_1_end_0, end_mask = x1_1_end_mask_0, x = var_153_cast_fp16)[name = string("x1_1_cast_fp16")];91 tensor<int32, [4]> x2_1_begin_0 = const()[name = string("x2_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];92 tensor<int32, [4]> x2_1_end_0 = const()[name = string("x2_1_end_0"), val = tensor<int32, [4]>([1, 16, 128, 128])];93 tensor<bool, [4]> x2_1_end_mask_0 = const()[name = string("x2_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];94 tensor<fp16, [1, 16, 128, 64]> x2_1_cast_fp16 = slice_by_index(begin = x2_1_begin_0, end = x2_1_end_0, end_mask = x2_1_end_mask_0, x = var_153_cast_fp16)[name = string("x2_1_cast_fp16")];95 tensor<fp16, [1, 16, 128, 64]> var_185_cast_fp16 = mul(x = x1_1_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_185_cast_fp16")];96 tensor<fp16, [1, 16, 128, 64]> var_186_cast_fp16 = mul(x = x2_1_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_186_cast_fp16")];97 tensor<fp16, [1, 16, 128, 64]> var_187_cast_fp16 = sub(x = var_185_cast_fp16, y = var_186_cast_fp16)[name = string("op_187_cast_fp16")];98 tensor<fp16, [1, 16, 128, 64]> var_188_cast_fp16 = mul(x = x2_1_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_188_cast_fp16")];99 tensor<fp16, [1, 16, 128, 64]> var_189_cast_fp16 = mul(x = x1_1_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_189_cast_fp16")];100 tensor<fp16, [1, 16, 128, 64]> var_190_cast_fp16 = add(x = var_188_cast_fp16, y = var_189_cast_fp16)[name = string("op_190_cast_fp16")];101 bool q_1_interleave_0 = const()[name = string("q_1_interleave_0"), val = bool(false)];102 tensor<fp16, [1, 16, 128, 128]> q_1_cast_fp16 = concat(axis = var_99, interleave = q_1_interleave_0, values = (var_187_cast_fp16, var_190_cast_fp16))[name = string("q_1_cast_fp16")];103 tensor<int32, [4]> x1_3_begin_0 = const()[name = string("x1_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];104 tensor<int32, [4]> x1_3_end_0 = const()[name = string("x1_3_end_0"), val = tensor<int32, [4]>([1, 8, 128, 64])];105 tensor<bool, [4]> x1_3_end_mask_0 = const()[name = string("x1_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];106 tensor<fp16, [1, 8, 128, 64]> x1_3_cast_fp16 = slice_by_index(begin = x1_3_begin_0, end = x1_3_end_0, end_mask = x1_3_end_mask_0, x = var_164_cast_fp16)[name = string("x1_3_cast_fp16")];107 tensor<int32, [4]> x2_3_begin_0 = const()[name = string("x2_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];108 tensor<int32, [4]> x2_3_end_0 = const()[name = string("x2_3_end_0"), val = tensor<int32, [4]>([1, 8, 128, 128])];109 tensor<bool, [4]> x2_3_end_mask_0 = const()[name = string("x2_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];110 tensor<fp16, [1, 8, 128, 64]> x2_3_cast_fp16 = slice_by_index(begin = x2_3_begin_0, end = x2_3_end_0, end_mask = x2_3_end_mask_0, x = var_164_cast_fp16)[name = string("x2_3_cast_fp16")];111 tensor<fp16, [1, 8, 128, 64]> var_212_cast_fp16 = mul(x = x1_3_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_212_cast_fp16")];112 tensor<fp16, [1, 8, 128, 64]> var_213_cast_fp16 = mul(x = x2_3_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_213_cast_fp16")];113 tensor<fp16, [1, 8, 128, 64]> var_214_cast_fp16 = sub(x = var_212_cast_fp16, y = var_213_cast_fp16)[name = string("op_214_cast_fp16")];114 tensor<fp16, [1, 8, 128, 64]> var_215_cast_fp16 = mul(x = x2_3_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_215_cast_fp16")];115 tensor<fp16, [1, 8, 128, 64]> var_216_cast_fp16 = mul(x = x1_3_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_216_cast_fp16")];116 tensor<fp16, [1, 8, 128, 64]> var_217_cast_fp16 = add(x = var_215_cast_fp16, y = var_216_cast_fp16)[name = string("op_217_cast_fp16")];117 bool var_219_interleave_0 = const()[name = string("op_219_interleave_0"), val = bool(false)];118 tensor<fp16, [1, 8, 128, 128]> var_219_cast_fp16 = concat(axis = var_99, interleave = var_219_interleave_0, values = (var_214_cast_fp16, var_217_cast_fp16))[name = string("op_219_cast_fp16")];119 tensor<int32, [4]> transpose_1_perm_0 = const()[name = string("transpose_1_perm_0"), val = tensor<int32, [4]>([2, 0, 1, 3])];120 tensor<int32, [2]> concat_4 = const()[name = string("concat_4"), val = tensor<int32, [2]>([128, 1024])];121 tensor<fp16, [128, 1, 8, 128]> transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = var_219_cast_fp16)[name = string("transpose_95")];122 tensor<fp16, [128, 1024]> reshape_1_cast_fp16 = reshape(shape = concat_4, x = transpose_1_cast_fp16)[name = string("reshape_1_cast_fp16")];123 bool matmul_0_transpose_x_1 = const()[name = string("matmul_0_transpose_x_1"), val = bool(true)];124 bool matmul_0_transpose_y_1 = const()[name = string("matmul_0_transpose_y_1"), val = bool(false)];125 tensor<fp16, [1024, 1024]> matmul_0_cast_fp16 = matmul(transpose_x = matmul_0_transpose_x_1, transpose_y = matmul_0_transpose_y_1, x = var_68_to_fp16, y = reshape_1_cast_fp16)[name = string("matmul_0_cast_fp16")];126 tensor<int32, [4]> concat_7 = const()[name = string("concat_7"), val = tensor<int32, [4]>([1024, 1, 8, 128])];127 tensor<fp16, [1024, 1, 8, 128]> reshape_2_cast_fp16 = reshape(shape = concat_7, x = matmul_0_cast_fp16)[name = string("reshape_2_cast_fp16")];128 tensor<int32, [4]> scattered_k_1_perm_0 = const()[name = string("scattered_k_1_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];129 tensor<int32, [2]> concat_12 = const()[name = string("concat_12"), val = tensor<int32, [2]>([128, 1024])];130 tensor<fp16, [128, 1, 8, 128]> transpose_56_cast_fp16 = transpose(perm = transpose_56_perm_0, x = var_142_cast_fp16)[name = string("transpose_94")];131 tensor<fp16, [128, 1024]> reshape_4_cast_fp16 = reshape(shape = concat_12, x = transpose_56_cast_fp16)[name = string("reshape_4_cast_fp16")];132 bool matmul_1_transpose_x_1 = const()[name = string("matmul_1_transpose_x_1"), val = bool(true)];133 bool matmul_1_transpose_y_1 = const()[name = string("matmul_1_transpose_y_1"), val = bool(false)];134 tensor<fp16, [1024, 1024]> matmul_1_cast_fp16 = matmul(transpose_x = matmul_1_transpose_x_1, transpose_y = matmul_1_transpose_y_1, x = var_68_to_fp16, y = reshape_4_cast_fp16)[name = string("matmul_1_cast_fp16")];135 tensor<int32, [4]> concat_15 = const()[name = string("concat_15"), val = tensor<int32, [4]>([1024, 1, 8, 128])];136 tensor<fp16, [1024, 1, 8, 128]> reshape_5_cast_fp16 = reshape(shape = concat_15, x = matmul_1_cast_fp16)[name = string("reshape_5_cast_fp16")];137 tensor<int32, [4]> scattered_v_1_perm_0 = const()[name = string("scattered_v_1_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];138 fp16 var_101_promoted_to_fp16 = const()[name = string("op_101_promoted_to_fp16"), val = fp16(0x1p+0)];139 tensor<fp16, [1, 1, 1024, 1]> var_224_cast_fp16 = sub(x = var_101_promoted_to_fp16, y = var_84_cast_fp16)[name = string("op_224_cast_fp16")];140 tensor<fp16, [1, 8, 1024, 128]> read_state_0 = read_state(input = k_cache_0)[name = string("read_state_0")];141 tensor<fp16, [1, 8, 1024, 128]> k_cache_3_cast_fp16 = mul(x = read_state_0, y = var_224_cast_fp16)[name = string("k_cache_3_cast_fp16")];142 write_state(data = k_cache_3_cast_fp16, input = k_cache_0)[name = string("coreml_update_state_56_write_state")];143 tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_56 = read_state(input = k_cache_0)[name = string("coreml_update_state_56")];144 tensor<fp16, [1, 8, 1024, 128]> scattered_k_1_cast_fp16 = transpose(perm = scattered_k_1_perm_0, x = reshape_2_cast_fp16)[name = string("transpose_93")];145 tensor<fp16, [1, 8, 1024, 128]> k_cache_5_cast_fp16 = add(x = coreml_update_state_56, y = scattered_k_1_cast_fp16)[name = string("k_cache_5_cast_fp16")];146 write_state(data = k_cache_5_cast_fp16, input = k_cache_0)[name = string("coreml_update_state_57_write_state")];147 tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_57 = read_state(input = k_cache_0)[name = string("coreml_update_state_57")];148 tensor<fp16, [1, 8, 1024, 128]> read_state_1 = read_state(input = v_cache_0)[name = string("read_state_1")];149 tensor<fp16, [1, 8, 1024, 128]> v_cache_3_cast_fp16 = mul(x = read_state_1, y = var_224_cast_fp16)[name = string("v_cache_3_cast_fp16")];150 write_state(data = v_cache_3_cast_fp16, input = v_cache_0)[name = string("coreml_update_state_58_write_state")];151 tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_58 = read_state(input = v_cache_0)[name = string("coreml_update_state_58")];152 tensor<fp16, [1, 8, 1024, 128]> scattered_v_1_cast_fp16 = transpose(perm = scattered_v_1_perm_0, x = reshape_5_cast_fp16)[name = string("transpose_92")];153 tensor<fp16, [1, 8, 1024, 128]> v_cache_5_cast_fp16 = add(x = coreml_update_state_58, y = scattered_v_1_cast_fp16)[name = string("v_cache_5_cast_fp16")];154 write_state(data = v_cache_5_cast_fp16, input = v_cache_0)[name = string("coreml_update_state_59_write_state")];155 tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_59 = read_state(input = v_cache_0)[name = string("coreml_update_state_59")];156 tensor<int32, [1]> var_230_axes_0 = const()[name = string("op_230_axes_0"), val = tensor<int32, [1]>([2])];157 tensor<fp16, [1, 8, 1, 1024, 128]> var_230_cast_fp16 = expand_dims(axes = var_230_axes_0, x = coreml_update_state_57)[name = string("op_230_cast_fp16")];158 tensor<int32, [5]> k_exp_1_reps_0 = const()[name = string("k_exp_1_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];159 tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_1_cast_fp16 = tile(reps = k_exp_1_reps_0, x = var_230_cast_fp16)[name = string("k_exp_1_cast_fp16")];160 tensor<int32, [4]> var_233 = const()[name = string("op_233"), val = tensor<int32, [4]>([1, 16, 1024, 128])];161 tensor<fp16, [1, 16, 1024, 128]> k_exp_3_cast_fp16 = reshape(shape = var_233, x = k_exp_1_cast_fp16)[name = string("k_exp_3_cast_fp16")];162 tensor<int32, [1]> var_235_axes_0 = const()[name = string("op_235_axes_0"), val = tensor<int32, [1]>([2])];163 tensor<fp16, [1, 8, 1, 1024, 128]> var_235_cast_fp16 = expand_dims(axes = var_235_axes_0, x = coreml_update_state_59)[name = string("op_235_cast_fp16")];164 tensor<int32, [5]> v_exp_1_reps_0 = const()[name = string("v_exp_1_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];165 tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_1_cast_fp16 = tile(reps = v_exp_1_reps_0, x = var_235_cast_fp16)[name = string("v_exp_1_cast_fp16")];166 tensor<int32, [4]> var_238 = const()[name = string("op_238"), val = tensor<int32, [4]>([1, 16, 1024, 128])];167 tensor<fp16, [1, 16, 1024, 128]> v_exp_3_cast_fp16 = reshape(shape = var_238, x = v_exp_1_cast_fp16)[name = string("v_exp_3_cast_fp16")];168 bool var_241_transpose_x_1 = const()[name = string("op_241_transpose_x_1"), val = bool(false)];169 bool var_241_transpose_y_1 = const()[name = string("op_241_transpose_y_1"), val = bool(true)];170 tensor<fp16, [1, 16, 128, 1024]> var_241_cast_fp16 = matmul(transpose_x = var_241_transpose_x_1, transpose_y = var_241_transpose_y_1, x = q_1_cast_fp16, y = k_exp_3_cast_fp16)[name = string("op_241_cast_fp16")];171 fp16 var_242_to_fp16 = const()[name = string("op_242_to_fp16"), val = fp16(0x1.6ap-4)];172 tensor<fp16, [1, 16, 128, 1024]> attn_1_cast_fp16 = mul(x = var_241_cast_fp16, y = var_242_to_fp16)[name = string("attn_1_cast_fp16")];173 string attention_mask_to_fp16_dtype_0 = const()[name = string("attention_mask_to_fp16_dtype_0"), val = string("fp16")];174 tensor<fp16, [1, 1, 128, 1024]> attention_mask_to_fp16 = cast(dtype = attention_mask_to_fp16_dtype_0, x = attention_mask)[name = string("cast_0")];175 tensor<fp16, [1, 16, 128, 1024]> input_1_cast_fp16 = add(x = attn_1_cast_fp16, y = attention_mask_to_fp16)[name = string("input_1_cast_fp16")];176 tensor<fp16, [1, 16, 128, 1024]> attn_3_cast_fp16 = softmax(axis = var_99, x = input_1_cast_fp16)[name = string("attn_3_cast_fp16")];177 bool out_1_transpose_x_0 = const()[name = string("out_1_transpose_x_0"), val = bool(false)];178 bool out_1_transpose_y_0 = const()[name = string("out_1_transpose_y_0"), val = bool(false)];179 tensor<fp16, [1, 16, 128, 128]> out_1_cast_fp16 = matmul(transpose_x = out_1_transpose_x_0, transpose_y = out_1_transpose_y_0, x = attn_3_cast_fp16, y = v_exp_3_cast_fp16)[name = string("out_1_cast_fp16")];180 tensor<int32, [4]> var_247_perm_0 = const()[name = string("op_247_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];181 tensor<int32, [3]> var_248 = const()[name = string("op_248"), val = tensor<int32, [3]>([1, 128, -1])];182 tensor<fp16, [1, 128, 16, 128]> var_247_cast_fp16 = transpose(perm = var_247_perm_0, x = out_1_cast_fp16)[name = string("transpose_91")];183 tensor<fp16, [1, 128, 2048]> input_3_cast_fp16 = reshape(shape = var_248, x = var_247_cast_fp16)[name = string("input_3_cast_fp16")];184 tensor<fp16, [1024, 2048]> layers_0_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4205504))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6302720))))[name = string("layers_0_self_attn_o_proj_weight_to_fp16_palettized")];185 tensor<fp16, [1, 128, 1024]> linear_3_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_0_self_attn_o_proj_weight_to_fp16_palettized, x = input_3_cast_fp16)[name = string("linear_3_cast_fp16")];186 tensor<fp16, [1, 128, 1024]> x_19_cast_fp16 = add(x = input_embeds_to_fp16, y = linear_3_cast_fp16)[name = string("x_19_cast_fp16")];187 fp16 var_98_promoted_3_to_fp16 = const()[name = string("op_98_promoted_3_to_fp16"), val = fp16(0x1p+1)];188 tensor<fp16, [1, 128, 1024]> var_255_cast_fp16 = pow(x = x_19_cast_fp16, y = var_98_promoted_3_to_fp16)[name = string("op_255_cast_fp16")];189 tensor<int32, [1]> var_257_axes_0 = const()[name = string("op_257_axes_0"), val = tensor<int32, [1]>([-1])];190 bool var_257_keep_dims_0 = const()[name = string("op_257_keep_dims_0"), val = bool(true)];191 tensor<fp16, [1, 128, 1]> var_257_cast_fp16 = reduce_mean(axes = var_257_axes_0, keep_dims = var_257_keep_dims_0, x = var_255_cast_fp16)[name = string("op_257_cast_fp16")];192 fp16 var_258_to_fp16 = const()[name = string("op_258_to_fp16"), val = fp16(0x1.1p-20)];193 tensor<fp16, [1, 128, 1]> var_259_cast_fp16 = add(x = var_257_cast_fp16, y = var_258_to_fp16)[name = string("op_259_cast_fp16")];194 fp32 norm_7_epsilon_0 = const()[name = string("norm_7_epsilon_0"), val = fp32(0x1.197998p-40)];195 tensor<fp16, [1, 128, 1]> norm_7_cast_fp16 = rsqrt(epsilon = norm_7_epsilon_0, x = var_259_cast_fp16)[name = string("norm_7_cast_fp16")];196 tensor<fp16, [1, 128, 1024]> var_261_cast_fp16 = mul(x = x_19_cast_fp16, y = norm_7_cast_fp16)[name = string("op_261_cast_fp16")];197 tensor<fp16, [1024]> layers_0_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_0_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6303296)))];198 tensor<fp16, [1, 128, 1024]> var_262_cast_fp16 = mul(x = var_261_cast_fp16, y = layers_0_post_attention_layernorm_weight_to_fp16)[name = string("op_262_cast_fp16")];199 tensor<fp16, [3072, 1024]> layers_0_mlp_gate_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6305408))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9451200))))[name = string("layers_0_mlp_gate_proj_weight_to_fp16_palettized")];200 tensor<fp16, [3072]> linear_4_bias_0_to_fp16 = const()[name = string("linear_4_bias_0_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9451776)))];201 tensor<fp16, [1, 128, 3072]> linear_4_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_0_mlp_gate_proj_weight_to_fp16_palettized, x = var_262_cast_fp16)[name = string("linear_4_cast_fp16")];202 tensor<fp16, [1, 128, 3072]> var_272_cast_fp16 = silu(x = linear_4_cast_fp16)[name = string("op_272_cast_fp16")];203 tensor<fp16, [3072, 1024]> layers_0_mlp_up_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9457984))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12603776))))[name = string("layers_0_mlp_up_proj_weight_to_fp16_palettized")];204 tensor<fp16, [1, 128, 3072]> linear_5_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_0_mlp_up_proj_weight_to_fp16_palettized, x = var_262_cast_fp16)[name = string("linear_5_cast_fp16")];205 tensor<fp16, [1, 128, 3072]> input_9_cast_fp16 = mul(x = var_272_cast_fp16, y = linear_5_cast_fp16)[name = string("input_9_cast_fp16")];206 tensor<fp16, [1024, 3072]> layers_0_mlp_down_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12604352))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15750144))))[name = string("layers_0_mlp_down_proj_weight_to_fp16_palettized")];207 tensor<fp16, [1, 128, 1024]> linear_6_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_0_mlp_down_proj_weight_to_fp16_palettized, x = input_9_cast_fp16)[name = string("linear_6_cast_fp16")];208 tensor<fp16, [1, 128, 1024]> x_25_cast_fp16 = add(x = x_19_cast_fp16, y = linear_6_cast_fp16)[name = string("x_25_cast_fp16")];209 int32 var_293 = const()[name = string("op_293"), val = int32(-1)];210 fp16 var_292_promoted_to_fp16 = const()[name = string("op_292_promoted_to_fp16"), val = fp16(0x1p+1)];211 tensor<fp16, [1, 128, 1024]> var_302_cast_fp16 = pow(x = x_25_cast_fp16, y = var_292_promoted_to_fp16)[name = string("op_302_cast_fp16")];212 tensor<int32, [1]> var_304_axes_0 = const()[name = string("op_304_axes_0"), val = tensor<int32, [1]>([-1])];213 bool var_304_keep_dims_0 = const()[name = string("op_304_keep_dims_0"), val = bool(true)];214 tensor<fp16, [1, 128, 1]> var_304_cast_fp16 = reduce_mean(axes = var_304_axes_0, keep_dims = var_304_keep_dims_0, x = var_302_cast_fp16)[name = string("op_304_cast_fp16")];215 fp16 var_305_to_fp16 = const()[name = string("op_305_to_fp16"), val = fp16(0x1.1p-20)];216 tensor<fp16, [1, 128, 1]> var_306_cast_fp16 = add(x = var_304_cast_fp16, y = var_305_to_fp16)[name = string("op_306_cast_fp16")];217 fp32 norm_9_epsilon_0 = const()[name = string("norm_9_epsilon_0"), val = fp32(0x1.197998p-40)];218 tensor<fp16, [1, 128, 1]> norm_9_cast_fp16 = rsqrt(epsilon = norm_9_epsilon_0, x = var_306_cast_fp16)[name = string("norm_9_cast_fp16")];219 tensor<fp16, [1, 128, 1024]> var_308_cast_fp16 = mul(x = x_25_cast_fp16, y = norm_9_cast_fp16)[name = string("op_308_cast_fp16")];220 tensor<fp16, [1024]> layers_1_input_layernorm_weight_to_fp16 = const()[name = string("layers_1_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15750720)))];221 tensor<fp16, [1, 128, 1024]> var_309_cast_fp16 = mul(x = var_308_cast_fp16, y = layers_1_input_layernorm_weight_to_fp16)[name = string("op_309_cast_fp16")];222 tensor<fp16, [2048, 1024]> layers_1_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15752832))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17850048))))[name = string("layers_1_self_attn_q_proj_weight_to_fp16_palettized")];223 tensor<fp16, [1, 128, 2048]> linear_7_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_1_self_attn_q_proj_weight_to_fp16_palettized, x = var_309_cast_fp16)[name = string("linear_7_cast_fp16")];224 tensor<int32, [4]> var_325 = const()[name = string("op_325"), val = tensor<int32, [4]>([1, 128, 16, 128])];225 tensor<fp16, [1, 128, 16, 128]> var_326_cast_fp16 = reshape(shape = var_325, x = linear_7_cast_fp16)[name = string("op_326_cast_fp16")];226 tensor<int32, [4]> x_31_perm_0 = const()[name = string("x_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];227 tensor<fp16, [1024, 1024]> layers_1_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17850624))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18899264))))[name = string("layers_1_self_attn_k_proj_weight_to_fp16_palettized")];228 tensor<fp16, [1, 128, 1024]> linear_8_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_1_self_attn_k_proj_weight_to_fp16_palettized, x = var_309_cast_fp16)[name = string("linear_8_cast_fp16")];229 tensor<int32, [4]> var_330 = const()[name = string("op_330"), val = tensor<int32, [4]>([1, 128, 8, 128])];230 tensor<fp16, [1, 128, 8, 128]> var_331_cast_fp16 = reshape(shape = var_330, x = linear_8_cast_fp16)[name = string("op_331_cast_fp16")];231 tensor<int32, [4]> x_35_perm_0 = const()[name = string("x_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];232 tensor<fp16, [1024, 1024]> layers_1_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18899840))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19948480))))[name = string("layers_1_self_attn_v_proj_weight_to_fp16_palettized")];233 tensor<fp16, [1, 128, 1024]> linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_1_self_attn_v_proj_weight_to_fp16_palettized, x = var_309_cast_fp16)[name = string("linear_9_cast_fp16")];234 tensor<int32, [4]> var_335 = const()[name = string("op_335"), val = tensor<int32, [4]>([1, 128, 8, 128])];235 tensor<fp16, [1, 128, 8, 128]> var_336_cast_fp16 = reshape(shape = var_335, x = linear_9_cast_fp16)[name = string("op_336_cast_fp16")];236 tensor<int32, [4]> transpose_57_perm_0 = const()[name = string("transpose_57_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];237 fp16 var_292_promoted_1_to_fp16 = const()[name = string("op_292_promoted_1_to_fp16"), val = fp16(0x1p+1)];238 tensor<fp16, [1, 16, 128, 128]> x_31_cast_fp16 = transpose(perm = x_31_perm_0, x = var_326_cast_fp16)[name = string("transpose_90")];239 tensor<fp16, [1, 16, 128, 128]> var_340_cast_fp16 = pow(x = x_31_cast_fp16, y = var_292_promoted_1_to_fp16)[name = string("op_340_cast_fp16")];240 tensor<int32, [1]> var_342_axes_0 = const()[name = string("op_342_axes_0"), val = tensor<int32, [1]>([-1])];241 bool var_342_keep_dims_0 = const()[name = string("op_342_keep_dims_0"), val = bool(true)];242 tensor<fp16, [1, 16, 128, 1]> var_342_cast_fp16 = reduce_mean(axes = var_342_axes_0, keep_dims = var_342_keep_dims_0, x = var_340_cast_fp16)[name = string("op_342_cast_fp16")];243 fp16 var_343_to_fp16 = const()[name = string("op_343_to_fp16"), val = fp16(0x1.1p-20)];244 tensor<fp16, [1, 16, 128, 1]> var_344_cast_fp16 = add(x = var_342_cast_fp16, y = var_343_to_fp16)[name = string("op_344_cast_fp16")];245 fp32 norm_11_epsilon_0 = const()[name = string("norm_11_epsilon_0"), val = fp32(0x1.197998p-40)];246 tensor<fp16, [1, 16, 128, 1]> norm_11_cast_fp16 = rsqrt(epsilon = norm_11_epsilon_0, x = var_344_cast_fp16)[name = string("norm_11_cast_fp16")];247 tensor<fp16, [1, 16, 128, 128]> var_346_cast_fp16 = mul(x = x_31_cast_fp16, y = norm_11_cast_fp16)[name = string("op_346_cast_fp16")];248 tensor<fp16, [128]> layers_1_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_1_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19949056)))];249 tensor<fp16, [1, 16, 128, 128]> var_347_cast_fp16 = mul(x = var_346_cast_fp16, y = layers_1_self_attn_q_norm_weight_to_fp16)[name = string("op_347_cast_fp16")];250 fp16 var_292_promoted_2_to_fp16 = const()[name = string("op_292_promoted_2_to_fp16"), val = fp16(0x1p+1)];251 tensor<fp16, [1, 8, 128, 128]> x_35_cast_fp16 = transpose(perm = x_35_perm_0, x = var_331_cast_fp16)[name = string("transpose_89")];252 tensor<fp16, [1, 8, 128, 128]> var_351_cast_fp16 = pow(x = x_35_cast_fp16, y = var_292_promoted_2_to_fp16)[name = string("op_351_cast_fp16")];253 tensor<int32, [1]> var_353_axes_0 = const()[name = string("op_353_axes_0"), val = tensor<int32, [1]>([-1])];254 bool var_353_keep_dims_0 = const()[name = string("op_353_keep_dims_0"), val = bool(true)];255 tensor<fp16, [1, 8, 128, 1]> var_353_cast_fp16 = reduce_mean(axes = var_353_axes_0, keep_dims = var_353_keep_dims_0, x = var_351_cast_fp16)[name = string("op_353_cast_fp16")];256 fp16 var_354_to_fp16 = const()[name = string("op_354_to_fp16"), val = fp16(0x1.1p-20)];257 tensor<fp16, [1, 8, 128, 1]> var_355_cast_fp16 = add(x = var_353_cast_fp16, y = var_354_to_fp16)[name = string("op_355_cast_fp16")];258 fp32 norm_13_epsilon_0 = const()[name = string("norm_13_epsilon_0"), val = fp32(0x1.197998p-40)];259 tensor<fp16, [1, 8, 128, 1]> norm_13_cast_fp16 = rsqrt(epsilon = norm_13_epsilon_0, x = var_355_cast_fp16)[name = string("norm_13_cast_fp16")];260 tensor<fp16, [1, 8, 128, 128]> var_357_cast_fp16 = mul(x = x_35_cast_fp16, y = norm_13_cast_fp16)[name = string("op_357_cast_fp16")];261 tensor<fp16, [128]> layers_1_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_1_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19949376)))];262 tensor<fp16, [1, 8, 128, 128]> var_358_cast_fp16 = mul(x = var_357_cast_fp16, y = layers_1_self_attn_k_norm_weight_to_fp16)[name = string("op_358_cast_fp16")];263 tensor<int32, [4]> x1_5_begin_0 = const()[name = string("x1_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];264 tensor<int32, [4]> x1_5_end_0 = const()[name = string("x1_5_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];265 tensor<bool, [4]> x1_5_end_mask_0 = const()[name = string("x1_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];266 tensor<fp16, [1, 16, 128, 64]> x1_5_cast_fp16 = slice_by_index(begin = x1_5_begin_0, end = x1_5_end_0, end_mask = x1_5_end_mask_0, x = var_347_cast_fp16)[name = string("x1_5_cast_fp16")];267 tensor<int32, [4]> x2_5_begin_0 = const()[name = string("x2_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];268 tensor<int32, [4]> x2_5_end_0 = const()[name = string("x2_5_end_0"), val = tensor<int32, [4]>([1, 16, 128, 128])];269 tensor<bool, [4]> x2_5_end_mask_0 = const()[name = string("x2_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];270 tensor<fp16, [1, 16, 128, 64]> x2_5_cast_fp16 = slice_by_index(begin = x2_5_begin_0, end = x2_5_end_0, end_mask = x2_5_end_mask_0, x = var_347_cast_fp16)[name = string("x2_5_cast_fp16")];271 tensor<fp16, [1, 16, 128, 64]> var_379_cast_fp16 = mul(x = x1_5_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_379_cast_fp16")];272 tensor<fp16, [1, 16, 128, 64]> var_380_cast_fp16 = mul(x = x2_5_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_380_cast_fp16")];273 tensor<fp16, [1, 16, 128, 64]> var_381_cast_fp16 = sub(x = var_379_cast_fp16, y = var_380_cast_fp16)[name = string("op_381_cast_fp16")];274 tensor<fp16, [1, 16, 128, 64]> var_382_cast_fp16 = mul(x = x2_5_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_382_cast_fp16")];275 tensor<fp16, [1, 16, 128, 64]> var_383_cast_fp16 = mul(x = x1_5_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_383_cast_fp16")];276 tensor<fp16, [1, 16, 128, 64]> var_384_cast_fp16 = add(x = var_382_cast_fp16, y = var_383_cast_fp16)[name = string("op_384_cast_fp16")];277 bool q_3_interleave_0 = const()[name = string("q_3_interleave_0"), val = bool(false)];278 tensor<fp16, [1, 16, 128, 128]> q_3_cast_fp16 = concat(axis = var_293, interleave = q_3_interleave_0, values = (var_381_cast_fp16, var_384_cast_fp16))[name = string("q_3_cast_fp16")];279 tensor<int32, [4]> x1_7_begin_0 = const()[name = string("x1_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];280 tensor<int32, [4]> x1_7_end_0 = const()[name = string("x1_7_end_0"), val = tensor<int32, [4]>([1, 8, 128, 64])];281 tensor<bool, [4]> x1_7_end_mask_0 = const()[name = string("x1_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];282 tensor<fp16, [1, 8, 128, 64]> x1_7_cast_fp16 = slice_by_index(begin = x1_7_begin_0, end = x1_7_end_0, end_mask = x1_7_end_mask_0, x = var_358_cast_fp16)[name = string("x1_7_cast_fp16")];283 tensor<int32, [4]> x2_7_begin_0 = const()[name = string("x2_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];284 tensor<int32, [4]> x2_7_end_0 = const()[name = string("x2_7_end_0"), val = tensor<int32, [4]>([1, 8, 128, 128])];285 tensor<bool, [4]> x2_7_end_mask_0 = const()[name = string("x2_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];286 tensor<fp16, [1, 8, 128, 64]> x2_7_cast_fp16 = slice_by_index(begin = x2_7_begin_0, end = x2_7_end_0, end_mask = x2_7_end_mask_0, x = var_358_cast_fp16)[name = string("x2_7_cast_fp16")];287 tensor<fp16, [1, 8, 128, 64]> var_406_cast_fp16 = mul(x = x1_7_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_406_cast_fp16")];288 tensor<fp16, [1, 8, 128, 64]> var_407_cast_fp16 = mul(x = x2_7_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_407_cast_fp16")];289 tensor<fp16, [1, 8, 128, 64]> var_408_cast_fp16 = sub(x = var_406_cast_fp16, y = var_407_cast_fp16)[name = string("op_408_cast_fp16")];290 tensor<fp16, [1, 8, 128, 64]> var_409_cast_fp16 = mul(x = x2_7_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_409_cast_fp16")];291 tensor<fp16, [1, 8, 128, 64]> var_410_cast_fp16 = mul(x = x1_7_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_410_cast_fp16")];292 tensor<fp16, [1, 8, 128, 64]> var_411_cast_fp16 = add(x = var_409_cast_fp16, y = var_410_cast_fp16)[name = string("op_411_cast_fp16")];293 bool var_413_interleave_0 = const()[name = string("op_413_interleave_0"), val = bool(false)];294 tensor<fp16, [1, 8, 128, 128]> var_413_cast_fp16 = concat(axis = var_293, interleave = var_413_interleave_0, values = (var_408_cast_fp16, var_411_cast_fp16))[name = string("op_413_cast_fp16")];295 tensor<int32, [4]> transpose_5_perm_0 = const()[name = string("transpose_5_perm_0"), val = tensor<int32, [4]>([2, 0, 1, 3])];296 tensor<int32, [2]> concat_22 = const()[name = string("concat_22"), val = tensor<int32, [2]>([128, 1024])];297 tensor<fp16, [128, 1, 8, 128]> transpose_5_cast_fp16 = transpose(perm = transpose_5_perm_0, x = var_413_cast_fp16)[name = string("transpose_88")];298 tensor<fp16, [128, 1024]> reshape_7_cast_fp16 = reshape(shape = concat_22, x = transpose_5_cast_fp16)[name = string("reshape_7_cast_fp16")];299 bool matmul_2_transpose_x_1 = const()[name = string("matmul_2_transpose_x_1"), val = bool(true)];300 bool matmul_2_transpose_y_1 = const()[name = string("matmul_2_transpose_y_1"), val = bool(false)];301 tensor<fp16, [1024, 1024]> matmul_2_cast_fp16 = matmul(transpose_x = matmul_2_transpose_x_1, transpose_y = matmul_2_transpose_y_1, x = var_68_to_fp16, y = reshape_7_cast_fp16)[name = string("matmul_2_cast_fp16")];302 tensor<int32, [4]> concat_25 = const()[name = string("concat_25"), val = tensor<int32, [4]>([1024, 1, 8, 128])];303 tensor<fp16, [1024, 1, 8, 128]> reshape_8_cast_fp16 = reshape(shape = concat_25, x = matmul_2_cast_fp16)[name = string("reshape_8_cast_fp16")];304 tensor<int32, [4]> scattered_k_3_perm_0 = const()[name = string("scattered_k_3_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];305 tensor<int32, [2]> concat_30 = const()[name = string("concat_30"), val = tensor<int32, [2]>([128, 1024])];306 tensor<fp16, [128, 1, 8, 128]> transpose_57_cast_fp16 = transpose(perm = transpose_57_perm_0, x = var_336_cast_fp16)[name = string("transpose_87")];307 tensor<fp16, [128, 1024]> reshape_10_cast_fp16 = reshape(shape = concat_30, x = transpose_57_cast_fp16)[name = string("reshape_10_cast_fp16")];308 bool matmul_3_transpose_x_1 = const()[name = string("matmul_3_transpose_x_1"), val = bool(true)];309 bool matmul_3_transpose_y_1 = const()[name = string("matmul_3_transpose_y_1"), val = bool(false)];310 tensor<fp16, [1024, 1024]> matmul_3_cast_fp16 = matmul(transpose_x = matmul_3_transpose_x_1, transpose_y = matmul_3_transpose_y_1, x = var_68_to_fp16, y = reshape_10_cast_fp16)[name = string("matmul_3_cast_fp16")];311 tensor<int32, [4]> concat_33 = const()[name = string("concat_33"), val = tensor<int32, [4]>([1024, 1, 8, 128])];312 tensor<fp16, [1024, 1, 8, 128]> reshape_11_cast_fp16 = reshape(shape = concat_33, x = matmul_3_cast_fp16)[name = string("reshape_11_cast_fp16")];313 tensor<int32, [4]> scattered_v_3_perm_0 = const()[name = string("scattered_v_3_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];314 tensor<fp16, [1, 8, 1024, 128]> read_state_2 = read_state(input = k_cache_1)[name = string("read_state_2")];315 tensor<fp16, [1, 8, 1024, 128]> k_cache_9_cast_fp16 = mul(x = read_state_2, y = var_224_cast_fp16)[name = string("k_cache_9_cast_fp16")];316 write_state(data = k_cache_9_cast_fp16, input = k_cache_1)[name = string("coreml_update_state_60_write_state")];317 tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_60 = read_state(input = k_cache_1)[name = string("coreml_update_state_60")];318 tensor<fp16, [1, 8, 1024, 128]> scattered_k_3_cast_fp16 = transpose(perm = scattered_k_3_perm_0, x = reshape_8_cast_fp16)[name = string("transpose_86")];319 tensor<fp16, [1, 8, 1024, 128]> k_cache_11_cast_fp16 = add(x = coreml_update_state_60, y = scattered_k_3_cast_fp16)[name = string("k_cache_11_cast_fp16")];320 write_state(data = k_cache_11_cast_fp16, input = k_cache_1)[name = string("coreml_update_state_61_write_state")];321 tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_61 = read_state(input = k_cache_1)[name = string("coreml_update_state_61")];322 tensor<fp16, [1, 8, 1024, 128]> read_state_3 = read_state(input = v_cache_1)[name = string("read_state_3")];323 tensor<fp16, [1, 8, 1024, 128]> v_cache_9_cast_fp16 = mul(x = read_state_3, y = var_224_cast_fp16)[name = string("v_cache_9_cast_fp16")];324 write_state(data = v_cache_9_cast_fp16, input = v_cache_1)[name = string("coreml_update_state_62_write_state")];325 tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_62 = read_state(input = v_cache_1)[name = string("coreml_update_state_62")];326 tensor<fp16, [1, 8, 1024, 128]> scattered_v_3_cast_fp16 = transpose(perm = scattered_v_3_perm_0, x = reshape_11_cast_fp16)[name = string("transpose_85")];327 tensor<fp16, [1, 8, 1024, 128]> v_cache_11_cast_fp16 = add(x = coreml_update_state_62, y = scattered_v_3_cast_fp16)[name = string("v_cache_11_cast_fp16")];328 write_state(data = v_cache_11_cast_fp16, input = v_cache_1)[name = string("coreml_update_state_63_write_state")];329 tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_63 = read_state(input = v_cache_1)[name = string("coreml_update_state_63")];330 tensor<int32, [1]> var_424_axes_0 = const()[name = string("op_424_axes_0"), val = tensor<int32, [1]>([2])];331 tensor<fp16, [1, 8, 1, 1024, 128]> var_424_cast_fp16 = expand_dims(axes = var_424_axes_0, x = coreml_update_state_61)[name = string("op_424_cast_fp16")];332 tensor<int32, [5]> k_exp_5_reps_0 = const()[name = string("k_exp_5_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];333 tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_5_cast_fp16 = tile(reps = k_exp_5_reps_0, x = var_424_cast_fp16)[name = string("k_exp_5_cast_fp16")];334 tensor<int32, [4]> var_427 = const()[name = string("op_427"), val = tensor<int32, [4]>([1, 16, 1024, 128])];335 tensor<fp16, [1, 16, 1024, 128]> k_exp_7_cast_fp16 = reshape(shape = var_427, x = k_exp_5_cast_fp16)[name = string("k_exp_7_cast_fp16")];336 tensor<int32, [1]> var_429_axes_0 = const()[name = string("op_429_axes_0"), val = tensor<int32, [1]>([2])];337 tensor<fp16, [1, 8, 1, 1024, 128]> var_429_cast_fp16 = expand_dims(axes = var_429_axes_0, x = coreml_update_state_63)[name = string("op_429_cast_fp16")];338 tensor<int32, [5]> v_exp_5_reps_0 = const()[name = string("v_exp_5_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];339 tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_5_cast_fp16 = tile(reps = v_exp_5_reps_0, x = var_429_cast_fp16)[name = string("v_exp_5_cast_fp16")];340 tensor<int32, [4]> var_432 = const()[name = string("op_432"), val = tensor<int32, [4]>([1, 16, 1024, 128])];341 tensor<fp16, [1, 16, 1024, 128]> v_exp_7_cast_fp16 = reshape(shape = var_432, x = v_exp_5_cast_fp16)[name = string("v_exp_7_cast_fp16")];342 bool var_435_transpose_x_1 = const()[name = string("op_435_transpose_x_1"), val = bool(false)];343 bool var_435_transpose_y_1 = const()[name = string("op_435_transpose_y_1"), val = bool(true)];344 tensor<fp16, [1, 16, 128, 1024]> var_435_cast_fp16 = matmul(transpose_x = var_435_transpose_x_1, transpose_y = var_435_transpose_y_1, x = q_3_cast_fp16, y = k_exp_7_cast_fp16)[name = string("op_435_cast_fp16")];345 fp16 var_436_to_fp16 = const()[name = string("op_436_to_fp16"), val = fp16(0x1.6ap-4)];346 tensor<fp16, [1, 16, 128, 1024]> attn_5_cast_fp16 = mul(x = var_435_cast_fp16, y = var_436_to_fp16)[name = string("attn_5_cast_fp16")];347 tensor<fp16, [1, 16, 128, 1024]> input_11_cast_fp16 = add(x = attn_5_cast_fp16, y = attention_mask_to_fp16)[name = string("input_11_cast_fp16")];348 tensor<fp16, [1, 16, 128, 1024]> attn_7_cast_fp16 = softmax(axis = var_293, x = input_11_cast_fp16)[name = string("attn_7_cast_fp16")];349 bool out_3_transpose_x_0 = const()[name = string("out_3_transpose_x_0"), val = bool(false)];350 bool out_3_transpose_y_0 = const()[name = string("out_3_transpose_y_0"), val = bool(false)];351 tensor<fp16, [1, 16, 128, 128]> out_3_cast_fp16 = matmul(transpose_x = out_3_transpose_x_0, transpose_y = out_3_transpose_y_0, x = attn_7_cast_fp16, y = v_exp_7_cast_fp16)[name = string("out_3_cast_fp16")];352 tensor<int32, [4]> var_441_perm_0 = const()[name = string("op_441_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];353 tensor<int32, [3]> var_442 = const()[name = string("op_442"), val = tensor<int32, [3]>([1, 128, -1])];354 tensor<fp16, [1, 128, 16, 128]> var_441_cast_fp16 = transpose(perm = var_441_perm_0, x = out_3_cast_fp16)[name = string("transpose_84")];355 tensor<fp16, [1, 128, 2048]> input_13_cast_fp16 = reshape(shape = var_442, x = var_441_cast_fp16)[name = string("input_13_cast_fp16")];356 tensor<fp16, [1024, 2048]> layers_1_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19949696))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22046912))))[name = string("layers_1_self_attn_o_proj_weight_to_fp16_palettized")];357 tensor<fp16, [1, 128, 1024]> linear_10_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_1_self_attn_o_proj_weight_to_fp16_palettized, x = input_13_cast_fp16)[name = string("linear_10_cast_fp16")];358 tensor<fp16, [1, 128, 1024]> x_45_cast_fp16 = add(x = x_25_cast_fp16, y = linear_10_cast_fp16)[name = string("x_45_cast_fp16")];359 fp16 var_292_promoted_3_to_fp16 = const()[name = string("op_292_promoted_3_to_fp16"), val = fp16(0x1p+1)];360 tensor<fp16, [1, 128, 1024]> var_449_cast_fp16 = pow(x = x_45_cast_fp16, y = var_292_promoted_3_to_fp16)[name = string("op_449_cast_fp16")];361 tensor<int32, [1]> var_451_axes_0 = const()[name = string("op_451_axes_0"), val = tensor<int32, [1]>([-1])];362 bool var_451_keep_dims_0 = const()[name = string("op_451_keep_dims_0"), val = bool(true)];363 tensor<fp16, [1, 128, 1]> var_451_cast_fp16 = reduce_mean(axes = var_451_axes_0, keep_dims = var_451_keep_dims_0, x = var_449_cast_fp16)[name = string("op_451_cast_fp16")];364 fp16 var_452_to_fp16 = const()[name = string("op_452_to_fp16"), val = fp16(0x1.1p-20)];365 tensor<fp16, [1, 128, 1]> var_453_cast_fp16 = add(x = var_451_cast_fp16, y = var_452_to_fp16)[name = string("op_453_cast_fp16")];366 fp32 norm_15_epsilon_0 = const()[name = string("norm_15_epsilon_0"), val = fp32(0x1.197998p-40)];367 tensor<fp16, [1, 128, 1]> norm_15_cast_fp16 = rsqrt(epsilon = norm_15_epsilon_0, x = var_453_cast_fp16)[name = string("norm_15_cast_fp16")];368 tensor<fp16, [1, 128, 1024]> var_455_cast_fp16 = mul(x = x_45_cast_fp16, y = norm_15_cast_fp16)[name = string("op_455_cast_fp16")];369 tensor<fp16, [1024]> layers_1_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_1_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22047488)))];370 tensor<fp16, [1, 128, 1024]> var_456_cast_fp16 = mul(x = var_455_cast_fp16, y = layers_1_post_attention_layernorm_weight_to_fp16)[name = string("op_456_cast_fp16")];371 tensor<fp16, [3072, 1024]> layers_1_mlp_gate_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22049600))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25195392))))[name = string("layers_1_mlp_gate_proj_weight_to_fp16_palettized")];372 tensor<fp16, [1, 128, 3072]> linear_11_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_1_mlp_gate_proj_weight_to_fp16_palettized, x = var_456_cast_fp16)[name = string("linear_11_cast_fp16")];373 tensor<fp16, [1, 128, 3072]> var_466_cast_fp16 = silu(x = linear_11_cast_fp16)[name = string("op_466_cast_fp16")];374 tensor<fp16, [3072, 1024]> layers_1_mlp_up_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25195968))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28341760))))[name = string("layers_1_mlp_up_proj_weight_to_fp16_palettized")];375 tensor<fp16, [1, 128, 3072]> linear_12_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_1_mlp_up_proj_weight_to_fp16_palettized, x = var_456_cast_fp16)[name = string("linear_12_cast_fp16")];376 tensor<fp16, [1, 128, 3072]> input_19_cast_fp16 = mul(x = var_466_cast_fp16, y = linear_12_cast_fp16)[name = string("input_19_cast_fp16")];377 tensor<fp16, [1024, 3072]> layers_1_mlp_down_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28342336))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31488128))))[name = string("layers_1_mlp_down_proj_weight_to_fp16_palettized")];378 tensor<fp16, [1, 128, 1024]> linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_1_mlp_down_proj_weight_to_fp16_palettized, x = input_19_cast_fp16)[name = string("linear_13_cast_fp16")];379 tensor<fp16, [1, 128, 1024]> x_51_cast_fp16 = add(x = x_45_cast_fp16, y = linear_13_cast_fp16)[name = string("x_51_cast_fp16")];380 int32 var_487 = const()[name = string("op_487"), val = int32(-1)];381 fp16 var_486_promoted_to_fp16 = const()[name = string("op_486_promoted_to_fp16"), val = fp16(0x1p+1)];382 tensor<fp16, [1, 128, 1024]> var_496_cast_fp16 = pow(x = x_51_cast_fp16, y = var_486_promoted_to_fp16)[name = string("op_496_cast_fp16")];383 tensor<int32, [1]> var_498_axes_0 = const()[name = string("op_498_axes_0"), val = tensor<int32, [1]>([-1])];384 bool var_498_keep_dims_0 = const()[name = string("op_498_keep_dims_0"), val = bool(true)];385 tensor<fp16, [1, 128, 1]> var_498_cast_fp16 = reduce_mean(axes = var_498_axes_0, keep_dims = var_498_keep_dims_0, x = var_496_cast_fp16)[name = string("op_498_cast_fp16")];386 fp16 var_499_to_fp16 = const()[name = string("op_499_to_fp16"), val = fp16(0x1.1p-20)];387 tensor<fp16, [1, 128, 1]> var_500_cast_fp16 = add(x = var_498_cast_fp16, y = var_499_to_fp16)[name = string("op_500_cast_fp16")];388 fp32 norm_17_epsilon_0 = const()[name = string("norm_17_epsilon_0"), val = fp32(0x1.197998p-40)];389 tensor<fp16, [1, 128, 1]> norm_17_cast_fp16 = rsqrt(epsilon = norm_17_epsilon_0, x = var_500_cast_fp16)[name = string("norm_17_cast_fp16")];390 tensor<fp16, [1, 128, 1024]> var_502_cast_fp16 = mul(x = x_51_cast_fp16, y = norm_17_cast_fp16)[name = string("op_502_cast_fp16")];391 tensor<fp16, [1024]> layers_2_input_layernorm_weight_to_fp16 = const()[name = string("layers_2_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31488704)))];392 tensor<fp16, [1, 128, 1024]> var_503_cast_fp16 = mul(x = var_502_cast_fp16, y = layers_2_input_layernorm_weight_to_fp16)[name = string("op_503_cast_fp16")];393 tensor<fp16, [2048, 1024]> layers_2_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31490816))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33588032))))[name = string("layers_2_self_attn_q_proj_weight_to_fp16_palettized")];394 tensor<fp16, [1, 128, 2048]> linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_2_self_attn_q_proj_weight_to_fp16_palettized, x = var_503_cast_fp16)[name = string("linear_14_cast_fp16")];395 tensor<int32, [4]> var_519 = const()[name = string("op_519"), val = tensor<int32, [4]>([1, 128, 16, 128])];396 tensor<fp16, [1, 128, 16, 128]> var_520_cast_fp16 = reshape(shape = var_519, x = linear_14_cast_fp16)[name = string("op_520_cast_fp16")];397 tensor<int32, [4]> x_57_perm_0 = const()[name = string("x_57_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];398 tensor<fp16, [1024, 1024]> layers_2_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33588608))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34637248))))[name = string("layers_2_self_attn_k_proj_weight_to_fp16_palettized")];399 tensor<fp16, [1, 128, 1024]> linear_15_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_2_self_attn_k_proj_weight_to_fp16_palettized, x = var_503_cast_fp16)[name = string("linear_15_cast_fp16")];400 tensor<int32, [4]> var_524 = const()[name = string("op_524"), val = tensor<int32, [4]>([1, 128, 8, 128])];401 tensor<fp16, [1, 128, 8, 128]> var_525_cast_fp16 = reshape(shape = var_524, x = linear_15_cast_fp16)[name = string("op_525_cast_fp16")];402 tensor<int32, [4]> x_61_perm_0 = const()[name = string("x_61_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];403 tensor<fp16, [1024, 1024]> layers_2_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34637824))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35686464))))[name = string("layers_2_self_attn_v_proj_weight_to_fp16_palettized")];404 tensor<fp16, [1, 128, 1024]> linear_16_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_2_self_attn_v_proj_weight_to_fp16_palettized, x = var_503_cast_fp16)[name = string("linear_16_cast_fp16")];405 tensor<int32, [4]> var_529 = const()[name = string("op_529"), val = tensor<int32, [4]>([1, 128, 8, 128])];406 tensor<fp16, [1, 128, 8, 128]> var_530_cast_fp16 = reshape(shape = var_529, x = linear_16_cast_fp16)[name = string("op_530_cast_fp16")];407 tensor<int32, [4]> transpose_58_perm_0 = const()[name = string("transpose_58_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];408 fp16 var_486_promoted_1_to_fp16 = const()[name = string("op_486_promoted_1_to_fp16"), val = fp16(0x1p+1)];409 tensor<fp16, [1, 16, 128, 128]> x_57_cast_fp16 = transpose(perm = x_57_perm_0, x = var_520_cast_fp16)[name = string("transpose_83")];410 tensor<fp16, [1, 16, 128, 128]> var_534_cast_fp16 = pow(x = x_57_cast_fp16, y = var_486_promoted_1_to_fp16)[name = string("op_534_cast_fp16")];411 tensor<int32, [1]> var_536_axes_0 = const()[name = string("op_536_axes_0"), val = tensor<int32, [1]>([-1])];412 bool var_536_keep_dims_0 = const()[name = string("op_536_keep_dims_0"), val = bool(true)];413 tensor<fp16, [1, 16, 128, 1]> var_536_cast_fp16 = reduce_mean(axes = var_536_axes_0, keep_dims = var_536_keep_dims_0, x = var_534_cast_fp16)[name = string("op_536_cast_fp16")];414 fp16 var_537_to_fp16 = const()[name = string("op_537_to_fp16"), val = fp16(0x1.1p-20)];415 tensor<fp16, [1, 16, 128, 1]> var_538_cast_fp16 = add(x = var_536_cast_fp16, y = var_537_to_fp16)[name = string("op_538_cast_fp16")];416 fp32 norm_19_epsilon_0 = const()[name = string("norm_19_epsilon_0"), val = fp32(0x1.197998p-40)];417 tensor<fp16, [1, 16, 128, 1]> norm_19_cast_fp16 = rsqrt(epsilon = norm_19_epsilon_0, x = var_538_cast_fp16)[name = string("norm_19_cast_fp16")];418 tensor<fp16, [1, 16, 128, 128]> var_540_cast_fp16 = mul(x = x_57_cast_fp16, y = norm_19_cast_fp16)[name = string("op_540_cast_fp16")];419 tensor<fp16, [128]> layers_2_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_2_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35687040)))];420 tensor<fp16, [1, 16, 128, 128]> var_541_cast_fp16 = mul(x = var_540_cast_fp16, y = layers_2_self_attn_q_norm_weight_to_fp16)[name = string("op_541_cast_fp16")];421 fp16 var_486_promoted_2_to_fp16 = const()[name = string("op_486_promoted_2_to_fp16"), val = fp16(0x1p+1)];422 tensor<fp16, [1, 8, 128, 128]> x_61_cast_fp16 = transpose(perm = x_61_perm_0, x = var_525_cast_fp16)[name = string("transpose_82")];423 tensor<fp16, [1, 8, 128, 128]> var_545_cast_fp16 = pow(x = x_61_cast_fp16, y = var_486_promoted_2_to_fp16)[name = string("op_545_cast_fp16")];424 tensor<int32, [1]> var_547_axes_0 = const()[name = string("op_547_axes_0"), val = tensor<int32, [1]>([-1])];425 bool var_547_keep_dims_0 = const()[name = string("op_547_keep_dims_0"), val = bool(true)];426 tensor<fp16, [1, 8, 128, 1]> var_547_cast_fp16 = reduce_mean(axes = var_547_axes_0, keep_dims = var_547_keep_dims_0, x = var_545_cast_fp16)[name = string("op_547_cast_fp16")];427 fp16 var_548_to_fp16 = const()[name = string("op_548_to_fp16"), val = fp16(0x1.1p-20)];428 tensor<fp16, [1, 8, 128, 1]> var_549_cast_fp16 = add(x = var_547_cast_fp16, y = var_548_to_fp16)[name = string("op_549_cast_fp16")];429 fp32 norm_21_epsilon_0 = const()[name = string("norm_21_epsilon_0"), val = fp32(0x1.197998p-40)];430 tensor<fp16, [1, 8, 128, 1]> norm_21_cast_fp16 = rsqrt(epsilon = norm_21_epsilon_0, x = var_549_cast_fp16)[name = string("norm_21_cast_fp16")];431 tensor<fp16, [1, 8, 128, 128]> var_551_cast_fp16 = mul(x = x_61_cast_fp16, y = norm_21_cast_fp16)[name = string("op_551_cast_fp16")];432 tensor<fp16, [128]> layers_2_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_2_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35687360)))];433 tensor<fp16, [1, 8, 128, 128]> var_552_cast_fp16 = mul(x = var_551_cast_fp16, y = layers_2_self_attn_k_norm_weight_to_fp16)[name = string("op_552_cast_fp16")];434 tensor<int32, [4]> x1_9_begin_0 = const()[name = string("x1_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];435 tensor<int32, [4]> x1_9_end_0 = const()[name = string("x1_9_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];436 tensor<bool, [4]> x1_9_end_mask_0 = const()[name = string("x1_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];437 tensor<fp16, [1, 16, 128, 64]> x1_9_cast_fp16 = slice_by_index(begin = x1_9_begin_0, end = x1_9_end_0, end_mask = x1_9_end_mask_0, x = var_541_cast_fp16)[name = string("x1_9_cast_fp16")];438 tensor<int32, [4]> x2_9_begin_0 = const()[name = string("x2_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];439 tensor<int32, [4]> x2_9_end_0 = const()[name = string("x2_9_end_0"), val = tensor<int32, [4]>([1, 16, 128, 128])];440 tensor<bool, [4]> x2_9_end_mask_0 = const()[name = string("x2_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];441 tensor<fp16, [1, 16, 128, 64]> x2_9_cast_fp16 = slice_by_index(begin = x2_9_begin_0, end = x2_9_end_0, end_mask = x2_9_end_mask_0, x = var_541_cast_fp16)[name = string("x2_9_cast_fp16")];442 tensor<fp16, [1, 16, 128, 64]> var_573_cast_fp16 = mul(x = x1_9_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_573_cast_fp16")];443 tensor<fp16, [1, 16, 128, 64]> var_574_cast_fp16 = mul(x = x2_9_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_574_cast_fp16")];444 tensor<fp16, [1, 16, 128, 64]> var_575_cast_fp16 = sub(x = var_573_cast_fp16, y = var_574_cast_fp16)[name = string("op_575_cast_fp16")];445 tensor<fp16, [1, 16, 128, 64]> var_576_cast_fp16 = mul(x = x2_9_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_576_cast_fp16")];446 tensor<fp16, [1, 16, 128, 64]> var_577_cast_fp16 = mul(x = x1_9_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_577_cast_fp16")];447 tensor<fp16, [1, 16, 128, 64]> var_578_cast_fp16 = add(x = var_576_cast_fp16, y = var_577_cast_fp16)[name = string("op_578_cast_fp16")];448 bool q_5_interleave_0 = const()[name = string("q_5_interleave_0"), val = bool(false)];449 tensor<fp16, [1, 16, 128, 128]> q_5_cast_fp16 = concat(axis = var_487, interleave = q_5_interleave_0, values = (var_575_cast_fp16, var_578_cast_fp16))[name = string("q_5_cast_fp16")];450 tensor<int32, [4]> x1_11_begin_0 = const()[name = string("x1_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];451 tensor<int32, [4]> x1_11_end_0 = const()[name = string("x1_11_end_0"), val = tensor<int32, [4]>([1, 8, 128, 64])];452 tensor<bool, [4]> x1_11_end_mask_0 = const()[name = string("x1_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];453 tensor<fp16, [1, 8, 128, 64]> x1_11_cast_fp16 = slice_by_index(begin = x1_11_begin_0, end = x1_11_end_0, end_mask = x1_11_end_mask_0, x = var_552_cast_fp16)[name = string("x1_11_cast_fp16")];454 tensor<int32, [4]> x2_11_begin_0 = const()[name = string("x2_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];455 tensor<int32, [4]> x2_11_end_0 = const()[name = string("x2_11_end_0"), val = tensor<int32, [4]>([1, 8, 128, 128])];456 tensor<bool, [4]> x2_11_end_mask_0 = const()[name = string("x2_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];457 tensor<fp16, [1, 8, 128, 64]> x2_11_cast_fp16 = slice_by_index(begin = x2_11_begin_0, end = x2_11_end_0, end_mask = x2_11_end_mask_0, x = var_552_cast_fp16)[name = string("x2_11_cast_fp16")];458 tensor<fp16, [1, 8, 128, 64]> var_600_cast_fp16 = mul(x = x1_11_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_600_cast_fp16")];459 tensor<fp16, [1, 8, 128, 64]> var_601_cast_fp16 = mul(x = x2_11_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_601_cast_fp16")];460 tensor<fp16, [1, 8, 128, 64]> var_602_cast_fp16 = sub(x = var_600_cast_fp16, y = var_601_cast_fp16)[name = string("op_602_cast_fp16")];461 tensor<fp16, [1, 8, 128, 64]> var_603_cast_fp16 = mul(x = x2_11_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_603_cast_fp16")];462 tensor<fp16, [1, 8, 128, 64]> var_604_cast_fp16 = mul(x = x1_11_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_604_cast_fp16")];463 tensor<fp16, [1, 8, 128, 64]> var_605_cast_fp16 = add(x = var_603_cast_fp16, y = var_604_cast_fp16)[name = string("op_605_cast_fp16")];464 bool var_607_interleave_0 = const()[name = string("op_607_interleave_0"), val = bool(false)];465 tensor<fp16, [1, 8, 128, 128]> var_607_cast_fp16 = concat(axis = var_487, interleave = var_607_interleave_0, values = (var_602_cast_fp16, var_605_cast_fp16))[name = string("op_607_cast_fp16")];466 tensor<int32, [4]> transpose_9_perm_0 = const()[name = string("transpose_9_perm_0"), val = tensor<int32, [4]>([2, 0, 1, 3])];467 tensor<int32, [2]> concat_40 = const()[name = string("concat_40"), val = tensor<int32, [2]>([128, 1024])];468 tensor<fp16, [128, 1, 8, 128]> transpose_9_cast_fp16 = transpose(perm = transpose_9_perm_0, x = var_607_cast_fp16)[name = string("transpose_81")];469 tensor<fp16, [128, 1024]> reshape_13_cast_fp16 = reshape(shape = concat_40, x = transpose_9_cast_fp16)[name = string("reshape_13_cast_fp16")];470 bool matmul_4_transpose_x_1 = const()[name = string("matmul_4_transpose_x_1"), val = bool(true)];471 bool matmul_4_transpose_y_1 = const()[name = string("matmul_4_transpose_y_1"), val = bool(false)];472 tensor<fp16, [1024, 1024]> matmul_4_cast_fp16 = matmul(transpose_x = matmul_4_transpose_x_1, transpose_y = matmul_4_transpose_y_1, x = var_68_to_fp16, y = reshape_13_cast_fp16)[name = string("matmul_4_cast_fp16")];473 tensor<int32, [4]> concat_43 = const()[name = string("concat_43"), val = tensor<int32, [4]>([1024, 1, 8, 128])];474 tensor<fp16, [1024, 1, 8, 128]> reshape_14_cast_fp16 = reshape(shape = concat_43, x = matmul_4_cast_fp16)[name = string("reshape_14_cast_fp16")];475 tensor<int32, [4]> scattered_k_5_perm_0 = const()[name = string("scattered_k_5_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];476 tensor<int32, [2]> concat_48 = const()[name = string("concat_48"), val = tensor<int32, [2]>([128, 1024])];477 tensor<fp16, [128, 1, 8, 128]> transpose_58_cast_fp16 = transpose(perm = transpose_58_perm_0, x = var_530_cast_fp16)[name = string("transpose_80")];478 tensor<fp16, [128, 1024]> reshape_16_cast_fp16 = reshape(shape = concat_48, x = transpose_58_cast_fp16)[name = string("reshape_16_cast_fp16")];479 bool matmul_5_transpose_x_1 = const()[name = string("matmul_5_transpose_x_1"), val = bool(true)];480 bool matmul_5_transpose_y_1 = const()[name = string("matmul_5_transpose_y_1"), val = bool(false)];481 tensor<fp16, [1024, 1024]> matmul_5_cast_fp16 = matmul(transpose_x = matmul_5_transpose_x_1, transpose_y = matmul_5_transpose_y_1, x = var_68_to_fp16, y = reshape_16_cast_fp16)[name = string("matmul_5_cast_fp16")];482 tensor<int32, [4]> concat_51 = const()[name = string("concat_51"), val = tensor<int32, [4]>([1024, 1, 8, 128])];483 tensor<fp16, [1024, 1, 8, 128]> reshape_17_cast_fp16 = reshape(shape = concat_51, x = matmul_5_cast_fp16)[name = string("reshape_17_cast_fp16")];484 tensor<int32, [4]> scattered_v_5_perm_0 = const()[name = string("scattered_v_5_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];485 tensor<fp16, [1, 8, 1024, 128]> read_state_4 = read_state(input = k_cache_2)[name = string("read_state_4")];486 tensor<fp16, [1, 8, 1024, 128]> k_cache_15_cast_fp16 = mul(x = read_state_4, y = var_224_cast_fp16)[name = string("k_cache_15_cast_fp16")];487 write_state(data = k_cache_15_cast_fp16, input = k_cache_2)[name = string("coreml_update_state_64_write_state")];488 tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_64 = read_state(input = k_cache_2)[name = string("coreml_update_state_64")];489 tensor<fp16, [1, 8, 1024, 128]> scattered_k_5_cast_fp16 = transpose(perm = scattered_k_5_perm_0, x = reshape_14_cast_fp16)[name = string("transpose_79")];490 tensor<fp16, [1, 8, 1024, 128]> k_cache_17_cast_fp16 = add(x = coreml_update_state_64, y = scattered_k_5_cast_fp16)[name = string("k_cache_17_cast_fp16")];491 write_state(data = k_cache_17_cast_fp16, input = k_cache_2)[name = string("coreml_update_state_65_write_state")];492 tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_65 = read_state(input = k_cache_2)[name = string("coreml_update_state_65")];493 tensor<fp16, [1, 8, 1024, 128]> read_state_5 = read_state(input = v_cache_2)[name = string("read_state_5")];494 tensor<fp16, [1, 8, 1024, 128]> v_cache_15_cast_fp16 = mul(x = read_state_5, y = var_224_cast_fp16)[name = string("v_cache_15_cast_fp16")];495 write_state(data = v_cache_15_cast_fp16, input = v_cache_2)[name = string("coreml_update_state_66_write_state")];496 tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_66 = read_state(input = v_cache_2)[name = string("coreml_update_state_66")];497 tensor<fp16, [1, 8, 1024, 128]> scattered_v_5_cast_fp16 = transpose(perm = scattered_v_5_perm_0, x = reshape_17_cast_fp16)[name = string("transpose_78")];498 tensor<fp16, [1, 8, 1024, 128]> v_cache_17_cast_fp16 = add(x = coreml_update_state_66, y = scattered_v_5_cast_fp16)[name = string("v_cache_17_cast_fp16")];499 write_state(data = v_cache_17_cast_fp16, input = v_cache_2)[name = string("coreml_update_state_67_write_state")];500 tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_67 = read_state(input = v_cache_2)[name = string("coreml_update_state_67")];501 tensor<int32, [1]> var_618_axes_0 = const()[name = string("op_618_axes_0"), val = tensor<int32, [1]>([2])];502 tensor<fp16, [1, 8, 1, 1024, 128]> var_618_cast_fp16 = expand_dims(axes = var_618_axes_0, x = coreml_update_state_65)[name = string("op_618_cast_fp16")];503 tensor<int32, [5]> k_exp_9_reps_0 = const()[name = string("k_exp_9_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];504 tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_9_cast_fp16 = tile(reps = k_exp_9_reps_0, x = var_618_cast_fp16)[name = string("k_exp_9_cast_fp16")];505 tensor<int32, [4]> var_621 = const()[name = string("op_621"), val = tensor<int32, [4]>([1, 16, 1024, 128])];506 tensor<fp16, [1, 16, 1024, 128]> k_exp_11_cast_fp16 = reshape(shape = var_621, x = k_exp_9_cast_fp16)[name = string("k_exp_11_cast_fp16")];507 tensor<int32, [1]> var_623_axes_0 = const()[name = string("op_623_axes_0"), val = tensor<int32, [1]>([2])];508 tensor<fp16, [1, 8, 1, 1024, 128]> var_623_cast_fp16 = expand_dims(axes = var_623_axes_0, x = coreml_update_state_67)[name = string("op_623_cast_fp16")];509 tensor<int32, [5]> v_exp_9_reps_0 = const()[name = string("v_exp_9_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];510 tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_9_cast_fp16 = tile(reps = v_exp_9_reps_0, x = var_623_cast_fp16)[name = string("v_exp_9_cast_fp16")];511 tensor<int32, [4]> var_626 = const()[name = string("op_626"), val = tensor<int32, [4]>([1, 16, 1024, 128])];512 tensor<fp16, [1, 16, 1024, 128]> v_exp_11_cast_fp16 = reshape(shape = var_626, x = v_exp_9_cast_fp16)[name = string("v_exp_11_cast_fp16")];513 bool var_629_transpose_x_1 = const()[name = string("op_629_transpose_x_1"), val = bool(false)];514 bool var_629_transpose_y_1 = const()[name = string("op_629_transpose_y_1"), val = bool(true)];515 tensor<fp16, [1, 16, 128, 1024]> var_629_cast_fp16 = matmul(transpose_x = var_629_transpose_x_1, transpose_y = var_629_transpose_y_1, x = q_5_cast_fp16, y = k_exp_11_cast_fp16)[name = string("op_629_cast_fp16")];516 fp16 var_630_to_fp16 = const()[name = string("op_630_to_fp16"), val = fp16(0x1.6ap-4)];517 tensor<fp16, [1, 16, 128, 1024]> attn_9_cast_fp16 = mul(x = var_629_cast_fp16, y = var_630_to_fp16)[name = string("attn_9_cast_fp16")];518 tensor<fp16, [1, 16, 128, 1024]> input_21_cast_fp16 = add(x = attn_9_cast_fp16, y = attention_mask_to_fp16)[name = string("input_21_cast_fp16")];519 tensor<fp16, [1, 16, 128, 1024]> attn_11_cast_fp16 = softmax(axis = var_487, x = input_21_cast_fp16)[name = string("attn_11_cast_fp16")];520 bool out_5_transpose_x_0 = const()[name = string("out_5_transpose_x_0"), val = bool(false)];521 bool out_5_transpose_y_0 = const()[name = string("out_5_transpose_y_0"), val = bool(false)];522 tensor<fp16, [1, 16, 128, 128]> out_5_cast_fp16 = matmul(transpose_x = out_5_transpose_x_0, transpose_y = out_5_transpose_y_0, x = attn_11_cast_fp16, y = v_exp_11_cast_fp16)[name = string("out_5_cast_fp16")];523 tensor<int32, [4]> var_635_perm_0 = const()[name = string("op_635_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];524 tensor<int32, [3]> var_636 = const()[name = string("op_636"), val = tensor<int32, [3]>([1, 128, -1])];525 tensor<fp16, [1, 128, 16, 128]> var_635_cast_fp16 = transpose(perm = var_635_perm_0, x = out_5_cast_fp16)[name = string("transpose_77")];526 tensor<fp16, [1, 128, 2048]> input_23_cast_fp16 = reshape(shape = var_636, x = var_635_cast_fp16)[name = string("input_23_cast_fp16")];527 tensor<fp16, [1024, 2048]> layers_2_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35687680))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37784896))))[name = string("layers_2_self_attn_o_proj_weight_to_fp16_palettized")];528 tensor<fp16, [1, 128, 1024]> linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_2_self_attn_o_proj_weight_to_fp16_palettized, x = input_23_cast_fp16)[name = string("linear_17_cast_fp16")];529 tensor<fp16, [1, 128, 1024]> x_71_cast_fp16 = add(x = x_51_cast_fp16, y = linear_17_cast_fp16)[name = string("x_71_cast_fp16")];530 fp16 var_486_promoted_3_to_fp16 = const()[name = string("op_486_promoted_3_to_fp16"), val = fp16(0x1p+1)];531 tensor<fp16, [1, 128, 1024]> var_643_cast_fp16 = pow(x = x_71_cast_fp16, y = var_486_promoted_3_to_fp16)[name = string("op_643_cast_fp16")];532 tensor<int32, [1]> var_645_axes_0 = const()[name = string("op_645_axes_0"), val = tensor<int32, [1]>([-1])];533 bool var_645_keep_dims_0 = const()[name = string("op_645_keep_dims_0"), val = bool(true)];534 tensor<fp16, [1, 128, 1]> var_645_cast_fp16 = reduce_mean(axes = var_645_axes_0, keep_dims = var_645_keep_dims_0, x = var_643_cast_fp16)[name = string("op_645_cast_fp16")];535 fp16 var_646_to_fp16 = const()[name = string("op_646_to_fp16"), val = fp16(0x1.1p-20)];536 tensor<fp16, [1, 128, 1]> var_647_cast_fp16 = add(x = var_645_cast_fp16, y = var_646_to_fp16)[name = string("op_647_cast_fp16")];537 fp32 norm_23_epsilon_0 = const()[name = string("norm_23_epsilon_0"), val = fp32(0x1.197998p-40)];538 tensor<fp16, [1, 128, 1]> norm_23_cast_fp16 = rsqrt(epsilon = norm_23_epsilon_0, x = var_647_cast_fp16)[name = string("norm_23_cast_fp16")];539 tensor<fp16, [1, 128, 1024]> var_649_cast_fp16 = mul(x = x_71_cast_fp16, y = norm_23_cast_fp16)[name = string("op_649_cast_fp16")];540 tensor<fp16, [1024]> layers_2_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_2_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37785472)))];541 tensor<fp16, [1, 128, 1024]> var_650_cast_fp16 = mul(x = var_649_cast_fp16, y = layers_2_post_attention_layernorm_weight_to_fp16)[name = string("op_650_cast_fp16")];542 tensor<fp16, [3072, 1024]> layers_2_mlp_gate_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37787584))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40933376))))[name = string("layers_2_mlp_gate_proj_weight_to_fp16_palettized")];543 tensor<fp16, [1, 128, 3072]> linear_18_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_2_mlp_gate_proj_weight_to_fp16_palettized, x = var_650_cast_fp16)[name = string("linear_18_cast_fp16")];544 tensor<fp16, [1, 128, 3072]> var_660_cast_fp16 = silu(x = linear_18_cast_fp16)[name = string("op_660_cast_fp16")];545 tensor<fp16, [3072, 1024]> layers_2_mlp_up_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40933952))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44079744))))[name = string("layers_2_mlp_up_proj_weight_to_fp16_palettized")];546 tensor<fp16, [1, 128, 3072]> linear_19_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_2_mlp_up_proj_weight_to_fp16_palettized, x = var_650_cast_fp16)[name = string("linear_19_cast_fp16")];547 tensor<fp16, [1, 128, 3072]> input_29_cast_fp16 = mul(x = var_660_cast_fp16, y = linear_19_cast_fp16)[name = string("input_29_cast_fp16")];548 tensor<fp16, [1024, 3072]> layers_2_mlp_down_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44080320))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47226112))))[name = string("layers_2_mlp_down_proj_weight_to_fp16_palettized")];549 tensor<fp16, [1, 128, 1024]> linear_20_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_2_mlp_down_proj_weight_to_fp16_palettized, x = input_29_cast_fp16)[name = string("linear_20_cast_fp16")];550 tensor<fp16, [1, 128, 1024]> x_77_cast_fp16 = add(x = x_71_cast_fp16, y = linear_20_cast_fp16)[name = string("x_77_cast_fp16")];551 int32 var_681 = const()[name = string("op_681"), val = int32(-1)];552 fp16 var_680_promoted_to_fp16 = const()[name = string("op_680_promoted_to_fp16"), val = fp16(0x1p+1)];553 tensor<fp16, [1, 128, 1024]> var_690_cast_fp16 = pow(x = x_77_cast_fp16, y = var_680_promoted_to_fp16)[name = string("op_690_cast_fp16")];554 tensor<int32, [1]> var_692_axes_0 = const()[name = string("op_692_axes_0"), val = tensor<int32, [1]>([-1])];555 bool var_692_keep_dims_0 = const()[name = string("op_692_keep_dims_0"), val = bool(true)];556 tensor<fp16, [1, 128, 1]> var_692_cast_fp16 = reduce_mean(axes = var_692_axes_0, keep_dims = var_692_keep_dims_0, x = var_690_cast_fp16)[name = string("op_692_cast_fp16")];557 fp16 var_693_to_fp16 = const()[name = string("op_693_to_fp16"), val = fp16(0x1.1p-20)];558 tensor<fp16, [1, 128, 1]> var_694_cast_fp16 = add(x = var_692_cast_fp16, y = var_693_to_fp16)[name = string("op_694_cast_fp16")];559 fp32 norm_25_epsilon_0 = const()[name = string("norm_25_epsilon_0"), val = fp32(0x1.197998p-40)];560 tensor<fp16, [1, 128, 1]> norm_25_cast_fp16 = rsqrt(epsilon = norm_25_epsilon_0, x = var_694_cast_fp16)[name = string("norm_25_cast_fp16")];561 tensor<fp16, [1, 128, 1024]> var_696_cast_fp16 = mul(x = x_77_cast_fp16, y = norm_25_cast_fp16)[name = string("op_696_cast_fp16")];562 tensor<fp16, [1024]> layers_3_input_layernorm_weight_to_fp16 = const()[name = string("layers_3_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47226688)))];563 tensor<fp16, [1, 128, 1024]> var_697_cast_fp16 = mul(x = var_696_cast_fp16, y = layers_3_input_layernorm_weight_to_fp16)[name = string("op_697_cast_fp16")];564 tensor<fp16, [2048, 1024]> layers_3_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47228800))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49326016))))[name = string("layers_3_self_attn_q_proj_weight_to_fp16_palettized")];565 tensor<fp16, [1, 128, 2048]> linear_21_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_3_self_attn_q_proj_weight_to_fp16_palettized, x = var_697_cast_fp16)[name = string("linear_21_cast_fp16")];566 tensor<int32, [4]> var_713 = const()[name = string("op_713"), val = tensor<int32, [4]>([1, 128, 16, 128])];567 tensor<fp16, [1, 128, 16, 128]> var_714_cast_fp16 = reshape(shape = var_713, x = linear_21_cast_fp16)[name = string("op_714_cast_fp16")];568 tensor<int32, [4]> x_83_perm_0 = const()[name = string("x_83_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];569 tensor<fp16, [1024, 1024]> layers_3_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49326592))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50375232))))[name = string("layers_3_self_attn_k_proj_weight_to_fp16_palettized")];570 tensor<fp16, [1, 128, 1024]> linear_22_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_3_self_attn_k_proj_weight_to_fp16_palettized, x = var_697_cast_fp16)[name = string("linear_22_cast_fp16")];571 tensor<int32, [4]> var_718 = const()[name = string("op_718"), val = tensor<int32, [4]>([1, 128, 8, 128])];572 tensor<fp16, [1, 128, 8, 128]> var_719_cast_fp16 = reshape(shape = var_718, x = linear_22_cast_fp16)[name = string("op_719_cast_fp16")];573 tensor<int32, [4]> x_87_perm_0 = const()[name = string("x_87_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];574 tensor<fp16, [1024, 1024]> layers_3_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50375808))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51424448))))[name = string("layers_3_self_attn_v_proj_weight_to_fp16_palettized")];575 tensor<fp16, [1, 128, 1024]> linear_23_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_3_self_attn_v_proj_weight_to_fp16_palettized, x = var_697_cast_fp16)[name = string("linear_23_cast_fp16")];576 tensor<int32, [4]> var_723 = const()[name = string("op_723"), val = tensor<int32, [4]>([1, 128, 8, 128])];577 tensor<fp16, [1, 128, 8, 128]> var_724_cast_fp16 = reshape(shape = var_723, x = linear_23_cast_fp16)[name = string("op_724_cast_fp16")];578 tensor<int32, [4]> transpose_59_perm_0 = const()[name = string("transpose_59_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];579 fp16 var_680_promoted_1_to_fp16 = const()[name = string("op_680_promoted_1_to_fp16"), val = fp16(0x1p+1)];580 tensor<fp16, [1, 16, 128, 128]> x_83_cast_fp16 = transpose(perm = x_83_perm_0, x = var_714_cast_fp16)[name = string("transpose_76")];581 tensor<fp16, [1, 16, 128, 128]> var_728_cast_fp16 = pow(x = x_83_cast_fp16, y = var_680_promoted_1_to_fp16)[name = string("op_728_cast_fp16")];582 tensor<int32, [1]> var_730_axes_0 = const()[name = string("op_730_axes_0"), val = tensor<int32, [1]>([-1])];583 bool var_730_keep_dims_0 = const()[name = string("op_730_keep_dims_0"), val = bool(true)];584 tensor<fp16, [1, 16, 128, 1]> var_730_cast_fp16 = reduce_mean(axes = var_730_axes_0, keep_dims = var_730_keep_dims_0, x = var_728_cast_fp16)[name = string("op_730_cast_fp16")];585 fp16 var_731_to_fp16 = const()[name = string("op_731_to_fp16"), val = fp16(0x1.1p-20)];586 tensor<fp16, [1, 16, 128, 1]> var_732_cast_fp16 = add(x = var_730_cast_fp16, y = var_731_to_fp16)[name = string("op_732_cast_fp16")];587 fp32 norm_27_epsilon_0 = const()[name = string("norm_27_epsilon_0"), val = fp32(0x1.197998p-40)];588 tensor<fp16, [1, 16, 128, 1]> norm_27_cast_fp16 = rsqrt(epsilon = norm_27_epsilon_0, x = var_732_cast_fp16)[name = string("norm_27_cast_fp16")];589 tensor<fp16, [1, 16, 128, 128]> var_734_cast_fp16 = mul(x = x_83_cast_fp16, y = norm_27_cast_fp16)[name = string("op_734_cast_fp16")];590 tensor<fp16, [128]> layers_3_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_3_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51425024)))];591 tensor<fp16, [1, 16, 128, 128]> var_735_cast_fp16 = mul(x = var_734_cast_fp16, y = layers_3_self_attn_q_norm_weight_to_fp16)[name = string("op_735_cast_fp16")];592 fp16 var_680_promoted_2_to_fp16 = const()[name = string("op_680_promoted_2_to_fp16"), val = fp16(0x1p+1)];593 tensor<fp16, [1, 8, 128, 128]> x_87_cast_fp16 = transpose(perm = x_87_perm_0, x = var_719_cast_fp16)[name = string("transpose_75")];594 tensor<fp16, [1, 8, 128, 128]> var_739_cast_fp16 = pow(x = x_87_cast_fp16, y = var_680_promoted_2_to_fp16)[name = string("op_739_cast_fp16")];595 tensor<int32, [1]> var_741_axes_0 = const()[name = string("op_741_axes_0"), val = tensor<int32, [1]>([-1])];596 bool var_741_keep_dims_0 = const()[name = string("op_741_keep_dims_0"), val = bool(true)];597 tensor<fp16, [1, 8, 128, 1]> var_741_cast_fp16 = reduce_mean(axes = var_741_axes_0, keep_dims = var_741_keep_dims_0, x = var_739_cast_fp16)[name = string("op_741_cast_fp16")];598 fp16 var_742_to_fp16 = const()[name = string("op_742_to_fp16"), val = fp16(0x1.1p-20)];599 tensor<fp16, [1, 8, 128, 1]> var_743_cast_fp16 = add(x = var_741_cast_fp16, y = var_742_to_fp16)[name = string("op_743_cast_fp16")];600 fp32 norm_29_epsilon_0 = const()[name = string("norm_29_epsilon_0"), val = fp32(0x1.197998p-40)];601 tensor<fp16, [1, 8, 128, 1]> norm_29_cast_fp16 = rsqrt(epsilon = norm_29_epsilon_0, x = var_743_cast_fp16)[name = string("norm_29_cast_fp16")];602 tensor<fp16, [1, 8, 128, 128]> var_745_cast_fp16 = mul(x = x_87_cast_fp16, y = norm_29_cast_fp16)[name = string("op_745_cast_fp16")];603 tensor<fp16, [128]> layers_3_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_3_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51425344)))];604 tensor<fp16, [1, 8, 128, 128]> var_746_cast_fp16 = mul(x = var_745_cast_fp16, y = layers_3_self_attn_k_norm_weight_to_fp16)[name = string("op_746_cast_fp16")];605 tensor<int32, [4]> x1_13_begin_0 = const()[name = string("x1_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];606 tensor<int32, [4]> x1_13_end_0 = const()[name = string("x1_13_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];607 tensor<bool, [4]> x1_13_end_mask_0 = const()[name = string("x1_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];608 tensor<fp16, [1, 16, 128, 64]> x1_13_cast_fp16 = slice_by_index(begin = x1_13_begin_0, end = x1_13_end_0, end_mask = x1_13_end_mask_0, x = var_735_cast_fp16)[name = string("x1_13_cast_fp16")];609 tensor<int32, [4]> x2_13_begin_0 = const()[name = string("x2_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];610 tensor<int32, [4]> x2_13_end_0 = const()[name = string("x2_13_end_0"), val = tensor<int32, [4]>([1, 16, 128, 128])];611 tensor<bool, [4]> x2_13_end_mask_0 = const()[name = string("x2_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];612 tensor<fp16, [1, 16, 128, 64]> x2_13_cast_fp16 = slice_by_index(begin = x2_13_begin_0, end = x2_13_end_0, end_mask = x2_13_end_mask_0, x = var_735_cast_fp16)[name = string("x2_13_cast_fp16")];613 tensor<fp16, [1, 16, 128, 64]> var_767_cast_fp16 = mul(x = x1_13_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_767_cast_fp16")];614 tensor<fp16, [1, 16, 128, 64]> var_768_cast_fp16 = mul(x = x2_13_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_768_cast_fp16")];615 tensor<fp16, [1, 16, 128, 64]> var_769_cast_fp16 = sub(x = var_767_cast_fp16, y = var_768_cast_fp16)[name = string("op_769_cast_fp16")];616 tensor<fp16, [1, 16, 128, 64]> var_770_cast_fp16 = mul(x = x2_13_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_770_cast_fp16")];617 tensor<fp16, [1, 16, 128, 64]> var_771_cast_fp16 = mul(x = x1_13_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_771_cast_fp16")];618 tensor<fp16, [1, 16, 128, 64]> var_772_cast_fp16 = add(x = var_770_cast_fp16, y = var_771_cast_fp16)[name = string("op_772_cast_fp16")];619 bool q_7_interleave_0 = const()[name = string("q_7_interleave_0"), val = bool(false)];620 tensor<fp16, [1, 16, 128, 128]> q_7_cast_fp16 = concat(axis = var_681, interleave = q_7_interleave_0, values = (var_769_cast_fp16, var_772_cast_fp16))[name = string("q_7_cast_fp16")];621 tensor<int32, [4]> x1_15_begin_0 = const()[name = string("x1_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];622 tensor<int32, [4]> x1_15_end_0 = const()[name = string("x1_15_end_0"), val = tensor<int32, [4]>([1, 8, 128, 64])];623 tensor<bool, [4]> x1_15_end_mask_0 = const()[name = string("x1_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];624 tensor<fp16, [1, 8, 128, 64]> x1_15_cast_fp16 = slice_by_index(begin = x1_15_begin_0, end = x1_15_end_0, end_mask = x1_15_end_mask_0, x = var_746_cast_fp16)[name = string("x1_15_cast_fp16")];625 tensor<int32, [4]> x2_15_begin_0 = const()[name = string("x2_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];626 tensor<int32, [4]> x2_15_end_0 = const()[name = string("x2_15_end_0"), val = tensor<int32, [4]>([1, 8, 128, 128])];627 tensor<bool, [4]> x2_15_end_mask_0 = const()[name = string("x2_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];628 tensor<fp16, [1, 8, 128, 64]> x2_15_cast_fp16 = slice_by_index(begin = x2_15_begin_0, end = x2_15_end_0, end_mask = x2_15_end_mask_0, x = var_746_cast_fp16)[name = string("x2_15_cast_fp16")];629 tensor<fp16, [1, 8, 128, 64]> var_794_cast_fp16 = mul(x = x1_15_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_794_cast_fp16")];630 tensor<fp16, [1, 8, 128, 64]> var_795_cast_fp16 = mul(x = x2_15_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_795_cast_fp16")];631 tensor<fp16, [1, 8, 128, 64]> var_796_cast_fp16 = sub(x = var_794_cast_fp16, y = var_795_cast_fp16)[name = string("op_796_cast_fp16")];632 tensor<fp16, [1, 8, 128, 64]> var_797_cast_fp16 = mul(x = x2_15_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_797_cast_fp16")];633 tensor<fp16, [1, 8, 128, 64]> var_798_cast_fp16 = mul(x = x1_15_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_798_cast_fp16")];634 tensor<fp16, [1, 8, 128, 64]> var_799_cast_fp16 = add(x = var_797_cast_fp16, y = var_798_cast_fp16)[name = string("op_799_cast_fp16")];635 bool var_801_interleave_0 = const()[name = string("op_801_interleave_0"), val = bool(false)];636 tensor<fp16, [1, 8, 128, 128]> var_801_cast_fp16 = concat(axis = var_681, interleave = var_801_interleave_0, values = (var_796_cast_fp16, var_799_cast_fp16))[name = string("op_801_cast_fp16")];637 tensor<int32, [4]> transpose_13_perm_0 = const()[name = string("transpose_13_perm_0"), val = tensor<int32, [4]>([2, 0, 1, 3])];638 tensor<int32, [2]> concat_58 = const()[name = string("concat_58"), val = tensor<int32, [2]>([128, 1024])];639 tensor<fp16, [128, 1, 8, 128]> transpose_13_cast_fp16 = transpose(perm = transpose_13_perm_0, x = var_801_cast_fp16)[name = string("transpose_74")];640 tensor<fp16, [128, 1024]> reshape_19_cast_fp16 = reshape(shape = concat_58, x = transpose_13_cast_fp16)[name = string("reshape_19_cast_fp16")];641 bool matmul_6_transpose_x_1 = const()[name = string("matmul_6_transpose_x_1"), val = bool(true)];642 bool matmul_6_transpose_y_1 = const()[name = string("matmul_6_transpose_y_1"), val = bool(false)];643 tensor<fp16, [1024, 1024]> matmul_6_cast_fp16 = matmul(transpose_x = matmul_6_transpose_x_1, transpose_y = matmul_6_transpose_y_1, x = var_68_to_fp16, y = reshape_19_cast_fp16)[name = string("matmul_6_cast_fp16")];644 tensor<int32, [4]> concat_61 = const()[name = string("concat_61"), val = tensor<int32, [4]>([1024, 1, 8, 128])];645 tensor<fp16, [1024, 1, 8, 128]> reshape_20_cast_fp16 = reshape(shape = concat_61, x = matmul_6_cast_fp16)[name = string("reshape_20_cast_fp16")];646 tensor<int32, [4]> scattered_k_7_perm_0 = const()[name = string("scattered_k_7_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];647 tensor<int32, [2]> concat_66 = const()[name = string("concat_66"), val = tensor<int32, [2]>([128, 1024])];648 tensor<fp16, [128, 1, 8, 128]> transpose_59_cast_fp16 = transpose(perm = transpose_59_perm_0, x = var_724_cast_fp16)[name = string("transpose_73")];649 tensor<fp16, [128, 1024]> reshape_22_cast_fp16 = reshape(shape = concat_66, x = transpose_59_cast_fp16)[name = string("reshape_22_cast_fp16")];650 bool matmul_7_transpose_x_1 = const()[name = string("matmul_7_transpose_x_1"), val = bool(true)];651 bool matmul_7_transpose_y_1 = const()[name = string("matmul_7_transpose_y_1"), val = bool(false)];652 tensor<fp16, [1024, 1024]> matmul_7_cast_fp16 = matmul(transpose_x = matmul_7_transpose_x_1, transpose_y = matmul_7_transpose_y_1, x = var_68_to_fp16, y = reshape_22_cast_fp16)[name = string("matmul_7_cast_fp16")];653 tensor<int32, [4]> concat_69 = const()[name = string("concat_69"), val = tensor<int32, [4]>([1024, 1, 8, 128])];654 tensor<fp16, [1024, 1, 8, 128]> reshape_23_cast_fp16 = reshape(shape = concat_69, x = matmul_7_cast_fp16)[name = string("reshape_23_cast_fp16")];655 tensor<int32, [4]> scattered_v_7_perm_0 = const()[name = string("scattered_v_7_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];656 tensor<fp16, [1, 8, 1024, 128]> read_state_6 = read_state(input = k_cache_3)[name = string("read_state_6")];657 tensor<fp16, [1, 8, 1024, 128]> k_cache_21_cast_fp16 = mul(x = read_state_6, y = var_224_cast_fp16)[name = string("k_cache_21_cast_fp16")];658 write_state(data = k_cache_21_cast_fp16, input = k_cache_3)[name = string("coreml_update_state_68_write_state")];659 tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_68 = read_state(input = k_cache_3)[name = string("coreml_update_state_68")];660 tensor<fp16, [1, 8, 1024, 128]> scattered_k_7_cast_fp16 = transpose(perm = scattered_k_7_perm_0, x = reshape_20_cast_fp16)[name = string("transpose_72")];661 tensor<fp16, [1, 8, 1024, 128]> k_cache_23_cast_fp16 = add(x = coreml_update_state_68, y = scattered_k_7_cast_fp16)[name = string("k_cache_23_cast_fp16")];662 write_state(data = k_cache_23_cast_fp16, input = k_cache_3)[name = string("coreml_update_state_69_write_state")];663 tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_69 = read_state(input = k_cache_3)[name = string("coreml_update_state_69")];664 tensor<fp16, [1, 8, 1024, 128]> read_state_7 = read_state(input = v_cache_3)[name = string("read_state_7")];665 tensor<fp16, [1, 8, 1024, 128]> v_cache_21_cast_fp16 = mul(x = read_state_7, y = var_224_cast_fp16)[name = string("v_cache_21_cast_fp16")];666 write_state(data = v_cache_21_cast_fp16, input = v_cache_3)[name = string("coreml_update_state_70_write_state")];667 tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_70 = read_state(input = v_cache_3)[name = string("coreml_update_state_70")];668 tensor<fp16, [1, 8, 1024, 128]> scattered_v_7_cast_fp16 = transpose(perm = scattered_v_7_perm_0, x = reshape_23_cast_fp16)[name = string("transpose_71")];669 tensor<fp16, [1, 8, 1024, 128]> v_cache_23_cast_fp16 = add(x = coreml_update_state_70, y = scattered_v_7_cast_fp16)[name = string("v_cache_23_cast_fp16")];670 write_state(data = v_cache_23_cast_fp16, input = v_cache_3)[name = string("coreml_update_state_71_write_state")];671 tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_71 = read_state(input = v_cache_3)[name = string("coreml_update_state_71")];672 tensor<int32, [1]> var_812_axes_0 = const()[name = string("op_812_axes_0"), val = tensor<int32, [1]>([2])];673 tensor<fp16, [1, 8, 1, 1024, 128]> var_812_cast_fp16 = expand_dims(axes = var_812_axes_0, x = coreml_update_state_69)[name = string("op_812_cast_fp16")];674 tensor<int32, [5]> k_exp_13_reps_0 = const()[name = string("k_exp_13_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];675 tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_13_cast_fp16 = tile(reps = k_exp_13_reps_0, x = var_812_cast_fp16)[name = string("k_exp_13_cast_fp16")];676 tensor<int32, [4]> var_815 = const()[name = string("op_815"), val = tensor<int32, [4]>([1, 16, 1024, 128])];677 tensor<fp16, [1, 16, 1024, 128]> k_exp_15_cast_fp16 = reshape(shape = var_815, x = k_exp_13_cast_fp16)[name = string("k_exp_15_cast_fp16")];678 tensor<int32, [1]> var_817_axes_0 = const()[name = string("op_817_axes_0"), val = tensor<int32, [1]>([2])];679 tensor<fp16, [1, 8, 1, 1024, 128]> var_817_cast_fp16 = expand_dims(axes = var_817_axes_0, x = coreml_update_state_71)[name = string("op_817_cast_fp16")];680 tensor<int32, [5]> v_exp_13_reps_0 = const()[name = string("v_exp_13_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];681 tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_13_cast_fp16 = tile(reps = v_exp_13_reps_0, x = var_817_cast_fp16)[name = string("v_exp_13_cast_fp16")];682 tensor<int32, [4]> var_820 = const()[name = string("op_820"), val = tensor<int32, [4]>([1, 16, 1024, 128])];683 tensor<fp16, [1, 16, 1024, 128]> v_exp_15_cast_fp16 = reshape(shape = var_820, x = v_exp_13_cast_fp16)[name = string("v_exp_15_cast_fp16")];684 bool var_823_transpose_x_1 = const()[name = string("op_823_transpose_x_1"), val = bool(false)];685 bool var_823_transpose_y_1 = const()[name = string("op_823_transpose_y_1"), val = bool(true)];686 tensor<fp16, [1, 16, 128, 1024]> var_823_cast_fp16 = matmul(transpose_x = var_823_transpose_x_1, transpose_y = var_823_transpose_y_1, x = q_7_cast_fp16, y = k_exp_15_cast_fp16)[name = string("op_823_cast_fp16")];687 fp16 var_824_to_fp16 = const()[name = string("op_824_to_fp16"), val = fp16(0x1.6ap-4)];688 tensor<fp16, [1, 16, 128, 1024]> attn_13_cast_fp16 = mul(x = var_823_cast_fp16, y = var_824_to_fp16)[name = string("attn_13_cast_fp16")];689 tensor<fp16, [1, 16, 128, 1024]> input_31_cast_fp16 = add(x = attn_13_cast_fp16, y = attention_mask_to_fp16)[name = string("input_31_cast_fp16")];690 tensor<fp16, [1, 16, 128, 1024]> attn_15_cast_fp16 = softmax(axis = var_681, x = input_31_cast_fp16)[name = string("attn_15_cast_fp16")];691 bool out_7_transpose_x_0 = const()[name = string("out_7_transpose_x_0"), val = bool(false)];692 bool out_7_transpose_y_0 = const()[name = string("out_7_transpose_y_0"), val = bool(false)];693 tensor<fp16, [1, 16, 128, 128]> out_7_cast_fp16 = matmul(transpose_x = out_7_transpose_x_0, transpose_y = out_7_transpose_y_0, x = attn_15_cast_fp16, y = v_exp_15_cast_fp16)[name = string("out_7_cast_fp16")];694 tensor<int32, [4]> var_829_perm_0 = const()[name = string("op_829_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];695 tensor<int32, [3]> var_830 = const()[name = string("op_830"), val = tensor<int32, [3]>([1, 128, -1])];696 tensor<fp16, [1, 128, 16, 128]> var_829_cast_fp16 = transpose(perm = var_829_perm_0, x = out_7_cast_fp16)[name = string("transpose_70")];697 tensor<fp16, [1, 128, 2048]> input_33_cast_fp16 = reshape(shape = var_830, x = var_829_cast_fp16)[name = string("input_33_cast_fp16")];698 tensor<fp16, [1024, 2048]> layers_3_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51425664))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53522880))))[name = string("layers_3_self_attn_o_proj_weight_to_fp16_palettized")];699 tensor<fp16, [1, 128, 1024]> linear_24_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_3_self_attn_o_proj_weight_to_fp16_palettized, x = input_33_cast_fp16)[name = string("linear_24_cast_fp16")];700 tensor<fp16, [1, 128, 1024]> x_97_cast_fp16 = add(x = x_77_cast_fp16, y = linear_24_cast_fp16)[name = string("x_97_cast_fp16")];701 fp16 var_680_promoted_3_to_fp16 = const()[name = string("op_680_promoted_3_to_fp16"), val = fp16(0x1p+1)];702 tensor<fp16, [1, 128, 1024]> var_837_cast_fp16 = pow(x = x_97_cast_fp16, y = var_680_promoted_3_to_fp16)[name = string("op_837_cast_fp16")];703 tensor<int32, [1]> var_839_axes_0 = const()[name = string("op_839_axes_0"), val = tensor<int32, [1]>([-1])];704 bool var_839_keep_dims_0 = const()[name = string("op_839_keep_dims_0"), val = bool(true)];705 tensor<fp16, [1, 128, 1]> var_839_cast_fp16 = reduce_mean(axes = var_839_axes_0, keep_dims = var_839_keep_dims_0, x = var_837_cast_fp16)[name = string("op_839_cast_fp16")];706 fp16 var_840_to_fp16 = const()[name = string("op_840_to_fp16"), val = fp16(0x1.1p-20)];707 tensor<fp16, [1, 128, 1]> var_841_cast_fp16 = add(x = var_839_cast_fp16, y = var_840_to_fp16)[name = string("op_841_cast_fp16")];708 fp32 norm_31_epsilon_0 = const()[name = string("norm_31_epsilon_0"), val = fp32(0x1.197998p-40)];709 tensor<fp16, [1, 128, 1]> norm_31_cast_fp16 = rsqrt(epsilon = norm_31_epsilon_0, x = var_841_cast_fp16)[name = string("norm_31_cast_fp16")];710 tensor<fp16, [1, 128, 1024]> var_843_cast_fp16 = mul(x = x_97_cast_fp16, y = norm_31_cast_fp16)[name = string("op_843_cast_fp16")];711 tensor<fp16, [1024]> layers_3_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_3_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53523456)))];712 tensor<fp16, [1, 128, 1024]> var_844_cast_fp16 = mul(x = var_843_cast_fp16, y = layers_3_post_attention_layernorm_weight_to_fp16)[name = string("op_844_cast_fp16")];713 tensor<fp16, [3072, 1024]> layers_3_mlp_gate_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53525568))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56671360))))[name = string("layers_3_mlp_gate_proj_weight_to_fp16_palettized")];714 tensor<fp16, [1, 128, 3072]> linear_25_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_3_mlp_gate_proj_weight_to_fp16_palettized, x = var_844_cast_fp16)[name = string("linear_25_cast_fp16")];715 tensor<fp16, [1, 128, 3072]> var_854_cast_fp16 = silu(x = linear_25_cast_fp16)[name = string("op_854_cast_fp16")];716 tensor<fp16, [3072, 1024]> layers_3_mlp_up_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56671936))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59817728))))[name = string("layers_3_mlp_up_proj_weight_to_fp16_palettized")];717 tensor<fp16, [1, 128, 3072]> linear_26_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_3_mlp_up_proj_weight_to_fp16_palettized, x = var_844_cast_fp16)[name = string("linear_26_cast_fp16")];718 tensor<fp16, [1, 128, 3072]> input_39_cast_fp16 = mul(x = var_854_cast_fp16, y = linear_26_cast_fp16)[name = string("input_39_cast_fp16")];719 tensor<fp16, [1024, 3072]> layers_3_mlp_down_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59818304))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62964096))))[name = string("layers_3_mlp_down_proj_weight_to_fp16_palettized")];720 tensor<fp16, [1, 128, 1024]> linear_27_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_3_mlp_down_proj_weight_to_fp16_palettized, x = input_39_cast_fp16)[name = string("linear_27_cast_fp16")];721 tensor<fp16, [1, 128, 1024]> x_103_cast_fp16 = add(x = x_97_cast_fp16, y = linear_27_cast_fp16)[name = string("x_103_cast_fp16")];722 int32 var_875 = const()[name = string("op_875"), val = int32(-1)];723 fp16 var_874_promoted_to_fp16 = const()[name = string("op_874_promoted_to_fp16"), val = fp16(0x1p+1)];724 tensor<fp16, [1, 128, 1024]> var_884_cast_fp16 = pow(x = x_103_cast_fp16, y = var_874_promoted_to_fp16)[name = string("op_884_cast_fp16")];725 tensor<int32, [1]> var_886_axes_0 = const()[name = string("op_886_axes_0"), val = tensor<int32, [1]>([-1])];726 bool var_886_keep_dims_0 = const()[name = string("op_886_keep_dims_0"), val = bool(true)];727 tensor<fp16, [1, 128, 1]> var_886_cast_fp16 = reduce_mean(axes = var_886_axes_0, keep_dims = var_886_keep_dims_0, x = var_884_cast_fp16)[name = string("op_886_cast_fp16")];728 fp16 var_887_to_fp16 = const()[name = string("op_887_to_fp16"), val = fp16(0x1.1p-20)];729 tensor<fp16, [1, 128, 1]> var_888_cast_fp16 = add(x = var_886_cast_fp16, y = var_887_to_fp16)[name = string("op_888_cast_fp16")];730 fp32 norm_33_epsilon_0 = const()[name = string("norm_33_epsilon_0"), val = fp32(0x1.197998p-40)];731 tensor<fp16, [1, 128, 1]> norm_33_cast_fp16 = rsqrt(epsilon = norm_33_epsilon_0, x = var_888_cast_fp16)[name = string("norm_33_cast_fp16")];732 tensor<fp16, [1, 128, 1024]> var_890_cast_fp16 = mul(x = x_103_cast_fp16, y = norm_33_cast_fp16)[name = string("op_890_cast_fp16")];733 tensor<fp16, [1024]> layers_4_input_layernorm_weight_to_fp16 = const()[name = string("layers_4_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62964672)))];734 tensor<fp16, [1, 128, 1024]> var_891_cast_fp16 = mul(x = var_890_cast_fp16, y = layers_4_input_layernorm_weight_to_fp16)[name = string("op_891_cast_fp16")];735 tensor<fp16, [2048, 1024]> layers_4_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62966784))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65064000))))[name = string("layers_4_self_attn_q_proj_weight_to_fp16_palettized")];736 tensor<fp16, [1, 128, 2048]> linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_4_self_attn_q_proj_weight_to_fp16_palettized, x = var_891_cast_fp16)[name = string("linear_28_cast_fp16")];737 tensor<int32, [4]> var_907 = const()[name = string("op_907"), val = tensor<int32, [4]>([1, 128, 16, 128])];738 tensor<fp16, [1, 128, 16, 128]> var_908_cast_fp16 = reshape(shape = var_907, x = linear_28_cast_fp16)[name = string("op_908_cast_fp16")];739 tensor<int32, [4]> x_109_perm_0 = const()[name = string("x_109_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];740 tensor<fp16, [1024, 1024]> layers_4_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65064576))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66113216))))[name = string("layers_4_self_attn_k_proj_weight_to_fp16_palettized")];741 tensor<fp16, [1, 128, 1024]> linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_4_self_attn_k_proj_weight_to_fp16_palettized, x = var_891_cast_fp16)[name = string("linear_29_cast_fp16")];742 tensor<int32, [4]> var_912 = const()[name = string("op_912"), val = tensor<int32, [4]>([1, 128, 8, 128])];743 tensor<fp16, [1, 128, 8, 128]> var_913_cast_fp16 = reshape(shape = var_912, x = linear_29_cast_fp16)[name = string("op_913_cast_fp16")];744 tensor<int32, [4]> x_113_perm_0 = const()[name = string("x_113_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];745 tensor<fp16, [1024, 1024]> layers_4_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66113792))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67162432))))[name = string("layers_4_self_attn_v_proj_weight_to_fp16_palettized")];746 tensor<fp16, [1, 128, 1024]> linear_30_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_4_self_attn_v_proj_weight_to_fp16_palettized, x = var_891_cast_fp16)[name = string("linear_30_cast_fp16")];747 tensor<int32, [4]> var_917 = const()[name = string("op_917"), val = tensor<int32, [4]>([1, 128, 8, 128])];748 tensor<fp16, [1, 128, 8, 128]> var_918_cast_fp16 = reshape(shape = var_917, x = linear_30_cast_fp16)[name = string("op_918_cast_fp16")];749 tensor<int32, [4]> transpose_60_perm_0 = const()[name = string("transpose_60_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];750 fp16 var_874_promoted_1_to_fp16 = const()[name = string("op_874_promoted_1_to_fp16"), val = fp16(0x1p+1)];751 tensor<fp16, [1, 16, 128, 128]> x_109_cast_fp16 = transpose(perm = x_109_perm_0, x = var_908_cast_fp16)[name = string("transpose_69")];752 tensor<fp16, [1, 16, 128, 128]> var_922_cast_fp16 = pow(x = x_109_cast_fp16, y = var_874_promoted_1_to_fp16)[name = string("op_922_cast_fp16")];753 tensor<int32, [1]> var_924_axes_0 = const()[name = string("op_924_axes_0"), val = tensor<int32, [1]>([-1])];754 bool var_924_keep_dims_0 = const()[name = string("op_924_keep_dims_0"), val = bool(true)];755 tensor<fp16, [1, 16, 128, 1]> var_924_cast_fp16 = reduce_mean(axes = var_924_axes_0, keep_dims = var_924_keep_dims_0, x = var_922_cast_fp16)[name = string("op_924_cast_fp16")];756 fp16 var_925_to_fp16 = const()[name = string("op_925_to_fp16"), val = fp16(0x1.1p-20)];757 tensor<fp16, [1, 16, 128, 1]> var_926_cast_fp16 = add(x = var_924_cast_fp16, y = var_925_to_fp16)[name = string("op_926_cast_fp16")];758 fp32 norm_35_epsilon_0 = const()[name = string("norm_35_epsilon_0"), val = fp32(0x1.197998p-40)];759 tensor<fp16, [1, 16, 128, 1]> norm_35_cast_fp16 = rsqrt(epsilon = norm_35_epsilon_0, x = var_926_cast_fp16)[name = string("norm_35_cast_fp16")];760 tensor<fp16, [1, 16, 128, 128]> var_928_cast_fp16 = mul(x = x_109_cast_fp16, y = norm_35_cast_fp16)[name = string("op_928_cast_fp16")];761 tensor<fp16, [128]> layers_4_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_4_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67163008)))];762 tensor<fp16, [1, 16, 128, 128]> var_929_cast_fp16 = mul(x = var_928_cast_fp16, y = layers_4_self_attn_q_norm_weight_to_fp16)[name = string("op_929_cast_fp16")];763 fp16 var_874_promoted_2_to_fp16 = const()[name = string("op_874_promoted_2_to_fp16"), val = fp16(0x1p+1)];764 tensor<fp16, [1, 8, 128, 128]> x_113_cast_fp16 = transpose(perm = x_113_perm_0, x = var_913_cast_fp16)[name = string("transpose_68")];765 tensor<fp16, [1, 8, 128, 128]> var_933_cast_fp16 = pow(x = x_113_cast_fp16, y = var_874_promoted_2_to_fp16)[name = string("op_933_cast_fp16")];766 tensor<int32, [1]> var_935_axes_0 = const()[name = string("op_935_axes_0"), val = tensor<int32, [1]>([-1])];767 bool var_935_keep_dims_0 = const()[name = string("op_935_keep_dims_0"), val = bool(true)];768 tensor<fp16, [1, 8, 128, 1]> var_935_cast_fp16 = reduce_mean(axes = var_935_axes_0, keep_dims = var_935_keep_dims_0, x = var_933_cast_fp16)[name = string("op_935_cast_fp16")];769 fp16 var_936_to_fp16 = const()[name = string("op_936_to_fp16"), val = fp16(0x1.1p-20)];770 tensor<fp16, [1, 8, 128, 1]> var_937_cast_fp16 = add(x = var_935_cast_fp16, y = var_936_to_fp16)[name = string("op_937_cast_fp16")];771 fp32 norm_37_epsilon_0 = const()[name = string("norm_37_epsilon_0"), val = fp32(0x1.197998p-40)];772 tensor<fp16, [1, 8, 128, 1]> norm_37_cast_fp16 = rsqrt(epsilon = norm_37_epsilon_0, x = var_937_cast_fp16)[name = string("norm_37_cast_fp16")];773 tensor<fp16, [1, 8, 128, 128]> var_939_cast_fp16 = mul(x = x_113_cast_fp16, y = norm_37_cast_fp16)[name = string("op_939_cast_fp16")];774 tensor<fp16, [128]> layers_4_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_4_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67163328)))];775 tensor<fp16, [1, 8, 128, 128]> var_940_cast_fp16 = mul(x = var_939_cast_fp16, y = layers_4_self_attn_k_norm_weight_to_fp16)[name = string("op_940_cast_fp16")];776 tensor<int32, [4]> x1_17_begin_0 = const()[name = string("x1_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];777 tensor<int32, [4]> x1_17_end_0 = const()[name = string("x1_17_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];778 tensor<bool, [4]> x1_17_end_mask_0 = const()[name = string("x1_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];779 tensor<fp16, [1, 16, 128, 64]> x1_17_cast_fp16 = slice_by_index(begin = x1_17_begin_0, end = x1_17_end_0, end_mask = x1_17_end_mask_0, x = var_929_cast_fp16)[name = string("x1_17_cast_fp16")];780 tensor<int32, [4]> x2_17_begin_0 = const()[name = string("x2_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];781 tensor<int32, [4]> x2_17_end_0 = const()[name = string("x2_17_end_0"), val = tensor<int32, [4]>([1, 16, 128, 128])];782 tensor<bool, [4]> x2_17_end_mask_0 = const()[name = string("x2_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];783 tensor<fp16, [1, 16, 128, 64]> x2_17_cast_fp16 = slice_by_index(begin = x2_17_begin_0, end = x2_17_end_0, end_mask = x2_17_end_mask_0, x = var_929_cast_fp16)[name = string("x2_17_cast_fp16")];784 tensor<fp16, [1, 16, 128, 64]> var_961_cast_fp16 = mul(x = x1_17_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_961_cast_fp16")];785 tensor<fp16, [1, 16, 128, 64]> var_962_cast_fp16 = mul(x = x2_17_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_962_cast_fp16")];786 tensor<fp16, [1, 16, 128, 64]> var_963_cast_fp16 = sub(x = var_961_cast_fp16, y = var_962_cast_fp16)[name = string("op_963_cast_fp16")];787 tensor<fp16, [1, 16, 128, 64]> var_964_cast_fp16 = mul(x = x2_17_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_964_cast_fp16")];788 tensor<fp16, [1, 16, 128, 64]> var_965_cast_fp16 = mul(x = x1_17_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_965_cast_fp16")];789 tensor<fp16, [1, 16, 128, 64]> var_966_cast_fp16 = add(x = var_964_cast_fp16, y = var_965_cast_fp16)[name = string("op_966_cast_fp16")];790 bool q_9_interleave_0 = const()[name = string("q_9_interleave_0"), val = bool(false)];791 tensor<fp16, [1, 16, 128, 128]> q_9_cast_fp16 = concat(axis = var_875, interleave = q_9_interleave_0, values = (var_963_cast_fp16, var_966_cast_fp16))[name = string("q_9_cast_fp16")];792 tensor<int32, [4]> x1_19_begin_0 = const()[name = string("x1_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];793 tensor<int32, [4]> x1_19_end_0 = const()[name = string("x1_19_end_0"), val = tensor<int32, [4]>([1, 8, 128, 64])];794 tensor<bool, [4]> x1_19_end_mask_0 = const()[name = string("x1_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];795 tensor<fp16, [1, 8, 128, 64]> x1_19_cast_fp16 = slice_by_index(begin = x1_19_begin_0, end = x1_19_end_0, end_mask = x1_19_end_mask_0, x = var_940_cast_fp16)[name = string("x1_19_cast_fp16")];796 tensor<int32, [4]> x2_19_begin_0 = const()[name = string("x2_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];797 tensor<int32, [4]> x2_19_end_0 = const()[name = string("x2_19_end_0"), val = tensor<int32, [4]>([1, 8, 128, 128])];798 tensor<bool, [4]> x2_19_end_mask_0 = const()[name = string("x2_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];799 tensor<fp16, [1, 8, 128, 64]> x2_19_cast_fp16 = slice_by_index(begin = x2_19_begin_0, end = x2_19_end_0, end_mask = x2_19_end_mask_0, x = var_940_cast_fp16)[name = string("x2_19_cast_fp16")];800 tensor<fp16, [1, 8, 128, 64]> var_988_cast_fp16 = mul(x = x1_19_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_988_cast_fp16")];801 tensor<fp16, [1, 8, 128, 64]> var_989_cast_fp16 = mul(x = x2_19_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_989_cast_fp16")];802 tensor<fp16, [1, 8, 128, 64]> var_990_cast_fp16 = sub(x = var_988_cast_fp16, y = var_989_cast_fp16)[name = string("op_990_cast_fp16")];803 tensor<fp16, [1, 8, 128, 64]> var_991_cast_fp16 = mul(x = x2_19_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_991_cast_fp16")];804 tensor<fp16, [1, 8, 128, 64]> var_992_cast_fp16 = mul(x = x1_19_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_992_cast_fp16")];805 tensor<fp16, [1, 8, 128, 64]> var_993_cast_fp16 = add(x = var_991_cast_fp16, y = var_992_cast_fp16)[name = string("op_993_cast_fp16")];806 bool var_995_interleave_0 = const()[name = string("op_995_interleave_0"), val = bool(false)];807 tensor<fp16, [1, 8, 128, 128]> var_995_cast_fp16 = concat(axis = var_875, interleave = var_995_interleave_0, values = (var_990_cast_fp16, var_993_cast_fp16))[name = string("op_995_cast_fp16")];808 tensor<int32, [4]> transpose_17_perm_0 = const()[name = string("transpose_17_perm_0"), val = tensor<int32, [4]>([2, 0, 1, 3])];809 tensor<int32, [2]> concat_76 = const()[name = string("concat_76"), val = tensor<int32, [2]>([128, 1024])];810 tensor<fp16, [128, 1, 8, 128]> transpose_17_cast_fp16 = transpose(perm = transpose_17_perm_0, x = var_995_cast_fp16)[name = string("transpose_67")];811 tensor<fp16, [128, 1024]> reshape_25_cast_fp16 = reshape(shape = concat_76, x = transpose_17_cast_fp16)[name = string("reshape_25_cast_fp16")];812 bool matmul_8_transpose_x_1 = const()[name = string("matmul_8_transpose_x_1"), val = bool(true)];813 bool matmul_8_transpose_y_1 = const()[name = string("matmul_8_transpose_y_1"), val = bool(false)];814 tensor<fp16, [1024, 1024]> matmul_8_cast_fp16 = matmul(transpose_x = matmul_8_transpose_x_1, transpose_y = matmul_8_transpose_y_1, x = var_68_to_fp16, y = reshape_25_cast_fp16)[name = string("matmul_8_cast_fp16")];815 tensor<int32, [4]> concat_79 = const()[name = string("concat_79"), val = tensor<int32, [4]>([1024, 1, 8, 128])];816 tensor<fp16, [1024, 1, 8, 128]> reshape_26_cast_fp16 = reshape(shape = concat_79, x = matmul_8_cast_fp16)[name = string("reshape_26_cast_fp16")];817 tensor<int32, [4]> scattered_k_9_perm_0 = const()[name = string("scattered_k_9_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];818 tensor<int32, [2]> concat_84 = const()[name = string("concat_84"), val = tensor<int32, [2]>([128, 1024])];819 tensor<fp16, [128, 1, 8, 128]> transpose_60_cast_fp16 = transpose(perm = transpose_60_perm_0, x = var_918_cast_fp16)[name = string("transpose_66")];820 tensor<fp16, [128, 1024]> reshape_28_cast_fp16 = reshape(shape = concat_84, x = transpose_60_cast_fp16)[name = string("reshape_28_cast_fp16")];821 bool matmul_9_transpose_x_1 = const()[name = string("matmul_9_transpose_x_1"), val = bool(true)];822 bool matmul_9_transpose_y_1 = const()[name = string("matmul_9_transpose_y_1"), val = bool(false)];823 tensor<fp16, [1024, 1024]> matmul_9_cast_fp16 = matmul(transpose_x = matmul_9_transpose_x_1, transpose_y = matmul_9_transpose_y_1, x = var_68_to_fp16, y = reshape_28_cast_fp16)[name = string("matmul_9_cast_fp16")];824 tensor<int32, [4]> concat_87 = const()[name = string("concat_87"), val = tensor<int32, [4]>([1024, 1, 8, 128])];825 tensor<fp16, [1024, 1, 8, 128]> reshape_29_cast_fp16 = reshape(shape = concat_87, x = matmul_9_cast_fp16)[name = string("reshape_29_cast_fp16")];826 tensor<int32, [4]> scattered_v_9_perm_0 = const()[name = string("scattered_v_9_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];827 tensor<fp16, [1, 8, 1024, 128]> read_state_8 = read_state(input = k_cache_4)[name = string("read_state_8")];828 tensor<fp16, [1, 8, 1024, 128]> k_cache_27_cast_fp16 = mul(x = read_state_8, y = var_224_cast_fp16)[name = string("k_cache_27_cast_fp16")];829 write_state(data = k_cache_27_cast_fp16, input = k_cache_4)[name = string("coreml_update_state_72_write_state")];830 tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_72 = read_state(input = k_cache_4)[name = string("coreml_update_state_72")];831 tensor<fp16, [1, 8, 1024, 128]> scattered_k_9_cast_fp16 = transpose(perm = scattered_k_9_perm_0, x = reshape_26_cast_fp16)[name = string("transpose_65")];832 tensor<fp16, [1, 8, 1024, 128]> k_cache_29_cast_fp16 = add(x = coreml_update_state_72, y = scattered_k_9_cast_fp16)[name = string("k_cache_29_cast_fp16")];833 write_state(data = k_cache_29_cast_fp16, input = k_cache_4)[name = string("coreml_update_state_73_write_state")];834 tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_73 = read_state(input = k_cache_4)[name = string("coreml_update_state_73")];835 tensor<fp16, [1, 8, 1024, 128]> read_state_9 = read_state(input = v_cache_4)[name = string("read_state_9")];836 tensor<fp16, [1, 8, 1024, 128]> v_cache_27_cast_fp16 = mul(x = read_state_9, y = var_224_cast_fp16)[name = string("v_cache_27_cast_fp16")];837 write_state(data = v_cache_27_cast_fp16, input = v_cache_4)[name = string("coreml_update_state_74_write_state")];838 tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_74 = read_state(input = v_cache_4)[name = string("coreml_update_state_74")];839 tensor<fp16, [1, 8, 1024, 128]> scattered_v_9_cast_fp16 = transpose(perm = scattered_v_9_perm_0, x = reshape_29_cast_fp16)[name = string("transpose_64")];840 tensor<fp16, [1, 8, 1024, 128]> v_cache_29_cast_fp16 = add(x = coreml_update_state_74, y = scattered_v_9_cast_fp16)[name = string("v_cache_29_cast_fp16")];841 write_state(data = v_cache_29_cast_fp16, input = v_cache_4)[name = string("coreml_update_state_75_write_state")];842 tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_75 = read_state(input = v_cache_4)[name = string("coreml_update_state_75")];843 tensor<int32, [1]> var_1006_axes_0 = const()[name = string("op_1006_axes_0"), val = tensor<int32, [1]>([2])];844 tensor<fp16, [1, 8, 1, 1024, 128]> var_1006_cast_fp16 = expand_dims(axes = var_1006_axes_0, x = coreml_update_state_73)[name = string("op_1006_cast_fp16")];845 tensor<int32, [5]> k_exp_17_reps_0 = const()[name = string("k_exp_17_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];846 tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_17_cast_fp16 = tile(reps = k_exp_17_reps_0, x = var_1006_cast_fp16)[name = string("k_exp_17_cast_fp16")];847 tensor<int32, [4]> var_1009 = const()[name = string("op_1009"), val = tensor<int32, [4]>([1, 16, 1024, 128])];848 tensor<fp16, [1, 16, 1024, 128]> k_exp_19_cast_fp16 = reshape(shape = var_1009, x = k_exp_17_cast_fp16)[name = string("k_exp_19_cast_fp16")];849 tensor<int32, [1]> var_1011_axes_0 = const()[name = string("op_1011_axes_0"), val = tensor<int32, [1]>([2])];850 tensor<fp16, [1, 8, 1, 1024, 128]> var_1011_cast_fp16 = expand_dims(axes = var_1011_axes_0, x = coreml_update_state_75)[name = string("op_1011_cast_fp16")];851 tensor<int32, [5]> v_exp_17_reps_0 = const()[name = string("v_exp_17_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];852 tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_17_cast_fp16 = tile(reps = v_exp_17_reps_0, x = var_1011_cast_fp16)[name = string("v_exp_17_cast_fp16")];853 tensor<int32, [4]> var_1014 = const()[name = string("op_1014"), val = tensor<int32, [4]>([1, 16, 1024, 128])];854 tensor<fp16, [1, 16, 1024, 128]> v_exp_19_cast_fp16 = reshape(shape = var_1014, x = v_exp_17_cast_fp16)[name = string("v_exp_19_cast_fp16")];855 bool var_1017_transpose_x_1 = const()[name = string("op_1017_transpose_x_1"), val = bool(false)];856 bool var_1017_transpose_y_1 = const()[name = string("op_1017_transpose_y_1"), val = bool(true)];857 tensor<fp16, [1, 16, 128, 1024]> var_1017_cast_fp16 = matmul(transpose_x = var_1017_transpose_x_1, transpose_y = var_1017_transpose_y_1, x = q_9_cast_fp16, y = k_exp_19_cast_fp16)[name = string("op_1017_cast_fp16")];858 fp16 var_1018_to_fp16 = const()[name = string("op_1018_to_fp16"), val = fp16(0x1.6ap-4)];859 tensor<fp16, [1, 16, 128, 1024]> attn_17_cast_fp16 = mul(x = var_1017_cast_fp16, y = var_1018_to_fp16)[name = string("attn_17_cast_fp16")];860 tensor<fp16, [1, 16, 128, 1024]> input_41_cast_fp16 = add(x = attn_17_cast_fp16, y = attention_mask_to_fp16)[name = string("input_41_cast_fp16")];861 tensor<fp16, [1, 16, 128, 1024]> attn_19_cast_fp16 = softmax(axis = var_875, x = input_41_cast_fp16)[name = string("attn_19_cast_fp16")];862 bool out_9_transpose_x_0 = const()[name = string("out_9_transpose_x_0"), val = bool(false)];863 bool out_9_transpose_y_0 = const()[name = string("out_9_transpose_y_0"), val = bool(false)];864 tensor<fp16, [1, 16, 128, 128]> out_9_cast_fp16 = matmul(transpose_x = out_9_transpose_x_0, transpose_y = out_9_transpose_y_0, x = attn_19_cast_fp16, y = v_exp_19_cast_fp16)[name = string("out_9_cast_fp16")];865 tensor<int32, [4]> var_1023_perm_0 = const()[name = string("op_1023_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];866 tensor<int32, [3]> var_1024 = const()[name = string("op_1024"), val = tensor<int32, [3]>([1, 128, -1])];867 tensor<fp16, [1, 128, 16, 128]> var_1023_cast_fp16 = transpose(perm = var_1023_perm_0, x = out_9_cast_fp16)[name = string("transpose_63")];868 tensor<fp16, [1, 128, 2048]> input_43_cast_fp16 = reshape(shape = var_1024, x = var_1023_cast_fp16)[name = string("input_43_cast_fp16")];869 tensor<fp16, [1024, 2048]> layers_4_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67163648))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69260864))))[name = string("layers_4_self_attn_o_proj_weight_to_fp16_palettized")];870 tensor<fp16, [1, 128, 1024]> linear_31_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_4_self_attn_o_proj_weight_to_fp16_palettized, x = input_43_cast_fp16)[name = string("linear_31_cast_fp16")];871 tensor<fp16, [1, 128, 1024]> x_123_cast_fp16 = add(x = x_103_cast_fp16, y = linear_31_cast_fp16)[name = string("x_123_cast_fp16")];872 fp16 var_874_promoted_3_to_fp16 = const()[name = string("op_874_promoted_3_to_fp16"), val = fp16(0x1p+1)];873 tensor<fp16, [1, 128, 1024]> var_1031_cast_fp16 = pow(x = x_123_cast_fp16, y = var_874_promoted_3_to_fp16)[name = string("op_1031_cast_fp16")];874 tensor<int32, [1]> var_1033_axes_0 = const()[name = string("op_1033_axes_0"), val = tensor<int32, [1]>([-1])];875 bool var_1033_keep_dims_0 = const()[name = string("op_1033_keep_dims_0"), val = bool(true)];876 tensor<fp16, [1, 128, 1]> var_1033_cast_fp16 = reduce_mean(axes = var_1033_axes_0, keep_dims = var_1033_keep_dims_0, x = var_1031_cast_fp16)[name = string("op_1033_cast_fp16")];877 fp16 var_1034_to_fp16 = const()[name = string("op_1034_to_fp16"), val = fp16(0x1.1p-20)];878 tensor<fp16, [1, 128, 1]> var_1035_cast_fp16 = add(x = var_1033_cast_fp16, y = var_1034_to_fp16)[name = string("op_1035_cast_fp16")];879 fp32 norm_39_epsilon_0 = const()[name = string("norm_39_epsilon_0"), val = fp32(0x1.197998p-40)];880 tensor<fp16, [1, 128, 1]> norm_39_cast_fp16 = rsqrt(epsilon = norm_39_epsilon_0, x = var_1035_cast_fp16)[name = string("norm_39_cast_fp16")];881 tensor<fp16, [1, 128, 1024]> var_1037_cast_fp16 = mul(x = x_123_cast_fp16, y = norm_39_cast_fp16)[name = string("op_1037_cast_fp16")];882 tensor<fp16, [1024]> layers_4_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_4_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69261440)))];883 tensor<fp16, [1, 128, 1024]> var_1038_cast_fp16 = mul(x = var_1037_cast_fp16, y = layers_4_post_attention_layernorm_weight_to_fp16)[name = string("op_1038_cast_fp16")];884 tensor<fp16, [3072, 1024]> layers_4_mlp_gate_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69263552))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72409344))))[name = string("layers_4_mlp_gate_proj_weight_to_fp16_palettized")];885 tensor<fp16, [1, 128, 3072]> linear_32_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_4_mlp_gate_proj_weight_to_fp16_palettized, x = var_1038_cast_fp16)[name = string("linear_32_cast_fp16")];886 tensor<fp16, [1, 128, 3072]> var_1048_cast_fp16 = silu(x = linear_32_cast_fp16)[name = string("op_1048_cast_fp16")];887 tensor<fp16, [3072, 1024]> layers_4_mlp_up_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72409920))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75555712))))[name = string("layers_4_mlp_up_proj_weight_to_fp16_palettized")];888 tensor<fp16, [1, 128, 3072]> linear_33_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_4_mlp_up_proj_weight_to_fp16_palettized, x = var_1038_cast_fp16)[name = string("linear_33_cast_fp16")];889 tensor<fp16, [1, 128, 3072]> input_49_cast_fp16 = mul(x = var_1048_cast_fp16, y = linear_33_cast_fp16)[name = string("input_49_cast_fp16")];890 tensor<fp16, [1024, 3072]> layers_4_mlp_down_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75556288))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78702080))))[name = string("layers_4_mlp_down_proj_weight_to_fp16_palettized")];891 tensor<fp16, [1, 128, 1024]> linear_34_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_4_mlp_down_proj_weight_to_fp16_palettized, x = input_49_cast_fp16)[name = string("linear_34_cast_fp16")];892 tensor<fp16, [1, 128, 1024]> x_129_cast_fp16 = add(x = x_123_cast_fp16, y = linear_34_cast_fp16)[name = string("x_129_cast_fp16")];893 int32 var_1069 = const()[name = string("op_1069"), val = int32(-1)];894 fp16 var_1068_promoted_to_fp16 = const()[name = string("op_1068_promoted_to_fp16"), val = fp16(0x1p+1)];895 tensor<fp16, [1, 128, 1024]> var_1078_cast_fp16 = pow(x = x_129_cast_fp16, y = var_1068_promoted_to_fp16)[name = string("op_1078_cast_fp16")];896 tensor<int32, [1]> var_1080_axes_0 = const()[name = string("op_1080_axes_0"), val = tensor<int32, [1]>([-1])];897 bool var_1080_keep_dims_0 = const()[name = string("op_1080_keep_dims_0"), val = bool(true)];898 tensor<fp16, [1, 128, 1]> var_1080_cast_fp16 = reduce_mean(axes = var_1080_axes_0, keep_dims = var_1080_keep_dims_0, x = var_1078_cast_fp16)[name = string("op_1080_cast_fp16")];899 fp16 var_1081_to_fp16 = const()[name = string("op_1081_to_fp16"), val = fp16(0x1.1p-20)];900 tensor<fp16, [1, 128, 1]> var_1082_cast_fp16 = add(x = var_1080_cast_fp16, y = var_1081_to_fp16)[name = string("op_1082_cast_fp16")];901 fp32 norm_41_epsilon_0 = const()[name = string("norm_41_epsilon_0"), val = fp32(0x1.197998p-40)];902 tensor<fp16, [1, 128, 1]> norm_41_cast_fp16 = rsqrt(epsilon = norm_41_epsilon_0, x = var_1082_cast_fp16)[name = string("norm_41_cast_fp16")];903 tensor<fp16, [1, 128, 1024]> var_1084_cast_fp16 = mul(x = x_129_cast_fp16, y = norm_41_cast_fp16)[name = string("op_1084_cast_fp16")];904 tensor<fp16, [1024]> layers_5_input_layernorm_weight_to_fp16 = const()[name = string("layers_5_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78702656)))];905 tensor<fp16, [1, 128, 1024]> var_1085_cast_fp16 = mul(x = var_1084_cast_fp16, y = layers_5_input_layernorm_weight_to_fp16)[name = string("op_1085_cast_fp16")];906 tensor<fp16, [2048, 1024]> layers_5_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78704768))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80801984))))[name = string("layers_5_self_attn_q_proj_weight_to_fp16_palettized")];907 tensor<fp16, [1, 128, 2048]> linear_35_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_5_self_attn_q_proj_weight_to_fp16_palettized, x = var_1085_cast_fp16)[name = string("linear_35_cast_fp16")];908 tensor<int32, [4]> var_1101 = const()[name = string("op_1101"), val = tensor<int32, [4]>([1, 128, 16, 128])];909 tensor<fp16, [1, 128, 16, 128]> var_1102_cast_fp16 = reshape(shape = var_1101, x = linear_35_cast_fp16)[name = string("op_1102_cast_fp16")];910 tensor<int32, [4]> x_135_perm_0 = const()[name = string("x_135_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];911 tensor<fp16, [1024, 1024]> layers_5_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80802560))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81851200))))[name = string("layers_5_self_attn_k_proj_weight_to_fp16_palettized")];912 tensor<fp16, [1, 128, 1024]> linear_36_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_5_self_attn_k_proj_weight_to_fp16_palettized, x = var_1085_cast_fp16)[name = string("linear_36_cast_fp16")];913 tensor<int32, [4]> var_1106 = const()[name = string("op_1106"), val = tensor<int32, [4]>([1, 128, 8, 128])];914 tensor<fp16, [1, 128, 8, 128]> var_1107_cast_fp16 = reshape(shape = var_1106, x = linear_36_cast_fp16)[name = string("op_1107_cast_fp16")];915 tensor<int32, [4]> x_139_perm_0 = const()[name = string("x_139_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];916 tensor<fp16, [1024, 1024]> layers_5_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81851776))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82900416))))[name = string("layers_5_self_attn_v_proj_weight_to_fp16_palettized")];917 tensor<fp16, [1, 128, 1024]> linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_5_self_attn_v_proj_weight_to_fp16_palettized, x = var_1085_cast_fp16)[name = string("linear_37_cast_fp16")];918 tensor<int32, [4]> var_1111 = const()[name = string("op_1111"), val = tensor<int32, [4]>([1, 128, 8, 128])];919 tensor<fp16, [1, 128, 8, 128]> var_1112_cast_fp16 = reshape(shape = var_1111, x = linear_37_cast_fp16)[name = string("op_1112_cast_fp16")];920 tensor<int32, [4]> transpose_61_perm_0 = const()[name = string("transpose_61_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];921 fp16 var_1068_promoted_1_to_fp16 = const()[name = string("op_1068_promoted_1_to_fp16"), val = fp16(0x1p+1)];922 tensor<fp16, [1, 16, 128, 128]> x_135_cast_fp16 = transpose(perm = x_135_perm_0, x = var_1102_cast_fp16)[name = string("transpose_62")];923 tensor<fp16, [1, 16, 128, 128]> var_1116_cast_fp16 = pow(x = x_135_cast_fp16, y = var_1068_promoted_1_to_fp16)[name = string("op_1116_cast_fp16")];924 tensor<int32, [1]> var_1118_axes_0 = const()[name = string("op_1118_axes_0"), val = tensor<int32, [1]>([-1])];925 bool var_1118_keep_dims_0 = const()[name = string("op_1118_keep_dims_0"), val = bool(true)];926 tensor<fp16, [1, 16, 128, 1]> var_1118_cast_fp16 = reduce_mean(axes = var_1118_axes_0, keep_dims = var_1118_keep_dims_0, x = var_1116_cast_fp16)[name = string("op_1118_cast_fp16")];927 fp16 var_1119_to_fp16 = const()[name = string("op_1119_to_fp16"), val = fp16(0x1.1p-20)];928 tensor<fp16, [1, 16, 128, 1]> var_1120_cast_fp16 = add(x = var_1118_cast_fp16, y = var_1119_to_fp16)[name = string("op_1120_cast_fp16")];929 fp32 norm_43_epsilon_0 = const()[name = string("norm_43_epsilon_0"), val = fp32(0x1.197998p-40)];930 tensor<fp16, [1, 16, 128, 1]> norm_43_cast_fp16 = rsqrt(epsilon = norm_43_epsilon_0, x = var_1120_cast_fp16)[name = string("norm_43_cast_fp16")];931 tensor<fp16, [1, 16, 128, 128]> var_1122_cast_fp16 = mul(x = x_135_cast_fp16, y = norm_43_cast_fp16)[name = string("op_1122_cast_fp16")];932 tensor<fp16, [128]> layers_5_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_5_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82900992)))];933 tensor<fp16, [1, 16, 128, 128]> var_1123_cast_fp16 = mul(x = var_1122_cast_fp16, y = layers_5_self_attn_q_norm_weight_to_fp16)[name = string("op_1123_cast_fp16")];934 fp16 var_1068_promoted_2_to_fp16 = const()[name = string("op_1068_promoted_2_to_fp16"), val = fp16(0x1p+1)];935 tensor<fp16, [1, 8, 128, 128]> x_139_cast_fp16 = transpose(perm = x_139_perm_0, x = var_1107_cast_fp16)[name = string("transpose_61")];936 tensor<fp16, [1, 8, 128, 128]> var_1127_cast_fp16 = pow(x = x_139_cast_fp16, y = var_1068_promoted_2_to_fp16)[name = string("op_1127_cast_fp16")];937 tensor<int32, [1]> var_1129_axes_0 = const()[name = string("op_1129_axes_0"), val = tensor<int32, [1]>([-1])];938 bool var_1129_keep_dims_0 = const()[name = string("op_1129_keep_dims_0"), val = bool(true)];939 tensor<fp16, [1, 8, 128, 1]> var_1129_cast_fp16 = reduce_mean(axes = var_1129_axes_0, keep_dims = var_1129_keep_dims_0, x = var_1127_cast_fp16)[name = string("op_1129_cast_fp16")];940 fp16 var_1130_to_fp16 = const()[name = string("op_1130_to_fp16"), val = fp16(0x1.1p-20)];941 tensor<fp16, [1, 8, 128, 1]> var_1131_cast_fp16 = add(x = var_1129_cast_fp16, y = var_1130_to_fp16)[name = string("op_1131_cast_fp16")];942 fp32 norm_45_epsilon_0 = const()[name = string("norm_45_epsilon_0"), val = fp32(0x1.197998p-40)];943 tensor<fp16, [1, 8, 128, 1]> norm_45_cast_fp16 = rsqrt(epsilon = norm_45_epsilon_0, x = var_1131_cast_fp16)[name = string("norm_45_cast_fp16")];944 tensor<fp16, [1, 8, 128, 128]> var_1133_cast_fp16 = mul(x = x_139_cast_fp16, y = norm_45_cast_fp16)[name = string("op_1133_cast_fp16")];945 tensor<fp16, [128]> layers_5_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_5_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82901312)))];946 tensor<fp16, [1, 8, 128, 128]> var_1134_cast_fp16 = mul(x = var_1133_cast_fp16, y = layers_5_self_attn_k_norm_weight_to_fp16)[name = string("op_1134_cast_fp16")];947 tensor<int32, [4]> x1_21_begin_0 = const()[name = string("x1_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];948 tensor<int32, [4]> x1_21_end_0 = const()[name = string("x1_21_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];949 tensor<bool, [4]> x1_21_end_mask_0 = const()[name = string("x1_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];950 tensor<fp16, [1, 16, 128, 64]> x1_21_cast_fp16 = slice_by_index(begin = x1_21_begin_0, end = x1_21_end_0, end_mask = x1_21_end_mask_0, x = var_1123_cast_fp16)[name = string("x1_21_cast_fp16")];951 tensor<int32, [4]> x2_21_begin_0 = const()[name = string("x2_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];952 tensor<int32, [4]> x2_21_end_0 = const()[name = string("x2_21_end_0"), val = tensor<int32, [4]>([1, 16, 128, 128])];953 tensor<bool, [4]> x2_21_end_mask_0 = const()[name = string("x2_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];954 tensor<fp16, [1, 16, 128, 64]> x2_21_cast_fp16 = slice_by_index(begin = x2_21_begin_0, end = x2_21_end_0, end_mask = x2_21_end_mask_0, x = var_1123_cast_fp16)[name = string("x2_21_cast_fp16")];955 tensor<fp16, [1, 16, 128, 64]> var_1155_cast_fp16 = mul(x = x1_21_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1155_cast_fp16")];956 tensor<fp16, [1, 16, 128, 64]> var_1156_cast_fp16 = mul(x = x2_21_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1156_cast_fp16")];957 tensor<fp16, [1, 16, 128, 64]> var_1157_cast_fp16 = sub(x = var_1155_cast_fp16, y = var_1156_cast_fp16)[name = string("op_1157_cast_fp16")];958 tensor<fp16, [1, 16, 128, 64]> var_1158_cast_fp16 = mul(x = x2_21_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1158_cast_fp16")];959 tensor<fp16, [1, 16, 128, 64]> var_1159_cast_fp16 = mul(x = x1_21_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1159_cast_fp16")];960 tensor<fp16, [1, 16, 128, 64]> var_1160_cast_fp16 = add(x = var_1158_cast_fp16, y = var_1159_cast_fp16)[name = string("op_1160_cast_fp16")];961 bool q_11_interleave_0 = const()[name = string("q_11_interleave_0"), val = bool(false)];962 tensor<fp16, [1, 16, 128, 128]> q_11_cast_fp16 = concat(axis = var_1069, interleave = q_11_interleave_0, values = (var_1157_cast_fp16, var_1160_cast_fp16))[name = string("q_11_cast_fp16")];963 tensor<int32, [4]> x1_23_begin_0 = const()[name = string("x1_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];964 tensor<int32, [4]> x1_23_end_0 = const()[name = string("x1_23_end_0"), val = tensor<int32, [4]>([1, 8, 128, 64])];965 tensor<bool, [4]> x1_23_end_mask_0 = const()[name = string("x1_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];966 tensor<fp16, [1, 8, 128, 64]> x1_23_cast_fp16 = slice_by_index(begin = x1_23_begin_0, end = x1_23_end_0, end_mask = x1_23_end_mask_0, x = var_1134_cast_fp16)[name = string("x1_23_cast_fp16")];967 tensor<int32, [4]> x2_23_begin_0 = const()[name = string("x2_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];968 tensor<int32, [4]> x2_23_end_0 = const()[name = string("x2_23_end_0"), val = tensor<int32, [4]>([1, 8, 128, 128])];969 tensor<bool, [4]> x2_23_end_mask_0 = const()[name = string("x2_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];970 tensor<fp16, [1, 8, 128, 64]> x2_23_cast_fp16 = slice_by_index(begin = x2_23_begin_0, end = x2_23_end_0, end_mask = x2_23_end_mask_0, x = var_1134_cast_fp16)[name = string("x2_23_cast_fp16")];971 tensor<fp16, [1, 8, 128, 64]> var_1182_cast_fp16 = mul(x = x1_23_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1182_cast_fp16")];972 tensor<fp16, [1, 8, 128, 64]> var_1183_cast_fp16 = mul(x = x2_23_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1183_cast_fp16")];973 tensor<fp16, [1, 8, 128, 64]> var_1184_cast_fp16 = sub(x = var_1182_cast_fp16, y = var_1183_cast_fp16)[name = string("op_1184_cast_fp16")];974 tensor<fp16, [1, 8, 128, 64]> var_1185_cast_fp16 = mul(x = x2_23_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1185_cast_fp16")];975 tensor<fp16, [1, 8, 128, 64]> var_1186_cast_fp16 = mul(x = x1_23_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1186_cast_fp16")];976 tensor<fp16, [1, 8, 128, 64]> var_1187_cast_fp16 = add(x = var_1185_cast_fp16, y = var_1186_cast_fp16)[name = string("op_1187_cast_fp16")];977 bool var_1189_interleave_0 = const()[name = string("op_1189_interleave_0"), val = bool(false)];978 tensor<fp16, [1, 8, 128, 128]> var_1189_cast_fp16 = concat(axis = var_1069, interleave = var_1189_interleave_0, values = (var_1184_cast_fp16, var_1187_cast_fp16))[name = string("op_1189_cast_fp16")];979 tensor<int32, [4]> transpose_21_perm_0 = const()[name = string("transpose_21_perm_0"), val = tensor<int32, [4]>([2, 0, 1, 3])];980 tensor<int32, [2]> concat_94 = const()[name = string("concat_94"), val = tensor<int32, [2]>([128, 1024])];981 tensor<fp16, [128, 1, 8, 128]> transpose_21_cast_fp16 = transpose(perm = transpose_21_perm_0, x = var_1189_cast_fp16)[name = string("transpose_60")];982 tensor<fp16, [128, 1024]> reshape_31_cast_fp16 = reshape(shape = concat_94, x = transpose_21_cast_fp16)[name = string("reshape_31_cast_fp16")];983 bool matmul_10_transpose_x_1 = const()[name = string("matmul_10_transpose_x_1"), val = bool(true)];984 bool matmul_10_transpose_y_1 = const()[name = string("matmul_10_transpose_y_1"), val = bool(false)];985 tensor<fp16, [1024, 1024]> matmul_10_cast_fp16 = matmul(transpose_x = matmul_10_transpose_x_1, transpose_y = matmul_10_transpose_y_1, x = var_68_to_fp16, y = reshape_31_cast_fp16)[name = string("matmul_10_cast_fp16")];986 tensor<int32, [4]> concat_97 = const()[name = string("concat_97"), val = tensor<int32, [4]>([1024, 1, 8, 128])];987 tensor<fp16, [1024, 1, 8, 128]> reshape_32_cast_fp16 = reshape(shape = concat_97, x = matmul_10_cast_fp16)[name = string("reshape_32_cast_fp16")];988 tensor<int32, [4]> scattered_k_11_perm_0 = const()[name = string("scattered_k_11_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];989 tensor<int32, [2]> concat_102 = const()[name = string("concat_102"), val = tensor<int32, [2]>([128, 1024])];990 tensor<fp16, [128, 1, 8, 128]> transpose_61_cast_fp16 = transpose(perm = transpose_61_perm_0, x = var_1112_cast_fp16)[name = string("transpose_59")];991 tensor<fp16, [128, 1024]> reshape_34_cast_fp16 = reshape(shape = concat_102, x = transpose_61_cast_fp16)[name = string("reshape_34_cast_fp16")];992 bool matmul_11_transpose_x_1 = const()[name = string("matmul_11_transpose_x_1"), val = bool(true)];993 bool matmul_11_transpose_y_1 = const()[name = string("matmul_11_transpose_y_1"), val = bool(false)];994 tensor<fp16, [1024, 1024]> matmul_11_cast_fp16 = matmul(transpose_x = matmul_11_transpose_x_1, transpose_y = matmul_11_transpose_y_1, x = var_68_to_fp16, y = reshape_34_cast_fp16)[name = string("matmul_11_cast_fp16")];995 tensor<int32, [4]> concat_105 = const()[name = string("concat_105"), val = tensor<int32, [4]>([1024, 1, 8, 128])];996 tensor<fp16, [1024, 1, 8, 128]> reshape_35_cast_fp16 = reshape(shape = concat_105, x = matmul_11_cast_fp16)[name = string("reshape_35_cast_fp16")];997 tensor<int32, [4]> scattered_v_11_perm_0 = const()[name = string("scattered_v_11_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];998 tensor<fp16, [1, 8, 1024, 128]> read_state_10 = read_state(input = k_cache_5)[name = string("read_state_10")];999 tensor<fp16, [1, 8, 1024, 128]> k_cache_33_cast_fp16 = mul(x = read_state_10, y = var_224_cast_fp16)[name = string("k_cache_33_cast_fp16")];1000 write_state(data = k_cache_33_cast_fp16, input = k_cache_5)[name = string("coreml_update_state_76_write_state")];1001 tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_76 = read_state(input = k_cache_5)[name = string("coreml_update_state_76")];1002 tensor<fp16, [1, 8, 1024, 128]> scattered_k_11_cast_fp16 = transpose(perm = scattered_k_11_perm_0, x = reshape_32_cast_fp16)[name = string("transpose_58")];1003 tensor<fp16, [1, 8, 1024, 128]> k_cache_35_cast_fp16 = add(x = coreml_update_state_76, y = scattered_k_11_cast_fp16)[name = string("k_cache_35_cast_fp16")];1004 write_state(data = k_cache_35_cast_fp16, input = k_cache_5)[name = string("coreml_update_state_77_write_state")];1005 tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_77 = read_state(input = k_cache_5)[name = string("coreml_update_state_77")];1006 tensor<fp16, [1, 8, 1024, 128]> read_state_11 = read_state(input = v_cache_5)[name = string("read_state_11")];1007 tensor<fp16, [1, 8, 1024, 128]> v_cache_33_cast_fp16 = mul(x = read_state_11, y = var_224_cast_fp16)[name = string("v_cache_33_cast_fp16")];1008 write_state(data = v_cache_33_cast_fp16, input = v_cache_5)[name = string("coreml_update_state_78_write_state")];1009 tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_78 = read_state(input = v_cache_5)[name = string("coreml_update_state_78")];1010 tensor<fp16, [1, 8, 1024, 128]> scattered_v_11_cast_fp16 = transpose(perm = scattered_v_11_perm_0, x = reshape_35_cast_fp16)[name = string("transpose_57")];1011 tensor<fp16, [1, 8, 1024, 128]> v_cache_35_cast_fp16 = add(x = coreml_update_state_78, y = scattered_v_11_cast_fp16)[name = string("v_cache_35_cast_fp16")];1012 write_state(data = v_cache_35_cast_fp16, input = v_cache_5)[name = string("coreml_update_state_79_write_state")];1013 tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_79 = read_state(input = v_cache_5)[name = string("coreml_update_state_79")];1014 tensor<int32, [1]> var_1200_axes_0 = const()[name = string("op_1200_axes_0"), val = tensor<int32, [1]>([2])];1015 tensor<fp16, [1, 8, 1, 1024, 128]> var_1200_cast_fp16 = expand_dims(axes = var_1200_axes_0, x = coreml_update_state_77)[name = string("op_1200_cast_fp16")];1016 tensor<int32, [5]> k_exp_21_reps_0 = const()[name = string("k_exp_21_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];1017 tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_21_cast_fp16 = tile(reps = k_exp_21_reps_0, x = var_1200_cast_fp16)[name = string("k_exp_21_cast_fp16")];1018 tensor<int32, [4]> var_1203 = const()[name = string("op_1203"), val = tensor<int32, [4]>([1, 16, 1024, 128])];1019 tensor<fp16, [1, 16, 1024, 128]> k_exp_23_cast_fp16 = reshape(shape = var_1203, x = k_exp_21_cast_fp16)[name = string("k_exp_23_cast_fp16")];1020 tensor<int32, [1]> var_1205_axes_0 = const()[name = string("op_1205_axes_0"), val = tensor<int32, [1]>([2])];1021 tensor<fp16, [1, 8, 1, 1024, 128]> var_1205_cast_fp16 = expand_dims(axes = var_1205_axes_0, x = coreml_update_state_79)[name = string("op_1205_cast_fp16")];1022 tensor<int32, [5]> v_exp_21_reps_0 = const()[name = string("v_exp_21_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];1023 tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_21_cast_fp16 = tile(reps = v_exp_21_reps_0, x = var_1205_cast_fp16)[name = string("v_exp_21_cast_fp16")];1024 tensor<int32, [4]> var_1208 = const()[name = string("op_1208"), val = tensor<int32, [4]>([1, 16, 1024, 128])];1025 tensor<fp16, [1, 16, 1024, 128]> v_exp_23_cast_fp16 = reshape(shape = var_1208, x = v_exp_21_cast_fp16)[name = string("v_exp_23_cast_fp16")];1026 bool var_1211_transpose_x_1 = const()[name = string("op_1211_transpose_x_1"), val = bool(false)];1027 bool var_1211_transpose_y_1 = const()[name = string("op_1211_transpose_y_1"), val = bool(true)];1028 tensor<fp16, [1, 16, 128, 1024]> var_1211_cast_fp16 = matmul(transpose_x = var_1211_transpose_x_1, transpose_y = var_1211_transpose_y_1, x = q_11_cast_fp16, y = k_exp_23_cast_fp16)[name = string("op_1211_cast_fp16")];1029 fp16 var_1212_to_fp16 = const()[name = string("op_1212_to_fp16"), val = fp16(0x1.6ap-4)];1030 tensor<fp16, [1, 16, 128, 1024]> attn_21_cast_fp16 = mul(x = var_1211_cast_fp16, y = var_1212_to_fp16)[name = string("attn_21_cast_fp16")];1031 tensor<fp16, [1, 16, 128, 1024]> input_51_cast_fp16 = add(x = attn_21_cast_fp16, y = attention_mask_to_fp16)[name = string("input_51_cast_fp16")];1032 tensor<fp16, [1, 16, 128, 1024]> attn_23_cast_fp16 = softmax(axis = var_1069, x = input_51_cast_fp16)[name = string("attn_23_cast_fp16")];1033 bool out_11_transpose_x_0 = const()[name = string("out_11_transpose_x_0"), val = bool(false)];1034 bool out_11_transpose_y_0 = const()[name = string("out_11_transpose_y_0"), val = bool(false)];1035 tensor<fp16, [1, 16, 128, 128]> out_11_cast_fp16 = matmul(transpose_x = out_11_transpose_x_0, transpose_y = out_11_transpose_y_0, x = attn_23_cast_fp16, y = v_exp_23_cast_fp16)[name = string("out_11_cast_fp16")];1036 tensor<int32, [4]> var_1217_perm_0 = const()[name = string("op_1217_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];1037 tensor<int32, [3]> var_1218 = const()[name = string("op_1218"), val = tensor<int32, [3]>([1, 128, -1])];1038 tensor<fp16, [1, 128, 16, 128]> var_1217_cast_fp16 = transpose(perm = var_1217_perm_0, x = out_11_cast_fp16)[name = string("transpose_56")];1039 tensor<fp16, [1, 128, 2048]> input_53_cast_fp16 = reshape(shape = var_1218, x = var_1217_cast_fp16)[name = string("input_53_cast_fp16")];1040 tensor<fp16, [1024, 2048]> layers_5_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82901632))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84998848))))[name = string("layers_5_self_attn_o_proj_weight_to_fp16_palettized")];1041 tensor<fp16, [1, 128, 1024]> linear_38_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_5_self_attn_o_proj_weight_to_fp16_palettized, x = input_53_cast_fp16)[name = string("linear_38_cast_fp16")];1042 tensor<fp16, [1, 128, 1024]> x_149_cast_fp16 = add(x = x_129_cast_fp16, y = linear_38_cast_fp16)[name = string("x_149_cast_fp16")];1043 fp16 var_1068_promoted_3_to_fp16 = const()[name = string("op_1068_promoted_3_to_fp16"), val = fp16(0x1p+1)];1044 tensor<fp16, [1, 128, 1024]> var_1225_cast_fp16 = pow(x = x_149_cast_fp16, y = var_1068_promoted_3_to_fp16)[name = string("op_1225_cast_fp16")];1045 tensor<int32, [1]> var_1227_axes_0 = const()[name = string("op_1227_axes_0"), val = tensor<int32, [1]>([-1])];1046 bool var_1227_keep_dims_0 = const()[name = string("op_1227_keep_dims_0"), val = bool(true)];1047 tensor<fp16, [1, 128, 1]> var_1227_cast_fp16 = reduce_mean(axes = var_1227_axes_0, keep_dims = var_1227_keep_dims_0, x = var_1225_cast_fp16)[name = string("op_1227_cast_fp16")];1048 fp16 var_1228_to_fp16 = const()[name = string("op_1228_to_fp16"), val = fp16(0x1.1p-20)];1049 tensor<fp16, [1, 128, 1]> var_1229_cast_fp16 = add(x = var_1227_cast_fp16, y = var_1228_to_fp16)[name = string("op_1229_cast_fp16")];1050 fp32 norm_47_epsilon_0 = const()[name = string("norm_47_epsilon_0"), val = fp32(0x1.197998p-40)];1051 tensor<fp16, [1, 128, 1]> norm_47_cast_fp16 = rsqrt(epsilon = norm_47_epsilon_0, x = var_1229_cast_fp16)[name = string("norm_47_cast_fp16")];1052 tensor<fp16, [1, 128, 1024]> var_1231_cast_fp16 = mul(x = x_149_cast_fp16, y = norm_47_cast_fp16)[name = string("op_1231_cast_fp16")];1053 tensor<fp16, [1024]> layers_5_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_5_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84999424)))];1054 tensor<fp16, [1, 128, 1024]> var_1232_cast_fp16 = mul(x = var_1231_cast_fp16, y = layers_5_post_attention_layernorm_weight_to_fp16)[name = string("op_1232_cast_fp16")];1055 tensor<fp16, [3072, 1024]> layers_5_mlp_gate_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85001536))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88147328))))[name = string("layers_5_mlp_gate_proj_weight_to_fp16_palettized")];1056 tensor<fp16, [1, 128, 3072]> linear_39_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_5_mlp_gate_proj_weight_to_fp16_palettized, x = var_1232_cast_fp16)[name = string("linear_39_cast_fp16")];1057 tensor<fp16, [1, 128, 3072]> var_1242_cast_fp16 = silu(x = linear_39_cast_fp16)[name = string("op_1242_cast_fp16")];1058 tensor<fp16, [3072, 1024]> layers_5_mlp_up_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88147904))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91293696))))[name = string("layers_5_mlp_up_proj_weight_to_fp16_palettized")];1059 tensor<fp16, [1, 128, 3072]> linear_40_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_5_mlp_up_proj_weight_to_fp16_palettized, x = var_1232_cast_fp16)[name = string("linear_40_cast_fp16")];1060 tensor<fp16, [1, 128, 3072]> input_59_cast_fp16 = mul(x = var_1242_cast_fp16, y = linear_40_cast_fp16)[name = string("input_59_cast_fp16")];1061 tensor<fp16, [1024, 3072]> layers_5_mlp_down_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91294272))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94440064))))[name = string("layers_5_mlp_down_proj_weight_to_fp16_palettized")];1062 tensor<fp16, [1, 128, 1024]> linear_41_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_5_mlp_down_proj_weight_to_fp16_palettized, x = input_59_cast_fp16)[name = string("linear_41_cast_fp16")];1063 tensor<fp16, [1, 128, 1024]> x_155_cast_fp16 = add(x = x_149_cast_fp16, y = linear_41_cast_fp16)[name = string("x_155_cast_fp16")];1064 int32 var_1263 = const()[name = string("op_1263"), val = int32(-1)];1065 fp16 var_1262_promoted_to_fp16 = const()[name = string("op_1262_promoted_to_fp16"), val = fp16(0x1p+1)];1066 tensor<fp16, [1, 128, 1024]> var_1272_cast_fp16 = pow(x = x_155_cast_fp16, y = var_1262_promoted_to_fp16)[name = string("op_1272_cast_fp16")];1067 tensor<int32, [1]> var_1274_axes_0 = const()[name = string("op_1274_axes_0"), val = tensor<int32, [1]>([-1])];1068 bool var_1274_keep_dims_0 = const()[name = string("op_1274_keep_dims_0"), val = bool(true)];1069 tensor<fp16, [1, 128, 1]> var_1274_cast_fp16 = reduce_mean(axes = var_1274_axes_0, keep_dims = var_1274_keep_dims_0, x = var_1272_cast_fp16)[name = string("op_1274_cast_fp16")];1070 fp16 var_1275_to_fp16 = const()[name = string("op_1275_to_fp16"), val = fp16(0x1.1p-20)];1071 tensor<fp16, [1, 128, 1]> var_1276_cast_fp16 = add(x = var_1274_cast_fp16, y = var_1275_to_fp16)[name = string("op_1276_cast_fp16")];1072 fp32 norm_49_epsilon_0 = const()[name = string("norm_49_epsilon_0"), val = fp32(0x1.197998p-40)];1073 tensor<fp16, [1, 128, 1]> norm_49_cast_fp16 = rsqrt(epsilon = norm_49_epsilon_0, x = var_1276_cast_fp16)[name = string("norm_49_cast_fp16")];1074 tensor<fp16, [1, 128, 1024]> var_1278_cast_fp16 = mul(x = x_155_cast_fp16, y = norm_49_cast_fp16)[name = string("op_1278_cast_fp16")];1075 tensor<fp16, [1024]> layers_6_input_layernorm_weight_to_fp16 = const()[name = string("layers_6_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94440640)))];1076 tensor<fp16, [1, 128, 1024]> var_1279_cast_fp16 = mul(x = var_1278_cast_fp16, y = layers_6_input_layernorm_weight_to_fp16)[name = string("op_1279_cast_fp16")];1077 tensor<fp16, [2048, 1024]> layers_6_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94442752))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96539968))))[name = string("layers_6_self_attn_q_proj_weight_to_fp16_palettized")];1078 tensor<fp16, [1, 128, 2048]> linear_42_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_6_self_attn_q_proj_weight_to_fp16_palettized, x = var_1279_cast_fp16)[name = string("linear_42_cast_fp16")];1079 tensor<int32, [4]> var_1295 = const()[name = string("op_1295"), val = tensor<int32, [4]>([1, 128, 16, 128])];1080 tensor<fp16, [1, 128, 16, 128]> var_1296_cast_fp16 = reshape(shape = var_1295, x = linear_42_cast_fp16)[name = string("op_1296_cast_fp16")];1081 tensor<int32, [4]> x_161_perm_0 = const()[name = string("x_161_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];1082 tensor<fp16, [1024, 1024]> layers_6_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96540544))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97589184))))[name = string("layers_6_self_attn_k_proj_weight_to_fp16_palettized")];1083 tensor<fp16, [1, 128, 1024]> linear_43_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_6_self_attn_k_proj_weight_to_fp16_palettized, x = var_1279_cast_fp16)[name = string("linear_43_cast_fp16")];1084 tensor<int32, [4]> var_1300 = const()[name = string("op_1300"), val = tensor<int32, [4]>([1, 128, 8, 128])];1085 tensor<fp16, [1, 128, 8, 128]> var_1301_cast_fp16 = reshape(shape = var_1300, x = linear_43_cast_fp16)[name = string("op_1301_cast_fp16")];1086 tensor<int32, [4]> x_165_perm_0 = const()[name = string("x_165_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];1087 tensor<fp16, [1024, 1024]> layers_6_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97589760))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98638400))))[name = string("layers_6_self_attn_v_proj_weight_to_fp16_palettized")];1088 tensor<fp16, [1, 128, 1024]> linear_44_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_6_self_attn_v_proj_weight_to_fp16_palettized, x = var_1279_cast_fp16)[name = string("linear_44_cast_fp16")];1089 tensor<int32, [4]> var_1305 = const()[name = string("op_1305"), val = tensor<int32, [4]>([1, 128, 8, 128])];1090 tensor<fp16, [1, 128, 8, 128]> var_1306_cast_fp16 = reshape(shape = var_1305, x = linear_44_cast_fp16)[name = string("op_1306_cast_fp16")];1091 tensor<int32, [4]> transpose_62_perm_0 = const()[name = string("transpose_62_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];1092 fp16 var_1262_promoted_1_to_fp16 = const()[name = string("op_1262_promoted_1_to_fp16"), val = fp16(0x1p+1)];1093 tensor<fp16, [1, 16, 128, 128]> x_161_cast_fp16 = transpose(perm = x_161_perm_0, x = var_1296_cast_fp16)[name = string("transpose_55")];1094 tensor<fp16, [1, 16, 128, 128]> var_1310_cast_fp16 = pow(x = x_161_cast_fp16, y = var_1262_promoted_1_to_fp16)[name = string("op_1310_cast_fp16")];1095 tensor<int32, [1]> var_1312_axes_0 = const()[name = string("op_1312_axes_0"), val = tensor<int32, [1]>([-1])];1096 bool var_1312_keep_dims_0 = const()[name = string("op_1312_keep_dims_0"), val = bool(true)];1097 tensor<fp16, [1, 16, 128, 1]> var_1312_cast_fp16 = reduce_mean(axes = var_1312_axes_0, keep_dims = var_1312_keep_dims_0, x = var_1310_cast_fp16)[name = string("op_1312_cast_fp16")];1098 fp16 var_1313_to_fp16 = const()[name = string("op_1313_to_fp16"), val = fp16(0x1.1p-20)];1099 tensor<fp16, [1, 16, 128, 1]> var_1314_cast_fp16 = add(x = var_1312_cast_fp16, y = var_1313_to_fp16)[name = string("op_1314_cast_fp16")];1100 fp32 norm_51_epsilon_0 = const()[name = string("norm_51_epsilon_0"), val = fp32(0x1.197998p-40)];1101 tensor<fp16, [1, 16, 128, 1]> norm_51_cast_fp16 = rsqrt(epsilon = norm_51_epsilon_0, x = var_1314_cast_fp16)[name = string("norm_51_cast_fp16")];1102 tensor<fp16, [1, 16, 128, 128]> var_1316_cast_fp16 = mul(x = x_161_cast_fp16, y = norm_51_cast_fp16)[name = string("op_1316_cast_fp16")];1103 tensor<fp16, [128]> layers_6_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_6_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98638976)))];1104 tensor<fp16, [1, 16, 128, 128]> var_1317_cast_fp16 = mul(x = var_1316_cast_fp16, y = layers_6_self_attn_q_norm_weight_to_fp16)[name = string("op_1317_cast_fp16")];1105 fp16 var_1262_promoted_2_to_fp16 = const()[name = string("op_1262_promoted_2_to_fp16"), val = fp16(0x1p+1)];1106 tensor<fp16, [1, 8, 128, 128]> x_165_cast_fp16 = transpose(perm = x_165_perm_0, x = var_1301_cast_fp16)[name = string("transpose_54")];1107 tensor<fp16, [1, 8, 128, 128]> var_1321_cast_fp16 = pow(x = x_165_cast_fp16, y = var_1262_promoted_2_to_fp16)[name = string("op_1321_cast_fp16")];1108 tensor<int32, [1]> var_1323_axes_0 = const()[name = string("op_1323_axes_0"), val = tensor<int32, [1]>([-1])];1109 bool var_1323_keep_dims_0 = const()[name = string("op_1323_keep_dims_0"), val = bool(true)];1110 tensor<fp16, [1, 8, 128, 1]> var_1323_cast_fp16 = reduce_mean(axes = var_1323_axes_0, keep_dims = var_1323_keep_dims_0, x = var_1321_cast_fp16)[name = string("op_1323_cast_fp16")];1111 fp16 var_1324_to_fp16 = const()[name = string("op_1324_to_fp16"), val = fp16(0x1.1p-20)];1112 tensor<fp16, [1, 8, 128, 1]> var_1325_cast_fp16 = add(x = var_1323_cast_fp16, y = var_1324_to_fp16)[name = string("op_1325_cast_fp16")];1113 fp32 norm_53_epsilon_0 = const()[name = string("norm_53_epsilon_0"), val = fp32(0x1.197998p-40)];1114 tensor<fp16, [1, 8, 128, 1]> norm_53_cast_fp16 = rsqrt(epsilon = norm_53_epsilon_0, x = var_1325_cast_fp16)[name = string("norm_53_cast_fp16")];1115 tensor<fp16, [1, 8, 128, 128]> var_1327_cast_fp16 = mul(x = x_165_cast_fp16, y = norm_53_cast_fp16)[name = string("op_1327_cast_fp16")];1116 tensor<fp16, [128]> layers_6_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_6_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98639296)))];1117 tensor<fp16, [1, 8, 128, 128]> var_1328_cast_fp16 = mul(x = var_1327_cast_fp16, y = layers_6_self_attn_k_norm_weight_to_fp16)[name = string("op_1328_cast_fp16")];1118 tensor<int32, [4]> x1_25_begin_0 = const()[name = string("x1_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];1119 tensor<int32, [4]> x1_25_end_0 = const()[name = string("x1_25_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];1120 tensor<bool, [4]> x1_25_end_mask_0 = const()[name = string("x1_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];1121 tensor<fp16, [1, 16, 128, 64]> x1_25_cast_fp16 = slice_by_index(begin = x1_25_begin_0, end = x1_25_end_0, end_mask = x1_25_end_mask_0, x = var_1317_cast_fp16)[name = string("x1_25_cast_fp16")];1122 tensor<int32, [4]> x2_25_begin_0 = const()[name = string("x2_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];1123 tensor<int32, [4]> x2_25_end_0 = const()[name = string("x2_25_end_0"), val = tensor<int32, [4]>([1, 16, 128, 128])];1124 tensor<bool, [4]> x2_25_end_mask_0 = const()[name = string("x2_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];1125 tensor<fp16, [1, 16, 128, 64]> x2_25_cast_fp16 = slice_by_index(begin = x2_25_begin_0, end = x2_25_end_0, end_mask = x2_25_end_mask_0, x = var_1317_cast_fp16)[name = string("x2_25_cast_fp16")];1126 tensor<fp16, [1, 16, 128, 64]> var_1349_cast_fp16 = mul(x = x1_25_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1349_cast_fp16")];1127 tensor<fp16, [1, 16, 128, 64]> var_1350_cast_fp16 = mul(x = x2_25_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1350_cast_fp16")];1128 tensor<fp16, [1, 16, 128, 64]> var_1351_cast_fp16 = sub(x = var_1349_cast_fp16, y = var_1350_cast_fp16)[name = string("op_1351_cast_fp16")];1129 tensor<fp16, [1, 16, 128, 64]> var_1352_cast_fp16 = mul(x = x2_25_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1352_cast_fp16")];1130 tensor<fp16, [1, 16, 128, 64]> var_1353_cast_fp16 = mul(x = x1_25_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1353_cast_fp16")];1131 tensor<fp16, [1, 16, 128, 64]> var_1354_cast_fp16 = add(x = var_1352_cast_fp16, y = var_1353_cast_fp16)[name = string("op_1354_cast_fp16")];1132 bool q_13_interleave_0 = const()[name = string("q_13_interleave_0"), val = bool(false)];1133 tensor<fp16, [1, 16, 128, 128]> q_13_cast_fp16 = concat(axis = var_1263, interleave = q_13_interleave_0, values = (var_1351_cast_fp16, var_1354_cast_fp16))[name = string("q_13_cast_fp16")];1134 tensor<int32, [4]> x1_27_begin_0 = const()[name = string("x1_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];1135 tensor<int32, [4]> x1_27_end_0 = const()[name = string("x1_27_end_0"), val = tensor<int32, [4]>([1, 8, 128, 64])];1136 tensor<bool, [4]> x1_27_end_mask_0 = const()[name = string("x1_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];1137 tensor<fp16, [1, 8, 128, 64]> x1_27_cast_fp16 = slice_by_index(begin = x1_27_begin_0, end = x1_27_end_0, end_mask = x1_27_end_mask_0, x = var_1328_cast_fp16)[name = string("x1_27_cast_fp16")];1138 tensor<int32, [4]> x2_27_begin_0 = const()[name = string("x2_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];1139 tensor<int32, [4]> x2_27_end_0 = const()[name = string("x2_27_end_0"), val = tensor<int32, [4]>([1, 8, 128, 128])];1140 tensor<bool, [4]> x2_27_end_mask_0 = const()[name = string("x2_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];1141 tensor<fp16, [1, 8, 128, 64]> x2_27_cast_fp16 = slice_by_index(begin = x2_27_begin_0, end = x2_27_end_0, end_mask = x2_27_end_mask_0, x = var_1328_cast_fp16)[name = string("x2_27_cast_fp16")];1142 tensor<fp16, [1, 8, 128, 64]> var_1376_cast_fp16 = mul(x = x1_27_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1376_cast_fp16")];1143 tensor<fp16, [1, 8, 128, 64]> var_1377_cast_fp16 = mul(x = x2_27_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1377_cast_fp16")];1144 tensor<fp16, [1, 8, 128, 64]> var_1378_cast_fp16 = sub(x = var_1376_cast_fp16, y = var_1377_cast_fp16)[name = string("op_1378_cast_fp16")];1145 tensor<fp16, [1, 8, 128, 64]> var_1379_cast_fp16 = mul(x = x2_27_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1379_cast_fp16")];1146 tensor<fp16, [1, 8, 128, 64]> var_1380_cast_fp16 = mul(x = x1_27_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1380_cast_fp16")];1147 tensor<fp16, [1, 8, 128, 64]> var_1381_cast_fp16 = add(x = var_1379_cast_fp16, y = var_1380_cast_fp16)[name = string("op_1381_cast_fp16")];1148 bool var_1383_interleave_0 = const()[name = string("op_1383_interleave_0"), val = bool(false)];1149 tensor<fp16, [1, 8, 128, 128]> var_1383_cast_fp16 = concat(axis = var_1263, interleave = var_1383_interleave_0, values = (var_1378_cast_fp16, var_1381_cast_fp16))[name = string("op_1383_cast_fp16")];1150 tensor<int32, [4]> transpose_25_perm_0 = const()[name = string("transpose_25_perm_0"), val = tensor<int32, [4]>([2, 0, 1, 3])];1151 tensor<int32, [2]> concat_112 = const()[name = string("concat_112"), val = tensor<int32, [2]>([128, 1024])];1152 tensor<fp16, [128, 1, 8, 128]> transpose_25_cast_fp16 = transpose(perm = transpose_25_perm_0, x = var_1383_cast_fp16)[name = string("transpose_53")];1153 tensor<fp16, [128, 1024]> reshape_37_cast_fp16 = reshape(shape = concat_112, x = transpose_25_cast_fp16)[name = string("reshape_37_cast_fp16")];1154 bool matmul_12_transpose_x_1 = const()[name = string("matmul_12_transpose_x_1"), val = bool(true)];1155 bool matmul_12_transpose_y_1 = const()[name = string("matmul_12_transpose_y_1"), val = bool(false)];1156 tensor<fp16, [1024, 1024]> matmul_12_cast_fp16 = matmul(transpose_x = matmul_12_transpose_x_1, transpose_y = matmul_12_transpose_y_1, x = var_68_to_fp16, y = reshape_37_cast_fp16)[name = string("matmul_12_cast_fp16")];1157 tensor<int32, [4]> concat_115 = const()[name = string("concat_115"), val = tensor<int32, [4]>([1024, 1, 8, 128])];1158 tensor<fp16, [1024, 1, 8, 128]> reshape_38_cast_fp16 = reshape(shape = concat_115, x = matmul_12_cast_fp16)[name = string("reshape_38_cast_fp16")];1159 tensor<int32, [4]> scattered_k_13_perm_0 = const()[name = string("scattered_k_13_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];1160 tensor<int32, [2]> concat_120 = const()[name = string("concat_120"), val = tensor<int32, [2]>([128, 1024])];1161 tensor<fp16, [128, 1, 8, 128]> transpose_62_cast_fp16 = transpose(perm = transpose_62_perm_0, x = var_1306_cast_fp16)[name = string("transpose_52")];1162 tensor<fp16, [128, 1024]> reshape_40_cast_fp16 = reshape(shape = concat_120, x = transpose_62_cast_fp16)[name = string("reshape_40_cast_fp16")];1163 bool matmul_13_transpose_x_1 = const()[name = string("matmul_13_transpose_x_1"), val = bool(true)];1164 bool matmul_13_transpose_y_1 = const()[name = string("matmul_13_transpose_y_1"), val = bool(false)];1165 tensor<fp16, [1024, 1024]> matmul_13_cast_fp16 = matmul(transpose_x = matmul_13_transpose_x_1, transpose_y = matmul_13_transpose_y_1, x = var_68_to_fp16, y = reshape_40_cast_fp16)[name = string("matmul_13_cast_fp16")];1166 tensor<int32, [4]> concat_123 = const()[name = string("concat_123"), val = tensor<int32, [4]>([1024, 1, 8, 128])];1167 tensor<fp16, [1024, 1, 8, 128]> reshape_41_cast_fp16 = reshape(shape = concat_123, x = matmul_13_cast_fp16)[name = string("reshape_41_cast_fp16")];1168 tensor<int32, [4]> scattered_v_13_perm_0 = const()[name = string("scattered_v_13_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];1169 tensor<fp16, [1, 8, 1024, 128]> read_state_12 = read_state(input = k_cache_6)[name = string("read_state_12")];1170 tensor<fp16, [1, 8, 1024, 128]> k_cache_39_cast_fp16 = mul(x = read_state_12, y = var_224_cast_fp16)[name = string("k_cache_39_cast_fp16")];1171 write_state(data = k_cache_39_cast_fp16, input = k_cache_6)[name = string("coreml_update_state_80_write_state")];1172 tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_80 = read_state(input = k_cache_6)[name = string("coreml_update_state_80")];1173 tensor<fp16, [1, 8, 1024, 128]> scattered_k_13_cast_fp16 = transpose(perm = scattered_k_13_perm_0, x = reshape_38_cast_fp16)[name = string("transpose_51")];1174 tensor<fp16, [1, 8, 1024, 128]> k_cache_41_cast_fp16 = add(x = coreml_update_state_80, y = scattered_k_13_cast_fp16)[name = string("k_cache_41_cast_fp16")];1175 write_state(data = k_cache_41_cast_fp16, input = k_cache_6)[name = string("coreml_update_state_81_write_state")];1176 tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_81 = read_state(input = k_cache_6)[name = string("coreml_update_state_81")];1177 tensor<fp16, [1, 8, 1024, 128]> read_state_13 = read_state(input = v_cache_6)[name = string("read_state_13")];1178 tensor<fp16, [1, 8, 1024, 128]> v_cache_39_cast_fp16 = mul(x = read_state_13, y = var_224_cast_fp16)[name = string("v_cache_39_cast_fp16")];1179 write_state(data = v_cache_39_cast_fp16, input = v_cache_6)[name = string("coreml_update_state_82_write_state")];1180 tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_82 = read_state(input = v_cache_6)[name = string("coreml_update_state_82")];1181 tensor<fp16, [1, 8, 1024, 128]> scattered_v_13_cast_fp16 = transpose(perm = scattered_v_13_perm_0, x = reshape_41_cast_fp16)[name = string("transpose_50")];1182 tensor<fp16, [1, 8, 1024, 128]> v_cache_41_cast_fp16 = add(x = coreml_update_state_82, y = scattered_v_13_cast_fp16)[name = string("v_cache_41_cast_fp16")];1183 write_state(data = v_cache_41_cast_fp16, input = v_cache_6)[name = string("coreml_update_state_83_write_state")];1184 tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_83 = read_state(input = v_cache_6)[name = string("coreml_update_state_83")];1185 tensor<int32, [1]> var_1394_axes_0 = const()[name = string("op_1394_axes_0"), val = tensor<int32, [1]>([2])];1186 tensor<fp16, [1, 8, 1, 1024, 128]> var_1394_cast_fp16 = expand_dims(axes = var_1394_axes_0, x = coreml_update_state_81)[name = string("op_1394_cast_fp16")];1187 tensor<int32, [5]> k_exp_25_reps_0 = const()[name = string("k_exp_25_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];1188 tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_25_cast_fp16 = tile(reps = k_exp_25_reps_0, x = var_1394_cast_fp16)[name = string("k_exp_25_cast_fp16")];1189 tensor<int32, [4]> var_1397 = const()[name = string("op_1397"), val = tensor<int32, [4]>([1, 16, 1024, 128])];1190 tensor<fp16, [1, 16, 1024, 128]> k_exp_27_cast_fp16 = reshape(shape = var_1397, x = k_exp_25_cast_fp16)[name = string("k_exp_27_cast_fp16")];1191 tensor<int32, [1]> var_1399_axes_0 = const()[name = string("op_1399_axes_0"), val = tensor<int32, [1]>([2])];1192 tensor<fp16, [1, 8, 1, 1024, 128]> var_1399_cast_fp16 = expand_dims(axes = var_1399_axes_0, x = coreml_update_state_83)[name = string("op_1399_cast_fp16")];1193 tensor<int32, [5]> v_exp_25_reps_0 = const()[name = string("v_exp_25_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];1194 tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_25_cast_fp16 = tile(reps = v_exp_25_reps_0, x = var_1399_cast_fp16)[name = string("v_exp_25_cast_fp16")];1195 tensor<int32, [4]> var_1402 = const()[name = string("op_1402"), val = tensor<int32, [4]>([1, 16, 1024, 128])];1196 tensor<fp16, [1, 16, 1024, 128]> v_exp_27_cast_fp16 = reshape(shape = var_1402, x = v_exp_25_cast_fp16)[name = string("v_exp_27_cast_fp16")];1197 bool var_1405_transpose_x_1 = const()[name = string("op_1405_transpose_x_1"), val = bool(false)];1198 bool var_1405_transpose_y_1 = const()[name = string("op_1405_transpose_y_1"), val = bool(true)];1199 tensor<fp16, [1, 16, 128, 1024]> var_1405_cast_fp16 = matmul(transpose_x = var_1405_transpose_x_1, transpose_y = var_1405_transpose_y_1, x = q_13_cast_fp16, y = k_exp_27_cast_fp16)[name = string("op_1405_cast_fp16")];1200 fp16 var_1406_to_fp16 = const()[name = string("op_1406_to_fp16"), val = fp16(0x1.6ap-4)];