diff --git "a/DeepSeek_FFN_PF_lut6_chunk_03of08.mlmodelc/model.mil" "b/DeepSeek_FFN_PF_lut6_chunk_03of08.mlmodelc/model.mil" new file mode 100644--- /dev/null +++ "b/DeepSeek_FFN_PF_lut6_chunk_03of08.mlmodelc/model.mil" @@ -0,0 +1,1687 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3404.16.1"}, {"coremlc-version", "3404.23.1"}})] +{ + func infer(tensor causal_mask, tensor current_pos, tensor hidden_states, state> model_model_kv_cache_0, tensor position_ids) { + tensor model_model_layers_8_self_attn_q_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12583040))))[name = string("model_model_layers_8_self_attn_q_proj_weight_palettized")]; + tensor model_model_layers_8_self_attn_k_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12648640))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15794432))))[name = string("model_model_layers_8_self_attn_k_proj_weight_palettized")]; + tensor model_model_layers_8_self_attn_v_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15810880))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18956672))))[name = string("model_model_layers_8_self_attn_v_proj_weight_palettized")]; + tensor model_model_layers_8_mlp_gate_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18973120))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63013376))))[name = string("model_model_layers_8_mlp_gate_proj_weight_palettized")]; + tensor model_model_layers_8_mlp_up_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63242816))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107283072))))[name = string("model_model_layers_8_mlp_up_proj_weight_palettized")]; + tensor model_model_layers_8_mlp_down_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107512512))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151552768))))[name = string("model_model_layers_8_mlp_down_proj_weight_palettized")]; + tensor model_model_layers_9_self_attn_q_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151618368))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164201344))))[name = string("model_model_layers_9_self_attn_q_proj_weight_palettized")]; + tensor model_model_layers_9_self_attn_k_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164266944))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167412736))))[name = string("model_model_layers_9_self_attn_k_proj_weight_palettized")]; + tensor model_model_layers_9_self_attn_v_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167429184))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170574976))))[name = string("model_model_layers_9_self_attn_v_proj_weight_palettized")]; + tensor model_model_layers_9_mlp_gate_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170591424))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214631680))))[name = string("model_model_layers_9_mlp_gate_proj_weight_palettized")]; + tensor model_model_layers_9_mlp_up_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214861120))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258901376))))[name = string("model_model_layers_9_mlp_up_proj_weight_palettized")]; + tensor model_model_layers_9_mlp_down_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259130816))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303171072))))[name = string("model_model_layers_9_mlp_down_proj_weight_palettized")]; + tensor model_model_layers_10_self_attn_q_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303236672))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315819648))))[name = string("model_model_layers_10_self_attn_q_proj_weight_palettized")]; + tensor model_model_layers_10_self_attn_k_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315885248))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319031040))))[name = string("model_model_layers_10_self_attn_k_proj_weight_palettized")]; + tensor model_model_layers_10_self_attn_v_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319047488))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322193280))))[name = string("model_model_layers_10_self_attn_v_proj_weight_palettized")]; + tensor model_model_layers_10_mlp_gate_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322209728))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366249984))))[name = string("model_model_layers_10_mlp_gate_proj_weight_palettized")]; + tensor model_model_layers_10_mlp_up_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366479424))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410519680))))[name = string("model_model_layers_10_mlp_up_proj_weight_palettized")]; + tensor model_model_layers_10_mlp_down_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410749120))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454789376))))[name = string("model_model_layers_10_mlp_down_proj_weight_palettized")]; + tensor model_model_layers_11_self_attn_q_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454854976))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467437952))))[name = string("model_model_layers_11_self_attn_q_proj_weight_palettized")]; + tensor model_model_layers_11_self_attn_k_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467503552))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(470649344))))[name = string("model_model_layers_11_self_attn_k_proj_weight_palettized")]; + tensor model_model_layers_11_self_attn_v_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(470665792))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473811584))))[name = string("model_model_layers_11_self_attn_v_proj_weight_palettized")]; + tensor model_model_layers_11_mlp_gate_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473828032))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(517868288))))[name = string("model_model_layers_11_mlp_gate_proj_weight_palettized")]; + tensor model_model_layers_11_mlp_up_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(518097728))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(562137984))))[name = string("model_model_layers_11_mlp_up_proj_weight_palettized")]; + tensor model_model_layers_11_mlp_down_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(562367424))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(606407680))))[name = string("model_model_layers_11_mlp_down_proj_weight_palettized")]; + int32 var_40 = const()[name = string("op_40"), val = int32(-1)]; + int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)]; + tensor greater_equal_0 = greater_equal(x = current_pos, y = greater_equal_0_y_0)[name = string("greater_equal_0")]; + int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(131072)]; + tensor add_0 = add(x = current_pos, y = slice_by_index_0)[name = string("add_0")]; + tensor select_0 = select(a = current_pos, b = add_0, cond = greater_equal_0)[name = string("select_0")]; + int32 var_149_axis_0 = const()[name = string("op_149_axis_0"), val = int32(1)]; + int32 var_149_batch_dims_0 = const()[name = string("op_149_batch_dims_0"), val = int32(0)]; + bool var_149_validate_indices_0 = const()[name = string("op_149_validate_indices_0"), val = bool(false)]; + tensor var_45_to_fp16 = const()[name = string("op_45_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(606473280)))]; + tensor var_149_cast_fp16 = gather(axis = var_149_axis_0, batch_dims = var_149_batch_dims_0, indices = select_0, validate_indices = var_149_validate_indices_0, x = var_45_to_fp16)[name = string("op_149_cast_fp16")]; + tensor var_150 = const()[name = string("op_150"), val = tensor([1, 1, 1, -1])]; + tensor sin_1_cast_fp16 = reshape(shape = var_150, x = var_149_cast_fp16)[name = string("sin_1_cast_fp16")]; + int32 var_154_axis_0 = const()[name = string("op_154_axis_0"), val = int32(1)]; + int32 var_154_batch_dims_0 = const()[name = string("op_154_batch_dims_0"), val = int32(0)]; + bool var_154_validate_indices_0 = const()[name = string("op_154_validate_indices_0"), val = bool(false)]; + tensor var_39_to_fp16 = const()[name = string("op_39_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(640027776)))]; + tensor var_154_cast_fp16 = gather(axis = var_154_axis_0, batch_dims = var_154_batch_dims_0, indices = select_0, validate_indices = var_154_validate_indices_0, x = var_39_to_fp16)[name = string("op_154_cast_fp16")]; + tensor var_155 = const()[name = string("op_155"), val = tensor([1, 1, 1, -1])]; + tensor cos_1_cast_fp16 = reshape(shape = var_155, x = var_154_cast_fp16)[name = string("cos_1_cast_fp16")]; + tensor mean_1_axes_0 = const()[name = string("mean_1_axes_0"), val = tensor([-1])]; + bool mean_1_keep_dims_0 = const()[name = string("mean_1_keep_dims_0"), val = bool(true)]; + tensor mean_1_cast_fp16 = reduce_mean(axes = mean_1_axes_0, keep_dims = mean_1_keep_dims_0, x = hidden_states)[name = string("mean_1_cast_fp16")]; + tensor input_1_cast_fp16 = sub(x = hidden_states, y = mean_1_cast_fp16)[name = string("input_1_cast_fp16")]; + tensor var_163_axes_0 = const()[name = string("op_163_axes_0"), val = tensor([-1])]; + tensor model_model_layers_8_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_8_input_layernorm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(673582272)))]; + fp16 var_35_to_fp16 = const()[name = string("op_35_to_fp16"), val = fp16(0x1.5p-17)]; + tensor var_163_cast_fp16 = layer_norm(axes = var_163_axes_0, epsilon = var_35_to_fp16, gamma = model_model_layers_8_input_layernorm_weight_to_fp16, x = input_1_cast_fp16)[name = string("op_163_cast_fp16")]; + tensor var_166 = const()[name = string("op_166"), val = tensor([0, 2, 1])]; + tensor var_168_axes_0 = const()[name = string("op_168_axes_0"), val = tensor([2])]; + tensor var_167 = transpose(perm = var_166, x = var_163_cast_fp16)[name = string("transpose_15")]; + tensor var_168 = expand_dims(axes = var_168_axes_0, x = var_167)[name = string("op_168")]; + string var_175_pad_type_0 = const()[name = string("op_175_pad_type_0"), val = string("valid")]; + tensor var_175_strides_0 = const()[name = string("op_175_strides_0"), val = tensor([1, 1])]; + tensor var_175_pad_0 = const()[name = string("op_175_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_175_dilations_0 = const()[name = string("op_175_dilations_0"), val = tensor([1, 1])]; + int32 var_175_groups_0 = const()[name = string("op_175_groups_0"), val = int32(1)]; + tensor var_175 = conv(dilations = var_175_dilations_0, groups = var_175_groups_0, pad = var_175_pad_0, pad_type = var_175_pad_type_0, strides = var_175_strides_0, weight = model_model_layers_8_self_attn_q_proj_weight_palettized, x = var_168)[name = string("op_175")]; + tensor var_176 = const()[name = string("op_176"), val = tensor([1, 32, 1, 128])]; + tensor var_177 = reshape(shape = var_176, x = var_175)[name = string("op_177")]; + string var_184_pad_type_0 = const()[name = string("op_184_pad_type_0"), val = string("valid")]; + tensor var_184_strides_0 = const()[name = string("op_184_strides_0"), val = tensor([1, 1])]; + tensor var_184_pad_0 = const()[name = string("op_184_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_184_dilations_0 = const()[name = string("op_184_dilations_0"), val = tensor([1, 1])]; + int32 var_184_groups_0 = const()[name = string("op_184_groups_0"), val = int32(1)]; + tensor var_184 = conv(dilations = var_184_dilations_0, groups = var_184_groups_0, pad = var_184_pad_0, pad_type = var_184_pad_type_0, strides = var_184_strides_0, weight = model_model_layers_8_self_attn_k_proj_weight_palettized, x = var_168)[name = string("op_184")]; + tensor var_185 = const()[name = string("op_185"), val = tensor([1, 8, 1, 128])]; + tensor var_186 = reshape(shape = var_185, x = var_184)[name = string("op_186")]; + string var_193_pad_type_0 = const()[name = string("op_193_pad_type_0"), val = string("valid")]; + tensor var_193_strides_0 = const()[name = string("op_193_strides_0"), val = tensor([1, 1])]; + tensor var_193_pad_0 = const()[name = string("op_193_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_193_dilations_0 = const()[name = string("op_193_dilations_0"), val = tensor([1, 1])]; + int32 var_193_groups_0 = const()[name = string("op_193_groups_0"), val = int32(1)]; + tensor var_193 = conv(dilations = var_193_dilations_0, groups = var_193_groups_0, pad = var_193_pad_0, pad_type = var_193_pad_type_0, strides = var_193_strides_0, weight = model_model_layers_8_self_attn_v_proj_weight_palettized, x = var_168)[name = string("op_193")]; + tensor var_194 = const()[name = string("op_194"), val = tensor([1, 8, 1, 128])]; + tensor var_195 = reshape(shape = var_194, x = var_193)[name = string("op_195")]; + tensor x1_1_begin_0 = const()[name = string("x1_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_1_end_0 = const()[name = string("x1_1_end_0"), val = tensor([1, 32, 1, 64])]; + tensor x1_1_end_mask_0 = const()[name = string("x1_1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_1 = slice_by_index(begin = x1_1_begin_0, end = x1_1_end_0, end_mask = x1_1_end_mask_0, x = var_177)[name = string("x1_1")]; + tensor x2_1_begin_0 = const()[name = string("x2_1_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_1_end_0 = const()[name = string("x2_1_end_0"), val = tensor([1, 32, 1, 128])]; + tensor x2_1_end_mask_0 = const()[name = string("x2_1_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_1 = slice_by_index(begin = x2_1_begin_0, end = x2_1_end_0, end_mask = x2_1_end_mask_0, x = var_177)[name = string("x2_1")]; + tensor cos_3_begin_0 = const()[name = string("cos_3_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor cos_3_end_0 = const()[name = string("cos_3_end_0"), val = tensor([1, 1, 1, 64])]; + tensor cos_3_end_mask_0 = const()[name = string("cos_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor cos_3_cast_fp16 = slice_by_index(begin = cos_3_begin_0, end = cos_3_end_0, end_mask = cos_3_end_mask_0, x = cos_1_cast_fp16)[name = string("cos_3_cast_fp16")]; + tensor sin_3_begin_0 = const()[name = string("sin_3_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor sin_3_end_0 = const()[name = string("sin_3_end_0"), val = tensor([1, 1, 1, 64])]; + tensor sin_3_end_mask_0 = const()[name = string("sin_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor sin_3_cast_fp16 = slice_by_index(begin = sin_3_begin_0, end = sin_3_end_0, end_mask = sin_3_end_mask_0, x = sin_1_cast_fp16)[name = string("sin_3_cast_fp16")]; + tensor var_209_cast_fp16 = mul(x = x1_1, y = cos_3_cast_fp16)[name = string("op_209_cast_fp16")]; + tensor var_210_cast_fp16 = mul(x = x2_1, y = sin_3_cast_fp16)[name = string("op_210_cast_fp16")]; + tensor var_211_cast_fp16 = sub(x = var_209_cast_fp16, y = var_210_cast_fp16)[name = string("op_211_cast_fp16")]; + tensor var_212_cast_fp16 = mul(x = x2_1, y = cos_3_cast_fp16)[name = string("op_212_cast_fp16")]; + tensor var_213_cast_fp16 = mul(x = x1_1, y = sin_3_cast_fp16)[name = string("op_213_cast_fp16")]; + tensor var_214_cast_fp16 = add(x = var_212_cast_fp16, y = var_213_cast_fp16)[name = string("op_214_cast_fp16")]; + bool rotated_1_interleave_0 = const()[name = string("rotated_1_interleave_0"), val = bool(false)]; + tensor rotated_1_cast_fp16 = concat(axis = var_40, interleave = rotated_1_interleave_0, values = (var_211_cast_fp16, var_214_cast_fp16))[name = string("rotated_1_cast_fp16")]; + tensor x1_3_begin_0 = const()[name = string("x1_3_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_3_end_0 = const()[name = string("x1_3_end_0"), val = tensor([1, 8, 1, 64])]; + tensor x1_3_end_mask_0 = const()[name = string("x1_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_3 = slice_by_index(begin = x1_3_begin_0, end = x1_3_end_0, end_mask = x1_3_end_mask_0, x = var_186)[name = string("x1_3")]; + tensor x2_3_begin_0 = const()[name = string("x2_3_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_3_end_0 = const()[name = string("x2_3_end_0"), val = tensor([1, 8, 1, 128])]; + tensor x2_3_end_mask_0 = const()[name = string("x2_3_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_3 = slice_by_index(begin = x2_3_begin_0, end = x2_3_end_0, end_mask = x2_3_end_mask_0, x = var_186)[name = string("x2_3")]; + tensor var_230_cast_fp16 = mul(x = x1_3, y = cos_3_cast_fp16)[name = string("op_230_cast_fp16")]; + tensor var_231_cast_fp16 = mul(x = x2_3, y = sin_3_cast_fp16)[name = string("op_231_cast_fp16")]; + tensor var_232_cast_fp16 = sub(x = var_230_cast_fp16, y = var_231_cast_fp16)[name = string("op_232_cast_fp16")]; + tensor var_233_cast_fp16 = mul(x = x2_3, y = cos_3_cast_fp16)[name = string("op_233_cast_fp16")]; + tensor var_234_cast_fp16 = mul(x = x1_3, y = sin_3_cast_fp16)[name = string("op_234_cast_fp16")]; + tensor var_235_cast_fp16 = add(x = var_233_cast_fp16, y = var_234_cast_fp16)[name = string("op_235_cast_fp16")]; + bool rotated_3_interleave_0 = const()[name = string("rotated_3_interleave_0"), val = bool(false)]; + tensor rotated_3_cast_fp16 = concat(axis = var_40, interleave = rotated_3_interleave_0, values = (var_232_cast_fp16, var_235_cast_fp16))[name = string("rotated_3_cast_fp16")]; + int32 var_239 = const()[name = string("op_239"), val = int32(1)]; + tensor var_240 = add(x = current_pos, y = var_239)[name = string("op_240")]; + tensor read_state_0 = read_state(input = model_model_kv_cache_0)[name = string("read_state_0")]; + tensor expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor([8])]; + tensor expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor([0])]; + tensor expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor([0])]; + tensor expand_dims_4 = const()[name = string("expand_dims_4"), val = tensor([9])]; + int32 concat_2_axis_0 = const()[name = string("concat_2_axis_0"), val = int32(0)]; + bool concat_2_interleave_0 = const()[name = string("concat_2_interleave_0"), val = bool(false)]; + tensor concat_2 = concat(axis = concat_2_axis_0, interleave = concat_2_interleave_0, values = (expand_dims_0, expand_dims_1, current_pos, expand_dims_3))[name = string("concat_2")]; + tensor concat_3_values1_0 = const()[name = string("concat_3_values1_0"), val = tensor([0])]; + tensor concat_3_values3_0 = const()[name = string("concat_3_values3_0"), val = tensor([0])]; + int32 concat_3_axis_0 = const()[name = string("concat_3_axis_0"), val = int32(0)]; + bool concat_3_interleave_0 = const()[name = string("concat_3_interleave_0"), val = bool(false)]; + tensor concat_3 = concat(axis = concat_3_axis_0, interleave = concat_3_interleave_0, values = (expand_dims_4, concat_3_values1_0, var_240, concat_3_values3_0))[name = string("concat_3")]; + tensor model_model_kv_cache_0_internal_tensor_assign_1_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1])]; + tensor model_model_kv_cache_0_internal_tensor_assign_1_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, false, false, false])]; + tensor model_model_kv_cache_0_internal_tensor_assign_1_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, false, true])]; + tensor model_model_kv_cache_0_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([false, false, false, false])]; + tensor model_model_kv_cache_0_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_2, begin_mask = model_model_kv_cache_0_internal_tensor_assign_1_begin_mask_0, end = concat_3, end_mask = model_model_kv_cache_0_internal_tensor_assign_1_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_1_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_1_stride_0, update = rotated_3_cast_fp16, x = read_state_0)[name = string("model_model_kv_cache_0_internal_tensor_assign_1_cast_fp16")]; + write_state(data = model_model_kv_cache_0_internal_tensor_assign_1_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_0_write_state")]; + tensor coreml_update_state_8 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_0")]; + tensor expand_dims_6 = const()[name = string("expand_dims_6"), val = tensor([40])]; + tensor expand_dims_7 = const()[name = string("expand_dims_7"), val = tensor([0])]; + tensor expand_dims_9 = const()[name = string("expand_dims_9"), val = tensor([0])]; + tensor expand_dims_10 = const()[name = string("expand_dims_10"), val = tensor([41])]; + int32 concat_6_axis_0 = const()[name = string("concat_6_axis_0"), val = int32(0)]; + bool concat_6_interleave_0 = const()[name = string("concat_6_interleave_0"), val = bool(false)]; + tensor concat_6 = concat(axis = concat_6_axis_0, interleave = concat_6_interleave_0, values = (expand_dims_6, expand_dims_7, current_pos, expand_dims_9))[name = string("concat_6")]; + tensor concat_7_values1_0 = const()[name = string("concat_7_values1_0"), val = tensor([0])]; + tensor concat_7_values3_0 = const()[name = string("concat_7_values3_0"), val = tensor([0])]; + int32 concat_7_axis_0 = const()[name = string("concat_7_axis_0"), val = int32(0)]; + bool concat_7_interleave_0 = const()[name = string("concat_7_interleave_0"), val = bool(false)]; + tensor concat_7 = concat(axis = concat_7_axis_0, interleave = concat_7_interleave_0, values = (expand_dims_10, concat_7_values1_0, var_240, concat_7_values3_0))[name = string("concat_7")]; + tensor model_model_kv_cache_0_internal_tensor_assign_2_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1])]; + tensor model_model_kv_cache_0_internal_tensor_assign_2_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, false, false, false])]; + tensor model_model_kv_cache_0_internal_tensor_assign_2_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, false, true])]; + tensor model_model_kv_cache_0_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([false, false, false, false])]; + tensor model_model_kv_cache_0_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_6, begin_mask = model_model_kv_cache_0_internal_tensor_assign_2_begin_mask_0, end = concat_7, end_mask = model_model_kv_cache_0_internal_tensor_assign_2_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_2_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_2_stride_0, update = var_195, x = coreml_update_state_8)[name = string("model_model_kv_cache_0_internal_tensor_assign_2_cast_fp16")]; + write_state(data = model_model_kv_cache_0_internal_tensor_assign_2_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_1_write_state")]; + tensor coreml_update_state_9 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_1")]; + tensor var_255_begin_0 = const()[name = string("op_255_begin_0"), val = tensor([8, 0, 0, 0])]; + tensor var_255_end_0 = const()[name = string("op_255_end_0"), val = tensor([9, 8, 1024, 128])]; + tensor var_255_end_mask_0 = const()[name = string("op_255_end_mask_0"), val = tensor([false, true, true, true])]; + tensor var_255_cast_fp16 = slice_by_index(begin = var_255_begin_0, end = var_255_end_0, end_mask = var_255_end_mask_0, x = coreml_update_state_9)[name = string("op_255_cast_fp16")]; + tensor K_layer_cache_1_axes_0 = const()[name = string("K_layer_cache_1_axes_0"), val = tensor([0])]; + tensor K_layer_cache_1_cast_fp16 = squeeze(axes = K_layer_cache_1_axes_0, x = var_255_cast_fp16)[name = string("K_layer_cache_1_cast_fp16")]; + tensor var_257_begin_0 = const()[name = string("op_257_begin_0"), val = tensor([40, 0, 0, 0])]; + tensor var_257_end_0 = const()[name = string("op_257_end_0"), val = tensor([41, 8, 1024, 128])]; + tensor var_257_end_mask_0 = const()[name = string("op_257_end_mask_0"), val = tensor([false, true, true, true])]; + tensor var_257_cast_fp16 = slice_by_index(begin = var_257_begin_0, end = var_257_end_0, end_mask = var_257_end_mask_0, x = coreml_update_state_9)[name = string("op_257_cast_fp16")]; + tensor V_layer_cache_1_axes_0 = const()[name = string("V_layer_cache_1_axes_0"), val = tensor([0])]; + tensor V_layer_cache_1_cast_fp16 = squeeze(axes = V_layer_cache_1_axes_0, x = var_257_cast_fp16)[name = string("V_layer_cache_1_cast_fp16")]; + tensor x_11_axes_0 = const()[name = string("x_11_axes_0"), val = tensor([1])]; + tensor x_11_cast_fp16 = expand_dims(axes = x_11_axes_0, x = K_layer_cache_1_cast_fp16)[name = string("x_11_cast_fp16")]; + tensor var_266 = const()[name = string("op_266"), val = tensor([1, 4, 1, 1])]; + tensor x_13_cast_fp16 = tile(reps = var_266, x = x_11_cast_fp16)[name = string("x_13_cast_fp16")]; + tensor var_270 = const()[name = string("op_270"), val = tensor([1, -1, 1024, 128])]; + tensor key_states_3_cast_fp16 = reshape(shape = var_270, x = x_13_cast_fp16)[name = string("key_states_3_cast_fp16")]; + tensor x_17_axes_0 = const()[name = string("x_17_axes_0"), val = tensor([1])]; + tensor x_17_cast_fp16 = expand_dims(axes = x_17_axes_0, x = V_layer_cache_1_cast_fp16)[name = string("x_17_cast_fp16")]; + tensor var_273 = const()[name = string("op_273"), val = tensor([1, 4, 1, 1])]; + tensor x_19_cast_fp16 = tile(reps = var_273, x = x_17_cast_fp16)[name = string("x_19_cast_fp16")]; + tensor var_277 = const()[name = string("op_277"), val = tensor([1, -1, 1024, 128])]; + tensor value_states_3_cast_fp16 = reshape(shape = var_277, x = x_19_cast_fp16)[name = string("value_states_3_cast_fp16")]; + bool var_280_transpose_x_1 = const()[name = string("op_280_transpose_x_1"), val = bool(false)]; + bool var_280_transpose_y_1 = const()[name = string("op_280_transpose_y_1"), val = bool(true)]; + tensor var_280_cast_fp16 = matmul(transpose_x = var_280_transpose_x_1, transpose_y = var_280_transpose_y_1, x = rotated_1_cast_fp16, y = key_states_3_cast_fp16)[name = string("op_280_cast_fp16")]; + fp16 var_281_to_fp16 = const()[name = string("op_281_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor attn_weights_1_cast_fp16 = mul(x = var_280_cast_fp16, y = var_281_to_fp16)[name = string("attn_weights_1_cast_fp16")]; + tensor x_21_cast_fp16 = add(x = attn_weights_1_cast_fp16, y = causal_mask)[name = string("x_21_cast_fp16")]; + tensor reduce_max_0_axes_0 = const()[name = string("reduce_max_0_axes_0"), val = tensor([-1])]; + bool reduce_max_0_keep_dims_0 = const()[name = string("reduce_max_0_keep_dims_0"), val = bool(true)]; + tensor reduce_max_0_cast_fp16 = reduce_max(axes = reduce_max_0_axes_0, keep_dims = reduce_max_0_keep_dims_0, x = x_21_cast_fp16)[name = string("reduce_max_0_cast_fp16")]; + tensor x_23_cast_fp16 = sub(x = x_21_cast_fp16, y = reduce_max_0_cast_fp16)[name = string("x_23_cast_fp16")]; + tensor exp_x_1_cast_fp16 = exp(x = x_23_cast_fp16)[name = string("exp_x_1_cast_fp16")]; + tensor var_292_axes_0 = const()[name = string("op_292_axes_0"), val = tensor([-1])]; + bool var_292_keep_dims_0 = const()[name = string("op_292_keep_dims_0"), val = bool(true)]; + tensor var_292_cast_fp16 = reduce_sum(axes = var_292_axes_0, keep_dims = var_292_keep_dims_0, x = exp_x_1_cast_fp16)[name = string("op_292_cast_fp16")]; + tensor attn_weights_3_cast_fp16 = real_div(x = exp_x_1_cast_fp16, y = var_292_cast_fp16)[name = string("attn_weights_3_cast_fp16")]; + bool attn_output_1_transpose_x_0 = const()[name = string("attn_output_1_transpose_x_0"), val = bool(false)]; + bool attn_output_1_transpose_y_0 = const()[name = string("attn_output_1_transpose_y_0"), val = bool(false)]; + tensor attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = attn_weights_3_cast_fp16, y = value_states_3_cast_fp16)[name = string("attn_output_1_cast_fp16")]; + tensor var_295_perm_0 = const()[name = string("op_295_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_297 = const()[name = string("op_297"), val = tensor([1, 1, 4096])]; + tensor var_295_cast_fp16 = transpose(perm = var_295_perm_0, x = attn_output_1_cast_fp16)[name = string("transpose_14")]; + tensor input_5_cast_fp16 = reshape(shape = var_297, x = var_295_cast_fp16)[name = string("input_5_cast_fp16")]; + tensor model_model_layers_8_self_attn_o_proj_weight_promoted_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(673590528))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(686173504))))[name = string("model_model_layers_8_self_attn_o_proj_weight_promoted_to_fp16_palettized")]; + tensor linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(686239104)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_8_self_attn_o_proj_weight_promoted_to_fp16_palettized, x = input_5_cast_fp16)[name = string("linear_0_cast_fp16")]; + tensor hidden_states_5_cast_fp16 = add(x = hidden_states, y = linear_0_cast_fp16)[name = string("hidden_states_5_cast_fp16")]; + tensor mean_3_axes_0 = const()[name = string("mean_3_axes_0"), val = tensor([-1])]; + bool mean_3_keep_dims_0 = const()[name = string("mean_3_keep_dims_0"), val = bool(true)]; + tensor mean_3_cast_fp16 = reduce_mean(axes = mean_3_axes_0, keep_dims = mean_3_keep_dims_0, x = hidden_states_5_cast_fp16)[name = string("mean_3_cast_fp16")]; + tensor input_7_cast_fp16 = sub(x = hidden_states_5_cast_fp16, y = mean_3_cast_fp16)[name = string("input_7_cast_fp16")]; + tensor var_308_axes_0 = const()[name = string("op_308_axes_0"), val = tensor([-1])]; + tensor model_model_layers_8_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_8_post_attention_layernorm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(686247360)))]; + tensor var_308_cast_fp16 = layer_norm(axes = var_308_axes_0, epsilon = var_35_to_fp16, gamma = model_model_layers_8_post_attention_layernorm_weight_to_fp16, x = input_7_cast_fp16)[name = string("op_308_cast_fp16")]; + tensor var_315 = const()[name = string("op_315"), val = tensor([0, 2, 1])]; + tensor input_9_axes_0 = const()[name = string("input_9_axes_0"), val = tensor([2])]; + tensor var_316 = transpose(perm = var_315, x = var_308_cast_fp16)[name = string("transpose_13")]; + tensor input_9 = expand_dims(axes = input_9_axes_0, x = var_316)[name = string("input_9")]; + string input_11_pad_type_0 = const()[name = string("input_11_pad_type_0"), val = string("valid")]; + tensor input_11_strides_0 = const()[name = string("input_11_strides_0"), val = tensor([1, 1])]; + tensor input_11_pad_0 = const()[name = string("input_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_11_dilations_0 = const()[name = string("input_11_dilations_0"), val = tensor([1, 1])]; + int32 input_11_groups_0 = const()[name = string("input_11_groups_0"), val = int32(1)]; + tensor input_11 = conv(dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = model_model_layers_8_mlp_gate_proj_weight_palettized, x = input_9)[name = string("input_11")]; + string up_states_1_pad_type_0 = const()[name = string("up_states_1_pad_type_0"), val = string("valid")]; + tensor up_states_1_strides_0 = const()[name = string("up_states_1_strides_0"), val = tensor([1, 1])]; + tensor up_states_1_pad_0 = const()[name = string("up_states_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor up_states_1_dilations_0 = const()[name = string("up_states_1_dilations_0"), val = tensor([1, 1])]; + int32 up_states_1_groups_0 = const()[name = string("up_states_1_groups_0"), val = int32(1)]; + tensor up_states_1 = conv(dilations = up_states_1_dilations_0, groups = up_states_1_groups_0, pad = up_states_1_pad_0, pad_type = up_states_1_pad_type_0, strides = up_states_1_strides_0, weight = model_model_layers_8_mlp_up_proj_weight_palettized, x = input_9)[name = string("up_states_1")]; + tensor gate_states_1 = silu(x = input_11)[name = string("gate_states_1")]; + tensor input_13 = mul(x = gate_states_1, y = up_states_1)[name = string("input_13")]; + string hidden_states_7_pad_type_0 = const()[name = string("hidden_states_7_pad_type_0"), val = string("valid")]; + tensor hidden_states_7_strides_0 = const()[name = string("hidden_states_7_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_7_pad_0 = const()[name = string("hidden_states_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_7_dilations_0 = const()[name = string("hidden_states_7_dilations_0"), val = tensor([1, 1])]; + int32 hidden_states_7_groups_0 = const()[name = string("hidden_states_7_groups_0"), val = int32(1)]; + tensor hidden_states_7 = conv(dilations = hidden_states_7_dilations_0, groups = hidden_states_7_groups_0, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = hidden_states_7_strides_0, weight = model_model_layers_8_mlp_down_proj_weight_palettized, x = input_13)[name = string("hidden_states_7")]; + tensor var_338_axes_0 = const()[name = string("op_338_axes_0"), val = tensor([2])]; + tensor var_338 = squeeze(axes = var_338_axes_0, x = hidden_states_7)[name = string("op_338")]; + tensor var_339 = const()[name = string("op_339"), val = tensor([0, 2, 1])]; + tensor var_340 = transpose(perm = var_339, x = var_338)[name = string("transpose_12")]; + tensor hidden_states_9_cast_fp16 = add(x = hidden_states_5_cast_fp16, y = var_340)[name = string("hidden_states_9_cast_fp16")]; + tensor mean_5_axes_0 = const()[name = string("mean_5_axes_0"), val = tensor([-1])]; + bool mean_5_keep_dims_0 = const()[name = string("mean_5_keep_dims_0"), val = bool(true)]; + tensor mean_5_cast_fp16 = reduce_mean(axes = mean_5_axes_0, keep_dims = mean_5_keep_dims_0, x = hidden_states_9_cast_fp16)[name = string("mean_5_cast_fp16")]; + tensor input_15_cast_fp16 = sub(x = hidden_states_9_cast_fp16, y = mean_5_cast_fp16)[name = string("input_15_cast_fp16")]; + tensor var_348_axes_0 = const()[name = string("op_348_axes_0"), val = tensor([-1])]; + tensor model_model_layers_9_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_9_input_layernorm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(686255616)))]; + tensor var_348_cast_fp16 = layer_norm(axes = var_348_axes_0, epsilon = var_35_to_fp16, gamma = model_model_layers_9_input_layernorm_weight_to_fp16, x = input_15_cast_fp16)[name = string("op_348_cast_fp16")]; + tensor var_351 = const()[name = string("op_351"), val = tensor([0, 2, 1])]; + tensor var_353_axes_0 = const()[name = string("op_353_axes_0"), val = tensor([2])]; + tensor var_352 = transpose(perm = var_351, x = var_348_cast_fp16)[name = string("transpose_11")]; + tensor var_353 = expand_dims(axes = var_353_axes_0, x = var_352)[name = string("op_353")]; + string var_360_pad_type_0 = const()[name = string("op_360_pad_type_0"), val = string("valid")]; + tensor var_360_strides_0 = const()[name = string("op_360_strides_0"), val = tensor([1, 1])]; + tensor var_360_pad_0 = const()[name = string("op_360_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_360_dilations_0 = const()[name = string("op_360_dilations_0"), val = tensor([1, 1])]; + int32 var_360_groups_0 = const()[name = string("op_360_groups_0"), val = int32(1)]; + tensor var_360 = conv(dilations = var_360_dilations_0, groups = var_360_groups_0, pad = var_360_pad_0, pad_type = var_360_pad_type_0, strides = var_360_strides_0, weight = model_model_layers_9_self_attn_q_proj_weight_palettized, x = var_353)[name = string("op_360")]; + tensor var_361 = const()[name = string("op_361"), val = tensor([1, 32, 1, 128])]; + tensor var_362 = reshape(shape = var_361, x = var_360)[name = string("op_362")]; + string var_369_pad_type_0 = const()[name = string("op_369_pad_type_0"), val = string("valid")]; + tensor var_369_strides_0 = const()[name = string("op_369_strides_0"), val = tensor([1, 1])]; + tensor var_369_pad_0 = const()[name = string("op_369_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_369_dilations_0 = const()[name = string("op_369_dilations_0"), val = tensor([1, 1])]; + int32 var_369_groups_0 = const()[name = string("op_369_groups_0"), val = int32(1)]; + tensor var_369 = conv(dilations = var_369_dilations_0, groups = var_369_groups_0, pad = var_369_pad_0, pad_type = var_369_pad_type_0, strides = var_369_strides_0, weight = model_model_layers_9_self_attn_k_proj_weight_palettized, x = var_353)[name = string("op_369")]; + tensor var_370 = const()[name = string("op_370"), val = tensor([1, 8, 1, 128])]; + tensor var_371 = reshape(shape = var_370, x = var_369)[name = string("op_371")]; + string var_378_pad_type_0 = const()[name = string("op_378_pad_type_0"), val = string("valid")]; + tensor var_378_strides_0 = const()[name = string("op_378_strides_0"), val = tensor([1, 1])]; + tensor var_378_pad_0 = const()[name = string("op_378_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_378_dilations_0 = const()[name = string("op_378_dilations_0"), val = tensor([1, 1])]; + int32 var_378_groups_0 = const()[name = string("op_378_groups_0"), val = int32(1)]; + tensor var_378 = conv(dilations = var_378_dilations_0, groups = var_378_groups_0, pad = var_378_pad_0, pad_type = var_378_pad_type_0, strides = var_378_strides_0, weight = model_model_layers_9_self_attn_v_proj_weight_palettized, x = var_353)[name = string("op_378")]; + tensor var_379 = const()[name = string("op_379"), val = tensor([1, 8, 1, 128])]; + tensor var_380 = reshape(shape = var_379, x = var_378)[name = string("op_380")]; + tensor x1_5_begin_0 = const()[name = string("x1_5_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_5_end_0 = const()[name = string("x1_5_end_0"), val = tensor([1, 32, 1, 64])]; + tensor x1_5_end_mask_0 = const()[name = string("x1_5_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_5 = slice_by_index(begin = x1_5_begin_0, end = x1_5_end_0, end_mask = x1_5_end_mask_0, x = var_362)[name = string("x1_5")]; + tensor x2_5_begin_0 = const()[name = string("x2_5_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_5_end_0 = const()[name = string("x2_5_end_0"), val = tensor([1, 32, 1, 128])]; + tensor x2_5_end_mask_0 = const()[name = string("x2_5_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_5 = slice_by_index(begin = x2_5_begin_0, end = x2_5_end_0, end_mask = x2_5_end_mask_0, x = var_362)[name = string("x2_5")]; + tensor var_394_cast_fp16 = mul(x = x1_5, y = cos_3_cast_fp16)[name = string("op_394_cast_fp16")]; + tensor var_395_cast_fp16 = mul(x = x2_5, y = sin_3_cast_fp16)[name = string("op_395_cast_fp16")]; + tensor var_396_cast_fp16 = sub(x = var_394_cast_fp16, y = var_395_cast_fp16)[name = string("op_396_cast_fp16")]; + tensor var_397_cast_fp16 = mul(x = x2_5, y = cos_3_cast_fp16)[name = string("op_397_cast_fp16")]; + tensor var_398_cast_fp16 = mul(x = x1_5, y = sin_3_cast_fp16)[name = string("op_398_cast_fp16")]; + tensor var_399_cast_fp16 = add(x = var_397_cast_fp16, y = var_398_cast_fp16)[name = string("op_399_cast_fp16")]; + bool rotated_5_interleave_0 = const()[name = string("rotated_5_interleave_0"), val = bool(false)]; + tensor rotated_5_cast_fp16 = concat(axis = var_40, interleave = rotated_5_interleave_0, values = (var_396_cast_fp16, var_399_cast_fp16))[name = string("rotated_5_cast_fp16")]; + tensor x1_7_begin_0 = const()[name = string("x1_7_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_7_end_0 = const()[name = string("x1_7_end_0"), val = tensor([1, 8, 1, 64])]; + tensor x1_7_end_mask_0 = const()[name = string("x1_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_7 = slice_by_index(begin = x1_7_begin_0, end = x1_7_end_0, end_mask = x1_7_end_mask_0, x = var_371)[name = string("x1_7")]; + tensor x2_7_begin_0 = const()[name = string("x2_7_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_7_end_0 = const()[name = string("x2_7_end_0"), val = tensor([1, 8, 1, 128])]; + tensor x2_7_end_mask_0 = const()[name = string("x2_7_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_7 = slice_by_index(begin = x2_7_begin_0, end = x2_7_end_0, end_mask = x2_7_end_mask_0, x = var_371)[name = string("x2_7")]; + tensor var_415_cast_fp16 = mul(x = x1_7, y = cos_3_cast_fp16)[name = string("op_415_cast_fp16")]; + tensor var_416_cast_fp16 = mul(x = x2_7, y = sin_3_cast_fp16)[name = string("op_416_cast_fp16")]; + tensor var_417_cast_fp16 = sub(x = var_415_cast_fp16, y = var_416_cast_fp16)[name = string("op_417_cast_fp16")]; + tensor var_418_cast_fp16 = mul(x = x2_7, y = cos_3_cast_fp16)[name = string("op_418_cast_fp16")]; + tensor var_419_cast_fp16 = mul(x = x1_7, y = sin_3_cast_fp16)[name = string("op_419_cast_fp16")]; + tensor var_420_cast_fp16 = add(x = var_418_cast_fp16, y = var_419_cast_fp16)[name = string("op_420_cast_fp16")]; + bool rotated_7_interleave_0 = const()[name = string("rotated_7_interleave_0"), val = bool(false)]; + tensor rotated_7_cast_fp16 = concat(axis = var_40, interleave = rotated_7_interleave_0, values = (var_417_cast_fp16, var_420_cast_fp16))[name = string("rotated_7_cast_fp16")]; + tensor expand_dims_12 = const()[name = string("expand_dims_12"), val = tensor([9])]; + tensor expand_dims_13 = const()[name = string("expand_dims_13"), val = tensor([0])]; + tensor expand_dims_15 = const()[name = string("expand_dims_15"), val = tensor([0])]; + tensor expand_dims_16 = const()[name = string("expand_dims_16"), val = tensor([10])]; + int32 concat_10_axis_0 = const()[name = string("concat_10_axis_0"), val = int32(0)]; + bool concat_10_interleave_0 = const()[name = string("concat_10_interleave_0"), val = bool(false)]; + tensor concat_10 = concat(axis = concat_10_axis_0, interleave = concat_10_interleave_0, values = (expand_dims_12, expand_dims_13, current_pos, expand_dims_15))[name = string("concat_10")]; + tensor concat_11_values1_0 = const()[name = string("concat_11_values1_0"), val = tensor([0])]; + tensor concat_11_values3_0 = const()[name = string("concat_11_values3_0"), val = tensor([0])]; + int32 concat_11_axis_0 = const()[name = string("concat_11_axis_0"), val = int32(0)]; + bool concat_11_interleave_0 = const()[name = string("concat_11_interleave_0"), val = bool(false)]; + tensor concat_11 = concat(axis = concat_11_axis_0, interleave = concat_11_interleave_0, values = (expand_dims_16, concat_11_values1_0, var_240, concat_11_values3_0))[name = string("concat_11")]; + tensor model_model_kv_cache_0_internal_tensor_assign_3_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_3_stride_0"), val = tensor([1, 1, 1, 1])]; + tensor model_model_kv_cache_0_internal_tensor_assign_3_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_3_begin_mask_0"), val = tensor([false, false, false, false])]; + tensor model_model_kv_cache_0_internal_tensor_assign_3_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_3_end_mask_0"), val = tensor([false, true, false, true])]; + tensor model_model_kv_cache_0_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_3_squeeze_mask_0"), val = tensor([false, false, false, false])]; + tensor model_model_kv_cache_0_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_10, begin_mask = model_model_kv_cache_0_internal_tensor_assign_3_begin_mask_0, end = concat_11, end_mask = model_model_kv_cache_0_internal_tensor_assign_3_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_3_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_3_stride_0, update = rotated_7_cast_fp16, x = coreml_update_state_9)[name = string("model_model_kv_cache_0_internal_tensor_assign_3_cast_fp16")]; + write_state(data = model_model_kv_cache_0_internal_tensor_assign_3_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_2_write_state")]; + tensor coreml_update_state_10 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_2")]; + tensor expand_dims_18 = const()[name = string("expand_dims_18"), val = tensor([41])]; + tensor expand_dims_19 = const()[name = string("expand_dims_19"), val = tensor([0])]; + tensor expand_dims_21 = const()[name = string("expand_dims_21"), val = tensor([0])]; + tensor expand_dims_22 = const()[name = string("expand_dims_22"), val = tensor([42])]; + int32 concat_14_axis_0 = const()[name = string("concat_14_axis_0"), val = int32(0)]; + bool concat_14_interleave_0 = const()[name = string("concat_14_interleave_0"), val = bool(false)]; + tensor concat_14 = concat(axis = concat_14_axis_0, interleave = concat_14_interleave_0, values = (expand_dims_18, expand_dims_19, current_pos, expand_dims_21))[name = string("concat_14")]; + tensor concat_15_values1_0 = const()[name = string("concat_15_values1_0"), val = tensor([0])]; + tensor concat_15_values3_0 = const()[name = string("concat_15_values3_0"), val = tensor([0])]; + int32 concat_15_axis_0 = const()[name = string("concat_15_axis_0"), val = int32(0)]; + bool concat_15_interleave_0 = const()[name = string("concat_15_interleave_0"), val = bool(false)]; + tensor concat_15 = concat(axis = concat_15_axis_0, interleave = concat_15_interleave_0, values = (expand_dims_22, concat_15_values1_0, var_240, concat_15_values3_0))[name = string("concat_15")]; + tensor model_model_kv_cache_0_internal_tensor_assign_4_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_4_stride_0"), val = tensor([1, 1, 1, 1])]; + tensor model_model_kv_cache_0_internal_tensor_assign_4_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_4_begin_mask_0"), val = tensor([false, false, false, false])]; + tensor model_model_kv_cache_0_internal_tensor_assign_4_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_4_end_mask_0"), val = tensor([false, true, false, true])]; + tensor model_model_kv_cache_0_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_4_squeeze_mask_0"), val = tensor([false, false, false, false])]; + tensor model_model_kv_cache_0_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_14, begin_mask = model_model_kv_cache_0_internal_tensor_assign_4_begin_mask_0, end = concat_15, end_mask = model_model_kv_cache_0_internal_tensor_assign_4_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_4_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_4_stride_0, update = var_380, x = coreml_update_state_10)[name = string("model_model_kv_cache_0_internal_tensor_assign_4_cast_fp16")]; + write_state(data = model_model_kv_cache_0_internal_tensor_assign_4_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_3_write_state")]; + tensor coreml_update_state_11 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_3")]; + tensor var_440_begin_0 = const()[name = string("op_440_begin_0"), val = tensor([9, 0, 0, 0])]; + tensor var_440_end_0 = const()[name = string("op_440_end_0"), val = tensor([10, 8, 1024, 128])]; + tensor var_440_end_mask_0 = const()[name = string("op_440_end_mask_0"), val = tensor([false, true, true, true])]; + tensor var_440_cast_fp16 = slice_by_index(begin = var_440_begin_0, end = var_440_end_0, end_mask = var_440_end_mask_0, x = coreml_update_state_11)[name = string("op_440_cast_fp16")]; + tensor K_layer_cache_3_axes_0 = const()[name = string("K_layer_cache_3_axes_0"), val = tensor([0])]; + tensor K_layer_cache_3_cast_fp16 = squeeze(axes = K_layer_cache_3_axes_0, x = var_440_cast_fp16)[name = string("K_layer_cache_3_cast_fp16")]; + tensor var_442_begin_0 = const()[name = string("op_442_begin_0"), val = tensor([41, 0, 0, 0])]; + tensor var_442_end_0 = const()[name = string("op_442_end_0"), val = tensor([42, 8, 1024, 128])]; + tensor var_442_end_mask_0 = const()[name = string("op_442_end_mask_0"), val = tensor([false, true, true, true])]; + tensor var_442_cast_fp16 = slice_by_index(begin = var_442_begin_0, end = var_442_end_0, end_mask = var_442_end_mask_0, x = coreml_update_state_11)[name = string("op_442_cast_fp16")]; + tensor V_layer_cache_3_axes_0 = const()[name = string("V_layer_cache_3_axes_0"), val = tensor([0])]; + tensor V_layer_cache_3_cast_fp16 = squeeze(axes = V_layer_cache_3_axes_0, x = var_442_cast_fp16)[name = string("V_layer_cache_3_cast_fp16")]; + tensor x_39_axes_0 = const()[name = string("x_39_axes_0"), val = tensor([1])]; + tensor x_39_cast_fp16 = expand_dims(axes = x_39_axes_0, x = K_layer_cache_3_cast_fp16)[name = string("x_39_cast_fp16")]; + tensor var_451 = const()[name = string("op_451"), val = tensor([1, 4, 1, 1])]; + tensor x_41_cast_fp16 = tile(reps = var_451, x = x_39_cast_fp16)[name = string("x_41_cast_fp16")]; + tensor var_455 = const()[name = string("op_455"), val = tensor([1, -1, 1024, 128])]; + tensor key_states_7_cast_fp16 = reshape(shape = var_455, x = x_41_cast_fp16)[name = string("key_states_7_cast_fp16")]; + tensor x_45_axes_0 = const()[name = string("x_45_axes_0"), val = tensor([1])]; + tensor x_45_cast_fp16 = expand_dims(axes = x_45_axes_0, x = V_layer_cache_3_cast_fp16)[name = string("x_45_cast_fp16")]; + tensor var_458 = const()[name = string("op_458"), val = tensor([1, 4, 1, 1])]; + tensor x_47_cast_fp16 = tile(reps = var_458, x = x_45_cast_fp16)[name = string("x_47_cast_fp16")]; + tensor var_462 = const()[name = string("op_462"), val = tensor([1, -1, 1024, 128])]; + tensor value_states_7_cast_fp16 = reshape(shape = var_462, x = x_47_cast_fp16)[name = string("value_states_7_cast_fp16")]; + bool var_465_transpose_x_1 = const()[name = string("op_465_transpose_x_1"), val = bool(false)]; + bool var_465_transpose_y_1 = const()[name = string("op_465_transpose_y_1"), val = bool(true)]; + tensor var_465_cast_fp16 = matmul(transpose_x = var_465_transpose_x_1, transpose_y = var_465_transpose_y_1, x = rotated_5_cast_fp16, y = key_states_7_cast_fp16)[name = string("op_465_cast_fp16")]; + fp16 var_466_to_fp16 = const()[name = string("op_466_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor attn_weights_5_cast_fp16 = mul(x = var_465_cast_fp16, y = var_466_to_fp16)[name = string("attn_weights_5_cast_fp16")]; + tensor x_49_cast_fp16 = add(x = attn_weights_5_cast_fp16, y = causal_mask)[name = string("x_49_cast_fp16")]; + tensor reduce_max_1_axes_0 = const()[name = string("reduce_max_1_axes_0"), val = tensor([-1])]; + bool reduce_max_1_keep_dims_0 = const()[name = string("reduce_max_1_keep_dims_0"), val = bool(true)]; + tensor reduce_max_1_cast_fp16 = reduce_max(axes = reduce_max_1_axes_0, keep_dims = reduce_max_1_keep_dims_0, x = x_49_cast_fp16)[name = string("reduce_max_1_cast_fp16")]; + tensor x_51_cast_fp16 = sub(x = x_49_cast_fp16, y = reduce_max_1_cast_fp16)[name = string("x_51_cast_fp16")]; + tensor exp_x_3_cast_fp16 = exp(x = x_51_cast_fp16)[name = string("exp_x_3_cast_fp16")]; + tensor var_477_axes_0 = const()[name = string("op_477_axes_0"), val = tensor([-1])]; + bool var_477_keep_dims_0 = const()[name = string("op_477_keep_dims_0"), val = bool(true)]; + tensor var_477_cast_fp16 = reduce_sum(axes = var_477_axes_0, keep_dims = var_477_keep_dims_0, x = exp_x_3_cast_fp16)[name = string("op_477_cast_fp16")]; + tensor attn_weights_7_cast_fp16 = real_div(x = exp_x_3_cast_fp16, y = var_477_cast_fp16)[name = string("attn_weights_7_cast_fp16")]; + bool attn_output_7_transpose_x_0 = const()[name = string("attn_output_7_transpose_x_0"), val = bool(false)]; + bool attn_output_7_transpose_y_0 = const()[name = string("attn_output_7_transpose_y_0"), val = bool(false)]; + tensor attn_output_7_cast_fp16 = matmul(transpose_x = attn_output_7_transpose_x_0, transpose_y = attn_output_7_transpose_y_0, x = attn_weights_7_cast_fp16, y = value_states_7_cast_fp16)[name = string("attn_output_7_cast_fp16")]; + tensor var_480_perm_0 = const()[name = string("op_480_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_482 = const()[name = string("op_482"), val = tensor([1, 1, 4096])]; + tensor var_480_cast_fp16 = transpose(perm = var_480_perm_0, x = attn_output_7_cast_fp16)[name = string("transpose_10")]; + tensor input_19_cast_fp16 = reshape(shape = var_482, x = var_480_cast_fp16)[name = string("input_19_cast_fp16")]; + tensor model_model_layers_9_self_attn_o_proj_weight_promoted_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(686263872))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(698846848))))[name = string("model_model_layers_9_self_attn_o_proj_weight_promoted_to_fp16_palettized")]; + tensor linear_1_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_9_self_attn_o_proj_weight_promoted_to_fp16_palettized, x = input_19_cast_fp16)[name = string("linear_1_cast_fp16")]; + tensor hidden_states_13_cast_fp16 = add(x = hidden_states_9_cast_fp16, y = linear_1_cast_fp16)[name = string("hidden_states_13_cast_fp16")]; + tensor mean_7_axes_0 = const()[name = string("mean_7_axes_0"), val = tensor([-1])]; + bool mean_7_keep_dims_0 = const()[name = string("mean_7_keep_dims_0"), val = bool(true)]; + tensor mean_7_cast_fp16 = reduce_mean(axes = mean_7_axes_0, keep_dims = mean_7_keep_dims_0, x = hidden_states_13_cast_fp16)[name = string("mean_7_cast_fp16")]; + tensor input_21_cast_fp16 = sub(x = hidden_states_13_cast_fp16, y = mean_7_cast_fp16)[name = string("input_21_cast_fp16")]; + tensor var_493_axes_0 = const()[name = string("op_493_axes_0"), val = tensor([-1])]; + tensor model_model_layers_9_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_9_post_attention_layernorm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(698912448)))]; + tensor var_493_cast_fp16 = layer_norm(axes = var_493_axes_0, epsilon = var_35_to_fp16, gamma = model_model_layers_9_post_attention_layernorm_weight_to_fp16, x = input_21_cast_fp16)[name = string("op_493_cast_fp16")]; + tensor var_500 = const()[name = string("op_500"), val = tensor([0, 2, 1])]; + tensor input_23_axes_0 = const()[name = string("input_23_axes_0"), val = tensor([2])]; + tensor var_501 = transpose(perm = var_500, x = var_493_cast_fp16)[name = string("transpose_9")]; + tensor input_23 = expand_dims(axes = input_23_axes_0, x = var_501)[name = string("input_23")]; + string input_25_pad_type_0 = const()[name = string("input_25_pad_type_0"), val = string("valid")]; + tensor input_25_strides_0 = const()[name = string("input_25_strides_0"), val = tensor([1, 1])]; + tensor input_25_pad_0 = const()[name = string("input_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_25_dilations_0 = const()[name = string("input_25_dilations_0"), val = tensor([1, 1])]; + int32 input_25_groups_0 = const()[name = string("input_25_groups_0"), val = int32(1)]; + tensor input_25 = conv(dilations = input_25_dilations_0, groups = input_25_groups_0, pad = input_25_pad_0, pad_type = input_25_pad_type_0, strides = input_25_strides_0, weight = model_model_layers_9_mlp_gate_proj_weight_palettized, x = input_23)[name = string("input_25")]; + string up_states_3_pad_type_0 = const()[name = string("up_states_3_pad_type_0"), val = string("valid")]; + tensor up_states_3_strides_0 = const()[name = string("up_states_3_strides_0"), val = tensor([1, 1])]; + tensor up_states_3_pad_0 = const()[name = string("up_states_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor up_states_3_dilations_0 = const()[name = string("up_states_3_dilations_0"), val = tensor([1, 1])]; + int32 up_states_3_groups_0 = const()[name = string("up_states_3_groups_0"), val = int32(1)]; + tensor up_states_3 = conv(dilations = up_states_3_dilations_0, groups = up_states_3_groups_0, pad = up_states_3_pad_0, pad_type = up_states_3_pad_type_0, strides = up_states_3_strides_0, weight = model_model_layers_9_mlp_up_proj_weight_palettized, x = input_23)[name = string("up_states_3")]; + tensor gate_states_3 = silu(x = input_25)[name = string("gate_states_3")]; + tensor input_27 = mul(x = gate_states_3, y = up_states_3)[name = string("input_27")]; + string hidden_states_15_pad_type_0 = const()[name = string("hidden_states_15_pad_type_0"), val = string("valid")]; + tensor hidden_states_15_strides_0 = const()[name = string("hidden_states_15_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_15_pad_0 = const()[name = string("hidden_states_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_15_dilations_0 = const()[name = string("hidden_states_15_dilations_0"), val = tensor([1, 1])]; + int32 hidden_states_15_groups_0 = const()[name = string("hidden_states_15_groups_0"), val = int32(1)]; + tensor hidden_states_15 = conv(dilations = hidden_states_15_dilations_0, groups = hidden_states_15_groups_0, pad = hidden_states_15_pad_0, pad_type = hidden_states_15_pad_type_0, strides = hidden_states_15_strides_0, weight = model_model_layers_9_mlp_down_proj_weight_palettized, x = input_27)[name = string("hidden_states_15")]; + tensor var_523_axes_0 = const()[name = string("op_523_axes_0"), val = tensor([2])]; + tensor var_523 = squeeze(axes = var_523_axes_0, x = hidden_states_15)[name = string("op_523")]; + tensor var_524 = const()[name = string("op_524"), val = tensor([0, 2, 1])]; + tensor var_525 = transpose(perm = var_524, x = var_523)[name = string("transpose_8")]; + tensor hidden_states_17_cast_fp16 = add(x = hidden_states_13_cast_fp16, y = var_525)[name = string("hidden_states_17_cast_fp16")]; + tensor mean_9_axes_0 = const()[name = string("mean_9_axes_0"), val = tensor([-1])]; + bool mean_9_keep_dims_0 = const()[name = string("mean_9_keep_dims_0"), val = bool(true)]; + tensor mean_9_cast_fp16 = reduce_mean(axes = mean_9_axes_0, keep_dims = mean_9_keep_dims_0, x = hidden_states_17_cast_fp16)[name = string("mean_9_cast_fp16")]; + tensor input_29_cast_fp16 = sub(x = hidden_states_17_cast_fp16, y = mean_9_cast_fp16)[name = string("input_29_cast_fp16")]; + tensor var_533_axes_0 = const()[name = string("op_533_axes_0"), val = tensor([-1])]; + tensor model_model_layers_10_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_10_input_layernorm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(698920704)))]; + tensor var_533_cast_fp16 = layer_norm(axes = var_533_axes_0, epsilon = var_35_to_fp16, gamma = model_model_layers_10_input_layernorm_weight_to_fp16, x = input_29_cast_fp16)[name = string("op_533_cast_fp16")]; + tensor var_536 = const()[name = string("op_536"), val = tensor([0, 2, 1])]; + tensor var_538_axes_0 = const()[name = string("op_538_axes_0"), val = tensor([2])]; + tensor var_537 = transpose(perm = var_536, x = var_533_cast_fp16)[name = string("transpose_7")]; + tensor var_538 = expand_dims(axes = var_538_axes_0, x = var_537)[name = string("op_538")]; + string var_545_pad_type_0 = const()[name = string("op_545_pad_type_0"), val = string("valid")]; + tensor var_545_strides_0 = const()[name = string("op_545_strides_0"), val = tensor([1, 1])]; + tensor var_545_pad_0 = const()[name = string("op_545_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_545_dilations_0 = const()[name = string("op_545_dilations_0"), val = tensor([1, 1])]; + int32 var_545_groups_0 = const()[name = string("op_545_groups_0"), val = int32(1)]; + tensor var_545 = conv(dilations = var_545_dilations_0, groups = var_545_groups_0, pad = var_545_pad_0, pad_type = var_545_pad_type_0, strides = var_545_strides_0, weight = model_model_layers_10_self_attn_q_proj_weight_palettized, x = var_538)[name = string("op_545")]; + tensor var_546 = const()[name = string("op_546"), val = tensor([1, 32, 1, 128])]; + tensor var_547 = reshape(shape = var_546, x = var_545)[name = string("op_547")]; + string var_554_pad_type_0 = const()[name = string("op_554_pad_type_0"), val = string("valid")]; + tensor var_554_strides_0 = const()[name = string("op_554_strides_0"), val = tensor([1, 1])]; + tensor var_554_pad_0 = const()[name = string("op_554_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_554_dilations_0 = const()[name = string("op_554_dilations_0"), val = tensor([1, 1])]; + int32 var_554_groups_0 = const()[name = string("op_554_groups_0"), val = int32(1)]; + tensor var_554 = conv(dilations = var_554_dilations_0, groups = var_554_groups_0, pad = var_554_pad_0, pad_type = var_554_pad_type_0, strides = var_554_strides_0, weight = model_model_layers_10_self_attn_k_proj_weight_palettized, x = var_538)[name = string("op_554")]; + tensor var_555 = const()[name = string("op_555"), val = tensor([1, 8, 1, 128])]; + tensor var_556 = reshape(shape = var_555, x = var_554)[name = string("op_556")]; + string var_563_pad_type_0 = const()[name = string("op_563_pad_type_0"), val = string("valid")]; + tensor var_563_strides_0 = const()[name = string("op_563_strides_0"), val = tensor([1, 1])]; + tensor var_563_pad_0 = const()[name = string("op_563_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_563_dilations_0 = const()[name = string("op_563_dilations_0"), val = tensor([1, 1])]; + int32 var_563_groups_0 = const()[name = string("op_563_groups_0"), val = int32(1)]; + tensor var_563 = conv(dilations = var_563_dilations_0, groups = var_563_groups_0, pad = var_563_pad_0, pad_type = var_563_pad_type_0, strides = var_563_strides_0, weight = model_model_layers_10_self_attn_v_proj_weight_palettized, x = var_538)[name = string("op_563")]; + tensor var_564 = const()[name = string("op_564"), val = tensor([1, 8, 1, 128])]; + tensor var_565 = reshape(shape = var_564, x = var_563)[name = string("op_565")]; + tensor x1_9_begin_0 = const()[name = string("x1_9_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_9_end_0 = const()[name = string("x1_9_end_0"), val = tensor([1, 32, 1, 64])]; + tensor x1_9_end_mask_0 = const()[name = string("x1_9_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_9 = slice_by_index(begin = x1_9_begin_0, end = x1_9_end_0, end_mask = x1_9_end_mask_0, x = var_547)[name = string("x1_9")]; + tensor x2_9_begin_0 = const()[name = string("x2_9_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_9_end_0 = const()[name = string("x2_9_end_0"), val = tensor([1, 32, 1, 128])]; + tensor x2_9_end_mask_0 = const()[name = string("x2_9_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_9 = slice_by_index(begin = x2_9_begin_0, end = x2_9_end_0, end_mask = x2_9_end_mask_0, x = var_547)[name = string("x2_9")]; + tensor var_579_cast_fp16 = mul(x = x1_9, y = cos_3_cast_fp16)[name = string("op_579_cast_fp16")]; + tensor var_580_cast_fp16 = mul(x = x2_9, y = sin_3_cast_fp16)[name = string("op_580_cast_fp16")]; + tensor var_581_cast_fp16 = sub(x = var_579_cast_fp16, y = var_580_cast_fp16)[name = string("op_581_cast_fp16")]; + tensor var_582_cast_fp16 = mul(x = x2_9, y = cos_3_cast_fp16)[name = string("op_582_cast_fp16")]; + tensor var_583_cast_fp16 = mul(x = x1_9, y = sin_3_cast_fp16)[name = string("op_583_cast_fp16")]; + tensor var_584_cast_fp16 = add(x = var_582_cast_fp16, y = var_583_cast_fp16)[name = string("op_584_cast_fp16")]; + bool rotated_9_interleave_0 = const()[name = string("rotated_9_interleave_0"), val = bool(false)]; + tensor rotated_9_cast_fp16 = concat(axis = var_40, interleave = rotated_9_interleave_0, values = (var_581_cast_fp16, var_584_cast_fp16))[name = string("rotated_9_cast_fp16")]; + tensor x1_11_begin_0 = const()[name = string("x1_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_11_end_0 = const()[name = string("x1_11_end_0"), val = tensor([1, 8, 1, 64])]; + tensor x1_11_end_mask_0 = const()[name = string("x1_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_11 = slice_by_index(begin = x1_11_begin_0, end = x1_11_end_0, end_mask = x1_11_end_mask_0, x = var_556)[name = string("x1_11")]; + tensor x2_11_begin_0 = const()[name = string("x2_11_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_11_end_0 = const()[name = string("x2_11_end_0"), val = tensor([1, 8, 1, 128])]; + tensor x2_11_end_mask_0 = const()[name = string("x2_11_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_11 = slice_by_index(begin = x2_11_begin_0, end = x2_11_end_0, end_mask = x2_11_end_mask_0, x = var_556)[name = string("x2_11")]; + tensor var_600_cast_fp16 = mul(x = x1_11, y = cos_3_cast_fp16)[name = string("op_600_cast_fp16")]; + tensor var_601_cast_fp16 = mul(x = x2_11, y = sin_3_cast_fp16)[name = string("op_601_cast_fp16")]; + tensor var_602_cast_fp16 = sub(x = var_600_cast_fp16, y = var_601_cast_fp16)[name = string("op_602_cast_fp16")]; + tensor var_603_cast_fp16 = mul(x = x2_11, y = cos_3_cast_fp16)[name = string("op_603_cast_fp16")]; + tensor var_604_cast_fp16 = mul(x = x1_11, y = sin_3_cast_fp16)[name = string("op_604_cast_fp16")]; + tensor var_605_cast_fp16 = add(x = var_603_cast_fp16, y = var_604_cast_fp16)[name = string("op_605_cast_fp16")]; + bool rotated_11_interleave_0 = const()[name = string("rotated_11_interleave_0"), val = bool(false)]; + tensor rotated_11_cast_fp16 = concat(axis = var_40, interleave = rotated_11_interleave_0, values = (var_602_cast_fp16, var_605_cast_fp16))[name = string("rotated_11_cast_fp16")]; + tensor expand_dims_24 = const()[name = string("expand_dims_24"), val = tensor([10])]; + tensor expand_dims_25 = const()[name = string("expand_dims_25"), val = tensor([0])]; + tensor expand_dims_27 = const()[name = string("expand_dims_27"), val = tensor([0])]; + tensor expand_dims_28 = const()[name = string("expand_dims_28"), val = tensor([11])]; + int32 concat_18_axis_0 = const()[name = string("concat_18_axis_0"), val = int32(0)]; + bool concat_18_interleave_0 = const()[name = string("concat_18_interleave_0"), val = bool(false)]; + tensor concat_18 = concat(axis = concat_18_axis_0, interleave = concat_18_interleave_0, values = (expand_dims_24, expand_dims_25, current_pos, expand_dims_27))[name = string("concat_18")]; + tensor concat_19_values1_0 = const()[name = string("concat_19_values1_0"), val = tensor([0])]; + tensor concat_19_values3_0 = const()[name = string("concat_19_values3_0"), val = tensor([0])]; + int32 concat_19_axis_0 = const()[name = string("concat_19_axis_0"), val = int32(0)]; + bool concat_19_interleave_0 = const()[name = string("concat_19_interleave_0"), val = bool(false)]; + tensor concat_19 = concat(axis = concat_19_axis_0, interleave = concat_19_interleave_0, values = (expand_dims_28, concat_19_values1_0, var_240, concat_19_values3_0))[name = string("concat_19")]; + tensor model_model_kv_cache_0_internal_tensor_assign_5_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_5_stride_0"), val = tensor([1, 1, 1, 1])]; + tensor model_model_kv_cache_0_internal_tensor_assign_5_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_5_begin_mask_0"), val = tensor([false, false, false, false])]; + tensor model_model_kv_cache_0_internal_tensor_assign_5_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_5_end_mask_0"), val = tensor([false, true, false, true])]; + tensor model_model_kv_cache_0_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_5_squeeze_mask_0"), val = tensor([false, false, false, false])]; + tensor model_model_kv_cache_0_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_18, begin_mask = model_model_kv_cache_0_internal_tensor_assign_5_begin_mask_0, end = concat_19, end_mask = model_model_kv_cache_0_internal_tensor_assign_5_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_5_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_5_stride_0, update = rotated_11_cast_fp16, x = coreml_update_state_11)[name = string("model_model_kv_cache_0_internal_tensor_assign_5_cast_fp16")]; + write_state(data = model_model_kv_cache_0_internal_tensor_assign_5_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_4_write_state")]; + tensor coreml_update_state_12 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_4")]; + tensor expand_dims_30 = const()[name = string("expand_dims_30"), val = tensor([42])]; + tensor expand_dims_31 = const()[name = string("expand_dims_31"), val = tensor([0])]; + tensor expand_dims_33 = const()[name = string("expand_dims_33"), val = tensor([0])]; + tensor expand_dims_34 = const()[name = string("expand_dims_34"), val = tensor([43])]; + int32 concat_22_axis_0 = const()[name = string("concat_22_axis_0"), val = int32(0)]; + bool concat_22_interleave_0 = const()[name = string("concat_22_interleave_0"), val = bool(false)]; + tensor concat_22 = concat(axis = concat_22_axis_0, interleave = concat_22_interleave_0, values = (expand_dims_30, expand_dims_31, current_pos, expand_dims_33))[name = string("concat_22")]; + tensor concat_23_values1_0 = const()[name = string("concat_23_values1_0"), val = tensor([0])]; + tensor concat_23_values3_0 = const()[name = string("concat_23_values3_0"), val = tensor([0])]; + int32 concat_23_axis_0 = const()[name = string("concat_23_axis_0"), val = int32(0)]; + bool concat_23_interleave_0 = const()[name = string("concat_23_interleave_0"), val = bool(false)]; + tensor concat_23 = concat(axis = concat_23_axis_0, interleave = concat_23_interleave_0, values = (expand_dims_34, concat_23_values1_0, var_240, concat_23_values3_0))[name = string("concat_23")]; + tensor model_model_kv_cache_0_internal_tensor_assign_6_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_6_stride_0"), val = tensor([1, 1, 1, 1])]; + tensor model_model_kv_cache_0_internal_tensor_assign_6_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_6_begin_mask_0"), val = tensor([false, false, false, false])]; + tensor model_model_kv_cache_0_internal_tensor_assign_6_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_6_end_mask_0"), val = tensor([false, true, false, true])]; + tensor model_model_kv_cache_0_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_6_squeeze_mask_0"), val = tensor([false, false, false, false])]; + tensor model_model_kv_cache_0_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_22, begin_mask = model_model_kv_cache_0_internal_tensor_assign_6_begin_mask_0, end = concat_23, end_mask = model_model_kv_cache_0_internal_tensor_assign_6_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_6_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_6_stride_0, update = var_565, x = coreml_update_state_12)[name = string("model_model_kv_cache_0_internal_tensor_assign_6_cast_fp16")]; + write_state(data = model_model_kv_cache_0_internal_tensor_assign_6_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_5_write_state")]; + tensor coreml_update_state_13 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_5")]; + tensor var_625_begin_0 = const()[name = string("op_625_begin_0"), val = tensor([10, 0, 0, 0])]; + tensor var_625_end_0 = const()[name = string("op_625_end_0"), val = tensor([11, 8, 1024, 128])]; + tensor var_625_end_mask_0 = const()[name = string("op_625_end_mask_0"), val = tensor([false, true, true, true])]; + tensor var_625_cast_fp16 = slice_by_index(begin = var_625_begin_0, end = var_625_end_0, end_mask = var_625_end_mask_0, x = coreml_update_state_13)[name = string("op_625_cast_fp16")]; + tensor K_layer_cache_5_axes_0 = const()[name = string("K_layer_cache_5_axes_0"), val = tensor([0])]; + tensor K_layer_cache_5_cast_fp16 = squeeze(axes = K_layer_cache_5_axes_0, x = var_625_cast_fp16)[name = string("K_layer_cache_5_cast_fp16")]; + tensor var_627_begin_0 = const()[name = string("op_627_begin_0"), val = tensor([42, 0, 0, 0])]; + tensor var_627_end_0 = const()[name = string("op_627_end_0"), val = tensor([43, 8, 1024, 128])]; + tensor var_627_end_mask_0 = const()[name = string("op_627_end_mask_0"), val = tensor([false, true, true, true])]; + tensor var_627_cast_fp16 = slice_by_index(begin = var_627_begin_0, end = var_627_end_0, end_mask = var_627_end_mask_0, x = coreml_update_state_13)[name = string("op_627_cast_fp16")]; + tensor V_layer_cache_5_axes_0 = const()[name = string("V_layer_cache_5_axes_0"), val = tensor([0])]; + tensor V_layer_cache_5_cast_fp16 = squeeze(axes = V_layer_cache_5_axes_0, x = var_627_cast_fp16)[name = string("V_layer_cache_5_cast_fp16")]; + tensor x_67_axes_0 = const()[name = string("x_67_axes_0"), val = tensor([1])]; + tensor x_67_cast_fp16 = expand_dims(axes = x_67_axes_0, x = K_layer_cache_5_cast_fp16)[name = string("x_67_cast_fp16")]; + tensor var_636 = const()[name = string("op_636"), val = tensor([1, 4, 1, 1])]; + tensor x_69_cast_fp16 = tile(reps = var_636, x = x_67_cast_fp16)[name = string("x_69_cast_fp16")]; + tensor var_640 = const()[name = string("op_640"), val = tensor([1, -1, 1024, 128])]; + tensor key_states_11_cast_fp16 = reshape(shape = var_640, x = x_69_cast_fp16)[name = string("key_states_11_cast_fp16")]; + tensor x_73_axes_0 = const()[name = string("x_73_axes_0"), val = tensor([1])]; + tensor x_73_cast_fp16 = expand_dims(axes = x_73_axes_0, x = V_layer_cache_5_cast_fp16)[name = string("x_73_cast_fp16")]; + tensor var_643 = const()[name = string("op_643"), val = tensor([1, 4, 1, 1])]; + tensor x_75_cast_fp16 = tile(reps = var_643, x = x_73_cast_fp16)[name = string("x_75_cast_fp16")]; + tensor var_647 = const()[name = string("op_647"), val = tensor([1, -1, 1024, 128])]; + tensor value_states_11_cast_fp16 = reshape(shape = var_647, x = x_75_cast_fp16)[name = string("value_states_11_cast_fp16")]; + bool var_650_transpose_x_1 = const()[name = string("op_650_transpose_x_1"), val = bool(false)]; + bool var_650_transpose_y_1 = const()[name = string("op_650_transpose_y_1"), val = bool(true)]; + tensor var_650_cast_fp16 = matmul(transpose_x = var_650_transpose_x_1, transpose_y = var_650_transpose_y_1, x = rotated_9_cast_fp16, y = key_states_11_cast_fp16)[name = string("op_650_cast_fp16")]; + fp16 var_651_to_fp16 = const()[name = string("op_651_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor attn_weights_9_cast_fp16 = mul(x = var_650_cast_fp16, y = var_651_to_fp16)[name = string("attn_weights_9_cast_fp16")]; + tensor x_77_cast_fp16 = add(x = attn_weights_9_cast_fp16, y = causal_mask)[name = string("x_77_cast_fp16")]; + tensor reduce_max_2_axes_0 = const()[name = string("reduce_max_2_axes_0"), val = tensor([-1])]; + bool reduce_max_2_keep_dims_0 = const()[name = string("reduce_max_2_keep_dims_0"), val = bool(true)]; + tensor reduce_max_2_cast_fp16 = reduce_max(axes = reduce_max_2_axes_0, keep_dims = reduce_max_2_keep_dims_0, x = x_77_cast_fp16)[name = string("reduce_max_2_cast_fp16")]; + tensor x_79_cast_fp16 = sub(x = x_77_cast_fp16, y = reduce_max_2_cast_fp16)[name = string("x_79_cast_fp16")]; + tensor exp_x_5_cast_fp16 = exp(x = x_79_cast_fp16)[name = string("exp_x_5_cast_fp16")]; + tensor var_662_axes_0 = const()[name = string("op_662_axes_0"), val = tensor([-1])]; + bool var_662_keep_dims_0 = const()[name = string("op_662_keep_dims_0"), val = bool(true)]; + tensor var_662_cast_fp16 = reduce_sum(axes = var_662_axes_0, keep_dims = var_662_keep_dims_0, x = exp_x_5_cast_fp16)[name = string("op_662_cast_fp16")]; + tensor attn_weights_11_cast_fp16 = real_div(x = exp_x_5_cast_fp16, y = var_662_cast_fp16)[name = string("attn_weights_11_cast_fp16")]; + bool attn_output_13_transpose_x_0 = const()[name = string("attn_output_13_transpose_x_0"), val = bool(false)]; + bool attn_output_13_transpose_y_0 = const()[name = string("attn_output_13_transpose_y_0"), val = bool(false)]; + tensor attn_output_13_cast_fp16 = matmul(transpose_x = attn_output_13_transpose_x_0, transpose_y = attn_output_13_transpose_y_0, x = attn_weights_11_cast_fp16, y = value_states_11_cast_fp16)[name = string("attn_output_13_cast_fp16")]; + tensor var_665_perm_0 = const()[name = string("op_665_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_667 = const()[name = string("op_667"), val = tensor([1, 1, 4096])]; + tensor var_665_cast_fp16 = transpose(perm = var_665_perm_0, x = attn_output_13_cast_fp16)[name = string("transpose_6")]; + tensor input_33_cast_fp16 = reshape(shape = var_667, x = var_665_cast_fp16)[name = string("input_33_cast_fp16")]; + tensor model_model_layers_10_self_attn_o_proj_weight_promoted_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(698928960))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(711511936))))[name = string("model_model_layers_10_self_attn_o_proj_weight_promoted_to_fp16_palettized")]; + tensor linear_2_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_10_self_attn_o_proj_weight_promoted_to_fp16_palettized, x = input_33_cast_fp16)[name = string("linear_2_cast_fp16")]; + tensor hidden_states_21_cast_fp16 = add(x = hidden_states_17_cast_fp16, y = linear_2_cast_fp16)[name = string("hidden_states_21_cast_fp16")]; + tensor mean_11_axes_0 = const()[name = string("mean_11_axes_0"), val = tensor([-1])]; + bool mean_11_keep_dims_0 = const()[name = string("mean_11_keep_dims_0"), val = bool(true)]; + tensor mean_11_cast_fp16 = reduce_mean(axes = mean_11_axes_0, keep_dims = mean_11_keep_dims_0, x = hidden_states_21_cast_fp16)[name = string("mean_11_cast_fp16")]; + tensor input_35_cast_fp16 = sub(x = hidden_states_21_cast_fp16, y = mean_11_cast_fp16)[name = string("input_35_cast_fp16")]; + tensor var_678_axes_0 = const()[name = string("op_678_axes_0"), val = tensor([-1])]; + tensor model_model_layers_10_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_10_post_attention_layernorm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(711577536)))]; + tensor var_678_cast_fp16 = layer_norm(axes = var_678_axes_0, epsilon = var_35_to_fp16, gamma = model_model_layers_10_post_attention_layernorm_weight_to_fp16, x = input_35_cast_fp16)[name = string("op_678_cast_fp16")]; + tensor var_685 = const()[name = string("op_685"), val = tensor([0, 2, 1])]; + tensor input_37_axes_0 = const()[name = string("input_37_axes_0"), val = tensor([2])]; + tensor var_686 = transpose(perm = var_685, x = var_678_cast_fp16)[name = string("transpose_5")]; + tensor input_37 = expand_dims(axes = input_37_axes_0, x = var_686)[name = string("input_37")]; + string input_39_pad_type_0 = const()[name = string("input_39_pad_type_0"), val = string("valid")]; + tensor input_39_strides_0 = const()[name = string("input_39_strides_0"), val = tensor([1, 1])]; + tensor input_39_pad_0 = const()[name = string("input_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_39_dilations_0 = const()[name = string("input_39_dilations_0"), val = tensor([1, 1])]; + int32 input_39_groups_0 = const()[name = string("input_39_groups_0"), val = int32(1)]; + tensor input_39 = conv(dilations = input_39_dilations_0, groups = input_39_groups_0, pad = input_39_pad_0, pad_type = input_39_pad_type_0, strides = input_39_strides_0, weight = model_model_layers_10_mlp_gate_proj_weight_palettized, x = input_37)[name = string("input_39")]; + string up_states_5_pad_type_0 = const()[name = string("up_states_5_pad_type_0"), val = string("valid")]; + tensor up_states_5_strides_0 = const()[name = string("up_states_5_strides_0"), val = tensor([1, 1])]; + tensor up_states_5_pad_0 = const()[name = string("up_states_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor up_states_5_dilations_0 = const()[name = string("up_states_5_dilations_0"), val = tensor([1, 1])]; + int32 up_states_5_groups_0 = const()[name = string("up_states_5_groups_0"), val = int32(1)]; + tensor up_states_5 = conv(dilations = up_states_5_dilations_0, groups = up_states_5_groups_0, pad = up_states_5_pad_0, pad_type = up_states_5_pad_type_0, strides = up_states_5_strides_0, weight = model_model_layers_10_mlp_up_proj_weight_palettized, x = input_37)[name = string("up_states_5")]; + tensor gate_states_5 = silu(x = input_39)[name = string("gate_states_5")]; + tensor input_41 = mul(x = gate_states_5, y = up_states_5)[name = string("input_41")]; + string hidden_states_23_pad_type_0 = const()[name = string("hidden_states_23_pad_type_0"), val = string("valid")]; + tensor hidden_states_23_strides_0 = const()[name = string("hidden_states_23_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_23_pad_0 = const()[name = string("hidden_states_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_23_dilations_0 = const()[name = string("hidden_states_23_dilations_0"), val = tensor([1, 1])]; + int32 hidden_states_23_groups_0 = const()[name = string("hidden_states_23_groups_0"), val = int32(1)]; + tensor hidden_states_23 = conv(dilations = hidden_states_23_dilations_0, groups = hidden_states_23_groups_0, pad = hidden_states_23_pad_0, pad_type = hidden_states_23_pad_type_0, strides = hidden_states_23_strides_0, weight = model_model_layers_10_mlp_down_proj_weight_palettized, x = input_41)[name = string("hidden_states_23")]; + tensor var_708_axes_0 = const()[name = string("op_708_axes_0"), val = tensor([2])]; + tensor var_708 = squeeze(axes = var_708_axes_0, x = hidden_states_23)[name = string("op_708")]; + tensor var_709 = const()[name = string("op_709"), val = tensor([0, 2, 1])]; + tensor var_710 = transpose(perm = var_709, x = var_708)[name = string("transpose_4")]; + tensor hidden_states_25_cast_fp16 = add(x = hidden_states_21_cast_fp16, y = var_710)[name = string("hidden_states_25_cast_fp16")]; + tensor mean_13_axes_0 = const()[name = string("mean_13_axes_0"), val = tensor([-1])]; + bool mean_13_keep_dims_0 = const()[name = string("mean_13_keep_dims_0"), val = bool(true)]; + tensor mean_13_cast_fp16 = reduce_mean(axes = mean_13_axes_0, keep_dims = mean_13_keep_dims_0, x = hidden_states_25_cast_fp16)[name = string("mean_13_cast_fp16")]; + tensor input_43_cast_fp16 = sub(x = hidden_states_25_cast_fp16, y = mean_13_cast_fp16)[name = string("input_43_cast_fp16")]; + tensor var_718_axes_0 = const()[name = string("op_718_axes_0"), val = tensor([-1])]; + tensor model_model_layers_11_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_11_input_layernorm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(711585792)))]; + tensor var_718_cast_fp16 = layer_norm(axes = var_718_axes_0, epsilon = var_35_to_fp16, gamma = model_model_layers_11_input_layernorm_weight_to_fp16, x = input_43_cast_fp16)[name = string("op_718_cast_fp16")]; + tensor var_721 = const()[name = string("op_721"), val = tensor([0, 2, 1])]; + tensor var_723_axes_0 = const()[name = string("op_723_axes_0"), val = tensor([2])]; + tensor var_722 = transpose(perm = var_721, x = var_718_cast_fp16)[name = string("transpose_3")]; + tensor var_723 = expand_dims(axes = var_723_axes_0, x = var_722)[name = string("op_723")]; + string var_730_pad_type_0 = const()[name = string("op_730_pad_type_0"), val = string("valid")]; + tensor var_730_strides_0 = const()[name = string("op_730_strides_0"), val = tensor([1, 1])]; + tensor var_730_pad_0 = const()[name = string("op_730_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_730_dilations_0 = const()[name = string("op_730_dilations_0"), val = tensor([1, 1])]; + int32 var_730_groups_0 = const()[name = string("op_730_groups_0"), val = int32(1)]; + tensor var_730 = conv(dilations = var_730_dilations_0, groups = var_730_groups_0, pad = var_730_pad_0, pad_type = var_730_pad_type_0, strides = var_730_strides_0, weight = model_model_layers_11_self_attn_q_proj_weight_palettized, x = var_723)[name = string("op_730")]; + tensor var_731 = const()[name = string("op_731"), val = tensor([1, 32, 1, 128])]; + tensor var_732 = reshape(shape = var_731, x = var_730)[name = string("op_732")]; + string var_739_pad_type_0 = const()[name = string("op_739_pad_type_0"), val = string("valid")]; + tensor var_739_strides_0 = const()[name = string("op_739_strides_0"), val = tensor([1, 1])]; + tensor var_739_pad_0 = const()[name = string("op_739_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_739_dilations_0 = const()[name = string("op_739_dilations_0"), val = tensor([1, 1])]; + int32 var_739_groups_0 = const()[name = string("op_739_groups_0"), val = int32(1)]; + tensor var_739 = conv(dilations = var_739_dilations_0, groups = var_739_groups_0, pad = var_739_pad_0, pad_type = var_739_pad_type_0, strides = var_739_strides_0, weight = model_model_layers_11_self_attn_k_proj_weight_palettized, x = var_723)[name = string("op_739")]; + tensor var_740 = const()[name = string("op_740"), val = tensor([1, 8, 1, 128])]; + tensor var_741 = reshape(shape = var_740, x = var_739)[name = string("op_741")]; + string var_748_pad_type_0 = const()[name = string("op_748_pad_type_0"), val = string("valid")]; + tensor var_748_strides_0 = const()[name = string("op_748_strides_0"), val = tensor([1, 1])]; + tensor var_748_pad_0 = const()[name = string("op_748_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_748_dilations_0 = const()[name = string("op_748_dilations_0"), val = tensor([1, 1])]; + int32 var_748_groups_0 = const()[name = string("op_748_groups_0"), val = int32(1)]; + tensor var_748 = conv(dilations = var_748_dilations_0, groups = var_748_groups_0, pad = var_748_pad_0, pad_type = var_748_pad_type_0, strides = var_748_strides_0, weight = model_model_layers_11_self_attn_v_proj_weight_palettized, x = var_723)[name = string("op_748")]; + tensor var_749 = const()[name = string("op_749"), val = tensor([1, 8, 1, 128])]; + tensor var_750 = reshape(shape = var_749, x = var_748)[name = string("op_750")]; + tensor x1_13_begin_0 = const()[name = string("x1_13_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_13_end_0 = const()[name = string("x1_13_end_0"), val = tensor([1, 32, 1, 64])]; + tensor x1_13_end_mask_0 = const()[name = string("x1_13_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_13 = slice_by_index(begin = x1_13_begin_0, end = x1_13_end_0, end_mask = x1_13_end_mask_0, x = var_732)[name = string("x1_13")]; + tensor x2_13_begin_0 = const()[name = string("x2_13_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_13_end_0 = const()[name = string("x2_13_end_0"), val = tensor([1, 32, 1, 128])]; + tensor x2_13_end_mask_0 = const()[name = string("x2_13_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_13 = slice_by_index(begin = x2_13_begin_0, end = x2_13_end_0, end_mask = x2_13_end_mask_0, x = var_732)[name = string("x2_13")]; + tensor var_764_cast_fp16 = mul(x = x1_13, y = cos_3_cast_fp16)[name = string("op_764_cast_fp16")]; + tensor var_765_cast_fp16 = mul(x = x2_13, y = sin_3_cast_fp16)[name = string("op_765_cast_fp16")]; + tensor var_766_cast_fp16 = sub(x = var_764_cast_fp16, y = var_765_cast_fp16)[name = string("op_766_cast_fp16")]; + tensor var_767_cast_fp16 = mul(x = x2_13, y = cos_3_cast_fp16)[name = string("op_767_cast_fp16")]; + tensor var_768_cast_fp16 = mul(x = x1_13, y = sin_3_cast_fp16)[name = string("op_768_cast_fp16")]; + tensor var_769_cast_fp16 = add(x = var_767_cast_fp16, y = var_768_cast_fp16)[name = string("op_769_cast_fp16")]; + bool rotated_13_interleave_0 = const()[name = string("rotated_13_interleave_0"), val = bool(false)]; + tensor rotated_13_cast_fp16 = concat(axis = var_40, interleave = rotated_13_interleave_0, values = (var_766_cast_fp16, var_769_cast_fp16))[name = string("rotated_13_cast_fp16")]; + tensor x1_begin_0 = const()[name = string("x1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_end_0 = const()[name = string("x1_end_0"), val = tensor([1, 8, 1, 64])]; + tensor x1_end_mask_0 = const()[name = string("x1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1 = slice_by_index(begin = x1_begin_0, end = x1_end_0, end_mask = x1_end_mask_0, x = var_741)[name = string("x1")]; + tensor x2_begin_0 = const()[name = string("x2_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_end_0 = const()[name = string("x2_end_0"), val = tensor([1, 8, 1, 128])]; + tensor x2_end_mask_0 = const()[name = string("x2_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2 = slice_by_index(begin = x2_begin_0, end = x2_end_0, end_mask = x2_end_mask_0, x = var_741)[name = string("x2")]; + tensor var_785_cast_fp16 = mul(x = x1, y = cos_3_cast_fp16)[name = string("op_785_cast_fp16")]; + tensor var_786_cast_fp16 = mul(x = x2, y = sin_3_cast_fp16)[name = string("op_786_cast_fp16")]; + tensor var_787_cast_fp16 = sub(x = var_785_cast_fp16, y = var_786_cast_fp16)[name = string("op_787_cast_fp16")]; + tensor var_788_cast_fp16 = mul(x = x2, y = cos_3_cast_fp16)[name = string("op_788_cast_fp16")]; + tensor var_789_cast_fp16 = mul(x = x1, y = sin_3_cast_fp16)[name = string("op_789_cast_fp16")]; + tensor var_790_cast_fp16 = add(x = var_788_cast_fp16, y = var_789_cast_fp16)[name = string("op_790_cast_fp16")]; + bool rotated_interleave_0 = const()[name = string("rotated_interleave_0"), val = bool(false)]; + tensor rotated_cast_fp16 = concat(axis = var_40, interleave = rotated_interleave_0, values = (var_787_cast_fp16, var_790_cast_fp16))[name = string("rotated_cast_fp16")]; + tensor expand_dims_36 = const()[name = string("expand_dims_36"), val = tensor([11])]; + tensor expand_dims_37 = const()[name = string("expand_dims_37"), val = tensor([0])]; + tensor expand_dims_39 = const()[name = string("expand_dims_39"), val = tensor([0])]; + tensor expand_dims_40 = const()[name = string("expand_dims_40"), val = tensor([12])]; + int32 concat_26_axis_0 = const()[name = string("concat_26_axis_0"), val = int32(0)]; + bool concat_26_interleave_0 = const()[name = string("concat_26_interleave_0"), val = bool(false)]; + tensor concat_26 = concat(axis = concat_26_axis_0, interleave = concat_26_interleave_0, values = (expand_dims_36, expand_dims_37, current_pos, expand_dims_39))[name = string("concat_26")]; + tensor concat_27_values1_0 = const()[name = string("concat_27_values1_0"), val = tensor([0])]; + tensor concat_27_values3_0 = const()[name = string("concat_27_values3_0"), val = tensor([0])]; + int32 concat_27_axis_0 = const()[name = string("concat_27_axis_0"), val = int32(0)]; + bool concat_27_interleave_0 = const()[name = string("concat_27_interleave_0"), val = bool(false)]; + tensor concat_27 = concat(axis = concat_27_axis_0, interleave = concat_27_interleave_0, values = (expand_dims_40, concat_27_values1_0, var_240, concat_27_values3_0))[name = string("concat_27")]; + tensor model_model_kv_cache_0_internal_tensor_assign_7_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_7_stride_0"), val = tensor([1, 1, 1, 1])]; + tensor model_model_kv_cache_0_internal_tensor_assign_7_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_7_begin_mask_0"), val = tensor([false, false, false, false])]; + tensor model_model_kv_cache_0_internal_tensor_assign_7_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_7_end_mask_0"), val = tensor([false, true, false, true])]; + tensor model_model_kv_cache_0_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_7_squeeze_mask_0"), val = tensor([false, false, false, false])]; + tensor model_model_kv_cache_0_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_26, begin_mask = model_model_kv_cache_0_internal_tensor_assign_7_begin_mask_0, end = concat_27, end_mask = model_model_kv_cache_0_internal_tensor_assign_7_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_7_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_7_stride_0, update = rotated_cast_fp16, x = coreml_update_state_13)[name = string("model_model_kv_cache_0_internal_tensor_assign_7_cast_fp16")]; + write_state(data = model_model_kv_cache_0_internal_tensor_assign_7_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_6_write_state")]; + tensor coreml_update_state_14 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_6")]; + tensor expand_dims_42 = const()[name = string("expand_dims_42"), val = tensor([43])]; + tensor expand_dims_43 = const()[name = string("expand_dims_43"), val = tensor([0])]; + tensor expand_dims_45 = const()[name = string("expand_dims_45"), val = tensor([0])]; + tensor expand_dims_46 = const()[name = string("expand_dims_46"), val = tensor([44])]; + int32 concat_30_axis_0 = const()[name = string("concat_30_axis_0"), val = int32(0)]; + bool concat_30_interleave_0 = const()[name = string("concat_30_interleave_0"), val = bool(false)]; + tensor concat_30 = concat(axis = concat_30_axis_0, interleave = concat_30_interleave_0, values = (expand_dims_42, expand_dims_43, current_pos, expand_dims_45))[name = string("concat_30")]; + tensor concat_31_values1_0 = const()[name = string("concat_31_values1_0"), val = tensor([0])]; + tensor concat_31_values3_0 = const()[name = string("concat_31_values3_0"), val = tensor([0])]; + int32 concat_31_axis_0 = const()[name = string("concat_31_axis_0"), val = int32(0)]; + bool concat_31_interleave_0 = const()[name = string("concat_31_interleave_0"), val = bool(false)]; + tensor concat_31 = concat(axis = concat_31_axis_0, interleave = concat_31_interleave_0, values = (expand_dims_46, concat_31_values1_0, var_240, concat_31_values3_0))[name = string("concat_31")]; + tensor model_model_kv_cache_0_internal_tensor_assign_8_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_8_stride_0"), val = tensor([1, 1, 1, 1])]; + tensor model_model_kv_cache_0_internal_tensor_assign_8_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_8_begin_mask_0"), val = tensor([false, false, false, false])]; + tensor model_model_kv_cache_0_internal_tensor_assign_8_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_8_end_mask_0"), val = tensor([false, true, false, true])]; + tensor model_model_kv_cache_0_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_8_squeeze_mask_0"), val = tensor([false, false, false, false])]; + tensor model_model_kv_cache_0_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_30, begin_mask = model_model_kv_cache_0_internal_tensor_assign_8_begin_mask_0, end = concat_31, end_mask = model_model_kv_cache_0_internal_tensor_assign_8_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_8_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_8_stride_0, update = var_750, x = coreml_update_state_14)[name = string("model_model_kv_cache_0_internal_tensor_assign_8_cast_fp16")]; + write_state(data = model_model_kv_cache_0_internal_tensor_assign_8_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_7_write_state")]; + tensor coreml_update_state_15 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_7")]; + tensor var_810_begin_0 = const()[name = string("op_810_begin_0"), val = tensor([11, 0, 0, 0])]; + tensor var_810_end_0 = const()[name = string("op_810_end_0"), val = tensor([12, 8, 1024, 128])]; + tensor var_810_end_mask_0 = const()[name = string("op_810_end_mask_0"), val = tensor([false, true, true, true])]; + tensor var_810_cast_fp16 = slice_by_index(begin = var_810_begin_0, end = var_810_end_0, end_mask = var_810_end_mask_0, x = coreml_update_state_15)[name = string("op_810_cast_fp16")]; + tensor K_layer_cache_axes_0 = const()[name = string("K_layer_cache_axes_0"), val = tensor([0])]; + tensor K_layer_cache_cast_fp16 = squeeze(axes = K_layer_cache_axes_0, x = var_810_cast_fp16)[name = string("K_layer_cache_cast_fp16")]; + tensor var_812_begin_0 = const()[name = string("op_812_begin_0"), val = tensor([43, 0, 0, 0])]; + tensor var_812_end_0 = const()[name = string("op_812_end_0"), val = tensor([44, 8, 1024, 128])]; + tensor var_812_end_mask_0 = const()[name = string("op_812_end_mask_0"), val = tensor([false, true, true, true])]; + tensor var_812_cast_fp16 = slice_by_index(begin = var_812_begin_0, end = var_812_end_0, end_mask = var_812_end_mask_0, x = coreml_update_state_15)[name = string("op_812_cast_fp16")]; + tensor V_layer_cache_axes_0 = const()[name = string("V_layer_cache_axes_0"), val = tensor([0])]; + tensor V_layer_cache_cast_fp16 = squeeze(axes = V_layer_cache_axes_0, x = var_812_cast_fp16)[name = string("V_layer_cache_cast_fp16")]; + tensor x_95_axes_0 = const()[name = string("x_95_axes_0"), val = tensor([1])]; + tensor x_95_cast_fp16 = expand_dims(axes = x_95_axes_0, x = K_layer_cache_cast_fp16)[name = string("x_95_cast_fp16")]; + tensor var_821 = const()[name = string("op_821"), val = tensor([1, 4, 1, 1])]; + tensor x_97_cast_fp16 = tile(reps = var_821, x = x_95_cast_fp16)[name = string("x_97_cast_fp16")]; + tensor var_825 = const()[name = string("op_825"), val = tensor([1, -1, 1024, 128])]; + tensor key_states_cast_fp16 = reshape(shape = var_825, x = x_97_cast_fp16)[name = string("key_states_cast_fp16")]; + tensor x_101_axes_0 = const()[name = string("x_101_axes_0"), val = tensor([1])]; + tensor x_101_cast_fp16 = expand_dims(axes = x_101_axes_0, x = V_layer_cache_cast_fp16)[name = string("x_101_cast_fp16")]; + tensor var_828 = const()[name = string("op_828"), val = tensor([1, 4, 1, 1])]; + tensor x_103_cast_fp16 = tile(reps = var_828, x = x_101_cast_fp16)[name = string("x_103_cast_fp16")]; + tensor var_832 = const()[name = string("op_832"), val = tensor([1, -1, 1024, 128])]; + tensor value_states_cast_fp16 = reshape(shape = var_832, x = x_103_cast_fp16)[name = string("value_states_cast_fp16")]; + bool var_835_transpose_x_1 = const()[name = string("op_835_transpose_x_1"), val = bool(false)]; + bool var_835_transpose_y_1 = const()[name = string("op_835_transpose_y_1"), val = bool(true)]; + tensor var_835_cast_fp16 = matmul(transpose_x = var_835_transpose_x_1, transpose_y = var_835_transpose_y_1, x = rotated_13_cast_fp16, y = key_states_cast_fp16)[name = string("op_835_cast_fp16")]; + fp16 var_836_to_fp16 = const()[name = string("op_836_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor attn_weights_13_cast_fp16 = mul(x = var_835_cast_fp16, y = var_836_to_fp16)[name = string("attn_weights_13_cast_fp16")]; + tensor x_105_cast_fp16 = add(x = attn_weights_13_cast_fp16, y = causal_mask)[name = string("x_105_cast_fp16")]; + tensor reduce_max_3_axes_0 = const()[name = string("reduce_max_3_axes_0"), val = tensor([-1])]; + bool reduce_max_3_keep_dims_0 = const()[name = string("reduce_max_3_keep_dims_0"), val = bool(true)]; + tensor reduce_max_3_cast_fp16 = reduce_max(axes = reduce_max_3_axes_0, keep_dims = reduce_max_3_keep_dims_0, x = x_105_cast_fp16)[name = string("reduce_max_3_cast_fp16")]; + tensor x_107_cast_fp16 = sub(x = x_105_cast_fp16, y = reduce_max_3_cast_fp16)[name = string("x_107_cast_fp16")]; + tensor exp_x_cast_fp16 = exp(x = x_107_cast_fp16)[name = string("exp_x_cast_fp16")]; + tensor var_847_axes_0 = const()[name = string("op_847_axes_0"), val = tensor([-1])]; + bool var_847_keep_dims_0 = const()[name = string("op_847_keep_dims_0"), val = bool(true)]; + tensor var_847_cast_fp16 = reduce_sum(axes = var_847_axes_0, keep_dims = var_847_keep_dims_0, x = exp_x_cast_fp16)[name = string("op_847_cast_fp16")]; + tensor attn_weights_cast_fp16 = real_div(x = exp_x_cast_fp16, y = var_847_cast_fp16)[name = string("attn_weights_cast_fp16")]; + bool attn_output_19_transpose_x_0 = const()[name = string("attn_output_19_transpose_x_0"), val = bool(false)]; + bool attn_output_19_transpose_y_0 = const()[name = string("attn_output_19_transpose_y_0"), val = bool(false)]; + tensor attn_output_19_cast_fp16 = matmul(transpose_x = attn_output_19_transpose_x_0, transpose_y = attn_output_19_transpose_y_0, x = attn_weights_cast_fp16, y = value_states_cast_fp16)[name = string("attn_output_19_cast_fp16")]; + tensor var_850_perm_0 = const()[name = string("op_850_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_852 = const()[name = string("op_852"), val = tensor([1, 1, 4096])]; + tensor var_850_cast_fp16 = transpose(perm = var_850_perm_0, x = attn_output_19_cast_fp16)[name = string("transpose_2")]; + tensor input_47_cast_fp16 = reshape(shape = var_852, x = var_850_cast_fp16)[name = string("input_47_cast_fp16")]; + tensor model_model_layers_11_self_attn_o_proj_weight_promoted_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(711594048))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(724177024))))[name = string("model_model_layers_11_self_attn_o_proj_weight_promoted_to_fp16_palettized")]; + tensor linear_3_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_11_self_attn_o_proj_weight_promoted_to_fp16_palettized, x = input_47_cast_fp16)[name = string("linear_3_cast_fp16")]; + tensor hidden_states_29_cast_fp16 = add(x = hidden_states_25_cast_fp16, y = linear_3_cast_fp16)[name = string("hidden_states_29_cast_fp16")]; + tensor mean_axes_0 = const()[name = string("mean_axes_0"), val = tensor([-1])]; + bool mean_keep_dims_0 = const()[name = string("mean_keep_dims_0"), val = bool(true)]; + tensor mean_cast_fp16 = reduce_mean(axes = mean_axes_0, keep_dims = mean_keep_dims_0, x = hidden_states_29_cast_fp16)[name = string("mean_cast_fp16")]; + tensor input_49_cast_fp16 = sub(x = hidden_states_29_cast_fp16, y = mean_cast_fp16)[name = string("input_49_cast_fp16")]; + tensor var_863_axes_0 = const()[name = string("op_863_axes_0"), val = tensor([-1])]; + tensor model_model_layers_11_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_11_post_attention_layernorm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(724242624)))]; + tensor var_863_cast_fp16 = layer_norm(axes = var_863_axes_0, epsilon = var_35_to_fp16, gamma = model_model_layers_11_post_attention_layernorm_weight_to_fp16, x = input_49_cast_fp16)[name = string("op_863_cast_fp16")]; + tensor var_870 = const()[name = string("op_870"), val = tensor([0, 2, 1])]; + tensor input_51_axes_0 = const()[name = string("input_51_axes_0"), val = tensor([2])]; + tensor var_871 = transpose(perm = var_870, x = var_863_cast_fp16)[name = string("transpose_1")]; + tensor input_51 = expand_dims(axes = input_51_axes_0, x = var_871)[name = string("input_51")]; + string input_53_pad_type_0 = const()[name = string("input_53_pad_type_0"), val = string("valid")]; + tensor input_53_strides_0 = const()[name = string("input_53_strides_0"), val = tensor([1, 1])]; + tensor input_53_pad_0 = const()[name = string("input_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_53_dilations_0 = const()[name = string("input_53_dilations_0"), val = tensor([1, 1])]; + int32 input_53_groups_0 = const()[name = string("input_53_groups_0"), val = int32(1)]; + tensor input_53 = conv(dilations = input_53_dilations_0, groups = input_53_groups_0, pad = input_53_pad_0, pad_type = input_53_pad_type_0, strides = input_53_strides_0, weight = model_model_layers_11_mlp_gate_proj_weight_palettized, x = input_51)[name = string("input_53")]; + string up_states_pad_type_0 = const()[name = string("up_states_pad_type_0"), val = string("valid")]; + tensor up_states_strides_0 = const()[name = string("up_states_strides_0"), val = tensor([1, 1])]; + tensor up_states_pad_0 = const()[name = string("up_states_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor up_states_dilations_0 = const()[name = string("up_states_dilations_0"), val = tensor([1, 1])]; + int32 up_states_groups_0 = const()[name = string("up_states_groups_0"), val = int32(1)]; + tensor up_states = conv(dilations = up_states_dilations_0, groups = up_states_groups_0, pad = up_states_pad_0, pad_type = up_states_pad_type_0, strides = up_states_strides_0, weight = model_model_layers_11_mlp_up_proj_weight_palettized, x = input_51)[name = string("up_states")]; + tensor gate_states = silu(x = input_53)[name = string("gate_states")]; + tensor input = mul(x = gate_states, y = up_states)[name = string("input")]; + string hidden_states_pad_type_0 = const()[name = string("hidden_states_pad_type_0"), val = string("valid")]; + tensor hidden_states_strides_0 = const()[name = string("hidden_states_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_pad_0 = const()[name = string("hidden_states_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_dilations_0 = const()[name = string("hidden_states_dilations_0"), val = tensor([1, 1])]; + int32 hidden_states_groups_0 = const()[name = string("hidden_states_groups_0"), val = int32(1)]; + tensor hidden_states_1 = conv(dilations = hidden_states_dilations_0, groups = hidden_states_groups_0, pad = hidden_states_pad_0, pad_type = hidden_states_pad_type_0, strides = hidden_states_strides_0, weight = model_model_layers_11_mlp_down_proj_weight_palettized, x = input)[name = string("hidden_states")]; + tensor var_893_axes_0 = const()[name = string("op_893_axes_0"), val = tensor([2])]; + tensor var_893 = squeeze(axes = var_893_axes_0, x = hidden_states_1)[name = string("op_893")]; + tensor var_894 = const()[name = string("op_894"), val = tensor([0, 2, 1])]; + tensor var_895 = transpose(perm = var_894, x = var_893)[name = string("transpose_0")]; + tensor output_hidden_states = add(x = hidden_states_29_cast_fp16, y = var_895)[name = string("op_896_cast_fp16")]; + tensor position_ids_tmp = identity(x = position_ids)[name = string("position_ids_tmp")]; + } -> (output_hidden_states); + func prefill(tensor causal_mask, tensor current_pos, tensor hidden_states, state> model_model_kv_cache_0, tensor position_ids) { + tensor model_model_layers_8_self_attn_q_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12583040))))[name = string("model_model_layers_8_self_attn_q_proj_weight_palettized")]; + tensor model_model_layers_8_self_attn_k_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12648640))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15794432))))[name = string("model_model_layers_8_self_attn_k_proj_weight_palettized")]; + tensor model_model_layers_8_self_attn_v_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15810880))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18956672))))[name = string("model_model_layers_8_self_attn_v_proj_weight_palettized")]; + tensor model_model_layers_8_mlp_gate_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18973120))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63013376))))[name = string("model_model_layers_8_mlp_gate_proj_weight_palettized")]; + tensor model_model_layers_8_mlp_up_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63242816))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107283072))))[name = string("model_model_layers_8_mlp_up_proj_weight_palettized")]; + tensor model_model_layers_8_mlp_down_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107512512))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151552768))))[name = string("model_model_layers_8_mlp_down_proj_weight_palettized")]; + tensor model_model_layers_9_self_attn_q_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151618368))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164201344))))[name = string("model_model_layers_9_self_attn_q_proj_weight_palettized")]; + tensor model_model_layers_9_self_attn_k_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164266944))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167412736))))[name = string("model_model_layers_9_self_attn_k_proj_weight_palettized")]; + tensor model_model_layers_9_self_attn_v_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167429184))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170574976))))[name = string("model_model_layers_9_self_attn_v_proj_weight_palettized")]; + tensor model_model_layers_9_mlp_gate_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170591424))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214631680))))[name = string("model_model_layers_9_mlp_gate_proj_weight_palettized")]; + tensor model_model_layers_9_mlp_up_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214861120))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258901376))))[name = string("model_model_layers_9_mlp_up_proj_weight_palettized")]; + tensor model_model_layers_9_mlp_down_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259130816))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303171072))))[name = string("model_model_layers_9_mlp_down_proj_weight_palettized")]; + tensor model_model_layers_10_self_attn_q_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303236672))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315819648))))[name = string("model_model_layers_10_self_attn_q_proj_weight_palettized")]; + tensor model_model_layers_10_self_attn_k_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315885248))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319031040))))[name = string("model_model_layers_10_self_attn_k_proj_weight_palettized")]; + tensor model_model_layers_10_self_attn_v_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319047488))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322193280))))[name = string("model_model_layers_10_self_attn_v_proj_weight_palettized")]; + tensor model_model_layers_10_mlp_gate_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322209728))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366249984))))[name = string("model_model_layers_10_mlp_gate_proj_weight_palettized")]; + tensor model_model_layers_10_mlp_up_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366479424))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410519680))))[name = string("model_model_layers_10_mlp_up_proj_weight_palettized")]; + tensor model_model_layers_10_mlp_down_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410749120))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454789376))))[name = string("model_model_layers_10_mlp_down_proj_weight_palettized")]; + tensor model_model_layers_11_self_attn_q_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454854976))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467437952))))[name = string("model_model_layers_11_self_attn_q_proj_weight_palettized")]; + tensor model_model_layers_11_self_attn_k_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467503552))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(470649344))))[name = string("model_model_layers_11_self_attn_k_proj_weight_palettized")]; + tensor model_model_layers_11_self_attn_v_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(470665792))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473811584))))[name = string("model_model_layers_11_self_attn_v_proj_weight_palettized")]; + tensor model_model_layers_11_mlp_gate_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473828032))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(517868288))))[name = string("model_model_layers_11_mlp_gate_proj_weight_palettized")]; + tensor model_model_layers_11_mlp_up_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(518097728))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(562137984))))[name = string("model_model_layers_11_mlp_up_proj_weight_palettized")]; + tensor model_model_layers_11_mlp_down_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(562367424))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(606407680))))[name = string("model_model_layers_11_mlp_down_proj_weight_palettized")]; + int32 var_35 = const()[name = string("op_35"), val = int32(-1)]; + int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)]; + tensor greater_equal_0 = greater_equal(x = position_ids, y = greater_equal_0_y_0)[name = string("greater_equal_0")]; + int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(131072)]; + tensor add_0 = add(x = position_ids, y = slice_by_index_0)[name = string("add_0")]; + tensor select_0 = select(a = position_ids, b = add_0, cond = greater_equal_0)[name = string("select_0")]; + int32 var_152_axis_0 = const()[name = string("op_152_axis_0"), val = int32(1)]; + int32 var_152_batch_dims_0 = const()[name = string("op_152_batch_dims_0"), val = int32(0)]; + bool var_152_validate_indices_0 = const()[name = string("op_152_validate_indices_0"), val = bool(false)]; + tensor var_46_to_fp16 = const()[name = string("op_46_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(640027776)))]; + tensor var_152_cast_fp16 = gather(axis = var_152_axis_0, batch_dims = var_152_batch_dims_0, indices = select_0, validate_indices = var_152_validate_indices_0, x = var_46_to_fp16)[name = string("op_152_cast_fp16")]; + tensor var_153 = const()[name = string("op_153"), val = tensor([1, 256, 1, 128])]; + tensor cos_1_cast_fp16 = reshape(shape = var_153, x = var_152_cast_fp16)[name = string("cos_1_cast_fp16")]; + int32 var_157_axis_0 = const()[name = string("op_157_axis_0"), val = int32(1)]; + int32 var_157_batch_dims_0 = const()[name = string("op_157_batch_dims_0"), val = int32(0)]; + bool var_157_validate_indices_0 = const()[name = string("op_157_validate_indices_0"), val = bool(false)]; + tensor var_41_to_fp16 = const()[name = string("op_41_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(606473280)))]; + tensor var_157_cast_fp16 = gather(axis = var_157_axis_0, batch_dims = var_157_batch_dims_0, indices = select_0, validate_indices = var_157_validate_indices_0, x = var_41_to_fp16)[name = string("op_157_cast_fp16")]; + tensor var_158 = const()[name = string("op_158"), val = tensor([1, 256, 1, 128])]; + tensor sin_1_cast_fp16 = reshape(shape = var_158, x = var_157_cast_fp16)[name = string("sin_1_cast_fp16")]; + tensor mean_1_axes_0 = const()[name = string("mean_1_axes_0"), val = tensor([-1])]; + bool mean_1_keep_dims_0 = const()[name = string("mean_1_keep_dims_0"), val = bool(true)]; + tensor mean_1_cast_fp16 = reduce_mean(axes = mean_1_axes_0, keep_dims = mean_1_keep_dims_0, x = hidden_states)[name = string("mean_1_cast_fp16")]; + tensor input_1_cast_fp16 = sub(x = hidden_states, y = mean_1_cast_fp16)[name = string("input_1_cast_fp16")]; + tensor var_168_axes_0 = const()[name = string("op_168_axes_0"), val = tensor([-1])]; + tensor model_model_layers_8_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_8_input_layernorm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(673582272)))]; + fp16 var_37_to_fp16 = const()[name = string("op_37_to_fp16"), val = fp16(0x1.5p-17)]; + tensor var_168_cast_fp16 = layer_norm(axes = var_168_axes_0, epsilon = var_37_to_fp16, gamma = model_model_layers_8_input_layernorm_weight_to_fp16, x = input_1_cast_fp16)[name = string("op_168_cast_fp16")]; + tensor var_172 = const()[name = string("op_172"), val = tensor([0, 2, 1])]; + tensor var_174_axes_0 = const()[name = string("op_174_axes_0"), val = tensor([2])]; + tensor var_173 = transpose(perm = var_172, x = var_168_cast_fp16)[name = string("transpose_29")]; + tensor var_174 = expand_dims(axes = var_174_axes_0, x = var_173)[name = string("op_174")]; + string query_states_1_pad_type_0 = const()[name = string("query_states_1_pad_type_0"), val = string("valid")]; + tensor query_states_1_strides_0 = const()[name = string("query_states_1_strides_0"), val = tensor([1, 1])]; + tensor query_states_1_pad_0 = const()[name = string("query_states_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_states_1_dilations_0 = const()[name = string("query_states_1_dilations_0"), val = tensor([1, 1])]; + int32 query_states_1_groups_0 = const()[name = string("query_states_1_groups_0"), val = int32(1)]; + tensor query_states_1 = conv(dilations = query_states_1_dilations_0, groups = query_states_1_groups_0, pad = query_states_1_pad_0, pad_type = query_states_1_pad_type_0, strides = query_states_1_strides_0, weight = model_model_layers_8_self_attn_q_proj_weight_palettized, x = var_174)[name = string("query_states_1")]; + string key_states_1_pad_type_0 = const()[name = string("key_states_1_pad_type_0"), val = string("valid")]; + tensor key_states_1_strides_0 = const()[name = string("key_states_1_strides_0"), val = tensor([1, 1])]; + tensor key_states_1_pad_0 = const()[name = string("key_states_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_states_1_dilations_0 = const()[name = string("key_states_1_dilations_0"), val = tensor([1, 1])]; + int32 key_states_1_groups_0 = const()[name = string("key_states_1_groups_0"), val = int32(1)]; + tensor key_states_1 = conv(dilations = key_states_1_dilations_0, groups = key_states_1_groups_0, pad = key_states_1_pad_0, pad_type = key_states_1_pad_type_0, strides = key_states_1_strides_0, weight = model_model_layers_8_self_attn_k_proj_weight_palettized, x = var_174)[name = string("key_states_1")]; + string value_states_1_pad_type_0 = const()[name = string("value_states_1_pad_type_0"), val = string("valid")]; + tensor value_states_1_strides_0 = const()[name = string("value_states_1_strides_0"), val = tensor([1, 1])]; + tensor value_states_1_pad_0 = const()[name = string("value_states_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_states_1_dilations_0 = const()[name = string("value_states_1_dilations_0"), val = tensor([1, 1])]; + int32 value_states_1_groups_0 = const()[name = string("value_states_1_groups_0"), val = int32(1)]; + tensor value_states_1 = conv(dilations = value_states_1_dilations_0, groups = value_states_1_groups_0, pad = value_states_1_pad_0, pad_type = value_states_1_pad_type_0, strides = value_states_1_strides_0, weight = model_model_layers_8_self_attn_v_proj_weight_palettized, x = var_174)[name = string("value_states_1")]; + tensor var_194 = const()[name = string("op_194"), val = tensor([1, 32, 128, 256])]; + tensor var_195 = reshape(shape = var_194, x = query_states_1)[name = string("op_195")]; + tensor var_196 = const()[name = string("op_196"), val = tensor([0, 1, 3, 2])]; + tensor var_198 = const()[name = string("op_198"), val = tensor([1, 8, 128, 256])]; + tensor var_199 = reshape(shape = var_198, x = key_states_1)[name = string("op_199")]; + tensor var_200 = const()[name = string("op_200"), val = tensor([0, 1, 3, 2])]; + tensor var_202 = const()[name = string("op_202"), val = tensor([1, 8, 128, 256])]; + tensor var_203 = reshape(shape = var_202, x = value_states_1)[name = string("op_203")]; + tensor var_204 = const()[name = string("op_204"), val = tensor([0, 1, 3, 2])]; + tensor var_206 = const()[name = string("op_206"), val = tensor([0, 2, 1, 3])]; + tensor var_208 = const()[name = string("op_208"), val = tensor([0, 2, 1, 3])]; + tensor x1_1_begin_0 = const()[name = string("x1_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_1_end_0 = const()[name = string("x1_1_end_0"), val = tensor([1, 32, 256, 64])]; + tensor x1_1_end_mask_0 = const()[name = string("x1_1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_1 = transpose(perm = var_196, x = var_195)[name = string("transpose_28")]; + tensor x1_1 = slice_by_index(begin = x1_1_begin_0, end = x1_1_end_0, end_mask = x1_1_end_mask_0, x = x_1)[name = string("x1_1")]; + tensor x2_1_begin_0 = const()[name = string("x2_1_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_1_end_0 = const()[name = string("x2_1_end_0"), val = tensor([1, 32, 256, 128])]; + tensor x2_1_end_mask_0 = const()[name = string("x2_1_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_1 = slice_by_index(begin = x2_1_begin_0, end = x2_1_end_0, end_mask = x2_1_end_mask_0, x = x_1)[name = string("x2_1")]; + tensor cos_7_begin_0 = const()[name = string("cos_7_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor cos_7_end_0 = const()[name = string("cos_7_end_0"), val = tensor([1, 1, 256, 64])]; + tensor cos_7_end_mask_0 = const()[name = string("cos_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor cos_5 = transpose(perm = var_206, x = cos_1_cast_fp16)[name = string("transpose_27")]; + tensor cos_7 = slice_by_index(begin = cos_7_begin_0, end = cos_7_end_0, end_mask = cos_7_end_mask_0, x = cos_5)[name = string("cos_7")]; + tensor sin_7_begin_0 = const()[name = string("sin_7_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor sin_7_end_0 = const()[name = string("sin_7_end_0"), val = tensor([1, 1, 256, 64])]; + tensor sin_7_end_mask_0 = const()[name = string("sin_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor sin_5 = transpose(perm = var_208, x = sin_1_cast_fp16)[name = string("transpose_26")]; + tensor sin_7 = slice_by_index(begin = sin_7_begin_0, end = sin_7_end_0, end_mask = sin_7_end_mask_0, x = sin_5)[name = string("sin_7")]; + tensor var_222 = mul(x = x1_1, y = cos_7)[name = string("op_222")]; + tensor var_223 = mul(x = x2_1, y = sin_7)[name = string("op_223")]; + tensor var_224 = sub(x = var_222, y = var_223)[name = string("op_224")]; + tensor var_225 = mul(x = x2_1, y = cos_7)[name = string("op_225")]; + tensor var_226 = mul(x = x1_1, y = sin_7)[name = string("op_226")]; + tensor var_227 = add(x = var_225, y = var_226)[name = string("op_227")]; + bool rotated_1_interleave_0 = const()[name = string("rotated_1_interleave_0"), val = bool(false)]; + tensor rotated_1 = concat(axis = var_35, interleave = rotated_1_interleave_0, values = (var_224, var_227))[name = string("rotated_1")]; + tensor x1_3_begin_0 = const()[name = string("x1_3_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_3_end_0 = const()[name = string("x1_3_end_0"), val = tensor([1, 8, 256, 64])]; + tensor x1_3_end_mask_0 = const()[name = string("x1_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_5 = transpose(perm = var_200, x = var_199)[name = string("transpose_25")]; + tensor x1_3 = slice_by_index(begin = x1_3_begin_0, end = x1_3_end_0, end_mask = x1_3_end_mask_0, x = x_5)[name = string("x1_3")]; + tensor x2_3_begin_0 = const()[name = string("x2_3_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_3_end_0 = const()[name = string("x2_3_end_0"), val = tensor([1, 8, 256, 128])]; + tensor x2_3_end_mask_0 = const()[name = string("x2_3_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_3 = slice_by_index(begin = x2_3_begin_0, end = x2_3_end_0, end_mask = x2_3_end_mask_0, x = x_5)[name = string("x2_3")]; + tensor var_243 = mul(x = x1_3, y = cos_7)[name = string("op_243")]; + tensor var_244 = mul(x = x2_3, y = sin_7)[name = string("op_244")]; + tensor var_245 = sub(x = var_243, y = var_244)[name = string("op_245")]; + tensor var_246 = mul(x = x2_3, y = cos_7)[name = string("op_246")]; + tensor var_247 = mul(x = x1_3, y = sin_7)[name = string("op_247")]; + tensor var_248 = add(x = var_246, y = var_247)[name = string("op_248")]; + bool rotated_3_interleave_0 = const()[name = string("rotated_3_interleave_0"), val = bool(false)]; + tensor rotated_3 = concat(axis = var_35, interleave = rotated_3_interleave_0, values = (var_245, var_248))[name = string("rotated_3")]; + tensor seq_length_1 = const()[name = string("seq_length_1"), val = tensor([256])]; + tensor var_257 = add(x = current_pos, y = seq_length_1)[name = string("op_257")]; + tensor read_state_0 = read_state(input = model_model_kv_cache_0)[name = string("read_state_0")]; + tensor expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor([8])]; + tensor expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor([0])]; + tensor expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor([0])]; + tensor expand_dims_4 = const()[name = string("expand_dims_4"), val = tensor([9])]; + int32 concat_2_axis_0 = const()[name = string("concat_2_axis_0"), val = int32(0)]; + bool concat_2_interleave_0 = const()[name = string("concat_2_interleave_0"), val = bool(false)]; + tensor concat_2 = concat(axis = concat_2_axis_0, interleave = concat_2_interleave_0, values = (expand_dims_0, expand_dims_1, current_pos, expand_dims_3))[name = string("concat_2")]; + tensor concat_3_values1_0 = const()[name = string("concat_3_values1_0"), val = tensor([0])]; + tensor concat_3_values3_0 = const()[name = string("concat_3_values3_0"), val = tensor([0])]; + int32 concat_3_axis_0 = const()[name = string("concat_3_axis_0"), val = int32(0)]; + bool concat_3_interleave_0 = const()[name = string("concat_3_interleave_0"), val = bool(false)]; + tensor concat_3 = concat(axis = concat_3_axis_0, interleave = concat_3_interleave_0, values = (expand_dims_4, concat_3_values1_0, var_257, concat_3_values3_0))[name = string("concat_3")]; + tensor model_model_kv_cache_0_internal_tensor_assign_1_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1])]; + tensor model_model_kv_cache_0_internal_tensor_assign_1_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, false, false, false])]; + tensor model_model_kv_cache_0_internal_tensor_assign_1_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, false, true])]; + tensor model_model_kv_cache_0_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([false, false, false, false])]; + tensor model_model_kv_cache_0_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_2, begin_mask = model_model_kv_cache_0_internal_tensor_assign_1_begin_mask_0, end = concat_3, end_mask = model_model_kv_cache_0_internal_tensor_assign_1_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_1_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_1_stride_0, update = rotated_3, x = read_state_0)[name = string("model_model_kv_cache_0_internal_tensor_assign_1_cast_fp16")]; + write_state(data = model_model_kv_cache_0_internal_tensor_assign_1_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_8_write_state")]; + tensor coreml_update_state_8 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_8")]; + tensor expand_dims_6 = const()[name = string("expand_dims_6"), val = tensor([40])]; + tensor expand_dims_7 = const()[name = string("expand_dims_7"), val = tensor([0])]; + tensor expand_dims_9 = const()[name = string("expand_dims_9"), val = tensor([0])]; + tensor expand_dims_10 = const()[name = string("expand_dims_10"), val = tensor([41])]; + int32 concat_6_axis_0 = const()[name = string("concat_6_axis_0"), val = int32(0)]; + bool concat_6_interleave_0 = const()[name = string("concat_6_interleave_0"), val = bool(false)]; + tensor concat_6 = concat(axis = concat_6_axis_0, interleave = concat_6_interleave_0, values = (expand_dims_6, expand_dims_7, current_pos, expand_dims_9))[name = string("concat_6")]; + tensor concat_7_values1_0 = const()[name = string("concat_7_values1_0"), val = tensor([0])]; + tensor concat_7_values3_0 = const()[name = string("concat_7_values3_0"), val = tensor([0])]; + int32 concat_7_axis_0 = const()[name = string("concat_7_axis_0"), val = int32(0)]; + bool concat_7_interleave_0 = const()[name = string("concat_7_interleave_0"), val = bool(false)]; + tensor concat_7 = concat(axis = concat_7_axis_0, interleave = concat_7_interleave_0, values = (expand_dims_10, concat_7_values1_0, var_257, concat_7_values3_0))[name = string("concat_7")]; + tensor model_model_kv_cache_0_internal_tensor_assign_2_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1])]; + tensor model_model_kv_cache_0_internal_tensor_assign_2_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, false, false, false])]; + tensor model_model_kv_cache_0_internal_tensor_assign_2_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, false, true])]; + tensor model_model_kv_cache_0_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([false, false, false, false])]; + tensor value_states_3 = transpose(perm = var_204, x = var_203)[name = string("transpose_24")]; + tensor model_model_kv_cache_0_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_6, begin_mask = model_model_kv_cache_0_internal_tensor_assign_2_begin_mask_0, end = concat_7, end_mask = model_model_kv_cache_0_internal_tensor_assign_2_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_2_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_2_stride_0, update = value_states_3, x = coreml_update_state_8)[name = string("model_model_kv_cache_0_internal_tensor_assign_2_cast_fp16")]; + write_state(data = model_model_kv_cache_0_internal_tensor_assign_2_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_9_write_state")]; + tensor coreml_update_state_9 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_9")]; + tensor var_271_begin_0 = const()[name = string("op_271_begin_0"), val = tensor([8, 0, 0, 0])]; + tensor var_271_end_0 = const()[name = string("op_271_end_0"), val = tensor([9, 8, 1024, 128])]; + tensor var_271_end_mask_0 = const()[name = string("op_271_end_mask_0"), val = tensor([false, true, true, true])]; + tensor var_271_cast_fp16 = slice_by_index(begin = var_271_begin_0, end = var_271_end_0, end_mask = var_271_end_mask_0, x = coreml_update_state_9)[name = string("op_271_cast_fp16")]; + tensor K_layer_cache_1_axes_0 = const()[name = string("K_layer_cache_1_axes_0"), val = tensor([0])]; + tensor K_layer_cache_1_cast_fp16 = squeeze(axes = K_layer_cache_1_axes_0, x = var_271_cast_fp16)[name = string("K_layer_cache_1_cast_fp16")]; + tensor var_273_begin_0 = const()[name = string("op_273_begin_0"), val = tensor([40, 0, 0, 0])]; + tensor var_273_end_0 = const()[name = string("op_273_end_0"), val = tensor([41, 8, 1024, 128])]; + tensor var_273_end_mask_0 = const()[name = string("op_273_end_mask_0"), val = tensor([false, true, true, true])]; + tensor var_273_cast_fp16 = slice_by_index(begin = var_273_begin_0, end = var_273_end_0, end_mask = var_273_end_mask_0, x = coreml_update_state_9)[name = string("op_273_cast_fp16")]; + tensor V_layer_cache_1_axes_0 = const()[name = string("V_layer_cache_1_axes_0"), val = tensor([0])]; + tensor V_layer_cache_1_cast_fp16 = squeeze(axes = V_layer_cache_1_axes_0, x = var_273_cast_fp16)[name = string("V_layer_cache_1_cast_fp16")]; + tensor x_11_axes_0 = const()[name = string("x_11_axes_0"), val = tensor([1])]; + tensor x_11_cast_fp16 = expand_dims(axes = x_11_axes_0, x = K_layer_cache_1_cast_fp16)[name = string("x_11_cast_fp16")]; + tensor var_282 = const()[name = string("op_282"), val = tensor([1, 4, 1, 1])]; + tensor x_13_cast_fp16 = tile(reps = var_282, x = x_11_cast_fp16)[name = string("x_13_cast_fp16")]; + tensor var_286 = const()[name = string("op_286"), val = tensor([1, -1, 1024, 128])]; + tensor var_287_cast_fp16 = reshape(shape = var_286, x = x_13_cast_fp16)[name = string("op_287_cast_fp16")]; + tensor x_17_axes_0 = const()[name = string("x_17_axes_0"), val = tensor([1])]; + tensor x_17_cast_fp16 = expand_dims(axes = x_17_axes_0, x = V_layer_cache_1_cast_fp16)[name = string("x_17_cast_fp16")]; + tensor var_289 = const()[name = string("op_289"), val = tensor([1, 4, 1, 1])]; + tensor x_19_cast_fp16 = tile(reps = var_289, x = x_17_cast_fp16)[name = string("x_19_cast_fp16")]; + bool var_296_transpose_x_0 = const()[name = string("op_296_transpose_x_0"), val = bool(false)]; + bool var_296_transpose_y_0 = const()[name = string("op_296_transpose_y_0"), val = bool(true)]; + tensor var_296_cast_fp16 = matmul(transpose_x = var_296_transpose_x_0, transpose_y = var_296_transpose_y_0, x = rotated_1, y = var_287_cast_fp16)[name = string("op_296_cast_fp16")]; + fp16 var_297_to_fp16 = const()[name = string("op_297_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor attn_weights_1_cast_fp16 = mul(x = var_296_cast_fp16, y = var_297_to_fp16)[name = string("attn_weights_1_cast_fp16")]; + tensor x_21_cast_fp16 = add(x = attn_weights_1_cast_fp16, y = causal_mask)[name = string("x_21_cast_fp16")]; + tensor reduce_max_0_axes_0 = const()[name = string("reduce_max_0_axes_0"), val = tensor([-1])]; + bool reduce_max_0_keep_dims_0 = const()[name = string("reduce_max_0_keep_dims_0"), val = bool(true)]; + tensor reduce_max_0_cast_fp16 = reduce_max(axes = reduce_max_0_axes_0, keep_dims = reduce_max_0_keep_dims_0, x = x_21_cast_fp16)[name = string("reduce_max_0_cast_fp16")]; + tensor x_23_cast_fp16 = sub(x = x_21_cast_fp16, y = reduce_max_0_cast_fp16)[name = string("x_23_cast_fp16")]; + tensor exp_x_1_cast_fp16 = exp(x = x_23_cast_fp16)[name = string("exp_x_1_cast_fp16")]; + tensor var_308_axes_0 = const()[name = string("op_308_axes_0"), val = tensor([-1])]; + bool var_308_keep_dims_0 = const()[name = string("op_308_keep_dims_0"), val = bool(true)]; + tensor var_308_cast_fp16 = reduce_sum(axes = var_308_axes_0, keep_dims = var_308_keep_dims_0, x = exp_x_1_cast_fp16)[name = string("op_308_cast_fp16")]; + tensor var_309_cast_fp16 = real_div(x = exp_x_1_cast_fp16, y = var_308_cast_fp16)[name = string("op_309_cast_fp16")]; + tensor concat_12 = const()[name = string("concat_12"), val = tensor([32, 256, 1024])]; + tensor reshape_0_cast_fp16 = reshape(shape = concat_12, x = var_309_cast_fp16)[name = string("reshape_0_cast_fp16")]; + tensor concat_13 = const()[name = string("concat_13"), val = tensor([32, 1024, 128])]; + tensor reshape_1_cast_fp16 = reshape(shape = concat_13, x = x_19_cast_fp16)[name = string("reshape_1_cast_fp16")]; + bool matmul_0_transpose_x_0 = const()[name = string("matmul_0_transpose_x_0"), val = bool(false)]; + bool matmul_0_transpose_y_0 = const()[name = string("matmul_0_transpose_y_0"), val = bool(false)]; + tensor matmul_0_cast_fp16 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = reshape_0_cast_fp16, y = reshape_1_cast_fp16)[name = string("matmul_0_cast_fp16")]; + tensor concat_17 = const()[name = string("concat_17"), val = tensor([1, 32, 256, 128])]; + tensor reshape_2_cast_fp16 = reshape(shape = concat_17, x = matmul_0_cast_fp16)[name = string("reshape_2_cast_fp16")]; + tensor var_312_perm_0 = const()[name = string("op_312_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_314 = const()[name = string("op_314"), val = tensor([1, 256, 4096])]; + tensor var_312_cast_fp16 = transpose(perm = var_312_perm_0, x = reshape_2_cast_fp16)[name = string("transpose_23")]; + tensor input_5_cast_fp16 = reshape(shape = var_314, x = var_312_cast_fp16)[name = string("input_5_cast_fp16")]; + tensor model_model_layers_8_self_attn_o_proj_weight_promoted_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(673590528))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(686173504))))[name = string("model_model_layers_8_self_attn_o_proj_weight_promoted_to_fp16_palettized")]; + tensor linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(686239104)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_8_self_attn_o_proj_weight_promoted_to_fp16_palettized, x = input_5_cast_fp16)[name = string("linear_0_cast_fp16")]; + tensor hidden_states_5_cast_fp16 = add(x = hidden_states, y = linear_0_cast_fp16)[name = string("hidden_states_5_cast_fp16")]; + tensor mean_3_axes_0 = const()[name = string("mean_3_axes_0"), val = tensor([-1])]; + bool mean_3_keep_dims_0 = const()[name = string("mean_3_keep_dims_0"), val = bool(true)]; + tensor mean_3_cast_fp16 = reduce_mean(axes = mean_3_axes_0, keep_dims = mean_3_keep_dims_0, x = hidden_states_5_cast_fp16)[name = string("mean_3_cast_fp16")]; + tensor input_7_cast_fp16 = sub(x = hidden_states_5_cast_fp16, y = mean_3_cast_fp16)[name = string("input_7_cast_fp16")]; + tensor var_325_axes_0 = const()[name = string("op_325_axes_0"), val = tensor([-1])]; + tensor model_model_layers_8_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_8_post_attention_layernorm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(686247360)))]; + tensor var_325_cast_fp16 = layer_norm(axes = var_325_axes_0, epsilon = var_37_to_fp16, gamma = model_model_layers_8_post_attention_layernorm_weight_to_fp16, x = input_7_cast_fp16)[name = string("op_325_cast_fp16")]; + tensor var_332 = const()[name = string("op_332"), val = tensor([0, 2, 1])]; + tensor input_9_axes_0 = const()[name = string("input_9_axes_0"), val = tensor([2])]; + tensor var_333 = transpose(perm = var_332, x = var_325_cast_fp16)[name = string("transpose_22")]; + tensor input_9 = expand_dims(axes = input_9_axes_0, x = var_333)[name = string("input_9")]; + string input_11_pad_type_0 = const()[name = string("input_11_pad_type_0"), val = string("valid")]; + tensor input_11_strides_0 = const()[name = string("input_11_strides_0"), val = tensor([1, 1])]; + tensor input_11_pad_0 = const()[name = string("input_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_11_dilations_0 = const()[name = string("input_11_dilations_0"), val = tensor([1, 1])]; + int32 input_11_groups_0 = const()[name = string("input_11_groups_0"), val = int32(1)]; + tensor input_11 = conv(dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = model_model_layers_8_mlp_gate_proj_weight_palettized, x = input_9)[name = string("input_11")]; + string up_states_1_pad_type_0 = const()[name = string("up_states_1_pad_type_0"), val = string("valid")]; + tensor up_states_1_strides_0 = const()[name = string("up_states_1_strides_0"), val = tensor([1, 1])]; + tensor up_states_1_pad_0 = const()[name = string("up_states_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor up_states_1_dilations_0 = const()[name = string("up_states_1_dilations_0"), val = tensor([1, 1])]; + int32 up_states_1_groups_0 = const()[name = string("up_states_1_groups_0"), val = int32(1)]; + tensor up_states_1 = conv(dilations = up_states_1_dilations_0, groups = up_states_1_groups_0, pad = up_states_1_pad_0, pad_type = up_states_1_pad_type_0, strides = up_states_1_strides_0, weight = model_model_layers_8_mlp_up_proj_weight_palettized, x = input_9)[name = string("up_states_1")]; + tensor gate_states_1 = silu(x = input_11)[name = string("gate_states_1")]; + tensor input_13 = mul(x = gate_states_1, y = up_states_1)[name = string("input_13")]; + string hidden_states_7_pad_type_0 = const()[name = string("hidden_states_7_pad_type_0"), val = string("valid")]; + tensor hidden_states_7_strides_0 = const()[name = string("hidden_states_7_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_7_pad_0 = const()[name = string("hidden_states_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_7_dilations_0 = const()[name = string("hidden_states_7_dilations_0"), val = tensor([1, 1])]; + int32 hidden_states_7_groups_0 = const()[name = string("hidden_states_7_groups_0"), val = int32(1)]; + tensor hidden_states_7 = conv(dilations = hidden_states_7_dilations_0, groups = hidden_states_7_groups_0, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = hidden_states_7_strides_0, weight = model_model_layers_8_mlp_down_proj_weight_palettized, x = input_13)[name = string("hidden_states_7")]; + tensor var_355_axes_0 = const()[name = string("op_355_axes_0"), val = tensor([2])]; + tensor var_355 = squeeze(axes = var_355_axes_0, x = hidden_states_7)[name = string("op_355")]; + tensor var_356 = const()[name = string("op_356"), val = tensor([0, 2, 1])]; + tensor var_357 = transpose(perm = var_356, x = var_355)[name = string("transpose_21")]; + tensor hidden_states_9_cast_fp16 = add(x = hidden_states_5_cast_fp16, y = var_357)[name = string("hidden_states_9_cast_fp16")]; + tensor mean_5_axes_0 = const()[name = string("mean_5_axes_0"), val = tensor([-1])]; + bool mean_5_keep_dims_0 = const()[name = string("mean_5_keep_dims_0"), val = bool(true)]; + tensor mean_5_cast_fp16 = reduce_mean(axes = mean_5_axes_0, keep_dims = mean_5_keep_dims_0, x = hidden_states_9_cast_fp16)[name = string("mean_5_cast_fp16")]; + tensor input_15_cast_fp16 = sub(x = hidden_states_9_cast_fp16, y = mean_5_cast_fp16)[name = string("input_15_cast_fp16")]; + tensor var_365_axes_0 = const()[name = string("op_365_axes_0"), val = tensor([-1])]; + tensor model_model_layers_9_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_9_input_layernorm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(686255616)))]; + tensor var_365_cast_fp16 = layer_norm(axes = var_365_axes_0, epsilon = var_37_to_fp16, gamma = model_model_layers_9_input_layernorm_weight_to_fp16, x = input_15_cast_fp16)[name = string("op_365_cast_fp16")]; + tensor var_369 = const()[name = string("op_369"), val = tensor([0, 2, 1])]; + tensor var_371_axes_0 = const()[name = string("op_371_axes_0"), val = tensor([2])]; + tensor var_370 = transpose(perm = var_369, x = var_365_cast_fp16)[name = string("transpose_20")]; + tensor var_371 = expand_dims(axes = var_371_axes_0, x = var_370)[name = string("op_371")]; + string query_states_5_pad_type_0 = const()[name = string("query_states_5_pad_type_0"), val = string("valid")]; + tensor query_states_5_strides_0 = const()[name = string("query_states_5_strides_0"), val = tensor([1, 1])]; + tensor query_states_5_pad_0 = const()[name = string("query_states_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_states_5_dilations_0 = const()[name = string("query_states_5_dilations_0"), val = tensor([1, 1])]; + int32 query_states_5_groups_0 = const()[name = string("query_states_5_groups_0"), val = int32(1)]; + tensor query_states_5 = conv(dilations = query_states_5_dilations_0, groups = query_states_5_groups_0, pad = query_states_5_pad_0, pad_type = query_states_5_pad_type_0, strides = query_states_5_strides_0, weight = model_model_layers_9_self_attn_q_proj_weight_palettized, x = var_371)[name = string("query_states_5")]; + string key_states_7_pad_type_0 = const()[name = string("key_states_7_pad_type_0"), val = string("valid")]; + tensor key_states_7_strides_0 = const()[name = string("key_states_7_strides_0"), val = tensor([1, 1])]; + tensor key_states_7_pad_0 = const()[name = string("key_states_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_states_7_dilations_0 = const()[name = string("key_states_7_dilations_0"), val = tensor([1, 1])]; + int32 key_states_7_groups_0 = const()[name = string("key_states_7_groups_0"), val = int32(1)]; + tensor key_states_7 = conv(dilations = key_states_7_dilations_0, groups = key_states_7_groups_0, pad = key_states_7_pad_0, pad_type = key_states_7_pad_type_0, strides = key_states_7_strides_0, weight = model_model_layers_9_self_attn_k_proj_weight_palettized, x = var_371)[name = string("key_states_7")]; + string value_states_7_pad_type_0 = const()[name = string("value_states_7_pad_type_0"), val = string("valid")]; + tensor value_states_7_strides_0 = const()[name = string("value_states_7_strides_0"), val = tensor([1, 1])]; + tensor value_states_7_pad_0 = const()[name = string("value_states_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_states_7_dilations_0 = const()[name = string("value_states_7_dilations_0"), val = tensor([1, 1])]; + int32 value_states_7_groups_0 = const()[name = string("value_states_7_groups_0"), val = int32(1)]; + tensor value_states_7 = conv(dilations = value_states_7_dilations_0, groups = value_states_7_groups_0, pad = value_states_7_pad_0, pad_type = value_states_7_pad_type_0, strides = value_states_7_strides_0, weight = model_model_layers_9_self_attn_v_proj_weight_palettized, x = var_371)[name = string("value_states_7")]; + tensor var_391 = const()[name = string("op_391"), val = tensor([1, 32, 128, 256])]; + tensor var_392 = reshape(shape = var_391, x = query_states_5)[name = string("op_392")]; + tensor var_393 = const()[name = string("op_393"), val = tensor([0, 1, 3, 2])]; + tensor var_395 = const()[name = string("op_395"), val = tensor([1, 8, 128, 256])]; + tensor var_396 = reshape(shape = var_395, x = key_states_7)[name = string("op_396")]; + tensor var_397 = const()[name = string("op_397"), val = tensor([0, 1, 3, 2])]; + tensor var_399 = const()[name = string("op_399"), val = tensor([1, 8, 128, 256])]; + tensor var_400 = reshape(shape = var_399, x = value_states_7)[name = string("op_400")]; + tensor var_401 = const()[name = string("op_401"), val = tensor([0, 1, 3, 2])]; + tensor x1_5_begin_0 = const()[name = string("x1_5_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_5_end_0 = const()[name = string("x1_5_end_0"), val = tensor([1, 32, 256, 64])]; + tensor x1_5_end_mask_0 = const()[name = string("x1_5_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_29 = transpose(perm = var_393, x = var_392)[name = string("transpose_19")]; + tensor x1_5 = slice_by_index(begin = x1_5_begin_0, end = x1_5_end_0, end_mask = x1_5_end_mask_0, x = x_29)[name = string("x1_5")]; + tensor x2_5_begin_0 = const()[name = string("x2_5_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_5_end_0 = const()[name = string("x2_5_end_0"), val = tensor([1, 32, 256, 128])]; + tensor x2_5_end_mask_0 = const()[name = string("x2_5_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_5 = slice_by_index(begin = x2_5_begin_0, end = x2_5_end_0, end_mask = x2_5_end_mask_0, x = x_29)[name = string("x2_5")]; + tensor var_419 = mul(x = x1_5, y = cos_7)[name = string("op_419")]; + tensor var_420 = mul(x = x2_5, y = sin_7)[name = string("op_420")]; + tensor var_421 = sub(x = var_419, y = var_420)[name = string("op_421")]; + tensor var_422 = mul(x = x2_5, y = cos_7)[name = string("op_422")]; + tensor var_423 = mul(x = x1_5, y = sin_7)[name = string("op_423")]; + tensor var_424 = add(x = var_422, y = var_423)[name = string("op_424")]; + bool rotated_5_interleave_0 = const()[name = string("rotated_5_interleave_0"), val = bool(false)]; + tensor rotated_5 = concat(axis = var_35, interleave = rotated_5_interleave_0, values = (var_421, var_424))[name = string("rotated_5")]; + tensor x1_7_begin_0 = const()[name = string("x1_7_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_7_end_0 = const()[name = string("x1_7_end_0"), val = tensor([1, 8, 256, 64])]; + tensor x1_7_end_mask_0 = const()[name = string("x1_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_33 = transpose(perm = var_397, x = var_396)[name = string("transpose_18")]; + tensor x1_7 = slice_by_index(begin = x1_7_begin_0, end = x1_7_end_0, end_mask = x1_7_end_mask_0, x = x_33)[name = string("x1_7")]; + tensor x2_7_begin_0 = const()[name = string("x2_7_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_7_end_0 = const()[name = string("x2_7_end_0"), val = tensor([1, 8, 256, 128])]; + tensor x2_7_end_mask_0 = const()[name = string("x2_7_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_7 = slice_by_index(begin = x2_7_begin_0, end = x2_7_end_0, end_mask = x2_7_end_mask_0, x = x_33)[name = string("x2_7")]; + tensor var_440 = mul(x = x1_7, y = cos_7)[name = string("op_440")]; + tensor var_441 = mul(x = x2_7, y = sin_7)[name = string("op_441")]; + tensor var_442 = sub(x = var_440, y = var_441)[name = string("op_442")]; + tensor var_443 = mul(x = x2_7, y = cos_7)[name = string("op_443")]; + tensor var_444 = mul(x = x1_7, y = sin_7)[name = string("op_444")]; + tensor var_445 = add(x = var_443, y = var_444)[name = string("op_445")]; + bool rotated_7_interleave_0 = const()[name = string("rotated_7_interleave_0"), val = bool(false)]; + tensor rotated_7 = concat(axis = var_35, interleave = rotated_7_interleave_0, values = (var_442, var_445))[name = string("rotated_7")]; + tensor expand_dims_12 = const()[name = string("expand_dims_12"), val = tensor([9])]; + tensor expand_dims_13 = const()[name = string("expand_dims_13"), val = tensor([0])]; + tensor expand_dims_15 = const()[name = string("expand_dims_15"), val = tensor([0])]; + tensor expand_dims_16 = const()[name = string("expand_dims_16"), val = tensor([10])]; + int32 concat_20_axis_0 = const()[name = string("concat_20_axis_0"), val = int32(0)]; + bool concat_20_interleave_0 = const()[name = string("concat_20_interleave_0"), val = bool(false)]; + tensor concat_20 = concat(axis = concat_20_axis_0, interleave = concat_20_interleave_0, values = (expand_dims_12, expand_dims_13, current_pos, expand_dims_15))[name = string("concat_20")]; + tensor concat_21_values1_0 = const()[name = string("concat_21_values1_0"), val = tensor([0])]; + tensor concat_21_values3_0 = const()[name = string("concat_21_values3_0"), val = tensor([0])]; + int32 concat_21_axis_0 = const()[name = string("concat_21_axis_0"), val = int32(0)]; + bool concat_21_interleave_0 = const()[name = string("concat_21_interleave_0"), val = bool(false)]; + tensor concat_21 = concat(axis = concat_21_axis_0, interleave = concat_21_interleave_0, values = (expand_dims_16, concat_21_values1_0, var_257, concat_21_values3_0))[name = string("concat_21")]; + tensor model_model_kv_cache_0_internal_tensor_assign_3_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_3_stride_0"), val = tensor([1, 1, 1, 1])]; + tensor model_model_kv_cache_0_internal_tensor_assign_3_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_3_begin_mask_0"), val = tensor([false, false, false, false])]; + tensor model_model_kv_cache_0_internal_tensor_assign_3_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_3_end_mask_0"), val = tensor([false, true, false, true])]; + tensor model_model_kv_cache_0_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_3_squeeze_mask_0"), val = tensor([false, false, false, false])]; + tensor model_model_kv_cache_0_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_20, begin_mask = model_model_kv_cache_0_internal_tensor_assign_3_begin_mask_0, end = concat_21, end_mask = model_model_kv_cache_0_internal_tensor_assign_3_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_3_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_3_stride_0, update = rotated_7, x = coreml_update_state_9)[name = string("model_model_kv_cache_0_internal_tensor_assign_3_cast_fp16")]; + write_state(data = model_model_kv_cache_0_internal_tensor_assign_3_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_10_write_state")]; + tensor coreml_update_state_10 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_10")]; + tensor expand_dims_18 = const()[name = string("expand_dims_18"), val = tensor([41])]; + tensor expand_dims_19 = const()[name = string("expand_dims_19"), val = tensor([0])]; + tensor expand_dims_21 = const()[name = string("expand_dims_21"), val = tensor([0])]; + tensor expand_dims_22 = const()[name = string("expand_dims_22"), val = tensor([42])]; + int32 concat_24_axis_0 = const()[name = string("concat_24_axis_0"), val = int32(0)]; + bool concat_24_interleave_0 = const()[name = string("concat_24_interleave_0"), val = bool(false)]; + tensor concat_24 = concat(axis = concat_24_axis_0, interleave = concat_24_interleave_0, values = (expand_dims_18, expand_dims_19, current_pos, expand_dims_21))[name = string("concat_24")]; + tensor concat_25_values1_0 = const()[name = string("concat_25_values1_0"), val = tensor([0])]; + tensor concat_25_values3_0 = const()[name = string("concat_25_values3_0"), val = tensor([0])]; + int32 concat_25_axis_0 = const()[name = string("concat_25_axis_0"), val = int32(0)]; + bool concat_25_interleave_0 = const()[name = string("concat_25_interleave_0"), val = bool(false)]; + tensor concat_25 = concat(axis = concat_25_axis_0, interleave = concat_25_interleave_0, values = (expand_dims_22, concat_25_values1_0, var_257, concat_25_values3_0))[name = string("concat_25")]; + tensor model_model_kv_cache_0_internal_tensor_assign_4_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_4_stride_0"), val = tensor([1, 1, 1, 1])]; + tensor model_model_kv_cache_0_internal_tensor_assign_4_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_4_begin_mask_0"), val = tensor([false, false, false, false])]; + tensor model_model_kv_cache_0_internal_tensor_assign_4_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_4_end_mask_0"), val = tensor([false, true, false, true])]; + tensor model_model_kv_cache_0_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_4_squeeze_mask_0"), val = tensor([false, false, false, false])]; + tensor value_states_9 = transpose(perm = var_401, x = var_400)[name = string("transpose_17")]; + tensor model_model_kv_cache_0_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_24, begin_mask = model_model_kv_cache_0_internal_tensor_assign_4_begin_mask_0, end = concat_25, end_mask = model_model_kv_cache_0_internal_tensor_assign_4_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_4_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_4_stride_0, update = value_states_9, x = coreml_update_state_10)[name = string("model_model_kv_cache_0_internal_tensor_assign_4_cast_fp16")]; + write_state(data = model_model_kv_cache_0_internal_tensor_assign_4_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_11_write_state")]; + tensor coreml_update_state_11 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_11")]; + tensor var_468_begin_0 = const()[name = string("op_468_begin_0"), val = tensor([9, 0, 0, 0])]; + tensor var_468_end_0 = const()[name = string("op_468_end_0"), val = tensor([10, 8, 1024, 128])]; + tensor var_468_end_mask_0 = const()[name = string("op_468_end_mask_0"), val = tensor([false, true, true, true])]; + tensor var_468_cast_fp16 = slice_by_index(begin = var_468_begin_0, end = var_468_end_0, end_mask = var_468_end_mask_0, x = coreml_update_state_11)[name = string("op_468_cast_fp16")]; + tensor K_layer_cache_3_axes_0 = const()[name = string("K_layer_cache_3_axes_0"), val = tensor([0])]; + tensor K_layer_cache_3_cast_fp16 = squeeze(axes = K_layer_cache_3_axes_0, x = var_468_cast_fp16)[name = string("K_layer_cache_3_cast_fp16")]; + tensor var_470_begin_0 = const()[name = string("op_470_begin_0"), val = tensor([41, 0, 0, 0])]; + tensor var_470_end_0 = const()[name = string("op_470_end_0"), val = tensor([42, 8, 1024, 128])]; + tensor var_470_end_mask_0 = const()[name = string("op_470_end_mask_0"), val = tensor([false, true, true, true])]; + tensor var_470_cast_fp16 = slice_by_index(begin = var_470_begin_0, end = var_470_end_0, end_mask = var_470_end_mask_0, x = coreml_update_state_11)[name = string("op_470_cast_fp16")]; + tensor V_layer_cache_3_axes_0 = const()[name = string("V_layer_cache_3_axes_0"), val = tensor([0])]; + tensor V_layer_cache_3_cast_fp16 = squeeze(axes = V_layer_cache_3_axes_0, x = var_470_cast_fp16)[name = string("V_layer_cache_3_cast_fp16")]; + tensor x_39_axes_0 = const()[name = string("x_39_axes_0"), val = tensor([1])]; + tensor x_39_cast_fp16 = expand_dims(axes = x_39_axes_0, x = K_layer_cache_3_cast_fp16)[name = string("x_39_cast_fp16")]; + tensor var_479 = const()[name = string("op_479"), val = tensor([1, 4, 1, 1])]; + tensor x_41_cast_fp16 = tile(reps = var_479, x = x_39_cast_fp16)[name = string("x_41_cast_fp16")]; + tensor var_483 = const()[name = string("op_483"), val = tensor([1, -1, 1024, 128])]; + tensor var_484_cast_fp16 = reshape(shape = var_483, x = x_41_cast_fp16)[name = string("op_484_cast_fp16")]; + tensor x_45_axes_0 = const()[name = string("x_45_axes_0"), val = tensor([1])]; + tensor x_45_cast_fp16 = expand_dims(axes = x_45_axes_0, x = V_layer_cache_3_cast_fp16)[name = string("x_45_cast_fp16")]; + tensor var_486 = const()[name = string("op_486"), val = tensor([1, 4, 1, 1])]; + tensor x_47_cast_fp16 = tile(reps = var_486, x = x_45_cast_fp16)[name = string("x_47_cast_fp16")]; + bool var_493_transpose_x_0 = const()[name = string("op_493_transpose_x_0"), val = bool(false)]; + bool var_493_transpose_y_0 = const()[name = string("op_493_transpose_y_0"), val = bool(true)]; + tensor var_493_cast_fp16 = matmul(transpose_x = var_493_transpose_x_0, transpose_y = var_493_transpose_y_0, x = rotated_5, y = var_484_cast_fp16)[name = string("op_493_cast_fp16")]; + fp16 var_494_to_fp16 = const()[name = string("op_494_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor attn_weights_3_cast_fp16 = mul(x = var_493_cast_fp16, y = var_494_to_fp16)[name = string("attn_weights_3_cast_fp16")]; + tensor x_49_cast_fp16 = add(x = attn_weights_3_cast_fp16, y = causal_mask)[name = string("x_49_cast_fp16")]; + tensor reduce_max_1_axes_0 = const()[name = string("reduce_max_1_axes_0"), val = tensor([-1])]; + bool reduce_max_1_keep_dims_0 = const()[name = string("reduce_max_1_keep_dims_0"), val = bool(true)]; + tensor reduce_max_1_cast_fp16 = reduce_max(axes = reduce_max_1_axes_0, keep_dims = reduce_max_1_keep_dims_0, x = x_49_cast_fp16)[name = string("reduce_max_1_cast_fp16")]; + tensor x_51_cast_fp16 = sub(x = x_49_cast_fp16, y = reduce_max_1_cast_fp16)[name = string("x_51_cast_fp16")]; + tensor exp_x_3_cast_fp16 = exp(x = x_51_cast_fp16)[name = string("exp_x_3_cast_fp16")]; + tensor var_505_axes_0 = const()[name = string("op_505_axes_0"), val = tensor([-1])]; + bool var_505_keep_dims_0 = const()[name = string("op_505_keep_dims_0"), val = bool(true)]; + tensor var_505_cast_fp16 = reduce_sum(axes = var_505_axes_0, keep_dims = var_505_keep_dims_0, x = exp_x_3_cast_fp16)[name = string("op_505_cast_fp16")]; + tensor var_506_cast_fp16 = real_div(x = exp_x_3_cast_fp16, y = var_505_cast_fp16)[name = string("op_506_cast_fp16")]; + tensor concat_30 = const()[name = string("concat_30"), val = tensor([32, 256, 1024])]; + tensor reshape_3_cast_fp16 = reshape(shape = concat_30, x = var_506_cast_fp16)[name = string("reshape_3_cast_fp16")]; + tensor concat_31 = const()[name = string("concat_31"), val = tensor([32, 1024, 128])]; + tensor reshape_4_cast_fp16 = reshape(shape = concat_31, x = x_47_cast_fp16)[name = string("reshape_4_cast_fp16")]; + bool matmul_1_transpose_x_0 = const()[name = string("matmul_1_transpose_x_0"), val = bool(false)]; + bool matmul_1_transpose_y_0 = const()[name = string("matmul_1_transpose_y_0"), val = bool(false)]; + tensor matmul_1_cast_fp16 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = reshape_3_cast_fp16, y = reshape_4_cast_fp16)[name = string("matmul_1_cast_fp16")]; + tensor concat_35 = const()[name = string("concat_35"), val = tensor([1, 32, 256, 128])]; + tensor reshape_5_cast_fp16 = reshape(shape = concat_35, x = matmul_1_cast_fp16)[name = string("reshape_5_cast_fp16")]; + tensor var_509_perm_0 = const()[name = string("op_509_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_511 = const()[name = string("op_511"), val = tensor([1, 256, 4096])]; + tensor var_509_cast_fp16 = transpose(perm = var_509_perm_0, x = reshape_5_cast_fp16)[name = string("transpose_16")]; + tensor input_19_cast_fp16 = reshape(shape = var_511, x = var_509_cast_fp16)[name = string("input_19_cast_fp16")]; + tensor model_model_layers_9_self_attn_o_proj_weight_promoted_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(686263872))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(698846848))))[name = string("model_model_layers_9_self_attn_o_proj_weight_promoted_to_fp16_palettized")]; + tensor linear_1_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_9_self_attn_o_proj_weight_promoted_to_fp16_palettized, x = input_19_cast_fp16)[name = string("linear_1_cast_fp16")]; + tensor hidden_states_13_cast_fp16 = add(x = hidden_states_9_cast_fp16, y = linear_1_cast_fp16)[name = string("hidden_states_13_cast_fp16")]; + tensor mean_7_axes_0 = const()[name = string("mean_7_axes_0"), val = tensor([-1])]; + bool mean_7_keep_dims_0 = const()[name = string("mean_7_keep_dims_0"), val = bool(true)]; + tensor mean_7_cast_fp16 = reduce_mean(axes = mean_7_axes_0, keep_dims = mean_7_keep_dims_0, x = hidden_states_13_cast_fp16)[name = string("mean_7_cast_fp16")]; + tensor input_21_cast_fp16 = sub(x = hidden_states_13_cast_fp16, y = mean_7_cast_fp16)[name = string("input_21_cast_fp16")]; + tensor var_522_axes_0 = const()[name = string("op_522_axes_0"), val = tensor([-1])]; + tensor model_model_layers_9_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_9_post_attention_layernorm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(698912448)))]; + tensor var_522_cast_fp16 = layer_norm(axes = var_522_axes_0, epsilon = var_37_to_fp16, gamma = model_model_layers_9_post_attention_layernorm_weight_to_fp16, x = input_21_cast_fp16)[name = string("op_522_cast_fp16")]; + tensor var_529 = const()[name = string("op_529"), val = tensor([0, 2, 1])]; + tensor input_23_axes_0 = const()[name = string("input_23_axes_0"), val = tensor([2])]; + tensor var_530 = transpose(perm = var_529, x = var_522_cast_fp16)[name = string("transpose_15")]; + tensor input_23 = expand_dims(axes = input_23_axes_0, x = var_530)[name = string("input_23")]; + string input_25_pad_type_0 = const()[name = string("input_25_pad_type_0"), val = string("valid")]; + tensor input_25_strides_0 = const()[name = string("input_25_strides_0"), val = tensor([1, 1])]; + tensor input_25_pad_0 = const()[name = string("input_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_25_dilations_0 = const()[name = string("input_25_dilations_0"), val = tensor([1, 1])]; + int32 input_25_groups_0 = const()[name = string("input_25_groups_0"), val = int32(1)]; + tensor input_25 = conv(dilations = input_25_dilations_0, groups = input_25_groups_0, pad = input_25_pad_0, pad_type = input_25_pad_type_0, strides = input_25_strides_0, weight = model_model_layers_9_mlp_gate_proj_weight_palettized, x = input_23)[name = string("input_25")]; + string up_states_3_pad_type_0 = const()[name = string("up_states_3_pad_type_0"), val = string("valid")]; + tensor up_states_3_strides_0 = const()[name = string("up_states_3_strides_0"), val = tensor([1, 1])]; + tensor up_states_3_pad_0 = const()[name = string("up_states_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor up_states_3_dilations_0 = const()[name = string("up_states_3_dilations_0"), val = tensor([1, 1])]; + int32 up_states_3_groups_0 = const()[name = string("up_states_3_groups_0"), val = int32(1)]; + tensor up_states_3 = conv(dilations = up_states_3_dilations_0, groups = up_states_3_groups_0, pad = up_states_3_pad_0, pad_type = up_states_3_pad_type_0, strides = up_states_3_strides_0, weight = model_model_layers_9_mlp_up_proj_weight_palettized, x = input_23)[name = string("up_states_3")]; + tensor gate_states_3 = silu(x = input_25)[name = string("gate_states_3")]; + tensor input_27 = mul(x = gate_states_3, y = up_states_3)[name = string("input_27")]; + string hidden_states_15_pad_type_0 = const()[name = string("hidden_states_15_pad_type_0"), val = string("valid")]; + tensor hidden_states_15_strides_0 = const()[name = string("hidden_states_15_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_15_pad_0 = const()[name = string("hidden_states_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_15_dilations_0 = const()[name = string("hidden_states_15_dilations_0"), val = tensor([1, 1])]; + int32 hidden_states_15_groups_0 = const()[name = string("hidden_states_15_groups_0"), val = int32(1)]; + tensor hidden_states_15 = conv(dilations = hidden_states_15_dilations_0, groups = hidden_states_15_groups_0, pad = hidden_states_15_pad_0, pad_type = hidden_states_15_pad_type_0, strides = hidden_states_15_strides_0, weight = model_model_layers_9_mlp_down_proj_weight_palettized, x = input_27)[name = string("hidden_states_15")]; + tensor var_552_axes_0 = const()[name = string("op_552_axes_0"), val = tensor([2])]; + tensor var_552 = squeeze(axes = var_552_axes_0, x = hidden_states_15)[name = string("op_552")]; + tensor var_553 = const()[name = string("op_553"), val = tensor([0, 2, 1])]; + tensor var_554 = transpose(perm = var_553, x = var_552)[name = string("transpose_14")]; + tensor hidden_states_17_cast_fp16 = add(x = hidden_states_13_cast_fp16, y = var_554)[name = string("hidden_states_17_cast_fp16")]; + tensor mean_9_axes_0 = const()[name = string("mean_9_axes_0"), val = tensor([-1])]; + bool mean_9_keep_dims_0 = const()[name = string("mean_9_keep_dims_0"), val = bool(true)]; + tensor mean_9_cast_fp16 = reduce_mean(axes = mean_9_axes_0, keep_dims = mean_9_keep_dims_0, x = hidden_states_17_cast_fp16)[name = string("mean_9_cast_fp16")]; + tensor input_29_cast_fp16 = sub(x = hidden_states_17_cast_fp16, y = mean_9_cast_fp16)[name = string("input_29_cast_fp16")]; + tensor var_562_axes_0 = const()[name = string("op_562_axes_0"), val = tensor([-1])]; + tensor model_model_layers_10_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_10_input_layernorm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(698920704)))]; + tensor var_562_cast_fp16 = layer_norm(axes = var_562_axes_0, epsilon = var_37_to_fp16, gamma = model_model_layers_10_input_layernorm_weight_to_fp16, x = input_29_cast_fp16)[name = string("op_562_cast_fp16")]; + tensor var_566 = const()[name = string("op_566"), val = tensor([0, 2, 1])]; + tensor var_568_axes_0 = const()[name = string("op_568_axes_0"), val = tensor([2])]; + tensor var_567 = transpose(perm = var_566, x = var_562_cast_fp16)[name = string("transpose_13")]; + tensor var_568 = expand_dims(axes = var_568_axes_0, x = var_567)[name = string("op_568")]; + string query_states_9_pad_type_0 = const()[name = string("query_states_9_pad_type_0"), val = string("valid")]; + tensor query_states_9_strides_0 = const()[name = string("query_states_9_strides_0"), val = tensor([1, 1])]; + tensor query_states_9_pad_0 = const()[name = string("query_states_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_states_9_dilations_0 = const()[name = string("query_states_9_dilations_0"), val = tensor([1, 1])]; + int32 query_states_9_groups_0 = const()[name = string("query_states_9_groups_0"), val = int32(1)]; + tensor query_states_9 = conv(dilations = query_states_9_dilations_0, groups = query_states_9_groups_0, pad = query_states_9_pad_0, pad_type = query_states_9_pad_type_0, strides = query_states_9_strides_0, weight = model_model_layers_10_self_attn_q_proj_weight_palettized, x = var_568)[name = string("query_states_9")]; + string key_states_13_pad_type_0 = const()[name = string("key_states_13_pad_type_0"), val = string("valid")]; + tensor key_states_13_strides_0 = const()[name = string("key_states_13_strides_0"), val = tensor([1, 1])]; + tensor key_states_13_pad_0 = const()[name = string("key_states_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_states_13_dilations_0 = const()[name = string("key_states_13_dilations_0"), val = tensor([1, 1])]; + int32 key_states_13_groups_0 = const()[name = string("key_states_13_groups_0"), val = int32(1)]; + tensor key_states_13 = conv(dilations = key_states_13_dilations_0, groups = key_states_13_groups_0, pad = key_states_13_pad_0, pad_type = key_states_13_pad_type_0, strides = key_states_13_strides_0, weight = model_model_layers_10_self_attn_k_proj_weight_palettized, x = var_568)[name = string("key_states_13")]; + string value_states_13_pad_type_0 = const()[name = string("value_states_13_pad_type_0"), val = string("valid")]; + tensor value_states_13_strides_0 = const()[name = string("value_states_13_strides_0"), val = tensor([1, 1])]; + tensor value_states_13_pad_0 = const()[name = string("value_states_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_states_13_dilations_0 = const()[name = string("value_states_13_dilations_0"), val = tensor([1, 1])]; + int32 value_states_13_groups_0 = const()[name = string("value_states_13_groups_0"), val = int32(1)]; + tensor value_states_13 = conv(dilations = value_states_13_dilations_0, groups = value_states_13_groups_0, pad = value_states_13_pad_0, pad_type = value_states_13_pad_type_0, strides = value_states_13_strides_0, weight = model_model_layers_10_self_attn_v_proj_weight_palettized, x = var_568)[name = string("value_states_13")]; + tensor var_588 = const()[name = string("op_588"), val = tensor([1, 32, 128, 256])]; + tensor var_589 = reshape(shape = var_588, x = query_states_9)[name = string("op_589")]; + tensor var_590 = const()[name = string("op_590"), val = tensor([0, 1, 3, 2])]; + tensor var_592 = const()[name = string("op_592"), val = tensor([1, 8, 128, 256])]; + tensor var_593 = reshape(shape = var_592, x = key_states_13)[name = string("op_593")]; + tensor var_594 = const()[name = string("op_594"), val = tensor([0, 1, 3, 2])]; + tensor var_596 = const()[name = string("op_596"), val = tensor([1, 8, 128, 256])]; + tensor var_597 = reshape(shape = var_596, x = value_states_13)[name = string("op_597")]; + tensor var_598 = const()[name = string("op_598"), val = tensor([0, 1, 3, 2])]; + tensor x1_9_begin_0 = const()[name = string("x1_9_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_9_end_0 = const()[name = string("x1_9_end_0"), val = tensor([1, 32, 256, 64])]; + tensor x1_9_end_mask_0 = const()[name = string("x1_9_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_57 = transpose(perm = var_590, x = var_589)[name = string("transpose_12")]; + tensor x1_9 = slice_by_index(begin = x1_9_begin_0, end = x1_9_end_0, end_mask = x1_9_end_mask_0, x = x_57)[name = string("x1_9")]; + tensor x2_9_begin_0 = const()[name = string("x2_9_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_9_end_0 = const()[name = string("x2_9_end_0"), val = tensor([1, 32, 256, 128])]; + tensor x2_9_end_mask_0 = const()[name = string("x2_9_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_9 = slice_by_index(begin = x2_9_begin_0, end = x2_9_end_0, end_mask = x2_9_end_mask_0, x = x_57)[name = string("x2_9")]; + tensor var_616 = mul(x = x1_9, y = cos_7)[name = string("op_616")]; + tensor var_617 = mul(x = x2_9, y = sin_7)[name = string("op_617")]; + tensor var_618 = sub(x = var_616, y = var_617)[name = string("op_618")]; + tensor var_619 = mul(x = x2_9, y = cos_7)[name = string("op_619")]; + tensor var_620 = mul(x = x1_9, y = sin_7)[name = string("op_620")]; + tensor var_621 = add(x = var_619, y = var_620)[name = string("op_621")]; + bool rotated_9_interleave_0 = const()[name = string("rotated_9_interleave_0"), val = bool(false)]; + tensor rotated_9 = concat(axis = var_35, interleave = rotated_9_interleave_0, values = (var_618, var_621))[name = string("rotated_9")]; + tensor x1_11_begin_0 = const()[name = string("x1_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_11_end_0 = const()[name = string("x1_11_end_0"), val = tensor([1, 8, 256, 64])]; + tensor x1_11_end_mask_0 = const()[name = string("x1_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_61 = transpose(perm = var_594, x = var_593)[name = string("transpose_11")]; + tensor x1_11 = slice_by_index(begin = x1_11_begin_0, end = x1_11_end_0, end_mask = x1_11_end_mask_0, x = x_61)[name = string("x1_11")]; + tensor x2_11_begin_0 = const()[name = string("x2_11_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_11_end_0 = const()[name = string("x2_11_end_0"), val = tensor([1, 8, 256, 128])]; + tensor x2_11_end_mask_0 = const()[name = string("x2_11_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_11 = slice_by_index(begin = x2_11_begin_0, end = x2_11_end_0, end_mask = x2_11_end_mask_0, x = x_61)[name = string("x2_11")]; + tensor var_637 = mul(x = x1_11, y = cos_7)[name = string("op_637")]; + tensor var_638 = mul(x = x2_11, y = sin_7)[name = string("op_638")]; + tensor var_639 = sub(x = var_637, y = var_638)[name = string("op_639")]; + tensor var_640 = mul(x = x2_11, y = cos_7)[name = string("op_640")]; + tensor var_641 = mul(x = x1_11, y = sin_7)[name = string("op_641")]; + tensor var_642 = add(x = var_640, y = var_641)[name = string("op_642")]; + bool rotated_11_interleave_0 = const()[name = string("rotated_11_interleave_0"), val = bool(false)]; + tensor rotated_11 = concat(axis = var_35, interleave = rotated_11_interleave_0, values = (var_639, var_642))[name = string("rotated_11")]; + tensor expand_dims_24 = const()[name = string("expand_dims_24"), val = tensor([10])]; + tensor expand_dims_25 = const()[name = string("expand_dims_25"), val = tensor([0])]; + tensor expand_dims_27 = const()[name = string("expand_dims_27"), val = tensor([0])]; + tensor expand_dims_28 = const()[name = string("expand_dims_28"), val = tensor([11])]; + int32 concat_38_axis_0 = const()[name = string("concat_38_axis_0"), val = int32(0)]; + bool concat_38_interleave_0 = const()[name = string("concat_38_interleave_0"), val = bool(false)]; + tensor concat_38 = concat(axis = concat_38_axis_0, interleave = concat_38_interleave_0, values = (expand_dims_24, expand_dims_25, current_pos, expand_dims_27))[name = string("concat_38")]; + tensor concat_39_values1_0 = const()[name = string("concat_39_values1_0"), val = tensor([0])]; + tensor concat_39_values3_0 = const()[name = string("concat_39_values3_0"), val = tensor([0])]; + int32 concat_39_axis_0 = const()[name = string("concat_39_axis_0"), val = int32(0)]; + bool concat_39_interleave_0 = const()[name = string("concat_39_interleave_0"), val = bool(false)]; + tensor concat_39 = concat(axis = concat_39_axis_0, interleave = concat_39_interleave_0, values = (expand_dims_28, concat_39_values1_0, var_257, concat_39_values3_0))[name = string("concat_39")]; + tensor model_model_kv_cache_0_internal_tensor_assign_5_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_5_stride_0"), val = tensor([1, 1, 1, 1])]; + tensor model_model_kv_cache_0_internal_tensor_assign_5_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_5_begin_mask_0"), val = tensor([false, false, false, false])]; + tensor model_model_kv_cache_0_internal_tensor_assign_5_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_5_end_mask_0"), val = tensor([false, true, false, true])]; + tensor model_model_kv_cache_0_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_5_squeeze_mask_0"), val = tensor([false, false, false, false])]; + tensor model_model_kv_cache_0_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_38, begin_mask = model_model_kv_cache_0_internal_tensor_assign_5_begin_mask_0, end = concat_39, end_mask = model_model_kv_cache_0_internal_tensor_assign_5_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_5_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_5_stride_0, update = rotated_11, x = coreml_update_state_11)[name = string("model_model_kv_cache_0_internal_tensor_assign_5_cast_fp16")]; + write_state(data = model_model_kv_cache_0_internal_tensor_assign_5_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_12_write_state")]; + tensor coreml_update_state_12 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_12")]; + tensor expand_dims_30 = const()[name = string("expand_dims_30"), val = tensor([42])]; + tensor expand_dims_31 = const()[name = string("expand_dims_31"), val = tensor([0])]; + tensor expand_dims_33 = const()[name = string("expand_dims_33"), val = tensor([0])]; + tensor expand_dims_34 = const()[name = string("expand_dims_34"), val = tensor([43])]; + int32 concat_42_axis_0 = const()[name = string("concat_42_axis_0"), val = int32(0)]; + bool concat_42_interleave_0 = const()[name = string("concat_42_interleave_0"), val = bool(false)]; + tensor concat_42 = concat(axis = concat_42_axis_0, interleave = concat_42_interleave_0, values = (expand_dims_30, expand_dims_31, current_pos, expand_dims_33))[name = string("concat_42")]; + tensor concat_43_values1_0 = const()[name = string("concat_43_values1_0"), val = tensor([0])]; + tensor concat_43_values3_0 = const()[name = string("concat_43_values3_0"), val = tensor([0])]; + int32 concat_43_axis_0 = const()[name = string("concat_43_axis_0"), val = int32(0)]; + bool concat_43_interleave_0 = const()[name = string("concat_43_interleave_0"), val = bool(false)]; + tensor concat_43 = concat(axis = concat_43_axis_0, interleave = concat_43_interleave_0, values = (expand_dims_34, concat_43_values1_0, var_257, concat_43_values3_0))[name = string("concat_43")]; + tensor model_model_kv_cache_0_internal_tensor_assign_6_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_6_stride_0"), val = tensor([1, 1, 1, 1])]; + tensor model_model_kv_cache_0_internal_tensor_assign_6_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_6_begin_mask_0"), val = tensor([false, false, false, false])]; + tensor model_model_kv_cache_0_internal_tensor_assign_6_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_6_end_mask_0"), val = tensor([false, true, false, true])]; + tensor model_model_kv_cache_0_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_6_squeeze_mask_0"), val = tensor([false, false, false, false])]; + tensor value_states_15 = transpose(perm = var_598, x = var_597)[name = string("transpose_10")]; + tensor model_model_kv_cache_0_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_42, begin_mask = model_model_kv_cache_0_internal_tensor_assign_6_begin_mask_0, end = concat_43, end_mask = model_model_kv_cache_0_internal_tensor_assign_6_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_6_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_6_stride_0, update = value_states_15, x = coreml_update_state_12)[name = string("model_model_kv_cache_0_internal_tensor_assign_6_cast_fp16")]; + write_state(data = model_model_kv_cache_0_internal_tensor_assign_6_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_13_write_state")]; + tensor coreml_update_state_13 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_13")]; + tensor var_665_begin_0 = const()[name = string("op_665_begin_0"), val = tensor([10, 0, 0, 0])]; + tensor var_665_end_0 = const()[name = string("op_665_end_0"), val = tensor([11, 8, 1024, 128])]; + tensor var_665_end_mask_0 = const()[name = string("op_665_end_mask_0"), val = tensor([false, true, true, true])]; + tensor var_665_cast_fp16 = slice_by_index(begin = var_665_begin_0, end = var_665_end_0, end_mask = var_665_end_mask_0, x = coreml_update_state_13)[name = string("op_665_cast_fp16")]; + tensor K_layer_cache_5_axes_0 = const()[name = string("K_layer_cache_5_axes_0"), val = tensor([0])]; + tensor K_layer_cache_5_cast_fp16 = squeeze(axes = K_layer_cache_5_axes_0, x = var_665_cast_fp16)[name = string("K_layer_cache_5_cast_fp16")]; + tensor var_667_begin_0 = const()[name = string("op_667_begin_0"), val = tensor([42, 0, 0, 0])]; + tensor var_667_end_0 = const()[name = string("op_667_end_0"), val = tensor([43, 8, 1024, 128])]; + tensor var_667_end_mask_0 = const()[name = string("op_667_end_mask_0"), val = tensor([false, true, true, true])]; + tensor var_667_cast_fp16 = slice_by_index(begin = var_667_begin_0, end = var_667_end_0, end_mask = var_667_end_mask_0, x = coreml_update_state_13)[name = string("op_667_cast_fp16")]; + tensor V_layer_cache_5_axes_0 = const()[name = string("V_layer_cache_5_axes_0"), val = tensor([0])]; + tensor V_layer_cache_5_cast_fp16 = squeeze(axes = V_layer_cache_5_axes_0, x = var_667_cast_fp16)[name = string("V_layer_cache_5_cast_fp16")]; + tensor x_67_axes_0 = const()[name = string("x_67_axes_0"), val = tensor([1])]; + tensor x_67_cast_fp16 = expand_dims(axes = x_67_axes_0, x = K_layer_cache_5_cast_fp16)[name = string("x_67_cast_fp16")]; + tensor var_676 = const()[name = string("op_676"), val = tensor([1, 4, 1, 1])]; + tensor x_69_cast_fp16 = tile(reps = var_676, x = x_67_cast_fp16)[name = string("x_69_cast_fp16")]; + tensor var_680 = const()[name = string("op_680"), val = tensor([1, -1, 1024, 128])]; + tensor var_681_cast_fp16 = reshape(shape = var_680, x = x_69_cast_fp16)[name = string("op_681_cast_fp16")]; + tensor x_73_axes_0 = const()[name = string("x_73_axes_0"), val = tensor([1])]; + tensor x_73_cast_fp16 = expand_dims(axes = x_73_axes_0, x = V_layer_cache_5_cast_fp16)[name = string("x_73_cast_fp16")]; + tensor var_683 = const()[name = string("op_683"), val = tensor([1, 4, 1, 1])]; + tensor x_75_cast_fp16 = tile(reps = var_683, x = x_73_cast_fp16)[name = string("x_75_cast_fp16")]; + bool var_690_transpose_x_0 = const()[name = string("op_690_transpose_x_0"), val = bool(false)]; + bool var_690_transpose_y_0 = const()[name = string("op_690_transpose_y_0"), val = bool(true)]; + tensor var_690_cast_fp16 = matmul(transpose_x = var_690_transpose_x_0, transpose_y = var_690_transpose_y_0, x = rotated_9, y = var_681_cast_fp16)[name = string("op_690_cast_fp16")]; + fp16 var_691_to_fp16 = const()[name = string("op_691_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor attn_weights_5_cast_fp16 = mul(x = var_690_cast_fp16, y = var_691_to_fp16)[name = string("attn_weights_5_cast_fp16")]; + tensor x_77_cast_fp16 = add(x = attn_weights_5_cast_fp16, y = causal_mask)[name = string("x_77_cast_fp16")]; + tensor reduce_max_2_axes_0 = const()[name = string("reduce_max_2_axes_0"), val = tensor([-1])]; + bool reduce_max_2_keep_dims_0 = const()[name = string("reduce_max_2_keep_dims_0"), val = bool(true)]; + tensor reduce_max_2_cast_fp16 = reduce_max(axes = reduce_max_2_axes_0, keep_dims = reduce_max_2_keep_dims_0, x = x_77_cast_fp16)[name = string("reduce_max_2_cast_fp16")]; + tensor x_79_cast_fp16 = sub(x = x_77_cast_fp16, y = reduce_max_2_cast_fp16)[name = string("x_79_cast_fp16")]; + tensor exp_x_5_cast_fp16 = exp(x = x_79_cast_fp16)[name = string("exp_x_5_cast_fp16")]; + tensor var_702_axes_0 = const()[name = string("op_702_axes_0"), val = tensor([-1])]; + bool var_702_keep_dims_0 = const()[name = string("op_702_keep_dims_0"), val = bool(true)]; + tensor var_702_cast_fp16 = reduce_sum(axes = var_702_axes_0, keep_dims = var_702_keep_dims_0, x = exp_x_5_cast_fp16)[name = string("op_702_cast_fp16")]; + tensor var_703_cast_fp16 = real_div(x = exp_x_5_cast_fp16, y = var_702_cast_fp16)[name = string("op_703_cast_fp16")]; + tensor concat_48 = const()[name = string("concat_48"), val = tensor([32, 256, 1024])]; + tensor reshape_6_cast_fp16 = reshape(shape = concat_48, x = var_703_cast_fp16)[name = string("reshape_6_cast_fp16")]; + tensor concat_49 = const()[name = string("concat_49"), val = tensor([32, 1024, 128])]; + tensor reshape_7_cast_fp16 = reshape(shape = concat_49, x = x_75_cast_fp16)[name = string("reshape_7_cast_fp16")]; + bool matmul_2_transpose_x_0 = const()[name = string("matmul_2_transpose_x_0"), val = bool(false)]; + bool matmul_2_transpose_y_0 = const()[name = string("matmul_2_transpose_y_0"), val = bool(false)]; + tensor matmul_2_cast_fp16 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = reshape_6_cast_fp16, y = reshape_7_cast_fp16)[name = string("matmul_2_cast_fp16")]; + tensor concat_53 = const()[name = string("concat_53"), val = tensor([1, 32, 256, 128])]; + tensor reshape_8_cast_fp16 = reshape(shape = concat_53, x = matmul_2_cast_fp16)[name = string("reshape_8_cast_fp16")]; + tensor var_706_perm_0 = const()[name = string("op_706_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_708 = const()[name = string("op_708"), val = tensor([1, 256, 4096])]; + tensor var_706_cast_fp16 = transpose(perm = var_706_perm_0, x = reshape_8_cast_fp16)[name = string("transpose_9")]; + tensor input_33_cast_fp16 = reshape(shape = var_708, x = var_706_cast_fp16)[name = string("input_33_cast_fp16")]; + tensor model_model_layers_10_self_attn_o_proj_weight_promoted_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(698928960))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(711511936))))[name = string("model_model_layers_10_self_attn_o_proj_weight_promoted_to_fp16_palettized")]; + tensor linear_2_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_10_self_attn_o_proj_weight_promoted_to_fp16_palettized, x = input_33_cast_fp16)[name = string("linear_2_cast_fp16")]; + tensor hidden_states_21_cast_fp16 = add(x = hidden_states_17_cast_fp16, y = linear_2_cast_fp16)[name = string("hidden_states_21_cast_fp16")]; + tensor mean_11_axes_0 = const()[name = string("mean_11_axes_0"), val = tensor([-1])]; + bool mean_11_keep_dims_0 = const()[name = string("mean_11_keep_dims_0"), val = bool(true)]; + tensor mean_11_cast_fp16 = reduce_mean(axes = mean_11_axes_0, keep_dims = mean_11_keep_dims_0, x = hidden_states_21_cast_fp16)[name = string("mean_11_cast_fp16")]; + tensor input_35_cast_fp16 = sub(x = hidden_states_21_cast_fp16, y = mean_11_cast_fp16)[name = string("input_35_cast_fp16")]; + tensor var_719_axes_0 = const()[name = string("op_719_axes_0"), val = tensor([-1])]; + tensor model_model_layers_10_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_10_post_attention_layernorm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(711577536)))]; + tensor var_719_cast_fp16 = layer_norm(axes = var_719_axes_0, epsilon = var_37_to_fp16, gamma = model_model_layers_10_post_attention_layernorm_weight_to_fp16, x = input_35_cast_fp16)[name = string("op_719_cast_fp16")]; + tensor var_726 = const()[name = string("op_726"), val = tensor([0, 2, 1])]; + tensor input_37_axes_0 = const()[name = string("input_37_axes_0"), val = tensor([2])]; + tensor var_727 = transpose(perm = var_726, x = var_719_cast_fp16)[name = string("transpose_8")]; + tensor input_37 = expand_dims(axes = input_37_axes_0, x = var_727)[name = string("input_37")]; + string input_39_pad_type_0 = const()[name = string("input_39_pad_type_0"), val = string("valid")]; + tensor input_39_strides_0 = const()[name = string("input_39_strides_0"), val = tensor([1, 1])]; + tensor input_39_pad_0 = const()[name = string("input_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_39_dilations_0 = const()[name = string("input_39_dilations_0"), val = tensor([1, 1])]; + int32 input_39_groups_0 = const()[name = string("input_39_groups_0"), val = int32(1)]; + tensor input_39 = conv(dilations = input_39_dilations_0, groups = input_39_groups_0, pad = input_39_pad_0, pad_type = input_39_pad_type_0, strides = input_39_strides_0, weight = model_model_layers_10_mlp_gate_proj_weight_palettized, x = input_37)[name = string("input_39")]; + string up_states_5_pad_type_0 = const()[name = string("up_states_5_pad_type_0"), val = string("valid")]; + tensor up_states_5_strides_0 = const()[name = string("up_states_5_strides_0"), val = tensor([1, 1])]; + tensor up_states_5_pad_0 = const()[name = string("up_states_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor up_states_5_dilations_0 = const()[name = string("up_states_5_dilations_0"), val = tensor([1, 1])]; + int32 up_states_5_groups_0 = const()[name = string("up_states_5_groups_0"), val = int32(1)]; + tensor up_states_5 = conv(dilations = up_states_5_dilations_0, groups = up_states_5_groups_0, pad = up_states_5_pad_0, pad_type = up_states_5_pad_type_0, strides = up_states_5_strides_0, weight = model_model_layers_10_mlp_up_proj_weight_palettized, x = input_37)[name = string("up_states_5")]; + tensor gate_states_5 = silu(x = input_39)[name = string("gate_states_5")]; + tensor input_41 = mul(x = gate_states_5, y = up_states_5)[name = string("input_41")]; + string hidden_states_23_pad_type_0 = const()[name = string("hidden_states_23_pad_type_0"), val = string("valid")]; + tensor hidden_states_23_strides_0 = const()[name = string("hidden_states_23_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_23_pad_0 = const()[name = string("hidden_states_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_23_dilations_0 = const()[name = string("hidden_states_23_dilations_0"), val = tensor([1, 1])]; + int32 hidden_states_23_groups_0 = const()[name = string("hidden_states_23_groups_0"), val = int32(1)]; + tensor hidden_states_23 = conv(dilations = hidden_states_23_dilations_0, groups = hidden_states_23_groups_0, pad = hidden_states_23_pad_0, pad_type = hidden_states_23_pad_type_0, strides = hidden_states_23_strides_0, weight = model_model_layers_10_mlp_down_proj_weight_palettized, x = input_41)[name = string("hidden_states_23")]; + tensor var_749_axes_0 = const()[name = string("op_749_axes_0"), val = tensor([2])]; + tensor var_749 = squeeze(axes = var_749_axes_0, x = hidden_states_23)[name = string("op_749")]; + tensor var_750 = const()[name = string("op_750"), val = tensor([0, 2, 1])]; + tensor var_751 = transpose(perm = var_750, x = var_749)[name = string("transpose_7")]; + tensor hidden_states_25_cast_fp16 = add(x = hidden_states_21_cast_fp16, y = var_751)[name = string("hidden_states_25_cast_fp16")]; + tensor mean_13_axes_0 = const()[name = string("mean_13_axes_0"), val = tensor([-1])]; + bool mean_13_keep_dims_0 = const()[name = string("mean_13_keep_dims_0"), val = bool(true)]; + tensor mean_13_cast_fp16 = reduce_mean(axes = mean_13_axes_0, keep_dims = mean_13_keep_dims_0, x = hidden_states_25_cast_fp16)[name = string("mean_13_cast_fp16")]; + tensor input_43_cast_fp16 = sub(x = hidden_states_25_cast_fp16, y = mean_13_cast_fp16)[name = string("input_43_cast_fp16")]; + tensor var_759_axes_0 = const()[name = string("op_759_axes_0"), val = tensor([-1])]; + tensor model_model_layers_11_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_11_input_layernorm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(711585792)))]; + tensor var_759_cast_fp16 = layer_norm(axes = var_759_axes_0, epsilon = var_37_to_fp16, gamma = model_model_layers_11_input_layernorm_weight_to_fp16, x = input_43_cast_fp16)[name = string("op_759_cast_fp16")]; + tensor var_763 = const()[name = string("op_763"), val = tensor([0, 2, 1])]; + tensor var_765_axes_0 = const()[name = string("op_765_axes_0"), val = tensor([2])]; + tensor var_764 = transpose(perm = var_763, x = var_759_cast_fp16)[name = string("transpose_6")]; + tensor var_765 = expand_dims(axes = var_765_axes_0, x = var_764)[name = string("op_765")]; + string query_states_13_pad_type_0 = const()[name = string("query_states_13_pad_type_0"), val = string("valid")]; + tensor query_states_13_strides_0 = const()[name = string("query_states_13_strides_0"), val = tensor([1, 1])]; + tensor query_states_13_pad_0 = const()[name = string("query_states_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_states_13_dilations_0 = const()[name = string("query_states_13_dilations_0"), val = tensor([1, 1])]; + int32 query_states_13_groups_0 = const()[name = string("query_states_13_groups_0"), val = int32(1)]; + tensor query_states_13 = conv(dilations = query_states_13_dilations_0, groups = query_states_13_groups_0, pad = query_states_13_pad_0, pad_type = query_states_13_pad_type_0, strides = query_states_13_strides_0, weight = model_model_layers_11_self_attn_q_proj_weight_palettized, x = var_765)[name = string("query_states_13")]; + string key_states_19_pad_type_0 = const()[name = string("key_states_19_pad_type_0"), val = string("valid")]; + tensor key_states_19_strides_0 = const()[name = string("key_states_19_strides_0"), val = tensor([1, 1])]; + tensor key_states_19_pad_0 = const()[name = string("key_states_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_states_19_dilations_0 = const()[name = string("key_states_19_dilations_0"), val = tensor([1, 1])]; + int32 key_states_19_groups_0 = const()[name = string("key_states_19_groups_0"), val = int32(1)]; + tensor key_states_19 = conv(dilations = key_states_19_dilations_0, groups = key_states_19_groups_0, pad = key_states_19_pad_0, pad_type = key_states_19_pad_type_0, strides = key_states_19_strides_0, weight = model_model_layers_11_self_attn_k_proj_weight_palettized, x = var_765)[name = string("key_states_19")]; + string value_states_19_pad_type_0 = const()[name = string("value_states_19_pad_type_0"), val = string("valid")]; + tensor value_states_19_strides_0 = const()[name = string("value_states_19_strides_0"), val = tensor([1, 1])]; + tensor value_states_19_pad_0 = const()[name = string("value_states_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_states_19_dilations_0 = const()[name = string("value_states_19_dilations_0"), val = tensor([1, 1])]; + int32 value_states_19_groups_0 = const()[name = string("value_states_19_groups_0"), val = int32(1)]; + tensor value_states_19 = conv(dilations = value_states_19_dilations_0, groups = value_states_19_groups_0, pad = value_states_19_pad_0, pad_type = value_states_19_pad_type_0, strides = value_states_19_strides_0, weight = model_model_layers_11_self_attn_v_proj_weight_palettized, x = var_765)[name = string("value_states_19")]; + tensor var_785 = const()[name = string("op_785"), val = tensor([1, 32, 128, 256])]; + tensor var_786 = reshape(shape = var_785, x = query_states_13)[name = string("op_786")]; + tensor var_787 = const()[name = string("op_787"), val = tensor([0, 1, 3, 2])]; + tensor var_789 = const()[name = string("op_789"), val = tensor([1, 8, 128, 256])]; + tensor var_790 = reshape(shape = var_789, x = key_states_19)[name = string("op_790")]; + tensor var_791 = const()[name = string("op_791"), val = tensor([0, 1, 3, 2])]; + tensor var_793 = const()[name = string("op_793"), val = tensor([1, 8, 128, 256])]; + tensor var_794 = reshape(shape = var_793, x = value_states_19)[name = string("op_794")]; + tensor var_795 = const()[name = string("op_795"), val = tensor([0, 1, 3, 2])]; + tensor x1_13_begin_0 = const()[name = string("x1_13_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_13_end_0 = const()[name = string("x1_13_end_0"), val = tensor([1, 32, 256, 64])]; + tensor x1_13_end_mask_0 = const()[name = string("x1_13_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_85 = transpose(perm = var_787, x = var_786)[name = string("transpose_5")]; + tensor x1_13 = slice_by_index(begin = x1_13_begin_0, end = x1_13_end_0, end_mask = x1_13_end_mask_0, x = x_85)[name = string("x1_13")]; + tensor x2_13_begin_0 = const()[name = string("x2_13_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_13_end_0 = const()[name = string("x2_13_end_0"), val = tensor([1, 32, 256, 128])]; + tensor x2_13_end_mask_0 = const()[name = string("x2_13_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_13 = slice_by_index(begin = x2_13_begin_0, end = x2_13_end_0, end_mask = x2_13_end_mask_0, x = x_85)[name = string("x2_13")]; + tensor var_813 = mul(x = x1_13, y = cos_7)[name = string("op_813")]; + tensor var_814 = mul(x = x2_13, y = sin_7)[name = string("op_814")]; + tensor var_815 = sub(x = var_813, y = var_814)[name = string("op_815")]; + tensor var_816 = mul(x = x2_13, y = cos_7)[name = string("op_816")]; + tensor var_817 = mul(x = x1_13, y = sin_7)[name = string("op_817")]; + tensor var_818 = add(x = var_816, y = var_817)[name = string("op_818")]; + bool rotated_13_interleave_0 = const()[name = string("rotated_13_interleave_0"), val = bool(false)]; + tensor rotated_13 = concat(axis = var_35, interleave = rotated_13_interleave_0, values = (var_815, var_818))[name = string("rotated_13")]; + tensor x1_begin_0 = const()[name = string("x1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_end_0 = const()[name = string("x1_end_0"), val = tensor([1, 8, 256, 64])]; + tensor x1_end_mask_0 = const()[name = string("x1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x_89 = transpose(perm = var_791, x = var_790)[name = string("transpose_4")]; + tensor x1 = slice_by_index(begin = x1_begin_0, end = x1_end_0, end_mask = x1_end_mask_0, x = x_89)[name = string("x1")]; + tensor x2_begin_0 = const()[name = string("x2_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_end_0 = const()[name = string("x2_end_0"), val = tensor([1, 8, 256, 128])]; + tensor x2_end_mask_0 = const()[name = string("x2_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2 = slice_by_index(begin = x2_begin_0, end = x2_end_0, end_mask = x2_end_mask_0, x = x_89)[name = string("x2")]; + tensor var_834 = mul(x = x1, y = cos_7)[name = string("op_834")]; + tensor var_835 = mul(x = x2, y = sin_7)[name = string("op_835")]; + tensor var_836 = sub(x = var_834, y = var_835)[name = string("op_836")]; + tensor var_837 = mul(x = x2, y = cos_7)[name = string("op_837")]; + tensor var_838 = mul(x = x1, y = sin_7)[name = string("op_838")]; + tensor var_839 = add(x = var_837, y = var_838)[name = string("op_839")]; + bool rotated_interleave_0 = const()[name = string("rotated_interleave_0"), val = bool(false)]; + tensor rotated = concat(axis = var_35, interleave = rotated_interleave_0, values = (var_836, var_839))[name = string("rotated")]; + tensor expand_dims_36 = const()[name = string("expand_dims_36"), val = tensor([11])]; + tensor expand_dims_37 = const()[name = string("expand_dims_37"), val = tensor([0])]; + tensor expand_dims_39 = const()[name = string("expand_dims_39"), val = tensor([0])]; + tensor expand_dims_40 = const()[name = string("expand_dims_40"), val = tensor([12])]; + int32 concat_56_axis_0 = const()[name = string("concat_56_axis_0"), val = int32(0)]; + bool concat_56_interleave_0 = const()[name = string("concat_56_interleave_0"), val = bool(false)]; + tensor concat_56 = concat(axis = concat_56_axis_0, interleave = concat_56_interleave_0, values = (expand_dims_36, expand_dims_37, current_pos, expand_dims_39))[name = string("concat_56")]; + tensor concat_57_values1_0 = const()[name = string("concat_57_values1_0"), val = tensor([0])]; + tensor concat_57_values3_0 = const()[name = string("concat_57_values3_0"), val = tensor([0])]; + int32 concat_57_axis_0 = const()[name = string("concat_57_axis_0"), val = int32(0)]; + bool concat_57_interleave_0 = const()[name = string("concat_57_interleave_0"), val = bool(false)]; + tensor concat_57 = concat(axis = concat_57_axis_0, interleave = concat_57_interleave_0, values = (expand_dims_40, concat_57_values1_0, var_257, concat_57_values3_0))[name = string("concat_57")]; + tensor model_model_kv_cache_0_internal_tensor_assign_7_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_7_stride_0"), val = tensor([1, 1, 1, 1])]; + tensor model_model_kv_cache_0_internal_tensor_assign_7_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_7_begin_mask_0"), val = tensor([false, false, false, false])]; + tensor model_model_kv_cache_0_internal_tensor_assign_7_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_7_end_mask_0"), val = tensor([false, true, false, true])]; + tensor model_model_kv_cache_0_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_7_squeeze_mask_0"), val = tensor([false, false, false, false])]; + tensor model_model_kv_cache_0_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_56, begin_mask = model_model_kv_cache_0_internal_tensor_assign_7_begin_mask_0, end = concat_57, end_mask = model_model_kv_cache_0_internal_tensor_assign_7_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_7_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_7_stride_0, update = rotated, x = coreml_update_state_13)[name = string("model_model_kv_cache_0_internal_tensor_assign_7_cast_fp16")]; + write_state(data = model_model_kv_cache_0_internal_tensor_assign_7_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_14_write_state")]; + tensor coreml_update_state_14 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_14")]; + tensor expand_dims_42 = const()[name = string("expand_dims_42"), val = tensor([43])]; + tensor expand_dims_43 = const()[name = string("expand_dims_43"), val = tensor([0])]; + tensor expand_dims_45 = const()[name = string("expand_dims_45"), val = tensor([0])]; + tensor expand_dims_46 = const()[name = string("expand_dims_46"), val = tensor([44])]; + int32 concat_60_axis_0 = const()[name = string("concat_60_axis_0"), val = int32(0)]; + bool concat_60_interleave_0 = const()[name = string("concat_60_interleave_0"), val = bool(false)]; + tensor concat_60 = concat(axis = concat_60_axis_0, interleave = concat_60_interleave_0, values = (expand_dims_42, expand_dims_43, current_pos, expand_dims_45))[name = string("concat_60")]; + tensor concat_61_values1_0 = const()[name = string("concat_61_values1_0"), val = tensor([0])]; + tensor concat_61_values3_0 = const()[name = string("concat_61_values3_0"), val = tensor([0])]; + int32 concat_61_axis_0 = const()[name = string("concat_61_axis_0"), val = int32(0)]; + bool concat_61_interleave_0 = const()[name = string("concat_61_interleave_0"), val = bool(false)]; + tensor concat_61 = concat(axis = concat_61_axis_0, interleave = concat_61_interleave_0, values = (expand_dims_46, concat_61_values1_0, var_257, concat_61_values3_0))[name = string("concat_61")]; + tensor model_model_kv_cache_0_internal_tensor_assign_8_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_8_stride_0"), val = tensor([1, 1, 1, 1])]; + tensor model_model_kv_cache_0_internal_tensor_assign_8_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_8_begin_mask_0"), val = tensor([false, false, false, false])]; + tensor model_model_kv_cache_0_internal_tensor_assign_8_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_8_end_mask_0"), val = tensor([false, true, false, true])]; + tensor model_model_kv_cache_0_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_8_squeeze_mask_0"), val = tensor([false, false, false, false])]; + tensor value_states_21 = transpose(perm = var_795, x = var_794)[name = string("transpose_3")]; + tensor model_model_kv_cache_0_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_60, begin_mask = model_model_kv_cache_0_internal_tensor_assign_8_begin_mask_0, end = concat_61, end_mask = model_model_kv_cache_0_internal_tensor_assign_8_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_8_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_8_stride_0, update = value_states_21, x = coreml_update_state_14)[name = string("model_model_kv_cache_0_internal_tensor_assign_8_cast_fp16")]; + write_state(data = model_model_kv_cache_0_internal_tensor_assign_8_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_15_write_state")]; + tensor coreml_update_state_15 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_15")]; + tensor var_862_begin_0 = const()[name = string("op_862_begin_0"), val = tensor([11, 0, 0, 0])]; + tensor var_862_end_0 = const()[name = string("op_862_end_0"), val = tensor([12, 8, 1024, 128])]; + tensor var_862_end_mask_0 = const()[name = string("op_862_end_mask_0"), val = tensor([false, true, true, true])]; + tensor var_862_cast_fp16 = slice_by_index(begin = var_862_begin_0, end = var_862_end_0, end_mask = var_862_end_mask_0, x = coreml_update_state_15)[name = string("op_862_cast_fp16")]; + tensor K_layer_cache_axes_0 = const()[name = string("K_layer_cache_axes_0"), val = tensor([0])]; + tensor K_layer_cache_cast_fp16 = squeeze(axes = K_layer_cache_axes_0, x = var_862_cast_fp16)[name = string("K_layer_cache_cast_fp16")]; + tensor var_864_begin_0 = const()[name = string("op_864_begin_0"), val = tensor([43, 0, 0, 0])]; + tensor var_864_end_0 = const()[name = string("op_864_end_0"), val = tensor([44, 8, 1024, 128])]; + tensor var_864_end_mask_0 = const()[name = string("op_864_end_mask_0"), val = tensor([false, true, true, true])]; + tensor var_864_cast_fp16 = slice_by_index(begin = var_864_begin_0, end = var_864_end_0, end_mask = var_864_end_mask_0, x = coreml_update_state_15)[name = string("op_864_cast_fp16")]; + tensor V_layer_cache_axes_0 = const()[name = string("V_layer_cache_axes_0"), val = tensor([0])]; + tensor V_layer_cache_cast_fp16 = squeeze(axes = V_layer_cache_axes_0, x = var_864_cast_fp16)[name = string("V_layer_cache_cast_fp16")]; + tensor x_95_axes_0 = const()[name = string("x_95_axes_0"), val = tensor([1])]; + tensor x_95_cast_fp16 = expand_dims(axes = x_95_axes_0, x = K_layer_cache_cast_fp16)[name = string("x_95_cast_fp16")]; + tensor var_873 = const()[name = string("op_873"), val = tensor([1, 4, 1, 1])]; + tensor x_97_cast_fp16 = tile(reps = var_873, x = x_95_cast_fp16)[name = string("x_97_cast_fp16")]; + tensor var_877 = const()[name = string("op_877"), val = tensor([1, -1, 1024, 128])]; + tensor var_878_cast_fp16 = reshape(shape = var_877, x = x_97_cast_fp16)[name = string("op_878_cast_fp16")]; + tensor x_101_axes_0 = const()[name = string("x_101_axes_0"), val = tensor([1])]; + tensor x_101_cast_fp16 = expand_dims(axes = x_101_axes_0, x = V_layer_cache_cast_fp16)[name = string("x_101_cast_fp16")]; + tensor var_880 = const()[name = string("op_880"), val = tensor([1, 4, 1, 1])]; + tensor x_103_cast_fp16 = tile(reps = var_880, x = x_101_cast_fp16)[name = string("x_103_cast_fp16")]; + bool var_887_transpose_x_0 = const()[name = string("op_887_transpose_x_0"), val = bool(false)]; + bool var_887_transpose_y_0 = const()[name = string("op_887_transpose_y_0"), val = bool(true)]; + tensor var_887_cast_fp16 = matmul(transpose_x = var_887_transpose_x_0, transpose_y = var_887_transpose_y_0, x = rotated_13, y = var_878_cast_fp16)[name = string("op_887_cast_fp16")]; + fp16 var_888_to_fp16 = const()[name = string("op_888_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor attn_weights_cast_fp16 = mul(x = var_887_cast_fp16, y = var_888_to_fp16)[name = string("attn_weights_cast_fp16")]; + tensor x_105_cast_fp16 = add(x = attn_weights_cast_fp16, y = causal_mask)[name = string("x_105_cast_fp16")]; + tensor reduce_max_3_axes_0 = const()[name = string("reduce_max_3_axes_0"), val = tensor([-1])]; + bool reduce_max_3_keep_dims_0 = const()[name = string("reduce_max_3_keep_dims_0"), val = bool(true)]; + tensor reduce_max_3_cast_fp16 = reduce_max(axes = reduce_max_3_axes_0, keep_dims = reduce_max_3_keep_dims_0, x = x_105_cast_fp16)[name = string("reduce_max_3_cast_fp16")]; + tensor x_107_cast_fp16 = sub(x = x_105_cast_fp16, y = reduce_max_3_cast_fp16)[name = string("x_107_cast_fp16")]; + tensor exp_x_cast_fp16 = exp(x = x_107_cast_fp16)[name = string("exp_x_cast_fp16")]; + tensor var_899_axes_0 = const()[name = string("op_899_axes_0"), val = tensor([-1])]; + bool var_899_keep_dims_0 = const()[name = string("op_899_keep_dims_0"), val = bool(true)]; + tensor var_899_cast_fp16 = reduce_sum(axes = var_899_axes_0, keep_dims = var_899_keep_dims_0, x = exp_x_cast_fp16)[name = string("op_899_cast_fp16")]; + tensor var_900_cast_fp16 = real_div(x = exp_x_cast_fp16, y = var_899_cast_fp16)[name = string("op_900_cast_fp16")]; + tensor concat_66 = const()[name = string("concat_66"), val = tensor([32, 256, 1024])]; + tensor reshape_9_cast_fp16 = reshape(shape = concat_66, x = var_900_cast_fp16)[name = string("reshape_9_cast_fp16")]; + tensor concat_67 = const()[name = string("concat_67"), val = tensor([32, 1024, 128])]; + tensor reshape_10_cast_fp16 = reshape(shape = concat_67, x = x_103_cast_fp16)[name = string("reshape_10_cast_fp16")]; + bool matmul_3_transpose_x_0 = const()[name = string("matmul_3_transpose_x_0"), val = bool(false)]; + bool matmul_3_transpose_y_0 = const()[name = string("matmul_3_transpose_y_0"), val = bool(false)]; + tensor matmul_3_cast_fp16 = matmul(transpose_x = matmul_3_transpose_x_0, transpose_y = matmul_3_transpose_y_0, x = reshape_9_cast_fp16, y = reshape_10_cast_fp16)[name = string("matmul_3_cast_fp16")]; + tensor concat_71 = const()[name = string("concat_71"), val = tensor([1, 32, 256, 128])]; + tensor reshape_11_cast_fp16 = reshape(shape = concat_71, x = matmul_3_cast_fp16)[name = string("reshape_11_cast_fp16")]; + tensor var_903_perm_0 = const()[name = string("op_903_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_905 = const()[name = string("op_905"), val = tensor([1, 256, 4096])]; + tensor var_903_cast_fp16 = transpose(perm = var_903_perm_0, x = reshape_11_cast_fp16)[name = string("transpose_2")]; + tensor input_47_cast_fp16 = reshape(shape = var_905, x = var_903_cast_fp16)[name = string("input_47_cast_fp16")]; + tensor model_model_layers_11_self_attn_o_proj_weight_promoted_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(711594048))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(724177024))))[name = string("model_model_layers_11_self_attn_o_proj_weight_promoted_to_fp16_palettized")]; + tensor linear_3_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_11_self_attn_o_proj_weight_promoted_to_fp16_palettized, x = input_47_cast_fp16)[name = string("linear_3_cast_fp16")]; + tensor hidden_states_29_cast_fp16 = add(x = hidden_states_25_cast_fp16, y = linear_3_cast_fp16)[name = string("hidden_states_29_cast_fp16")]; + tensor mean_axes_0 = const()[name = string("mean_axes_0"), val = tensor([-1])]; + bool mean_keep_dims_0 = const()[name = string("mean_keep_dims_0"), val = bool(true)]; + tensor mean_cast_fp16 = reduce_mean(axes = mean_axes_0, keep_dims = mean_keep_dims_0, x = hidden_states_29_cast_fp16)[name = string("mean_cast_fp16")]; + tensor input_49_cast_fp16 = sub(x = hidden_states_29_cast_fp16, y = mean_cast_fp16)[name = string("input_49_cast_fp16")]; + tensor var_916_axes_0 = const()[name = string("op_916_axes_0"), val = tensor([-1])]; + tensor model_model_layers_11_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_11_post_attention_layernorm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(724242624)))]; + tensor var_916_cast_fp16 = layer_norm(axes = var_916_axes_0, epsilon = var_37_to_fp16, gamma = model_model_layers_11_post_attention_layernorm_weight_to_fp16, x = input_49_cast_fp16)[name = string("op_916_cast_fp16")]; + tensor var_923 = const()[name = string("op_923"), val = tensor([0, 2, 1])]; + tensor input_51_axes_0 = const()[name = string("input_51_axes_0"), val = tensor([2])]; + tensor var_924 = transpose(perm = var_923, x = var_916_cast_fp16)[name = string("transpose_1")]; + tensor input_51 = expand_dims(axes = input_51_axes_0, x = var_924)[name = string("input_51")]; + string input_53_pad_type_0 = const()[name = string("input_53_pad_type_0"), val = string("valid")]; + tensor input_53_strides_0 = const()[name = string("input_53_strides_0"), val = tensor([1, 1])]; + tensor input_53_pad_0 = const()[name = string("input_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_53_dilations_0 = const()[name = string("input_53_dilations_0"), val = tensor([1, 1])]; + int32 input_53_groups_0 = const()[name = string("input_53_groups_0"), val = int32(1)]; + tensor input_53 = conv(dilations = input_53_dilations_0, groups = input_53_groups_0, pad = input_53_pad_0, pad_type = input_53_pad_type_0, strides = input_53_strides_0, weight = model_model_layers_11_mlp_gate_proj_weight_palettized, x = input_51)[name = string("input_53")]; + string up_states_pad_type_0 = const()[name = string("up_states_pad_type_0"), val = string("valid")]; + tensor up_states_strides_0 = const()[name = string("up_states_strides_0"), val = tensor([1, 1])]; + tensor up_states_pad_0 = const()[name = string("up_states_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor up_states_dilations_0 = const()[name = string("up_states_dilations_0"), val = tensor([1, 1])]; + int32 up_states_groups_0 = const()[name = string("up_states_groups_0"), val = int32(1)]; + tensor up_states = conv(dilations = up_states_dilations_0, groups = up_states_groups_0, pad = up_states_pad_0, pad_type = up_states_pad_type_0, strides = up_states_strides_0, weight = model_model_layers_11_mlp_up_proj_weight_palettized, x = input_51)[name = string("up_states")]; + tensor gate_states = silu(x = input_53)[name = string("gate_states")]; + tensor input = mul(x = gate_states, y = up_states)[name = string("input")]; + string hidden_states_pad_type_0 = const()[name = string("hidden_states_pad_type_0"), val = string("valid")]; + tensor hidden_states_strides_0 = const()[name = string("hidden_states_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_pad_0 = const()[name = string("hidden_states_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_dilations_0 = const()[name = string("hidden_states_dilations_0"), val = tensor([1, 1])]; + int32 hidden_states_groups_0 = const()[name = string("hidden_states_groups_0"), val = int32(1)]; + tensor hidden_states_1 = conv(dilations = hidden_states_dilations_0, groups = hidden_states_groups_0, pad = hidden_states_pad_0, pad_type = hidden_states_pad_type_0, strides = hidden_states_strides_0, weight = model_model_layers_11_mlp_down_proj_weight_palettized, x = input)[name = string("hidden_states")]; + tensor var_946_axes_0 = const()[name = string("op_946_axes_0"), val = tensor([2])]; + tensor var_946 = squeeze(axes = var_946_axes_0, x = hidden_states_1)[name = string("op_946")]; + tensor var_947 = const()[name = string("op_947"), val = tensor([0, 2, 1])]; + tensor var_948 = transpose(perm = var_947, x = var_946)[name = string("transpose_0")]; + tensor output_hidden_states = add(x = hidden_states_29_cast_fp16, y = var_948)[name = string("op_949_cast_fp16")]; + } -> (output_hidden_states); +} \ No newline at end of file