diff --git "a/encoder.mlmodelc/model.mil" "b/encoder.mlmodelc/model.mil"
--- "a/encoder.mlmodelc/model.mil"
+++ "b/encoder.mlmodelc/model.mil"
@@ -1,172 +1,177 @@
 program(1.3)
-[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.8.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})]
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}})]
 {
     func main<ios19>(tensor<fp16, [1, 128]> attention_mask, tensor<int32, [1, 128]> input_ids) {
-            tensor<fp16, [768, 768, 1, 1]> encoder_layers_0_self_attn_q_proj_weight = const()[name = string("encoder_layers_0_self_attn_q_proj_weight"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_0_self_attn_k_proj_weight = const()[name = string("encoder_layers_0_self_attn_k_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1179776)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_0_self_attn_v_proj_weight = const()[name = string("encoder_layers_0_self_attn_v_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1573056)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_0_mlp_gate_proj_weight = const()[name = string("encoder_layers_0_mlp_gate_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1966336)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_0_mlp_up_proj_weight = const()[name = string("encoder_layers_0_mlp_up_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3735872)))];
-            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_0_mlp_down_proj_weight = const()[name = string("encoder_layers_0_mlp_down_proj_weight"), val = tensor<fp16, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5505408)))];
-            tensor<fp16, [768, 768, 1, 1]> encoder_layers_1_self_attn_q_proj_weight = const()[name = string("encoder_layers_1_self_attn_q_proj_weight"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7274944)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_1_self_attn_k_proj_weight = const()[name = string("encoder_layers_1_self_attn_k_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8454656)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_1_self_attn_v_proj_weight = const()[name = string("encoder_layers_1_self_attn_v_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8847936)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_1_mlp_gate_proj_weight = const()[name = string("encoder_layers_1_mlp_gate_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9241216)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_1_mlp_up_proj_weight = const()[name = string("encoder_layers_1_mlp_up_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11010752)))];
-            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_1_mlp_down_proj_weight = const()[name = string("encoder_layers_1_mlp_down_proj_weight"), val = tensor<fp16, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12780288)))];
-            tensor<fp16, [768, 768, 1, 1]> encoder_layers_2_self_attn_q_proj_weight = const()[name = string("encoder_layers_2_self_attn_q_proj_weight"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14549824)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_2_self_attn_k_proj_weight = const()[name = string("encoder_layers_2_self_attn_k_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15729536)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_2_self_attn_v_proj_weight = const()[name = string("encoder_layers_2_self_attn_v_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16122816)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_2_mlp_gate_proj_weight = const()[name = string("encoder_layers_2_mlp_gate_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16516096)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_2_mlp_up_proj_weight = const()[name = string("encoder_layers_2_mlp_up_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18285632)))];
-            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_2_mlp_down_proj_weight = const()[name = string("encoder_layers_2_mlp_down_proj_weight"), val = tensor<fp16, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20055168)))];
-            tensor<fp16, [768, 768, 1, 1]> encoder_layers_3_self_attn_q_proj_weight = const()[name = string("encoder_layers_3_self_attn_q_proj_weight"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21824704)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_3_self_attn_k_proj_weight = const()[name = string("encoder_layers_3_self_attn_k_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23004416)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_3_self_attn_v_proj_weight = const()[name = string("encoder_layers_3_self_attn_v_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23397696)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_3_mlp_gate_proj_weight = const()[name = string("encoder_layers_3_mlp_gate_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23790976)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_3_mlp_up_proj_weight = const()[name = string("encoder_layers_3_mlp_up_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25560512)))];
-            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_3_mlp_down_proj_weight = const()[name = string("encoder_layers_3_mlp_down_proj_weight"), val = tensor<fp16, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27330048)))];
-            tensor<fp16, [768, 768, 1, 1]> encoder_layers_4_self_attn_q_proj_weight = const()[name = string("encoder_layers_4_self_attn_q_proj_weight"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29099584)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_4_self_attn_k_proj_weight = const()[name = string("encoder_layers_4_self_attn_k_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30279296)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_4_self_attn_v_proj_weight = const()[name = string("encoder_layers_4_self_attn_v_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30672576)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_4_mlp_gate_proj_weight = const()[name = string("encoder_layers_4_mlp_gate_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31065856)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_4_mlp_up_proj_weight = const()[name = string("encoder_layers_4_mlp_up_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32835392)))];
-            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_4_mlp_down_proj_weight = const()[name = string("encoder_layers_4_mlp_down_proj_weight"), val = tensor<fp16, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34604928)))];
-            tensor<fp16, [768, 768, 1, 1]> encoder_layers_5_self_attn_q_proj_weight = const()[name = string("encoder_layers_5_self_attn_q_proj_weight"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36374464)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_5_self_attn_k_proj_weight = const()[name = string("encoder_layers_5_self_attn_k_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37554176)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_5_self_attn_v_proj_weight = const()[name = string("encoder_layers_5_self_attn_v_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37947456)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_5_mlp_gate_proj_weight = const()[name = string("encoder_layers_5_mlp_gate_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38340736)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_5_mlp_up_proj_weight = const()[name = string("encoder_layers_5_mlp_up_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40110272)))];
-            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_5_mlp_down_proj_weight = const()[name = string("encoder_layers_5_mlp_down_proj_weight"), val = tensor<fp16, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41879808)))];
-            tensor<fp16, [768, 768, 1, 1]> encoder_layers_6_self_attn_q_proj_weight = const()[name = string("encoder_layers_6_self_attn_q_proj_weight"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43649344)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_6_self_attn_k_proj_weight = const()[name = string("encoder_layers_6_self_attn_k_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44829056)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_6_self_attn_v_proj_weight = const()[name = string("encoder_layers_6_self_attn_v_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45222336)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_6_mlp_gate_proj_weight = const()[name = string("encoder_layers_6_mlp_gate_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45615616)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_6_mlp_up_proj_weight = const()[name = string("encoder_layers_6_mlp_up_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47385152)))];
-            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_6_mlp_down_proj_weight = const()[name = string("encoder_layers_6_mlp_down_proj_weight"), val = tensor<fp16, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49154688)))];
-            tensor<fp16, [768, 768, 1, 1]> encoder_layers_7_self_attn_q_proj_weight = const()[name = string("encoder_layers_7_self_attn_q_proj_weight"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50924224)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_7_self_attn_k_proj_weight = const()[name = string("encoder_layers_7_self_attn_k_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52103936)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_7_self_attn_v_proj_weight = const()[name = string("encoder_layers_7_self_attn_v_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52497216)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_7_mlp_gate_proj_weight = const()[name = string("encoder_layers_7_mlp_gate_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52890496)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_7_mlp_up_proj_weight = const()[name = string("encoder_layers_7_mlp_up_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54660032)))];
-            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_7_mlp_down_proj_weight = const()[name = string("encoder_layers_7_mlp_down_proj_weight"), val = tensor<fp16, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56429568)))];
-            tensor<fp16, [768, 768, 1, 1]> encoder_layers_8_self_attn_q_proj_weight = const()[name = string("encoder_layers_8_self_attn_q_proj_weight"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58199104)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_8_self_attn_k_proj_weight = const()[name = string("encoder_layers_8_self_attn_k_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59378816)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_8_self_attn_v_proj_weight = const()[name = string("encoder_layers_8_self_attn_v_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59772096)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_8_mlp_gate_proj_weight = const()[name = string("encoder_layers_8_mlp_gate_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60165376)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_8_mlp_up_proj_weight = const()[name = string("encoder_layers_8_mlp_up_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61934912)))];
-            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_8_mlp_down_proj_weight = const()[name = string("encoder_layers_8_mlp_down_proj_weight"), val = tensor<fp16, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63704448)))];
-            tensor<fp16, [768, 768, 1, 1]> encoder_layers_9_self_attn_q_proj_weight = const()[name = string("encoder_layers_9_self_attn_q_proj_weight"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65473984)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_9_self_attn_k_proj_weight = const()[name = string("encoder_layers_9_self_attn_k_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66653696)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_9_self_attn_v_proj_weight = const()[name = string("encoder_layers_9_self_attn_v_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67046976)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_9_mlp_gate_proj_weight = const()[name = string("encoder_layers_9_mlp_gate_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67440256)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_9_mlp_up_proj_weight = const()[name = string("encoder_layers_9_mlp_up_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69209792)))];
-            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_9_mlp_down_proj_weight = const()[name = string("encoder_layers_9_mlp_down_proj_weight"), val = tensor<fp16, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70979328)))];
-            tensor<fp16, [768, 768, 1, 1]> encoder_layers_10_self_attn_q_proj_weight = const()[name = string("encoder_layers_10_self_attn_q_proj_weight"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72748864)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_10_self_attn_k_proj_weight = const()[name = string("encoder_layers_10_self_attn_k_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73928576)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_10_self_attn_v_proj_weight = const()[name = string("encoder_layers_10_self_attn_v_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74321856)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_10_mlp_gate_proj_weight = const()[name = string("encoder_layers_10_mlp_gate_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74715136)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_10_mlp_up_proj_weight = const()[name = string("encoder_layers_10_mlp_up_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76484672)))];
-            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_10_mlp_down_proj_weight = const()[name = string("encoder_layers_10_mlp_down_proj_weight"), val = tensor<fp16, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78254208)))];
-            tensor<fp16, [768, 768, 1, 1]> encoder_layers_11_self_attn_q_proj_weight = const()[name = string("encoder_layers_11_self_attn_q_proj_weight"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80023744)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_11_self_attn_k_proj_weight = const()[name = string("encoder_layers_11_self_attn_k_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81203456)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_11_self_attn_v_proj_weight = const()[name = string("encoder_layers_11_self_attn_v_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81596736)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_11_mlp_gate_proj_weight = const()[name = string("encoder_layers_11_mlp_gate_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81990016)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_11_mlp_up_proj_weight = const()[name = string("encoder_layers_11_mlp_up_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83759552)))];
-            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_11_mlp_down_proj_weight = const()[name = string("encoder_layers_11_mlp_down_proj_weight"), val = tensor<fp16, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85529088)))];
-            tensor<fp16, [768, 768, 1, 1]> encoder_layers_12_self_attn_q_proj_weight = const()[name = string("encoder_layers_12_self_attn_q_proj_weight"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87298624)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_12_self_attn_k_proj_weight = const()[name = string("encoder_layers_12_self_attn_k_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88478336)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_12_self_attn_v_proj_weight = const()[name = string("encoder_layers_12_self_attn_v_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88871616)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_12_mlp_gate_proj_weight = const()[name = string("encoder_layers_12_mlp_gate_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89264896)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_12_mlp_up_proj_weight = const()[name = string("encoder_layers_12_mlp_up_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91034432)))];
-            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_12_mlp_down_proj_weight = const()[name = string("encoder_layers_12_mlp_down_proj_weight"), val = tensor<fp16, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92803968)))];
-            tensor<fp16, [768, 768, 1, 1]> encoder_layers_13_self_attn_q_proj_weight = const()[name = string("encoder_layers_13_self_attn_q_proj_weight"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94573504)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_13_self_attn_k_proj_weight = const()[name = string("encoder_layers_13_self_attn_k_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95753216)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_13_self_attn_v_proj_weight = const()[name = string("encoder_layers_13_self_attn_v_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96146496)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_13_mlp_gate_proj_weight = const()[name = string("encoder_layers_13_mlp_gate_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96539776)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_13_mlp_up_proj_weight = const()[name = string("encoder_layers_13_mlp_up_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98309312)))];
-            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_13_mlp_down_proj_weight = const()[name = string("encoder_layers_13_mlp_down_proj_weight"), val = tensor<fp16, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100078848)))];
-            tensor<fp16, [768, 768, 1, 1]> encoder_layers_14_self_attn_q_proj_weight = const()[name = string("encoder_layers_14_self_attn_q_proj_weight"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101848384)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_14_self_attn_k_proj_weight = const()[name = string("encoder_layers_14_self_attn_k_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103028096)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_14_self_attn_v_proj_weight = const()[name = string("encoder_layers_14_self_attn_v_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103421376)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_14_mlp_gate_proj_weight = const()[name = string("encoder_layers_14_mlp_gate_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103814656)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_14_mlp_up_proj_weight = const()[name = string("encoder_layers_14_mlp_up_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105584192)))];
-            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_14_mlp_down_proj_weight = const()[name = string("encoder_layers_14_mlp_down_proj_weight"), val = tensor<fp16, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107353728)))];
-            tensor<fp16, [768, 768, 1, 1]> encoder_layers_15_self_attn_q_proj_weight = const()[name = string("encoder_layers_15_self_attn_q_proj_weight"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109123264)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_15_self_attn_k_proj_weight = const()[name = string("encoder_layers_15_self_attn_k_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110302976)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_15_self_attn_v_proj_weight = const()[name = string("encoder_layers_15_self_attn_v_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110696256)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_15_mlp_gate_proj_weight = const()[name = string("encoder_layers_15_mlp_gate_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111089536)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_15_mlp_up_proj_weight = const()[name = string("encoder_layers_15_mlp_up_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(112859072)))];
-            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_15_mlp_down_proj_weight = const()[name = string("encoder_layers_15_mlp_down_proj_weight"), val = tensor<fp16, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114628608)))];
-            tensor<fp16, [768, 768, 1, 1]> encoder_layers_16_self_attn_q_proj_weight = const()[name = string("encoder_layers_16_self_attn_q_proj_weight"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116398144)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_16_self_attn_k_proj_weight = const()[name = string("encoder_layers_16_self_attn_k_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117577856)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_16_self_attn_v_proj_weight = const()[name = string("encoder_layers_16_self_attn_v_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117971136)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_16_mlp_gate_proj_weight = const()[name = string("encoder_layers_16_mlp_gate_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118364416)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_16_mlp_up_proj_weight = const()[name = string("encoder_layers_16_mlp_up_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120133952)))];
-            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_16_mlp_down_proj_weight = const()[name = string("encoder_layers_16_mlp_down_proj_weight"), val = tensor<fp16, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121903488)))];
-            tensor<fp16, [768, 768, 1, 1]> encoder_layers_17_self_attn_q_proj_weight = const()[name = string("encoder_layers_17_self_attn_q_proj_weight"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123673024)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_17_self_attn_k_proj_weight = const()[name = string("encoder_layers_17_self_attn_k_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124852736)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_17_self_attn_v_proj_weight = const()[name = string("encoder_layers_17_self_attn_v_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125246016)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_17_mlp_gate_proj_weight = const()[name = string("encoder_layers_17_mlp_gate_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125639296)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_17_mlp_up_proj_weight = const()[name = string("encoder_layers_17_mlp_up_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127408832)))];
-            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_17_mlp_down_proj_weight = const()[name = string("encoder_layers_17_mlp_down_proj_weight"), val = tensor<fp16, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129178368)))];
-            tensor<fp16, [768, 768, 1, 1]> encoder_layers_18_self_attn_q_proj_weight = const()[name = string("encoder_layers_18_self_attn_q_proj_weight"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130947904)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_18_self_attn_k_proj_weight = const()[name = string("encoder_layers_18_self_attn_k_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132127616)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_18_self_attn_v_proj_weight = const()[name = string("encoder_layers_18_self_attn_v_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132520896)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_18_mlp_gate_proj_weight = const()[name = string("encoder_layers_18_mlp_gate_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132914176)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_18_mlp_up_proj_weight = const()[name = string("encoder_layers_18_mlp_up_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134683712)))];
-            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_18_mlp_down_proj_weight = const()[name = string("encoder_layers_18_mlp_down_proj_weight"), val = tensor<fp16, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136453248)))];
-            tensor<fp16, [768, 768, 1, 1]> encoder_layers_19_self_attn_q_proj_weight = const()[name = string("encoder_layers_19_self_attn_q_proj_weight"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138222784)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_19_self_attn_k_proj_weight = const()[name = string("encoder_layers_19_self_attn_k_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139402496)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_19_self_attn_v_proj_weight = const()[name = string("encoder_layers_19_self_attn_v_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139795776)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_19_mlp_gate_proj_weight = const()[name = string("encoder_layers_19_mlp_gate_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140189056)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_19_mlp_up_proj_weight = const()[name = string("encoder_layers_19_mlp_up_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141958592)))];
-            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_19_mlp_down_proj_weight = const()[name = string("encoder_layers_19_mlp_down_proj_weight"), val = tensor<fp16, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143728128)))];
-            tensor<fp16, [768, 768, 1, 1]> encoder_layers_20_self_attn_q_proj_weight = const()[name = string("encoder_layers_20_self_attn_q_proj_weight"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145497664)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_20_self_attn_k_proj_weight = const()[name = string("encoder_layers_20_self_attn_k_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146677376)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_20_self_attn_v_proj_weight = const()[name = string("encoder_layers_20_self_attn_v_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147070656)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_20_mlp_gate_proj_weight = const()[name = string("encoder_layers_20_mlp_gate_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147463936)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_20_mlp_up_proj_weight = const()[name = string("encoder_layers_20_mlp_up_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149233472)))];
-            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_20_mlp_down_proj_weight = const()[name = string("encoder_layers_20_mlp_down_proj_weight"), val = tensor<fp16, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151003008)))];
-            tensor<fp16, [768, 768, 1, 1]> encoder_layers_21_self_attn_q_proj_weight = const()[name = string("encoder_layers_21_self_attn_q_proj_weight"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152772544)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_21_self_attn_k_proj_weight = const()[name = string("encoder_layers_21_self_attn_k_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153952256)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_21_self_attn_v_proj_weight = const()[name = string("encoder_layers_21_self_attn_v_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154345536)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_21_mlp_gate_proj_weight = const()[name = string("encoder_layers_21_mlp_gate_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154738816)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_21_mlp_up_proj_weight = const()[name = string("encoder_layers_21_mlp_up_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156508352)))];
-            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_21_mlp_down_proj_weight = const()[name = string("encoder_layers_21_mlp_down_proj_weight"), val = tensor<fp16, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158277888)))];
-            tensor<fp16, [768, 768, 1, 1]> encoder_layers_22_self_attn_q_proj_weight = const()[name = string("encoder_layers_22_self_attn_q_proj_weight"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160047424)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_22_self_attn_k_proj_weight = const()[name = string("encoder_layers_22_self_attn_k_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161227136)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_22_self_attn_v_proj_weight = const()[name = string("encoder_layers_22_self_attn_v_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161620416)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_22_mlp_gate_proj_weight = const()[name = string("encoder_layers_22_mlp_gate_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162013696)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_22_mlp_up_proj_weight = const()[name = string("encoder_layers_22_mlp_up_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163783232)))];
-            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_22_mlp_down_proj_weight = const()[name = string("encoder_layers_22_mlp_down_proj_weight"), val = tensor<fp16, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165552768)))];
-            tensor<fp16, [768, 768, 1, 1]> encoder_layers_23_self_attn_q_proj_weight = const()[name = string("encoder_layers_23_self_attn_q_proj_weight"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167322304)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_23_self_attn_k_proj_weight = const()[name = string("encoder_layers_23_self_attn_k_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168502016)))];
-            tensor<fp16, [256, 768, 1, 1]> encoder_layers_23_self_attn_v_proj_weight = const()[name = string("encoder_layers_23_self_attn_v_proj_weight"), val = tensor<fp16, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168895296)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_23_mlp_gate_proj_weight = const()[name = string("encoder_layers_23_mlp_gate_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169288576)))];
-            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_23_mlp_up_proj_weight = const()[name = string("encoder_layers_23_mlp_up_proj_weight"), val = tensor<fp16, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171058112)))];
-            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_23_mlp_down_proj_weight = const()[name = string("encoder_layers_23_mlp_down_proj_weight"), val = tensor<fp16, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172827648)))];
-            tensor<fp16, [3072]> dense1_bias = const()[name = string("dense1_bias"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174597184)))];
-            tensor<fp16, [3072, 768, 1, 1]> dense1_weight = const()[name = string("dense1_weight"), val = tensor<fp16, [3072, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174603392)))];
-            tensor<fp16, [768]> dense2_bias = const()[name = string("dense2_bias"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179322048)))];
-            tensor<fp16, [768, 3072, 1, 1]> dense2_weight = const()[name = string("dense2_weight"), val = tensor<fp16, [768, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179323648)))];
+            tensor<fp16, [768, 768, 1, 1]> encoder_layers_0_self_attn_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(589952))))[name = string("encoder_layers_0_self_attn_q_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_0_self_attn_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(591552))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(788224))))[name = string("encoder_layers_0_self_attn_k_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_0_self_attn_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(788800))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(985472))))[name = string("encoder_layers_0_self_attn_v_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_0_mlp_gate_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(986048))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1870848))))[name = string("encoder_layers_0_mlp_gate_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_0_mlp_up_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1873216))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2758016))))[name = string("encoder_layers_0_mlp_up_proj_weight_quantized")];
+            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_0_mlp_down_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2760384))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3645184))))[name = string("encoder_layers_0_mlp_down_proj_weight_quantized")];
+            tensor<fp16, [768, 768, 1, 1]> encoder_layers_1_self_attn_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3646784))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4236672))))[name = string("encoder_layers_1_self_attn_q_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_1_self_attn_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4238272))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4434944))))[name = string("encoder_layers_1_self_attn_k_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_1_self_attn_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4435520))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4632192))))[name = string("encoder_layers_1_self_attn_v_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_1_mlp_gate_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4632768))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5517568))))[name = string("encoder_layers_1_mlp_gate_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_1_mlp_up_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5519936))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6404736))))[name = string("encoder_layers_1_mlp_up_proj_weight_quantized")];
+            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_1_mlp_down_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6407104))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7291904))))[name = string("encoder_layers_1_mlp_down_proj_weight_quantized")];
+            tensor<fp16, [768, 768, 1, 1]> encoder_layers_2_self_attn_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7293504))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7883392))))[name = string("encoder_layers_2_self_attn_q_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_2_self_attn_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7884992))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8081664))))[name = string("encoder_layers_2_self_attn_k_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_2_self_attn_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8082240))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8278912))))[name = string("encoder_layers_2_self_attn_v_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_2_mlp_gate_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8279488))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9164288))))[name = string("encoder_layers_2_mlp_gate_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_2_mlp_up_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9166656))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10051456))))[name = string("encoder_layers_2_mlp_up_proj_weight_quantized")];
+            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_2_mlp_down_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10053824))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10938624))))[name = string("encoder_layers_2_mlp_down_proj_weight_quantized")];
+            tensor<fp16, [768, 768, 1, 1]> encoder_layers_3_self_attn_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10940224))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11530112))))[name = string("encoder_layers_3_self_attn_q_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_3_self_attn_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11531712))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11728384))))[name = string("encoder_layers_3_self_attn_k_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_3_self_attn_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11728960))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11925632))))[name = string("encoder_layers_3_self_attn_v_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_3_mlp_gate_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11926208))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12811008))))[name = string("encoder_layers_3_mlp_gate_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_3_mlp_up_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12813376))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13698176))))[name = string("encoder_layers_3_mlp_up_proj_weight_quantized")];
+            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_3_mlp_down_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13700544))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14585344))))[name = string("encoder_layers_3_mlp_down_proj_weight_quantized")];
+            tensor<fp16, [768, 768, 1, 1]> encoder_layers_4_self_attn_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14586944))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15176832))))[name = string("encoder_layers_4_self_attn_q_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_4_self_attn_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15178432))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15375104))))[name = string("encoder_layers_4_self_attn_k_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_4_self_attn_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15375680))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15572352))))[name = string("encoder_layers_4_self_attn_v_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_4_mlp_gate_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15572928))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16457728))))[name = string("encoder_layers_4_mlp_gate_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_4_mlp_up_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16460096))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17344896))))[name = string("encoder_layers_4_mlp_up_proj_weight_quantized")];
+            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_4_mlp_down_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17347264))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18232064))))[name = string("encoder_layers_4_mlp_down_proj_weight_quantized")];
+            tensor<fp16, [768, 768, 1, 1]> encoder_layers_5_self_attn_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18233664))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18823552))))[name = string("encoder_layers_5_self_attn_q_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_5_self_attn_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18825152))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19021824))))[name = string("encoder_layers_5_self_attn_k_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_5_self_attn_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19022400))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19219072))))[name = string("encoder_layers_5_self_attn_v_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_5_mlp_gate_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19219648))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20104448))))[name = string("encoder_layers_5_mlp_gate_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_5_mlp_up_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20106816))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20991616))))[name = string("encoder_layers_5_mlp_up_proj_weight_quantized")];
+            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_5_mlp_down_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20993984))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21878784))))[name = string("encoder_layers_5_mlp_down_proj_weight_quantized")];
+            tensor<fp16, [768, 768, 1, 1]> encoder_layers_6_self_attn_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21880384))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22470272))))[name = string("encoder_layers_6_self_attn_q_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_6_self_attn_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22471872))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22668544))))[name = string("encoder_layers_6_self_attn_k_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_6_self_attn_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22669120))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22865792))))[name = string("encoder_layers_6_self_attn_v_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_6_mlp_gate_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22866368))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23751168))))[name = string("encoder_layers_6_mlp_gate_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_6_mlp_up_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23753536))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24638336))))[name = string("encoder_layers_6_mlp_up_proj_weight_quantized")];
+            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_6_mlp_down_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24640704))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25525504))))[name = string("encoder_layers_6_mlp_down_proj_weight_quantized")];
+            tensor<fp16, [768, 768, 1, 1]> encoder_layers_7_self_attn_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25527104))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26116992))))[name = string("encoder_layers_7_self_attn_q_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_7_self_attn_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26118592))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26315264))))[name = string("encoder_layers_7_self_attn_k_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_7_self_attn_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26315840))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26512512))))[name = string("encoder_layers_7_self_attn_v_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_7_mlp_gate_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26513088))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27397888))))[name = string("encoder_layers_7_mlp_gate_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_7_mlp_up_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27400256))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28285056))))[name = string("encoder_layers_7_mlp_up_proj_weight_quantized")];
+            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_7_mlp_down_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28287424))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29172224))))[name = string("encoder_layers_7_mlp_down_proj_weight_quantized")];
+            tensor<fp16, [768, 768, 1, 1]> encoder_layers_8_self_attn_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29173824))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29763712))))[name = string("encoder_layers_8_self_attn_q_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_8_self_attn_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29765312))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29961984))))[name = string("encoder_layers_8_self_attn_k_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_8_self_attn_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29962560))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30159232))))[name = string("encoder_layers_8_self_attn_v_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_8_mlp_gate_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30159808))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31044608))))[name = string("encoder_layers_8_mlp_gate_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_8_mlp_up_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31046976))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31931776))))[name = string("encoder_layers_8_mlp_up_proj_weight_quantized")];
+            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_8_mlp_down_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31934144))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32818944))))[name = string("encoder_layers_8_mlp_down_proj_weight_quantized")];
+            tensor<fp16, [768, 768, 1, 1]> encoder_layers_9_self_attn_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32820544))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33410432))))[name = string("encoder_layers_9_self_attn_q_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_9_self_attn_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33412032))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33608704))))[name = string("encoder_layers_9_self_attn_k_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_9_self_attn_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33609280))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33805952))))[name = string("encoder_layers_9_self_attn_v_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_9_mlp_gate_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33806528))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34691328))))[name = string("encoder_layers_9_mlp_gate_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_9_mlp_up_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34693696))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35578496))))[name = string("encoder_layers_9_mlp_up_proj_weight_quantized")];
+            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_9_mlp_down_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35580864))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36465664))))[name = string("encoder_layers_9_mlp_down_proj_weight_quantized")];
+            tensor<fp16, [768, 768, 1, 1]> encoder_layers_10_self_attn_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36467264))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37057152))))[name = string("encoder_layers_10_self_attn_q_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_10_self_attn_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37058752))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37255424))))[name = string("encoder_layers_10_self_attn_k_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_10_self_attn_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37256000))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37452672))))[name = string("encoder_layers_10_self_attn_v_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_10_mlp_gate_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37453248))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38338048))))[name = string("encoder_layers_10_mlp_gate_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_10_mlp_up_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38340416))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39225216))))[name = string("encoder_layers_10_mlp_up_proj_weight_quantized")];
+            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_10_mlp_down_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39227584))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40112384))))[name = string("encoder_layers_10_mlp_down_proj_weight_quantized")];
+            tensor<fp16, [768, 768, 1, 1]> encoder_layers_11_self_attn_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40113984))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40703872))))[name = string("encoder_layers_11_self_attn_q_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_11_self_attn_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40705472))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40902144))))[name = string("encoder_layers_11_self_attn_k_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_11_self_attn_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40902720))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41099392))))[name = string("encoder_layers_11_self_attn_v_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_11_mlp_gate_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41099968))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41984768))))[name = string("encoder_layers_11_mlp_gate_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_11_mlp_up_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41987136))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42871936))))[name = string("encoder_layers_11_mlp_up_proj_weight_quantized")];
+            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_11_mlp_down_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42874304))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43759104))))[name = string("encoder_layers_11_mlp_down_proj_weight_quantized")];
+            tensor<fp16, [768, 768, 1, 1]> encoder_layers_12_self_attn_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43760704))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44350592))))[name = string("encoder_layers_12_self_attn_q_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_12_self_attn_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44352192))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44548864))))[name = string("encoder_layers_12_self_attn_k_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_12_self_attn_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44549440))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44746112))))[name = string("encoder_layers_12_self_attn_v_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_12_mlp_gate_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44746688))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45631488))))[name = string("encoder_layers_12_mlp_gate_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_12_mlp_up_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45633856))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46518656))))[name = string("encoder_layers_12_mlp_up_proj_weight_quantized")];
+            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_12_mlp_down_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46521024))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47405824))))[name = string("encoder_layers_12_mlp_down_proj_weight_quantized")];
+            tensor<fp16, [768, 768, 1, 1]> encoder_layers_13_self_attn_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47407424))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47997312))))[name = string("encoder_layers_13_self_attn_q_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_13_self_attn_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47998912))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48195584))))[name = string("encoder_layers_13_self_attn_k_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_13_self_attn_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48196160))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48392832))))[name = string("encoder_layers_13_self_attn_v_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_13_mlp_gate_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48393408))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49278208))))[name = string("encoder_layers_13_mlp_gate_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_13_mlp_up_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49280576))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50165376))))[name = string("encoder_layers_13_mlp_up_proj_weight_quantized")];
+            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_13_mlp_down_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50167744))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51052544))))[name = string("encoder_layers_13_mlp_down_proj_weight_quantized")];
+            tensor<fp16, [768, 768, 1, 1]> encoder_layers_14_self_attn_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51054144))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51644032))))[name = string("encoder_layers_14_self_attn_q_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_14_self_attn_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51645632))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51842304))))[name = string("encoder_layers_14_self_attn_k_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_14_self_attn_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51842880))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52039552))))[name = string("encoder_layers_14_self_attn_v_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_14_mlp_gate_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52040128))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52924928))))[name = string("encoder_layers_14_mlp_gate_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_14_mlp_up_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52927296))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53812096))))[name = string("encoder_layers_14_mlp_up_proj_weight_quantized")];
+            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_14_mlp_down_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53814464))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54699264))))[name = string("encoder_layers_14_mlp_down_proj_weight_quantized")];
+            tensor<fp16, [768, 768, 1, 1]> encoder_layers_15_self_attn_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54700864))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55290752))))[name = string("encoder_layers_15_self_attn_q_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_15_self_attn_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55292352))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55489024))))[name = string("encoder_layers_15_self_attn_k_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_15_self_attn_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55489600))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55686272))))[name = string("encoder_layers_15_self_attn_v_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_15_mlp_gate_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55686848))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56571648))))[name = string("encoder_layers_15_mlp_gate_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_15_mlp_up_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56574016))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57458816))))[name = string("encoder_layers_15_mlp_up_proj_weight_quantized")];
+            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_15_mlp_down_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57461184))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58345984))))[name = string("encoder_layers_15_mlp_down_proj_weight_quantized")];
+            tensor<fp16, [768, 768, 1, 1]> encoder_layers_16_self_attn_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58347584))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58937472))))[name = string("encoder_layers_16_self_attn_q_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_16_self_attn_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58939072))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59135744))))[name = string("encoder_layers_16_self_attn_k_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_16_self_attn_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59136320))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59332992))))[name = string("encoder_layers_16_self_attn_v_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_16_mlp_gate_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59333568))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60218368))))[name = string("encoder_layers_16_mlp_gate_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_16_mlp_up_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60220736))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61105536))))[name = string("encoder_layers_16_mlp_up_proj_weight_quantized")];
+            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_16_mlp_down_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61107904))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61992704))))[name = string("encoder_layers_16_mlp_down_proj_weight_quantized")];
+            tensor<fp16, [768, 768, 1, 1]> encoder_layers_17_self_attn_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61994304))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62584192))))[name = string("encoder_layers_17_self_attn_q_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_17_self_attn_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62585792))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62782464))))[name = string("encoder_layers_17_self_attn_k_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_17_self_attn_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62783040))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62979712))))[name = string("encoder_layers_17_self_attn_v_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_17_mlp_gate_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62980288))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63865088))))[name = string("encoder_layers_17_mlp_gate_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_17_mlp_up_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63867456))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64752256))))[name = string("encoder_layers_17_mlp_up_proj_weight_quantized")];
+            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_17_mlp_down_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64754624))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65639424))))[name = string("encoder_layers_17_mlp_down_proj_weight_quantized")];
+            tensor<fp16, [768, 768, 1, 1]> encoder_layers_18_self_attn_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65641024))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66230912))))[name = string("encoder_layers_18_self_attn_q_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_18_self_attn_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66232512))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66429184))))[name = string("encoder_layers_18_self_attn_k_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_18_self_attn_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66429760))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66626432))))[name = string("encoder_layers_18_self_attn_v_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_18_mlp_gate_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66627008))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67511808))))[name = string("encoder_layers_18_mlp_gate_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_18_mlp_up_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67514176))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68398976))))[name = string("encoder_layers_18_mlp_up_proj_weight_quantized")];
+            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_18_mlp_down_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68401344))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69286144))))[name = string("encoder_layers_18_mlp_down_proj_weight_quantized")];
+            tensor<fp16, [768, 768, 1, 1]> encoder_layers_19_self_attn_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69287744))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69877632))))[name = string("encoder_layers_19_self_attn_q_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_19_self_attn_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69879232))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70075904))))[name = string("encoder_layers_19_self_attn_k_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_19_self_attn_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70076480))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70273152))))[name = string("encoder_layers_19_self_attn_v_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_19_mlp_gate_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70273728))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71158528))))[name = string("encoder_layers_19_mlp_gate_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_19_mlp_up_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71160896))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72045696))))[name = string("encoder_layers_19_mlp_up_proj_weight_quantized")];
+            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_19_mlp_down_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72048064))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72932864))))[name = string("encoder_layers_19_mlp_down_proj_weight_quantized")];
+            tensor<fp16, [768, 768, 1, 1]> encoder_layers_20_self_attn_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72934464))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73524352))))[name = string("encoder_layers_20_self_attn_q_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_20_self_attn_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73525952))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73722624))))[name = string("encoder_layers_20_self_attn_k_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_20_self_attn_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73723200))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73919872))))[name = string("encoder_layers_20_self_attn_v_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_20_mlp_gate_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73920448))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74805248))))[name = string("encoder_layers_20_mlp_gate_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_20_mlp_up_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74807616))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75692416))))[name = string("encoder_layers_20_mlp_up_proj_weight_quantized")];
+            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_20_mlp_down_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75694784))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76579584))))[name = string("encoder_layers_20_mlp_down_proj_weight_quantized")];
+            tensor<fp16, [768, 768, 1, 1]> encoder_layers_21_self_attn_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76581184))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77171072))))[name = string("encoder_layers_21_self_attn_q_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_21_self_attn_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77172672))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77369344))))[name = string("encoder_layers_21_self_attn_k_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_21_self_attn_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77369920))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77566592))))[name = string("encoder_layers_21_self_attn_v_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_21_mlp_gate_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77567168))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78451968))))[name = string("encoder_layers_21_mlp_gate_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_21_mlp_up_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78454336))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79339136))))[name = string("encoder_layers_21_mlp_up_proj_weight_quantized")];
+            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_21_mlp_down_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79341504))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80226304))))[name = string("encoder_layers_21_mlp_down_proj_weight_quantized")];
+            tensor<fp16, [768, 768, 1, 1]> encoder_layers_22_self_attn_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80227904))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80817792))))[name = string("encoder_layers_22_self_attn_q_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_22_self_attn_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80819392))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81016064))))[name = string("encoder_layers_22_self_attn_k_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_22_self_attn_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81016640))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81213312))))[name = string("encoder_layers_22_self_attn_v_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_22_mlp_gate_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81213888))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82098688))))[name = string("encoder_layers_22_mlp_gate_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_22_mlp_up_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82101056))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82985856))))[name = string("encoder_layers_22_mlp_up_proj_weight_quantized")];
+            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_22_mlp_down_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82988224))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83873024))))[name = string("encoder_layers_22_mlp_down_proj_weight_quantized")];
+            tensor<fp16, [768, 768, 1, 1]> encoder_layers_23_self_attn_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83874624))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84464512))))[name = string("encoder_layers_23_self_attn_q_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_23_self_attn_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84466112))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84662784))))[name = string("encoder_layers_23_self_attn_k_proj_weight_quantized")];
+            tensor<fp16, [256, 768, 1, 1]> encoder_layers_23_self_attn_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84663360))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84860032))))[name = string("encoder_layers_23_self_attn_v_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_23_mlp_gate_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84860608))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85745408))))[name = string("encoder_layers_23_mlp_gate_proj_weight_quantized")];
+            tensor<fp16, [1152, 768, 1, 1]> encoder_layers_23_mlp_up_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85747776))), scale = tensor<fp16, [1152, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86632576))))[name = string("encoder_layers_23_mlp_up_proj_weight_quantized")];
+            tensor<fp16, [768, 1152, 1, 1]> encoder_layers_23_mlp_down_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86634944))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87519744))))[name = string("encoder_layers_23_mlp_down_proj_weight_quantized")];
+            tensor<fp16, [3072]> dense1_bias = const()[name = string("dense1_bias"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87521344)))];
+            tensor<fp16, [3072, 768, 1, 1]> dense1_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87527552))), scale = tensor<fp16, [3072, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89886912))))[name = string("dense1_weight_quantized")];
+            tensor<fp16, [768]> dense2_bias = const()[name = string("dense2_bias"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89893120)))];
+            tensor<fp16, [768, 3072, 1, 1]> dense2_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89894720))), scale = tensor<fp16, [768, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92254080))))[name = string("dense2_weight_quantized")];
             int32 var_23 = const()[name = string("op_23"), val = int32(-1)];
             int32 var_81_batch_dims_0 = const()[name = string("op_81_batch_dims_0"), val = int32(0)];
             bool var_81_validate_indices_0 = const()[name = string("op_81_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [262144, 768]> encoder_embed_tokens_weight_to_fp16 = const()[name = string("encoder_embed_tokens_weight_to_fp16"), val = tensor<fp16, [262144, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184042304)))];
+            tensor<fp16, [262144, 768]> encoder_embed_tokens_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [262144, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92255680))), scale = tensor<fp16, [262144, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293582336))))[name = string("encoder_embed_tokens_weight_to_fp16_quantized")];
             int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
             tensor<bool, [1, 128]> greater_equal_0 = greater_equal(x = input_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(262144)];
             tensor<int32, [1, 128]> add_0 = add(x = input_ids, y = slice_by_index_0)[name = string("add_0")];
             tensor<int32, [1, 128]> select_0 = select(a = input_ids, b = add_0, cond = greater_equal_0)[name = string("select_0")];
-            int32 var_81_cast_fp16_axis_0 = const()[name = string("op_81_cast_fp16_axis_0"), val = int32(0)];
-            tensor<fp16, [1, 128, 768]> var_81_cast_fp16 = gather(axis = var_81_cast_fp16_axis_0, batch_dims = var_81_batch_dims_0, indices = select_0, validate_indices = var_81_validate_indices_0, x = encoder_embed_tokens_weight_to_fp16)[name = string("op_81_cast_fp16")];
+            int32 greater_equal_0_y_0_1 = const()[name = string("greater_equal_0_y_0_1"), val = int32(0)];
+            tensor<bool, [1, 128]> greater_equal_0_1 = greater_equal(x = select_0, y = greater_equal_0_y_0_1)[name = string("greater_equal_0_1")];
+            int32 slice_by_index_0_1 = const()[name = string("slice_by_index_0_1"), val = int32(262144)];
+            tensor<int32, [1, 128]> add_0_1 = add(x = select_0, y = slice_by_index_0_1)[name = string("add_0_1")];
+            tensor<int32, [1, 128]> select_0_1 = select(a = select_0, b = add_0_1, cond = greater_equal_0_1)[name = string("select_0_1")];
+            int32 op_81_cast_fp16_axis_0 = const()[name = string("op_81_cast_fp16_axis_0"), val = int32(0)];
+            tensor<fp16, [1, 128, 768]> op_81_cast_fp16 = gather(axis = op_81_cast_fp16_axis_0, batch_dims = var_81_batch_dims_0, indices = select_0_1, validate_indices = var_81_validate_indices_0, x = encoder_embed_tokens_weight_to_fp16_quantized)[name = string("op_81_cast_fp16")];
             fp16 var_83_to_fp16 = const()[name = string("op_83_to_fp16"), val = fp16(0x1.bb8p+4)];
-            tensor<fp16, [1, 128, 768]> x_1_cast_fp16 = mul(x = var_81_cast_fp16, y = var_83_to_fp16)[name = string("x_1_cast_fp16")];
-            tensor<fp16, [1, 1, 128, 256]> cos_1 = const()[name = string("cos_1"), val = tensor<fp16, [1, 1, 128, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(586695552)))];
-            tensor<fp16, [1, 1, 128, 256]> sin_1 = const()[name = string("sin_1"), val = tensor<fp16, [1, 1, 128, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(586761152)))];
-            tensor<fp16, [1, 1, 128, 256]> cos = const()[name = string("cos"), val = tensor<fp16, [1, 1, 128, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(586826752)))];
-            tensor<fp16, [1, 1, 128, 256]> sin = const()[name = string("sin"), val = tensor<fp16, [1, 1, 128, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(586892352)))];
+            tensor<fp16, [1, 128, 768]> x_1_cast_fp16 = mul(x = op_81_cast_fp16, y = var_83_to_fp16)[name = string("x_1_cast_fp16")];
+            tensor<fp16, [1, 1, 128, 256]> cos_1_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 1, 128, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294106688))), scale = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.02p-7]]]]))[name = string("cos_1_quantized")];
+            tensor<fp16, [1, 1, 128, 256]> sin_1_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 1, 128, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294139520))), scale = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.02p-7]]]]))[name = string("sin_1_quantized")];
+            tensor<fp16, [1, 1, 128, 256]> cos_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 1, 128, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294172352))), scale = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.02p-7]]]]))[name = string("cos_quantized")];
+            tensor<fp16, [1, 1, 128, 256]> sin_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 1, 128, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294205184))), scale = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.02p-7]]]]))[name = string("sin_quantized")];
             fp16 var_17_to_fp16 = const()[name = string("op_17_to_fp16"), val = fp16(0x1p+0)];
             tensor<fp16, [1, 128]> var_93_cast_fp16 = sub(x = var_17_to_fp16, y = attention_mask)[name = string("op_93_cast_fp16")];
             fp16 var_95_to_fp16 = const()[name = string("op_95_to_fp16"), val = fp16(-0x1.388p+13)];
@@ -185,7 +190,7 @@ program(1.3)
             tensor<int32, [2]> var_133_split_sizes_0 = const()[name = string("op_133_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_133_axis_0 = const()[name = string("op_133_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_133_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_133_cast_fp16_1 = split(axis = var_133_axis_0, split_sizes = var_133_split_sizes_0, x = normed_1_cast_fp16)[name = string("op_133_cast_fp16")];
-            tensor<fp16, [768]> var_137_to_fp16 = const()[name = string("op_137_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(586957952)))];
+            tensor<fp16, [768]> var_137_to_fp16 = const()[name = string("op_137_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294238016)))];
             tensor<fp16, [1, 128, 768]> out_1_cast_fp16 = mul(x = var_133_cast_fp16_0, y = var_137_to_fp16)[name = string("out_1_cast_fp16")];
             tensor<int32, [3]> var_143 = const()[name = string("op_143"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> var_145_axes_0 = const()[name = string("op_145_axes_0"), val = tensor<int32, [1]>([2])];
@@ -196,7 +201,7 @@ program(1.3)
             tensor<int32, [4]> var_152_pad_0 = const()[name = string("op_152_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_152_dilations_0 = const()[name = string("op_152_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_152_groups_0 = const()[name = string("op_152_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_152 = conv(dilations = var_152_dilations_0, groups = var_152_groups_0, pad = var_152_pad_0, pad_type = var_152_pad_type_0, strides = var_152_strides_0, weight = encoder_layers_0_self_attn_q_proj_weight, x = var_145_cast_fp16)[name = string("op_152")];
+            tensor<fp16, [1, 768, 1, 128]> var_152 = conv(dilations = var_152_dilations_0, groups = var_152_groups_0, pad = var_152_pad_0, pad_type = var_152_pad_type_0, strides = var_152_strides_0, weight = encoder_layers_0_self_attn_q_proj_weight_quantized, x = var_145_cast_fp16)[name = string("op_152")];
             tensor<int32, [4]> var_153 = const()[name = string("op_153"), val = tensor<int32, [4]>([1, 3, 256, 128])];
             tensor<fp16, [1, 3, 256, 128]> var_154 = reshape(shape = var_153, x = var_152)[name = string("op_154")];
             tensor<int32, [4]> var_155 = const()[name = string("op_155"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -205,7 +210,7 @@ program(1.3)
             tensor<int32, [4]> var_162_pad_0 = const()[name = string("op_162_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_162_dilations_0 = const()[name = string("op_162_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_162_groups_0 = const()[name = string("op_162_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_162 = conv(dilations = var_162_dilations_0, groups = var_162_groups_0, pad = var_162_pad_0, pad_type = var_162_pad_type_0, strides = var_162_strides_0, weight = encoder_layers_0_self_attn_k_proj_weight, x = var_145_cast_fp16)[name = string("op_162")];
+            tensor<fp16, [1, 256, 1, 128]> var_162 = conv(dilations = var_162_dilations_0, groups = var_162_groups_0, pad = var_162_pad_0, pad_type = var_162_pad_type_0, strides = var_162_strides_0, weight = encoder_layers_0_self_attn_k_proj_weight_quantized, x = var_145_cast_fp16)[name = string("op_162")];
             tensor<int32, [4]> var_163 = const()[name = string("op_163"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_164 = reshape(shape = var_163, x = var_162)[name = string("op_164")];
             tensor<int32, [4]> var_165 = const()[name = string("op_165"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -214,7 +219,7 @@ program(1.3)
             tensor<int32, [4]> var_172_pad_0 = const()[name = string("op_172_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_172_dilations_0 = const()[name = string("op_172_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_172_groups_0 = const()[name = string("op_172_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_172 = conv(dilations = var_172_dilations_0, groups = var_172_groups_0, pad = var_172_pad_0, pad_type = var_172_pad_type_0, strides = var_172_strides_0, weight = encoder_layers_0_self_attn_v_proj_weight, x = var_145_cast_fp16)[name = string("op_172")];
+            tensor<fp16, [1, 256, 1, 128]> var_172 = conv(dilations = var_172_dilations_0, groups = var_172_groups_0, pad = var_172_pad_0, pad_type = var_172_pad_type_0, strides = var_172_strides_0, weight = encoder_layers_0_self_attn_v_proj_weight_quantized, x = var_145_cast_fp16)[name = string("op_172")];
             tensor<int32, [4]> var_173 = const()[name = string("op_173"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_174 = reshape(shape = var_173, x = var_172)[name = string("op_174")];
             tensor<int32, [4]> var_175 = const()[name = string("op_175"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -228,7 +233,7 @@ program(1.3)
             tensor<int32, [2]> var_186_split_sizes_0 = const()[name = string("op_186_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_186_axis_0 = const()[name = string("op_186_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 256]> var_186_cast_fp16_0, tensor<fp16, [1, 3, 128, 256]> var_186_cast_fp16_1 = split(axis = var_186_axis_0, split_sizes = var_186_split_sizes_0, x = normed_7_cast_fp16)[name = string("op_186_cast_fp16")];
-            tensor<fp16, [256]> var_190_to_fp16 = const()[name = string("op_190_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(586959552)))];
+            tensor<fp16, [256]> var_190_to_fp16 = const()[name = string("op_190_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294239616)))];
             tensor<fp16, [1, 3, 128, 256]> out_3_cast_fp16 = mul(x = var_186_cast_fp16_0, y = var_190_to_fp16)[name = string("out_3_cast_fp16")];
             fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 1, 128, 256]> k_1 = transpose(perm = var_165, x = var_164)[name = string("transpose_213")];
@@ -240,9 +245,9 @@ program(1.3)
             tensor<int32, [2]> var_202_split_sizes_0 = const()[name = string("op_202_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_202_axis_0 = const()[name = string("op_202_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 256]> var_202_cast_fp16_0, tensor<fp16, [1, 1, 128, 256]> var_202_cast_fp16_1 = split(axis = var_202_axis_0, split_sizes = var_202_split_sizes_0, x = normed_11_cast_fp16)[name = string("op_202_cast_fp16")];
-            tensor<fp16, [256]> var_206_to_fp16 = const()[name = string("op_206_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(586960128)))];
+            tensor<fp16, [256]> var_206_to_fp16 = const()[name = string("op_206_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294240192)))];
             tensor<fp16, [1, 1, 128, 256]> out_5_cast_fp16 = mul(x = var_202_cast_fp16_0, y = var_206_to_fp16)[name = string("out_5_cast_fp16")];
-            tensor<fp16, [1, 3, 128, 256]> var_209 = mul(x = out_3_cast_fp16, y = cos_1)[name = string("op_209")];
+            tensor<fp16, [1, 3, 128, 256]> var_209 = mul(x = out_3_cast_fp16, y = cos_1_quantized)[name = string("op_209")];
             tensor<int32, [2]> var_210_split_sizes_0 = const()[name = string("op_210_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_210_axis_0 = const()[name = string("op_210_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 128]> var_210_0, tensor<fp16, [1, 3, 128, 128]> var_210_1 = split(axis = var_210_axis_0, split_sizes = var_210_split_sizes_0, x = out_3_cast_fp16)[name = string("op_210")];
@@ -250,9 +255,9 @@ program(1.3)
             tensor<fp16, [1, 3, 128, 128]> var_212 = mul(x = var_210_1, y = const_6_promoted)[name = string("op_212")];
             bool var_214_interleave_0 = const()[name = string("op_214_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 3, 128, 256]> var_214 = concat(axis = var_23, interleave = var_214_interleave_0, values = (var_212, var_210_0))[name = string("op_214")];
-            tensor<fp16, [1, 3, 128, 256]> var_215 = mul(x = var_214, y = sin_1)[name = string("op_215")];
+            tensor<fp16, [1, 3, 128, 256]> var_215 = mul(x = var_214, y = sin_1_quantized)[name = string("op_215")];
             tensor<fp16, [1, 3, 128, 256]> q_5 = add(x = var_209, y = var_215)[name = string("q_5")];
-            tensor<fp16, [1, 1, 128, 256]> var_217 = mul(x = out_5_cast_fp16, y = cos_1)[name = string("op_217")];
+            tensor<fp16, [1, 1, 128, 256]> var_217 = mul(x = out_5_cast_fp16, y = cos_1_quantized)[name = string("op_217")];
             tensor<int32, [2]> var_218_split_sizes_0 = const()[name = string("op_218_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_218_axis_0 = const()[name = string("op_218_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 128]> var_218_0, tensor<fp16, [1, 1, 128, 128]> var_218_1 = split(axis = var_218_axis_0, split_sizes = var_218_split_sizes_0, x = out_5_cast_fp16)[name = string("op_218")];
@@ -260,7 +265,7 @@ program(1.3)
             tensor<fp16, [1, 1, 128, 128]> var_220 = mul(x = var_218_1, y = const_7_promoted)[name = string("op_220")];
             bool var_222_interleave_0 = const()[name = string("op_222_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 1, 128, 256]> var_222 = concat(axis = var_23, interleave = var_222_interleave_0, values = (var_220, var_218_0))[name = string("op_222")];
-            tensor<fp16, [1, 1, 128, 256]> var_223 = mul(x = var_222, y = sin_1)[name = string("op_223")];
+            tensor<fp16, [1, 1, 128, 256]> var_223 = mul(x = var_222, y = sin_1_quantized)[name = string("op_223")];
             tensor<fp16, [1, 1, 128, 256]> hidden_states_1 = add(x = var_217, y = var_223)[name = string("hidden_states_1")];
             tensor<int32, [1]> hidden_states_3_axes_0 = const()[name = string("hidden_states_3_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 1, 1, 128, 256]> hidden_states_3 = expand_dims(axes = hidden_states_3_axes_0, x = hidden_states_1)[name = string("hidden_states_3")];
@@ -290,14 +295,14 @@ program(1.3)
             tensor<fp16, [1, 128, 3, 256]> var_250 = transpose(perm = var_249, x = var_247_cast_fp16)[name = string("transpose_211")];
             tensor<fp16, [1, 128, 768]> attn_out_3 = reshape(shape = var_252, x = var_250)[name = string("attn_out_3")];
             tensor<int32, [3]> var_254 = const()[name = string("op_254"), val = tensor<int32, [3]>([0, 2, 1])];
-            tensor<fp16, [768, 768, 1]> squeeze_0 = const()[name = string("squeeze_0"), val = tensor<fp16, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(586960704)))];
+            tensor<fp16, [768, 768, 1]> squeeze_0_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294240768))), scale = tensor<fp16, [768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294830656))))[name = string("squeeze_0_quantized")];
             string var_263_pad_type_0 = const()[name = string("op_263_pad_type_0"), val = string("valid")];
             int32 var_263_groups_0 = const()[name = string("op_263_groups_0"), val = int32(1)];
             tensor<int32, [1]> var_263_strides_0 = const()[name = string("op_263_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> var_263_pad_0 = const()[name = string("op_263_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> var_263_dilations_0 = const()[name = string("op_263_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, [1, 768, 128]> var_255 = transpose(perm = var_254, x = attn_out_3)[name = string("transpose_210")];
-            tensor<fp16, [1, 768, 128]> var_263 = conv(dilations = var_263_dilations_0, groups = var_263_groups_0, pad = var_263_pad_0, pad_type = var_263_pad_type_0, strides = var_263_strides_0, weight = squeeze_0, x = var_255)[name = string("op_263")];
+            tensor<fp16, [1, 768, 128]> var_263 = conv(dilations = var_263_dilations_0, groups = var_263_groups_0, pad = var_263_pad_0, pad_type = var_263_pad_type_0, strides = var_263_strides_0, weight = squeeze_0_quantized, x = var_255)[name = string("op_263")];
             tensor<int32, [3]> var_264 = const()[name = string("op_264"), val = tensor<int32, [3]>([0, 2, 1])];
             fp16 const_8_promoted_to_fp16 = const()[name = string("const_8_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 128, 768]> x_9 = transpose(perm = var_264, x = var_263)[name = string("transpose_209")];
@@ -309,7 +314,7 @@ program(1.3)
             tensor<int32, [2]> var_273_split_sizes_0 = const()[name = string("op_273_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_273_axis_0 = const()[name = string("op_273_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_273_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_273_cast_fp16_1 = split(axis = var_273_axis_0, split_sizes = var_273_split_sizes_0, x = normed_15_cast_fp16)[name = string("op_273_cast_fp16")];
-            tensor<fp16, [768]> var_277_to_fp16 = const()[name = string("op_277_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(588140416)))];
+            tensor<fp16, [768]> var_277_to_fp16 = const()[name = string("op_277_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294832256)))];
             tensor<fp16, [1, 128, 768]> out_7_cast_fp16 = mul(x = var_273_cast_fp16_0, y = var_277_to_fp16)[name = string("out_7_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_11_cast_fp16 = add(x = x_1_cast_fp16, y = out_7_cast_fp16)[name = string("x_11_cast_fp16")];
             fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -321,7 +326,7 @@ program(1.3)
             tensor<int32, [2]> var_289_split_sizes_0 = const()[name = string("op_289_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_289_axis_0 = const()[name = string("op_289_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_289_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_289_cast_fp16_1 = split(axis = var_289_axis_0, split_sizes = var_289_split_sizes_0, x = normed_19_cast_fp16)[name = string("op_289_cast_fp16")];
-            tensor<fp16, [768]> var_293_to_fp16 = const()[name = string("op_293_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(588142016)))];
+            tensor<fp16, [768]> var_293_to_fp16 = const()[name = string("op_293_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294833856)))];
             tensor<fp16, [1, 128, 768]> out_9_cast_fp16 = mul(x = var_289_cast_fp16_0, y = var_293_to_fp16)[name = string("out_9_cast_fp16")];
             tensor<int32, [3]> var_300 = const()[name = string("op_300"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> input_15_axes_0 = const()[name = string("input_15_axes_0"), val = tensor<int32, [1]>([2])];
@@ -332,13 +337,13 @@ program(1.3)
             tensor<int32, [4]> gate_1_pad_0 = const()[name = string("gate_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> gate_1_dilations_0 = const()[name = string("gate_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 gate_1_groups_0 = const()[name = string("gate_1_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> gate_1 = conv(dilations = gate_1_dilations_0, groups = gate_1_groups_0, pad = gate_1_pad_0, pad_type = gate_1_pad_type_0, strides = gate_1_strides_0, weight = encoder_layers_0_mlp_gate_proj_weight, x = input_15)[name = string("gate_1")];
+            tensor<fp16, [1, 1152, 1, 128]> gate_1 = conv(dilations = gate_1_dilations_0, groups = gate_1_groups_0, pad = gate_1_pad_0, pad_type = gate_1_pad_type_0, strides = gate_1_strides_0, weight = encoder_layers_0_mlp_gate_proj_weight_quantized, x = input_15)[name = string("gate_1")];
             string up_1_pad_type_0 = const()[name = string("up_1_pad_type_0"), val = string("valid")];
             tensor<int32, [2]> up_1_strides_0 = const()[name = string("up_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> up_1_pad_0 = const()[name = string("up_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> up_1_dilations_0 = const()[name = string("up_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 up_1_groups_0 = const()[name = string("up_1_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> up_1 = conv(dilations = up_1_dilations_0, groups = up_1_groups_0, pad = up_1_pad_0, pad_type = up_1_pad_type_0, strides = up_1_strides_0, weight = encoder_layers_0_mlp_up_proj_weight, x = input_15)[name = string("up_1")];
+            tensor<fp16, [1, 1152, 1, 128]> up_1 = conv(dilations = up_1_dilations_0, groups = up_1_groups_0, pad = up_1_pad_0, pad_type = up_1_pad_type_0, strides = up_1_strides_0, weight = encoder_layers_0_mlp_up_proj_weight_quantized, x = input_15)[name = string("up_1")];
             string gate_3_mode_0 = const()[name = string("gate_3_mode_0"), val = string("TANH_APPROXIMATION")];
             tensor<fp16, [1, 1152, 1, 128]> gate_3 = gelu(mode = gate_3_mode_0, x = gate_1)[name = string("gate_3")];
             tensor<fp16, [1, 1152, 1, 128]> input_17 = mul(x = gate_3, y = up_1)[name = string("input_17")];
@@ -347,7 +352,7 @@ program(1.3)
             tensor<int32, [4]> var_322_pad_0 = const()[name = string("op_322_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_322_dilations_0 = const()[name = string("op_322_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_322_groups_0 = const()[name = string("op_322_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_322 = conv(dilations = var_322_dilations_0, groups = var_322_groups_0, pad = var_322_pad_0, pad_type = var_322_pad_type_0, strides = var_322_strides_0, weight = encoder_layers_0_mlp_down_proj_weight, x = input_17)[name = string("op_322")];
+            tensor<fp16, [1, 768, 1, 128]> var_322 = conv(dilations = var_322_dilations_0, groups = var_322_groups_0, pad = var_322_pad_0, pad_type = var_322_pad_type_0, strides = var_322_strides_0, weight = encoder_layers_0_mlp_down_proj_weight_quantized, x = input_17)[name = string("op_322")];
             tensor<int32, [1]> var_323_axes_0 = const()[name = string("op_323_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 768, 128]> var_323 = squeeze(axes = var_323_axes_0, x = var_322)[name = string("op_323")];
             tensor<int32, [3]> var_324 = const()[name = string("op_324"), val = tensor<int32, [3]>([0, 2, 1])];
@@ -361,7 +366,7 @@ program(1.3)
             tensor<int32, [2]> var_333_split_sizes_0 = const()[name = string("op_333_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_333_axis_0 = const()[name = string("op_333_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_333_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_333_cast_fp16_1 = split(axis = var_333_axis_0, split_sizes = var_333_split_sizes_0, x = normed_25_cast_fp16)[name = string("op_333_cast_fp16")];
-            tensor<fp16, [768]> var_337_to_fp16 = const()[name = string("op_337_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(588143616)))];
+            tensor<fp16, [768]> var_337_to_fp16 = const()[name = string("op_337_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294835456)))];
             tensor<fp16, [1, 128, 768]> out_11_cast_fp16 = mul(x = var_333_cast_fp16_0, y = var_337_to_fp16)[name = string("out_11_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_17_cast_fp16 = add(x = x_11_cast_fp16, y = out_11_cast_fp16)[name = string("x_17_cast_fp16")];
             fp16 const_14_promoted_to_fp16 = const()[name = string("const_14_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -373,7 +378,7 @@ program(1.3)
             tensor<int32, [2]> var_371_split_sizes_0 = const()[name = string("op_371_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_371_axis_0 = const()[name = string("op_371_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_371_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_371_cast_fp16_1 = split(axis = var_371_axis_0, split_sizes = var_371_split_sizes_0, x = normed_29_cast_fp16)[name = string("op_371_cast_fp16")];
-            tensor<fp16, [768]> var_375_to_fp16 = const()[name = string("op_375_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(588145216)))];
+            tensor<fp16, [768]> var_375_to_fp16 = const()[name = string("op_375_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294837056)))];
             tensor<fp16, [1, 128, 768]> out_13_cast_fp16 = mul(x = var_371_cast_fp16_0, y = var_375_to_fp16)[name = string("out_13_cast_fp16")];
             tensor<int32, [3]> var_381 = const()[name = string("op_381"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> var_383_axes_0 = const()[name = string("op_383_axes_0"), val = tensor<int32, [1]>([2])];
@@ -384,7 +389,7 @@ program(1.3)
             tensor<int32, [4]> var_390_pad_0 = const()[name = string("op_390_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_390_dilations_0 = const()[name = string("op_390_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_390_groups_0 = const()[name = string("op_390_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_390 = conv(dilations = var_390_dilations_0, groups = var_390_groups_0, pad = var_390_pad_0, pad_type = var_390_pad_type_0, strides = var_390_strides_0, weight = encoder_layers_1_self_attn_q_proj_weight, x = var_383_cast_fp16)[name = string("op_390")];
+            tensor<fp16, [1, 768, 1, 128]> var_390 = conv(dilations = var_390_dilations_0, groups = var_390_groups_0, pad = var_390_pad_0, pad_type = var_390_pad_type_0, strides = var_390_strides_0, weight = encoder_layers_1_self_attn_q_proj_weight_quantized, x = var_383_cast_fp16)[name = string("op_390")];
             tensor<int32, [4]> var_391 = const()[name = string("op_391"), val = tensor<int32, [4]>([1, 3, 256, 128])];
             tensor<fp16, [1, 3, 256, 128]> var_392 = reshape(shape = var_391, x = var_390)[name = string("op_392")];
             tensor<int32, [4]> var_393 = const()[name = string("op_393"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -393,7 +398,7 @@ program(1.3)
             tensor<int32, [4]> var_400_pad_0 = const()[name = string("op_400_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_400_dilations_0 = const()[name = string("op_400_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_400_groups_0 = const()[name = string("op_400_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_400 = conv(dilations = var_400_dilations_0, groups = var_400_groups_0, pad = var_400_pad_0, pad_type = var_400_pad_type_0, strides = var_400_strides_0, weight = encoder_layers_1_self_attn_k_proj_weight, x = var_383_cast_fp16)[name = string("op_400")];
+            tensor<fp16, [1, 256, 1, 128]> var_400 = conv(dilations = var_400_dilations_0, groups = var_400_groups_0, pad = var_400_pad_0, pad_type = var_400_pad_type_0, strides = var_400_strides_0, weight = encoder_layers_1_self_attn_k_proj_weight_quantized, x = var_383_cast_fp16)[name = string("op_400")];
             tensor<int32, [4]> var_401 = const()[name = string("op_401"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_402 = reshape(shape = var_401, x = var_400)[name = string("op_402")];
             tensor<int32, [4]> var_403 = const()[name = string("op_403"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -402,7 +407,7 @@ program(1.3)
             tensor<int32, [4]> var_410_pad_0 = const()[name = string("op_410_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_410_dilations_0 = const()[name = string("op_410_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_410_groups_0 = const()[name = string("op_410_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_410 = conv(dilations = var_410_dilations_0, groups = var_410_groups_0, pad = var_410_pad_0, pad_type = var_410_pad_type_0, strides = var_410_strides_0, weight = encoder_layers_1_self_attn_v_proj_weight, x = var_383_cast_fp16)[name = string("op_410")];
+            tensor<fp16, [1, 256, 1, 128]> var_410 = conv(dilations = var_410_dilations_0, groups = var_410_groups_0, pad = var_410_pad_0, pad_type = var_410_pad_type_0, strides = var_410_strides_0, weight = encoder_layers_1_self_attn_v_proj_weight_quantized, x = var_383_cast_fp16)[name = string("op_410")];
             tensor<int32, [4]> var_411 = const()[name = string("op_411"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_412 = reshape(shape = var_411, x = var_410)[name = string("op_412")];
             tensor<int32, [4]> var_413 = const()[name = string("op_413"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -416,7 +421,7 @@ program(1.3)
             tensor<int32, [2]> var_424_split_sizes_0 = const()[name = string("op_424_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_424_axis_0 = const()[name = string("op_424_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 256]> var_424_cast_fp16_0, tensor<fp16, [1, 3, 128, 256]> var_424_cast_fp16_1 = split(axis = var_424_axis_0, split_sizes = var_424_split_sizes_0, x = normed_35_cast_fp16)[name = string("op_424_cast_fp16")];
-            tensor<fp16, [256]> var_428_to_fp16 = const()[name = string("op_428_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(588146816)))];
+            tensor<fp16, [256]> var_428_to_fp16 = const()[name = string("op_428_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294838656)))];
             tensor<fp16, [1, 3, 128, 256]> out_15_cast_fp16 = mul(x = var_424_cast_fp16_0, y = var_428_to_fp16)[name = string("out_15_cast_fp16")];
             fp16 const_18_promoted_to_fp16 = const()[name = string("const_18_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 1, 128, 256]> k_7 = transpose(perm = var_403, x = var_402)[name = string("transpose_204")];
@@ -428,9 +433,9 @@ program(1.3)
             tensor<int32, [2]> var_440_split_sizes_0 = const()[name = string("op_440_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_440_axis_0 = const()[name = string("op_440_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 256]> var_440_cast_fp16_0, tensor<fp16, [1, 1, 128, 256]> var_440_cast_fp16_1 = split(axis = var_440_axis_0, split_sizes = var_440_split_sizes_0, x = normed_39_cast_fp16)[name = string("op_440_cast_fp16")];
-            tensor<fp16, [256]> var_444_to_fp16 = const()[name = string("op_444_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(588147392)))];
+            tensor<fp16, [256]> var_444_to_fp16 = const()[name = string("op_444_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294839232)))];
             tensor<fp16, [1, 1, 128, 256]> out_17_cast_fp16 = mul(x = var_440_cast_fp16_0, y = var_444_to_fp16)[name = string("out_17_cast_fp16")];
-            tensor<fp16, [1, 3, 128, 256]> var_447 = mul(x = out_15_cast_fp16, y = cos_1)[name = string("op_447")];
+            tensor<fp16, [1, 3, 128, 256]> var_447 = mul(x = out_15_cast_fp16, y = cos_1_quantized)[name = string("op_447")];
             tensor<int32, [2]> var_448_split_sizes_0 = const()[name = string("op_448_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_448_axis_0 = const()[name = string("op_448_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 128]> var_448_0, tensor<fp16, [1, 3, 128, 128]> var_448_1 = split(axis = var_448_axis_0, split_sizes = var_448_split_sizes_0, x = out_15_cast_fp16)[name = string("op_448")];
@@ -438,9 +443,9 @@ program(1.3)
             tensor<fp16, [1, 3, 128, 128]> var_450 = mul(x = var_448_1, y = const_20_promoted)[name = string("op_450")];
             bool var_452_interleave_0 = const()[name = string("op_452_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 3, 128, 256]> var_452 = concat(axis = var_23, interleave = var_452_interleave_0, values = (var_450, var_448_0))[name = string("op_452")];
-            tensor<fp16, [1, 3, 128, 256]> var_453 = mul(x = var_452, y = sin_1)[name = string("op_453")];
+            tensor<fp16, [1, 3, 128, 256]> var_453 = mul(x = var_452, y = sin_1_quantized)[name = string("op_453")];
             tensor<fp16, [1, 3, 128, 256]> q_11 = add(x = var_447, y = var_453)[name = string("q_11")];
-            tensor<fp16, [1, 1, 128, 256]> var_455 = mul(x = out_17_cast_fp16, y = cos_1)[name = string("op_455")];
+            tensor<fp16, [1, 1, 128, 256]> var_455 = mul(x = out_17_cast_fp16, y = cos_1_quantized)[name = string("op_455")];
             tensor<int32, [2]> var_456_split_sizes_0 = const()[name = string("op_456_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_456_axis_0 = const()[name = string("op_456_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 128]> var_456_0, tensor<fp16, [1, 1, 128, 128]> var_456_1 = split(axis = var_456_axis_0, split_sizes = var_456_split_sizes_0, x = out_17_cast_fp16)[name = string("op_456")];
@@ -448,7 +453,7 @@ program(1.3)
             tensor<fp16, [1, 1, 128, 128]> var_458 = mul(x = var_456_1, y = const_21_promoted)[name = string("op_458")];
             bool var_460_interleave_0 = const()[name = string("op_460_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 1, 128, 256]> var_460 = concat(axis = var_23, interleave = var_460_interleave_0, values = (var_458, var_456_0))[name = string("op_460")];
-            tensor<fp16, [1, 1, 128, 256]> var_461 = mul(x = var_460, y = sin_1)[name = string("op_461")];
+            tensor<fp16, [1, 1, 128, 256]> var_461 = mul(x = var_460, y = sin_1_quantized)[name = string("op_461")];
             tensor<fp16, [1, 1, 128, 256]> hidden_states_13 = add(x = var_455, y = var_461)[name = string("hidden_states_13")];
             tensor<int32, [1]> hidden_states_15_axes_0 = const()[name = string("hidden_states_15_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 1, 1, 128, 256]> hidden_states_15 = expand_dims(axes = hidden_states_15_axes_0, x = hidden_states_13)[name = string("hidden_states_15")];
@@ -478,14 +483,14 @@ program(1.3)
             tensor<fp16, [1, 128, 3, 256]> var_488 = transpose(perm = var_487, x = var_485_cast_fp16)[name = string("transpose_202")];
             tensor<fp16, [1, 128, 768]> attn_out_9 = reshape(shape = var_490, x = var_488)[name = string("attn_out_9")];
             tensor<int32, [3]> var_492 = const()[name = string("op_492"), val = tensor<int32, [3]>([0, 2, 1])];
-            tensor<fp16, [768, 768, 1]> squeeze_1 = const()[name = string("squeeze_1"), val = tensor<fp16, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(588147968)))];
+            tensor<fp16, [768, 768, 1]> squeeze_1_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294839808))), scale = tensor<fp16, [768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295429696))))[name = string("squeeze_1_quantized")];
             string var_501_pad_type_0 = const()[name = string("op_501_pad_type_0"), val = string("valid")];
             int32 var_501_groups_0 = const()[name = string("op_501_groups_0"), val = int32(1)];
             tensor<int32, [1]> var_501_strides_0 = const()[name = string("op_501_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> var_501_pad_0 = const()[name = string("op_501_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> var_501_dilations_0 = const()[name = string("op_501_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, [1, 768, 128]> var_493 = transpose(perm = var_492, x = attn_out_9)[name = string("transpose_201")];
-            tensor<fp16, [1, 768, 128]> var_501 = conv(dilations = var_501_dilations_0, groups = var_501_groups_0, pad = var_501_pad_0, pad_type = var_501_pad_type_0, strides = var_501_strides_0, weight = squeeze_1, x = var_493)[name = string("op_501")];
+            tensor<fp16, [1, 768, 128]> var_501 = conv(dilations = var_501_dilations_0, groups = var_501_groups_0, pad = var_501_pad_0, pad_type = var_501_pad_type_0, strides = var_501_strides_0, weight = squeeze_1_quantized, x = var_493)[name = string("op_501")];
             tensor<int32, [3]> var_502 = const()[name = string("op_502"), val = tensor<int32, [3]>([0, 2, 1])];
             fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 128, 768]> x_25 = transpose(perm = var_502, x = var_501)[name = string("transpose_200")];
@@ -497,7 +502,7 @@ program(1.3)
             tensor<int32, [2]> var_511_split_sizes_0 = const()[name = string("op_511_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_511_axis_0 = const()[name = string("op_511_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_511_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_511_cast_fp16_1 = split(axis = var_511_axis_0, split_sizes = var_511_split_sizes_0, x = normed_43_cast_fp16)[name = string("op_511_cast_fp16")];
-            tensor<fp16, [768]> var_515_to_fp16 = const()[name = string("op_515_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(589327680)))];
+            tensor<fp16, [768]> var_515_to_fp16 = const()[name = string("op_515_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295431296)))];
             tensor<fp16, [1, 128, 768]> out_19_cast_fp16 = mul(x = var_511_cast_fp16_0, y = var_515_to_fp16)[name = string("out_19_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_27_cast_fp16 = add(x = x_17_cast_fp16, y = out_19_cast_fp16)[name = string("x_27_cast_fp16")];
             fp16 const_24_promoted_to_fp16 = const()[name = string("const_24_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -509,7 +514,7 @@ program(1.3)
             tensor<int32, [2]> var_527_split_sizes_0 = const()[name = string("op_527_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_527_axis_0 = const()[name = string("op_527_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_527_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_527_cast_fp16_1 = split(axis = var_527_axis_0, split_sizes = var_527_split_sizes_0, x = normed_47_cast_fp16)[name = string("op_527_cast_fp16")];
-            tensor<fp16, [768]> var_531_to_fp16 = const()[name = string("op_531_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(589329280)))];
+            tensor<fp16, [768]> var_531_to_fp16 = const()[name = string("op_531_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295432896)))];
             tensor<fp16, [1, 128, 768]> out_21_cast_fp16 = mul(x = var_527_cast_fp16_0, y = var_531_to_fp16)[name = string("out_21_cast_fp16")];
             tensor<int32, [3]> var_538 = const()[name = string("op_538"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> input_35_axes_0 = const()[name = string("input_35_axes_0"), val = tensor<int32, [1]>([2])];
@@ -520,13 +525,13 @@ program(1.3)
             tensor<int32, [4]> gate_5_pad_0 = const()[name = string("gate_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> gate_5_dilations_0 = const()[name = string("gate_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 gate_5_groups_0 = const()[name = string("gate_5_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> gate_5 = conv(dilations = gate_5_dilations_0, groups = gate_5_groups_0, pad = gate_5_pad_0, pad_type = gate_5_pad_type_0, strides = gate_5_strides_0, weight = encoder_layers_1_mlp_gate_proj_weight, x = input_35)[name = string("gate_5")];
+            tensor<fp16, [1, 1152, 1, 128]> gate_5 = conv(dilations = gate_5_dilations_0, groups = gate_5_groups_0, pad = gate_5_pad_0, pad_type = gate_5_pad_type_0, strides = gate_5_strides_0, weight = encoder_layers_1_mlp_gate_proj_weight_quantized, x = input_35)[name = string("gate_5")];
             string up_3_pad_type_0 = const()[name = string("up_3_pad_type_0"), val = string("valid")];
             tensor<int32, [2]> up_3_strides_0 = const()[name = string("up_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> up_3_pad_0 = const()[name = string("up_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> up_3_dilations_0 = const()[name = string("up_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 up_3_groups_0 = const()[name = string("up_3_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> up_3 = conv(dilations = up_3_dilations_0, groups = up_3_groups_0, pad = up_3_pad_0, pad_type = up_3_pad_type_0, strides = up_3_strides_0, weight = encoder_layers_1_mlp_up_proj_weight, x = input_35)[name = string("up_3")];
+            tensor<fp16, [1, 1152, 1, 128]> up_3 = conv(dilations = up_3_dilations_0, groups = up_3_groups_0, pad = up_3_pad_0, pad_type = up_3_pad_type_0, strides = up_3_strides_0, weight = encoder_layers_1_mlp_up_proj_weight_quantized, x = input_35)[name = string("up_3")];
             string gate_7_mode_0 = const()[name = string("gate_7_mode_0"), val = string("TANH_APPROXIMATION")];
             tensor<fp16, [1, 1152, 1, 128]> gate_7 = gelu(mode = gate_7_mode_0, x = gate_5)[name = string("gate_7")];
             tensor<fp16, [1, 1152, 1, 128]> input_37 = mul(x = gate_7, y = up_3)[name = string("input_37")];
@@ -535,7 +540,7 @@ program(1.3)
             tensor<int32, [4]> var_560_pad_0 = const()[name = string("op_560_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_560_dilations_0 = const()[name = string("op_560_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_560_groups_0 = const()[name = string("op_560_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_560 = conv(dilations = var_560_dilations_0, groups = var_560_groups_0, pad = var_560_pad_0, pad_type = var_560_pad_type_0, strides = var_560_strides_0, weight = encoder_layers_1_mlp_down_proj_weight, x = input_37)[name = string("op_560")];
+            tensor<fp16, [1, 768, 1, 128]> var_560 = conv(dilations = var_560_dilations_0, groups = var_560_groups_0, pad = var_560_pad_0, pad_type = var_560_pad_type_0, strides = var_560_strides_0, weight = encoder_layers_1_mlp_down_proj_weight_quantized, x = input_37)[name = string("op_560")];
             tensor<int32, [1]> var_561_axes_0 = const()[name = string("op_561_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 768, 128]> var_561 = squeeze(axes = var_561_axes_0, x = var_560)[name = string("op_561")];
             tensor<int32, [3]> var_562 = const()[name = string("op_562"), val = tensor<int32, [3]>([0, 2, 1])];
@@ -549,7 +554,7 @@ program(1.3)
             tensor<int32, [2]> var_571_split_sizes_0 = const()[name = string("op_571_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_571_axis_0 = const()[name = string("op_571_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_571_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_571_cast_fp16_1 = split(axis = var_571_axis_0, split_sizes = var_571_split_sizes_0, x = normed_53_cast_fp16)[name = string("op_571_cast_fp16")];
-            tensor<fp16, [768]> var_575_to_fp16 = const()[name = string("op_575_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(589330880)))];
+            tensor<fp16, [768]> var_575_to_fp16 = const()[name = string("op_575_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295434496)))];
             tensor<fp16, [1, 128, 768]> out_23_cast_fp16 = mul(x = var_571_cast_fp16_0, y = var_575_to_fp16)[name = string("out_23_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_33_cast_fp16 = add(x = x_27_cast_fp16, y = out_23_cast_fp16)[name = string("x_33_cast_fp16")];
             fp16 const_28_promoted_to_fp16 = const()[name = string("const_28_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -561,7 +566,7 @@ program(1.3)
             tensor<int32, [2]> var_609_split_sizes_0 = const()[name = string("op_609_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_609_axis_0 = const()[name = string("op_609_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_609_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_609_cast_fp16_1 = split(axis = var_609_axis_0, split_sizes = var_609_split_sizes_0, x = normed_57_cast_fp16)[name = string("op_609_cast_fp16")];
-            tensor<fp16, [768]> var_613_to_fp16 = const()[name = string("op_613_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(589332480)))];
+            tensor<fp16, [768]> var_613_to_fp16 = const()[name = string("op_613_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295436096)))];
             tensor<fp16, [1, 128, 768]> out_25_cast_fp16 = mul(x = var_609_cast_fp16_0, y = var_613_to_fp16)[name = string("out_25_cast_fp16")];
             tensor<int32, [3]> var_619 = const()[name = string("op_619"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> var_621_axes_0 = const()[name = string("op_621_axes_0"), val = tensor<int32, [1]>([2])];
@@ -572,7 +577,7 @@ program(1.3)
             tensor<int32, [4]> var_628_pad_0 = const()[name = string("op_628_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_628_dilations_0 = const()[name = string("op_628_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_628_groups_0 = const()[name = string("op_628_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_628 = conv(dilations = var_628_dilations_0, groups = var_628_groups_0, pad = var_628_pad_0, pad_type = var_628_pad_type_0, strides = var_628_strides_0, weight = encoder_layers_2_self_attn_q_proj_weight, x = var_621_cast_fp16)[name = string("op_628")];
+            tensor<fp16, [1, 768, 1, 128]> var_628 = conv(dilations = var_628_dilations_0, groups = var_628_groups_0, pad = var_628_pad_0, pad_type = var_628_pad_type_0, strides = var_628_strides_0, weight = encoder_layers_2_self_attn_q_proj_weight_quantized, x = var_621_cast_fp16)[name = string("op_628")];
             tensor<int32, [4]> var_629 = const()[name = string("op_629"), val = tensor<int32, [4]>([1, 3, 256, 128])];
             tensor<fp16, [1, 3, 256, 128]> var_630 = reshape(shape = var_629, x = var_628)[name = string("op_630")];
             tensor<int32, [4]> var_631 = const()[name = string("op_631"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -581,7 +586,7 @@ program(1.3)
             tensor<int32, [4]> var_638_pad_0 = const()[name = string("op_638_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_638_dilations_0 = const()[name = string("op_638_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_638_groups_0 = const()[name = string("op_638_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_638 = conv(dilations = var_638_dilations_0, groups = var_638_groups_0, pad = var_638_pad_0, pad_type = var_638_pad_type_0, strides = var_638_strides_0, weight = encoder_layers_2_self_attn_k_proj_weight, x = var_621_cast_fp16)[name = string("op_638")];
+            tensor<fp16, [1, 256, 1, 128]> var_638 = conv(dilations = var_638_dilations_0, groups = var_638_groups_0, pad = var_638_pad_0, pad_type = var_638_pad_type_0, strides = var_638_strides_0, weight = encoder_layers_2_self_attn_k_proj_weight_quantized, x = var_621_cast_fp16)[name = string("op_638")];
             tensor<int32, [4]> var_639 = const()[name = string("op_639"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_640 = reshape(shape = var_639, x = var_638)[name = string("op_640")];
             tensor<int32, [4]> var_641 = const()[name = string("op_641"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -590,7 +595,7 @@ program(1.3)
             tensor<int32, [4]> var_648_pad_0 = const()[name = string("op_648_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_648_dilations_0 = const()[name = string("op_648_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_648_groups_0 = const()[name = string("op_648_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_648 = conv(dilations = var_648_dilations_0, groups = var_648_groups_0, pad = var_648_pad_0, pad_type = var_648_pad_type_0, strides = var_648_strides_0, weight = encoder_layers_2_self_attn_v_proj_weight, x = var_621_cast_fp16)[name = string("op_648")];
+            tensor<fp16, [1, 256, 1, 128]> var_648 = conv(dilations = var_648_dilations_0, groups = var_648_groups_0, pad = var_648_pad_0, pad_type = var_648_pad_type_0, strides = var_648_strides_0, weight = encoder_layers_2_self_attn_v_proj_weight_quantized, x = var_621_cast_fp16)[name = string("op_648")];
             tensor<int32, [4]> var_649 = const()[name = string("op_649"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_650 = reshape(shape = var_649, x = var_648)[name = string("op_650")];
             tensor<int32, [4]> var_651 = const()[name = string("op_651"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -604,7 +609,7 @@ program(1.3)
             tensor<int32, [2]> var_662_split_sizes_0 = const()[name = string("op_662_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_662_axis_0 = const()[name = string("op_662_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 256]> var_662_cast_fp16_0, tensor<fp16, [1, 3, 128, 256]> var_662_cast_fp16_1 = split(axis = var_662_axis_0, split_sizes = var_662_split_sizes_0, x = normed_63_cast_fp16)[name = string("op_662_cast_fp16")];
-            tensor<fp16, [256]> var_666_to_fp16 = const()[name = string("op_666_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(589334080)))];
+            tensor<fp16, [256]> var_666_to_fp16 = const()[name = string("op_666_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295437696)))];
             tensor<fp16, [1, 3, 128, 256]> out_27_cast_fp16 = mul(x = var_662_cast_fp16_0, y = var_666_to_fp16)[name = string("out_27_cast_fp16")];
             fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 1, 128, 256]> k_13 = transpose(perm = var_641, x = var_640)[name = string("transpose_195")];
@@ -616,9 +621,9 @@ program(1.3)
             tensor<int32, [2]> var_678_split_sizes_0 = const()[name = string("op_678_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_678_axis_0 = const()[name = string("op_678_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 256]> var_678_cast_fp16_0, tensor<fp16, [1, 1, 128, 256]> var_678_cast_fp16_1 = split(axis = var_678_axis_0, split_sizes = var_678_split_sizes_0, x = normed_67_cast_fp16)[name = string("op_678_cast_fp16")];
-            tensor<fp16, [256]> var_682_to_fp16 = const()[name = string("op_682_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(589334656)))];
+            tensor<fp16, [256]> var_682_to_fp16 = const()[name = string("op_682_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295438272)))];
             tensor<fp16, [1, 1, 128, 256]> out_29_cast_fp16 = mul(x = var_678_cast_fp16_0, y = var_682_to_fp16)[name = string("out_29_cast_fp16")];
-            tensor<fp16, [1, 3, 128, 256]> var_685 = mul(x = out_27_cast_fp16, y = cos_1)[name = string("op_685")];
+            tensor<fp16, [1, 3, 128, 256]> var_685 = mul(x = out_27_cast_fp16, y = cos_1_quantized)[name = string("op_685")];
             tensor<int32, [2]> var_686_split_sizes_0 = const()[name = string("op_686_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_686_axis_0 = const()[name = string("op_686_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 128]> var_686_0, tensor<fp16, [1, 3, 128, 128]> var_686_1 = split(axis = var_686_axis_0, split_sizes = var_686_split_sizes_0, x = out_27_cast_fp16)[name = string("op_686")];
@@ -626,9 +631,9 @@ program(1.3)
             tensor<fp16, [1, 3, 128, 128]> var_688 = mul(x = var_686_1, y = const_34_promoted)[name = string("op_688")];
             bool var_690_interleave_0 = const()[name = string("op_690_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 3, 128, 256]> var_690 = concat(axis = var_23, interleave = var_690_interleave_0, values = (var_688, var_686_0))[name = string("op_690")];
-            tensor<fp16, [1, 3, 128, 256]> var_691 = mul(x = var_690, y = sin_1)[name = string("op_691")];
+            tensor<fp16, [1, 3, 128, 256]> var_691 = mul(x = var_690, y = sin_1_quantized)[name = string("op_691")];
             tensor<fp16, [1, 3, 128, 256]> q_17 = add(x = var_685, y = var_691)[name = string("q_17")];
-            tensor<fp16, [1, 1, 128, 256]> var_693 = mul(x = out_29_cast_fp16, y = cos_1)[name = string("op_693")];
+            tensor<fp16, [1, 1, 128, 256]> var_693 = mul(x = out_29_cast_fp16, y = cos_1_quantized)[name = string("op_693")];
             tensor<int32, [2]> var_694_split_sizes_0 = const()[name = string("op_694_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_694_axis_0 = const()[name = string("op_694_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 128]> var_694_0, tensor<fp16, [1, 1, 128, 128]> var_694_1 = split(axis = var_694_axis_0, split_sizes = var_694_split_sizes_0, x = out_29_cast_fp16)[name = string("op_694")];
@@ -636,7 +641,7 @@ program(1.3)
             tensor<fp16, [1, 1, 128, 128]> var_696 = mul(x = var_694_1, y = const_35_promoted)[name = string("op_696")];
             bool var_698_interleave_0 = const()[name = string("op_698_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 1, 128, 256]> var_698 = concat(axis = var_23, interleave = var_698_interleave_0, values = (var_696, var_694_0))[name = string("op_698")];
-            tensor<fp16, [1, 1, 128, 256]> var_699 = mul(x = var_698, y = sin_1)[name = string("op_699")];
+            tensor<fp16, [1, 1, 128, 256]> var_699 = mul(x = var_698, y = sin_1_quantized)[name = string("op_699")];
             tensor<fp16, [1, 1, 128, 256]> hidden_states_25 = add(x = var_693, y = var_699)[name = string("hidden_states_25")];
             tensor<int32, [1]> hidden_states_27_axes_0 = const()[name = string("hidden_states_27_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 1, 1, 128, 256]> hidden_states_27 = expand_dims(axes = hidden_states_27_axes_0, x = hidden_states_25)[name = string("hidden_states_27")];
@@ -666,14 +671,14 @@ program(1.3)
             tensor<fp16, [1, 128, 3, 256]> var_726 = transpose(perm = var_725, x = var_723_cast_fp16)[name = string("transpose_193")];
             tensor<fp16, [1, 128, 768]> attn_out_15 = reshape(shape = var_728, x = var_726)[name = string("attn_out_15")];
             tensor<int32, [3]> var_730 = const()[name = string("op_730"), val = tensor<int32, [3]>([0, 2, 1])];
-            tensor<fp16, [768, 768, 1]> squeeze_2 = const()[name = string("squeeze_2"), val = tensor<fp16, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(589335232)))];
+            tensor<fp16, [768, 768, 1]> squeeze_2_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295438848))), scale = tensor<fp16, [768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296028736))))[name = string("squeeze_2_quantized")];
             string var_739_pad_type_0 = const()[name = string("op_739_pad_type_0"), val = string("valid")];
             int32 var_739_groups_0 = const()[name = string("op_739_groups_0"), val = int32(1)];
             tensor<int32, [1]> var_739_strides_0 = const()[name = string("op_739_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> var_739_pad_0 = const()[name = string("op_739_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> var_739_dilations_0 = const()[name = string("op_739_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, [1, 768, 128]> var_731 = transpose(perm = var_730, x = attn_out_15)[name = string("transpose_192")];
-            tensor<fp16, [1, 768, 128]> 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 = squeeze_2, x = var_731)[name = string("op_739")];
+            tensor<fp16, [1, 768, 128]> 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 = squeeze_2_quantized, x = var_731)[name = string("op_739")];
             tensor<int32, [3]> var_740 = const()[name = string("op_740"), val = tensor<int32, [3]>([0, 2, 1])];
             fp16 const_36_promoted_to_fp16 = const()[name = string("const_36_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 128, 768]> x_41 = transpose(perm = var_740, x = var_739)[name = string("transpose_191")];
@@ -685,7 +690,7 @@ program(1.3)
             tensor<int32, [2]> var_749_split_sizes_0 = const()[name = string("op_749_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_749_axis_0 = const()[name = string("op_749_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_749_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_749_cast_fp16_1 = split(axis = var_749_axis_0, split_sizes = var_749_split_sizes_0, x = normed_71_cast_fp16)[name = string("op_749_cast_fp16")];
-            tensor<fp16, [768]> var_753_to_fp16 = const()[name = string("op_753_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(590514944)))];
+            tensor<fp16, [768]> var_753_to_fp16 = const()[name = string("op_753_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296030336)))];
             tensor<fp16, [1, 128, 768]> out_31_cast_fp16 = mul(x = var_749_cast_fp16_0, y = var_753_to_fp16)[name = string("out_31_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_43_cast_fp16 = add(x = x_33_cast_fp16, y = out_31_cast_fp16)[name = string("x_43_cast_fp16")];
             fp16 const_38_promoted_to_fp16 = const()[name = string("const_38_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -697,7 +702,7 @@ program(1.3)
             tensor<int32, [2]> var_765_split_sizes_0 = const()[name = string("op_765_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_765_axis_0 = const()[name = string("op_765_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_765_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_765_cast_fp16_1 = split(axis = var_765_axis_0, split_sizes = var_765_split_sizes_0, x = normed_75_cast_fp16)[name = string("op_765_cast_fp16")];
-            tensor<fp16, [768]> var_769_to_fp16 = const()[name = string("op_769_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(590516544)))];
+            tensor<fp16, [768]> var_769_to_fp16 = const()[name = string("op_769_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296031936)))];
             tensor<fp16, [1, 128, 768]> out_33_cast_fp16 = mul(x = var_765_cast_fp16_0, y = var_769_to_fp16)[name = string("out_33_cast_fp16")];
             tensor<int32, [3]> var_776 = const()[name = string("op_776"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> input_55_axes_0 = const()[name = string("input_55_axes_0"), val = tensor<int32, [1]>([2])];
@@ -708,13 +713,13 @@ program(1.3)
             tensor<int32, [4]> gate_9_pad_0 = const()[name = string("gate_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> gate_9_dilations_0 = const()[name = string("gate_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 gate_9_groups_0 = const()[name = string("gate_9_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> gate_9 = conv(dilations = gate_9_dilations_0, groups = gate_9_groups_0, pad = gate_9_pad_0, pad_type = gate_9_pad_type_0, strides = gate_9_strides_0, weight = encoder_layers_2_mlp_gate_proj_weight, x = input_55)[name = string("gate_9")];
+            tensor<fp16, [1, 1152, 1, 128]> gate_9 = conv(dilations = gate_9_dilations_0, groups = gate_9_groups_0, pad = gate_9_pad_0, pad_type = gate_9_pad_type_0, strides = gate_9_strides_0, weight = encoder_layers_2_mlp_gate_proj_weight_quantized, x = input_55)[name = string("gate_9")];
             string up_5_pad_type_0 = const()[name = string("up_5_pad_type_0"), val = string("valid")];
             tensor<int32, [2]> up_5_strides_0 = const()[name = string("up_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> up_5_pad_0 = const()[name = string("up_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> up_5_dilations_0 = const()[name = string("up_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 up_5_groups_0 = const()[name = string("up_5_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> up_5 = conv(dilations = up_5_dilations_0, groups = up_5_groups_0, pad = up_5_pad_0, pad_type = up_5_pad_type_0, strides = up_5_strides_0, weight = encoder_layers_2_mlp_up_proj_weight, x = input_55)[name = string("up_5")];
+            tensor<fp16, [1, 1152, 1, 128]> up_5 = conv(dilations = up_5_dilations_0, groups = up_5_groups_0, pad = up_5_pad_0, pad_type = up_5_pad_type_0, strides = up_5_strides_0, weight = encoder_layers_2_mlp_up_proj_weight_quantized, x = input_55)[name = string("up_5")];
             string gate_11_mode_0 = const()[name = string("gate_11_mode_0"), val = string("TANH_APPROXIMATION")];
             tensor<fp16, [1, 1152, 1, 128]> gate_11 = gelu(mode = gate_11_mode_0, x = gate_9)[name = string("gate_11")];
             tensor<fp16, [1, 1152, 1, 128]> input_57 = mul(x = gate_11, y = up_5)[name = string("input_57")];
@@ -723,7 +728,7 @@ program(1.3)
             tensor<int32, [4]> var_798_pad_0 = const()[name = string("op_798_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_798_dilations_0 = const()[name = string("op_798_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_798_groups_0 = const()[name = string("op_798_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_798 = conv(dilations = var_798_dilations_0, groups = var_798_groups_0, pad = var_798_pad_0, pad_type = var_798_pad_type_0, strides = var_798_strides_0, weight = encoder_layers_2_mlp_down_proj_weight, x = input_57)[name = string("op_798")];
+            tensor<fp16, [1, 768, 1, 128]> var_798 = conv(dilations = var_798_dilations_0, groups = var_798_groups_0, pad = var_798_pad_0, pad_type = var_798_pad_type_0, strides = var_798_strides_0, weight = encoder_layers_2_mlp_down_proj_weight_quantized, x = input_57)[name = string("op_798")];
             tensor<int32, [1]> var_799_axes_0 = const()[name = string("op_799_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 768, 128]> var_799 = squeeze(axes = var_799_axes_0, x = var_798)[name = string("op_799")];
             tensor<int32, [3]> var_800 = const()[name = string("op_800"), val = tensor<int32, [3]>([0, 2, 1])];
@@ -737,7 +742,7 @@ program(1.3)
             tensor<int32, [2]> var_809_split_sizes_0 = const()[name = string("op_809_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_809_axis_0 = const()[name = string("op_809_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_809_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_809_cast_fp16_1 = split(axis = var_809_axis_0, split_sizes = var_809_split_sizes_0, x = normed_81_cast_fp16)[name = string("op_809_cast_fp16")];
-            tensor<fp16, [768]> var_813_to_fp16 = const()[name = string("op_813_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(590518144)))];
+            tensor<fp16, [768]> var_813_to_fp16 = const()[name = string("op_813_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296033536)))];
             tensor<fp16, [1, 128, 768]> out_35_cast_fp16 = mul(x = var_809_cast_fp16_0, y = var_813_to_fp16)[name = string("out_35_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_49_cast_fp16 = add(x = x_43_cast_fp16, y = out_35_cast_fp16)[name = string("x_49_cast_fp16")];
             fp16 const_42_promoted_to_fp16 = const()[name = string("const_42_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -749,7 +754,7 @@ program(1.3)
             tensor<int32, [2]> var_847_split_sizes_0 = const()[name = string("op_847_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_847_axis_0 = const()[name = string("op_847_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_847_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_847_cast_fp16_1 = split(axis = var_847_axis_0, split_sizes = var_847_split_sizes_0, x = normed_85_cast_fp16)[name = string("op_847_cast_fp16")];
-            tensor<fp16, [768]> var_851_to_fp16 = const()[name = string("op_851_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(590519744)))];
+            tensor<fp16, [768]> var_851_to_fp16 = const()[name = string("op_851_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296035136)))];
             tensor<fp16, [1, 128, 768]> out_37_cast_fp16 = mul(x = var_847_cast_fp16_0, y = var_851_to_fp16)[name = string("out_37_cast_fp16")];
             tensor<int32, [3]> var_857 = const()[name = string("op_857"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> var_859_axes_0 = const()[name = string("op_859_axes_0"), val = tensor<int32, [1]>([2])];
@@ -760,7 +765,7 @@ program(1.3)
             tensor<int32, [4]> var_866_pad_0 = const()[name = string("op_866_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_866_dilations_0 = const()[name = string("op_866_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_866_groups_0 = const()[name = string("op_866_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_866 = conv(dilations = var_866_dilations_0, groups = var_866_groups_0, pad = var_866_pad_0, pad_type = var_866_pad_type_0, strides = var_866_strides_0, weight = encoder_layers_3_self_attn_q_proj_weight, x = var_859_cast_fp16)[name = string("op_866")];
+            tensor<fp16, [1, 768, 1, 128]> var_866 = conv(dilations = var_866_dilations_0, groups = var_866_groups_0, pad = var_866_pad_0, pad_type = var_866_pad_type_0, strides = var_866_strides_0, weight = encoder_layers_3_self_attn_q_proj_weight_quantized, x = var_859_cast_fp16)[name = string("op_866")];
             tensor<int32, [4]> var_867 = const()[name = string("op_867"), val = tensor<int32, [4]>([1, 3, 256, 128])];
             tensor<fp16, [1, 3, 256, 128]> var_868 = reshape(shape = var_867, x = var_866)[name = string("op_868")];
             tensor<int32, [4]> var_869 = const()[name = string("op_869"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -769,7 +774,7 @@ program(1.3)
             tensor<int32, [4]> var_876_pad_0 = const()[name = string("op_876_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_876_dilations_0 = const()[name = string("op_876_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_876_groups_0 = const()[name = string("op_876_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_876 = conv(dilations = var_876_dilations_0, groups = var_876_groups_0, pad = var_876_pad_0, pad_type = var_876_pad_type_0, strides = var_876_strides_0, weight = encoder_layers_3_self_attn_k_proj_weight, x = var_859_cast_fp16)[name = string("op_876")];
+            tensor<fp16, [1, 256, 1, 128]> var_876 = conv(dilations = var_876_dilations_0, groups = var_876_groups_0, pad = var_876_pad_0, pad_type = var_876_pad_type_0, strides = var_876_strides_0, weight = encoder_layers_3_self_attn_k_proj_weight_quantized, x = var_859_cast_fp16)[name = string("op_876")];
             tensor<int32, [4]> var_877 = const()[name = string("op_877"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_878 = reshape(shape = var_877, x = var_876)[name = string("op_878")];
             tensor<int32, [4]> var_879 = const()[name = string("op_879"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -778,7 +783,7 @@ program(1.3)
             tensor<int32, [4]> var_886_pad_0 = const()[name = string("op_886_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_886_dilations_0 = const()[name = string("op_886_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_886_groups_0 = const()[name = string("op_886_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_886 = conv(dilations = var_886_dilations_0, groups = var_886_groups_0, pad = var_886_pad_0, pad_type = var_886_pad_type_0, strides = var_886_strides_0, weight = encoder_layers_3_self_attn_v_proj_weight, x = var_859_cast_fp16)[name = string("op_886")];
+            tensor<fp16, [1, 256, 1, 128]> var_886 = conv(dilations = var_886_dilations_0, groups = var_886_groups_0, pad = var_886_pad_0, pad_type = var_886_pad_type_0, strides = var_886_strides_0, weight = encoder_layers_3_self_attn_v_proj_weight_quantized, x = var_859_cast_fp16)[name = string("op_886")];
             tensor<int32, [4]> var_887 = const()[name = string("op_887"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_888 = reshape(shape = var_887, x = var_886)[name = string("op_888")];
             tensor<int32, [4]> var_889 = const()[name = string("op_889"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -792,7 +797,7 @@ program(1.3)
             tensor<int32, [2]> var_900_split_sizes_0 = const()[name = string("op_900_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_900_axis_0 = const()[name = string("op_900_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 256]> var_900_cast_fp16_0, tensor<fp16, [1, 3, 128, 256]> var_900_cast_fp16_1 = split(axis = var_900_axis_0, split_sizes = var_900_split_sizes_0, x = normed_91_cast_fp16)[name = string("op_900_cast_fp16")];
-            tensor<fp16, [256]> var_904_to_fp16 = const()[name = string("op_904_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(590521344)))];
+            tensor<fp16, [256]> var_904_to_fp16 = const()[name = string("op_904_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296036736)))];
             tensor<fp16, [1, 3, 128, 256]> out_39_cast_fp16 = mul(x = var_900_cast_fp16_0, y = var_904_to_fp16)[name = string("out_39_cast_fp16")];
             fp16 const_46_promoted_to_fp16 = const()[name = string("const_46_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 1, 128, 256]> k_19 = transpose(perm = var_879, x = var_878)[name = string("transpose_186")];
@@ -804,9 +809,9 @@ program(1.3)
             tensor<int32, [2]> var_916_split_sizes_0 = const()[name = string("op_916_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_916_axis_0 = const()[name = string("op_916_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 256]> var_916_cast_fp16_0, tensor<fp16, [1, 1, 128, 256]> var_916_cast_fp16_1 = split(axis = var_916_axis_0, split_sizes = var_916_split_sizes_0, x = normed_95_cast_fp16)[name = string("op_916_cast_fp16")];
-            tensor<fp16, [256]> var_920_to_fp16 = const()[name = string("op_920_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(590521920)))];
+            tensor<fp16, [256]> var_920_to_fp16 = const()[name = string("op_920_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296037312)))];
             tensor<fp16, [1, 1, 128, 256]> out_41_cast_fp16 = mul(x = var_916_cast_fp16_0, y = var_920_to_fp16)[name = string("out_41_cast_fp16")];
-            tensor<fp16, [1, 3, 128, 256]> var_923 = mul(x = out_39_cast_fp16, y = cos_1)[name = string("op_923")];
+            tensor<fp16, [1, 3, 128, 256]> var_923 = mul(x = out_39_cast_fp16, y = cos_1_quantized)[name = string("op_923")];
             tensor<int32, [2]> var_924_split_sizes_0 = const()[name = string("op_924_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_924_axis_0 = const()[name = string("op_924_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 128]> var_924_0, tensor<fp16, [1, 3, 128, 128]> var_924_1 = split(axis = var_924_axis_0, split_sizes = var_924_split_sizes_0, x = out_39_cast_fp16)[name = string("op_924")];
@@ -814,9 +819,9 @@ program(1.3)
             tensor<fp16, [1, 3, 128, 128]> var_926 = mul(x = var_924_1, y = const_48_promoted)[name = string("op_926")];
             bool var_928_interleave_0 = const()[name = string("op_928_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 3, 128, 256]> var_928 = concat(axis = var_23, interleave = var_928_interleave_0, values = (var_926, var_924_0))[name = string("op_928")];
-            tensor<fp16, [1, 3, 128, 256]> var_929 = mul(x = var_928, y = sin_1)[name = string("op_929")];
+            tensor<fp16, [1, 3, 128, 256]> var_929 = mul(x = var_928, y = sin_1_quantized)[name = string("op_929")];
             tensor<fp16, [1, 3, 128, 256]> q_23 = add(x = var_923, y = var_929)[name = string("q_23")];
-            tensor<fp16, [1, 1, 128, 256]> var_931 = mul(x = out_41_cast_fp16, y = cos_1)[name = string("op_931")];
+            tensor<fp16, [1, 1, 128, 256]> var_931 = mul(x = out_41_cast_fp16, y = cos_1_quantized)[name = string("op_931")];
             tensor<int32, [2]> var_932_split_sizes_0 = const()[name = string("op_932_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_932_axis_0 = const()[name = string("op_932_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 128]> var_932_0, tensor<fp16, [1, 1, 128, 128]> var_932_1 = split(axis = var_932_axis_0, split_sizes = var_932_split_sizes_0, x = out_41_cast_fp16)[name = string("op_932")];
@@ -824,7 +829,7 @@ program(1.3)
             tensor<fp16, [1, 1, 128, 128]> var_934 = mul(x = var_932_1, y = const_49_promoted)[name = string("op_934")];
             bool var_936_interleave_0 = const()[name = string("op_936_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 1, 128, 256]> var_936 = concat(axis = var_23, interleave = var_936_interleave_0, values = (var_934, var_932_0))[name = string("op_936")];
-            tensor<fp16, [1, 1, 128, 256]> var_937 = mul(x = var_936, y = sin_1)[name = string("op_937")];
+            tensor<fp16, [1, 1, 128, 256]> var_937 = mul(x = var_936, y = sin_1_quantized)[name = string("op_937")];
             tensor<fp16, [1, 1, 128, 256]> hidden_states_37 = add(x = var_931, y = var_937)[name = string("hidden_states_37")];
             tensor<int32, [1]> hidden_states_39_axes_0 = const()[name = string("hidden_states_39_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 1, 1, 128, 256]> hidden_states_39 = expand_dims(axes = hidden_states_39_axes_0, x = hidden_states_37)[name = string("hidden_states_39")];
@@ -854,14 +859,14 @@ program(1.3)
             tensor<fp16, [1, 128, 3, 256]> var_964 = transpose(perm = var_963, x = var_961_cast_fp16)[name = string("transpose_184")];
             tensor<fp16, [1, 128, 768]> attn_out_21 = reshape(shape = var_966, x = var_964)[name = string("attn_out_21")];
             tensor<int32, [3]> var_968 = const()[name = string("op_968"), val = tensor<int32, [3]>([0, 2, 1])];
-            tensor<fp16, [768, 768, 1]> squeeze_3 = const()[name = string("squeeze_3"), val = tensor<fp16, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(590522496)))];
+            tensor<fp16, [768, 768, 1]> squeeze_3_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296037888))), scale = tensor<fp16, [768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296627776))))[name = string("squeeze_3_quantized")];
             string var_977_pad_type_0 = const()[name = string("op_977_pad_type_0"), val = string("valid")];
             int32 var_977_groups_0 = const()[name = string("op_977_groups_0"), val = int32(1)];
             tensor<int32, [1]> var_977_strides_0 = const()[name = string("op_977_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> var_977_pad_0 = const()[name = string("op_977_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> var_977_dilations_0 = const()[name = string("op_977_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, [1, 768, 128]> var_969 = transpose(perm = var_968, x = attn_out_21)[name = string("transpose_183")];
-            tensor<fp16, [1, 768, 128]> var_977 = conv(dilations = var_977_dilations_0, groups = var_977_groups_0, pad = var_977_pad_0, pad_type = var_977_pad_type_0, strides = var_977_strides_0, weight = squeeze_3, x = var_969)[name = string("op_977")];
+            tensor<fp16, [1, 768, 128]> var_977 = conv(dilations = var_977_dilations_0, groups = var_977_groups_0, pad = var_977_pad_0, pad_type = var_977_pad_type_0, strides = var_977_strides_0, weight = squeeze_3_quantized, x = var_969)[name = string("op_977")];
             tensor<int32, [3]> var_978 = const()[name = string("op_978"), val = tensor<int32, [3]>([0, 2, 1])];
             fp16 const_50_promoted_to_fp16 = const()[name = string("const_50_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 128, 768]> x_57 = transpose(perm = var_978, x = var_977)[name = string("transpose_182")];
@@ -873,7 +878,7 @@ program(1.3)
             tensor<int32, [2]> var_987_split_sizes_0 = const()[name = string("op_987_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_987_axis_0 = const()[name = string("op_987_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_987_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_987_cast_fp16_1 = split(axis = var_987_axis_0, split_sizes = var_987_split_sizes_0, x = normed_99_cast_fp16)[name = string("op_987_cast_fp16")];
-            tensor<fp16, [768]> var_991_to_fp16 = const()[name = string("op_991_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(591702208)))];
+            tensor<fp16, [768]> var_991_to_fp16 = const()[name = string("op_991_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296629376)))];
             tensor<fp16, [1, 128, 768]> out_43_cast_fp16 = mul(x = var_987_cast_fp16_0, y = var_991_to_fp16)[name = string("out_43_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_59_cast_fp16 = add(x = x_49_cast_fp16, y = out_43_cast_fp16)[name = string("x_59_cast_fp16")];
             fp16 const_52_promoted_to_fp16 = const()[name = string("const_52_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -885,7 +890,7 @@ program(1.3)
             tensor<int32, [2]> var_1003_split_sizes_0 = const()[name = string("op_1003_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_1003_axis_0 = const()[name = string("op_1003_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_1003_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_1003_cast_fp16_1 = split(axis = var_1003_axis_0, split_sizes = var_1003_split_sizes_0, x = normed_103_cast_fp16)[name = string("op_1003_cast_fp16")];
-            tensor<fp16, [768]> var_1007_to_fp16 = const()[name = string("op_1007_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(591703808)))];
+            tensor<fp16, [768]> var_1007_to_fp16 = const()[name = string("op_1007_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296630976)))];
             tensor<fp16, [1, 128, 768]> out_45_cast_fp16 = mul(x = var_1003_cast_fp16_0, y = var_1007_to_fp16)[name = string("out_45_cast_fp16")];
             tensor<int32, [3]> var_1014 = const()[name = string("op_1014"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> input_75_axes_0 = const()[name = string("input_75_axes_0"), val = tensor<int32, [1]>([2])];
@@ -896,13 +901,13 @@ program(1.3)
             tensor<int32, [4]> gate_13_pad_0 = const()[name = string("gate_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> gate_13_dilations_0 = const()[name = string("gate_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 gate_13_groups_0 = const()[name = string("gate_13_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> gate_13 = conv(dilations = gate_13_dilations_0, groups = gate_13_groups_0, pad = gate_13_pad_0, pad_type = gate_13_pad_type_0, strides = gate_13_strides_0, weight = encoder_layers_3_mlp_gate_proj_weight, x = input_75)[name = string("gate_13")];
+            tensor<fp16, [1, 1152, 1, 128]> gate_13 = conv(dilations = gate_13_dilations_0, groups = gate_13_groups_0, pad = gate_13_pad_0, pad_type = gate_13_pad_type_0, strides = gate_13_strides_0, weight = encoder_layers_3_mlp_gate_proj_weight_quantized, x = input_75)[name = string("gate_13")];
             string up_7_pad_type_0 = const()[name = string("up_7_pad_type_0"), val = string("valid")];
             tensor<int32, [2]> up_7_strides_0 = const()[name = string("up_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> up_7_pad_0 = const()[name = string("up_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> up_7_dilations_0 = const()[name = string("up_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 up_7_groups_0 = const()[name = string("up_7_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> up_7 = conv(dilations = up_7_dilations_0, groups = up_7_groups_0, pad = up_7_pad_0, pad_type = up_7_pad_type_0, strides = up_7_strides_0, weight = encoder_layers_3_mlp_up_proj_weight, x = input_75)[name = string("up_7")];
+            tensor<fp16, [1, 1152, 1, 128]> up_7 = conv(dilations = up_7_dilations_0, groups = up_7_groups_0, pad = up_7_pad_0, pad_type = up_7_pad_type_0, strides = up_7_strides_0, weight = encoder_layers_3_mlp_up_proj_weight_quantized, x = input_75)[name = string("up_7")];
             string gate_15_mode_0 = const()[name = string("gate_15_mode_0"), val = string("TANH_APPROXIMATION")];
             tensor<fp16, [1, 1152, 1, 128]> gate_15 = gelu(mode = gate_15_mode_0, x = gate_13)[name = string("gate_15")];
             tensor<fp16, [1, 1152, 1, 128]> input_77 = mul(x = gate_15, y = up_7)[name = string("input_77")];
@@ -911,7 +916,7 @@ program(1.3)
             tensor<int32, [4]> var_1036_pad_0 = const()[name = string("op_1036_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_1036_dilations_0 = const()[name = string("op_1036_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_1036_groups_0 = const()[name = string("op_1036_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_1036 = conv(dilations = var_1036_dilations_0, groups = var_1036_groups_0, pad = var_1036_pad_0, pad_type = var_1036_pad_type_0, strides = var_1036_strides_0, weight = encoder_layers_3_mlp_down_proj_weight, x = input_77)[name = string("op_1036")];
+            tensor<fp16, [1, 768, 1, 128]> var_1036 = conv(dilations = var_1036_dilations_0, groups = var_1036_groups_0, pad = var_1036_pad_0, pad_type = var_1036_pad_type_0, strides = var_1036_strides_0, weight = encoder_layers_3_mlp_down_proj_weight_quantized, x = input_77)[name = string("op_1036")];
             tensor<int32, [1]> var_1037_axes_0 = const()[name = string("op_1037_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 768, 128]> var_1037 = squeeze(axes = var_1037_axes_0, x = var_1036)[name = string("op_1037")];
             tensor<int32, [3]> var_1038 = const()[name = string("op_1038"), val = tensor<int32, [3]>([0, 2, 1])];
@@ -925,7 +930,7 @@ program(1.3)
             tensor<int32, [2]> var_1047_split_sizes_0 = const()[name = string("op_1047_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_1047_axis_0 = const()[name = string("op_1047_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_1047_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_1047_cast_fp16_1 = split(axis = var_1047_axis_0, split_sizes = var_1047_split_sizes_0, x = normed_109_cast_fp16)[name = string("op_1047_cast_fp16")];
-            tensor<fp16, [768]> var_1051_to_fp16 = const()[name = string("op_1051_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(591705408)))];
+            tensor<fp16, [768]> var_1051_to_fp16 = const()[name = string("op_1051_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296632576)))];
             tensor<fp16, [1, 128, 768]> out_47_cast_fp16 = mul(x = var_1047_cast_fp16_0, y = var_1051_to_fp16)[name = string("out_47_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_65_cast_fp16 = add(x = x_59_cast_fp16, y = out_47_cast_fp16)[name = string("x_65_cast_fp16")];
             fp16 const_56_promoted_to_fp16 = const()[name = string("const_56_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -937,7 +942,7 @@ program(1.3)
             tensor<int32, [2]> var_1085_split_sizes_0 = const()[name = string("op_1085_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_1085_axis_0 = const()[name = string("op_1085_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_1085_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_1085_cast_fp16_1 = split(axis = var_1085_axis_0, split_sizes = var_1085_split_sizes_0, x = normed_113_cast_fp16)[name = string("op_1085_cast_fp16")];
-            tensor<fp16, [768]> var_1089_to_fp16 = const()[name = string("op_1089_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(591707008)))];
+            tensor<fp16, [768]> var_1089_to_fp16 = const()[name = string("op_1089_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296634176)))];
             tensor<fp16, [1, 128, 768]> out_49_cast_fp16 = mul(x = var_1085_cast_fp16_0, y = var_1089_to_fp16)[name = string("out_49_cast_fp16")];
             tensor<int32, [3]> var_1095 = const()[name = string("op_1095"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> var_1097_axes_0 = const()[name = string("op_1097_axes_0"), val = tensor<int32, [1]>([2])];
@@ -948,7 +953,7 @@ program(1.3)
             tensor<int32, [4]> var_1104_pad_0 = const()[name = string("op_1104_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_1104_dilations_0 = const()[name = string("op_1104_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_1104_groups_0 = const()[name = string("op_1104_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_1104 = conv(dilations = var_1104_dilations_0, groups = var_1104_groups_0, pad = var_1104_pad_0, pad_type = var_1104_pad_type_0, strides = var_1104_strides_0, weight = encoder_layers_4_self_attn_q_proj_weight, x = var_1097_cast_fp16)[name = string("op_1104")];
+            tensor<fp16, [1, 768, 1, 128]> var_1104 = conv(dilations = var_1104_dilations_0, groups = var_1104_groups_0, pad = var_1104_pad_0, pad_type = var_1104_pad_type_0, strides = var_1104_strides_0, weight = encoder_layers_4_self_attn_q_proj_weight_quantized, x = var_1097_cast_fp16)[name = string("op_1104")];
             tensor<int32, [4]> var_1105 = const()[name = string("op_1105"), val = tensor<int32, [4]>([1, 3, 256, 128])];
             tensor<fp16, [1, 3, 256, 128]> var_1106 = reshape(shape = var_1105, x = var_1104)[name = string("op_1106")];
             tensor<int32, [4]> var_1107 = const()[name = string("op_1107"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -957,7 +962,7 @@ program(1.3)
             tensor<int32, [4]> var_1114_pad_0 = const()[name = string("op_1114_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_1114_dilations_0 = const()[name = string("op_1114_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_1114_groups_0 = const()[name = string("op_1114_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_1114 = conv(dilations = var_1114_dilations_0, groups = var_1114_groups_0, pad = var_1114_pad_0, pad_type = var_1114_pad_type_0, strides = var_1114_strides_0, weight = encoder_layers_4_self_attn_k_proj_weight, x = var_1097_cast_fp16)[name = string("op_1114")];
+            tensor<fp16, [1, 256, 1, 128]> var_1114 = conv(dilations = var_1114_dilations_0, groups = var_1114_groups_0, pad = var_1114_pad_0, pad_type = var_1114_pad_type_0, strides = var_1114_strides_0, weight = encoder_layers_4_self_attn_k_proj_weight_quantized, x = var_1097_cast_fp16)[name = string("op_1114")];
             tensor<int32, [4]> var_1115 = const()[name = string("op_1115"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_1116 = reshape(shape = var_1115, x = var_1114)[name = string("op_1116")];
             tensor<int32, [4]> var_1117 = const()[name = string("op_1117"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -966,7 +971,7 @@ program(1.3)
             tensor<int32, [4]> var_1124_pad_0 = const()[name = string("op_1124_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_1124_dilations_0 = const()[name = string("op_1124_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_1124_groups_0 = const()[name = string("op_1124_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_1124 = conv(dilations = var_1124_dilations_0, groups = var_1124_groups_0, pad = var_1124_pad_0, pad_type = var_1124_pad_type_0, strides = var_1124_strides_0, weight = encoder_layers_4_self_attn_v_proj_weight, x = var_1097_cast_fp16)[name = string("op_1124")];
+            tensor<fp16, [1, 256, 1, 128]> var_1124 = conv(dilations = var_1124_dilations_0, groups = var_1124_groups_0, pad = var_1124_pad_0, pad_type = var_1124_pad_type_0, strides = var_1124_strides_0, weight = encoder_layers_4_self_attn_v_proj_weight_quantized, x = var_1097_cast_fp16)[name = string("op_1124")];
             tensor<int32, [4]> var_1125 = const()[name = string("op_1125"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_1126 = reshape(shape = var_1125, x = var_1124)[name = string("op_1126")];
             tensor<int32, [4]> var_1127 = const()[name = string("op_1127"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -980,7 +985,7 @@ program(1.3)
             tensor<int32, [2]> var_1138_split_sizes_0 = const()[name = string("op_1138_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_1138_axis_0 = const()[name = string("op_1138_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 256]> var_1138_cast_fp16_0, tensor<fp16, [1, 3, 128, 256]> var_1138_cast_fp16_1 = split(axis = var_1138_axis_0, split_sizes = var_1138_split_sizes_0, x = normed_119_cast_fp16)[name = string("op_1138_cast_fp16")];
-            tensor<fp16, [256]> var_1142_to_fp16 = const()[name = string("op_1142_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(591708608)))];
+            tensor<fp16, [256]> var_1142_to_fp16 = const()[name = string("op_1142_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296635776)))];
             tensor<fp16, [1, 3, 128, 256]> out_51_cast_fp16 = mul(x = var_1138_cast_fp16_0, y = var_1142_to_fp16)[name = string("out_51_cast_fp16")];
             fp16 const_60_promoted_to_fp16 = const()[name = string("const_60_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 1, 128, 256]> k_25 = transpose(perm = var_1117, x = var_1116)[name = string("transpose_177")];
@@ -992,9 +997,9 @@ program(1.3)
             tensor<int32, [2]> var_1154_split_sizes_0 = const()[name = string("op_1154_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_1154_axis_0 = const()[name = string("op_1154_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 256]> var_1154_cast_fp16_0, tensor<fp16, [1, 1, 128, 256]> var_1154_cast_fp16_1 = split(axis = var_1154_axis_0, split_sizes = var_1154_split_sizes_0, x = normed_123_cast_fp16)[name = string("op_1154_cast_fp16")];
-            tensor<fp16, [256]> var_1158_to_fp16 = const()[name = string("op_1158_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(591709184)))];
+            tensor<fp16, [256]> var_1158_to_fp16 = const()[name = string("op_1158_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296636352)))];
             tensor<fp16, [1, 1, 128, 256]> out_53_cast_fp16 = mul(x = var_1154_cast_fp16_0, y = var_1158_to_fp16)[name = string("out_53_cast_fp16")];
-            tensor<fp16, [1, 3, 128, 256]> var_1161 = mul(x = out_51_cast_fp16, y = cos_1)[name = string("op_1161")];
+            tensor<fp16, [1, 3, 128, 256]> var_1161 = mul(x = out_51_cast_fp16, y = cos_1_quantized)[name = string("op_1161")];
             tensor<int32, [2]> var_1162_split_sizes_0 = const()[name = string("op_1162_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_1162_axis_0 = const()[name = string("op_1162_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 128]> var_1162_0, tensor<fp16, [1, 3, 128, 128]> var_1162_1 = split(axis = var_1162_axis_0, split_sizes = var_1162_split_sizes_0, x = out_51_cast_fp16)[name = string("op_1162")];
@@ -1002,9 +1007,9 @@ program(1.3)
             tensor<fp16, [1, 3, 128, 128]> var_1164 = mul(x = var_1162_1, y = const_62_promoted)[name = string("op_1164")];
             bool var_1166_interleave_0 = const()[name = string("op_1166_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 3, 128, 256]> var_1166 = concat(axis = var_23, interleave = var_1166_interleave_0, values = (var_1164, var_1162_0))[name = string("op_1166")];
-            tensor<fp16, [1, 3, 128, 256]> var_1167 = mul(x = var_1166, y = sin_1)[name = string("op_1167")];
+            tensor<fp16, [1, 3, 128, 256]> var_1167 = mul(x = var_1166, y = sin_1_quantized)[name = string("op_1167")];
             tensor<fp16, [1, 3, 128, 256]> q_29 = add(x = var_1161, y = var_1167)[name = string("q_29")];
-            tensor<fp16, [1, 1, 128, 256]> var_1169 = mul(x = out_53_cast_fp16, y = cos_1)[name = string("op_1169")];
+            tensor<fp16, [1, 1, 128, 256]> var_1169 = mul(x = out_53_cast_fp16, y = cos_1_quantized)[name = string("op_1169")];
             tensor<int32, [2]> var_1170_split_sizes_0 = const()[name = string("op_1170_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_1170_axis_0 = const()[name = string("op_1170_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 128]> var_1170_0, tensor<fp16, [1, 1, 128, 128]> var_1170_1 = split(axis = var_1170_axis_0, split_sizes = var_1170_split_sizes_0, x = out_53_cast_fp16)[name = string("op_1170")];
@@ -1012,7 +1017,7 @@ program(1.3)
             tensor<fp16, [1, 1, 128, 128]> var_1172 = mul(x = var_1170_1, y = const_63_promoted)[name = string("op_1172")];
             bool var_1174_interleave_0 = const()[name = string("op_1174_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 1, 128, 256]> var_1174 = concat(axis = var_23, interleave = var_1174_interleave_0, values = (var_1172, var_1170_0))[name = string("op_1174")];
-            tensor<fp16, [1, 1, 128, 256]> var_1175 = mul(x = var_1174, y = sin_1)[name = string("op_1175")];
+            tensor<fp16, [1, 1, 128, 256]> var_1175 = mul(x = var_1174, y = sin_1_quantized)[name = string("op_1175")];
             tensor<fp16, [1, 1, 128, 256]> hidden_states_49 = add(x = var_1169, y = var_1175)[name = string("hidden_states_49")];
             tensor<int32, [1]> hidden_states_51_axes_0 = const()[name = string("hidden_states_51_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 1, 1, 128, 256]> hidden_states_51 = expand_dims(axes = hidden_states_51_axes_0, x = hidden_states_49)[name = string("hidden_states_51")];
@@ -1042,14 +1047,14 @@ program(1.3)
             tensor<fp16, [1, 128, 3, 256]> var_1202 = transpose(perm = var_1201, x = var_1199_cast_fp16)[name = string("transpose_175")];
             tensor<fp16, [1, 128, 768]> attn_out_27 = reshape(shape = var_1204, x = var_1202)[name = string("attn_out_27")];
             tensor<int32, [3]> var_1206 = const()[name = string("op_1206"), val = tensor<int32, [3]>([0, 2, 1])];
-            tensor<fp16, [768, 768, 1]> squeeze_4 = const()[name = string("squeeze_4"), val = tensor<fp16, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(591709760)))];
+            tensor<fp16, [768, 768, 1]> squeeze_4_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296636928))), scale = tensor<fp16, [768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297226816))))[name = string("squeeze_4_quantized")];
             string var_1215_pad_type_0 = const()[name = string("op_1215_pad_type_0"), val = string("valid")];
             int32 var_1215_groups_0 = const()[name = string("op_1215_groups_0"), val = int32(1)];
             tensor<int32, [1]> var_1215_strides_0 = const()[name = string("op_1215_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> var_1215_pad_0 = const()[name = string("op_1215_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> var_1215_dilations_0 = const()[name = string("op_1215_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, [1, 768, 128]> var_1207 = transpose(perm = var_1206, x = attn_out_27)[name = string("transpose_174")];
-            tensor<fp16, [1, 768, 128]> var_1215 = conv(dilations = var_1215_dilations_0, groups = var_1215_groups_0, pad = var_1215_pad_0, pad_type = var_1215_pad_type_0, strides = var_1215_strides_0, weight = squeeze_4, x = var_1207)[name = string("op_1215")];
+            tensor<fp16, [1, 768, 128]> var_1215 = conv(dilations = var_1215_dilations_0, groups = var_1215_groups_0, pad = var_1215_pad_0, pad_type = var_1215_pad_type_0, strides = var_1215_strides_0, weight = squeeze_4_quantized, x = var_1207)[name = string("op_1215")];
             tensor<int32, [3]> var_1216 = const()[name = string("op_1216"), val = tensor<int32, [3]>([0, 2, 1])];
             fp16 const_64_promoted_to_fp16 = const()[name = string("const_64_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 128, 768]> x_73 = transpose(perm = var_1216, x = var_1215)[name = string("transpose_173")];
@@ -1061,7 +1066,7 @@ program(1.3)
             tensor<int32, [2]> var_1225_split_sizes_0 = const()[name = string("op_1225_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_1225_axis_0 = const()[name = string("op_1225_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_1225_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_1225_cast_fp16_1 = split(axis = var_1225_axis_0, split_sizes = var_1225_split_sizes_0, x = normed_127_cast_fp16)[name = string("op_1225_cast_fp16")];
-            tensor<fp16, [768]> var_1229_to_fp16 = const()[name = string("op_1229_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(592889472)))];
+            tensor<fp16, [768]> var_1229_to_fp16 = const()[name = string("op_1229_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297228416)))];
             tensor<fp16, [1, 128, 768]> out_55_cast_fp16 = mul(x = var_1225_cast_fp16_0, y = var_1229_to_fp16)[name = string("out_55_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_75_cast_fp16 = add(x = x_65_cast_fp16, y = out_55_cast_fp16)[name = string("x_75_cast_fp16")];
             fp16 const_66_promoted_to_fp16 = const()[name = string("const_66_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -1073,7 +1078,7 @@ program(1.3)
             tensor<int32, [2]> var_1241_split_sizes_0 = const()[name = string("op_1241_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_1241_axis_0 = const()[name = string("op_1241_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_1241_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_1241_cast_fp16_1 = split(axis = var_1241_axis_0, split_sizes = var_1241_split_sizes_0, x = normed_131_cast_fp16)[name = string("op_1241_cast_fp16")];
-            tensor<fp16, [768]> var_1245_to_fp16 = const()[name = string("op_1245_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(592891072)))];
+            tensor<fp16, [768]> var_1245_to_fp16 = const()[name = string("op_1245_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297230016)))];
             tensor<fp16, [1, 128, 768]> out_57_cast_fp16 = mul(x = var_1241_cast_fp16_0, y = var_1245_to_fp16)[name = string("out_57_cast_fp16")];
             tensor<int32, [3]> var_1252 = const()[name = string("op_1252"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> input_95_axes_0 = const()[name = string("input_95_axes_0"), val = tensor<int32, [1]>([2])];
@@ -1084,13 +1089,13 @@ program(1.3)
             tensor<int32, [4]> gate_17_pad_0 = const()[name = string("gate_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> gate_17_dilations_0 = const()[name = string("gate_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 gate_17_groups_0 = const()[name = string("gate_17_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> gate_17 = conv(dilations = gate_17_dilations_0, groups = gate_17_groups_0, pad = gate_17_pad_0, pad_type = gate_17_pad_type_0, strides = gate_17_strides_0, weight = encoder_layers_4_mlp_gate_proj_weight, x = input_95)[name = string("gate_17")];
+            tensor<fp16, [1, 1152, 1, 128]> gate_17 = conv(dilations = gate_17_dilations_0, groups = gate_17_groups_0, pad = gate_17_pad_0, pad_type = gate_17_pad_type_0, strides = gate_17_strides_0, weight = encoder_layers_4_mlp_gate_proj_weight_quantized, x = input_95)[name = string("gate_17")];
             string up_9_pad_type_0 = const()[name = string("up_9_pad_type_0"), val = string("valid")];
             tensor<int32, [2]> up_9_strides_0 = const()[name = string("up_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> up_9_pad_0 = const()[name = string("up_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> up_9_dilations_0 = const()[name = string("up_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 up_9_groups_0 = const()[name = string("up_9_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> up_9 = conv(dilations = up_9_dilations_0, groups = up_9_groups_0, pad = up_9_pad_0, pad_type = up_9_pad_type_0, strides = up_9_strides_0, weight = encoder_layers_4_mlp_up_proj_weight, x = input_95)[name = string("up_9")];
+            tensor<fp16, [1, 1152, 1, 128]> up_9 = conv(dilations = up_9_dilations_0, groups = up_9_groups_0, pad = up_9_pad_0, pad_type = up_9_pad_type_0, strides = up_9_strides_0, weight = encoder_layers_4_mlp_up_proj_weight_quantized, x = input_95)[name = string("up_9")];
             string gate_19_mode_0 = const()[name = string("gate_19_mode_0"), val = string("TANH_APPROXIMATION")];
             tensor<fp16, [1, 1152, 1, 128]> gate_19 = gelu(mode = gate_19_mode_0, x = gate_17)[name = string("gate_19")];
             tensor<fp16, [1, 1152, 1, 128]> input_97 = mul(x = gate_19, y = up_9)[name = string("input_97")];
@@ -1099,7 +1104,7 @@ program(1.3)
             tensor<int32, [4]> var_1274_pad_0 = const()[name = string("op_1274_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_1274_dilations_0 = const()[name = string("op_1274_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_1274_groups_0 = const()[name = string("op_1274_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_1274 = conv(dilations = var_1274_dilations_0, groups = var_1274_groups_0, pad = var_1274_pad_0, pad_type = var_1274_pad_type_0, strides = var_1274_strides_0, weight = encoder_layers_4_mlp_down_proj_weight, x = input_97)[name = string("op_1274")];
+            tensor<fp16, [1, 768, 1, 128]> var_1274 = conv(dilations = var_1274_dilations_0, groups = var_1274_groups_0, pad = var_1274_pad_0, pad_type = var_1274_pad_type_0, strides = var_1274_strides_0, weight = encoder_layers_4_mlp_down_proj_weight_quantized, x = input_97)[name = string("op_1274")];
             tensor<int32, [1]> var_1275_axes_0 = const()[name = string("op_1275_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 768, 128]> var_1275 = squeeze(axes = var_1275_axes_0, x = var_1274)[name = string("op_1275")];
             tensor<int32, [3]> var_1276 = const()[name = string("op_1276"), val = tensor<int32, [3]>([0, 2, 1])];
@@ -1113,7 +1118,7 @@ program(1.3)
             tensor<int32, [2]> var_1285_split_sizes_0 = const()[name = string("op_1285_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_1285_axis_0 = const()[name = string("op_1285_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_1285_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_1285_cast_fp16_1 = split(axis = var_1285_axis_0, split_sizes = var_1285_split_sizes_0, x = normed_137_cast_fp16)[name = string("op_1285_cast_fp16")];
-            tensor<fp16, [768]> var_1289_to_fp16 = const()[name = string("op_1289_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(592892672)))];
+            tensor<fp16, [768]> var_1289_to_fp16 = const()[name = string("op_1289_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297231616)))];
             tensor<fp16, [1, 128, 768]> out_59_cast_fp16 = mul(x = var_1285_cast_fp16_0, y = var_1289_to_fp16)[name = string("out_59_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_81_cast_fp16 = add(x = x_75_cast_fp16, y = out_59_cast_fp16)[name = string("x_81_cast_fp16")];
             fp16 const_70_promoted_to_fp16 = const()[name = string("const_70_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -1125,7 +1130,7 @@ program(1.3)
             tensor<int32, [2]> var_1323_split_sizes_0 = const()[name = string("op_1323_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_1323_axis_0 = const()[name = string("op_1323_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_1323_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_1323_cast_fp16_1 = split(axis = var_1323_axis_0, split_sizes = var_1323_split_sizes_0, x = normed_141_cast_fp16)[name = string("op_1323_cast_fp16")];
-            tensor<fp16, [768]> var_1327_to_fp16 = const()[name = string("op_1327_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(592894272)))];
+            tensor<fp16, [768]> var_1327_to_fp16 = const()[name = string("op_1327_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297233216)))];
             tensor<fp16, [1, 128, 768]> out_61_cast_fp16 = mul(x = var_1323_cast_fp16_0, y = var_1327_to_fp16)[name = string("out_61_cast_fp16")];
             tensor<int32, [3]> var_1333 = const()[name = string("op_1333"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> var_1335_axes_0 = const()[name = string("op_1335_axes_0"), val = tensor<int32, [1]>([2])];
@@ -1136,7 +1141,7 @@ program(1.3)
             tensor<int32, [4]> var_1342_pad_0 = const()[name = string("op_1342_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_1342_dilations_0 = const()[name = string("op_1342_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_1342_groups_0 = const()[name = string("op_1342_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_1342 = conv(dilations = var_1342_dilations_0, groups = var_1342_groups_0, pad = var_1342_pad_0, pad_type = var_1342_pad_type_0, strides = var_1342_strides_0, weight = encoder_layers_5_self_attn_q_proj_weight, x = var_1335_cast_fp16)[name = string("op_1342")];
+            tensor<fp16, [1, 768, 1, 128]> var_1342 = conv(dilations = var_1342_dilations_0, groups = var_1342_groups_0, pad = var_1342_pad_0, pad_type = var_1342_pad_type_0, strides = var_1342_strides_0, weight = encoder_layers_5_self_attn_q_proj_weight_quantized, x = var_1335_cast_fp16)[name = string("op_1342")];
             tensor<int32, [4]> var_1343 = const()[name = string("op_1343"), val = tensor<int32, [4]>([1, 3, 256, 128])];
             tensor<fp16, [1, 3, 256, 128]> var_1344 = reshape(shape = var_1343, x = var_1342)[name = string("op_1344")];
             tensor<int32, [4]> var_1345 = const()[name = string("op_1345"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -1145,7 +1150,7 @@ program(1.3)
             tensor<int32, [4]> var_1352_pad_0 = const()[name = string("op_1352_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_1352_dilations_0 = const()[name = string("op_1352_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_1352_groups_0 = const()[name = string("op_1352_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_1352 = conv(dilations = var_1352_dilations_0, groups = var_1352_groups_0, pad = var_1352_pad_0, pad_type = var_1352_pad_type_0, strides = var_1352_strides_0, weight = encoder_layers_5_self_attn_k_proj_weight, x = var_1335_cast_fp16)[name = string("op_1352")];
+            tensor<fp16, [1, 256, 1, 128]> var_1352 = conv(dilations = var_1352_dilations_0, groups = var_1352_groups_0, pad = var_1352_pad_0, pad_type = var_1352_pad_type_0, strides = var_1352_strides_0, weight = encoder_layers_5_self_attn_k_proj_weight_quantized, x = var_1335_cast_fp16)[name = string("op_1352")];
             tensor<int32, [4]> var_1353 = const()[name = string("op_1353"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_1354 = reshape(shape = var_1353, x = var_1352)[name = string("op_1354")];
             tensor<int32, [4]> var_1355 = const()[name = string("op_1355"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -1154,7 +1159,7 @@ program(1.3)
             tensor<int32, [4]> var_1362_pad_0 = const()[name = string("op_1362_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_1362_dilations_0 = const()[name = string("op_1362_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_1362_groups_0 = const()[name = string("op_1362_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_1362 = conv(dilations = var_1362_dilations_0, groups = var_1362_groups_0, pad = var_1362_pad_0, pad_type = var_1362_pad_type_0, strides = var_1362_strides_0, weight = encoder_layers_5_self_attn_v_proj_weight, x = var_1335_cast_fp16)[name = string("op_1362")];
+            tensor<fp16, [1, 256, 1, 128]> var_1362 = conv(dilations = var_1362_dilations_0, groups = var_1362_groups_0, pad = var_1362_pad_0, pad_type = var_1362_pad_type_0, strides = var_1362_strides_0, weight = encoder_layers_5_self_attn_v_proj_weight_quantized, x = var_1335_cast_fp16)[name = string("op_1362")];
             tensor<int32, [4]> var_1363 = const()[name = string("op_1363"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_1364 = reshape(shape = var_1363, x = var_1362)[name = string("op_1364")];
             tensor<int32, [4]> var_1365 = const()[name = string("op_1365"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -1168,7 +1173,7 @@ program(1.3)
             tensor<int32, [2]> var_1376_split_sizes_0 = const()[name = string("op_1376_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_1376_axis_0 = const()[name = string("op_1376_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 256]> var_1376_cast_fp16_0, tensor<fp16, [1, 3, 128, 256]> var_1376_cast_fp16_1 = split(axis = var_1376_axis_0, split_sizes = var_1376_split_sizes_0, x = normed_147_cast_fp16)[name = string("op_1376_cast_fp16")];
-            tensor<fp16, [256]> var_1380_to_fp16 = const()[name = string("op_1380_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(592895872)))];
+            tensor<fp16, [256]> var_1380_to_fp16 = const()[name = string("op_1380_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297234816)))];
             tensor<fp16, [1, 3, 128, 256]> out_63_cast_fp16 = mul(x = var_1376_cast_fp16_0, y = var_1380_to_fp16)[name = string("out_63_cast_fp16")];
             fp16 const_74_promoted_to_fp16 = const()[name = string("const_74_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 1, 128, 256]> k_31 = transpose(perm = var_1355, x = var_1354)[name = string("transpose_168")];
@@ -1180,9 +1185,9 @@ program(1.3)
             tensor<int32, [2]> var_1392_split_sizes_0 = const()[name = string("op_1392_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_1392_axis_0 = const()[name = string("op_1392_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 256]> var_1392_cast_fp16_0, tensor<fp16, [1, 1, 128, 256]> var_1392_cast_fp16_1 = split(axis = var_1392_axis_0, split_sizes = var_1392_split_sizes_0, x = normed_151_cast_fp16)[name = string("op_1392_cast_fp16")];
-            tensor<fp16, [256]> var_1396_to_fp16 = const()[name = string("op_1396_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(592896448)))];
+            tensor<fp16, [256]> var_1396_to_fp16 = const()[name = string("op_1396_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297235392)))];
             tensor<fp16, [1, 1, 128, 256]> out_65_cast_fp16 = mul(x = var_1392_cast_fp16_0, y = var_1396_to_fp16)[name = string("out_65_cast_fp16")];
-            tensor<fp16, [1, 3, 128, 256]> var_1399 = mul(x = out_63_cast_fp16, y = cos)[name = string("op_1399")];
+            tensor<fp16, [1, 3, 128, 256]> var_1399 = mul(x = out_63_cast_fp16, y = cos_quantized)[name = string("op_1399")];
             tensor<int32, [2]> var_1400_split_sizes_0 = const()[name = string("op_1400_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_1400_axis_0 = const()[name = string("op_1400_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 128]> var_1400_0, tensor<fp16, [1, 3, 128, 128]> var_1400_1 = split(axis = var_1400_axis_0, split_sizes = var_1400_split_sizes_0, x = out_63_cast_fp16)[name = string("op_1400")];
@@ -1190,9 +1195,9 @@ program(1.3)
             tensor<fp16, [1, 3, 128, 128]> var_1402 = mul(x = var_1400_1, y = const_76_promoted)[name = string("op_1402")];
             bool var_1404_interleave_0 = const()[name = string("op_1404_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 3, 128, 256]> var_1404 = concat(axis = var_23, interleave = var_1404_interleave_0, values = (var_1402, var_1400_0))[name = string("op_1404")];
-            tensor<fp16, [1, 3, 128, 256]> var_1405 = mul(x = var_1404, y = sin)[name = string("op_1405")];
+            tensor<fp16, [1, 3, 128, 256]> var_1405 = mul(x = var_1404, y = sin_quantized)[name = string("op_1405")];
             tensor<fp16, [1, 3, 128, 256]> q_35 = add(x = var_1399, y = var_1405)[name = string("q_35")];
-            tensor<fp16, [1, 1, 128, 256]> var_1407 = mul(x = out_65_cast_fp16, y = cos)[name = string("op_1407")];
+            tensor<fp16, [1, 1, 128, 256]> var_1407 = mul(x = out_65_cast_fp16, y = cos_quantized)[name = string("op_1407")];
             tensor<int32, [2]> var_1408_split_sizes_0 = const()[name = string("op_1408_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_1408_axis_0 = const()[name = string("op_1408_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 128]> var_1408_0, tensor<fp16, [1, 1, 128, 128]> var_1408_1 = split(axis = var_1408_axis_0, split_sizes = var_1408_split_sizes_0, x = out_65_cast_fp16)[name = string("op_1408")];
@@ -1200,7 +1205,7 @@ program(1.3)
             tensor<fp16, [1, 1, 128, 128]> var_1410 = mul(x = var_1408_1, y = const_77_promoted)[name = string("op_1410")];
             bool var_1412_interleave_0 = const()[name = string("op_1412_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 1, 128, 256]> var_1412 = concat(axis = var_23, interleave = var_1412_interleave_0, values = (var_1410, var_1408_0))[name = string("op_1412")];
-            tensor<fp16, [1, 1, 128, 256]> var_1413 = mul(x = var_1412, y = sin)[name = string("op_1413")];
+            tensor<fp16, [1, 1, 128, 256]> var_1413 = mul(x = var_1412, y = sin_quantized)[name = string("op_1413")];
             tensor<fp16, [1, 1, 128, 256]> hidden_states_61 = add(x = var_1407, y = var_1413)[name = string("hidden_states_61")];
             tensor<int32, [1]> hidden_states_63_axes_0 = const()[name = string("hidden_states_63_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 1, 1, 128, 256]> hidden_states_63 = expand_dims(axes = hidden_states_63_axes_0, x = hidden_states_61)[name = string("hidden_states_63")];
@@ -1230,14 +1235,14 @@ program(1.3)
             tensor<fp16, [1, 128, 3, 256]> var_1440 = transpose(perm = var_1439, x = var_1437_cast_fp16)[name = string("transpose_166")];
             tensor<fp16, [1, 128, 768]> attn_out_33 = reshape(shape = var_1442, x = var_1440)[name = string("attn_out_33")];
             tensor<int32, [3]> var_1444 = const()[name = string("op_1444"), val = tensor<int32, [3]>([0, 2, 1])];
-            tensor<fp16, [768, 768, 1]> squeeze_5 = const()[name = string("squeeze_5"), val = tensor<fp16, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(592897024)))];
+            tensor<fp16, [768, 768, 1]> squeeze_5_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297235968))), scale = tensor<fp16, [768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297825856))))[name = string("squeeze_5_quantized")];
             string var_1453_pad_type_0 = const()[name = string("op_1453_pad_type_0"), val = string("valid")];
             int32 var_1453_groups_0 = const()[name = string("op_1453_groups_0"), val = int32(1)];
             tensor<int32, [1]> var_1453_strides_0 = const()[name = string("op_1453_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> var_1453_pad_0 = const()[name = string("op_1453_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> var_1453_dilations_0 = const()[name = string("op_1453_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, [1, 768, 128]> var_1445 = transpose(perm = var_1444, x = attn_out_33)[name = string("transpose_165")];
-            tensor<fp16, [1, 768, 128]> var_1453 = conv(dilations = var_1453_dilations_0, groups = var_1453_groups_0, pad = var_1453_pad_0, pad_type = var_1453_pad_type_0, strides = var_1453_strides_0, weight = squeeze_5, x = var_1445)[name = string("op_1453")];
+            tensor<fp16, [1, 768, 128]> var_1453 = conv(dilations = var_1453_dilations_0, groups = var_1453_groups_0, pad = var_1453_pad_0, pad_type = var_1453_pad_type_0, strides = var_1453_strides_0, weight = squeeze_5_quantized, x = var_1445)[name = string("op_1453")];
             tensor<int32, [3]> var_1454 = const()[name = string("op_1454"), val = tensor<int32, [3]>([0, 2, 1])];
             fp16 const_78_promoted_to_fp16 = const()[name = string("const_78_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 128, 768]> x_89 = transpose(perm = var_1454, x = var_1453)[name = string("transpose_164")];
@@ -1249,7 +1254,7 @@ program(1.3)
             tensor<int32, [2]> var_1463_split_sizes_0 = const()[name = string("op_1463_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_1463_axis_0 = const()[name = string("op_1463_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_1463_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_1463_cast_fp16_1 = split(axis = var_1463_axis_0, split_sizes = var_1463_split_sizes_0, x = normed_155_cast_fp16)[name = string("op_1463_cast_fp16")];
-            tensor<fp16, [768]> var_1467_to_fp16 = const()[name = string("op_1467_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(594076736)))];
+            tensor<fp16, [768]> var_1467_to_fp16 = const()[name = string("op_1467_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297827456)))];
             tensor<fp16, [1, 128, 768]> out_67_cast_fp16 = mul(x = var_1463_cast_fp16_0, y = var_1467_to_fp16)[name = string("out_67_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_91_cast_fp16 = add(x = x_81_cast_fp16, y = out_67_cast_fp16)[name = string("x_91_cast_fp16")];
             fp16 const_80_promoted_to_fp16 = const()[name = string("const_80_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -1261,7 +1266,7 @@ program(1.3)
             tensor<int32, [2]> var_1479_split_sizes_0 = const()[name = string("op_1479_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_1479_axis_0 = const()[name = string("op_1479_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_1479_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_1479_cast_fp16_1 = split(axis = var_1479_axis_0, split_sizes = var_1479_split_sizes_0, x = normed_159_cast_fp16)[name = string("op_1479_cast_fp16")];
-            tensor<fp16, [768]> var_1483_to_fp16 = const()[name = string("op_1483_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(594078336)))];
+            tensor<fp16, [768]> var_1483_to_fp16 = const()[name = string("op_1483_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297829056)))];
             tensor<fp16, [1, 128, 768]> out_69_cast_fp16 = mul(x = var_1479_cast_fp16_0, y = var_1483_to_fp16)[name = string("out_69_cast_fp16")];
             tensor<int32, [3]> var_1490 = const()[name = string("op_1490"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> input_115_axes_0 = const()[name = string("input_115_axes_0"), val = tensor<int32, [1]>([2])];
@@ -1272,13 +1277,13 @@ program(1.3)
             tensor<int32, [4]> gate_21_pad_0 = const()[name = string("gate_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> gate_21_dilations_0 = const()[name = string("gate_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 gate_21_groups_0 = const()[name = string("gate_21_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> gate_21 = conv(dilations = gate_21_dilations_0, groups = gate_21_groups_0, pad = gate_21_pad_0, pad_type = gate_21_pad_type_0, strides = gate_21_strides_0, weight = encoder_layers_5_mlp_gate_proj_weight, x = input_115)[name = string("gate_21")];
+            tensor<fp16, [1, 1152, 1, 128]> gate_21 = conv(dilations = gate_21_dilations_0, groups = gate_21_groups_0, pad = gate_21_pad_0, pad_type = gate_21_pad_type_0, strides = gate_21_strides_0, weight = encoder_layers_5_mlp_gate_proj_weight_quantized, x = input_115)[name = string("gate_21")];
             string up_11_pad_type_0 = const()[name = string("up_11_pad_type_0"), val = string("valid")];
             tensor<int32, [2]> up_11_strides_0 = const()[name = string("up_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> up_11_pad_0 = const()[name = string("up_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> up_11_dilations_0 = const()[name = string("up_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 up_11_groups_0 = const()[name = string("up_11_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> up_11 = conv(dilations = up_11_dilations_0, groups = up_11_groups_0, pad = up_11_pad_0, pad_type = up_11_pad_type_0, strides = up_11_strides_0, weight = encoder_layers_5_mlp_up_proj_weight, x = input_115)[name = string("up_11")];
+            tensor<fp16, [1, 1152, 1, 128]> up_11 = conv(dilations = up_11_dilations_0, groups = up_11_groups_0, pad = up_11_pad_0, pad_type = up_11_pad_type_0, strides = up_11_strides_0, weight = encoder_layers_5_mlp_up_proj_weight_quantized, x = input_115)[name = string("up_11")];
             string gate_23_mode_0 = const()[name = string("gate_23_mode_0"), val = string("TANH_APPROXIMATION")];
             tensor<fp16, [1, 1152, 1, 128]> gate_23 = gelu(mode = gate_23_mode_0, x = gate_21)[name = string("gate_23")];
             tensor<fp16, [1, 1152, 1, 128]> input_117 = mul(x = gate_23, y = up_11)[name = string("input_117")];
@@ -1287,7 +1292,7 @@ program(1.3)
             tensor<int32, [4]> var_1512_pad_0 = const()[name = string("op_1512_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_1512_dilations_0 = const()[name = string("op_1512_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_1512_groups_0 = const()[name = string("op_1512_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_1512 = conv(dilations = var_1512_dilations_0, groups = var_1512_groups_0, pad = var_1512_pad_0, pad_type = var_1512_pad_type_0, strides = var_1512_strides_0, weight = encoder_layers_5_mlp_down_proj_weight, x = input_117)[name = string("op_1512")];
+            tensor<fp16, [1, 768, 1, 128]> var_1512 = conv(dilations = var_1512_dilations_0, groups = var_1512_groups_0, pad = var_1512_pad_0, pad_type = var_1512_pad_type_0, strides = var_1512_strides_0, weight = encoder_layers_5_mlp_down_proj_weight_quantized, x = input_117)[name = string("op_1512")];
             tensor<int32, [1]> var_1513_axes_0 = const()[name = string("op_1513_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 768, 128]> var_1513 = squeeze(axes = var_1513_axes_0, x = var_1512)[name = string("op_1513")];
             tensor<int32, [3]> var_1514 = const()[name = string("op_1514"), val = tensor<int32, [3]>([0, 2, 1])];
@@ -1301,7 +1306,7 @@ program(1.3)
             tensor<int32, [2]> var_1523_split_sizes_0 = const()[name = string("op_1523_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_1523_axis_0 = const()[name = string("op_1523_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_1523_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_1523_cast_fp16_1 = split(axis = var_1523_axis_0, split_sizes = var_1523_split_sizes_0, x = normed_165_cast_fp16)[name = string("op_1523_cast_fp16")];
-            tensor<fp16, [768]> var_1527_to_fp16 = const()[name = string("op_1527_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(594079936)))];
+            tensor<fp16, [768]> var_1527_to_fp16 = const()[name = string("op_1527_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297830656)))];
             tensor<fp16, [1, 128, 768]> out_71_cast_fp16 = mul(x = var_1523_cast_fp16_0, y = var_1527_to_fp16)[name = string("out_71_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_97_cast_fp16 = add(x = x_91_cast_fp16, y = out_71_cast_fp16)[name = string("x_97_cast_fp16")];
             fp16 const_84_promoted_to_fp16 = const()[name = string("const_84_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -1313,7 +1318,7 @@ program(1.3)
             tensor<int32, [2]> var_1561_split_sizes_0 = const()[name = string("op_1561_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_1561_axis_0 = const()[name = string("op_1561_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_1561_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_1561_cast_fp16_1 = split(axis = var_1561_axis_0, split_sizes = var_1561_split_sizes_0, x = normed_169_cast_fp16)[name = string("op_1561_cast_fp16")];
-            tensor<fp16, [768]> var_1565_to_fp16 = const()[name = string("op_1565_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(594081536)))];
+            tensor<fp16, [768]> var_1565_to_fp16 = const()[name = string("op_1565_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297832256)))];
             tensor<fp16, [1, 128, 768]> out_73_cast_fp16 = mul(x = var_1561_cast_fp16_0, y = var_1565_to_fp16)[name = string("out_73_cast_fp16")];
             tensor<int32, [3]> var_1571 = const()[name = string("op_1571"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> var_1573_axes_0 = const()[name = string("op_1573_axes_0"), val = tensor<int32, [1]>([2])];
@@ -1324,7 +1329,7 @@ program(1.3)
             tensor<int32, [4]> var_1580_pad_0 = const()[name = string("op_1580_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_1580_dilations_0 = const()[name = string("op_1580_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_1580_groups_0 = const()[name = string("op_1580_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_1580 = conv(dilations = var_1580_dilations_0, groups = var_1580_groups_0, pad = var_1580_pad_0, pad_type = var_1580_pad_type_0, strides = var_1580_strides_0, weight = encoder_layers_6_self_attn_q_proj_weight, x = var_1573_cast_fp16)[name = string("op_1580")];
+            tensor<fp16, [1, 768, 1, 128]> var_1580 = conv(dilations = var_1580_dilations_0, groups = var_1580_groups_0, pad = var_1580_pad_0, pad_type = var_1580_pad_type_0, strides = var_1580_strides_0, weight = encoder_layers_6_self_attn_q_proj_weight_quantized, x = var_1573_cast_fp16)[name = string("op_1580")];
             tensor<int32, [4]> var_1581 = const()[name = string("op_1581"), val = tensor<int32, [4]>([1, 3, 256, 128])];
             tensor<fp16, [1, 3, 256, 128]> var_1582 = reshape(shape = var_1581, x = var_1580)[name = string("op_1582")];
             tensor<int32, [4]> var_1583 = const()[name = string("op_1583"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -1333,7 +1338,7 @@ program(1.3)
             tensor<int32, [4]> var_1590_pad_0 = const()[name = string("op_1590_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_1590_dilations_0 = const()[name = string("op_1590_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_1590_groups_0 = const()[name = string("op_1590_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_1590 = conv(dilations = var_1590_dilations_0, groups = var_1590_groups_0, pad = var_1590_pad_0, pad_type = var_1590_pad_type_0, strides = var_1590_strides_0, weight = encoder_layers_6_self_attn_k_proj_weight, x = var_1573_cast_fp16)[name = string("op_1590")];
+            tensor<fp16, [1, 256, 1, 128]> var_1590 = conv(dilations = var_1590_dilations_0, groups = var_1590_groups_0, pad = var_1590_pad_0, pad_type = var_1590_pad_type_0, strides = var_1590_strides_0, weight = encoder_layers_6_self_attn_k_proj_weight_quantized, x = var_1573_cast_fp16)[name = string("op_1590")];
             tensor<int32, [4]> var_1591 = const()[name = string("op_1591"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_1592 = reshape(shape = var_1591, x = var_1590)[name = string("op_1592")];
             tensor<int32, [4]> var_1593 = const()[name = string("op_1593"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -1342,7 +1347,7 @@ program(1.3)
             tensor<int32, [4]> var_1600_pad_0 = const()[name = string("op_1600_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_1600_dilations_0 = const()[name = string("op_1600_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_1600_groups_0 = const()[name = string("op_1600_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_1600 = conv(dilations = var_1600_dilations_0, groups = var_1600_groups_0, pad = var_1600_pad_0, pad_type = var_1600_pad_type_0, strides = var_1600_strides_0, weight = encoder_layers_6_self_attn_v_proj_weight, x = var_1573_cast_fp16)[name = string("op_1600")];
+            tensor<fp16, [1, 256, 1, 128]> var_1600 = conv(dilations = var_1600_dilations_0, groups = var_1600_groups_0, pad = var_1600_pad_0, pad_type = var_1600_pad_type_0, strides = var_1600_strides_0, weight = encoder_layers_6_self_attn_v_proj_weight_quantized, x = var_1573_cast_fp16)[name = string("op_1600")];
             tensor<int32, [4]> var_1601 = const()[name = string("op_1601"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_1602 = reshape(shape = var_1601, x = var_1600)[name = string("op_1602")];
             tensor<int32, [4]> var_1603 = const()[name = string("op_1603"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -1356,7 +1361,7 @@ program(1.3)
             tensor<int32, [2]> var_1614_split_sizes_0 = const()[name = string("op_1614_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_1614_axis_0 = const()[name = string("op_1614_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 256]> var_1614_cast_fp16_0, tensor<fp16, [1, 3, 128, 256]> var_1614_cast_fp16_1 = split(axis = var_1614_axis_0, split_sizes = var_1614_split_sizes_0, x = normed_175_cast_fp16)[name = string("op_1614_cast_fp16")];
-            tensor<fp16, [256]> var_1618_to_fp16 = const()[name = string("op_1618_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(594083136)))];
+            tensor<fp16, [256]> var_1618_to_fp16 = const()[name = string("op_1618_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297833856)))];
             tensor<fp16, [1, 3, 128, 256]> out_75_cast_fp16 = mul(x = var_1614_cast_fp16_0, y = var_1618_to_fp16)[name = string("out_75_cast_fp16")];
             fp16 const_88_promoted_to_fp16 = const()[name = string("const_88_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 1, 128, 256]> k_37 = transpose(perm = var_1593, x = var_1592)[name = string("transpose_159")];
@@ -1368,9 +1373,9 @@ program(1.3)
             tensor<int32, [2]> var_1630_split_sizes_0 = const()[name = string("op_1630_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_1630_axis_0 = const()[name = string("op_1630_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 256]> var_1630_cast_fp16_0, tensor<fp16, [1, 1, 128, 256]> var_1630_cast_fp16_1 = split(axis = var_1630_axis_0, split_sizes = var_1630_split_sizes_0, x = normed_179_cast_fp16)[name = string("op_1630_cast_fp16")];
-            tensor<fp16, [256]> var_1634_to_fp16 = const()[name = string("op_1634_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(594083712)))];
+            tensor<fp16, [256]> var_1634_to_fp16 = const()[name = string("op_1634_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297834432)))];
             tensor<fp16, [1, 1, 128, 256]> out_77_cast_fp16 = mul(x = var_1630_cast_fp16_0, y = var_1634_to_fp16)[name = string("out_77_cast_fp16")];
-            tensor<fp16, [1, 3, 128, 256]> var_1637 = mul(x = out_75_cast_fp16, y = cos_1)[name = string("op_1637")];
+            tensor<fp16, [1, 3, 128, 256]> var_1637 = mul(x = out_75_cast_fp16, y = cos_1_quantized)[name = string("op_1637")];
             tensor<int32, [2]> var_1638_split_sizes_0 = const()[name = string("op_1638_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_1638_axis_0 = const()[name = string("op_1638_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 128]> var_1638_0, tensor<fp16, [1, 3, 128, 128]> var_1638_1 = split(axis = var_1638_axis_0, split_sizes = var_1638_split_sizes_0, x = out_75_cast_fp16)[name = string("op_1638")];
@@ -1378,9 +1383,9 @@ program(1.3)
             tensor<fp16, [1, 3, 128, 128]> var_1640 = mul(x = var_1638_1, y = const_90_promoted)[name = string("op_1640")];
             bool var_1642_interleave_0 = const()[name = string("op_1642_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 3, 128, 256]> var_1642 = concat(axis = var_23, interleave = var_1642_interleave_0, values = (var_1640, var_1638_0))[name = string("op_1642")];
-            tensor<fp16, [1, 3, 128, 256]> var_1643 = mul(x = var_1642, y = sin_1)[name = string("op_1643")];
+            tensor<fp16, [1, 3, 128, 256]> var_1643 = mul(x = var_1642, y = sin_1_quantized)[name = string("op_1643")];
             tensor<fp16, [1, 3, 128, 256]> q_41 = add(x = var_1637, y = var_1643)[name = string("q_41")];
-            tensor<fp16, [1, 1, 128, 256]> var_1645 = mul(x = out_77_cast_fp16, y = cos_1)[name = string("op_1645")];
+            tensor<fp16, [1, 1, 128, 256]> var_1645 = mul(x = out_77_cast_fp16, y = cos_1_quantized)[name = string("op_1645")];
             tensor<int32, [2]> var_1646_split_sizes_0 = const()[name = string("op_1646_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_1646_axis_0 = const()[name = string("op_1646_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 128]> var_1646_0, tensor<fp16, [1, 1, 128, 128]> var_1646_1 = split(axis = var_1646_axis_0, split_sizes = var_1646_split_sizes_0, x = out_77_cast_fp16)[name = string("op_1646")];
@@ -1388,7 +1393,7 @@ program(1.3)
             tensor<fp16, [1, 1, 128, 128]> var_1648 = mul(x = var_1646_1, y = const_91_promoted)[name = string("op_1648")];
             bool var_1650_interleave_0 = const()[name = string("op_1650_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 1, 128, 256]> var_1650 = concat(axis = var_23, interleave = var_1650_interleave_0, values = (var_1648, var_1646_0))[name = string("op_1650")];
-            tensor<fp16, [1, 1, 128, 256]> var_1651 = mul(x = var_1650, y = sin_1)[name = string("op_1651")];
+            tensor<fp16, [1, 1, 128, 256]> var_1651 = mul(x = var_1650, y = sin_1_quantized)[name = string("op_1651")];
             tensor<fp16, [1, 1, 128, 256]> hidden_states_73 = add(x = var_1645, y = var_1651)[name = string("hidden_states_73")];
             tensor<int32, [1]> hidden_states_75_axes_0 = const()[name = string("hidden_states_75_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 1, 1, 128, 256]> hidden_states_75 = expand_dims(axes = hidden_states_75_axes_0, x = hidden_states_73)[name = string("hidden_states_75")];
@@ -1418,14 +1423,14 @@ program(1.3)
             tensor<fp16, [1, 128, 3, 256]> var_1678 = transpose(perm = var_1677, x = var_1675_cast_fp16)[name = string("transpose_157")];
             tensor<fp16, [1, 128, 768]> attn_out_39 = reshape(shape = var_1680, x = var_1678)[name = string("attn_out_39")];
             tensor<int32, [3]> var_1682 = const()[name = string("op_1682"), val = tensor<int32, [3]>([0, 2, 1])];
-            tensor<fp16, [768, 768, 1]> squeeze_6 = const()[name = string("squeeze_6"), val = tensor<fp16, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(594084288)))];
+            tensor<fp16, [768, 768, 1]> squeeze_6_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297835008))), scale = tensor<fp16, [768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298424896))))[name = string("squeeze_6_quantized")];
             string var_1691_pad_type_0 = const()[name = string("op_1691_pad_type_0"), val = string("valid")];
             int32 var_1691_groups_0 = const()[name = string("op_1691_groups_0"), val = int32(1)];
             tensor<int32, [1]> var_1691_strides_0 = const()[name = string("op_1691_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> var_1691_pad_0 = const()[name = string("op_1691_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> var_1691_dilations_0 = const()[name = string("op_1691_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, [1, 768, 128]> var_1683 = transpose(perm = var_1682, x = attn_out_39)[name = string("transpose_156")];
-            tensor<fp16, [1, 768, 128]> var_1691 = conv(dilations = var_1691_dilations_0, groups = var_1691_groups_0, pad = var_1691_pad_0, pad_type = var_1691_pad_type_0, strides = var_1691_strides_0, weight = squeeze_6, x = var_1683)[name = string("op_1691")];
+            tensor<fp16, [1, 768, 128]> var_1691 = conv(dilations = var_1691_dilations_0, groups = var_1691_groups_0, pad = var_1691_pad_0, pad_type = var_1691_pad_type_0, strides = var_1691_strides_0, weight = squeeze_6_quantized, x = var_1683)[name = string("op_1691")];
             tensor<int32, [3]> var_1692 = const()[name = string("op_1692"), val = tensor<int32, [3]>([0, 2, 1])];
             fp16 const_92_promoted_to_fp16 = const()[name = string("const_92_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 128, 768]> x_105 = transpose(perm = var_1692, x = var_1691)[name = string("transpose_155")];
@@ -1437,7 +1442,7 @@ program(1.3)
             tensor<int32, [2]> var_1701_split_sizes_0 = const()[name = string("op_1701_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_1701_axis_0 = const()[name = string("op_1701_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_1701_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_1701_cast_fp16_1 = split(axis = var_1701_axis_0, split_sizes = var_1701_split_sizes_0, x = normed_183_cast_fp16)[name = string("op_1701_cast_fp16")];
-            tensor<fp16, [768]> var_1705_to_fp16 = const()[name = string("op_1705_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(595264000)))];
+            tensor<fp16, [768]> var_1705_to_fp16 = const()[name = string("op_1705_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298426496)))];
             tensor<fp16, [1, 128, 768]> out_79_cast_fp16 = mul(x = var_1701_cast_fp16_0, y = var_1705_to_fp16)[name = string("out_79_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_107_cast_fp16 = add(x = x_97_cast_fp16, y = out_79_cast_fp16)[name = string("x_107_cast_fp16")];
             fp16 const_94_promoted_to_fp16 = const()[name = string("const_94_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -1449,7 +1454,7 @@ program(1.3)
             tensor<int32, [2]> var_1717_split_sizes_0 = const()[name = string("op_1717_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_1717_axis_0 = const()[name = string("op_1717_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_1717_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_1717_cast_fp16_1 = split(axis = var_1717_axis_0, split_sizes = var_1717_split_sizes_0, x = normed_187_cast_fp16)[name = string("op_1717_cast_fp16")];
-            tensor<fp16, [768]> var_1721_to_fp16 = const()[name = string("op_1721_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(595265600)))];
+            tensor<fp16, [768]> var_1721_to_fp16 = const()[name = string("op_1721_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298428096)))];
             tensor<fp16, [1, 128, 768]> out_81_cast_fp16 = mul(x = var_1717_cast_fp16_0, y = var_1721_to_fp16)[name = string("out_81_cast_fp16")];
             tensor<int32, [3]> var_1728 = const()[name = string("op_1728"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> input_135_axes_0 = const()[name = string("input_135_axes_0"), val = tensor<int32, [1]>([2])];
@@ -1460,13 +1465,13 @@ program(1.3)
             tensor<int32, [4]> gate_25_pad_0 = const()[name = string("gate_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> gate_25_dilations_0 = const()[name = string("gate_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 gate_25_groups_0 = const()[name = string("gate_25_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> gate_25 = conv(dilations = gate_25_dilations_0, groups = gate_25_groups_0, pad = gate_25_pad_0, pad_type = gate_25_pad_type_0, strides = gate_25_strides_0, weight = encoder_layers_6_mlp_gate_proj_weight, x = input_135)[name = string("gate_25")];
+            tensor<fp16, [1, 1152, 1, 128]> gate_25 = conv(dilations = gate_25_dilations_0, groups = gate_25_groups_0, pad = gate_25_pad_0, pad_type = gate_25_pad_type_0, strides = gate_25_strides_0, weight = encoder_layers_6_mlp_gate_proj_weight_quantized, x = input_135)[name = string("gate_25")];
             string up_13_pad_type_0 = const()[name = string("up_13_pad_type_0"), val = string("valid")];
             tensor<int32, [2]> up_13_strides_0 = const()[name = string("up_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> up_13_pad_0 = const()[name = string("up_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> up_13_dilations_0 = const()[name = string("up_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 up_13_groups_0 = const()[name = string("up_13_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> up_13 = conv(dilations = up_13_dilations_0, groups = up_13_groups_0, pad = up_13_pad_0, pad_type = up_13_pad_type_0, strides = up_13_strides_0, weight = encoder_layers_6_mlp_up_proj_weight, x = input_135)[name = string("up_13")];
+            tensor<fp16, [1, 1152, 1, 128]> up_13 = conv(dilations = up_13_dilations_0, groups = up_13_groups_0, pad = up_13_pad_0, pad_type = up_13_pad_type_0, strides = up_13_strides_0, weight = encoder_layers_6_mlp_up_proj_weight_quantized, x = input_135)[name = string("up_13")];
             string gate_27_mode_0 = const()[name = string("gate_27_mode_0"), val = string("TANH_APPROXIMATION")];
             tensor<fp16, [1, 1152, 1, 128]> gate_27 = gelu(mode = gate_27_mode_0, x = gate_25)[name = string("gate_27")];
             tensor<fp16, [1, 1152, 1, 128]> input_137 = mul(x = gate_27, y = up_13)[name = string("input_137")];
@@ -1475,7 +1480,7 @@ program(1.3)
             tensor<int32, [4]> var_1750_pad_0 = const()[name = string("op_1750_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_1750_dilations_0 = const()[name = string("op_1750_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_1750_groups_0 = const()[name = string("op_1750_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_1750 = conv(dilations = var_1750_dilations_0, groups = var_1750_groups_0, pad = var_1750_pad_0, pad_type = var_1750_pad_type_0, strides = var_1750_strides_0, weight = encoder_layers_6_mlp_down_proj_weight, x = input_137)[name = string("op_1750")];
+            tensor<fp16, [1, 768, 1, 128]> var_1750 = conv(dilations = var_1750_dilations_0, groups = var_1750_groups_0, pad = var_1750_pad_0, pad_type = var_1750_pad_type_0, strides = var_1750_strides_0, weight = encoder_layers_6_mlp_down_proj_weight_quantized, x = input_137)[name = string("op_1750")];
             tensor<int32, [1]> var_1751_axes_0 = const()[name = string("op_1751_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 768, 128]> var_1751 = squeeze(axes = var_1751_axes_0, x = var_1750)[name = string("op_1751")];
             tensor<int32, [3]> var_1752 = const()[name = string("op_1752"), val = tensor<int32, [3]>([0, 2, 1])];
@@ -1489,7 +1494,7 @@ program(1.3)
             tensor<int32, [2]> var_1761_split_sizes_0 = const()[name = string("op_1761_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_1761_axis_0 = const()[name = string("op_1761_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_1761_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_1761_cast_fp16_1 = split(axis = var_1761_axis_0, split_sizes = var_1761_split_sizes_0, x = normed_193_cast_fp16)[name = string("op_1761_cast_fp16")];
-            tensor<fp16, [768]> var_1765_to_fp16 = const()[name = string("op_1765_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(595267200)))];
+            tensor<fp16, [768]> var_1765_to_fp16 = const()[name = string("op_1765_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298429696)))];
             tensor<fp16, [1, 128, 768]> out_83_cast_fp16 = mul(x = var_1761_cast_fp16_0, y = var_1765_to_fp16)[name = string("out_83_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_113_cast_fp16 = add(x = x_107_cast_fp16, y = out_83_cast_fp16)[name = string("x_113_cast_fp16")];
             fp16 const_98_promoted_to_fp16 = const()[name = string("const_98_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -1501,7 +1506,7 @@ program(1.3)
             tensor<int32, [2]> var_1799_split_sizes_0 = const()[name = string("op_1799_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_1799_axis_0 = const()[name = string("op_1799_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_1799_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_1799_cast_fp16_1 = split(axis = var_1799_axis_0, split_sizes = var_1799_split_sizes_0, x = normed_197_cast_fp16)[name = string("op_1799_cast_fp16")];
-            tensor<fp16, [768]> var_1803_to_fp16 = const()[name = string("op_1803_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(595268800)))];
+            tensor<fp16, [768]> var_1803_to_fp16 = const()[name = string("op_1803_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298431296)))];
             tensor<fp16, [1, 128, 768]> out_85_cast_fp16 = mul(x = var_1799_cast_fp16_0, y = var_1803_to_fp16)[name = string("out_85_cast_fp16")];
             tensor<int32, [3]> var_1809 = const()[name = string("op_1809"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> var_1811_axes_0 = const()[name = string("op_1811_axes_0"), val = tensor<int32, [1]>([2])];
@@ -1512,7 +1517,7 @@ program(1.3)
             tensor<int32, [4]> var_1818_pad_0 = const()[name = string("op_1818_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_1818_dilations_0 = const()[name = string("op_1818_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_1818_groups_0 = const()[name = string("op_1818_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_1818 = conv(dilations = var_1818_dilations_0, groups = var_1818_groups_0, pad = var_1818_pad_0, pad_type = var_1818_pad_type_0, strides = var_1818_strides_0, weight = encoder_layers_7_self_attn_q_proj_weight, x = var_1811_cast_fp16)[name = string("op_1818")];
+            tensor<fp16, [1, 768, 1, 128]> var_1818 = conv(dilations = var_1818_dilations_0, groups = var_1818_groups_0, pad = var_1818_pad_0, pad_type = var_1818_pad_type_0, strides = var_1818_strides_0, weight = encoder_layers_7_self_attn_q_proj_weight_quantized, x = var_1811_cast_fp16)[name = string("op_1818")];
             tensor<int32, [4]> var_1819 = const()[name = string("op_1819"), val = tensor<int32, [4]>([1, 3, 256, 128])];
             tensor<fp16, [1, 3, 256, 128]> var_1820 = reshape(shape = var_1819, x = var_1818)[name = string("op_1820")];
             tensor<int32, [4]> var_1821 = const()[name = string("op_1821"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -1521,7 +1526,7 @@ program(1.3)
             tensor<int32, [4]> var_1828_pad_0 = const()[name = string("op_1828_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_1828_dilations_0 = const()[name = string("op_1828_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_1828_groups_0 = const()[name = string("op_1828_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_1828 = conv(dilations = var_1828_dilations_0, groups = var_1828_groups_0, pad = var_1828_pad_0, pad_type = var_1828_pad_type_0, strides = var_1828_strides_0, weight = encoder_layers_7_self_attn_k_proj_weight, x = var_1811_cast_fp16)[name = string("op_1828")];
+            tensor<fp16, [1, 256, 1, 128]> var_1828 = conv(dilations = var_1828_dilations_0, groups = var_1828_groups_0, pad = var_1828_pad_0, pad_type = var_1828_pad_type_0, strides = var_1828_strides_0, weight = encoder_layers_7_self_attn_k_proj_weight_quantized, x = var_1811_cast_fp16)[name = string("op_1828")];
             tensor<int32, [4]> var_1829 = const()[name = string("op_1829"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_1830 = reshape(shape = var_1829, x = var_1828)[name = string("op_1830")];
             tensor<int32, [4]> var_1831 = const()[name = string("op_1831"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -1530,7 +1535,7 @@ program(1.3)
             tensor<int32, [4]> var_1838_pad_0 = const()[name = string("op_1838_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_1838_dilations_0 = const()[name = string("op_1838_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_1838_groups_0 = const()[name = string("op_1838_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_1838 = conv(dilations = var_1838_dilations_0, groups = var_1838_groups_0, pad = var_1838_pad_0, pad_type = var_1838_pad_type_0, strides = var_1838_strides_0, weight = encoder_layers_7_self_attn_v_proj_weight, x = var_1811_cast_fp16)[name = string("op_1838")];
+            tensor<fp16, [1, 256, 1, 128]> var_1838 = conv(dilations = var_1838_dilations_0, groups = var_1838_groups_0, pad = var_1838_pad_0, pad_type = var_1838_pad_type_0, strides = var_1838_strides_0, weight = encoder_layers_7_self_attn_v_proj_weight_quantized, x = var_1811_cast_fp16)[name = string("op_1838")];
             tensor<int32, [4]> var_1839 = const()[name = string("op_1839"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_1840 = reshape(shape = var_1839, x = var_1838)[name = string("op_1840")];
             tensor<int32, [4]> var_1841 = const()[name = string("op_1841"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -1544,7 +1549,7 @@ program(1.3)
             tensor<int32, [2]> var_1852_split_sizes_0 = const()[name = string("op_1852_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_1852_axis_0 = const()[name = string("op_1852_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 256]> var_1852_cast_fp16_0, tensor<fp16, [1, 3, 128, 256]> var_1852_cast_fp16_1 = split(axis = var_1852_axis_0, split_sizes = var_1852_split_sizes_0, x = normed_203_cast_fp16)[name = string("op_1852_cast_fp16")];
-            tensor<fp16, [256]> var_1856_to_fp16 = const()[name = string("op_1856_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(595270400)))];
+            tensor<fp16, [256]> var_1856_to_fp16 = const()[name = string("op_1856_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298432896)))];
             tensor<fp16, [1, 3, 128, 256]> out_87_cast_fp16 = mul(x = var_1852_cast_fp16_0, y = var_1856_to_fp16)[name = string("out_87_cast_fp16")];
             fp16 const_102_promoted_to_fp16 = const()[name = string("const_102_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 1, 128, 256]> k_43 = transpose(perm = var_1831, x = var_1830)[name = string("transpose_150")];
@@ -1556,9 +1561,9 @@ program(1.3)
             tensor<int32, [2]> var_1868_split_sizes_0 = const()[name = string("op_1868_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_1868_axis_0 = const()[name = string("op_1868_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 256]> var_1868_cast_fp16_0, tensor<fp16, [1, 1, 128, 256]> var_1868_cast_fp16_1 = split(axis = var_1868_axis_0, split_sizes = var_1868_split_sizes_0, x = normed_207_cast_fp16)[name = string("op_1868_cast_fp16")];
-            tensor<fp16, [256]> var_1872_to_fp16 = const()[name = string("op_1872_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(595270976)))];
+            tensor<fp16, [256]> var_1872_to_fp16 = const()[name = string("op_1872_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298433472)))];
             tensor<fp16, [1, 1, 128, 256]> out_89_cast_fp16 = mul(x = var_1868_cast_fp16_0, y = var_1872_to_fp16)[name = string("out_89_cast_fp16")];
-            tensor<fp16, [1, 3, 128, 256]> var_1875 = mul(x = out_87_cast_fp16, y = cos_1)[name = string("op_1875")];
+            tensor<fp16, [1, 3, 128, 256]> var_1875 = mul(x = out_87_cast_fp16, y = cos_1_quantized)[name = string("op_1875")];
             tensor<int32, [2]> var_1876_split_sizes_0 = const()[name = string("op_1876_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_1876_axis_0 = const()[name = string("op_1876_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 128]> var_1876_0, tensor<fp16, [1, 3, 128, 128]> var_1876_1 = split(axis = var_1876_axis_0, split_sizes = var_1876_split_sizes_0, x = out_87_cast_fp16)[name = string("op_1876")];
@@ -1566,9 +1571,9 @@ program(1.3)
             tensor<fp16, [1, 3, 128, 128]> var_1878 = mul(x = var_1876_1, y = const_104_promoted)[name = string("op_1878")];
             bool var_1880_interleave_0 = const()[name = string("op_1880_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 3, 128, 256]> var_1880 = concat(axis = var_23, interleave = var_1880_interleave_0, values = (var_1878, var_1876_0))[name = string("op_1880")];
-            tensor<fp16, [1, 3, 128, 256]> var_1881 = mul(x = var_1880, y = sin_1)[name = string("op_1881")];
+            tensor<fp16, [1, 3, 128, 256]> var_1881 = mul(x = var_1880, y = sin_1_quantized)[name = string("op_1881")];
             tensor<fp16, [1, 3, 128, 256]> q_47 = add(x = var_1875, y = var_1881)[name = string("q_47")];
-            tensor<fp16, [1, 1, 128, 256]> var_1883 = mul(x = out_89_cast_fp16, y = cos_1)[name = string("op_1883")];
+            tensor<fp16, [1, 1, 128, 256]> var_1883 = mul(x = out_89_cast_fp16, y = cos_1_quantized)[name = string("op_1883")];
             tensor<int32, [2]> var_1884_split_sizes_0 = const()[name = string("op_1884_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_1884_axis_0 = const()[name = string("op_1884_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 128]> var_1884_0, tensor<fp16, [1, 1, 128, 128]> var_1884_1 = split(axis = var_1884_axis_0, split_sizes = var_1884_split_sizes_0, x = out_89_cast_fp16)[name = string("op_1884")];
@@ -1576,7 +1581,7 @@ program(1.3)
             tensor<fp16, [1, 1, 128, 128]> var_1886 = mul(x = var_1884_1, y = const_105_promoted)[name = string("op_1886")];
             bool var_1888_interleave_0 = const()[name = string("op_1888_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 1, 128, 256]> var_1888 = concat(axis = var_23, interleave = var_1888_interleave_0, values = (var_1886, var_1884_0))[name = string("op_1888")];
-            tensor<fp16, [1, 1, 128, 256]> var_1889 = mul(x = var_1888, y = sin_1)[name = string("op_1889")];
+            tensor<fp16, [1, 1, 128, 256]> var_1889 = mul(x = var_1888, y = sin_1_quantized)[name = string("op_1889")];
             tensor<fp16, [1, 1, 128, 256]> hidden_states_85 = add(x = var_1883, y = var_1889)[name = string("hidden_states_85")];
             tensor<int32, [1]> hidden_states_87_axes_0 = const()[name = string("hidden_states_87_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 1, 1, 128, 256]> hidden_states_87 = expand_dims(axes = hidden_states_87_axes_0, x = hidden_states_85)[name = string("hidden_states_87")];
@@ -1606,14 +1611,14 @@ program(1.3)
             tensor<fp16, [1, 128, 3, 256]> var_1916 = transpose(perm = var_1915, x = var_1913_cast_fp16)[name = string("transpose_148")];
             tensor<fp16, [1, 128, 768]> attn_out_45 = reshape(shape = var_1918, x = var_1916)[name = string("attn_out_45")];
             tensor<int32, [3]> var_1920 = const()[name = string("op_1920"), val = tensor<int32, [3]>([0, 2, 1])];
-            tensor<fp16, [768, 768, 1]> squeeze_7 = const()[name = string("squeeze_7"), val = tensor<fp16, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(595271552)))];
+            tensor<fp16, [768, 768, 1]> squeeze_7_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298434048))), scale = tensor<fp16, [768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299023936))))[name = string("squeeze_7_quantized")];
             string var_1929_pad_type_0 = const()[name = string("op_1929_pad_type_0"), val = string("valid")];
             int32 var_1929_groups_0 = const()[name = string("op_1929_groups_0"), val = int32(1)];
             tensor<int32, [1]> var_1929_strides_0 = const()[name = string("op_1929_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> var_1929_pad_0 = const()[name = string("op_1929_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> var_1929_dilations_0 = const()[name = string("op_1929_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, [1, 768, 128]> var_1921 = transpose(perm = var_1920, x = attn_out_45)[name = string("transpose_147")];
-            tensor<fp16, [1, 768, 128]> var_1929 = conv(dilations = var_1929_dilations_0, groups = var_1929_groups_0, pad = var_1929_pad_0, pad_type = var_1929_pad_type_0, strides = var_1929_strides_0, weight = squeeze_7, x = var_1921)[name = string("op_1929")];
+            tensor<fp16, [1, 768, 128]> var_1929 = conv(dilations = var_1929_dilations_0, groups = var_1929_groups_0, pad = var_1929_pad_0, pad_type = var_1929_pad_type_0, strides = var_1929_strides_0, weight = squeeze_7_quantized, x = var_1921)[name = string("op_1929")];
             tensor<int32, [3]> var_1930 = const()[name = string("op_1930"), val = tensor<int32, [3]>([0, 2, 1])];
             fp16 const_106_promoted_to_fp16 = const()[name = string("const_106_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 128, 768]> x_121 = transpose(perm = var_1930, x = var_1929)[name = string("transpose_146")];
@@ -1625,7 +1630,7 @@ program(1.3)
             tensor<int32, [2]> var_1939_split_sizes_0 = const()[name = string("op_1939_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_1939_axis_0 = const()[name = string("op_1939_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_1939_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_1939_cast_fp16_1 = split(axis = var_1939_axis_0, split_sizes = var_1939_split_sizes_0, x = normed_211_cast_fp16)[name = string("op_1939_cast_fp16")];
-            tensor<fp16, [768]> var_1943_to_fp16 = const()[name = string("op_1943_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(596451264)))];
+            tensor<fp16, [768]> var_1943_to_fp16 = const()[name = string("op_1943_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299025536)))];
             tensor<fp16, [1, 128, 768]> out_91_cast_fp16 = mul(x = var_1939_cast_fp16_0, y = var_1943_to_fp16)[name = string("out_91_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_123_cast_fp16 = add(x = x_113_cast_fp16, y = out_91_cast_fp16)[name = string("x_123_cast_fp16")];
             fp16 const_108_promoted_to_fp16 = const()[name = string("const_108_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -1637,7 +1642,7 @@ program(1.3)
             tensor<int32, [2]> var_1955_split_sizes_0 = const()[name = string("op_1955_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_1955_axis_0 = const()[name = string("op_1955_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_1955_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_1955_cast_fp16_1 = split(axis = var_1955_axis_0, split_sizes = var_1955_split_sizes_0, x = normed_215_cast_fp16)[name = string("op_1955_cast_fp16")];
-            tensor<fp16, [768]> var_1959_to_fp16 = const()[name = string("op_1959_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(596452864)))];
+            tensor<fp16, [768]> var_1959_to_fp16 = const()[name = string("op_1959_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299027136)))];
             tensor<fp16, [1, 128, 768]> out_93_cast_fp16 = mul(x = var_1955_cast_fp16_0, y = var_1959_to_fp16)[name = string("out_93_cast_fp16")];
             tensor<int32, [3]> var_1966 = const()[name = string("op_1966"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> input_155_axes_0 = const()[name = string("input_155_axes_0"), val = tensor<int32, [1]>([2])];
@@ -1648,13 +1653,13 @@ program(1.3)
             tensor<int32, [4]> gate_29_pad_0 = const()[name = string("gate_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> gate_29_dilations_0 = const()[name = string("gate_29_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 gate_29_groups_0 = const()[name = string("gate_29_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> gate_29 = conv(dilations = gate_29_dilations_0, groups = gate_29_groups_0, pad = gate_29_pad_0, pad_type = gate_29_pad_type_0, strides = gate_29_strides_0, weight = encoder_layers_7_mlp_gate_proj_weight, x = input_155)[name = string("gate_29")];
+            tensor<fp16, [1, 1152, 1, 128]> gate_29 = conv(dilations = gate_29_dilations_0, groups = gate_29_groups_0, pad = gate_29_pad_0, pad_type = gate_29_pad_type_0, strides = gate_29_strides_0, weight = encoder_layers_7_mlp_gate_proj_weight_quantized, x = input_155)[name = string("gate_29")];
             string up_15_pad_type_0 = const()[name = string("up_15_pad_type_0"), val = string("valid")];
             tensor<int32, [2]> up_15_strides_0 = const()[name = string("up_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> up_15_pad_0 = const()[name = string("up_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> up_15_dilations_0 = const()[name = string("up_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 up_15_groups_0 = const()[name = string("up_15_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> up_15 = conv(dilations = up_15_dilations_0, groups = up_15_groups_0, pad = up_15_pad_0, pad_type = up_15_pad_type_0, strides = up_15_strides_0, weight = encoder_layers_7_mlp_up_proj_weight, x = input_155)[name = string("up_15")];
+            tensor<fp16, [1, 1152, 1, 128]> up_15 = conv(dilations = up_15_dilations_0, groups = up_15_groups_0, pad = up_15_pad_0, pad_type = up_15_pad_type_0, strides = up_15_strides_0, weight = encoder_layers_7_mlp_up_proj_weight_quantized, x = input_155)[name = string("up_15")];
             string gate_31_mode_0 = const()[name = string("gate_31_mode_0"), val = string("TANH_APPROXIMATION")];
             tensor<fp16, [1, 1152, 1, 128]> gate_31 = gelu(mode = gate_31_mode_0, x = gate_29)[name = string("gate_31")];
             tensor<fp16, [1, 1152, 1, 128]> input_157 = mul(x = gate_31, y = up_15)[name = string("input_157")];
@@ -1663,7 +1668,7 @@ program(1.3)
             tensor<int32, [4]> var_1988_pad_0 = const()[name = string("op_1988_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_1988_dilations_0 = const()[name = string("op_1988_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_1988_groups_0 = const()[name = string("op_1988_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_1988 = conv(dilations = var_1988_dilations_0, groups = var_1988_groups_0, pad = var_1988_pad_0, pad_type = var_1988_pad_type_0, strides = var_1988_strides_0, weight = encoder_layers_7_mlp_down_proj_weight, x = input_157)[name = string("op_1988")];
+            tensor<fp16, [1, 768, 1, 128]> var_1988 = conv(dilations = var_1988_dilations_0, groups = var_1988_groups_0, pad = var_1988_pad_0, pad_type = var_1988_pad_type_0, strides = var_1988_strides_0, weight = encoder_layers_7_mlp_down_proj_weight_quantized, x = input_157)[name = string("op_1988")];
             tensor<int32, [1]> var_1989_axes_0 = const()[name = string("op_1989_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 768, 128]> var_1989 = squeeze(axes = var_1989_axes_0, x = var_1988)[name = string("op_1989")];
             tensor<int32, [3]> var_1990 = const()[name = string("op_1990"), val = tensor<int32, [3]>([0, 2, 1])];
@@ -1677,7 +1682,7 @@ program(1.3)
             tensor<int32, [2]> var_1999_split_sizes_0 = const()[name = string("op_1999_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_1999_axis_0 = const()[name = string("op_1999_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_1999_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_1999_cast_fp16_1 = split(axis = var_1999_axis_0, split_sizes = var_1999_split_sizes_0, x = normed_221_cast_fp16)[name = string("op_1999_cast_fp16")];
-            tensor<fp16, [768]> var_2003_to_fp16 = const()[name = string("op_2003_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(596454464)))];
+            tensor<fp16, [768]> var_2003_to_fp16 = const()[name = string("op_2003_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299028736)))];
             tensor<fp16, [1, 128, 768]> out_95_cast_fp16 = mul(x = var_1999_cast_fp16_0, y = var_2003_to_fp16)[name = string("out_95_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_129_cast_fp16 = add(x = x_123_cast_fp16, y = out_95_cast_fp16)[name = string("x_129_cast_fp16")];
             fp16 const_112_promoted_to_fp16 = const()[name = string("const_112_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -1689,7 +1694,7 @@ program(1.3)
             tensor<int32, [2]> var_2037_split_sizes_0 = const()[name = string("op_2037_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_2037_axis_0 = const()[name = string("op_2037_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_2037_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_2037_cast_fp16_1 = split(axis = var_2037_axis_0, split_sizes = var_2037_split_sizes_0, x = normed_225_cast_fp16)[name = string("op_2037_cast_fp16")];
-            tensor<fp16, [768]> var_2041_to_fp16 = const()[name = string("op_2041_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(596456064)))];
+            tensor<fp16, [768]> var_2041_to_fp16 = const()[name = string("op_2041_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299030336)))];
             tensor<fp16, [1, 128, 768]> out_97_cast_fp16 = mul(x = var_2037_cast_fp16_0, y = var_2041_to_fp16)[name = string("out_97_cast_fp16")];
             tensor<int32, [3]> var_2047 = const()[name = string("op_2047"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> var_2049_axes_0 = const()[name = string("op_2049_axes_0"), val = tensor<int32, [1]>([2])];
@@ -1700,7 +1705,7 @@ program(1.3)
             tensor<int32, [4]> var_2056_pad_0 = const()[name = string("op_2056_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_2056_dilations_0 = const()[name = string("op_2056_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_2056_groups_0 = const()[name = string("op_2056_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_2056 = conv(dilations = var_2056_dilations_0, groups = var_2056_groups_0, pad = var_2056_pad_0, pad_type = var_2056_pad_type_0, strides = var_2056_strides_0, weight = encoder_layers_8_self_attn_q_proj_weight, x = var_2049_cast_fp16)[name = string("op_2056")];
+            tensor<fp16, [1, 768, 1, 128]> var_2056 = conv(dilations = var_2056_dilations_0, groups = var_2056_groups_0, pad = var_2056_pad_0, pad_type = var_2056_pad_type_0, strides = var_2056_strides_0, weight = encoder_layers_8_self_attn_q_proj_weight_quantized, x = var_2049_cast_fp16)[name = string("op_2056")];
             tensor<int32, [4]> var_2057 = const()[name = string("op_2057"), val = tensor<int32, [4]>([1, 3, 256, 128])];
             tensor<fp16, [1, 3, 256, 128]> var_2058 = reshape(shape = var_2057, x = var_2056)[name = string("op_2058")];
             tensor<int32, [4]> var_2059 = const()[name = string("op_2059"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -1709,7 +1714,7 @@ program(1.3)
             tensor<int32, [4]> var_2066_pad_0 = const()[name = string("op_2066_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_2066_dilations_0 = const()[name = string("op_2066_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_2066_groups_0 = const()[name = string("op_2066_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_2066 = conv(dilations = var_2066_dilations_0, groups = var_2066_groups_0, pad = var_2066_pad_0, pad_type = var_2066_pad_type_0, strides = var_2066_strides_0, weight = encoder_layers_8_self_attn_k_proj_weight, x = var_2049_cast_fp16)[name = string("op_2066")];
+            tensor<fp16, [1, 256, 1, 128]> var_2066 = conv(dilations = var_2066_dilations_0, groups = var_2066_groups_0, pad = var_2066_pad_0, pad_type = var_2066_pad_type_0, strides = var_2066_strides_0, weight = encoder_layers_8_self_attn_k_proj_weight_quantized, x = var_2049_cast_fp16)[name = string("op_2066")];
             tensor<int32, [4]> var_2067 = const()[name = string("op_2067"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_2068 = reshape(shape = var_2067, x = var_2066)[name = string("op_2068")];
             tensor<int32, [4]> var_2069 = const()[name = string("op_2069"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -1718,7 +1723,7 @@ program(1.3)
             tensor<int32, [4]> var_2076_pad_0 = const()[name = string("op_2076_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_2076_dilations_0 = const()[name = string("op_2076_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_2076_groups_0 = const()[name = string("op_2076_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_2076 = conv(dilations = var_2076_dilations_0, groups = var_2076_groups_0, pad = var_2076_pad_0, pad_type = var_2076_pad_type_0, strides = var_2076_strides_0, weight = encoder_layers_8_self_attn_v_proj_weight, x = var_2049_cast_fp16)[name = string("op_2076")];
+            tensor<fp16, [1, 256, 1, 128]> var_2076 = conv(dilations = var_2076_dilations_0, groups = var_2076_groups_0, pad = var_2076_pad_0, pad_type = var_2076_pad_type_0, strides = var_2076_strides_0, weight = encoder_layers_8_self_attn_v_proj_weight_quantized, x = var_2049_cast_fp16)[name = string("op_2076")];
             tensor<int32, [4]> var_2077 = const()[name = string("op_2077"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_2078 = reshape(shape = var_2077, x = var_2076)[name = string("op_2078")];
             tensor<int32, [4]> var_2079 = const()[name = string("op_2079"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -1732,7 +1737,7 @@ program(1.3)
             tensor<int32, [2]> var_2090_split_sizes_0 = const()[name = string("op_2090_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_2090_axis_0 = const()[name = string("op_2090_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 256]> var_2090_cast_fp16_0, tensor<fp16, [1, 3, 128, 256]> var_2090_cast_fp16_1 = split(axis = var_2090_axis_0, split_sizes = var_2090_split_sizes_0, x = normed_231_cast_fp16)[name = string("op_2090_cast_fp16")];
-            tensor<fp16, [256]> var_2094_to_fp16 = const()[name = string("op_2094_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(596457664)))];
+            tensor<fp16, [256]> var_2094_to_fp16 = const()[name = string("op_2094_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299031936)))];
             tensor<fp16, [1, 3, 128, 256]> out_99_cast_fp16 = mul(x = var_2090_cast_fp16_0, y = var_2094_to_fp16)[name = string("out_99_cast_fp16")];
             fp16 const_116_promoted_to_fp16 = const()[name = string("const_116_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 1, 128, 256]> k_49 = transpose(perm = var_2069, x = var_2068)[name = string("transpose_141")];
@@ -1744,9 +1749,9 @@ program(1.3)
             tensor<int32, [2]> var_2106_split_sizes_0 = const()[name = string("op_2106_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_2106_axis_0 = const()[name = string("op_2106_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 256]> var_2106_cast_fp16_0, tensor<fp16, [1, 1, 128, 256]> var_2106_cast_fp16_1 = split(axis = var_2106_axis_0, split_sizes = var_2106_split_sizes_0, x = normed_235_cast_fp16)[name = string("op_2106_cast_fp16")];
-            tensor<fp16, [256]> var_2110_to_fp16 = const()[name = string("op_2110_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(596458240)))];
+            tensor<fp16, [256]> var_2110_to_fp16 = const()[name = string("op_2110_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299032512)))];
             tensor<fp16, [1, 1, 128, 256]> out_101_cast_fp16 = mul(x = var_2106_cast_fp16_0, y = var_2110_to_fp16)[name = string("out_101_cast_fp16")];
-            tensor<fp16, [1, 3, 128, 256]> var_2113 = mul(x = out_99_cast_fp16, y = cos_1)[name = string("op_2113")];
+            tensor<fp16, [1, 3, 128, 256]> var_2113 = mul(x = out_99_cast_fp16, y = cos_1_quantized)[name = string("op_2113")];
             tensor<int32, [2]> var_2114_split_sizes_0 = const()[name = string("op_2114_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_2114_axis_0 = const()[name = string("op_2114_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 128]> var_2114_0, tensor<fp16, [1, 3, 128, 128]> var_2114_1 = split(axis = var_2114_axis_0, split_sizes = var_2114_split_sizes_0, x = out_99_cast_fp16)[name = string("op_2114")];
@@ -1754,9 +1759,9 @@ program(1.3)
             tensor<fp16, [1, 3, 128, 128]> var_2116 = mul(x = var_2114_1, y = const_118_promoted)[name = string("op_2116")];
             bool var_2118_interleave_0 = const()[name = string("op_2118_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 3, 128, 256]> var_2118 = concat(axis = var_23, interleave = var_2118_interleave_0, values = (var_2116, var_2114_0))[name = string("op_2118")];
-            tensor<fp16, [1, 3, 128, 256]> var_2119 = mul(x = var_2118, y = sin_1)[name = string("op_2119")];
+            tensor<fp16, [1, 3, 128, 256]> var_2119 = mul(x = var_2118, y = sin_1_quantized)[name = string("op_2119")];
             tensor<fp16, [1, 3, 128, 256]> q_53 = add(x = var_2113, y = var_2119)[name = string("q_53")];
-            tensor<fp16, [1, 1, 128, 256]> var_2121 = mul(x = out_101_cast_fp16, y = cos_1)[name = string("op_2121")];
+            tensor<fp16, [1, 1, 128, 256]> var_2121 = mul(x = out_101_cast_fp16, y = cos_1_quantized)[name = string("op_2121")];
             tensor<int32, [2]> var_2122_split_sizes_0 = const()[name = string("op_2122_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_2122_axis_0 = const()[name = string("op_2122_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 128]> var_2122_0, tensor<fp16, [1, 1, 128, 128]> var_2122_1 = split(axis = var_2122_axis_0, split_sizes = var_2122_split_sizes_0, x = out_101_cast_fp16)[name = string("op_2122")];
@@ -1764,7 +1769,7 @@ program(1.3)
             tensor<fp16, [1, 1, 128, 128]> var_2124 = mul(x = var_2122_1, y = const_119_promoted)[name = string("op_2124")];
             bool var_2126_interleave_0 = const()[name = string("op_2126_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 1, 128, 256]> var_2126 = concat(axis = var_23, interleave = var_2126_interleave_0, values = (var_2124, var_2122_0))[name = string("op_2126")];
-            tensor<fp16, [1, 1, 128, 256]> var_2127 = mul(x = var_2126, y = sin_1)[name = string("op_2127")];
+            tensor<fp16, [1, 1, 128, 256]> var_2127 = mul(x = var_2126, y = sin_1_quantized)[name = string("op_2127")];
             tensor<fp16, [1, 1, 128, 256]> hidden_states_97 = add(x = var_2121, y = var_2127)[name = string("hidden_states_97")];
             tensor<int32, [1]> hidden_states_99_axes_0 = const()[name = string("hidden_states_99_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 1, 1, 128, 256]> hidden_states_99 = expand_dims(axes = hidden_states_99_axes_0, x = hidden_states_97)[name = string("hidden_states_99")];
@@ -1794,14 +1799,14 @@ program(1.3)
             tensor<fp16, [1, 128, 3, 256]> var_2154 = transpose(perm = var_2153, x = var_2151_cast_fp16)[name = string("transpose_139")];
             tensor<fp16, [1, 128, 768]> attn_out_51 = reshape(shape = var_2156, x = var_2154)[name = string("attn_out_51")];
             tensor<int32, [3]> var_2158 = const()[name = string("op_2158"), val = tensor<int32, [3]>([0, 2, 1])];
-            tensor<fp16, [768, 768, 1]> squeeze_8 = const()[name = string("squeeze_8"), val = tensor<fp16, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(596458816)))];
+            tensor<fp16, [768, 768, 1]> squeeze_8_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299033088))), scale = tensor<fp16, [768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299622976))))[name = string("squeeze_8_quantized")];
             string var_2167_pad_type_0 = const()[name = string("op_2167_pad_type_0"), val = string("valid")];
             int32 var_2167_groups_0 = const()[name = string("op_2167_groups_0"), val = int32(1)];
             tensor<int32, [1]> var_2167_strides_0 = const()[name = string("op_2167_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> var_2167_pad_0 = const()[name = string("op_2167_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> var_2167_dilations_0 = const()[name = string("op_2167_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, [1, 768, 128]> var_2159 = transpose(perm = var_2158, x = attn_out_51)[name = string("transpose_138")];
-            tensor<fp16, [1, 768, 128]> var_2167 = conv(dilations = var_2167_dilations_0, groups = var_2167_groups_0, pad = var_2167_pad_0, pad_type = var_2167_pad_type_0, strides = var_2167_strides_0, weight = squeeze_8, x = var_2159)[name = string("op_2167")];
+            tensor<fp16, [1, 768, 128]> var_2167 = conv(dilations = var_2167_dilations_0, groups = var_2167_groups_0, pad = var_2167_pad_0, pad_type = var_2167_pad_type_0, strides = var_2167_strides_0, weight = squeeze_8_quantized, x = var_2159)[name = string("op_2167")];
             tensor<int32, [3]> var_2168 = const()[name = string("op_2168"), val = tensor<int32, [3]>([0, 2, 1])];
             fp16 const_120_promoted_to_fp16 = const()[name = string("const_120_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 128, 768]> x_137 = transpose(perm = var_2168, x = var_2167)[name = string("transpose_137")];
@@ -1813,7 +1818,7 @@ program(1.3)
             tensor<int32, [2]> var_2177_split_sizes_0 = const()[name = string("op_2177_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_2177_axis_0 = const()[name = string("op_2177_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_2177_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_2177_cast_fp16_1 = split(axis = var_2177_axis_0, split_sizes = var_2177_split_sizes_0, x = normed_239_cast_fp16)[name = string("op_2177_cast_fp16")];
-            tensor<fp16, [768]> var_2181_to_fp16 = const()[name = string("op_2181_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(597638528)))];
+            tensor<fp16, [768]> var_2181_to_fp16 = const()[name = string("op_2181_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299624576)))];
             tensor<fp16, [1, 128, 768]> out_103_cast_fp16 = mul(x = var_2177_cast_fp16_0, y = var_2181_to_fp16)[name = string("out_103_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_139_cast_fp16 = add(x = x_129_cast_fp16, y = out_103_cast_fp16)[name = string("x_139_cast_fp16")];
             fp16 const_122_promoted_to_fp16 = const()[name = string("const_122_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -1825,7 +1830,7 @@ program(1.3)
             tensor<int32, [2]> var_2193_split_sizes_0 = const()[name = string("op_2193_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_2193_axis_0 = const()[name = string("op_2193_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_2193_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_2193_cast_fp16_1 = split(axis = var_2193_axis_0, split_sizes = var_2193_split_sizes_0, x = normed_243_cast_fp16)[name = string("op_2193_cast_fp16")];
-            tensor<fp16, [768]> var_2197_to_fp16 = const()[name = string("op_2197_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(597640128)))];
+            tensor<fp16, [768]> var_2197_to_fp16 = const()[name = string("op_2197_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299626176)))];
             tensor<fp16, [1, 128, 768]> out_105_cast_fp16 = mul(x = var_2193_cast_fp16_0, y = var_2197_to_fp16)[name = string("out_105_cast_fp16")];
             tensor<int32, [3]> var_2204 = const()[name = string("op_2204"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> input_175_axes_0 = const()[name = string("input_175_axes_0"), val = tensor<int32, [1]>([2])];
@@ -1836,13 +1841,13 @@ program(1.3)
             tensor<int32, [4]> gate_33_pad_0 = const()[name = string("gate_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> gate_33_dilations_0 = const()[name = string("gate_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 gate_33_groups_0 = const()[name = string("gate_33_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> gate_33 = conv(dilations = gate_33_dilations_0, groups = gate_33_groups_0, pad = gate_33_pad_0, pad_type = gate_33_pad_type_0, strides = gate_33_strides_0, weight = encoder_layers_8_mlp_gate_proj_weight, x = input_175)[name = string("gate_33")];
+            tensor<fp16, [1, 1152, 1, 128]> gate_33 = conv(dilations = gate_33_dilations_0, groups = gate_33_groups_0, pad = gate_33_pad_0, pad_type = gate_33_pad_type_0, strides = gate_33_strides_0, weight = encoder_layers_8_mlp_gate_proj_weight_quantized, x = input_175)[name = string("gate_33")];
             string up_17_pad_type_0 = const()[name = string("up_17_pad_type_0"), val = string("valid")];
             tensor<int32, [2]> up_17_strides_0 = const()[name = string("up_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> up_17_pad_0 = const()[name = string("up_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> up_17_dilations_0 = const()[name = string("up_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 up_17_groups_0 = const()[name = string("up_17_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> up_17 = conv(dilations = up_17_dilations_0, groups = up_17_groups_0, pad = up_17_pad_0, pad_type = up_17_pad_type_0, strides = up_17_strides_0, weight = encoder_layers_8_mlp_up_proj_weight, x = input_175)[name = string("up_17")];
+            tensor<fp16, [1, 1152, 1, 128]> up_17 = conv(dilations = up_17_dilations_0, groups = up_17_groups_0, pad = up_17_pad_0, pad_type = up_17_pad_type_0, strides = up_17_strides_0, weight = encoder_layers_8_mlp_up_proj_weight_quantized, x = input_175)[name = string("up_17")];
             string gate_35_mode_0 = const()[name = string("gate_35_mode_0"), val = string("TANH_APPROXIMATION")];
             tensor<fp16, [1, 1152, 1, 128]> gate_35 = gelu(mode = gate_35_mode_0, x = gate_33)[name = string("gate_35")];
             tensor<fp16, [1, 1152, 1, 128]> input_177 = mul(x = gate_35, y = up_17)[name = string("input_177")];
@@ -1851,7 +1856,7 @@ program(1.3)
             tensor<int32, [4]> var_2226_pad_0 = const()[name = string("op_2226_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_2226_dilations_0 = const()[name = string("op_2226_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_2226_groups_0 = const()[name = string("op_2226_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_2226 = conv(dilations = var_2226_dilations_0, groups = var_2226_groups_0, pad = var_2226_pad_0, pad_type = var_2226_pad_type_0, strides = var_2226_strides_0, weight = encoder_layers_8_mlp_down_proj_weight, x = input_177)[name = string("op_2226")];
+            tensor<fp16, [1, 768, 1, 128]> var_2226 = conv(dilations = var_2226_dilations_0, groups = var_2226_groups_0, pad = var_2226_pad_0, pad_type = var_2226_pad_type_0, strides = var_2226_strides_0, weight = encoder_layers_8_mlp_down_proj_weight_quantized, x = input_177)[name = string("op_2226")];
             tensor<int32, [1]> var_2227_axes_0 = const()[name = string("op_2227_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 768, 128]> var_2227 = squeeze(axes = var_2227_axes_0, x = var_2226)[name = string("op_2227")];
             tensor<int32, [3]> var_2228 = const()[name = string("op_2228"), val = tensor<int32, [3]>([0, 2, 1])];
@@ -1865,7 +1870,7 @@ program(1.3)
             tensor<int32, [2]> var_2237_split_sizes_0 = const()[name = string("op_2237_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_2237_axis_0 = const()[name = string("op_2237_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_2237_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_2237_cast_fp16_1 = split(axis = var_2237_axis_0, split_sizes = var_2237_split_sizes_0, x = normed_249_cast_fp16)[name = string("op_2237_cast_fp16")];
-            tensor<fp16, [768]> var_2241_to_fp16 = const()[name = string("op_2241_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(597641728)))];
+            tensor<fp16, [768]> var_2241_to_fp16 = const()[name = string("op_2241_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299627776)))];
             tensor<fp16, [1, 128, 768]> out_107_cast_fp16 = mul(x = var_2237_cast_fp16_0, y = var_2241_to_fp16)[name = string("out_107_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_145_cast_fp16 = add(x = x_139_cast_fp16, y = out_107_cast_fp16)[name = string("x_145_cast_fp16")];
             fp16 const_126_promoted_to_fp16 = const()[name = string("const_126_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -1877,7 +1882,7 @@ program(1.3)
             tensor<int32, [2]> var_2275_split_sizes_0 = const()[name = string("op_2275_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_2275_axis_0 = const()[name = string("op_2275_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_2275_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_2275_cast_fp16_1 = split(axis = var_2275_axis_0, split_sizes = var_2275_split_sizes_0, x = normed_253_cast_fp16)[name = string("op_2275_cast_fp16")];
-            tensor<fp16, [768]> var_2279_to_fp16 = const()[name = string("op_2279_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(597643328)))];
+            tensor<fp16, [768]> var_2279_to_fp16 = const()[name = string("op_2279_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299629376)))];
             tensor<fp16, [1, 128, 768]> out_109_cast_fp16 = mul(x = var_2275_cast_fp16_0, y = var_2279_to_fp16)[name = string("out_109_cast_fp16")];
             tensor<int32, [3]> var_2285 = const()[name = string("op_2285"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> var_2287_axes_0 = const()[name = string("op_2287_axes_0"), val = tensor<int32, [1]>([2])];
@@ -1888,7 +1893,7 @@ program(1.3)
             tensor<int32, [4]> var_2294_pad_0 = const()[name = string("op_2294_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_2294_dilations_0 = const()[name = string("op_2294_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_2294_groups_0 = const()[name = string("op_2294_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_2294 = conv(dilations = var_2294_dilations_0, groups = var_2294_groups_0, pad = var_2294_pad_0, pad_type = var_2294_pad_type_0, strides = var_2294_strides_0, weight = encoder_layers_9_self_attn_q_proj_weight, x = var_2287_cast_fp16)[name = string("op_2294")];
+            tensor<fp16, [1, 768, 1, 128]> var_2294 = conv(dilations = var_2294_dilations_0, groups = var_2294_groups_0, pad = var_2294_pad_0, pad_type = var_2294_pad_type_0, strides = var_2294_strides_0, weight = encoder_layers_9_self_attn_q_proj_weight_quantized, x = var_2287_cast_fp16)[name = string("op_2294")];
             tensor<int32, [4]> var_2295 = const()[name = string("op_2295"), val = tensor<int32, [4]>([1, 3, 256, 128])];
             tensor<fp16, [1, 3, 256, 128]> var_2296 = reshape(shape = var_2295, x = var_2294)[name = string("op_2296")];
             tensor<int32, [4]> var_2297 = const()[name = string("op_2297"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -1897,7 +1902,7 @@ program(1.3)
             tensor<int32, [4]> var_2304_pad_0 = const()[name = string("op_2304_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_2304_dilations_0 = const()[name = string("op_2304_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_2304_groups_0 = const()[name = string("op_2304_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_2304 = conv(dilations = var_2304_dilations_0, groups = var_2304_groups_0, pad = var_2304_pad_0, pad_type = var_2304_pad_type_0, strides = var_2304_strides_0, weight = encoder_layers_9_self_attn_k_proj_weight, x = var_2287_cast_fp16)[name = string("op_2304")];
+            tensor<fp16, [1, 256, 1, 128]> var_2304 = conv(dilations = var_2304_dilations_0, groups = var_2304_groups_0, pad = var_2304_pad_0, pad_type = var_2304_pad_type_0, strides = var_2304_strides_0, weight = encoder_layers_9_self_attn_k_proj_weight_quantized, x = var_2287_cast_fp16)[name = string("op_2304")];
             tensor<int32, [4]> var_2305 = const()[name = string("op_2305"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_2306 = reshape(shape = var_2305, x = var_2304)[name = string("op_2306")];
             tensor<int32, [4]> var_2307 = const()[name = string("op_2307"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -1906,7 +1911,7 @@ program(1.3)
             tensor<int32, [4]> var_2314_pad_0 = const()[name = string("op_2314_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_2314_dilations_0 = const()[name = string("op_2314_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_2314_groups_0 = const()[name = string("op_2314_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_2314 = conv(dilations = var_2314_dilations_0, groups = var_2314_groups_0, pad = var_2314_pad_0, pad_type = var_2314_pad_type_0, strides = var_2314_strides_0, weight = encoder_layers_9_self_attn_v_proj_weight, x = var_2287_cast_fp16)[name = string("op_2314")];
+            tensor<fp16, [1, 256, 1, 128]> var_2314 = conv(dilations = var_2314_dilations_0, groups = var_2314_groups_0, pad = var_2314_pad_0, pad_type = var_2314_pad_type_0, strides = var_2314_strides_0, weight = encoder_layers_9_self_attn_v_proj_weight_quantized, x = var_2287_cast_fp16)[name = string("op_2314")];
             tensor<int32, [4]> var_2315 = const()[name = string("op_2315"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_2316 = reshape(shape = var_2315, x = var_2314)[name = string("op_2316")];
             tensor<int32, [4]> var_2317 = const()[name = string("op_2317"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -1920,7 +1925,7 @@ program(1.3)
             tensor<int32, [2]> var_2328_split_sizes_0 = const()[name = string("op_2328_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_2328_axis_0 = const()[name = string("op_2328_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 256]> var_2328_cast_fp16_0, tensor<fp16, [1, 3, 128, 256]> var_2328_cast_fp16_1 = split(axis = var_2328_axis_0, split_sizes = var_2328_split_sizes_0, x = normed_259_cast_fp16)[name = string("op_2328_cast_fp16")];
-            tensor<fp16, [256]> var_2332_to_fp16 = const()[name = string("op_2332_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(597644928)))];
+            tensor<fp16, [256]> var_2332_to_fp16 = const()[name = string("op_2332_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299630976)))];
             tensor<fp16, [1, 3, 128, 256]> out_111_cast_fp16 = mul(x = var_2328_cast_fp16_0, y = var_2332_to_fp16)[name = string("out_111_cast_fp16")];
             fp16 const_130_promoted_to_fp16 = const()[name = string("const_130_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 1, 128, 256]> k_55 = transpose(perm = var_2307, x = var_2306)[name = string("transpose_132")];
@@ -1932,9 +1937,9 @@ program(1.3)
             tensor<int32, [2]> var_2344_split_sizes_0 = const()[name = string("op_2344_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_2344_axis_0 = const()[name = string("op_2344_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 256]> var_2344_cast_fp16_0, tensor<fp16, [1, 1, 128, 256]> var_2344_cast_fp16_1 = split(axis = var_2344_axis_0, split_sizes = var_2344_split_sizes_0, x = normed_263_cast_fp16)[name = string("op_2344_cast_fp16")];
-            tensor<fp16, [256]> var_2348_to_fp16 = const()[name = string("op_2348_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(597645504)))];
+            tensor<fp16, [256]> var_2348_to_fp16 = const()[name = string("op_2348_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299631552)))];
             tensor<fp16, [1, 1, 128, 256]> out_113_cast_fp16 = mul(x = var_2344_cast_fp16_0, y = var_2348_to_fp16)[name = string("out_113_cast_fp16")];
-            tensor<fp16, [1, 3, 128, 256]> var_2351 = mul(x = out_111_cast_fp16, y = cos_1)[name = string("op_2351")];
+            tensor<fp16, [1, 3, 128, 256]> var_2351 = mul(x = out_111_cast_fp16, y = cos_1_quantized)[name = string("op_2351")];
             tensor<int32, [2]> var_2352_split_sizes_0 = const()[name = string("op_2352_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_2352_axis_0 = const()[name = string("op_2352_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 128]> var_2352_0, tensor<fp16, [1, 3, 128, 128]> var_2352_1 = split(axis = var_2352_axis_0, split_sizes = var_2352_split_sizes_0, x = out_111_cast_fp16)[name = string("op_2352")];
@@ -1942,9 +1947,9 @@ program(1.3)
             tensor<fp16, [1, 3, 128, 128]> var_2354 = mul(x = var_2352_1, y = const_132_promoted)[name = string("op_2354")];
             bool var_2356_interleave_0 = const()[name = string("op_2356_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 3, 128, 256]> var_2356 = concat(axis = var_23, interleave = var_2356_interleave_0, values = (var_2354, var_2352_0))[name = string("op_2356")];
-            tensor<fp16, [1, 3, 128, 256]> var_2357 = mul(x = var_2356, y = sin_1)[name = string("op_2357")];
+            tensor<fp16, [1, 3, 128, 256]> var_2357 = mul(x = var_2356, y = sin_1_quantized)[name = string("op_2357")];
             tensor<fp16, [1, 3, 128, 256]> q_59 = add(x = var_2351, y = var_2357)[name = string("q_59")];
-            tensor<fp16, [1, 1, 128, 256]> var_2359 = mul(x = out_113_cast_fp16, y = cos_1)[name = string("op_2359")];
+            tensor<fp16, [1, 1, 128, 256]> var_2359 = mul(x = out_113_cast_fp16, y = cos_1_quantized)[name = string("op_2359")];
             tensor<int32, [2]> var_2360_split_sizes_0 = const()[name = string("op_2360_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_2360_axis_0 = const()[name = string("op_2360_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 128]> var_2360_0, tensor<fp16, [1, 1, 128, 128]> var_2360_1 = split(axis = var_2360_axis_0, split_sizes = var_2360_split_sizes_0, x = out_113_cast_fp16)[name = string("op_2360")];
@@ -1952,7 +1957,7 @@ program(1.3)
             tensor<fp16, [1, 1, 128, 128]> var_2362 = mul(x = var_2360_1, y = const_133_promoted)[name = string("op_2362")];
             bool var_2364_interleave_0 = const()[name = string("op_2364_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 1, 128, 256]> var_2364 = concat(axis = var_23, interleave = var_2364_interleave_0, values = (var_2362, var_2360_0))[name = string("op_2364")];
-            tensor<fp16, [1, 1, 128, 256]> var_2365 = mul(x = var_2364, y = sin_1)[name = string("op_2365")];
+            tensor<fp16, [1, 1, 128, 256]> var_2365 = mul(x = var_2364, y = sin_1_quantized)[name = string("op_2365")];
             tensor<fp16, [1, 1, 128, 256]> hidden_states_109 = add(x = var_2359, y = var_2365)[name = string("hidden_states_109")];
             tensor<int32, [1]> hidden_states_111_axes_0 = const()[name = string("hidden_states_111_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 1, 1, 128, 256]> hidden_states_111 = expand_dims(axes = hidden_states_111_axes_0, x = hidden_states_109)[name = string("hidden_states_111")];
@@ -1982,14 +1987,14 @@ program(1.3)
             tensor<fp16, [1, 128, 3, 256]> var_2392 = transpose(perm = var_2391, x = var_2389_cast_fp16)[name = string("transpose_130")];
             tensor<fp16, [1, 128, 768]> attn_out_57 = reshape(shape = var_2394, x = var_2392)[name = string("attn_out_57")];
             tensor<int32, [3]> var_2396 = const()[name = string("op_2396"), val = tensor<int32, [3]>([0, 2, 1])];
-            tensor<fp16, [768, 768, 1]> squeeze_9 = const()[name = string("squeeze_9"), val = tensor<fp16, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(597646080)))];
+            tensor<fp16, [768, 768, 1]> squeeze_9_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299632128))), scale = tensor<fp16, [768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300222016))))[name = string("squeeze_9_quantized")];
             string var_2405_pad_type_0 = const()[name = string("op_2405_pad_type_0"), val = string("valid")];
             int32 var_2405_groups_0 = const()[name = string("op_2405_groups_0"), val = int32(1)];
             tensor<int32, [1]> var_2405_strides_0 = const()[name = string("op_2405_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> var_2405_pad_0 = const()[name = string("op_2405_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> var_2405_dilations_0 = const()[name = string("op_2405_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, [1, 768, 128]> var_2397 = transpose(perm = var_2396, x = attn_out_57)[name = string("transpose_129")];
-            tensor<fp16, [1, 768, 128]> var_2405 = conv(dilations = var_2405_dilations_0, groups = var_2405_groups_0, pad = var_2405_pad_0, pad_type = var_2405_pad_type_0, strides = var_2405_strides_0, weight = squeeze_9, x = var_2397)[name = string("op_2405")];
+            tensor<fp16, [1, 768, 128]> var_2405 = conv(dilations = var_2405_dilations_0, groups = var_2405_groups_0, pad = var_2405_pad_0, pad_type = var_2405_pad_type_0, strides = var_2405_strides_0, weight = squeeze_9_quantized, x = var_2397)[name = string("op_2405")];
             tensor<int32, [3]> var_2406 = const()[name = string("op_2406"), val = tensor<int32, [3]>([0, 2, 1])];
             fp16 const_134_promoted_to_fp16 = const()[name = string("const_134_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 128, 768]> x_153 = transpose(perm = var_2406, x = var_2405)[name = string("transpose_128")];
@@ -2001,7 +2006,7 @@ program(1.3)
             tensor<int32, [2]> var_2415_split_sizes_0 = const()[name = string("op_2415_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_2415_axis_0 = const()[name = string("op_2415_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_2415_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_2415_cast_fp16_1 = split(axis = var_2415_axis_0, split_sizes = var_2415_split_sizes_0, x = normed_267_cast_fp16)[name = string("op_2415_cast_fp16")];
-            tensor<fp16, [768]> var_2419_to_fp16 = const()[name = string("op_2419_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(598825792)))];
+            tensor<fp16, [768]> var_2419_to_fp16 = const()[name = string("op_2419_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300223616)))];
             tensor<fp16, [1, 128, 768]> out_115_cast_fp16 = mul(x = var_2415_cast_fp16_0, y = var_2419_to_fp16)[name = string("out_115_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_155_cast_fp16 = add(x = x_145_cast_fp16, y = out_115_cast_fp16)[name = string("x_155_cast_fp16")];
             fp16 const_136_promoted_to_fp16 = const()[name = string("const_136_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -2013,7 +2018,7 @@ program(1.3)
             tensor<int32, [2]> var_2431_split_sizes_0 = const()[name = string("op_2431_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_2431_axis_0 = const()[name = string("op_2431_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_2431_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_2431_cast_fp16_1 = split(axis = var_2431_axis_0, split_sizes = var_2431_split_sizes_0, x = normed_271_cast_fp16)[name = string("op_2431_cast_fp16")];
-            tensor<fp16, [768]> var_2435_to_fp16 = const()[name = string("op_2435_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(598827392)))];
+            tensor<fp16, [768]> var_2435_to_fp16 = const()[name = string("op_2435_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300225216)))];
             tensor<fp16, [1, 128, 768]> out_117_cast_fp16 = mul(x = var_2431_cast_fp16_0, y = var_2435_to_fp16)[name = string("out_117_cast_fp16")];
             tensor<int32, [3]> var_2442 = const()[name = string("op_2442"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> input_195_axes_0 = const()[name = string("input_195_axes_0"), val = tensor<int32, [1]>([2])];
@@ -2024,13 +2029,13 @@ program(1.3)
             tensor<int32, [4]> gate_37_pad_0 = const()[name = string("gate_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> gate_37_dilations_0 = const()[name = string("gate_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 gate_37_groups_0 = const()[name = string("gate_37_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> gate_37 = conv(dilations = gate_37_dilations_0, groups = gate_37_groups_0, pad = gate_37_pad_0, pad_type = gate_37_pad_type_0, strides = gate_37_strides_0, weight = encoder_layers_9_mlp_gate_proj_weight, x = input_195)[name = string("gate_37")];
+            tensor<fp16, [1, 1152, 1, 128]> gate_37 = conv(dilations = gate_37_dilations_0, groups = gate_37_groups_0, pad = gate_37_pad_0, pad_type = gate_37_pad_type_0, strides = gate_37_strides_0, weight = encoder_layers_9_mlp_gate_proj_weight_quantized, x = input_195)[name = string("gate_37")];
             string up_19_pad_type_0 = const()[name = string("up_19_pad_type_0"), val = string("valid")];
             tensor<int32, [2]> up_19_strides_0 = const()[name = string("up_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> up_19_pad_0 = const()[name = string("up_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> up_19_dilations_0 = const()[name = string("up_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 up_19_groups_0 = const()[name = string("up_19_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> up_19 = conv(dilations = up_19_dilations_0, groups = up_19_groups_0, pad = up_19_pad_0, pad_type = up_19_pad_type_0, strides = up_19_strides_0, weight = encoder_layers_9_mlp_up_proj_weight, x = input_195)[name = string("up_19")];
+            tensor<fp16, [1, 1152, 1, 128]> up_19 = conv(dilations = up_19_dilations_0, groups = up_19_groups_0, pad = up_19_pad_0, pad_type = up_19_pad_type_0, strides = up_19_strides_0, weight = encoder_layers_9_mlp_up_proj_weight_quantized, x = input_195)[name = string("up_19")];
             string gate_39_mode_0 = const()[name = string("gate_39_mode_0"), val = string("TANH_APPROXIMATION")];
             tensor<fp16, [1, 1152, 1, 128]> gate_39 = gelu(mode = gate_39_mode_0, x = gate_37)[name = string("gate_39")];
             tensor<fp16, [1, 1152, 1, 128]> input_197 = mul(x = gate_39, y = up_19)[name = string("input_197")];
@@ -2039,7 +2044,7 @@ program(1.3)
             tensor<int32, [4]> var_2464_pad_0 = const()[name = string("op_2464_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_2464_dilations_0 = const()[name = string("op_2464_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_2464_groups_0 = const()[name = string("op_2464_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_2464 = conv(dilations = var_2464_dilations_0, groups = var_2464_groups_0, pad = var_2464_pad_0, pad_type = var_2464_pad_type_0, strides = var_2464_strides_0, weight = encoder_layers_9_mlp_down_proj_weight, x = input_197)[name = string("op_2464")];
+            tensor<fp16, [1, 768, 1, 128]> var_2464 = conv(dilations = var_2464_dilations_0, groups = var_2464_groups_0, pad = var_2464_pad_0, pad_type = var_2464_pad_type_0, strides = var_2464_strides_0, weight = encoder_layers_9_mlp_down_proj_weight_quantized, x = input_197)[name = string("op_2464")];
             tensor<int32, [1]> var_2465_axes_0 = const()[name = string("op_2465_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 768, 128]> var_2465 = squeeze(axes = var_2465_axes_0, x = var_2464)[name = string("op_2465")];
             tensor<int32, [3]> var_2466 = const()[name = string("op_2466"), val = tensor<int32, [3]>([0, 2, 1])];
@@ -2053,7 +2058,7 @@ program(1.3)
             tensor<int32, [2]> var_2475_split_sizes_0 = const()[name = string("op_2475_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_2475_axis_0 = const()[name = string("op_2475_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_2475_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_2475_cast_fp16_1 = split(axis = var_2475_axis_0, split_sizes = var_2475_split_sizes_0, x = normed_277_cast_fp16)[name = string("op_2475_cast_fp16")];
-            tensor<fp16, [768]> var_2479_to_fp16 = const()[name = string("op_2479_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(598828992)))];
+            tensor<fp16, [768]> var_2479_to_fp16 = const()[name = string("op_2479_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300226816)))];
             tensor<fp16, [1, 128, 768]> out_119_cast_fp16 = mul(x = var_2475_cast_fp16_0, y = var_2479_to_fp16)[name = string("out_119_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_161_cast_fp16 = add(x = x_155_cast_fp16, y = out_119_cast_fp16)[name = string("x_161_cast_fp16")];
             fp16 const_140_promoted_to_fp16 = const()[name = string("const_140_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -2065,7 +2070,7 @@ program(1.3)
             tensor<int32, [2]> var_2513_split_sizes_0 = const()[name = string("op_2513_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_2513_axis_0 = const()[name = string("op_2513_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_2513_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_2513_cast_fp16_1 = split(axis = var_2513_axis_0, split_sizes = var_2513_split_sizes_0, x = normed_281_cast_fp16)[name = string("op_2513_cast_fp16")];
-            tensor<fp16, [768]> var_2517_to_fp16 = const()[name = string("op_2517_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(598830592)))];
+            tensor<fp16, [768]> var_2517_to_fp16 = const()[name = string("op_2517_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300228416)))];
             tensor<fp16, [1, 128, 768]> out_121_cast_fp16 = mul(x = var_2513_cast_fp16_0, y = var_2517_to_fp16)[name = string("out_121_cast_fp16")];
             tensor<int32, [3]> var_2523 = const()[name = string("op_2523"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> var_2525_axes_0 = const()[name = string("op_2525_axes_0"), val = tensor<int32, [1]>([2])];
@@ -2076,7 +2081,7 @@ program(1.3)
             tensor<int32, [4]> var_2532_pad_0 = const()[name = string("op_2532_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_2532_dilations_0 = const()[name = string("op_2532_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_2532_groups_0 = const()[name = string("op_2532_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_2532 = conv(dilations = var_2532_dilations_0, groups = var_2532_groups_0, pad = var_2532_pad_0, pad_type = var_2532_pad_type_0, strides = var_2532_strides_0, weight = encoder_layers_10_self_attn_q_proj_weight, x = var_2525_cast_fp16)[name = string("op_2532")];
+            tensor<fp16, [1, 768, 1, 128]> var_2532 = conv(dilations = var_2532_dilations_0, groups = var_2532_groups_0, pad = var_2532_pad_0, pad_type = var_2532_pad_type_0, strides = var_2532_strides_0, weight = encoder_layers_10_self_attn_q_proj_weight_quantized, x = var_2525_cast_fp16)[name = string("op_2532")];
             tensor<int32, [4]> var_2533 = const()[name = string("op_2533"), val = tensor<int32, [4]>([1, 3, 256, 128])];
             tensor<fp16, [1, 3, 256, 128]> var_2534 = reshape(shape = var_2533, x = var_2532)[name = string("op_2534")];
             tensor<int32, [4]> var_2535 = const()[name = string("op_2535"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -2085,7 +2090,7 @@ program(1.3)
             tensor<int32, [4]> var_2542_pad_0 = const()[name = string("op_2542_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_2542_dilations_0 = const()[name = string("op_2542_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_2542_groups_0 = const()[name = string("op_2542_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_2542 = conv(dilations = var_2542_dilations_0, groups = var_2542_groups_0, pad = var_2542_pad_0, pad_type = var_2542_pad_type_0, strides = var_2542_strides_0, weight = encoder_layers_10_self_attn_k_proj_weight, x = var_2525_cast_fp16)[name = string("op_2542")];
+            tensor<fp16, [1, 256, 1, 128]> var_2542 = conv(dilations = var_2542_dilations_0, groups = var_2542_groups_0, pad = var_2542_pad_0, pad_type = var_2542_pad_type_0, strides = var_2542_strides_0, weight = encoder_layers_10_self_attn_k_proj_weight_quantized, x = var_2525_cast_fp16)[name = string("op_2542")];
             tensor<int32, [4]> var_2543 = const()[name = string("op_2543"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_2544 = reshape(shape = var_2543, x = var_2542)[name = string("op_2544")];
             tensor<int32, [4]> var_2545 = const()[name = string("op_2545"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -2094,7 +2099,7 @@ program(1.3)
             tensor<int32, [4]> var_2552_pad_0 = const()[name = string("op_2552_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_2552_dilations_0 = const()[name = string("op_2552_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_2552_groups_0 = const()[name = string("op_2552_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_2552 = conv(dilations = var_2552_dilations_0, groups = var_2552_groups_0, pad = var_2552_pad_0, pad_type = var_2552_pad_type_0, strides = var_2552_strides_0, weight = encoder_layers_10_self_attn_v_proj_weight, x = var_2525_cast_fp16)[name = string("op_2552")];
+            tensor<fp16, [1, 256, 1, 128]> var_2552 = conv(dilations = var_2552_dilations_0, groups = var_2552_groups_0, pad = var_2552_pad_0, pad_type = var_2552_pad_type_0, strides = var_2552_strides_0, weight = encoder_layers_10_self_attn_v_proj_weight_quantized, x = var_2525_cast_fp16)[name = string("op_2552")];
             tensor<int32, [4]> var_2553 = const()[name = string("op_2553"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_2554 = reshape(shape = var_2553, x = var_2552)[name = string("op_2554")];
             tensor<int32, [4]> var_2555 = const()[name = string("op_2555"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -2108,7 +2113,7 @@ program(1.3)
             tensor<int32, [2]> var_2566_split_sizes_0 = const()[name = string("op_2566_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_2566_axis_0 = const()[name = string("op_2566_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 256]> var_2566_cast_fp16_0, tensor<fp16, [1, 3, 128, 256]> var_2566_cast_fp16_1 = split(axis = var_2566_axis_0, split_sizes = var_2566_split_sizes_0, x = normed_287_cast_fp16)[name = string("op_2566_cast_fp16")];
-            tensor<fp16, [256]> var_2570_to_fp16 = const()[name = string("op_2570_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(598832192)))];
+            tensor<fp16, [256]> var_2570_to_fp16 = const()[name = string("op_2570_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300230016)))];
             tensor<fp16, [1, 3, 128, 256]> out_123_cast_fp16 = mul(x = var_2566_cast_fp16_0, y = var_2570_to_fp16)[name = string("out_123_cast_fp16")];
             fp16 const_144_promoted_to_fp16 = const()[name = string("const_144_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 1, 128, 256]> k_61 = transpose(perm = var_2545, x = var_2544)[name = string("transpose_123")];
@@ -2120,9 +2125,9 @@ program(1.3)
             tensor<int32, [2]> var_2582_split_sizes_0 = const()[name = string("op_2582_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_2582_axis_0 = const()[name = string("op_2582_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 256]> var_2582_cast_fp16_0, tensor<fp16, [1, 1, 128, 256]> var_2582_cast_fp16_1 = split(axis = var_2582_axis_0, split_sizes = var_2582_split_sizes_0, x = normed_291_cast_fp16)[name = string("op_2582_cast_fp16")];
-            tensor<fp16, [256]> var_2586_to_fp16 = const()[name = string("op_2586_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(598832768)))];
+            tensor<fp16, [256]> var_2586_to_fp16 = const()[name = string("op_2586_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300230592)))];
             tensor<fp16, [1, 1, 128, 256]> out_125_cast_fp16 = mul(x = var_2582_cast_fp16_0, y = var_2586_to_fp16)[name = string("out_125_cast_fp16")];
-            tensor<fp16, [1, 3, 128, 256]> var_2589 = mul(x = out_123_cast_fp16, y = cos_1)[name = string("op_2589")];
+            tensor<fp16, [1, 3, 128, 256]> var_2589 = mul(x = out_123_cast_fp16, y = cos_1_quantized)[name = string("op_2589")];
             tensor<int32, [2]> var_2590_split_sizes_0 = const()[name = string("op_2590_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_2590_axis_0 = const()[name = string("op_2590_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 128]> var_2590_0, tensor<fp16, [1, 3, 128, 128]> var_2590_1 = split(axis = var_2590_axis_0, split_sizes = var_2590_split_sizes_0, x = out_123_cast_fp16)[name = string("op_2590")];
@@ -2130,9 +2135,9 @@ program(1.3)
             tensor<fp16, [1, 3, 128, 128]> var_2592 = mul(x = var_2590_1, y = const_146_promoted)[name = string("op_2592")];
             bool var_2594_interleave_0 = const()[name = string("op_2594_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 3, 128, 256]> var_2594 = concat(axis = var_23, interleave = var_2594_interleave_0, values = (var_2592, var_2590_0))[name = string("op_2594")];
-            tensor<fp16, [1, 3, 128, 256]> var_2595 = mul(x = var_2594, y = sin_1)[name = string("op_2595")];
+            tensor<fp16, [1, 3, 128, 256]> var_2595 = mul(x = var_2594, y = sin_1_quantized)[name = string("op_2595")];
             tensor<fp16, [1, 3, 128, 256]> q_65 = add(x = var_2589, y = var_2595)[name = string("q_65")];
-            tensor<fp16, [1, 1, 128, 256]> var_2597 = mul(x = out_125_cast_fp16, y = cos_1)[name = string("op_2597")];
+            tensor<fp16, [1, 1, 128, 256]> var_2597 = mul(x = out_125_cast_fp16, y = cos_1_quantized)[name = string("op_2597")];
             tensor<int32, [2]> var_2598_split_sizes_0 = const()[name = string("op_2598_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_2598_axis_0 = const()[name = string("op_2598_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 128]> var_2598_0, tensor<fp16, [1, 1, 128, 128]> var_2598_1 = split(axis = var_2598_axis_0, split_sizes = var_2598_split_sizes_0, x = out_125_cast_fp16)[name = string("op_2598")];
@@ -2140,7 +2145,7 @@ program(1.3)
             tensor<fp16, [1, 1, 128, 128]> var_2600 = mul(x = var_2598_1, y = const_147_promoted)[name = string("op_2600")];
             bool var_2602_interleave_0 = const()[name = string("op_2602_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 1, 128, 256]> var_2602 = concat(axis = var_23, interleave = var_2602_interleave_0, values = (var_2600, var_2598_0))[name = string("op_2602")];
-            tensor<fp16, [1, 1, 128, 256]> var_2603 = mul(x = var_2602, y = sin_1)[name = string("op_2603")];
+            tensor<fp16, [1, 1, 128, 256]> var_2603 = mul(x = var_2602, y = sin_1_quantized)[name = string("op_2603")];
             tensor<fp16, [1, 1, 128, 256]> hidden_states_121 = add(x = var_2597, y = var_2603)[name = string("hidden_states_121")];
             tensor<int32, [1]> hidden_states_123_axes_0 = const()[name = string("hidden_states_123_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 1, 1, 128, 256]> hidden_states_123 = expand_dims(axes = hidden_states_123_axes_0, x = hidden_states_121)[name = string("hidden_states_123")];
@@ -2170,14 +2175,14 @@ program(1.3)
             tensor<fp16, [1, 128, 3, 256]> var_2630 = transpose(perm = var_2629, x = var_2627_cast_fp16)[name = string("transpose_121")];
             tensor<fp16, [1, 128, 768]> attn_out_63 = reshape(shape = var_2632, x = var_2630)[name = string("attn_out_63")];
             tensor<int32, [3]> var_2634 = const()[name = string("op_2634"), val = tensor<int32, [3]>([0, 2, 1])];
-            tensor<fp16, [768, 768, 1]> squeeze_10 = const()[name = string("squeeze_10"), val = tensor<fp16, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(598833344)))];
+            tensor<fp16, [768, 768, 1]> squeeze_10_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300231168))), scale = tensor<fp16, [768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300821056))))[name = string("squeeze_10_quantized")];
             string var_2643_pad_type_0 = const()[name = string("op_2643_pad_type_0"), val = string("valid")];
             int32 var_2643_groups_0 = const()[name = string("op_2643_groups_0"), val = int32(1)];
             tensor<int32, [1]> var_2643_strides_0 = const()[name = string("op_2643_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> var_2643_pad_0 = const()[name = string("op_2643_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> var_2643_dilations_0 = const()[name = string("op_2643_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, [1, 768, 128]> var_2635 = transpose(perm = var_2634, x = attn_out_63)[name = string("transpose_120")];
-            tensor<fp16, [1, 768, 128]> var_2643 = conv(dilations = var_2643_dilations_0, groups = var_2643_groups_0, pad = var_2643_pad_0, pad_type = var_2643_pad_type_0, strides = var_2643_strides_0, weight = squeeze_10, x = var_2635)[name = string("op_2643")];
+            tensor<fp16, [1, 768, 128]> var_2643 = conv(dilations = var_2643_dilations_0, groups = var_2643_groups_0, pad = var_2643_pad_0, pad_type = var_2643_pad_type_0, strides = var_2643_strides_0, weight = squeeze_10_quantized, x = var_2635)[name = string("op_2643")];
             tensor<int32, [3]> var_2644 = const()[name = string("op_2644"), val = tensor<int32, [3]>([0, 2, 1])];
             fp16 const_148_promoted_to_fp16 = const()[name = string("const_148_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 128, 768]> x_169 = transpose(perm = var_2644, x = var_2643)[name = string("transpose_119")];
@@ -2189,7 +2194,7 @@ program(1.3)
             tensor<int32, [2]> var_2653_split_sizes_0 = const()[name = string("op_2653_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_2653_axis_0 = const()[name = string("op_2653_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_2653_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_2653_cast_fp16_1 = split(axis = var_2653_axis_0, split_sizes = var_2653_split_sizes_0, x = normed_295_cast_fp16)[name = string("op_2653_cast_fp16")];
-            tensor<fp16, [768]> var_2657_to_fp16 = const()[name = string("op_2657_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(600013056)))];
+            tensor<fp16, [768]> var_2657_to_fp16 = const()[name = string("op_2657_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300822656)))];
             tensor<fp16, [1, 128, 768]> out_127_cast_fp16 = mul(x = var_2653_cast_fp16_0, y = var_2657_to_fp16)[name = string("out_127_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_171_cast_fp16 = add(x = x_161_cast_fp16, y = out_127_cast_fp16)[name = string("x_171_cast_fp16")];
             fp16 const_150_promoted_to_fp16 = const()[name = string("const_150_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -2201,7 +2206,7 @@ program(1.3)
             tensor<int32, [2]> var_2669_split_sizes_0 = const()[name = string("op_2669_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_2669_axis_0 = const()[name = string("op_2669_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_2669_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_2669_cast_fp16_1 = split(axis = var_2669_axis_0, split_sizes = var_2669_split_sizes_0, x = normed_299_cast_fp16)[name = string("op_2669_cast_fp16")];
-            tensor<fp16, [768]> var_2673_to_fp16 = const()[name = string("op_2673_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(600014656)))];
+            tensor<fp16, [768]> var_2673_to_fp16 = const()[name = string("op_2673_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300824256)))];
             tensor<fp16, [1, 128, 768]> out_129_cast_fp16 = mul(x = var_2669_cast_fp16_0, y = var_2673_to_fp16)[name = string("out_129_cast_fp16")];
             tensor<int32, [3]> var_2680 = const()[name = string("op_2680"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> input_215_axes_0 = const()[name = string("input_215_axes_0"), val = tensor<int32, [1]>([2])];
@@ -2212,13 +2217,13 @@ program(1.3)
             tensor<int32, [4]> gate_41_pad_0 = const()[name = string("gate_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> gate_41_dilations_0 = const()[name = string("gate_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 gate_41_groups_0 = const()[name = string("gate_41_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> gate_41 = conv(dilations = gate_41_dilations_0, groups = gate_41_groups_0, pad = gate_41_pad_0, pad_type = gate_41_pad_type_0, strides = gate_41_strides_0, weight = encoder_layers_10_mlp_gate_proj_weight, x = input_215)[name = string("gate_41")];
+            tensor<fp16, [1, 1152, 1, 128]> gate_41 = conv(dilations = gate_41_dilations_0, groups = gate_41_groups_0, pad = gate_41_pad_0, pad_type = gate_41_pad_type_0, strides = gate_41_strides_0, weight = encoder_layers_10_mlp_gate_proj_weight_quantized, x = input_215)[name = string("gate_41")];
             string up_21_pad_type_0 = const()[name = string("up_21_pad_type_0"), val = string("valid")];
             tensor<int32, [2]> up_21_strides_0 = const()[name = string("up_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> up_21_pad_0 = const()[name = string("up_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> up_21_dilations_0 = const()[name = string("up_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 up_21_groups_0 = const()[name = string("up_21_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> up_21 = conv(dilations = up_21_dilations_0, groups = up_21_groups_0, pad = up_21_pad_0, pad_type = up_21_pad_type_0, strides = up_21_strides_0, weight = encoder_layers_10_mlp_up_proj_weight, x = input_215)[name = string("up_21")];
+            tensor<fp16, [1, 1152, 1, 128]> up_21 = conv(dilations = up_21_dilations_0, groups = up_21_groups_0, pad = up_21_pad_0, pad_type = up_21_pad_type_0, strides = up_21_strides_0, weight = encoder_layers_10_mlp_up_proj_weight_quantized, x = input_215)[name = string("up_21")];
             string gate_43_mode_0 = const()[name = string("gate_43_mode_0"), val = string("TANH_APPROXIMATION")];
             tensor<fp16, [1, 1152, 1, 128]> gate_43 = gelu(mode = gate_43_mode_0, x = gate_41)[name = string("gate_43")];
             tensor<fp16, [1, 1152, 1, 128]> input_217 = mul(x = gate_43, y = up_21)[name = string("input_217")];
@@ -2227,7 +2232,7 @@ program(1.3)
             tensor<int32, [4]> var_2702_pad_0 = const()[name = string("op_2702_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_2702_dilations_0 = const()[name = string("op_2702_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_2702_groups_0 = const()[name = string("op_2702_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_2702 = conv(dilations = var_2702_dilations_0, groups = var_2702_groups_0, pad = var_2702_pad_0, pad_type = var_2702_pad_type_0, strides = var_2702_strides_0, weight = encoder_layers_10_mlp_down_proj_weight, x = input_217)[name = string("op_2702")];
+            tensor<fp16, [1, 768, 1, 128]> var_2702 = conv(dilations = var_2702_dilations_0, groups = var_2702_groups_0, pad = var_2702_pad_0, pad_type = var_2702_pad_type_0, strides = var_2702_strides_0, weight = encoder_layers_10_mlp_down_proj_weight_quantized, x = input_217)[name = string("op_2702")];
             tensor<int32, [1]> var_2703_axes_0 = const()[name = string("op_2703_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 768, 128]> var_2703 = squeeze(axes = var_2703_axes_0, x = var_2702)[name = string("op_2703")];
             tensor<int32, [3]> var_2704 = const()[name = string("op_2704"), val = tensor<int32, [3]>([0, 2, 1])];
@@ -2241,7 +2246,7 @@ program(1.3)
             tensor<int32, [2]> var_2713_split_sizes_0 = const()[name = string("op_2713_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_2713_axis_0 = const()[name = string("op_2713_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_2713_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_2713_cast_fp16_1 = split(axis = var_2713_axis_0, split_sizes = var_2713_split_sizes_0, x = normed_305_cast_fp16)[name = string("op_2713_cast_fp16")];
-            tensor<fp16, [768]> var_2717_to_fp16 = const()[name = string("op_2717_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(600016256)))];
+            tensor<fp16, [768]> var_2717_to_fp16 = const()[name = string("op_2717_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300825856)))];
             tensor<fp16, [1, 128, 768]> out_131_cast_fp16 = mul(x = var_2713_cast_fp16_0, y = var_2717_to_fp16)[name = string("out_131_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_177_cast_fp16 = add(x = x_171_cast_fp16, y = out_131_cast_fp16)[name = string("x_177_cast_fp16")];
             fp16 const_154_promoted_to_fp16 = const()[name = string("const_154_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -2253,7 +2258,7 @@ program(1.3)
             tensor<int32, [2]> var_2751_split_sizes_0 = const()[name = string("op_2751_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_2751_axis_0 = const()[name = string("op_2751_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_2751_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_2751_cast_fp16_1 = split(axis = var_2751_axis_0, split_sizes = var_2751_split_sizes_0, x = normed_309_cast_fp16)[name = string("op_2751_cast_fp16")];
-            tensor<fp16, [768]> var_2755_to_fp16 = const()[name = string("op_2755_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(600017856)))];
+            tensor<fp16, [768]> var_2755_to_fp16 = const()[name = string("op_2755_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300827456)))];
             tensor<fp16, [1, 128, 768]> out_133_cast_fp16 = mul(x = var_2751_cast_fp16_0, y = var_2755_to_fp16)[name = string("out_133_cast_fp16")];
             tensor<int32, [3]> var_2761 = const()[name = string("op_2761"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> var_2763_axes_0 = const()[name = string("op_2763_axes_0"), val = tensor<int32, [1]>([2])];
@@ -2264,7 +2269,7 @@ program(1.3)
             tensor<int32, [4]> var_2770_pad_0 = const()[name = string("op_2770_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_2770_dilations_0 = const()[name = string("op_2770_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_2770_groups_0 = const()[name = string("op_2770_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_2770 = conv(dilations = var_2770_dilations_0, groups = var_2770_groups_0, pad = var_2770_pad_0, pad_type = var_2770_pad_type_0, strides = var_2770_strides_0, weight = encoder_layers_11_self_attn_q_proj_weight, x = var_2763_cast_fp16)[name = string("op_2770")];
+            tensor<fp16, [1, 768, 1, 128]> var_2770 = conv(dilations = var_2770_dilations_0, groups = var_2770_groups_0, pad = var_2770_pad_0, pad_type = var_2770_pad_type_0, strides = var_2770_strides_0, weight = encoder_layers_11_self_attn_q_proj_weight_quantized, x = var_2763_cast_fp16)[name = string("op_2770")];
             tensor<int32, [4]> var_2771 = const()[name = string("op_2771"), val = tensor<int32, [4]>([1, 3, 256, 128])];
             tensor<fp16, [1, 3, 256, 128]> var_2772 = reshape(shape = var_2771, x = var_2770)[name = string("op_2772")];
             tensor<int32, [4]> var_2773 = const()[name = string("op_2773"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -2273,7 +2278,7 @@ program(1.3)
             tensor<int32, [4]> var_2780_pad_0 = const()[name = string("op_2780_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_2780_dilations_0 = const()[name = string("op_2780_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_2780_groups_0 = const()[name = string("op_2780_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_2780 = conv(dilations = var_2780_dilations_0, groups = var_2780_groups_0, pad = var_2780_pad_0, pad_type = var_2780_pad_type_0, strides = var_2780_strides_0, weight = encoder_layers_11_self_attn_k_proj_weight, x = var_2763_cast_fp16)[name = string("op_2780")];
+            tensor<fp16, [1, 256, 1, 128]> var_2780 = conv(dilations = var_2780_dilations_0, groups = var_2780_groups_0, pad = var_2780_pad_0, pad_type = var_2780_pad_type_0, strides = var_2780_strides_0, weight = encoder_layers_11_self_attn_k_proj_weight_quantized, x = var_2763_cast_fp16)[name = string("op_2780")];
             tensor<int32, [4]> var_2781 = const()[name = string("op_2781"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_2782 = reshape(shape = var_2781, x = var_2780)[name = string("op_2782")];
             tensor<int32, [4]> var_2783 = const()[name = string("op_2783"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -2282,7 +2287,7 @@ program(1.3)
             tensor<int32, [4]> var_2790_pad_0 = const()[name = string("op_2790_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_2790_dilations_0 = const()[name = string("op_2790_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_2790_groups_0 = const()[name = string("op_2790_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_2790 = conv(dilations = var_2790_dilations_0, groups = var_2790_groups_0, pad = var_2790_pad_0, pad_type = var_2790_pad_type_0, strides = var_2790_strides_0, weight = encoder_layers_11_self_attn_v_proj_weight, x = var_2763_cast_fp16)[name = string("op_2790")];
+            tensor<fp16, [1, 256, 1, 128]> var_2790 = conv(dilations = var_2790_dilations_0, groups = var_2790_groups_0, pad = var_2790_pad_0, pad_type = var_2790_pad_type_0, strides = var_2790_strides_0, weight = encoder_layers_11_self_attn_v_proj_weight_quantized, x = var_2763_cast_fp16)[name = string("op_2790")];
             tensor<int32, [4]> var_2791 = const()[name = string("op_2791"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_2792 = reshape(shape = var_2791, x = var_2790)[name = string("op_2792")];
             tensor<int32, [4]> var_2793 = const()[name = string("op_2793"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -2296,7 +2301,7 @@ program(1.3)
             tensor<int32, [2]> var_2804_split_sizes_0 = const()[name = string("op_2804_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_2804_axis_0 = const()[name = string("op_2804_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 256]> var_2804_cast_fp16_0, tensor<fp16, [1, 3, 128, 256]> var_2804_cast_fp16_1 = split(axis = var_2804_axis_0, split_sizes = var_2804_split_sizes_0, x = normed_315_cast_fp16)[name = string("op_2804_cast_fp16")];
-            tensor<fp16, [256]> var_2808_to_fp16 = const()[name = string("op_2808_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(600019456)))];
+            tensor<fp16, [256]> var_2808_to_fp16 = const()[name = string("op_2808_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300829056)))];
             tensor<fp16, [1, 3, 128, 256]> out_135_cast_fp16 = mul(x = var_2804_cast_fp16_0, y = var_2808_to_fp16)[name = string("out_135_cast_fp16")];
             fp16 const_158_promoted_to_fp16 = const()[name = string("const_158_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 1, 128, 256]> k_67 = transpose(perm = var_2783, x = var_2782)[name = string("transpose_114")];
@@ -2308,9 +2313,9 @@ program(1.3)
             tensor<int32, [2]> var_2820_split_sizes_0 = const()[name = string("op_2820_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_2820_axis_0 = const()[name = string("op_2820_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 256]> var_2820_cast_fp16_0, tensor<fp16, [1, 1, 128, 256]> var_2820_cast_fp16_1 = split(axis = var_2820_axis_0, split_sizes = var_2820_split_sizes_0, x = normed_319_cast_fp16)[name = string("op_2820_cast_fp16")];
-            tensor<fp16, [256]> var_2824_to_fp16 = const()[name = string("op_2824_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(600020032)))];
+            tensor<fp16, [256]> var_2824_to_fp16 = const()[name = string("op_2824_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300829632)))];
             tensor<fp16, [1, 1, 128, 256]> out_137_cast_fp16 = mul(x = var_2820_cast_fp16_0, y = var_2824_to_fp16)[name = string("out_137_cast_fp16")];
-            tensor<fp16, [1, 3, 128, 256]> var_2827 = mul(x = out_135_cast_fp16, y = cos)[name = string("op_2827")];
+            tensor<fp16, [1, 3, 128, 256]> var_2827 = mul(x = out_135_cast_fp16, y = cos_quantized)[name = string("op_2827")];
             tensor<int32, [2]> var_2828_split_sizes_0 = const()[name = string("op_2828_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_2828_axis_0 = const()[name = string("op_2828_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 128]> var_2828_0, tensor<fp16, [1, 3, 128, 128]> var_2828_1 = split(axis = var_2828_axis_0, split_sizes = var_2828_split_sizes_0, x = out_135_cast_fp16)[name = string("op_2828")];
@@ -2318,9 +2323,9 @@ program(1.3)
             tensor<fp16, [1, 3, 128, 128]> var_2830 = mul(x = var_2828_1, y = const_160_promoted)[name = string("op_2830")];
             bool var_2832_interleave_0 = const()[name = string("op_2832_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 3, 128, 256]> var_2832 = concat(axis = var_23, interleave = var_2832_interleave_0, values = (var_2830, var_2828_0))[name = string("op_2832")];
-            tensor<fp16, [1, 3, 128, 256]> var_2833 = mul(x = var_2832, y = sin)[name = string("op_2833")];
+            tensor<fp16, [1, 3, 128, 256]> var_2833 = mul(x = var_2832, y = sin_quantized)[name = string("op_2833")];
             tensor<fp16, [1, 3, 128, 256]> q_71 = add(x = var_2827, y = var_2833)[name = string("q_71")];
-            tensor<fp16, [1, 1, 128, 256]> var_2835 = mul(x = out_137_cast_fp16, y = cos)[name = string("op_2835")];
+            tensor<fp16, [1, 1, 128, 256]> var_2835 = mul(x = out_137_cast_fp16, y = cos_quantized)[name = string("op_2835")];
             tensor<int32, [2]> var_2836_split_sizes_0 = const()[name = string("op_2836_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_2836_axis_0 = const()[name = string("op_2836_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 128]> var_2836_0, tensor<fp16, [1, 1, 128, 128]> var_2836_1 = split(axis = var_2836_axis_0, split_sizes = var_2836_split_sizes_0, x = out_137_cast_fp16)[name = string("op_2836")];
@@ -2328,7 +2333,7 @@ program(1.3)
             tensor<fp16, [1, 1, 128, 128]> var_2838 = mul(x = var_2836_1, y = const_161_promoted)[name = string("op_2838")];
             bool var_2840_interleave_0 = const()[name = string("op_2840_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 1, 128, 256]> var_2840 = concat(axis = var_23, interleave = var_2840_interleave_0, values = (var_2838, var_2836_0))[name = string("op_2840")];
-            tensor<fp16, [1, 1, 128, 256]> var_2841 = mul(x = var_2840, y = sin)[name = string("op_2841")];
+            tensor<fp16, [1, 1, 128, 256]> var_2841 = mul(x = var_2840, y = sin_quantized)[name = string("op_2841")];
             tensor<fp16, [1, 1, 128, 256]> hidden_states_133 = add(x = var_2835, y = var_2841)[name = string("hidden_states_133")];
             tensor<int32, [1]> hidden_states_135_axes_0 = const()[name = string("hidden_states_135_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 1, 1, 128, 256]> hidden_states_135 = expand_dims(axes = hidden_states_135_axes_0, x = hidden_states_133)[name = string("hidden_states_135")];
@@ -2358,14 +2363,14 @@ program(1.3)
             tensor<fp16, [1, 128, 3, 256]> var_2868 = transpose(perm = var_2867, x = var_2865_cast_fp16)[name = string("transpose_112")];
             tensor<fp16, [1, 128, 768]> attn_out_69 = reshape(shape = var_2870, x = var_2868)[name = string("attn_out_69")];
             tensor<int32, [3]> var_2872 = const()[name = string("op_2872"), val = tensor<int32, [3]>([0, 2, 1])];
-            tensor<fp16, [768, 768, 1]> squeeze_11 = const()[name = string("squeeze_11"), val = tensor<fp16, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(600020608)))];
+            tensor<fp16, [768, 768, 1]> squeeze_11_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300830208))), scale = tensor<fp16, [768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301420096))))[name = string("squeeze_11_quantized")];
             string var_2881_pad_type_0 = const()[name = string("op_2881_pad_type_0"), val = string("valid")];
             int32 var_2881_groups_0 = const()[name = string("op_2881_groups_0"), val = int32(1)];
             tensor<int32, [1]> var_2881_strides_0 = const()[name = string("op_2881_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> var_2881_pad_0 = const()[name = string("op_2881_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> var_2881_dilations_0 = const()[name = string("op_2881_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, [1, 768, 128]> var_2873 = transpose(perm = var_2872, x = attn_out_69)[name = string("transpose_111")];
-            tensor<fp16, [1, 768, 128]> var_2881 = conv(dilations = var_2881_dilations_0, groups = var_2881_groups_0, pad = var_2881_pad_0, pad_type = var_2881_pad_type_0, strides = var_2881_strides_0, weight = squeeze_11, x = var_2873)[name = string("op_2881")];
+            tensor<fp16, [1, 768, 128]> var_2881 = conv(dilations = var_2881_dilations_0, groups = var_2881_groups_0, pad = var_2881_pad_0, pad_type = var_2881_pad_type_0, strides = var_2881_strides_0, weight = squeeze_11_quantized, x = var_2873)[name = string("op_2881")];
             tensor<int32, [3]> var_2882 = const()[name = string("op_2882"), val = tensor<int32, [3]>([0, 2, 1])];
             fp16 const_162_promoted_to_fp16 = const()[name = string("const_162_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 128, 768]> x_185 = transpose(perm = var_2882, x = var_2881)[name = string("transpose_110")];
@@ -2377,7 +2382,7 @@ program(1.3)
             tensor<int32, [2]> var_2891_split_sizes_0 = const()[name = string("op_2891_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_2891_axis_0 = const()[name = string("op_2891_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_2891_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_2891_cast_fp16_1 = split(axis = var_2891_axis_0, split_sizes = var_2891_split_sizes_0, x = normed_323_cast_fp16)[name = string("op_2891_cast_fp16")];
-            tensor<fp16, [768]> var_2895_to_fp16 = const()[name = string("op_2895_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601200320)))];
+            tensor<fp16, [768]> var_2895_to_fp16 = const()[name = string("op_2895_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301421696)))];
             tensor<fp16, [1, 128, 768]> out_139_cast_fp16 = mul(x = var_2891_cast_fp16_0, y = var_2895_to_fp16)[name = string("out_139_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_187_cast_fp16 = add(x = x_177_cast_fp16, y = out_139_cast_fp16)[name = string("x_187_cast_fp16")];
             fp16 const_164_promoted_to_fp16 = const()[name = string("const_164_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -2389,7 +2394,7 @@ program(1.3)
             tensor<int32, [2]> var_2907_split_sizes_0 = const()[name = string("op_2907_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_2907_axis_0 = const()[name = string("op_2907_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_2907_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_2907_cast_fp16_1 = split(axis = var_2907_axis_0, split_sizes = var_2907_split_sizes_0, x = normed_327_cast_fp16)[name = string("op_2907_cast_fp16")];
-            tensor<fp16, [768]> var_2911_to_fp16 = const()[name = string("op_2911_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601201920)))];
+            tensor<fp16, [768]> var_2911_to_fp16 = const()[name = string("op_2911_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301423296)))];
             tensor<fp16, [1, 128, 768]> out_141_cast_fp16 = mul(x = var_2907_cast_fp16_0, y = var_2911_to_fp16)[name = string("out_141_cast_fp16")];
             tensor<int32, [3]> var_2918 = const()[name = string("op_2918"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> input_235_axes_0 = const()[name = string("input_235_axes_0"), val = tensor<int32, [1]>([2])];
@@ -2400,13 +2405,13 @@ program(1.3)
             tensor<int32, [4]> gate_45_pad_0 = const()[name = string("gate_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> gate_45_dilations_0 = const()[name = string("gate_45_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 gate_45_groups_0 = const()[name = string("gate_45_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> gate_45 = conv(dilations = gate_45_dilations_0, groups = gate_45_groups_0, pad = gate_45_pad_0, pad_type = gate_45_pad_type_0, strides = gate_45_strides_0, weight = encoder_layers_11_mlp_gate_proj_weight, x = input_235)[name = string("gate_45")];
+            tensor<fp16, [1, 1152, 1, 128]> gate_45 = conv(dilations = gate_45_dilations_0, groups = gate_45_groups_0, pad = gate_45_pad_0, pad_type = gate_45_pad_type_0, strides = gate_45_strides_0, weight = encoder_layers_11_mlp_gate_proj_weight_quantized, x = input_235)[name = string("gate_45")];
             string up_23_pad_type_0 = const()[name = string("up_23_pad_type_0"), val = string("valid")];
             tensor<int32, [2]> up_23_strides_0 = const()[name = string("up_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> up_23_pad_0 = const()[name = string("up_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> up_23_dilations_0 = const()[name = string("up_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 up_23_groups_0 = const()[name = string("up_23_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> up_23 = conv(dilations = up_23_dilations_0, groups = up_23_groups_0, pad = up_23_pad_0, pad_type = up_23_pad_type_0, strides = up_23_strides_0, weight = encoder_layers_11_mlp_up_proj_weight, x = input_235)[name = string("up_23")];
+            tensor<fp16, [1, 1152, 1, 128]> up_23 = conv(dilations = up_23_dilations_0, groups = up_23_groups_0, pad = up_23_pad_0, pad_type = up_23_pad_type_0, strides = up_23_strides_0, weight = encoder_layers_11_mlp_up_proj_weight_quantized, x = input_235)[name = string("up_23")];
             string gate_47_mode_0 = const()[name = string("gate_47_mode_0"), val = string("TANH_APPROXIMATION")];
             tensor<fp16, [1, 1152, 1, 128]> gate_47 = gelu(mode = gate_47_mode_0, x = gate_45)[name = string("gate_47")];
             tensor<fp16, [1, 1152, 1, 128]> input_237 = mul(x = gate_47, y = up_23)[name = string("input_237")];
@@ -2415,7 +2420,7 @@ program(1.3)
             tensor<int32, [4]> var_2940_pad_0 = const()[name = string("op_2940_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_2940_dilations_0 = const()[name = string("op_2940_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_2940_groups_0 = const()[name = string("op_2940_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_2940 = conv(dilations = var_2940_dilations_0, groups = var_2940_groups_0, pad = var_2940_pad_0, pad_type = var_2940_pad_type_0, strides = var_2940_strides_0, weight = encoder_layers_11_mlp_down_proj_weight, x = input_237)[name = string("op_2940")];
+            tensor<fp16, [1, 768, 1, 128]> var_2940 = conv(dilations = var_2940_dilations_0, groups = var_2940_groups_0, pad = var_2940_pad_0, pad_type = var_2940_pad_type_0, strides = var_2940_strides_0, weight = encoder_layers_11_mlp_down_proj_weight_quantized, x = input_237)[name = string("op_2940")];
             tensor<int32, [1]> var_2941_axes_0 = const()[name = string("op_2941_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 768, 128]> var_2941 = squeeze(axes = var_2941_axes_0, x = var_2940)[name = string("op_2941")];
             tensor<int32, [3]> var_2942 = const()[name = string("op_2942"), val = tensor<int32, [3]>([0, 2, 1])];
@@ -2429,7 +2434,7 @@ program(1.3)
             tensor<int32, [2]> var_2951_split_sizes_0 = const()[name = string("op_2951_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_2951_axis_0 = const()[name = string("op_2951_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_2951_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_2951_cast_fp16_1 = split(axis = var_2951_axis_0, split_sizes = var_2951_split_sizes_0, x = normed_333_cast_fp16)[name = string("op_2951_cast_fp16")];
-            tensor<fp16, [768]> var_2955_to_fp16 = const()[name = string("op_2955_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601203520)))];
+            tensor<fp16, [768]> var_2955_to_fp16 = const()[name = string("op_2955_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301424896)))];
             tensor<fp16, [1, 128, 768]> out_143_cast_fp16 = mul(x = var_2951_cast_fp16_0, y = var_2955_to_fp16)[name = string("out_143_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_193_cast_fp16 = add(x = x_187_cast_fp16, y = out_143_cast_fp16)[name = string("x_193_cast_fp16")];
             fp16 const_168_promoted_to_fp16 = const()[name = string("const_168_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -2441,7 +2446,7 @@ program(1.3)
             tensor<int32, [2]> var_2989_split_sizes_0 = const()[name = string("op_2989_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_2989_axis_0 = const()[name = string("op_2989_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_2989_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_2989_cast_fp16_1 = split(axis = var_2989_axis_0, split_sizes = var_2989_split_sizes_0, x = normed_337_cast_fp16)[name = string("op_2989_cast_fp16")];
-            tensor<fp16, [768]> var_2993_to_fp16 = const()[name = string("op_2993_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601205120)))];
+            tensor<fp16, [768]> var_2993_to_fp16 = const()[name = string("op_2993_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301426496)))];
             tensor<fp16, [1, 128, 768]> out_145_cast_fp16 = mul(x = var_2989_cast_fp16_0, y = var_2993_to_fp16)[name = string("out_145_cast_fp16")];
             tensor<int32, [3]> var_2999 = const()[name = string("op_2999"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> var_3001_axes_0 = const()[name = string("op_3001_axes_0"), val = tensor<int32, [1]>([2])];
@@ -2452,7 +2457,7 @@ program(1.3)
             tensor<int32, [4]> var_3008_pad_0 = const()[name = string("op_3008_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_3008_dilations_0 = const()[name = string("op_3008_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_3008_groups_0 = const()[name = string("op_3008_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_3008 = conv(dilations = var_3008_dilations_0, groups = var_3008_groups_0, pad = var_3008_pad_0, pad_type = var_3008_pad_type_0, strides = var_3008_strides_0, weight = encoder_layers_12_self_attn_q_proj_weight, x = var_3001_cast_fp16)[name = string("op_3008")];
+            tensor<fp16, [1, 768, 1, 128]> var_3008 = conv(dilations = var_3008_dilations_0, groups = var_3008_groups_0, pad = var_3008_pad_0, pad_type = var_3008_pad_type_0, strides = var_3008_strides_0, weight = encoder_layers_12_self_attn_q_proj_weight_quantized, x = var_3001_cast_fp16)[name = string("op_3008")];
             tensor<int32, [4]> var_3009 = const()[name = string("op_3009"), val = tensor<int32, [4]>([1, 3, 256, 128])];
             tensor<fp16, [1, 3, 256, 128]> var_3010 = reshape(shape = var_3009, x = var_3008)[name = string("op_3010")];
             tensor<int32, [4]> var_3011 = const()[name = string("op_3011"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -2461,7 +2466,7 @@ program(1.3)
             tensor<int32, [4]> var_3018_pad_0 = const()[name = string("op_3018_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_3018_dilations_0 = const()[name = string("op_3018_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_3018_groups_0 = const()[name = string("op_3018_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_3018 = conv(dilations = var_3018_dilations_0, groups = var_3018_groups_0, pad = var_3018_pad_0, pad_type = var_3018_pad_type_0, strides = var_3018_strides_0, weight = encoder_layers_12_self_attn_k_proj_weight, x = var_3001_cast_fp16)[name = string("op_3018")];
+            tensor<fp16, [1, 256, 1, 128]> var_3018 = conv(dilations = var_3018_dilations_0, groups = var_3018_groups_0, pad = var_3018_pad_0, pad_type = var_3018_pad_type_0, strides = var_3018_strides_0, weight = encoder_layers_12_self_attn_k_proj_weight_quantized, x = var_3001_cast_fp16)[name = string("op_3018")];
             tensor<int32, [4]> var_3019 = const()[name = string("op_3019"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_3020 = reshape(shape = var_3019, x = var_3018)[name = string("op_3020")];
             tensor<int32, [4]> var_3021 = const()[name = string("op_3021"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -2470,7 +2475,7 @@ program(1.3)
             tensor<int32, [4]> var_3028_pad_0 = const()[name = string("op_3028_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_3028_dilations_0 = const()[name = string("op_3028_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_3028_groups_0 = const()[name = string("op_3028_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_3028 = conv(dilations = var_3028_dilations_0, groups = var_3028_groups_0, pad = var_3028_pad_0, pad_type = var_3028_pad_type_0, strides = var_3028_strides_0, weight = encoder_layers_12_self_attn_v_proj_weight, x = var_3001_cast_fp16)[name = string("op_3028")];
+            tensor<fp16, [1, 256, 1, 128]> var_3028 = conv(dilations = var_3028_dilations_0, groups = var_3028_groups_0, pad = var_3028_pad_0, pad_type = var_3028_pad_type_0, strides = var_3028_strides_0, weight = encoder_layers_12_self_attn_v_proj_weight_quantized, x = var_3001_cast_fp16)[name = string("op_3028")];
             tensor<int32, [4]> var_3029 = const()[name = string("op_3029"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_3030 = reshape(shape = var_3029, x = var_3028)[name = string("op_3030")];
             tensor<int32, [4]> var_3031 = const()[name = string("op_3031"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -2484,7 +2489,7 @@ program(1.3)
             tensor<int32, [2]> var_3042_split_sizes_0 = const()[name = string("op_3042_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_3042_axis_0 = const()[name = string("op_3042_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 256]> var_3042_cast_fp16_0, tensor<fp16, [1, 3, 128, 256]> var_3042_cast_fp16_1 = split(axis = var_3042_axis_0, split_sizes = var_3042_split_sizes_0, x = normed_343_cast_fp16)[name = string("op_3042_cast_fp16")];
-            tensor<fp16, [256]> var_3046_to_fp16 = const()[name = string("op_3046_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601206720)))];
+            tensor<fp16, [256]> var_3046_to_fp16 = const()[name = string("op_3046_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301428096)))];
             tensor<fp16, [1, 3, 128, 256]> out_147_cast_fp16 = mul(x = var_3042_cast_fp16_0, y = var_3046_to_fp16)[name = string("out_147_cast_fp16")];
             fp16 const_172_promoted_to_fp16 = const()[name = string("const_172_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 1, 128, 256]> k_73 = transpose(perm = var_3021, x = var_3020)[name = string("transpose_105")];
@@ -2496,9 +2501,9 @@ program(1.3)
             tensor<int32, [2]> var_3058_split_sizes_0 = const()[name = string("op_3058_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_3058_axis_0 = const()[name = string("op_3058_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 256]> var_3058_cast_fp16_0, tensor<fp16, [1, 1, 128, 256]> var_3058_cast_fp16_1 = split(axis = var_3058_axis_0, split_sizes = var_3058_split_sizes_0, x = normed_347_cast_fp16)[name = string("op_3058_cast_fp16")];
-            tensor<fp16, [256]> var_3062_to_fp16 = const()[name = string("op_3062_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601207296)))];
+            tensor<fp16, [256]> var_3062_to_fp16 = const()[name = string("op_3062_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301428672)))];
             tensor<fp16, [1, 1, 128, 256]> out_149_cast_fp16 = mul(x = var_3058_cast_fp16_0, y = var_3062_to_fp16)[name = string("out_149_cast_fp16")];
-            tensor<fp16, [1, 3, 128, 256]> var_3065 = mul(x = out_147_cast_fp16, y = cos_1)[name = string("op_3065")];
+            tensor<fp16, [1, 3, 128, 256]> var_3065 = mul(x = out_147_cast_fp16, y = cos_1_quantized)[name = string("op_3065")];
             tensor<int32, [2]> var_3066_split_sizes_0 = const()[name = string("op_3066_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_3066_axis_0 = const()[name = string("op_3066_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 128]> var_3066_0, tensor<fp16, [1, 3, 128, 128]> var_3066_1 = split(axis = var_3066_axis_0, split_sizes = var_3066_split_sizes_0, x = out_147_cast_fp16)[name = string("op_3066")];
@@ -2506,9 +2511,9 @@ program(1.3)
             tensor<fp16, [1, 3, 128, 128]> var_3068 = mul(x = var_3066_1, y = const_174_promoted)[name = string("op_3068")];
             bool var_3070_interleave_0 = const()[name = string("op_3070_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 3, 128, 256]> var_3070 = concat(axis = var_23, interleave = var_3070_interleave_0, values = (var_3068, var_3066_0))[name = string("op_3070")];
-            tensor<fp16, [1, 3, 128, 256]> var_3071 = mul(x = var_3070, y = sin_1)[name = string("op_3071")];
+            tensor<fp16, [1, 3, 128, 256]> var_3071 = mul(x = var_3070, y = sin_1_quantized)[name = string("op_3071")];
             tensor<fp16, [1, 3, 128, 256]> q_77 = add(x = var_3065, y = var_3071)[name = string("q_77")];
-            tensor<fp16, [1, 1, 128, 256]> var_3073 = mul(x = out_149_cast_fp16, y = cos_1)[name = string("op_3073")];
+            tensor<fp16, [1, 1, 128, 256]> var_3073 = mul(x = out_149_cast_fp16, y = cos_1_quantized)[name = string("op_3073")];
             tensor<int32, [2]> var_3074_split_sizes_0 = const()[name = string("op_3074_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_3074_axis_0 = const()[name = string("op_3074_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 128]> var_3074_0, tensor<fp16, [1, 1, 128, 128]> var_3074_1 = split(axis = var_3074_axis_0, split_sizes = var_3074_split_sizes_0, x = out_149_cast_fp16)[name = string("op_3074")];
@@ -2516,7 +2521,7 @@ program(1.3)
             tensor<fp16, [1, 1, 128, 128]> var_3076 = mul(x = var_3074_1, y = const_175_promoted)[name = string("op_3076")];
             bool var_3078_interleave_0 = const()[name = string("op_3078_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 1, 128, 256]> var_3078 = concat(axis = var_23, interleave = var_3078_interleave_0, values = (var_3076, var_3074_0))[name = string("op_3078")];
-            tensor<fp16, [1, 1, 128, 256]> var_3079 = mul(x = var_3078, y = sin_1)[name = string("op_3079")];
+            tensor<fp16, [1, 1, 128, 256]> var_3079 = mul(x = var_3078, y = sin_1_quantized)[name = string("op_3079")];
             tensor<fp16, [1, 1, 128, 256]> hidden_states_145 = add(x = var_3073, y = var_3079)[name = string("hidden_states_145")];
             tensor<int32, [1]> hidden_states_147_axes_0 = const()[name = string("hidden_states_147_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 1, 1, 128, 256]> hidden_states_147 = expand_dims(axes = hidden_states_147_axes_0, x = hidden_states_145)[name = string("hidden_states_147")];
@@ -2546,14 +2551,14 @@ program(1.3)
             tensor<fp16, [1, 128, 3, 256]> var_3106 = transpose(perm = var_3105, x = var_3103_cast_fp16)[name = string("transpose_103")];
             tensor<fp16, [1, 128, 768]> attn_out_75 = reshape(shape = var_3108, x = var_3106)[name = string("attn_out_75")];
             tensor<int32, [3]> var_3110 = const()[name = string("op_3110"), val = tensor<int32, [3]>([0, 2, 1])];
-            tensor<fp16, [768, 768, 1]> squeeze_12 = const()[name = string("squeeze_12"), val = tensor<fp16, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601207872)))];
+            tensor<fp16, [768, 768, 1]> squeeze_12_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301429248))), scale = tensor<fp16, [768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302019136))))[name = string("squeeze_12_quantized")];
             string var_3119_pad_type_0 = const()[name = string("op_3119_pad_type_0"), val = string("valid")];
             int32 var_3119_groups_0 = const()[name = string("op_3119_groups_0"), val = int32(1)];
             tensor<int32, [1]> var_3119_strides_0 = const()[name = string("op_3119_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> var_3119_pad_0 = const()[name = string("op_3119_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> var_3119_dilations_0 = const()[name = string("op_3119_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, [1, 768, 128]> var_3111 = transpose(perm = var_3110, x = attn_out_75)[name = string("transpose_102")];
-            tensor<fp16, [1, 768, 128]> var_3119 = conv(dilations = var_3119_dilations_0, groups = var_3119_groups_0, pad = var_3119_pad_0, pad_type = var_3119_pad_type_0, strides = var_3119_strides_0, weight = squeeze_12, x = var_3111)[name = string("op_3119")];
+            tensor<fp16, [1, 768, 128]> var_3119 = conv(dilations = var_3119_dilations_0, groups = var_3119_groups_0, pad = var_3119_pad_0, pad_type = var_3119_pad_type_0, strides = var_3119_strides_0, weight = squeeze_12_quantized, x = var_3111)[name = string("op_3119")];
             tensor<int32, [3]> var_3120 = const()[name = string("op_3120"), val = tensor<int32, [3]>([0, 2, 1])];
             fp16 const_176_promoted_to_fp16 = const()[name = string("const_176_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 128, 768]> x_201 = transpose(perm = var_3120, x = var_3119)[name = string("transpose_101")];
@@ -2565,7 +2570,7 @@ program(1.3)
             tensor<int32, [2]> var_3129_split_sizes_0 = const()[name = string("op_3129_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_3129_axis_0 = const()[name = string("op_3129_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_3129_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_3129_cast_fp16_1 = split(axis = var_3129_axis_0, split_sizes = var_3129_split_sizes_0, x = normed_351_cast_fp16)[name = string("op_3129_cast_fp16")];
-            tensor<fp16, [768]> var_3133_to_fp16 = const()[name = string("op_3133_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(602387584)))];
+            tensor<fp16, [768]> var_3133_to_fp16 = const()[name = string("op_3133_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302020736)))];
             tensor<fp16, [1, 128, 768]> out_151_cast_fp16 = mul(x = var_3129_cast_fp16_0, y = var_3133_to_fp16)[name = string("out_151_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_203_cast_fp16 = add(x = x_193_cast_fp16, y = out_151_cast_fp16)[name = string("x_203_cast_fp16")];
             fp16 const_178_promoted_to_fp16 = const()[name = string("const_178_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -2577,7 +2582,7 @@ program(1.3)
             tensor<int32, [2]> var_3145_split_sizes_0 = const()[name = string("op_3145_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_3145_axis_0 = const()[name = string("op_3145_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_3145_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_3145_cast_fp16_1 = split(axis = var_3145_axis_0, split_sizes = var_3145_split_sizes_0, x = normed_355_cast_fp16)[name = string("op_3145_cast_fp16")];
-            tensor<fp16, [768]> var_3149_to_fp16 = const()[name = string("op_3149_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(602389184)))];
+            tensor<fp16, [768]> var_3149_to_fp16 = const()[name = string("op_3149_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302022336)))];
             tensor<fp16, [1, 128, 768]> out_153_cast_fp16 = mul(x = var_3145_cast_fp16_0, y = var_3149_to_fp16)[name = string("out_153_cast_fp16")];
             tensor<int32, [3]> var_3156 = const()[name = string("op_3156"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> input_255_axes_0 = const()[name = string("input_255_axes_0"), val = tensor<int32, [1]>([2])];
@@ -2588,13 +2593,13 @@ program(1.3)
             tensor<int32, [4]> gate_49_pad_0 = const()[name = string("gate_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> gate_49_dilations_0 = const()[name = string("gate_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 gate_49_groups_0 = const()[name = string("gate_49_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> gate_49 = conv(dilations = gate_49_dilations_0, groups = gate_49_groups_0, pad = gate_49_pad_0, pad_type = gate_49_pad_type_0, strides = gate_49_strides_0, weight = encoder_layers_12_mlp_gate_proj_weight, x = input_255)[name = string("gate_49")];
+            tensor<fp16, [1, 1152, 1, 128]> gate_49 = conv(dilations = gate_49_dilations_0, groups = gate_49_groups_0, pad = gate_49_pad_0, pad_type = gate_49_pad_type_0, strides = gate_49_strides_0, weight = encoder_layers_12_mlp_gate_proj_weight_quantized, x = input_255)[name = string("gate_49")];
             string up_25_pad_type_0 = const()[name = string("up_25_pad_type_0"), val = string("valid")];
             tensor<int32, [2]> up_25_strides_0 = const()[name = string("up_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> up_25_pad_0 = const()[name = string("up_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> up_25_dilations_0 = const()[name = string("up_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 up_25_groups_0 = const()[name = string("up_25_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> up_25 = conv(dilations = up_25_dilations_0, groups = up_25_groups_0, pad = up_25_pad_0, pad_type = up_25_pad_type_0, strides = up_25_strides_0, weight = encoder_layers_12_mlp_up_proj_weight, x = input_255)[name = string("up_25")];
+            tensor<fp16, [1, 1152, 1, 128]> up_25 = conv(dilations = up_25_dilations_0, groups = up_25_groups_0, pad = up_25_pad_0, pad_type = up_25_pad_type_0, strides = up_25_strides_0, weight = encoder_layers_12_mlp_up_proj_weight_quantized, x = input_255)[name = string("up_25")];
             string gate_51_mode_0 = const()[name = string("gate_51_mode_0"), val = string("TANH_APPROXIMATION")];
             tensor<fp16, [1, 1152, 1, 128]> gate_51 = gelu(mode = gate_51_mode_0, x = gate_49)[name = string("gate_51")];
             tensor<fp16, [1, 1152, 1, 128]> input_257 = mul(x = gate_51, y = up_25)[name = string("input_257")];
@@ -2603,7 +2608,7 @@ program(1.3)
             tensor<int32, [4]> var_3178_pad_0 = const()[name = string("op_3178_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_3178_dilations_0 = const()[name = string("op_3178_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_3178_groups_0 = const()[name = string("op_3178_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_3178 = conv(dilations = var_3178_dilations_0, groups = var_3178_groups_0, pad = var_3178_pad_0, pad_type = var_3178_pad_type_0, strides = var_3178_strides_0, weight = encoder_layers_12_mlp_down_proj_weight, x = input_257)[name = string("op_3178")];
+            tensor<fp16, [1, 768, 1, 128]> var_3178 = conv(dilations = var_3178_dilations_0, groups = var_3178_groups_0, pad = var_3178_pad_0, pad_type = var_3178_pad_type_0, strides = var_3178_strides_0, weight = encoder_layers_12_mlp_down_proj_weight_quantized, x = input_257)[name = string("op_3178")];
             tensor<int32, [1]> var_3179_axes_0 = const()[name = string("op_3179_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 768, 128]> var_3179 = squeeze(axes = var_3179_axes_0, x = var_3178)[name = string("op_3179")];
             tensor<int32, [3]> var_3180 = const()[name = string("op_3180"), val = tensor<int32, [3]>([0, 2, 1])];
@@ -2617,7 +2622,7 @@ program(1.3)
             tensor<int32, [2]> var_3189_split_sizes_0 = const()[name = string("op_3189_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_3189_axis_0 = const()[name = string("op_3189_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_3189_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_3189_cast_fp16_1 = split(axis = var_3189_axis_0, split_sizes = var_3189_split_sizes_0, x = normed_361_cast_fp16)[name = string("op_3189_cast_fp16")];
-            tensor<fp16, [768]> var_3193_to_fp16 = const()[name = string("op_3193_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(602390784)))];
+            tensor<fp16, [768]> var_3193_to_fp16 = const()[name = string("op_3193_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302023936)))];
             tensor<fp16, [1, 128, 768]> out_155_cast_fp16 = mul(x = var_3189_cast_fp16_0, y = var_3193_to_fp16)[name = string("out_155_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_209_cast_fp16 = add(x = x_203_cast_fp16, y = out_155_cast_fp16)[name = string("x_209_cast_fp16")];
             fp16 const_182_promoted_to_fp16 = const()[name = string("const_182_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -2629,7 +2634,7 @@ program(1.3)
             tensor<int32, [2]> var_3227_split_sizes_0 = const()[name = string("op_3227_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_3227_axis_0 = const()[name = string("op_3227_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_3227_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_3227_cast_fp16_1 = split(axis = var_3227_axis_0, split_sizes = var_3227_split_sizes_0, x = normed_365_cast_fp16)[name = string("op_3227_cast_fp16")];
-            tensor<fp16, [768]> var_3231_to_fp16 = const()[name = string("op_3231_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(602392384)))];
+            tensor<fp16, [768]> var_3231_to_fp16 = const()[name = string("op_3231_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302025536)))];
             tensor<fp16, [1, 128, 768]> out_157_cast_fp16 = mul(x = var_3227_cast_fp16_0, y = var_3231_to_fp16)[name = string("out_157_cast_fp16")];
             tensor<int32, [3]> var_3237 = const()[name = string("op_3237"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> var_3239_axes_0 = const()[name = string("op_3239_axes_0"), val = tensor<int32, [1]>([2])];
@@ -2640,7 +2645,7 @@ program(1.3)
             tensor<int32, [4]> var_3246_pad_0 = const()[name = string("op_3246_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_3246_dilations_0 = const()[name = string("op_3246_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_3246_groups_0 = const()[name = string("op_3246_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_3246 = conv(dilations = var_3246_dilations_0, groups = var_3246_groups_0, pad = var_3246_pad_0, pad_type = var_3246_pad_type_0, strides = var_3246_strides_0, weight = encoder_layers_13_self_attn_q_proj_weight, x = var_3239_cast_fp16)[name = string("op_3246")];
+            tensor<fp16, [1, 768, 1, 128]> var_3246 = conv(dilations = var_3246_dilations_0, groups = var_3246_groups_0, pad = var_3246_pad_0, pad_type = var_3246_pad_type_0, strides = var_3246_strides_0, weight = encoder_layers_13_self_attn_q_proj_weight_quantized, x = var_3239_cast_fp16)[name = string("op_3246")];
             tensor<int32, [4]> var_3247 = const()[name = string("op_3247"), val = tensor<int32, [4]>([1, 3, 256, 128])];
             tensor<fp16, [1, 3, 256, 128]> var_3248 = reshape(shape = var_3247, x = var_3246)[name = string("op_3248")];
             tensor<int32, [4]> var_3249 = const()[name = string("op_3249"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -2649,7 +2654,7 @@ program(1.3)
             tensor<int32, [4]> var_3256_pad_0 = const()[name = string("op_3256_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_3256_dilations_0 = const()[name = string("op_3256_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_3256_groups_0 = const()[name = string("op_3256_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_3256 = conv(dilations = var_3256_dilations_0, groups = var_3256_groups_0, pad = var_3256_pad_0, pad_type = var_3256_pad_type_0, strides = var_3256_strides_0, weight = encoder_layers_13_self_attn_k_proj_weight, x = var_3239_cast_fp16)[name = string("op_3256")];
+            tensor<fp16, [1, 256, 1, 128]> var_3256 = conv(dilations = var_3256_dilations_0, groups = var_3256_groups_0, pad = var_3256_pad_0, pad_type = var_3256_pad_type_0, strides = var_3256_strides_0, weight = encoder_layers_13_self_attn_k_proj_weight_quantized, x = var_3239_cast_fp16)[name = string("op_3256")];
             tensor<int32, [4]> var_3257 = const()[name = string("op_3257"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_3258 = reshape(shape = var_3257, x = var_3256)[name = string("op_3258")];
             tensor<int32, [4]> var_3259 = const()[name = string("op_3259"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -2658,7 +2663,7 @@ program(1.3)
             tensor<int32, [4]> var_3266_pad_0 = const()[name = string("op_3266_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_3266_dilations_0 = const()[name = string("op_3266_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_3266_groups_0 = const()[name = string("op_3266_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_3266 = conv(dilations = var_3266_dilations_0, groups = var_3266_groups_0, pad = var_3266_pad_0, pad_type = var_3266_pad_type_0, strides = var_3266_strides_0, weight = encoder_layers_13_self_attn_v_proj_weight, x = var_3239_cast_fp16)[name = string("op_3266")];
+            tensor<fp16, [1, 256, 1, 128]> var_3266 = conv(dilations = var_3266_dilations_0, groups = var_3266_groups_0, pad = var_3266_pad_0, pad_type = var_3266_pad_type_0, strides = var_3266_strides_0, weight = encoder_layers_13_self_attn_v_proj_weight_quantized, x = var_3239_cast_fp16)[name = string("op_3266")];
             tensor<int32, [4]> var_3267 = const()[name = string("op_3267"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_3268 = reshape(shape = var_3267, x = var_3266)[name = string("op_3268")];
             tensor<int32, [4]> var_3269 = const()[name = string("op_3269"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -2672,7 +2677,7 @@ program(1.3)
             tensor<int32, [2]> var_3280_split_sizes_0 = const()[name = string("op_3280_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_3280_axis_0 = const()[name = string("op_3280_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 256]> var_3280_cast_fp16_0, tensor<fp16, [1, 3, 128, 256]> var_3280_cast_fp16_1 = split(axis = var_3280_axis_0, split_sizes = var_3280_split_sizes_0, x = normed_371_cast_fp16)[name = string("op_3280_cast_fp16")];
-            tensor<fp16, [256]> var_3284_to_fp16 = const()[name = string("op_3284_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(602393984)))];
+            tensor<fp16, [256]> var_3284_to_fp16 = const()[name = string("op_3284_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302027136)))];
             tensor<fp16, [1, 3, 128, 256]> out_159_cast_fp16 = mul(x = var_3280_cast_fp16_0, y = var_3284_to_fp16)[name = string("out_159_cast_fp16")];
             fp16 const_186_promoted_to_fp16 = const()[name = string("const_186_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 1, 128, 256]> k_79 = transpose(perm = var_3259, x = var_3258)[name = string("transpose_96")];
@@ -2684,9 +2689,9 @@ program(1.3)
             tensor<int32, [2]> var_3296_split_sizes_0 = const()[name = string("op_3296_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_3296_axis_0 = const()[name = string("op_3296_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 256]> var_3296_cast_fp16_0, tensor<fp16, [1, 1, 128, 256]> var_3296_cast_fp16_1 = split(axis = var_3296_axis_0, split_sizes = var_3296_split_sizes_0, x = normed_375_cast_fp16)[name = string("op_3296_cast_fp16")];
-            tensor<fp16, [256]> var_3300_to_fp16 = const()[name = string("op_3300_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(602394560)))];
+            tensor<fp16, [256]> var_3300_to_fp16 = const()[name = string("op_3300_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302027712)))];
             tensor<fp16, [1, 1, 128, 256]> out_161_cast_fp16 = mul(x = var_3296_cast_fp16_0, y = var_3300_to_fp16)[name = string("out_161_cast_fp16")];
-            tensor<fp16, [1, 3, 128, 256]> var_3303 = mul(x = out_159_cast_fp16, y = cos_1)[name = string("op_3303")];
+            tensor<fp16, [1, 3, 128, 256]> var_3303 = mul(x = out_159_cast_fp16, y = cos_1_quantized)[name = string("op_3303")];
             tensor<int32, [2]> var_3304_split_sizes_0 = const()[name = string("op_3304_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_3304_axis_0 = const()[name = string("op_3304_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 128]> var_3304_0, tensor<fp16, [1, 3, 128, 128]> var_3304_1 = split(axis = var_3304_axis_0, split_sizes = var_3304_split_sizes_0, x = out_159_cast_fp16)[name = string("op_3304")];
@@ -2694,9 +2699,9 @@ program(1.3)
             tensor<fp16, [1, 3, 128, 128]> var_3306 = mul(x = var_3304_1, y = const_188_promoted)[name = string("op_3306")];
             bool var_3308_interleave_0 = const()[name = string("op_3308_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 3, 128, 256]> var_3308 = concat(axis = var_23, interleave = var_3308_interleave_0, values = (var_3306, var_3304_0))[name = string("op_3308")];
-            tensor<fp16, [1, 3, 128, 256]> var_3309 = mul(x = var_3308, y = sin_1)[name = string("op_3309")];
+            tensor<fp16, [1, 3, 128, 256]> var_3309 = mul(x = var_3308, y = sin_1_quantized)[name = string("op_3309")];
             tensor<fp16, [1, 3, 128, 256]> q_83 = add(x = var_3303, y = var_3309)[name = string("q_83")];
-            tensor<fp16, [1, 1, 128, 256]> var_3311 = mul(x = out_161_cast_fp16, y = cos_1)[name = string("op_3311")];
+            tensor<fp16, [1, 1, 128, 256]> var_3311 = mul(x = out_161_cast_fp16, y = cos_1_quantized)[name = string("op_3311")];
             tensor<int32, [2]> var_3312_split_sizes_0 = const()[name = string("op_3312_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_3312_axis_0 = const()[name = string("op_3312_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 128]> var_3312_0, tensor<fp16, [1, 1, 128, 128]> var_3312_1 = split(axis = var_3312_axis_0, split_sizes = var_3312_split_sizes_0, x = out_161_cast_fp16)[name = string("op_3312")];
@@ -2704,7 +2709,7 @@ program(1.3)
             tensor<fp16, [1, 1, 128, 128]> var_3314 = mul(x = var_3312_1, y = const_189_promoted)[name = string("op_3314")];
             bool var_3316_interleave_0 = const()[name = string("op_3316_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 1, 128, 256]> var_3316 = concat(axis = var_23, interleave = var_3316_interleave_0, values = (var_3314, var_3312_0))[name = string("op_3316")];
-            tensor<fp16, [1, 1, 128, 256]> var_3317 = mul(x = var_3316, y = sin_1)[name = string("op_3317")];
+            tensor<fp16, [1, 1, 128, 256]> var_3317 = mul(x = var_3316, y = sin_1_quantized)[name = string("op_3317")];
             tensor<fp16, [1, 1, 128, 256]> hidden_states_157 = add(x = var_3311, y = var_3317)[name = string("hidden_states_157")];
             tensor<int32, [1]> hidden_states_159_axes_0 = const()[name = string("hidden_states_159_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 1, 1, 128, 256]> hidden_states_159 = expand_dims(axes = hidden_states_159_axes_0, x = hidden_states_157)[name = string("hidden_states_159")];
@@ -2734,14 +2739,14 @@ program(1.3)
             tensor<fp16, [1, 128, 3, 256]> var_3344 = transpose(perm = var_3343, x = var_3341_cast_fp16)[name = string("transpose_94")];
             tensor<fp16, [1, 128, 768]> attn_out_81 = reshape(shape = var_3346, x = var_3344)[name = string("attn_out_81")];
             tensor<int32, [3]> var_3348 = const()[name = string("op_3348"), val = tensor<int32, [3]>([0, 2, 1])];
-            tensor<fp16, [768, 768, 1]> squeeze_13 = const()[name = string("squeeze_13"), val = tensor<fp16, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(602395136)))];
+            tensor<fp16, [768, 768, 1]> squeeze_13_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302028288))), scale = tensor<fp16, [768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302618176))))[name = string("squeeze_13_quantized")];
             string var_3357_pad_type_0 = const()[name = string("op_3357_pad_type_0"), val = string("valid")];
             int32 var_3357_groups_0 = const()[name = string("op_3357_groups_0"), val = int32(1)];
             tensor<int32, [1]> var_3357_strides_0 = const()[name = string("op_3357_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> var_3357_pad_0 = const()[name = string("op_3357_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> var_3357_dilations_0 = const()[name = string("op_3357_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, [1, 768, 128]> var_3349 = transpose(perm = var_3348, x = attn_out_81)[name = string("transpose_93")];
-            tensor<fp16, [1, 768, 128]> var_3357 = conv(dilations = var_3357_dilations_0, groups = var_3357_groups_0, pad = var_3357_pad_0, pad_type = var_3357_pad_type_0, strides = var_3357_strides_0, weight = squeeze_13, x = var_3349)[name = string("op_3357")];
+            tensor<fp16, [1, 768, 128]> var_3357 = conv(dilations = var_3357_dilations_0, groups = var_3357_groups_0, pad = var_3357_pad_0, pad_type = var_3357_pad_type_0, strides = var_3357_strides_0, weight = squeeze_13_quantized, x = var_3349)[name = string("op_3357")];
             tensor<int32, [3]> var_3358 = const()[name = string("op_3358"), val = tensor<int32, [3]>([0, 2, 1])];
             fp16 const_190_promoted_to_fp16 = const()[name = string("const_190_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 128, 768]> x_217 = transpose(perm = var_3358, x = var_3357)[name = string("transpose_92")];
@@ -2753,7 +2758,7 @@ program(1.3)
             tensor<int32, [2]> var_3367_split_sizes_0 = const()[name = string("op_3367_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_3367_axis_0 = const()[name = string("op_3367_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_3367_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_3367_cast_fp16_1 = split(axis = var_3367_axis_0, split_sizes = var_3367_split_sizes_0, x = normed_379_cast_fp16)[name = string("op_3367_cast_fp16")];
-            tensor<fp16, [768]> var_3371_to_fp16 = const()[name = string("op_3371_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(603574848)))];
+            tensor<fp16, [768]> var_3371_to_fp16 = const()[name = string("op_3371_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302619776)))];
             tensor<fp16, [1, 128, 768]> out_163_cast_fp16 = mul(x = var_3367_cast_fp16_0, y = var_3371_to_fp16)[name = string("out_163_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_219_cast_fp16 = add(x = x_209_cast_fp16, y = out_163_cast_fp16)[name = string("x_219_cast_fp16")];
             fp16 const_192_promoted_to_fp16 = const()[name = string("const_192_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -2765,7 +2770,7 @@ program(1.3)
             tensor<int32, [2]> var_3383_split_sizes_0 = const()[name = string("op_3383_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_3383_axis_0 = const()[name = string("op_3383_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_3383_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_3383_cast_fp16_1 = split(axis = var_3383_axis_0, split_sizes = var_3383_split_sizes_0, x = normed_383_cast_fp16)[name = string("op_3383_cast_fp16")];
-            tensor<fp16, [768]> var_3387_to_fp16 = const()[name = string("op_3387_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(603576448)))];
+            tensor<fp16, [768]> var_3387_to_fp16 = const()[name = string("op_3387_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302621376)))];
             tensor<fp16, [1, 128, 768]> out_165_cast_fp16 = mul(x = var_3383_cast_fp16_0, y = var_3387_to_fp16)[name = string("out_165_cast_fp16")];
             tensor<int32, [3]> var_3394 = const()[name = string("op_3394"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> input_275_axes_0 = const()[name = string("input_275_axes_0"), val = tensor<int32, [1]>([2])];
@@ -2776,13 +2781,13 @@ program(1.3)
             tensor<int32, [4]> gate_53_pad_0 = const()[name = string("gate_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> gate_53_dilations_0 = const()[name = string("gate_53_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 gate_53_groups_0 = const()[name = string("gate_53_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> gate_53 = conv(dilations = gate_53_dilations_0, groups = gate_53_groups_0, pad = gate_53_pad_0, pad_type = gate_53_pad_type_0, strides = gate_53_strides_0, weight = encoder_layers_13_mlp_gate_proj_weight, x = input_275)[name = string("gate_53")];
+            tensor<fp16, [1, 1152, 1, 128]> gate_53 = conv(dilations = gate_53_dilations_0, groups = gate_53_groups_0, pad = gate_53_pad_0, pad_type = gate_53_pad_type_0, strides = gate_53_strides_0, weight = encoder_layers_13_mlp_gate_proj_weight_quantized, x = input_275)[name = string("gate_53")];
             string up_27_pad_type_0 = const()[name = string("up_27_pad_type_0"), val = string("valid")];
             tensor<int32, [2]> up_27_strides_0 = const()[name = string("up_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> up_27_pad_0 = const()[name = string("up_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> up_27_dilations_0 = const()[name = string("up_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 up_27_groups_0 = const()[name = string("up_27_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> up_27 = conv(dilations = up_27_dilations_0, groups = up_27_groups_0, pad = up_27_pad_0, pad_type = up_27_pad_type_0, strides = up_27_strides_0, weight = encoder_layers_13_mlp_up_proj_weight, x = input_275)[name = string("up_27")];
+            tensor<fp16, [1, 1152, 1, 128]> up_27 = conv(dilations = up_27_dilations_0, groups = up_27_groups_0, pad = up_27_pad_0, pad_type = up_27_pad_type_0, strides = up_27_strides_0, weight = encoder_layers_13_mlp_up_proj_weight_quantized, x = input_275)[name = string("up_27")];
             string gate_55_mode_0 = const()[name = string("gate_55_mode_0"), val = string("TANH_APPROXIMATION")];
             tensor<fp16, [1, 1152, 1, 128]> gate_55 = gelu(mode = gate_55_mode_0, x = gate_53)[name = string("gate_55")];
             tensor<fp16, [1, 1152, 1, 128]> input_277 = mul(x = gate_55, y = up_27)[name = string("input_277")];
@@ -2791,7 +2796,7 @@ program(1.3)
             tensor<int32, [4]> var_3416_pad_0 = const()[name = string("op_3416_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_3416_dilations_0 = const()[name = string("op_3416_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_3416_groups_0 = const()[name = string("op_3416_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_3416 = conv(dilations = var_3416_dilations_0, groups = var_3416_groups_0, pad = var_3416_pad_0, pad_type = var_3416_pad_type_0, strides = var_3416_strides_0, weight = encoder_layers_13_mlp_down_proj_weight, x = input_277)[name = string("op_3416")];
+            tensor<fp16, [1, 768, 1, 128]> var_3416 = conv(dilations = var_3416_dilations_0, groups = var_3416_groups_0, pad = var_3416_pad_0, pad_type = var_3416_pad_type_0, strides = var_3416_strides_0, weight = encoder_layers_13_mlp_down_proj_weight_quantized, x = input_277)[name = string("op_3416")];
             tensor<int32, [1]> var_3417_axes_0 = const()[name = string("op_3417_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 768, 128]> var_3417 = squeeze(axes = var_3417_axes_0, x = var_3416)[name = string("op_3417")];
             tensor<int32, [3]> var_3418 = const()[name = string("op_3418"), val = tensor<int32, [3]>([0, 2, 1])];
@@ -2805,7 +2810,7 @@ program(1.3)
             tensor<int32, [2]> var_3427_split_sizes_0 = const()[name = string("op_3427_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_3427_axis_0 = const()[name = string("op_3427_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_3427_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_3427_cast_fp16_1 = split(axis = var_3427_axis_0, split_sizes = var_3427_split_sizes_0, x = normed_389_cast_fp16)[name = string("op_3427_cast_fp16")];
-            tensor<fp16, [768]> var_3431_to_fp16 = const()[name = string("op_3431_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(603578048)))];
+            tensor<fp16, [768]> var_3431_to_fp16 = const()[name = string("op_3431_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302622976)))];
             tensor<fp16, [1, 128, 768]> out_167_cast_fp16 = mul(x = var_3427_cast_fp16_0, y = var_3431_to_fp16)[name = string("out_167_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_225_cast_fp16 = add(x = x_219_cast_fp16, y = out_167_cast_fp16)[name = string("x_225_cast_fp16")];
             fp16 const_196_promoted_to_fp16 = const()[name = string("const_196_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -2817,7 +2822,7 @@ program(1.3)
             tensor<int32, [2]> var_3465_split_sizes_0 = const()[name = string("op_3465_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_3465_axis_0 = const()[name = string("op_3465_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_3465_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_3465_cast_fp16_1 = split(axis = var_3465_axis_0, split_sizes = var_3465_split_sizes_0, x = normed_393_cast_fp16)[name = string("op_3465_cast_fp16")];
-            tensor<fp16, [768]> var_3469_to_fp16 = const()[name = string("op_3469_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(603579648)))];
+            tensor<fp16, [768]> var_3469_to_fp16 = const()[name = string("op_3469_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302624576)))];
             tensor<fp16, [1, 128, 768]> out_169_cast_fp16 = mul(x = var_3465_cast_fp16_0, y = var_3469_to_fp16)[name = string("out_169_cast_fp16")];
             tensor<int32, [3]> var_3475 = const()[name = string("op_3475"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> var_3477_axes_0 = const()[name = string("op_3477_axes_0"), val = tensor<int32, [1]>([2])];
@@ -2828,7 +2833,7 @@ program(1.3)
             tensor<int32, [4]> var_3484_pad_0 = const()[name = string("op_3484_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_3484_dilations_0 = const()[name = string("op_3484_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_3484_groups_0 = const()[name = string("op_3484_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_3484 = conv(dilations = var_3484_dilations_0, groups = var_3484_groups_0, pad = var_3484_pad_0, pad_type = var_3484_pad_type_0, strides = var_3484_strides_0, weight = encoder_layers_14_self_attn_q_proj_weight, x = var_3477_cast_fp16)[name = string("op_3484")];
+            tensor<fp16, [1, 768, 1, 128]> var_3484 = conv(dilations = var_3484_dilations_0, groups = var_3484_groups_0, pad = var_3484_pad_0, pad_type = var_3484_pad_type_0, strides = var_3484_strides_0, weight = encoder_layers_14_self_attn_q_proj_weight_quantized, x = var_3477_cast_fp16)[name = string("op_3484")];
             tensor<int32, [4]> var_3485 = const()[name = string("op_3485"), val = tensor<int32, [4]>([1, 3, 256, 128])];
             tensor<fp16, [1, 3, 256, 128]> var_3486 = reshape(shape = var_3485, x = var_3484)[name = string("op_3486")];
             tensor<int32, [4]> var_3487 = const()[name = string("op_3487"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -2837,7 +2842,7 @@ program(1.3)
             tensor<int32, [4]> var_3494_pad_0 = const()[name = string("op_3494_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_3494_dilations_0 = const()[name = string("op_3494_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_3494_groups_0 = const()[name = string("op_3494_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_3494 = conv(dilations = var_3494_dilations_0, groups = var_3494_groups_0, pad = var_3494_pad_0, pad_type = var_3494_pad_type_0, strides = var_3494_strides_0, weight = encoder_layers_14_self_attn_k_proj_weight, x = var_3477_cast_fp16)[name = string("op_3494")];
+            tensor<fp16, [1, 256, 1, 128]> var_3494 = conv(dilations = var_3494_dilations_0, groups = var_3494_groups_0, pad = var_3494_pad_0, pad_type = var_3494_pad_type_0, strides = var_3494_strides_0, weight = encoder_layers_14_self_attn_k_proj_weight_quantized, x = var_3477_cast_fp16)[name = string("op_3494")];
             tensor<int32, [4]> var_3495 = const()[name = string("op_3495"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_3496 = reshape(shape = var_3495, x = var_3494)[name = string("op_3496")];
             tensor<int32, [4]> var_3497 = const()[name = string("op_3497"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -2846,7 +2851,7 @@ program(1.3)
             tensor<int32, [4]> var_3504_pad_0 = const()[name = string("op_3504_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_3504_dilations_0 = const()[name = string("op_3504_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_3504_groups_0 = const()[name = string("op_3504_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_3504 = conv(dilations = var_3504_dilations_0, groups = var_3504_groups_0, pad = var_3504_pad_0, pad_type = var_3504_pad_type_0, strides = var_3504_strides_0, weight = encoder_layers_14_self_attn_v_proj_weight, x = var_3477_cast_fp16)[name = string("op_3504")];
+            tensor<fp16, [1, 256, 1, 128]> var_3504 = conv(dilations = var_3504_dilations_0, groups = var_3504_groups_0, pad = var_3504_pad_0, pad_type = var_3504_pad_type_0, strides = var_3504_strides_0, weight = encoder_layers_14_self_attn_v_proj_weight_quantized, x = var_3477_cast_fp16)[name = string("op_3504")];
             tensor<int32, [4]> var_3505 = const()[name = string("op_3505"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_3506 = reshape(shape = var_3505, x = var_3504)[name = string("op_3506")];
             tensor<int32, [4]> var_3507 = const()[name = string("op_3507"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -2860,7 +2865,7 @@ program(1.3)
             tensor<int32, [2]> var_3518_split_sizes_0 = const()[name = string("op_3518_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_3518_axis_0 = const()[name = string("op_3518_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 256]> var_3518_cast_fp16_0, tensor<fp16, [1, 3, 128, 256]> var_3518_cast_fp16_1 = split(axis = var_3518_axis_0, split_sizes = var_3518_split_sizes_0, x = normed_399_cast_fp16)[name = string("op_3518_cast_fp16")];
-            tensor<fp16, [256]> var_3522_to_fp16 = const()[name = string("op_3522_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(603581248)))];
+            tensor<fp16, [256]> var_3522_to_fp16 = const()[name = string("op_3522_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302626176)))];
             tensor<fp16, [1, 3, 128, 256]> out_171_cast_fp16 = mul(x = var_3518_cast_fp16_0, y = var_3522_to_fp16)[name = string("out_171_cast_fp16")];
             fp16 const_200_promoted_to_fp16 = const()[name = string("const_200_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 1, 128, 256]> k_85 = transpose(perm = var_3497, x = var_3496)[name = string("transpose_87")];
@@ -2872,9 +2877,9 @@ program(1.3)
             tensor<int32, [2]> var_3534_split_sizes_0 = const()[name = string("op_3534_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_3534_axis_0 = const()[name = string("op_3534_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 256]> var_3534_cast_fp16_0, tensor<fp16, [1, 1, 128, 256]> var_3534_cast_fp16_1 = split(axis = var_3534_axis_0, split_sizes = var_3534_split_sizes_0, x = normed_403_cast_fp16)[name = string("op_3534_cast_fp16")];
-            tensor<fp16, [256]> var_3538_to_fp16 = const()[name = string("op_3538_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(603581824)))];
+            tensor<fp16, [256]> var_3538_to_fp16 = const()[name = string("op_3538_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302626752)))];
             tensor<fp16, [1, 1, 128, 256]> out_173_cast_fp16 = mul(x = var_3534_cast_fp16_0, y = var_3538_to_fp16)[name = string("out_173_cast_fp16")];
-            tensor<fp16, [1, 3, 128, 256]> var_3541 = mul(x = out_171_cast_fp16, y = cos_1)[name = string("op_3541")];
+            tensor<fp16, [1, 3, 128, 256]> var_3541 = mul(x = out_171_cast_fp16, y = cos_1_quantized)[name = string("op_3541")];
             tensor<int32, [2]> var_3542_split_sizes_0 = const()[name = string("op_3542_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_3542_axis_0 = const()[name = string("op_3542_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 128]> var_3542_0, tensor<fp16, [1, 3, 128, 128]> var_3542_1 = split(axis = var_3542_axis_0, split_sizes = var_3542_split_sizes_0, x = out_171_cast_fp16)[name = string("op_3542")];
@@ -2882,9 +2887,9 @@ program(1.3)
             tensor<fp16, [1, 3, 128, 128]> var_3544 = mul(x = var_3542_1, y = const_202_promoted)[name = string("op_3544")];
             bool var_3546_interleave_0 = const()[name = string("op_3546_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 3, 128, 256]> var_3546 = concat(axis = var_23, interleave = var_3546_interleave_0, values = (var_3544, var_3542_0))[name = string("op_3546")];
-            tensor<fp16, [1, 3, 128, 256]> var_3547 = mul(x = var_3546, y = sin_1)[name = string("op_3547")];
+            tensor<fp16, [1, 3, 128, 256]> var_3547 = mul(x = var_3546, y = sin_1_quantized)[name = string("op_3547")];
             tensor<fp16, [1, 3, 128, 256]> q_89 = add(x = var_3541, y = var_3547)[name = string("q_89")];
-            tensor<fp16, [1, 1, 128, 256]> var_3549 = mul(x = out_173_cast_fp16, y = cos_1)[name = string("op_3549")];
+            tensor<fp16, [1, 1, 128, 256]> var_3549 = mul(x = out_173_cast_fp16, y = cos_1_quantized)[name = string("op_3549")];
             tensor<int32, [2]> var_3550_split_sizes_0 = const()[name = string("op_3550_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_3550_axis_0 = const()[name = string("op_3550_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 128]> var_3550_0, tensor<fp16, [1, 1, 128, 128]> var_3550_1 = split(axis = var_3550_axis_0, split_sizes = var_3550_split_sizes_0, x = out_173_cast_fp16)[name = string("op_3550")];
@@ -2892,7 +2897,7 @@ program(1.3)
             tensor<fp16, [1, 1, 128, 128]> var_3552 = mul(x = var_3550_1, y = const_203_promoted)[name = string("op_3552")];
             bool var_3554_interleave_0 = const()[name = string("op_3554_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 1, 128, 256]> var_3554 = concat(axis = var_23, interleave = var_3554_interleave_0, values = (var_3552, var_3550_0))[name = string("op_3554")];
-            tensor<fp16, [1, 1, 128, 256]> var_3555 = mul(x = var_3554, y = sin_1)[name = string("op_3555")];
+            tensor<fp16, [1, 1, 128, 256]> var_3555 = mul(x = var_3554, y = sin_1_quantized)[name = string("op_3555")];
             tensor<fp16, [1, 1, 128, 256]> hidden_states_169 = add(x = var_3549, y = var_3555)[name = string("hidden_states_169")];
             tensor<int32, [1]> hidden_states_171_axes_0 = const()[name = string("hidden_states_171_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 1, 1, 128, 256]> hidden_states_171 = expand_dims(axes = hidden_states_171_axes_0, x = hidden_states_169)[name = string("hidden_states_171")];
@@ -2922,14 +2927,14 @@ program(1.3)
             tensor<fp16, [1, 128, 3, 256]> var_3582 = transpose(perm = var_3581, x = var_3579_cast_fp16)[name = string("transpose_85")];
             tensor<fp16, [1, 128, 768]> attn_out_87 = reshape(shape = var_3584, x = var_3582)[name = string("attn_out_87")];
             tensor<int32, [3]> var_3586 = const()[name = string("op_3586"), val = tensor<int32, [3]>([0, 2, 1])];
-            tensor<fp16, [768, 768, 1]> squeeze_14 = const()[name = string("squeeze_14"), val = tensor<fp16, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(603582400)))];
+            tensor<fp16, [768, 768, 1]> squeeze_14_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302627328))), scale = tensor<fp16, [768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303217216))))[name = string("squeeze_14_quantized")];
             string var_3595_pad_type_0 = const()[name = string("op_3595_pad_type_0"), val = string("valid")];
             int32 var_3595_groups_0 = const()[name = string("op_3595_groups_0"), val = int32(1)];
             tensor<int32, [1]> var_3595_strides_0 = const()[name = string("op_3595_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> var_3595_pad_0 = const()[name = string("op_3595_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> var_3595_dilations_0 = const()[name = string("op_3595_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, [1, 768, 128]> var_3587 = transpose(perm = var_3586, x = attn_out_87)[name = string("transpose_84")];
-            tensor<fp16, [1, 768, 128]> var_3595 = conv(dilations = var_3595_dilations_0, groups = var_3595_groups_0, pad = var_3595_pad_0, pad_type = var_3595_pad_type_0, strides = var_3595_strides_0, weight = squeeze_14, x = var_3587)[name = string("op_3595")];
+            tensor<fp16, [1, 768, 128]> var_3595 = conv(dilations = var_3595_dilations_0, groups = var_3595_groups_0, pad = var_3595_pad_0, pad_type = var_3595_pad_type_0, strides = var_3595_strides_0, weight = squeeze_14_quantized, x = var_3587)[name = string("op_3595")];
             tensor<int32, [3]> var_3596 = const()[name = string("op_3596"), val = tensor<int32, [3]>([0, 2, 1])];
             fp16 const_204_promoted_to_fp16 = const()[name = string("const_204_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 128, 768]> x_233 = transpose(perm = var_3596, x = var_3595)[name = string("transpose_83")];
@@ -2941,7 +2946,7 @@ program(1.3)
             tensor<int32, [2]> var_3605_split_sizes_0 = const()[name = string("op_3605_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_3605_axis_0 = const()[name = string("op_3605_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_3605_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_3605_cast_fp16_1 = split(axis = var_3605_axis_0, split_sizes = var_3605_split_sizes_0, x = normed_407_cast_fp16)[name = string("op_3605_cast_fp16")];
-            tensor<fp16, [768]> var_3609_to_fp16 = const()[name = string("op_3609_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604762112)))];
+            tensor<fp16, [768]> var_3609_to_fp16 = const()[name = string("op_3609_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303218816)))];
             tensor<fp16, [1, 128, 768]> out_175_cast_fp16 = mul(x = var_3605_cast_fp16_0, y = var_3609_to_fp16)[name = string("out_175_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_235_cast_fp16 = add(x = x_225_cast_fp16, y = out_175_cast_fp16)[name = string("x_235_cast_fp16")];
             fp16 const_206_promoted_to_fp16 = const()[name = string("const_206_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -2953,7 +2958,7 @@ program(1.3)
             tensor<int32, [2]> var_3621_split_sizes_0 = const()[name = string("op_3621_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_3621_axis_0 = const()[name = string("op_3621_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_3621_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_3621_cast_fp16_1 = split(axis = var_3621_axis_0, split_sizes = var_3621_split_sizes_0, x = normed_411_cast_fp16)[name = string("op_3621_cast_fp16")];
-            tensor<fp16, [768]> var_3625_to_fp16 = const()[name = string("op_3625_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604763712)))];
+            tensor<fp16, [768]> var_3625_to_fp16 = const()[name = string("op_3625_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303220416)))];
             tensor<fp16, [1, 128, 768]> out_177_cast_fp16 = mul(x = var_3621_cast_fp16_0, y = var_3625_to_fp16)[name = string("out_177_cast_fp16")];
             tensor<int32, [3]> var_3632 = const()[name = string("op_3632"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> input_295_axes_0 = const()[name = string("input_295_axes_0"), val = tensor<int32, [1]>([2])];
@@ -2964,13 +2969,13 @@ program(1.3)
             tensor<int32, [4]> gate_57_pad_0 = const()[name = string("gate_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> gate_57_dilations_0 = const()[name = string("gate_57_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 gate_57_groups_0 = const()[name = string("gate_57_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> gate_57 = conv(dilations = gate_57_dilations_0, groups = gate_57_groups_0, pad = gate_57_pad_0, pad_type = gate_57_pad_type_0, strides = gate_57_strides_0, weight = encoder_layers_14_mlp_gate_proj_weight, x = input_295)[name = string("gate_57")];
+            tensor<fp16, [1, 1152, 1, 128]> gate_57 = conv(dilations = gate_57_dilations_0, groups = gate_57_groups_0, pad = gate_57_pad_0, pad_type = gate_57_pad_type_0, strides = gate_57_strides_0, weight = encoder_layers_14_mlp_gate_proj_weight_quantized, x = input_295)[name = string("gate_57")];
             string up_29_pad_type_0 = const()[name = string("up_29_pad_type_0"), val = string("valid")];
             tensor<int32, [2]> up_29_strides_0 = const()[name = string("up_29_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> up_29_pad_0 = const()[name = string("up_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> up_29_dilations_0 = const()[name = string("up_29_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 up_29_groups_0 = const()[name = string("up_29_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> up_29 = conv(dilations = up_29_dilations_0, groups = up_29_groups_0, pad = up_29_pad_0, pad_type = up_29_pad_type_0, strides = up_29_strides_0, weight = encoder_layers_14_mlp_up_proj_weight, x = input_295)[name = string("up_29")];
+            tensor<fp16, [1, 1152, 1, 128]> up_29 = conv(dilations = up_29_dilations_0, groups = up_29_groups_0, pad = up_29_pad_0, pad_type = up_29_pad_type_0, strides = up_29_strides_0, weight = encoder_layers_14_mlp_up_proj_weight_quantized, x = input_295)[name = string("up_29")];
             string gate_59_mode_0 = const()[name = string("gate_59_mode_0"), val = string("TANH_APPROXIMATION")];
             tensor<fp16, [1, 1152, 1, 128]> gate_59 = gelu(mode = gate_59_mode_0, x = gate_57)[name = string("gate_59")];
             tensor<fp16, [1, 1152, 1, 128]> input_297 = mul(x = gate_59, y = up_29)[name = string("input_297")];
@@ -2979,7 +2984,7 @@ program(1.3)
             tensor<int32, [4]> var_3654_pad_0 = const()[name = string("op_3654_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_3654_dilations_0 = const()[name = string("op_3654_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_3654_groups_0 = const()[name = string("op_3654_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_3654 = conv(dilations = var_3654_dilations_0, groups = var_3654_groups_0, pad = var_3654_pad_0, pad_type = var_3654_pad_type_0, strides = var_3654_strides_0, weight = encoder_layers_14_mlp_down_proj_weight, x = input_297)[name = string("op_3654")];
+            tensor<fp16, [1, 768, 1, 128]> var_3654 = conv(dilations = var_3654_dilations_0, groups = var_3654_groups_0, pad = var_3654_pad_0, pad_type = var_3654_pad_type_0, strides = var_3654_strides_0, weight = encoder_layers_14_mlp_down_proj_weight_quantized, x = input_297)[name = string("op_3654")];
             tensor<int32, [1]> var_3655_axes_0 = const()[name = string("op_3655_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 768, 128]> var_3655 = squeeze(axes = var_3655_axes_0, x = var_3654)[name = string("op_3655")];
             tensor<int32, [3]> var_3656 = const()[name = string("op_3656"), val = tensor<int32, [3]>([0, 2, 1])];
@@ -2993,7 +2998,7 @@ program(1.3)
             tensor<int32, [2]> var_3665_split_sizes_0 = const()[name = string("op_3665_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_3665_axis_0 = const()[name = string("op_3665_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_3665_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_3665_cast_fp16_1 = split(axis = var_3665_axis_0, split_sizes = var_3665_split_sizes_0, x = normed_417_cast_fp16)[name = string("op_3665_cast_fp16")];
-            tensor<fp16, [768]> var_3669_to_fp16 = const()[name = string("op_3669_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604765312)))];
+            tensor<fp16, [768]> var_3669_to_fp16 = const()[name = string("op_3669_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303222016)))];
             tensor<fp16, [1, 128, 768]> out_179_cast_fp16 = mul(x = var_3665_cast_fp16_0, y = var_3669_to_fp16)[name = string("out_179_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_241_cast_fp16 = add(x = x_235_cast_fp16, y = out_179_cast_fp16)[name = string("x_241_cast_fp16")];
             fp16 const_210_promoted_to_fp16 = const()[name = string("const_210_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -3005,7 +3010,7 @@ program(1.3)
             tensor<int32, [2]> var_3703_split_sizes_0 = const()[name = string("op_3703_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_3703_axis_0 = const()[name = string("op_3703_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_3703_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_3703_cast_fp16_1 = split(axis = var_3703_axis_0, split_sizes = var_3703_split_sizes_0, x = normed_421_cast_fp16)[name = string("op_3703_cast_fp16")];
-            tensor<fp16, [768]> var_3707_to_fp16 = const()[name = string("op_3707_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604766912)))];
+            tensor<fp16, [768]> var_3707_to_fp16 = const()[name = string("op_3707_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303223616)))];
             tensor<fp16, [1, 128, 768]> out_181_cast_fp16 = mul(x = var_3703_cast_fp16_0, y = var_3707_to_fp16)[name = string("out_181_cast_fp16")];
             tensor<int32, [3]> var_3713 = const()[name = string("op_3713"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> var_3715_axes_0 = const()[name = string("op_3715_axes_0"), val = tensor<int32, [1]>([2])];
@@ -3016,7 +3021,7 @@ program(1.3)
             tensor<int32, [4]> var_3722_pad_0 = const()[name = string("op_3722_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_3722_dilations_0 = const()[name = string("op_3722_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_3722_groups_0 = const()[name = string("op_3722_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_3722 = conv(dilations = var_3722_dilations_0, groups = var_3722_groups_0, pad = var_3722_pad_0, pad_type = var_3722_pad_type_0, strides = var_3722_strides_0, weight = encoder_layers_15_self_attn_q_proj_weight, x = var_3715_cast_fp16)[name = string("op_3722")];
+            tensor<fp16, [1, 768, 1, 128]> var_3722 = conv(dilations = var_3722_dilations_0, groups = var_3722_groups_0, pad = var_3722_pad_0, pad_type = var_3722_pad_type_0, strides = var_3722_strides_0, weight = encoder_layers_15_self_attn_q_proj_weight_quantized, x = var_3715_cast_fp16)[name = string("op_3722")];
             tensor<int32, [4]> var_3723 = const()[name = string("op_3723"), val = tensor<int32, [4]>([1, 3, 256, 128])];
             tensor<fp16, [1, 3, 256, 128]> var_3724 = reshape(shape = var_3723, x = var_3722)[name = string("op_3724")];
             tensor<int32, [4]> var_3725 = const()[name = string("op_3725"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -3025,7 +3030,7 @@ program(1.3)
             tensor<int32, [4]> var_3732_pad_0 = const()[name = string("op_3732_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_3732_dilations_0 = const()[name = string("op_3732_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_3732_groups_0 = const()[name = string("op_3732_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_3732 = conv(dilations = var_3732_dilations_0, groups = var_3732_groups_0, pad = var_3732_pad_0, pad_type = var_3732_pad_type_0, strides = var_3732_strides_0, weight = encoder_layers_15_self_attn_k_proj_weight, x = var_3715_cast_fp16)[name = string("op_3732")];
+            tensor<fp16, [1, 256, 1, 128]> var_3732 = conv(dilations = var_3732_dilations_0, groups = var_3732_groups_0, pad = var_3732_pad_0, pad_type = var_3732_pad_type_0, strides = var_3732_strides_0, weight = encoder_layers_15_self_attn_k_proj_weight_quantized, x = var_3715_cast_fp16)[name = string("op_3732")];
             tensor<int32, [4]> var_3733 = const()[name = string("op_3733"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_3734 = reshape(shape = var_3733, x = var_3732)[name = string("op_3734")];
             tensor<int32, [4]> var_3735 = const()[name = string("op_3735"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -3034,7 +3039,7 @@ program(1.3)
             tensor<int32, [4]> var_3742_pad_0 = const()[name = string("op_3742_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_3742_dilations_0 = const()[name = string("op_3742_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_3742_groups_0 = const()[name = string("op_3742_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_3742 = conv(dilations = var_3742_dilations_0, groups = var_3742_groups_0, pad = var_3742_pad_0, pad_type = var_3742_pad_type_0, strides = var_3742_strides_0, weight = encoder_layers_15_self_attn_v_proj_weight, x = var_3715_cast_fp16)[name = string("op_3742")];
+            tensor<fp16, [1, 256, 1, 128]> var_3742 = conv(dilations = var_3742_dilations_0, groups = var_3742_groups_0, pad = var_3742_pad_0, pad_type = var_3742_pad_type_0, strides = var_3742_strides_0, weight = encoder_layers_15_self_attn_v_proj_weight_quantized, x = var_3715_cast_fp16)[name = string("op_3742")];
             tensor<int32, [4]> var_3743 = const()[name = string("op_3743"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_3744 = reshape(shape = var_3743, x = var_3742)[name = string("op_3744")];
             tensor<int32, [4]> var_3745 = const()[name = string("op_3745"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -3048,7 +3053,7 @@ program(1.3)
             tensor<int32, [2]> var_3756_split_sizes_0 = const()[name = string("op_3756_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_3756_axis_0 = const()[name = string("op_3756_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 256]> var_3756_cast_fp16_0, tensor<fp16, [1, 3, 128, 256]> var_3756_cast_fp16_1 = split(axis = var_3756_axis_0, split_sizes = var_3756_split_sizes_0, x = normed_427_cast_fp16)[name = string("op_3756_cast_fp16")];
-            tensor<fp16, [256]> var_3760_to_fp16 = const()[name = string("op_3760_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604768512)))];
+            tensor<fp16, [256]> var_3760_to_fp16 = const()[name = string("op_3760_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303225216)))];
             tensor<fp16, [1, 3, 128, 256]> out_183_cast_fp16 = mul(x = var_3756_cast_fp16_0, y = var_3760_to_fp16)[name = string("out_183_cast_fp16")];
             fp16 const_214_promoted_to_fp16 = const()[name = string("const_214_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 1, 128, 256]> k_91 = transpose(perm = var_3735, x = var_3734)[name = string("transpose_78")];
@@ -3060,9 +3065,9 @@ program(1.3)
             tensor<int32, [2]> var_3772_split_sizes_0 = const()[name = string("op_3772_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_3772_axis_0 = const()[name = string("op_3772_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 256]> var_3772_cast_fp16_0, tensor<fp16, [1, 1, 128, 256]> var_3772_cast_fp16_1 = split(axis = var_3772_axis_0, split_sizes = var_3772_split_sizes_0, x = normed_431_cast_fp16)[name = string("op_3772_cast_fp16")];
-            tensor<fp16, [256]> var_3776_to_fp16 = const()[name = string("op_3776_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604769088)))];
+            tensor<fp16, [256]> var_3776_to_fp16 = const()[name = string("op_3776_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303225792)))];
             tensor<fp16, [1, 1, 128, 256]> out_185_cast_fp16 = mul(x = var_3772_cast_fp16_0, y = var_3776_to_fp16)[name = string("out_185_cast_fp16")];
-            tensor<fp16, [1, 3, 128, 256]> var_3779 = mul(x = out_183_cast_fp16, y = cos_1)[name = string("op_3779")];
+            tensor<fp16, [1, 3, 128, 256]> var_3779 = mul(x = out_183_cast_fp16, y = cos_1_quantized)[name = string("op_3779")];
             tensor<int32, [2]> var_3780_split_sizes_0 = const()[name = string("op_3780_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_3780_axis_0 = const()[name = string("op_3780_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 128]> var_3780_0, tensor<fp16, [1, 3, 128, 128]> var_3780_1 = split(axis = var_3780_axis_0, split_sizes = var_3780_split_sizes_0, x = out_183_cast_fp16)[name = string("op_3780")];
@@ -3070,9 +3075,9 @@ program(1.3)
             tensor<fp16, [1, 3, 128, 128]> var_3782 = mul(x = var_3780_1, y = const_216_promoted)[name = string("op_3782")];
             bool var_3784_interleave_0 = const()[name = string("op_3784_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 3, 128, 256]> var_3784 = concat(axis = var_23, interleave = var_3784_interleave_0, values = (var_3782, var_3780_0))[name = string("op_3784")];
-            tensor<fp16, [1, 3, 128, 256]> var_3785 = mul(x = var_3784, y = sin_1)[name = string("op_3785")];
+            tensor<fp16, [1, 3, 128, 256]> var_3785 = mul(x = var_3784, y = sin_1_quantized)[name = string("op_3785")];
             tensor<fp16, [1, 3, 128, 256]> q_95 = add(x = var_3779, y = var_3785)[name = string("q_95")];
-            tensor<fp16, [1, 1, 128, 256]> var_3787 = mul(x = out_185_cast_fp16, y = cos_1)[name = string("op_3787")];
+            tensor<fp16, [1, 1, 128, 256]> var_3787 = mul(x = out_185_cast_fp16, y = cos_1_quantized)[name = string("op_3787")];
             tensor<int32, [2]> var_3788_split_sizes_0 = const()[name = string("op_3788_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_3788_axis_0 = const()[name = string("op_3788_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 128]> var_3788_0, tensor<fp16, [1, 1, 128, 128]> var_3788_1 = split(axis = var_3788_axis_0, split_sizes = var_3788_split_sizes_0, x = out_185_cast_fp16)[name = string("op_3788")];
@@ -3080,7 +3085,7 @@ program(1.3)
             tensor<fp16, [1, 1, 128, 128]> var_3790 = mul(x = var_3788_1, y = const_217_promoted)[name = string("op_3790")];
             bool var_3792_interleave_0 = const()[name = string("op_3792_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 1, 128, 256]> var_3792 = concat(axis = var_23, interleave = var_3792_interleave_0, values = (var_3790, var_3788_0))[name = string("op_3792")];
-            tensor<fp16, [1, 1, 128, 256]> var_3793 = mul(x = var_3792, y = sin_1)[name = string("op_3793")];
+            tensor<fp16, [1, 1, 128, 256]> var_3793 = mul(x = var_3792, y = sin_1_quantized)[name = string("op_3793")];
             tensor<fp16, [1, 1, 128, 256]> hidden_states_181 = add(x = var_3787, y = var_3793)[name = string("hidden_states_181")];
             tensor<int32, [1]> hidden_states_183_axes_0 = const()[name = string("hidden_states_183_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 1, 1, 128, 256]> hidden_states_183 = expand_dims(axes = hidden_states_183_axes_0, x = hidden_states_181)[name = string("hidden_states_183")];
@@ -3110,14 +3115,14 @@ program(1.3)
             tensor<fp16, [1, 128, 3, 256]> var_3820 = transpose(perm = var_3819, x = var_3817_cast_fp16)[name = string("transpose_76")];
             tensor<fp16, [1, 128, 768]> attn_out_93 = reshape(shape = var_3822, x = var_3820)[name = string("attn_out_93")];
             tensor<int32, [3]> var_3824 = const()[name = string("op_3824"), val = tensor<int32, [3]>([0, 2, 1])];
-            tensor<fp16, [768, 768, 1]> squeeze_15 = const()[name = string("squeeze_15"), val = tensor<fp16, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604769664)))];
+            tensor<fp16, [768, 768, 1]> squeeze_15_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303226368))), scale = tensor<fp16, [768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303816256))))[name = string("squeeze_15_quantized")];
             string var_3833_pad_type_0 = const()[name = string("op_3833_pad_type_0"), val = string("valid")];
             int32 var_3833_groups_0 = const()[name = string("op_3833_groups_0"), val = int32(1)];
             tensor<int32, [1]> var_3833_strides_0 = const()[name = string("op_3833_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> var_3833_pad_0 = const()[name = string("op_3833_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> var_3833_dilations_0 = const()[name = string("op_3833_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, [1, 768, 128]> var_3825 = transpose(perm = var_3824, x = attn_out_93)[name = string("transpose_75")];
-            tensor<fp16, [1, 768, 128]> var_3833 = conv(dilations = var_3833_dilations_0, groups = var_3833_groups_0, pad = var_3833_pad_0, pad_type = var_3833_pad_type_0, strides = var_3833_strides_0, weight = squeeze_15, x = var_3825)[name = string("op_3833")];
+            tensor<fp16, [1, 768, 128]> var_3833 = conv(dilations = var_3833_dilations_0, groups = var_3833_groups_0, pad = var_3833_pad_0, pad_type = var_3833_pad_type_0, strides = var_3833_strides_0, weight = squeeze_15_quantized, x = var_3825)[name = string("op_3833")];
             tensor<int32, [3]> var_3834 = const()[name = string("op_3834"), val = tensor<int32, [3]>([0, 2, 1])];
             fp16 const_218_promoted_to_fp16 = const()[name = string("const_218_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 128, 768]> x_249 = transpose(perm = var_3834, x = var_3833)[name = string("transpose_74")];
@@ -3129,7 +3134,7 @@ program(1.3)
             tensor<int32, [2]> var_3843_split_sizes_0 = const()[name = string("op_3843_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_3843_axis_0 = const()[name = string("op_3843_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_3843_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_3843_cast_fp16_1 = split(axis = var_3843_axis_0, split_sizes = var_3843_split_sizes_0, x = normed_435_cast_fp16)[name = string("op_3843_cast_fp16")];
-            tensor<fp16, [768]> var_3847_to_fp16 = const()[name = string("op_3847_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(605949376)))];
+            tensor<fp16, [768]> var_3847_to_fp16 = const()[name = string("op_3847_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303817856)))];
             tensor<fp16, [1, 128, 768]> out_187_cast_fp16 = mul(x = var_3843_cast_fp16_0, y = var_3847_to_fp16)[name = string("out_187_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_251_cast_fp16 = add(x = x_241_cast_fp16, y = out_187_cast_fp16)[name = string("x_251_cast_fp16")];
             fp16 const_220_promoted_to_fp16 = const()[name = string("const_220_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -3141,7 +3146,7 @@ program(1.3)
             tensor<int32, [2]> var_3859_split_sizes_0 = const()[name = string("op_3859_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_3859_axis_0 = const()[name = string("op_3859_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_3859_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_3859_cast_fp16_1 = split(axis = var_3859_axis_0, split_sizes = var_3859_split_sizes_0, x = normed_439_cast_fp16)[name = string("op_3859_cast_fp16")];
-            tensor<fp16, [768]> var_3863_to_fp16 = const()[name = string("op_3863_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(605950976)))];
+            tensor<fp16, [768]> var_3863_to_fp16 = const()[name = string("op_3863_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303819456)))];
             tensor<fp16, [1, 128, 768]> out_189_cast_fp16 = mul(x = var_3859_cast_fp16_0, y = var_3863_to_fp16)[name = string("out_189_cast_fp16")];
             tensor<int32, [3]> var_3870 = const()[name = string("op_3870"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> input_315_axes_0 = const()[name = string("input_315_axes_0"), val = tensor<int32, [1]>([2])];
@@ -3152,13 +3157,13 @@ program(1.3)
             tensor<int32, [4]> gate_61_pad_0 = const()[name = string("gate_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> gate_61_dilations_0 = const()[name = string("gate_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 gate_61_groups_0 = const()[name = string("gate_61_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> gate_61 = conv(dilations = gate_61_dilations_0, groups = gate_61_groups_0, pad = gate_61_pad_0, pad_type = gate_61_pad_type_0, strides = gate_61_strides_0, weight = encoder_layers_15_mlp_gate_proj_weight, x = input_315)[name = string("gate_61")];
+            tensor<fp16, [1, 1152, 1, 128]> gate_61 = conv(dilations = gate_61_dilations_0, groups = gate_61_groups_0, pad = gate_61_pad_0, pad_type = gate_61_pad_type_0, strides = gate_61_strides_0, weight = encoder_layers_15_mlp_gate_proj_weight_quantized, x = input_315)[name = string("gate_61")];
             string up_31_pad_type_0 = const()[name = string("up_31_pad_type_0"), val = string("valid")];
             tensor<int32, [2]> up_31_strides_0 = const()[name = string("up_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> up_31_pad_0 = const()[name = string("up_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> up_31_dilations_0 = const()[name = string("up_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 up_31_groups_0 = const()[name = string("up_31_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> up_31 = conv(dilations = up_31_dilations_0, groups = up_31_groups_0, pad = up_31_pad_0, pad_type = up_31_pad_type_0, strides = up_31_strides_0, weight = encoder_layers_15_mlp_up_proj_weight, x = input_315)[name = string("up_31")];
+            tensor<fp16, [1, 1152, 1, 128]> up_31 = conv(dilations = up_31_dilations_0, groups = up_31_groups_0, pad = up_31_pad_0, pad_type = up_31_pad_type_0, strides = up_31_strides_0, weight = encoder_layers_15_mlp_up_proj_weight_quantized, x = input_315)[name = string("up_31")];
             string gate_63_mode_0 = const()[name = string("gate_63_mode_0"), val = string("TANH_APPROXIMATION")];
             tensor<fp16, [1, 1152, 1, 128]> gate_63 = gelu(mode = gate_63_mode_0, x = gate_61)[name = string("gate_63")];
             tensor<fp16, [1, 1152, 1, 128]> input_317 = mul(x = gate_63, y = up_31)[name = string("input_317")];
@@ -3167,7 +3172,7 @@ program(1.3)
             tensor<int32, [4]> var_3892_pad_0 = const()[name = string("op_3892_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_3892_dilations_0 = const()[name = string("op_3892_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_3892_groups_0 = const()[name = string("op_3892_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_3892 = conv(dilations = var_3892_dilations_0, groups = var_3892_groups_0, pad = var_3892_pad_0, pad_type = var_3892_pad_type_0, strides = var_3892_strides_0, weight = encoder_layers_15_mlp_down_proj_weight, x = input_317)[name = string("op_3892")];
+            tensor<fp16, [1, 768, 1, 128]> var_3892 = conv(dilations = var_3892_dilations_0, groups = var_3892_groups_0, pad = var_3892_pad_0, pad_type = var_3892_pad_type_0, strides = var_3892_strides_0, weight = encoder_layers_15_mlp_down_proj_weight_quantized, x = input_317)[name = string("op_3892")];
             tensor<int32, [1]> var_3893_axes_0 = const()[name = string("op_3893_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 768, 128]> var_3893 = squeeze(axes = var_3893_axes_0, x = var_3892)[name = string("op_3893")];
             tensor<int32, [3]> var_3894 = const()[name = string("op_3894"), val = tensor<int32, [3]>([0, 2, 1])];
@@ -3181,7 +3186,7 @@ program(1.3)
             tensor<int32, [2]> var_3903_split_sizes_0 = const()[name = string("op_3903_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_3903_axis_0 = const()[name = string("op_3903_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_3903_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_3903_cast_fp16_1 = split(axis = var_3903_axis_0, split_sizes = var_3903_split_sizes_0, x = normed_445_cast_fp16)[name = string("op_3903_cast_fp16")];
-            tensor<fp16, [768]> var_3907_to_fp16 = const()[name = string("op_3907_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(605952576)))];
+            tensor<fp16, [768]> var_3907_to_fp16 = const()[name = string("op_3907_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303821056)))];
             tensor<fp16, [1, 128, 768]> out_191_cast_fp16 = mul(x = var_3903_cast_fp16_0, y = var_3907_to_fp16)[name = string("out_191_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_257_cast_fp16 = add(x = x_251_cast_fp16, y = out_191_cast_fp16)[name = string("x_257_cast_fp16")];
             fp16 const_224_promoted_to_fp16 = const()[name = string("const_224_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -3193,7 +3198,7 @@ program(1.3)
             tensor<int32, [2]> var_3941_split_sizes_0 = const()[name = string("op_3941_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_3941_axis_0 = const()[name = string("op_3941_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_3941_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_3941_cast_fp16_1 = split(axis = var_3941_axis_0, split_sizes = var_3941_split_sizes_0, x = normed_449_cast_fp16)[name = string("op_3941_cast_fp16")];
-            tensor<fp16, [768]> var_3945_to_fp16 = const()[name = string("op_3945_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(605954176)))];
+            tensor<fp16, [768]> var_3945_to_fp16 = const()[name = string("op_3945_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303822656)))];
             tensor<fp16, [1, 128, 768]> out_193_cast_fp16 = mul(x = var_3941_cast_fp16_0, y = var_3945_to_fp16)[name = string("out_193_cast_fp16")];
             tensor<int32, [3]> var_3951 = const()[name = string("op_3951"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> var_3953_axes_0 = const()[name = string("op_3953_axes_0"), val = tensor<int32, [1]>([2])];
@@ -3204,7 +3209,7 @@ program(1.3)
             tensor<int32, [4]> var_3960_pad_0 = const()[name = string("op_3960_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_3960_dilations_0 = const()[name = string("op_3960_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_3960_groups_0 = const()[name = string("op_3960_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_3960 = conv(dilations = var_3960_dilations_0, groups = var_3960_groups_0, pad = var_3960_pad_0, pad_type = var_3960_pad_type_0, strides = var_3960_strides_0, weight = encoder_layers_16_self_attn_q_proj_weight, x = var_3953_cast_fp16)[name = string("op_3960")];
+            tensor<fp16, [1, 768, 1, 128]> var_3960 = conv(dilations = var_3960_dilations_0, groups = var_3960_groups_0, pad = var_3960_pad_0, pad_type = var_3960_pad_type_0, strides = var_3960_strides_0, weight = encoder_layers_16_self_attn_q_proj_weight_quantized, x = var_3953_cast_fp16)[name = string("op_3960")];
             tensor<int32, [4]> var_3961 = const()[name = string("op_3961"), val = tensor<int32, [4]>([1, 3, 256, 128])];
             tensor<fp16, [1, 3, 256, 128]> var_3962 = reshape(shape = var_3961, x = var_3960)[name = string("op_3962")];
             tensor<int32, [4]> var_3963 = const()[name = string("op_3963"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -3213,7 +3218,7 @@ program(1.3)
             tensor<int32, [4]> var_3970_pad_0 = const()[name = string("op_3970_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_3970_dilations_0 = const()[name = string("op_3970_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_3970_groups_0 = const()[name = string("op_3970_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_3970 = conv(dilations = var_3970_dilations_0, groups = var_3970_groups_0, pad = var_3970_pad_0, pad_type = var_3970_pad_type_0, strides = var_3970_strides_0, weight = encoder_layers_16_self_attn_k_proj_weight, x = var_3953_cast_fp16)[name = string("op_3970")];
+            tensor<fp16, [1, 256, 1, 128]> var_3970 = conv(dilations = var_3970_dilations_0, groups = var_3970_groups_0, pad = var_3970_pad_0, pad_type = var_3970_pad_type_0, strides = var_3970_strides_0, weight = encoder_layers_16_self_attn_k_proj_weight_quantized, x = var_3953_cast_fp16)[name = string("op_3970")];
             tensor<int32, [4]> var_3971 = const()[name = string("op_3971"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_3972 = reshape(shape = var_3971, x = var_3970)[name = string("op_3972")];
             tensor<int32, [4]> var_3973 = const()[name = string("op_3973"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -3222,7 +3227,7 @@ program(1.3)
             tensor<int32, [4]> var_3980_pad_0 = const()[name = string("op_3980_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_3980_dilations_0 = const()[name = string("op_3980_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_3980_groups_0 = const()[name = string("op_3980_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_3980 = conv(dilations = var_3980_dilations_0, groups = var_3980_groups_0, pad = var_3980_pad_0, pad_type = var_3980_pad_type_0, strides = var_3980_strides_0, weight = encoder_layers_16_self_attn_v_proj_weight, x = var_3953_cast_fp16)[name = string("op_3980")];
+            tensor<fp16, [1, 256, 1, 128]> var_3980 = conv(dilations = var_3980_dilations_0, groups = var_3980_groups_0, pad = var_3980_pad_0, pad_type = var_3980_pad_type_0, strides = var_3980_strides_0, weight = encoder_layers_16_self_attn_v_proj_weight_quantized, x = var_3953_cast_fp16)[name = string("op_3980")];
             tensor<int32, [4]> var_3981 = const()[name = string("op_3981"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_3982 = reshape(shape = var_3981, x = var_3980)[name = string("op_3982")];
             tensor<int32, [4]> var_3983 = const()[name = string("op_3983"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -3236,7 +3241,7 @@ program(1.3)
             tensor<int32, [2]> var_3994_split_sizes_0 = const()[name = string("op_3994_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_3994_axis_0 = const()[name = string("op_3994_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 256]> var_3994_cast_fp16_0, tensor<fp16, [1, 3, 128, 256]> var_3994_cast_fp16_1 = split(axis = var_3994_axis_0, split_sizes = var_3994_split_sizes_0, x = normed_455_cast_fp16)[name = string("op_3994_cast_fp16")];
-            tensor<fp16, [256]> var_3998_to_fp16 = const()[name = string("op_3998_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(605955776)))];
+            tensor<fp16, [256]> var_3998_to_fp16 = const()[name = string("op_3998_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303824256)))];
             tensor<fp16, [1, 3, 128, 256]> out_195_cast_fp16 = mul(x = var_3994_cast_fp16_0, y = var_3998_to_fp16)[name = string("out_195_cast_fp16")];
             fp16 const_228_promoted_to_fp16 = const()[name = string("const_228_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 1, 128, 256]> k_97 = transpose(perm = var_3973, x = var_3972)[name = string("transpose_69")];
@@ -3248,9 +3253,9 @@ program(1.3)
             tensor<int32, [2]> var_4010_split_sizes_0 = const()[name = string("op_4010_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_4010_axis_0 = const()[name = string("op_4010_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 256]> var_4010_cast_fp16_0, tensor<fp16, [1, 1, 128, 256]> var_4010_cast_fp16_1 = split(axis = var_4010_axis_0, split_sizes = var_4010_split_sizes_0, x = normed_459_cast_fp16)[name = string("op_4010_cast_fp16")];
-            tensor<fp16, [256]> var_4014_to_fp16 = const()[name = string("op_4014_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(605956352)))];
+            tensor<fp16, [256]> var_4014_to_fp16 = const()[name = string("op_4014_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303824832)))];
             tensor<fp16, [1, 1, 128, 256]> out_197_cast_fp16 = mul(x = var_4010_cast_fp16_0, y = var_4014_to_fp16)[name = string("out_197_cast_fp16")];
-            tensor<fp16, [1, 3, 128, 256]> var_4017 = mul(x = out_195_cast_fp16, y = cos_1)[name = string("op_4017")];
+            tensor<fp16, [1, 3, 128, 256]> var_4017 = mul(x = out_195_cast_fp16, y = cos_1_quantized)[name = string("op_4017")];
             tensor<int32, [2]> var_4018_split_sizes_0 = const()[name = string("op_4018_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_4018_axis_0 = const()[name = string("op_4018_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 128]> var_4018_0, tensor<fp16, [1, 3, 128, 128]> var_4018_1 = split(axis = var_4018_axis_0, split_sizes = var_4018_split_sizes_0, x = out_195_cast_fp16)[name = string("op_4018")];
@@ -3258,9 +3263,9 @@ program(1.3)
             tensor<fp16, [1, 3, 128, 128]> var_4020 = mul(x = var_4018_1, y = const_230_promoted)[name = string("op_4020")];
             bool var_4022_interleave_0 = const()[name = string("op_4022_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 3, 128, 256]> var_4022 = concat(axis = var_23, interleave = var_4022_interleave_0, values = (var_4020, var_4018_0))[name = string("op_4022")];
-            tensor<fp16, [1, 3, 128, 256]> var_4023 = mul(x = var_4022, y = sin_1)[name = string("op_4023")];
+            tensor<fp16, [1, 3, 128, 256]> var_4023 = mul(x = var_4022, y = sin_1_quantized)[name = string("op_4023")];
             tensor<fp16, [1, 3, 128, 256]> q_101 = add(x = var_4017, y = var_4023)[name = string("q_101")];
-            tensor<fp16, [1, 1, 128, 256]> var_4025 = mul(x = out_197_cast_fp16, y = cos_1)[name = string("op_4025")];
+            tensor<fp16, [1, 1, 128, 256]> var_4025 = mul(x = out_197_cast_fp16, y = cos_1_quantized)[name = string("op_4025")];
             tensor<int32, [2]> var_4026_split_sizes_0 = const()[name = string("op_4026_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_4026_axis_0 = const()[name = string("op_4026_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 128]> var_4026_0, tensor<fp16, [1, 1, 128, 128]> var_4026_1 = split(axis = var_4026_axis_0, split_sizes = var_4026_split_sizes_0, x = out_197_cast_fp16)[name = string("op_4026")];
@@ -3268,7 +3273,7 @@ program(1.3)
             tensor<fp16, [1, 1, 128, 128]> var_4028 = mul(x = var_4026_1, y = const_231_promoted)[name = string("op_4028")];
             bool var_4030_interleave_0 = const()[name = string("op_4030_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 1, 128, 256]> var_4030 = concat(axis = var_23, interleave = var_4030_interleave_0, values = (var_4028, var_4026_0))[name = string("op_4030")];
-            tensor<fp16, [1, 1, 128, 256]> var_4031 = mul(x = var_4030, y = sin_1)[name = string("op_4031")];
+            tensor<fp16, [1, 1, 128, 256]> var_4031 = mul(x = var_4030, y = sin_1_quantized)[name = string("op_4031")];
             tensor<fp16, [1, 1, 128, 256]> hidden_states_193 = add(x = var_4025, y = var_4031)[name = string("hidden_states_193")];
             tensor<int32, [1]> hidden_states_195_axes_0 = const()[name = string("hidden_states_195_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 1, 1, 128, 256]> hidden_states_195 = expand_dims(axes = hidden_states_195_axes_0, x = hidden_states_193)[name = string("hidden_states_195")];
@@ -3298,14 +3303,14 @@ program(1.3)
             tensor<fp16, [1, 128, 3, 256]> var_4058 = transpose(perm = var_4057, x = var_4055_cast_fp16)[name = string("transpose_67")];
             tensor<fp16, [1, 128, 768]> attn_out_99 = reshape(shape = var_4060, x = var_4058)[name = string("attn_out_99")];
             tensor<int32, [3]> var_4062 = const()[name = string("op_4062"), val = tensor<int32, [3]>([0, 2, 1])];
-            tensor<fp16, [768, 768, 1]> squeeze_16 = const()[name = string("squeeze_16"), val = tensor<fp16, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(605956928)))];
+            tensor<fp16, [768, 768, 1]> squeeze_16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303825408))), scale = tensor<fp16, [768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304415296))))[name = string("squeeze_16_quantized")];
             string var_4071_pad_type_0 = const()[name = string("op_4071_pad_type_0"), val = string("valid")];
             int32 var_4071_groups_0 = const()[name = string("op_4071_groups_0"), val = int32(1)];
             tensor<int32, [1]> var_4071_strides_0 = const()[name = string("op_4071_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> var_4071_pad_0 = const()[name = string("op_4071_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> var_4071_dilations_0 = const()[name = string("op_4071_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, [1, 768, 128]> var_4063 = transpose(perm = var_4062, x = attn_out_99)[name = string("transpose_66")];
-            tensor<fp16, [1, 768, 128]> var_4071 = conv(dilations = var_4071_dilations_0, groups = var_4071_groups_0, pad = var_4071_pad_0, pad_type = var_4071_pad_type_0, strides = var_4071_strides_0, weight = squeeze_16, x = var_4063)[name = string("op_4071")];
+            tensor<fp16, [1, 768, 128]> var_4071 = conv(dilations = var_4071_dilations_0, groups = var_4071_groups_0, pad = var_4071_pad_0, pad_type = var_4071_pad_type_0, strides = var_4071_strides_0, weight = squeeze_16_quantized, x = var_4063)[name = string("op_4071")];
             tensor<int32, [3]> var_4072 = const()[name = string("op_4072"), val = tensor<int32, [3]>([0, 2, 1])];
             fp16 const_232_promoted_to_fp16 = const()[name = string("const_232_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 128, 768]> x_265 = transpose(perm = var_4072, x = var_4071)[name = string("transpose_65")];
@@ -3317,7 +3322,7 @@ program(1.3)
             tensor<int32, [2]> var_4081_split_sizes_0 = const()[name = string("op_4081_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_4081_axis_0 = const()[name = string("op_4081_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_4081_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_4081_cast_fp16_1 = split(axis = var_4081_axis_0, split_sizes = var_4081_split_sizes_0, x = normed_463_cast_fp16)[name = string("op_4081_cast_fp16")];
-            tensor<fp16, [768]> var_4085_to_fp16 = const()[name = string("op_4085_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(607136640)))];
+            tensor<fp16, [768]> var_4085_to_fp16 = const()[name = string("op_4085_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304416896)))];
             tensor<fp16, [1, 128, 768]> out_199_cast_fp16 = mul(x = var_4081_cast_fp16_0, y = var_4085_to_fp16)[name = string("out_199_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_267_cast_fp16 = add(x = x_257_cast_fp16, y = out_199_cast_fp16)[name = string("x_267_cast_fp16")];
             fp16 const_234_promoted_to_fp16 = const()[name = string("const_234_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -3329,7 +3334,7 @@ program(1.3)
             tensor<int32, [2]> var_4097_split_sizes_0 = const()[name = string("op_4097_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_4097_axis_0 = const()[name = string("op_4097_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_4097_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_4097_cast_fp16_1 = split(axis = var_4097_axis_0, split_sizes = var_4097_split_sizes_0, x = normed_467_cast_fp16)[name = string("op_4097_cast_fp16")];
-            tensor<fp16, [768]> var_4101_to_fp16 = const()[name = string("op_4101_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(607138240)))];
+            tensor<fp16, [768]> var_4101_to_fp16 = const()[name = string("op_4101_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304418496)))];
             tensor<fp16, [1, 128, 768]> out_201_cast_fp16 = mul(x = var_4097_cast_fp16_0, y = var_4101_to_fp16)[name = string("out_201_cast_fp16")];
             tensor<int32, [3]> var_4108 = const()[name = string("op_4108"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> input_335_axes_0 = const()[name = string("input_335_axes_0"), val = tensor<int32, [1]>([2])];
@@ -3340,13 +3345,13 @@ program(1.3)
             tensor<int32, [4]> gate_65_pad_0 = const()[name = string("gate_65_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> gate_65_dilations_0 = const()[name = string("gate_65_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 gate_65_groups_0 = const()[name = string("gate_65_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> gate_65 = conv(dilations = gate_65_dilations_0, groups = gate_65_groups_0, pad = gate_65_pad_0, pad_type = gate_65_pad_type_0, strides = gate_65_strides_0, weight = encoder_layers_16_mlp_gate_proj_weight, x = input_335)[name = string("gate_65")];
+            tensor<fp16, [1, 1152, 1, 128]> gate_65 = conv(dilations = gate_65_dilations_0, groups = gate_65_groups_0, pad = gate_65_pad_0, pad_type = gate_65_pad_type_0, strides = gate_65_strides_0, weight = encoder_layers_16_mlp_gate_proj_weight_quantized, x = input_335)[name = string("gate_65")];
             string up_33_pad_type_0 = const()[name = string("up_33_pad_type_0"), val = string("valid")];
             tensor<int32, [2]> up_33_strides_0 = const()[name = string("up_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> up_33_pad_0 = const()[name = string("up_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> up_33_dilations_0 = const()[name = string("up_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 up_33_groups_0 = const()[name = string("up_33_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> up_33 = conv(dilations = up_33_dilations_0, groups = up_33_groups_0, pad = up_33_pad_0, pad_type = up_33_pad_type_0, strides = up_33_strides_0, weight = encoder_layers_16_mlp_up_proj_weight, x = input_335)[name = string("up_33")];
+            tensor<fp16, [1, 1152, 1, 128]> up_33 = conv(dilations = up_33_dilations_0, groups = up_33_groups_0, pad = up_33_pad_0, pad_type = up_33_pad_type_0, strides = up_33_strides_0, weight = encoder_layers_16_mlp_up_proj_weight_quantized, x = input_335)[name = string("up_33")];
             string gate_67_mode_0 = const()[name = string("gate_67_mode_0"), val = string("TANH_APPROXIMATION")];
             tensor<fp16, [1, 1152, 1, 128]> gate_67 = gelu(mode = gate_67_mode_0, x = gate_65)[name = string("gate_67")];
             tensor<fp16, [1, 1152, 1, 128]> input_337 = mul(x = gate_67, y = up_33)[name = string("input_337")];
@@ -3355,7 +3360,7 @@ program(1.3)
             tensor<int32, [4]> var_4130_pad_0 = const()[name = string("op_4130_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_4130_dilations_0 = const()[name = string("op_4130_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_4130_groups_0 = const()[name = string("op_4130_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_4130 = conv(dilations = var_4130_dilations_0, groups = var_4130_groups_0, pad = var_4130_pad_0, pad_type = var_4130_pad_type_0, strides = var_4130_strides_0, weight = encoder_layers_16_mlp_down_proj_weight, x = input_337)[name = string("op_4130")];
+            tensor<fp16, [1, 768, 1, 128]> var_4130 = conv(dilations = var_4130_dilations_0, groups = var_4130_groups_0, pad = var_4130_pad_0, pad_type = var_4130_pad_type_0, strides = var_4130_strides_0, weight = encoder_layers_16_mlp_down_proj_weight_quantized, x = input_337)[name = string("op_4130")];
             tensor<int32, [1]> var_4131_axes_0 = const()[name = string("op_4131_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 768, 128]> var_4131 = squeeze(axes = var_4131_axes_0, x = var_4130)[name = string("op_4131")];
             tensor<int32, [3]> var_4132 = const()[name = string("op_4132"), val = tensor<int32, [3]>([0, 2, 1])];
@@ -3369,7 +3374,7 @@ program(1.3)
             tensor<int32, [2]> var_4141_split_sizes_0 = const()[name = string("op_4141_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_4141_axis_0 = const()[name = string("op_4141_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_4141_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_4141_cast_fp16_1 = split(axis = var_4141_axis_0, split_sizes = var_4141_split_sizes_0, x = normed_473_cast_fp16)[name = string("op_4141_cast_fp16")];
-            tensor<fp16, [768]> var_4145_to_fp16 = const()[name = string("op_4145_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(607139840)))];
+            tensor<fp16, [768]> var_4145_to_fp16 = const()[name = string("op_4145_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304420096)))];
             tensor<fp16, [1, 128, 768]> out_203_cast_fp16 = mul(x = var_4141_cast_fp16_0, y = var_4145_to_fp16)[name = string("out_203_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_273_cast_fp16 = add(x = x_267_cast_fp16, y = out_203_cast_fp16)[name = string("x_273_cast_fp16")];
             fp16 const_238_promoted_to_fp16 = const()[name = string("const_238_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -3381,7 +3386,7 @@ program(1.3)
             tensor<int32, [2]> var_4179_split_sizes_0 = const()[name = string("op_4179_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_4179_axis_0 = const()[name = string("op_4179_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_4179_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_4179_cast_fp16_1 = split(axis = var_4179_axis_0, split_sizes = var_4179_split_sizes_0, x = normed_477_cast_fp16)[name = string("op_4179_cast_fp16")];
-            tensor<fp16, [768]> var_4183_to_fp16 = const()[name = string("op_4183_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(607141440)))];
+            tensor<fp16, [768]> var_4183_to_fp16 = const()[name = string("op_4183_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304421696)))];
             tensor<fp16, [1, 128, 768]> out_205_cast_fp16 = mul(x = var_4179_cast_fp16_0, y = var_4183_to_fp16)[name = string("out_205_cast_fp16")];
             tensor<int32, [3]> var_4189 = const()[name = string("op_4189"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> var_4191_axes_0 = const()[name = string("op_4191_axes_0"), val = tensor<int32, [1]>([2])];
@@ -3392,7 +3397,7 @@ program(1.3)
             tensor<int32, [4]> var_4198_pad_0 = const()[name = string("op_4198_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_4198_dilations_0 = const()[name = string("op_4198_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_4198_groups_0 = const()[name = string("op_4198_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_4198 = conv(dilations = var_4198_dilations_0, groups = var_4198_groups_0, pad = var_4198_pad_0, pad_type = var_4198_pad_type_0, strides = var_4198_strides_0, weight = encoder_layers_17_self_attn_q_proj_weight, x = var_4191_cast_fp16)[name = string("op_4198")];
+            tensor<fp16, [1, 768, 1, 128]> var_4198 = conv(dilations = var_4198_dilations_0, groups = var_4198_groups_0, pad = var_4198_pad_0, pad_type = var_4198_pad_type_0, strides = var_4198_strides_0, weight = encoder_layers_17_self_attn_q_proj_weight_quantized, x = var_4191_cast_fp16)[name = string("op_4198")];
             tensor<int32, [4]> var_4199 = const()[name = string("op_4199"), val = tensor<int32, [4]>([1, 3, 256, 128])];
             tensor<fp16, [1, 3, 256, 128]> var_4200 = reshape(shape = var_4199, x = var_4198)[name = string("op_4200")];
             tensor<int32, [4]> var_4201 = const()[name = string("op_4201"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -3401,7 +3406,7 @@ program(1.3)
             tensor<int32, [4]> var_4208_pad_0 = const()[name = string("op_4208_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_4208_dilations_0 = const()[name = string("op_4208_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_4208_groups_0 = const()[name = string("op_4208_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_4208 = conv(dilations = var_4208_dilations_0, groups = var_4208_groups_0, pad = var_4208_pad_0, pad_type = var_4208_pad_type_0, strides = var_4208_strides_0, weight = encoder_layers_17_self_attn_k_proj_weight, x = var_4191_cast_fp16)[name = string("op_4208")];
+            tensor<fp16, [1, 256, 1, 128]> var_4208 = conv(dilations = var_4208_dilations_0, groups = var_4208_groups_0, pad = var_4208_pad_0, pad_type = var_4208_pad_type_0, strides = var_4208_strides_0, weight = encoder_layers_17_self_attn_k_proj_weight_quantized, x = var_4191_cast_fp16)[name = string("op_4208")];
             tensor<int32, [4]> var_4209 = const()[name = string("op_4209"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_4210 = reshape(shape = var_4209, x = var_4208)[name = string("op_4210")];
             tensor<int32, [4]> var_4211 = const()[name = string("op_4211"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -3410,7 +3415,7 @@ program(1.3)
             tensor<int32, [4]> var_4218_pad_0 = const()[name = string("op_4218_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_4218_dilations_0 = const()[name = string("op_4218_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_4218_groups_0 = const()[name = string("op_4218_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_4218 = conv(dilations = var_4218_dilations_0, groups = var_4218_groups_0, pad = var_4218_pad_0, pad_type = var_4218_pad_type_0, strides = var_4218_strides_0, weight = encoder_layers_17_self_attn_v_proj_weight, x = var_4191_cast_fp16)[name = string("op_4218")];
+            tensor<fp16, [1, 256, 1, 128]> var_4218 = conv(dilations = var_4218_dilations_0, groups = var_4218_groups_0, pad = var_4218_pad_0, pad_type = var_4218_pad_type_0, strides = var_4218_strides_0, weight = encoder_layers_17_self_attn_v_proj_weight_quantized, x = var_4191_cast_fp16)[name = string("op_4218")];
             tensor<int32, [4]> var_4219 = const()[name = string("op_4219"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_4220 = reshape(shape = var_4219, x = var_4218)[name = string("op_4220")];
             tensor<int32, [4]> var_4221 = const()[name = string("op_4221"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -3424,7 +3429,7 @@ program(1.3)
             tensor<int32, [2]> var_4232_split_sizes_0 = const()[name = string("op_4232_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_4232_axis_0 = const()[name = string("op_4232_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 256]> var_4232_cast_fp16_0, tensor<fp16, [1, 3, 128, 256]> var_4232_cast_fp16_1 = split(axis = var_4232_axis_0, split_sizes = var_4232_split_sizes_0, x = normed_483_cast_fp16)[name = string("op_4232_cast_fp16")];
-            tensor<fp16, [256]> var_4236_to_fp16 = const()[name = string("op_4236_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(607143040)))];
+            tensor<fp16, [256]> var_4236_to_fp16 = const()[name = string("op_4236_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304423296)))];
             tensor<fp16, [1, 3, 128, 256]> out_207_cast_fp16 = mul(x = var_4232_cast_fp16_0, y = var_4236_to_fp16)[name = string("out_207_cast_fp16")];
             fp16 const_242_promoted_to_fp16 = const()[name = string("const_242_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 1, 128, 256]> k_103 = transpose(perm = var_4211, x = var_4210)[name = string("transpose_60")];
@@ -3436,9 +3441,9 @@ program(1.3)
             tensor<int32, [2]> var_4248_split_sizes_0 = const()[name = string("op_4248_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_4248_axis_0 = const()[name = string("op_4248_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 256]> var_4248_cast_fp16_0, tensor<fp16, [1, 1, 128, 256]> var_4248_cast_fp16_1 = split(axis = var_4248_axis_0, split_sizes = var_4248_split_sizes_0, x = normed_487_cast_fp16)[name = string("op_4248_cast_fp16")];
-            tensor<fp16, [256]> var_4252_to_fp16 = const()[name = string("op_4252_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(607143616)))];
+            tensor<fp16, [256]> var_4252_to_fp16 = const()[name = string("op_4252_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304423872)))];
             tensor<fp16, [1, 1, 128, 256]> out_209_cast_fp16 = mul(x = var_4248_cast_fp16_0, y = var_4252_to_fp16)[name = string("out_209_cast_fp16")];
-            tensor<fp16, [1, 3, 128, 256]> var_4255 = mul(x = out_207_cast_fp16, y = cos)[name = string("op_4255")];
+            tensor<fp16, [1, 3, 128, 256]> var_4255 = mul(x = out_207_cast_fp16, y = cos_quantized)[name = string("op_4255")];
             tensor<int32, [2]> var_4256_split_sizes_0 = const()[name = string("op_4256_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_4256_axis_0 = const()[name = string("op_4256_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 128]> var_4256_0, tensor<fp16, [1, 3, 128, 128]> var_4256_1 = split(axis = var_4256_axis_0, split_sizes = var_4256_split_sizes_0, x = out_207_cast_fp16)[name = string("op_4256")];
@@ -3446,9 +3451,9 @@ program(1.3)
             tensor<fp16, [1, 3, 128, 128]> var_4258 = mul(x = var_4256_1, y = const_244_promoted)[name = string("op_4258")];
             bool var_4260_interleave_0 = const()[name = string("op_4260_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 3, 128, 256]> var_4260 = concat(axis = var_23, interleave = var_4260_interleave_0, values = (var_4258, var_4256_0))[name = string("op_4260")];
-            tensor<fp16, [1, 3, 128, 256]> var_4261 = mul(x = var_4260, y = sin)[name = string("op_4261")];
+            tensor<fp16, [1, 3, 128, 256]> var_4261 = mul(x = var_4260, y = sin_quantized)[name = string("op_4261")];
             tensor<fp16, [1, 3, 128, 256]> q_107 = add(x = var_4255, y = var_4261)[name = string("q_107")];
-            tensor<fp16, [1, 1, 128, 256]> var_4263 = mul(x = out_209_cast_fp16, y = cos)[name = string("op_4263")];
+            tensor<fp16, [1, 1, 128, 256]> var_4263 = mul(x = out_209_cast_fp16, y = cos_quantized)[name = string("op_4263")];
             tensor<int32, [2]> var_4264_split_sizes_0 = const()[name = string("op_4264_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_4264_axis_0 = const()[name = string("op_4264_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 128]> var_4264_0, tensor<fp16, [1, 1, 128, 128]> var_4264_1 = split(axis = var_4264_axis_0, split_sizes = var_4264_split_sizes_0, x = out_209_cast_fp16)[name = string("op_4264")];
@@ -3456,7 +3461,7 @@ program(1.3)
             tensor<fp16, [1, 1, 128, 128]> var_4266 = mul(x = var_4264_1, y = const_245_promoted)[name = string("op_4266")];
             bool var_4268_interleave_0 = const()[name = string("op_4268_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 1, 128, 256]> var_4268 = concat(axis = var_23, interleave = var_4268_interleave_0, values = (var_4266, var_4264_0))[name = string("op_4268")];
-            tensor<fp16, [1, 1, 128, 256]> var_4269 = mul(x = var_4268, y = sin)[name = string("op_4269")];
+            tensor<fp16, [1, 1, 128, 256]> var_4269 = mul(x = var_4268, y = sin_quantized)[name = string("op_4269")];
             tensor<fp16, [1, 1, 128, 256]> hidden_states_205 = add(x = var_4263, y = var_4269)[name = string("hidden_states_205")];
             tensor<int32, [1]> hidden_states_207_axes_0 = const()[name = string("hidden_states_207_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 1, 1, 128, 256]> hidden_states_207 = expand_dims(axes = hidden_states_207_axes_0, x = hidden_states_205)[name = string("hidden_states_207")];
@@ -3486,14 +3491,14 @@ program(1.3)
             tensor<fp16, [1, 128, 3, 256]> var_4296 = transpose(perm = var_4295, x = var_4293_cast_fp16)[name = string("transpose_58")];
             tensor<fp16, [1, 128, 768]> attn_out_105 = reshape(shape = var_4298, x = var_4296)[name = string("attn_out_105")];
             tensor<int32, [3]> var_4300 = const()[name = string("op_4300"), val = tensor<int32, [3]>([0, 2, 1])];
-            tensor<fp16, [768, 768, 1]> squeeze_17 = const()[name = string("squeeze_17"), val = tensor<fp16, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(607144192)))];
+            tensor<fp16, [768, 768, 1]> squeeze_17_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304424448))), scale = tensor<fp16, [768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305014336))))[name = string("squeeze_17_quantized")];
             string var_4309_pad_type_0 = const()[name = string("op_4309_pad_type_0"), val = string("valid")];
             int32 var_4309_groups_0 = const()[name = string("op_4309_groups_0"), val = int32(1)];
             tensor<int32, [1]> var_4309_strides_0 = const()[name = string("op_4309_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> var_4309_pad_0 = const()[name = string("op_4309_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> var_4309_dilations_0 = const()[name = string("op_4309_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, [1, 768, 128]> var_4301 = transpose(perm = var_4300, x = attn_out_105)[name = string("transpose_57")];
-            tensor<fp16, [1, 768, 128]> var_4309 = conv(dilations = var_4309_dilations_0, groups = var_4309_groups_0, pad = var_4309_pad_0, pad_type = var_4309_pad_type_0, strides = var_4309_strides_0, weight = squeeze_17, x = var_4301)[name = string("op_4309")];
+            tensor<fp16, [1, 768, 128]> var_4309 = conv(dilations = var_4309_dilations_0, groups = var_4309_groups_0, pad = var_4309_pad_0, pad_type = var_4309_pad_type_0, strides = var_4309_strides_0, weight = squeeze_17_quantized, x = var_4301)[name = string("op_4309")];
             tensor<int32, [3]> var_4310 = const()[name = string("op_4310"), val = tensor<int32, [3]>([0, 2, 1])];
             fp16 const_246_promoted_to_fp16 = const()[name = string("const_246_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 128, 768]> x_281 = transpose(perm = var_4310, x = var_4309)[name = string("transpose_56")];
@@ -3505,7 +3510,7 @@ program(1.3)
             tensor<int32, [2]> var_4319_split_sizes_0 = const()[name = string("op_4319_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_4319_axis_0 = const()[name = string("op_4319_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_4319_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_4319_cast_fp16_1 = split(axis = var_4319_axis_0, split_sizes = var_4319_split_sizes_0, x = normed_491_cast_fp16)[name = string("op_4319_cast_fp16")];
-            tensor<fp16, [768]> var_4323_to_fp16 = const()[name = string("op_4323_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(608323904)))];
+            tensor<fp16, [768]> var_4323_to_fp16 = const()[name = string("op_4323_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305015936)))];
             tensor<fp16, [1, 128, 768]> out_211_cast_fp16 = mul(x = var_4319_cast_fp16_0, y = var_4323_to_fp16)[name = string("out_211_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_283_cast_fp16 = add(x = x_273_cast_fp16, y = out_211_cast_fp16)[name = string("x_283_cast_fp16")];
             fp16 const_248_promoted_to_fp16 = const()[name = string("const_248_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -3517,7 +3522,7 @@ program(1.3)
             tensor<int32, [2]> var_4335_split_sizes_0 = const()[name = string("op_4335_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_4335_axis_0 = const()[name = string("op_4335_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_4335_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_4335_cast_fp16_1 = split(axis = var_4335_axis_0, split_sizes = var_4335_split_sizes_0, x = normed_495_cast_fp16)[name = string("op_4335_cast_fp16")];
-            tensor<fp16, [768]> var_4339_to_fp16 = const()[name = string("op_4339_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(608325504)))];
+            tensor<fp16, [768]> var_4339_to_fp16 = const()[name = string("op_4339_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305017536)))];
             tensor<fp16, [1, 128, 768]> out_213_cast_fp16 = mul(x = var_4335_cast_fp16_0, y = var_4339_to_fp16)[name = string("out_213_cast_fp16")];
             tensor<int32, [3]> var_4346 = const()[name = string("op_4346"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> input_355_axes_0 = const()[name = string("input_355_axes_0"), val = tensor<int32, [1]>([2])];
@@ -3528,13 +3533,13 @@ program(1.3)
             tensor<int32, [4]> gate_69_pad_0 = const()[name = string("gate_69_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> gate_69_dilations_0 = const()[name = string("gate_69_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 gate_69_groups_0 = const()[name = string("gate_69_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> gate_69 = conv(dilations = gate_69_dilations_0, groups = gate_69_groups_0, pad = gate_69_pad_0, pad_type = gate_69_pad_type_0, strides = gate_69_strides_0, weight = encoder_layers_17_mlp_gate_proj_weight, x = input_355)[name = string("gate_69")];
+            tensor<fp16, [1, 1152, 1, 128]> gate_69 = conv(dilations = gate_69_dilations_0, groups = gate_69_groups_0, pad = gate_69_pad_0, pad_type = gate_69_pad_type_0, strides = gate_69_strides_0, weight = encoder_layers_17_mlp_gate_proj_weight_quantized, x = input_355)[name = string("gate_69")];
             string up_35_pad_type_0 = const()[name = string("up_35_pad_type_0"), val = string("valid")];
             tensor<int32, [2]> up_35_strides_0 = const()[name = string("up_35_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> up_35_pad_0 = const()[name = string("up_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> up_35_dilations_0 = const()[name = string("up_35_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 up_35_groups_0 = const()[name = string("up_35_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> up_35 = conv(dilations = up_35_dilations_0, groups = up_35_groups_0, pad = up_35_pad_0, pad_type = up_35_pad_type_0, strides = up_35_strides_0, weight = encoder_layers_17_mlp_up_proj_weight, x = input_355)[name = string("up_35")];
+            tensor<fp16, [1, 1152, 1, 128]> up_35 = conv(dilations = up_35_dilations_0, groups = up_35_groups_0, pad = up_35_pad_0, pad_type = up_35_pad_type_0, strides = up_35_strides_0, weight = encoder_layers_17_mlp_up_proj_weight_quantized, x = input_355)[name = string("up_35")];
             string gate_71_mode_0 = const()[name = string("gate_71_mode_0"), val = string("TANH_APPROXIMATION")];
             tensor<fp16, [1, 1152, 1, 128]> gate_71 = gelu(mode = gate_71_mode_0, x = gate_69)[name = string("gate_71")];
             tensor<fp16, [1, 1152, 1, 128]> input_357 = mul(x = gate_71, y = up_35)[name = string("input_357")];
@@ -3543,7 +3548,7 @@ program(1.3)
             tensor<int32, [4]> var_4368_pad_0 = const()[name = string("op_4368_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_4368_dilations_0 = const()[name = string("op_4368_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_4368_groups_0 = const()[name = string("op_4368_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_4368 = conv(dilations = var_4368_dilations_0, groups = var_4368_groups_0, pad = var_4368_pad_0, pad_type = var_4368_pad_type_0, strides = var_4368_strides_0, weight = encoder_layers_17_mlp_down_proj_weight, x = input_357)[name = string("op_4368")];
+            tensor<fp16, [1, 768, 1, 128]> var_4368 = conv(dilations = var_4368_dilations_0, groups = var_4368_groups_0, pad = var_4368_pad_0, pad_type = var_4368_pad_type_0, strides = var_4368_strides_0, weight = encoder_layers_17_mlp_down_proj_weight_quantized, x = input_357)[name = string("op_4368")];
             tensor<int32, [1]> var_4369_axes_0 = const()[name = string("op_4369_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 768, 128]> var_4369 = squeeze(axes = var_4369_axes_0, x = var_4368)[name = string("op_4369")];
             tensor<int32, [3]> var_4370 = const()[name = string("op_4370"), val = tensor<int32, [3]>([0, 2, 1])];
@@ -3557,7 +3562,7 @@ program(1.3)
             tensor<int32, [2]> var_4379_split_sizes_0 = const()[name = string("op_4379_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_4379_axis_0 = const()[name = string("op_4379_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_4379_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_4379_cast_fp16_1 = split(axis = var_4379_axis_0, split_sizes = var_4379_split_sizes_0, x = normed_501_cast_fp16)[name = string("op_4379_cast_fp16")];
-            tensor<fp16, [768]> var_4383_to_fp16 = const()[name = string("op_4383_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(608327104)))];
+            tensor<fp16, [768]> var_4383_to_fp16 = const()[name = string("op_4383_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305019136)))];
             tensor<fp16, [1, 128, 768]> out_215_cast_fp16 = mul(x = var_4379_cast_fp16_0, y = var_4383_to_fp16)[name = string("out_215_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_289_cast_fp16 = add(x = x_283_cast_fp16, y = out_215_cast_fp16)[name = string("x_289_cast_fp16")];
             fp16 const_252_promoted_to_fp16 = const()[name = string("const_252_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -3569,7 +3574,7 @@ program(1.3)
             tensor<int32, [2]> var_4417_split_sizes_0 = const()[name = string("op_4417_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_4417_axis_0 = const()[name = string("op_4417_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_4417_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_4417_cast_fp16_1 = split(axis = var_4417_axis_0, split_sizes = var_4417_split_sizes_0, x = normed_505_cast_fp16)[name = string("op_4417_cast_fp16")];
-            tensor<fp16, [768]> var_4421_to_fp16 = const()[name = string("op_4421_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(608328704)))];
+            tensor<fp16, [768]> var_4421_to_fp16 = const()[name = string("op_4421_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305020736)))];
             tensor<fp16, [1, 128, 768]> out_217_cast_fp16 = mul(x = var_4417_cast_fp16_0, y = var_4421_to_fp16)[name = string("out_217_cast_fp16")];
             tensor<int32, [3]> var_4427 = const()[name = string("op_4427"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> var_4429_axes_0 = const()[name = string("op_4429_axes_0"), val = tensor<int32, [1]>([2])];
@@ -3580,7 +3585,7 @@ program(1.3)
             tensor<int32, [4]> var_4436_pad_0 = const()[name = string("op_4436_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_4436_dilations_0 = const()[name = string("op_4436_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_4436_groups_0 = const()[name = string("op_4436_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_4436 = conv(dilations = var_4436_dilations_0, groups = var_4436_groups_0, pad = var_4436_pad_0, pad_type = var_4436_pad_type_0, strides = var_4436_strides_0, weight = encoder_layers_18_self_attn_q_proj_weight, x = var_4429_cast_fp16)[name = string("op_4436")];
+            tensor<fp16, [1, 768, 1, 128]> var_4436 = conv(dilations = var_4436_dilations_0, groups = var_4436_groups_0, pad = var_4436_pad_0, pad_type = var_4436_pad_type_0, strides = var_4436_strides_0, weight = encoder_layers_18_self_attn_q_proj_weight_quantized, x = var_4429_cast_fp16)[name = string("op_4436")];
             tensor<int32, [4]> var_4437 = const()[name = string("op_4437"), val = tensor<int32, [4]>([1, 3, 256, 128])];
             tensor<fp16, [1, 3, 256, 128]> var_4438 = reshape(shape = var_4437, x = var_4436)[name = string("op_4438")];
             tensor<int32, [4]> var_4439 = const()[name = string("op_4439"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -3589,7 +3594,7 @@ program(1.3)
             tensor<int32, [4]> var_4446_pad_0 = const()[name = string("op_4446_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_4446_dilations_0 = const()[name = string("op_4446_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_4446_groups_0 = const()[name = string("op_4446_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_4446 = conv(dilations = var_4446_dilations_0, groups = var_4446_groups_0, pad = var_4446_pad_0, pad_type = var_4446_pad_type_0, strides = var_4446_strides_0, weight = encoder_layers_18_self_attn_k_proj_weight, x = var_4429_cast_fp16)[name = string("op_4446")];
+            tensor<fp16, [1, 256, 1, 128]> var_4446 = conv(dilations = var_4446_dilations_0, groups = var_4446_groups_0, pad = var_4446_pad_0, pad_type = var_4446_pad_type_0, strides = var_4446_strides_0, weight = encoder_layers_18_self_attn_k_proj_weight_quantized, x = var_4429_cast_fp16)[name = string("op_4446")];
             tensor<int32, [4]> var_4447 = const()[name = string("op_4447"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_4448 = reshape(shape = var_4447, x = var_4446)[name = string("op_4448")];
             tensor<int32, [4]> var_4449 = const()[name = string("op_4449"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -3598,7 +3603,7 @@ program(1.3)
             tensor<int32, [4]> var_4456_pad_0 = const()[name = string("op_4456_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_4456_dilations_0 = const()[name = string("op_4456_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_4456_groups_0 = const()[name = string("op_4456_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_4456 = conv(dilations = var_4456_dilations_0, groups = var_4456_groups_0, pad = var_4456_pad_0, pad_type = var_4456_pad_type_0, strides = var_4456_strides_0, weight = encoder_layers_18_self_attn_v_proj_weight, x = var_4429_cast_fp16)[name = string("op_4456")];
+            tensor<fp16, [1, 256, 1, 128]> var_4456 = conv(dilations = var_4456_dilations_0, groups = var_4456_groups_0, pad = var_4456_pad_0, pad_type = var_4456_pad_type_0, strides = var_4456_strides_0, weight = encoder_layers_18_self_attn_v_proj_weight_quantized, x = var_4429_cast_fp16)[name = string("op_4456")];
             tensor<int32, [4]> var_4457 = const()[name = string("op_4457"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_4458 = reshape(shape = var_4457, x = var_4456)[name = string("op_4458")];
             tensor<int32, [4]> var_4459 = const()[name = string("op_4459"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -3612,7 +3617,7 @@ program(1.3)
             tensor<int32, [2]> var_4470_split_sizes_0 = const()[name = string("op_4470_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_4470_axis_0 = const()[name = string("op_4470_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 256]> var_4470_cast_fp16_0, tensor<fp16, [1, 3, 128, 256]> var_4470_cast_fp16_1 = split(axis = var_4470_axis_0, split_sizes = var_4470_split_sizes_0, x = normed_511_cast_fp16)[name = string("op_4470_cast_fp16")];
-            tensor<fp16, [256]> var_4474_to_fp16 = const()[name = string("op_4474_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(608330304)))];
+            tensor<fp16, [256]> var_4474_to_fp16 = const()[name = string("op_4474_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305022336)))];
             tensor<fp16, [1, 3, 128, 256]> out_219_cast_fp16 = mul(x = var_4470_cast_fp16_0, y = var_4474_to_fp16)[name = string("out_219_cast_fp16")];
             fp16 const_256_promoted_to_fp16 = const()[name = string("const_256_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 1, 128, 256]> k_109 = transpose(perm = var_4449, x = var_4448)[name = string("transpose_51")];
@@ -3624,9 +3629,9 @@ program(1.3)
             tensor<int32, [2]> var_4486_split_sizes_0 = const()[name = string("op_4486_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_4486_axis_0 = const()[name = string("op_4486_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 256]> var_4486_cast_fp16_0, tensor<fp16, [1, 1, 128, 256]> var_4486_cast_fp16_1 = split(axis = var_4486_axis_0, split_sizes = var_4486_split_sizes_0, x = normed_515_cast_fp16)[name = string("op_4486_cast_fp16")];
-            tensor<fp16, [256]> var_4490_to_fp16 = const()[name = string("op_4490_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(608330880)))];
+            tensor<fp16, [256]> var_4490_to_fp16 = const()[name = string("op_4490_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305022912)))];
             tensor<fp16, [1, 1, 128, 256]> out_221_cast_fp16 = mul(x = var_4486_cast_fp16_0, y = var_4490_to_fp16)[name = string("out_221_cast_fp16")];
-            tensor<fp16, [1, 3, 128, 256]> var_4493 = mul(x = out_219_cast_fp16, y = cos_1)[name = string("op_4493")];
+            tensor<fp16, [1, 3, 128, 256]> var_4493 = mul(x = out_219_cast_fp16, y = cos_1_quantized)[name = string("op_4493")];
             tensor<int32, [2]> var_4494_split_sizes_0 = const()[name = string("op_4494_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_4494_axis_0 = const()[name = string("op_4494_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 128]> var_4494_0, tensor<fp16, [1, 3, 128, 128]> var_4494_1 = split(axis = var_4494_axis_0, split_sizes = var_4494_split_sizes_0, x = out_219_cast_fp16)[name = string("op_4494")];
@@ -3634,9 +3639,9 @@ program(1.3)
             tensor<fp16, [1, 3, 128, 128]> var_4496 = mul(x = var_4494_1, y = const_258_promoted)[name = string("op_4496")];
             bool var_4498_interleave_0 = const()[name = string("op_4498_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 3, 128, 256]> var_4498 = concat(axis = var_23, interleave = var_4498_interleave_0, values = (var_4496, var_4494_0))[name = string("op_4498")];
-            tensor<fp16, [1, 3, 128, 256]> var_4499 = mul(x = var_4498, y = sin_1)[name = string("op_4499")];
+            tensor<fp16, [1, 3, 128, 256]> var_4499 = mul(x = var_4498, y = sin_1_quantized)[name = string("op_4499")];
             tensor<fp16, [1, 3, 128, 256]> q_113 = add(x = var_4493, y = var_4499)[name = string("q_113")];
-            tensor<fp16, [1, 1, 128, 256]> var_4501 = mul(x = out_221_cast_fp16, y = cos_1)[name = string("op_4501")];
+            tensor<fp16, [1, 1, 128, 256]> var_4501 = mul(x = out_221_cast_fp16, y = cos_1_quantized)[name = string("op_4501")];
             tensor<int32, [2]> var_4502_split_sizes_0 = const()[name = string("op_4502_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_4502_axis_0 = const()[name = string("op_4502_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 128]> var_4502_0, tensor<fp16, [1, 1, 128, 128]> var_4502_1 = split(axis = var_4502_axis_0, split_sizes = var_4502_split_sizes_0, x = out_221_cast_fp16)[name = string("op_4502")];
@@ -3644,7 +3649,7 @@ program(1.3)
             tensor<fp16, [1, 1, 128, 128]> var_4504 = mul(x = var_4502_1, y = const_259_promoted)[name = string("op_4504")];
             bool var_4506_interleave_0 = const()[name = string("op_4506_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 1, 128, 256]> var_4506 = concat(axis = var_23, interleave = var_4506_interleave_0, values = (var_4504, var_4502_0))[name = string("op_4506")];
-            tensor<fp16, [1, 1, 128, 256]> var_4507 = mul(x = var_4506, y = sin_1)[name = string("op_4507")];
+            tensor<fp16, [1, 1, 128, 256]> var_4507 = mul(x = var_4506, y = sin_1_quantized)[name = string("op_4507")];
             tensor<fp16, [1, 1, 128, 256]> hidden_states_217 = add(x = var_4501, y = var_4507)[name = string("hidden_states_217")];
             tensor<int32, [1]> hidden_states_219_axes_0 = const()[name = string("hidden_states_219_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 1, 1, 128, 256]> hidden_states_219 = expand_dims(axes = hidden_states_219_axes_0, x = hidden_states_217)[name = string("hidden_states_219")];
@@ -3674,14 +3679,14 @@ program(1.3)
             tensor<fp16, [1, 128, 3, 256]> var_4534 = transpose(perm = var_4533, x = var_4531_cast_fp16)[name = string("transpose_49")];
             tensor<fp16, [1, 128, 768]> attn_out_111 = reshape(shape = var_4536, x = var_4534)[name = string("attn_out_111")];
             tensor<int32, [3]> var_4538 = const()[name = string("op_4538"), val = tensor<int32, [3]>([0, 2, 1])];
-            tensor<fp16, [768, 768, 1]> squeeze_18 = const()[name = string("squeeze_18"), val = tensor<fp16, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(608331456)))];
+            tensor<fp16, [768, 768, 1]> squeeze_18_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305023488))), scale = tensor<fp16, [768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305613376))))[name = string("squeeze_18_quantized")];
             string var_4547_pad_type_0 = const()[name = string("op_4547_pad_type_0"), val = string("valid")];
             int32 var_4547_groups_0 = const()[name = string("op_4547_groups_0"), val = int32(1)];
             tensor<int32, [1]> var_4547_strides_0 = const()[name = string("op_4547_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> var_4547_pad_0 = const()[name = string("op_4547_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> var_4547_dilations_0 = const()[name = string("op_4547_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, [1, 768, 128]> var_4539 = transpose(perm = var_4538, x = attn_out_111)[name = string("transpose_48")];
-            tensor<fp16, [1, 768, 128]> var_4547 = conv(dilations = var_4547_dilations_0, groups = var_4547_groups_0, pad = var_4547_pad_0, pad_type = var_4547_pad_type_0, strides = var_4547_strides_0, weight = squeeze_18, x = var_4539)[name = string("op_4547")];
+            tensor<fp16, [1, 768, 128]> var_4547 = conv(dilations = var_4547_dilations_0, groups = var_4547_groups_0, pad = var_4547_pad_0, pad_type = var_4547_pad_type_0, strides = var_4547_strides_0, weight = squeeze_18_quantized, x = var_4539)[name = string("op_4547")];
             tensor<int32, [3]> var_4548 = const()[name = string("op_4548"), val = tensor<int32, [3]>([0, 2, 1])];
             fp16 const_260_promoted_to_fp16 = const()[name = string("const_260_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 128, 768]> x_297 = transpose(perm = var_4548, x = var_4547)[name = string("transpose_47")];
@@ -3693,7 +3698,7 @@ program(1.3)
             tensor<int32, [2]> var_4557_split_sizes_0 = const()[name = string("op_4557_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_4557_axis_0 = const()[name = string("op_4557_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_4557_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_4557_cast_fp16_1 = split(axis = var_4557_axis_0, split_sizes = var_4557_split_sizes_0, x = normed_519_cast_fp16)[name = string("op_4557_cast_fp16")];
-            tensor<fp16, [768]> var_4561_to_fp16 = const()[name = string("op_4561_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(609511168)))];
+            tensor<fp16, [768]> var_4561_to_fp16 = const()[name = string("op_4561_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305614976)))];
             tensor<fp16, [1, 128, 768]> out_223_cast_fp16 = mul(x = var_4557_cast_fp16_0, y = var_4561_to_fp16)[name = string("out_223_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_299_cast_fp16 = add(x = x_289_cast_fp16, y = out_223_cast_fp16)[name = string("x_299_cast_fp16")];
             fp16 const_262_promoted_to_fp16 = const()[name = string("const_262_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -3705,7 +3710,7 @@ program(1.3)
             tensor<int32, [2]> var_4573_split_sizes_0 = const()[name = string("op_4573_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_4573_axis_0 = const()[name = string("op_4573_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_4573_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_4573_cast_fp16_1 = split(axis = var_4573_axis_0, split_sizes = var_4573_split_sizes_0, x = normed_523_cast_fp16)[name = string("op_4573_cast_fp16")];
-            tensor<fp16, [768]> var_4577_to_fp16 = const()[name = string("op_4577_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(609512768)))];
+            tensor<fp16, [768]> var_4577_to_fp16 = const()[name = string("op_4577_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305616576)))];
             tensor<fp16, [1, 128, 768]> out_225_cast_fp16 = mul(x = var_4573_cast_fp16_0, y = var_4577_to_fp16)[name = string("out_225_cast_fp16")];
             tensor<int32, [3]> var_4584 = const()[name = string("op_4584"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> input_375_axes_0 = const()[name = string("input_375_axes_0"), val = tensor<int32, [1]>([2])];
@@ -3716,13 +3721,13 @@ program(1.3)
             tensor<int32, [4]> gate_73_pad_0 = const()[name = string("gate_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> gate_73_dilations_0 = const()[name = string("gate_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 gate_73_groups_0 = const()[name = string("gate_73_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> gate_73 = conv(dilations = gate_73_dilations_0, groups = gate_73_groups_0, pad = gate_73_pad_0, pad_type = gate_73_pad_type_0, strides = gate_73_strides_0, weight = encoder_layers_18_mlp_gate_proj_weight, x = input_375)[name = string("gate_73")];
+            tensor<fp16, [1, 1152, 1, 128]> gate_73 = conv(dilations = gate_73_dilations_0, groups = gate_73_groups_0, pad = gate_73_pad_0, pad_type = gate_73_pad_type_0, strides = gate_73_strides_0, weight = encoder_layers_18_mlp_gate_proj_weight_quantized, x = input_375)[name = string("gate_73")];
             string up_37_pad_type_0 = const()[name = string("up_37_pad_type_0"), val = string("valid")];
             tensor<int32, [2]> up_37_strides_0 = const()[name = string("up_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> up_37_pad_0 = const()[name = string("up_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> up_37_dilations_0 = const()[name = string("up_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 up_37_groups_0 = const()[name = string("up_37_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> up_37 = conv(dilations = up_37_dilations_0, groups = up_37_groups_0, pad = up_37_pad_0, pad_type = up_37_pad_type_0, strides = up_37_strides_0, weight = encoder_layers_18_mlp_up_proj_weight, x = input_375)[name = string("up_37")];
+            tensor<fp16, [1, 1152, 1, 128]> up_37 = conv(dilations = up_37_dilations_0, groups = up_37_groups_0, pad = up_37_pad_0, pad_type = up_37_pad_type_0, strides = up_37_strides_0, weight = encoder_layers_18_mlp_up_proj_weight_quantized, x = input_375)[name = string("up_37")];
             string gate_75_mode_0 = const()[name = string("gate_75_mode_0"), val = string("TANH_APPROXIMATION")];
             tensor<fp16, [1, 1152, 1, 128]> gate_75 = gelu(mode = gate_75_mode_0, x = gate_73)[name = string("gate_75")];
             tensor<fp16, [1, 1152, 1, 128]> input_377 = mul(x = gate_75, y = up_37)[name = string("input_377")];
@@ -3731,7 +3736,7 @@ program(1.3)
             tensor<int32, [4]> var_4606_pad_0 = const()[name = string("op_4606_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_4606_dilations_0 = const()[name = string("op_4606_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_4606_groups_0 = const()[name = string("op_4606_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_4606 = conv(dilations = var_4606_dilations_0, groups = var_4606_groups_0, pad = var_4606_pad_0, pad_type = var_4606_pad_type_0, strides = var_4606_strides_0, weight = encoder_layers_18_mlp_down_proj_weight, x = input_377)[name = string("op_4606")];
+            tensor<fp16, [1, 768, 1, 128]> var_4606 = conv(dilations = var_4606_dilations_0, groups = var_4606_groups_0, pad = var_4606_pad_0, pad_type = var_4606_pad_type_0, strides = var_4606_strides_0, weight = encoder_layers_18_mlp_down_proj_weight_quantized, x = input_377)[name = string("op_4606")];
             tensor<int32, [1]> var_4607_axes_0 = const()[name = string("op_4607_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 768, 128]> var_4607 = squeeze(axes = var_4607_axes_0, x = var_4606)[name = string("op_4607")];
             tensor<int32, [3]> var_4608 = const()[name = string("op_4608"), val = tensor<int32, [3]>([0, 2, 1])];
@@ -3745,7 +3750,7 @@ program(1.3)
             tensor<int32, [2]> var_4617_split_sizes_0 = const()[name = string("op_4617_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_4617_axis_0 = const()[name = string("op_4617_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_4617_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_4617_cast_fp16_1 = split(axis = var_4617_axis_0, split_sizes = var_4617_split_sizes_0, x = normed_529_cast_fp16)[name = string("op_4617_cast_fp16")];
-            tensor<fp16, [768]> var_4621_to_fp16 = const()[name = string("op_4621_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(609514368)))];
+            tensor<fp16, [768]> var_4621_to_fp16 = const()[name = string("op_4621_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305618176)))];
             tensor<fp16, [1, 128, 768]> out_227_cast_fp16 = mul(x = var_4617_cast_fp16_0, y = var_4621_to_fp16)[name = string("out_227_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_305_cast_fp16 = add(x = x_299_cast_fp16, y = out_227_cast_fp16)[name = string("x_305_cast_fp16")];
             fp16 const_266_promoted_to_fp16 = const()[name = string("const_266_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -3757,7 +3762,7 @@ program(1.3)
             tensor<int32, [2]> var_4655_split_sizes_0 = const()[name = string("op_4655_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_4655_axis_0 = const()[name = string("op_4655_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_4655_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_4655_cast_fp16_1 = split(axis = var_4655_axis_0, split_sizes = var_4655_split_sizes_0, x = normed_533_cast_fp16)[name = string("op_4655_cast_fp16")];
-            tensor<fp16, [768]> var_4659_to_fp16 = const()[name = string("op_4659_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(609515968)))];
+            tensor<fp16, [768]> var_4659_to_fp16 = const()[name = string("op_4659_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305619776)))];
             tensor<fp16, [1, 128, 768]> out_229_cast_fp16 = mul(x = var_4655_cast_fp16_0, y = var_4659_to_fp16)[name = string("out_229_cast_fp16")];
             tensor<int32, [3]> var_4665 = const()[name = string("op_4665"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> var_4667_axes_0 = const()[name = string("op_4667_axes_0"), val = tensor<int32, [1]>([2])];
@@ -3768,7 +3773,7 @@ program(1.3)
             tensor<int32, [4]> var_4674_pad_0 = const()[name = string("op_4674_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_4674_dilations_0 = const()[name = string("op_4674_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_4674_groups_0 = const()[name = string("op_4674_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_4674 = conv(dilations = var_4674_dilations_0, groups = var_4674_groups_0, pad = var_4674_pad_0, pad_type = var_4674_pad_type_0, strides = var_4674_strides_0, weight = encoder_layers_19_self_attn_q_proj_weight, x = var_4667_cast_fp16)[name = string("op_4674")];
+            tensor<fp16, [1, 768, 1, 128]> var_4674 = conv(dilations = var_4674_dilations_0, groups = var_4674_groups_0, pad = var_4674_pad_0, pad_type = var_4674_pad_type_0, strides = var_4674_strides_0, weight = encoder_layers_19_self_attn_q_proj_weight_quantized, x = var_4667_cast_fp16)[name = string("op_4674")];
             tensor<int32, [4]> var_4675 = const()[name = string("op_4675"), val = tensor<int32, [4]>([1, 3, 256, 128])];
             tensor<fp16, [1, 3, 256, 128]> var_4676 = reshape(shape = var_4675, x = var_4674)[name = string("op_4676")];
             tensor<int32, [4]> var_4677 = const()[name = string("op_4677"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -3777,7 +3782,7 @@ program(1.3)
             tensor<int32, [4]> var_4684_pad_0 = const()[name = string("op_4684_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_4684_dilations_0 = const()[name = string("op_4684_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_4684_groups_0 = const()[name = string("op_4684_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_4684 = conv(dilations = var_4684_dilations_0, groups = var_4684_groups_0, pad = var_4684_pad_0, pad_type = var_4684_pad_type_0, strides = var_4684_strides_0, weight = encoder_layers_19_self_attn_k_proj_weight, x = var_4667_cast_fp16)[name = string("op_4684")];
+            tensor<fp16, [1, 256, 1, 128]> var_4684 = conv(dilations = var_4684_dilations_0, groups = var_4684_groups_0, pad = var_4684_pad_0, pad_type = var_4684_pad_type_0, strides = var_4684_strides_0, weight = encoder_layers_19_self_attn_k_proj_weight_quantized, x = var_4667_cast_fp16)[name = string("op_4684")];
             tensor<int32, [4]> var_4685 = const()[name = string("op_4685"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_4686 = reshape(shape = var_4685, x = var_4684)[name = string("op_4686")];
             tensor<int32, [4]> var_4687 = const()[name = string("op_4687"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -3786,7 +3791,7 @@ program(1.3)
             tensor<int32, [4]> var_4694_pad_0 = const()[name = string("op_4694_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_4694_dilations_0 = const()[name = string("op_4694_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_4694_groups_0 = const()[name = string("op_4694_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_4694 = conv(dilations = var_4694_dilations_0, groups = var_4694_groups_0, pad = var_4694_pad_0, pad_type = var_4694_pad_type_0, strides = var_4694_strides_0, weight = encoder_layers_19_self_attn_v_proj_weight, x = var_4667_cast_fp16)[name = string("op_4694")];
+            tensor<fp16, [1, 256, 1, 128]> var_4694 = conv(dilations = var_4694_dilations_0, groups = var_4694_groups_0, pad = var_4694_pad_0, pad_type = var_4694_pad_type_0, strides = var_4694_strides_0, weight = encoder_layers_19_self_attn_v_proj_weight_quantized, x = var_4667_cast_fp16)[name = string("op_4694")];
             tensor<int32, [4]> var_4695 = const()[name = string("op_4695"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_4696 = reshape(shape = var_4695, x = var_4694)[name = string("op_4696")];
             tensor<int32, [4]> var_4697 = const()[name = string("op_4697"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -3800,7 +3805,7 @@ program(1.3)
             tensor<int32, [2]> var_4708_split_sizes_0 = const()[name = string("op_4708_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_4708_axis_0 = const()[name = string("op_4708_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 256]> var_4708_cast_fp16_0, tensor<fp16, [1, 3, 128, 256]> var_4708_cast_fp16_1 = split(axis = var_4708_axis_0, split_sizes = var_4708_split_sizes_0, x = normed_539_cast_fp16)[name = string("op_4708_cast_fp16")];
-            tensor<fp16, [256]> var_4712_to_fp16 = const()[name = string("op_4712_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(609517568)))];
+            tensor<fp16, [256]> var_4712_to_fp16 = const()[name = string("op_4712_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305621376)))];
             tensor<fp16, [1, 3, 128, 256]> out_231_cast_fp16 = mul(x = var_4708_cast_fp16_0, y = var_4712_to_fp16)[name = string("out_231_cast_fp16")];
             fp16 const_270_promoted_to_fp16 = const()[name = string("const_270_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 1, 128, 256]> k_115 = transpose(perm = var_4687, x = var_4686)[name = string("transpose_42")];
@@ -3812,9 +3817,9 @@ program(1.3)
             tensor<int32, [2]> var_4724_split_sizes_0 = const()[name = string("op_4724_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_4724_axis_0 = const()[name = string("op_4724_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 256]> var_4724_cast_fp16_0, tensor<fp16, [1, 1, 128, 256]> var_4724_cast_fp16_1 = split(axis = var_4724_axis_0, split_sizes = var_4724_split_sizes_0, x = normed_543_cast_fp16)[name = string("op_4724_cast_fp16")];
-            tensor<fp16, [256]> var_4728_to_fp16 = const()[name = string("op_4728_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(609518144)))];
+            tensor<fp16, [256]> var_4728_to_fp16 = const()[name = string("op_4728_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305621952)))];
             tensor<fp16, [1, 1, 128, 256]> out_233_cast_fp16 = mul(x = var_4724_cast_fp16_0, y = var_4728_to_fp16)[name = string("out_233_cast_fp16")];
-            tensor<fp16, [1, 3, 128, 256]> var_4731 = mul(x = out_231_cast_fp16, y = cos_1)[name = string("op_4731")];
+            tensor<fp16, [1, 3, 128, 256]> var_4731 = mul(x = out_231_cast_fp16, y = cos_1_quantized)[name = string("op_4731")];
             tensor<int32, [2]> var_4732_split_sizes_0 = const()[name = string("op_4732_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_4732_axis_0 = const()[name = string("op_4732_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 128]> var_4732_0, tensor<fp16, [1, 3, 128, 128]> var_4732_1 = split(axis = var_4732_axis_0, split_sizes = var_4732_split_sizes_0, x = out_231_cast_fp16)[name = string("op_4732")];
@@ -3822,9 +3827,9 @@ program(1.3)
             tensor<fp16, [1, 3, 128, 128]> var_4734 = mul(x = var_4732_1, y = const_272_promoted)[name = string("op_4734")];
             bool var_4736_interleave_0 = const()[name = string("op_4736_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 3, 128, 256]> var_4736 = concat(axis = var_23, interleave = var_4736_interleave_0, values = (var_4734, var_4732_0))[name = string("op_4736")];
-            tensor<fp16, [1, 3, 128, 256]> var_4737 = mul(x = var_4736, y = sin_1)[name = string("op_4737")];
+            tensor<fp16, [1, 3, 128, 256]> var_4737 = mul(x = var_4736, y = sin_1_quantized)[name = string("op_4737")];
             tensor<fp16, [1, 3, 128, 256]> q_119 = add(x = var_4731, y = var_4737)[name = string("q_119")];
-            tensor<fp16, [1, 1, 128, 256]> var_4739 = mul(x = out_233_cast_fp16, y = cos_1)[name = string("op_4739")];
+            tensor<fp16, [1, 1, 128, 256]> var_4739 = mul(x = out_233_cast_fp16, y = cos_1_quantized)[name = string("op_4739")];
             tensor<int32, [2]> var_4740_split_sizes_0 = const()[name = string("op_4740_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_4740_axis_0 = const()[name = string("op_4740_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 128]> var_4740_0, tensor<fp16, [1, 1, 128, 128]> var_4740_1 = split(axis = var_4740_axis_0, split_sizes = var_4740_split_sizes_0, x = out_233_cast_fp16)[name = string("op_4740")];
@@ -3832,7 +3837,7 @@ program(1.3)
             tensor<fp16, [1, 1, 128, 128]> var_4742 = mul(x = var_4740_1, y = const_273_promoted)[name = string("op_4742")];
             bool var_4744_interleave_0 = const()[name = string("op_4744_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 1, 128, 256]> var_4744 = concat(axis = var_23, interleave = var_4744_interleave_0, values = (var_4742, var_4740_0))[name = string("op_4744")];
-            tensor<fp16, [1, 1, 128, 256]> var_4745 = mul(x = var_4744, y = sin_1)[name = string("op_4745")];
+            tensor<fp16, [1, 1, 128, 256]> var_4745 = mul(x = var_4744, y = sin_1_quantized)[name = string("op_4745")];
             tensor<fp16, [1, 1, 128, 256]> hidden_states_229 = add(x = var_4739, y = var_4745)[name = string("hidden_states_229")];
             tensor<int32, [1]> hidden_states_231_axes_0 = const()[name = string("hidden_states_231_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 1, 1, 128, 256]> hidden_states_231 = expand_dims(axes = hidden_states_231_axes_0, x = hidden_states_229)[name = string("hidden_states_231")];
@@ -3862,14 +3867,14 @@ program(1.3)
             tensor<fp16, [1, 128, 3, 256]> var_4772 = transpose(perm = var_4771, x = var_4769_cast_fp16)[name = string("transpose_40")];
             tensor<fp16, [1, 128, 768]> attn_out_117 = reshape(shape = var_4774, x = var_4772)[name = string("attn_out_117")];
             tensor<int32, [3]> var_4776 = const()[name = string("op_4776"), val = tensor<int32, [3]>([0, 2, 1])];
-            tensor<fp16, [768, 768, 1]> squeeze_19 = const()[name = string("squeeze_19"), val = tensor<fp16, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(609518720)))];
+            tensor<fp16, [768, 768, 1]> squeeze_19_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305622528))), scale = tensor<fp16, [768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306212416))))[name = string("squeeze_19_quantized")];
             string var_4785_pad_type_0 = const()[name = string("op_4785_pad_type_0"), val = string("valid")];
             int32 var_4785_groups_0 = const()[name = string("op_4785_groups_0"), val = int32(1)];
             tensor<int32, [1]> var_4785_strides_0 = const()[name = string("op_4785_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> var_4785_pad_0 = const()[name = string("op_4785_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> var_4785_dilations_0 = const()[name = string("op_4785_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, [1, 768, 128]> var_4777 = transpose(perm = var_4776, x = attn_out_117)[name = string("transpose_39")];
-            tensor<fp16, [1, 768, 128]> var_4785 = conv(dilations = var_4785_dilations_0, groups = var_4785_groups_0, pad = var_4785_pad_0, pad_type = var_4785_pad_type_0, strides = var_4785_strides_0, weight = squeeze_19, x = var_4777)[name = string("op_4785")];
+            tensor<fp16, [1, 768, 128]> var_4785 = conv(dilations = var_4785_dilations_0, groups = var_4785_groups_0, pad = var_4785_pad_0, pad_type = var_4785_pad_type_0, strides = var_4785_strides_0, weight = squeeze_19_quantized, x = var_4777)[name = string("op_4785")];
             tensor<int32, [3]> var_4786 = const()[name = string("op_4786"), val = tensor<int32, [3]>([0, 2, 1])];
             fp16 const_274_promoted_to_fp16 = const()[name = string("const_274_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 128, 768]> x_313 = transpose(perm = var_4786, x = var_4785)[name = string("transpose_38")];
@@ -3881,7 +3886,7 @@ program(1.3)
             tensor<int32, [2]> var_4795_split_sizes_0 = const()[name = string("op_4795_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_4795_axis_0 = const()[name = string("op_4795_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_4795_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_4795_cast_fp16_1 = split(axis = var_4795_axis_0, split_sizes = var_4795_split_sizes_0, x = normed_547_cast_fp16)[name = string("op_4795_cast_fp16")];
-            tensor<fp16, [768]> var_4799_to_fp16 = const()[name = string("op_4799_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(610698432)))];
+            tensor<fp16, [768]> var_4799_to_fp16 = const()[name = string("op_4799_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306214016)))];
             tensor<fp16, [1, 128, 768]> out_235_cast_fp16 = mul(x = var_4795_cast_fp16_0, y = var_4799_to_fp16)[name = string("out_235_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_315_cast_fp16 = add(x = x_305_cast_fp16, y = out_235_cast_fp16)[name = string("x_315_cast_fp16")];
             fp16 const_276_promoted_to_fp16 = const()[name = string("const_276_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -3893,7 +3898,7 @@ program(1.3)
             tensor<int32, [2]> var_4811_split_sizes_0 = const()[name = string("op_4811_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_4811_axis_0 = const()[name = string("op_4811_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_4811_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_4811_cast_fp16_1 = split(axis = var_4811_axis_0, split_sizes = var_4811_split_sizes_0, x = normed_551_cast_fp16)[name = string("op_4811_cast_fp16")];
-            tensor<fp16, [768]> var_4815_to_fp16 = const()[name = string("op_4815_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(610700032)))];
+            tensor<fp16, [768]> var_4815_to_fp16 = const()[name = string("op_4815_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306215616)))];
             tensor<fp16, [1, 128, 768]> out_237_cast_fp16 = mul(x = var_4811_cast_fp16_0, y = var_4815_to_fp16)[name = string("out_237_cast_fp16")];
             tensor<int32, [3]> var_4822 = const()[name = string("op_4822"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> input_395_axes_0 = const()[name = string("input_395_axes_0"), val = tensor<int32, [1]>([2])];
@@ -3904,13 +3909,13 @@ program(1.3)
             tensor<int32, [4]> gate_77_pad_0 = const()[name = string("gate_77_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> gate_77_dilations_0 = const()[name = string("gate_77_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 gate_77_groups_0 = const()[name = string("gate_77_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> gate_77 = conv(dilations = gate_77_dilations_0, groups = gate_77_groups_0, pad = gate_77_pad_0, pad_type = gate_77_pad_type_0, strides = gate_77_strides_0, weight = encoder_layers_19_mlp_gate_proj_weight, x = input_395)[name = string("gate_77")];
+            tensor<fp16, [1, 1152, 1, 128]> gate_77 = conv(dilations = gate_77_dilations_0, groups = gate_77_groups_0, pad = gate_77_pad_0, pad_type = gate_77_pad_type_0, strides = gate_77_strides_0, weight = encoder_layers_19_mlp_gate_proj_weight_quantized, x = input_395)[name = string("gate_77")];
             string up_39_pad_type_0 = const()[name = string("up_39_pad_type_0"), val = string("valid")];
             tensor<int32, [2]> up_39_strides_0 = const()[name = string("up_39_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> up_39_pad_0 = const()[name = string("up_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> up_39_dilations_0 = const()[name = string("up_39_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 up_39_groups_0 = const()[name = string("up_39_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> up_39 = conv(dilations = up_39_dilations_0, groups = up_39_groups_0, pad = up_39_pad_0, pad_type = up_39_pad_type_0, strides = up_39_strides_0, weight = encoder_layers_19_mlp_up_proj_weight, x = input_395)[name = string("up_39")];
+            tensor<fp16, [1, 1152, 1, 128]> up_39 = conv(dilations = up_39_dilations_0, groups = up_39_groups_0, pad = up_39_pad_0, pad_type = up_39_pad_type_0, strides = up_39_strides_0, weight = encoder_layers_19_mlp_up_proj_weight_quantized, x = input_395)[name = string("up_39")];
             string gate_79_mode_0 = const()[name = string("gate_79_mode_0"), val = string("TANH_APPROXIMATION")];
             tensor<fp16, [1, 1152, 1, 128]> gate_79 = gelu(mode = gate_79_mode_0, x = gate_77)[name = string("gate_79")];
             tensor<fp16, [1, 1152, 1, 128]> input_397 = mul(x = gate_79, y = up_39)[name = string("input_397")];
@@ -3919,7 +3924,7 @@ program(1.3)
             tensor<int32, [4]> var_4844_pad_0 = const()[name = string("op_4844_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_4844_dilations_0 = const()[name = string("op_4844_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_4844_groups_0 = const()[name = string("op_4844_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_4844 = conv(dilations = var_4844_dilations_0, groups = var_4844_groups_0, pad = var_4844_pad_0, pad_type = var_4844_pad_type_0, strides = var_4844_strides_0, weight = encoder_layers_19_mlp_down_proj_weight, x = input_397)[name = string("op_4844")];
+            tensor<fp16, [1, 768, 1, 128]> var_4844 = conv(dilations = var_4844_dilations_0, groups = var_4844_groups_0, pad = var_4844_pad_0, pad_type = var_4844_pad_type_0, strides = var_4844_strides_0, weight = encoder_layers_19_mlp_down_proj_weight_quantized, x = input_397)[name = string("op_4844")];
             tensor<int32, [1]> var_4845_axes_0 = const()[name = string("op_4845_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 768, 128]> var_4845 = squeeze(axes = var_4845_axes_0, x = var_4844)[name = string("op_4845")];
             tensor<int32, [3]> var_4846 = const()[name = string("op_4846"), val = tensor<int32, [3]>([0, 2, 1])];
@@ -3933,7 +3938,7 @@ program(1.3)
             tensor<int32, [2]> var_4855_split_sizes_0 = const()[name = string("op_4855_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_4855_axis_0 = const()[name = string("op_4855_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_4855_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_4855_cast_fp16_1 = split(axis = var_4855_axis_0, split_sizes = var_4855_split_sizes_0, x = normed_557_cast_fp16)[name = string("op_4855_cast_fp16")];
-            tensor<fp16, [768]> var_4859_to_fp16 = const()[name = string("op_4859_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(610701632)))];
+            tensor<fp16, [768]> var_4859_to_fp16 = const()[name = string("op_4859_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306217216)))];
             tensor<fp16, [1, 128, 768]> out_239_cast_fp16 = mul(x = var_4855_cast_fp16_0, y = var_4859_to_fp16)[name = string("out_239_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_321_cast_fp16 = add(x = x_315_cast_fp16, y = out_239_cast_fp16)[name = string("x_321_cast_fp16")];
             fp16 const_280_promoted_to_fp16 = const()[name = string("const_280_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -3945,7 +3950,7 @@ program(1.3)
             tensor<int32, [2]> var_4893_split_sizes_0 = const()[name = string("op_4893_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_4893_axis_0 = const()[name = string("op_4893_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_4893_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_4893_cast_fp16_1 = split(axis = var_4893_axis_0, split_sizes = var_4893_split_sizes_0, x = normed_561_cast_fp16)[name = string("op_4893_cast_fp16")];
-            tensor<fp16, [768]> var_4897_to_fp16 = const()[name = string("op_4897_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(610703232)))];
+            tensor<fp16, [768]> var_4897_to_fp16 = const()[name = string("op_4897_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306218816)))];
             tensor<fp16, [1, 128, 768]> out_241_cast_fp16 = mul(x = var_4893_cast_fp16_0, y = var_4897_to_fp16)[name = string("out_241_cast_fp16")];
             tensor<int32, [3]> var_4903 = const()[name = string("op_4903"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> var_4905_axes_0 = const()[name = string("op_4905_axes_0"), val = tensor<int32, [1]>([2])];
@@ -3956,7 +3961,7 @@ program(1.3)
             tensor<int32, [4]> var_4912_pad_0 = const()[name = string("op_4912_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_4912_dilations_0 = const()[name = string("op_4912_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_4912_groups_0 = const()[name = string("op_4912_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_4912 = conv(dilations = var_4912_dilations_0, groups = var_4912_groups_0, pad = var_4912_pad_0, pad_type = var_4912_pad_type_0, strides = var_4912_strides_0, weight = encoder_layers_20_self_attn_q_proj_weight, x = var_4905_cast_fp16)[name = string("op_4912")];
+            tensor<fp16, [1, 768, 1, 128]> var_4912 = conv(dilations = var_4912_dilations_0, groups = var_4912_groups_0, pad = var_4912_pad_0, pad_type = var_4912_pad_type_0, strides = var_4912_strides_0, weight = encoder_layers_20_self_attn_q_proj_weight_quantized, x = var_4905_cast_fp16)[name = string("op_4912")];
             tensor<int32, [4]> var_4913 = const()[name = string("op_4913"), val = tensor<int32, [4]>([1, 3, 256, 128])];
             tensor<fp16, [1, 3, 256, 128]> var_4914 = reshape(shape = var_4913, x = var_4912)[name = string("op_4914")];
             tensor<int32, [4]> var_4915 = const()[name = string("op_4915"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -3965,7 +3970,7 @@ program(1.3)
             tensor<int32, [4]> var_4922_pad_0 = const()[name = string("op_4922_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_4922_dilations_0 = const()[name = string("op_4922_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_4922_groups_0 = const()[name = string("op_4922_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_4922 = conv(dilations = var_4922_dilations_0, groups = var_4922_groups_0, pad = var_4922_pad_0, pad_type = var_4922_pad_type_0, strides = var_4922_strides_0, weight = encoder_layers_20_self_attn_k_proj_weight, x = var_4905_cast_fp16)[name = string("op_4922")];
+            tensor<fp16, [1, 256, 1, 128]> var_4922 = conv(dilations = var_4922_dilations_0, groups = var_4922_groups_0, pad = var_4922_pad_0, pad_type = var_4922_pad_type_0, strides = var_4922_strides_0, weight = encoder_layers_20_self_attn_k_proj_weight_quantized, x = var_4905_cast_fp16)[name = string("op_4922")];
             tensor<int32, [4]> var_4923 = const()[name = string("op_4923"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_4924 = reshape(shape = var_4923, x = var_4922)[name = string("op_4924")];
             tensor<int32, [4]> var_4925 = const()[name = string("op_4925"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -3974,7 +3979,7 @@ program(1.3)
             tensor<int32, [4]> var_4932_pad_0 = const()[name = string("op_4932_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_4932_dilations_0 = const()[name = string("op_4932_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_4932_groups_0 = const()[name = string("op_4932_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_4932 = conv(dilations = var_4932_dilations_0, groups = var_4932_groups_0, pad = var_4932_pad_0, pad_type = var_4932_pad_type_0, strides = var_4932_strides_0, weight = encoder_layers_20_self_attn_v_proj_weight, x = var_4905_cast_fp16)[name = string("op_4932")];
+            tensor<fp16, [1, 256, 1, 128]> var_4932 = conv(dilations = var_4932_dilations_0, groups = var_4932_groups_0, pad = var_4932_pad_0, pad_type = var_4932_pad_type_0, strides = var_4932_strides_0, weight = encoder_layers_20_self_attn_v_proj_weight_quantized, x = var_4905_cast_fp16)[name = string("op_4932")];
             tensor<int32, [4]> var_4933 = const()[name = string("op_4933"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_4934 = reshape(shape = var_4933, x = var_4932)[name = string("op_4934")];
             tensor<int32, [4]> var_4935 = const()[name = string("op_4935"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -3988,7 +3993,7 @@ program(1.3)
             tensor<int32, [2]> var_4946_split_sizes_0 = const()[name = string("op_4946_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_4946_axis_0 = const()[name = string("op_4946_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 256]> var_4946_cast_fp16_0, tensor<fp16, [1, 3, 128, 256]> var_4946_cast_fp16_1 = split(axis = var_4946_axis_0, split_sizes = var_4946_split_sizes_0, x = normed_567_cast_fp16)[name = string("op_4946_cast_fp16")];
-            tensor<fp16, [256]> var_4950_to_fp16 = const()[name = string("op_4950_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(610704832)))];
+            tensor<fp16, [256]> var_4950_to_fp16 = const()[name = string("op_4950_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306220416)))];
             tensor<fp16, [1, 3, 128, 256]> out_243_cast_fp16 = mul(x = var_4946_cast_fp16_0, y = var_4950_to_fp16)[name = string("out_243_cast_fp16")];
             fp16 const_284_promoted_to_fp16 = const()[name = string("const_284_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 1, 128, 256]> k_121 = transpose(perm = var_4925, x = var_4924)[name = string("transpose_33")];
@@ -4000,9 +4005,9 @@ program(1.3)
             tensor<int32, [2]> var_4962_split_sizes_0 = const()[name = string("op_4962_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_4962_axis_0 = const()[name = string("op_4962_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 256]> var_4962_cast_fp16_0, tensor<fp16, [1, 1, 128, 256]> var_4962_cast_fp16_1 = split(axis = var_4962_axis_0, split_sizes = var_4962_split_sizes_0, x = normed_571_cast_fp16)[name = string("op_4962_cast_fp16")];
-            tensor<fp16, [256]> var_4966_to_fp16 = const()[name = string("op_4966_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(610705408)))];
+            tensor<fp16, [256]> var_4966_to_fp16 = const()[name = string("op_4966_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306220992)))];
             tensor<fp16, [1, 1, 128, 256]> out_245_cast_fp16 = mul(x = var_4962_cast_fp16_0, y = var_4966_to_fp16)[name = string("out_245_cast_fp16")];
-            tensor<fp16, [1, 3, 128, 256]> var_4969 = mul(x = out_243_cast_fp16, y = cos_1)[name = string("op_4969")];
+            tensor<fp16, [1, 3, 128, 256]> var_4969 = mul(x = out_243_cast_fp16, y = cos_1_quantized)[name = string("op_4969")];
             tensor<int32, [2]> var_4970_split_sizes_0 = const()[name = string("op_4970_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_4970_axis_0 = const()[name = string("op_4970_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 128]> var_4970_0, tensor<fp16, [1, 3, 128, 128]> var_4970_1 = split(axis = var_4970_axis_0, split_sizes = var_4970_split_sizes_0, x = out_243_cast_fp16)[name = string("op_4970")];
@@ -4010,9 +4015,9 @@ program(1.3)
             tensor<fp16, [1, 3, 128, 128]> var_4972 = mul(x = var_4970_1, y = const_286_promoted)[name = string("op_4972")];
             bool var_4974_interleave_0 = const()[name = string("op_4974_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 3, 128, 256]> var_4974 = concat(axis = var_23, interleave = var_4974_interleave_0, values = (var_4972, var_4970_0))[name = string("op_4974")];
-            tensor<fp16, [1, 3, 128, 256]> var_4975 = mul(x = var_4974, y = sin_1)[name = string("op_4975")];
+            tensor<fp16, [1, 3, 128, 256]> var_4975 = mul(x = var_4974, y = sin_1_quantized)[name = string("op_4975")];
             tensor<fp16, [1, 3, 128, 256]> q_125 = add(x = var_4969, y = var_4975)[name = string("q_125")];
-            tensor<fp16, [1, 1, 128, 256]> var_4977 = mul(x = out_245_cast_fp16, y = cos_1)[name = string("op_4977")];
+            tensor<fp16, [1, 1, 128, 256]> var_4977 = mul(x = out_245_cast_fp16, y = cos_1_quantized)[name = string("op_4977")];
             tensor<int32, [2]> var_4978_split_sizes_0 = const()[name = string("op_4978_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_4978_axis_0 = const()[name = string("op_4978_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 128]> var_4978_0, tensor<fp16, [1, 1, 128, 128]> var_4978_1 = split(axis = var_4978_axis_0, split_sizes = var_4978_split_sizes_0, x = out_245_cast_fp16)[name = string("op_4978")];
@@ -4020,7 +4025,7 @@ program(1.3)
             tensor<fp16, [1, 1, 128, 128]> var_4980 = mul(x = var_4978_1, y = const_287_promoted)[name = string("op_4980")];
             bool var_4982_interleave_0 = const()[name = string("op_4982_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 1, 128, 256]> var_4982 = concat(axis = var_23, interleave = var_4982_interleave_0, values = (var_4980, var_4978_0))[name = string("op_4982")];
-            tensor<fp16, [1, 1, 128, 256]> var_4983 = mul(x = var_4982, y = sin_1)[name = string("op_4983")];
+            tensor<fp16, [1, 1, 128, 256]> var_4983 = mul(x = var_4982, y = sin_1_quantized)[name = string("op_4983")];
             tensor<fp16, [1, 1, 128, 256]> hidden_states_241 = add(x = var_4977, y = var_4983)[name = string("hidden_states_241")];
             tensor<int32, [1]> hidden_states_243_axes_0 = const()[name = string("hidden_states_243_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 1, 1, 128, 256]> hidden_states_243 = expand_dims(axes = hidden_states_243_axes_0, x = hidden_states_241)[name = string("hidden_states_243")];
@@ -4050,14 +4055,14 @@ program(1.3)
             tensor<fp16, [1, 128, 3, 256]> var_5010 = transpose(perm = var_5009, x = var_5007_cast_fp16)[name = string("transpose_31")];
             tensor<fp16, [1, 128, 768]> attn_out_123 = reshape(shape = var_5012, x = var_5010)[name = string("attn_out_123")];
             tensor<int32, [3]> var_5014 = const()[name = string("op_5014"), val = tensor<int32, [3]>([0, 2, 1])];
-            tensor<fp16, [768, 768, 1]> squeeze_20 = const()[name = string("squeeze_20"), val = tensor<fp16, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(610705984)))];
+            tensor<fp16, [768, 768, 1]> squeeze_20_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306221568))), scale = tensor<fp16, [768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306811456))))[name = string("squeeze_20_quantized")];
             string var_5023_pad_type_0 = const()[name = string("op_5023_pad_type_0"), val = string("valid")];
             int32 var_5023_groups_0 = const()[name = string("op_5023_groups_0"), val = int32(1)];
             tensor<int32, [1]> var_5023_strides_0 = const()[name = string("op_5023_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> var_5023_pad_0 = const()[name = string("op_5023_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> var_5023_dilations_0 = const()[name = string("op_5023_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, [1, 768, 128]> var_5015 = transpose(perm = var_5014, x = attn_out_123)[name = string("transpose_30")];
-            tensor<fp16, [1, 768, 128]> var_5023 = conv(dilations = var_5023_dilations_0, groups = var_5023_groups_0, pad = var_5023_pad_0, pad_type = var_5023_pad_type_0, strides = var_5023_strides_0, weight = squeeze_20, x = var_5015)[name = string("op_5023")];
+            tensor<fp16, [1, 768, 128]> var_5023 = conv(dilations = var_5023_dilations_0, groups = var_5023_groups_0, pad = var_5023_pad_0, pad_type = var_5023_pad_type_0, strides = var_5023_strides_0, weight = squeeze_20_quantized, x = var_5015)[name = string("op_5023")];
             tensor<int32, [3]> var_5024 = const()[name = string("op_5024"), val = tensor<int32, [3]>([0, 2, 1])];
             fp16 const_288_promoted_to_fp16 = const()[name = string("const_288_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 128, 768]> x_329 = transpose(perm = var_5024, x = var_5023)[name = string("transpose_29")];
@@ -4069,7 +4074,7 @@ program(1.3)
             tensor<int32, [2]> var_5033_split_sizes_0 = const()[name = string("op_5033_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_5033_axis_0 = const()[name = string("op_5033_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_5033_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_5033_cast_fp16_1 = split(axis = var_5033_axis_0, split_sizes = var_5033_split_sizes_0, x = normed_575_cast_fp16)[name = string("op_5033_cast_fp16")];
-            tensor<fp16, [768]> var_5037_to_fp16 = const()[name = string("op_5037_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(611885696)))];
+            tensor<fp16, [768]> var_5037_to_fp16 = const()[name = string("op_5037_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306813056)))];
             tensor<fp16, [1, 128, 768]> out_247_cast_fp16 = mul(x = var_5033_cast_fp16_0, y = var_5037_to_fp16)[name = string("out_247_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_331_cast_fp16 = add(x = x_321_cast_fp16, y = out_247_cast_fp16)[name = string("x_331_cast_fp16")];
             fp16 const_290_promoted_to_fp16 = const()[name = string("const_290_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -4081,7 +4086,7 @@ program(1.3)
             tensor<int32, [2]> var_5049_split_sizes_0 = const()[name = string("op_5049_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_5049_axis_0 = const()[name = string("op_5049_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_5049_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_5049_cast_fp16_1 = split(axis = var_5049_axis_0, split_sizes = var_5049_split_sizes_0, x = normed_579_cast_fp16)[name = string("op_5049_cast_fp16")];
-            tensor<fp16, [768]> var_5053_to_fp16 = const()[name = string("op_5053_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(611887296)))];
+            tensor<fp16, [768]> var_5053_to_fp16 = const()[name = string("op_5053_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306814656)))];
             tensor<fp16, [1, 128, 768]> out_249_cast_fp16 = mul(x = var_5049_cast_fp16_0, y = var_5053_to_fp16)[name = string("out_249_cast_fp16")];
             tensor<int32, [3]> var_5060 = const()[name = string("op_5060"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> input_415_axes_0 = const()[name = string("input_415_axes_0"), val = tensor<int32, [1]>([2])];
@@ -4092,13 +4097,13 @@ program(1.3)
             tensor<int32, [4]> gate_81_pad_0 = const()[name = string("gate_81_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> gate_81_dilations_0 = const()[name = string("gate_81_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 gate_81_groups_0 = const()[name = string("gate_81_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> gate_81 = conv(dilations = gate_81_dilations_0, groups = gate_81_groups_0, pad = gate_81_pad_0, pad_type = gate_81_pad_type_0, strides = gate_81_strides_0, weight = encoder_layers_20_mlp_gate_proj_weight, x = input_415)[name = string("gate_81")];
+            tensor<fp16, [1, 1152, 1, 128]> gate_81 = conv(dilations = gate_81_dilations_0, groups = gate_81_groups_0, pad = gate_81_pad_0, pad_type = gate_81_pad_type_0, strides = gate_81_strides_0, weight = encoder_layers_20_mlp_gate_proj_weight_quantized, x = input_415)[name = string("gate_81")];
             string up_41_pad_type_0 = const()[name = string("up_41_pad_type_0"), val = string("valid")];
             tensor<int32, [2]> up_41_strides_0 = const()[name = string("up_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> up_41_pad_0 = const()[name = string("up_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> up_41_dilations_0 = const()[name = string("up_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 up_41_groups_0 = const()[name = string("up_41_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> up_41 = conv(dilations = up_41_dilations_0, groups = up_41_groups_0, pad = up_41_pad_0, pad_type = up_41_pad_type_0, strides = up_41_strides_0, weight = encoder_layers_20_mlp_up_proj_weight, x = input_415)[name = string("up_41")];
+            tensor<fp16, [1, 1152, 1, 128]> up_41 = conv(dilations = up_41_dilations_0, groups = up_41_groups_0, pad = up_41_pad_0, pad_type = up_41_pad_type_0, strides = up_41_strides_0, weight = encoder_layers_20_mlp_up_proj_weight_quantized, x = input_415)[name = string("up_41")];
             string gate_83_mode_0 = const()[name = string("gate_83_mode_0"), val = string("TANH_APPROXIMATION")];
             tensor<fp16, [1, 1152, 1, 128]> gate_83 = gelu(mode = gate_83_mode_0, x = gate_81)[name = string("gate_83")];
             tensor<fp16, [1, 1152, 1, 128]> input_417 = mul(x = gate_83, y = up_41)[name = string("input_417")];
@@ -4107,7 +4112,7 @@ program(1.3)
             tensor<int32, [4]> var_5082_pad_0 = const()[name = string("op_5082_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_5082_dilations_0 = const()[name = string("op_5082_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_5082_groups_0 = const()[name = string("op_5082_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_5082 = conv(dilations = var_5082_dilations_0, groups = var_5082_groups_0, pad = var_5082_pad_0, pad_type = var_5082_pad_type_0, strides = var_5082_strides_0, weight = encoder_layers_20_mlp_down_proj_weight, x = input_417)[name = string("op_5082")];
+            tensor<fp16, [1, 768, 1, 128]> var_5082 = conv(dilations = var_5082_dilations_0, groups = var_5082_groups_0, pad = var_5082_pad_0, pad_type = var_5082_pad_type_0, strides = var_5082_strides_0, weight = encoder_layers_20_mlp_down_proj_weight_quantized, x = input_417)[name = string("op_5082")];
             tensor<int32, [1]> var_5083_axes_0 = const()[name = string("op_5083_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 768, 128]> var_5083 = squeeze(axes = var_5083_axes_0, x = var_5082)[name = string("op_5083")];
             tensor<int32, [3]> var_5084 = const()[name = string("op_5084"), val = tensor<int32, [3]>([0, 2, 1])];
@@ -4121,7 +4126,7 @@ program(1.3)
             tensor<int32, [2]> var_5093_split_sizes_0 = const()[name = string("op_5093_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_5093_axis_0 = const()[name = string("op_5093_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_5093_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_5093_cast_fp16_1 = split(axis = var_5093_axis_0, split_sizes = var_5093_split_sizes_0, x = normed_585_cast_fp16)[name = string("op_5093_cast_fp16")];
-            tensor<fp16, [768]> var_5097_to_fp16 = const()[name = string("op_5097_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(611888896)))];
+            tensor<fp16, [768]> var_5097_to_fp16 = const()[name = string("op_5097_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306816256)))];
             tensor<fp16, [1, 128, 768]> out_251_cast_fp16 = mul(x = var_5093_cast_fp16_0, y = var_5097_to_fp16)[name = string("out_251_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_337_cast_fp16 = add(x = x_331_cast_fp16, y = out_251_cast_fp16)[name = string("x_337_cast_fp16")];
             fp16 const_294_promoted_to_fp16 = const()[name = string("const_294_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -4133,7 +4138,7 @@ program(1.3)
             tensor<int32, [2]> var_5131_split_sizes_0 = const()[name = string("op_5131_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_5131_axis_0 = const()[name = string("op_5131_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_5131_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_5131_cast_fp16_1 = split(axis = var_5131_axis_0, split_sizes = var_5131_split_sizes_0, x = normed_589_cast_fp16)[name = string("op_5131_cast_fp16")];
-            tensor<fp16, [768]> var_5135_to_fp16 = const()[name = string("op_5135_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(611890496)))];
+            tensor<fp16, [768]> var_5135_to_fp16 = const()[name = string("op_5135_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306817856)))];
             tensor<fp16, [1, 128, 768]> out_253_cast_fp16 = mul(x = var_5131_cast_fp16_0, y = var_5135_to_fp16)[name = string("out_253_cast_fp16")];
             tensor<int32, [3]> var_5141 = const()[name = string("op_5141"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> var_5143_axes_0 = const()[name = string("op_5143_axes_0"), val = tensor<int32, [1]>([2])];
@@ -4144,7 +4149,7 @@ program(1.3)
             tensor<int32, [4]> var_5150_pad_0 = const()[name = string("op_5150_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_5150_dilations_0 = const()[name = string("op_5150_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_5150_groups_0 = const()[name = string("op_5150_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_5150 = conv(dilations = var_5150_dilations_0, groups = var_5150_groups_0, pad = var_5150_pad_0, pad_type = var_5150_pad_type_0, strides = var_5150_strides_0, weight = encoder_layers_21_self_attn_q_proj_weight, x = var_5143_cast_fp16)[name = string("op_5150")];
+            tensor<fp16, [1, 768, 1, 128]> var_5150 = conv(dilations = var_5150_dilations_0, groups = var_5150_groups_0, pad = var_5150_pad_0, pad_type = var_5150_pad_type_0, strides = var_5150_strides_0, weight = encoder_layers_21_self_attn_q_proj_weight_quantized, x = var_5143_cast_fp16)[name = string("op_5150")];
             tensor<int32, [4]> var_5151 = const()[name = string("op_5151"), val = tensor<int32, [4]>([1, 3, 256, 128])];
             tensor<fp16, [1, 3, 256, 128]> var_5152 = reshape(shape = var_5151, x = var_5150)[name = string("op_5152")];
             tensor<int32, [4]> var_5153 = const()[name = string("op_5153"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -4153,7 +4158,7 @@ program(1.3)
             tensor<int32, [4]> var_5160_pad_0 = const()[name = string("op_5160_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_5160_dilations_0 = const()[name = string("op_5160_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_5160_groups_0 = const()[name = string("op_5160_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_5160 = conv(dilations = var_5160_dilations_0, groups = var_5160_groups_0, pad = var_5160_pad_0, pad_type = var_5160_pad_type_0, strides = var_5160_strides_0, weight = encoder_layers_21_self_attn_k_proj_weight, x = var_5143_cast_fp16)[name = string("op_5160")];
+            tensor<fp16, [1, 256, 1, 128]> var_5160 = conv(dilations = var_5160_dilations_0, groups = var_5160_groups_0, pad = var_5160_pad_0, pad_type = var_5160_pad_type_0, strides = var_5160_strides_0, weight = encoder_layers_21_self_attn_k_proj_weight_quantized, x = var_5143_cast_fp16)[name = string("op_5160")];
             tensor<int32, [4]> var_5161 = const()[name = string("op_5161"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_5162 = reshape(shape = var_5161, x = var_5160)[name = string("op_5162")];
             tensor<int32, [4]> var_5163 = const()[name = string("op_5163"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -4162,7 +4167,7 @@ program(1.3)
             tensor<int32, [4]> var_5170_pad_0 = const()[name = string("op_5170_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_5170_dilations_0 = const()[name = string("op_5170_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_5170_groups_0 = const()[name = string("op_5170_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_5170 = conv(dilations = var_5170_dilations_0, groups = var_5170_groups_0, pad = var_5170_pad_0, pad_type = var_5170_pad_type_0, strides = var_5170_strides_0, weight = encoder_layers_21_self_attn_v_proj_weight, x = var_5143_cast_fp16)[name = string("op_5170")];
+            tensor<fp16, [1, 256, 1, 128]> var_5170 = conv(dilations = var_5170_dilations_0, groups = var_5170_groups_0, pad = var_5170_pad_0, pad_type = var_5170_pad_type_0, strides = var_5170_strides_0, weight = encoder_layers_21_self_attn_v_proj_weight_quantized, x = var_5143_cast_fp16)[name = string("op_5170")];
             tensor<int32, [4]> var_5171 = const()[name = string("op_5171"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_5172 = reshape(shape = var_5171, x = var_5170)[name = string("op_5172")];
             tensor<int32, [4]> var_5173 = const()[name = string("op_5173"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -4176,7 +4181,7 @@ program(1.3)
             tensor<int32, [2]> var_5184_split_sizes_0 = const()[name = string("op_5184_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_5184_axis_0 = const()[name = string("op_5184_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 256]> var_5184_cast_fp16_0, tensor<fp16, [1, 3, 128, 256]> var_5184_cast_fp16_1 = split(axis = var_5184_axis_0, split_sizes = var_5184_split_sizes_0, x = normed_595_cast_fp16)[name = string("op_5184_cast_fp16")];
-            tensor<fp16, [256]> var_5188_to_fp16 = const()[name = string("op_5188_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(611892096)))];
+            tensor<fp16, [256]> var_5188_to_fp16 = const()[name = string("op_5188_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306819456)))];
             tensor<fp16, [1, 3, 128, 256]> out_255_cast_fp16 = mul(x = var_5184_cast_fp16_0, y = var_5188_to_fp16)[name = string("out_255_cast_fp16")];
             fp16 const_298_promoted_to_fp16 = const()[name = string("const_298_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 1, 128, 256]> k_127 = transpose(perm = var_5163, x = var_5162)[name = string("transpose_24")];
@@ -4188,9 +4193,9 @@ program(1.3)
             tensor<int32, [2]> var_5200_split_sizes_0 = const()[name = string("op_5200_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_5200_axis_0 = const()[name = string("op_5200_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 256]> var_5200_cast_fp16_0, tensor<fp16, [1, 1, 128, 256]> var_5200_cast_fp16_1 = split(axis = var_5200_axis_0, split_sizes = var_5200_split_sizes_0, x = normed_599_cast_fp16)[name = string("op_5200_cast_fp16")];
-            tensor<fp16, [256]> var_5204_to_fp16 = const()[name = string("op_5204_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(611892672)))];
+            tensor<fp16, [256]> var_5204_to_fp16 = const()[name = string("op_5204_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306820032)))];
             tensor<fp16, [1, 1, 128, 256]> out_257_cast_fp16 = mul(x = var_5200_cast_fp16_0, y = var_5204_to_fp16)[name = string("out_257_cast_fp16")];
-            tensor<fp16, [1, 3, 128, 256]> var_5207 = mul(x = out_255_cast_fp16, y = cos_1)[name = string("op_5207")];
+            tensor<fp16, [1, 3, 128, 256]> var_5207 = mul(x = out_255_cast_fp16, y = cos_1_quantized)[name = string("op_5207")];
             tensor<int32, [2]> var_5208_split_sizes_0 = const()[name = string("op_5208_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_5208_axis_0 = const()[name = string("op_5208_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 128]> var_5208_0, tensor<fp16, [1, 3, 128, 128]> var_5208_1 = split(axis = var_5208_axis_0, split_sizes = var_5208_split_sizes_0, x = out_255_cast_fp16)[name = string("op_5208")];
@@ -4198,9 +4203,9 @@ program(1.3)
             tensor<fp16, [1, 3, 128, 128]> var_5210 = mul(x = var_5208_1, y = const_300_promoted)[name = string("op_5210")];
             bool var_5212_interleave_0 = const()[name = string("op_5212_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 3, 128, 256]> var_5212 = concat(axis = var_23, interleave = var_5212_interleave_0, values = (var_5210, var_5208_0))[name = string("op_5212")];
-            tensor<fp16, [1, 3, 128, 256]> var_5213 = mul(x = var_5212, y = sin_1)[name = string("op_5213")];
+            tensor<fp16, [1, 3, 128, 256]> var_5213 = mul(x = var_5212, y = sin_1_quantized)[name = string("op_5213")];
             tensor<fp16, [1, 3, 128, 256]> q_131 = add(x = var_5207, y = var_5213)[name = string("q_131")];
-            tensor<fp16, [1, 1, 128, 256]> var_5215 = mul(x = out_257_cast_fp16, y = cos_1)[name = string("op_5215")];
+            tensor<fp16, [1, 1, 128, 256]> var_5215 = mul(x = out_257_cast_fp16, y = cos_1_quantized)[name = string("op_5215")];
             tensor<int32, [2]> var_5216_split_sizes_0 = const()[name = string("op_5216_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_5216_axis_0 = const()[name = string("op_5216_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 128]> var_5216_0, tensor<fp16, [1, 1, 128, 128]> var_5216_1 = split(axis = var_5216_axis_0, split_sizes = var_5216_split_sizes_0, x = out_257_cast_fp16)[name = string("op_5216")];
@@ -4208,7 +4213,7 @@ program(1.3)
             tensor<fp16, [1, 1, 128, 128]> var_5218 = mul(x = var_5216_1, y = const_301_promoted)[name = string("op_5218")];
             bool var_5220_interleave_0 = const()[name = string("op_5220_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 1, 128, 256]> var_5220 = concat(axis = var_23, interleave = var_5220_interleave_0, values = (var_5218, var_5216_0))[name = string("op_5220")];
-            tensor<fp16, [1, 1, 128, 256]> var_5221 = mul(x = var_5220, y = sin_1)[name = string("op_5221")];
+            tensor<fp16, [1, 1, 128, 256]> var_5221 = mul(x = var_5220, y = sin_1_quantized)[name = string("op_5221")];
             tensor<fp16, [1, 1, 128, 256]> hidden_states_253 = add(x = var_5215, y = var_5221)[name = string("hidden_states_253")];
             tensor<int32, [1]> hidden_states_255_axes_0 = const()[name = string("hidden_states_255_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 1, 1, 128, 256]> hidden_states_255 = expand_dims(axes = hidden_states_255_axes_0, x = hidden_states_253)[name = string("hidden_states_255")];
@@ -4238,14 +4243,14 @@ program(1.3)
             tensor<fp16, [1, 128, 3, 256]> var_5248 = transpose(perm = var_5247, x = var_5245_cast_fp16)[name = string("transpose_22")];
             tensor<fp16, [1, 128, 768]> attn_out_129 = reshape(shape = var_5250, x = var_5248)[name = string("attn_out_129")];
             tensor<int32, [3]> var_5252 = const()[name = string("op_5252"), val = tensor<int32, [3]>([0, 2, 1])];
-            tensor<fp16, [768, 768, 1]> squeeze_21 = const()[name = string("squeeze_21"), val = tensor<fp16, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(611893248)))];
+            tensor<fp16, [768, 768, 1]> squeeze_21_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306820608))), scale = tensor<fp16, [768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307410496))))[name = string("squeeze_21_quantized")];
             string var_5261_pad_type_0 = const()[name = string("op_5261_pad_type_0"), val = string("valid")];
             int32 var_5261_groups_0 = const()[name = string("op_5261_groups_0"), val = int32(1)];
             tensor<int32, [1]> var_5261_strides_0 = const()[name = string("op_5261_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> var_5261_pad_0 = const()[name = string("op_5261_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> var_5261_dilations_0 = const()[name = string("op_5261_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, [1, 768, 128]> var_5253 = transpose(perm = var_5252, x = attn_out_129)[name = string("transpose_21")];
-            tensor<fp16, [1, 768, 128]> var_5261 = conv(dilations = var_5261_dilations_0, groups = var_5261_groups_0, pad = var_5261_pad_0, pad_type = var_5261_pad_type_0, strides = var_5261_strides_0, weight = squeeze_21, x = var_5253)[name = string("op_5261")];
+            tensor<fp16, [1, 768, 128]> var_5261 = conv(dilations = var_5261_dilations_0, groups = var_5261_groups_0, pad = var_5261_pad_0, pad_type = var_5261_pad_type_0, strides = var_5261_strides_0, weight = squeeze_21_quantized, x = var_5253)[name = string("op_5261")];
             tensor<int32, [3]> var_5262 = const()[name = string("op_5262"), val = tensor<int32, [3]>([0, 2, 1])];
             fp16 const_302_promoted_to_fp16 = const()[name = string("const_302_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 128, 768]> x_345 = transpose(perm = var_5262, x = var_5261)[name = string("transpose_20")];
@@ -4257,7 +4262,7 @@ program(1.3)
             tensor<int32, [2]> var_5271_split_sizes_0 = const()[name = string("op_5271_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_5271_axis_0 = const()[name = string("op_5271_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_5271_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_5271_cast_fp16_1 = split(axis = var_5271_axis_0, split_sizes = var_5271_split_sizes_0, x = normed_603_cast_fp16)[name = string("op_5271_cast_fp16")];
-            tensor<fp16, [768]> var_5275_to_fp16 = const()[name = string("op_5275_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(613072960)))];
+            tensor<fp16, [768]> var_5275_to_fp16 = const()[name = string("op_5275_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307412096)))];
             tensor<fp16, [1, 128, 768]> out_259_cast_fp16 = mul(x = var_5271_cast_fp16_0, y = var_5275_to_fp16)[name = string("out_259_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_347_cast_fp16 = add(x = x_337_cast_fp16, y = out_259_cast_fp16)[name = string("x_347_cast_fp16")];
             fp16 const_304_promoted_to_fp16 = const()[name = string("const_304_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -4269,7 +4274,7 @@ program(1.3)
             tensor<int32, [2]> var_5287_split_sizes_0 = const()[name = string("op_5287_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_5287_axis_0 = const()[name = string("op_5287_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_5287_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_5287_cast_fp16_1 = split(axis = var_5287_axis_0, split_sizes = var_5287_split_sizes_0, x = normed_607_cast_fp16)[name = string("op_5287_cast_fp16")];
-            tensor<fp16, [768]> var_5291_to_fp16 = const()[name = string("op_5291_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(613074560)))];
+            tensor<fp16, [768]> var_5291_to_fp16 = const()[name = string("op_5291_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307413696)))];
             tensor<fp16, [1, 128, 768]> out_261_cast_fp16 = mul(x = var_5287_cast_fp16_0, y = var_5291_to_fp16)[name = string("out_261_cast_fp16")];
             tensor<int32, [3]> var_5298 = const()[name = string("op_5298"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> input_435_axes_0 = const()[name = string("input_435_axes_0"), val = tensor<int32, [1]>([2])];
@@ -4280,13 +4285,13 @@ program(1.3)
             tensor<int32, [4]> gate_85_pad_0 = const()[name = string("gate_85_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> gate_85_dilations_0 = const()[name = string("gate_85_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 gate_85_groups_0 = const()[name = string("gate_85_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> gate_85 = conv(dilations = gate_85_dilations_0, groups = gate_85_groups_0, pad = gate_85_pad_0, pad_type = gate_85_pad_type_0, strides = gate_85_strides_0, weight = encoder_layers_21_mlp_gate_proj_weight, x = input_435)[name = string("gate_85")];
+            tensor<fp16, [1, 1152, 1, 128]> gate_85 = conv(dilations = gate_85_dilations_0, groups = gate_85_groups_0, pad = gate_85_pad_0, pad_type = gate_85_pad_type_0, strides = gate_85_strides_0, weight = encoder_layers_21_mlp_gate_proj_weight_quantized, x = input_435)[name = string("gate_85")];
             string up_43_pad_type_0 = const()[name = string("up_43_pad_type_0"), val = string("valid")];
             tensor<int32, [2]> up_43_strides_0 = const()[name = string("up_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> up_43_pad_0 = const()[name = string("up_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> up_43_dilations_0 = const()[name = string("up_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 up_43_groups_0 = const()[name = string("up_43_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> up_43 = conv(dilations = up_43_dilations_0, groups = up_43_groups_0, pad = up_43_pad_0, pad_type = up_43_pad_type_0, strides = up_43_strides_0, weight = encoder_layers_21_mlp_up_proj_weight, x = input_435)[name = string("up_43")];
+            tensor<fp16, [1, 1152, 1, 128]> up_43 = conv(dilations = up_43_dilations_0, groups = up_43_groups_0, pad = up_43_pad_0, pad_type = up_43_pad_type_0, strides = up_43_strides_0, weight = encoder_layers_21_mlp_up_proj_weight_quantized, x = input_435)[name = string("up_43")];
             string gate_87_mode_0 = const()[name = string("gate_87_mode_0"), val = string("TANH_APPROXIMATION")];
             tensor<fp16, [1, 1152, 1, 128]> gate_87 = gelu(mode = gate_87_mode_0, x = gate_85)[name = string("gate_87")];
             tensor<fp16, [1, 1152, 1, 128]> input_437 = mul(x = gate_87, y = up_43)[name = string("input_437")];
@@ -4295,7 +4300,7 @@ program(1.3)
             tensor<int32, [4]> var_5320_pad_0 = const()[name = string("op_5320_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_5320_dilations_0 = const()[name = string("op_5320_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_5320_groups_0 = const()[name = string("op_5320_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_5320 = conv(dilations = var_5320_dilations_0, groups = var_5320_groups_0, pad = var_5320_pad_0, pad_type = var_5320_pad_type_0, strides = var_5320_strides_0, weight = encoder_layers_21_mlp_down_proj_weight, x = input_437)[name = string("op_5320")];
+            tensor<fp16, [1, 768, 1, 128]> var_5320 = conv(dilations = var_5320_dilations_0, groups = var_5320_groups_0, pad = var_5320_pad_0, pad_type = var_5320_pad_type_0, strides = var_5320_strides_0, weight = encoder_layers_21_mlp_down_proj_weight_quantized, x = input_437)[name = string("op_5320")];
             tensor<int32, [1]> var_5321_axes_0 = const()[name = string("op_5321_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 768, 128]> var_5321 = squeeze(axes = var_5321_axes_0, x = var_5320)[name = string("op_5321")];
             tensor<int32, [3]> var_5322 = const()[name = string("op_5322"), val = tensor<int32, [3]>([0, 2, 1])];
@@ -4309,7 +4314,7 @@ program(1.3)
             tensor<int32, [2]> var_5331_split_sizes_0 = const()[name = string("op_5331_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_5331_axis_0 = const()[name = string("op_5331_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_5331_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_5331_cast_fp16_1 = split(axis = var_5331_axis_0, split_sizes = var_5331_split_sizes_0, x = normed_613_cast_fp16)[name = string("op_5331_cast_fp16")];
-            tensor<fp16, [768]> var_5335_to_fp16 = const()[name = string("op_5335_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(613076160)))];
+            tensor<fp16, [768]> var_5335_to_fp16 = const()[name = string("op_5335_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307415296)))];
             tensor<fp16, [1, 128, 768]> out_263_cast_fp16 = mul(x = var_5331_cast_fp16_0, y = var_5335_to_fp16)[name = string("out_263_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_353_cast_fp16 = add(x = x_347_cast_fp16, y = out_263_cast_fp16)[name = string("x_353_cast_fp16")];
             fp16 const_308_promoted_to_fp16 = const()[name = string("const_308_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -4321,7 +4326,7 @@ program(1.3)
             tensor<int32, [2]> var_5369_split_sizes_0 = const()[name = string("op_5369_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_5369_axis_0 = const()[name = string("op_5369_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_5369_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_5369_cast_fp16_1 = split(axis = var_5369_axis_0, split_sizes = var_5369_split_sizes_0, x = normed_617_cast_fp16)[name = string("op_5369_cast_fp16")];
-            tensor<fp16, [768]> var_5373_to_fp16 = const()[name = string("op_5373_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(613077760)))];
+            tensor<fp16, [768]> var_5373_to_fp16 = const()[name = string("op_5373_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307416896)))];
             tensor<fp16, [1, 128, 768]> out_265_cast_fp16 = mul(x = var_5369_cast_fp16_0, y = var_5373_to_fp16)[name = string("out_265_cast_fp16")];
             tensor<int32, [3]> var_5379 = const()[name = string("op_5379"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> var_5381_axes_0 = const()[name = string("op_5381_axes_0"), val = tensor<int32, [1]>([2])];
@@ -4332,7 +4337,7 @@ program(1.3)
             tensor<int32, [4]> var_5388_pad_0 = const()[name = string("op_5388_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_5388_dilations_0 = const()[name = string("op_5388_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_5388_groups_0 = const()[name = string("op_5388_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_5388 = conv(dilations = var_5388_dilations_0, groups = var_5388_groups_0, pad = var_5388_pad_0, pad_type = var_5388_pad_type_0, strides = var_5388_strides_0, weight = encoder_layers_22_self_attn_q_proj_weight, x = var_5381_cast_fp16)[name = string("op_5388")];
+            tensor<fp16, [1, 768, 1, 128]> var_5388 = conv(dilations = var_5388_dilations_0, groups = var_5388_groups_0, pad = var_5388_pad_0, pad_type = var_5388_pad_type_0, strides = var_5388_strides_0, weight = encoder_layers_22_self_attn_q_proj_weight_quantized, x = var_5381_cast_fp16)[name = string("op_5388")];
             tensor<int32, [4]> var_5389 = const()[name = string("op_5389"), val = tensor<int32, [4]>([1, 3, 256, 128])];
             tensor<fp16, [1, 3, 256, 128]> var_5390 = reshape(shape = var_5389, x = var_5388)[name = string("op_5390")];
             tensor<int32, [4]> var_5391 = const()[name = string("op_5391"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -4341,7 +4346,7 @@ program(1.3)
             tensor<int32, [4]> var_5398_pad_0 = const()[name = string("op_5398_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_5398_dilations_0 = const()[name = string("op_5398_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_5398_groups_0 = const()[name = string("op_5398_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_5398 = conv(dilations = var_5398_dilations_0, groups = var_5398_groups_0, pad = var_5398_pad_0, pad_type = var_5398_pad_type_0, strides = var_5398_strides_0, weight = encoder_layers_22_self_attn_k_proj_weight, x = var_5381_cast_fp16)[name = string("op_5398")];
+            tensor<fp16, [1, 256, 1, 128]> var_5398 = conv(dilations = var_5398_dilations_0, groups = var_5398_groups_0, pad = var_5398_pad_0, pad_type = var_5398_pad_type_0, strides = var_5398_strides_0, weight = encoder_layers_22_self_attn_k_proj_weight_quantized, x = var_5381_cast_fp16)[name = string("op_5398")];
             tensor<int32, [4]> var_5399 = const()[name = string("op_5399"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_5400 = reshape(shape = var_5399, x = var_5398)[name = string("op_5400")];
             tensor<int32, [4]> var_5401 = const()[name = string("op_5401"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -4350,7 +4355,7 @@ program(1.3)
             tensor<int32, [4]> var_5408_pad_0 = const()[name = string("op_5408_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_5408_dilations_0 = const()[name = string("op_5408_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_5408_groups_0 = const()[name = string("op_5408_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_5408 = conv(dilations = var_5408_dilations_0, groups = var_5408_groups_0, pad = var_5408_pad_0, pad_type = var_5408_pad_type_0, strides = var_5408_strides_0, weight = encoder_layers_22_self_attn_v_proj_weight, x = var_5381_cast_fp16)[name = string("op_5408")];
+            tensor<fp16, [1, 256, 1, 128]> var_5408 = conv(dilations = var_5408_dilations_0, groups = var_5408_groups_0, pad = var_5408_pad_0, pad_type = var_5408_pad_type_0, strides = var_5408_strides_0, weight = encoder_layers_22_self_attn_v_proj_weight_quantized, x = var_5381_cast_fp16)[name = string("op_5408")];
             tensor<int32, [4]> var_5409 = const()[name = string("op_5409"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_5410 = reshape(shape = var_5409, x = var_5408)[name = string("op_5410")];
             tensor<int32, [4]> var_5411 = const()[name = string("op_5411"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -4364,7 +4369,7 @@ program(1.3)
             tensor<int32, [2]> var_5422_split_sizes_0 = const()[name = string("op_5422_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_5422_axis_0 = const()[name = string("op_5422_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 256]> var_5422_cast_fp16_0, tensor<fp16, [1, 3, 128, 256]> var_5422_cast_fp16_1 = split(axis = var_5422_axis_0, split_sizes = var_5422_split_sizes_0, x = normed_623_cast_fp16)[name = string("op_5422_cast_fp16")];
-            tensor<fp16, [256]> var_5426_to_fp16 = const()[name = string("op_5426_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(613079360)))];
+            tensor<fp16, [256]> var_5426_to_fp16 = const()[name = string("op_5426_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307418496)))];
             tensor<fp16, [1, 3, 128, 256]> out_267_cast_fp16 = mul(x = var_5422_cast_fp16_0, y = var_5426_to_fp16)[name = string("out_267_cast_fp16")];
             fp16 const_312_promoted_to_fp16 = const()[name = string("const_312_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 1, 128, 256]> k_133 = transpose(perm = var_5401, x = var_5400)[name = string("transpose_15")];
@@ -4376,9 +4381,9 @@ program(1.3)
             tensor<int32, [2]> var_5438_split_sizes_0 = const()[name = string("op_5438_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_5438_axis_0 = const()[name = string("op_5438_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 256]> var_5438_cast_fp16_0, tensor<fp16, [1, 1, 128, 256]> var_5438_cast_fp16_1 = split(axis = var_5438_axis_0, split_sizes = var_5438_split_sizes_0, x = normed_627_cast_fp16)[name = string("op_5438_cast_fp16")];
-            tensor<fp16, [256]> var_5442_to_fp16 = const()[name = string("op_5442_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(613079936)))];
+            tensor<fp16, [256]> var_5442_to_fp16 = const()[name = string("op_5442_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307419072)))];
             tensor<fp16, [1, 1, 128, 256]> out_269_cast_fp16 = mul(x = var_5438_cast_fp16_0, y = var_5442_to_fp16)[name = string("out_269_cast_fp16")];
-            tensor<fp16, [1, 3, 128, 256]> var_5445 = mul(x = out_267_cast_fp16, y = cos_1)[name = string("op_5445")];
+            tensor<fp16, [1, 3, 128, 256]> var_5445 = mul(x = out_267_cast_fp16, y = cos_1_quantized)[name = string("op_5445")];
             tensor<int32, [2]> var_5446_split_sizes_0 = const()[name = string("op_5446_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_5446_axis_0 = const()[name = string("op_5446_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 128]> var_5446_0, tensor<fp16, [1, 3, 128, 128]> var_5446_1 = split(axis = var_5446_axis_0, split_sizes = var_5446_split_sizes_0, x = out_267_cast_fp16)[name = string("op_5446")];
@@ -4386,9 +4391,9 @@ program(1.3)
             tensor<fp16, [1, 3, 128, 128]> var_5448 = mul(x = var_5446_1, y = const_314_promoted)[name = string("op_5448")];
             bool var_5450_interleave_0 = const()[name = string("op_5450_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 3, 128, 256]> var_5450 = concat(axis = var_23, interleave = var_5450_interleave_0, values = (var_5448, var_5446_0))[name = string("op_5450")];
-            tensor<fp16, [1, 3, 128, 256]> var_5451 = mul(x = var_5450, y = sin_1)[name = string("op_5451")];
+            tensor<fp16, [1, 3, 128, 256]> var_5451 = mul(x = var_5450, y = sin_1_quantized)[name = string("op_5451")];
             tensor<fp16, [1, 3, 128, 256]> q_137 = add(x = var_5445, y = var_5451)[name = string("q_137")];
-            tensor<fp16, [1, 1, 128, 256]> var_5453 = mul(x = out_269_cast_fp16, y = cos_1)[name = string("op_5453")];
+            tensor<fp16, [1, 1, 128, 256]> var_5453 = mul(x = out_269_cast_fp16, y = cos_1_quantized)[name = string("op_5453")];
             tensor<int32, [2]> var_5454_split_sizes_0 = const()[name = string("op_5454_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_5454_axis_0 = const()[name = string("op_5454_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 128]> var_5454_0, tensor<fp16, [1, 1, 128, 128]> var_5454_1 = split(axis = var_5454_axis_0, split_sizes = var_5454_split_sizes_0, x = out_269_cast_fp16)[name = string("op_5454")];
@@ -4396,7 +4401,7 @@ program(1.3)
             tensor<fp16, [1, 1, 128, 128]> var_5456 = mul(x = var_5454_1, y = const_315_promoted)[name = string("op_5456")];
             bool var_5458_interleave_0 = const()[name = string("op_5458_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 1, 128, 256]> var_5458 = concat(axis = var_23, interleave = var_5458_interleave_0, values = (var_5456, var_5454_0))[name = string("op_5458")];
-            tensor<fp16, [1, 1, 128, 256]> var_5459 = mul(x = var_5458, y = sin_1)[name = string("op_5459")];
+            tensor<fp16, [1, 1, 128, 256]> var_5459 = mul(x = var_5458, y = sin_1_quantized)[name = string("op_5459")];
             tensor<fp16, [1, 1, 128, 256]> hidden_states_265 = add(x = var_5453, y = var_5459)[name = string("hidden_states_265")];
             tensor<int32, [1]> hidden_states_267_axes_0 = const()[name = string("hidden_states_267_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 1, 1, 128, 256]> hidden_states_267 = expand_dims(axes = hidden_states_267_axes_0, x = hidden_states_265)[name = string("hidden_states_267")];
@@ -4426,14 +4431,14 @@ program(1.3)
             tensor<fp16, [1, 128, 3, 256]> var_5486 = transpose(perm = var_5485, x = var_5483_cast_fp16)[name = string("transpose_13")];
             tensor<fp16, [1, 128, 768]> attn_out_135 = reshape(shape = var_5488, x = var_5486)[name = string("attn_out_135")];
             tensor<int32, [3]> var_5490 = const()[name = string("op_5490"), val = tensor<int32, [3]>([0, 2, 1])];
-            tensor<fp16, [768, 768, 1]> squeeze_22 = const()[name = string("squeeze_22"), val = tensor<fp16, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(613080512)))];
+            tensor<fp16, [768, 768, 1]> squeeze_22_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307419648))), scale = tensor<fp16, [768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308009536))))[name = string("squeeze_22_quantized")];
             string var_5499_pad_type_0 = const()[name = string("op_5499_pad_type_0"), val = string("valid")];
             int32 var_5499_groups_0 = const()[name = string("op_5499_groups_0"), val = int32(1)];
             tensor<int32, [1]> var_5499_strides_0 = const()[name = string("op_5499_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> var_5499_pad_0 = const()[name = string("op_5499_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> var_5499_dilations_0 = const()[name = string("op_5499_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, [1, 768, 128]> var_5491 = transpose(perm = var_5490, x = attn_out_135)[name = string("transpose_12")];
-            tensor<fp16, [1, 768, 128]> var_5499 = conv(dilations = var_5499_dilations_0, groups = var_5499_groups_0, pad = var_5499_pad_0, pad_type = var_5499_pad_type_0, strides = var_5499_strides_0, weight = squeeze_22, x = var_5491)[name = string("op_5499")];
+            tensor<fp16, [1, 768, 128]> var_5499 = conv(dilations = var_5499_dilations_0, groups = var_5499_groups_0, pad = var_5499_pad_0, pad_type = var_5499_pad_type_0, strides = var_5499_strides_0, weight = squeeze_22_quantized, x = var_5491)[name = string("op_5499")];
             tensor<int32, [3]> var_5500 = const()[name = string("op_5500"), val = tensor<int32, [3]>([0, 2, 1])];
             fp16 const_316_promoted_to_fp16 = const()[name = string("const_316_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 128, 768]> x_361 = transpose(perm = var_5500, x = var_5499)[name = string("transpose_11")];
@@ -4445,7 +4450,7 @@ program(1.3)
             tensor<int32, [2]> var_5509_split_sizes_0 = const()[name = string("op_5509_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_5509_axis_0 = const()[name = string("op_5509_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_5509_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_5509_cast_fp16_1 = split(axis = var_5509_axis_0, split_sizes = var_5509_split_sizes_0, x = normed_631_cast_fp16)[name = string("op_5509_cast_fp16")];
-            tensor<fp16, [768]> var_5513_to_fp16 = const()[name = string("op_5513_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(614260224)))];
+            tensor<fp16, [768]> var_5513_to_fp16 = const()[name = string("op_5513_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308011136)))];
             tensor<fp16, [1, 128, 768]> out_271_cast_fp16 = mul(x = var_5509_cast_fp16_0, y = var_5513_to_fp16)[name = string("out_271_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_363_cast_fp16 = add(x = x_353_cast_fp16, y = out_271_cast_fp16)[name = string("x_363_cast_fp16")];
             fp16 const_318_promoted_to_fp16 = const()[name = string("const_318_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -4457,7 +4462,7 @@ program(1.3)
             tensor<int32, [2]> var_5525_split_sizes_0 = const()[name = string("op_5525_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_5525_axis_0 = const()[name = string("op_5525_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_5525_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_5525_cast_fp16_1 = split(axis = var_5525_axis_0, split_sizes = var_5525_split_sizes_0, x = normed_635_cast_fp16)[name = string("op_5525_cast_fp16")];
-            tensor<fp16, [768]> var_5529_to_fp16 = const()[name = string("op_5529_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(614261824)))];
+            tensor<fp16, [768]> var_5529_to_fp16 = const()[name = string("op_5529_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308012736)))];
             tensor<fp16, [1, 128, 768]> out_273_cast_fp16 = mul(x = var_5525_cast_fp16_0, y = var_5529_to_fp16)[name = string("out_273_cast_fp16")];
             tensor<int32, [3]> var_5536 = const()[name = string("op_5536"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> input_455_axes_0 = const()[name = string("input_455_axes_0"), val = tensor<int32, [1]>([2])];
@@ -4468,13 +4473,13 @@ program(1.3)
             tensor<int32, [4]> gate_89_pad_0 = const()[name = string("gate_89_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> gate_89_dilations_0 = const()[name = string("gate_89_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 gate_89_groups_0 = const()[name = string("gate_89_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> gate_89 = conv(dilations = gate_89_dilations_0, groups = gate_89_groups_0, pad = gate_89_pad_0, pad_type = gate_89_pad_type_0, strides = gate_89_strides_0, weight = encoder_layers_22_mlp_gate_proj_weight, x = input_455)[name = string("gate_89")];
+            tensor<fp16, [1, 1152, 1, 128]> gate_89 = conv(dilations = gate_89_dilations_0, groups = gate_89_groups_0, pad = gate_89_pad_0, pad_type = gate_89_pad_type_0, strides = gate_89_strides_0, weight = encoder_layers_22_mlp_gate_proj_weight_quantized, x = input_455)[name = string("gate_89")];
             string up_45_pad_type_0 = const()[name = string("up_45_pad_type_0"), val = string("valid")];
             tensor<int32, [2]> up_45_strides_0 = const()[name = string("up_45_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> up_45_pad_0 = const()[name = string("up_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> up_45_dilations_0 = const()[name = string("up_45_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 up_45_groups_0 = const()[name = string("up_45_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> up_45 = conv(dilations = up_45_dilations_0, groups = up_45_groups_0, pad = up_45_pad_0, pad_type = up_45_pad_type_0, strides = up_45_strides_0, weight = encoder_layers_22_mlp_up_proj_weight, x = input_455)[name = string("up_45")];
+            tensor<fp16, [1, 1152, 1, 128]> up_45 = conv(dilations = up_45_dilations_0, groups = up_45_groups_0, pad = up_45_pad_0, pad_type = up_45_pad_type_0, strides = up_45_strides_0, weight = encoder_layers_22_mlp_up_proj_weight_quantized, x = input_455)[name = string("up_45")];
             string gate_91_mode_0 = const()[name = string("gate_91_mode_0"), val = string("TANH_APPROXIMATION")];
             tensor<fp16, [1, 1152, 1, 128]> gate_91 = gelu(mode = gate_91_mode_0, x = gate_89)[name = string("gate_91")];
             tensor<fp16, [1, 1152, 1, 128]> input_457 = mul(x = gate_91, y = up_45)[name = string("input_457")];
@@ -4483,7 +4488,7 @@ program(1.3)
             tensor<int32, [4]> var_5558_pad_0 = const()[name = string("op_5558_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_5558_dilations_0 = const()[name = string("op_5558_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_5558_groups_0 = const()[name = string("op_5558_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_5558 = conv(dilations = var_5558_dilations_0, groups = var_5558_groups_0, pad = var_5558_pad_0, pad_type = var_5558_pad_type_0, strides = var_5558_strides_0, weight = encoder_layers_22_mlp_down_proj_weight, x = input_457)[name = string("op_5558")];
+            tensor<fp16, [1, 768, 1, 128]> var_5558 = conv(dilations = var_5558_dilations_0, groups = var_5558_groups_0, pad = var_5558_pad_0, pad_type = var_5558_pad_type_0, strides = var_5558_strides_0, weight = encoder_layers_22_mlp_down_proj_weight_quantized, x = input_457)[name = string("op_5558")];
             tensor<int32, [1]> var_5559_axes_0 = const()[name = string("op_5559_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 768, 128]> var_5559 = squeeze(axes = var_5559_axes_0, x = var_5558)[name = string("op_5559")];
             tensor<int32, [3]> var_5560 = const()[name = string("op_5560"), val = tensor<int32, [3]>([0, 2, 1])];
@@ -4497,7 +4502,7 @@ program(1.3)
             tensor<int32, [2]> var_5569_split_sizes_0 = const()[name = string("op_5569_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_5569_axis_0 = const()[name = string("op_5569_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_5569_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_5569_cast_fp16_1 = split(axis = var_5569_axis_0, split_sizes = var_5569_split_sizes_0, x = normed_641_cast_fp16)[name = string("op_5569_cast_fp16")];
-            tensor<fp16, [768]> var_5573_to_fp16 = const()[name = string("op_5573_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(614263424)))];
+            tensor<fp16, [768]> var_5573_to_fp16 = const()[name = string("op_5573_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308014336)))];
             tensor<fp16, [1, 128, 768]> out_275_cast_fp16 = mul(x = var_5569_cast_fp16_0, y = var_5573_to_fp16)[name = string("out_275_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_369_cast_fp16 = add(x = x_363_cast_fp16, y = out_275_cast_fp16)[name = string("x_369_cast_fp16")];
             fp16 const_322_promoted_to_fp16 = const()[name = string("const_322_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -4509,7 +4514,7 @@ program(1.3)
             tensor<int32, [2]> var_5607_split_sizes_0 = const()[name = string("op_5607_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_5607_axis_0 = const()[name = string("op_5607_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_5607_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_5607_cast_fp16_1 = split(axis = var_5607_axis_0, split_sizes = var_5607_split_sizes_0, x = normed_645_cast_fp16)[name = string("op_5607_cast_fp16")];
-            tensor<fp16, [768]> var_5611_to_fp16 = const()[name = string("op_5611_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(614265024)))];
+            tensor<fp16, [768]> var_5611_to_fp16 = const()[name = string("op_5611_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308015936)))];
             tensor<fp16, [1, 128, 768]> out_277_cast_fp16 = mul(x = var_5607_cast_fp16_0, y = var_5611_to_fp16)[name = string("out_277_cast_fp16")];
             tensor<int32, [3]> var_5617 = const()[name = string("op_5617"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> var_5619_axes_0 = const()[name = string("op_5619_axes_0"), val = tensor<int32, [1]>([2])];
@@ -4520,7 +4525,7 @@ program(1.3)
             tensor<int32, [4]> var_5626_pad_0 = const()[name = string("op_5626_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_5626_dilations_0 = const()[name = string("op_5626_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_5626_groups_0 = const()[name = string("op_5626_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_5626 = conv(dilations = var_5626_dilations_0, groups = var_5626_groups_0, pad = var_5626_pad_0, pad_type = var_5626_pad_type_0, strides = var_5626_strides_0, weight = encoder_layers_23_self_attn_q_proj_weight, x = var_5619_cast_fp16)[name = string("op_5626")];
+            tensor<fp16, [1, 768, 1, 128]> var_5626 = conv(dilations = var_5626_dilations_0, groups = var_5626_groups_0, pad = var_5626_pad_0, pad_type = var_5626_pad_type_0, strides = var_5626_strides_0, weight = encoder_layers_23_self_attn_q_proj_weight_quantized, x = var_5619_cast_fp16)[name = string("op_5626")];
             tensor<int32, [4]> var_5627 = const()[name = string("op_5627"), val = tensor<int32, [4]>([1, 3, 256, 128])];
             tensor<fp16, [1, 3, 256, 128]> var_5628 = reshape(shape = var_5627, x = var_5626)[name = string("op_5628")];
             tensor<int32, [4]> var_5629 = const()[name = string("op_5629"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -4529,7 +4534,7 @@ program(1.3)
             tensor<int32, [4]> var_5636_pad_0 = const()[name = string("op_5636_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_5636_dilations_0 = const()[name = string("op_5636_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_5636_groups_0 = const()[name = string("op_5636_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_5636 = conv(dilations = var_5636_dilations_0, groups = var_5636_groups_0, pad = var_5636_pad_0, pad_type = var_5636_pad_type_0, strides = var_5636_strides_0, weight = encoder_layers_23_self_attn_k_proj_weight, x = var_5619_cast_fp16)[name = string("op_5636")];
+            tensor<fp16, [1, 256, 1, 128]> var_5636 = conv(dilations = var_5636_dilations_0, groups = var_5636_groups_0, pad = var_5636_pad_0, pad_type = var_5636_pad_type_0, strides = var_5636_strides_0, weight = encoder_layers_23_self_attn_k_proj_weight_quantized, x = var_5619_cast_fp16)[name = string("op_5636")];
             tensor<int32, [4]> var_5637 = const()[name = string("op_5637"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_5638 = reshape(shape = var_5637, x = var_5636)[name = string("op_5638")];
             tensor<int32, [4]> var_5639 = const()[name = string("op_5639"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -4538,7 +4543,7 @@ program(1.3)
             tensor<int32, [4]> var_5646_pad_0 = const()[name = string("op_5646_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_5646_dilations_0 = const()[name = string("op_5646_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_5646_groups_0 = const()[name = string("op_5646_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 256, 1, 128]> var_5646 = conv(dilations = var_5646_dilations_0, groups = var_5646_groups_0, pad = var_5646_pad_0, pad_type = var_5646_pad_type_0, strides = var_5646_strides_0, weight = encoder_layers_23_self_attn_v_proj_weight, x = var_5619_cast_fp16)[name = string("op_5646")];
+            tensor<fp16, [1, 256, 1, 128]> var_5646 = conv(dilations = var_5646_dilations_0, groups = var_5646_groups_0, pad = var_5646_pad_0, pad_type = var_5646_pad_type_0, strides = var_5646_strides_0, weight = encoder_layers_23_self_attn_v_proj_weight_quantized, x = var_5619_cast_fp16)[name = string("op_5646")];
             tensor<int32, [4]> var_5647 = const()[name = string("op_5647"), val = tensor<int32, [4]>([1, 1, 256, 128])];
             tensor<fp16, [1, 1, 256, 128]> var_5648 = reshape(shape = var_5647, x = var_5646)[name = string("op_5648")];
             tensor<int32, [4]> var_5649 = const()[name = string("op_5649"), val = tensor<int32, [4]>([0, 1, 3, 2])];
@@ -4552,7 +4557,7 @@ program(1.3)
             tensor<int32, [2]> var_5660_split_sizes_0 = const()[name = string("op_5660_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_5660_axis_0 = const()[name = string("op_5660_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 256]> var_5660_cast_fp16_0, tensor<fp16, [1, 3, 128, 256]> var_5660_cast_fp16_1 = split(axis = var_5660_axis_0, split_sizes = var_5660_split_sizes_0, x = normed_651_cast_fp16)[name = string("op_5660_cast_fp16")];
-            tensor<fp16, [256]> var_5664_to_fp16 = const()[name = string("op_5664_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(614266624)))];
+            tensor<fp16, [256]> var_5664_to_fp16 = const()[name = string("op_5664_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308017536)))];
             tensor<fp16, [1, 3, 128, 256]> out_279_cast_fp16 = mul(x = var_5660_cast_fp16_0, y = var_5664_to_fp16)[name = string("out_279_cast_fp16")];
             fp16 const_326_promoted_to_fp16 = const()[name = string("const_326_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 1, 128, 256]> k_139 = transpose(perm = var_5639, x = var_5638)[name = string("transpose_6")];
@@ -4564,9 +4569,9 @@ program(1.3)
             tensor<int32, [2]> var_5676_split_sizes_0 = const()[name = string("op_5676_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
             int32 var_5676_axis_0 = const()[name = string("op_5676_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 256]> var_5676_cast_fp16_0, tensor<fp16, [1, 1, 128, 256]> var_5676_cast_fp16_1 = split(axis = var_5676_axis_0, split_sizes = var_5676_split_sizes_0, x = normed_655_cast_fp16)[name = string("op_5676_cast_fp16")];
-            tensor<fp16, [256]> var_5680_to_fp16 = const()[name = string("op_5680_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(614267200)))];
+            tensor<fp16, [256]> var_5680_to_fp16 = const()[name = string("op_5680_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308018112)))];
             tensor<fp16, [1, 1, 128, 256]> out_281_cast_fp16 = mul(x = var_5676_cast_fp16_0, y = var_5680_to_fp16)[name = string("out_281_cast_fp16")];
-            tensor<fp16, [1, 3, 128, 256]> var_5683 = mul(x = out_279_cast_fp16, y = cos)[name = string("op_5683")];
+            tensor<fp16, [1, 3, 128, 256]> var_5683 = mul(x = out_279_cast_fp16, y = cos_quantized)[name = string("op_5683")];
             tensor<int32, [2]> var_5684_split_sizes_0 = const()[name = string("op_5684_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_5684_axis_0 = const()[name = string("op_5684_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 3, 128, 128]> var_5684_0, tensor<fp16, [1, 3, 128, 128]> var_5684_1 = split(axis = var_5684_axis_0, split_sizes = var_5684_split_sizes_0, x = out_279_cast_fp16)[name = string("op_5684")];
@@ -4574,9 +4579,9 @@ program(1.3)
             tensor<fp16, [1, 3, 128, 128]> var_5686 = mul(x = var_5684_1, y = const_328_promoted)[name = string("op_5686")];
             bool var_5688_interleave_0 = const()[name = string("op_5688_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 3, 128, 256]> var_5688 = concat(axis = var_23, interleave = var_5688_interleave_0, values = (var_5686, var_5684_0))[name = string("op_5688")];
-            tensor<fp16, [1, 3, 128, 256]> var_5689 = mul(x = var_5688, y = sin)[name = string("op_5689")];
+            tensor<fp16, [1, 3, 128, 256]> var_5689 = mul(x = var_5688, y = sin_quantized)[name = string("op_5689")];
             tensor<fp16, [1, 3, 128, 256]> q = add(x = var_5683, y = var_5689)[name = string("q")];
-            tensor<fp16, [1, 1, 128, 256]> var_5691 = mul(x = out_281_cast_fp16, y = cos)[name = string("op_5691")];
+            tensor<fp16, [1, 1, 128, 256]> var_5691 = mul(x = out_281_cast_fp16, y = cos_quantized)[name = string("op_5691")];
             tensor<int32, [2]> var_5692_split_sizes_0 = const()[name = string("op_5692_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
             int32 var_5692_axis_0 = const()[name = string("op_5692_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 1, 128, 128]> var_5692_0, tensor<fp16, [1, 1, 128, 128]> var_5692_1 = split(axis = var_5692_axis_0, split_sizes = var_5692_split_sizes_0, x = out_281_cast_fp16)[name = string("op_5692")];
@@ -4584,7 +4589,7 @@ program(1.3)
             tensor<fp16, [1, 1, 128, 128]> var_5694 = mul(x = var_5692_1, y = const_329_promoted)[name = string("op_5694")];
             bool var_5696_interleave_0 = const()[name = string("op_5696_interleave_0"), val = bool(false)];
             tensor<fp16, [1, 1, 128, 256]> var_5696 = concat(axis = var_23, interleave = var_5696_interleave_0, values = (var_5694, var_5692_0))[name = string("op_5696")];
-            tensor<fp16, [1, 1, 128, 256]> var_5697 = mul(x = var_5696, y = sin)[name = string("op_5697")];
+            tensor<fp16, [1, 1, 128, 256]> var_5697 = mul(x = var_5696, y = sin_quantized)[name = string("op_5697")];
             tensor<fp16, [1, 1, 128, 256]> hidden_states_277 = add(x = var_5691, y = var_5697)[name = string("hidden_states_277")];
             tensor<int32, [1]> hidden_states_279_axes_0 = const()[name = string("hidden_states_279_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 1, 1, 128, 256]> hidden_states_279 = expand_dims(axes = hidden_states_279_axes_0, x = hidden_states_277)[name = string("hidden_states_279")];
@@ -4614,14 +4619,14 @@ program(1.3)
             tensor<fp16, [1, 128, 3, 256]> var_5724 = transpose(perm = var_5723, x = var_5721_cast_fp16)[name = string("transpose_4")];
             tensor<fp16, [1, 128, 768]> attn_out_141 = reshape(shape = var_5726, x = var_5724)[name = string("attn_out_141")];
             tensor<int32, [3]> var_5728 = const()[name = string("op_5728"), val = tensor<int32, [3]>([0, 2, 1])];
-            tensor<fp16, [768, 768, 1]> squeeze_23 = const()[name = string("squeeze_23"), val = tensor<fp16, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(614267776)))];
+            tensor<fp16, [768, 768, 1]> squeeze_23_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308018688))), scale = tensor<fp16, [768, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308608576))))[name = string("squeeze_23_quantized")];
             string var_5737_pad_type_0 = const()[name = string("op_5737_pad_type_0"), val = string("valid")];
             int32 var_5737_groups_0 = const()[name = string("op_5737_groups_0"), val = int32(1)];
             tensor<int32, [1]> var_5737_strides_0 = const()[name = string("op_5737_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> var_5737_pad_0 = const()[name = string("op_5737_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> var_5737_dilations_0 = const()[name = string("op_5737_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, [1, 768, 128]> var_5729 = transpose(perm = var_5728, x = attn_out_141)[name = string("transpose_3")];
-            tensor<fp16, [1, 768, 128]> var_5737 = conv(dilations = var_5737_dilations_0, groups = var_5737_groups_0, pad = var_5737_pad_0, pad_type = var_5737_pad_type_0, strides = var_5737_strides_0, weight = squeeze_23, x = var_5729)[name = string("op_5737")];
+            tensor<fp16, [1, 768, 128]> var_5737 = conv(dilations = var_5737_dilations_0, groups = var_5737_groups_0, pad = var_5737_pad_0, pad_type = var_5737_pad_type_0, strides = var_5737_strides_0, weight = squeeze_23_quantized, x = var_5729)[name = string("op_5737")];
             tensor<int32, [3]> var_5738 = const()[name = string("op_5738"), val = tensor<int32, [3]>([0, 2, 1])];
             fp16 const_330_promoted_to_fp16 = const()[name = string("const_330_promoted_to_fp16"), val = fp16(-0x1p+0)];
             tensor<fp16, [1, 128, 768]> x_377 = transpose(perm = var_5738, x = var_5737)[name = string("transpose_2")];
@@ -4633,7 +4638,7 @@ program(1.3)
             tensor<int32, [2]> var_5747_split_sizes_0 = const()[name = string("op_5747_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_5747_axis_0 = const()[name = string("op_5747_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_5747_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_5747_cast_fp16_1 = split(axis = var_5747_axis_0, split_sizes = var_5747_split_sizes_0, x = normed_659_cast_fp16)[name = string("op_5747_cast_fp16")];
-            tensor<fp16, [768]> var_5751_to_fp16 = const()[name = string("op_5751_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(615447488)))];
+            tensor<fp16, [768]> var_5751_to_fp16 = const()[name = string("op_5751_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308610176)))];
             tensor<fp16, [1, 128, 768]> out_283_cast_fp16 = mul(x = var_5747_cast_fp16_0, y = var_5751_to_fp16)[name = string("out_283_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_379_cast_fp16 = add(x = x_369_cast_fp16, y = out_283_cast_fp16)[name = string("x_379_cast_fp16")];
             fp16 const_332_promoted_to_fp16 = const()[name = string("const_332_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -4645,7 +4650,7 @@ program(1.3)
             tensor<int32, [2]> var_5763_split_sizes_0 = const()[name = string("op_5763_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_5763_axis_0 = const()[name = string("op_5763_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_5763_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_5763_cast_fp16_1 = split(axis = var_5763_axis_0, split_sizes = var_5763_split_sizes_0, x = normed_663_cast_fp16)[name = string("op_5763_cast_fp16")];
-            tensor<fp16, [768]> var_5767_to_fp16 = const()[name = string("op_5767_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(615449088)))];
+            tensor<fp16, [768]> var_5767_to_fp16 = const()[name = string("op_5767_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308611776)))];
             tensor<fp16, [1, 128, 768]> out_285_cast_fp16 = mul(x = var_5763_cast_fp16_0, y = var_5767_to_fp16)[name = string("out_285_cast_fp16")];
             tensor<int32, [3]> var_5774 = const()[name = string("op_5774"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, [1]> input_475_axes_0 = const()[name = string("input_475_axes_0"), val = tensor<int32, [1]>([2])];
@@ -4656,13 +4661,13 @@ program(1.3)
             tensor<int32, [4]> gate_93_pad_0 = const()[name = string("gate_93_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> gate_93_dilations_0 = const()[name = string("gate_93_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 gate_93_groups_0 = const()[name = string("gate_93_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> gate_93 = conv(dilations = gate_93_dilations_0, groups = gate_93_groups_0, pad = gate_93_pad_0, pad_type = gate_93_pad_type_0, strides = gate_93_strides_0, weight = encoder_layers_23_mlp_gate_proj_weight, x = input_475)[name = string("gate_93")];
+            tensor<fp16, [1, 1152, 1, 128]> gate_93 = conv(dilations = gate_93_dilations_0, groups = gate_93_groups_0, pad = gate_93_pad_0, pad_type = gate_93_pad_type_0, strides = gate_93_strides_0, weight = encoder_layers_23_mlp_gate_proj_weight_quantized, x = input_475)[name = string("gate_93")];
             string up_pad_type_0 = const()[name = string("up_pad_type_0"), val = string("valid")];
             tensor<int32, [2]> up_strides_0 = const()[name = string("up_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> up_pad_0 = const()[name = string("up_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> up_dilations_0 = const()[name = string("up_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 up_groups_0 = const()[name = string("up_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 1152, 1, 128]> up = conv(dilations = up_dilations_0, groups = up_groups_0, pad = up_pad_0, pad_type = up_pad_type_0, strides = up_strides_0, weight = encoder_layers_23_mlp_up_proj_weight, x = input_475)[name = string("up")];
+            tensor<fp16, [1, 1152, 1, 128]> up = conv(dilations = up_dilations_0, groups = up_groups_0, pad = up_pad_0, pad_type = up_pad_type_0, strides = up_strides_0, weight = encoder_layers_23_mlp_up_proj_weight_quantized, x = input_475)[name = string("up")];
             string gate_mode_0 = const()[name = string("gate_mode_0"), val = string("TANH_APPROXIMATION")];
             tensor<fp16, [1, 1152, 1, 128]> gate = gelu(mode = gate_mode_0, x = gate_93)[name = string("gate")];
             tensor<fp16, [1, 1152, 1, 128]> input_477 = mul(x = gate, y = up)[name = string("input_477")];
@@ -4671,7 +4676,7 @@ program(1.3)
             tensor<int32, [4]> var_5796_pad_0 = const()[name = string("op_5796_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> var_5796_dilations_0 = const()[name = string("op_5796_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 var_5796_groups_0 = const()[name = string("op_5796_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 128]> var_5796 = conv(dilations = var_5796_dilations_0, groups = var_5796_groups_0, pad = var_5796_pad_0, pad_type = var_5796_pad_type_0, strides = var_5796_strides_0, weight = encoder_layers_23_mlp_down_proj_weight, x = input_477)[name = string("op_5796")];
+            tensor<fp16, [1, 768, 1, 128]> var_5796 = conv(dilations = var_5796_dilations_0, groups = var_5796_groups_0, pad = var_5796_pad_0, pad_type = var_5796_pad_type_0, strides = var_5796_strides_0, weight = encoder_layers_23_mlp_down_proj_weight_quantized, x = input_477)[name = string("op_5796")];
             tensor<int32, [1]> var_5797_axes_0 = const()[name = string("op_5797_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 768, 128]> var_5797 = squeeze(axes = var_5797_axes_0, x = var_5796)[name = string("op_5797")];
             tensor<int32, [3]> var_5798 = const()[name = string("op_5798"), val = tensor<int32, [3]>([0, 2, 1])];
@@ -4685,7 +4690,7 @@ program(1.3)
             tensor<int32, [2]> var_5807_split_sizes_0 = const()[name = string("op_5807_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_5807_axis_0 = const()[name = string("op_5807_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_5807_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_5807_cast_fp16_1 = split(axis = var_5807_axis_0, split_sizes = var_5807_split_sizes_0, x = normed_669_cast_fp16)[name = string("op_5807_cast_fp16")];
-            tensor<fp16, [768]> var_5811_to_fp16 = const()[name = string("op_5811_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(615450688)))];
+            tensor<fp16, [768]> var_5811_to_fp16 = const()[name = string("op_5811_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308613376)))];
             tensor<fp16, [1, 128, 768]> out_287_cast_fp16 = mul(x = var_5807_cast_fp16_0, y = var_5811_to_fp16)[name = string("out_287_cast_fp16")];
             tensor<fp16, [1, 128, 768]> x_385_cast_fp16 = add(x = x_379_cast_fp16, y = out_287_cast_fp16)[name = string("x_385_cast_fp16")];
             fp16 const_336_promoted_to_fp16 = const()[name = string("const_336_promoted_to_fp16"), val = fp16(-0x1p+0)];
@@ -4697,7 +4702,7 @@ program(1.3)
             tensor<int32, [2]> var_5823_split_sizes_0 = const()[name = string("op_5823_split_sizes_0"), val = tensor<int32, [2]>([768, 768])];
             int32 var_5823_axis_0 = const()[name = string("op_5823_axis_0"), val = int32(-1)];
             tensor<fp16, [1, 128, 768]> var_5823_cast_fp16_0, tensor<fp16, [1, 128, 768]> var_5823_cast_fp16_1 = split(axis = var_5823_axis_0, split_sizes = var_5823_split_sizes_0, x = normed_673_cast_fp16)[name = string("op_5823_cast_fp16")];
-            tensor<fp16, [768]> var_5827_to_fp16 = const()[name = string("op_5827_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(615452288)))];
+            tensor<fp16, [768]> var_5827_to_fp16 = const()[name = string("op_5827_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308614976)))];
             tensor<fp16, [1, 128, 768]> out_cast_fp16 = mul(x = var_5823_cast_fp16_0, y = var_5827_to_fp16)[name = string("out_cast_fp16")];
             tensor<int32, [1]> mask_axes_0 = const()[name = string("mask_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1, 128, 1]> mask_cast_fp16 = expand_dims(axes = mask_axes_0, x = attention_mask)[name = string("mask_cast_fp16")];
@@ -4720,13 +4725,13 @@ program(1.3)
             tensor<int32, [4]> input_pad_0 = const()[name = string("input_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> input_dilations_0 = const()[name = string("input_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 input_groups_0 = const()[name = string("input_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 3072, 1, 1]> input = conv(bias = dense1_bias, dilations = input_dilations_0, groups = input_groups_0, pad = input_pad_0, pad_type = input_pad_type_0, strides = input_strides_0, weight = dense1_weight, x = input_483)[name = string("input")];
+            tensor<fp16, [1, 3072, 1, 1]> input = conv(bias = dense1_bias, dilations = input_dilations_0, groups = input_groups_0, pad = input_pad_0, pad_type = input_pad_type_0, strides = input_strides_0, weight = dense1_weight_quantized, x = input_483)[name = string("input")];
             string x_pad_type_0 = const()[name = string("x_pad_type_0"), val = string("valid")];
             tensor<int32, [2]> x_strides_0 = const()[name = string("x_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_pad_0 = const()[name = string("x_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_dilations_0 = const()[name = string("x_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             int32 x_groups_0 = const()[name = string("x_groups_0"), val = int32(1)];
-            tensor<fp16, [1, 768, 1, 1]> x = conv(bias = dense2_bias, dilations = x_dilations_0, groups = x_groups_0, pad = x_pad_0, pad_type = x_pad_type_0, strides = x_strides_0, weight = dense2_weight, x = input)[name = string("x")];
+            tensor<fp16, [1, 768, 1, 1]> x = conv(bias = dense2_bias, dilations = x_dilations_0, groups = x_groups_0, pad = x_pad_0, pad_type = x_pad_type_0, strides = x_strides_0, weight = dense2_weight_quantized, x = input)[name = string("x")];
             tensor<int32, [2]> var_5884 = const()[name = string("op_5884"), val = tensor<int32, [2]>([1, 768])];
             tensor<fp16, [1, 768]> pooled = reshape(shape = var_5884, x = x)[name = string("pooled")];
             fp16 const_338 = const()[name = string("const_338"), val = fp16(0x1.1p-20)];