diff --git a/python/bitblas/__init__.py b/bitblas/__init__.py
similarity index 97%
rename from python/bitblas/__init__.py
rename to bitblas/__init__.py
index 14b510845..172c4cbf1 100644
--- a/python/bitblas/__init__.py
+++ b/bitblas/__init__.py
@@ -11,7 +11,7 @@
     sys.path.insert(0, install_tvm_path)
 
 develop_tvm_path = os.path.join(
-    os.path.dirname(os.path.abspath(__file__)), "..", "..", "3rdparty", "tvm", "python")
+    os.path.dirname(os.path.abspath(__file__)), "..", "3rdparty", "tvm", "python")
 if os.path.exists(develop_tvm_path) and develop_tvm_path not in sys.path:
     os.environ["PYTHONPATH"] = develop_tvm_path + ":" + os.environ.get("PYTHONPATH", "")
     sys.path.insert(0, develop_tvm_path)
diff --git a/python/bitblas/base/__init__.py b/bitblas/base/__init__.py
similarity index 100%
rename from python/bitblas/base/__init__.py
rename to bitblas/base/__init__.py
diff --git a/python/bitblas/base/analysis.py b/bitblas/base/analysis.py
similarity index 100%
rename from python/bitblas/base/analysis.py
rename to bitblas/base/analysis.py
diff --git a/python/bitblas/base/common_schedules.py b/bitblas/base/common_schedules.py
similarity index 100%
rename from python/bitblas/base/common_schedules.py
rename to bitblas/base/common_schedules.py
diff --git a/python/bitblas/base/roller/__init__.py b/bitblas/base/roller/__init__.py
similarity index 100%
rename from python/bitblas/base/roller/__init__.py
rename to bitblas/base/roller/__init__.py
diff --git a/python/bitblas/base/roller/arch/__init__.py b/bitblas/base/roller/arch/__init__.py
similarity index 100%
rename from python/bitblas/base/roller/arch/__init__.py
rename to bitblas/base/roller/arch/__init__.py
diff --git a/python/bitblas/base/roller/arch/arch_base.py b/bitblas/base/roller/arch/arch_base.py
similarity index 100%
rename from python/bitblas/base/roller/arch/arch_base.py
rename to bitblas/base/roller/arch/arch_base.py
diff --git a/python/bitblas/base/roller/arch/cpu.py b/bitblas/base/roller/arch/cpu.py
similarity index 100%
rename from python/bitblas/base/roller/arch/cpu.py
rename to bitblas/base/roller/arch/cpu.py
diff --git a/python/bitblas/base/roller/arch/cuda.py b/bitblas/base/roller/arch/cuda.py
similarity index 100%
rename from python/bitblas/base/roller/arch/cuda.py
rename to bitblas/base/roller/arch/cuda.py
diff --git a/python/bitblas/base/roller/bestfit.py b/bitblas/base/roller/bestfit.py
similarity index 100%
rename from python/bitblas/base/roller/bestfit.py
rename to bitblas/base/roller/bestfit.py
diff --git a/python/bitblas/base/roller/hint.py b/bitblas/base/roller/hint.py
similarity index 100%
rename from python/bitblas/base/roller/hint.py
rename to bitblas/base/roller/hint.py
diff --git a/python/bitblas/base/roller/node.py b/bitblas/base/roller/node.py
similarity index 100%
rename from python/bitblas/base/roller/node.py
rename to bitblas/base/roller/node.py
diff --git a/python/bitblas/base/roller/policy/__init__.py b/bitblas/base/roller/policy/__init__.py
similarity index 100%
rename from python/bitblas/base/roller/policy/__init__.py
rename to bitblas/base/roller/policy/__init__.py
diff --git a/python/bitblas/base/roller/policy/common.py b/bitblas/base/roller/policy/common.py
similarity index 100%
rename from python/bitblas/base/roller/policy/common.py
rename to bitblas/base/roller/policy/common.py
diff --git a/python/bitblas/base/roller/policy/default.py b/bitblas/base/roller/policy/default.py
similarity index 100%
rename from python/bitblas/base/roller/policy/default.py
rename to bitblas/base/roller/policy/default.py
diff --git a/python/bitblas/base/roller/policy/tensorcore.py b/bitblas/base/roller/policy/tensorcore.py
similarity index 100%
rename from python/bitblas/base/roller/policy/tensorcore.py
rename to bitblas/base/roller/policy/tensorcore.py
diff --git a/python/bitblas/base/roller/rasterization.py b/bitblas/base/roller/rasterization.py
similarity index 100%
rename from python/bitblas/base/roller/rasterization.py
rename to bitblas/base/roller/rasterization.py
diff --git a/python/bitblas/base/roller/shape_inference/__init__.py b/bitblas/base/roller/shape_inference/__init__.py
similarity index 100%
rename from python/bitblas/base/roller/shape_inference/__init__.py
rename to bitblas/base/roller/shape_inference/__init__.py
diff --git a/python/bitblas/base/roller/shape_inference/common.py b/bitblas/base/roller/shape_inference/common.py
similarity index 100%
rename from python/bitblas/base/roller/shape_inference/common.py
rename to bitblas/base/roller/shape_inference/common.py
diff --git a/python/bitblas/base/roller/shape_inference/tir.py b/bitblas/base/roller/shape_inference/tir.py
similarity index 100%
rename from python/bitblas/base/roller/shape_inference/tir.py
rename to bitblas/base/roller/shape_inference/tir.py
diff --git a/python/bitblas/base/schedule_rule.py b/bitblas/base/schedule_rule.py
similarity index 100%
rename from python/bitblas/base/schedule_rule.py
rename to bitblas/base/schedule_rule.py
diff --git a/python/bitblas/base/transform.py b/bitblas/base/transform.py
similarity index 100%
rename from python/bitblas/base/transform.py
rename to bitblas/base/transform.py
diff --git a/python/bitblas/base/utils.py b/bitblas/base/utils.py
similarity index 100%
rename from python/bitblas/base/utils.py
rename to bitblas/base/utils.py
diff --git a/python/bitblas/cache/__init__.py b/bitblas/cache/__init__.py
similarity index 100%
rename from python/bitblas/cache/__init__.py
rename to bitblas/cache/__init__.py
diff --git a/python/bitblas/cache/operator.py b/bitblas/cache/operator.py
similarity index 100%
rename from python/bitblas/cache/operator.py
rename to bitblas/cache/operator.py
diff --git a/python/bitblas/generator.py b/bitblas/generator.py
similarity index 100%
rename from python/bitblas/generator.py
rename to bitblas/generator.py
diff --git a/python/bitblas/gpu/__init__.py b/bitblas/gpu/__init__.py
similarity index 100%
rename from python/bitblas/gpu/__init__.py
rename to bitblas/gpu/__init__.py
diff --git a/python/bitblas/gpu/base.py b/bitblas/gpu/base.py
similarity index 100%
rename from python/bitblas/gpu/base.py
rename to bitblas/gpu/base.py
diff --git a/python/bitblas/gpu/element_wise.py b/bitblas/gpu/element_wise.py
similarity index 100%
rename from python/bitblas/gpu/element_wise.py
rename to bitblas/gpu/element_wise.py
diff --git a/python/bitblas/gpu/fallback.py b/bitblas/gpu/fallback.py
similarity index 100%
rename from python/bitblas/gpu/fallback.py
rename to bitblas/gpu/fallback.py
diff --git a/python/bitblas/gpu/gemv.py b/bitblas/gpu/gemv.py
similarity index 100%
rename from python/bitblas/gpu/gemv.py
rename to bitblas/gpu/gemv.py
diff --git a/python/bitblas/gpu/gemv_dequantize.py b/bitblas/gpu/gemv_dequantize.py
similarity index 100%
rename from python/bitblas/gpu/gemv_dequantize.py
rename to bitblas/gpu/gemv_dequantize.py
diff --git a/python/bitblas/gpu/general_reduction.py b/bitblas/gpu/general_reduction.py
similarity index 100%
rename from python/bitblas/gpu/general_reduction.py
rename to bitblas/gpu/general_reduction.py
diff --git a/python/bitblas/gpu/intrin/__init__.py b/bitblas/gpu/intrin/__init__.py
similarity index 100%
rename from python/bitblas/gpu/intrin/__init__.py
rename to bitblas/gpu/intrin/__init__.py
diff --git a/python/bitblas/gpu/intrin/lop3.py b/bitblas/gpu/intrin/lop3.py
similarity index 100%
rename from python/bitblas/gpu/intrin/lop3.py
rename to bitblas/gpu/intrin/lop3.py
diff --git a/python/bitblas/gpu/matmul.py b/bitblas/gpu/matmul.py
similarity index 100%
rename from python/bitblas/gpu/matmul.py
rename to bitblas/gpu/matmul.py
diff --git a/python/bitblas/gpu/matmul_analysis.py b/bitblas/gpu/matmul_analysis.py
similarity index 100%
rename from python/bitblas/gpu/matmul_analysis.py
rename to bitblas/gpu/matmul_analysis.py
diff --git a/python/bitblas/gpu/matmul_mma.py b/bitblas/gpu/matmul_mma.py
similarity index 100%
rename from python/bitblas/gpu/matmul_mma.py
rename to bitblas/gpu/matmul_mma.py
diff --git a/python/bitblas/gpu/matmul_mma_dequantize.py b/bitblas/gpu/matmul_mma_dequantize.py
similarity index 100%
rename from python/bitblas/gpu/matmul_mma_dequantize.py
rename to bitblas/gpu/matmul_mma_dequantize.py
diff --git a/python/bitblas/gpu/matmul_wmma.py b/bitblas/gpu/matmul_wmma.py
similarity index 100%
rename from python/bitblas/gpu/matmul_wmma.py
rename to bitblas/gpu/matmul_wmma.py
diff --git a/python/bitblas/gpu/reduction.py b/bitblas/gpu/reduction.py
similarity index 100%
rename from python/bitblas/gpu/reduction.py
rename to bitblas/gpu/reduction.py
diff --git a/python/bitblas/gpu/rmsnorm.py b/bitblas/gpu/rmsnorm.py
similarity index 100%
rename from python/bitblas/gpu/rmsnorm.py
rename to bitblas/gpu/rmsnorm.py
diff --git a/python/bitblas/gpu/transpose.py b/bitblas/gpu/transpose.py
similarity index 100%
rename from python/bitblas/gpu/transpose.py
rename to bitblas/gpu/transpose.py
diff --git a/python/bitblas/gpu/utils.py b/bitblas/gpu/utils.py
similarity index 100%
rename from python/bitblas/gpu/utils.py
rename to bitblas/gpu/utils.py
diff --git a/python/bitblas/module/__init__.py b/bitblas/module/__init__.py
similarity index 100%
rename from python/bitblas/module/__init__.py
rename to bitblas/module/__init__.py
diff --git a/python/bitblas/ops/__init__.py b/bitblas/ops/__init__.py
similarity index 100%
rename from python/bitblas/ops/__init__.py
rename to bitblas/ops/__init__.py
diff --git a/python/bitblas/ops/general_matmul.py b/bitblas/ops/general_matmul.py
similarity index 100%
rename from python/bitblas/ops/general_matmul.py
rename to bitblas/ops/general_matmul.py
diff --git a/python/bitblas/ops/general_matmul_splitk.py b/bitblas/ops/general_matmul_splitk.py
similarity index 100%
rename from python/bitblas/ops/general_matmul_splitk.py
rename to bitblas/ops/general_matmul_splitk.py
diff --git a/python/bitblas/ops/impl/__init__.py b/bitblas/ops/impl/__init__.py
similarity index 100%
rename from python/bitblas/ops/impl/__init__.py
rename to bitblas/ops/impl/__init__.py
diff --git a/python/bitblas/ops/impl/batch_matmul_dequantize_impl.py b/bitblas/ops/impl/batch_matmul_dequantize_impl.py
similarity index 100%
rename from python/bitblas/ops/impl/batch_matmul_dequantize_impl.py
rename to bitblas/ops/impl/batch_matmul_dequantize_impl.py
diff --git a/python/bitblas/ops/impl/batch_matmul_impl.py b/bitblas/ops/impl/batch_matmul_impl.py
similarity index 100%
rename from python/bitblas/ops/impl/batch_matmul_impl.py
rename to bitblas/ops/impl/batch_matmul_impl.py
diff --git a/python/bitblas/ops/impl/convolution2d_impl.py b/bitblas/ops/impl/convolution2d_impl.py
similarity index 100%
rename from python/bitblas/ops/impl/convolution2d_impl.py
rename to bitblas/ops/impl/convolution2d_impl.py
diff --git a/python/bitblas/ops/impl/ladder_permutate_impl.py b/bitblas/ops/impl/ladder_permutate_impl.py
similarity index 100%
rename from python/bitblas/ops/impl/ladder_permutate_impl.py
rename to bitblas/ops/impl/ladder_permutate_impl.py
diff --git a/python/bitblas/ops/impl/lop3_permutate_impl.py b/bitblas/ops/impl/lop3_permutate_impl.py
similarity index 100%
rename from python/bitblas/ops/impl/lop3_permutate_impl.py
rename to bitblas/ops/impl/lop3_permutate_impl.py
diff --git a/python/bitblas/ops/impl/matmul_dequantize_impl.py b/bitblas/ops/impl/matmul_dequantize_impl.py
similarity index 100%
rename from python/bitblas/ops/impl/matmul_dequantize_impl.py
rename to bitblas/ops/impl/matmul_dequantize_impl.py
diff --git a/python/bitblas/ops/impl/matmul_dequantize_splitk_impl.py b/bitblas/ops/impl/matmul_dequantize_splitk_impl.py
similarity index 100%
rename from python/bitblas/ops/impl/matmul_dequantize_splitk_impl.py
rename to bitblas/ops/impl/matmul_dequantize_splitk_impl.py
diff --git a/python/bitblas/ops/impl/matmul_impl.py b/bitblas/ops/impl/matmul_impl.py
similarity index 100%
rename from python/bitblas/ops/impl/matmul_impl.py
rename to bitblas/ops/impl/matmul_impl.py
diff --git a/python/bitblas/ops/impl/matmul_splitk_impl.py b/bitblas/ops/impl/matmul_splitk_impl.py
similarity index 100%
rename from python/bitblas/ops/impl/matmul_splitk_impl.py
rename to bitblas/ops/impl/matmul_splitk_impl.py
diff --git a/python/bitblas/ops/impl/param_permutate_impl.py b/bitblas/ops/impl/param_permutate_impl.py
similarity index 100%
rename from python/bitblas/ops/impl/param_permutate_impl.py
rename to bitblas/ops/impl/param_permutate_impl.py
diff --git a/python/bitblas/ops/ladder_permutate.py b/bitblas/ops/ladder_permutate.py
similarity index 100%
rename from python/bitblas/ops/ladder_permutate.py
rename to bitblas/ops/ladder_permutate.py
diff --git a/python/bitblas/ops/lop3_permutate.py b/bitblas/ops/lop3_permutate.py
similarity index 100%
rename from python/bitblas/ops/lop3_permutate.py
rename to bitblas/ops/lop3_permutate.py
diff --git a/python/bitblas/ops/matmul.py b/bitblas/ops/matmul.py
similarity index 100%
rename from python/bitblas/ops/matmul.py
rename to bitblas/ops/matmul.py
diff --git a/python/bitblas/ops/matmul_dequantize.py b/bitblas/ops/matmul_dequantize.py
similarity index 100%
rename from python/bitblas/ops/matmul_dequantize.py
rename to bitblas/ops/matmul_dequantize.py
diff --git a/python/bitblas/ops/operator.py b/bitblas/ops/operator.py
similarity index 100%
rename from python/bitblas/ops/operator.py
rename to bitblas/ops/operator.py
diff --git a/python/bitblas/ops/param_permutate.py b/bitblas/ops/param_permutate.py
similarity index 100%
rename from python/bitblas/ops/param_permutate.py
rename to bitblas/ops/param_permutate.py
diff --git a/python/bitblas/quantization/__init__.py b/bitblas/quantization/__init__.py
similarity index 100%
rename from python/bitblas/quantization/__init__.py
rename to bitblas/quantization/__init__.py
diff --git a/python/bitblas/quantization/quantization.py b/bitblas/quantization/quantization.py
similarity index 100%
rename from python/bitblas/quantization/quantization.py
rename to bitblas/quantization/quantization.py
diff --git a/python/bitblas/quantization/utils.py b/bitblas/quantization/utils.py
similarity index 100%
rename from python/bitblas/quantization/utils.py
rename to bitblas/quantization/utils.py
diff --git a/python/bitblas/relax/op/interleave_weight.py b/bitblas/relax/op/interleave_weight.py
similarity index 100%
rename from python/bitblas/relax/op/interleave_weight.py
rename to bitblas/relax/op/interleave_weight.py
diff --git a/python/bitblas/relax/transform/__init__.py b/bitblas/relax/transform/__init__.py
similarity index 100%
rename from python/bitblas/relax/transform/__init__.py
rename to bitblas/relax/transform/__init__.py
diff --git a/python/bitblas/relax/transform/annotate_decode_block.py b/bitblas/relax/transform/annotate_decode_block.py
similarity index 100%
rename from python/bitblas/relax/transform/annotate_decode_block.py
rename to bitblas/relax/transform/annotate_decode_block.py
diff --git a/python/bitblas/relax/transform/weight_only_propagate.py b/bitblas/relax/transform/weight_only_propagate.py
similarity index 100%
rename from python/bitblas/relax/transform/weight_only_propagate.py
rename to bitblas/relax/transform/weight_only_propagate.py
diff --git a/python/bitblas/testing/__init__.py b/bitblas/testing/__init__.py
similarity index 100%
rename from python/bitblas/testing/__init__.py
rename to bitblas/testing/__init__.py
diff --git a/python/bitblas/utils/__init__.py b/bitblas/utils/__init__.py
similarity index 100%
rename from python/bitblas/utils/__init__.py
rename to bitblas/utils/__init__.py
diff --git a/python/bitblas/utils/post_process.py b/bitblas/utils/post_process.py
similarity index 100%
rename from python/bitblas/utils/post_process.py
rename to bitblas/utils/post_process.py
diff --git a/python/bitblas/utils/target_detector.py b/bitblas/utils/target_detector.py
similarity index 100%
rename from python/bitblas/utils/target_detector.py
rename to bitblas/utils/target_detector.py
diff --git a/python/bitblas/utils/tensor_adapter.py b/bitblas/utils/tensor_adapter.py
similarity index 100%
rename from python/bitblas/utils/tensor_adapter.py
rename to bitblas/utils/tensor_adapter.py
diff --git a/python/bitblas/wrapper/__init__.py b/bitblas/wrapper/__init__.py
similarity index 100%
rename from python/bitblas/wrapper/__init__.py
rename to bitblas/wrapper/__init__.py
diff --git a/python/bitblas/wrapper/general.py b/bitblas/wrapper/general.py
similarity index 100%
rename from python/bitblas/wrapper/general.py
rename to bitblas/wrapper/general.py
diff --git a/install.sh b/install.sh
new file mode 100755
index 000000000..584392820
--- /dev/null
+++ b/install.sh
@@ -0,0 +1,26 @@
+#!/bin/bash
+
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT License.
+
+# install requirements
+pip install -r requirements.txt
+
+# install llvm
+apt-get install llvm-10
+
+# clone and build tvm
+git submodule update --init --recursive
+
+cd 3rdparty/tvm
+mkdir build
+cp cmake/config.cmake build
+cd build
+echo "set(USE_LLVM llvm-config-10)" >> config.cmake && echo "set(USE_CUDA ON)" >> config.cmake
+
+cmake .. && make -j && cd ../../..
+
+echo "export TVM_HOME=$(pwd)/3rdparty/tvm" >> ~/.bashrc
+echo "export PYTHONPATH=\$TVM_HOME/python:$(pwd)/python:\$PYTHONPATH" >> ~/.bashrc
+
+source ~/.bashrc
diff --git a/python/bitblas_cli.py b/python/bitblas_cli.py
deleted file mode 100644
index 59e481eb9..000000000
--- a/python/bitblas_cli.py
+++ /dev/null
@@ -1,2 +0,0 @@
-# Copyright (c) Microsoft Corporation.
-# Licensed under the MIT License.
diff --git a/testing/python/dsl/test_auto_normalized_tensorcore.py b/testing/python/dsl/test_auto_normalized_tensorcore.py
deleted file mode 100644
index eb6e0baef..000000000
--- a/testing/python/dsl/test_auto_normalized_tensorcore.py
+++ /dev/null
@@ -1,112 +0,0 @@
-# Copyright (c) Microsoft Corporation.
-# Licensed under the MIT License.
-import numpy as np
-import tvm
-from bitblas.base.roller.policy import TensorCorePolicy, DefaultPolicy
-from bitblas.base.roller.arch import CUDA
-from bitblas.gpu.matmul_analysis import get_tensorized_func_and_tags
-from bitblas.gpu import Matmul
-from bitblas.ops.impl.convolution2d_impl import conv2d_nhwc_hwio, conv2d_nhwc_ohwi
-from bitblas.base.utils import apply_and_build
-import time
-
-benchmark_sets = [
-    # (prim_func, input_args, default_bitblas_schedule),
-    (conv2d_nhwc_hwio, (128, 64, 224, 224, 3, 7, 7, 2, 1, 3, "float16", "float16"), Matmul),
-    (conv2d_nhwc_ohwi, (128, 64, 56, 56, 64, 3, 3, 1, 1, 1, "float16", "float16"), Matmul),
-    (conv2d_nhwc_hwio, (128, 64, 56, 56, 64, 1, 1, 1, 1, 1, "float16", "float16"), Matmul),
-    (conv2d_nhwc_ohwi, (128, 64, 56, 56, 64, 1, 1, 1, 1, 1, "float16", "float16"), Matmul),
-    (conv2d_nhwc_ohwi, (128, 128, 28, 28, 128, 3, 3, 1, 1, 1, "float16", "float16"), Matmul),
-    (conv2d_nhwc_hwio, (128, 256, 14, 14, 128, 3, 3, 2, 1, 1, "float16", "float16"), Matmul),
-    (conv2d_nhwc_ohwi, (128, 256, 14, 14, 128, 1, 1, 2, 1, 1, "float16", "float16"), Matmul),
-]
-benchmark_results = {}
-for get_prim_func, input_args, d_schedule in benchmark_sets:
-    ir_module = get_prim_func(*input_args)
-    func = ir_module["main"]
-    target = tvm.target.Target("nvidia/nvidia-a100")
-    arch = CUDA(target)
-    policy = DefaultPolicy(func=func, arch=arch)
-    tensorized_func, tags = get_tensorized_func_and_tags(func, arch.target)
-    try:
-        tensorized_func, tags = get_tensorized_func_and_tags(func, arch.target)
-    except Exception as e:
-        print(f"Failed to get tensorized function and tags: {e}")
-        tags = None
-    if tags:
-        policy = TensorCorePolicy(func=tensorized_func, arch=arch, tags=tags)
-
-    configs = policy.emit_config(20)
-
-    tune_start = time.time()
-    cpresults, best = apply_and_build(func, configs, arch, parallel_build=True)
-    fast_tune_time = time.time() - tune_start
-    print("[BitBLAS] The best latency of top 1 is {:.3f} ms".format(cpresults[0].latency * 1e3))
-    print("[BitBLAS] The best latency of top 20 is {:.3f} ms".format(best.latency * 1e3))
-
-    # evaluate the performance of the default schedule
-
-    rule = d_schedule()
-    default_tune_start = time.time()
-    sch_default = rule.apply(func, target, False)
-    with tvm.transform.PassContext(config={"tir.use_async_copy": True}):
-        mod_default = tvm.build(sch_default.mod["main"], target="cuda")
-    default_tune_time = time.time() - default_tune_start
-
-    args = func.buffer_map.values()
-
-    profile_tensors = []
-    for arg in args:
-        profile_tensors.append(
-            tvm.nd.array(
-                np.random.uniform(0, 1, [int(i) for i in arg.shape]).astype(arg.dtype),
-                device=arch.device,
-            ))
-
-    timer_cuda_mod = mod_default.time_evaluator(mod_default.entry_name, arch.device, number=5)
-    t = timer_cuda_mod(*profile_tensors).mean
-
-    print("Time cost of BitBLAS default schedule: {:.3f} ms".format(t * 1e3))
-
-    profile_config = {
-        f"{get_prim_func.__name__}-{'-'.join([str(i) for i in input_args])}": {
-            "fast_bitblas_top20_tune_time": fast_tune_time,
-            "fast_bitblas_top1_latency": cpresults[0].latency * 1e3,
-            "fast_bitblas_top20_latency": best.latency * 1e3,
-            "default_bitblas_tune_time": default_tune_time,
-            "default_bitblas_latency": t * 1e3,
-        }
-    }
-    benchmark_results.update(profile_config)
-
-headers = [
-    "PrimFunc",
-    "Input Arguments",
-    "FastDLight Top20 Tune Time",
-    "FastDLight Top1 Latency",
-    "FastDLight Top20 Latency",
-    "DefaultDLight Tune Time",
-    "DefaultDLight Latency",
-]
-
-col_width = (max(len(word) for row in [headers] + list(profile_config.values()) for word in row) + 2
-            )  # padding
-
-print("".join(word.ljust(col_width) for word in headers))
-
-print("-" * col_width * len(headers))
-
-for config, values in benchmark_results.items():
-    args = config.split("-")
-    func_name = args[0]
-    input_args = "-".join(args[1:])
-    row = [
-        func_name,
-        input_args,
-        f" {str(values['fast_bitblas_top20_tune_time'])} s",
-        f"{values['fast_bitblas_top1_latency']:.3f} ms",
-        f"{values['fast_bitblas_top20_latency']:.3f} ms",
-        str(values["default_bitblas_tune_time"]),
-        f"{values['default_bitblas_latency']:.3f} ms",
-    ]
-    print("".join(word.ljust(col_width) for word in row))