diff --git a/python/CMakeLists.txt b/python/CMakeLists.txt
index 7f8548c..6d37350 100644
--- a/python/CMakeLists.txt
+++ b/python/CMakeLists.txt
@@ -19,7 +19,9 @@ include(${_cpm_download_location})
 CPMAddPackage(
    NAME nvbench
    SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/..
-   OPTIONS "NVBench_INSTALL_RULES ON"
+   OPTIONS
+     "NVBench_INSTALL_RULES ON"
+     "NVBench_ENABLE_CUPTI OFF"
    FIND_PACKAGE_ARGS CONFIG REQUIRED
 )
 
diff --git a/python/cuda/bench/__init__.py b/python/cuda/bench/__init__.py
index 1913376..e6e7753 100644
--- a/python/cuda/bench/__init__.py
+++ b/python/cuda/bench/__init__.py
@@ -18,10 +18,6 @@ import importlib
 import importlib.metadata
 import warnings
 
-from cuda.pathfinder import (  # type: ignore[import-not-found]
-    load_nvidia_dynamic_lib,
-)
-
 try:
     __version__ = importlib.metadata.version("cuda-bench")
 except Exception as e:
@@ -65,10 +61,6 @@ except ImportError as e:
         f"Original error: {e}"
     )
 
-# Load required NVIDIA libraries
-for libname in ("cupti", "nvperf_target", "nvperf_host"):
-    load_nvidia_dynamic_lib(libname)
-
 # Import and expose all public symbols from the CUDA-specific extension
 Benchmark = _nvbench_module.Benchmark
 CudaStream = _nvbench_module.CudaStream
@@ -85,7 +77,6 @@ _nvbench = _nvbench_module
 
 # Clean up internal symbols
 del (
-    load_nvidia_dynamic_lib,
     _nvbench_module,
     _cuda_major,
     _extra_name,
diff --git a/python/cuda/bench/__init__.pyi b/python/cuda/bench/__init__.pyi
index 63b5190..6b0076f 100644
--- a/python/cuda/bench/__init__.pyi
+++ b/python/cuda/bench/__init__.pyi
@@ -96,8 +96,6 @@ class State:
     def set_timeout(self, duration: SupportsFloat) -> None: ...
     def get_blocking_kernel_timeout(self) -> float: ...
     def set_blocking_kernel_timeout(self, duration: SupportsFloat) -> None: ...
-    def collect_cupti_metrics(self) -> None: ...
-    def is_cupti_required(self) -> bool: ...
     def exec(
         self,
         fn: Callable[[Launch], None],
diff --git a/python/examples/auto_throughput.py b/python/examples/auto_throughput.py
index db4fa19..5d41b09 100644
--- a/python/examples/auto_throughput.py
+++ b/python/examples/auto_throughput.py
@@ -51,7 +51,6 @@ def throughput_bench(state: bench.State) -> None:
     out_arr = cuda.device_array(elements * ipt, dtype=np.int32, stream=alloc_stream)
 
     state.add_element_count(elements, column_name="Elements")
-    state.collect_cupti_metrics()
 
     threads_per_block = 256
     blocks_in_grid = (elements + threads_per_block - 1) // threads_per_block
diff --git a/python/pyproject.toml b/python/pyproject.toml
index 4d288ac..f7ddf78 100644
--- a/python/pyproject.toml
+++ b/python/pyproject.toml
@@ -21,19 +21,15 @@ classifiers = [
   "Operating System :: POSIX :: Linux",
 ]
 requires-python = ">=3.10"
-dependencies = [
-  # pathfinder for finding CUDA libraries
-  "cuda-pathfinder",
-]
 dynamic = ["version"]
 readme = { file = "README.md", content-type = "text/markdown" }
 
 [project.optional-dependencies]
 # CUDA 12.x dependencies
-cu12 = ["cuda-bindings>=12.0.0,<13.0.0", "nvidia-cuda-cupti-cu12"]
+cu12 = ["cuda-bindings>=12.0.0,<13.0.0"]
 
 # CUDA 13.x dependencies
-cu13 = ["cuda-bindings>=13.0.0,<14.0.0", "nvidia-cuda-cupti>=13.0"]
+cu13 = ["cuda-bindings>=13.0.0,<14.0.0"]
 
 # Test dependencies for CUDA 12
 test-cu12 = ["cuda-bench[cu12]", "pytest", "cupy-cuda12x", "numba"]
diff --git a/python/src/py_nvbench.cpp b/python/src/py_nvbench.cpp
index 97e4011..aeadcc2 100644
--- a/python/src/py_nvbench.cpp
+++ b/python/src/py_nvbench.cpp
@@ -676,13 +676,11 @@ void def_class_State(py::module_ m)
   //        nvbench::state::collect_stores_efficiency
   //        nvbench::state::collect_loads_efficiency
   //        nvbench::state::collect_dram_throughput
-  //        nvbench::state::collect_cupti_metrics
   //        nvbench::state::is_l1_hit_rate_collected
   //        nvbench::state::is_l2_hit_rate_collected
   //        nvbench::state::is_stores_efficiency_collected
   //        nvbench::state::is_loads_efficiency_collected
   //        nvbench::state::is_dram_throughput_collected
-  //        nvbench::state::is_cupti_required
   //        nvbench::state::add_summary
   //        nvbench::state::get_summary
   //        nvbench::state::get_summaries
@@ -972,20 +970,6 @@ Use argument True to disable use of blocking kernel by NVBench"
                   method_set_blocking_kernel_timeout_doc,
                   py::arg("duration_seconds"));
 
-  // method State.collect_cupti_metrics
-  static constexpr const char *method_collect_cupti_metrics_doc =
-    R"XXXX(Request NVBench to record CUPTI metrics while running benchmark for this configuration)XXXX";
-  pystate_cls.def("collect_cupti_metrics",
-                  &nvbench::state::collect_cupti_metrics,
-                  method_collect_cupti_metrics_doc);
-
-  // method State.is_cupti_required
-  static constexpr const char *method_is_cupti_required_doc =
-    R"XXXX(True if (some) CUPTI metrics are being collected)XXXX";
-  pystate_cls.def("is_cupti_required",
-                  &nvbench::state::is_cupti_required,
-                  method_is_cupti_required_doc);
-
   // method State.exec
   auto method_exec_impl =
     [](nvbench::state &state, py::object py_launcher_fn, bool batched, bool sync) -> void {