Use libnuma instead of hwloc

It's way, way simpler.

README.md

@@ -5,7 +5,7 @@ Usermode Driver for Tenstorrent AI Accelerators
 ## Dependencies
 Required Ubuntu dependencies:
 ```
-sudo apt install -y libhwloc-dev cmake ninja-build
+sudo apt install -y cmake ninja-build libnuma-dev
 ```
 
 Suggested third-party dependency is Clang 17:
@@ -17,7 +17,7 @@ sudo ./llvm.sh 17
 
 ## Build flow
 
-To build `libdevice.so`: 
+To build `libdevice.so`:
 ```
 cmake -B build -G Ninja
 cmake --build build
@@ -70,7 +70,7 @@ add_subdirectory(<path to umd>)
 
 ### Ubuntu
 ```
-apt install ./umd-dev-x.y.z-Linux.deb 
+apt install ./umd-dev-x.y.z-Linux.deb
 ```
 
 ## Simulator Integration
@@ -100,7 +100,7 @@ To set up pre-commit on your local machine, follow these steps:
    Ensure you have Python installed, then run:
    ```bash
    pip install pre-commit
-   ```  
+   ```
 2. **Install the Git Hook Scripts**:
    In your local repository, run the following command to install the pre-commit hooks:
    ```bash
@@ -113,10 +113,10 @@ To set up pre-commit on your local machine, follow these steps:
    pre-commit run --all-files
    ```
 ## Why You Should Use Pre-commit
-By setting up pre-commit locally, you can help maintain the quality of the codebase and ensure that commits consistently meet the project's formatting standards. This saves time during code reviews and reduces the likelihood of code formatting issues slipping into the repository.  
-  
-Since the hooks run automatically before each commit, you don't need to remember to manually format or check your code, making it easier to maintain consistency.  
-  
+By setting up pre-commit locally, you can help maintain the quality of the codebase and ensure that commits consistently meet the project's formatting standards. This saves time during code reviews and reduces the likelihood of code formatting issues slipping into the repository.
+
+Since the hooks run automatically before each commit, you don't need to remember to manually format or check your code, making it easier to maintain consistency.
+
 We strongly encourage all developers to integrate pre-commit into their workflow.
 
 # Formatting C++ code

device/CMakeLists.txt

@@ -45,7 +45,7 @@ target_link_libraries(
     PUBLIC
         flatbuffers
     PRIVATE
-        hwloc
+        numa
         nng
         rt
         uv_a

device/pcie/pci_device.cpp

@@ -6,9 +6,9 @@
 
 #include "pci_device.hpp"
 
-#include <fcntl.h>  // for ::open
-#include <hwloc.h>
+#include <fcntl.h>      // for ::open
 #include <linux/pci.h>  // for PCI_SLOT, PCI_FUNC
+#include <numa.h>
 #include <sys/ioctl.h>  // for ioctl
 #include <sys/mman.h>   // for mmap, munmap
 #include <sys/stat.h>   // for fstat
@@ -46,53 +46,6 @@ const uint32_t HUGEPAGE_REGION_SIZE = 1 << 30;  // 1GB
 using namespace tt;
 using namespace tt::umd;
 
-/**
- * Configures the NUMA memory binding for the current process.
- *
- * Sets the memory allocation policy to bind memory allocations to the specified
- * NUMA node for the lifetime of the process.
- *
- * @param numa_node The NUMA node to which memory allocations should be bound
- */
-static void configure_numa_binding(int numa_node) {
-    if (numa_node < 0) {
-        // Don't bother if we're not on a NUMA system.
-        return;
-    }
-
-    hwloc_topology_t topology;
-
-    if (hwloc_topology_init(&topology) < 0) {
-        log_warning(LogSiliconDriver, "hwloc_topology_init failed; will not bind to NUMA node.");
-        return;
-    }
-
-    if (hwloc_topology_load(topology) < 0) {
-        log_warning(LogSiliconDriver, "hwloc_topology_load failed; will not bind to NUMA node.");
-        hwloc_topology_destroy(topology);
-        return;
-    }
-
-    hwloc_nodeset_t nodeset = hwloc_bitmap_alloc();
-    if (!nodeset) {
-        log_warning(LogSiliconDriver, "hwloc_bitmap_alloc failed; will not bind to NUMA node.");
-        hwloc_topology_destroy(topology);
-        return;
-    }
-
-    hwloc_bitmap_only(nodeset, numa_node);
-    if (hwloc_set_membind(topology, nodeset, HWLOC_MEMBIND_BIND, HWLOC_MEMBIND_THREAD) < 0) {
-        log_warning(LogSiliconDriver, "hwloc_set_membind failed: {}", strerror(errno));
-        log_warning(LogSiliconDriver, "Your hugepage memory may not be on the same NUMA node as the device.");
-        hwloc_bitmap_free(nodeset);
-        hwloc_topology_destroy(topology);
-        return;
-    }
-
-    hwloc_bitmap_free(nodeset);
-    hwloc_topology_destroy(topology);
-}
-
 template <typename T>
 static T read_sysfs(const PciDeviceInfo &device_info, const std::string &attribute_name) {
     const auto sysfs_path = fmt::format(
@@ -735,7 +688,12 @@ tt::umd::architecture_implementation *PCIDevice::get_architecture_implementation
 bool PCIDevice::init_hugepage(uint32_t num_host_mem_channels) {
     const size_t hugepage_size = HUGEPAGE_REGION_SIZE;
 
-    configure_numa_binding(this->numa_node);
+    if (numa_node > numa_max_node()) {
+        log_warning(LogSiliconDriver, "numa_node: {} is greater than numa_max_node: {}.", numa_node, numa_max_node());
+    }
+
+    // Prefer allocations on the NUMA node associated with the device.
+    numa_set_preferred(numa_node);
 
     // Convert from logical (device_id in netlist) to physical device_id (in case of virtualization)
     auto physical_device_id = get_device_num();