|
| 1 | +#!/bin/bash |
| 2 | +# |
| 3 | +# Fix NCU Permissions - Enable NVIDIA GPU Performance Counter Access |
| 4 | +# |
| 5 | +# NCU requires special permissions to access GPU performance counters. |
| 6 | +# This script enables those permissions. |
| 7 | +# |
| 8 | + |
| 9 | +set -e |
| 10 | + |
| 11 | +echo "==========================================" |
| 12 | +echo "Fixing NCU Permissions" |
| 13 | +echo "==========================================" |
| 14 | +echo "" |
| 15 | + |
| 16 | +# Check if running as root |
| 17 | +if [ "$EUID" -eq 0 ]; then |
| 18 | + echo "✓ Running as root" |
| 19 | +else |
| 20 | + echo "⚠ Not running as root. You may need sudo for some operations." |
| 21 | +fi |
| 22 | + |
| 23 | +echo "" |
| 24 | +echo "Enabling GPU performance counter access..." |
| 25 | +echo "" |
| 26 | + |
| 27 | +# Method 1: Set profiling mode to unrestricted (temporary, lost on reboot) |
| 28 | +echo "Method 1: Temporary fix (until reboot)" |
| 29 | +echo "-----------------------------------------" |
| 30 | +if [ -f /proc/driver/nvidia/params ]; then |
| 31 | + echo "Setting NVreg_RestrictProfilingToAdminUsers=0..." |
| 32 | + if sudo sh -c 'echo "options nvidia NVreg_RestrictProfilingToAdminUsers=0" > /etc/modprobe.d/nvidia-profiling.conf'; then |
| 33 | + echo "✓ Modprobe config updated" |
| 34 | + echo "" |
| 35 | + echo "Reloading NVIDIA kernel module..." |
| 36 | + if sudo modprobe -r nvidia_uvm nvidia_drm nvidia_modeset nvidia && sudo modprobe nvidia; then |
| 37 | + echo "✓ NVIDIA module reloaded" |
| 38 | + else |
| 39 | + echo "⚠ Could not reload module. You may need to reboot." |
| 40 | + fi |
| 41 | + else |
| 42 | + echo "✗ Failed to update modprobe config" |
| 43 | + fi |
| 44 | +else |
| 45 | + echo "⚠ NVIDIA driver not found at /proc/driver/nvidia/params" |
| 46 | +fi |
| 47 | + |
| 48 | +echo "" |
| 49 | +echo "Method 2: Immediate fix (current session only)" |
| 50 | +echo "-----------------------------------------" |
| 51 | +if [ -f /sys/module/nvidia/parameters/NVreg_RestrictProfilingToAdminUsers ]; then |
| 52 | + echo "Current value:" |
| 53 | + cat /sys/module/nvidia/parameters/NVreg_RestrictProfilingToAdminUsers |
| 54 | + echo "" |
| 55 | + |
| 56 | + echo "Note: Cannot modify this sysfs parameter directly." |
| 57 | + echo "The modprobe configuration above will take effect after module reload or reboot." |
| 58 | +else |
| 59 | + echo "⚠ Parameter file not found" |
| 60 | +fi |
| 61 | + |
| 62 | +echo "" |
| 63 | +echo "Method 3: Using nvidia-modprobe (if available)" |
| 64 | +echo "-----------------------------------------" |
| 65 | +if command -v nvidia-modprobe &> /dev/null; then |
| 66 | + echo "Running nvidia-modprobe..." |
| 67 | + sudo nvidia-modprobe || true |
| 68 | + echo "✓ Done" |
| 69 | +else |
| 70 | + echo "⚠ nvidia-modprobe not found" |
| 71 | +fi |
| 72 | + |
| 73 | +echo "" |
| 74 | +echo "==========================================" |
| 75 | +echo "Verification" |
| 76 | +echo "==========================================" |
| 77 | +echo "" |
| 78 | + |
| 79 | +# Test NCU access |
| 80 | +if command -v ncu &> /dev/null; then |
| 81 | + echo "Testing NCU access with a simple command..." |
| 82 | + if ncu --query-metrics 2>&1 | grep -q "dram__bytes"; then |
| 83 | + echo "✓ NCU can access performance counters!" |
| 84 | + else |
| 85 | + echo "⚠ NCU may still have permission issues" |
| 86 | + echo "" |
| 87 | + echo "Output from ncu --query-metrics:" |
| 88 | + ncu --query-metrics 2>&1 | head -20 |
| 89 | + fi |
| 90 | +else |
| 91 | + echo "⚠ ncu command not found" |
| 92 | +fi |
| 93 | + |
| 94 | +echo "" |
| 95 | +echo "==========================================" |
| 96 | +echo "Next Steps" |
| 97 | +echo "==========================================" |
| 98 | +echo "" |
| 99 | +echo "1. If the temporary fix worked, you can now run NCU profiling:" |
| 100 | +echo " ./run_ncu_bandwidth_test.sh" |
| 101 | +echo "" |
| 102 | +echo "2. To make the fix permanent across reboots:" |
| 103 | +echo " - The modprobe config has been created at:" |
| 104 | +echo " /etc/modprobe.d/nvidia-profiling.conf" |
| 105 | +echo " - It will be loaded on next boot" |
| 106 | +echo "" |
| 107 | +echo "3. If you still see permission errors, you may need to:" |
| 108 | +echo " - Reboot the system for changes to take effect" |
| 109 | +echo " - OR run the profiling command with sudo:" |
| 110 | +echo " sudo ./run_ncu_bandwidth_test.sh" |
| 111 | +echo "" |
| 112 | +echo "4. Alternative: Run the microbench directly with sudo:" |
| 113 | +echo " sudo python3 tools/profiling/run_nwor_microbench.py \\" |
| 114 | +echo " --scenario short --requests 8 --batches 2 --draft-tokens 4 \\" |
| 115 | +echo " --temperature 0.7 --nwor-modes off --scv-modes off \\" |
| 116 | +echo " --enable-ncu --ncu-metrics \"dram__bytes_write.sum\" \\" |
| 117 | +echo " --output test_ncu.json" |
| 118 | +echo "" |
| 119 | + |
| 120 | +# Show current NVIDIA driver version |
| 121 | +echo "Current NVIDIA Driver Info:" |
| 122 | +echo "----------------------------" |
| 123 | +nvidia-smi --query-gpu=driver_version,name --format=csv,noheader 2>/dev/null || echo "nvidia-smi not available" |
| 124 | +echo "" |
| 125 | + |
| 126 | +echo "Done!" |
0 commit comments