light-heart-labs--dreamserver
wehub-resource-sync
9e8f1bbeed
Dashboard / frontend (push) Failing after 0s
Dashboard / api (push) Failing after 0s
Lint PowerShell / powershell-lint (ubuntu-latest) (push) Failing after 1s
Python Lint / Lint Python with Ruff (push) Failing after 1s
ShellCheck / Lint shell scripts (push) Failing after 1s
Matrix Smoke / linux-smoke (push) Failing after 1s
Matrix Smoke / distro: cachyos (push) Failing after 15s
Matrix Smoke / distro: linux-mint-21.3 (push) Failing after 15s
Matrix Smoke / distro: debian-12 (push) Failing after 5m21s
Matrix Smoke / distro: fedora-41 (push) Failing after 4m56s
Matrix Smoke / distro: ubuntu-24.04 (push) Failing after 2m13s
Matrix Smoke / distro: rocky-9 (push) Failing after 10m39s
Matrix Smoke / distro: manjaro (push) Failing after 12m11s
Matrix Smoke / distro: opensuse-tw (push) Failing after 11m53s
Matrix Smoke / distro: archlinux (push) Failing after 20m3s
Matrix Smoke / distro: ubuntu-22.04 (push) Failing after 13m49s
Validate .env Schema / tier-1-env-validation (push) Successful in 52s
Validate .env Schema / tier-2-env-validation (push) Successful in 44s
Validate .env Schema / tier-3-env-validation (push) Successful in 52s
Validate .env Schema / tier-4-env-validation (push) Successful in 51s
Validate Extensions Catalog / Check catalog is up-to-date (push) Failing after 9m47s
Secret Scan / Scan for secrets (push) Failing after 21m4s
Validate Docker Compose / Validate Docker Compose files (push) Has been cancelled
Python Type Check / Type check with mypy (push) Has been cancelled
Validate .env Schema / tier-0-env-validation (push) Has been cancelled
Test Linux / integration-smoke (push) Has been cancelled
Lint PowerShell / powershell-lint (windows-latest) (push) Has been cancelled
Matrix Smoke / macos-smoke (push) Has been cancelled
2.0 KiB
2.0 KiB
System Tuning for Strix Halo
These files optimize the system for LLM inference on AMD Strix Halo.
Apply all tuning (requires reboot for GRUB/modprobe):
# 1. Kernel boot parameters (GRUB)
# amd_iommu=off gives 2-6% improvement (iommu=pt does NOT give the same benefit)
sudo sed -i 's/GRUB_CMDLINE_LINUX_DEFAULT="quiet splash"/GRUB_CMDLINE_LINUX_DEFAULT="quiet splash amd_iommu=off"/' /etc/default/grub
sudo update-grub
# 2. AMD GPU module options
sudo cp amdgpu.conf /etc/modprobe.d/amdgpu.conf
sudo cp amdgpu_llm_optimized.conf /etc/modprobe.d/amdgpu_llm_optimized.conf
sudo update-initramfs -u
# 3. Memory tuning (applies immediately + persists)
sudo cp 99-ods.conf /etc/sysctl.d/99-ods.conf
sudo sysctl --system
# 4. Enable tuned for CPU governor optimization (5-8% prompt processing improvement)
sudo apt install tuned # or: sudo dnf install tuned
sudo systemctl enable --now tuned
sudo tuned-adm profile accelerator-performance
# 5. Reboot for GRUB + modprobe changes
sudo reboot
What each setting does:
GRUB parameters
amd_iommu=off— disable IOMMU for lower GPU memory access overhead (2-6% improvement)
modprobe (amdgpu.conf)
ppfeaturemask=0xffffffff— enable all power management featuresgpu_recovery=1— enable GPU hang recovery
modprobe (amdgpu_llm_optimized.conf)
gttsize=120000— allocate 120GB as GPU GTT memory (where HIP puts model weights)pages_limit=31457280— max 4KiB pages for GPU memory (120 GB)page_pool_size=15728640— pre-cache ~60GB for GPU usage (reduces allocation latency)
sysctl (99-ods.conf)
vm.swappiness=10— prefer keeping data in RAM (default 60 is too aggressive at swapping)vm.vfs_cache_pressure=50— keep directory/inode caches longer
tuned (accelerator-performance)
- Sets CPU governor to
performance(no power-saving throttling during inference) - Disables CPU idle states for lowest latency
- 5-8% prompt processing improvement measured on Strix Halo