feat(optimize): add Phase 2 power profile and system tuning
Add `make optimize-power` (ryzenadj 85W, sysctl, THP, RADV nogttspill) with systemd services for boot/resume persistence. Integrate into `make optimize --all` as Phase 2. Update optimization log with RyzenAdj results (+46% tg at 70W sustained), KV sweep data, and quant shootout. Add Qwen3-Coder-30B and Nemotron-Cascade-2 to model catalog.
This commit is contained in:
Felipe Cardoso
3
Makefile
3
Makefile
@@ -48,6 +48,9 @@ optimize-kernel: ## Configure kernel boot parameters
|
||||
optimize-tuned: ## Switch to accelerator-performance profile
|
||||
@bash scripts/optimize/tuned-profile.sh
|
||||
|
||||
optimize-power: ## Apply Phase 2 tuning (ryzenadj, sysctl, THP, RADV)
|
||||
@bash scripts/optimize/power-profile.sh
|
||||
|
||||
optimize-vram: ## BIOS VRAM guidance + GTT verification
|
||||
@bash scripts/optimize/vram-gtt.sh
|
||||
|
||||
|
||||
13
bin/optimize
13
bin/optimize
@@ -6,24 +6,31 @@ SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
|
||||
|
||||
case "${1:---all}" in
|
||||
--all|-a)
|
||||
echo "Running optimization walkthrough..."
|
||||
echo "Running full optimization walkthrough..."
|
||||
echo ""
|
||||
echo "=== Phase 1: Core System ==="
|
||||
bash "$SCRIPT_DIR/scripts/optimize/tuned-profile.sh"
|
||||
bash "$SCRIPT_DIR/scripts/optimize/kernel-params.sh"
|
||||
bash "$SCRIPT_DIR/scripts/optimize/vram-gtt.sh"
|
||||
echo ""
|
||||
echo "=== Phase 2: System Tuning ==="
|
||||
bash "$SCRIPT_DIR/scripts/optimize/power-profile.sh"
|
||||
echo ""
|
||||
bash "$SCRIPT_DIR/scripts/optimize/verify.sh"
|
||||
;;
|
||||
--kernel|-k) exec bash "$SCRIPT_DIR/scripts/optimize/kernel-params.sh" ;;
|
||||
--tuned|-t) exec bash "$SCRIPT_DIR/scripts/optimize/tuned-profile.sh" ;;
|
||||
--vram|-v) exec bash "$SCRIPT_DIR/scripts/optimize/vram-gtt.sh" ;;
|
||||
--power|-p) exec bash "$SCRIPT_DIR/scripts/optimize/power-profile.sh" ;;
|
||||
--verify) exec bash "$SCRIPT_DIR/scripts/optimize/verify.sh" ;;
|
||||
--rollback) exec bash "$SCRIPT_DIR/scripts/optimize/rollback.sh" ;;
|
||||
*)
|
||||
echo "Usage: optimize [--all|--kernel|--tuned|--vram|--verify|--rollback]"
|
||||
echo " --all Full optimization walkthrough (default)"
|
||||
echo "Usage: optimize [--all|--kernel|--tuned|--vram|--power|--verify|--rollback]"
|
||||
echo " --all Full optimization walkthrough (Phase 1 + 2)"
|
||||
echo " --kernel Configure kernel boot parameters"
|
||||
echo " --tuned Switch tuned profile"
|
||||
echo " --vram BIOS VRAM + GTT guidance"
|
||||
echo " --power Phase 2: ryzenadj, sysctl, THP, RADV"
|
||||
echo " --verify Post-optimization checklist"
|
||||
echo " --rollback Revert changes"
|
||||
exit 1
|
||||
|
||||
@@ -21,3 +21,10 @@ qwen3.5-27b-q4|unsloth/Qwen3.5-27B-GGUF|Qwen3.5-27B-Q4_K_M.gguf|17|dense|Dense 2
|
||||
qwen3.5-35b-a3b-q4|unsloth/Qwen3.5-35B-A3B-GGUF|Qwen3.5-35B-A3B-UD-Q4_K_L.gguf|19|moe|MoE 35B, 3B active, Unsloth dynamic
|
||||
qwen3.5-35b-a3b-q8|unsloth/Qwen3.5-35B-A3B-GGUF|Qwen3.5-35B-A3B-Q8_0.gguf|37|moe|MoE 35B Q8, near-full precision
|
||||
nemotron-30b-a3b-q4|lmstudio-community/NVIDIA-Nemotron-3-Nano-30B-A3B-GGUF|NVIDIA-Nemotron-3-Nano-30B-A3B-Q4_K_M.gguf|23|moe|Nemotron MoE 30B, 3B active
|
||||
nemotron-cascade2-q8|bartowski/nvidia_Nemotron-Cascade-2-30B-A3B-GGUF|nvidia_Nemotron-Cascade-2-30B-A3B-Q8_0.gguf|31|moe|Nemotron Cascade 2, Mamba-2 hybrid
|
||||
|
||||
# ── Coding models ─────────────────────────────────────────
|
||||
qwen3-coder-30b-a3b-q6|unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF|Qwen3-Coder-30B-A3B-Instruct-UD-Q6_K_XL.gguf|26|moe|Agentic coding MoE, pure Transformer
|
||||
|
||||
# ── Draft models (speculative decoding) ───────────────────
|
||||
qwen3.5-0.8b-q8-draft|unsloth/Qwen3.5-0.8B-GGUF|Qwen3.5-0.8B-Q8_0.gguf|0.8|draft|Draft for Qwen3.5 speculative decoding
|
||||
|
||||
14
configs/ryzenadj-llm.service
Normal file
14
configs/ryzenadj-llm.service
Normal file
@@ -0,0 +1,14 @@
|
||||
[Unit]
|
||||
Description=Apply RyzenAdj power limits for LLM inference
|
||||
After=multi-user.target
|
||||
|
||||
[Service]
|
||||
Type=oneshot
|
||||
ExecStart=/usr/local/bin/ryzenadj --stapm-limit=85000 --fast-limit=85000 --slow-limit=85000 --apu-slow-limit=85000
|
||||
RemainAfterExit=yes
|
||||
|
||||
# Re-apply after resume from sleep/hibernate (HP firmware resets limits)
|
||||
ExecStartPost=/bin/sleep 2
|
||||
|
||||
[Install]
|
||||
WantedBy=multi-user.target
|
||||
10
configs/ryzenadj-resume.service
Normal file
10
configs/ryzenadj-resume.service
Normal file
@@ -0,0 +1,10 @@
|
||||
[Unit]
|
||||
Description=Re-apply RyzenAdj power limits after resume
|
||||
After=suspend.target hibernate.target hybrid-sleep.target suspend-then-hibernate.target
|
||||
|
||||
[Service]
|
||||
Type=oneshot
|
||||
ExecStart=/usr/local/bin/ryzenadj --stapm-limit=85000 --fast-limit=85000 --slow-limit=85000 --apu-slow-limit=85000
|
||||
|
||||
[Install]
|
||||
WantedBy=suspend.target hibernate.target hybrid-sleep.target suspend-then-hibernate.target
|
||||
@@ -37,38 +37,42 @@ Living document tracking what was applied, tested, and the actual results. Each
|
||||
|
||||
## Phase 2: System Tuning
|
||||
|
||||
### 2.1 RyzenAdj 85W PPT
|
||||
### 2.1 RyzenAdj PPT Increase
|
||||
|
||||
- **Date**: PENDING
|
||||
- **Change**: `sudo ryzenadj --stapm-limit=85000 --fast-limit=85000 --slow-limit=85000`
|
||||
- **Expected**: +12-19% CPU/GPU throughput (community data from Strix Halo Wiki)
|
||||
- **Benchmark**: Not yet run
|
||||
- **Notes**: HP ZBook ships at 60W. 85W is the community-recommended sweet spot.
|
||||
- **Verdict**: PENDING
|
||||
- **Date**: 2026-03-30
|
||||
- **Change**: `sudo ryzenadj --stapm-limit=85000 --fast-limit=85000 --slow-limit=85000 --apu-slow-limit=85000`
|
||||
- **Result**: STAPM raised from 59W→81W. PPT Fast raised to 81W. **However, PPT SLOW and APU SLOW stuck at 70W** — HP ZBook BIOS EC overrides these limits. Effective sustained power: ~70W (was ~59W).
|
||||
- **Benchmark**: `data/benchmarks/qwen35-shootout-v2-*` (Vulkan, q4_0 KV, pp2048/tg1024)
|
||||
- UD-Q4_K_L: **57.0 t/s** (was ~39 t/s before RyzenAdj = **+46%**)
|
||||
- UD-Q4_K_XL: **56.4 t/s**
|
||||
- Q8_0: **51.4 t/s** (was ~39-41 t/s before = **+25%**)
|
||||
- **Thermals**: 70-73C under load, 30C headroom. Cooling handles it easily.
|
||||
- **Notes**: Settings are volatile (reset on reboot/sleep). Use `sudo make optimize-power` or install systemd service for persistence. HP firmware hard-caps slow PPT at 70W regardless.
|
||||
- **Verdict**: KEEP — significant real-world improvement despite HP firmware limit
|
||||
|
||||
### 2.2 VM Sysctl Tuning
|
||||
|
||||
- **Date**: PENDING
|
||||
- **Change**: `vm.swappiness=1, vm.dirty_ratio=40, vm.max_map_count=500000`
|
||||
- **Expected**: Prevent model weight eviction, reduce I/O disruption
|
||||
- **Benchmark**: Not yet run
|
||||
- **Verdict**: PENDING
|
||||
- **Date**: 2026-03-30
|
||||
- **Change**: `vm.swappiness=1, vm.dirty_ratio=40, vm.dirty_background_ratio=10, vm.max_map_count=500000, vm.zone_reclaim_mode=0`
|
||||
- **Applied via**: `sudo make optimize-power` (persists to `/etc/sysctl.d/99-llm-inference.conf`)
|
||||
- **Notes**: Hard to isolate impact — applied together with other Phase 2 changes. Prevents model weight eviction and I/O disruption.
|
||||
- **Verdict**: KEEP — low risk, persists across reboots
|
||||
|
||||
### 2.3 Transparent Huge Pages
|
||||
|
||||
- **Date**: PENDING
|
||||
- **Change**: `transparent_hugepage=always`
|
||||
- **Expected**: Faster model load time, possible 1-5% tg improvement from reduced TLB misses
|
||||
- **Benchmark**: Not yet run
|
||||
- **Verdict**: PENDING
|
||||
- **Date**: 2026-03-30
|
||||
- **Change**: `echo always > /sys/kernel/mm/transparent_hugepage/enabled`
|
||||
- **Applied via**: `sudo make optimize-power` (volatile — add `transparent_hugepage=always` to kernel cmdline for persistence)
|
||||
- **Notes**: Reduces TLB misses for mmap'd model files. Hard to isolate impact.
|
||||
- **Verdict**: KEEP — low risk
|
||||
|
||||
### 2.4 RADV_PERFTEST=nogttspill
|
||||
|
||||
- **Date**: PENDING
|
||||
- **Change**: `export RADV_PERFTEST=nogttspill`
|
||||
- **Expected**: Fix pp degradation on Vulkan RADV (community-reported fix for Strix Halo)
|
||||
- **Benchmark**: Not yet run — needs Vulkan-specific benchmark comparison
|
||||
- **Verdict**: PENDING
|
||||
- **Date**: 2026-03-30
|
||||
- **Change**: `RADV_PERFTEST=nogttspill` persisted to `/etc/environment.d/radv-llm.conf`
|
||||
- **Applied via**: `sudo make optimize-power`
|
||||
- **Notes**: Prevents GTT spill management overhead on unified memory Vulkan. Takes effect on next login. For current session: `export RADV_PERFTEST=nogttspill`
|
||||
- **Verdict**: KEEP — persists across reboots
|
||||
|
||||
### 2.5 amdgpu.noretry=0
|
||||
|
||||
@@ -84,11 +88,19 @@ Living document tracking what was applied, tested, and the actual results. Each
|
||||
|
||||
### 3.1 KV Cache Quantization
|
||||
|
||||
- **Date**: PENDING (sweep running)
|
||||
- **Change**: `-ctk q8_0 -ctv q8_0` / `-ctk q4_0 -ctv q4_0`
|
||||
- **Benchmark**: `data/benchmarks/kv-sweep-128k-*` (in progress)
|
||||
- **Expected**: Q8_0: ~50% less KV memory, negligible quality loss. Q4_0: ~75% less, noticeable quality impact.
|
||||
- **Verdict**: PENDING
|
||||
- **Date**: 2026-03-27
|
||||
- **Change**: `--kv-types f16,q8_0,q4_0` sweep
|
||||
- **Benchmark**: `data/benchmarks/kv-sweep-256k-*`
|
||||
- **Result** (Vulkan RADV, Qwen3.5-35B-A3B Q8, pp2048/tg1024):
|
||||
- f16: 456 pp, 39.8 tg
|
||||
- q8_0: 418 pp, 38.5 tg (slight Vulkan regression — unexpected)
|
||||
- **q4_0: 460 pp, 41.1 tg** (fastest overall, +3% tg over f16)
|
||||
- **Result** (ROCm, same model):
|
||||
- f16: 445 pp, 21.5 tg
|
||||
- q8_0: 495 pp, 21.7 tg (+11% pp, same tg)
|
||||
- q4_0: 494 pp, 21.8 tg (+11% pp, same tg)
|
||||
- **Conclusion**: q4_0 is the sweet spot on Vulkan (fastest tg + 75% less KV memory). On ROCm, KV quant helps pp but not tg.
|
||||
- **Verdict**: KEEP — use q4_0 KV as default for serving
|
||||
|
||||
### 3.2 MoE Batch Size `-b 256`
|
||||
|
||||
@@ -161,6 +173,24 @@ Living document tracking what was applied, tested, and the actual results. Each
|
||||
|
||||
---
|
||||
|
||||
## Model Quant Shootout
|
||||
|
||||
### Qwen3.5-35B-A3B — Q4_K_L vs Q4_K_XL vs Q8 (2026-03-30)
|
||||
|
||||
- **Benchmark**: `data/benchmarks/qwen35-shootout-v2-*`
|
||||
- **Config**: Vulkan RADV, q4_0 KV cache, pp2048/tg1024, 2 reps
|
||||
- **RyzenAdj**: STAPM=81W (sustained ~70W due to HP firmware cap)
|
||||
|
||||
| Quant | File Size | pp2048 (t/s) | tg1024 (t/s) | Recommendation |
|
||||
|-------|-----------|-------------|-------------|----------------|
|
||||
| UD-Q4_K_L | 18.8 GB | 825 | **57.0** | Fastest. Good quality. |
|
||||
| **UD-Q4_K_XL** | 20.7 GB | 835 | **56.4** | **Daily driver** — best quality/speed. |
|
||||
| Q8_0 | 34.4 GB | 850 | 51.4 | Best quality, 10% slower tg. |
|
||||
|
||||
**Decision**: Keep UD-Q4_K_XL (daily driver) and Q8_0 (quality fallback). Q4_K_L can be deleted — Q4_K_XL is strictly better at only +2 GB.
|
||||
|
||||
---
|
||||
|
||||
## How to Add Entries
|
||||
|
||||
When testing a new optimization:
|
||||
|
||||
111
scripts/optimize/power-profile.sh
Normal file
111
scripts/optimize/power-profile.sh
Normal file
@@ -0,0 +1,111 @@
|
||||
#!/usr/bin/env bash
|
||||
# Apply power profile and system tuning for LLM inference workloads
|
||||
# Requires root. Settings are volatile — use the systemd service for persistence.
|
||||
set -euo pipefail
|
||||
|
||||
SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
|
||||
source "$SCRIPT_DIR/../../lib/common.sh"
|
||||
source "$SCRIPT_DIR/../../lib/format.sh"
|
||||
|
||||
require_root
|
||||
|
||||
# ── Power limits via ryzenadj ─────────────────────────────
|
||||
STAPM=85000
|
||||
FAST=85000
|
||||
SLOW=85000
|
||||
APU_SLOW=85000
|
||||
|
||||
if is_cmd ryzenadj; then
|
||||
log_header "Power Profile (ryzenadj)"
|
||||
log_info "Setting STAPM=${STAPM}mW, Fast=${FAST}mW, Slow=${SLOW}mW, APU=${APU_SLOW}mW"
|
||||
ryzenadj \
|
||||
--stapm-limit=$STAPM \
|
||||
--fast-limit=$FAST \
|
||||
--slow-limit=$SLOW \
|
||||
--apu-slow-limit=$APU_SLOW 2>&1 | grep -E 'Successfully|Error|not supported' || true
|
||||
|
||||
# Verify what actually took effect
|
||||
log_info "Verifying limits..."
|
||||
ryzenadj -i 2>&1 | grep -E 'LIMIT|VALUE' | head -8
|
||||
echo ""
|
||||
log_warn "Note: HP firmware may cap PPT SLOW/APU at 70W regardless of setting"
|
||||
else
|
||||
log_error "ryzenadj not found. Install: cd /tmp && git clone https://github.com/FlyGoat/RyzenAdj.git && cd RyzenAdj && mkdir build && cd build && cmake .. && make && sudo cp ryzenadj /usr/local/bin/"
|
||||
exit 1
|
||||
fi
|
||||
|
||||
# ── VM sysctl tuning ──────────────────────────────────────
|
||||
log_header "VM Sysctl Tuning"
|
||||
|
||||
declare -A SYSCTLS=(
|
||||
[vm.swappiness]=1
|
||||
[vm.dirty_ratio]=40
|
||||
[vm.dirty_background_ratio]=10
|
||||
[vm.max_map_count]=500000
|
||||
[vm.zone_reclaim_mode]=0
|
||||
)
|
||||
|
||||
for KEY in "${!SYSCTLS[@]}"; do
|
||||
VAL="${SYSCTLS[$KEY]}"
|
||||
CURRENT=$(sysctl -n "$KEY" 2>/dev/null || echo "?")
|
||||
if [[ "$CURRENT" == "$VAL" ]]; then
|
||||
log_success "$KEY = $VAL (already set)"
|
||||
else
|
||||
sysctl -w "$KEY=$VAL" > /dev/null 2>&1
|
||||
log_success "$KEY = $VAL (was $CURRENT)"
|
||||
fi
|
||||
done
|
||||
|
||||
# Persist sysctl settings
|
||||
SYSCTL_CONF="/etc/sysctl.d/99-llm-inference.conf"
|
||||
if [[ ! -f "$SYSCTL_CONF" ]]; then
|
||||
log_info "Persisting to $SYSCTL_CONF"
|
||||
cat > "$SYSCTL_CONF" << 'EOF'
|
||||
# LLM inference optimizations
|
||||
vm.swappiness = 1
|
||||
vm.dirty_ratio = 40
|
||||
vm.dirty_background_ratio = 10
|
||||
vm.max_map_count = 500000
|
||||
vm.zone_reclaim_mode = 0
|
||||
EOF
|
||||
log_success "Sysctl config saved (persists across reboots)"
|
||||
else
|
||||
log_info "Sysctl config already exists at $SYSCTL_CONF"
|
||||
fi
|
||||
|
||||
# ── Transparent Huge Pages ────────────────────────────────
|
||||
log_header "Transparent Huge Pages"
|
||||
|
||||
THP_ENABLED=$(cat /sys/kernel/mm/transparent_hugepage/enabled 2>/dev/null || echo "unknown")
|
||||
if [[ "$THP_ENABLED" == *"[always]"* ]]; then
|
||||
log_success "THP = always (already set)"
|
||||
else
|
||||
echo always > /sys/kernel/mm/transparent_hugepage/enabled 2>/dev/null || true
|
||||
echo defer+madvise > /sys/kernel/mm/transparent_hugepage/defrag 2>/dev/null || true
|
||||
log_success "THP = always, defrag = defer+madvise"
|
||||
fi
|
||||
log_info "For persistence, add to kernel cmdline: transparent_hugepage=always"
|
||||
|
||||
# ── RADV nogttspill ───────────────────────────────────────
|
||||
log_header "Vulkan RADV Environment"
|
||||
|
||||
RADV_CONF="/etc/environment.d/radv-llm.conf"
|
||||
if [[ ! -f "$RADV_CONF" ]]; then
|
||||
mkdir -p /etc/environment.d
|
||||
echo 'RADV_PERFTEST=nogttspill' > "$RADV_CONF"
|
||||
log_success "RADV_PERFTEST=nogttspill persisted to $RADV_CONF"
|
||||
log_info "Takes effect on next login. For this session: export RADV_PERFTEST=nogttspill"
|
||||
else
|
||||
log_success "RADV config already exists at $RADV_CONF"
|
||||
fi
|
||||
|
||||
# ── Summary ───────────────────────────────────────────────
|
||||
log_header "Phase 2 Optimization Summary"
|
||||
log_success "Power profile: ryzenadj limits applied (volatile — resets on reboot)"
|
||||
log_success "VM tuning: sysctl applied and persisted"
|
||||
log_success "THP: enabled (volatile — add to kernel cmdline for persistence)"
|
||||
log_success "RADV: nogttspill persisted"
|
||||
echo ""
|
||||
log_info "To persist ryzenadj across reboots:"
|
||||
log_info " sudo cp $SCRIPT_DIR/../../configs/ryzenadj-llm.service /etc/systemd/system/"
|
||||
log_info " sudo systemctl enable --now ryzenadj-llm.service"
|
||||
Reference in New Issue
Block a user