红尘
d73ff9b0fb
Deploy Site / deploy-vercel (push) Has been cancelled
Deploy Site / deploy-docs (push) Has been cancelled
Docker / shell lint / Lint Dockerfile (hadolint) (push) Has been cancelled
Docker / shell lint / Lint docker/ shell scripts (shellcheck) (push) Has been cancelled
Docker Build and Publish / build-amd64 (push) Has been cancelled
Docker Build and Publish / build-arm64 (push) Has been cancelled
Lint (ruff + ty) / ruff + ty diff (push) Has been cancelled
Lint (ruff + ty) / ruff enforcement (blocking) (push) Has been cancelled
Lint (ruff + ty) / Windows footguns (blocking) (push) Has been cancelled
Nix Lockfile Fix / auto-fix-main (push) Has been cancelled
Nix Lockfile Fix / fix (push) Has been cancelled
Nix / nix (macos-latest) (push) Has been cancelled
Nix / nix (ubuntu-latest) (push) Has been cancelled
OSV-Scanner / Scan lockfiles (push) Has been cancelled
Build Skills Index / build-index (push) Has been cancelled
Tests / test (1) (push) Has been cancelled
Tests / test (2) (push) Has been cancelled
Tests / test (3) (push) Has been cancelled
Tests / test (4) (push) Has been cancelled
Tests / test (5) (push) Has been cancelled
Tests / test (6) (push) Has been cancelled
Tests / e2e (push) Has been cancelled
uv.lock check / uv lock --check (push) Has been cancelled
Docker Build and Publish / merge (push) Has been cancelled
Build Skills Index / trigger-deploy (push) Has been cancelled
Tests / save-durations (push) Has been cancelled
1.6 KiB
1.6 KiB
Performance Optimization Guide
Maximize llama.cpp inference speed and efficiency.
CPU Optimization
Thread tuning
# Set threads (default: physical cores)
./llama-cli -m model.gguf -t 8
# For AMD Ryzen 9 7950X (16 cores, 32 threads)
-t 16 # Best: physical cores
# Avoid hyperthreading (slower for matrix ops)
BLAS acceleration
# OpenBLAS (faster matrix ops)
make LLAMA_OPENBLAS=1
# BLAS gives 2-3× speedup
GPU Offloading
Layer offloading
# Offload 35 layers to GPU (hybrid mode)
./llama-cli -m model.gguf -ngl 35
# Offload all layers
./llama-cli -m model.gguf -ngl 999
# Find optimal value:
# Start with -ngl 999
# If OOM, reduce by 5 until fits
Memory usage
# Check VRAM usage
nvidia-smi dmon
# Reduce context if needed
./llama-cli -m model.gguf -c 2048 # 2K context instead of 4K
Batch Processing
# Increase batch size for throughput
./llama-cli -m model.gguf -b 512 # Default: 512
# Physical batch (GPU)
--ubatch 128 # Process 128 tokens at once
Context Management
# Default context (512 tokens)
-c 512
# Longer context (slower, more memory)
-c 4096
# Very long context (if model supports)
-c 32768
Benchmarks
CPU Performance (Llama 2-7B Q4_K_M)
| Setup | Speed | Notes |
|---|---|---|
| Apple M3 Max | 50 tok/s | Metal acceleration |
| AMD 7950X (16c) | 35 tok/s | OpenBLAS |
| Intel i9-13900K | 30 tok/s | AVX2 |
GPU Offloading (RTX 4090)
| Layers GPU | Speed | VRAM |
|---|---|---|
| 0 (CPU only) | 30 tok/s | 0 GB |
| 20 (hybrid) | 80 tok/s | 8 GB |
| 35 (all) | 120 tok/s | 12 GB |