Local LLMs can be 2–3× faster with proper optimization. Techniques include: disabling logging, reducing batch size, optimizing quantization, using faster inference engines, and GPU memory tuning. As of April 2026, combining these techniques can achieve 2× speed improvement with no quality loss.

重要なポイント

Disable logging/debugging (easy): ~10% speed gain.
Use Q4 quantization (easy): Same speed, smaller VRAM.
Optimize batch size (medium): 2–3× speed for batch processing.
Use vLLM instead of Ollama (hard): 2–5× speed for concurrent requests.
GPU memory utilization 90%+ (medium): 15–20% speed gain.
Combining all techniques: ~2–3× total speedup.

GPU Memory Utilization: The Hidden Speed Dial

By default, most tools use 70–80% of GPU VRAM. Increasing to 90%+ improves speed by 15–20%:

bash

# vLLM: increase GPU memory utilization
vllm serve meta-llama/Llama-2-7b-hf \
  --gpu-memory-utilization 0.95

# Ollama: environment variable
export OLLAMA_GPU_THRESHOLD=0.95  # Use 95% of GPU
ollama run llama3.2:3b

# LM Studio: Settings → GPU acceleration slider (move to 100%)

Batch Size: The Multiplier for Throughput

For batch processing (multiple prompts), increasing batch size dramatically improves throughput.

Single request = limited pipeline utilization. Batch 32 requests = 2–4× throughput.

Trade-off: Higher latency per individual request (they wait for batch to complete).

Batch Size	Throughput	Latency/Request	Use Case
1 (single)	50 tokens/sec	Minimum	Real-time chat
8	120 tokens/sec	Acceptable	Light concurrency
32	200 tokens/sec	High	Batch API
64+	250+ tokens/sec	Very high	Offline batch

Inference Engine Selection and Tuning

vLLM: 5–10× faster than Ollama for batch processing (concurrent requests).

llama.cpp: Fastest for single requests on consumer hardware.

Text-Generation-WebUI: Slower, but more features for experimentation.

For production APIs, vLLM is optimal.

Quantization Impact on Speed

Q4 and Q5 are approximately the same speed as FP16 on modern GPUs. Older GPUs may benefit from quantization speed-ups.

Benefits of quantization for speed:

- Smaller model file = faster loading

- Reduced memory bandwidth = slightly faster (10–15%) on some hardware

Quantization is primarily for VRAM reduction, not speed.

Realistic Speed Improvements With Tuning

Example: Optimizing a 7B model on RTX 4090:

Change	Speed	Cumulative Gain
—	120 tokens/sec	—
—	132 tokens/sec	—
—	150 tokens/sec	—
—	300 tokens/sec (batch)	—
—	300 tokens/sec	—

Common Speed Optimization Mistakes

Pushing GPU memory to 100%. Risks out-of-memory crashes. Safe max is 90–95%.
Lowering batch size for speed. Batch size does not affect single-request latency. Only helps throughput.
Over-quantizing for speed. Q4 is roughly the same speed as FP16. Quantize for VRAM, not speed.
Changing inference engine mid-deployment. Switching Ollama → vLLM → llama.cpp introduces bugs. Pick one, optimize it.

Sources

vLLM Optimization Guide — docs.vllm.ai/en/dev_guide/performance_tuning.html
Ollama Performance Tips — github.com/ollama/ollama/blob/main/docs/troubleshooting.md

How to Double Local LLM Speed: Optimization Techniques