Qwen2.5-14B OOM on RTX 3080 Ti (12GB)

Problem

When attempting to run or fine-tune Qwen2.5-14B on an NVIDIA RTX 3080 Ti with 12GB of VRAM, the process fails with an Out of Memory (OOM) error:

torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate X GiB (GPU 0; 12.00 GiB total capacity; Y GiB already allocated; Z GiB free; ...)

The 12GB VRAM limit is hit during the initial model load or immediately upon starting the first training step.

Root Causes

Model Size: A 14B parameter model in FP16/BF16 requires ~28GB of VRAM just for the weights.
Context Length: High context lengths (e.g., 4096+) significantly increase VRAM usage during training.
Training Overhead: Even with QLoRA (4-bit quantization), the overhead of gradients, optimizer states, and activations can exceed 12GB for a 14B model.

Solutions

1. Pivot to a 7B Model (Recommended)

For a 12GB GPU, a 7B parameter model (like Qwen2.5-7B-Instruct) is the sweet spot. It provides excellent performance while leaving enough VRAM for high context lengths and larger batch sizes.

VRAM Usage (7B QLoRA): ~6-8GB
Pros: Stable, fast, supports long context.
Cons: Slightly lower reasoning capability than 14B.

2. Aggressive Quantization

If you MUST run 14B, use 4-bit quantization (GGUF or EXL2) for inference only. Training 14B on 12GB is not reliably possible even with extreme offloading.

# Example Ollama run (uses 4-bit quantization by default)
ollama run qwen2.5:14b

3. Training Optimizations (if attempting 14B)

If you have no choice but to try 14B training:

Set max_seq_length to 512 or 1024.
Use Unsloth (it is highly memory-efficient).
Enable gradient_checkpointing.
Set per_device_train_batch_size = 1.

Maintenance

Keep your NVIDIA drivers and CUDA toolkit updated. On Windows (MajorRig), ensure WSL2 has sufficient memory allocation in .wslconfig.

2.0 KiB Raw Blame History