The author has built MOLT, a Windows-first runtime for thermally aware QLoRA fine-tuning on consumer NVIDIA GPUs, and benchmarked it against Unsloth using one million Qwen 1.5B training targets on an RTX 4060 Laptop GPU.
- End-to-end session time was 1,103 seconds for MOLT versus 1,433 seconds for Unsloth, a 23% reduction in total training time.
- Training throughput reached 1,063 targets/s with MOLT compared to 768 targets/s with Unsloth, representing a 38% increase.
- Peak allocated VRAM was 1.63 GB for MOLT versus 1.85 GB for Unsloth, a 12% decrease in memory usage.
- Measured board energy consumption was nearly identical at 63.1 kJ for MOLT and 63.8 kJ for Unsloth.
The author notes this is single-machine, single-seed development evidence and invites Windows NVIDIA GPU users to reproduce the benchmark and test their own workloads.