TrimLLM: Progressive Layer Dropping for Domain-Specific LLMs
This addresses the need for efficient, private domain-specific LLMs, offering a practical compression method that works across hardware without requiring specialized support, though it is incremental as it builds on existing compression techniques.
The paper tackles the problem of specializing large language models for domain-specific deployment by introducing TrimLLM, a method that reduces model depth via progressive layer dropping, achieving 2.1-5.7x inference speedup on consumer GPUs with no accuracy loss at 50-60% compression ratios.
Specializing large language models (LLMs) for local deployment in domain-specific use cases is necessary for strong performance while meeting latency and privacy constraints. However, conventional task-specific adaptation approaches do not show simultaneous memory saving and inference speedup at deployment time. Practical compression techniques like quantization and pruning require dedicated hardware or kernel support to achieve measured inference speedup. We develop TrimLLM based on the layer-wise specialization phenomenon we empirically observed and verified on contemporary LLMs. TrimLLM reduces the depth of LLMs via progressive layer dropping. We show it retains LLMs' capacity in specific domains and achieves inference speedup irrespective of hardware and deep learning frameworks. We evaluated TrimLLM on LLMs of various sizes for inference; models adapted on medical, legal, and financial datasets all demonstrate $2.1-5.7\times$ inference speedup on consumer GPUs and up to $3.1\times$ speedup on A100 when compared to state-of-the-art model compression algorithms, with no loss in accuracy at 50$\sim$60\% model compression ratio.