Efficient Knowledge Distillation for LLMs: Offline Top-K Logits and a Fused Chunked KL Loss

Small language models are often the only option for deployment under tight latency, cost, and on-premises constraints, but they are rarely trained from scratch: a compressed model is usually recovered through knowledge distillation (KD). This recovery step largely decides the final quality, yet it is expensive. We present a practitioner's study of how to make distillation training efficient, organised around two systems contributions. First, we show that offline KD (caching the teacher's top-$K$ logits once and training the student against the cache) matches online distillation at near-identical training loss while removing the teacher from memory, running about 29\% faster per iteration, and reaching up to 41\% higher throughput on a single H200 GPU. Second, we introduce a \emph{fused, chunked KL loss} that never materialises the full vocabulary-sized logit tensor, making peak memory linear in the sequence length. This removes the memory spike that otherwise caps context length and lets us train at four times the context (32{,}768 tokens) on a single GPU. A separate output-head-only toy benchmark isolates the loss kernel and confirms its memory and iteration-rate scaling from 4K to 256K tokens. Together these make large-scale healing and hundreds of ablations affordable. We also report supporting ablations on loss design and sequence packing. We release our chunked-loss implementation: https://github.com/CompactifAI/Full-Chunked-KL-Loss.

Block removal for large language models through constrained binary optimization

Compressing resource-intensive large language models by removing whole transformer blocks is a seemingly simple idea, but identifying which blocks to remove constitutes an exponentially difficult combinatorial problem. In this paper, we formulate block removal as a constrained binary optimization problem that can be mapped to a physical system (Ising model), whose energies are a strong proxy for downstream model performance. This formulation enables an efficient ranking of a large number of candidate block-removal configurations and yields many high-quality, non-trivial solutions beyond consecutive regions. We demonstrate that our approach outperforms state-of-the-art block-removal methods across several benchmarks, with performance gains persisting after short retraining, and reaching improvements of up to 6 points on the MMLU benchmark. Our method requires only forward and backward passes for a few active parameters, together with an (at least approximate) Ising solver, and can be readily applied to any architecture. We illustrate this generality on the recent NVIDIA-Nemotron-3-Nano-30B-A3B-FP8 model, which exhibits a highly inhomogeneous and challenging block structure.