This work investigates distillation methods for large language models (LLMs) with the goal of developing compact models that preserve high performance. Several existing approaches are reviewed, with a discussion of their respective strengths and limitations. An improved method based on the ShortGPT approach has been developed, building upon the idea of incorporating iterative evaluation of layer importance. At each step, importance is assessed by measuring performance degradation when individual layers are removed, using a set of representative datasets. This process is combined with further training using a joint loss function based on KL divergence and mean squared error. Experiments on the Qwen2.5-3B model show that the number of layers can be reduced from 36 to 28 (resulting in a 2.47 billion parameter model) with only a 9.7% quality loss, and to 24 layers with an 18% loss. The findings suggest that the middle transformer layers contribute less to inference, underscoring the potential of the proposed method for creating efficient models. The results demonstrate the effectiveness of iterative distillation and fine-tuning, making the approach suitable for deployment in resource-limited settings.

Knowledge graphs represent complex data using nodes, relationships, and properties. Cypher, a powerful query language for graph databases, enables efficient modeling and querying. Recent advancements in large language models allow translation of natural language questions into Cypher queries - Text2Cypher. A common approach is incorporating database schema into prompts. However, complex schemas can introduce noise, increase hallucinations, and raise computational costs. Schema filtering addresses these challenges by including only relevant schema elements, improving query generation while reducing token costs. This work explores various schema filtering methods for Text2Cypher task and analyzes their impact on token length, performance, and cost. Results show that schema filtering effectively optimizes Text2Cypher, especially for smaller models. Consistent with prior research, we find that larger models benefit less from schema filtering due to their longer context capabilities. However, schema filtering remains valuable for both larger and smaller models in cost reduction.

Neural network pruning is essential for reducing model complexity to enable deployment on resource constrained hardware. While performance loss of pruned networks is often attributed to the removal of critical parameters, we identify signal collapse a reduction in activation variance across layers as the root cause. Existing one shot pruning methods focus on weight selection strategies and rely on computationally expensive second order approximations. In contrast, we demonstrate that mitigating signal collapse, rather than optimizing weight selection, is key to improving accuracy of pruned networks. We propose REFLOW that addresses signal collapse without updating trainable weights, revealing high quality sparse sub networks within the original parameter space. REFLOW enables magnitude pruning to achieve state of the art performance, restoring ResNeXt101 accuracy from under 4.1% to 78.9% on ImageNet with only 20% of the weights retained, surpassing state of the art approaches.

We show for the first time that large-scale generative pretrained transformer (GPT) family models can be pruned to at least 50% sparsity in one-shot, without any retraining, at minimal loss of accuracy. This is achieved via a new pruning method called SparseGPT, specifically designed to work efficiently and accurately on massive GPT-family models. We can execute SparseGPT on the largest available open-source models, OPT-175B and BLOOM-176B, in under 4.5 hours, and can reach 60% unstructured sparsity with negligible increase in perplexity: remarkably, more than 100 billion weights from these models can be ignored at inference time. SparseGPT generalizes to semi-structured (2:4 and 4:8) patterns, and is compatible with weight quantization approaches. The code is available at: https://github.com/IST-DASLab/sparsegpt.

The hls4ml library [1, 2] is an open source software designed to facilitate the deployment of machine learning (ML) models on field-programmable gate arrays (FPGAs), targeting low-latency and low-power edge applications. Taking as input a neural network model, hls4ml generates C/C code designed to be transpiled into FPGA firmware by processing it with a high-level synthesis (HLS) library. The development of hls4ml was historically driven by the need to integrate ML algorithms in the first stage of the real-time data processing of particle physics experiments operating at the CERN Large Hadron Collider (LHC). The LHC produces high-energy proton collisions (or events) every 25 ns, each consisting of about 1 MB of raw data. Since this throughput is overwhelming for the currently available processing and storage resources, the LHC experiments run a real-time event selection system, the so-called Level-1 trigger (L1T), to reduce the event rate from 40 MHz to 100 kHz [3-6]. Due to the size of the buffering system, the L1T system operates with a fixed latency of O(1 µs). While hls4ml excels as a tool to automatically generate low-latency ML firmware for L1T applications, it also offers interesting opportunities for edge-computing applications beyond particle physics whenever efficient, e.g.

Modern speech enhancement algorithms achieve remarkable noise suppression by means of large recurrent neural networks (RNNs). However, large RNNs limit practical deployment in hearing aid hardware (HW) form-factors, which are battery powered and run on resource-constrained microcontroller units (MCUs) with limited memory capacity and compute capability. In this work, we use model compression techniques to bridge this gap. We define the constraints imposed on the RNN by the HW and describe a method to satisfy them. Although model compression techniques are an active area of research, we are the first to demonstrate their efficacy for RNN speech enhancement, using pruning and integer quantization of weights/activations. We also demonstrate state update skipping, which reduces the computational load. Finally, we conduct a perceptual evaluation of the compressed models to verify audio quality on human raters. Results show a reduction in model size and operations of 11.9$\times$ and 2.9$\times$, respectively, over the baseline for compressed models, without a statistical difference in listening preference and only exhibiting a loss of 0.55dB SDR. Our model achieves a computational latency of 2.39ms, well within the 10ms target and 351$\times$ better than previous work.