An official website of the United States government
raining modern large language models (LLMs) is extremely resource-intensive, and repeatedly customizing them for deployment scenarios with limited compute and memory is impractical. This paper introduces Flextron, a network architecture and post-training model optimization framework that supports flexible model deployment. Flextron uses a nested elastic structure that adapts rapidly to user-defined latency and accuracy targets during inference without requiring additional fine-tuning. It is also input-adaptive, automatically routing tokens through sub-networks for improved efficiency and performance. The authors propose a sample-efficient training method and routing algorithms to systematically transform an already-trained LLM into a Flextron model. Evaluation on the GPT-3 and LLaMA-2 families demonstrates Flextron’s superior performance over end-to-end trained variants and other state-of-the-art elastic networks, all with a single pretraining run that consumes only 7.63% of the tokens compared to original pretraining.
more »
« less
This content will become publicly available on January 1, 2026
Nudging: Inference-time Alignment of LLMs via Guided Decoding
Large language models (LLMs) require alignment to effectively and safely follow user instructions. This process necessitates training an aligned version for every base model, resulting in significant computational overhead. In this work, we propose NUDGING, a simple, training-free algorithm that aligns any base model at inference time using a small aligned model. NUDGING is motivated by recent findings that alignment primarily alters the model’s behavior on a small subset of stylistic tokens (e.g., discourse markers). We find that base models are significantly more uncertain when generating these tokens. Building on this insight, NUDGING employs a small aligned model to generate nudging tokens to guide the base model’s output during decoding when the base model’s uncertainty is high, with only a minor additional inference overhead. We evaluate NUDGING across 3 model families on a diverse range of open-instruction tasks. Without any training, nudging a large base model with a 7×-14× smaller aligned model achieves zero-shot performance comparable to, and sometimes surpassing, that of large aligned models. By operating at the token level, NUDGING enables off-the-shelf collaboration between model families. For instance, nudging Gemma-2-27b with Llama-27b-chat outperforms Llama-2-70b-chat on various tasks. Overall, our work offers a modular and cost-efficient solution to LLM alignment. Our code and demo are available at: https://fywalter.github.io/nudging/.
more »
« less
- Award ID(s):
- 2046873
- PAR ID:
- 10635512
- Publisher / Repository:
- Association for Computational Linguistics
- Date Published:
- Page Range / eLocation ID:
- 12702 to 12739
- Format(s):
- Medium: X
- Location:
- Vienna, Austria
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Nowadays, parameter-efficient fine-tuning (PEFT) large pre-trained models (LPMs) for downstream task have gained significant popularity, since it could significantly minimize the training computational overhead. The representative work, LoRA [1], learns a low-rank adaptor for a new downstream task, rather than fine-tuning the whole backbone model. However, for inference, the large size of the learned model remains unchanged, leading to in-efficient inference computation. To mitigate this, in this work, we are the first to propose a learning-to-prune methodology specially designed for fine-tuning downstream tasks based on LPMs with low-rank adaptation. Unlike prior low-rank adaptation approaches that only learn the low-rank adaptors for downstream tasks, our method further leverages the Gumbel-Sigmoid tricks to learn a set of trainable binary channel-wise masks that automatically prune the backbone LPMs. Therefore, our method could leverage the benefits of low-rank adaptation to reduce the training parameters size and smaller pruned backbone LPM size for efficient inference computation. Extensive experiments show that the Pruned-RoBbase model with our method achieves an average channel-wise structured pruning ratio of 24.5% across the popular GLUE Benchmark, coupled with an average of 18% inference time speed-up in real NVIDIA A5000 GPU. The Pruned-DistilBERT shows an average of 13% inference time improvement with 17% sparsity. The Pruned-LLaMA-7B model achieves up to 18.2% inference time improvement with 24.5% sparsity, demonstrating the effectiveness of our learnable pruning approach across different models and tasks.more » « less
-
The authors introduce DataComp for Language Models (DCLM), a testbed for controlled dataset experiments aimed at improving language models. DCLM provides a standardized corpus of 240T tokens extracted from Common Crawl, effective pretraining recipes based on the OpenLM framework, and a broad suite of 53 downstream evaluations. Participants can experiment with dataset curation strategies such as deduplication, filtering, and data mixing at model scales ranging from 412M to 7B parameters. As a baseline, the authors find that model-based filtering is critical for assembling a high-quality training set. Their resulting dataset, DCLM-Baseline, enables training a 7B parameter model from scratch to achieve 64% 5-shot accuracy on MMLU with 2.6T training tokens. This represents a 6.6 percentage point improvement over MAP-Neo (the previous state-of-the-art in open-data LMs), while using 40% less compute. The baseline model is also comparable to Mistral-7B-v0.3 and Llama 3 8B on MMLU (63% and 66%), and performs similarly on an average of 53 NLU tasks, while using 6.6x less compute than Llama 3 8B. These findings emphasize the importance of dataset design for training LMs and establish a foundation for further research on data curation.more » « less
-
A model is considered well-calibrated when its probability estimate aligns with the actual likelihood of the output being correct. Calibrating language models (LMs) is crucial, as it plays a vital role in detecting and mitigating hallucinations of LMs as well as building more trustworthy models. However, standard calibration techniques may not be suited for LM calibration. For instance, post-processing methods such as temperature scaling do not reorder the candidate generations. On the other hand, training-based methods require fine-tuning the entire model, which is impractical for LMs of large scale. We present LITCAB, a lightweight calibration mechanism consisting of a single linear layer that takes the input text representation and predicts a bias term, which is then added to the LM output logits. LITCAB improves model calibration by only adding < 2% of the original model parameters. For evaluation, we construct CAT, a benchmark consisting of eight text generation tasks, covering responses ranging from short phrases to paragraphs. We test LITCAB with Llama2-7B, where it improves calibration across all tasks, reducing the average ECE score by as large as 30%. We further conduct a comprehensive evaluation with multiple popular open-sourced LMs from GPT and LLaMA families, yielding the following key findings: (i) Larger models within the same family exhibit better calibration on tasks with short generation tasks, but not necessarily for longer ones. (ii) GPT-family models show superior calibration compared to LLaMA, Llama2, and Vicuna models, despite having much fewer parameters. (iii) Fine-tuning pretrained model (e.g., LLaMA) with samples of limited purpose (e.g., conversations) may lead to worse calibration, highlighting the importance of fine-tuning setups for calibrating LMs.more » « less
-
Linear RNN architectures, like Mamba, can be competitive with Transformer models in language modeling while having advantageous deployment characteristics. Given the focus on training large-scale Transformer models, we consider the challenge of converting these pretrained models for deployment. We demonstrate that it is feasible to distill large Transformers into linear RNNs by reusing the linear projection weights from attention layers with academic GPU resources. The resulting hybrid model, which incorporates a quarter of the attention layers, achieves performance comparable to the original Transformer in chat benchmarks and outperforms open-source hybrid Mamba models trained from scratch with trillions of tokens in both chat benchmarks and general benchmarks. Moreover, we introduce a hardware-aware speculative decoding algorithm that accelerates the inference speed of Mamba and hybrid models. Overall we show how, with limited computation resources, we can remove many of the original attention layers and generate from the resulting model more efficiently. Our top-performing model, distilled from Llama3-8B-Instruct, achieves a 29.61 length-controlled win rate on AlpacaEval 2 against GPT-4 and 7.35 on MT-Bench, surpassing the best 8B scale instruction-tuned linear RNN model. We also find that the distilled model has natural length extrapolation, showing almost perfect accuracy in the needle-in-a-haystack test at 20x the distillation length.more » « less
