Searcharxiv⌕ Search

arXiv subjects

Pranav Sastry

Publications and source records attributed to Pranav Sastry.

2 recordsLinked to original sources

BridgeKV: Large-Model Prefill and Small-Model Decoding for LLM Inference via KV Cache Translation

LLM inference has two phases with different bottlenecks: LLM prefill runs once per request and is compute-intensive. Decoding repeats for every output token and is limited by memory bandwidth. To address the challenges associated with this issue, different solutions have been proposed. All of these solutions use the same model for both prefill and decode phases. We propose BridgeKV, which uses a larger model for prefill and a smaller model of the same family for decoding. A lightweight KV-stitching module translates the larger model's prompt cache into the smaller model's cache layout. It handles differences in cache shape and representation through learned local layer mixing and separate nonlinear maps for keys and values. With both models frozen, we train the module to align their caches and make the smaller model follow the larger model's next-token predictions. A second training stage adds low-rank adapters when further task adaptation is needed. Across six model pairs and five datasets, BridgeKV outperforms fine-tuned small-model baselines with matched trainable parameter budgets in all 30 settings. It closes up to 80% of the quality gap to using the larger model for both phases. On one A40 GPU, Llama-3.1-Instruct-8B to Llama-3.2-Instruct-1B achieves 3.0X faster requests than the larger model alone, with latency close to the smaller model alone. BridgeKV can improve the hardware and system solutions that separate prefill and decode phases across GPUs. A separate serving simulation models request arrivals, queues, and cache transfer. For Llama-3.1-Instruct-70B to Llama-3.2-Instruct-1B, it predicts up to 33.86X faster requests and 27.5X lower energy use than using the larger model for both prefill and decode phases.

cs.CL↗

Fusing Conditional Submodular GAN and Programmatic Weak Supervision

Programmatic Weak Supervision (PWS) and generative models serve as crucial tools that enable researchers to maximize the utility of existing datasets without resorting to laborious data gathering and manual annotation processes. PWS uses various weak supervision techniques to estimate the underlying class labels of data, while generative models primarily concentrate on sampling from the underlying distribution of the given dataset. Although these methods have the potential to complement each other, they have mostly been studied independently. Recently, WSGAN proposed a mechanism to fuse these two models. Their approach utilizes the discrete latent factors of InfoGAN to train the label model and leverages the class-dependent information of the label model to generate images of specific classes. However, the disentangled latent factors learned by InfoGAN might not necessarily be class-specific and could potentially affect the label model's accuracy. Moreover, prediction made by the label model is often noisy in nature and can have a detrimental impact on the quality of images generated by GAN. In our work, we address these challenges by (i) implementing a noise-aware classifier using the pseudo labels generated by the label model (ii) utilizing the noise-aware classifier's prediction to train the label model and generate class-conditional images. Additionally, we also investigate the effect of training the classifier with a subset of the dataset within a defined uncertainty budget on pseudo labels. We accomplish this by formalizing the subset selection problem as a submodular maximization objective with a knapsack constraint on the entropy of pseudo labels. We conduct experiments on multiple datasets and demonstrate the efficacy of our methods on several tasks vis-a-vis the current state-of-the-art methods.

cs.CV↗