Bo Yuan

Cheng Luo, Zefan Cai, Hanshi Sun et al.

Transformer-based large language models (LLMs) demonstrate impressive performance in long context generation. Extending the context length has disproportionately shifted the memory footprint of LLMs during inference to the key-value cache (KV cache). In this paper, we propose HEADINFER, which offloads the KV cache to CPU RAM while avoiding the need to fully store the KV cache for any transformer layer on the GPU. HEADINFER employs a fine-grained, head-wise offloading strategy, maintaining only selective attention heads KV cache on the GPU while computing attention output dynamically. Through roofline analysis, we demonstrate that HEADINFER maintains computational efficiency while significantly reducing memory footprint. We evaluate HEADINFER on the Llama-3-8B model with a 1-million-token sequence, reducing the GPU memory footprint of the KV cache from 128 GB to 1 GB and the total GPU memory usage from 207 GB to 17 GB, achieving a 92% reduction compared to BF16 baseline inference. Notably, HEADINFER enables 4-million-token inference with an 8B model on a single consumer GPU with 24GB memory (e.g., NVIDIA RTX 4090) without approximation methods.

Weiqi Ji, Bo Yuan, Ciyue Shen et al.

Data-driven dynamic models of cell biology can be used to predict cell response to unseen perturbations. Recent work (CellBox) had demonstrated the derivation of interpretable models with explicit interaction terms, in which the parameters were optimized using machine learning techniques. While the previous work was tested only in a single biological setting, this work aims to extend the range of applicability of this model inference approach to a diversity of biological systems. Here we adapted CellBox in Julia differential programming and augmented the method with adjoint algorithms, which has recently been used in the context of neural ODEs. We trained the models using simulated data from both abstract and biology-inspired networks, which afford the ability to evaluate the recovery of the ground truth network structure. The resulting accuracy of prediction by these models is high both in terms of low error against data and excellent agreement with the network structure used for the simulated training data. While there is no analogous ground truth for real life biological systems, this work demonstrates the ability to construct and parameterize a considerable diversity of network models with high predictive ability. The expectation is that this kind of procedure can be used on real perturbation-response data to derive models applicable to diverse biological systems.

Yu Wang, Weikang Qian, Shuchang Zhang et al.

Computational inference of causal relationships underlying complex networks, such as gene-regulatory pathways, is NP-complete due to its combinatorial nature when permuting all possible interactions. Markov chain Monte Carlo (MCMC) has been introduced to sample only part of the combinations while still guaranteeing convergence and traversability, which therefore becomes widely used. However, MCMC is not able to perform efficiently enough for networks that have more than 15~20 nodes because of the computational complexity. In this paper, we use general purpose processor (GPP) and general purpose graphics processing unit (GPGPU) to implement and accelerate a novel Bayesian network learning algorithm. With a hash-table-based memory-saving strategy and a novel task assigning strategy, we achieve a 10-fold acceleration per iteration than using a serial GPP. Specially, we use a greedy method to search for the best graph from a given order. We incorporate a prior component in the current scoring function, which further facilitates the searching. Overall, we are able to apply this system to networks with more than 60 nodes, allowing inferences and modeling of bigger and more complex networks than current methods.