Hongming Huang

Jiongjiong Gu, Jianfeng Wang, Zidong Han et al.

Modern AI serving increasingly relies on NPUs for conventional inference and large language model serving. However, current NPU deployments commonly expose physical devices directly to applications, which limits runtime control over scheduling and makes it difficult to adapt execution to phase-level workload behavior. This limitation is particularly evident in LLM serving, where the prefill phase is compute-intensive while the decode phase is often constrained by memory bandwidth and KV-cache accesses. Static prefill-decode (PD) disaggregation reduces phase interference, but can introduce resource imbalance and unnecessary data movement. We present FlexNPU, a transparent user-space virtualization layer for Ascend NPUs. FlexNPU interposes on AscendCL APIs and routes NPU operations through per-device daemons, decoupling unmodified from physical NPU devices without modifying model code, AI frameworks, or NPU drivers. This runtime boundary allows FlexNPU to virtualize NPU objects, control operator dispatch, and support phase-aware scheduling for LLM serving. In particular, FlexNPU enables dynamic PD co-location, which adapts scheduling between prefill and decode according to their complementary resource characteristics. We implement FlexNPU on Huawei Ascend NPUs and evaluate it with typical LLM workloads. Compared with direct NPU passthrough, FlexNPU introduces no measurable inference overhead and slightly improves throughput in some scenarios. On a 384-card Ascend 910C deployment of DeepSeek-R1, FlexNPU improves throughput over static PD disaggregation by 5.15% and 26.33%. On Qwen2.5-7B, compared with static PD co-location, FlexNPU maintains comparable throughput while reducing TTFT by over 92% across tested workloads with nearly unchanged TPOT. These results show that transparent NPU virtualization is a practical substrate for efficient and responsive LLM serving.

Hongming Huang, Joe Suzuki

We propose a globally optimal Bayesian network structure discovery algorithm based on a progressively leveled scoring approach. Bayesian network structure discovery is a fundamental yet NP-hard problem in the field of probabilistic graphical models, and as the number of variables increases, memory usage grows exponentially. The simple and effective method proposed by Silander and Myllymäki has been widely applied in this field, as it incrementally calculates local scores to achieve global optimality. However, existing methods that utilize disk storage, while capable of handling networks with a larger number of variables, introduce issues such as latency, fragmentation, and additional overhead associated with disk I/O operations. To avoid these problems, we explore how to further enhance computational efficiency and reduce peak memory usage using only memory. We introduce an efficient hierarchical computation method that requires only a single traversal of all local structures, retaining only the data and information necessary for the current computation, thereby improving efficiency and significantly reducing memory requirements. Experimental results indicate that our method, when using only memory, not only reduces peak memory usage but also improves computational efficiency compared to existing methods, demonstrating good scalability for handling larger networks and exhibiting stable experimental results. Ultimately, we successfully achieved the processing of a Bayesian network with 28 variables using only memory.

Hongming Huang, Peng Cheng, Hong Xu et al.

We advocate that simulation based on offline profiling is a promising approach to better understand and improve the complex ML systems. Our approach uses operation-level profiling and dataflow based simulation to ensure it offers a unified and automated solution for all frameworks and ML models, and is also accurate by considering the various parallelization strategies in a real system.