- Beijing, China
- Senior
Responsible for the full-link deployment of on-device models, including model conversion, quantization, compilation, runtime integration, operator adaptation, memory management, version compatibility, and phased release. Focus on performance profiling, bottleneck identification, and inference optimization for CPU/GPU/NPU heterogeneous computing environments (Windows/Linux/Android).
Responsibilities
- Responsible for full-link on-device model deployment: model conversion, quantization, compilation, runtime integration, operator adaptation, memory management, version compatibility, and phased release.
- Perform performance profiling, bottleneck identification, and inference optimization for CPU/GPU/NPU heterogeneous computing and Windows/Linux/Android environments.
- Implement token routing, device-cloud task orchestration, low-latency multimodal inference, and privacy-preserving personalization capabilities.
- Implement algorithm optimizations such as FP4/INT4 QAT, speculative decoding, and on-policy distillation to balance accuracy, latency, throughput, memory, and stability.
- Align with Agentic OS framework metrics, building performance regression, stability monitoring, and engineering best practices.
Requirements
- Master's degree or above in Artificial Intelligence, Computer Science, or Electronic Information Technology
- Nice to have: Quantization
- Nice to have: Graph optimization
- Nice to have: Operator optimization
- Nice to have: Memory optimization
