Nvidia · posted · checked
Deep Learning Model and System Optimization Engineering Intern – 2027
C++PythonPytorchOn-site · ShanghaiIntern
Who can apply
On-site · Shanghai
Location: 2 Locations / China, Shanghai
We are now looking for a Deep Learning Performance Engineer Intern to join our team. Large language models are transforming how people interact with computing, while placing unprecedented demands on inference performance, memory efficiency, scalability, and deployment cost.
Our team develops GPU-accelerated software for deep learning inference. We work across model architectures, algorithms, numerical representations, inference systems, and GPU architectures to continuously improve the performance and deployment capabilities of large language models on NVIDIA GPUs. Our work contributes to inference frameworks such as TensorRT-LLM and helps bring state-of-the-art models to production efficiently. We are looking for outstanding interns who are passionate about large language model inference, low-precision computing, GPU performance optimization, and hardware-software co-design.
What you’ll be doing:
Analyze and optimize inference performance across model architecture, numerical precision, GPU kernels, memory systems, and end-to-end serving frameworks.
Develop low-bit quantization techniques for model weights, activations, attention operands, and KV caches.
Explore numerical formats and mixed-precision strategies involving FP8, NVFP4, MXFP4, and other emerging low-precision formats while preserving model accuracy and capabilities.
Design and evaluate techniques for KV cache compression, sparsification, hierarchical storage, and reuse across inference requests.
Optimize the inference of model architectures including MHA, GQA, MQA, MLA, sliding-window attention, sparse attention, linear attention, and state space models.
Investigate model-system co-optimization techniques such as speculative decoding, state reuse, and computation reuse.
Build performance models and optimization strategies for long-context inference and agentic workloads, with a focus on throughput, time to first token, inter-token latency, memory efficiency, and deployment cost.
Collaborate closely with model, framework, kernel, library, system software, and GPU architecture teams to influence the design of future models, software platforms, and hardware architectures.