3 days ago
Shanghai, ChinaMid Level
Responsibilities
- Deploy and optimize LLM and multimodal inference workloads on Qualcomm Cloud AI accelerator platforms from single accelerators through multi-node and rack-scale configurations.
- Profile and improve prefill and decode metrics including TTFT, TPOT, inter-token latency, tokens per second, throughput per user, and performance per watt.
- Analyze and optimize compute, accelerator memory, host memory, KV cache, interconnect, networking, and software-stack bottlenecks.
- Apply quantization, batching, KV-cache management, paged attention, context-length tuning, speculative decoding, and kernel, runtime, graph, and compiler optimizations.
- Optimize distributed inference using tensor, pipeline, data, context, and expert parallelism and characterize collective communication performance.
- Design reproducible benchmarks and stress tests, identify saturation points and regressions, and establish performance baselines and production-readiness reports.
- Collaborate with runtime, compiler, firmware, driver, platform, networking, and model teams to resolve cross-layer performance issues.
- Build Python and Bash profiling, benchmarking, automation, and data-analysis tools and document optimization recipes and reference configurations.
- Work directly with customers during evaluations, PoCs, design wins, and production deployments, translating workload requirements into engineering actions.
Requirements
- Master's degree or higher in Computer Science, Electrical or Computer Engineering, AI/ML, or a related technical field, with strong hands-on systems and performance engineering experience.
- Bachelor's degree plus 4 or more years, master's degree plus 3 or more years, or PhD plus 2 or more years of relevant Software Applications Engineering, software development, or related experience.
- Strong understanding of Transformer and LLM inference, including attention, MHA, GQA, MQA, feed-forward networks, KV cache, and prefill versus decode behavior.
- Hands-on experience deploying and profiling LLM workloads with Python and at least one major ML framework such as PyTorch.
- Experience with quantization, batching, memory optimization, kernel and runtime tuning, and graph or compiler optimization.
- Understanding of distributed AI, multi-accelerator execution, collective communication, Scale-Up and Scale-Out fabrics, and parallelism strategies.
- Strong Linux, Python, Bash, container, orchestration, automation, debugging, and performance-data analysis skills.
- At least 2 years of experience with a programming language such as C, C++, Java, or Python and at least 1 year of debugging experience.
- Preferred experience with vLLM, TensorRT-LLM, SGLang, Hugging Face Transformers, DeepSpeed, Llama, Qwen, DeepSeek, Mixtral, MoE architectures, RDMA/RoCE/InfiniBand, MLPerf Inference, observability, CI/CD performance regression testing, capacity planning, and customer AI accelerator deployments.
Categories
Solutions Engineering
About Qualcomm
Qualcomm is a public semiconductor company headquartered in San Diego, founded in 1985, that designs and sells wireless chipsets and platforms for mobile devices, automotive, IoT, and networking, notably the Snapdragon application processors and 5G modems. It also licenses a large portfolio of cellular patents to device makers, generating revenue alongside chip sales; its technology underpins many Android smartphones and emerging automotive and edge-compute systems, and it trades on NASDAQ as QCOM.