JOBSEARCHER

Machine Learning Engineer

We are seeking a high-caliber Machine Learning Engineer (Mid to Senior level) specializing in lower-level GPU acceleration, hardware-aware algorithm design, and custom kernel development. In this role, you will sit at the intersection of deep learning theory and high-performance hardware execution, directly driving latency reduction, memory optimization, and throughput scaling for modern Large Language Models (LLMs) and multimodal architectures. You will write high-performance CUDA, Triton, and C++ extensions to break past standard framework bottlenecks, optimizing compute graphs, memory access patterns, and distributed inference/training pipelines across state-of-the-art GPU clusters (NVIDIA Hopper/Blackwell, AMD Instinct).
Key Responsibilities: Custom Kernel Development: Design, implement, and profile custom GPU kernels in CUDA and OpenAI Triton to optimize core LLM primitives (e.g., FlashAttention variants, fused linear-activation layers, MoE routing, and quantized matrix multiplications).Hardware-Aware LLM Optimization: Drive low-level optimizations for modern attention mechanisms, speculative decoding, KV-cache management (Paged Attention), FP8/INT4/FP4 quantization, and structured sparsity. Framework Integration: Seamlessly integrate high-performance kernels into mainstream AI frameworks and inference engines (PyTorch, vLLM, TensorRT-LLM, DeepSpeed, Megatron-LM).Profiling & Bottleneck Identification: Utilize deep-dive profiling tools (NVIDIA Nsight Systems, Nsight Compute, PyTorch Profiler, Rocprof) to identify hardware stall conditions, memory bandwidth limitations, register pressure, and warp divergence. Distributed Execution & Communication: Optimize inter-GPU and inter-node communication primitives (NCCL/RCCL) for Tensor, Pipeline, and Sequence Parallelism to reduce latency overhead in large-scale cluster deployments. Benchmark & CI/CD Pipelines: Establish automated performance regression suites, profiling pipelines, and accuracy checks to ensure kernel optimizations preserve numerical precision while maximizing TFLOPS utilization. Required Qualifications & Experience
Education: Bachelor’s, Master’s, or Ph.D. in Computer Science, Computer Engineering, Electrical Engineering, or a related quantitative field.
Experience: 3+ years (Mid) or 6+ years (Senior) of hands-on experience in high-performance computing (HPC), deep learning performance engineering, or low-level systems engineering. Programming Mastery: Fluent in C++17/20 and Python, with strong familiarity with PyTorch C++ extensions/PyBind11.GPU Programming: Proven experience writing production-grade CUDA C/C++ and/or OpenAI Triton code. LLM Architecture Familiarity: Deep domain knowledge of Transformer mechanics (multi-head/grouped-query attention, rotary position embeddings, mixture-of-experts, KV-caching).Profiling Tooling: Hands-on expertise using NVIDIA Nsight Systems (nsys) and Nsight Compute (ncu) to analyze roofline models, shared memory usage, and memory bandwidth occupancy.
Preferred / Nice-to-Have Qualifications: Open-Source Contributions: Track record of contributions to vLLM, TensorRT-LLM, FlashAttention, PyTorch, DeepSpeed, or Triton. Quantization & Numerics: Hands-on experience with post-training quantization (AWQ, GPTQ) and FP8/FP4 low-precision GEMMs. Alternative Silicon: Familiarity with AMD ROCm/HIP, custom ASICs (Google TPU, AWS Trainium), or custom hardware accelerators. Distributed Training/Inference: Experience optimizing large-scale distributed runs using Megatron-LM or DeepSpeed ZeRO-3.