Member of Technical Staff, Performance Optimization

About Us:At Fireworks, we’re building the future of generative AI infrastructure. Our platform delivers the highest-quality models with the fastest and most scalable inference in the industry. We’ve been independently benchmarked as the leader in LLM inference speed and are driving cutting-edge innovation through projects like our own function calling and multimodal models. Fireworks is a Series C company valued at $4 billion and backed by top investors including Benchmark, Sequoia, Lightspeed, Index, and Evantic. We’re an ambitious, collaborative team of builders, founded by veterans of Meta PyTorch and Google Vertex AI.In the last few months alone we launched Fireworks Training, partnered with Microsoft Azure Foundry, and published research straight from our production systems. A few examples of what that looks like in practice:Frontier RL is cheaper than the mega-cluster narrative suggests: we ran cross-region rollouts using 98% sparse weight deltas and published what we learned. (blog)Open source agents with frontier advisors: matching frontier performance through training and harness engineering. (blog)The fine-tuning bottleneck is not the algorithm: integration friction and iteration speed are what actually stall teams; we documented the patterns across dozens of customer engagements. (blog)The Role: We're looking for a Software Engineer focused on Performance Optimization to help push the boundaries of speed and efficiency across our AI infrastructure. In this role, you'll take ownership of optimizing performance at every layer of the stack—from low-level GPU kernels to large-scale distributed systems. A key focus will be maximizing the performance of our most demanding workloads, including large language models (LLMs), vision-language models (VLMs), and next-generation video models.You’ll work closely with teams across research, infrastructure, and systems to identify performance bottlenecks, implement cutting-edge optimizations, and scale our AI systems to meet the demands of real-world production use cases. Your work will directly impact the speed, scalability, and cost-effectiveness of some of the most advanced generative AI models in the world.Key Responsibilities:Optimize system and GPU performance for high-throughput AI workloads across training and inferenceAnalyze and improve latency, throughput, memory usage, and compute efficiencyProfile system performance to detect and resolve GPU- and kernel-level bottlenecksImplement low-level optimizations using CUDA, Triton, and other performance toolingDrive improvements in execution speed and resource utilization for large-scale model workloads (LLMs, VLMs, and video models)Collaborate with ML researchers to co-design and tune model architectures for hardware efficiencyImprove support for mixed precision, quantization, and model graph optimizationBuild and maintain performance benchmarking and monitoring infrastructureScale inference and training systems across multi-GPU, multi-node environmentsEvaluate and integrate optimizations for emerging hardware accelerators and specialized runtimesMinimum Qualifications:Bachelor’s degree in Computer Science, Computer Engineering, Electrical Engineering, or equivalent practical experience5+ years of experience working on performance optimization or high-performance computing systemsProficiency in CUDA or ROCm and experience with GPU profiling tools (e.g., Nsight, nvprof, CUPTI)Familiarity with PyTorch and performance-critical model executionExperience with distributed system debugging and optimization in multi-GPU environmentsDeep understanding of GPU architecture, parallel programming models, and compute kernelsPreferred Qualifications:Master’s or PhD in Computer Science, Electrical Engineering, or a related fieldExperience optimizing large models for training and inference (LLMs, VLMs, or video models)Knowledge of compiler stacks or ML compilers (e.g., torch.compile, Triton, XLA)Contributions to open-source ML or HPC infrastructureFamiliarity with cloud-scale AI infrastructure and orchestration tools (e.g., Kubernetes)Background in ML systems engineering or hardware-aware model designExample projects:Implement fully asynchronous low-latency sampling for large language models integrated with structured outputsImplement GPU kernels for the new low-precision scheme and run experiments to find optimal speed-quality tradeoffBuild a distributed router with a custom load-balancing algorithm to optimize LLM cache efficiencyDefine metrics and build harness for finding optimal performance configuration (e.g. sharding, precision) for a given class of modelDetermine and implement in PyTorch an optimal sharding scheme for a novel attention variantOptimize communication patterns in RDMA networks (Infiniband, RoCE)Debug numerical instabilities for a given model for a small portion of requests when deployed at scaleWhy Fireworks AI?Solve Hard Problems: Tackle challenges at the forefront of AI infrastructure, from low-latency inference to scalable model serving.Build What’s Next: Work with bleeding-edge technology that impacts how businesses and developers harness AI globally.Ownership & Impact: Join a fast-growing, passionate team where your work directly shapes the future of AI—no bureaucracy, just results.Learn from the Best: Collaborate with world-class engineers and AI researchers who thrive on curiosity and innovation.Fireworks AI is an equal-opportunity employer. We celebrate diversity and are committed to creating an inclusive environment for all innovators.