About Us:

Fireworks is the platform for specialized intelligence, enabling companies to build, train, and serve AI models tailored to their own data, workflows, and products. Founded by the team behind PyTorch and backed by AMD, Atreides, Benchmark Capital, Index Ventures, Lightspeed, NVIDIA, Sequoia Capital, and TCV, Fireworks powers production AI with hundreds of state-of-the-art open models across text, image, embedding, audio, and multimodal workloads. Today, Fireworks is a Series D company valued at $17.5 billion, bringing together an ambitious, collaborative team that's building the future of enterprise AI.

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 inference

• Analyze and improve latency, throughput, memory usage, and compute efficiency

• Profile system performance to detect and resolve GPU- and kernel-level bottlenecks

• Implement low-level optimizations using CUDA, Triton, and other performance tooling

• Drive 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 efficiency

• Improve support for mixed precision, quantization, and model graph optimization

• Build and maintain performance benchmarking and monitoring infrastructure

• Scale inference and training systems across multi-GPU, multi-node environments

• Evaluate and integrate optimizations for emerging hardware accelerators and specialized runtimes

Minimum Qualifications:

• Bachelor’s degree in Computer Science, Computer Engineering, Electrical Engineering, or equivalent practical experience

• 5+ years of experience working on performance optimization or high-performance computing systems

• Proficiency in CUDA or ROCm and experience with GPU profiling tools (e.g., Nsight, nvprof, CUPTI)

• Familiarity with PyTorch and performance-critical model execution

• Experience with distributed system debugging and optimization in multi-GPU environments

• Deep understanding of GPU architecture, parallel programming models, and compute kernels

Preferred Qualifications:

• Master’s or PhD in Computer Science, Electrical Engineering, or a related field

• Experience 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 infrastructure

• Familiarity with cloud-scale AI infrastructure and orchestration tools (e.g., Kubernetes)

• Background in ML systems engineering or hardware-aware model design

Example projects:

• Implement fully asynchronous low-latency sampling for large language models integrated with structured outputs

• Implement GPU kernels for the new low-precision scheme and run experiments to find optimal speed-quality tradeoff

• Build a distributed router with a custom load-balancing algorithm to optimize LLM cache efficiency

• Define metrics and build harness for finding optimal performance configuration (e.g. sharding, precision) for a given class of model

• Determine and implement in PyTorch an optimal sharding scheme for a novel attention variant

• Optimize communication patterns in RDMA networks (Infiniband, RoCE)

• Debug numerical instabilities for a given model for a small portion of requests when deployed at scale

Why Fireworks?

• 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.