Black Forest Labs is a cutting-edge startup pioneering generative image and video models. Our team, which invented Stable Diffusion, Stable Video Diffusion, and FLUX.1, is currently seeking a strong researcher / engineer to work closely with our research team on pretraining and inference optimization.

Role:

• Finding ideal training strategies (parallelism, precision trade-offs) for a variety of model sizes and compute loads
• Profiling, debugging, and optimizing single and multi-GPU operations using tools such as Nsight or stack trace viewers
• Reasoning about the speed and quality trade-offs of quantization for model inference
• Developing and improving low-level kernel optimizations for state-of-the-art inference and training
• Innovating new ideas that bring us closer to the limits of a GPU

Ideal Experiences:

• Being familiar with the latest and the most effective techniques in optimizing inference and training workloads
• Optimizing for both memory-bound and compute-bound operations
• Understanding GPU memory hierarchy and computation capabilities
• Deep understanding of efficient attention algorithms
• Implementing both forward and backward Triton kernels and ensuring their correctness while considering floating point errors
• Using, for example, pybind to integrate custom-written kernels into a PyTorch framework

Nice to have:

• Experience with Diffusion and Autoregressive models
• Experience in low-level CUDA kernel optimizations