2 days ago
Paris, FranceSenior
Responsibilities
- Write, tune, and debug GPU kernels using CUDA, Triton, or similar technologies for LLM training and inference.
- Optimize NVIDIA GPU workloads involving memory hierarchy, tensor cores, FP8, and FP4.
- Reduce all-to-all and all-reduce communication costs and overlap communication with computation across NVLink and multi-node systems.
- Profile end-to-end training and rollout workloads to identify compute, memory, and communication bottlenecks.
- Build benchmarks, profiling tools, and regression tests to guide optimization decisions.
- Collaborate with Research Scientists to implement efficient versions of new model architectures.
- Explore low-precision training and serving, speculative decoding, and fused or persistent kernels.
Requirements
- Depth in at least one of GPU kernel development and optimization, low-latency communication and collective operations, or large-scale training infrastructure.
- Ability to reason about memory, bandwidth, and interconnect topology and diagnose performance bottlenecks.
- Proficiency in a systems language such as C++, Rust, or CUDA C, along with Python.
- Strong debugging and profiling skills, including experience with tools such as Nsight.
- Ability to explain design and performance trade-offs to technical and non-technical audiences.
- Preferred experience includes Mixture-of-Experts models, expert parallelism, sparse model inference, NCCL, NVSHMEM, Triton, CUTLASS, Megatron-LM, TorchTitan, DeepSpeed, or GPU-intensive systems in graphics, rendering, simulation, or scientific computing.
- Open-source contributions to ML or systems infrastructure are a plus.
Benefits
- Hybrid work arrangement, indicated by #LI-Hybrid.
About Datadog
Datadog builds a SaaS observability and security platform that monitors infrastructure, applications, logs, and services for engineering and DevOps teams. Its products include infrastructure monitoring, APM, log management, real user monitoring, and cloud security, sold via subscriptions and used across cloud-native and hybrid environments. Founded in 2010 and headquartered in New York City, Datadog is a public company trading on NASDAQ under the ticker DDOG.