This is a fully remote position, open to applicants in United States.

📋 Description

• Analyze customer training and inference workloads to pinpoint inefficiencies in data loading, collective communication, memory bandwidth, kernel efficiency, batching, and parallelism layout.

• Set performance benchmarks and determine optimal performance criteria for each hardware generation, beginning with the latest NVIDIA hardware.

• Transform optimization knowledge into agent-executable solutions that can be applied across customer clusters.

• Collaborate with committed-capacity customers, either directly or via operating teams, to fine-tune workloads.

• Take ownership of the internal model serving as a secondary responsibility, which includes cost- and latency-aware routing and managing model version migrations.

• Provide guidance on GPU resource allocation between internal inference tasks and customer workloads.

⛳️ Requirements

• Extensive experience in GPU performance engineering, encompassing profiling, memory and bandwidth behavior, and a kernel-level understanding.

• Proficiency in serving stack technologies such as vLLM, SGLang, or TensorRT-LLM, with expertise in batching, KV cache management, and quantization trade-offs.

• Background in distributed training and inference at scale, covering tensor, pipeline, and data parallelism.

• Familiarity with tuning collective communication frameworks such as NCCL.

• Proficient with profiling tools like Nsight, DCGM, and framework-level profilers.

• Capability to articulate performance insights to customers and encode this information for an agent.

• Desirable: Experience with CUDA kernel development, open-weight model deployment at scale, benchmark design, and prior technical experience in customer-facing roles.