Principal GPU Inference Optimization Engineer

digitalocean98; Stati Uniti; Ibrido; USD 250.000 - 312.000; Tempo pieno

Vantaggi offerti da questo lavoro

Hybrid work model

Equity compensation

Conference & education reimbursement

Flexible time off

Descrizione del lavoro

digitalocean98 is seeking a principal-level engineer to own the model-optimization discipline across a heterogeneous GPU fleet powering production inference. The role spans quantization, kernel authorship, and parallelism topology across NVIDIA and AMD stacks.

You will set technical direction, drive spec decoding and upstream contributions, mentor engineers, and represent the group in conferences and customer discussions. Hybrid work is offered with strong compensation.

Competenze

• Twelve or more years in performance-critical systems with extensive production ML inference.
• Deep knowledge of GPU architecture and inference performance modeling.
• Hands-on kernel-level experience on at least one vendor stack and cross-stack ability.
• Quantization expertise with accuracy-budget judgment.
• Familiarity with major serving engines (e.g., vLLM, SGLang, TensorRT-LLM).

Mansioni

• Set the technical direction for model optimization across the fleet.
• Own end-to-end quantization, including FP8, FP4/MXFP4, INT8, and weight-only schemes.
• Drive kernel-level perf for modern architectures and various attention variants.
• Lead speculative decoding efforts and tuning of acceptance rates.
• Author and tune kernels in CUDA/Triton/CUTLASS and HIP/CK stacks.
• Choose parallelism layouts per model, GPU family, and traffic shape.
• Build benchmarking/regression infra to validate TTFT/ITL throughput and accuracy.
• Make AMD a first-class inference target and contribute upstream projects.
• Mentor senior/staff engineers and engage in external communities.

Conoscenze

GPU architecture

Performance optimization

Quantization

CUDA/CUTLASS

Triton

HIP/CK

Serving engines

Strumenti

CUDA

CUTLASS

Triton

HIP

Composable Kernel

Descrizione del lavoro

digitalocean98 is seeking a principal-level engineer to own the model-optimization discipline across a heterogeneous GPU fleet powering production inference. The role spans quantization, kernel authorship, and parallelism topology across NVIDIA and AMD stacks.

You will set technical direction, drive spec decoding and upstream contributions, mentor engineers, and represent the group in conferences and customer discussions. Hybrid work is offered with strong compensation.