Senior Machine Learning Engineer, LLM Inference Optimization

As a Senior Machine Learning Engineer on the Applied AI team at Nebius Token Factory, you will own model and endpoint optimization from model artifacts through production deployment. Your work will span model internals, quantization and model compression, speculative decoding, KV-cache optimization, inference engines, serving architecture, and benchmarking to improve latency, throughput, memory efficiency, GPU utilization, and cost per token while preserving quality and reliability.

You will work with kernel and platform engineers to investigate serving problems, compare configurations, and resolve performance and quality regressions. You will evaluate model- and engine-level optimizations together with distributed inference designs, including request routing, scheduling, prefill/decode coordination, and multi-node GPU execution. Your improvements will be validated through reproducible benchmarks and safe production rollouts under real-world workloads.

• Own optimization projects for specific model families, customer endpoints, or serving backends.
• Compare inference engines and recommend serving configurations suited to individual workloads.
• Diagnose model-quality and performance regressions during production rollouts.
• Improve LLM and VLM endpoint latency, throughput, memory efficiency, GPU utilization, quality, and cost per token.
• Deploy, configure, benchmark, and extend engines such as vLLM, SGLang, TensorRT-LLM, Triton Inference Server, and NVIDIA Dynamo.
• Develop production model-compression workflows covering quantization, quantization-aware training, distillation, low-bit serving, and accuracy recovery.
• Implement or integrate speculative decoding, draft-model approaches, KV-cache optimization, prefix caching, chunked prefill, continuous batching, and disaggregated prefill/decode serving. For prefill–decode disaggregation (PDD), evaluate KV-cache transfer, worker placement, and capacity balancing to determine which workloads benefit from the architecture.
• Design and improve LLM request routers and scheduling policies that balance worker utilization, queueing, request characteristics, and KV-cache locality while meeting latency and reliability targets.
• Scale dense and mixture-of-experts inference across multiple GPU nodes. Select and tune tensor, pipeline, data, and expert parallelism—including wide expert parallelism (WideEP)—with attention to hardware topology, expert load balance, and communication overhead.
• Build reproducible benchmark harnesses measuring TTFT, TPOT, tokens per second per GPU, p95/p99 latency, GPU memory, reliability, and cost per token. Validate architecture choices under representative traffic and consistent GPU budgets.
• Collaborate with GPU kernel and platform engineers to trace bottlenecks across model code, kernels, runtime, scheduler, gateway, and cluster layers.
• Produce design documents, performance reports, rollout plans, and customer-facing technical explanations.

• Strong engineering skills in Python and PyTorch.
• Hands-on experience deploying or optimizing LLM, VLM, or high-throughput transformer inference.
• Practical knowledge of at least one modern inference stack, such as vLLM, SGLang, TensorRT-LLM, Triton Inference Server, NVIDIA Dynamo, Ray Serve, KServe, or an equivalent internal system.
• A strong understanding of transformer inference bottlenecks involving KV cache, attention, memory bandwidth, batching, parallelism, and long-context serving.
• Ability to quantify trade-offs among latency, throughput, quality, utilization, and cost.
• Experience designing or optimizing distributed inference system architecture, with hands-on work in one or more areas such as request routing, distributed scheduling, PDD, multi-node inference, or expert parallelism.
• Strong communication and collaboration skills across research, kernel, infrastructure, product, and customer teams.

• Experience with quantization-aware training, post-training quantization, FP8, INT8, INT4, NVFP4, MXFP4, AWQ, GPTQ, or SmoothQuant.
• Experience with distillation, speculative decoding, EAGLE, Medusa, multi-token prediction, or related inference acceleration methods.
• Experience supporting agentic workloads involving tool calling, structured outputs, streaming APIs, high concurrency, and multi-step orchestration.
• Familiarity with CUDA or Triton; the role does not require kernel engineering to be the candidate’s primary specialization.
• Contributions to open-source projects such as vLLM, SGLang, TensorRT-LLM, FlashInfer, LMCache, PyTorch, Triton, Ray, or KServe.
• Implementation experience with cache-aware request routing, PDD, or WideEP, including diagnosing communication bottlenecks and load imbalance.
• Familiarity with GPU communication libraries and interconnects, such as NCCL, NVLink, InfiniBand, or RoCE.