A senior machine learning systems engineer will optimize LLM inference performance with Python, PyTorch, and AWS Trainium remotely anywhere in Canada. The focus includes decode latency and throughput.

At Capgemini Engineering, this hands-on engineer improves production-grade model serving across Trainium2 and Trainium3. The work covers vLLM on AWS Neuron, Qwen-family models, speculative decoding, multi-token prediction, and reproducible benchmarking against GPUs. You will address long-context and multi-chip scaling.

Key Responsibilities:

- Improve decode latency, throughput, cost, and correctness on Trainium
- Implement speculative decoding and native multi-token prediction
- Design matched tests controlling prompts, concurrency, cache, and warm-up




- Scale long-context serving across multi-chip Trainium systems
- Create runbooks, guides, documentation, and hands-on workshops

Requirements:

- Five-plus years in machine learning or systems engineering
- Two-plus years in production model inference or serving
- Solid Python and PyTorch proficiency modifying model implementations
- Experience with vLLM, TensorRT-LLM, SGLang, or Triton
- Understanding of prefill, decode, batching, parallelism, and cost

Apply machine learning systems expertise to optimize AWS Trainium inference, benchmarking, and serving performance at Capgemini Engineering.

📌 Senior LLM Inference Performance Engineer (Toronto)
🏢 Capgemini Engineering
📍 Toronto