Oferta pracy od partnera
NapiszOpis oferty pracy
The Role
We are seeking a senior AI engineer passionate about performance optimization to join our team and revolutionize our inference systems. You will be at the heart of technical innovation, transforming the latest research advances into ultra-high-performance production solutions.
Core Mission
Your mission will be to radically optimize the performance of our systems by leveraging and improving current leading inference technologies. You will work on cutting-edge technical challenges, reducing latency and maximizing throughput to serve millions of users.
Key Responsibilities
Analysis and Benchmarking
Analyze performance bottlenecks (CPU scheduling, GPU memory, kernel efficiency) - Establish detailed performance metrics (TTFT, tokens/s, P99 latency, GPU utilization) - Design and implement comprehensive benchmarks comparing SOTA inference solutions performance
Document trade-offs between different optimization approaches
Inference Systems Optimization
Optimize current inference engines/frameworks performance for our specific use cases - Implement advanced techniques: RadixAttention, PagedAttention, continuous batching - Develop optimized CUDA kernels (Triton, FlashInfer integration)
Integrate torch.compile and CUDA graphs to maximize performance
Optimize KV cache management and batching strategies
Research Paper Implementation
Transform the latest academic innovations into production code
Implement optimization techniques like MLA (Multi-head Latent Attention) - Adapt MoE (Mixture of Experts) architectures for efficient inference
Integrate model-specific optimizations (DeepSeek, Llama, etc.)
Infrastructure and Scalability
Architect distributed multi-GPU solutions with tensor parallelism
Optimize GPU fleet utilization (H100, H200, ...)
Implement advanced monitoring and profiling systems
Develop debugging tools to identify performance issues
Desired Profile
Essential Technical Skills
Deep expertise in PyTorch and NVIDIA ecosystem (CUDA, NCCL, cuDNN)
Mastery of inference frameworks: SGLang, vLLM, Dynamo, or equivalents
Solid experience (5+ years) in ML systems optimization in production
Practical knowledge of Transformer architectures and attention techniques
Skills in GPU programming (CUDA, Triton) and kernel optimization
Advanced Technical Skills (Strong Plus)
Experience with inference optimization techniques:
Quantization (INT8, INT4, FP8)
KV cache optimization (MQA, GQA, MLA)
Speculative decoding, multi-token prediction
Structured generation and constrained decoding
Knowledge of frameworks: FlashAttention, FlashInfer, xFormers - Experience with high-performance distributed systems
Contributions to open-source ML inference projects
Personal Qualities
Passion for optimization and performance
Ability to read and implement complex research papers
Excellent analytical and problem-solving skills
Autonomy and ability to work on unstructured problems
Clear communication of technical results
What We Offer
Technical Impact
Work on systems serving billions of tokens per day
Access to latest GPUs (H100, H200) and compute resources - Direct collaboration with research teams
Open-source contributions and technical publications
Work Environment
Team of experts passionate about performance
Culture of innovation and technical experimentation
Flexibility and autonomy in technical approaches
Continuous training on latest advances
Compensation Package
- Competitive salary aligned with expertise
- Significant equity
- Conference and training budget
- Cutting-edge hardware
Key Technologies
Inference frameworks: SGLang, vLLM, TensorRT-LLM
GPU optimization: CUDA, Triton, FlashInfer, FlashAttention
Deep Learning: PyTorch, torch.compile
Architectures: Transformers, MoE, MLA, attention variants
Infrastructure: Multi-GPU, tensor parallelism, distributed systems
How to Apply
Send your resume along with examples of optimization projects you have completed. We particularly value:
Open-source contributions to inference projects
Benchmarks or performance analyses you have conducted
Implementations of innovative optimization techniques
