مهندس نشر متقدم (Forward Deployed Engineer) لتحسين استدلال AI و GenAI

Job Summary

This role involves working directly with customers and strategic partners to enable, optimize, and deploy GenAI inference workloads on AMD data center CPUs and GPUs. The engineer will focus on bridging AI models and AMD hardware to enhance inference performance.

Responsibilities

• Collaborate with customers to enable and optimize large language and multimodal models on AMD Instinct GPUs, achieving performance and accuracy goals.
• Optimize GenAI inference pipelines using vLLM and SGLang, including techniques like PD disaggregation, expert parallelism, tensor/data parallelism, speculative decoding, and KV cache management.
• Develop and tune GPU kernels with Triton, HIP, and composable kernels for operations such as attention, MoE, GEMM, quantization, and normalization.
• Profile workloads end-to-end, identify bottlenecks across GPU/CPU, and improve throughput and latency against competitive benchmarks.
• Implement advanced data types and methods like FP8, MXFP4, and sparse attention focusing on performance and numerical accuracy.
• Debug runtime, stability, and accuracy issues in production, collaborating with core engineering teams to find root causes and fixes.
• Contribute optimizations to open-source frameworks (vLLM, SGLang, ATOM, AITER) and internal libraries, documenting best practices and delivering production-ready solutions.
• Use AI-assisted coding tools and agentic workflows to speed up development, debugging, optimization, and code review across the inference stack.

Must haves

• BS, MS, or PhD in Computer Science, Electrical Engineering, or related field with 8+ years of relevant experience.
• Expertise in GPU architecture and performance optimization, including hands-on experience with AMD ROCm or CUDA and GPU kernel development.
• Strong skills in Python and C++, with experience in Triton, HIP, or CUDA kernel development and optimization.
• Experience using AI-assisted coding and agentic development tools such as GitHub Copilot, Claude, or Cursor for kernel development and debugging.
• Familiarity with LLM inference frameworks like vLLM and SGLang and serving optimization techniques.
• Understanding of transformer and MoE architectures, quantization, and distributed inference strategies including TP, DP, and EP parallelism.
• Proficiency with profiling and debugging tools like rocProfiler and ROCdebug-agent.

Nice to haves

• Experience optimizing frontier models such as DeepSeek, Qwen, Kimi, MiniMax, or GLM in production.
• Contributions to open-source AI inference frameworks.
• Experience integrating AI-driven automation into engineering workflows and developer tools.
• Strong communication skills for customer-facing interactions and managing technical deliverables on schedule.

We refresh listings regularly, but some roles close early on the source platform.