校招 北京-LPU Operator / Compiler / Performance Engineer

岗位职责 / Responsibilities

Build LLM inference software for the LPU (Language Processing Unit). Experience in any ONE of the following areas is sufficient: operator development, compiler development, or performance optimization.

Core Responsibilities

• Develop and optimize inference operators/kernels such as GEMM and attention; tune tiling, data layouts, and data movement; validate correctness and performance across tensor shapes and precisions.
• Implement and optimize lowering from tensor graphs/IR to LPU ISA: operator fusion, on-chip buffer allocation, instruction scheduling, and code generation.
• Run microbenchmarks and end-to-end profiling to identify compute, memory-access, and synchronization bottlenecks; validate first-token latency, per-token latency, and throughput gains across input lengths and batch sizes.
• Co-design with architecture and runtime teams, evaluate trade-offs among compiler scheduling, operator implementations, and hardware resource usage, and add key workloads to regression tests.

任职要求 / Requirements

Requirements

• Recent/new PhD in CS, EE, Automation, or related fields. Exceptional final-year Master’s students with strong relevant research or engineering track records are also welcome.
• Proficient in C/C++ and Python scripting; experienced in Linux-based development and debugging.
• Strong computer architecture fundamentals: parallelism, memory hierarchy, and data dependencies.
• Able to independently break down problems, design reproducible experiments, and communicate technical contributions clearly.

Preferred

• Experience with LLVM, MLIR, TVM, or Triton for compiler backends, code generation, or operator optimization; or low-level programming with CUDA or other accelerators.
• LLM inference optimization experience, e.g., prefill/decode profiling, KV cache management, low-precision computation, or cross-chip communication.

What We Offer

• Mentorship from architecture, compiler, and runtime experts.
• End-to-end ownership of a core LPU inference operator or compiler optimization, from prototype to regression validation.
• Access to state-of-the-art foundation models; leverage LLM agents to accelerate R&D, generate tests, and validate results through correctness tests and benchmarks.