The opportunity
Cerebras Systems builds the world's largest AI chip, 56 times larger than GPUs. This architecture allows Cerebras to deliver industry-leading training and inference speeds; over 10 times faster than GPU-based hyperscale cloud inference services.
What you'll do
Bring up new models: Understand model architectures, load and convert weights, implement supported execution paths, and establish correctness against reference implementations.
Lower models to hardware using MLIR: Develop and extend dialects, graph transformations, lowering passes, and hardware-specific mappings.
Support model operations: Enable attention, matrix multiplication, normalization, positional embeddings, and other operators through compiler and kernel changes.
Optimize execution: Improve operator fusion, tensor layouts, tiling, memory allocation, data movement, and parallel execution.
Optimize inference: Tune prefill and decode performance, KV cache management, batching, and quantization to improve latency, throughput, and memory efficiency.
Validate model quality: Investigate numerical differences and measure the accuracy impact of precision changes and compiler optimizations.
What they're looking for
- Diagnose bottlenecks: Use profiling, execution traces, and hardware counters to identify compute, memory, communication, and runtime limitations.
- Work closely with hardware, compiler, kernel, and runtime teams to deliver: reliable model support and repeatable performance benchmarks.
- Strong programming skills in C++ and Python .
- Hands-on experience bringing up and debugging ML models in PyTorch or a comparable framework.