Intern, ML Compiler (Kernel Auto-Tuning)

Build a cost-model-guided auto-tuning framework for PyTorch and Triton kernels on Samsung’s compute-near-memory AI accelerator.

Originally posted 13 November 2026 by the employer.

About the role

This internship role focuses on building a cost-model-guided auto-tuning framework for PyTorch and Triton kernels on Samsung’s compute-near-memory AI accelerator. You will integrate the tuner with torch.compile Inductor autotuning paths and measure performance gains on LLM inference kernels.

What you'll do

• Design a search space for Triton kernel configurations on the accelerator, covering tile sizes, launch-grid mapping, on-chip buffer budgets, and load style.
• Build legality rules with compiler engineers to prune configurations that fail compilation or exceed on-chip memory limits.
• Develop a kernel cost model calibrated against simulation and profiling data to rank candidate configurations.
• Build a benchmarking harness and persistent tuning cache, integrating the tuner with Triton's autotuner and the torch.compile Inductor autotuning path.
• Extend tuning to data placement to optimize kernels for specific tensor layouts and export tuned costs for placement passes.
• Measure kernel-level and end-to-end LLM inference gains on simulators and hardware, sharing findings in presentations and technical syncs.

What you'll need

• Currently pursuing a Master's or PhD in Computer Science, Computer Engineering, or Electrical Engineering, with at least 1 academic quarter/semester remaining.
• Strong Python and C++ programming skills.
• Experience writing or tuning Triton kernels or working with PyTorch compiler internals such as torch.compile, FX, or Inductor.
• Grounding in compiler fundamentals and performance engineering, including IR and passes, cost models, roofline analysis, and profiling.

Nice to have

• PhD in Computer Science, Computer Engineering, or Electrical Engineering.
• Experience with auto-tuning or auto-sharding frameworks (e.g., TVM/Ansor, Halide autoscheduler, Kernel Tuner, Alpa, GSPMD) or search methods such as Bayesian optimization or ILP.
• Experience with MLIR/LLVM back ends, scratchpad or near-memory architectures, or LLM inference kernels.
• Publications or open-source PyTorch/Triton contributions.

More semiconductor jobs: Samsung Semiconductor jobs · Embedded software jobs · San Jose jobs