You and the process

With a person
Anna, in-house recruiter
9 years hiring engineers
30 minutes with a real recruiter
They read your CV with you, on a call, and say where the offers are being lost.
Didn’t find what you were looking for? Tell us what to build
Samsung Semiconductor

Intern, ML Compiler (Kernel Auto-Tuning)

  • Office
  • 0-3 years

Salary

Not stated

Similar roles pay $100K - 170K a year · our estimate

AI summary

For members

The whole posting in a few lines. Sign up to read it here and on every role you open.

Sign Up to Read

Description

You'll join a team building the software that runs PyTorch models on a compute-near-memory AI accelerator from Samsung's technology group, the company behind devices from phones to electric vehicles and data centers.

The project is a cost-model-guided auto-tuner for PyTorch and Triton kernels that accounts for where each kernel's data sits. It connects to torch.compile, and you'll measure its speedups on LLM inference.

You'll learn performance tuning for a DMA-driven accelerator without a cache, including search spaces, legality rules and cost models. You'll also work with PyTorch compiler internals (torch.compile, FX, Inductor autotuning) and Triton kernel compilation, and benchmark GEMM and attention kernels on simulators and hardware.

Responsibilities

  • Work with the ML Compiler team, which builds a PyTorch device and compiler backend, a Triton-based kernel compiler and an LLVM back end, alongside the runtime and hardware teams and the PyTorch and Triton open-source communities.
  • Over the summer, own one part of making standard PyTorch models run fast on new memory-centric hardware, from research idea to merged, measured code.
  • Define a search space for Triton kernel settings on the accelerator, covering tile sizes, launch-grid mapping, on-chip buffer budgets and load style.
  • With the compiler engineers, write legality rules that remove configurations which fail to compile or exceed on-chip memory.
  • Build a kernel cost model, calibrated against the simulator and profiler, that ranks candidates so that only the strongest get benchmarked.
  • Create a dependable benchmarking harness and a persistent tuning cache to keep tuning time manageable.
  • Hook the tuner into the torch.compile Inductor autotuning path and Triton's autotuner.
  • Extend tuning to data placement: tune kernels for a given tensor layout and export their costs for a placement pass.
  • Measure kernel-level and end-to-end LLM inference gains, and present results in weekly syncs and a final presentation.
  • Work with PyTorch and Triton developers on upstream autotuning hooks, and deliver documented, tested code.
  • Take on other tasks as assigned.

Requirements

  • Be pursuing a Master's or PhD in Computer Science, Computer Engineering or Electrical Engineering. A PhD is preferred.
  • Have at least one academic quarter or semester left after the internship.
  • Write strong Python and C++, and have written or tuned Triton kernels or worked with PyTorch compiler internals (torch.compile, FX, Inductor).
  • Understand compilers or performance engineering, including IR and passes, cost models, roofline analysis and profiling.
  • Adapt your communication style to the situation and to the varied global norms of colleagues.
  • Learn eagerly, approach problems with curiosity and resilience, and look for data to build understanding.
  • Build working relationships, offer help with humility, and welcome other people's approaches.
  • Come up with new ideas on your own, and adjust quickly when things change.
  • Preferred: experience with auto-tuning or auto-sharding, such as TVM/Ansor, the Halide autoscheduler, Kernel Tuner, Alpa or GSPMD, or with search methods such as Bayesian optimisation or ILP.
  • Preferred: experience with MLIR or LLVM back ends, scratchpad or near-memory architectures, LLM inference kernels, publications, or open-source PyTorch or Triton contributions.

Conditions

This is a summer internship based at the San Jose headquarters, with daily onsite presence under the Flexible Work policy.

You will report to the Director.

The company welcomes applicants of all backgrounds and does not discriminate on protected characteristics.

Benefits

  • The hourly base pay range is $28 to $68 USD for all roles at this level across US locations. Where you fall in that range depends on your work location, job-related knowledge, skills and experience.
  • Medical, dental and vision coverage, plus an inclusive rewards plan that supports your whole self.
  • Relocation stipend if you live more than 50 miles from your new office.
  • Monthly housing stipend to help cover a lease for your summer internship.
  • Charitable giving match, along with regular volunteering opportunities.
  • On-demand wellness apps and paid therapy sessions, wherever you are.
  • Onsite café and gym, plus virtual fitness classes.
  • A flexible environment that lets you balance your time.

Where you’d work

From the office

Relocation offered

About the company

Samsung Semiconductor

Office in San Jose, United States

Also hiring in United States

Your chances

Worth a look before you spend an evening tailoring a CV for it.

  • 21 checks run
  • 2 red flags

Still hiring?

14 checks

No red flags

How crowded?

7 checks

2 red flags

Fits Me

How well does this role fit you?

Answer a few questions or drop your CV, and every role gets a fit score with the reasons, this one first.

  • Your field
  • Level
  • Stack
  • Work model
  • Salary floor
  • Must-haves
Details12 facts · Role, Location, Compensation, Employment
Tech stack
  • Python
  • PyTorch
  • LLMs
Seniority
Intern
Type
Internship
Specialty
ML
Region
United States
Pay period
Annual
Show 6 more factsShow less

Role

Category
Data & Analytics
Specialty
ML
Seniority
Intern
Experience
0-3 years
Tech stack
  • Python
  • PyTorch
  • LLMs

Location

Work model
Office
Region
United States
Office
  • San Jose, United States
Relocation
Offered

Compensation

Salary
Salary by agreement
Pay period
Annual

Employment

Type
Internship

Something wrong with this vacancy?

Similar vacancies

Share this vacancy

What's wrong with it?

The employer never sees who reported.

Reason