Machine Learning Performance Engineer - CUDA Python

Job Order #: 10540Title: Machine Learning Performance Engineer - CUDA Python

Position/Sheet Notes: MUST HAVE SOME SORT OF RECORDED VIDEO ON WITH NON-USC's FOR INTERNAL USE. This job is open to W2, C2C, USC, GC and other legal status (H1B).

Duration: 6 month contract with the likelihood to extend

Location: Remote but candidates must be willing to travel to different customer sites.

Position Category: Infrastructure

Job Description: *Must be willing to travel 50% of the time

*Must have strong pre-sales abilities i.e. presentation skills, communication skills, etc.

*Must be willing to help train employees and customers

Your part here is optimizing the performance of our models – both training and inference. We care about efficient large-scale training, low-latency inference in real-time systems, and high-throughput inference in research.

Part of this is improving straightforward CUDA, but the interesting part needs a whole-systems approach, including storage systems, networking, and host- and GPU-level considerations. Zooming in, we also want to ensure our platform makes sense even at the lowest level – is all that throughput actually goodput? Does loading that vector from the L2 cache really take that long?

• An understanding of modern ML techniques and toolsets

• The experience and systems knowledge required to debug a training run's performance end to end

• Low-level GPU knowledge of PTX, SASS, warps, cooperative groups, Tensor Cores, and the memory hierarchy

• Debugging and optimization experience using tools like CUDA GDB, NSight Systems, NSight Compute

• Library knowledge of Triton, CUTLASS, CUB, Thrust, cuDNN, and cuBLAS

• Intuition about the latency and throughput characteristics of CUDA graph launch, tensor core arithmetic, warp-level synchronization, and asynchronous memory loads

• Background in Infiniband, RoCE, GPUDirect, PXN, rail optimization, and NVLink, and how to use these networking technologies to link up GPU clusters

• An understanding of the collective algorithms supporting distributed GPU training in NCCL or MPI

• An inventive approach and the willingness to ask hard questions about whether we're taking the right approaches and using the right tools