Senior ML Engineer - Model Compression

General Motors•Washington, DC

5d•$128,700 - $261,300•Remote

About The Position

The Compression and Parity team in GM’s Autonomous Vehicle (AV) Organization enables repeatable, high-velocity model deployments through principled and automated model compression under strict safety guarantees. We partner closely with model developers and deployment and infra engineers to ship numerically robust, low-latency models to the car, blending rigorous analysis with state-of-the-art methods and our own innovations. Over time, you will help grow and evolve the Compression and Parity function through developing and iterating on quantization and compression strategies for AV models, advancing numerical sensitivity analyses, evolving sensitivity analysis, compression, and parity tooling into a connected, automated flow, and bridging the gap between state-of-the-art model compression research and safety-constrained deployment.

Requirements

Bachelor's degree in Computer Science, Electrical Engineering, Physics, Mathematics, Data Science / ML, or a closely related quantitative field (or equivalent experience)
3+ years of industry experience focused on model optimization and deployment, with significant hands‑on work in neural network quantization / model compression / efficient inference or relevant experience
Strong proficiency in PyTorch and experience with graph‑level representations (e.g., PyTorch FX, ONNX) for capture and manipulation
Background in numerical linear algebra and optimization (conditioning, spectral properties, Jacobians, Hessians) and how they relate to quantization robustness

Nice To Haves

Master's or PhD degree in related quantitative fields
Deep experience with PTQ and QAT, compression frameworks (e.g., PT2E, ModelOpt, torchao) and advanced quantization algorithms (e.g., GPTQ, AWQ, SmoothQuant, QuIP, SparseGPT), as well as with building or extending quantization toolchains
Hands‑on experience designing numerics observability and sensitivity tooling integrated into training or evaluation pipelines (logging ranges, saturation, quant noise, etc.)
A track record of collaboration, including leading cross-functional initiatives and mentoring others
Experience with additional compression techniques such as structured/unstructured pruning, low‑rank decomposition, or knowledge distillation
Experience with perception and/or transformer‑based models (e.g., multi‑view encoders, BEV backbones, detection/segmentation heads, trajectory or planning networks), ideally in AV / ADAS
General understanding of kernel performance and optimization for reduced precision formats
Direct experience with specialized hardware accelerators for edge deployment on tight latency and memory budgets (automotive SoCs, robotics platforms, or similar)
Published research, open‑source contributions, or other notable, intellectually curious work in quantization, compression, or efficient inference

Responsibilities

Developing and iterating on quantization and compression strategies for our AV models, considering model numerical properties, safety and latency constraints, and hardware performance, and partnering on deployment of quantized models to NVIDIA‑based AV hardware with our deployment, compiler, and kernel teams
Advancing our numerical sensitivity analyses to recommend safe compression policies per op/layer/block, using AV-relevant metrics (perception, trajectory, etc.) to evaluate compressed models, and collaborating with Embodied AI to support compression-aware modeling
Evolving sensitivity analysis, compression, and parity tooling into a connected, automated flow that makes low‑precision deployments repeatable, reliable, and low‑touch, with an emphasis on robust execution and maintainability
Bridging the gap between state-of-the-art model compression research and safety-constrained deployment while making strong technical contributions in cross-functional projects and educating others on best practices