ML Engineer (Training Infra), Foundational Models

Own and optimize the distributed training infrastructure for Sarvam's foundational AI models, including large GPU clusters and custom kernel development.

Design and implement advanced parallelism strategies (data, tensor, pipeline, sequence, expert) tailored to various model architectures and scales.

Ensure reliability and fault tolerance of long-running training jobs, addressing performance bottlenecks and operational stability to prevent costly failures.

Bachelor's or Master's in Computer Science or closely related technical field, or equivalent experience.

3+ years experience in ML training infrastructure or large-scale distributed systems; exceptional early-career candidates with strong systems background considered.

Hands-on experience with distributed training frameworks such as Megatron-LM, DeepSpeed, FSDP, NeMo; experience managing real pretraining runs including on-call responsibilities.

Deep knowledge of GPU architecture and CUDA programming; proficiency with GPU profiling tools and strong PyTorch internals expertise.

Experienced with designing and tuning high-performance distributed training systems at scale, including custom CUDA/Triton kernel development.

Background in training very large models (10B+ parameters) on extensive GPU clusters (1000+ GPUs) with strong operational ownership.

Proven contributions to open-source training infrastructure projects and familiarity with cluster orchestration/job scheduling systems at scale.