Member of Technical Staff - Efficient ML
Embedding VC · San Francisco, CA · 9 mo ago
On-siteEngineeringFull-time
Scope of Work
- Training efficiency: Dataloaders, fusion, activation remat, gradient checkpointing.
- FSDP/ZeRO/tensor+pipeline parallel; NCCL tuning.
- GPU + kernel performance: Nsight profiling, Triton/CUDA kernels, fused ops.
- Flash-attention–style speedups, sequence packing, KV-cache tricks.
Inference Optimization
- Low-latency serving, continuous batching, speculative decoding.
- Quantization (GPTQ/AWQ), distillation, pruning.
Infrastructure & Reliability
- SLURM/K8s multi-node jobs, checkpoint hygiene.
- Determinism, env pinning, GPU failure handling.