saloni-garg.github.io
Portrait of Saloni Garg
San Jose, CA

Saloni Garg

Staff Machine Learning Engineer, Adobe Firefly

San Jose, California

Saloni Garg is a Staff Machine Learning Engineer at Adobe (Firefly) in San Jose, CA. She leads post-training and inference for multimodal foundation models up to 80B, building SFT, DPO, RLHF and GRPO pipelines, reward models and preference data infrastructure, and the inference systems that serve them: KV cache design, speculative decoding and prefill-decode disaggregation. She authors CUDA kernels in C++ and Triton, works across PyTorch and JAX framework internals, and builds distributed training and automated evaluation infrastructure at scale.

She previously worked at Wayfair as a Senior ML Infrastructure Engineer in Mountain View, CA, and before that as a Founding ML Engineer at Brisko in Bengaluru, India. She is co-author of a textbook on retrieval-augmented generation (De Gruyter, 2026) and recipient of the 2026 CITRIS Award by UC Berkeley.

Focus: post-training and RLHF (SFT, DPO, GRPO, RLVR, reward models, verifiers) / GPU inference optimization (KV cache, speculative decoding, prefill-decode disaggregation, FP8, dynamic batching) / CUDA and Triton kernel authoring in C++ / distributed GPU training (PyTorch, JAX, FSDP, NCCL) / evaluation infrastructure

01Experience

02Selected work

03Open source contributions

Pull requests I have sent upstream, grouped by the bug or feature. Statuses as of November 2026; every title links to the PR itself. Three have landed so far: the Triton interpreter fix (merged), and both torch_tpu fixes (landed on main via Copybara, which closes the GitHub PR without a merged flag).

One bug, five codebases: uniform-reward GRPO groups

In GRPO training, when every response in a group earns the same reward, the advantage should be exactly zero. Float32 group statistics invented a nonzero one instead. I found the bug class, fixed it, and sent the same fix upstream five times:

RLHF reward correctness

Compilers, TPU, and Triton

Verifiers and evaluation tooling

Inference serving

Also open: a listing PR adding memfix to punkpeye/awesome-mcp-servers (PR #16020).

04Publications and IP

05Talks

06Awards and recognition

07Contact

LinkedIn: linkedin.com/in/saloni-garg
Google Scholar: Google Scholar profile
GitHub: github.com/saloni-garg

Email salonigarg2318@gmail.com