Sequoia GitHub repository (tree-based speculative decoding, Infini-AI-Lab)

README describes Sequoia as a 'scalable and robust tree-based speculative decoding algorithm', paper arXiv 2402.12374. CRITICAL: the README explicitly redirects all current work to a refactored successor repo, UMbreLLa (https://github.com/Infini-AI-Lab/UMbreLLa), offering up-to-date models (Llama3, Qwen, DeepSeek), AWQ support, and Gradio/API/CLI chatbots. So this repo is effectively a frozen research snapshot. Environment pins old versions: torch 2.1.2, transformers 4.36.2, accelerate 0.26.1, datasets 2.16.1, CUDA 12.1 wheels. Evaluation requires per-hardware 'growmaps' (tree structures) with prepared growmaps for A100 and L40; tests are separate scripts (testbed.py for stochastic, testbed_greedy.py for greedy, test_specinfer.py for specinfer sampling, test_accept.py for preparing the accepting-rate vector). Requires --M at least '#tree + 256' (384 suggested). Dataset restricted to cnn, openwebtext, c4. Seed fixed at 17 by default for exact reproduction. Supported models Llama family only (Llama2, Sheared-LLaMA, Vicuna, TinyLlama). API metadata 2026-10-04: created 2024-02-29, last push 2025-01-28, 376 stars, 36 forks, 10 open issues, NO LICENSE DETECTED by the GitHub API (license field null - a genuine caveat for a research repo), not archived.