eduardoslonski/telescope
Scalable high-performance async RL post-training framework for LLMs with real-time observability dashboard
What's novel
Scalable high-performance async RL post-training framework for LLMs with real-time observability dashboard
Code Analysis
10 files read · 3 roundsA production-grade RL post-training framework that coordinates vLLM inference servers and FSDP/Megatron trainers on Ray clusters to train LLMs with GRPO, CISPO, GSPO, SAPO, and other policy gradient algorithms across single-turn and multi-turn environments.
Strengths
Genuinely deep implementation of multiple RL algorithms (GSPO's geometric-mean ratio trick, DR-GRPO's length-bias removal, ICE-POP TIS) with production-grade distributed training (FSDP2 + ring attention, Megatron), clever performance engineering (thread-pool HTTP for 2000+ concurrent requests, FFD sequence packing, interleaved tokenization for multi-turn), and a clean environment abstraction supporting both single-turn and multi-turn agent tasks.
Weaknesses
No visible test suite in the file tree (only pytest in dev dependencies), and the orchestrator at 149KB is a monolithic class that would benefit from further decomposition; the thread-pool HTTP approach, while clever, is a non-standard pattern that could be harder to maintain.
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
25
Contributors
1
Files
98
Active weeks
1
Repository
Language
Python
Stars
2
Forks
0
License
MIT