https://jybsuper.github.io/Yanbin's BlogExploring LLM infrastructure, reinforcement learning, and machine learning engineering. Tech blog by Yanbin Jiang, an LLM engineer and RL researcher. 2026-04-28T16:25:26-07:00 Yanbin Jiang https://jybsuper.github.io/ Jekyll © 2026 Yanbin Jiang /assets/img/favicons/favicon.ico /assets/img/favicons/favicon-96x96.png SGLang + LoRA Deep Dive — Qwen3-30B-A3B-Instruct-25072026-04-21T14:00:00-07:00 2026-04-28T16:24:57-07:00 https://jybsuper.github.io/posts/sglang-lora-deep-dive/ yanbin_jiang Audited against SGLang main: 1ebe1c57eddd0ea7915b408e35a1b9b33cd10c41 (2026-04-19) Audited against HF transformers main: 631d45082cbd23f4146f1e79c37b3875a3dbc4f4 (2026-04-20) Model: Qwen/Qwen3-30B-A3B-Instruct-2507 — bfloat16, 30.5 B total / ~3 B active, 48 layers, 128 experts per layer, top-8 routing, 262 144 native context. All line numbers, function names, and tensor shapes in this document ... Efficient Forward Pass for Agent RL: Solving Multi-Turn Context Consistency (Part 2)2025-07-07T09:00:00-07:00 2025-07-10T00:46:43-07:00 https://jybsuper.github.io/posts/multiturn_forward_pass_cont/ yanbin_jiang In Part 1, I explored the fundamental challenge of training-inference context mismatch in reasoning models and prototyped three solutions. While those initial experiments on a single conversation demonstrated promising accuracy improvements (300-700× better KL divergence), the real question remained: how do these approaches perform at scale? This post benchmarks the three approaches from Part ... Efficient Forward Pass for Agent RL: Solving Multi-Turn Context Consistency (Part 1)2025-06-29T09:00:00-07:00 2025-07-01T10:33:31-07:00 https://jybsuper.github.io/posts/multiturn_forward_pass/ yanbin_jiang After implementing correct and scalable tokenization and masking1 for multi-turn rollout, there remains a critical challenge to achieve full consistency between training and inference: the context discrepancy problem. The Training-Inference Context Mismatch In my previous post, I briefly mentioned this issue. Now, let’s dive deep into why it matters and how to solve it efficiently. Consider ... LangGraph Rollout: Evolving VeRL's Multi-Turn Capabilities for Agent RL2025-06-21T09:00:00-07:00 2025-06-30T13:27:53-07:00 https://jybsuper.github.io/posts/langgraph_rollout/ yanbin_jiang After completing our multi-turn tokenization and masking refactoring, we eliminated a critical bottleneck that was preventing us from building a more consistent and flexible rollout system for our Agent RL research. This breakthrough enabled us to implement a LangGraph-based rollout for VeRL in just a few days, which we’ve already successfully deployed in our Agent RL experiments. In this artic... When Reasoning Models Break Tokenization: The Hidden Complexity of Multiturn Training2025-06-11T09:00:00-07:00 2025-06-11T09:00:00-07:00 https://jybsuper.github.io/posts/multiturn_tokenization/ yanbin_jiang I recently spent two weeks refactoring multiturn tokenization and masking for VeRL. While VeRL already had a functional implementation, what initially seemed like a straightforward refactor turned out to be surprisingly nuanced. Through multiple iterations, we arrived at a solution that is both robust and flexible for VeRL users. This post shares the key learnings and design choices from that j...