Feiyuan Zhang

Pengbo Li, Feiyuan Zhang, Guangming Sheng et al.

GRPO- and PPO-style LLM post-training commonly sample multiple trajectories from the same prompt and then train on the resulting group. In long-context RL workloads, this shared prompt-side prefix can contain retrieved passages, visual tokens, tool schemas, system instructions, or task context, while the full rollout group is still too large to pack into one training microbatch. Standard dense trainers therefore recompute the same prefix forward and backward for every trajectory. We present a schedule-level reuse mechanism that decouples prefix and suffix computation. The schedule runs prefix forward once, executes suffixes as ordinary microbatches while reading prefix K/V and accumulating prefix-side gK/gV , and then runs prefix backward once on the accumulated gradient cache. This reordered schedule is equivalent to baseline training over real arithmetic and aligns numerically within finite-precision tolerance. Because only K/V and gK/gV are hot during suffix computation, the approach offloads dormant prefix activations, integrates with TP/EP/CP/PP and DP-style placement at the execution level, and preserves aux-loss-based MoE router semantics through logical prefix-token accounting. On dense Llama3-8B, Qwen3-8B, and MoE Qwen3-MoE-30B-A3B configurations, the schedule matches optimizer updates across TP/CP/PP/EP combinations, aligns on a 100-step real RL trace replay, reaches up to 4.395x speedup (2.930x under a conservative compile-on comparison) as prefix ratio and rollout group size grow, and reduces Phase-B peak HBM by up to 59.1%, extending the Llama3-8B capacity frontier from 17,920 to 29,696 total tokens.

Feiyuan Zhang, Dezhi Zhu, James Ming et al.

Retrieval-Augmented Generation (RAG) systems have shown substantial benefits in applications such as question answering and multi-turn dialogue \citep{lewis2020retrieval}. However, traditional RAG methods, while leveraging static knowledge bases, often overlook the potential of dynamic historical information in ongoing conversations. To bridge this gap, we introduce DH-RAG, a Dynamic Historical Context-Powered Retrieval-Augmented Generation Method for Multi-Turn Dialogue. DH-RAG is inspired by human cognitive processes that utilize both long-term memory and immediate historical context in conversational responses \citep{stafford1987conversational}. DH-RAG is structured around two principal components: a History-Learning based Query Reconstruction Module, designed to generate effective queries by synthesizing current and prior interactions, and a Dynamic History Information Updating Module, which continually refreshes historical context throughout the dialogue. The center of DH-RAG is a Dynamic Historical Information database, which is further refined by three strategies within the Query Reconstruction Module: Historical Query Clustering, Hierarchical Matching, and Chain of Thought Tracking. Experimental evaluations show that DH-RAG significantly surpasses conventional models on several benchmarks, enhancing response relevance, coherence, and dialogue quality.

Di Chai, Pengbo Li, Feiyuan Zhang et al.

Token filtering has been proposed to enhance utility of large language models (LLMs) by eliminating inconsequential tokens during training. While using fewer tokens should reduce computational workloads, existing studies have not succeeded in achieving higher efficiency. This is primarily due to the insufficient sparsity caused by filtering tokens only in the output layers, as well as inefficient sparse GEMM (General Matrix Multiplication), even when having sufficient sparsity. This paper presents Collider, a system unleashing the full efficiency of token filtering in LLM training. At its core, Collider filters activations of inconsequential tokens across all layers to maintain sparsity. Additionally, it features an automatic workflow that transforms sparse GEMM into dimension-reduced dense GEMM for optimized efficiency. Evaluations on three LLMs-TinyLlama-1.1B, Qwen2.5-1.5B, and Phi1.5-1.4B-demonstrate that Collider reduces backpropagation time by up to 35.1% and end-to-end training time by up to 22.0% when filtering 40% of tokens. Utility assessments of training TinyLlama on 15B tokens indicate that Collider sustains the utility advancements of token filtering by relatively improving model utility by 16.3% comparing to regular training, and reduces training time from 4.7 days to 3.5 days using 8 GPUs. Collider is designed for easy integration into existing LLM training frameworks, allowing systems already using token filtering to accelerate training with just one line of code.