Author

Xiaohongshu Tech REDtech

Official account of the Xiaohongshu tech team, sharing tech innovations and problem insights, advancing together.

119

Articles

Likes

410

Views

Comments

Latest from Xiaohongshu Tech REDtech

100 recent articles max

Xiaohongshu Tech REDtech

Jan 2, 2025 · Artificial Intelligence

Xiaohongshu's Self-developed RLHF System for Multimodal Large Language Models: Design, Optimization, and Performance

Xiaohongshu’s team unveiled a self‑developed RLHF system that trains multimodal large language models using heterogeneous and homogeneous network architectures, extensive PPO optimizations, and Medusa speculative sampling, achieving over 50% throughput gains, reduced hardware needs, and 5‑20% performance improvements on zero‑shot benchmarks.

PPOPRMRLHF

0 likes · 21 min read

Xiaohongshu's Self-developed RLHF System for Multimodal Large Language Models: Design, Optimization, and Performance

Xiaohongshu Tech REDtech

Dec 26, 2024 · Artificial Intelligence

Focused Large Language Models are Stable Many-Shot Learners

FocusICL mitigates the reverse‑scaling of in‑context learning by masking irrelevant tokens and applying hierarchical batch attention, cutting attention complexity, and delivering consistent query focus that yields average accuracy gains of about 5 % across multiple LLMs and benchmarks.

FocusICLIn-Context Learningattention mechanisms

0 likes · 16 min read

Focused Large Language Models are Stable Many-Shot Learners

Xiaohongshu Tech REDtech

Dec 26, 2024 · Artificial Intelligence

Instruction Embedding: Latent Representations of Instructions for Task Identification

The paper introduces Instruction Embedding—a task‑focused text representation learned on the new Instruction Embedding Benchmark—and shows that Prompt‑based Instruction Embedding (PIE) outperforms standard embeddings in clustering, similarity, and downstream tasks such as data selection, in‑context example retrieval, test‑set compression, and task‑correlation analysis.

contrastive learningfine-tuninginstruction embedding

0 likes · 15 min read

Instruction Embedding: Latent Representations of Instructions for Task Identification

Xiaohongshu Tech REDtech

Dec 25, 2024 · Industry Insights

How Xiaohongshu’s Security Team Achieved Zero Defense Losses in Shanghai’s 2024 “Panshi Action”

In December 2024, Xiaohongshu’s information security team topped the Shanghai “Panshi Action” competition, earning top blue‑team honors and a zero‑loss defense record by leveraging real‑time traffic monitoring, big‑data analytics, rapid incident response, and successful attacker attribution.

Incident ResponseInformation Securitybig data analysis

0 likes · 3 min read

How Xiaohongshu’s Security Team Achieved Zero Defense Losses in Shanghai’s 2024 “Panshi Action”

Xiaohongshu Tech REDtech

Dec 19, 2024 · Databases

Data Consistency Verification Practices and Implementation at Xiaohongshu

Xiaohongshu built a lock‑free, non‑disruptive data‑consistency verification tool that automatically selects optimal methods, handles heterogeneous sources and dynamic changes, performs full and incremental checks via chunked checksums or row‑by‑row comparison, quickly isolates mismatches, and supports automatic remediation, ensuring reliable migrations and sharding.

Data ConsistencyDistributed SystemsMySQL

0 likes · 16 min read

Data Consistency Verification Practices and Implementation at Xiaohongshu

Xiaohongshu Tech REDtech

Dec 5, 2024 · Big Data

Interview with Jianchen: Journey from Open Source Contributor to Data Engineer at Xiaohongshu

In this interview, Xiaohongshu data engineer Jianchen recounts his evolution from a computer‑science student discovering open‑source through MIT6.824 to contributing to SOFAJRaft and Apache RocketMQ, detailing his OSPP projects, the decision to join Xiaohongshu, and his work on a cloud‑native Kafka engine that cut storage and compute usage by half.

Apache RocketMQBig DataCloud Native

0 likes · 11 min read

Interview with Jianchen: Journey from Open Source Contributor to Data Engineer at Xiaohongshu

Xiaohongshu Tech REDtech

Nov 7, 2024 · Artificial Intelligence

RTAMS-GANNS: A Real-Time Adaptive Multi-Stream GPU System for Online Approximate Nearest Neighbor Search

RTAMS‑GANNS, the award‑winning real‑time adaptive multi‑stream GPU system for online approximate nearest neighbor search, eliminates costly memory allocations and serial execution by using a dynamic memory‑block insertion algorithm and separate CUDA streams, cutting latency by 40‑80% and reliably serving over 100 million daily users in production.

GPUPerformance EvaluationVector Insertion

0 likes · 19 min read

RTAMS-GANNS: A Real-Time Adaptive Multi-Stream GPU System for Online Approximate Nearest Neighbor Search

Xiaohongshu Tech REDtech

Oct 23, 2024 · Backend Development

Pulsar vs RocketMQ: Architecture, Cost Benefits, and Migration Strategy for Xiaohongshu Online Messaging

Xiaohongshu replaced its RocketMQ‑based online messaging platform with Apache Pulsar, achieving up to 48% total cost reduction, 43% higher CPU utilization, 30% resource savings, and a latency drop from 20.2 ms to 5.7 ms through cloud‑native, elastic scaling and a phased migration strategy.

Apache PulsarMessage QueueRocketMQ

0 likes · 12 min read

Pulsar vs RocketMQ: Architecture, Cost Benefits, and Migration Strategy for Xiaohongshu Online Messaging

Xiaohongshu Tech REDtech

Oct 11, 2024 · Artificial Intelligence

Harmonized Speculative Sampling (HASS): Aligning Training and Decoding for Efficient Large Language Model Inference

HASS aligns training and decoding contexts and objectives for speculative sampling, using harmonized objective distillation and multi-step context alignment, achieving 2.81–4.05× speedup and 8%–20% improvement over EAGLE‑2 while preserving generation quality in real-world deployments at Xiaohongshu.

AIHASSInference Acceleration

0 likes · 11 min read

Harmonized Speculative Sampling (HASS): Aligning Training and Decoding for Efficient Large Language Model Inference

Xiaohongshu Tech REDtech

Oct 9, 2024 · Operations

AIOps Implementation at Xiaohongshu: Fault Localization and Intelligent Operations

Xiaohongshu’s AIOps initiative builds a four‑layer framework that leverages machine‑learning‑driven anomaly detection, causal analysis, and trace‑based fault localization to automatically identify root‑cause services in micro‑service environments, achieving over 80 % accuracy across 1000 daily diagnoses while guiding future enhancements in change correlation and automated remediation.

Anomaly DetectionDevOpsFault Localization

0 likes · 28 min read

AIOps Implementation at Xiaohongshu: Fault Localization and Intelligent Operations