Skip to content
View ITerydh's full-sized avatar
😄
happy
😄
happy

Block or report ITerydh

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
ITerydh/README.md
terminal typing banner

profile viewsfollowerslocation

root@iterhui-gpu-node:~# ./profile --scan[ok] identity : Yang Denghui / iterhui[ok] role : Multimodal Generation Engineer @ Taotian Group (Alibaba)[ok] previously : Video Generation Model Engineer @ Baidu[ok] focus : video / image generation, inference acceleration, training infra[ok] preferred io : profile first, verify quality, then ship to production[ok] status : squeezing every millisecond out of diffusion models

current_work

$ ydhctl top --sort impact
PID SYSTEM MODE SIGNAL
001 Image Generation Serving infra quantization, attention kernels, cache reuse
002 Multimodal Retrieval algorithm 2B dual-tower, contrastive learning + RL
003 Generation Eval Harness platform multi-node T2I / I2V benchmarks, GSB & FID
004 Training Pipeline infra profiler, async metrics, step-time regressions

recent_stack

root@iterhui-gpu-node:~# cat highlights.log[retrieval] 2B multimodal dual-tower: contrastive -> reasoning loss -> GRPO RL[video] MuseSteamer: i2v model ranked #1 on the VBench image-to-video leaderboard[scale] thousand-GPU clusters trained stably for months, minute-level auto recovery[cache] TeaCache, CFGCache, EasyCache, DiCache - plus auto coefficient prediction[distill] CFG / DMD / LoRA distillation with near-lossless quality[hetero] NVIDIA H & B series, Kunlunxin, AMD MI308X - all shipped to production[method] every win backed by profiler evidence and an offline + online quality gate

toolchain

runtime_stats

profile summary

background

root@iterhui-gpu-node:~# cat /etc/backgroundedu : Sichuan University - Cyber & Information Securitypaper : First author, CCF-B conference - pyramid features + ViT for Android malware detectionpaper : Co-author, SCI Q2 journal - multi-feature fusion for malicious code detectionhonors : PPDE (PaddlePaddle Developer Expert), CCF BDCI national award, CSDN blog expertwriting : CSDN AI domain expert, 30k+ followers, writing on generation and inference

contact

root@iterhui-gpu-node:~# cat /etc/contactGitHub : https://github.com/ITerydhCSDN : https://blog.csdn.net/qq_41976613AI Studio : https://aistudio.baidu.com/aistudio/personalcenter/thirdview/643467Work : Multimodal GEN & INFRA & ALGORITHM

Pinned Loading

  1. OCRandQPTandISOCRandQPTandISPublic

    采用PPOCR和QPT完成一个疫情信息统计小工具

    Python 11

  2. OnnxOCROnnxOCRPublic

    Forked from jingsongliujing/OnnxOCR

    基于PaddleOCR重构,并且脱离PaddlePaddle深度学习训练框架的轻量级OCR,推理速度超快 —— A lightweight OCR system based on PaddleOCR, decoupled from the PaddlePaddle deep learning training framework, with ultra-fast inference speed.

    Python

  3. sglangsglangPublic

    Forked from sgl-project/sglang

    SGLang is a high-performance serving framework for large language models and multimodal models.

    Python