你好，我是 KOO SHWAH，华南理工大学软件工程专业学生，GitHub 用户名是 ProIg-Chaa。

我目前关注的核心问题是：

如何让大模型在真实推理场景中更高效、更稳定、更容易被理解和复现。

我的兴趣集中在 LLM / MLLM 推理系统优化，包括 KV Cache、Prefix / Segment Reuse、serving 调度、量化压缩、CUDA kernel、benchmark，以及多模态推理过程中的效率与稳定性问题。

我更希望把研究想法落到可以运行、可以测量、可以继续演化的系统里，而不是只停留在 prompt、概念或离线指标层面。

• 面向 LLM / MLLM 的轻量推理与 serving 系统
• KV cache、prefix reuse、segment-level context reuse
• Agent / RAG / multi-turn workload 下的上下文复用与性能分析
• 推理质量、输出长度、latency、memory 之间的权衡
• CUDA operator、profiling、benchmark 与可复现实验

我希望把自己的工作统一到一条主线上：

面向多模态与 Agent 场景的大模型推理系统优化：从 CUDA kernel、KV cache、serving 调度，到 reasoning path 的效率与稳定性分析。

这条主线连接了三个层面：

flowchart LR
    A[Workload<br/>Chat / RAG / Agent / MLLM] --> B[Serving System<br/>Scheduling / Batching / Cache Reuse]
    B --> C[Model Inference<br/>Prefill / Decode / Reasoning Path]
    C --> D[Low-level Execution<br/>CUDA / Memory Access / Profiling]
    D --> E[Validation<br/>Benchmark / Ablation / Reproducibility]

我也在参与 多模态大模型隐式推理优化 相关工作。

我更关心的不只是模型“能不能推理”，而是：

• 推理过程是否能被观测、量化和解释
• 视觉信息是否真的在关键阶段被使用
• 隐式推理是否降低了显式 token 成本
• routing / latent reasoning / visual injection 是否能跨数据集稳定迁移
• 准确率提升是否值得额外的 latency 和 memory pressure

因此，我会把 MLLM reasoning 放在系统视角下看：不仅分析 accuracy，也分析输出长度、截断率、失败类型、推理阶段、prefill/decode 成本和整体 serving 影响。

• 读懂 LLM serving pipeline，而不是只调用 API
• 从 request lifecycle 视角分析 prefill、decode、cache、scheduler 的关系
• 用 benchmark 和 profiling 判断优化是否真实有效
• 将论文想法改造成可复现的小型系统
• 在 MLLM reasoning 实验中同时关注质量、成本和失败模式
• 使用 coding agents 辅助读代码、改实验、整理结果，但保留人工 review 和测试纪律

Python PyTorch CUDA C++ CMake Transformers LLM Serving Benchmarking Profiling

• Agent / RAG workload 下的 KV cache 复用与性能分析
• prefix reuse 到 segment-level context reuse 的小型原型
• vLLM / SGLang / LMCache 相关 serving 机制学习与实验
• CUDA kernel 与 Nsight profiling 能力建设
• MLLM reasoning 方法的机制分析、失败模式分析与成本收益评估

• LLM inference、KV cache、serving systems 方向的交流
• Agent / RAG workload 下 context reuse 与推理效率相关讨论
• MLLM reasoning efficiency / stability 方向的合作或建议
• 小而扎实、可复现、可继续演化的工程实验

• 对于中文读者，我整理了一些个人的学习笔记在*@Chaa的AI学习笔记*

• GitHub: @ProIg-Chaa

Hi, I'm KOO SHWAH, a Software Engineering student at South China University of Technology, and my GitHub handle is ProIg-Chaa.

I focus on one core problem:

How to make large model inference more efficient, stable, understandable, and reproducible in realistic workloads.

My current interests are centered around LLM / MLLM inference system optimization, including KV Cache, Prefix / Segment Reuse, serving scheduling, quantization, CUDA kernels, benchmarking, and the efficiency-stability tradeoffs in multimodal reasoning.

I enjoy turning research ideas into small but real systems that are runnable, measurable, and easy to iterate on.

• Lightweight LLM / MLLM inference and serving systems
• KV cache, prefix reuse, and segment-level context reuse
• Context reuse and performance analysis under Agent / RAG / multi-turn workloads
• Tradeoffs among reasoning quality, output length, latency, and memory
• CUDA operators, profiling, benchmarking, and reproducible experiments

I try to connect my work into one technical direction:

Inference system optimization for multimodal and agentic workloads, spanning CUDA kernels, KV cache, serving scheduling, and reasoning-path efficiency.

flowchart LR
    A[Workload<br/>Chat / RAG / Agent / MLLM] --> B[Serving System<br/>Scheduling / Batching / Cache Reuse]
    B --> C[Model Inference<br/>Prefill / Decode / Reasoning Path]
    C --> D[Low-level Execution<br/>CUDA / Memory Access / Profiling]
    D --> E[Validation<br/>Benchmark / Ablation / Reproducibility]

I am also working on implicit reasoning optimization for multimodal large models.

What matters to me is not only whether a model can reason, but whether the reasoning process can be observed, measured, and made efficient in practice.

I care about:

• whether visual information is actually used during key generation stages
• whether implicit reasoning can reduce explicit token cost
• whether routing, latent reasoning, or visual injection strategies transfer across datasets
• whether accuracy gains justify additional latency and memory pressure
• how reasoning optimization interacts with cache reuse, batching, compression, and serving

I therefore study MLLM reasoning from a systems perspective: not only accuracy, but also output length, truncation rate, failure modes, inference stage, prefill/decode cost, and serving impact.

• Understanding LLM serving pipelines instead of only calling APIs
• Analyzing prefill, decode, cache, and scheduler interactions from the request lifecycle
• Using benchmark and profiling to verify whether an optimization is real
• Turning research ideas into reproducible mini-systems
• Evaluating MLLM reasoning by quality, cost, and failure modes together
• Using coding agents to assist code reading, experimentation, and result analysis, while keeping human review and testing discipline

Python PyTorch CUDA C++ CMake Transformers LLM Serving Benchmarking Profiling

• KV cache reuse and performance analysis under Agent / RAG workloads
• A small prototype moving from prefix reuse toward segment-level context reuse
• Studying vLLM / SGLang / LMCache serving mechanisms through experiments
• CUDA kernel practice with Nsight-based profiling
• Mechanism analysis, failure mode analysis, and cost-benefit evaluation for MLLM reasoning methods

• Conversations around LLM inference, KV cache, and serving systems
• Discussions on context reuse and inference efficiency under Agent / RAG workloads
• Collaboration or feedback on MLLM reasoning efficiency and stability
• Small, serious, reproducible engineering experiments

• GitHub: @ProIg-Chaa