http://45.207.215.150:8090/古咩私人博客Halo v2.25.4zh-cnWed, 22 Jul 2026 07:52:52 GMThttp://45.207.215.150:8090/archives/deervla-dynamic-early-exit-inference-acceleration从多出口 MLLM、动作一致性判据与预算约束出发，拆解 NeurIPS 2024 的 DeeR-VLA 如何动态跳过 LLM 层，实现多模态机器人推理加速。]]>/archives/deervla-dynamic-early-exit-inference-accelerationAdministratorAIWed, 22 Jul 2026 07:18:29 GMThttp://45.207.215.150:8090/archives/tdmpc2-scalable-world-models从隐式潜在动力学、SimNorm、离散回归、策略先验与 MPPI 规划出发，拆解 ICLR 2024 Spotlight 论文 TD-MPC2 如何用统一超参数稳定训练并扩展到 317M 参数、80 个任务。]]>/archives/tdmpc2-scalable-world-modelsAdministratorAIWed, 22 Jul 2026 01:15:00 GMThttp://45.207.215.150:8090/archives/yollava-personalized-multimodal-assistant从可学习身份 Token、视觉属性软提示与难负样本出发，拆解 NeurIPS 2024 的 Yo'LLaVA 如何用 5 张图完成多模态大模型个性化。]]>/archives/yollava-personalized-multimodal-assistantAdministratorAITue, 21 Jul 2026 07:07:30 GMThttp://45.207.215.150:8090/archives/flashattention3-hopper-asynchronous-attention从 Hopper 的 TMA、WGMMA 和 FP8 出发，拆解 FlashAttention-3 如何用 warp specialization、GEMM-Softmax 流水与低误差量化提升精确注意力吞吐。]]>/archives/flashattention3-hopper-asynchronous-attentionAdministratorAITue, 21 Jul 2026 01:15:00 GMThttp://45.207.215.150:8090/archives/dccf-multimodal-fake-news-detection从事实-情感解耦、动态特征演化与冲突-共识度量出发，拆解 DCCF 的多模态假新闻检测方法、实验结果、复现缺口与结论边界。]]>/archives/dccf-multimodal-fake-news-detectionAdministratorAIMon, 20 Jul 2026 10:53:52 GMThttp://45.207.215.150:8090/archives/matryoshka-multimodal-models-inference-acceleration从嵌套视觉 Token、多尺度训练目标与预填充成本出发，拆解 ICLR 2025 的 M3 如何在同一多模态大模型中按需切换推理粒度。]]>/archives/matryoshka-multimodal-models-inference-accelerationAdministratorAIMon, 20 Jul 2026 07:15:49 GMThttp://45.207.215.150:8090/archives/blip2-q-former-multimodal从 Q-Former、三种表征学习目标与两阶段预训练出发，拆解 BLIP-2 如何用冻结视觉编码器和冻结 LLM 构建多模态模型。]]>/archives/blip2-q-former-multimodalAdministratorAIMon, 20 Jul 2026 01:15:00 GMThttp://45.207.215.150:8090/archives/dpo-direct-preference-optimization从 KL 约束的 RLHF 目标出发，推导 DPO 的闭式策略映射、偏好损失、训练流程、实验边界与复现要点。]]>/archives/dpo-direct-preference-optimizationAdministratorAISun, 19 Jul 2026 01:15:00 GMThttp://45.207.215.150:8090/archives/qlora-efficient-finetuning从 NF4、双重量化、Paged Optimizer 与全层 LoRA 出发，拆解 QLoRA 的显存账本、训练流程、实验边界和复现要点。]]>/archives/qlora-efficient-finetuningAdministratorAISat, 18 Jul 2026 01:15:00 GMThttp://45.207.215.150:8090/archives/transformer-vit-clip-interview-guide以三篇代表论文为主线，串起 Transformer、Vision Transformer 与 CLIP 的结构、目标和面试高频题。]]>/archives/transformer-vit-clip-interview-guideAdministratorAIMon, 13 Jul 2026 02:00:00 GMThttp://45.207.215.150:8090/archives/llm-roadmap一份系统化 LLM 学习路线，覆盖 Transformer、预训练、SFT、RLHF、RAG、Agent、推理优化、评测和工程项目。]]>/archives/llm-roadmapAdministrator技术Thu, 9 Jul 2026 03:00:00 GMThttp://45.207.215.150:8090/archives/ai-infra-roadmap一份面向工程同学的 AI Infra 学习路线，覆盖后端基础、RAG、Agent Runtime、模型服务、评测观测、稳定性工程和项目落地。]]>/archives/ai-infra-roadmapAdministrator技术Thu, 9 Jul 2026 02:00:00 GMThttp://45.207.215.150:8090/archives/multimodal-rag-from-scratch用纯 NumPy 手写 BM25、Dense 检索、RRF 融合和 MMR 重排，构建可评测的多模态 RAG。]]>/archives/multimodal-rag-from-scratchAdministratorAITue, 30 Jun 2026 10:00:00 GMThttp://45.207.215.150:8090/archives/mini-llm-engine-from-scratch通过可测试的 Python 实现，讲清 Paged KV Cache、Continuous Batching、Chunked Prefill 等六项推理优化。]]>/archives/mini-llm-engine-from-scratchAdministratorAITue, 30 Jun 2026 02:00:00 GMThttp://45.207.215.150:8090/archives/kafka-producer-broker-consumer用推荐系统里的用户行为日志为例，讲清楚 Kafka 的作用、Producer、Broker、Consumer、Topic、Partition、Offset 和消费语义。]]>/archives/kafka-producer-broker-consumerAdministrator技术Mon, 8 Jun 2026 16:00:00 GMThttp://45.207.215.150:8090/archives/leetcode-lru-merge-k-reverse-list面试高频算法题速记，整理 LRU Cache、合并 K 个有序链表、反转链表的核心思路、复杂度和 C++ 代码。]]>/archives/leetcode-lru-merge-k-reverse-listAdministrator技术Mon, 8 Jun 2026 16:00:00 GMThttp://45.207.215.150:8090/archives/high-concurrency-rate-limiting-algorithms面试和工程都常见的限流算法总结，讲清楚固定窗口、滑动窗口、漏桶、令牌桶、并发数限制以及分布式限流如何落地。]]>/archives/high-concurrency-rate-limiting-algorithmsAdministrator技术Mon, 8 Jun 2026 16:00:00 GMThttp://45.207.215.150:8090/archives/meddococr-interpreter-source-analysis从 OCR、版面恢复、字段归一化到安全解读，梳理医疗文档多模态原型的完整流水线。]]>/archives/meddococr-interpreter-source-analysisAdministrator技术Fri, 22 May 2026 02:30:00 GMThttp://45.207.215.150:8090/archives/cet6-writing-model-essays10 个六级写作高频话题范文，使用可复用的三段式结构并配套打印版 PDF。]]>/archives/cet6-writing-model-essaysAdministrator学习笔记Wed, 20 May 2026 01:00:00 GMThttp://45.207.215.150:8090/archives/claude-code-context-management基于源码区分请求前优化、高水位压缩和 API 报错恢复，厘清 Microcompact、Full Compact 与 Auto Compact 的关系。]]>/archives/claude-code-context-managementAdministrator技术Tue, 19 May 2026 02:00:00 GMT