<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"><channel><title>OneCrewDev</title><description>技术笔记、项目记录、随想。</description><link>https://onecrewdev.com/</link><item><title>QNN W8A16 的 INT32 Bias Overflow：从极小 Scale 到稳定部署</title><link>https://onecrewdev.com/posts/qnn-w8a16-int32-bias-overflow/</link><guid isPermaLink="true">https://onecrewdev.com/posts/qnn-w8a16-int32-bias-overflow/</guid><description>记录 QNN HTP W8A16 中 INT32 bias 溢出的根因、公式、QueOpt 修复方式，以及在真实板端的验证结果。</description><pubDate>Wed, 26 Aug 2026 02:00:00 GMT</pubDate></item><item><title>从 LoRA 到 QLoRA、QA-LoRA：低秩更新如何进入量化模型</title><link>https://onecrewdev.com/posts/lora-qlora-qa-lora-explained/</link><guid isPermaLink="true">https://onecrewdev.com/posts/lora-qlora-qa-lora-explained/</guid><description>用一个完整的数值例子解释 LoRA、QLoRA 和 QA-LoRA 的训练、推理与合并流程，并拆开 QA-LoRA 官方实现中的 group pooling 和 qzeros 写回。</description><pubDate>Mon, 10 Aug 2026 00:00:00 GMT</pubDate></item><item><title>从权重角度理解均匀量化与向量量化</title><link>https://onecrewdev.com/posts/uniform-vs-vector-quantization/</link><guid isPermaLink="true">https://onecrewdev.com/posts/uniform-vs-vector-quantization/</guid><description>用权重矩阵、scale/offset、码本和索引，拆开解释均匀量化与向量量化到底在压缩什么、如何反量化，以及为什么码本大小会决定 bit 数。</description><pubDate>Fri, 07 Aug 2026 04:00:00 GMT</pubDate></item><item><title>HQQ 算法解读：不用校准数据，怎样把大模型权重压到 4 bit</title><link>https://onecrewdev.com/posts/hqq-half-quadratic-quantization/</link><guid isPermaLink="true">https://onecrewdev.com/posts/hqq-half-quadratic-quantization/</guid><description>从 Dropbox HQQ 源码出发，拆解分组非对称量化、half-quadratic 零点优化、bit packing 与推理路径，并说明 axis、group size 和 backend 该怎样选。</description><pubDate>Tue, 28 Jul 2026 06:00:00 GMT</pubDate></item><item><title>DINOv3 做 W8A8 PTQ：一次 FP 激活分布定位，把问题从猜测变成证据</title><link>https://onecrewdev.com/posts/dinov3-w8a8-ptq-activation-outlier-diagnosis/</link><guid isPermaLink="true">https://onecrewdev.com/posts/dinov3-w8a8-ptq-activation-outlier-diagnosis/</guid><description>DINOv3 ViT-B/16 的标准全图 W8A8 为何失效？64 张图的逐层 FP32 激活 profile 表明，block 2 MLP 的 learned register 与 patch token 出现数百倍动态范围差。</description><pubDate>Wed, 22 Jul 2026 15:00:00 GMT</pubDate></item><item><title>【arXiv】Activation Quantization of Vision Encoders Needs Prefixing Registers</title><link>https://onecrewdev.com/posts/regcache-activation-quantization-vision-encoders-paper-reading/</link><guid isPermaLink="true">https://onecrewdev.com/posts/regcache-activation-quantization-vision-encoders-paper-reading/</guid><description>RegCache 发现视觉编码器中的高范数 outlier 会严重放大量化误差，并通过中间层 prefix registers 和 token deletion 改善 CLIP、SigLIP、DINOv2 等模型的低比特量化。</description><pubDate>Tue, 21 Jul 2026 10:30:00 GMT</pubDate></item><item><title>【ICLR 2024】Vision Transformers Need Registers</title><link>https://onecrewdev.com/posts/vision-transformers-need-registers-paper-reading/</link><guid isPermaLink="true">https://onecrewdev.com/posts/vision-transformers-need-registers-paper-reading/</guid><description>Vision Transformers Need Registers 解释 ViT 中背景高范数 token 的来源，并用额外的 register tokens 把模型的隐式工作空间显式化。</description><pubDate>Tue, 21 Jul 2026 10:30:00 GMT</pubDate></item><item><title>project 的 GitLab CI/CD 与 Python 包发版流程</title><link>https://onecrewdev.com/posts/queopt-gitlab-ci-cd-release/</link><guid isPermaLink="true">https://onecrewdev.com/posts/queopt-gitlab-ci-cd-release/</guid><description>记录 project 从专用 GitLab Runner、部署回归，到 wheel 构建、隔离验证、包发布和 Release 创建的完整流程。</description><pubDate>Mon, 20 Jul 2026 02:30:00 GMT</pubDate></item><item><title>Agent 时代的基础设施：从 lark-cli 看执行 Harness 应该怎么建</title><link>https://onecrewdev.com/posts/agent-infrastructure-harness-lark-cli/</link><guid isPermaLink="true">https://onecrewdev.com/posts/agent-infrastructure-harness-lark-cli/</guid><description>Agent 需要的不是更多工具，而是一条可验证、可约束、可恢复的执行链。以 lark-cli 源码为例，拆解命令层、身份、策略、输出契约与测试如何组成 Harness。</description><pubDate>Mon, 20 Jul 2026 02:00:00 GMT</pubDate></item><item><title>QAIRT 2.42 适配 Streaming Spatial Speech Enhancement Network 的 GRU</title><link>https://onecrewdev.com/posts/qnn-gru-streaming-spatial-speech-enhancement-htp-v73/</link><guid isPermaLink="true">https://onecrewdev.com/posts/qnn-gru-streaming-spatial-speech-enhancement-htp-v73/</guid><description>Streaming Spatial Speech Enhancement Network 的 HTP 部署排查：GRU 展开、长序列 native cell、QAIRT 源码开关、Einsum layout bug 与流式 I/O 对齐。</description><pubDate>Wed, 15 Jul 2026 01:00:00 GMT</pubDate></item><item><title>MNN QNN 里为什么 chunk size 会影响 PPL</title><link>https://onecrewdev.com/posts/mnn-qnn-chunk-size-ppl/</link><guid isPermaLink="true">https://onecrewdev.com/posts/mnn-qnn-chunk-size-ppl/</guid><description>记录一次 Qwen3-0.6B 在 MNN QNN 上排查 PPL 异常的过程，解释为什么 chunk size 不只是输入切块参数，而是同时决定 prefill 图 shape、padding 路径和 ppl_eval 的上下文窗口，因此会真实影响困惑度。</description><pubDate>Thu, 09 Jul 2026 16:00:00 GMT</pubDate></item><item><title>【arXiv】INP-Former++: Advancing Universal Anomaly Detection via Intrinsic Normal Prototypes and Residual Learning</title><link>https://onecrewdev.com/posts/inp-former-plus-plus-paper-reading/</link><guid isPermaLink="true">https://onecrewdev.com/posts/inp-former-plus-plus-paper-reading/</guid><description>INP-Former++ 在 INP-Former 的单图内在正常原型上加入 Soft INP Coherence Loss 和 Residual Learning，把方法扩展到 semi-supervised、few-shot、multi-class 和一定 zero-shot 场景。</description><pubDate>Thu, 09 Jul 2026 02:16:00 GMT</pubDate></item><item><title>【arXiv】One Dinomaly2 Detect Them All: A Unified Framework for Full-Spectrum Unsupervised Anomaly Detection</title><link>https://onecrewdev.com/posts/dinomaly-v2-paper-reading/</link><guid isPermaLink="true">https://onecrewdev.com/posts/dinomaly-v2-paper-reading/</guid><description>Dinomaly2 在 Dinomaly 的特征重建框架上加入 Context-Aware Recentering，并把同一套最小化设计扩展到多视角、多模态、few-shot 和 inference-unified MUAD。</description><pubDate>Thu, 09 Jul 2026 02:08:00 GMT</pubDate></item><item><title>【CVPR 2025】Exploring Intrinsic Normal Prototypes within a Single Image for Universal Anomaly Detection</title><link>https://onecrewdev.com/posts/inp-former-paper-reading/</link><guid isPermaLink="true">https://onecrewdev.com/posts/inp-former-paper-reading/</guid><description>INP-Former 从测试图像自身动态提取 Intrinsic Normal Prototypes，并用这些正常原型指导特征重建，缓解训练集正常原型与测试图不对齐的问题。</description><pubDate>Thu, 09 Jul 2026 01:58:00 GMT</pubDate></item><item><title>【CVPR 2025】Dinomaly: The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection</title><link>https://onecrewdev.com/posts/dinomaly-paper-reading/</link><guid isPermaLink="true">https://onecrewdev.com/posts/dinomaly-paper-reading/</guid><description>Dinomaly 用强预训练 ViT 特征、Noisy Bottleneck、Linear Attention 和 Loose Reconstruction，缓解多类别无监督异常检测中的 identity mapping 问题。</description><pubDate>Thu, 09 Jul 2026 01:50:00 GMT</pubDate></item><item><title>SeqMSE 原理：用输出重构误差选择量化 Encoding</title><link>https://onecrewdev.com/posts/seqmse-principle/</link><guid isPermaLink="true">https://onecrewdev.com/posts/seqmse-principle/</guid><description>SeqMSE 的核心思想、候选 encoding 搜索、reconstruction loss 计算、block-wise 优化方式，以及它和 min/max、AdaRound、GPTQ 的区别。</description><pubDate>Wed, 08 Jul 2026 01:30:00 GMT</pubDate></item><item><title>RKNN OCR Rec INT8 量化：Weight Outlier、CLE 与 Hybrid 的取舍</title><link>https://onecrewdev.com/posts/rknn-ocr-rec-int8-cle/</link><guid isPermaLink="true">https://onecrewdev.com/posts/rknn-ocr-rec-int8-cle/</guid><description>记录一次 OCR recognition presoftmax 模型在 RKNN W8A8 量化下的精度排查：对比 normal、KL、MMSE，尝试手动 Cross-Layer Equalization 处理 weight outlier，并分析为什么最终仍需要 hybrid。</description><pubDate>Wed, 08 Jul 2026 00:00:00 GMT</pubDate></item><item><title>MNN QNN 离线模式跑通 Qwen3-4B 大模型</title><link>https://onecrewdev.com/posts/mnn-qnn-qwen3-4b-bringup/</link><guid isPermaLink="true">https://onecrewdev.com/posts/mnn-qnn-qwen3-4b-bringup/</guid><description>完整记录从本地编译 MNN、生成 QNN 离线模型、准备高通 QNN 依赖，到在 QCS8550/8 Gen 2 板端按 MNN 文档跑通 Qwen3-4B 的全过程，以及中间遇到的权限、路径和库问题。</description><pubDate>Fri, 26 Jun 2026 12:00:00 GMT</pubDate></item><item><title>llama.cpp Hexagon HTP 长 Prompt 默认 ubatch=512 卡住的定位与修复</title><link>https://onecrewdev.com/posts/llama-cpp-hexagon-ubatch-prefill-hang/</link><guid isPermaLink="true">https://onecrewdev.com/posts/llama-cpp-hexagon-ubatch-prefill-hang/</guid><description>本来只是想在 Hexagon HTP 上跑一条稍长一点的总结题，结果 `llama.cpp` 默认 `ubatch=512` 直接卡在 prefill。最后一路排到 HMX matmul，发现是小 remainder batch 走异步 pipeline 后没回来。</description><pubDate>Tue, 23 Jun 2026 08:30:00 GMT</pubDate></item><item><title>llama.cpp 里的 lm-head：Q6_K、Q4_0，以及为什么 4B 不一定提速</title><link>https://onecrewdev.com/posts/llama-cpp-lm-head-q6k-q4/</link><guid isPermaLink="true">https://onecrewdev.com/posts/llama-cpp-lm-head-q6k-q4/</guid><description>从 Qwen3/Qwen3.5 的 tied embedding 看 lm-head 性能瓶颈：0.6B 改 output-q4_0 有收益，4B 上收益有限甚至可能变慢。</description><pubDate>Fri, 05 Jun 2026 10:00:00 GMT</pubDate></item><item><title>llama.cpp HTP 调优：少用 CPU 为什么反而慢</title><link>https://onecrewdev.com/posts/llama-cpp-htp-cpu-fallback/</link><guid isPermaLink="true">https://onecrewdev.com/posts/llama-cpp-htp-cpu-fallback/</guid><description>用 Qwen3-0.6B 的 profile 解释一个朴素误区：减少 CPU 参与不等于更快。HTP attention、SET_ROWS、图切分和真实 token/s 的关系。</description><pubDate>Fri, 05 Jun 2026 09:00:00 GMT</pubDate></item><item><title>llama.cpp 跑 Qwen3-0.6B：HTP 上从 20 到 70 token/s</title><link>https://onecrewdev.com/posts/llama-cpp-qwen3-0-6b-htp-70tps/</link><guid isPermaLink="true">https://onecrewdev.com/posts/llama-cpp-qwen3-0-6b-htp-70tps/</guid><description>一次 Snapdragon Hexagon HTP 上的真实调优记录：OPPOLL、算子过滤、lm-head 量化，以及为什么 70 token/s 不是“全上 NPU”跑出来的。</description><pubDate>Fri, 05 Jun 2026 08:00:00 GMT</pubDate></item><item><title>llama.cpp Hexagon NPU 量化：Q4_0 / IQ4_NL / MXFP4 / Q8_0</title><link>https://onecrewdev.com/posts/llama-cpp-quantization/</link><guid isPermaLink="true">https://onecrewdev.com/posts/llama-cpp-quantization/</guid><description>Q4_0 / Q4_1 / IQ4_NL / MXFP4 / Q8_0 五种 HTP 原生量化格式的原理、反量化路径与对比。</description><pubDate>Sat, 30 May 2026 00:00:00 GMT</pubDate></item><item><title>TVM 是什么：层级架构与编译流水线</title><link>https://onecrewdev.com/posts/tvm-compilation-pipeline/</link><guid isPermaLink="true">https://onecrewdev.com/posts/tvm-compilation-pipeline/</guid><description>从整体架构到端到端流水线，理解 TVM 是什么、分几层、一个模型如何从 Relax 走到机器码。</description><pubDate>Fri, 29 May 2026 06:00:00 GMT</pubDate></item><item><title>llama.cpp 高通 Hexagon NPU 初跑 Qwen3.5 4B 模型</title><link>https://onecrewdev.com/posts/llama-cpp-hexagon-npu/</link><guid isPermaLink="true">https://onecrewdev.com/posts/llama-cpp-hexagon-npu/</guid><description>从 GGUF 量化、交叉编译到板端推理，完整记录 llama.cpp 在骁龙 Hexagon NPU 上的部署流程与调优经验。</description><pubDate>Fri, 29 May 2026 00:00:00 GMT</pubDate></item><item><title>TVM FFI（三）：多语言绑定与工具链</title><link>https://onecrewdev.com/posts/tvm-ffi-ecosystem/</link><guid isPermaLink="true">https://onecrewdev.com/posts/tvm-ffi-ecosystem/</guid><description>Python/Rust 绑定 SDK、JIT 即时编译、stubgen 类型标注生成、全局注册表、stream 管理及 addons。</description><pubDate>Thu, 28 May 2026 16:00:00 GMT</pubDate></item><item><title>TVM FFI（二）：类型系统与容器</title><link>https://onecrewdev.com/posts/tvm-ffi-type-system/</link><guid isPermaLink="true">https://onecrewdev.com/posts/tvm-ffi-type-system/</guid><description>Any/AnyView 所有权语义、Array/Map/Variant/Expected 容器、结构相等性与哈希、反射系统与 Dataclass。</description><pubDate>Thu, 28 May 2026 14:00:00 GMT</pubDate></item><item><title>TVM FFI（一）：对象系统与调用约定</title><link>https://onecrewdev.com/posts/tvm-ffi-deep-dive/</link><guid isPermaLink="true">https://onecrewdev.com/posts/tvm-ffi-deep-dive/</guid><description>TVM FFI 的核心骨架：引用计数对象系统、类型擦除容器 TVMFFIAny、Packed Function 调用约定，以及 DLPack 零拷贝张量传递。</description><pubDate>Thu, 28 May 2026 12:00:00 GMT</pubDate></item></channel></rss>