| model-training |
07_Model_Training |
损失函数、优化器、学习率 |
| model-evaluation |
08_Model_Evaluation |
指标、基准、A/B 测试 |
| model-deployment |
09_Deployment_Inference |
部署策略、蓝绿、金丝雀 |
| model-serving |
09_Deployment_Inference |
vLLM、SGLang、模型服务 |
| model-inference |
09_Deployment_Inference |
自回归生成、条件概率、前向传播、采样策略 |
| model-compression |
09_Deployment_Inference |
量化、蒸馏、剪枝 |
| model-precision |
09_Deployment_Inference |
数值精度、模型准确性、FP32/FP16/BF16/FP8/FP4/INT8/INT4 |
| knowledge-distillation |
09_Deployment_Inference |
Teacher-Student、logit蒸馏、DeepSeek-R1蒸馏 |
| mlops |
10_MLOps_Pipeline |
CI/CD、实验追踪、特征存储 |
| rag-systems |
11_RAG_Systems |
向量数据库、混合检索 |
| embedding-models |
11_RAG_Systems |
GTE、bge、MTEB、双塔、交叉编码器 |
| vector-database |
11_RAG_Systems |
Milvus、Qdrant、Chroma |
| ai-architecture |
12_Architecture_Infrastructure |
四层模型、多租户、高可用 |
| llm-infrastructure |
12_Architecture_Infrastructure |
AI Gateway、推理集群 |
| multi-head-latent-attention |
12_Architecture_Infrastructure |
MLA、FlashMLA、KV Cache压缩、DeepSeek |
| kv-cache |
12_Architecture_Infrastructure |
KV Cache、显存墙、五大优化技术族 |
| paged-attention |
12_Architecture_Infrastructure |
PagedAttention、虚拟内存、vLLM |
| radix-attention |
12_Architecture_Infrastructure |
RadixAttention、基数树、SGLang |
| speculative-decoding |
12_Architecture_Infrastructure |
投机解码、Draft-Verify、MTP |
| continuous-batching |
12_Architecture_Infrastructure |
Continuous Batching、动态调度、Orca |
| prefix-caching |
12_Architecture_Infrastructure |
前缀缓存、System Prompt 复用 |
| attention-variants |
12_Architecture_Infrastructure |
GQA、MQA、SWA、注意力变体 |
| training-inference-unification |
12_Architecture_Infrastructure |
训推一体、LeMix、共置调度 |
| heterogeneous-gpu |
12_Architecture_Infrastructure |
异构GPU、国产芯片、统一纳管 |
| flash-attention-kernels |
12_Architecture_Infrastructure |
FlashMLA、FlashInfer、FlashAttention |
| inference-performance |
09_Deployment_Inference |
TTFT、TPOT、吞吐、推理优化 |
| inference-performance-gaps |
09_Deployment_Inference |
推理性能缺口、边缘、异构、能耗 |
| expert-parallelism |
09_Deployment_Inference |
MoE、All-to-All、专家并行 |
| request-scheduling |
09_Deployment_Inference |
Continuous Batching、抢占、SLO-aware |
| inference-autoscaling |
09_Deployment_Inference |
HPA、负载均衡、扩缩容 |
| grouped-query-attention |
12_Architecture_Infrastructure |
GQA、MQA、KV Cache 压缩 |
| flops |
01_基础入门 |
GPU 算力、FLOPS |
| ttft |
09_Deployment_Inference |
首字等待时间、TTFT |
| quantization |
09_Deployment_Inference |
FP8/INT8/INT4、量化 |
| prefill-decode-disaggregation |
09_Deployment_Inference |
PD 分离、Disaggregated Serving |
| rdma-roce |
12_Architecture_Infrastructure |
RDMA、RoCE、GPU 高速网络 |
| gpu-interconnect |
12_Architecture_Infrastructure |
NVLink、NVSwitch、PCIe、HCCS |
| prefill-decode |
12_Architecture_Infrastructure |
Prefill/Decode阶段、TTFT、TPS |
| mixed-precision |
07_Model_Training |
BF16、FP8、AMP、混合精度 |
| rbac |
12_Architecture_Infrastructure |
RBAC、三权分立、访问控制 |
| model-gateway |
12_Architecture_Infrastructure |
AI Gateway、Synapse、负载均衡 |
| rope |
04_NLP_LLMs |
RoPE、旋转位置编码、长度外推 |
| ai-for-science |
20_AI_Applications_Industry |
AlphaFold、药物发现、气象预测、材料设计 |
| distributed-parallelism |
07_Model_Training |
TP/PP/DP/EP、Megatron、DeepSpeed |
| gpu-virtualization |
12_Architecture_Infrastructure |
MIG、GPU共享、算力/显存隔离 |
| federated-learning |
19_Ethics_Safety |
FedAvg、差分隐私、安全聚合、联邦LLM |
| data-cleaning-pipeline |
07_Model_Training |
数据清洗、去重、质量过滤、配比 |
| dora |
05_NLP_LLMs / 07_Model_Training |
权重分解 LoRA、方向微调 |
| rs-lora |
05_NLP_LLMs / 07_Model_Training |
Rank-Stabilized LoRA、小 rank 稳定训练 |
| sglang |
10_Deployment_Inference |
RadixAttention、结构化生成 |
| dynamic-batch-scheduling |
10_Deployment_Inference |
动态批调度、Continuous Batching |
| gguf |
10_Deployment_Inference |
llama.cpp、单文件量化格式 |
| smoothquant |
10_Deployment_Inference |
INT8 量化、激活平滑 |
| tensorrt-llm |
10_Deployment_Inference |
NVIDIA 编译优化、FP8、端到端 |
| code-generation |
16_AI_Coding |
AI 代码生成、补全、测试生成 |
| llm-safety |
17_Ethics_Safety |
LLM 安全、护栏、对齐、红队 |
| bbh |
08_Model_Evaluation |
Big-Bench Hard、复杂推理基准 |
| llm-arena |
08_Model_Evaluation |
Chatbot Arena、人类偏好、Elo 排名 |
| red-teaming |
17_Ethics_Safety / 08_Model_Evaluation |
红队测试、越狱、安全评估 |
| ci-integrated-evaluation |
11_MLOps_Pipeline / 08_Model_Evaluation |
CI 集成评估、回归测试 |
| ab-testing-framework |
11_MLOps_Pipeline / 08_Model_Evaluation |
A/B 测试、在线评估、统计检验 |
| online-evaluation |
08_Model_Evaluation |
在线评估、影子部署、金丝雀 |
| llm-production-pipeline |
11_MLOps_Pipeline |
LLM 生产流水线、MLOps |
| cuda-platform |
12_Architecture_Infrastructure |
CUDA、Tensor Core、NVCC、cuDNN |
| checkpoint |
07_Model_Training |
检查点、分布式容错、Sharded/Full |
| single-tenant-architecture |
12_Architecture_Infrastructure |
单租户、物理隔离、AI Stack |
| sso-saml |
12_Architecture_Infrastructure |
SSO、SAML2、AzureAD、企业认证 |
| apsara-stack |
12_Architecture_Infrastructure |
飞天企业版、Apsara Stack、全栈私有云 |
| model-registry |
12_Architecture_Infrastructure |
模型仓库、版本管理、一键部署 |
| modelscope |
04_NLP_LLMs |
ModelScope 魔搭、SWIFT、中文模型社区 |
| a-speed |
12_Architecture_Infrastructure |
A-Speed 加速推理套件、AI Stack 核心引擎 |
| bailian-exclusive |
12_Architecture_Infrastructure |
百炼专属版、RAG、智能体平台 |
| qwen3-pro |
12_Architecture_Infrastructure |
Qwen3-Pro 专有优化、1.9× 性能 |
| deepseek-models |
04_NLP_LLMs |
DeepSeek R1/V3/V4、MLA/MoE/MTP |
| nvidia-smi |
12_Architecture_Infrastructure |
GPU 监控、nvidia-smi/ppu-smi/rocm-smi |
| torchrun |
07_Model_Training |
分布式训练启动器、弹性训练、DDP |
| ollama |
09_Deployment_Inference |
本地 LLM 运行、GGUF、OpenAI 兼容 API |
| nerdctl |
12_Architecture_Infrastructure |
容器管理 CLI、containerd、Docker 替代 |
| synapse-gateway |
12_Architecture_Infrastructure |
Synapse 模型网关、负载均衡、API-Key |
| apg-gpu |
12_Architecture_Infrastructure |
APG 自研加速卡、CUDA 兼容、700GB/s |
| ascend-npu |
12_Architecture_Infrastructure |
华为昇腾 NPU、CANN、910B/910C |
| deepgemm |
12_Architecture_Infrastructure |
DeepGEMM FP8 算子、Hopper 优化 |
| huggingface-cli |
04_NLP_LLMs |
HF Hub CLI、模型下载/上传/管理 |
| git-lfs |
09_Deployment_Inference |
Git LFS 大文件存储、模型权重版本控制 |
| accelerate |
07_Model_Training |
HF Accelerate、5行代码分布式、FSDP |
| kubectl |
12_Architecture_Infrastructure |
Kubernetes CLI、K8s 运维、Pod 管理 |
| moonshot-kimi |
12_Architecture_Infrastructure |
Moonshot AI / Kimi 长上下文模型 |
| zhipu-glm |
12_Architecture_Infrastructure |
智谱 AI / GLM 模型(ChatGLM 起家) |
| reranker |
12_Architecture_Infrastructure |
重排序模型、Cross-Encoder、bge-reranker |
| qwq |
12_Architecture_Infrastructure |
QwQ-32B 推理模型、CoT 思维链 |
| hygon |
12_Architecture_Infrastructure |
海光国产 x86 CPU、AMD Zen 授权 |
| dualpipe |
12_Architecture_Infrastructure |
DualPipe 双向流水线、DeepSeek 开源 |
| fp8 |
09_Deployment_Inference |
FP8 浮点精度、E4M3/E5M2、Hopper 原生 |
| safetensors |
09_Deployment_Inference |
安全模型格式、替代 pickle、零拷贝 |
| flashinfer |
09_Deployment_Inference |
FlashInfer 注意力算子库、MLSys 2025 Best Paper |
| flashmla |
09_Deployment_Inference |
FlashMLA 注意力加速、DeepSeek MLA 内核 |
| crictl |
12_Architecture_Infrastructure |
CRI 容器调试 CLI、底层容器排查 |
| ppu-smi |
12_Architecture_Infrastructure |
APG GPU 监控工具、对标 nvidia-smi |
| stackops |
12_Architecture_Infrastructure |
AI Stack 专属运维工具、一键部署/诊断 |
| swift |
07_Model_Training |
ModelScope SWIFT 微调框架、100+ 模型 |
| docling |
11_RAG_Systems |
IBM 文档解析工具、PDF/DOCX 结构化提取 |
| mtp |
09_Deployment_Inference |
Multi-Token Prediction、DeepSeek-V3 加速 |
| gradio |
13_Agent_Production |
Gradio ML 应用框架、模型 Web UI |
| 3fs |
12_Architecture_Infrastructure |
DeepSeek 3FS 分布式文件系统 |
| lemix |
12_Architecture_Infrastructure |
LeMix 训推统一调度 |
| rocm-smi |
12_Architecture_Infrastructure |
AMD GPU 监控工具 (ROCm) |
| sentencepiece |
04_NLP_LLMs |
SentencePiece 分词库 (BPE/Unigram) |
| eagle |
09_Deployment_Inference |
EAGLE 特征级推测解码 |
| langflow |
11_RAG_Systems |
LangFlow 可视化 LLM 编排 |
| reward-model |
07_Model_Training |
奖励模型 (RLHF/GRPO 偏好评估) |
| dify |
11_RAG_Systems |
Dify 开源 LLM 应用平台 |
| ragflow |
11_RAG_Systems |
RAGFlow 深度文档理解 RAG 引擎 |
| llama-index |
11_RAG_Systems |
LlamaIndex 数据框架 (索引/查询) |
| medusa |
09_Deployment_Inference |
Medusa 多头推测解码 |
| simpo |
07_Model_Training |
SimPO 简化偏好优化 (无参考模型) |
| qlora |
07_Model_Training |
QLoRA 4-bit 量化 LoRA 微调 |
| ctr |
12_Architecture_Infrastructure |
ctr containerd 原生 CLI |
| streamlit |
13_Agent_Production |
Streamlit 数据应用框架 |
| haystack |
11_RAG_Systems |
Haystack (deepset) Pipeline RAG 框架 |
| flowise |
11_RAG_Systems |
Flowise Node.js 可视化 LLM 编排 |
| opik |
16_AI_Ops |
Opik LLM 可观测性平台 (Comet) |
| chainlit |
13_Agent_Production |
Chainlit 生产级 AI 聊天界面 |
| pissa |
07_Model_Training |
PiSSA 奇异值适配 (SVD 初始化) |
| bitsandbytes |
07_Model_Training |
bitsandbytes 量化优化库 (NF4/8bit) |
| peft |
07_Model_Training |
PEFT 参数高效微调统一框架 |
| onnx |
09_Deployment_Inference |
ONNX 开放神经网络交换格式 |
| openvino |
09_Deployment_Inference |
OpenVINO Intel 推理优化工具包 |
| triton-server |
09_Deployment_Inference |
NVIDIA Triton 推理服务器 |
| exllama |
09_Deployment_Inference |
ExLlamaV2 量化 LLM 推理引擎 |
| colossalai |
07_Model_Training |
ColossalAI 分布式训练框架 |
| rslora |
07_Model_Training |
rsLoRA 秩稳定 LoRA |
| langsmith |
16_AI_Ops |
LangSmith LLM 可观测性平台 |
| ragas |
11_RAG_Systems |
Ragas RAG 评估框架 |
| deepeval |
08_Model_Evaluation |
DeepEval LLM 评估框架 |
| mlflow |
10_MLOps_Pipeline |
MLflow 实验追踪与模型管理 |
| wandb |
10_MLOps_Pipeline |
Weights & Biases 实验追踪 |
| litellm |
12_架构基建/11_AI网关 |
LiteLLM 统一 LLM API 代理 |
| outlines |
09_Deployment_Inference |
Outlines 结构化 LLM 生成 |
| helicone |
16_AI_Ops |
Helicone LLM API 监控 |
| trulens |
08_Model_Evaluation |
TruLens LLM 评估反馈 |
| promptfoo |
08_Model_Evaluation |
Promptfoo Prompt 测试框架 |
| ray-tune |
10_MLOps_Pipeline |
Ray Tune 分布式超参数调优 |
| guidance |
04_NLP_LLMs |
Microsoft Guidance 结构化生成库 |
| lm-format-enforcer |
09_Deployment_Inference |
LM Format Enforcer LLM 输出格式约束 |
| ne-mo |
07_Model_Training |
NVIDIA NeMo 训练与推理框架 |
| lisa |
07_Model_Training |
LISA 层级采样高效微调 |
| miniconda |
01_基础入门 |
Miniconda 轻量级 Python 环境管理 |
| flash-attn |
03_Deep_Learning |
Flash Attention 高效注意力内核 |
| guardrails-ai |
19_Ethics_Safety |
Guardrails AI 安全防护框架 |
| presidio |
19_Ethics_Safety |
Microsoft Presidio PII 检测与脱敏 |
| recursive-self-improvement |
19_Ethics_Safety |
RSI 递归自我改进、智能爆炸、Seed AI |
| sglang-frontend |
09_Deployment_Inference |
SGLang API 服务层 |
| vllm-tp-attention |
09_Deployment_Inference |
vLLM 张量并行注意力机制 |
| detect-secrets |
19_Ethics_Safety |
Yelp detect-secrets 密钥泄露检测 |
| llm-guard |
19_Ethics_Safety |
LLM Guard 安全防护中间件 |
| nemo-guardrails |
19_Ethics_Safety |
NVIDIA NeMo Guardrails 对话控制 |
| torch-tensorrt |
09_Deployment_Inference |
Torch-TensorRT PyTorch 编译器 |
| lm-eval-harness |
08_Model_Evaluation |
LM Evaluation Harness 标准化评估 |
| giskard |
08_Model_Evaluation |
Giskard AI 模型测试与评估平台 |
| huggingface-hub |
07_Model_Training |
Hugging Face Hub AI 模型托管平台 |
| gptcache |
09_Deployment_Inference |
GPTCache LLM 语义缓存引擎 |
| langserve |
12_Architecture_Infrastructure |
LangServe LangChain 一键部署 |
| zep |
13_Agent_Production |
Zep LLM 长期记忆平台 |
| langfuse |
16_AI_Ops |
Langfuse 开源 LLM 可观测性 |
| transformers-js |
09_Deployment_Inference |
Transformers.js 浏览器端 AI 推理 |
| llamaindex-cloud |
11_RAG_Systems |
LlamaIndex Cloud 云端 RAG 平台 |
| phoenix-langsmith |
16_AI_Ops |
Arize Phoenix LLM 可观测性 |
| mem0 |
13_Agent_Production |
Mem0 AI 记忆层基础设施 |
| letta |
13_Agent_Production |
Letta (MemGPT) 有状态 Agent 框架 |
| agentops |
16_AI_Ops |
AgentOps AI Agent 可观测性 |
| humanloop |
04_NLP_LLMs |
Humanloop Prompt 工程与评估 |
| promptlayer |
04_NLP_LLMs |
Promptlayer Prompt 版本管理 |
| arthur-ai |
19_Ethics_Safety |
Arthur AI LLM 安全监控平台 |
| whylogs |
16_AI_Ops |
whylogs 数据质量与 ML 可观测性 |
| feast |
10_MLOps_Pipeline |
Feast 开源特征存储平台 |
| label-studio |
01_基础入门 |
Label Studio 开源数据标注平台 |
| scale-ai |
01_基础入门 |
Scale AI 数据标注与 RLHF 平台 |
| snorkel-ai |
01_基础入门 |
Snorkel AI 弱监督数据编程平台 |
| dataherald |
11_RAG_Systems |
DataHerald 自然语言转 SQL 引擎 |
| dspy |
04_NLP_LLMs |
DSPy Stanford LLM 编程框架 |
| autogen-studio |
13_Agent_Production |
AutoGen Studio 多 Agent 可视化 IDE |
| crewai-tools |
13_Agent_Production |
CrewAI Tools Agent 工具集 |
| smolagents |
13_Agent_Production |
SmolAgents HuggingFace 轻量 Agent |