LangGraph is a production-grade Python framework for building stateful, resumable, multi-step LLM agent workflows, distributed as a monorepo of Python packages (core `langgraph`, `langgraph-sdk`, `langgraph-cli`, checkpoint backends, prebuilt agents) plus JS/TS SDK examples. The core is a Pregel-inspired execution engine: `StateGraph` compiles user-defined nodes and conditional edges into a `Pregel` runnable, where typed state channels (`LastValue`, `Topic`, `DeltaChannel`) apply reducer-based updates across supersteps. The engine supports durable checkpointing (in-memory, SQLite, Postgres), interrupts for human-in-the-loop, retry/timeout policies, subgraph composition, and v3 streaming. The CLI can scaffold projects and deploy graphs to LangGraph Cloud or self-hosted runtimes via Docker; the SDK provides sync/async clients that drive remote graphs over HTTP/SSE/WebSocket with typed stream projections (messages, tool calls, lifecycle). This is purely an application/framework repo; the `examples/` notebooks are usage demonstrations, not research pipelines. Primary data flow: user code builds a graph and compiles it; invocation reads channel state from a `BaseCheckpointSaver`, runs eligible nodes (with optional caching/retry), applies channel updates via reducers, writes the new checkpoint per superstep, and emits streamed events through a transformer pipeline (`StreamMux`). The SDK's v3 streaming protocol adds reconnection, subscription fan-out, and scoped subgraph handles over an SSE/WebSocket transport. External dependencies center on `langchain-core` (messages, tools, runnables), Pydantic, `httpx`, `psycopg`, `msgpack`, `orjson`, with optional `redis` and `pycryptodome` for cache/encryption. The codebase exhibits exceptional engineering: extensive unit/integration/conformance tests, a formal conformance suite that pins checkpoint-backend behavior across implementations, security-conscious serialization (strict msgpack/json module allowlists, a patched path-traversal vulnerability in the SDK), and benchmarks. Notable weaknesses are real but contained: heavy sync/async duplication in SDK clients and transports, a fast-moving deprecation path (many renamed APIs, `MessageGraph` slated for removal), and beta features (`DeltaChannel`, `ServerRuntime`) marked unstable. The `@task`/`@entrypoint` functional API, `create_react_agent` prebuilt, and conformance test framework are the most stable and reusable surfaces.
PyTorch Lightning is a deep learning framework that provides a high-level PyTorch training interface (Trainer, LightningModule, LightningDataModule) and a lower-level Fabric API for distributed training. It is a library, not a standalone application. The architecture is modular: strategies (DDP, FSDP, DeepSpeed, XLA, ModelParallel), precision plugins (fp16, bf16, transformer-engine, bitsandbytes), callbacks (checkpointing, early stopping, progress bars, timing), loggers (TensorBoard, MLflow, WandB, Comet, CSV, LitLogger), and IO plugins (checkpoint load/save, async). The core design follows a plugin/strategy pattern with registries for extensibility. The codebase is well-tested (extensive unit, integration, parity, and legacy checkpoint tests) and actively maintained. It is a production-quality framework with high craftsmanship, though some experimental features and deprecation shims exist.
Open-Sora is a video generation framework implementing text-to-video, image-to-video, and video-to-video capabilities using diffusion models and variational autoencoders. The system architecture consists of modular components for models (MMDiT, DC-AE, Hunyuan VAE), datasets with bucket-based sampling, acceleration utilities for tensor/sequence parallelism, and training/inference scripts. Key entry points include training scripts in `scripts/diffusion/train/` and `scripts/vae/train.py`, inference scripts in `scripts/diffusion/inference.py` and `scripts/vae/inference.py`, and a Gradio demo at `gradio/app.py`. The system relies on PyTorch, ColossalAI for distributed training, xformers/flash-attention for efficient attention, and safetensors for model checkpointing. Data flows involve loading video datasets, applying spatiotemporal transforms, training diffusion models with classifier-free guidance, and generating videos through denoising loops.
Jaxformer is a JAX-based library for training large language models (LLMs) on TPUs with data and model parallelism. It provides a complete training loop supporting distributed training across TPU cores, checkpointing, and resumable data loading from TFRecord files. The codebase includes model definitions for transformer-based architectures (like CodeGen), utilities for learning rate scheduling and optimization, and integration with Hugging Face for model conversion and sampling. The system is designed for scalability, supporting models up to 6B parameters via sharding strategies, and includes tools for preprocessing large datasets into TFRecord format. While the core training functionality is tightly coupled to the TPU environment, several components like the resumable data iterator and checkpointing mechanism demonstrate reusable patterns.
This repository is a curated collection of research papers on machine learning systems, organized by topic into directories such as Checkpoint&Memory, Cluster Scheduling, Serving, and Training. Each directory contains PDF files of relevant papers, often accompanied by README.md files providing context or summaries. The repository serves as a reference literature review for researchers and practitioners in the field of ML systems, covering topics like checkpointing, cluster scheduling, model serving, training optimizations, and more. It does not contain any source code or executable components, focusing solely on disseminating academic research.