How higgsfield-ai/higgsfield works
Higgsfield is a distributed GPU orchestration and large-model training framework. A developer defines an experiment, the CLI parses and builds it, optional CI setup provisions GPU nodes, and queued runs execute a distributed data-loader/model training loop. Models use FSDP sharding and checkpoint helpers to persist state or publish to Hugging Face. The tree supports dataset, checkpoint, LLaMA/Mistral, SSH setup, and experiment-parsing stages; queue, workflow publication, and monitoring details are primarily README-documented and their wiring is therefore conservative.
Main entry points: fsdp_checkpoint.py FSDP Checkpointing · main.py Higgsfield CLI · setup.py Node Setup · llama.py LLaMA Models
Components 20
Orchestration
- Higgsfield CLI
higgsfield/internal/main.py - Experiment Queue
higgsfield/internal/launch.py - Node Setup
higgsfield/internal/ci/setup.py
Experiments
- Experiment API
higgsfield/experiment.py - Experiment Parser
higgsfield/internal/experiment/ast_parser.py - Workflow Builder
higgsfield/internal/experiment/builder.py
Distributed Training
- Dataset Transforms
higgsfield/dataset/dataset.py - Distributed Loader
higgsfield/loaders/llama_loader.py - LLaMA Models
higgsfield/llama/llama.py - Mistral Models
higgsfield/mistral/mistral.py
State Persistence
- Checkpoint API
higgsfield/checkpoint/fsdp_checkpoint.py - FSDP Checkpointing
higgsfield/checkpoint/fsdp_checkpoint.py - Experiment Paths
higgsfield/path.py - Checkpoint Store
higgsfield/path.py
External Integrations
- Run Monitoring
higgsfield/internal/launch.py
Other components
- Developer
- Training Script
- GitHub
- GPU Nodes
- Hugging Face
Connections 23
- Developer → Higgsfield CLI: runs commands
- Developer → Experiment API: defines experiment
- Higgsfield CLI → Experiment Parser: parses definitions
- Experiment Parser → Workflow Builder: describes runs
- Workflow Builder → GitHub: generates workflows
- Higgsfield CLI → Node Setup: sets up nodes
- Node Setup → GPU Nodes: installs tooling
- Node Setup → GitHub: configures deploy key
- Developer → Experiment Queue: submits experiment
- Experiment Queue → Training Script: dispatches run
- Training Script → Distributed Loader: loads batches
- Distributed Loader → Dataset Transforms: formats samples
- Training Script → LLaMA Models: trains model
- Training Script → Mistral Models: trains model
- LLaMA Models → FSDP Checkpointing: saves state
- Mistral Models → FSDP Checkpointing: saves state
- Checkpoint API → FSDP Checkpointing: writes state
- Checkpoint API → Experiment Paths: resolves paths
- Experiment Paths → Checkpoint Store: creates directories
- FSDP Checkpointing → Checkpoint Store: writes checkpoints
- LLaMA Models → Hugging Face: loads or publishes
- Mistral Models → Hugging Face: loads or publishes
- GitHub → Run Monitoring: exposes run UI