Choose a maintained stack that supports your exact checkpoint, objective, accelerator, artifact format, and deployment path. Feature lists change faster than the concepts.
Composable Python APIs when you need control over loading, adapters, trainers, and alignment objectives.
Verify: Pin mutually compatible versions and test the exact model architecture, quantizer, and distributed launcher.
Open official documentationConfiguration-driven SFT, preference tuning, and reinforcement-learning recipes across supported models and hardware.
Verify: Start from a maintained example for your model; validate optional kernels and distributed settings on your cluster.
Open official documentationCLI and web interfaces for data preparation, training, evaluation, and export across a broad model matrix.
Verify: Confirm the repository support table and license for the precise model and method you plan to use.
Open official documentationA toolkit focused on optimized local and cloud post-training workflows for its supported combinations.
Verify: Treat vendor speed and memory numbers as workload-specific; reproduce quality and throughput locally.
Open official documentationPyTorch FSDP and DeepSpeed can shard model state and optimizer work. They are scaling layers, not substitutes for data, evaluation, or release management.
MLX and MLX LM support local experimentation on Apple silicon. Check the current model and tuning support before assuming parity with a CUDA recipe.
MLX LM repositoryRecord code, environment, dataset revision, parameters, checkpoints, metrics, and evaluation artifacts. A dashboard does not replace reproducible inputs.
Torchtune examples remain useful historical references, but its official repository says active maintenance wound down in 2025. Do not start a new production dependency on it without accepting that maintenance risk. Read the maintenance notice.