Create MoE model skeletons for PithTrain and wire them into training and inference workflows.
npx skills add https://github.com/mlc-ai/pith-train --skill add-new-model
Please help me install this Agent Skill. Skill: add-new-model Source: https://github.com/mlc-ai/pith-train/tree/main/.agents/skills/add-new-model Command: npx skills add https://github.com/mlc-ai/pith-train --skill add-new-model
Enables seamless addition of a new MoE language model to PithTrain, covering model implementation, framework wiring, and optional checkpoint conversion to run full training and inference cycles.
Add a new MoE model to PithTrain by creating the model skeleton, wiring it into the training framework, optionally adding a checkpoint converter, and validating with the standard phase ladder.
High-intent search queries and answers about installing and using this skill.
Adding a new MoE model to PithTrain requires creating the model skeleton, wiring it into the training framework, optionally adding an HF2dcp checkpoint converter, and validating the workflow using the standard phase ladder from single-GPU to distributed setups.
Converting Hugging Face checkpoints for distributed training involves implementing an optional HF2dcp or dcp2hf converter to align model layouts or quantify weights, ensuring correct round-trips and naming parity between frameworks.
Checkpoint converters are optional when adding a new model to PithTrain, but you must implement an HF2dcp or dcp2hf converter when necessary to align layouts or quantify weights for quantized checkpoints.
PithTrain validates MoE model training and inference by executing a standard phase ladder that starts with an HF reference inventory, progresses through distributed training checks, and enforces diagnostic checks to ensure correct round-trips.
Model integration fails phase-based gating checks when the implementation lacks naming parity with Hugging Face or violates the diagnostic checks described in the reference files required to ensure correct round-trips.
nebiusConfigure the Cosmos3 framework by validating source repositories and staging model checkpoints.
zhuangbiaoweiTrain Mixture of Experts models using DeepSpeed and HuggingFace Transformers.
Supporter09Optimize large-scale MoE reinforcement learning training with Megatron parallelism and FP8 or INT4 quantization.
richardnguyen0715Simplify distributed training in PyTorch models with a unified API.
