Get a dataset into the right shape for a Halo training method, and pre-process it (tokenize / pack / shard) when worth it. USER-INVOKED ONLY. Use when the user asks "what format does SFT/DPO/GRPO/reward/embedding expect", "prepare/tokenize/pack/shard my…
Use when writing, editing, or reviewing any page under agent-docs/, or when a code change in src/ needs its owning doc page updated (doc actualization). Authors and maintains the Halo agent-docs/ tree and enforces the anti-AI-slop style charter.
Recommend throughput/memory levers to raise tokens/s/GPU or cut peak memory for a given Halo training config — and call out the levers that do NOT help at fine-grained MoE shapes (low-precision fp8/fp4 compute, native DeepGEMM, torch.compile on EP MoE,…
Wire up online GRPO (RLVR) or async GRPO with environments (multi-turn tool-use) for Halo: bring up the separate rollout container — vLLM 0.26.0 (cu13), or SGLang 0.5.17 for async GRPO with environments (rollout_backend: sglang, the families its loaders can…
Add support for a new model family to Halo — the EP (expert-parallel) wrapper, the Liger coverage spec, router-balancing choice, and vendoring (with a removal-on-merge policy). USER-INVOKED ONLY. Use when the user asks to "add a new model", "support <model>…
Turn a finished Halo training run into a usable artifact, and resume correctly. USER-INVOKED ONLY. Use when the user asks to "merge EP shards", "load my EP/MoE checkpoint into vLLM", "merge a LoRA/QLoRA adapter", "convert to bf16", "quantize to fp8/fp4",…
Diagnose distributed-training failures in Halo — hangs/deadlocks, CUDA OOM, loss=NaN/Inf, DeepEP faults, and NCCL collective timeouts — by routing the symptom to the exact opt-in helper, env var, or known failure mode. Auto-fire when a training run…
Assemble and run a Halo training or test job inside the prebuilt Docker image. Picks the image (halo:blackwell for B200/B300 vs halo:hopper for H100/H200), sets GPU count, mounts ($(pwd)->/workspace, the host's large scratch volume, ~/.aws), passes --env-file…