Diagnoses and tunes Relax on-policy distillation teacher engine configurations and GPU splits.
npx skills add https://github.com/redai-infra/Relax --skill opd-tuning
Please help me install this Agent Skill. Skill: opd-tuning Source: https://github.com/redai-infra/Relax/tree/main/skills/opd-tuning Command: npx skills add https://github.com/redai-infra/Relax --skill opd-tuning
On-policy distillation (OPD/MOPD) runs in Relax often suffer from teacher engine OOMs, teacher timeouts, and idle student or teacher GPUs caused by misconfigured SGLang flags, logits memory peaks, and unbalanced GPU partitioning. This Skill analyzes an OPD launch script and produces a cited diagnosis report with concrete, ordered fixes.
Run the opd-tuning skill on my OPD launch script examples/on_policy_distillation/mopd/run-mopd-qwen35-9b-8xgpu-colocate.sh and diagnose why the teacher engine is OOMing.
High-intent search queries and answers about installing and using this skill.
Teacher OOM in OPD usually comes from the full-vocabulary logits tensor sized N_positions times vocab/TP times 4 bytes. Enable SGLang logits chunking first, then raise chunked-prefill-size, and only then consider raising teacher TP or repartitioning GPUs.
Measure both sides with nvidia-smi sampling or engine log time windows, compute the work ratio r, then assign teacher GPUs approximately G*r/(1+r). The split must satisfy rollout plus teacher equals actor GPUs and both sides' per-engine divisibility constraints.
Relax rebuilds each actor's runtime env and only copies variables explicitly listed in RELAX_PROPAGATE_ENV_VARS. Scripts submitted via ray job submit must also merge the variable into the runtime-env JSON, since the driver does not inherit the submitting shell.
The teacher is pure prefill, so decode CUDA graphs are never replayed and disabling them saves capture time and memory. Prefill graphs are actually used, so keep them unless measurement shows otherwise; newer SGLang versions split these into per-phase flags.
It only covers OPD-specific teacher-side issues; general actor training performance, TP/PP/CP, recompute, and hang debugging belong to other skills. It also cannot rely on teacher-side metrics because none exist, so diagnosis depends on logs and GPU sampling.
dongg622Monitor, configure, and troubleshoot GPU hardware in data center environments.
BaneeishaqueCategorize IDE noise files in a commit and restore them via commit edit.
NVIDIA-NeMoGuides context parallelism sizing and recompute strategies for long-context MoE training in Megatron Bridge.
EnkeZMigrate RL training recipes from verl to Relax frameworks.
