training-evaluation
Run and troubleshoot HRM CUDA training, checkpoint evaluation, W&B logging, and ARC prediction post-processing.
- License 许可
- Apache-2.0
- 最近更新
- 2026-10-08
- Stars
- ★ 11
- Forks
- 3
01 前言 Front Matter
- name
- training-evaluation
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Run and troubleshoot HRM CUDA training, checkpoint evaluation, W&B logging, and ARC prediction post-processing.
- disable-model-invocation
- true
02 资源文件 3
- 参考文档references/configuration.md2.9 KB
- 参考文档references/troubleshooting.md3.4 KB
- 参考文档references/workflows.md3.1 KB
03 SKILL.md 正文
HRM Training and Evaluation
Use this sub-skill when the task is to launch HRM experiments, adapt Hydra training overrides, evaluate checkpoints, interpret saved prediction shards, or troubleshoot CUDA/W&B/checkpoint failures.
When to use
- The user wants to run
pretrain.py,torchrun ... pretrain.py, orevaluate.py checkpoint=<CHECKPOINT_PATH>. - The task mentions
global_batch_size,epochs,eval_interval, W&B, checkpoint paths,all_config.yaml,*_all_preds.<rank>,eval/exact_accuracy, or ARC top-K aggregation. - A run fails due to CUDA, FlashAttention,
adam_atan2_backend, Hydra overrides, distributed launch, dataset batch sizing, W&B login/offline mode, or checkpoint/config mismatch.
Route map
- Read references/workflows.md for verified train, evaluate, checkpoint, and ARC post-processing commands.
- Read references/configuration.md for Hydra defaults and important overrides.
- Read references/troubleshooting.md for CUDA, W&B, checkpoint, distributed, and evaluation failures.
- Run scripts/check_training_env.py before a long run to validate imports and CUDA dependency readiness without training.
- Use scripts/arc_postprocess.py after
evaluate.pysaves ARC prediction shards.
Baseline commands
Single-GPU Sudoku quick demo after dataset creation:
OMP_NUM_THREADS=8 python pretrain.py \
data_path=data/sudoku-extreme-1k-aug-1000 \
epochs=20000 eval_interval=2000 global_batch_size=384 \
lr=7e-5 puzzle_emb_lr=7e-5 \
weight_decay=1.0 puzzle_emb_weight_decay=1.0
Eight-GPU ARC-1 default training:
OMP_NUM_THREADS=8 torchrun --nproc-per-node 8 pretrain.py
ARC-2:
OMP_NUM_THREADS=8 torchrun --nproc-per-node 8 pretrain.py data_path=data/arc-2-aug-1000
Evaluate an ARC checkpoint:
OMP_NUM_THREADS=8 torchrun --nproc-per-node 8 evaluate.py checkpoint=<CHECKPOINT_PATH>
Bounded readiness check
python <skill>/sub-skills/training-evaluation/scripts/check_training_env.py \
--repo-root /path/to/HRM --require-cuda
This imports pretrain.py and evaluate.py, checks PyTorch CUDA, and verifies
FlashAttention/adam-atan2 backend imports. It does not start training, download
data, contact W&B, or require checkpoints.
Boundaries
- Use
data-preparationfor dataset creation and validation before training. - Use
model-architecturefor model internals, losses, identifiers, and architecture-level CUDA import issues. - This sub-skill treats real HRM training/evaluation as CUDA-required. CPU-only runs are not a verified substitute.
- Do not run full training or checkpoint evaluation as a smoke test unless the user explicitly accepts runtime, GPU, data, storage, and W&B implications.
04 同仓库更多 Skills
obsidian-clipper-template-creator
Guide for creating templates for the Obsidian Web Clipper. Use when you want to create a new clipping template, understand available variables, or format clipped content.
obsidian-clipper-template-creator
Guide for creating templates for the Obsidian Web Clipper. Use when you want to create a new clipping template, understand available variables, or format clipped content.
python-api
Programmatic Object-Detection-Metrics API guidance for BoundingBox, BoundingBoxes, Evaluator, enums, coordinate conversion, AP/mAP semantics, and source-style import diagnostics.
file-evaluation
Folder-based PASCAL VOC AP/mAP evaluation from ground-truth and detection text files.
object-detection-metrics
Guides legacy Object-Detection-Metrics PASCAL VOC AP/mAP evaluation from detection text folders and direct Python API classes.
nx-workspace-patterns
Configure and optimize Nx monorepo workspaces. Use when setting up Nx, configuring project boundaries, optimizing build caching, or implementing affected commands.