القائمة
RLinf: Validate RLinf training examples end-to-end from container tests to single-step training
يُحدَّد Prompt الذي يراجع المصدر أولًا بشكل افتراضي. يمكنك التبديل إلى أمر مباشر أو تنزيل نسخة محلية.
اقرأ SKILL.md وأي ملفات مرافقة يعرضها SkillsMP قبل أن تقرر التثبيت.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
يتجاوز الأمر المباشر Prompt المخصّص للمراجعة. افحص المصدر قبل تشغيله.
npx skills add https://github.com/aws-samples/sample-rlinf-on-eks --skill rlinf-functional-testingيبقى الأمر في سطر واحد. مرّر أفقيًا لمراجعته كاملًا قبل النسخ.
تفضّل نسخة محلية؟ نزّل الملفات المتاحة حاليًا لدى SkillsMP.
Select optimal EC2 GPU instance types for Physical AI RL training based on VRAM, interconnect, and cost requirements
Leverage Kubernetes-native features for GPU training including GPU Operator, topology scheduling, and priority classes
Deploy and manage Physical AI RL training jobs on EKS including CodeBuild CI, GitOps, and experiment management
استنادا إلى تصنيف SOC المهني
عرض SKILL.md
| name | rlinf-functional-testing |
| description | RLinf: Validate RLinf training examples end-to-end from container tests to single-step training |
RLinf-specific test configurations for the validation harness (L2-L6). These complement the generic testing framework in the shared functional-testing skill with concrete RLinf values.
The container test (infrastructure/manifests/container-test.yaml) runs 10 tests verifying:
import verl, from flash_attn import flash_attn_func)MUJOCO_GL=osmesa)The model download jobs (examples/<example>/manifests/model-download.yaml) download models to FSx:
| Model | FSx Path | Size |
|---|---|---|
RLinf/Openvla-oft-SFT-libero10-trajall | /fsx/models/openvla-oft-sft-libero10 | ~14 GB |
RLinf/RLinf-Pi0-SFT-Spatial-Object-Goal | /fsx/models/rlinf-pi0-sft-spatial | ~14 GB |
Each example runs 1 PPO training step via HYDRA_OVERRIDES="runner.max_steps=1 runner.save_interval=-1 runner.val_check_interval=-1".
Defined in tests/lib/cluster.sh L5_EXAMPLE_LIST:
| Example Name | VENV_NAME | CONFIG_NAME | MODEL_PATH |
|---|---|---|---|
maniskill-openvlaoft | openvla-oft | maniskill_ppo_openvlaoft_quickstart | /fsx/models/openvla-oft-sft-libero10 |
libero-pi0 | openpi | libero_spatial_ppo_openpi_quickstart | /fsx/models/rlinf-pi0-sft-spatial |
./validate.sh --mode cluster --level 5 --skip-to 5 --example maniskill-openvlaoft
In the pod logs, look for:
TRAINING_STEP_TEST_PASSED -- explicit success markerGlobal Step: 1/1 -- training step completedTraceback, ERROR:, or FAILED in output| Symptom | Cause | Fix |
|---|---|---|
| OOM during FSDP init | GPU too small for 7B model | Use p-class instances (not g6) |
ModuleNotFoundError: verl | Wrong venv activated | Check VENV_NAME matches example |
input() hang (no output) | LIBERO import without LIBERO_DATASET_DIR | Set env var and pre-create config |
FileNotFoundError: config.json | Model not downloaded | Run L4 first |
Uses infrastructure/manifests/training-statefulset.yaml + training-service.yaml. Requires 2+ GPU nodes. Ray head starts on pod-0, workers connect via headless service training-svc.
The test patches HYDRA_OVERRIDES onto the StatefulSet for single-step validation.
functional-testing -- generic testing framework and validation levelsrlinf-pytorch-training-scripts -- Hydra configs and training entrypointsrlinf-dataset-preparation -- model download details