8000 mini-verl/PROJECT_STATE.md at main · thinglinks/mini-verl · GitHub
[go: up one dir, main page]

Skip to content

Latest commit

 

History

History
74 lines (59 loc) · 3.87 KB

PROJECT_STATE.md

File metadata and controls

74 lines (59 loc) · 3.87 KB

PROJECT_STATE

Current maintainer handoff for miniVERL (mini-verl package, miniverl CLI). release-state.yaml is the canonical version source; this page indexes current product and evidence state rather than repeating release history.

Last updated: 2026-08-16.

Canonical release state: stable v0.10.1 (2364e9b8e8b550f44d1b66a77fc2d407b76b05b5), development 0.10.2.dev0.

Release state

Current product boundary

The primary product is verl-opd-v0.8-single-gpu-v1: a typed, resolved subset of official verl v0.8.0 at commit 7aed6b230776f963fa09509c10d9c3a767d1102c. It runs one actor, one teacher, n=1, reward-free direct GKD forward_kl_topk and token-mean reduction in one local process on one NVIDIA CUDA GPU.

The stable v0.9.1 repair makes dtype, quantization, attention, student adapter input, logical/physical batch limits and placement legality explicit. Executable compatibility claims are mutation-tested and recorded in docs/generated/verl-opd-v0.8-field-effects.json. Quantized roles cannot swap; unknown-size quantized roles require proof instead of receiving an executable plan.

Development 0.10.2.dev0 has a closed typed profile registry and torch-free compatibility introspection. Profile-scoped plans, caches, checkpoints and exports bind an independent identity. The direct-GKD and sampled-k1 vanilla policy-loss profiles both have pinned conformance and measured RTX 4080 paths. Portable hardware records preserve measured, estimated and unknown states; schema-valid community submissions remain unreviewed until maintainer validation. The release chain builds one hosted-runner candidate and binds the maintainer's RTX 4080 qualification to that exact manifest and wheel. Release jobs may publish only the same verified bytes and never rebuild them. Future GitHub Releases use a canonical, explicit evidence layout with four principal JSON records and one deterministic subordinate-evidence archive; historical v0.10.1 asset names remain immutable.

Arbitrary verl YAML, other policy-gradient modes, rewards, PPO/GRPO, Ray, FSDP, Megatron, multi-GPU and distributed execution remain unsupported. The legacy environment/PPO artifact bridge is migration-only.

Current evidence

Evidence Status
Qwen3-0.6B/1.7B developer workload 32 prompts, 8 current-policy updates, 3.1914 GiB peak reserved on one RTX 4080; matched interruption/resume was byte-identical
Qwen3 sampled-k1 PG 32 prompts, 8 updates, exact interruption/resume, 3.1914 GiB peak reserved; no quality comparison
SmolLM2-360M/1.7B direct GKD 32 prompts, 8 updates, 1.4961 GiB peak reserved; exact resume, PEFT reload and materialized export passed
Ubuntu 26.04 WSL2 plan, bounded probe, rollout, teacher scoring, one update and PEFT reload measured on the same RTX 4080
external alignment v1 preregistered early stop: 0 selected checkpoints, teachers, continuation arms or final-test accesses
distributed verl execution not tested

Frozen benchmark and task-level result JSON/JSONL remain immutable. The calculator schema-v2 source stays at SHA-256 53fc1d4d5b7adee09618d77ad62d4086ba56b78569832d6fc7c3bcd5c2695bbc.

Maintainer entry points

0