8000 Release miniVERL v0.4.0 by DaoyuanLi2816 · Pull Request #31 · DaoyuanLi2816/mini-verl · GitHub
[go: up one dir, main page]

Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
5 changes: 4 additions & 1 deletion CHANGELOG.md
Original file line number Diff line number Diff line change
Expand Up @@ -6,6 +6,8 @@ All notable changes to miniVERL are recorded here. The format follows

## [Unreleased]

## [0.4.0] - 2026-08-02

### Added

- Typed, mask-isolated padded update batches for SFT, off 8000 line KD and strict OPD,
Expand Down Expand Up @@ -493,7 +495,8 @@ Same-tokenizer only; one trajectory per forward pass; `swap` unavailable for
quantized models; only Qwen3 and Qwen2 architectures tested; single-seed GPU
results. The full list is in `docs/limitations.md`.

[Unreleased]: https://github.com/DaoyuanLi2816/mini-verl/compare/v0.3.0...HEAD
[Unreleased]: https://github.com/DaoyuanLi2816/mini-verl/compare/v0.4.0...HEAD
[0.4.0]: https://github.com/DaoyuanLi2816/mini-verl/compare/v0.3.0...v0.4.0
[0.3.0]: https://github.com/DaoyuanLi2816/mini-verl/compare/v0.2.6...v0.3.0
[0.2.6]: https://github.com/DaoyuanLi2816/mini-verl/compare/v0.2.5...v0.2.6
[0.2.5]: https://github.com/DaoyuanLi2816/mini-verl/compare/v0.2.4...v0.2.5
Expand Down
6 changes: 3 additions & 3 deletions CITATION.cff
10000
Original file line number Diff line number Diff line change
Expand Up @@ -2,8 +2,8 @@ cff-version: 1.2.0
title: "miniVERL: On-policy distillation for tool-using agents on one GPU"
message: "If you use miniVERL in your work, please cite it as below."
type: software
version: 0.3.0
date-released: 2026-08-01
version: 0.4.0
date-released: 2026-08-02
license: Apache-2.0
repository-code: "https://github.com/DaoyuanLi2816/mini-verl"
url: "https://github.com/DaoyuanLi2816/mini-verl"
Expand Down Expand Up @@ -98,4 +98,4 @@ references:
url: "https://arxiv.org/abs/2603.07079"
notes: >-
Motivates recording per-token teacher entropy. Entropy-aware divergence
mixing is a roadmap item and is not implemented in miniVERL v0.3.0.
mixing is a roadmap item and is not implemented in miniVERL v0.4.0.
3 changes: 2 additions & 1 deletion PROJECT_STATE.md
Original file line number Diff line number Diff line change
Expand Up @@ -20,7 +20,8 @@ Last updated: 2026-08-02.
| performance result | dual batch-4 reached 3.866 trajectories/s at 3.035 GiB reserved (1.63× sequential); shared batch-4 reached 3.475 trajectories/s at 2.227 GiB (1.54× sequential and 26.6% less memory than dual batch-4, but 10.1% slower) |
| frozen artifacts | result `a302da31af99f1d29f1efd4e6b3dbeb6ea4ac956bba102ca8a1bee8dff0319eb`; profiler `66111cd7fc876cf1befea3297a1a51bcd99252c0bf8989c029381e1dc155a98b`; SVG `98645a668a7832423d28b621262292619615917f037adf7219ff1bf071fb2fea` |
| immutable baseline | calculator benchmark remains byte-identical at `53fc1d4d5b7adee09618d77ad62d4086ba56b78569832d6fc7c3bcd5c2695bbc`; every v0.3 artifact and negative result remains unchanged |
| release state | development PR #30 is open; merge, full post-merge validation, release metadata, `v0.4.0` publication and `0.5.0.dev0` state sync remain pending |
| integration source | Consumer Runtime PR [#30](https://github.com/DaoyuanLi2816/mini-verl/pull/30) was squash-merged as `9914c6d358fd6b0acc5a945c0eab67aebb1b2b51`; synchronized main CI and build runs `30776659196` and `30776659178` are green |
| release state | development integration and post-merge validation are complete; this focused release-metadata change binds exact `0.4.0`, after which annotated-tag publication and `0.5.0.dev0` state sync remain |

## v0.3.0 RecoveryBench release

Expand Down
70 changes: 35 additions & 35 deletions PYPI.md
Original file line number Diff line number Diff line change
@@ -1,5 +1,5 @@
<p align="center">
<img src="https://raw.githubusercontent.com/DaoyuanLi2816/mini-verl/main/docs/banner.svg" alt="miniVERL — auditable online post-training on one GPU" width="880">
<img src="https://raw.githubusercontent.com/DaoyuanLi2816/mini-verl/v0.4.0/docs/banner.svg" alt="miniVERL — auditable online post-training on one GPU" width="880">
</p>

<div align="center">
Expand All @@ -8,21 +8,21 @@
[![Build](https://github.com/DaoyuanLi2816/mini-verl/actions/workflows/build.yml/badge.svg)](https://github.com/DaoyuanLi2816/mini-verl/actions/workflows/build.yml)
[![PyPI](https://img.shields.io/pypi/v/miniverl.svg)](https://pypi.org/project/miniverl/)
[![Python](https://img.shields.io/badge/python-3.10%20%7C%203.11%20%7C%203.12%20%7C%203.13-blue)](https://www.python.org)
[![License](https://img.shields.io/badge/license-Apache--2.0-blue)](https://github.com/DaoyuanLi2816/mini-verl/blob/main/LICENSE)
[![License](https://img.shields.io/badge/license-Apache--2.0-blue)](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.4.0/LICENSE)

</div>

<p align="center">
<a href="https://pypi.org/project/miniverl/"><strong>PyPI package</strong></a> ·
<a href="#single-gpu-quickstart">Install &amp; train</a> ·
<a href="https://github.com/DaoyuanLi2816/mini-verl/blob/main/docs/single-gpu-guide.md">Bring your own GPU</a> ·
<a href="https://github.com/DaoyuanLi2816/mini-verl/blob/v0.4.0/docs/single-gpu-guide.md">Bring your own GPU</a> ·
<a href="#recoverybench-do-fresh-on-policy-states-justify-their-cost">Measured result</a>
</p>

**The independent one-GPU companion for prototyping, diagnosing and validating
online post-training workflows before scaling selected artifacts to verl.**

PyPI `v0.3.0` is the stable release; `main` is development and may be ahead.
PyPI `v0.4.0` is the stable release; `main` is development and may be ahead.

miniVERL is a compact, auditable lab for online teacher-student training on
multi-turn tool trajectories. SFT establishes task and protocol competence;
Expand Down Expand Up @@ -59,7 +59,7 @@ validate artifacts without downloading a multi-gigabyte ML stack; use
[Run the local demo](#local-toy-demo) ·
[Train on your GPU](#single-gpu-quickstart) ·
[Inspect the measured result](#recoverybench-do-fresh-on-policy-states-justify-their-cost) ·
[Read the math](https://github.com/DaoyuanLi2816/mini-verl/blob/main/docs/math.md)
[Read the math](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.4.0/docs/math.md)

## Why miniVERL exists

Expand Down Expand Up @@ -100,7 +100,7 @@ keeps the whole lifecycle in one readable single-GPU process.
| Calculator, JSON-navigation and SQLite environments | yes, deterministic with exact verifiers |
| Exact checkpoint/resume | yes, asserted parameter-for-parameter |
| Self-contained offline HTML report with token-level divergence | yes |
| Ray, FSDP, DeepSpeed, vLLM, VLMs, cross-tokenizer, PPO/GRPO | **no** — see [limitations](https://github.com/DaoyuanLi2816/mini-verl/blob/main/docs/limitations.md) |
| Ray, FSDP, DeepSpeed, vLLM, VLMs, cross-tokenizer, PPO/GRPO | **no** — see [limitations](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.4.0/docs/limitations.md) |

## Consumer Runtime: batch speed without a cluster

Expand All @@ -114,7 +114,7 @@ student adapter, a frozen teacher adapter and an optional frozen reference
adapter. The default remains `dual_model` plus sequential physical batches for
backward compatibility.

![Consumer-runtime throughput versus VRAM](https://raw.githubusercontent.com/DaoyuanLi2816/mini-verl/main/docs/consumer-runtime-v1-pareto.svg)
![Consumer-runtime throughput versus VRAM](https://raw.githubusercontent.com/DaoyuanLi2816/mini-verl/v0.4.0/docs/consumer-runtime-v1-pareto.svg)

On the preregistered RTX 4080 systems workload, physical batch-4 improved
end-to-end throughput by 1.63× for dual models and 1.54× for the shared
Expand All @@ -133,8 +133,8 @@ rollout server or distributed-runtime parity.
Set `train.trajectory_batch_size` to `1`, an integer, or `auto`; choose
`models.runtime: shared_backbone` only when student, teacher and optional
reference use the same pinned base and distinct adapters. See the
[data-bound report](https://github.com/DaoyuanLi2816/mini-verl/blob/main/docs/consumer-runtime-v1.md), [preregistration](https://github.com/DaoyuanLi2816/mini-verl/blob/main/benchmarks/preregistration/consumer-runtime-v1.yaml)
and [frozen result](https://github.com/DaoyuanLi2816/mini-verl/blob/main/benchmarks/results/consumer-runtime-v1.json).
[data-bound report](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.4.0/docs/consumer-runtime-v1.md), [preregistration](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.4.0/benchmarks/preregistration/consumer-runtime-v1.yaml)
and [frozen result](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.4.0/benchmarks/results/consumer-runtime-v1.json).

## RecoveryBench: do fresh on-policy states justify their cost?

Expand All @@ -158,7 +158,7 @@ All three seeds and all completed negative results are retained.
| strict fresh-state OPD | 10.9% | 9.1% | 686.8 s |
| budget-50 fresh-state OPD | 27.3% | 20.7% | 720.8 s |

![RecoveryBench three-seed result](https://raw.githubusercontent.com/DaoyuanLi2816/mini-verl/main/docs/recoverybench/recovery-success.svg)
![RecoveryBench three-seed result](https://raw.githubusercontent.com/DaoyuanLi2816/mini-verl/v0.4.0/docs/recoverybench/recovery-success.svg)

The equal-selected-position view reached the 6,224-position boundary after
eight updates for every core method, so its quality result matches the primary
Expand All @@ -168,9 +168,9 @@ did not reduce wall time because teacher backbone forwards were unchanged. The
evidence**: SFT and frozen KD completed their eight-cycle ceiling, while fresh
OPD crossed the target in one indivisible 88-121 second update.

Read the [full analysis](https://github.com/DaoyuanLi2816/mini-verl/blob/main/docs/recoverybench/recoverybench-v1.md), the
[data-bound technical report](https://github.com/DaoyuanLi2816/mini-verl/blob/main/paper/recoverybench-v1/recoverybench-v1.pdf), or
the [immutable schema-v3 artifacts](https://github.com/DaoyuanLi2816/mini-verl/blob/main/benchmarks/README.md#recoverybench-v1).
Read the [full analysis](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.4.0/docs/recoverybench/recoverybench-v1.md), the
[data-bound technical report](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.4.0/paper/recoverybench-v1/recoverybench-v1.pdf), or
the [immutable schema-v3 artifacts](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.4.0/benchmarks/README.md#recoverybench-v1).
The result is scoped to one Qwen3 pair, one task family, three seeds and one RTX
4080. It does not show that OPD is universally ineffective or that offline KD
always wins.
Expand All @@ -184,18 +184,18 @@ time. Two protocol-naive controls completed normally at 0%; they were not
configuration failures. Both used the ambiguous historical protocol-v1 prompt,
so the failure cannot be attributed solely to intrinsic teacher behavior.

![Two-seed protocol-teacher benchmark](https://raw.githubusercontent.com/DaoyuanLi2816/mini-verl/main/docs/gpu-calc-hard-equal-update-v2.svg)
![Two-seed protocol-teacher benchmark](https://raw.githubusercontent.com/DaoyuanLi2816/mini-verl/v0.4.0/docs/gpu-calc-hard-equal-update-v2.svg)

| Artifact | Role |
| --- | --- |
| [Default recipe](https://github.com/DaoyuanLi2816/mini-verl/blob/main/recipes/qwen_consumer_gpu_calc.yaml) | protocol-qualified default |
| [Schema-v2 benchmark](https://github.com/DaoyuanLi2816/mini-verl/blob/main/benchmarks/results/gpu-calc-hard-equal-update-v2.json) | frozen five-arm result |
| [Raw-teacher recipe](https://github.com/DaoyuanLi2816/mini-verl/blob/main/recipes/qwen_consumer_gpu_calc_raw_teacher.yaml) | historical control; not default |
| [Default recipe](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.4.0/recipes/qwen_consumer_gpu_calc.yaml) | protocol-qualified default |
| [Schema-v2 benchmark](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.4.0/benchmarks/results/gpu-calc-hard-equal-update-v2.json) | frozen five-arm result |
| [Raw-teacher recipe](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.4.0/recipes/qwen_consumer_gpu_calc_raw_teacher.yaml) | historical control; not default |

The teacher gate and downstream comparison reused the same 24-task v0.2 test
set, so this is evidence for qualification in that setup, not a general OPD
advantage. The separate schema-v1 481-second smoke proves the pipeline, not OPD
over SFT. [Full diagnosis and caveats](https://github.com/DaoyuanLi2816/mini-verl/blob/main/docs/rtx4080-baselines.md).
over SFT. [Full diagnosis and caveats](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.4.0/docs/rtx4080-baselines.md).

</details>

Expand Down Expand Up @@ -272,7 +272,7 @@ bf16, while older CUDA cards such as Titan V use fp16. RTX 3070, Titan V,
RTX 4080 and RTX 5090-class cards all enter the same code path; only the
RTX 4080 result is measured here. Exact fit is governed by VRAM, model sizes,
drivers and token budgets, not the card's marketing name. See the
[`single-GPU guide`](https://github.com/DaoyuanLi2816/mini-verl/blob/main/docs/single-gpu-guide.md) before changing the recipe.
[`single-GPU guide`](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.4.0/docs/single-gpu-guide.md) before changing the recipe.

```bash
git clone https://github.com/DaoyuanLi2816/mini-verl.git
Expand Down Expand Up @@ -329,7 +329,7 @@ Layer boundaries are strict, and the first layer never imports torch:
6. `evaluation/`, `reporting/` — measurement.
7. `cli.py` — a thin shell that calls one library function per command.

See [`docs/design.md`](https://github.com/DaoyuanLi2816/mini-verl/blob/main/docs/design.md).
See [`docs/design.md`](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.4.0/docs/design.md).

## Exact versus top-k + tail

Expand Down Expand Up @@ -359,7 +359,7 @@ the teacher still runs a full forward pass to produce th 10000 e hidden states. Reports
therefore say `teacher_queried_position_ratio`, never "teacher compute saved".

Top-k + tail targets are not a new idea — TRL's `ServerDistillationTrainer` has
`loss_top_k` with an optional tail bucket. See [`docs/math.md`](https://github.com/DaoyuanLi2816/mini-verl/blob/main/docs/math.md).
`loss_top_k` with an optional tail bucket. See [`docs/math.md`](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.4.0/docs/math.md).

## Tool-token masking

Expand All @@ -384,18 +384,18 @@ documented — see `tests/unit/test_token_provenance.py`.
## Benchmark results

Every number below was produced by the commands in
[`docs/benchmarking.md`](https://github.com/DaoyuanLi2816/mini-verl/blob/main/docs/benchmarking.md) on the hardware recorded in each
[`docs/benchmarking.md`](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.4.0/docs/benchmarking.md) on the hardware recorded in each
result file. Nothing is estimated or extrapolated.

* **RTX 4080, real models** — [`docs/rtx4080-baselines.md`](https://github.com/DaoyuanLi2816/mini-verl/blob/main/docs/rtx4080-baselines.md)
* **RTX 4080, real models** — [`docs/rtx4080-baselines.md`](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.4.0/docs/rtx4080-baselines.md)
has measured peak VRAM, decode throughput, the full-recipe run, the two-seed
schema-v2 protocol-teacher comparison, and the preserved legacy comparison.
* **CPU, toy models** — `recipes/toy_cpu.yaml` moves task success from 0.0% to
91.7% in 192 s, and `benchmarks/results/` holds the legacy equal-update parity
run.
The parity run's accuracy differences are **within noise**; it exists to show
that all seven arms run to completion under identical budgets, not to rank
them. See [`benchmarks/README.md`](https://github.com/DaoyuanLi2816/mini-verl/blob/main/benchmarks/README.md) for why the toy
them. See [`benchmarks/README.md`](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.4.0/benchmarks/README.md) for why the toy
backend cannot rank methods.

## Installation
Expand Down Expand Up @@ -500,11 +500,11 @@ distribution before handing it over, and asserts that the result still trains.

For a standard frozen PEFT teacher adapter, including the Qwen3 protocol-SFT
recipe, export command, compatibility checks and policy-competence gate, see
[`docs/teacher-adapters.md`](https://github.com/DaoyuanLi2816/mini-verl/blob/main/docs/teacher-adapters.md).
[`docs/teacher-adapters.md`](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.4.0/docs/teacher-adapters.md).

## Limitations

The short version; the full list is in [`docs/limitations.md`](https://github.com/DaoyuanLi2816/mini-verl/blob/main/docs/limitations.md).
The short version; the full list is in [`docs/limitations.md`](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.4.0/docs/limitations.md).

* Same tokenizer only. Cross-tokenizer distillation is rejected with an error.
* Rollout decoding is one sequence at a time. The update path supports padded
Expand Down Expand Up @@ -549,8 +549,8 @@ still retain the local state required for exact resume. Redaction is a
best-effort sharing defense, not permission to place real credentials in any
config, run artifact or report.

See [`docs/reproducibility.md`](https://github.com/DaoyuanLi2816/mini-verl/blob/main/docs/reproducibility.md) and the concise
[`compatibility policy`](https://github.com/DaoyuanLi2816/mini-verl/blob/main/docs/compatibility.md).
See [`docs/reproducibility.md`](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.4.0/docs/reproducibility.md) and the concise
[`compatibility policy`](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.4.0/docs/compatibility.md).

## Roadmap

Expand All @@ -571,7 +571,7 @@ multi-turn tool use — at cluster scale, with Ray. If you have a cluster, use i
miniVERL exists for the case where you have one personal GPU and want to read
every line of what is happening. That can be an older 12 GiB card or a current
high-end card; the repository claims measured performance only for hardware it
actually ran. See [`docs/comparisons.md`](https://github.com/DaoyuanLi2816/mini-verl/blob/main/docs/comparisons.md).
actually ran. See [`docs/comparisons.md`](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.4.0/docs/comparisons.md).

## Citation

Expand All @@ -585,13 +585,13 @@ actually ran. See [`docs/comparisons.md`](https://github.com/DaoyuanLi2816/mini-
}
```

See [CITATION.cff](https://github.com/DaoyuanLi2816/mini-verl/blob/main/CITATION.cff) and [CHANGELOG.md](https://github.com/DaoyuanLi2816/mini-verl/blob/main/CHANGELOG.md).
Contributions: [CONTRIBUTING.md](https://github.com/DaoyuanLi2816/mini-verl/blob/main/CONTRIBUTING.md). Security:
[SECURITY.md](https://github.com/DaoyuanLi2816/mini-verl/blob/main/SECURITY.md).
See [CITATION.cff](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.4.0/CITATION.cff) and [CHANGELOG.md](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.4.0/CHANGELOG.md).
Contributions: [CONTRIBUTING.md](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.4.0/CONTRIBUTING.md). Security:
[SECURITY.md](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.4.0/SECURITY.md).

## License

Apache-2.0. See [LICENSE](https://github.com/DaoyuanLi2816/mini-verl/blob/main/LICENSE) and
[THIRD_PARTY_NOTICES.md](https://github.com/DaoyuanLi2816/mini-verl/blob/main/THIRD_PARTY_NOTICES.md).
Apache-2.0. See [LICENSE](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.4.0/LICENSE) and
[THIRD_PARTY_NOTICES.md](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.4.0/THIRD_PARTY_NOTICES.md).

Chinese translation: [README.zh-CN.md](https://github.com/DaoyuanLi2816/mini-verl/blob/main/README.zh-CN.md).
Chinese translation: [README.zh-CN.md](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.4.0/README.zh-CN.md).
2 changes: 1 addition & 1 deletion README.md
Original file line number Diff line number Diff line change
Expand Up @@ -22,7 +22,7 @@
**The independent one-GPU companion for prototyping, diagnosing and validating
online post-training workflows before scaling selected artifacts to verl.**

PyPI `v0.3.0` is the stable release; `main` is development and may be ahead.
PyPI `v0.4.0` is the stable release; `main` is development and may be ahead.

miniVERL is a compact, auditable lab for online teacher-student training on
multi-turn tool trajectories. SFT establishes task and protocol competence;
Expand Down
2 changes: 1 addition & 1 deletion README.zh-CN.md
Original file line number Diff line number Diff line change
Expand Up @@ -24,7 +24,7 @@
**miniVERL 是一个独立的单卡配套工具:先在本地原型化、诊断和验证在线后训练
流程,再把选定产物扩展到 verl。**

PyPI `v0.3.0` 是稳定发布版;`main` 是开发分支,可能领先于稳定版。
PyPI `v0.4.0` 是稳定发布版;`main` 是开发分支,可能领先于稳定版。

miniVERL 是一个紧凑、可审计的在线师生训练实验室。SFT 负责建立任务能力和
协议能力;OPD 则是在此基础上转移教师的推理、策略、风格或其他行为的在线
Expand Down
Loading
0