Metadata-Version: 2.4
Name: verl-speco
Version: 0.1.0
Summary: Speculative decoding co-training extensions for verl
Author: verl-SpeCo contributors
License: 
                                         Apache License
                                   Version 2.0, January 2004
                                http://www.apache.org/licenses/
        
           TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
        
           1. Definitions.
        
              "License" shall mean the terms and conditions for use, reproduction,
              and distribution as defined by Sections 1 through 9 of this document.
        
              "Licensor" shall mean the copyright owner or entity authorized by
              the copyright owner that is granting the License.
        
              "Legal Entity" shall mean the union of the acting entity and all
              other entities that control, are controlled by, or are under common
              control with that entity. For the purposes of this definition,
              "control" means (i) the power, direct or indirect, to cause the
              direction or management of such entity, whether by contract or
              otherwise, or (ii) ownership of fifty percent (50%) or more of the
              outstanding shares, or (iii) beneficial ownership of such entity.
        
              "You" (or "Your") shall mean an individual or Legal Entity
              exercising permissions granted by this License.
        
              "Source" form shall mean the preferred form for making modifications,
              including but not limited to software source code, documentation
              source, and configuration files.
        
              "Object" form shall mean any form resulting from mechanical
              transformation or translation of a Source form, including but
              not limited to compiled object code, generated documentation,
              and conversions to other media types.
        
              "Work" shall mean the work of authorship, whether in Source or
              Object form, made available under the License, as indicated by a
              copyright notice that is included in or attached to the work
              (an example is provided in the Appendix below).
        
              "Derivative Works" shall mean any work, whether in Source or Object
              form, that is based on (or derived from) the Work and for which the
              editorial revisions, annotations, elaborations, or other modifications
              represent, as a whole, an original work of authorship. For the purposes
              of this License, Derivative Works shall not include works that remain
              separable from, or merely link (or bind by name) to the interfaces of,
              the Work and Derivative Works thereof.
        
              "Contribution" shall mean any work of authorship, including
              the original version of the Work and any modifications or additions
              to that Work or Derivative Works thereof, that is intentionally
              submitted to Licensor for inclusion in the Work by the copyright owner
              or by an individual or Legal Entity authorized to submit on behalf of
              the copyright owner. For the purposes of this definition, "submitted"
              means any form of electronic, verbal, or written communication sent
              to the Licensor or its representatives, including but not limited to
              communication on electronic mailing lists, source code control systems,
              and issue tracking systems that are managed by, or on behalf of, the
              Licensor for the purpose of discussing and improving the Work, but
              excluding communication that is conspicuously marked or otherwise
              designated in writing by the copyright owner as "Not a Contribution."
        
              "Contributor" shall mean Licensor and any individual or Legal Entity
              on behalf of whom a Contribution has been received by Licensor and
              subsequently incorporated within the Work.
        
           2. Grant of Copyright License. Subject to the terms and conditions of
              this License, each Contributor hereby grants to You a perpetual,
              worldwide, non-exclusive, no-charge, royalty-free, irrevocable
              copyright license to reproduce, prepare Derivative Works of,
              publicly display, publicly perform, sublicense, and distribute the
              Work and such Derivative Works in Source or Object form.
        
           3. Grant of Patent License. Subject to the terms and conditions of
              this License, each Contributor hereby grants to You a perpetual,
              worldwide, non-exclusive, no-charge, royalty-free, irrevocable
              (except as stated in this section) patent license to make, have made,
              use, offer to sell, sell, import, and otherwise transfer the Work,
              where such license applies only to those patent claims licensable
              by such Contributor that are necessarily infringed by their
              Contribution(s) alone or by combination of their Contribution(s)
              with the Work to which such Contribution(s) was submitted. If You
              institute patent litigation against any entity (including a
              cross-claim or counterclaim in a lawsuit) alleging that the Work
              or a Contribution incorporated within the Work constitutes direct
              or contributory patent infringement, then any patent licenses
              granted to You under this License for that Work shall terminate
              as of the date such litigation is filed.
        
           4. Redistribution. You may reproduce and distribute copies of the
              Work or Derivative Works thereof in any medium, with or without
              modifications, and in Source or Object form, provided that You
              meet the following conditions:
        
              (a) You must give any other recipients of the Work or
                  Derivative Works a copy of this License; and
        
              (b) You must cause any modified files to carry prominent notices
                  stating that You changed the files; and
        
              (c) You must retain, in the Source form of any Derivative Works
                  that You distribute, all copyright, patent, trademark, and
                  attribution notices from the Source form of the Work,
                  excluding those notices that do not pertain to any part of
                  the Derivative Works; and
        
              (d) If the Work includes a "NOTICE" text file as part of its
                  distribution, then any Derivative Works that You distribute must
                  include a readable copy of the attribution notices contained
                  within such NOTICE file, excluding those notices that do not
                  pertain to any part of the Derivative Works, in at least one
                  of the following places: within a NOTICE text file distributed
                  as part of the Derivative Works; within the Source form or
                  documentation, if provided along with the Derivative Works; or,
                  within a display generated by the Derivative Works, if and
                  wherever such third-party notices normally appear. The contents
                  of the NOTICE file are for informational purposes only and
                  do not modify the License. You may add Your own attribution
                  notices within Derivative Works that You distribute, alongside
                  or as an addendum to the NOTICE text from the Work, provided
                  that such additional attribution notices cannot be construed
                  as modifying the License.
        
              You may add Your own copyright statement to Your modifications and
              may provide additional or different license terms and conditions
              for use, reproduction, or distribution of Your modifications, or
              for any such Derivative Works as a whole, provided Your use,
              reproduction, and distribution of the Work otherwise complies with
              the conditions stated in this License.
        
           5. Submission of Contributions. Unless You explicitly state otherwise,
              any Contribution intentionally submitted for inclusion in the Work
              by You to the Licensor shall be under the terms and conditions of
              this License, without any additional terms or conditions.
              Notwithstanding the above, nothing herein shall supersede or modify
              the terms of any separate license agreement you may have executed
              with Licensor regarding such Contributions.
        
           6. Trademarks. This License does not grant permission to use the trade
              names, trademarks, service marks, or product names of the Licensor,
              except as required for reasonable and customary use in describing the
              origin of the Work and reproducing the content of the NOTICE file.
        
           7. Disclaimer of Warranty. Unless required by applicable law or
              agreed to in writing, Licensor provides the Work (and each
              Contributor provides its Contributions) on an "AS IS" BASIS,
              WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
              implied, including, without limitation, any warranties or conditions
              of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
              PARTICULAR PURPOSE. You are solely responsible for determining the
              appropriateness of using or redistributing the Work and assume any
              risks associated with Your exercise of permissions under this License.
        
           8. Limitation of Liability. In no event and under no legal theory,
              whether in tort (including negligence), contract, or otherwise,
              unless required by applicable law (such as deliberate and grossly
              negligent acts) or agreed to in writing, shall any Contributor be
              liable to You for damages, including any direct, indirect, special,
              incidental, or consequential damages of any character arising as a
              result of this License or out of the use or inability to use the
              Work (including but not limited to damages for loss of goodwill,
              work stoppage, computer failure or malfunction, or any and all
              other commercial damages or losses), even if such Contributor
              has been advised of the possibility of such damages.
        
           9. Accepting Warranty or Additional Liability. While redistributing
              the Work or Derivative Works thereof, You may choose to offer,
              and charge a fee for, acceptance of support, warranty, indemnity,
              or other liability obligations and/or rights consistent with this
              License. However, in accepting such obligations, You may act only
              on Your own behalf and on Your sole responsibility, not on behalf
              of any other Contributor, and only if You agree to indemnify,
              defend, and hold each Contributor harmless for any liability
              incurred by, or claims asserted against, such Contributor by reason
              of your accepting any such warranty or additional liability.
        
           END OF TERMS AND CONDITIONS
        
           APPENDIX: How to apply the Apache License to your work.
        
              To apply the Apache License to your work, attach the following
              boilerplate notice, with the fields enclosed by brackets "[]"
              replaced with your own identifying information. (Don't include
              the brackets!)  The text should be enclosed in the appropriate
              comment syntax for the file format. We also recommend that a
              file or class name and description of purpose be included on the
              same "printed page" as the copyright notice for easier
              identification within third-party archives.
        
           Copyright [yyyy] [name of copyright owner]
        
           Licensed under the Apache License, Version 2.0 (the "License");
           you may not use this file except in compliance with the License.
           You may obtain a copy of the License at
        
               http://www.apache.org/licenses/LICENSE-2.0
        
           Unless required by applicable law or agreed to in writing, software
           distributed under the License is distributed on an "AS IS" BASIS,
           WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
           See the License for the specific language governing permissions and
           limitations under the License.
        
Project-URL: Repository, https://github.com/verl-project/verl-SpeCo
Requires-Python: >=3.10
Description-Content-Type: text/markdown
License-File: LICENSE
Requires-Dist: hydra-core<1.4,>=1.3.2
Requires-Dist: omegaconf<2.4,>=2.3.0
Requires-Dist: packaging>=24
Dynamic: license-file

# verl-SpeCo: Co-Train to Accelerate RL and Inference

`verl-SpeCo` is a lightweight SPECO drafter-training overlay for
[verl](https://github.com/verl-project/verl). It keeps upstream `verl` as an
import-only dependency and adds speculative decoding drafter collection,
training, and hot-update logic through `verl_speco`.

## Highlights

- **Import-only verl overlay**: composes upstream `verl` PPO/GRPO config and
  runs through `python -m verl_speco.main` without patching the installed `verl`
  tree.
- **Drafter Co-Training in the RL loop**: collects hidden states during rollout or
  old-logprob computation, trains a drafter periodically, and publishes updated
  drafter weights back to the rollout engine.
- **Multiple drafter backends**: includes EAGLE-1, EAGLE-2, EAGLE3, DFlash,
  DSpark, Domino, and P-EAGLE trainer backends under `verl_speco.backends`.
- **vLLM and SGLang integration**: supports EAGLE-1, EAGLE-2, EAGLE3, DFlash,
  and DSpark speculative decoding on vLLM, plus EAGLE3 and DFlash on SGLang,
  with drafter collection and hot-update logic integrated through the rollout
  engine.
- **GPU and NPU examples**: provides example scripts for vLLM, SGLang, and
  vLLM-Ascend style graph settings.
- **Step-level observability**: exposes drafter timing and vLLM speculative
  decoding acceptance metrics, including
  `drafter/spec_decode/mean_acceptance_length`.

## Architecture

![verl-SpeCo architecture](https://raw.githubusercontent.com/verl-project/verl-SpeCo/main/docs/assets/speco-architecture.svg)

## Performance Preview

The current results focus on EAGLE3 with the vLLM rollout engine, where
verl-SpeCo supports both GPU and NPU deployments. The figures below show a
Qwen3-8B EAGLE3 run on vLLM-Ascend/NPU; DFlash support is available, and DFlash
figures will be added in a later update.

On Qwen3-8B with an EAGLE3 drafter on vLLM-Ascend/NPU, a 100-step run shows
that co-training increases mean acceptance length over the fixed-drafter setting
and, compared with the baseline, delivers about 20% faster rollout and 11%
faster end-to-end training without accuracy regression.

| Mean Acceptance Length | Generation Time |
| --- | --- |
| ![Qwen3-8B EAGLE3 mean acceptance length on vLLM-Ascend](https://raw.githubusercontent.com/verl-project/verl-SpeCo/main/docs/assets/qwen3-8b_eagle3_npu_accept-len.png) | ![Qwen3-8B EAGLE3 generation time on vLLM-Ascend](https://raw.githubusercontent.com/verl-project/verl-SpeCo/main/docs/assets/qwen3-8b_eagle3_npu_gen.png) |

| Step Time | Critic Reward |
| --- | --- |
| ![Qwen3-8B EAGLE3 step time on vLLM-Ascend](https://raw.githubusercontent.com/verl-project/verl-SpeCo/main/docs/assets/qwen3-8b_eagle3_npu_step.png) | ![Qwen3-8B EAGLE3 critic reward on vLLM-Ascend](https://raw.githubusercontent.com/verl-project/verl-SpeCo/main/docs/assets/qwen3-8b_eagle3_npu_critic-reward.png) |

## Draft Model Support

| Draft model | Rollout engines | Training engine | Status |
| :---: | :---: | :---: | :---: |
| EAGLE-1 | vLLM | FSDP | Available |
| EAGLE-2 | vLLM | FSDP | Available |
| EAGLE3 | vLLM, SGLang | FSDP | Available |
| DFlash | vLLM, SGLang | FSDP | Available |
| DSpark | vLLM | FSDP | Available |
| Domino | vLLM, SGLang via DFlash | FSDP | Available |
| P-EAGLE | Not wired in this overlay | FSDP | Training only |

EAGLE-1 and EAGLE-2 share vLLM's native EAGLE draft method; EAGLE-2 adds the
dynamic-tree decoding policy over the same draft head.

Standalone GPU training smoke tests for EAGLE-1/EAGLE-2, Domino, and P-EAGLE
are kept under `tests/special_standalone/`. The scheduled/manual
`gpu_drafter_training_smoke` workflow runs them with a configurable target
model, optimizer step count, and learning rate.

Domino is trained with `speculative_algorithm=DOMINO`, but it is served as a
DFlash projector sub-mode. For rollout, use `speculative_algorithm=DFLASH`
with a Domino checkpoint on an engine version that supports the Domino
projector.

P-EAGLE training is available, but its vLLM parallel-drafting rollout runtime
is not wired into this overlay yet. Keep rollout drafter serving disabled and
train or serve the checkpoint separately.

## Runtime Compatibility

The runtime requirements are backend-specific. `REQUIRED_VERL.txt` only pins
the upstream `verl` version; install the matching rollout runtime for the
drafter backend you use.

| Draft model | vLLM | vLLM-Ascend | SGLang |
| :---: | :---: | :---: | :---: |
| EAGLE-1 / EAGLE-2 | Engine version with native EAGLE support | Runtime-specific | - |
| EAGLE3 | &gt;= 0.18.0 | &gt;= 0.18.0 | &gt;= 0.5.10 |
| DFlash | &gt;= 0.20.2 | &gt;= 0.20.2 | &gt;= 0.5.12 |
| DSpark | GPU: [main](https://github.com/vllm-project/vllm/tree/main)<br>NPU: [`dc68bd8`](https://github.com/vllm-project/vllm/tree/dc68bd8c4199b00631fe71eb37313f406cc66ac1) | NPU: [`8214d19`](https://github.com/vllm-project/vllm-ascend/tree/8214d19f8b505484b839469444887b404db2e3a8) | - |
| Domino | DFlash-compatible runtime with Domino projector support | Runtime-specific | Runtime-specific |
| P-EAGLE | Not wired | Not wired | Not wired |

For vLLM DFlash, the drafter checkpoint must use the DFlash draft model config
expected by the runtime.

For vLLM DSpark on GPU, use vLLM main. For vLLM DSpark on NPU, follow the
version pairing documented by
[vLLM-Ascend PR #11153](https://github.com/vllm-project/vllm-ascend/pull/11153):
vLLM
[`dc68bd8c4199b00631fe71eb37313f406cc66ac1`](https://github.com/vllm-project/vllm/tree/dc68bd8c4199b00631fe71eb37313f406cc66ac1)
and vLLM-Ascend
[`8214d19f8b505484b839469444887b404db2e3a8`](https://github.com/vllm-project/vllm-ascend/tree/8214d19f8b505484b839469444887b404db2e3a8).
SpeCo keeps the user-facing algorithm as `DSPARK`; on vLLM-Ascend/NPU it
follows that PR's DSpark path.

## Repository Layout

```text
verl_speco/
  main.py                         # Hydra entrypoint
  config/speco_base.yaml          # shared SPECO/drafter defaults
  config/speco_trainer.yaml       # online PPO primary config
  config/draft_trainer.yaml       # standalone drafter primary config
  trainer/speco_ray_trainer.py    # RayPPOTrainer adapter
  workers/speco_worker.py         # drafter trainer worker
  integration/                    # vLLM, SGLang, old-logprob, publish adapters
  backends/                       # drafter-specific trainer backends
  models/                         # drafter model definitions

examples/                         # end-to-end command examples
tests/                            # CPU-light contract tests
ci/                               # smoke-test helpers and CI notes
```

## Installation

Install the upstream `verl` release branch specified in
[`REQUIRED_VERL.txt`](./REQUIRED_VERL.txt), which is mirrored in
[`verl_speco/config/speco_base.yaml`](./verl_speco/config/speco_base.yaml).
By default, unsupported `verl` versions produce a warning. Set
`VERL_SPECO_STRICT_VERL=1` to fail closed when the importable `verl` does not
match the release/v0.8.0 version and API contract.

One typical editable setup is:

```bash
git clone https://github.com/verl-project/verl.git
cd verl
git checkout release/v0.8.0
pip install -e .

cd ..
git clone https://github.com/verl-project/verl-SpeCo.git
cd verl-SpeCo
pip install -e .
```

The editable install exposes the `verl_speco` package without modifying
`PYTHONPATH`. It also installs the `verl-speco`, `verl-speco-draft-train`, and
`verl-speco-inspect-features` command-line entry points. Install the matching
GPU or NPU rollout runtime separately; `verl-SpeCo` intentionally does not let
pip replace accelerator-specific PyTorch, vLLM, SGLang, or vLLM-Ascend builds.

### Docker Images

You can also build GPU runtime images from the official `verlai/verl`
development images and then use the importable upstream `verl` checkout from
the release/v0.8.0 branch. The Dockerfiles below target GPU deployments; use the
matching accelerator image for NPU or other accelerator runtimes.

For GPU vLLM-based examples, use this Dockerfile:

```dockerfile
# GPU vLLM runtime image.
FROM verlai/verl:vllm023.dev1

ARG VERL_REF=release/v0.8.0
ARG VERL_REPO=https://github.com/verl-project/verl.git

WORKDIR /workspace

RUN git clone ${VERL_REPO} /workspace/verl \
    && cd /workspace/verl \
    && git checkout ${VERL_REF} \
    && pip install -e .

COPY . /workspace/verl-SpeCo

WORKDIR /workspace/verl-SpeCo
RUN pip install -e .
```

Build it from the `verl-SpeCo` repository root:

```bash
docker build -f docker/verl0.8.0/Dockerfile.vllm \
  -t verl-speco:vllm023-verl080 .
```

For GPU SGLang-based examples, use the same layout with the SGLang base image:

```dockerfile
# GPU SGLang runtime image.
FROM verlai/verl:sgl0512.dev1

ARG VERL_REF=release/v0.8.0
ARG VERL_REPO=https://github.com/verl-project/verl.git

WORKDIR /workspace

RUN git clone ${VERL_REPO} /workspace/verl \
    && cd /workspace/verl \
    && git checkout ${VERL_REF} \
    && pip install -e .

COPY . /workspace/verl-SpeCo

WORKDIR /workspace/verl-SpeCo
RUN pip install -e .
```

Build it from the `verl-SpeCo` repository root:

```bash
docker build -f docker/verl0.8.0/Dockerfile.sglang \
  -t verl-speco:sgl0512-verl080 .
```

Install the rollout engine and accelerator runtime that match the script you
intend to run, for example vLLM on GPU, SGLang on GPU, or vLLM-Ascend on NPU.
Those runtime packages are intentionally not pinned by this repository.

## Quickstart

Start from one of the example scripts and replace the model, drafter, dataset,
and checkpoint paths:

```bash
bash examples/run_qwen3-8b_drafter_eagle3_vllm.sh
```

For NPU with vLLM-Ascend-style graph settings:

```bash
bash examples/run_qwen3-8b_drafter_eagle3_vllm_npu.sh
```

The vLLM-Ascend examples keep `FULL_DECODE_ONLY` and dense cudagraph capture
sizes in the launch script so graph behavior is explicit.

All examples use the same entrypoint:

```bash
python -m verl_speco.main
```

The main drafter switches are:

```bash
actor_rollout_ref.rollout.drafter.enable=True
actor_rollout_ref.rollout.drafter.enable_drafter_training=True
actor_rollout_ref.rollout.drafter.model_path=/path/to/drafter
actor_rollout_ref.rollout.drafter.speculative_algorithm=EAGLE3
```

## Separate Draft Model Training

verl-SpeCo also supports a separate draft model training workflow. In this
mode, rollout workers collect drafter training features into a feature store,
and the draft model can be trained separately after feature collection.

Quickstart:

```bash
bash examples/run_qwen3-8b_drafter_separate_training.sh
```

Replace the model, drafter, dataset, feature-store, and checkpoint paths in
the script before running it. The script uses `collect_only` mode for rollout
feature collection and `offline` mode for standalone drafter training.

The main mode values are:

| Mode | Meaning |
| --- | --- |
| `online` | Default. Collects rollout features, trains the drafter inside the online PPO/Ray workflow, and can publish updated drafter weights back to the rollout engine. |
| `collect_only` | Collects rollout features into `feature_store.path` without running drafter training in the PPO/Ray workflow. |
| `offline` | Reads collected features from `feature_store.path` and trains the drafter with the standalone multi-GPU workflow. |

Offline training supports every drafter family the online workers support:
EAGLE-1, EAGLE-2, EAGLE-3, DFlash, DSpark, Domino and P-EAGLE.

Domino and P-EAGLE are training-time families with no engine-level speculative
method of their own (engines serve Domino as a DFlash projector sub-mode, and
P-EAGLE needs the parallel-drafting runtime), so the rollout stage collects
features with the engine algorithm whose hidden-state layout they consume, and
the offline stage trains them from that same feature store:

| Drafter to train | Stage 1 `speculative_algorithm` | Feature layout | Stage 2 `speculative_algorithm` |
| --- | --- | --- | --- |
| Domino | `DFLASH` | `dflash_aux` | `DOMINO` |
| P-EAGLE | `EAGLE3` | `eagle3_aux_plus_last` | `PEAGLE` |

```bash
DRAFT_ALGO=domino bash examples/run_qwen3-8b_drafter_domino_peagle_separate_training.sh
DRAFT_ALGO=peagle bash examples/run_qwen3-8b_drafter_domino_peagle_separate_training.sh
```

Feature stores collected with `speculative_algorithm=DOMINO` before Domino was
mapped to the DFlash layout carry `hidden_states_layout=eagle3_aux_plus_last`
in their sample metadata. DFlash preprocessing fails closed on that layout, so
those stores have to be collected again with `DFLASH`.

Collected feature stores can be inspected before offline training:

```bash
python -m verl_speco.inspect_feature_store /path/to/features \
  --max-samples 200 \
  --show-ok \
  --strict-exit
```

## Configuration

SPECO-specific options live under:

```text
actor_rollout_ref.rollout.drafter.*
```

Important groups:

- `drafter.enable`: enables speculative decoding at rollout time.
- `drafter.enable_drafter_training`: enables online drafter trainer workers.
- `drafter.rollout.*`: controls speculative steps, top-k, and verify tokens.
- `drafter.training.*`: controls hidden-state collection, training interval,
  publish interval, update mode, and DFlash/DSpark-specific training options.
- `drafter.vllm.*`: contains vLLM-specific drafter overrides.

Shared SPECO and drafter defaults are in
[`verl_speco/config/speco_base.yaml`](./verl_speco/config/speco_base.yaml).
The online PPO entrypoint composes them through
[`speco_trainer.yaml`](./verl_speco/config/speco_trainer.yaml), while standalone
feature-store training uses
[`draft_trainer.yaml`](./verl_speco/config/draft_trainer.yaml).

## Testing

CPU-light contract tests can be run with:

```bash
pip install -r ci/requirements-ci.txt
pytest tests
```

Some tests require an upstream `verl` checkout from `release/v0.8.0`. Set
`VERL_SPECO_UPSTREAM_ROOT` to the root of that checkout when running the config
composition contract:

```bash
export VERL_SPECO_UPSTREAM_ROOT=/path/to/verl
pytest tests/config/test_speco_config_overlay.py
```

Hardware smoke tests are kept under `ci/` and are intended for self-hosted GPU
or NPU runners with matching model paths and runtime packages.

## Community

Scan the QR code below to join the verl-SpeCo Lark user group.

<p align="center">
  <img src="https://raw.githubusercontent.com/verl-project/verl-SpeCo/main/docs/assets/Lark_QR_code.jpg" alt="verl-SpeCo Lark user group QR code" width="220">
</p>

## Contributing

Keep changes scoped to the overlay whenever possible. If a change requires
upstream `verl` behavior, prefer adding a compatibility adapter in
`verl_speco.integration` and document the supported `verl` version in
`REQUIRED_VERL.txt`.

Before proposing changes upstream or opening a PR, follow the repository rules
in [`AGENTS.md`](./AGENTS.md), including duplicate-work checks and test
reporting.

## Acknowledgements

This project builds on [verl](https://github.com/verl-project/verl),
[vLLM](https://github.com/vllm-project/vllm),
[SGLang](https://github.com/sgl-project/sglang), and
[vLLM-Ascend](https://github.com/vllm-project/vllm-ascend).
