Decide on the engines before judging setup
THUDM/slime connects Megatron training to SGLang rollouts. Its README describes custom generation, rewards, verifier feedback and environment interaction flowing through a shared training, rollout and Data Buffer path. This makes the framework relevant to teams that want to extend that particular stack.
Megatron arguments remain directly available. Installed SGLang arguments pass through with an --sglang- prefix. That access is useful when your engineers need to configure the underlying engines, but it also means those engines belong in your evaluation plan. The README's model-support and production-use statements are project claims, not compatibility results from our sandbox.
What our recorded run establishes
Our recorded run took place on 13 October 2026 at commit 05e3b017fed9e02de418c2fbac9d75602a534be5. The configured sandbox had 3 CPUs, 8 GB of memory and image ghcr.io/astral-sh/uv:python3.12-bookworm. The install log identifies Python 3.12.12.
| Recorded check | Observation | Limit |
|---|---|---|
| Install step | Exit 0 after 135.515 seconds, rounded to 136 seconds | The record lists 196 dependencies and 6,639 MB; this is not an installed-binary or GPU check |
| Build step | Exit 0 after 23.845 seconds, rounded to 24 seconds | The available build tail is empty and does not establish which artifacts were produced |
| Test command | Timed out after 900.023 seconds, exit 124 | No completed suite summary or exhaustive passing, failing or skipped counts |
| Dependency audit | Recorded pip-audit advisory total: 0 | Coverage is unknown; this is not proof of safety |
The checkout record lists 715 files, approximately 110,639 lines of source and a 12.6 MB repository footprint. These measurements describe the checkout and setup environment, not training memory or throughput.
The test tail contains a timeout marker, an F marker and later test progress reaching 38%. It also contains thread stacks for a claude_code-adapter. Those fragments do not identify the overall cause of the timeout or establish complete test totals. We cannot infer a product-wide failure, successful GPU execution or convergence from them. No new execution or complete-log reconciliation was performed for this review.
The documented environment goes beyond package installation
The supplied quick start recommends Docker images with a compatible SGLang, Megatron and patch stack. It lists CUDA 13.0 images for H100/H200 and B200/B300 hardware, while stating that GPU CI primarily covers H100/H200. A listed image is not evidence that our run validated any of those accelerators.
The host needs Linux x86_64, Git, Docker, a compatible NVIDIA driver and NVIDIA Container Toolkit. The container command exposes GPUs, uses host IPC and requests 16 GB of shared memory. Inside the container, the guide installs the mounted checkout without replacing pinned dependencies, then checks imports and GPU visibility.
Model weights, data preparation, conversion and training follow those environment steps. For multiple Ray hosts, the guide also requires consistent paths, matching dependencies and reachable networking. The single-host container command does not configure a multi-node network. Hardware and parallelism requirements depend on the selected recipe; the supplied guide does not establish a universal eight-GPU minimum.
Read failure reports in their version and configuration
Historical reports can help define validation work, but their original symptoms should not become blanket claims about the measured checkout. Issue 2209 illustrates the distinction. Its reporter described NaN or infinite rollout weights with delta updates over NCCL on Qwen3.5-122B-A10B. An August 23, 2026 follow-up says that distributed delta-plus-NCCL implementation had been removed and that argument validation then accepted delta updates only with disk transport. Neither report establishes the behavior of our October checkout.
Release v0.3.1 separately documents memory, weight-update and rollout-alignment improvements. Release notes describe changes; our sandbox did not benchmark their effect. For adoption, check the selected image, model, transport and dependency versions together, and reproduce relevant failure cases on that configuration.
Compare workflow fit before assuming an easier alternative
veRL's README lists FSDP, FSDP2 and Megatron-LM training, with vLLM, SGLang and HF Transformers rollout generation. It is a relevant comparison when engine choice drives the decision.
OpenRLHF documents Ray and vLLM infrastructure alongside supervised fine-tuning, reward-model training and RL training. Compare it when that workflow matches your requirements. We did not run either alternative, so this review makes no claim that either needs less hardware, installs faster or produces better results.
How to read the ratings
These are editorial ratings of the supplied evidence, not benchmark results. The rubric is 1 for a substantial barrier or insufficient evidence, 2 for documented capability with substantial setup or validation gaps, 3 for useful supporting evidence with material limits, 4 for detailed evidence across the relevant workflow, and 5 for extensive corroborated evidence with few material gaps. Unknown adoption, compatibility and test coverage do not earn a high score.
Setup receives 2/5 because the documented environment requires GPU infrastructure and model preparation beyond our successful setup steps. Documentation receives 4/5 for concrete environment instructions, verification steps, hardware caveats and multi-node requirements. Community receives 3/5 for visible contributor activity and technical issue follow-ups, without inferring adoption or support response times. Maturity receives 3/5 for documented operating controls and release work, tempered by the absence of completed suite and GPU-workload validation in our run.
Start with the documented import and GPU checks. Then validate one model recipe through a complete rollout and training update, record its resource use, and test checkpoint recovery before committing to a longer job.

