mrkeyoor.comAsk

THUDM/slimetested 13 Oct 2026commit 05e3b01

slime review

Evaluate slime if you want direct Megatron and SGLang integration with custom rollout and reward workflows.

  • Build 24 s
  • Tests timed out
136 sto install
  • 196 packages
LLM Toolsevaluationupdated 26 Oct 2026
+26stars / 7d

THUDM/slime is an LLM post-training framework that connects Megatron training with SGLang rollout generation. Custom generation, rewards, verifier feedback and environment interactions share its training, rollout and Data Buffer path. It is worth evaluating when those engines fit your infrastructure; our recorded sandbox run establishes setup-step results, not successful GPU training.

Verdict

Evaluate slime if you want direct Megatron and SGLang integration with custom rollout and reward workflows. In our 13 October 2026 run at 05e3b01, install and build steps succeeded, while the test command timed out after 900 seconds. That result leaves suite completion and GPU training unverified; it does not prove slime is broken. Before adoption, run the documented import and GPU checks, then validate a selected model recipe through rollout, reward computation, a training update and checkpoint recovery on your hardware.

We ran it

Lab card: what happened when we ran slime
Lab card from our sandbox run
Screenshot of slime (thudm.github.io/slime)
thudm.github.io/slime
Install✓ · 136s196 packages · 6639 MB
Build✓ · 24s
Tests✗ timed out · 900sran, no count parsed
Known vulns0(pip-audit)
Repo715 files~110,639 lines of source · 12.6 MB · 5 CI workflows · tests dir

Answers from our run

Does slime build from source?

Dependencies installed in 136 seconds (196 packages), and the build succeeded in 24 seconds. We cloned commit 05e3b01 into a clean Debian container with 3 CPUs and no project-specific setup.

Do slime's tests pass?

We could not finish them: the suite was still running after 15 minutes in our container.

Does slime have known vulnerabilities in its dependencies?

pip-audit found none in the dependency tree at the time of our run.

Who should not use slime?

Readers seeking a laptop-only introduction to reinforcement learning: the documented environment starts with a GPU host and a prepared software stack.

What are the alternatives to slime?

veRL, OpenRLHF. Evaluate slime if you want direct Megatron and SGLang integration with custom rollout and reward workflows.

Setup2/5The documented GPU environment, model preparation and networking require substantial work beyond our 136-second install and 24-second build steps. Those steps did not validate training.
Docs4/5The supplied quick start gives Docker and native setup paths, import and GPU checks, persistent-path instructions, hardware caveats and multi-node requirements. We did not execute that full workflow.
Community3/5The supplied release lists multiple contributors, and issue extracts contain detailed technical follow-ups. These support visible participation, but do not measure adoption, support quality or response-time guarantees.
Maturity3/5The README documents debugging, recovery and CPU/GPU testing, and v0.3.1 records maintenance work. Our October run did not complete the suite or validate GPU training, so operational reliability remains unverified here.

Who it’s for

Training teams evaluating Megatron and SGLang together.
Researchers who need custom generation functions, rewards or interactive environments within a distributed post-training workflow.
Engineers prepared to validate model configuration, GPU placement, checkpoint conversion and dependency compatibility on their intended hardware.

Who it’s NOT for

Readers seeking a laptop-only introduction to reinforcement learning: the documented environment starts with a GPU host and a prepared software stack.
Teams that need demonstrated training throughput, convergence or GPU compatibility from this review: our recorded run measured none of those.
Operators who need a completed test-suite result before proceeding: our recorded test command timed out, and exhaustive test counts are unavailable.
Teams choosing primarily for interchangeable rollout engines: slime documents direct SGLang integration; veRL provides a relevant comparison for engine choice.

Setup reality

In our recorded run on 13 October 2026 at commit 05e3b017fed9e02de418c2fbac9d75602a534be5, the install step exited successfully after 135.515 seconds, rounded to 136 seconds in the summary. The record lists 196 dependencies and 6,639 MB of installation footprint. The build step exited successfully after 23.845 seconds, rounded to 24 seconds. The test command timed out after 900.023 seconds with exit code 124; no complete test totals were recorded.

The sandbox configuration was 3 CPUs, 8 GB of memory and ghcr.io/astral-sh/uv:python3.12-bookworm. The install log identifies Python 3.12.12. Successful install and build steps do not establish an installed CLI, working GPU imports, a completed rollout or a training update. The record contains no separate installed-binary check or measured training workload.

The recorded pip-audit advisory total is 0, but audit coverage is unknown. This does not establish dependency safety.

The supplied quick start recommends prepared Docker images containing the compatible SGLang, Megatron and patch stack. Its container command requests GPUs and 16 GB of shared memory. Users must still check imports and GPU visibility, prepare model and data paths, and follow the model download, conversion and training steps. Multi-node networking requires separate configuration.

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.

Alternatives

ProjectWhat it isPick it when
veRL gh↗An LLM reinforcement-learning library documenting multiple training and rollout engines.Evaluate veRL when engine choice matters: its README lists FSDP, FSDP2 and Megatron-LM for training, and vLLM, SGLang and HF Transformers for rollout generation. This is a documented capability comparison, not a measured setup or performance advantage.
OpenRLHFAn RLHF framework documenting Ray and vLLM infrastructure, supervised fine-tuning, reward-model training and RL training.Evaluate OpenRLHF when you want to compare a Ray and vLLM workflow covering SFT, reward models and RL training. The supplied README supports that workflow scope; we did not measure whether it is easier or faster.

What people are saying

  1. [github-trending] THUDM/slime

Sources

  1. slime README
  2. slime quick start
  3. slime v0.3.1 release
  4. Full-vocabulary logits OOM report
  5. Delta weight sync NaN report
  6. GLM-5.2 nonfinite gradients report

More llm tools reviews