Step-5 Preview 部署指南：vLLM 专用镜像与 verl RL 边界

Step-5 Preview: vLLM Serving and verl RL Guide

Tech-News

Oct 15, 2026

更新于 Oct 15, 2026

🇨🇳 中文

Step-5 Preview 已有专用 vLLM 镜像和服务器部署配方；verl 的强化学习适配目前仍是待合并的草稿 PR。想用它提供 API，可以沿推理配方验证；想训练 LoRA，还需要另一套定制依赖与多节点环境。

你提供的链接被换行截断了，完整入口是：

vLLM 配方：https://recipes.vllm.ai/stepfun-ai/Step-5-Preview
Docker 标签：vllm/vllm-openai:step-5-preview
verl 实现：https://github.com/verl-project/verl/pull/8246
官方 FP8 权重：https://huggingface.co/stepfun-ai/Step-5-Preview
官方 NVFP4 权重：https://huggingface.co/stepfun-ai/Step-5-Preview-NVFP4

本文以 2026 年 10 月 15 日抓取的官方模型卡、vLLM Recipes 源文件与 PR 为依据。专用镜像存在，不代表通用发行版已包含全部能力；PR 的实验结果也不等于框架已经正式支持。

Step-5 Preview 是多大的模型？

官方模型卡描述为约 600B 总参数、每 token 激活约 27B 的稀疏 MoE；vLLM 配方按 checkpoint 张量形状计数标为约 604B。两者统计口径不同，部署时应看实际 checkpoint 与显存规划。

27B 激活参数主要描述计算路径，不能据此按一个 27B 稠密模型估算权重占用。未在某个 token 上激活的专家仍需被部署系统保存或分布到设备，长上下文还要另外容纳 KV cache。

小M把专家权重块分配到四张加速卡，并为 KV cache 留出独立空间，提醒激活参数不等于全部权重

官方资料介绍混合稀疏／滑动窗口注意力与多 token 预测（MTP）。配方给出的上下文上限为 1,048,576 tokens，包含输入与输出；“支持 1M”不意味着可以同时开满上下文、并发和输出长度。

模型介绍还涉及图像和视频能力，但本文只提供文本服务与文本 RL 的验证路径；下面的 curl 请求不能作为多模态接入已经验收的证据。

应该选择哪种硬件与并行方式？

配方路径	官方示例硬件	配置重点
FP8	单节点 4×B300	TP4、TEP4 或 DEP4
FP8	单节点 8×H200	对应 TP8、TEP8 或 DEP8，去掉 Blackwell 的 MegaMoE 参数
NVFP4	单节点 4×B300	切换专用 checkpoint，示例上下文改为 320000
verl LoRA RL	作者报告 4 节点×4 张 GB300	FSDP2 训练、两组 TP8 vLLM rollout、专门的 Step5 支持

TP 是把张量计算分片；EP 分布专家；DP 是多组数据并行处理。TEP 组合 TP 与 EP；DEP 示例使用 TP=1、DP=4，再启用 EP。这些布局要整体选择，不能把几行启动参数随意叠加。

B300 上 TEP/DEP 使用 deep_gemm_mega_moe；纯 TP 配方不选择它，因为该后端需要专家并行。--max-num-seqs 按 DP rank 生效，所以 DEP4 的 64 与 TP4 的 64，不能直接当作同一全局并发预算。

NVFP4 checkpoint 含 NVFP4 和 BF16 层，由配置检测量化方式。官方配方把它的示例上下文设为 320000；不能只替换模型名，就承诺完整 1M 上下文能以同样并发运行。这里没有跨精度质量与速度实测，因而不作性能排名。

如何启动一个可检查的文本 API？

先准备 Linux NVIDIA GPU 主机、可用驱动与容器 GPU 支持，并确认指定卡型、互联和存储容量。以下按官方 4×B300 FP8 的 TP4 示例整理；本文没有实际启动该服务器。

docker pull vllm/vllm-openai:step-5-preview

docker run --rm -it \
  --gpus all --ipc=host --network=host \
  --volume "$HOME/.cache/huggingface:/root/.cache/huggingface" \
  --entrypoint /bin/bash \
  vllm/vllm-openai:step-5-preview

进入容器后，确认权重访问权限，执行：

export CUDA_VISIBLE_DEVICES=0,1,2,3
vllm serve stepfun-ai/Step-5-Preview \
  --host 127.0.0.1 --port 18000 --served-model-name step5 \
  --tensor-parallel-size 4 \
  --data-parallel-backend mp --distributed-executor-backend mp \
  --disable-custom-all-reduce \
  --max-model-len 1048576 --max-num-batched-tokens 16384 \
  --max-num-seqs 64 --gpu-memory-utilization 0.85 \
  --enable-prefix-caching --spec-method mtp --spec-tokens 3 \
  --reasoning-parser step5 \
  --enable-auto-tool-choice --tool-call-parser step5

这保留模型卡的计算参数，监听地址改为本机以便先验收。将服务提供给其他机器时，应按自己的认证与网络方案配置入口。显存不足先降低 --max-model-len 或并发，不能把示例参数当作适用于所有负载的最佳值。

curl -fsS http://127.0.0.1:18000/v1/models
curl -N http://127.0.0.1:18000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  --data '{"model":"step5","messages":[{"role":"user","content":"Hello!"}],"max_tokens":4096,"stream":true}'

注意端口和模型名：这里显式设置 18000 与 step5，因此请求也要对应。Recipes 通用请求示例则使用 8000 和完整仓库名；混用两套示例会产生连接或模型名错误。

验收应依次覆盖短文本、流式结束、reasoning 字段解析、一个无副作用的工具调用，再逐步增加输入长度与并发。启用了 parser 只是配置成立；工具名、参数 JSON 和客户端实际调用仍需端到端验证。MTP 的三个候选 token 也不等于三倍吞吐，收益取决于接受率和硬件负载。

建议保存模型 revision、容器 digest、驱动版本、启动命令与首个成功请求。本文核对 Docker Hub 标签时取得摘要 sha256:8bc189bb3f57183bcaf4f01405b29d4d92d782a182b384aac79b4b2885fb1fda；标签可以更新，复现时应重新核对并固定摘要。

小M匹配 step-5-preview 容器镜像、18000 启动端口与 step5 请求模型，核对成功回执后再继续

verl #8246 已经能直接安装使用了吗？

截至核验时，PR 为 open、Draft、未合并，head 为 136e22c7d02683d90c48a8ff76d1292de43ff394。作者说明，上游通用 vLLM 当时尚不能服务 Step5；rollout 需要独立提交中的模型支持、融合注意力与 sparse-indexer 的 LoRA 服务，以及避免 sleep/wake 主机内存持续增长的备份处理。

因此，推理镜像能运行，不足以证明它同时满足此 PR 的训练依赖。不要只 pip install verl 后就执行脚本，也不要把草稿分支描述成已发布版本。

适配路径是文本模型上的 FP8 checkpoint、FSDP2、GRPO 与 LoRA。训练器从权重构建所需精度的投影，训练和同步 adapter；R3 记录 rollout 的专家路由，并在训练侧重放，以减少两侧路径不一致。它不是对全部 600B 参数进行全量训练，也没有在该实验中验证视觉 RL。

入口脚本为 examples/grpo_trainer/run_step5_preview_dapo_lora_fsdp.sh。需要 Ray 集群、各节点可访问的模型、DAPO-Math-17K 与 AIME-2024 的 Parquet 文件，以及匹配的 vLLM 和 Python 依赖。作者配置 LoRA rank 32、alpha 64、学习率 1e-5、4096-token 响应预算。

固定源码：https://github.com/aoshen02/verl/tree/136e22c7d02683d90c48a8ff76d1292de43ff394

怎样读懂作者报告的 RL 提升？

作者报告在 16 张 GB300 上运行 194 步，并经历两次从 LoRA checkpoint 恢复；AIME-2024 mean@8 从 step 0 的 0.375 到 step 190 的 0.658。这里的限定条件是 4096-token 预算下的正确率。

起点只有 40% 的答案在预算内完成，step 190 提升至 75%；已经完成的答案正确率则由约 94% 变为约 88%，且完成题目集合变大、难度构成变化。两组条件准确率不是同一子集，不能直接推断模型退化；整体分数提升也不能直接概括为一般数学能力同比增长。

更准确的理解是，这次训练让更多答案在指定预算内完成并被判对。作者称平均每训练步约 13 分钟、一次 AIME 验证约 4 分钟；这些是该集群与配置的报告，不能用于估算单卡成本。

小M对齐两把相同的 4096-token 预算尺，对比之前较多未完成解答与之后较多完成解答，RL 实验旁标有草稿旗帜

PR 还报告 131 项测试通过、5 项跳过，但同时说明测试未随 PR 提交。本文没有独立取得并重跑那套测试。示例脚本的数据洗牌方式也与参考实验不同，复现前需要按脚本注释准备相同数据顺序。

常见问题与本文核验范围

是 Apache 2.0 模型吗？

不是。官方权重使用自定义 StepFun Model License 1.0，HF 标记为 other。框架许可不能代替模型许可；商业部署应阅读对应版本全文与适用条件。

许可：https://huggingface.co/stepfun-ai/Step-5-Preview/blob/main/LICENSE.txt

有 Apple Silicon 或普通单卡配方吗？

本文核对的官方 Recipes 路径是大型 NVIDIA 服务器配置，不是 Mac 部署指南。27B 激活参数不意味着只需装下 27B 权重。

本文做了哪些验证？

我们抓取全文，对照 Recipes YAML、模型卡、模型配置、Docker Hub tag 元数据与 PR API，并静态检查启动参数及脚本。没有下载数百 B 权重、启动 GPU 容器、运行推理或复现 RL；硬件与成绩均标明来源。

可结合这篇文章理解预算受限的决策系统，但不能把不同任务指标相互排名：https://blog.mushroom.cv/blog/decision30-multimodal-system-one/

源版本：Recipes 2f6a61c1d2d7f25e4224d3e0b8dd5c1478e7b556；FP8 模型 364b3b03fed2477aa0953fe855a2c4c8c48ee3cc。原材料中的组织首页已补全为具体配方与 PR，避免读者落到不相关页面。

© 2026 Author: Mycelium Protocol. 本文采用 CC BY 4.0 授权——欢迎转载和引用，须注明作者姓名及原文链接，不得去除署名后以原创发布。

🇬🇧 English

Step-5 Preview has a dedicated vLLM container and server recipes. Its verl reinforcement-learning integration remains an unmerged draft PR. Serving an API and training LoRA therefore require separate dependency checks.

The normalized source URLs are:

Recipe: https://recipes.vllm.ai/stepfun-ai/Step-5-Preview
Image: vllm/vllm-openai:step-5-preview
RL PR: https://github.com/verl-project/verl/pull/8246
FP8: https://huggingface.co/stepfun-ai/Step-5-Preview
NVFP4: https://huggingface.co/stepfun-ai/Step-5-Preview-NVFP4

This review cross-checks official sources captured on October 15, 2026. An available dedicated image does not establish feature parity with a generic release, and PR experiments do not establish merged framework support.

How large is Step-5 Preview?

The official model card describes roughly 600B total parameters and 27B active parameters per token. Recipes labels it approximately 604B based on checkpoint tensor shapes. The active count describes computation, while the deployment still has to store or distribute the experts and allocate KV cache.

Xiao-M distributes expert weight blocks across four accelerator cards while reserving separate KV-cache space, showing that active parameters do not represent all stored weights

The model uses hybrid sparse/sliding-window attention and MTP. The recipe specifies 1,048,576 context tokens including input and output. Full context, maximum concurrency and maximum output cannot be assumed to fit together.

Official material also discusses image and video capabilities. The examples here validate a text-serving path and discuss text RL; a successful text request does not validate multimodal integration.

Which hardware and parallel layout should you choose?

Path	Official example	Key details
FP8	4×B300 in one node	TP4, TEP4 or DEP4
FP8	8×H200 in one node	TP8, TEP8 or DEP8; omit Blackwell MegaMoE settings
NVFP4	4×B300 in one node	Separate checkpoint; example context320000
verl LoRA RL	Author reports 4 nodes×4 GB300	FSDP2 trainer, two TP8 rollout engines and custom Step5 support

TP shards tensor computation; EP distributes experts; DP supplies data-parallel ranks. TEP combines TP and EP. DEP4 uses TP1, DP4 and expert parallelism. These are alternative layouts, not a set of independent switches to combine arbitrarily.

Blackwell TEP/DEP uses deep_gemm_mega_moe. Pure TP leaves backend selection to the default because MegaMoE requires expert parallelism. Sequence limits apply per DP rank, which changes the interpretation of concurrency across layouts.

The NVFP4 checkpoint contains NVFP4 and BF16 layers. Its recipe uses a 320000-token context. Switching checkpoints alone does not prove that full 1M context and identical concurrency will fit. We provide no cross-precision performance or quality ranking.

How can you bring up a checkable text API?

Prepare a Linux NVIDIA server with a working driver, container GPU support and the appropriate GPU layout. The following follows the official FP8 TP4 B300 example; we did not execute it on a GPU server.

docker pull vllm/vllm-openai:step-5-preview
docker run --rm -it \
  --gpus all --ipc=host --network=host \
  --volume "$HOME/.cache/huggingface:/root/.cache/huggingface" \
  --entrypoint /bin/bash \
  vllm/vllm-openai:step-5-preview

Inside the container, ensure checkpoint access, then run:

export CUDA_VISIBLE_DEVICES=0,1,2,3
vllm serve stepfun-ai/Step-5-Preview \
  --host 127.0.0.1 --port 18000 --served-model-name step5 \
  --tensor-parallel-size 4 \
  --data-parallel-backend mp --distributed-executor-backend mp \
  --disable-custom-all-reduce \
  --max-model-len 1048576 --max-num-batched-tokens 16384 \
  --max-num-seqs 64 --gpu-memory-utilization 0.85 \
  --enable-prefix-caching --spec-method mtp --spec-tokens 3 \
  --reasoning-parser step5 \
  --enable-auto-tool-choice --tool-call-parser step5

The compute arguments follow the model card; we change the bind address to loopback for initial acceptance. Configure an authenticated network entry point when serving other machines. Reduce context or concurrency when memory is insufficient.

curl -fsS http://127.0.0.1:18000/v1/models
curl -N http://127.0.0.1:18000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  --data '{"model":"step5","messages":[{"role":"user","content":"Hello!"}],"max_tokens":4096,"stream":true}'

Use the matching port and model alias. The generic Recipes request instead uses port 8000 and the full checkpoint name; mixing conventions creates avoidable failures.

Check short text, stream completion, reasoning parsing and a harmless tool call before increasing length and load. Parser flags do not prove client integration. Three speculative MTP tokens do not guarantee triple throughput: acceptance rate and hardware utilization matter.

Save the model revision, image digest, driver, command and successful request. Docker Hub returned sha256:8bc189bb3f57183bcaf4f01405b29d4d92d782a182b384aac79b4b2885fb1fda for the tag during this review. Tags can move; verify and pin a digest for reproduction.

Xiao-M matches the step-5-preview container image, startup port 18000 and request model step5, then checks a successful receipt before advancing

Is verl #8246 ready in a normal release?

At verification, the PR was open, draft and unmerged, at head 136e22c7d02683d90c48a8ff76d1292de43ff394. Its author says generic upstream vLLM did not yet serve Step5 and describes separately upstreamed support for the model, LoRA on fused attention/indexer projections, and host backups that avoid repeated sleep/wake memory growth.

A working serving image alone does not prove these training dependencies are present. Installing a generic verl release is insufficient evidence that the example will run.

The path uses a verl-native text model, FSDP2, GRPO and LoRA over an FP8 checkpoint. It trains and synchronizes adapters. R3 records expert routing during rollout and replays it during training. This is neither full-parameter 600B training nor demonstrated visual RL.

The entry point is examples/grpo_trainer/run_step5_preview_dapo_lora_fsdp.sh. It needs a Ray cluster, model access on every node, DAPO-Math-17K/AIME-2024 Parquet files and matching dependencies. The example uses rank 32, alpha 64, a 1e-5 learning rate and 4096-token responses.

Pinned source: https://github.com/aoshen02/verl/tree/136e22c7d02683d90c48a8ff76d1292de43ff394

What do the reported RL results establish?

The author reports 194 steps on 16 GB300 GPUs, with two resumptions from LoRA checkpoints. AIME-2024 mean@8 rises from 0.375 at step 0 to 0.658 at step 190, under a 4096-token response budget.

Only 40% of initial answers finished within budget, rising to 75%. Accuracy among finished answers changes from roughly 94% to 88%, over a larger and differently composed set. Those conditional scores do not isolate regression; the overall improvement also should not be presented as a general reasoning gain of the same proportion.

The supported interpretation is that more answers finish and score correctly within the specified budget. The reported roughly 13 minutes per step and 4 minutes per validation are specific to that cluster. They are not single-GPU cost estimates.

Xiao-M aligns two identical 4096-token budget rulers, comparing more unfinished solutions earlier with more completed solutions later, beside an RL experiment flag marked draft in Chinese

The author reports 131 tests passing and 5 skipped, while explicitly noting those tests are not included in the PR. We did not obtain and rerun them. The shipped example also shuffles data differently from the reference run; follow its comments before claiming reproduction.

FAQ and verification scope

Is the model Apache 2.0?

No. The official weights use StepFun Model License 1.0; Hugging Face marks it as other. Framework licenses do not replace the checkpoint license. Read the versioned terms for the intended deployment.

License: https://huggingface.co/stepfun-ai/Step-5-Preview/blob/main/LICENSE.txt

Is there a Mac or ordinary single-GPU recipe here?

The verified source describes large NVIDIA server deployments. The 27B active parameter count does not mean only 27B parameters need storage.

What did we actually verify?

We fetched full source text and cross-checked Recipes YAML, the model card/configuration, Docker Hub metadata and the PR API, with static command/script checks. We did not download the full weights, start a GPU container, run inference or reproduce RL.

Related reading on budget-constrained decisions, not a comparable benchmark: https://blog.mushroom.cv/blog/decision30-multimodal-system-one/

Source pins: Recipes 2f6a61c1d2d7f25e4224d3e0b8dd5c1478e7b556; FP8 model 364b3b03fed2477aa0953fe855a2c4c8c48ee3cc.

© 2026 Author: Mycelium Protocol. Licensed under CC BY 4.0 — free to share and adapt with attribution. You must credit the author and link to the original; removing attribution and republishing as original is not permitted.

💬 评论与讨论

使用 GitHub 账号登录后发表评论

关于本站 · 免责声明

🍄 Mushroom Research Blog 是非营利、免费公开的个人科技观察博客与公众号 XStack18，不接受商业合作、不代表任何企业或机构立场，也不谋求商业利益。我们以个人视角客观中立地记录和分析 AI、Web3 等领域的最新模型发布与技术动态——不止转述新闻标题或二手信息，而是给出有独立思考的深入分析，希望帮更多人获得有价值的一手科技认知。

⚠️ 文中介绍的开源代码与模型，仅供学习交流与技术借鉴。它们大多仍处于早期阶段，有待进一步研究和验证，请勿直接用于工作或生产环境；如需采用，请先自行充分测试，并核实其许可证与安全性。
Open-source code and models featured here are shared for learning and reference only. Most are early-stage and still need further study and verification — please don't use them directly in your work or in production. Test them thoroughly and check their licenses and security first.

本站文章均为作者基于公开信息的个人研究与观点整理，不代表文中提及的任何公司、产品、模型的官方立场，未与其构成商业关联或合作关系。
科技行业信息更新极快，我们尽力保证内容准确、及时，但不对完整性、实时性做绝对保证，具体请以相关企业/项目官方公告为准。
文中引用的第三方商标、产品名称、图片、数据等版权归原权利人所有，我们会尽量注明来源；如你认为存在版权疑问或侵权，请通过下方邮箱联系我们，收到通知后会尽快核实处理（更正、加注来源或删除）。
文章内容仅为技术科普与个人观点，不构成投资、法律或其他专业建议，据此进行任何决策的后果需自行判断和承担。

📮 侵权 / 勘误 / 合作咨询：[email protected]