docs(cookbook): add Qwen3.8 GB300 long-context recipe

[Ankur-singh]

Integrate the single-GPU GB300 long-context serving configuration qualified by InferenceX #3816 into the existing Qwen3.8-Flash-Next command builder.

Replace the unverified GB300 / NVFP4 (NVDA) / Low Latency cell with the qualified environment and serving arguments: embedded MTP, ReplaySSM, HiCache DRAM offload, and pinned-memory PLE offload. Pin the stock SGLang image digest and target/draft checkpoint revisions. Keep the memory budget, startup output and qualification notes in the MDX page, and remove its separate static launch command. PLE Auto/On retains the qualified configuration; Off removes the offload flags and marks the modified configuration unverified.

Omit 15 redundant arguments after verifying the pinned SGLang build's CLI defaults and model/backend inheritance. The generated command retains the qualified non-default settings, image digest, checkpoint revisions and serving environment.

This PR changes the per-model config and its MDX page. It can merge independently of #43083.

Run 37714835620, attempt 1 passed all eight concurrency points and the matching real-verification GSM8K eval: 97.35% on 1,319 questions. The generated command uses normal MTP verification; simulated-acceptance throughput is not presented as production performance. The builder's Docker wrapper is adapted from the validated Slurm/Enroot launch and has not been run separately on a GB300 host.

Qualification evidence

• Source head: e67ca6fc9ecf4cd57f27654bda8a3e9059923820 in InferenceX How to contribute an optimized R1 operator in SGlang? #3816.
• Image: lmsysorg/sglang:v0.5.21-cu130@sha256:1f176b3a1dd1205d2a462ab80ea9fc17769a32e53cb6f3c690efa2dc2c384977, built from SGLang e00930c5489053f26d86b179cee0d087f846acbb.
• Target and embedded draft: nvidia/Qwen3.8-Flash-Next-NVFP4 revision fc694b54fb0174e0913e6adf86691ef85a4ead47.
• AgentX concurrency: 1, 2, 4, 8, 16, 32, 48, 64. Matching GSM8K eval job.
• Benchmark-only trace loading, simulated acceptance, client concurrency overrides, result collection, and Slurm allocation controls are excluded from the serving recipe.

Contributor checklist

• Follow the native cookbook config and command-builder conventions.
• Run the applicable pre-commit checks.
• Update documentation; runtime implementation is unchanged.
• Unit tests: not applicable to this documentation-only change.
• Link the existing accuracy and workload qualification evidence, with the simulated-acceptance limitation stated explicitly.

中文

将 InferenceX #3816 已验证的 GB300 单 GPU 长上下文配置集成到现有 Qwen3.8-Flash-Next 命令生成器。

用已验证的环境变量及非默认服务端参数替换尚未验证的 GB300 / NVFP4 (NVDA) / Low Latency 配置单元，保留内置 MTP、ReplaySSM、HiCache DRAM 卸载及固定内存 PLE 卸载，并固定官方镜像摘要和目标及草稿检查点版本。MDX 页面保留内存预算、真实启动日志与验证说明，移除独立的静态启动命令。PLE 选择 Auto 或 On 时保留已验证配置；选择 Off 时移除卸载参数，并将修改后的配置标记为未验证。

根据固定 SGLang 版本的命令行默认值以及模型和后端参数继承规则，省略 15 个重复参数。生成的命令仍保留已验证的非默认设置、镜像摘要、检查点版本和服务环境变量。

本 PR 仅修改该模型的配置文件与 MDX 页面，可独立于 #43083 合并。

运行 37714835620，attempt 1 的八个并发点及对应 GSM8K 评测均通过；使用真实验证的 GSM8K 在 1,319 道题上取得 97.35%。生成的命令使用正常 MTP 验证，不将模拟接受率测试的吞吐量视为生产流量性能。命令生成器的 Docker 包装方式改编自已验证的 Slurm/Enroot 启动方式，尚未在 GB300 主机上单独执行。

AI model disclosure

Prepared with an AI assistant identified by the runtime as the GPT-6 family. The exact model/version could not be verified. Role: source/evidence review, recipe authoring and feedback revisions, local documentation validation, and PR preparation. No delegated agents contributed.

由运行时标识为 GPT-6 系列的 AI 助手协助完成；无法验证精确的模型及版本。工作包括来源与验证证据审查、配方编写与反馈修改、本地文档验证及 PR 准备。没有委派其他代理参与。

---

CI States

Latest PR Test (Base): ✅ Run #37835395177
Latest PR Test (Extra): ❌ Run #37835394473
Latest PR Test (AMD ROCm 10): ➖ No AMD PR run found for this commit.

[Ankur-singh]

@faradawn address the comment above plus some clean-up. Feel free to push directly to the PR. Appreciate your help and feedback.