docs(waiver): add the PR #3631 Mooncake EFA package waiver

The B300 DSXE disaggregated recipes replace the image's
mooncake-transfer-engine-cuda13 0.3.13 with the public EFA build
mooncake-transfer-engine-efa-cuda13 0.3.13.post1 at container setup, as
SGLang's AWS EFA documentation prescribes, because the nodes connect over
AWS EFA and the standard build lacks the EFA transport. Document what is
changed, why the unmodified image cannot run the disaggregated points,
the upstream links, and the removal plan.

B300 DSXE 分离式配方在容器初始化时，按照 SGLang 的 AWS EFA 文档，将镜像中的
mooncake-transfer-engine-cuda13 0.3.13 替换为公开的 EFA 版本
mooncake-transfer-engine-efa-cuda13 0.3.13.post1：节点之间通过 AWS EFA 互联，
而标准版本不包含 EFA 传输。本豁免说明所做的改动、未经修改的镜像无法运行分离式
测试点的原因、上游链接以及移除计划。

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Author: nvpohanh

inferencex-e2e/docs/waiver/3631.md

@@ -0,0 +1,50 @@
+# Serving-stack package waiver — PR #3631
+
+Filed per [`docs/PR_REVIEW_CHECKLIST.md`](../PR_REVIEW_CHECKLIST.md) (no-patch item): the disaggregated
+recipes in this PR replace one package in the pinned serving image at container setup. The checklist
+requires the pinned image to run as shipped unless such a change is covered by a filled-out waiver.
+
+## Config covered
+
+- **Master config entry:** `dsv4-fp4-b300-dynamo-sglang-agentic-disagg` in
+  [`configs/nvidia-master.yaml`](../../configs/nvidia-master.yaml) (1P1D DEP8 c64 and c240, 1P2D DEP8 c64).
+  The aggregate entry `dsv4-fp4-b300-dynamo-sglang-agentic-agg` does not run the script.
+- **Pinned image:** `lmsysorg/sglang:nightly-dev-20260916-c9a8fba9`
+- **Setup entrypoint:** `setup_script: install-mooncake-efa-cu13.sh` in
+  [`disagg-1p1d-dep8-dep8-c64-mtp-kvoffload.yaml`](../../benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/b300-fp4/agentx/disagg-1p1d-dep8-dep8-c64-mtp-kvoffload.yaml),
+  [`disagg-1p1d-dep8-dep8-c240-mtp-kvoffload.yaml`](../../benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/b300-fp4/agentx/disagg-1p1d-dep8-dep8-c240-mtp-kvoffload.yaml), and
+  [`disagg-1p2d-dep8-dep8-c64-mtp-kvoffload.yaml`](../../benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/b300-fp4/agentx/disagg-1p2d-dep8-dep8-c64-mtp-kvoffload.yaml)
+- **Script:** [`benchmarks/multi_node/srt-slurm-recipes/configs/install-mooncake-efa-cu13.sh`](../../benchmarks/multi_node/srt-slurm-recipes/configs/install-mooncake-efa-cu13.sh)
+
+## What is changed
+
+The script uninstalls the image's `mooncake-transfer-engine-cuda13` 0.3.13 and installs the public EFA
+build of the same Mooncake release line, `mooncake-transfer-engine-efa-cuda13==0.3.13.post1`, from PyPI
+with `--no-deps`. Both packages provide the same `mooncake` Python module; the EFA build adds the AWS EFA
+(libfabric) transport. The commands are the ones in SGLang's documentation.
+
+Nothing else changes. No SGLang, Dynamo, or Mooncake source is patched, no forked or rebuilt wheel is
+installed, and no other package is added or removed. Mooncake only moves KV-cache pages from prefill to
+decode workers; it performs no model computation, so the change cannot alter model numerics.
+
+## Why the unmodified upstream image cannot run this benchmark
+
+The B300 DSXE nodes connect to each other over AWS EFA; they have no InfiniBand fabric. The image's
+standard `mooncake-transfer-engine-cuda13` build does not include the EFA transport, so the disaggregated
+recipes cannot transfer KV cache between the prefill and decode nodes with it. SGLang's PD-disaggregation
+documentation prescribes replacing the standard CUDA 13 Mooncake package with the EFA build on AWS EFA
+instances, using these exact commands, and the recipes set `MOONCAKE_PROTOCOL=efa` as that documentation
+describes.
+
+## Upstream PR / issue links
+
+- SGLang documentation, AWS EFA section: https://docs.sglang.io/docs/advanced_features/pd_disaggregation#aws-efa,
+  added in https://github.com/sgl-project/sglang/pull/39973
+- Mooncake EFA wheel: https://pypi.org/project/mooncake-transfer-engine-efa-cuda13/0.3.13.post1/
+
+## Removal plan
+
+Retire `install-mooncake-efa-cu13.sh` once an upstream `lmsysorg/sglang` image ships an EFA-capable
+Mooncake build. At that point, bump the pinned `image:` for `dsv4-fp4-b300-dynamo-sglang-agentic-disagg`
+to that image, remove the `setup_script` lines from the three disaggregated recipes, delete the script,
+and delete this waiver in the same PR.