Conversation
新增 ATOM 单聚合 worker 的直接服务模式，复用 SGLang 生命周期及原生端口接口。 Signed-off-by: adibarra <93070681+adibarra@users.noreply.github.com>
保留复用的 NVIDIA frontend 代码署名。 Signed-off-by: adibarra <93070681+adibarra@users.noreply.github.com>
cquil11
force-pushed
the
agent/atom-engine-atomesh
branch
6 times, most recently
from
September 23, 2026 17:18
c4058c5 to
8081c5b
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Single-node ATOM recipes currently have to select
frontend.type: atomesh, which adds a router to workloads that previously called the ATOM server directly. Addfrontend.type: atomso one aggregate worker owns the public endpoint, with no AToMesh process or package requirement.The ATOM backend uses the existing frontend port contract to select the public port or its allocated worker port. Direct ATOM and SGLang share the existing single-worker lifecycle; replica and prefill/decode layouts still require a router. Readiness checks
/healthand a nonempty/v1/models, matching ATOM v0.1.2 source.Stacked on #7, which already carries #11 and #1. This is the direct-serving addition only; keep this draft and do not merge the stack yet. Required by InferenceX #3352. Independent of #26.
Validation: 91 focused backend/frontend tests pass; Ruff and schema-doc regeneration checks pass. Full local suite: 3,102 passed, 2 skipped, 6 deselected, 13 failed. All 13 failures reproduce on the unchanged #7 base on macOS (Linux CPU-affinity and profiling-wrapper cases); The fork runs full Linux CI only for PRs targeting main/master, so this stacked PR currently receives title/copyright checks only; full Linux CPU validation remains outstanding. No ATOM hardware qualification is claimed.
中文
目前单节点 ATOM 配方必须选择
frontend.type: atomesh，导致原先直接请求 ATOM 服务的任务多出一层路由。新增frontend.type: atom，由单个聚合 worker 直接监听公共端口，无需启动或安装 AToMesh。ATOM 后端通过已有的 frontend 端口接口选择公共端口或分配给 worker 的端口。ATOM 与 SGLang 复用单 worker 生命周期；多副本和 prefill/decode 拓扑仍需路由。就绪检查要求
/health成功且/v1/models返回非空模型列表，与上述 ATOM v0.1.2 源码一致。此 PR 基于 #7，后者包含 #11 和 #1；本层仅增加直接服务模式。保持草稿，暂不合并依赖链。它用于 InferenceX #3352，与 #26 无关。
验证：91 项后端和 frontend 定向测试通过，Ruff 和 schema 文档再生成检查通过。完整本地测试中 3,102 项通过、2 项跳过、6 项未选中、13 项失败。这 13 项在未修改的 #7 基线上均可复现，涉及 macOS 上的 Linux CPU affinity 和 profiling wrapper；分叉仅对以 main/master 为目标的 PR 运行完整 Linux CI，因此当前依赖链上的 PR 只有标题和版权检查，完整 Linux CPU 验证仍待完成。尚未完成 ATOM 真机验收。