Repository navigation
[AgentX] Enable unified memory for B200 Muse Spark / [AgentX] 为 B200 Muse Spark 启用 unified memory - #4366
[AgentX] Enable unified memory for B200 Muse Spark / [AgentX] 为 B200 Muse Spark 启用 unified memory#4366Fridge003 wants to merge 3 commits into
Conversation
更新 B200 Muse Spark 镜像，保留 cc64，并在 cc48/64 启用 unified memory、移除显式 SWA ratio。
将本次 B200 unified memory changelog 条目关联到 PR #4366。
Fridge003
added
the
full-sweep-fail-fast
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
There was a problem hiding this comment.
I reviewed this PR and didn't find any bugs. The NVIDIA master config and the B200 srt-slurm recipes have designated code owners, so one of them should still take a look. One small thing: the new perf-changelog entry still has the placeholder pull/XXX as its pr-link.
What was reviewed:
- The image bump for
musespark-fp4-b200-sglang-agentic-dflashin nvidia-master.yaml. - The c48/c64 overrides in agentic.yaml.
swa-full-tokens-ratio: nulluses the same null-clears-inherited-arg pattern asmoe-runner-backend: nullin the dsv4 sglang recipes, and the c16–c32 points keep the base ratio of 0.2. - The changelog append. It sits only at the tail, with the existing bytes unchanged, and it is one English entry covering the affected key.
Extended reasoning...
The change bumps the SGLang nightly image for one B200 Muse Spark AgentX config key, adds enable-unified-memory and clears the inherited SWA ratio at c48 and c64 in its single-node recipe, and appends one perf-changelog entry. It touches no security-sensitive surface: it is benchmark configuration only. It is small and follows existing patterns. Approval was withheld only because .github/CODEOWNERS assigns both nvidia-master.yaml and the b200-* srt-slurm recipe paths to specific owners.
This review covers commit 80cd137, which is no longer the latest commit on this pull request; later commits are not covered by it.
将 B200 Muse Spark cc48/64 的 mem-fraction-static 调整为 0.85，为 CUDA graph 和 MegaMoE buffer 留出更多显存。
This branch has not been deployed
1 participant
Summary
Update the Muse Spark Open B200 AgentX curve to
lmsysorg/sglang:nightly-dev-20261021-c1c287a4. At concurrency 48 and 64, enable--enable-unified-memoryand remove the inherited--swa-full-tokens-ratioargument so SGLang can use its unified-memory path.Keep concurrency 64 from #4221, which is already present on main. Lower concurrency points retain
swa-full-tokens-ratio: 0.2; c48/c64 retain HiCache size 160 and MegaMoE token capacity 8192.Set
mem-fraction-static: 0.85only at c48/c64 to leave more GPU memory for graph capture; lower concurrency points retain0.9.The image changes every point, so one regular changelog entry refreshes the complete 10-point curve.
The previous sweep failed with a CUDA OOM during MegaMoE buffer allocation at prefill graph capture with
mem-fraction-static: 0.9. The0.85setting is a candidate mitigation; runtime validation has not established that it resolves the OOM.AI model disclosure
PR checklist and change type
Type of Change
Checklist
python -m infx.matrix.changed --base origin/mainlists affected config keys, I have appended one new entry covering them to the physical end ofinferencex-e2e/perf-changelog.yaml; if it lists none, I have added none. I have not edited historical entriesinferencex-e2e/perf-changelog.yamlentry, it carries exactly one primary sweep label (a maintainer applies it on fork PRs):full-sweep-fail-fast(recommended),full-sweep-enabled, ornon-canary-full-sweep-enabled. Optional modifiersall-evals,evals-only, andagentx-fastrequire a primary label; the last two block reuse while applied.OWNER/MEMBER/COLLABORATOR) has commented/use <run_id>(or the legacy/reuse-sweep-run) on this PR. Do this only once there is a final full sweep that is all green with evals passing, since after this comment the primary sweep label will no longer automatically kick off new sweeps. Remove and re-add the primary sweep label to force a new sweep.中文
更新 Muse Spark Open B200 AgentX 曲线，使用
lmsysorg/sglang:nightly-dev-20261021-c1c287a4。cc=48、64 启用--enable-unified-memory，同时移除继承的--swa-full-tokens-ratio参数。保留 #4221 中的 cc=64，该点已经存在于 main。cc<48 继续使用
swa-full-tokens-ratio: 0.2；cc=48、64 的 HiCache size 保持 160，MegaMoE token capacity 保持 8192。仅将 cc=48、64 的
mem-fraction-static设为0.85，为 graph capture 留出更多显存；更低并发仍使用0.9。镜像影响全部数据点，因此使用一个常规 changelog 条目重新运行完整的 10 点曲线。改动类型为配置变更。
此前 sweep 在
mem-fraction-static: 0.9下，于 prefill graph capture 阶段分配 MegaMoE buffer 时发生 CUDA OOM。0.85是待验证的缓解方案，尚未通过运行确认解决 OOM。AI 参与实现、本地验证和独立审查；审查 agent 继承主 agent 的模型。运行环境未暴露可核实的精确模型标识，无法确认具体版本。