Repository navigation
[NVIDIA][AgentX] Add GLM-5.2 NVFP4 B300 Dynamo+SGLang aggregated TP4 AgentX key / 新增 GLM-5.2 NVFP4 B300 Dynamo+SGLang 聚合 TP4 AgentX 配置 - #3953
Conversation
…AgentX key / 新增 GLM-5.2 NVFP4 B300 Dynamo+SGLang 聚合 TP4 AgentX 配置 Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
hshrivastava-droid
added
the
full-sweep-enabled
|
Thanks for the contribution!
中文感谢你的贡献！
|
![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. Because it adds a new multi-node SLURM benchmark recipe plus matching master-config and changelog entries across files owned by the NVIDIA srt-slurm-recipes CODEOWNERS, a human look (from one of the listed owners) would still be worthwhile.
What was reviewed:
- The new agg-tp4-variants.yaml recipe (base + 7 concurrency overrides) against the matching nvidia-master.yaml search-space entries for config-key/override wiring consistency.
- Container image consistency between the recipe's
identity.container.imageand the master config'simage:field — both pinlmsysorg/sglang:nightly-dev-cu13-20260901-07c8f729. SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE: '0'against CONTRIBUTING.md's draft-precision rule and sibling GLM-5.2 FP4 recipes — all set it to'0', consistent, not an enable.- health_check timing (1440 attempts × 10s = 4h) against
slurm.time_limit: 4:00:00and sibling GLM-5.2 recipes — matches the established pattern exactly. - perf-changelog.yaml entry — single new append-only block, correct config-keys, matches the one-PR-one-block convention.
Extended reasoning...
The diff is a purely additive three-file change: a new multi-node srt-slurm recipe (GLM-5.2 NVFP4 AgentX on B300, Dynamo+SGLang aggregated TP4, 7 concurrency variants), a matching nvidia-master.yaml search-space entry, and one perf-changelog.yaml append. It touches no auth, crypto, or data-exposure surface — it's infrastructure/benchmark configuration. The deciding factor for deferring rather than approving is that both changed paths (inferencex-e2e/configs/nvidia-master.yaml and the b300-* srt-slurm-recipes tree) are covered by non-core CODEOWNERS entries in .github/CODEOWNERS, and the visible PR timeline shows no sign-off from those owners yet.
|
View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=38356572105 View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=38356572105 |
|
On my radar. I will update the cookbook tomorrow and run the sign-off process |
|
[by Claude Code] The SGLang cookbook integration of this PR's AgentX config ( This PR has no authorized |
…-sglang-agentx-tp4 Resolve the conflicts in configs/nvidia-master.yaml and perf-changelog.yaml. Main now selects srt-slurm recipes with srt-recipe-dir and srt-recipe and composes AgentX recipes from fragments (#3805, #3818), so this entry and its bundle move to that format: - The master entry sets srt-recipe-dir, and each row selects its override with srt-recipe instead of a CONFIG_FILE setting. - The bundle drops the keys the binder and launcher now write: the model block, the identity repo and image, the client command, benchmark.concurrencies, RESULT_DIR, IS_MULTINODE, ENABLE_AGENTX_POWER, AIPERF_DATASET_MMAP_CACHE_DIR and HF_HUB_CACHE. It also drops the unread INFMAX_CONTAINER_WORKSPACE and PORT. No serving setting changes: engine args and env, frontend, Dynamo, services and Slurm options bind exactly as the PR's sweep submitted them. The perf-changelog entry stays appended at the end of the file. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
nvpohanh
removed
the
full-sweep-enabled
This branch has not been deployed
2 participants
Summary
Adds config
glm5.2-fp4-b300-dynamo-sglang-agentic-agg-tp4(GLM-5.2 NVFP4, B300, Dynamo + SGLang, aggregated TP4, AgentX).Image:
lmsysorg/sglang:nightly-dev-cu13-20260901-07c8f729Type of change
Checklist
nvidia-master.yamlperf-changelog.yamlentry added中文
新增配置
glm5.2-fp4-b300-dynamo-sglang-agentic-agg-tp4（GLM-5.2 NVFP4，B300，Dynamo + SGLang，聚合 TP4，AgentX）。AI model disclosure
AI-assisted. The exact model identifier could not be verified from the runtime.
🤖 Generated with Claude Code