Skip to content

Dynamo v1.3.0-glm-5.2-dev.1

Pre-release
Pre-release

Choose a tag to compare

View all tags
@dagil-nvidia dagil-nvidia released this 21 Jul 03:49
v1.3.0-glm-5.2-dev.1
8e08e34
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Dynamo v1.3.0-glm-5.2-dev.1 — Release Notes

Dynamo v1.3.0-glm-5.2-dev.1 is an experimental snapshot build with early GLM-5.2 support on the Dynamo SGLang backend. It is not recommended for production and is not a QA-gated release. APIs, behavior, and defaults may change before stable support. Use it for evaluation, testing, and early feedback only.

Summary

Dynamo v1.3.0-glm-5.2-dev.1 adds experimental SGLang runtime changes intended to enable GLM-5.2 serving. The image applies six focused SGLang patches for MoE routing-bias dtype handling, speculative-token simulation, NIXL decode-abort handling, KV-cache capacity validation, deprecated option aliases, and Transformers 5.12.1 compatibility (#11859). This tag publishes only the multi-architecture SGLang container.

The release branch also includes GLM-5.2 SGLang aggregated and disaggregated Kubernetes recipes for B200 and H200, model-cache setup, and an AIPerf trace-replay benchmark workflow (#11937, #11940).

Base Branch: release/1.3.0
Base Commit: d2aed312b1f (2026-07-17)
Preview Branch: release/1.3.0-glm-5.2-dev.1
Preview Tip: 8e08e343cc2b20422938ceec420785f23ee64dd4
SGLang: lmsysorg/sglang:v0.5.14-cu130-runtime
NIXL: v1.0.1 (UCX v1.20.x)

Container Images

Backend Arch Image
SGLang (CUDA 13) multi-arch (amd64 + arm64) nvcr.io/nvidia/ai-dynamo/sglang-runtime:1.3.0-glm-5.2-dev.1

Backend Versions

Backend Source CUDA Python Notes
SGLang Dynamo-built runtime on lmsysorg/sglang:v0.5.14-cu130-runtime, with six vendored patches applied during the image build 13.0 3.12 Pins transformers==5.12.1 and mistral_common==1.11.5; NIXL v1.0.1

About GLM-5.2

GLM-5.2 is Z.ai's long-context mixture-of-experts model for reasoning and agentic workloads. This preview focuses on the SGLang runtime integration needed to evaluate the model with Dynamo.

Full Changelog

SGLang Backend

  • Added GLM-5.2 runtime compatibility: Vendored six focused SGLang patches, installed matching transformers and mistral_common versions, handled upstream images that already contain the dynamo user, and kept DCP/KV-capacity checks compatible across the supported runtime base (#11859).

Recipes

  • Added GLM-5.2 deployment recipes: Added B200 and H200 aggregated and disaggregated SGLang profiles, model-cache manifests, and AIPerf benchmarking under recipes/glm-5.2 (#11937, #11940).

Known Issues / Limitations

  • This is a branch-specific snapshot build, not a QA-gated stable release.
  • The SGLang patches are temporary compatibility changes and may be removed as fixes land in upstream runtime images.
  • B200 recipes support context lengths up to 500K; H200 recipes support up to 250K. n > 1 requests are not supported by the disaggregated recipes.
Assets 2
Loading