Dynamo v1.3.0-nemotron-super-dev.1
Pre-release
Release Notes
Dynamo v1.3.0-nemotron-super-dev.1 is an experimental dev build giving an early look at NVIDIA Nemotron-3 Super support on the Dynamo vLLM backend. It is not recommended for production — features may be incomplete and APIs, behaviors, and defaults may change before the stable release. Use it for evaluation, testing, and early feedback only.
Summary
Dynamo v1.3.0-nemotron-super-dev.1 adds NVIDIA Nemotron-3 Super (120B-A12B) serving on the Dynamo vLLM backend. It builds on a vLLM v0.21.0 runtime base and backports vLLM #40984 — semantic KV-cache event metadata — into the image (#10230), so the Dynamo KV-aware router can index per-prefix worker affinity on Nemotron-3 Super's hybrid Mamba/Attention architecture, which the pinned v0.21.0 base does not otherwise expose. Aggregated deployment recipes for B200 (NVFP4) and H200 (FP8) ship alongside (#10223). Only the vLLM container is published.
Branch: release/1.3.0-nemotron-super-dev.1, built from commit e0edf5a
vLLM: v0.21.0 runtime base + backported KV-events PR #40984
Container Images
| Backend | Arch | Image |
|---|---|---|
| vLLM (CUDA 13) | multi-arch (amd64 + arm64) |
nvcr.io/nvidia/ai-dynamo/vllm-runtime:1.3.0-nemotron-super-dev.1 |
Backend Versions
| Backend | Source | CUDA | Python | Notes |
|---|---|---|---|---|
| vLLM | Dynamo-built runtime on a vLLM v0.21.0 base, with vLLM #40984 backported (vendored patch under container/deps/vllm/patches/v0.21.0/, applied in templates/vllm_runtime.Dockerfile) |
13.0 | 3.12 | KV-cache events carry semantic cache-group metadata for the KV-aware router on hybrid Mamba/Attention models |
Models
| Model | Precision | GPU | Notes |
|---|---|---|---|
| NVIDIA-Nemotron-3-Super-120B-A12B | NVFP4 | B200 | aggregated serving; chat + agentic profiles |
| NVIDIA-Nemotron-3-Super-120B-A12B-FP8 | FP8 | H200 | aggregated serving; chat + agentic profiles |
Full Changelog
vLLM Backend (Nemotron-3 Super)
- KV-Events Metadata Backport: Backported vLLM PR #40984 ("emit KV cache metadata") onto the v0.21.0 runtime image so KV-cache events carry semantic cache-group metadata, enabling the Dynamo KV-aware router to index per-prefix worker affinity on the hybrid Mamba/Attention architecture (#10230). Vendored as a patch under the per-ref convention and applied via a build step in the runtime Dockerfile; intended to drop once it lands upstream in the base image.
Recipes
- Nemotron-3-Super Deployment Recipes: Added aggregated vLLM serving recipes for Nemotron-3-Super-120B-A12B across B200 (NVFP4) and H200 (FP8), each with chat and agentic profiles, plus model-cache PVC/download manifests and an AIPerf trace-replay performance job (recipe folder, README, #10223).