Pin the ModelExpress version tuple for Dynamo 1.2.1
Nobody has claimed this yet.
- Dominant language
- Go
- Stars
- 218
- Forks
- 72
- Avg merge
- 1d 9h
- Merged PRs (30d)
- 417
Description
Description
Establish which ModelExpress versions the NVCF-supported Dynamo runtime actually
ships, so every downstream sample, guide, and manifest entry pins the same tuple.
The Dynamo 1.2.1 build recipe references ModelExpress client 0.4.0 while the
current public chart and server are 0.5.1. The delivered version has to be read
from the worker image rather than the recipe, because the recipe is not what
ships.
Definition of Done
- The ModelExpress client, vLLM, and NIXL versions inside
nvcr.io/nvidia/ai-dynamo/vllm-runtime:1.2.1are recorded from the image. - A server version is chosen, with the reason stated.
- Whether a derivative worker image is needed is settled.
- No downstream task pins a version before this closes.
Resources
The image contains ai-dynamo 1.2.1, modelexpress 0.4.0, nixl 0.10.1, and vllm
0.20.1. modelexpress-server:0.4.0 exists, so client and server are matched by
pinning the server back to 0.4.0 rather than taking the current 0.5.1 chart
default. No derivative image is required.
The tuple is identical on linux/amd64 (digest
sha256:b91c824abd16ab0de57e67ee57602a660fc8635728250ae5a9b7f9bb2a36223e) and
linux/arm64. The amd64 read is authoritative, since the GPU nodes are x86_64.
One consequence for later tasks: the CRDs must come from the tag matching the
server image, not the chart. The 0.5.1 schemas add sourceType and
artifactSource, which a 0.4.0 server never writes.
A second image variant, vllm-runtime:1.2.1-efa-amd64, carries the same
component versions plus the AWS libfabric stack. It is the image the sample uses,
for reasons recorded in the NIXL backend task.
Part of #1113.
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Inspect nvcr.io/nvidia/ai-dynamo/vllm-runtime:1.2.1, using the amd64 digest as authoritative, and verify the recorded ModelExpress, vLLM, and NIXL versions. Review the matching ModelExpress server image and CRD tag, then document the server rationale and whether a derivative image is needed. Done means the tuple and decisions are recorded before downstream tasks pin versions.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- docker, helm, kubernetes
- Domain
- devops, infrastructure
- Issue type
- Refactor
- Difficulty
- 3/5
- Estimated time
- 1-2 days
- Activity status
- Active
- Clarity
- Clearly specified
- Newbie friendliness
- 68/100