Repository navigation
Conversation
Backfill the installed-package boundary fixed by vLLM 0.33 and Dynamo ai-dynamo#17862 (ai-dynamo#18271 on release/1.7.0). Decode native add_lora messages with the installed Python schema before stubbing weight loading, and check the load/list/unload lifecycle without a GPU. Signed-off-by: Molly Sun <diweis@nvidia.com>
DiweiSun
temporarily deployed
to
external_collaborator
November 12, 2026 08:08 — with
GitHub Actions
Inactive
DiweiSun
temporarily deployed
to
external_collaborator
November 12, 2026 08:08 — with
GitHub Actions
Inactive
Contributor
Unsigned Commits DetectedThe following commits are not GPG-signed and must be signed before CI can run:
Please sign your commits and push again. See the GitHub docs on commit signature verification for help. |
Contributor
|
👋 Hi DiweiSun! Thank you for contributing to ai-dynamo/dynamo. Just a reminder: The 🚀 |
This branch was previously deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Overview
Backfill regression coverage for native runtime LoRA loading fixed by #17862 and its 1.7.0 cherry-pick #18271. The purpose is to catch an incompatible Rust/Python protocol pair in the vLLM image we ship when updating backend dependencies, before running GPU deployment tests.
The underlying fix is vllm-project/vllm#63295. Python had removed
tensorizer_config_dictfrom the positional MessagePackLoRARequest, while vllm-rs 0.32.0 still serialized that slot. The extranulloccupied Python'sload_inplace: boolposition, so runtime adapter loading failed withExpected bool, got Noneat$[4]. vLLM 0.33.0 restores the matching layout.Key Changes
Production logic added: 0 lines. One new test file; no workflow or dependency changes.
add_loraargument bytes using the installed PythonLoRARequestdefinition, rather than a copied schema or a fabricated request.test_native_lora_matches_installed_python_schemachecks successful load, adapter visibility, and unload, along with the decoded name, ID, path, and boolean fields. A negative control verifies that the historical extra-null layout is rejected by the same decoder. Assertions use response data and decoded fields; logs are diagnostic only.ManagedProcessand the shared port allocator.pre_merge,vllm,integration, andgpu_0markers select the case in the existing vLLM CPU test job (pre_merge and vllm and gpu_0 and not e2e), with no new job or Rust test binary to distribute.Existing Python-worker LoRA tests exercise the Python path; Dynamo's fake-gRPC LoRA lifecycle tests return before the native Rust-to-Python serialization boundary. This case covers that missing boundary using the packaged producer and consumer together.
Validation
Ran the same pytest file, using the repository's actual conftests, default strict configuration, and the existing CPU-job marker expression:
add_lorapayload at$[4].Both runs used amd64 CPU containers with networking disabled, no GPU, no model downloads, and no model weights. The verification images add pytest plugins already declared in
container/deps/requirements.test.txt; the backend and protocol packages remain at their respective RC versions. The passing case takes approximately 1.3 seconds (approximately 5.5 seconds including pytest startup and collection).python3 -m pytest tests/serve/test_vllm_lora_protocol.py \ -m 'pre_merge and vllm and gpu_0 and not e2e' -vvThe peer replaces weight loading only after validating the real wire request. This is a protocol-compatibility test; GPU adapter computation remains covered by serving E2E tests. Upstream CI has not yet run for this branch.
Issues and Ongoing Work
Pull Request Checklist
/pr-checklistskill.