[faresobeid]

Summary

• Bump the router version from 0.2.2 to 0.2.3 in Cargo.toml, Cargo.lock, pyproject.toml, and py_src/vllm_router/version.py.
• Release fix(pd): honor routed_experts dtype in inline merge #61: P/D routed-experts merge requires one dtype on prefill and decode (uint8 or uint16, so models with more than 256 experts work), cuts decode rows by their absolute start position instead of relying on zero-padded rows, and uses checked arithmetic for row ranges and shapes. Zero-padded decode rows from current prime-rl are still accepted.

What merging this does

The release.yml workflow will:

1. tag the merge commit as v0.2.3 and create a GitHub Release with generated notes
2. build and attach manylinux_2_28 wheels for amd64 and arm64
3. publish ghcr.io/primeintellect-ai/router:0.2.3 and update :latest

Validation

• cargo test on this branch (735 passed), cargo fmt --check clean.
• GPU end to end (prime-rl #3830 + #3900, Qwen3-30B-A3B, 1P1D NIXL vs colocated, router replay + sampling masks): 10/10 identical at concurrency 1 and 20/20 at concurrency 10; forced uint16 routing 10/10 and 20/20; decode preemption 20/20. Router v0.2.2 on the same setup fails 8/10 requests and returns wrong rows for the rest. Details in fix(pd): honor routed_experts dtype in inline merge #61.

Follow-up

After both wheel assets are attached to the release, update prime-rl's amd64 and arm64 wheel pins from v0.2.2 to v0.2.3 and regenerate uv.lock (in prime-rl #3830, which needs v0.2.3).

Note

Release v0.2.3: bump version from 0.2.2

Bumps the package version to 0.2.3 in pyproject.toml, version.py, and Cargo.toml. Also updates a dependency entry in Cargo.lock. vllm_router.__version__ now evaluates to 0.2.3.

^{Macroscope summarized 331d91c.}

---

Note

Low Risk
Version-only bump with no runtime logic changes; risk is limited to packaging and release automation.

Overview
Release v0.2.3 aligns the published version to 0.2.3 across the Rust crate (Cargo.toml / Cargo.lock for vllm_router_rs), the Python distribution (pyproject.toml), and vllm_router.__version__ in py_src/vllm_router/version.py.

There are no functional code changes in this diff; it is the version pin that triggers the release workflow (tag, wheels, container image) for the fixes described in release #61.

^{Reviewed by Cursor Bugbot for commit 331d91c. Bugbot is set up for automated code reviews on this repo. Configure here.}

[macroscopeapp]

Approvability

Verdict: Approved at 331d91c

Macroscope's review found this PR approvable — This is an isolated 0.2.2-to-0.2.3 metadata bump across the Rust and Python packages, with no routing or request-path logic changes. Merging invokes the existing release automation for packaging and image publication, but the PR does not modify deployment behavior or infrastructure.

^{You can add or adjust custom eligibility rules. Learn more.}