[Thangnguyenvn98]

Overview

Enrolls every remaining step whose pytest commands set PYTHONPATH, with its AMD mirror, in automatic sharding. Draft until ci-infra#825 merges: setting PYTHONPATH replaces the path the sharding plugin is installed on, so today every shard of these steps fails with No module named 'runtime_shard_plugin' (seen in #62513's build 97706, which dropped Spec Decode for this reason). ci-infra#825 installs the plugin into site-packages, where a command's own PYTHONPATH can't hide it.

Claims

Step	Was	Planned shards
lm-eval-spec-decode-4xb200	one job	3
entrypoints-integration-api-server	4 fixed shards	5
amd-entrypoints-integration-api-server	4 fixed shards	4
amd-lm-eval-spec-decode-4xb200	one job	2
rust-frontend-serve-admin-coverage	one job	2
amd-rust-frontend-serve-admin-coverage	one job	1

• Every test runs exactly once across the shards, with the same environment and filters as before.
• The API Server step's shard-1-only dev/rpc command and the --num-shards/--shard-id flags give way to the planner, so new test files no longer need placing by hand.

Validation

• Planned each step and mirror on this branch with ci-infra main's planner and main's timings; the plans match the table.
• Rendered each with the pipeline generator and ran every shard's commands against a mock test tree, as a PR build and in main's OTel mode: every test ran in exactly one shard (4/4 NVIDIA cases, 2/2 AMD mirror cases). That harness imports the plugin from an editable install, so it cannot show the PYTHONPATH problem; the /ci run after ci-infra#825 merges is the check for it.
• Not run yet in CI: /ci run, plus a build scoped to the optional Spec Decode step and the AMD mirrors /ci run leaves out, once ci-infra#825 is merged.

Details

• Not duplicating an open PR. [CI] Shard Humming B200/H100 evals and two entrypoints AMD mirrors automatically #62513 dropped Spec Decode in favour of this one; no other open PR touches these keys.
• Spec Decode has three configs, and one runs about 25 minutes on its own, so its slowest shard stays over the 20-minute target; the step's wait still roughly halves.
• API Server goes from 4 fixed shards to 5 planned ones on the NVIDIA side. entrypoints/serve/dev/rpc now runs in whichever shard the plan gives it rather than always in shard 1. One file, scale_out/ec_integration/test_scale_out_ec_e2e.py, has no timing on main and is counted at the planner's default.
• Timeouts unchanged (human). timeout_in_minutes stays per job; fallback-aware timeouts belong in a central timeout setting.
• One PR for the blocker. Every enrollment that waits on ci-infra#825 is here, so they can all be validated with one /ci run once it merges.

🤖 Generated with Claude Code

---

Pull Request Checklist

• I used vLLM's /pr-checklist skill. (Mandatory for agents, optional for humans).

• AI assistance was used during the creation of this PR.

• Design Fit: Minimizes impact on core components, reuses existing functionality, and justifies added complexity.

• Testing and Validation: Validates the change and ensures any added tests are meaningful and reliable, with CI coverage or documented CI resource constraints and validation performed outside CI.

• Code Quality and Style: Keeps code and comments clear and concise, and updates relevant documentation and examples.

• Pull Request Contents: Includes a brief summary and relevant links, supports claims with evidence, explains root causes and implementation trade-offs, and follows the contributing guide.