fix(test): cap RL worker GPU memory so test_rl_worker_discovery survives parallel bin-packing

[nv-tusharma]

Problem

tests/rl/test_worker_discovery.py::test_rl_worker_discovery_and_engine_admin_routes has been failing 77.9% of the time on main since it was introduced in #9681, blocking post-merge/PR pipelines. Datadog shows 26 failed pipelines and ~265h of CI time lost. It fails with:

RuntimeError: Main server process exited with code 1 while waiting for health check

Root cause

The test launches its vLLM worker through a hardcoded RLVllmWorkerProcess command that omits the GPU-memory budget every other vLLM e2e test injects. The parallel GPU scheduler bin-packs this worker onto a shared GPU and sets _PROFILE_OVERRIDE_VLLM_KV_CACHE_BYTES from the test's requested_vllm_kv_cache_bytes(331_801_000) marker — but the worker never consumed it, so vLLM fell back to its default gpu_memory_utilization=0.9, demanded ~90% of the whole card, and crashed during engine init:

ValueError: Free memory on device cuda:0 (3.51/47.37 GiB) on startup is less than
desired GPU memory utilization (0.92, 43.58 GiB). ...
RuntimeError: Engine core initialization failed.

This crashes the worker (exit 1) only when it's co-located with other workers on the GPU, which is why it passed ~22% of the time (scheduled first/alone). tests/README.md documents honoring _PROFILE_OVERRIDE_VLLM_KV_CACHE_BYTES as a hard requirement for vLLM launch paths.

Fix

Route the worker command through build_gpu_mem_args("build_vllm_gpu_mem_args") — the same helper the standard vLLM workers (tests/mm_router/*, tests/router/test_router_e2e_with_vllm.py, etc.) use. Under parallel runs it emits --kv-cache-memory-bytes 331801000 --gpu-memory-utilization 0.01 (caps the worker to its profiled budget); serially it falls back to --gpu-memory-utilization 0.4.

Verification (local A/B, RTX 6000 Ada, simulated bin-packing: 43 GiB held, ~3.5 GiB free)

• Unfixed: dies at the vLLM free-memory check — exit_code=1 after 27s, same 503 → process-death pattern as CI.
• Fixed: passes the memory gate — reserved 0.31 GiB memory for KV Cache ... skipped memory profiling, GPU KV cache size: 2,880 tokens, init engine ... took 0.32 s — engine initializes within its budget and proceeds well past the crash point.

One-file, test-only change that brings this worker in line with the harness's documented GPU-budget contract.

🤖 Generated with Claude Code

Summary by CodeRabbit

• Tests
  • Updated RL worker discovery test to use shared GPU memory configuration helper, improving consistency in GPU memory settings across workers.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 2a336bfc-be9d-43c2-9203-fd5f289d65f3

📥 Commits

Reviewing files that changed from the base of the PR and between 78dab60 and 2f4d2ca.

📒 Files selected for processing (1)

• tests/rl/test_worker_discovery.py

---

Walkthrough

The PR refactors vLLM GPU memory configuration in the RL worker discovery test by extracting a reusable helper function. It imports a shared build_gpu_mem_args utility, creates _vllm_gpu_mem_args() to conditionally derive GPU memory CLI arguments, and applies the helper to the worker process launch command.

Changes

vLLM GPU Memory Configuration

Layer / File(s)	Summary
GPU memory settings helper and worker integration tests/rl/test_worker_discovery.py	Imports build_gpu_mem_args, introduces _vllm_gpu_mem_args(default_utilization) to build GPU memory CLI arguments from a shared budget utility or fall back to a utilization percentage, and updates the vLLM worker process launch command to expand those arguments into the subprocess call.

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~8 minutes

🚥 Pre-merge checks | ✅ 3 | ❌ 2

❌ Failed checks (1 warning, 1 inconclusive)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 0.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.
Description check	❓ Inconclusive	The description provides comprehensive Problem, Root cause, Fix, and Verification sections. However, the Related Issues section is missing—the description does not indicate whether this PR is linked to or closes an issue.	Add a Related Issues section specifying whether this PR closes #9681 or another issue, or explicitly confirm no related issue exists.

✅ Passed checks (3 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly and concisely summarizes the main change: capping RL worker GPU memory to fix a failing test in parallel bin-packing scenarios.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

---

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[biswapanda]

lgtm - thanks @nv-tusharma !