ci: exclude vllm_inference and megatron from nightly recipe CI

[svc-bionemo]

Problem

The 26.03 container update brought transformers==5.5.4, which breaks the vllm_inference recipe CI:

vLLM 0.15.1 (current) + transformers 5.x

• vLLM 0.15.1 hard-requires transformers < 5
• install_vllm.sh installs vLLM (which pins transformers <5), then upgrades transformers back to 5.x
• This creates a broken environment — vLLM is installed against transformers 4.x but runs with 5.x
• Additionally, the nvidia/esm2_* Hub models have "tokenizer_class": "TokenizersBackend" in their tokenizer_config.json, a class that only exists in transformers 5.x. If vLLM downgrades to <5, tokenizer loading fails

vLLM 0.19.1 (latest, supports transformers 5.x)

• Requires transformers >= 5.5.1 ✅
• But uses register_opaque_type(hoist=True) in vllm/utils/torch_utils.py, gated behind is_torch_equal_or_newer("2.11.0.dev")
• The NGC 26.03 torch build (2.11.0a0+nv26.03) matches that version check, but does not have the upstream hoist parameter on register_opaque_type yet
• Result: TypeError: register_opaque_type() got an unexpected keyword argument 'hoist' at import time

Megatron duplicates

• Megatron recipes (eden_megatron, evo2_megatron) already have a dedicated CI workflow (unit-tests-mbridge-recipes.yaml) but were also running as duplicates in this recipes workflow on nightly

Fix

Exclude both vllm_inference and megatron from the ALL_DIRS nightly enumeration in unit-tests-recipes.yml:

• vllm_inference: blocked until either NGC ships torch with the upstream hoist API (enabling vLLM 0.19.1) or vLLM releases a version compatible with both the NGC torch build and transformers 5.x
• megatron: removes duplicate nightly runs (already covered by unit-tests-mbridge-recipes.yaml)

Both were already excluded from PR changed-files detection but scheduled runs bypass that filter.

Related

• Fix HFInferenceParams missing is_compileable for transformers 5.x compatibility #1553 — Fixed HFInferenceParams.is_compileable for transformers 5.x (merged)
• Fix ESM2 tokenizer export: use PreTrainedTokenizerFast instead of TokenizersBackend #1555 — Fixes ESM2 tokenizer export to write PreTrainedTokenizerFast instead of TokenizersBackend

Failing CI run

https://github.com/NVIDIA/bionemo-framework/actions/runs/24601629903

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro Plus

Run ID: 7b77b361-f013-44e1-8d9c-92952ec4427e

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

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