vllm - 💡(How to fix) Fix [Bug]: [ROCm CI] HuggingFace dataset loading fails with "Feature type 'List' not found"

Your current environment

collect_env.py could not complete on the local machine due to a local CUDA driver/runtime mismatch during torch.cuda.init().

Relevant local test environment:

• Ubuntu Linux
• Python 3.12.3
• HuggingFace datasets successfully loads philschmid/mt-bench
• List exists in datasets.features._FEATURE_TYPES

🐛 Describe the bug

Summary

Multiple ROCm CI suites are failing during HuggingFace dataset loading with:

ValueError: Feature type 'List' not found

Affected suites observed:

• mi250_1: Multi-Modal Accuracy Eval (Small Models)
• mi355_1: Examples

The failure occurs while loading:

load_dataset("philschmid/mt-bench")

inside:

vllm/benchmarks/datasets/datasets.py

Investigation

The failure appears environment-specific rather than dataset-specific.

I verified in a clean Linux environment that:

from datasets import load_dataset
ds = load_dataset("philschmid/mt-bench")

loads successfully using a modern HuggingFace datasets installation.

Additionally:

from datasets.features import features
print("List" in features._FEATURE_TYPES)

returns True.

However, the ROCm CI environment fails with:

ValueError: Feature type 'List' not found

during dataset loading.

Likely cause

This suggests the issue is likely related to the ROCm CI environment itself, potentially:

• outdated/inconsistent datasets package
• stale HuggingFace dataset cache
• dependency mismatch
• CI image drift between ROCm and upstream environments

The issue does not appear to be related to ROCm kernel execution or model runtime initialization.

Additional context

The engine/model initialization succeeds successfully before the failure:

• graph capture completes
• engine warmup succeeds
• failure occurs specifically during load_dataset(...)

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.