[Misc] Clean up utils

[DarkLight1337]

Purpose

• Clean up some duplicate code caused by bad merge
• Remove unnecessary back-compatibility code
• Remove utils that are only used in V0

Part of #26900

Test Plan

Test Result

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.

[gemini-code-assist]

Code Review

This pull request cleans up the utility modules by refactoring code into more specific files, removing unused functions, and eliminating backward compatibility shims. The changes are well-structured and improve the organization of the codebase. I've found one critical issue where argument pre-processing can lead to an unhandled exception, and I've provided a recommendation to add a test case to expose this bug and guide its resolution.

[gemini-code-assist]

The FlexibleArgumentParser.parse_args method has a bug that causes an IndexError if vllm serve is called with --model as the last argument without a value. This should be handled gracefully by argparse (which would raise a SystemExit), but the current pre-processing logic crashes instead.

The crash occurs because the code at vllm/utils/argparse_utils.py:227 accesses args[model_idx + 1] without checking if the index is valid.

To ensure robust error handling and prevent this crash, I recommend adding a test case to expose this bug. This will help in fixing the underlying issue in FlexibleArgumentParser.parse_args and prevent future regressions.

Here is a suggested test case to add to this file:

def test_model_option_no_value(parser_with_config):
    with pytest.raises(SystemExit):
        parser_with_config.parse_args(["serve", "--model"])

[chatgpt-codex-connector]

💡 Codex Review

vllm/vllm/entrypoints/cli/serve.py

Lines 18 to 23 in 536298f

	from vllm.entrypoints.utils import VLLM_SUBCMD_PARSER_EPILOG
	from vllm.logger import init_logger
	from vllm.usage.usage_lib import UsageContext
	from vllm.utils import FlexibleArgumentParser
	from vllm.utils.network_utils import get_tcp_uri
	from vllm.utils.system_utils import decorate_logs, set_process_title

Update CLI FlexibleArgumentParser import

The utils refactor removed the re‑export of argparse helpers from vllm.utils, but the CLI entrypoint still imports FlexibleArgumentParser from the old module. Running vllm serve will now raise ImportError: cannot import name 'FlexibleArgumentParser' from 'vllm.utils' before any parsing happens. This needs to import from vllm.utils.argparse_utils like the other entrypoints.

---

vllm/vllm/distributed/kv_transfer/kv_connector/v1/decode_bench_connector.py

Lines 39 to 46 in 536298f

	from vllm.distributed.kv_transfer.kv_connector.v1 import (
	KVConnectorBase_V1,
	KVConnectorRole,
	)
	from vllm.distributed.kv_transfer.kv_connector.v1.base import KVConnectorMetadata
	from vllm.logger import init_logger
	from vllm.utils import cdiv

Fix cdiv import after utils cleanup

The commit removes the math helper aliases from vllm.utils, yet this file still imports cdiv from that module. Importing the connector will now fail with ImportError when running decode benchmarking or any KV‑transfer workflow that loads the module. The import should point to vllm.utils.math_utils.cdiv (and similar changes may be needed in lmcache_integration/vllm_v1_adapter.py).

---

vllm/vllm/lora/punica_wrapper/punica_gpu.py

Lines 14 to 17 in 536298f

	from vllm.lora.layers import LoRAMapping
	from vllm.triton_utils import HAS_TRITON, triton
	from vllm.utils import round_up

Update round_up import to new location

After the refactor, round_up is no longer re‑exported from vllm.utils, so this module (and other users such as layers/quantization/mxfp4.py) will fail to import with ImportError. Replace the import with from vllm.utils.math_utils import round_up or similar to match the new organization.

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[DarkLight1337]

/gemini review

[gemini-code-assist]

Code Review

This pull request provides a good cleanup of the vllm.utils module. The changes primarily involve refactoring by moving utilities into more specific submodules like argparse_utils and math_utils, and removing several unused or legacy (V0) functions and backward compatibility shims. The import paths across the codebase have been updated consistently to reflect these changes. The refactoring improves code organization and maintainability. I have reviewed the changes and found no issues.

[mergify]

Documentation preview: https://vllm--27552.org.readthedocs.build/en/27552/