[Doc] Fix %% rendering in CLI reference for --safetensors-load-strategy

[lizzzcai]

Summary

While exploring the --safetensors-load-strategy feature, I noticed the CLI reference docs showed 90%%%% instead of 90%.

Root cause

arg_utils.py escapes % → %% in docstring-derived help text for argparse. MarkdownFormatter.add_arguments in generate_argparse.py reads action.help directly, bypassing _expand_help which would normally unescape %% → %. Additionally, the load.py docstring had 90%% instead of 90%, causing a double-escape to 90%%%% in the rendered output.

Changes

• vllm/config/load.py: 90%% → 90% in docstring
• docs/mkdocs/hooks/generate_argparse.py: unescape %% → % in MarkdownFormatter.add_arguments before writing markdown

As a side effect, this also fixes --bin-by in the benchmark CLI reference which was showing request_throughput%%1 instead of the correct request_throughput%1.

Verification

Served docs locally with API_AUTONAV_EXCLUDE=vllm mkdocs serve and confirmed:

None (default): Uses memory-mapped (lazy) loading. When an NFS filesystem is
detected and the total checkpoint size fits within 90% of available RAM,
prefetching is enabled automatically.

Notes

• No existing open PR addresses this (searched for safetensors, generate_argparse, %%)
• AI assistance (Claude) was used to investigate and implement this fix

[mergify]

Documentation preview: https://vllm--46335.org.readthedocs.build/en/46335/

[yewentao256]

Thanks for the work!

[yewentao256]

Why this is needed? Shall we fix the source instead of editing text directly here?

[lizzzcai]

Hi @yewentao256 Thanks for the review.

The root cause is that arg_utils.py:319 unconditionally escapes % → %% for all docstring-derived help text:

help = help.replace("%", "%%")

This is needed for terminal --help because argparse's _expand_help does help % params, which collapses %% → % as a side effect:

"90%% of RAM" % {"default": None}  # → "90% of RAM" 

But MarkdownFormatter.add_arguments bypasses _expand_help and reads action.help directly, so %% is never unescaped. This affects multiple args today — --safetensors-load-strategy, --gpu-memory-utilization, and --cp-kv-cache-interleave-size.

Two options:

Option 1 — Fix in generate_argparse.py (current approach): unescape %% → % in MarkdownFormatter. Systemic — covers all current and future docstrings with %.

Option 2 — Fix in arg_utils.py: skip escaping when running under mkdocs, since _expand_help is never called in that path:

IS_MKDOCS = (
    sys.argv[0].endswith("mkdocs")
    or sys.argv[0].endswith("mkdocs/__main__.py")
)

if not IS_MKDOCS:
    help = help.replace("%", "%%")

Happy to switch to Option 2 if preferred.