Fix #367: reportgen crashes with AttributeError on Namespace.file

[idevasena]

Fixes #367.

mlpstorage reports reportgen (and any other non-benchmark subcommand) crashed at startup with AttributeError: 'Namespace' object has no attribute 'file'. The CLI parser's data-access-protocol consolidation block read parsed_args.file and parsed_args.object unconditionally, but those attributes only exist on subcommands that call add_storage_type_arguments() — namely training, checkpointing, vectordb, and kvcache. The reports, history, and lockfile subcommands never define those flags, so their Namespace doesn't carry them.

This change guards the consolidation on attribute presence. Behavior on benchmark subcommands is preserved bit-for-bit; non-benchmark subcommands now parse cleanly and simply don't get a data_access_protocol attribute (downstream code already reads it via getattr(self.args, 'data_access_protocol', None) in mlpstorage_py/benchmarks/dlio.py:136, so nothing else has to change).

Regression context

The bug entered with commit 39e657d (PR #359, "Bug fixes and performance enhancements: object storage, checkpointing, Parquet loading"), which introduced the --file / --object storage-type flags and the consolidation block but didn't account for subparsers that don't opt into those flags.

Root cause

In mlpstorage_py/cli_parser.py, inside parse_arguments() after parser.parse_args():

Consolidate the data access protocol into a single field

if parsed_args.file:
    parsed_args.data_access_protocol = "file"
else:
    parsed_args.data_access_protocol = parsed_args.object
del parsed_args.file
del parsed_args.object

A grep across the package confirms add_storage_type_arguments() (in mlpstorage_py/cli/common_args.py) is only invoked from training_args.py, checkpointing_args.py, vectordb_args.py, and kvcache_args.py. The reports, history, and lockfile subparsers never call it, so their parsed Namespace has no file or object attribute and the block above raises before any subcommand logic runs. The docstring of add_storage_type_arguments() already states the intended contract — benchmarks that don't use object storage "simply ignore the flags; those that do can check args.file / args.object" — but the parser's consolidation step didn't honor it.

Fix

Gate the consolidation block on attribute presence, and use getattr for the reads so a benchmark subcommand that defines only one of the two flags still works:

Consolidate the data access protocol into a single field.

# The --file / --object flags are only defined on benchmark subcommands
# that call add_storage_type_arguments() (training, checkpointing,
# vectordb, kvcache). Other subcommands (reports, history, lockfile)
# do not define them, so guard the consolidation on attribute presence.
if hasattr(parsed_args, "file") or hasattr(parsed_args, "object"):
    if getattr(parsed_args, "file", False):
        parsed_args.data_access_protocol = "file"
    else:
        parsed_args.data_access_protocol = getattr(parsed_args, "object", None)
    # Clean up the raw flags so downstream code uses data_access_protocol.
    for _attr in ("file", "object"):
        if hasattr(parsed_args, _attr):
            delattr(parsed_args, _attr)

Files changed

• mlpstorage_py/cli_parser.py — guarded consolidation block (7 lines removed, 14 added)
• tests/unit/test_cli.py — added TestParseArgumentsStorageFlagConsolidation with 5 tests (3 regression tests for non-benchmark subcommands, 2 lock-in tests for benchmark subcommands)

[github-actions]

MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅