Summary

• Fixes a silent argparse bug in main.py: --do-sample False parses as True, --enable-bucketing False parses as True.
• Switches both flags to argparse.BooleanOptionalAction so the standard --flag / --no-flag idiom works.
• Affected lines: main.py:113 (--do-sample) and main.py:157 (--enable-bucketing).

The bug

argparse(type=bool) is not a string→bool converter. The type= callable is invoked on the raw CLI argument string, and bool(non_empty_string) is always True in Python. So every textual "off" value the user might reasonably pass is silently coerced to True.

Empirically (with the unpatched main.py):

--do-sample ''       -> args.do_sample = False
--do-sample 'False'  -> args.do_sample = True   <-- silently wrong!
--do-sample 'false'  -> args.do_sample = True
--do-sample 'no'     -> args.do_sample = True
--do-sample '0'      -> args.do_sample = True
--do-sample 'True'   -> args.do_sample = True
--do-sample 'yes'    -> args.do_sample = True
no flag              -> args.do_sample = True   (default)

The only reachable "off" path is --do-sample '', which is awkward to type from most shells and obviously not what users intend. --no-do-sample is rejected outright because the action doesn't declare it. The same is true for --enable-bucketing.

Why it matters for the contest

This isn't a stylistic concern — it materially affects accuracy validation in evaluate_all:

1. args.do_sample=True (the default, and what every reasonable string also resolves to) is propagated into OnDeviceSamplingConfig(do_sample=True, ...) at main.py:380-383.
2. OnDeviceSamplingConfig is consumed at trace time by prepare_inference, so do_sample=True is baked into the compiled graph. There's no runtime override.
3. Inside the validation harness (run_accuracy_check → logit_validation → generate_fn), the call at main.py:604 is model.generate(..., do_sample=False, ...). That kwarg is silently ignored: when self.on_device_sampling=True, hf_adapter._sample uses outputs.tokens from the on-device multinomial sampler (hf_adapter.py:229-230), not the host-side do_sample argument.
4. Net effect: every accuracy iteration falls into logit_validation's divergence-handling fallback path, which is slower and non-deterministic. Multiple teams have had to programmatically work around this by mutating args.do_sample = False before prepare_inference.

The fix converts the validation accuracy run into the deterministic greedy/argmax path it was always supposed to use, simply by allowing --no-do-sample to actually parse as False.

The fix

# Before
parser.add_argument("--do-sample", type=bool, default=True)
parser.add_argument("--enable-bucketing", type=bool, default=True)

# After
parser.add_argument(
    "--do-sample",
    action=argparse.BooleanOptionalAction,
    default=True,
    help="Enable multinomial sampling on the on-device sampler. "
         "Pass --no-do-sample to force greedy/argmax (required for "
         "deterministic logit_validation accuracy checks).",
)
parser.add_argument(
    "--enable-bucketing",
    action=argparse.BooleanOptionalAction,
    default=True,
    help="Enable bucketed compilation. Pass --no-enable-bucketing to disable.",
)

argparse.BooleanOptionalAction is the standard library idiom for paired --flag / --no-flag boolean options. It has been part of argparse since Python 3.9; the contest harness already requires Python ≥3.10 transitively via neuronx_distributed.

Test plan

• Empirically verified that --no-do-sample → False, --do-sample → True, no-flag → True (default).
• Empirically verified that --no-enable-bucketing → False, --enable-bucketing → True, no-flag → True (default).
• No change to any other CLI flag, function body, or import (argparse was already imported at main.py:3).