Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@examples/megatron_bridge/README.md`:
- Around line 103-111: The bash example uses inline comments after a
backslash-continued line (e.g., the lines containing `--json_keys text \  #
change to your JSON key if needed` and `--tokenizer Qwen/Qwen3-0.6B \  # fyi,
all models...`), which breaks line continuation; remove those trailing inline
comments or move them to their own lines above the command, or place comments
only on the final non-continued line, and ensure every `\` is the last character
on the line so the `python -m
modelopt.torch.utils.plugins.megatron_preprocess_data` command and flags
(`--jsonl_paths`, `--json_keys`, `--tokenizer`, `--output_dir`, `--workers`,
`--max_sequence_length`) are valid when copy-pasted.

In `@modelopt/torch/utils/plugins/megatron_preprocess_data.py`:
- Around line 301-315: Move the creation of the output directory to before
calling _download_hf_dataset so the downloader can write files; specifically,
call Path(output_dir).mkdir(parents=True, exist_ok=True) prior to invoking
_download_hf_dataset(dataset, output_dir, json_keys, subset=subset, split=split)
in the code block that contains the dataset/download logic, leaving the rest of
the file discovery logic for input_dir/jsonl_paths unchanged.
- Around line 229-255: _download_hf_dataset currently builds jsonl_file_path
using ds_path which may contain slashes and never ensures parent directories
exist, causing FileNotFoundError when ds.to_json is called; fix by sanitizing or
using only the dataset name (not the HF org prefix) when composing
jsonl_file_path and ensure the parent directory is created before writing:
derive a safe filename from ds_path (or replace '/' with '_' or use
Path(ds_path).name), compute parent = Path(output_dir)/"raw"/<safe_name> and
call parent.mkdir(parents=True, exist_ok=True) before checking/creating
jsonl_file_path and before calling ds.to_json; also ensure output_dir is created
earlier (or rely on the same mkdir(parents=True)) so writes never fail.

In `@tests/_test_utils/torch/megatron/utils.py`:
- Line 23: The shared utils module uses skip_if_no_megatron(te_required=False)
which allows import to run TE-dependent megatron.core model imports (e.g.,
GPTModel, MambaModel, GPTInferenceWrapper) and causes ImportError when TE is
missing; update the guard so the module is only imported when TE is available by
changing the skip invocation to skip_if_no_megatron(te_required=True) (or, if
utility functions truly don't need TE, move the skip_if_no_megatron call to
after the TE-free helper definitions and keep TE imports behind a separate
guarded block); ensure initialize_for_megatron and any TE-dependent symbols are
only loaded under the TE-required guard to keep imports consistent with
tests/models.py.

In `@tests/unit/torch/utils/test_megatron_preprocess_data.py`:
- Around line 98-121: The test test_megatron_preprocess_data_with_hf_dataset
makes live network calls and also writes JSON into a non-existent raw/
subdirectory in _download_hf_dataset; update the test to skip in offline CI by
adding a guard (e.g., pytest.mark.skipif or an env-check like
SKIP_NETWORK_TESTS) before running megatron_preprocess_data, and in
_download_hf_dataset ensure the target directory (the constructed raw/ path) is
created (os.makedirs(..., exist_ok=True)) before calling ds.to_json(); reference
the test function name test_megatron_preprocess_data_with_hf_dataset, the helper
_download_hf_dataset, and the call sites around load_dataset and ds.to_json when
implementing these changes.

---

Outside diff comments:
In `@modelopt/torch/utils/plugins/megatron_preprocess_data.py`:
- Around line 172-179: The loop variable i is undefined when encoded_docs is
empty, causing UnboundLocalError at the final self._print_processing_stats(i,
...) call; fix by initializing i (e.g., i = 0) before the for i, (doc,
sentence_lens, (doc_len, enc_len)) in enumerate(...) loop or by using a safe
fallback when calling self._print_processing_stats (e.g., pass i if defined else
0); update the code around the enumerate block in megatron_preprocess_data.py so
that i is always defined before the final force_print call.

---

Nitpick comments:
In `@modelopt/torch/prune/plugins/mcore_minitron.py`:
- Line 320: The comment TODO notes that the name hybrid_override_pattern should
be renamed to hybrid_layer_pattern after Megatron-LM#3377; create a tracking
issue in your repo (or project board) documenting this rename, include the file
and symbol names (hybrid_override_pattern in
modelopt/torch/prune/plugins/mcore_minitron.py), mention the PR that triggers it
(NVIDIA/Megatron-LM#3377), and list required follow-ups (update the TODO
comment, rename the symbol across codebase, adjust any docs/tests that reference
hybrid_override_pattern) so the change won't be lost.

In `@modelopt/torch/utils/plugins/megatron_preprocess_data.py`:
- Line 271: The parameter json_keys is defined with a mutable default list
(`json_keys: list[str] = ["text"]`), so change the function signature to use
None as the default (e.g., `json_keys: list[str] | None = None`) and add a guard
at the top of the function (in
modelopt/torch/utils/plugins/megatron_preprocess_data.py) that sets `json_keys =
["text"]` when `json_keys is None`; reference the parameter name `json_keys` to
locate and update the code and ensure any type hints are adjusted accordingly.