* catch StrictDataclassClassValidationError

Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com>

* test: limit nightly to stepfun recipe for stepfun CI repro

Temporarily prune nightly_recipes.yml to the single
stepfun/step_3.5_flash_hellaswag_pp.yaml recipe to iterate on the
StrictDataclassClassValidationError fix without paying for the full
nightly matrix. Not intended for merge.

Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com>

* fix: retry NeMoAutoTokenizer load when config trips layer_types validator

AutoTokenizer.from_pretrained internally calls AutoConfig.from_pretrained
to resolve the tokenizer class. For checkpoints whose config has
layer_types longer than num_hidden_layers (e.g. stepfun-ai/Step-3.5-Flash),
newer transformers rejects the config and huggingface_hub wraps the
ValueError in StrictDataclassClassValidationError (not a ValueError
subclass). The previous get_hf_config fix only covered the model-load
path; the tokenizer path hit the same failure independently.

On that specific validator failure, preload a config via get_hf_config
(which truncates layer_types) and retry the tokenizer load with an
explicit config=, bypassing the internal AutoConfig call.

Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com>

* refactor: relax validate_layer_type globally instead of preloading config

The previous tokenizer retry preloaded a fixed config via get_hf_config
and re-entered AutoTokenizer.from_pretrained with an explicit config=.
That round-trip is brittle (reconstructs a config the tokenizer does not
use) and only fixes the tokenizer call site.

Replace it with relax_layer_types_validator(): a one-shot monkey-patch
that swaps PretrainedConfig.validate_layer_type with a no-op and rewrites
the already-frozen validator entries in every live subclass's
__class_validators__ list. After that, any downstream call that instantiates
a config with mismatched layer_types/num_hidden_layers skips the check.

The tokenizer retry now just applies the patch and re-invokes
super().from_pretrained(...) with the original kwargs.

Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com>

* fix: retry VLM AutoProcessor load on layer_types validation failure

AutoProcessor.from_pretrained internally loads AutoConfig, so configs
whose layer_types length differs from num_hidden_layers trip
validate_layer_type through the processor path too. Previously the
VLM build_dataloader caught the error under a broad except and silently
set processor=None, producing a cryptic downstream failure.

On the specific validator signature, call relax_layer_types_validator()
and retry AutoProcessor.from_pretrained once. Unrelated exceptions keep
the original fall-through to processor=None with a warning. LLM tokenizer
path is already covered via NeMoAutoTokenizer.

Also pass --force-exclude to the ruff pre-commit hooks so the
tests/ exclusion already declared in pyproject.toml takes effect when
pre-commit passes files explicitly.

Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com>

* revert

Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com>

* revert

Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com>

* fmt

Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com>

---------

Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com>
Signed-off-by: NeMo Bot <nemo-bot@nvidia.com>