Support target_modules: all in YAML configs

[marksverdhei]

Problem

When using target_modules: all in a YAML config, bakery fails with:

ValueError: Target modules all not found in the base model.

The LoraConfig.target_modules field is typed as List[str], so YAML scalar all gets parsed as the string "all" and passed directly to PEFT, which doesn't recognize it.

Expected behavior

target_modules: all (or target_modules: all-linear) in YAML should work and map to PEFT's "all-linear" string, which targets all linear layers.

Suggested fix

In config.py, add a __post_init__ to LoraConfig that normalizes the value:

def __post_init__(self):
    if self.target_modules == "all" or self.target_modules == ["all"]:
        self.target_modules = "all-linear"
    elif self.target_modules == ["all-linear"]:
        self.target_modules = "all-linear"

The type annotation may also need to accept Optional[Union[str, List[str]]] or similar to handle both forms from HfArgumentParser.

Workaround

Use a list with the PEFT-recognized string:

target_modules:
  - all-linear

[marksverdhei]

Resolved — this landed in #35 (commit c524189) and is already on main.

LoraConfig.__post_init__ (src/bakery/config.py) normalizes the scalar/list forms to PEFT's "all-linear":

def __post_init__(self):
    # Support "all" as shorthand for "all-linear" (PEFT convention)
    if self.target_modules == "all" or self.target_modules == ["all"]:
        self.target_modules = "all-linear"
    elif self.target_modules == ["all-linear"]:
        self.target_modules = "all-linear"

Verified:

• tests/test_config.py — LoraConfig(target_modules="all") → "all-linear" (direct-instantiation tests pass here).
• tests/test_cli.py::test_target_modules_all_normalized / test_target_modules_list_all_normalized — exercise the real HfArgumentParser.parse_yaml_file path for both target_modules: all (scalar) and [all] (list). These skip only on Python ≥3.14 (an unrelated transformers use_liger_kernel help-text %-format bug, guarded by _HF_PARSER_BROKEN); they run on supported Pythons.

Note on the issue's secondary suggestion: the annotation was intentionally left Optional[List[str]] rather than widened to Union[str, List[str]]. The __post_init__ approach already works end-to-end with HfArgumentParser, and changing the annotation would alter how HfArgumentParser parses the field — unnecessary risk for no behavioral gain.

Closing as resolved; reopen if the YAML path misbehaves on your runtime.