refactor(automodel)!: delete the inert log_every_n_steps

It described itself as "Logging frequency in steps. Controls how often training
metrics are logged" and controlled nothing. No code read it, it never reached
the recipe config, and `config.py` builds step_scheduler with no logging cadence
at all. A repo-wide grep found only the declaration.

The design doc called it "documented to users" and weighed deletion against
wiring it to StepScheduler.log_remote_every_steps, which gates W&B and MLflow
from inside log_train_metrics and so would have left our wrapper firing every
step. That wiring had a real argument behind it — it is the automodel analogue
of unsloth's logging_steps, which we deliberately keep separate from our own
reporting knob.

Checking the reachability settled it the other way. `api/v2/jobs/` holds a
schema file and no routes; CustomizationJobOutput is an internal intermediate
between the plugin adapter and the compiler; the submitter-facing
AutomodelJobOutput never carried the field; and regenerating the specs after
removing it produces no diff, because it was in none of them. No request could
ever set it. So there is no existing behaviour to preserve and no user
expectation to honour, and wiring it would have been adding a feature under
cover of a cleanup — a second field that sounds like "how often are metrics
logged", sitting next to progress_reporting.max_points, which is the knob it
was reaching for and which now exists.

Removal is invisible: the model ignores extras, so a stored spec that still
carries the key parses exactly as before. That property is the whole reason
this is safe rather than breaking, so it is pinned by a test — adding
extra="forbid" here later would otherwise turn silent tolerance into a hard
failure for precisely those specs.

Marked as a breaking change by convention, since the field was public in shape
even though it was unreachable in practice.

Signed-off-by: Albert Cui <albcui@nvidia.com>

Author: albcui

services/automodel/src/nmp/automodel/api/v2/jobs/schemas.py

@@ -211,14 +211,17 @@ class _TrainingBase(BaseModel):
         gt=0,
         description="Max training steps. Overrides epochs if set.",
     )
-    log_every_n_steps: Optional[int] = Field(
-        default=None,
-        description="Logging frequency in steps. Controls how often training metrics are logged.",
-    )
     val_check_interval: Optional[float] = Field(
         default=None,
         description="Validation interval. Float <= 1.0 is fraction of epoch; > 1.0 is step count.",
     )
+    # `log_every_n_steps` used to sit here, described as "Logging frequency in steps.
+    # Controls how often training metrics are logged." It controlled nothing: no
+    # code read it, it never reached the recipe config, and it was absent from the
+    # submitter-facing plugin schema and from every generated spec, so no request
+    # could set it in the first place. `progress_reporting` below is the knob it
+    # was reaching for. Removing it is invisible -- this model ignores extras, so a
+    # stored spec that still carries the key parses exactly as it did before.
     progress_reporting: ProgressReportingConfig = Field(default_factory=ProgressReportingConfig)
 
     # --- Batch ---

services/automodel/tests/test_compiler.py

@@ -126,6 +126,35 @@ def test_the_reporting_budget_reaches_the_training_step_config() -> None:
     assert cfg["schedule"]["progress_reporting"]["curves"] == ["loss"]
 
 
+def test_a_spec_still_carrying_log_every_n_steps_compiles() -> None:
+    """The removed field was inert, and removing it has to stay invisible.
+
+    It described itself as controlling how often training metrics are logged and
+    controlled nothing: nothing read it, it never reached the recipe config, and
+    it was in neither the submitter-facing plugin schema nor any generated spec.
+    What makes deleting it safe rather than breaking is that this model ignores
+    extras -- so a stored spec that still carries the key parses as it always
+    did. Pinned because a later `extra="forbid"` here would turn that silent
+    tolerance into a hard failure for exactly those specs.
+    """
+    from nmp.automodel.app.jobs.training.compiler import compile_training_step
+    from nmp.customization_common.training.reporting import DEFAULT_MAX_POINTS
+
+    training = SFTTraining.model_validate({"learning_rate": 1e-4, "log_every_n_steps": 10})
+    assert not hasattr(training, "log_every_n_steps")
+
+    job_output = CustomizationJobOutput(
+        model="default/test-target",
+        dataset="default/my-dataset",
+        training=training,
+        output=OutputResponse(name="out", type="adapter", fileset="out-fs"),
+    )
+    step = compile_training_step(job_output, base_env=[], me=_make_mock_model_entity())
+    cfg = step.config if hasattr(step, "config") else step["config"]
+
+    assert cfg["schedule"]["progress_reporting"]["max_points"] == DEFAULT_MAX_POINTS
+
+
 def test_the_reporting_budget_survives_the_plugin_adapter() -> None:
     """The adapter flattens the plugin's schedule block and is easy to drop a field from."""
     from nmp.automodel.adapter import automodel_spec_to_compiler_output