feat(customizer): expose all hyperparams supported in backend lib#371
Open
anubhutivyas wants to merge 1 commit into
Open
feat(customizer): expose all hyperparams supported in backend lib#371anubhutivyas wants to merge 1 commit into
anubhutivyas wants to merge 1 commit into
Conversation
Signed-off-by: anubhutiv <anubhutiv@nvidia.com>
📝 WalkthroughWalkthroughTwo parallel expansions add optional hyperparameter fields to automodel and unsloth services. Automodel gains LoRA exclusions, Triton toggle, attention backend selection, sequence-packing controls, and optimizer/LR-decay fields. Unsloth gains rope scaling, DoRA/LoRA init options, ChangesAutomodel pass-2 hyperparameter expansion
Unsloth pass-2 hyperparameter expansion
Hyperparameter reference documentation
Possibly related PRs
Suggested labels
Suggested reviewers
🚥 Pre-merge checks | ✅ 4 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (4 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing Touches📝 Generate docstrings
🧪 Generate unit tests (beta)
Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 2
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@services/automodel/src/nmp/automodel/tasks/training/backends/config.py`:
- Around line 257-261: The optimizer_targets.get() call with a default fallback
to Adam silently masks configuration errors when an unknown optimizer_name is
provided. Instead of using the get method with a default, explicitly check if
customizer_config.optimizer.optimizer_name exists in the optimizer_targets
dictionary and raise a clear error if it does not, ensuring invalid optimizer
configurations fail immediately rather than silently defaulting to Adam.
In `@services/unsloth/src/nmp/unsloth/tasks/training/backends/unsloth_sft.py`:
- Around line 85-96: The issue is that the `build_peft_kwargs` function relies
on a runtime assertion to check that `lora` is not None, but the schema allows
`finetuning_type="lora"` with `lora=None`, which will cause a crash at runtime.
Instead of relying on runtime assertions, enforce this invariant at the schema
level in `TrainingSpec` by either making `lora` required when
`finetuning_type="lora"` or providing an appropriate default value. This ensures
backend helper functions like `build_peft_kwargs` never receive invalid input
shapes without needing runtime assertions.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Enterprise
Run ID: af7a496d-dc75-454e-8083-359c419f6870
📒 Files selected for processing (12)
plugins/nemo-automodel/src/nemo_automodel_plugin/schema.pyplugins/nemo-customizer/src/nemo_customizer/skills/nemo-customizer/references/hyperparameters.mdservices/automodel/src/nmp/automodel/adapter.pyservices/automodel/src/nmp/automodel/api/v2/jobs/schemas.pyservices/automodel/src/nmp/automodel/app/jobs/training/compiler.pyservices/automodel/src/nmp/automodel/app/jobs/training/schemas.pyservices/automodel/src/nmp/automodel/tasks/training/backends/config.pyservices/automodel/tests/test_adapter.pyservices/automodel/tests/test_compiler.pyservices/unsloth/src/nmp/unsloth/schemas.pyservices/unsloth/src/nmp/unsloth/tasks/training/backends/unsloth_sft.pyservices/unsloth/tests/test_model_load_kwargs.py
Comment on lines
+257
to
261
| # Map the optimizer choice to its torch class. Defaults to Adam | ||
| optimizer_targets = {"adam": "torch.optim.Adam", "adamw": "torch.optim.AdamW"} | ||
| cfg["optimizer"] = { | ||
| "_target_": "torch.optim.Adam", | ||
| "_target_": optimizer_targets.get(customizer_config.optimizer.optimizer_name, "torch.optim.Adam"), | ||
| "lr": customizer_config.optimizer.learning_rate, |
There was a problem hiding this comment.
Fail fast on unknown optimizer names instead of silently using Adam.
Line 260 masks bad config values by defaulting to Adam. That can run with a different optimizer than requested without any error.
Suggested fix
# Map the optimizer choice to its torch class. Defaults to Adam
optimizer_targets = {"adam": "torch.optim.Adam", "adamw": "torch.optim.AdamW"}
+ optimizer_target = optimizer_targets.get(customizer_config.optimizer.optimizer_name)
+ if optimizer_target is None:
+ raise ValueError(
+ f"Unsupported optimizer '{customizer_config.optimizer.optimizer_name}'. "
+ f"Expected one of: {', '.join(sorted(optimizer_targets))}."
+ )
cfg["optimizer"] = {
- "_target_": optimizer_targets.get(customizer_config.optimizer.optimizer_name, "torch.optim.Adam"),
+ "_target_": optimizer_target,
"lr": customizer_config.optimizer.learning_rate,
"weight_decay": customizer_config.optimizer.weight_decay,
"betas": [customizer_config.optimizer.beta1, customizer_config.optimizer.beta2],
"eps": customizer_config.optimizer.eps, # Adam epsilon for numerical stability
}🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
In `@services/automodel/src/nmp/automodel/tasks/training/backends/config.py`
around lines 257 - 261, The optimizer_targets.get() call with a default fallback
to Adam silently masks configuration errors when an unknown optimizer_name is
provided. Instead of using the get method with a default, explicitly check if
customizer_config.optimizer.optimizer_name exists in the optimizer_targets
dictionary and raise a clear error if it does not, ensuring invalid optimizer
configurations fail immediately rather than silently defaulting to Adam.
Comment on lines
+85
to
+96
| def build_peft_kwargs(spec: UnslothJobOutput, *, gradient_checkpointing: bool | str) -> dict[str, Any]: | ||
| """Assemble ``FastLanguageModel.get_peft_model`` kwargs for a LoRA run. | ||
|
|
||
| Torch-free (unit-testable). Caller resolves ``gradient_checkpointing`` from | ||
| ``spec.training.use_gradient_checkpointing`` (the JSON literal → ``True`` / | ||
| ``False`` / ``"unsloth"`` mapping). Optional knobs (``loftq_config``, | ||
| ``modules_to_save``, ``layers_to_transform``, ``layer_replication``) are only | ||
| emitted when set so PEFT/Unsloth see absence, not ``None``. | ||
| """ | ||
| lora = spec.training.lora | ||
| assert lora is not None # validated by UnslothJobInput | ||
| kwargs: dict[str, Any] = { |
There was a problem hiding this comment.
Enforce the LoRA invariant in schema instead of relying on runtime assertions.
Line 95 assumes training.lora is present, but the schema default allows finetuning_type="lora" with lora=None. That path will crash at runtime. Enforce/default this in TrainingSpec so backend helpers never receive an invalid shape.
Proposed fix (schema-level invariant/defaulting)
diff --git a/services/unsloth/src/nmp/unsloth/schemas.py b/services/unsloth/src/nmp/unsloth/schemas.py
@@
-from pydantic import BaseModel, ConfigDict, Field
+from pydantic import BaseModel, ConfigDict, Field, model_validator
@@
class TrainingSpec(BaseModel):
@@
use_gradient_checkpointing: Literal["unsloth", "true", "false"] = "unsloth"
+
+ `@model_validator`(mode="after")
+ def _enforce_lora_invariant(self) -> TrainingSpec:
+ if self.finetuning_type == "lora" and self.lora is None:
+ self.lora = LoRAParams()
+ if self.finetuning_type == "all_weights" and self.lora is not None:
+ raise ValueError("training.lora must be omitted when finetuning_type='all_weights'")
+ return self🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
In `@services/unsloth/src/nmp/unsloth/tasks/training/backends/unsloth_sft.py`
around lines 85 - 96, The issue is that the `build_peft_kwargs` function relies
on a runtime assertion to check that `lora` is not None, but the schema allows
`finetuning_type="lora"` with `lora=None`, which will cause a crash at runtime.
Instead of relying on runtime assertions, enforce this invariant at the schema
level in `TrainingSpec` by either making `lora` required when
`finetuning_type="lora"` or providing an appropriate default value. This ensures
backend helper functions like `build_peft_kwargs` never receive invalid input
shapes without needing runtime assertions.
Contributor
|
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary by CodeRabbit
Release Notes
New Features
Documentation