feat: Add config serialization and loading support

[rahul-tuli]

PR 1: Add Config Serialization and Loading Support

Summary

Implements from_pretrained, to_dict, and from_dict methods for SpeculatorModelConfig to enable loading and saving configurations from the HuggingFace Hub and local paths. This is a prerequisite for model implementations that need these serialization capabilities.

Changes

Core Implementation

• SpeculatorModelConfig.from_pretrained(): Load configs from HF Hub or local paths with automatic subclass detection
• SpeculatorModelConfig.from_dict(): Polymorphic instantiation from dictionaries using registry pattern
• EagleSpeculatorConfig.to_dict(): Proper serialization of all config attributes
• EagleSpeculatorConfig.from_dict(): Handle nested config reconstruction (transformer_layer_config, etc.)

Technical Details

Registry Pattern

# Base class uses registry for polymorphic instantiation
model_type = config_dict.get("speculators_model_type")
config_class = cls.registry.get(model_type)
return config_class(**config_dict)

Nested Config Handling

• Automatically converts dict representations to proper config objects:
  • transformer_layer_config → LlamaConfig
  • speculators_config → SpeculatorsConfig
  • token_proposals → TokenProposalConfig

Comment Improvements

• Updated comments to explain "why" rather than "what"
• Removed redundant documentation

Testing

• test_config_loading.py: Comprehensive unit tests for serialization roundtrips
• Test Coverage:
  • to_dict/from_dict roundtrip preservation
  • from_pretrained with architecture inference
  • Nested config handling
  • Extra kwargs override behavior

API Changes

• No breaking changes - only adds new functionality
• Maintains compatibility with HuggingFace PretrainedConfig

Example Usage

# Load from HuggingFace Hub
config = SpeculatorModelConfig.from_pretrained("org/model-name")

# Load from local path
config = SpeculatorModelConfig.from_pretrained("/path/to/model")

# Serialize and deserialize
config_dict = config.to_dict()
new_config = SpeculatorModelConfig.from_dict(config_dict)

Review Notes

• The registry pattern allows future speculator types to be added without modifying base class
• All changes maintain backward compatibility
• Test coverage includes edge cases like missing model_type fields

[nm-red-hat-upstream-automation-bot]

📦 Build Artifacts Available
The build artifacts (.whl and .tar.gz) have been successfully generated and are available for download: https://github.com/neuralmagic/speculators/actions/runs/15896923777/artifacts/3408415086.
They will be retained for up to 30 days.

[nm-red-hat-upstream-automation-bot]

📦 Build Artifacts Available
The build artifacts (.whl and .tar.gz) have been successfully generated and are available for download: https://github.com/neuralmagic/speculators/actions/runs/15934725470/artifacts/3422000105.
They will be retained for up to 30 days.

[markurtz]

Noting here, working through simplifying this diff

[nm-red-hat-upstream-automation-bot]

📦 Build Artifacts Available
The build artifacts (.whl and .tar.gz) have been successfully generated and are available for download: https://github.com/neuralmagic/speculators/actions/runs/15974113330/artifacts/3432197377.
They will be retained for up to 30 days.

[nm-red-hat-upstream-automation-bot]

📦 Build Artifacts Available
The build artifacts (.whl and .tar.gz) have been successfully generated and are available for download: https://github.com/neuralmagic/speculators/actions/runs/15974267522/artifacts/3432251988.
They will be retained for up to 30 days.

[markurtz]

Removing my request changes block now that it's been reworked to utilize Pydantic builtins and reduce the complexity of the logic

[Copilot]

Pull Request Overview

This PR adds configuration serialization and deserialization support for speculator models, including methods for loading from local paths or the HuggingFace Hub as well as proper handling of nested configurations. Key changes include:

• Implementing from_pretrained, from_dict, and to_dict methods with registry support.
• Adding new field serialization and validation logic for EagleSpeculatorConfig.
• Updating and expanding unit and integration tests for configuration marshalling.

Reviewed Changes

Copilot reviewed 7 out of 8 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
tox.ini	Updated Python environment list to drop py38.
tests/unit/test_config.py	Refactored tests with updated markers and added additional marshalling tests.
tests/unit/models/test_eagle.py	Expanded coverage for EagleSpeculatorConfig including backwards compatibility.
tests/integration/test_config.py	Replaced a temporary failure test for from_pretrained with a stub.
src/speculators/models/eagle.py	Added field serialization and validation for transformer_layer_config.
src/speculators/config.py	Revised configuration loading (from_pretrained, from_dict) and init logic.
pyproject.toml	Updated lint rule codes to match new guidelines.

Comments suppressed due to low confidence (1)

tests/integration/test_config.py:44

• This placeholder test for from_pretrained does not currently validate actual loading behavior. Once the functionality is fully implemented, please add comprehensive tests to ensure proper integration.

    assert True

[brian-dellabetta]

nit to remove extraneous test, otherwise LGTM