feat: support parsing dtype field from config.

[DongheJin]

No description provided.

[gemini-code-assist]

Code Review

This pull request introduces a normalize_config_torch_dtype function to standardize the handling of dtype and torch_dtype in model configuration files. Specifically, it ensures that if a configuration only provides dtype, it is also used for torch_dtype when loading model and quantization arguments. A new test case has been added to validate this behavior. The reviewer pointed out that the current implementation of normalize_config_torch_dtype involves an inefficient serialization/deserialization roundtrip for JsonReader creation, suggesting a more direct construction method for performance improvement.

[gemini-code-assist]

The current implementation of normalize_config_torch_dtype creates a new JsonReader by first serializing the nlohmann::json object to a string (config.dump()) and then deserializing it (normalized_reader.parse_text()). This roundtrip serialization and deserialization is inefficient and introduces unnecessary overhead. It would be more performant if the JsonReader class provided a constructor that directly accepts an nlohmann::json object, allowing for direct construction without the intermediate string conversion.

[yq33victor]

LGTM