Transformers

Transforms a simple kernel_mapping of the form: { “RMSNorm”: “kernels-community/layer_norm:LlamaRMSNorm”, … },

or

{ “RMSNorm”: { “cuda”: “kernels-community/layer_norm:LlamaRMSNorm”, “rocm”: “kernels-community/layer_norm:LlamaRMSNorm”, … }, … }

into a nested mapping:

{ “RMSNorm”: { “cuda”: { Mode.INFERENCE: LayerRepository( repo_id=“kernels-community/layer_norm”, layer_name=“LlamaRMSNorm”, ) } } }

that’s compatible with the kernels library.

The device is inferred from the model’s parameters if not provided. The Mode is inferred from the model’s training state.

Validates the kernel_mapping to ensure that:

1. Each layer_name in the mapping is registered in the model (i.e., the model contains a module with a matching kernel_layer_name).
2. Each kernel value is either a string of the form ‘org/repo:layer_name’ or a dict mapping device types (“cuda”, “rocm”, “xpu”) to such strings.
3. Each device key in a dict is one of “cuda”, “rocm”, or “xpu”.
4. Each repo_name is a valid repository and layer name in the format ‘org/repo:layer_name’ (i.e., a string containing both a slash and a colon).