Do the forward pass. If LoRA's weights are merged with pretrained ones then it's a simple matrix multiplication. If not, then multipl

Kernel for computing the matmul C = A x B.T. A has shape (M, K), B has shape (N, K) and C has shape (M, N)

Generates text samples based on a pre-trained LLaMA model and tokenizer. Args: prompt: The prompt string to use for generating the sample

Generates text samples based on a pre-trained LLaMA model and tokenizer. Args: datasets: The datasets to use as a comma separated string

Generates text samples based on a pre-trained LLaMA model and tokenizer. Args: datasets: The datasets to use as a comma separated string

Generates text samples based on a pre-trained LLaMA model and tokenizer finetuned with LoRA. Args: datasets: The datasets to use a

Generates text samples based on a pre-trained LLaMA model and tokenizer. Args: datasets: The datasets to use as a comma separated string

Generates text samples based on a pre-trained LLaMA model and tokenizer. Args: checkpoint_path: The checkpoint path to load. outp

Merges lora weights to base model. Args: accelerator: The hardware to run on. Possible choices are: ``"cpu"``, ``"cuda"``, ``

Generates a response based on a given instruction and an optional input_sentence. This script will only work with checkpoints from the instruction

Generates text samples based on a pre-trained LLaMA model and tokenizer. Args: prompt: The prompt string to use for generating the sample

Generates a response based on a given instruction and an optional input_sentence. This script will only work with checkpoints from the instruction

Generates a response based on a given instruction and an optional input_sentence. This script will only work with checkpoints from the instruction

Prepare any dataset for finetuning (akin to Shakespheare full tuning). The output is a training and validation dataset saved as `train.pt` and `v

Prepare the "Red Pajama" dataset. We assume tokenizer has been trained (i.e. we reuse LLaMA's tokenizer model).

Prepare the Dolly dataset for instruction tuning. The output is a training and validation dataset saved as `train.pt` and `val.pt`, which

Prepare the Alpaca dataset for instruction tuning. The output is a training and validation dataset saved as `train.pt` and `val.pt`, whic

Prepare the "Tiny Shakespeare" dataset.

Tests backward-compatible loading of checkpoints after the `gating_factor` was extended per-head in PR #297.

Test parity between our implementation of LLaMA-Adapter and the reference code.