generate/adapter_v2.py

import sys
import time
import warnings
from pathlib import Path
from typing import Optional

import lightning as L
import torch

# support running without installing as a package
wd = Path(__file__).parent.parent.resolve()
sys.path.append(str(wd))

from generate import generate
from lit_llama import Tokenizer
from lit_llama.adapter import LLaMA
from lit_llama.utils import lazy_load, llama_model_lookup, quantization
from lit_llama.adapter_v2 import add_adapter_v2_parameters_to_linear_layers
from scripts.prepare_alpaca import generate_prompt


def main(
    prompt: str = "What food do lamas eat?",
    input: str = "",
    adapter_path: Path = Path("out/adapter_v2/alpaca/lit-llama-adapter-finetuned.pth"),
    pretrained_path: Path = Path("checkpoints/lit-llama/7B/lit-llama.pth"),
    tokenizer_path: Path = Path("checkpoints/lit-llama/tokenizer.model"),
    quantize: Optional[str] = None,
    max_new_tokens: int = 100,
    top_k: int = 200,
    temperature: float = 0.8,
) -> None:
    """Generates a response based on a given instruction and an optional input.
    This script will only work with checkpoints from the instruction-tuned LLaMA-Adapter model.
    See `finetune_adapter_v2.py`.

    Args:
        prompt: The prompt/instruction (Alpaca style).
        adapter_path: Path to the checkpoint with trained adapter weights, which are the output of
            `finetune_adapter_v2.py`.
        input: Optional input (Alpaca style).
        pretrained_path: The path to the checkpoint with pretrained LLaMA weights.
        tokenizer_path: The tokenizer path to load.
        quantize: Whether to quantize the model and using which method:
            ``"llm.int8"``: LLM.int8() mode,
            ``"gptq.int4"``: GPTQ 4-bit mode.
        max_new_tokens: The number of generation steps to take.
        top_k: The number of top most probable tokens to consider in the sampling process.
        temperature: A value controlling the randomness of the sampling process. Higher values result in more random
            samples.
    """
    assert adapter_path.is_file()
    assert pretrained_path.is_file()
    assert tokenizer_path.is_file()

    precision = "bf16-true" if torch.cuda.is_available() and torch.cuda.is_bf16_supported() else "32-true"
    fabric = L.Fabric(devices=1, precision=precision)

    print("Loading model ...", file=sys.stderr)
    t0 = time.time()
    with lazy_load(pretrained_path) as pretrained_checkpoint, lazy_load(adapter_path) as adapter_checkpoint:
        name = llama_model_lookup(pretrained_checkpoint)

        with fabric.init_module(empty_init=True), quantization(mode=quantize):
            model = LLaMA.from_name(name)
            add_adapter_v2_parameters_to_linear_layers(model)

        # 1. Load the pretrained weights
        model.load_state_dict(pretrained_checkpoint, strict=False)
        # 2. Load the fine-tuned adapter weights
        model.load_state_dict(adapter_checkpoint, strict=False)

    print(f"Time to load model: {time.time() - t0:.02f} seconds.", file=sys.stderr)

    model.eval()
    model = fabric.setup(model)

    tokenizer = Tokenizer(tokenizer_path)
    sample = {"instruction": prompt, "input": input}
    prompt = generate_prompt(sample)
    encoded = tokenizer.encode(prompt, bos=True, eos=False, device=model.device)
    prompt_length = encoded.size(0)

    t0 = time.perf_counter()
    y = generate(model, encoded, max_new_tokens, temperature=temperature, top_k=top_k, eos_id=tokenizer.eos_id)
    t = time.perf_counter() - t0

    model.reset_cache()
    output = tokenizer.decode(y)
    output = output.split("### Response:")[1].strip()
    print(output)

    tokens_generated = y.size(0) - prompt_length
    print(f"\n\nTime for inference: {t:.02f} sec total, {tokens_generated / t:.02f} tokens/sec", file=sys.stderr)
    if fabric.device.type == "cuda":
        print(f"Memory used: {torch.cuda.max_memory_reserved() / 1e9:.02f} GB", file=sys.stderr)


if __name__ == "__main__":
    from jsonargparse import CLI

    torch.set_float32_matmul_precision("high")
    warnings.filterwarnings(
        # Triggered internally at ../aten/src/ATen/EmptyTensor.cpp:31
        "ignore", 
        message="ComplexHalf support is experimental and many operators don't support it yet"
    )
    CLI(main)