run-tiny-merge.py

#!/usr/bin/env python3

import transformers
import torch
import logging
from ddare.merge import merge_tensors
from ddare.tensor import (
    dare_ties_sparsification,
    relative_norm,
    divide_tensor_into_sets
)
from ddare.util import get_device
import re
from typing import Dict, Tuple, List

# If you want to fine-tune, here's an example Unsloth fine tuning guide for:
# Alpaca + TinyLlama + RoPE Scaling full example.ipynb
# https://colab.research.google.com/drive/1AZghoNBQaMDgWJpi4RbffGM1h6raLUj9?usp=sharing

# code here was refactored from gist:
# https://gist.github.com/maldevide/08829eada04ad9bd78e46c1a3787d42b

logging.basicConfig(level=logging.INFO)
log = logging.getLogger(__name__)


def get_models(
    models: List[str],
    trust_remote_code: bool,
):
    config = {
        'torch_dtype': torch.float16,
        'low_cpu_mem_usage': False,
        'trust_remote_code': trust_remote_code,
    }
    loaded_models = []
    num_models = len(models)
    for midx, model_path in enumerate(models):
        log.info(
            f"loading model={midx + 1}/{num_models} "
            f"model={model_path} "
        )
        loaded_models.append(
            transformers.AutoModelForCausalLM.from_pretrained(
                model_path,
                **config
            )
        )
    return loaded_models


def pm(
    model,
):
    keys = model.state_dict().keys()
    log.info(f"model keys={len(keys)}")
    for i, k in enumerate(keys):
        tensor = model.state_dict()[k]
        log.info(
            f"{i:3d} {k} shape={tensor.shape} "
            f"type={tensor.dtype} dev={tensor.device} "
            f"contig={tensor.is_contiguous()}")


def run_text_test(
    model,
    tokenizer_path,
    question: str,
    device: str = "cuda",
):
    base_model = model.to(device)
    log.info(
        f"loading tokenizer={tokenizer_path}"
    )
    tokenizer = transformers.AutoTokenizer.from_pretrained(
        tokenizer_path,
        torch_dtype=torch.float16,
    )

    inputs = tokenizer(
        question,
        return_tensors="pt"
    ).to(device)
    with torch.backends.cuda.sdp_kernel(
        enable_flash=True,
        enable_math=False,
        enable_mem_efficient=False
    ):
        outputs = base_model.generate(
            **inputs,
            max_new_tokens=1000,
        )
    answer = tokenizer.decode(outputs[0], skip_special_tokens=True)
    log.info(
        "\n"
        "----------"
        "\n"
        f"tokenizer={tokenizer}\n "
        f"question:\n{question}\n"
        f"answer:\n{answer}\n"
        "----------"
    )
    base_model = base_model.to(device)


def get_layer_type(
    key: str
) -> Tuple[int, str]:
    matcher = re.compile(r"model.layers.(\d+).(.+)")
    m = matcher.match(key)
    if m is None:
        if "model.norm.weight" == key:
            return -1, "norm"
        if "model.embed_tokens.weight" == key:
            return -1, "embed"
        if "lm_head.weight" == key:
            return -1, "head"
        log.info(f"Unknown key {key}")
        return -1, "unknown"
    return int(m.group(1)), m.group(2)


def merge_model_with_ties(
    models: List[str],
    model_dst: str,
    trust_remote_code: bool = True
):
    models = get_models(
        models=models,
        trust_remote_code=trust_remote_code,
    )
    config = {}
    result_dict: Dict[str, torch.Tensor] = {}
    device = get_device()
    keys = models[0].state_dict().keys()
    num_keys = len(keys)
    for k in keys:
        block, layer_type = get_layer_type(k)
        m0: torch.Tensor = models[0].state_dict()[k]
        result = m0.clone()
        sets = divide_tensor_into_sets(tensor=m0, n_sets=4)

        # get the src layers to merge
        m = [
            models[1].state_dict()[k],
            models[2].state_dict()[k],
            models[3].state_dict()[k],
            models[4].state_dict()[k],
        ]

        # build a ratio
        ratio = {
            'to_q': 0.0,
            'to_k': 0.0,
            'to_v': 0.0,
        }.get(layer_type, .5)

        norm_ratio = 0.68
        log.info(
            f"model={k} {num_keys} shape={m0.shape} "
            f"dtype={m0.dtype} {m0.device} "
            f"raio={ratio} "
            f"contig={m0.is_contiguous()} "
            f"norm={norm_ratio}")

        # for all tensors
        for i, tensor in enumerate(m):
            if layer_type == "to_k":
                # Get to_q key
                q_base = models[0].state_dict()[k.replace("to_k", "to_q")]
                q_merge = models[i].state_dict()[k.replace("to_k", "to_q")]
                scale = relative_norm(q_merge, q_base)
                tensor = tensor.to(device) / scale
                del scale
            elif layer_type == "to_q":
                scale = relative_norm(tensor, m0)
                tensor = tensor.to(device) * scale
                del scale
            slice_mask = (
                sets == i
            ).bool()
            new_tensor = dare_ties_sparsification(
                model_a_param=m0,
                model_b_param=tensor,
                drop_rate=norm_ratio,
                ties="sum",
                rescale="off",
                device=device,
                **config)
            new_tensor = merge_tensors("slerp", m0, tensor, ratio)
            result = torch.where(slice_mask, new_tensor, result)
            del new_tensor, slice_mask

        result_dict[k] = result
    # end of merge

    log.info(
        f"done merge saving to file: {model_dst}"
    )
    out_model = (
        transformers.AutoModelForCausalLM.from_pretrained(
            model_dst,
            **config
        )
    )
    out_model.state_dict = lambda: result_dict
    out_model.save_pretrained(model_dst)


def run():
    question = (
        "why is the sky blue?"
    )
    log.info(f"merging models and asking the question: {question}")
    model_src = "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T"
    model_dst = "matlok/tinyllama-cinder-openhermes-32k"
    device = "cuda"
    config = {
        'torch_dtype': torch.float16,
        'low_cpu_mem_usage': False,
        'trust_remote_code': True,
    }
    models = [
        model_src,
        "Doctor-Shotgun/TinyLlama-1.1B-32k-Instruct",
        "Doctor-Shotgun/TinyLlama-1.1B-32k",
        "Tensoic/TinyLlama-1.1B-3T-openhermes",
        "Josephgflowers/TinyLlama-3T-Cinder-v1.3",
    ]
    merge_model_with_ties(
        models=models,
        model_dst=model_dst
    )
    log.info(f"loading newly-created file: {model_dst}")
    model = transformers.AutoModelForCausalLM.from_pretrained(
        model_dst,
        **config
    )
    log.info(
        f"loaded new model file: {model_dst} "
        f"asking question: {question} "
    )
    run_text_test(
        model=model,
        tokenizer_path=model_src,
        question=question,
        device=device,
    )
    log.info(f"done loading new model: {model} file: {model_dst}")


if __name__ == "__main__":
    run()