tinyllama-cinder-openhermes-32k

Merging AI Models like Lego Blocks

This model was merged with the following Hugging Face TinyLlama models using ties:

• TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T
• 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

How do I fine-tune this model?

Fine-tuning using Hugging Face SFTTrainer

• Fine-tuning using Hugging Face SFTTrainer

Fine-tuning using Unsloth

2024-02-07 was unable to use unsloth due to pip install issues. Maybe others in the future will have more luck:

• Alpaca + TinyLlama + RoPE Scaling full example.ipynb

How do I generate my own model merges?

This requires setting up your Hugging Face User Account Access Tokens before it will work:

If you're using the command line you can use:

huggingface-cli login

time ./run-tiny-merge.py

What's this code doing?

Here's the latest version:

#!/usr/bin/env python3

import os
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


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


def get_models(
    models: List[str],
    trust_remote_code: bool,
):
    """
    get the models

    :param models: model names to download
    :param trust_remote_code: are you sure??? True/False
    """
    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,
):
    """
    pretty print model

    :param model: show me the 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: str,
    question: str,
    device: str = "cuda",
):
    """
    run a question on the model and return the answer

    :param model: initialized model
    :param tokenizer_path: tokenizer path/name
    :param question: what are you asking?
    :param device: where do you want to run "cpu"/"gpu"?
    """
    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=True,
    ):
        outputs = base_model.generate(
            **inputs,
            max_new_tokens=256,
        )
    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)
    return tokenizer


def get_layer_type(key: str) -> Tuple[int, str]:
    """
    get the layer type

    :param key: name of the layer
    :return: layer id and name
    """
    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,
):
    """
    merge the list of models into one model
    called model_dst

    :param models: list of models to merge
    :param model_dst: name of the new model
    :param trust_remote_code: are you sure? True/False
    """
    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, 0.5)

        norm_ratio = 0.68
        log.info(
            f"model={k} {num_keys} shape={m0.shape} "
            f"dtype={m0.dtype} {m0.device} "
            f"ratio={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():
    """
    run the merge and upload the model and tokenizer

    This requires having the Hugging Face token
    set before it will work:
    ```huggingface-cli login```
    """
    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,
    )

    # clean the temp merge dir
    # remove model dir to prevent issues with the tokenizer upload
    model_org = model_dst.split("/")[0]
    if os.path.exists(model_org):
        os.system(f"rm -rf ./{model_org}")

    log.info(f"uploading model: {model_dst}")
    model.push_to_hub(model_dst)

    log.info(f"uploading src tokenizer: {model_src}")
    # reload tokenizer to save it and found on:
    # https://colab.research.google.com/drive/1PEQyJO1-f6j0S_XJ8DV50NkpzasXkrzd?usp=sharing#scrollTo=QQn30cRtAZ-P
    tokenizer = transformers.AutoTokenizer.from_pretrained(
        model_src, trust_remote_code=True
    )
    # https://huggingface.co/docs/transformers/model_sharing#use-the-pushtohub-function
    # tokenizer.push_to_hub("my-awesome-model")
    tokenizer.push_to_hub(model_dst)
    log.info(
        f"done loading new model: {model} "
        f"file: {model_dst}"
    )


if __name__ == "__main__":
    run()

Logs

Here's the logs from the code above:

time ./run-tiny-merge.py
Total VRAM 12282 MB, total RAM 85434 MB
Set vram state to: NORMAL_VRAM
Device: cuda:0 NVIDIA GeForce RTX 4070 Ti : native
VAE dtype: torch.bfloat16
INFO:__main__:merging models and asking the question: why is the sky blue?
INFO:__main__:loading model=1/5 model=TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T
config.json: 100%|█████████████████████████████████████| 560/560 [00:00<00:00, 5.23MB/s]
model.safetensors: 100%|███████████████████████████| 4.40G/4.40G [00:48<00:00, 90.2MB/s]
generation_config.json: 100%|███████████████████████████| 129/129 [00:00<00:00, 721kB/s]
INFO:__main__:loading model=2/5 model=Doctor-Shotgun/TinyLlama-1.1B-32k-Instruct
config.json: 100%|█████████████████████████████████████| 695/695 [00:00<00:00, 3.04MB/s]
pytorch_model.bin: 100%|███████████████████████████| 2.20G/2.20G [00:23<00:00, 92.6MB/s]
generation_config.json: 100%|███████████████████████████| 129/129 [00:00<00:00, 566kB/s]
INFO:__main__:loading model=3/5 model=Doctor-Shotgun/TinyLlama-1.1B-32k
config.json: 100%|█████████████████████████████████████| 686/686 [00:00<00:00, 3.57MB/s]
model.safetensors: 100%|███████████████████████████| 2.20G/2.20G [00:24<00:00, 90.5MB/s]
generation_config.json: 100%|██████████████████████████| 124/124 [00:00<00:00, 1.80MB/s]
INFO:__main__:loading model=4/5 model=Tensoic/TinyLlama-1.1B-3T-openhermes
config.json: 100%|█████████████████████████████████████| 702/702 [00:00<00:00, 2.97MB/s]
pytorch_model.bin: 100%|███████████████████████████| 2.20G/2.20G [00:23<00:00, 92.7MB/s]
generation_config.json: 100%|███████████████████████████| 124/124 [00:00<00:00, 671kB/s]
INFO:__main__:loading model=5/5 model=Josephgflowers/TinyLlama-3T-Cinder-v1.3
config.json: 100%|█████████████████████████████████████| 713/713 [00:00<00:00, 9.35MB/s]
model.safetensors: 100%|███████████████████████████| 2.20G/2.20G [00:24<00:00, 91.5MB/s]
generation_config.json: 100%|██████████████████████████| 138/138 [00:00<00:00, 1.86MB/s]
INFO:__main__:model=model.embed_tokens.weight 201 shape=torch.Size([32000, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.0.self_attn.q_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.0.self_attn.k_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.0.self_attn.v_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.0.self_attn.o_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.0.mlp.gate_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.0.mlp.up_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.0.mlp.down_proj.weight 201 shape=torch.Size([2048, 5632]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.0.input_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.0.post_attention_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.1.self_attn.q_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.1.self_attn.k_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.1.self_attn.v_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.1.self_attn.o_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.1.mlp.gate_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.1.mlp.up_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.1.mlp.down_proj.weight 201 shape=torch.Size([2048, 5632]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.1.input_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.1.post_attention_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.2.self_attn.q_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.2.self_attn.k_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.2.self_attn.v_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.2.self_attn.o_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.2.mlp.gate_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.2.mlp.up_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.2.mlp.down_proj.weight 201 shape=torch.Size([2048, 5632]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.2.input_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.2.post_attention_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.3.self_attn.q_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.3.self_attn.k_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.3.self_attn.v_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.3.self_attn.o_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.3.mlp.gate_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.3.mlp.up_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.3.mlp.down_proj.weight 201 shape=torch.Size([2048, 5632]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.3.input_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.3.post_attention_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.4.self_attn.q_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.4.self_attn.k_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.4.self_attn.v_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.4.self_attn.o_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.4.mlp.gate_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.4.mlp.up_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.4.mlp.down_proj.weight 201 shape=torch.Size([2048, 5632]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.4.input_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.4.post_attention_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.5.self_attn.q_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.5.self_attn.k_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.5.self_attn.v_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.5.self_attn.o_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.5.mlp.gate_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.5.mlp.up_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.5.mlp.down_proj.weight 201 shape=torch.Size([2048, 5632]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.5.input_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.5.post_attention_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.6.self_attn.q_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.6.self_attn.k_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.6.self_attn.v_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.6.self_attn.o_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.6.mlp.gate_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.6.mlp.up_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.6.mlp.down_proj.weight 201 shape=torch.Size([2048, 5632]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.6.input_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.6.post_attention_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.7.self_attn.q_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.7.self_attn.k_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.7.self_attn.v_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.7.self_attn.o_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.7.mlp.gate_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.7.mlp.up_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.7.mlp.down_proj.weight 201 shape=torch.Size([2048, 5632]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.7.input_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.7.post_attention_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.8.self_attn.q_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.8.self_attn.k_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.8.self_attn.v_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.8.self_attn.o_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.8.mlp.gate_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.8.mlp.up_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.8.mlp.down_proj.weight 201 shape=torch.Size([2048, 5632]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.8.input_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.8.post_attention_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.9.self_attn.q_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.9.self_attn.k_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.9.self_attn.v_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.9.self_attn.o_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.9.mlp.gate_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.9.mlp.up_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.9.mlp.down_proj.weight 201 shape=torch.Size([2048, 5632]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.9.input_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.9.post_attention_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.10.self_attn.q_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.10.self_attn.k_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.10.self_attn.v_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.10.self_attn.o_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.10.mlp.gate_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.10.mlp.up_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.10.mlp.down_proj.weight 201 shape=torch.Size([2048, 5632]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.10.input_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.10.post_attention_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.11.self_attn.q_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.11.self_attn.k_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.11.self_attn.v_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.11.self_attn.o_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.11.mlp.gate_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.11.mlp.up_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.11.mlp.down_proj.weight 201 shape=torch.Size([2048, 5632]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.11.input_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.11.post_attention_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.12.self_attn.q_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.12.self_attn.k_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.12.self_attn.v_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.12.self_attn.o_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.12.mlp.gate_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.12.mlp.up_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.12.mlp.down_proj.weight 201 shape=torch.Size([2048, 5632]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.12.input_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.12.post_attention_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.13.self_attn.q_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.13.self_attn.k_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.13.self_attn.v_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.13.self_attn.o_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.13.mlp.gate_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.13.mlp.up_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.13.mlp.down_proj.weight 201 shape=torch.Size([2048, 5632]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.13.input_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.13.post_attention_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.14.self_attn.q_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.14.self_attn.k_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.14.self_attn.v_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.14.self_attn.o_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.14.mlp.gate_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.14.mlp.up_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.14.mlp.down_proj.weight 201 shape=torch.Size([2048, 5632]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.14.input_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.14.post_attention_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.15.self_attn.q_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.15.self_attn.k_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.15.self_attn.v_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.15.self_attn.o_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.15.mlp.gate_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.15.mlp.up_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.15.mlp.down_proj.weight 201 shape=torch.Size([2048, 5632]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.15.input_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.15.post_attention_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.16.self_attn.q_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.16.self_attn.k_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.16.self_attn.v_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.16.self_attn.o_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.16.mlp.gate_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.16.mlp.up_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.16.mlp.down_proj.weight 201 shape=torch.Size([2048, 5632]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.16.input_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.16.post_attention_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.17.self_attn.q_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.17.self_attn.k_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.17.self_attn.v_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.17.self_attn.o_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.17.mlp.gate_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.17.mlp.up_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.17.mlp.down_proj.weight 201 shape=torch.Size([2048, 5632]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.17.input_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.17.post_attention_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.18.self_attn.q_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.18.self_attn.k_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.18.self_attn.v_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.18.self_attn.o_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.18.mlp.gate_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.18.mlp.up_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.18.mlp.down_proj.weight 201 shape=torch.Size([2048, 5632]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.18.input_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.18.post_attention_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.19.self_attn.q_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.19.self_attn.k_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.19.self_attn.v_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.19.self_attn.o_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.19.mlp.gate_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.19.mlp.up_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.19.mlp.down_proj.weight 201 shape=torch.Size([2048, 5632]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.19.input_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.19.post_attention_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.20.self_attn.q_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.20.self_attn.k_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.20.self_attn.v_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.20.self_attn.o_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.20.mlp.gate_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.20.mlp.up_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.20.mlp.down_proj.weight 201 shape=torch.Size([2048, 5632]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.20.input_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.20.post_attention_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.21.self_attn.q_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.21.self_attn.k_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.21.self_attn.v_proj.weight 201 shape=torch.Size([256, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.21.self_attn.o_proj.weight 201 shape=torch.Size([2048, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.21.mlp.gate_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.21.mlp.up_proj.weight 201 shape=torch.Size([5632, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.21.mlp.down_proj.weight 201 shape=torch.Size([2048, 5632]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.21.input_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.layers.21.post_attention_layernorm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=model.norm.weight 201 shape=torch.Size([2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:model=lm_head.weight 201 shape=torch.Size([32000, 2048]) dtype=torch.float16 cpu ratio=0.5 contig=True norm=0.68
INFO:__main__:done merge saving to file: matlok/tinyllama-cinder-openhermes-32k
config.json: 100%|█████████████████████████████████████| 724/724 [00:00<00:00, 7.75MB/s]
model.safetensors: 100%|███████████████████████████| 2.20G/2.20G [00:23<00:00, 91.8MB/s]
generation_config.json: 100%|██████████████████████████| 133/133 [00:00<00:00, 1.58MB/s]
INFO:__main__:loading newly-created file: matlok/tinyllama-cinder-openhermes-32k
INFO:__main__:loaded new model file: matlok/tinyllama-cinder-openhermes-32k asking question: why is the sky blue?
INFO:__main__:loading tokenizer=TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T
tokenizer_config.json: 100%|███████████████████████████| 776/776 [00:00<00:00, 8.26MB/s]
tokenizer.model: 100%|███████████████████████████████| 500k/500k [00:00<00:00, 64.6MB/s]
tokenizer.json: 100%|██████████████████████████████| 1.84M/1.84M [00:01<00:00, 1.57MB/s]
special_tokens_map.json: 100%|█████████████████████████| 414/414 [00:00<00:00, 2.47MB/s]
Setting `pad_token_id` to `eos_token_id`:2 for open-end generation.
INFO:__main__:
----------
tokenizer=LlamaTokenizerFast(name_or_path='TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T', vocab_size=32000, model_max_length=1000000000000000019884624838656, is_fast=True, padding_side='right', truncation_side='right', special_tokens={'bos_token': '<s>', 'eos_token': '</s>', 'unk_token': '<unk>'}, clean_up_tokenization_spaces=False),  added_tokens_decoder={
	0: AddedToken("<unk>", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),
	1: AddedToken("<s>", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),
	2: AddedToken("</s>", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),
}
 question:
why is the sky blue?
answer:
why is the sky blue?
Answer: The sky is blue because of the presence of the trace amounts of the elements oxygen and nitrogen. These elements are present in the atmosphere in very small amounts. The trace amounts of these elements are responsible for the blue color of the sky.
Why is the sky blue?
Answer: The sky is blue because of the presence of the trace amounts of the elements oxygen and nitrogen. These elements are present in the atmosphere in very small amounts. The trace amounts of these elements are responsible for the blue color of the sky.
Why is the sky blue?
Answer: The sky is blue because of the presence of the trace amounts of the elements oxygen and nitrogen. These elements are present in the atmosphere in very small amounts. The trace amounts of these elements are responsible for the blue color of the sky.
Why is the sky blue?
Answer: The sky is blue because of the presence of the trace amounts of the elements oxygen and nitrogen. These elements are present in the atmosphere in very small amounts. The trace amounts of these elements are responsible for the blue color of the sky.
Why is the sky blue?
Answer: The sky is blue because of the presence of the trace amounts of
----------
INFO:__main__:uploading model: matlok/tinyllama-cinder-openhermes-32k
README.md: 100%|████████████████████████████████████| 45.6k/45.6k [00:00<00:00, 297MB/s]
model.safetensors: 100%|███████████████████████████| 2.20G/2.20G [01:18<00:00, 28.0MB/s]
INFO:__main__:uploading src tokenizer: TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T
INFO:__main__:done loading new model: LlamaForCausalLM(
  (model): LlamaModel(
    (embed_tokens): Embedding(32000, 2048)
    (layers): ModuleList(
      (0-21): 22 x LlamaDecoderLayer(
        (self_attn): LlamaSdpaAttention(
          (q_proj): Linear(in_features=2048, out_features=2048, bias=False)
          (k_proj): Linear(in_features=2048, out_features=256, bias=False)
          (v_proj): Linear(in_features=2048, out_features=256, bias=False)
          (o_proj): Linear(in_features=2048, out_features=2048, bias=False)
          (rotary_emb): LlamaRotaryEmbedding()
        )
        (mlp): LlamaMLP(
          (gate_proj): Linear(in_features=2048, out_features=5632, bias=False)
          (up_proj): Linear(in_features=2048, out_features=5632, bias=False)
          (down_proj): Linear(in_features=5632, out_features=2048, bias=False)
          (act_fn): SiLU()
        )
        (input_layernorm): LlamaRMSNorm()
        (post_attention_layernorm): LlamaRMSNorm()
      )
    )
    (norm): LlamaRMSNorm()
  )
  (lm_head): Linear(in_features=2048, out_features=32000, bias=False)
) file: matlok/tinyllama-cinder-openhermes-32k

real 4m44.626s
user 2m54.434s
sys	 0m25.981s

Acknowlegdements

• Code sample above was modified from this very helpful GitHub gist
• Fine tuning example
• CodeLlama example