_Create_Instance.py

## CREATE MODEL FROM SCRATCH

## TOBE REMOVED
# pip install reportlab

from transformers import AutoTokenizer, AutoModelForCausalLM, TextGenerationPipeline, AutoConfig, BitsAndBytesConfig,AutoConfig
import time
import torch
torch.backends.cuda.matmul.allow_tf32 = True
import random
from datasets import load_dataset
from transformers import TrainingArguments
from trl import SFTTrainer
from peft import LoraConfig
# from accelerate import infer_auto_device_map, init_empty_weights, dispatch_model
from torch.nn import CrossEntropyLoss
torch.autograd.set_detect_anomaly(True)
random_seed = 42
torch.manual_seed(random_seed)
random.seed(random_seed)
# Set the device for each process
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# torch.cuda.set_device(device)



n_ahead_talk_global = 4
n_passes_global = 2
n_ahead_global = 8
n_examples = 0

def model_init(params):
    original = False
    if params is None:
        params = {}
    else:
        params = params.params
    # save params to file
    n_ahead = params.get("n_ahead", n_ahead_global if not original else 1)
    n_ahead_talk = params.get("n_ahead_talk", n_ahead_talk_global if not original else 1)
    n_passes = params.get("n_passes", n_passes_global if not original else 1)
    gumbel_temperature = params.get("gumbel_temperature", 1)
    use_start_thought_token = params.get("use_start_thought_token", True)
    use_end_thought_token = params.get("use_end_thought_token", True)
    include_policy_loss = params.get("include_policy_loss", True)
    gumbel_detach = params.get("gumbel_detach", True)
    merged_talk_heads = params.get("merged_talk_heads", True)
    residual_think_head = params.get("residual_think_head", False)
    optimize_lm_head_only_at_start = params.get("optimize_lm_head_only_at_start", False)

    model_id = "LeroyDyer/_Spydaz_Web_AI_V2_Aligned"
    tokenizer_id = model_id
    print("Loading model")

    model = AutoModelForCausalLM.from_pretrained(
        model_id,
        torch_dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32,
        max_thoughts=n_ahead + n_ahead_talk + 1,
        merged_talk_heads=merged_talk_heads,
        merged_lm_and_talk_heads=False,
        merged_lm_and_think_heads=True,
        use_concat_talk_head=True,
        use_shallow_think=True,
        use_shallow_talk=False,
        use_complex_think_head=False,
        use_complex_talk_head=True,
        use_weighted_talk_head=True,
        trust_remote_code=True,
        device_map="auto",
    )
    print("Loaded model")

    tokenizer = AutoTokenizer.from_pretrained(tokenizer_id, truncation=True, padding_side="right")
    tokenizer.pad_token_id = tokenizer.eos_token_id

    special_tokens_to_add = []
    if model.use_start_thought_token:
        special_tokens_to_add.append("<|startthought|>")
    if model.use_end_thought_token:
        special_tokens_to_add.append("<|endthought|>")
    if special_tokens_to_add:
        tokenizer.add_special_tokens({"additional_special_tokens": special_tokens_to_add})
        model.resize_token_embeddings(len(tokenizer))
    model.tokenizer = tokenizer
    for name, module in model.named_modules():
        if "embed" in name:
            print(module, flush=True)

    model.gumbel_detach = gumbel_detach
    model.include_policy_loss = include_policy_loss
    model.use_end_thought_token = use_end_thought_token
    model.use_start_thought_token = use_start_thought_token
    model.n_ahead = n_ahead
    model.n_ahead_talk = n_ahead_talk
    model.n_passes = n_passes
    model.residual_think_head = residual_think_head
    model.optimize_lm_head_only_at_start = optimize_lm_head_only_at_start
    model.gumbel_temperature = gumbel_temperature
    model.original_mode = original
    model.config_params = params
    return model,tokenizer




model,tokenizer = model_init(None)

model
tokenizer.save_pretrained("IModel")
model.save_pretrained("IModel")

import os
import huggingface_hub
from huggingface_hub import notebook_login
from huggingface_hub import create_repo, HfApi
from huggingface_hub import hf_hub_download
from huggingface_hub import create_repo, HfApi
from huggingface_hub import snapshot_download
WRITE_TOKEN=""
username = "LeroyDyer"
huggingface_hub.login(WRITE_TOKEN)
api = HfApi(token=WRITE_TOKEN)

MODEL_NAME = "_Spydaz_Web_AI_MistralStar"
Folderinput = "IModel"


# Create empty repo
api.create_repo(
    repo_id = f"{username}/{MODEL_NAME}",
    repo_type="model",
    exist_ok=True,
)

api.upload_folder(
    repo_id = f"{username}/{MODEL_NAME}",
    folder_path = Folderinput
)




import huggingface_hub
from trl import SFTTrainer
from transformers import TrainingArguments
from datasets import load_dataset
from unsloth import FastLanguageModel
import torch
WRITE_TOKEN = ""
username = "LeroyDyer"
huggingface_hub.login(WRITE_TOKEN)

MODEL_ID = "LeroyDyer/_Spydaz_Web_AI_MistralStar"
max_seq_length = 1512 # Choose any! We auto support RoPE Scaling internally!
dtype = None # None for auto detection. Float16 for Tesla T4, V100, Bfloat16 for Ampere+
load_in_4bit = True # Use 4bit quantization to reduce memory usage. Can be False.

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = MODEL_ID, # Choose ANY! eg teknium/OpenHermes-2.5-Mistral-7B
    max_seq_length = max_seq_length,
    dtype = dtype,
    load_in_4bit = load_in_4bit,
    #token = "", # use one if using gated models like meta-llama/Llama-2-7b-hf
)
model = FastLanguageModel.get_peft_model(
    model,
    r = 32, # Choose any number > 0 ! Suggested 8, 16, 32, 64, 128
    target_modules = ["q_proj", "k_proj", "v_proj", "o_proj",
                      "gate_proj", "up_proj", "down_proj"],
    lora_alpha = 64,
    lora_dropout = 0, # Supports any, but = 0 is optimized
    bias = "none",    # Supports any, but = "none" is optimized
    # [NEW] "unsloth" uses 30% less VRAM, fits 2x larger batch sizes!
    use_gradient_checkpointing = "unsloth", # True or "unsloth" for very long context
    random_state = 644993,
    use_rslora = False,  # We support rank stabilized LoRA
    loftq_config = None, # And LoftQ
)



alpaca_prompt = """Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.

### Instruction:
{}

### Input:
{}

### Response:
{}"""


EOS_TOKEN = tokenizer.eos_token # Must add EOS_TOKEN
def formatting_prompts_func(examples):
    instructions = examples["instruction"]
    inputs       = examples["input"]
    outputs      = examples["output"]
    texts = []
    for instruction, input, output in zip(instructions, inputs, outputs):
        # Must add EOS_TOKEN, otherwise your generation will go on forever!
        text = alpaca_prompt.format(instruction, input, output) + EOS_TOKEN
        texts.append(text)
    return { "text" : texts, }
pass

from datasets import load_dataset
dataset = load_dataset("gate369/Alpaca-Star", split = "train[:1000]")
dataset = dataset.shuffle(seed=9969)
dataset = dataset.map(formatting_prompts_func, batched = True,)


from trl import SFTTrainer
from transformers import TrainingArguments
from unsloth import is_bfloat16_supported
from unsloth import UnslothTrainer, UnslothTrainingArguments

trainer = UnslothTrainer(
    model = model,
    tokenizer = tokenizer,
    train_dataset = dataset,
    dataset_text_field = "text",
    max_seq_length = max_seq_length,
    dataset_num_proc = 8,
    args = UnslothTrainingArguments(
        per_device_train_batch_size = 10,
        gradient_accumulation_steps = 8,

        warmup_ratio = 0.1,
        num_train_epochs = 2,

        learning_rate = 2e-4,
        embedding_learning_rate = 2e-5,
        output_dir = "outputs",
        save_strategy = "steps",
        save_steps = 50,
        fp16 = not is_bfloat16_supported(),
        bf16 = is_bfloat16_supported(),
        logging_steps = 1,
        optim = "adamw_8bit",
        weight_decay = 0.00,
        lr_scheduler_type = "cosine",
        seed = 3607,
    ),
)

trainer_stats = trainer.train()



# Merge to 16bit
if False: model.save_pretrained_merged("LCARS_AI_015", tokenizer, save_method = "merged_16bit",)
if True: model.push_to_hub_merged("_Spydaz_Web_AI_STAR_Aligned", tokenizer, save_method = "merged_16bit", token = "")

# Merge to 4bit
if False: model.save_pretrained_merged("model", tokenizer, save_method = "merged_4bit_forced",)
if True: model.push_to_hub_merged("_Spydaz_Web_AI_STAR_Aligned_4_BIT", tokenizer, save_method = "merged_4bit_forced", token = "")

# Just LoRA adapters
if False: model.save_pretrained_merged("model", tokenizer, save_method = "lora",)
if False: model.push_to_hub_merged("Test_Lora", tokenizer, save_method = "lora", token = "")