Llama-3.2-1B-Instruct-int8-int4

Uploaded model

• Developed by: metascroy
• License: apache-2.0
• Finetuned from model : unsloth/Llama-3.2-1B-Instruct

This llama model was trained 2x faster with Unsloth and Huggingface's TRL library.

<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>

Script

# =========================================================================================
#  Fine-tuning script based on https://colab.research.google.com/github/unslothai/notebooks/blob/main/nb/Llama3.2_%281B_and_3B%29-Conversational.ipynb
#
# This script and HF checkpoint are only intended to showcase how to do finetuning in a way compatible with ExecuTorch
# Only 100 steps are done, and quality of the finetuned model is not evaluated
# =========================================================================================

from unsloth import FastLanguageModel
from unsloth.chat_templates import (
    get_chat_template,
    standardize_data_formats,
    standardize_sharegpt,
    train_on_responses_only,
)

from datasets import load_dataset
from trl import SFTConfig, SFTTrainer
from transformers import DataCollatorForSeq2Seq
import torch
import torch.nn as nn

batch_size = 2
learning_rate = 2e-5
gradient_accumulation_steps = 4
max_steps = 100
full_finetuning = True
qat_scheme = "int8-int4"
output_dir = "/tmp/unsloth_example"


model_id = "unsloth/Llama-3.2-1B-Instruct"
chat_template = "llama-3.1"
max_seq_length = 2048
dtype = torch.bfloat16
load_in_4bit = False

################################################################################
# Define model/tokenizer
################################################################################

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name=model_id,
    max_seq_length=max_seq_length,
    dtype=dtype,
    load_in_4bit =load_in_4bit,
    full_finetuning=full_finetuning,
    qat_scheme=qat_scheme,
)
tokenizer = get_chat_template(tokenizer, chat_template = chat_template)
data_collator = DataCollatorForSeq2Seq(tokenizer = tokenizer)

print("MODEL AFTER LOADING")
print(model)

################################################################################
# Untie model weights
################################################################################

def untie_word_embeddings_(model):
    """Untie input and output embeddings in a Hugging Face causal LM."""
    # 1) Persist setting in config
    if hasattr(model.config, "tie_word_embeddings"):
        model.config.tie_word_embeddings = False

    # 2) Find input and output embeddings
    in_emb = model.get_input_embeddings()         # nn.Embedding
    out_proj = model.get_output_embeddings() or getattr(model, "lm_head", None)
    if out_proj is None:
        raise AttributeError("Couldn't locate output projection (lm_head).")

    # (Optional) sanity: shapes should match [vocab, hidden]
    assert out_proj.weight.shape == in_emb.weight.shape, (
        f"Shape mismatch: out_proj {out_proj.weight.shape} vs in_emb {in_emb.weight.shape}"
    )

    # 3) Only clone if they are actually tied (shared storage)
    if out_proj.weight.data_ptr() == in_emb.weight.data_ptr():
        with torch.no_grad():
            W = in_emb.weight.detach().clone()
        out_proj.weight = nn.Parameter(W)  # new storage, keeps dtype/device

    # 4) Prevent future automatic re-tying
    def _no_tie(self):
        return
    model.tie_weights = _no_tie.__get__(model, model.__class__)

    # 5) Verify no shared storage
    assert out_proj.weight.data_ptr() != in_emb.weight.data_ptr(), "Embeddings still tied!"

    return model

model = untie_word_embeddings_(model)

print("MODEL AFTER UNTYING")
print(model)
print(model.config)


################################################################################
# Process dataset
################################################################################

def formatting_prompts_func(examples):
    convos = examples["conversations"]
    texts = [tokenizer.apply_chat_template(convo, tokenize = False, add_generation_prompt = False) for convo in convos]
    return { "text" : texts, }
dataset = load_dataset("mlabonne/FineTome-100k", split = "train")
dataset = standardize_sharegpt(dataset)
dataset = dataset.map(formatting_prompts_func, batched = True,)

print("DATASET ENTRY")
print(dataset[0])
print("\n\n")

################################################################################
# Define trainer
################################################################################

trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    train_dataset=dataset,
    dataset_text_field="text",
    max_seq_length=max_seq_length,
    data_collator=DataCollatorForSeq2Seq(tokenizer=tokenizer),
    packing=False,
    args=SFTConfig(
        per_device_train_batch_size=batch_size,
        gradient_accumulation_steps=gradient_accumulation_steps,
        warmup_steps=5,
        num_train_epochs=1,
        max_steps=max_steps,
        learning_rate=learning_rate,
        logging_steps=1,
        optim="adamw_8bit",
        weight_decay=0.01,
        lr_scheduler_type="linear",
        seed=3407,
        output_dir="outputs",
        report_to="none",
    ),
)
trainer = train_on_responses_only(
    trainer,
    instruction_part = "<|start_header_id|>user<|end_header_id|>\n\n",
    response_part = "<|start_header_id|>assistant<|end_header_id|>\n\n",
)

print("VERIFYING PROMPT MASKING ON EXAMPLE")
idx = 5
print("Original: ", tokenizer.decode(trainer.train_dataset[idx]["input_ids"]))
space = tokenizer(" ", add_special_tokens = False).input_ids[0]
print("Masked: ", tokenizer.decode([space if x == -100 else x for x in trainer.train_dataset[idx]["labels"]]))
print("\n\n")


################################################################################
# Do fine tuning
################################################################################
print("DOING FINETUNING")
trainer_stats = trainer.train()
print(f"{trainer_stats.metrics['train_runtime']} seconds used for training.")
print(
    f"{round(trainer_stats.metrics['train_runtime']/60, 2)} minutes used for training."
)

################################################################################
# Save model
################################################################################
model.save_pretrained(output_dir)
tokenizer.save_pretrained(output_dir)

################################################################################
# Convert model
################################################################################
from torchao.quantization import Int8DynamicActivationIntxWeightConfig, IntxWeightOnlyConfig, ModuleFqnToConfig, quantize_
from torchao.quantization.qat import QATConfig
from torchao.quantization.granularity import PerGroup, PerAxis
from transformers import TorchAoConfig

base_config = Int8DynamicActivationIntxWeightConfig(weight_dtype=torch.int4, weight_granularity=PerGroup(32))
quantize_(model, QATConfig(base_config, step="convert"))
 
################################################################################
# Quantize embeddings to 8-bit with PTQ since they are not supported by QAT yet
################################################################################

embedding_fqn = "model.embed_tokens"
embedding_config = IntxWeightOnlyConfig(weight_dtype=torch.int8, granularity=PerAxis(0))
quantize_(model, embedding_config, lambda m, fqn: fqn == embedding_fqn)

################################################################################
# Attach quantization config to model
################################################################################

quant_config = ModuleFqnToConfig({"_default": base_config, embedding_fqn: embedding_config})
quantization_config = TorchAoConfig(quant_type=quant_config, include_input_output_embeddings=True, modules_to_not_convert=[])
model.config.quantization_config = TorchAoConfig(base_config)

print('MODEL AFTER CONVERT', model)

################################################################################
# Push converted model to hub
################################################################################
from huggingface_hub import get_token, whoami

def _get_username():
    token = get_token()
    username = whoami(token=token)["name"]
    return username

username = _get_username()
model_name = model_id.split("/")[-1]
save_to = f"{username}/{model_name}-{qat_scheme}"
model.push_to_hub(save_to, safe_serialization=False)
tokenizer.push_to_hub(save_to)

################################################################################
# Load converted from hub and inspect
################################################################################
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained(save_to)
print('model', model)
print("model.embed_tokens.weight", model.model.embed_tokens.weight)
print("model.layers[0].self_attn.q_proj.weight", model.model.layers[0].self_attn.q_proj.weight)
print("lm_head.weight", model.lm_head.weight)