|
--- |
|
license: apache-2.0 |
|
language: |
|
- en |
|
tags: |
|
- bitnet |
|
datasets: |
|
- abideen/Cosmopedia-100k-pretrain |
|
--- |
|
|
|
|
|
# Bitnet-Nous-Llama3-225M 🚀 |
|
|
|
Este modelo es una variante optimizada del **Llama3** utilizando la arquitectura **BitNet**, lo que reduce los pesos a los valores `-1`, `0`, y `1` para mejorar la eficiencia en el cómputo sin perder precisión. |
|
|
|
 |
|
|
|
## Modelo Base 🦙 |
|
|
|
- **Modelo Original**: [Meta-Llama3-8B](https://huggingface.co./meta-llama/Meta-Llama-3-8B) |
|
- **Parámetros Reducidos**: 225M |
|
|
|
## Arquitectura 🔧 |
|
|
|
El modelo transforma las capas lineales de Llama3 en capas **BitLinear**, aprovechando las siguientes técnicas de cuantización: |
|
|
|
- **Cuantización de activaciones**: Escala a ±127 |
|
- **Cuantización de pesos**: Escala a ±1 |
|
|
|
### Especificaciones Técnicas 📋 |
|
|
|
- **Dimensiones**: 768 |
|
- **Capas**: 6 |
|
- **Contexto**: 256 tokens |
|
- **Tamaño intermedio**: 1024 |
|
- **Número de cabezas de atención**: 6 |
|
|
|
|
|
## Dataset 📚 |
|
|
|
El modelo fue entrenado usando el dataset [Cosmopedia-100k-pretrain](https://huggingface.co./datasets/abideen/Cosmopedia-100k-pretrain), que contiene una variedad de datos de texto. |
|
|
|
## Entrenamiento ⚙️ |
|
|
|
El modelo fue entrenado con la siguiente configuración: |
|
|
|
- **Lote**: 16 |
|
- **Tasa de aprendizaje**: 1.5e-4 |
|
- **Épocas**: 2 |
|
- **Acumulación de gradientes**: 2 pasos |
|
- **Decaimiento de pesos**: 0.01 |
|
- **Precisión Mixta**: FP16 |
|
|
|
### Monitoreo 📊 |
|
|
|
El proceso de entrenamiento fue monitoreado usando **Weights & Biases**. |
|
|
|
## Uso del Modelo 💻 |
|
|
|
Para usar este modelo, puedes cargarlo desde Hugging Face con el siguiente código: |
|
```python |
|
from transformers import AutoModelForCausalLM, AutoTokenizer |
|
from transformers.models.llama.modeling_llama import * |
|
import torch |
|
from torch import nn |
|
import torch.nn.functional as F |
|
import coloredlogs |
|
import logging |
|
|
|
from utils.utils import count_parameters |
|
|
|
coloredlogs.install(level='INFO', fmt='%(asctime)s - %(levelname)s - %(message)s', logger=logging.getLogger()) |
|
logger = logging.getLogger(__name__) |
|
|
|
|
|
|
|
|
|
HF_TOKEN = "tuclaveaqui" |
|
#model = "ejbejaranos/Bitnet-Llama3-from8BM-now2B" |
|
model = "ejbejaranos/Bitnet-Nous-Llama3-225M" ## Working |
|
|
|
# Load a pretrained BitNet model |
|
tokenizer = AutoTokenizer.from_pretrained(model) |
|
|
|
model = AutoModelForCausalLM.from_pretrained( |
|
model, |
|
token=HF_TOKEN |
|
) |
|
|
|
|
|
def count_parameters(model): |
|
# Calculate the number of parameters in billions |
|
num_params = sum(p.numel() for p in model.parameters() if p.requires_grad) / 10**9 |
|
print(f"Model size: {num_params:.3f}B parameters") |
|
return int(num_params) |
|
|
|
|
|
|
|
# Establece el pad_token_id |
|
model.config.pad_token_id = tokenizer.eos_token_id |
|
|
|
def activation_quant(x): |
|
scale = 127.0 / x.abs().max(dim=-1, keepdim=True).values.clamp_(min=1e-5) |
|
y = (x * scale).round().clamp_(-128, 127) |
|
y = y / scale |
|
return y |
|
|
|
def weight_quant(w): |
|
scale = 1.0 / w.abs().mean().clamp_(min=1e-5) |
|
u = (w * scale).round().clamp_(-1, 1) |
|
u = u / scale |
|
return u |
|
|
|
class BitLinear(nn.Linear): |
|
def forward(self, x): |
|
w = self.weight # a weight tensor with shape [d, k] |
|
x = x.to(w.device) |
|
RMSNorm = LlamaRMSNorm(x.shape[-1]).to(w.device) |
|
x_norm = RMSNorm(x) |
|
x_quant = x_norm + (activation_quant(x_norm) - x_norm).detach() |
|
w_quant = w + (weight_quant(w) - w).detach() |
|
y = F.linear(x_quant, w_quant) |
|
return y |
|
|
|
def convert_to_bitnet(model, copy_weights): |
|
for name, module in model.named_modules(): |
|
if isinstance(module, LlamaSdpaAttention) or isinstance(module, LlamaMLP): |
|
for child_name, child_module in module.named_children(): |
|
if isinstance(child_module, nn.Linear): |
|
bitlinear = BitLinear(child_module.in_features, child_module.out_features, child_module.bias is not None).to(device="cuda:0") |
|
if copy_weights: |
|
bitlinear.weight = child_module.weight |
|
if child_module.bias is not None: |
|
bitlinear.bias = child_module.bias |
|
setattr(module, child_name, bitlinear) |
|
elif isinstance(module, LlamaDecoderLayer): |
|
for child_name, child_module in module.named_children(): |
|
if isinstance(child_module, LlamaRMSNorm) and child_name == "input_layernorm": |
|
setattr(module, child_name, nn.Identity().to(device="cuda:0")) |
|
|
|
convert_to_bitnet(model, copy_weights=True) |
|
model.to(device="cuda:0") |
|
|
|
|
|
logger.info(f"🔢 Number of parameters in the model after extracting weights: {count_parameters(model)}") |
|
logger.info(f"📏 Reduced model structure:\n{model}") |
|
|
|
|
|
|
|
|
|
|
|
prompt = "What is Machine Learning?" |
|
inputs = tokenizer(prompt, return_tensors="pt", padding=True, truncation=True).to(model.device) |
|
inputs['attention_mask'] = inputs['input_ids'] != model.config.pad_token_id |
|
|
|
generate_ids = model.generate(inputs.input_ids, attention_mask=inputs['attention_mask'], max_length=250) |
|
decoded_output = tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False) |
|
|
|
print(decoded_output[0]) # Print the generated response |
|
|
|
``` |
