Efficiency: Emphasis on reducing memory usage and accelerating download speeds through 4bit quantization.
Adaptability: Auto detection of data types and support for advanced configuration options like RoPE scaling, LoRA, and gradient checkpointing.

Uploaded Model

Max Sequence Length: Trained at 16384 (via RoPE Scaling)
Data Type: Auto detection, with options for Float16 and Bfloat16
Quantization: 4bit, to reduce memory usage

Developed by: ndebuhr
License: apache-2.0
Finetuned from model : unsloth/gemma-2-27b-it-bnb-4bit

Configuration and Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
import torch

input_text = ""

# Set device based on CUDA availability
device = "cuda" if torch.cuda.is_available() else "cpu"

# Load the model and tokenizer
model_name = "ndebuhr/Gemma-2-27B-Technical-Tutorial-Summarization-QLoRA"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name).to(device)

instruction = "Clarify and summarize this tutorial transcript"
prompt = """{}

### Raw Transcript:
{}

### Summary:
"""

# Tokenize the input text
inputs = tokenizer(
    prompt.format(instruction, input_text),
    return_tensors="pt",
    truncation=True,
    max_length=16384
).to(device)

# Generate outputs
outputs = model.generate(
    **inputs,
    max_length=16384,
    num_return_sequences=1,
    use_cache=True
)

# Decode the generated text
generated_text = tokenizer.batch_decode(outputs, skip_special_tokens=True)

Compute Infrastructure

Fine-tuning: used 1xA100 (40GB)
Inference: recommend 1xL4 (24GB)

This gemma2 model was trained 2x faster with Unsloth and Huggingface's TRL library.