michaelfeil / ct2fast-Phind-CodeLlama-34B-v2

# Fast-Inference with Ctranslate2

Speedup inference while reducing memory by 2x-4x using int8 inference in C++ on CPU or GPU.

quantized version of Phind/Phind-CodeLlama-34B-v2

pip install hf-hub-ctranslate2>=2.12.0 ctranslate2>=3.17.1

# from transformers import AutoTokenizer
model_name = "michaelfeil/ct2fast-Phind-CodeLlama-34B-v2"


from hf_hub_ctranslate2 import GeneratorCT2fromHfHub
model = GeneratorCT2fromHfHub(
        # load in int8 on CUDA
        model_name_or_path=model_name,
        device="cuda",
        compute_type="int8_float16",
        # tokenizer=AutoTokenizer.from_pretrained("{ORG}/{NAME}")
)
outputs = model.generate(
    text=["def fibonnaci(", "User: How are you doing? Bot:"],
    max_length=64,
    include_prompt_in_result=False
)
print(outputs)

Checkpoint compatible to ctranslate2>=3.17.1 and hf-hub-ctranslate2>=2.12.0

• compute_type=int8_float16 for device="cuda"
• compute_type=int8 for device="cpu"

Converted on 2023-10-08 using

LLama-2 -> removed <pad> token.

Licence and other remarks:

This is just a quantized version. Licence conditions are intended to be idential to original huggingface repo.

Original description

Phind-CodeLlama-34B-v2

We've fine-tuned Phind-CodeLlama-34B-v1 on an additional 1.5B tokens high-quality programming-related data, achieving 73.8% pass@1 on HumanEval. It's the current state-of-the-art amongst open-source models.

Furthermore, this model is instruction-tuned on the Alpaca/Vicuna format to be steerable and easy-to-use.

More details can be found on our blog post .

Model Details

This model is fine-tuned from Phind-CodeLlama-34B-v1 and achieves 73.8% pass@1 on HumanEval.

Phind-CodeLlama-34B-v2 is multi-lingual and is proficient in Python, C/C++, TypeScript, Java, and more.

Dataset Details

We fined-tuned on a proprietary dataset of 1.5B tokens of high quality programming problems and solutions. This dataset consists of instruction-answer pairs instead of code completion examples, making it structurally different from HumanEval. LoRA was not used -- both models are a native finetune. We used DeepSpeed ZeRO 3 and Flash Attention 2 to train these models in 15 hours on 32 A100-80GB GPUs. We used a sequence length of 4096 tokens.

How to Get Started with the Model

Make sure to install Transformers from the main git branch:

pip install git+https://github.com/huggingface/transformers.git

How to Prompt the Model

This model accepts the Alpaca/Vicuna instruction format.

For example:

### System Prompt
You are an intelligent programming assistant.

### User Message
Implement a linked list in C++

### Assistant
...

How to reproduce HumanEval Results

To reproduce our results:

from transformers import AutoTokenizer, LlamaForCausalLM
from human_eval.data import write_jsonl, read_problems
from tqdm import tqdm

# initialize the model

model_path = "Phind/Phind-CodeLlama-34B-v2"
model = LlamaForCausalLM.from_pretrained(model_path, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(model_path)

# HumanEval helper

def generate_one_completion(prompt: str):
    tokenizer.pad_token = tokenizer.eos_token
    inputs = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=4096)

    # Generate
    generate_ids = model.generate(inputs.input_ids.to("cuda"), max_new_tokens=384, do_sample=True, top_p=0.75, top_k=40, temperature=0.1)
    completion = tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
    completion = completion.replace(prompt, "").split("\n\n\n")[0]

    return completion

# perform HumanEval
problems = read_problems()

num_samples_per_task = 1
samples = [
    dict(task_id=task_id, completion=generate_one_completion(problems[task_id]["prompt"]))
    for task_id in tqdm(problems)
    for _ in range(num_samples_per_task)
]
write_jsonl("samples.jsonl", samples)

# run `evaluate_functional_correctness samples.jsonl` in your HumanEval code sandbox

Bias, Risks, and Limitations

This model has undergone very limited testing. Additional safety testing should be performed before any real-world deployments.

Training details

• Hardware Type: 32x A100-80GB
• Hours used: 480 GPU-hours
• Cloud Provider: AWS
• Compute Region: us-east-1