app.py

import os
import shutil
import subprocess
import signal
import time
import torch
from torch.nn.utils import prune
from transformers import GPT2LMHeadModel, AutoTokenizer, AutoModelForCausalLM, DistilBertModel
from huggingface_hub import create_repo, HfApi, snapshot_download, whoami, ModelCard
from gradio_huggingfacehub_search import HuggingfaceHubSearch
from apscheduler.schedulers.background import BackgroundScheduler
from textwrap import dedent
import gradio as gr

os.environ["GRADIO_ANALYTICS_ENABLED"] = "False"
HF_TOKEN = os.environ.get("HF_TOKEN")

def generate_importance_matrix(model_path, train_data_path):
    imatrix_command = f"./llama-imatrix -m ../{model_path} -f {train_data_path} -ngl 99 --output-frequency 10"
    os.chdir("llama.cpp")
    if not os.path.isfile(f"../{model_path}"):
        raise Exception(f"Model file not found: {model_path}")
    process = subprocess.Popen(imatrix_command, shell=True)
    try:
        process.wait(timeout=60)
    except subprocess.TimeoutExpired:
        process.send_signal(signal.SIGINT)
        try:
            process.wait(timeout=5)
        except subprocess.TimeoutExpired:
            process.kill()
    os.chdir("..")

def split_upload_model(model_path, repo_id, oauth_token, split_max_tensors=256, split_max_size=None):
    if oauth_token.token is None:
        raise ValueError("You have to be logged in.")
    split_cmd = f"llama.cpp/llama-gguf-split --split --split-max-tensors {split_max_tensors}"
    if split_max_size:
        split_cmd += f" --split-max-size {split_max_size}"
    split_cmd += f" {model_path} {model_path.split('.')[0]}"
    result = subprocess.run(split_cmd, shell=True, capture_output=True, text=True)
    if result.returncode != 0:
        raise Exception(f"Error splitting the model: {result.stderr}")
    sharded_model_files = [f for f in os.listdir('.') if f.startswith(model_path.split('.')[0])]
    if sharded_model_files:
        api = HfApi(token=oauth_token.token)
        for file in sharded_model_files:
            file_path = os.path.join('.', file)
            try:
                api.upload_file(path_or_fileobj=file_path, path_in_repo=file, repo_id=repo_id)
            except Exception as e:
                raise Exception(f"Error uploading file {file_path}: {e}")
    else:
        raise Exception("No sharded files found.")

def prune_model(model, amount=0.5):
    for name, module in model.named_modules():
        if isinstance(module, (torch.nn.Linear, torch.nn.Conv2d)):
            prune.l1_unstructured(module, name='weight', amount=amount)
            prune.remove(module, 'weight')
    return model

def quantize_to_q1_with_min(tensor, min_value=-1):
    tensor = torch.sign(tensor)
    tensor[tensor < min_value] = min_value
    return tensor

def quantize_model_to_q1_with_min(model, min_value=-1):
    for name, param in model.named_parameters():
        if param.dtype in [torch.float32, torch.float16]:
            with torch.no_grad():
                param.copy_(quantize_to_q1_with_min(param.data, min_value))

def disable_unnecessary_components(model):
    for name, module in model.named_modules():
        if isinstance(module, torch.nn.Dropout):
            module.p = 0.0
        elif isinstance(module, torch.nn.BatchNorm1d):
            module.eval()

def ultra_max_compress(model):
    model = prune_model(model, amount=0.8)
    quantize_model_to_q1_with_min(model, min_value=-0.05)
    disable_unnecessary_components(model)
    with torch.no_grad():
        for name, param in model.named_parameters():
            if param.requires_grad:
                param.requires_grad = False
                param.data = torch.nn.functional.hardtanh(param.data, min_val=-1.0, max_val=1.0)
                param.data = param.data.half()
    try:
        model = torch.jit.script(model)
    except Exception:
        pass
    prune_model(model, amount=0.9)
    model.eval()
    for buffer_name, buffer in model.named_buffers():
        if buffer.numel() == 0:
            model._buffers.pop(buffer_name)
    return model

def optimize_model_resources(model):
    torch.set_grad_enabled(False)
    model.eval()
    for name, param in model.named_parameters():
        param.requires_grad = False
        if param.dtype == torch.float32:
            param.data = param.data.half()
    if hasattr(model, 'config'):
        if hasattr(model.config, 'max_position_embeddings'):
            model.config.max_position_embeddings = min(model.config.max_position_embeddings, 512)
        if hasattr(model.config, 'hidden_size'):
            model.config.hidden_size = min(model.config.hidden_size, 768)
    model = torch.jit.optimize_for_inference(model)
    return model

def process_model(model_id, q_method, use_imatrix, imatrix_q_method, private_repo, train_data_file, split_model, split_max_tensors, split_max_size, oauth_token: gr.OAuthToken | None):
    if oauth_token.token is None:
        raise ValueError("You must be logged in to use GGUF-my-repo")
    model_name = model_id.split('/')[-1]
    fp16 = f"{model_name}.fp16.gguf"

    try:
        api = HfApi(token=oauth_token.token)
        dl_pattern = ["*.safetensors", "*.bin", "*.pt", "*.onnx", "*.h5", "*.tflite", "*.ckpt", "*.pb", "*.tar", "*.xml", "*.caffemodel", "*.md", "*.json", "*.model"]
        pattern = "*.safetensors" if any(file.path.endswith(".safetensors") for file in api.list_repo_tree(repo_id=model_id, recursive=True)) else "*.bin"
        dl_pattern += pattern
        api.snapshot_download(repo_id=model_id, local_dir=model_name, local_dir_use_symlinks=False, allow_patterns=dl_pattern)
        conversion_script = "convert_hf_to_gguf.py"
        fp16_conversion = f"python llama.cpp/{conversion_script} {model_name} --outtype f16 --outfile {fp16}"
        result = subprocess.run(fp16_conversion, shell=True, capture_output=True)
        if result.returncode != 0:
            raise Exception(f"Error converting to fp16: {result.stderr}")
        imatrix_path = "llama.cpp/imatrix.dat"
        if use_imatrix:
            if train_data_file:
                train_data_path = train_data_file.name
            else:
                train_data_path = "groups_merged.txt"
            if not os.path.isfile(train_data_path):
                raise Exception(f"Training data file not found: {train_data_path}")
            generate_importance_matrix(fp16, train_data_path)
        username = whoami(oauth_token.token)["name"]
        quantized_gguf_name = f"{model_name.lower()}-{imatrix_q_method.lower()}-imat.gguf" if use_imatrix else f"{model_name.lower()}-{q_method.lower()}.gguf"
        quantized_gguf_path = quantized_gguf_name
        if use_imatrix:
            quantise_ggml = f"./llama.cpp/llama-quantize --imatrix {imatrix_path} {fp16} {quantized_gguf_path} {imatrix_q_method}"
        else:
            quantise_ggml = f"./llama.cpp/llama-quantize {fp16} {quantized_gguf_path} {q_method}"
        result = subprocess.run(quantise_ggml, shell=True, capture_output=True)
        if result.returncode != 0:
            raise Exception(f"Error quantizing: {result.stderr}")
        new_repo_url = api.create_repo(repo_id=f"{username}/{model_name}-{imatrix_q_method if use_imatrix else q_method}-GGUF", exist_ok=True, private=private_repo)
        new_repo_id = new_repo_url.repo_id
        try:
            card = ModelCard.load(model_id, token=oauth_token.token)
        except:
            card = ModelCard("")
        if card.data.tags is None:
            card.data.tags = []
        card.data.tags.append("llama-cpp")
        card.data.tags.append("gguf-my-repo")
        card.data.base_model = model_id
        card.text = dedent(
            f"""
            # {new_repo_id}
            This model was converted to GGUF format from [`{model_id}`](https://huggingface.co/{model_id}) using llama.cpp via the ggml.ai's [GGUF-my-repo](https://huggingface.co/spaces/ggml-org/gguf-my-repo) space.
            Refer to the [original model card](https://huggingface.co/{model_id}) for more details on the model.
            
            ## Use with llama.cpp
            Install llama.cpp through brew (works on Mac and Linux)
            
            ```bash
            brew install llama.cpp
            
            ```
            Invoke the llama.cpp server or the CLI.
            
            ### CLI:
            ```bash
            llama-cli --hf-repo {new_repo_id} --hf-file {quantized_gguf_name} -p "The meaning to life and the universe is"
            ```
            
            ### Server:
            ```bash
            llama-server --hf-repo {new_repo_id} --hf-file {quantized_gguf_name} -c 2048
            ```
            
            Note: You can also use this checkpoint directly through the [usage steps](https://github.com/ggerganov/llama.cpp?tab=readme-ov-file#usage) listed in the Llama.cpp repo as well.
            Step 1: Clone llama.cpp from GitHub.
            ```
            git clone https://github.com/ggerganov/llama.cpp
            ```
            Step 2: Move into the llama.cpp folder and build it with `LLAMA_CURL=1` flag along with other hardware-specific flags (for ex: LLAMA_CUDA=1 for Nvidia GPUs on Linux).
            ```
            cd llama.cpp && LLAMA_CURL=1 make
            ```
            Step 3: Run inference through the main binary.
            ```
            ./llama-cli --hf-repo {new_repo_id} --hf-file {quantized_gguf_name} -p "The meaning to life and the universe is"
            ```
            or 
            ```
            ./llama-server --hf-repo {new_repo_id} --hf-file {quantized_gguf_name} -c 2048
            ```
            """
        )
        card.save(f"README.md")

        if split_model:
            split_upload_model(quantized_gguf_path, new_repo_id, oauth_token, split_max_tensors, split_max_size)
        else:
            try:
                api.upload_file(path_or_fileobj=quantized_gguf_path, path_in_repo=quantized_gguf_name, repo_id=new_repo_id)
            except Exception as e:
                raise Exception(f"Error uploading quantized model: {e}")
        
        if os.path.isfile(imatrix_path):
            try:
                api.upload_file(path_or_fileobj=imatrix_path, path_in_repo="imatrix.dat", repo_id=new_repo_id)
            except Exception as e:
                raise Exception(f"Error uploading imatrix.dat: {e}")

        api.upload_file(path_or_fileobj=f"README.md", path_in_repo=f"README.md", repo_id=new_repo_id)
        return (f'Find your repo <a href=\'{new_repo_url}\' target="_blank" style="text-decoration:underline">here</a>', "llama.png")
    except Exception as e:
        return (f"Error: {e}", "error.png")
    finally:
        shutil.rmtree(model_name, ignore_errors=True)

css="""/* Custom CSS to allow scrolling */ .gradio-container {overflow-y: auto;}"""
with gr.Blocks(css=css) as demo: 
    gr.Markdown("You must be logged in to use GGUF-my-repo.")
    gr.LoginButton(min_width=250)
    model_id = HuggingfaceHubSearch(label="Hub Model ID", placeholder="Search for model id on Huggingface", search_type="model")
    q_method = gr.Dropdown(["Q2_K", "Q3_K_S", "Q3_K_M", "Q3_K_L", "Q4_0", "Q4_K_S", "Q4_K_M", "Q5_0", "Q5_K_S", "Q5_K_M", "Q6_K", "Q8_0"], label="Quantization Method", info="GGML quantization type", value="Q4_K_M", filterable=False, visible=True)
    imatrix_q_method = gr.Dropdown(["IQ3_M", "IQ3_XXS", "Q4_K_M", "Q4_K_S", "IQ4_NL", "IQ4_XS", "Q5_K_M", "Q5_K_S"], label="Imatrix Quantization Method", info="GGML imatrix quants type", value="IQ4_NL", filterable=False, visible=False)
    use_imatrix = gr.Checkbox(value=False, label="Use Imatrix Quantization", info="Use importance matrix for quantization.")
    private_repo = gr.Checkbox(value=False, label="Private Repo", info="Create a private repo under your username.")
    train_data_file = gr.File(label="Training Data File", file_types=["txt"], visible=False)
    split_model = gr.Checkbox(value=False, label="Split Model", info="Shard the model using gguf-split.")
    split_max_tensors = gr.Number(value=256, label="Max Tensors per File", info="Maximum number of tensors per file when splitting model.", visible=False)
    split_max_size = gr.Textbox(label="Max File Size", info="Maximum file size when splitting model (--split-max-size). May leave empty to use the default.", visible=False)

    use_imatrix.change(fn=lambda use_imatrix:  gr.update(visible=not use_imatrix), inputs=use_imatrix, outputs=q_method)
    use_imatrix.change(fn=lambda use_imatrix:  gr.update(visible=use_imatrix), inputs=use_imatrix, outputs=imatrix_q_method)
    use_imatrix.change(fn=lambda use_imatrix:  gr.update(visible=use_imatrix), inputs=use_imatrix, outputs=train_data_file)
    split_model.change(fn=lambda split_model:  gr.update(visible=split_model), inputs=split_model, outputs=split_max_tensors)
    split_model.change(fn=lambda split_model:  gr.update(visible=split_model), inputs=split_model, outputs=split_max_size)

    iface = gr.Interface(
        fn=process_model,
        inputs=[
            model_id,
            q_method,
            use_imatrix,
            imatrix_q_method,
            private_repo,
            train_data_file,
            split_model,
            split_max_tensors,
            split_max_size,
        ],
        outputs=[
            gr.Markdown(label="output"),
            gr.Image(show_label=False),
        ],
        title="Create your own GGUF Quants, blazingly fast ⚡!",
        description="The space takes an HF repo as an input, quantizes it and creates a Public repo containing the selected quant under your HF user namespace.",
        api_name=False
    )

def restart_space():
    HfApi().restart_space(repo_id="ggml-org/gguf-my-repo", token=HF_TOKEN, factory_reboot=True)

scheduler = BackgroundScheduler()
scheduler.add_job(restart_space, "interval", seconds=21600)
scheduler.start()

demo.queue(default_concurrency_limit=100, max_size=100).launch(debug=True, show_api=False)