import os import subprocess import signal os.environ["GRADIO_ANALYTICS_ENABLED"] = "False" import gradio as gr import tempfile from huggingface_hub import HfApi, ModelCard, whoami from gradio_huggingfacehub_search import HuggingfaceHubSearch from pathlib import Path from textwrap import dedent from apscheduler.schedulers.background import BackgroundScheduler # used for restarting the space HF_TOKEN = os.environ.get("HF_TOKEN") CONVERSION_SCRIPT = "./llama.cpp/convert_hf_to_gguf.py" # escape HTML for logging def escape(s: str) -> str: s = s.replace("&", "&") # Must be done first! s = s.replace("<", "<") s = s.replace(">", ">") s = s.replace('"', """) s = s.replace("\n", "
") return s def generate_importance_matrix(model_path: str, train_data_path: str, output_path: str): imatrix_command = [ "./llama.cpp/llama-imatrix", "-m", model_path, "-f", train_data_path, "-ngl", "99", "--output-frequency", "10", "-o", output_path, ] if not os.path.isfile(model_path): raise Exception(f"Model file not found: {model_path}") print("Running imatrix command...") process = subprocess.Popen(imatrix_command, shell=False) try: process.wait(timeout=60) # added wait except subprocess.TimeoutExpired: print("Imatrix computation timed out. Sending SIGINT to allow graceful termination...") process.send_signal(signal.SIGINT) try: process.wait(timeout=5) # grace period except subprocess.TimeoutExpired: print("Imatrix proc still didn't term. Forecfully terming process...") process.kill() print("Importance matrix generation completed.") def split_upload_model(model_path: str, outdir: str, repo_id: str, oauth_token: gr.OAuthToken | None, split_max_tensors=256, split_max_size=None): print(f"Model path: {model_path}") print(f"Output dir: {outdir}") if oauth_token.token is None: raise ValueError("You have to be logged in.") split_cmd = [ "./llama.cpp/llama-gguf-split", "--split", ] if split_max_size: split_cmd.append("--split-max-size") split_cmd.append(split_max_size) else: split_cmd.append("--split-max-tensors") split_cmd.append(str(split_max_tensors)) # args for output model_path_prefix = '.'.join(model_path.split('.')[:-1]) # remove the file extension split_cmd.append(model_path) split_cmd.append(model_path_prefix) print(f"Split command: {split_cmd}") result = subprocess.run(split_cmd, shell=False, capture_output=True, text=True) print(f"Split command stdout: {result.stdout}") print(f"Split command stderr: {result.stderr}") if result.returncode != 0: stderr_str = result.stderr.decode("utf-8") raise Exception(f"Error splitting the model: {stderr_str}") print("Model split successfully!") # remove the original model file if needed if os.path.exists(model_path): os.remove(model_path) model_file_prefix = model_path_prefix.split('/')[-1] print(f"Model file name prefix: {model_file_prefix}") sharded_model_files = [f for f in os.listdir(outdir) if f.startswith(model_file_prefix) and f.endswith(".gguf")] if sharded_model_files: print(f"Sharded model files: {sharded_model_files}") api = HfApi(token=oauth_token.token) for file in sharded_model_files: file_path = os.path.join(outdir, file) print(f"Uploading file: {file_path}") 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.") print("Sharded model has been uploaded successfully!") 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 is None or oauth_token.token is None: raise ValueError("You must be logged in to use GGUF-my-repo") model_name = model_id.split('/')[-1] try: api = HfApi(token=oauth_token.token) dl_pattern = ["*.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] if not os.path.exists("downloads"): os.makedirs("downloads") if not os.path.exists("outputs"): os.makedirs("outputs") with tempfile.TemporaryDirectory(dir="outputs") as outdir: fp16 = str(Path(outdir)/f"{model_name}.fp16.gguf") with tempfile.TemporaryDirectory(dir="downloads") as tmpdir: # Keep the model name as the dirname so the model name metadata is populated correctly local_dir = Path(tmpdir)/model_name print(local_dir) api.snapshot_download(repo_id=model_id, local_dir=local_dir, local_dir_use_symlinks=False, allow_patterns=dl_pattern) print("Model downloaded successfully!") print(f"Current working directory: {os.getcwd()}") print(f"Model directory contents: {os.listdir(local_dir)}") config_dir = local_dir/"config.json" adapter_config_dir = local_dir/"adapter_config.json" if os.path.exists(adapter_config_dir) and not os.path.exists(config_dir): raise Exception('adapter_config.json is present.

If you are converting a LoRA adapter to GGUF, please use GGUF-my-lora.') result = subprocess.run([ "python", CONVERSION_SCRIPT, local_dir, "--outtype", "f16", "--outfile", fp16 ], shell=False, capture_output=True) print(result) if result.returncode != 0: stderr_str = result.stderr.decode("utf-8") raise Exception(f"Error converting to fp16: {stderr_str}") print("Model converted to fp16 successfully!") print(f"Converted model path: {fp16}") imatrix_path = Path(outdir)/"imatrix.dat" if use_imatrix: if train_data_file: train_data_path = train_data_file.name else: train_data_path = "llama.cpp/groups_merged.txt" #fallback calibration dataset print(f"Training data file path: {train_data_path}") 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, imatrix_path) else: print("Not using imatrix quantization.") # Quantize the model 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 = str(Path(outdir)/quantized_gguf_name) if use_imatrix: quantise_ggml = [ "./llama.cpp/llama-quantize", "--imatrix", imatrix_path, fp16, quantized_gguf_path, imatrix_q_method ] else: quantise_ggml = [ "./llama.cpp/llama-quantize", fp16, quantized_gguf_path, q_method ] result = subprocess.run(quantise_ggml, shell=False, capture_output=True) if result.returncode != 0: stderr_str = result.stderr.decode("utf-8") raise Exception(f"Error quantizing: {stderr_str}") print(f"Quantized successfully with {imatrix_q_method if use_imatrix else q_method} option!") print(f"Quantized model path: {quantized_gguf_path}") # Create empty repo username = whoami(oauth_token.token)["name"] new_repo_url = api.create_repo(repo_id=f"{username}/{model_name}-GGUF", exist_ok=True, private=private_repo) new_repo_id = new_repo_url.repo_id print("Repo created successfully!", new_repo_url) 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("matrixportal") 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 [all-gguf-same-where](https://huggingface.co/spaces/matrixportal/all-gguf-same-where) space. Refer to the [original model card](https://huggingface.co/{model_id}) for more details on the model. """ ) readme_path = Path(outdir)/"README.md" card.save(readme_path) # Quant listesi oluşturma quant_list = f""" ## ✅ Quantized Models Download List ### 🔍 Recommended Quantizations | 🚀 Download | 🔢 Type | 📝 Notes | |:---------|:-----|:------| | [Download](https://huggingface.co/{new_repo_id}/resolve/main/{model_name.lower()}-q2_k.gguf) | ![Q2_K](https://img.shields.io/badge/Q2_K-1A73E8) | Basic quantization | | [Download](https://huggingface.co/{new_repo_id}/resolve/main/{model_name.lower()}-q3_k_s.gguf) | ![Q3_K_S](https://img.shields.io/badge/Q3_K_S-34A853) | Small size | | [Download](https://huggingface.co/{new_repo_id}/resolve/main/{model_name.lower()}-q3_k_m.gguf) | ![Q3_K_M](https://img.shields.io/badge/Q3_K_M-FBBC05) | Balanced quality | | [Download](https://huggingface.co/{new_repo_id}/resolve/main/{model_name.lower()}-q3_k_l.gguf) | ![Q3_K_L](https://img.shields.io/badge/Q3_K_L-4285F4) | Better quality | | [Download](https://huggingface.co/{new_repo_id}/resolve/main/{model_name.lower()}-q4_0.gguf) | ![Q4_0](https://img.shields.io/badge/Q4_0-EA4335) | Fast on ARM | | [Download](https://huggingface.co/{new_repo_id}/resolve/main/{model_name.lower()}-q4_k_s.gguf) | ![Q4_K_S](https://img.shields.io/badge/Q4_K_S-673AB7) | Fast, recommended | | [Download](https://huggingface.co/{new_repo_id}/resolve/main/{model_name.lower()}-q4_k_m.gguf) | ![Q4_K_M](https://img.shields.io/badge/Q4_K_M-673AB7) ⭐ | Best balance | | [Download](https://huggingface.co/{new_repo_id}/resolve/main/{model_name.lower()}-q5_0.gguf) | ![Q5_0](https://img.shields.io/badge/Q5_0-FF6D01) | Good quality | | [Download](https://huggingface.co/{new_repo_id}/resolve/main/{model_name.lower()}-q5_k_s.gguf) | ![Q5_K_S](https://img.shields.io/badge/Q5_K_S-0F9D58) | Balanced | | [Download](https://huggingface.co/{new_repo_id}/resolve/main/{model_name.lower()}-q5_k_m.gguf) | ![Q5_K_M](https://img.shields.io/badge/Q5_K_M-0F9D58) | High quality | | [Download](https://huggingface.co/{new_repo_id}/resolve/main/{model_name.lower()}-q6_k.gguf) | ![Q6_K](https://img.shields.io/badge/Q6_K-4285F4) 🏆 | Very good quality | | [Download](https://huggingface.co/{new_repo_id}/resolve/main/{model_name.lower()}-q8_0.gguf) | ![Q8_0](https://img.shields.io/badge/Q8_0-EA4335) ⚡ | Fast, best quality | | [Download](https://huggingface.co/{new_repo_id}/resolve/main/{model_name.lower()}-f16.gguf) | ![F16](https://img.shields.io/badge/F16-000000) | Maximum accuracy | 💡 **Tip:** Use `F16` for maximum precision when quality is critical # GGUF Model Quantization & Usage Guide with llama.cpp ## What is GGUF and Quantization? **GGUF** (GPT-Generated Unified Format) is an efficient model file format developed by the `llama.cpp` team that: - Supports multiple quantization levels - Works cross-platform - Enables fast loading and inference **Quantization** converts model weights to lower precision data types (e.g., 4-bit integers instead of 32-bit floats) to: - Reduce model size - Decrease memory usage - Speed up inference - (With minor accuracy trade-offs) ## Step-by-Step Guide ### 1. Prerequisites ```bash # System updates sudo apt update && sudo apt upgrade -y # Dependencies sudo apt install -y build-essential cmake python3-pip # Clone and build llama.cpp git clone https://github.com/ggerganov/llama.cpp cd llama.cpp make -j4 ``` ### 2. Using Quantized Models from Hugging Face My automated quantization script produces models in this format: ``` https://huggingface.co/{new_repo_id}/resolve/main/{model_name.lower()}-q4_k_m.gguf ``` Download your quantized model directly: ```bash wget https://huggingface.co/{new_repo_id}/resolve/main/{model_name.lower()}-q4_k_m.gguf ``` ### 3. Running the Quantized Model Basic usage: ```bash ./main -m {model_name.lower()}-q4_k_m.gguf -p "Your prompt here" -n 128 ``` Example with a creative writing prompt: ```bash ./main -m {model_name.lower()}-q4_k_m.gguf \ -p "[INST] Write a short poem about AI quantization in the style of Shakespeare [/INST]" \ -n 256 -c 2048 -t 8 --temp 0.7 ``` Advanced parameters: ```bash ./main -m {model_name.lower()}-q4_k_m.gguf \ -p "Question: What is the GGUF format?\nAnswer:" \ -n 256 -c 2048 -t 8 --temp 0.7 --top-k 40 --top-p 0.9 ``` ### 4. Python Integration Install the Python package: ```bash pip install llama-cpp-python ``` Example script: ```python from llama_cpp import Llama # Initialize the model llm = Llama( model_path="{model_name.lower()}-q4_k_m.gguf", n_ctx=2048, n_threads=8 ) # Run inference response = llm( "[INST] Explain GGUF quantization to a beginner [/INST]", max_tokens=256, temperature=0.7, top_p=0.9 ) print(response["choices"][0]["text"]) ``` ## Performance Tips 1. **Hardware Utilization**: - Set thread count with `-t` (typically CPU core count) - Compile with CUDA/OpenCL for GPU support 2. **Memory Optimization**: - Lower quantization (like q4_k_m) uses less RAM - Adjust context size with `-c` parameter 3. **Speed/Accuracy Balance**: - Higher bit quantization is slower but more accurate - Reduce randomness with `--temp 0` for consistent results ## FAQ **Q: What quantization levels are available?** A: Common options include q4_0, q4_k_m, q5_0, q5_k_m, q8_0 (my script uses q4_k_m by default) **Q: How much performance loss occurs with q4_k_m?** A: Typically 2-5% accuracy reduction but 4x smaller size **Q: How to enable GPU support?** A: Build with `make LLAMA_CUBLAS=1` for NVIDIA GPUs ## Useful Resources 1. [llama.cpp GitHub](https://github.com/ggerganov/llama.cpp) 2. [GGUF Format Specs](https://github.com/ggerganov/ggml/blob/master/docs/gguf.md) 3. [Hugging Face Model Hub](https://huggingface.co/models) """ # README'yi güncelle (ModelCard kullanarak) card.text += quant_list readme_path = Path(outdir)/"README.md" card.save(readme_path) if split_model: split_upload_model(str(quantized_gguf_path), outdir, new_repo_id, oauth_token, split_max_tensors, split_max_size) else: try: print(f"Uploading quantized model: {quantized_gguf_path}") 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: print(f"Uploading imatrix.dat: {imatrix_path}") 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=readme_path, path_in_repo="README.md", repo_id=new_repo_id, ) print(f"Uploaded successfully with {imatrix_q_method if use_imatrix else q_method} option!") # end of the TemporaryDirectory(dir="outputs") block; temporary outputs are deleted here return ( f'

✅ DONE


Find your repo here: {new_repo_id}', "llama.png", ) except Exception as e: return (f'

❌ ERROR


{escape(str(e))}
', "error.png") css="""/* Custom CSS to allow scrolling */ .gradio-container {overflow-y: auto;} """ # Create Gradio interface 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", "F16"], 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_0", "Q4_K_M", "Q4_K_S", "IQ4_NL", "IQ4_XS", "Q5_K_M", "Q5_K_S", "Q6_K", "Q8_0", "F16"], 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. Accepted suffixes: M, G. Example: 256M, 5G", visible=False ) def update_visibility(use_imatrix): return gr.update(visible=not use_imatrix), gr.update(visible=use_imatrix), gr.update(visible=use_imatrix) use_imatrix.change( fn=update_visibility, inputs=use_imatrix, outputs=[q_method, imatrix_q_method, train_data_file] ) 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 update_split_visibility(split_model): return gr.update(visible=split_model), gr.update(visible=split_model) split_model.change( fn=update_split_visibility, inputs=split_model, outputs=[split_max_tensors, split_max_size] ) def restart_space(): HfApi().restart_space(repo_id="matrixportal/all-gguf-same-where", token=HF_TOKEN, factory_reboot=True) scheduler = BackgroundScheduler() scheduler.add_job(restart_space, "interval", seconds=21600) scheduler.start() # Launch the interface demo.queue(default_concurrency_limit=1, max_size=5).launch(debug=True, show_api=False)