Create app.py

app.py

@@ -0,0 +1,672 @@
+--- START OF FILE app.py ---
+
+import os
+import shutil
+import subprocess
+import torch
+from transformers import AutoConfig, AutoModelForCausalLM
+from huggingface_hub import HfApi, whoami, ModelCard, list_models
+from gradio_huggingfacehub_search import HuggingfaceHubSearch
+from apscheduler.schedulers.background import BackgroundScheduler
+from textwrap import dedent
+import gradio as gr
+import hashlib
+import torch.nn.utils.prune as prune
+import torch.nn.functional as F
+from torch.utils.checkpoint import checkpoint
+import logging
+from datetime import datetime
+from typing import List, Dict
+
+logging.basicConfig(level=logging.INFO)
+
+os.environ["GRADIO_ANALYTICS_ENABLED"] = "False"
+HF_TOKEN = os.environ.get("HF_TOKEN")
+SPACE_ID = "Ffftdtd5dtft/gguf-my-repo" # Replace with your space ID if different
+
+def generate_importance_matrix(model_path, train_data_path):
+    os.chdir("llama.cpp")
+    if not os.path.isfile(f"../{model_path}"):
+        raise Exception(f"Model file not found: {model_path}")
+    imatrix_command = f"./llama-imatrix -m ../{model_path} -f {train_data_path} -ngl 99 --output-frequency 10"
+    process = subprocess.Popen(imatrix_command, shell=True)
+    try:
+        process.wait(timeout=3600)
+    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 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 = 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()
+    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)
+    return model
+
+def aggressive_optimize(model, reduce_layers_factor=0.5):
+    if hasattr(model.config, 'num_attention_heads'):
+        model.config.num_attention_heads = int(model.config.num_attention_heads * reduce_layers_factor)
+    if hasattr(model.config, 'hidden_size'):
+        model.config.hidden_size = int(model.config.hidden_size * reduce_layers_factor)
+    return model
+
+def apply_quantization(model, use_int8_inference):
+    if use_int8_inference:
+        quantized_model = torch.quantization.quantize_dynamic(
+            model, {torch.nn.Linear}, dtype=torch.qint8
+        )
+        return quantized_model
+    else:
+        return model
+
+def reduce_layers(model, reduction_factor=0.5):
+    if hasattr(model, 'transformer') and hasattr(model.transformer, 'h'):
+        original_num_layers = len(model.transformer.h)
+        new_num_layers = int(original_num_layers * reduction_factor)
+        model.transformer.h = torch.nn.ModuleList(model.transformer.h[:new_num_layers])
+    return model
+
+def use_smaller_embeddings(model, reduction_factor=0.75):
+    if hasattr(model, 'config'):
+        original_embedding_dim = model.config.hidden_size
+        new_embedding_dim = int(original_embedding_dim * reduction_factor)
+        model.config.hidden_size = new_embedding_dim
+        if hasattr(model, 'resize_token_embeddings'):
+            model.resize_token_embeddings(int(model.config.vocab_size * reduction_factor))
+    return model
+
+def use_fp16_embeddings(model):
+    if hasattr(model, 'transformer') and hasattr(model.transformer, 'wte'):
+        model.transformer.wte = model.transformer.wte.half()
+    return model
+
+def quantize_embeddings(model):
+    if hasattr(model, 'transformer') and hasattr(model.transformer, 'wte'):
+        model.transformer.wte = torch.quantization.quantize_dynamic(
+            model.transformer.wte, {torch.nn.Embedding}, dtype=torch.qint8
+        )
+    return model
+
+def use_bnb_f16(model):
+    if torch.cuda.is_available() and torch.cuda.is_bf16_supported():
+        model = model.to(dtype=torch.bfloat16)
+    return model
+
+def use_group_quantization(model):
+    for module in model.modules():
+        if isinstance(module, torch.nn.Linear):
+            torch.quantization.fuse_modules(module, ['weight'], inplace=True)
+            torch.quantization.quantize_dynamic(module, {torch.nn.Linear}, dtype=torch.qint8, inplace=True)
+    return model
+
+def apply_layer_norm_trick(model):
+    for name, module in model.named_modules():
+        if isinstance(module, torch.nn.LayerNorm):
+            module.elementwise_affine = False
+    return model
+
+def remove_padding(inputs, attention_mask):
+    last_non_padded = attention_mask.sum(dim=1) - 1
+    gathered_inputs = torch.gather(inputs, dim=1, index=last_non_padded.unsqueeze(1).unsqueeze(2).expand(-1, -1, inputs.size(2)))
+    return gathered_inputs
+
+def use_selective_quantization(model):
+    for module in model.modules():
+        if isinstance(module, torch.nn.MultiheadAttention):
+            torch.quantization.quantize_dynamic(module, {torch.nn.Linear}, dtype=torch.qint8, inplace=True)
+    return model
+
+def use_mixed_precision(model):
+    if hasattr(model, 'transformer') and hasattr(model.transformer, 'wte'):
+        model.transformer.wte = model.transformer.wte.half()
+    return model
+
+def use_pruning_after_training(model, prune_amount=0.1):
+    from torch import nn as nn
+    for name, module in model.named_modules():
+        if isinstance(module, (nn.Linear, nn.Conv2d)):
+            prune.l1_unstructured(module, name='weight', amount=prune_amount)
+            prune.remove(module, 'weight')
+    return model
+
+def use_knowledge_distillation(model, teacher_model, temperature=2.0, alpha=0.5):
+    teacher_model.eval()
+    criterion = torch.nn.KLDivLoss(reduction='batchmean')
+
+    def distillation_loss(student_logits, teacher_logits):
+        student_probs = F.log_softmax(student_logits / temperature, dim=-1)
+        teacher_probs = F.softmax(teacher_logits / temperature, dim=-1)
+        return criterion(student_probs, teacher_probs) * (temperature**2)
+
+    def train_step(inputs, labels):
+        student_outputs = model(**inputs, labels=labels)
+        student_logits = student_outputs.logits
+        with torch.no_grad():
+            teacher_outputs = teacher_model(**inputs)
+            teacher_logits = teacher_outputs.logits
+        loss = alpha * student_outputs.loss + (1 - alpha) * distillation_loss(student_logits, teacher_logits)
+        return loss
+
+    return train_step
+
+def use_weight_sharing(model):
+    if hasattr(model, 'transformer') and hasattr(model.transformer, 'h'):
+        if len(model.transformer.h) > 1:
+            model.transformer.h[-1].load_state_dict(model.transformer.h[0].state_dict())
+    return model
+
+def use_low_rank_approximation(model, rank_factor=0.5):
+    for module in model.modules():
+        if isinstance(module, torch.nn.Linear):
+            original_weight = module.weight.data
+            U, S, V = torch.linalg.svd(original_weight)
+            rank = int(S.size(0) * rank_factor)
+            module.weight.data = U[:, :rank] @ torch.diag(S[:rank]) @ V[:rank, :]
+    return model
+
+def use_hashing_trick(model, num_hashes=1024):
+    def hash_features(features):
+        features_bytes = features.cpu().numpy().tobytes()
+        hash_object = hashlib.sha256(features_bytes)
+        hash_value = hash_object.hexdigest()
+        hashed_features = int(hash_value, 16) % num_hashes
+        return torch.tensor(hashed_features, device=features.device)
+
+    original_forward = model.forward
+
+    def forward(*args, **kwargs):
+        inputs = args[0]
+        hashed_inputs = hash_features(inputs)
+        return original_forward(hashed_inputs, *args[1:], **kwargs)
+
+def use_quantization_aware_training(model):
+    model.qconfig = torch.quantization.get_default_qat_qconfig('fbgemm')
+    torch.quantization.prepare_qat(model, inplace=True)
+    torch.quantization.convert(model, inplace=True)
+    return model
+
+def use_gradient_checkpointing(model):
+    def custom_forward(*inputs):
+        return checkpoint(model, *inputs)
+    model.forward = custom_forward
+    return model
+
+def use_channel_pruning(model, prune_amount=0.1):
+    from torch import nn as nn
+    for module in model.modules():
+        if isinstance(module, nn.Conv2d):
+            prune.ln_structured(module, name="weight", amount=prune_amount, n=2, dim=0)
+            prune.remove(module, 'weight')
+    return model
+
+def use_sparse_tensors(model, sparsity_threshold=0.01):
+    for name, param in model.named_parameters():
+        if param.dim() >= 2 and param.is_floating_point():
+            sparse_param = param.to_sparse()
+            sparse_param._values()[sparse_param._values().abs() < sparsity_threshold] = 0
+            param.data = sparse_param.to_dense()
+    return model
+
+def use_lora(model, r=8, lora_alpha=16, lora_dropout=0.05, target_modules=None):
+    from peft import LoraConfig, get_peft_model
+    config = LoraConfig(
+        r=r,
+        lora_alpha=lora_alpha,
+        lora_dropout=lora_dropout,
+        target_modules=target_modules if target_modules else ["q_proj", "v_proj"], # Example target modules
+        bias="none",
+        task_type="CAUSAL_LM"
+    )
+    model = get_peft_model(model, config)
+    return model
+
+def use_adalora(model, target_r=8, init_r=12, tmask_init=0.01, beta1=0.85, beta2=0.99, loha=False, **kwargs):
+    from peft import AdaLoraConfig, get_peft_model
+    config = AdaLoraConfig(
+        target_r=target_r,
+        init_r=init_r,
+        tmask_init=tmask_init,
+        beta1=beta1,
+        beta2=beta2,
+        loha=loha,
+        task_type="CAUSAL_LM",
+        **kwargs
+    )
+    model = get_peft_model(model, config)
+    return model
+
+def use_ia3(model, target_modules=None):
+    from peft import IA3Config, get_peft_model
+    config = IA3Config(
+        target_modules=target_modules if target_modules else ["k_proj", "v_proj", "down_proj"], # Example target modules
+        feedforward_modules=None,
+        task_type="CAUSAL_LM"
+    )
+    model = get_peft_model(model, config)
+    return model
+
+def use_prompt_tuning(model, num_virtual_tokens=8, prompt_tuning_init_text="You are a helpful assistant."):
+    from peft import PromptTuningConfig, get_peft_model, TaskType
+    config = PromptTuningConfig(
+        task_type=TaskType.CAUSAL_LM,
+        num_virtual_tokens=num_virtual_tokens,
+        prompt_tuning_init_text=prompt_tuning_init_text,
+        tokenizer_name_or_path=model.config.tokenizer_class if hasattr(model.config, 'tokenizer_class') else None
+    )
+    model = get_peft_model(model, config)
+    return model
+
+def apply_moe_layer_splitting(model, num_experts: int = 4, expert_capacity_factor: float = 2.0, moe_layer_freq: int = 2):
+    # Assumes a standard transformer block structure
+    if not hasattr(model, 'transformer') or not hasattr(model.transformer, 'h'):
+        logging.warning("Model does not have the expected transformer structure for MoE splitting.")
+        return model
+
+    from transformers.models.mixtral.modeling_mixtral import MixtralSparseMoeBlock, MixtralBLock
+
+    for i in range(len(model.transformer.h)):
+        if (i + 1) % moe_layer_freq == 0:
+            original_layer = model.transformer.h[i]
+            # Extract necessary components, handling different layer structures
+            if isinstance(original_layer, MixtralBLock):
+                config = original_layer.config
+                new_moe_block = MixtralSparseMoeBlock(config, num_experts=num_experts, capacity_factor=expert_capacity_factor)
+                # Copy relevant weights - this might need adjustments based on the model
+                new_moe_block.load_state_dict(original_layer.mlp.state_dict(), strict=False)
+                model.transformer.h[i] = new_moe_block
+            else:
+                logging.warning(f"Skipping layer {i} for MoE, not a recognized block type.")
+    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, apply_aggressive_optimization, apply_reduce_layers, apply_smaller_embeddings,
+                 apply_weight_sharing, apply_low_rank_approx, use_lora_opt, use_adalora_opt, use_ia3_opt, use_prompt_tuning_opt,
+                 apply_moe_splitting, num_experts_moe, expert_capacity_factor_moe, moe_layer_freq_moe,
+                 is_automated=False):
+    if oauth_token.token is None and not is_automated:
+        raise ValueError("You must be logged in to use GGUF-my-repo")
+    elif oauth_token.token is None and is_automated:
+        logging.warning("Running in automated mode without user authentication.")
+
+    model_name = model_id.split('/')[-1]
+    fp16 = f"{model_name}.fp16.gguf"
+
+    try:
+        api = HfApi(token=oauth_token.token if oauth_token else None)
+        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)
+
+        config = AutoConfig.from_pretrained(model_id, trust_remote_code=True)
+        model = AutoModelForCausalLM.from_pretrained(model_id, config=config, torch_dtype=torch.float16, trust_remote_code=True)
+
+        if apply_aggressive_optimization:
+            model = aggressive_optimize(model)
+        if apply_reduce_layers:
+            model = reduce_layers(model)
+        if apply_smaller_embeddings:
+            model = use_smaller_embeddings(model)
+        if apply_weight_sharing:
+            model = use_weight_sharing(model)
+        if apply_low_rank_approx:
+            model = use_low_rank_approximation(model)
+        if use_lora_opt:
+            model = use_lora(model)
+        if use_adalora_opt:
+            model = use_adalora(model)
+        if use_ia3_opt:
+            model = use_ia3(model)
+        if use_prompt_tuning_opt:
+            model = use_prompt_tuning(model)
+        if apply_moe_splitting:
+            model = apply_moe_layer_splitting(model, num_experts_moe, expert_capacity_factor_moe, moe_layer_freq_moe)
+
+        optimized_model_path = f"{model_name}_optimized"
+        model.save_pretrained(optimized_model_path)
+
+        conversion_script = "convert_hf_to_gguf.py"
+        fp16_conversion = f"python llama.cpp/{conversion_script} {optimized_model_path} --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"] if oauth_token and oauth_token.token else "automated-gguf"
+        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}")
+
+        try:
+            subprocess.run(["llama.cpp/llama", "-m", quantized_gguf_path, "-p", "Test prompt"], check=True)
+        except Exception as e:
+            raise Exception(f"Model verification failed: {e}")
+
+        new_repo_id = f"{username}/{model_name}-{imatrix_q_method if use_imatrix else q_method}-GGUF"
+        new_repo_url = api.create_repo(repo_id=new_repo_id, exist_ok=True, private=private_repo)
+
+        try:
+            card = ModelCard.load(model_id, token=oauth_token.token if oauth_token else None)
+        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
+        optimization_notes = []
+        if apply_aggressive_optimization:
+            optimization_notes.append("Aggressive optimization applied.")
+        if apply_reduce_layers:
+            optimization_notes.append("Number of layers reduced.")
+        if apply_smaller_embeddings:
+            optimization_notes.append("Embedding size reduced.")
+        if apply_weight_sharing:
+            optimization_notes.append("Weight sharing applied.")
+        if apply_low_rank_approx:
+            optimization_notes.append(f"Low-rank approximation applied.")
+        if use_lora_opt:
+            optimization_notes.append("LoRA applied.")
+        if use_adalora_opt:
+            optimization_notes.append("AdaLoRA applied.")
+        if use_ia3_opt:
+            optimization_notes.append("IA3 applied.")
+        if use_prompt_tuning_opt:
+            optimization_notes.append("Prompt Tuning applied.")
+        if apply_moe_splitting:
+            optimization_notes.append(f"Mixture-of-Experts (MoE) layer splitting applied with {num_experts_moe} experts every {moe_layer_freq_moe} layers.")
+
+        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.
+
+            {' '.join(optimization_notes)}
+
+            ## 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)
+
+        log_message = f"Successfully processed and uploaded GGUF model for {model_id} to {new_repo_url}"
+        logging.info(log_message)
+        return (f'Find your repo <a href=\'{new_repo_url}\' target="_blank" style="text-decoration:underline">here</a>', "llama.png")
+    except Exception as e:
+        error_message = f"Error processing model {model_id}: {e}"
+        logging.error(error_message)
+        return (f"Error: {e}", "error.png")
+    finally:
+        shutil.rmtree(model_name, ignore_errors=True)
+        shutil.rmtree(optimized_model_path, ignore_errors=True)
+
+def select_models_for_automation():
+    # Example logic: Select top N most downloaded models
+    models = list_models(sort="downloads", direction=-1, limit=5)
+    return [model.modelId for model in models]
+
+def get_automation_parameters():
+    # Example logic: Define default parameters or load from a config
+    return {
+        "q_method": "Q4_K_M",
+        "use_imatrix": False,
+        "imatrix_q_method": "IQ4_NL",
+        "private_repo": True,
+        "train_data_file": None,
+        "split_model": False,
+        "split_max_tensors": 256,
+        "split_max_size": None,
+        "apply_aggressive_optimization": True,
+        "apply_reduce_layers": True,
+        "apply_smaller_embeddings": True,
+        "apply_weight_sharing": False,
+        "apply_low_rank_approx": False,
+        "use_lora_opt": False,
+        "use_adalora_opt": False,
+        "use_ia3_opt": False,
+        "use_prompt_tuning_opt": False,
+        "apply_moe_splitting": False,
+        "num_experts_moe": 4,
+        "expert_capacity_factor_moe": 2.0,
+        "moe_layer_freq_moe": 2,
+    }
+
+def automate_gguf_creation():
+    logging.info(f"Starting automated GGUF creation at {datetime.now()}")
+    api = HfApi(token=HF_TOKEN)
+    try:
+        whoami(token=HF_TOKEN) # Check if the token is valid
+    except Exception as e:
+        logging.error(f"Error with Hugging Face token: {e}")
+        return
+
+    models_to_process = select_models_for_automation()
+    automation_params = get_automation_parameters()
+
+    for model_id in models_to_process:
+        logging.info(f"Attempting to process model: {model_id}")
+        try:
+            process_model(model_id=model_id, oauth_token=None, is_automated=True, **automation_params)
+        except Exception as e:
+            logging.error(f"Failed to process model {model_id} automatically: {e}")
+
+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 for manual processing. Automation runs in the background.")
+    oauth_token = gr.OAuthButton(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(["IQ1", "IQ1_S", "IQ1_XXS", "IQ2_S", "IQ2_XXS", "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.")
+    train_data_file = gr.File(label="Training Data File", file_types=["txt"], visible=False)
+
+    size_reduction_accordion = gr.Accordion("Additional Size Reduction Techniques", open=False)
+    with size_reduction_accordion:
+        apply_aggressive_optimization = gr.Checkbox(value=True, label="Apply Aggressive Optimization", info="Reduces attention heads and hidden size.")
+        apply_reduce_layers = gr.Checkbox(value=True, label="Reduce Layers", info="Reduces the number of layers in the model.")
+        apply_smaller_embeddings = gr.Checkbox(value=True, label="Use Smaller Embeddings", info="Reduces the size of the embedding layer.")
+        apply_weight_sharing = gr.Checkbox(value=False, label="Apply Weight Sharing", info="Shares weights across layers to reduce parameters.")
+        apply_low_rank_approx = gr.Checkbox(value=False, label="Apply Low-Rank Approximation", info="Approximates weight matrices with lower rank.")
+        use_lora_opt = gr.Checkbox(value=False, label="Use LoRA", info="Applies Low-Rank Adaptation.")
+        use_adalora_opt = gr.Checkbox(value=False, label="Use AdaLoRA", info="Applies Adaptive Low-Rank Adaptation.")
+        use_ia3_opt = gr.Checkbox(value=False, label="Use IA3", info="Applies Infused Adapter by Inhibiting and Amplifying Inner Activations.")
+        use_prompt_tuning_opt = gr.Checkbox(value=False, label="Use Prompt Tuning", info="Adds trainable virtual tokens to the input embeddings.")
+        apply_moe_splitting = gr.Checkbox(value=False, label="Apply MoE Layer Splitting", info="Splits layers into a mixture-of-experts (MoE).", visible=False)
+        with gr.Row(visible=False) as moe_params:
+            num_experts_moe = gr.Number(value=4, label="Number of Experts", info="Number of experts to use in the MoE layers.", precision=0)
+            expert_capacity_factor_moe = gr.Number(value=2.0, label="Expert Capacity Factor", info="Capacity factor for each expert in the MoE layer.", precision=1)
+            moe_layer_freq_moe = gr.Number(value=2, label="MoE Layer Frequency", info="Apply MoE every N layers", precision=0)
+
+    private_repo = gr.Checkbox(value=True, label="Private Repo", info="Create a private repo under your username.")
+    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)
+    apply_moe_splitting.change(fn=lambda apply_moe_splitting: gr.update(visible=apply_moe_splitting), inputs=apply_moe_splitting, outputs=moe_params)
+
+
+    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,
+            oauth_token,
+            apply_aggressive_optimization,
+            apply_reduce_layers,
+            apply_smaller_embeddings,
+            apply_weight_sharing,
+            apply_low_rank_approx,
+            use_lora_opt,
+            use_adalora_opt,
+            use_ia3_opt,
+            use_prompt_tuning_opt,
+            apply_moe_splitting,
+            num_experts_moe,
+            expert_capacity_factor_moe,
+            moe_layer_freq_moe,
+        ],
+        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, applies size reduction techniques, quantizes it and creates a Public or Private repo containing the selected quant under your HF user namespace. It also automates the creation of GGUF quants for popular models in the background.",
+        api_name=False
+    )
+
+def restart_space():
+    HfApi().restart_space(repo_id=SPACE_ID, token=HF_TOKEN, factory_reboot=True)
+
+scheduler = BackgroundScheduler()
+scheduler.add_job(restart_space, "interval", seconds=21600)
+scheduler.add_job(automate_gguf_creation, "interval", hours=6) # Run automation every 6 hours
+scheduler.start()
+
+demo.queue(default_concurrency_limit=1, max_size=5).launch(debug=True, show_api=False)