Upload 3 files

Replace the config files in your Llama-3 model with the ones in the ´llama-3-config-files´ dir.

extra-files/llama-3-config-files/config.json

@@ -0,0 +1,38 @@
+{
+    "architectures": [
+        "LlamaForCausalLM"
+    ],
+    "attention_bias": false,
+    "attention_dropout": 0.0,
+    "bos_token_id": 128000,
+    "eos_token_id": 128009,
+    "hidden_act": "silu",
+    "hidden_size": 4096,
+    "initializer_range": 0.02,
+    "intermediate_size": 14336,
+    "max_position_embeddings": 8192,
+    "model_type": "llama",
+    "num_attention_heads": 32,
+    "num_hidden_layers": 32,
+    "num_key_value_heads": 8,
+    "pretraining_tp": 1,
+    "rms_norm_eps": 1e-05,
+    "rope_scaling": null,
+    "rope_theta": 500000.0,
+    "tie_word_embeddings": false,
+    "torch_dtype": "bfloat16",
+    "transformers_version": "4.40.0.dev0",
+    "use_cache": true,
+    "vocab_size": 128256,
+    "quantization_config": {
+        "quant_method": "exl2",
+        "version": "0.0.18",
+        "bits": 5.0,
+        "head_bits": 6,
+        "calibration": {
+            "rows": 100,
+            "length": 2048,
+            "dataset": "(default)"
+        }
+    }
+}

extra-files/llama-3-config-files/generation_config.json

@@ -0,0 +1,6 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 128000,
+  "eos_token_id": 128009,
+  "transformers_version": "4.40.0.dev0"
+}

gguf-imat-llama-3.py

@@ -0,0 +1,186 @@
+import os
+import requests
+import zipfile
+import subprocess
+import shutil
+from huggingface_hub import snapshot_download
+
+# Clone or update the llama.cpp repository with --depth 1
+def clone_or_update_llama_cpp():
+    print("Preparing...")
+    base_dir = os.path.dirname(os.path.abspath(__file__))
+    os.chdir(base_dir)
+    if not os.path.exists("llama.cpp"):
+        subprocess.run(["git", "clone", "--depth", "1", "https://github.com/ggerganov/llama.cpp"])
+    else:
+        os.chdir("llama.cpp")
+        subprocess.run(["git", "pull"])
+    os.chdir(base_dir)
+    print("The 'llama.cpp' repository is ready.")
+
+# Download and extract the latest release of llama.cpp Windows binaries
+def download_llama_release():
+    base_dir = os.path.dirname(os.path.abspath(__file__))
+    dl_dir = os.path.join(base_dir, "bin", "dl")
+    if not os.path.exists(dl_dir):
+        os.makedirs(dl_dir)
+
+    os.chdir(dl_dir)
+    latest_release_url = "https://github.com/ggerganov/llama.cpp/releases/latest"
+    response = requests.get(latest_release_url)
+    if response.status_code == 200:
+        latest_release_tag = response.url.split("/")[-1]
+        download_url = f"https://github.com/ggerganov/llama.cpp/releases/download/{latest_release_tag}/llama-{latest_release_tag}-bin-win-cuda-cu12.2.0-x64.zip"
+        response = requests.get(download_url)
+        if response.status_code == 200:
+            with open(f"llama-{latest_release_tag}-bin-win-cuda-cu12.2.0-x64.zip", "wb") as f:
+                f.write(response.content)
+            with zipfile.ZipFile(f"llama-{latest_release_tag}-bin-win-cuda-cu12.2.0-x64.zip", "r") as zip_ref:
+                zip_ref.extractall(os.path.join(base_dir, "bin"))
+            print("Downloading latest 'llama.cpp' prebuilt Windows binaries...")
+            print("Download and extraction completed successfully.")
+            return latest_release_tag
+        else:
+            print("Failed to download the release file.")
+    else:
+        print("Failed to fetch the latest release information.")
+
+# Download and extract the Cuda .dll resources if they aren't present in the bin folder
+def download_cudart_if_necessary(latest_release_tag):
+    base_dir = os.path.dirname(os.path.abspath(__file__))
+    cudart_dl_dir = os.path.join(base_dir, "bin", "dl")
+    if not os.path.exists(cudart_dl_dir):
+        os.makedirs(cudart_dl_dir)
+
+    cudart_zip_file = os.path.join(cudart_dl_dir, "cudart-llama-bin-win-cu12.2.0-x64.zip")
+    cudart_extracted_files = ["cublas64_12.dll", "cublasLt64_12.dll", "cudart64_12.dll"]
+
+    # Check if all required files exist
+    if all(os.path.exists(os.path.join(base_dir, "bin", file)) for file in cudart_extracted_files):
+        print("Cuda resources already exist. Skipping download.")
+    else:
+        cudart_download_url = f"https://github.com/ggerganov/llama.cpp/releases/download/{latest_release_tag}/cudart-llama-bin-win-cu12.2.0-x64.zip"
+        response = requests.get(cudart_download_url)
+        if response.status_code == 200:
+            with open(cudart_zip_file, "wb") as f:
+                f.write(response.content)
+            with zipfile.ZipFile(cudart_zip_file, "r") as zip_ref:
+                zip_ref.extractall(os.path.join(base_dir, "bin"))
+            print("Preparing 'cuda' resources...")
+            print("Download and extraction of cudart completed successfully.")
+        else:
+            print("Failed to download the cudart release file.")
+
+# Ask for user input to download or fetch from cache the specified model repository if it doesn't exist
+def download_model_repo():
+    base_dir = os.path.dirname(os.path.abspath(__file__))
+    models_dir = os.path.join(base_dir, "models")
+    if not os.path.exists(models_dir):
+        os.makedirs(models_dir)
+
+    model_id = input("Enter the model ID to download (e.g., huggingface/transformers): ")
+    model_name = model_id.split("/")[-1]
+    model_dir = os.path.join(models_dir, model_name)
+
+    # Check if the model repository already exists
+    if os.path.exists(model_dir):
+        print("Model repository already exists. Using existing repository.")
+
+        # If the model already exists, prompt the user if they want to delete the model directory
+        delete_model_dir = input("Remove HF model folder after converting original model to GGUF? (yes/no) (default: no): ").strip().lower()
+
+        # Ask for the name of the imatrix.txt file
+        imatrix_file_name = input("Enter the name of the imatrix.txt file (default: imatrix.txt): ").strip() or "imatrix.txt"
+
+        # Convert the existing model to GGUF F16 format and generate imatrix.dat
+        convert_model_to_gguf_f16(base_dir, model_dir, model_name, delete_model_dir, imatrix_file_name)
+
+    else:
+        revision = input("Enter the revision (branch, tag, or commit) to download (default: main): ") or "main"
+
+        # Ask the user if they want to remove the HF model folder after conversion
+        delete_model_dir = input("Remove HF model folder after converting original model to GGUF? (yes/no) (default: no): ").strip().lower()
+
+        print("Downloading model repository...")
+        snapshot_download(repo_id=model_id, local_dir=model_dir, revision=revision)
+        print("Model repository downloaded successfully.")
+
+        # Ask for the name of the imatrix.txt file
+        imatrix_file_name = input("Enter the name of the imatrix.txt file (default: imatrix.txt): ").strip() or "imatrix.txt"
+
+        # Convert the downloaded model to GGUF F16 format and generate imatrix.dat
+        convert_model_to_gguf_f16(base_dir, model_dir, model_name, delete_model_dir, imatrix_file_name)
+
+# Convert the downloaded model to GGUF F16 format
+def convert_model_to_gguf_f16(base_dir, model_dir, model_name, delete_model_dir, imatrix_file_name):
+    convert_script = os.path.join(base_dir, "llama.cpp", "convert.py")
+    gguf_dir = os.path.join(base_dir, "models", f"{model_name}-GGUF")
+    gguf_model_path = os.path.join(gguf_dir, f"{model_name}-F16.gguf")
+
+    if not os.path.exists(gguf_dir):
+        os.makedirs(gguf_dir)
+
+    # Check if F16 file already exists
+    if not os.path.exists(gguf_model_path):
+        # Execute the conversion command
+        subprocess.run(["python", convert_script, model_dir, "--outfile", gguf_model_path, "--outtype", "f16", "--vocab-type", "bpe"])
+
+        # Delete the original model directory under conditions
+        if delete_model_dir == 'yes' or delete_model_dir == 'y':
+            shutil.rmtree(model_dir)
+            print(f"Original model directory '{model_dir}' deleted.")
+        else:
+            print(f"Original model directory '{model_dir}' was not deleted. You can remove it manually.")
+
+    # Generate imatrix.dat if it doesn't exist
+    imatrix_exe = os.path.join(base_dir, "bin", "imatrix.exe")
+    imatrix_output = os.path.join(gguf_dir, "imatrix.dat")
+    imatrix_txt = os.path.join(base_dir, "imatrix", imatrix_file_name)
+    if not os.path.exists(imatrix_output):
+        # Execute the imatrix command
+        subprocess.run([imatrix_exe, "-m", gguf_model_path, "-f", imatrix_txt, "-ngl", "8"], cwd=gguf_dir)
+        # Move the imatrix.dat file to the GGUF folder
+        if os.path.exists(os.path.join(gguf_dir, "imatrix.dat")):
+            shutil.move(os.path.join(gguf_dir, "imatrix.dat"), gguf_dir)
+            print("imatrix.dat generated successfully.")
+        else:
+            print("Failed to generate imatrix.dat file.")
+    else:
+        print("Skipping imatrix generation as imatrix.dat already exists.")
+
+    # Quantize the models
+    quantize_models(base_dir, model_name)
+
+# Quantize models with different options
+def quantize_models(base_dir, model_name):
+    gguf_dir = os.path.join(base_dir, "models", f"{model_name}-GGUF")
+    f16_gguf_path = os.path.join(gguf_dir, f"{model_name}-F16.gguf")
+
+    quantization_options = [
+        "IQ3_M", "IQ3_XXS",
+        "Q4_K_M", "Q4_K_S", "IQ4_NL", "IQ4_XS",
+        "Q5_K_M", "Q5_K_S",
+        "Q6_K",
+        "Q8_0"
+    ]
+
+    for quant_option in quantization_options:
+        quantized_gguf_name = f"{model_name}-{quant_option}-imat.gguf"
+        quantized_gguf_path = os.path.join(gguf_dir, quantized_gguf_name)
+        quantize_command = os.path.join(base_dir, "bin", "quantize.exe")
+        imatrix_path = os.path.join(gguf_dir, "imatrix.dat")
+
+        subprocess.run([quantize_command, "--imatrix", imatrix_path, 
+                        f16_gguf_path, quantized_gguf_path, quant_option], cwd=gguf_dir)
+        print(f"Model quantized with {quant_option} option.")
+
+# Main function - Steps
+def main():
+    clone_or_update_llama_cpp()
+    latest_release_tag = download_llama_release()
+    download_cudart_if_necessary(latest_release_tag)
+    download_model_repo()
+    print("Finished preparing resources.")
+
+if __name__ == "__main__":
+    main()