File size: 8,438 Bytes
a60c3fc |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 |
#!/usr/bin/env python3
"""
Model Download and Setup for FinEE v2.0
========================================
Downloads and prepares base models for fine-tuning:
- Llama 3.1 8B Instruct (Primary)
- Qwen2.5 7B Instruct (Backup)
Supports:
- MLX format for Apple Silicon
- PyTorch/Transformers format
- GGUF for llama.cpp
"""
import argparse
import os
import subprocess
import sys
from pathlib import Path
MODELS = {
"llama-3.1-8b": {
"hf_name": "meta-llama/Llama-3.1-8B-Instruct",
"mlx_name": "mlx-community/Meta-Llama-3.1-8B-Instruct-4bit",
"gguf_name": "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF",
"description": "Llama 3.1 8B Instruct - Best instruction-following",
"size": "8B",
"context": "128K",
},
"qwen2.5-7b": {
"hf_name": "Qwen/Qwen2.5-7B-Instruct",
"mlx_name": "mlx-community/Qwen2.5-7B-Instruct-4bit",
"gguf_name": "Qwen/Qwen2.5-7B-Instruct-GGUF",
"description": "Qwen 2.5 7B - Excellent multilingual support",
"size": "7B",
"context": "128K",
},
"mistral-7b": {
"hf_name": "mistralai/Mistral-7B-Instruct-v0.3",
"mlx_name": "mlx-community/Mistral-7B-Instruct-v0.3-4bit",
"gguf_name": "bartowski/Mistral-7B-Instruct-v0.3-GGUF",
"description": "Mistral 7B - Fast and efficient",
"size": "7B",
"context": "32K",
},
"phi-3-medium": {
"hf_name": "microsoft/Phi-3-medium-128k-instruct",
"mlx_name": "mlx-community/Phi-3-medium-128k-instruct-4bit",
"description": "Phi-3 Medium - Compact but powerful",
"size": "14B",
"context": "128K",
},
}
def download_mlx_model(model_key: str, output_dir: Path):
"""Download model in MLX format."""
model = MODELS[model_key]
mlx_name = model.get("mlx_name")
if not mlx_name:
print(f"❌ No MLX version available for {model_key}")
return False
print(f"\n📥 Downloading {model_key} (MLX format)...")
print(f" From: {mlx_name}")
output_path = output_dir / model_key / "mlx"
output_path.mkdir(parents=True, exist_ok=True)
try:
from huggingface_hub import snapshot_download
snapshot_download(
repo_id=mlx_name,
local_dir=str(output_path),
local_dir_use_symlinks=False,
)
print(f"✅ Downloaded to: {output_path}")
return True
except Exception as e:
print(f"❌ Download failed: {e}")
return False
def download_hf_model(model_key: str, output_dir: Path):
"""Download model in HuggingFace format."""
model = MODELS[model_key]
hf_name = model["hf_name"]
print(f"\n📥 Downloading {model_key} (HuggingFace format)...")
print(f" From: {hf_name}")
output_path = output_dir / model_key / "hf"
output_path.mkdir(parents=True, exist_ok=True)
try:
from huggingface_hub import snapshot_download
snapshot_download(
repo_id=hf_name,
local_dir=str(output_path),
local_dir_use_symlinks=False,
ignore_patterns=["*.bin", "*.h5"], # Prefer safetensors
)
print(f"✅ Downloaded to: {output_path}")
return True
except Exception as e:
print(f"❌ Download failed: {e}")
print(" Note: Some models require HuggingFace login")
print(" Run: huggingface-cli login")
return False
def download_gguf_model(model_key: str, output_dir: Path, quant: str = "Q4_K_M"):
"""Download GGUF quantized model."""
model = MODELS[model_key]
gguf_name = model.get("gguf_name")
if not gguf_name:
print(f"❌ No GGUF version available for {model_key}")
return False
print(f"\n📥 Downloading {model_key} (GGUF {quant} format)...")
print(f" From: {gguf_name}")
output_path = output_dir / model_key / "gguf"
output_path.mkdir(parents=True, exist_ok=True)
try:
from huggingface_hub import hf_hub_download
# Find the right quantization file
filename = f"*{quant}*.gguf"
hf_hub_download(
repo_id=gguf_name,
filename=filename,
local_dir=str(output_path),
local_dir_use_symlinks=False,
)
print(f"✅ Downloaded to: {output_path}")
return True
except Exception as e:
print(f"❌ Download failed: {e}")
return False
def convert_to_mlx(model_path: Path, output_path: Path, quantize: bool = True):
"""Convert HuggingFace model to MLX format."""
print(f"\n🔄 Converting to MLX format...")
cmd = [
sys.executable, "-m", "mlx_lm.convert",
"--hf-path", str(model_path),
"--mlx-path", str(output_path),
]
if quantize:
cmd.extend(["--quantize", "--q-bits", "4"])
try:
subprocess.run(cmd, check=True)
print(f"✅ Converted to: {output_path}")
return True
except subprocess.CalledProcessError as e:
print(f"❌ Conversion failed: {e}")
return False
def verify_model(model_path: Path, backend: str = "mlx"):
"""Verify model can be loaded."""
print(f"\n🔍 Verifying model at {model_path}...")
if backend == "mlx":
try:
from mlx_lm import load, generate
model, tokenizer = load(str(model_path))
# Quick test
output = generate(model, tokenizer, "Hello", max_tokens=10)
print(f"✅ Model loaded successfully!")
print(f" Test output: {output[:50]}...")
return True
except Exception as e:
print(f"❌ Verification failed: {e}")
return False
elif backend == "transformers":
try:
from transformers import AutoModelForCausalLM, AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained(str(model_path))
model = AutoModelForCausalLM.from_pretrained(str(model_path))
print(f"✅ Model loaded successfully!")
return True
except Exception as e:
print(f"❌ Verification failed: {e}")
return False
def list_models():
"""List available models."""
print("\n📋 Available Models:\n")
print(f"{'Model':<20} {'Size':<8} {'Context':<10} {'Description'}")
print("-" * 80)
for key, model in MODELS.items():
print(f"{key:<20} {model['size']:<8} {model['context']:<10} {model['description']}")
def main():
parser = argparse.ArgumentParser(description="Download and setup base models")
parser.add_argument("action", choices=["download", "convert", "verify", "list"],
help="Action to perform")
parser.add_argument("-m", "--model", choices=list(MODELS.keys()),
default="llama-3.1-8b", help="Model to download")
parser.add_argument("-f", "--format", choices=["mlx", "hf", "gguf", "all"],
default="mlx", help="Model format")
parser.add_argument("-o", "--output", default="models/base",
help="Output directory")
parser.add_argument("-q", "--quant", default="Q4_K_M",
help="GGUF quantization level")
args = parser.parse_args()
output_dir = Path(args.output)
if args.action == "list":
list_models()
return
if args.action == "download":
if args.format in ["mlx", "all"]:
download_mlx_model(args.model, output_dir)
if args.format in ["hf", "all"]:
download_hf_model(args.model, output_dir)
if args.format in ["gguf", "all"]:
download_gguf_model(args.model, output_dir, args.quant)
elif args.action == "convert":
hf_path = output_dir / args.model / "hf"
mlx_path = output_dir / args.model / "mlx-converted"
convert_to_mlx(hf_path, mlx_path)
elif args.action == "verify":
model_path = output_dir / args.model
if args.format == "mlx":
model_path = model_path / "mlx"
elif args.format == "hf":
model_path = model_path / "hf"
verify_model(model_path, args.format)
if __name__ == "__main__":
main()
|