Spaces:
Sleeping
Sleeping
File size: 7,468 Bytes
d03f587 61e7d9a d03f587 7cd0e22 d03f587 7cd0e22 d03f587 7cd0e22 d03f587 7cd0e22 d03f587 7cd0e22 d03f587 7cd0e22 61e7d9a 7cd0e22 d03f587 61e7d9a d03f587 61e7d9a d03f587 61e7d9a d03f587 61e7d9a d03f587 61e7d9a d03f587 7cd0e22 d03f587 61e7d9a d03f587 61e7d9a d03f587 |
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 |
# model_loader.py
import os
import sys
from transformers import (
AutoTokenizer,
AutoModelForCausalLM,
AutoConfig,
pipeline
)
import torch
import warnings
warnings.filterwarnings("ignore")
MODEL_NAME = "RayyanAhmed9477/med-coding"
def load_model_and_tokenizer():
"""
Loads Phi-3 model with multiple fallback strategies.
Handles safetensors loading issues with robust error recovery.
"""
device = "cuda" if torch.cuda.is_available() else "cpu"
print(f"🔧 Using device: {device}")
print(f"🔧 PyTorch version: {torch.__version__}")
print(f"🔧 Transformers version: {sys.modules['transformers'].__version__}")
# Get HuggingFace token from environment
hf_token = os.getenv("HF_TOKEN")
if hf_token:
print("🔑 HuggingFace token found")
else:
print("⚠️ No HuggingFace token - assuming public model")
try:
# ===== STEP 1: Load Tokenizer =====
print(f"📥 Loading tokenizer: {MODEL_NAME}")
tokenizer = AutoTokenizer.from_pretrained(
MODEL_NAME,
trust_remote_code=True,
token=hf_token,
use_fast=True
)
# Configure tokenizer
if tokenizer.pad_token is None:
tokenizer.pad_token = tokenizer.eos_token
if not hasattr(tokenizer, 'padding_side') or tokenizer.padding_side is None:
tokenizer.padding_side = "left"
print("✅ Tokenizer loaded successfully")
# ===== STEP 2: Load Configuration =====
print(f"📥 Loading model configuration: {MODEL_NAME}")
config = AutoConfig.from_pretrained(
MODEL_NAME,
trust_remote_code=True,
token=hf_token
)
# Handle LongRoPE configuration
if hasattr(config, 'rope_scaling') and config.rope_scaling is not None:
rope_type = config.rope_scaling.get('type', 'default')
print(f"📐 RoPE scaling type detected: {rope_type}")
if rope_type == 'longrope':
print("✅ LongRoPE configuration detected and supported")
print(f"✅ Config loaded: {config.model_type}")
# ===== STEP 3: Load Model with Multiple Strategies =====
print(f"📥 Loading model: {MODEL_NAME}")
print("⏳ This may take 2-5 minutes on first load...")
model = None
loading_strategies = []
if device == "cuda":
loading_strategies = [
# Strategy 1: Standard GPU loading
{
"name": "GPU Standard",
"params": {
"trust_remote_code": True,
"torch_dtype": torch.bfloat16,
"device_map": "auto",
"token": hf_token,
"low_cpu_mem_usage": True
}
}
]
else:
loading_strategies = [
# Strategy 1: CPU with safetensors (preferred)
{
"name": "CPU with safetensors",
"params": {
"trust_remote_code": True,
"torch_dtype": torch.float32,
"device_map": {"": "cpu"},
"token": hf_token,
"low_cpu_mem_usage": True,
"use_safetensors": True
}
},
# Strategy 2: CPU without explicit safetensors
{
"name": "CPU standard",
"params": {
"trust_remote_code": True,
"torch_dtype": torch.float32,
"token": hf_token,
"low_cpu_mem_usage": True
}
},
# Strategy 3: CPU with PyTorch weights fallback
{
"name": "CPU PyTorch weights",
"params": {
"trust_remote_code": True,
"torch_dtype": torch.float32,
"token": hf_token,
"low_cpu_mem_usage": True,
"use_safetensors": False
}
},
# Strategy 4: Minimal parameters
{
"name": "CPU minimal",
"params": {
"trust_remote_code": True,
"token": hf_token
}
}
]
# Try each loading strategy
for idx, strategy in enumerate(loading_strategies, 1):
try:
print(f"\n🔄 Attempt {idx}/{len(loading_strategies)}: {strategy['name']}")
model = AutoModelForCausalLM.from_pretrained(
MODEL_NAME,
config=config,
**strategy['params']
)
# Move to CPU explicitly if needed
if device == "cpu" and not strategy['params'].get('device_map'):
model = model.to("cpu")
print(f"✅ Model loaded successfully using: {strategy['name']}")
break
except Exception as e:
print(f"❌ Strategy '{strategy['name']}' failed: {str(e)}")
if idx == len(loading_strategies):
# All strategies failed
raise
else:
print(f"⏭️ Trying next strategy...")
continue
if model is None:
raise RuntimeError("All loading strategies failed")
# Set model to evaluation mode
model.eval()
# Disable gradients to save memory
for param in model.parameters():
param.requires_grad = False
print("\n✅ Model fully loaded and ready!")
# ===== STEP 4: Create Pipeline =====
print("🔧 Creating text generation pipeline...")
gen_pipeline = pipeline(
"text-generation",
model=model,
tokenizer=tokenizer,
device=0 if device == "cuda" else -1,
torch_dtype=torch.bfloat16 if device == "cuda" else torch.float32
)
print("✅ Pipeline created successfully!")
print("=" * 60)
print("🎉 MODEL READY FOR INFERENCE")
print("=" * 60)
return gen_pipeline, tokenizer
except Exception as e:
print(f"\n❌ Error during model loading: {str(e)}")
print("\n🔍 Diagnostic Information:")
print(f" - Model: {MODEL_NAME}")
print(f" - Device: {device}")
print(f" - Token available: {hf_token is not None}")
import traceback
traceback.print_exc()
raise RuntimeError(
f"Failed to load model {MODEL_NAME}. "
"All loading strategies exhausted. "
"This could be due to: "
"1) Model file corruption during download, "
"2) Insufficient memory, "
"3) Model incompatibility. "
"Try upgrading Space to GPU or use a different model."
) from e
|