File size: 11,279 Bytes
3581a58 | 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 | import os
import logging
import torch
from pathlib import Path
from tqdm import tqdm
from peft import LoraConfig, get_peft_model, PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
from torch.utils.data import DataLoader, Dataset
os.environ["PYTORCH_MPS_HIGH_WATERMARK_RATIO"] = "0.0"
logger = logging.getLogger(__name__)
class SpiderDataset(Dataset):
"""Formats Spider JSON data into causal LM training examples (Question -> MongoDB Query)."""
def __init__(self, data, tokenizer, max_length=512):
self.data = data
self.tokenizer = tokenizer
self.max_length = max_length
def __len__(self):
return len(self.data)
def __getitem__(self, idx):
item = self.data[idx]
question = item.get("question", "")
mongodb_query = item.get("generated_mongodb_query", "")
# Handle if stored as a list
if isinstance(mongodb_query, list) and len(mongodb_query) > 0:
mongodb_query = mongodb_query[0]
prompt = f"### Task: Generate MongoDB Query (MQL)\n### Question:\n{question}\n\n### MongoDB Query:\n"
prompt_enc = self.tokenizer(prompt, truncation=True, max_length=self.max_length, return_tensors="pt")
completion_enc = self.tokenizer(mongodb_query, truncation=True, max_length=self.max_length, return_tensors="pt")
input_ids = torch.cat([prompt_enc.input_ids[0], completion_enc.input_ids[0]], dim=0)
attention_mask = torch.cat([prompt_enc.attention_mask[0], completion_enc.attention_mask[0]], dim=0)
prompt_len = prompt_enc.input_ids.shape[1]
labels = torch.cat([torch.full((prompt_len,), -100, dtype=torch.long), completion_enc.input_ids[0]], dim=0)
return {
"input_ids": input_ids,
"attention_mask": attention_mask,
"labels": labels
}
class ModelLoader:
def __init__(self, config, device_override=None):
self.config = config
self.device = device_override or self._detect_device()
if self.device == "cuda":
self.dtype = torch.float16
elif self.device == "mps":
self.dtype = torch.float16
else:
self.dtype = torch.float32
self.tokenizer = None
self.base_model = None
self.lora_model = None
logger.info(f"🖥️ Initialized ModelLoader | Device: {self.device} | Dtype: {self.dtype}")
def _detect_device(self):
if torch.cuda.is_available():
return "cuda"
if hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
return "mps"
return "cpu"
def load_tokenizer(self):
model_name = getattr(self.config, "model_name", "Qwen/Qwen2.5-Coder-0.5B-Instruct")
logger.info(f"🔄 Loading tokenizer: {model_name}")
self.tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
if self.tokenizer.pad_token is None:
self.tokenizer.pad_token = self.tokenizer.eos_token
return self.tokenizer
def load_base_model(self):
model_name = getattr(self.config, "model_name", "Qwen/Qwen2.5-Coder-0.5B-Instruct")
logger.info(f"🔄 Loading base model: {model_name}")
if self.device == "cuda":
self.base_model = AutoModelForCausalLM.from_pretrained(
model_name,
torch_dtype=self.dtype,
trust_remote_code=True,
low_cpu_mem_usage=True,
device_map="auto"
)
else:
self.base_model = AutoModelForCausalLM.from_pretrained(
model_name,
torch_dtype=self.dtype,
trust_remote_code=True,
low_cpu_mem_usage=True
)
self.base_model.to(self.device)
self.base_model.eval()
return self.base_model
def setup_lora(self, trainable_fraction=1.0, r=16, alpha=32, dropout=0.05):
if self.base_model is None:
self.load_base_model()
target_modules = ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"]
peft_config = LoraConfig(
r=r,
lora_alpha=alpha,
target_modules=target_modules,
lora_dropout=dropout,
bias="none",
task_type="CAUSAL_LM"
)
self.lora_model = get_peft_model(self.base_model, peft_config)
if trainable_fraction < 1.0:
self._apply_trainable_fraction(trainable_fraction)
else:
logger.info("✅ LoRA configured with 100% trainable parameters.")
self.lora_model.print_trainable_parameters()
return self.lora_model
def _apply_trainable_fraction(self, fraction):
lora_params = [(n, p) for n, p in self.lora_model.named_parameters() if "lora_" in n and p.requires_grad]
total_lora = len(lora_params)
num_to_freeze = int(total_lora * (1.0 - fraction))
if num_to_freeze > 0:
for i, (name, param) in enumerate(lora_params):
if i < num_to_freeze:
param.requires_grad = False
trainable_count = sum(1 for p in self.lora_model.parameters() if p.requires_grad)
logger.info(f"🔒 Froze {num_to_freeze}/{total_lora} LoRA parameters. Remaining trainable: {trainable_count} ({fraction*100}%).")
def autodetect_saved_model(self, output_dir):
output_dir = Path(output_dir)
if output_dir.exists() and (output_dir / "adapter_config.json").exists():
logger.info(f"✅ Found saved LoRA adapter at {output_dir}. Loading...")
if self.base_model is None:
self.load_base_model()
# NOTE (ZeroGPU compatibility): peft's from_pretrained() infers the load
# device via infer_device(), NOT from the base model's actual .device.
# On HF Spaces ZeroGPU, `spaces` monkey-patches torch.cuda.is_available()
# to report True even before a real GPU is attached to the process, so
# infer_device() always says "cuda" here -- which makes peft try to load
# the adapter's .safetensors weights with a raw, direct CUDA file load.
# That bypasses spaces' safe .to("cuda") staging and crashes with
# "No CUDA GPUs are available". Forcing torch_device="cpu" here loads the
# adapter safely on CPU; the single .to(self.device) call right after is
# a normal nn.Module.to() move, which spaces *does* stage/support.
self.lora_model = PeftModel.from_pretrained(self.base_model, str(output_dir), torch_device="cpu")
self.lora_model.to(self.device)
self.lora_model.eval()
return self.lora_model
logger.info(f"ℹ️ No saved adapter found at {output_dir}.")
return None
def train_lora(self, train_data, output_dir, epochs=3, batch_size=1, grad_accum=4, lr=2e-4, max_length=512):
lr = float(lr) # Ensure lr is a float to prevent PyTorch optimizer TypeError
if self.lora_model is None:
self.setup_lora()
dataset = SpiderDataset(train_data, self.tokenizer, max_length)
def collate_fn(batch):
max_len = max(len(b["input_ids"]) for b in batch)
input_ids, attention_mask, labels = [], [], []
pad_id = self.tokenizer.pad_token_id
for b in batch:
pad_len = max_len - len(b["input_ids"])
input_ids.append(b["input_ids"].tolist() + [pad_id] * pad_len)
attention_mask.append(b["attention_mask"].tolist() + [0] * pad_len)
labels.append(b["labels"].tolist() + [-100] * pad_len)
return {
"input_ids": torch.tensor(input_ids, dtype=torch.long).to(self.device),
"attention_mask": torch.tensor(attention_mask, dtype=torch.long).to(self.device),
"labels": torch.tensor(labels, dtype=torch.long).to(self.device)
}
loader = DataLoader(dataset, batch_size=batch_size, shuffle=True, collate_fn=collate_fn)
optimizer = torch.optim.AdamW(
filter(lambda p: p.requires_grad, self.lora_model.parameters()),
lr=lr
)
self.lora_model.train()
logger.info(f"🚀 Starting LoRA training | Device: {self.device} | Epochs: {epochs} | Batch: {batch_size} | Grad Accum: {grad_accum} | LR: {lr}")
for epoch in range(epochs):
epoch_loss = 0.0
progress_bar = tqdm(loader, desc=f"Epoch {epoch+1}/{epochs}", leave=False)
optimizer.zero_grad()
for step, batch in enumerate(progress_bar):
outputs = self.lora_model(**batch)
loss = outputs.loss / grad_accum
loss.backward()
epoch_loss += loss.item() * grad_accum
if (step + 1) % grad_accum == 0:
optimizer.step()
optimizer.zero_grad()
progress_bar.set_postfix({"loss": f"{loss.item() * grad_accum:.4f}"})
# Clear MPS cache periodically to prevent memory leaks on Mac
if self.device == "mps" and (step + 1) % 10 == 0:
torch.mps.empty_cache()
optimizer.step()
optimizer.zero_grad()
avg_loss = epoch_loss / len(loader)
logger.info(f"✅ Epoch {epoch+1} completed | Avg Loss: {avg_loss:.4f}")
output_dir = Path(output_dir)
output_dir.mkdir(parents=True, exist_ok=True)
self.lora_model.save_pretrained(str(output_dir))
self.tokenizer.save_pretrained(str(output_dir))
logger.info(f"💾 LoRA adapter saved to {output_dir}")
self.lora_model.eval()
return self.lora_model
def load_models(self, lora_path=None):
"""Returns a dictionary of models and the tokenizer."""
if self.tokenizer is None:
self.load_tokenizer()
if self.base_model is None:
self.load_base_model()
models = {"base": self.base_model, "lora": None}
if self.lora_model is not None:
logger.info("Reusing already-loaded LoRA model in memory.")
models["lora"] = self.lora_model
elif lora_path:
lora_path = Path(lora_path)
if lora_path.exists() and (lora_path / "adapter_config.json").exists():
logger.info(f"Loading LoRA model from {lora_path}")
self.lora_model = PeftModel.from_pretrained(self.base_model, str(lora_path))
self.lora_model.to(self.device)
self.lora_model.eval()
models["lora"] = self.lora_model
else:
logger.warning(f"LoRA path {lora_path} does not exist or is invalid. Returning base model as 'lora'.")
models["lora"] = self.base_model
return models, self.tokenizer |