Spaces:
Sleeping
Sleeping
File size: 5,410 Bytes
a952b69 ab8f3fb a952b69 |
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 |
#!/usr/bin/env python3
"""
train_lora.py
- Fine-tune DeepSeek 1.3B with LoRA (QLoRA-ish setup)
- Save adapters using safe_serialization=True -> adapter_model.safetensors
- Upload adapter folder to Hugging Face Hub (VaibhavHD/deepseek-lora-monthly)
- Log metrics/artifact to Weights & Biases
"""
import os
import json
import wandb
import torch
from huggingface_hub import HfApi
from datasets import load_dataset
from transformers import (
AutoTokenizer, AutoModelForCausalLM,
TrainingArguments, Trainer, DataCollatorForLanguageModeling
)
from peft import LoraConfig, get_peft_model
# -----------------------------
# Config (edit if needed)
# -----------------------------
HF_REPO = "VaibhavHD/deepseek-lora-monthly" # your HF model repo
MODEL_NAME = "deepseek-ai/deepseek-coder-1.3b-base"
OUT_DIR = "out"
ADAPTER_DIR = os.path.join(OUT_DIR, "lora_adapters")
# env secrets expected:
HF_TOKEN = os.getenv("HF_TOKEN")
WANDB_API_KEY = os.getenv("WANDB_API_KEY")
if WANDB_API_KEY:
wandb.login(key=WANDB_API_KEY)
else:
print("⚠️ WANDB_API_KEY not found in env; continuing without W&B logging.")
# -----------------------------
# Load dataset
# -----------------------------
print("Loading dataset...")
dataset = {}
dataset['train'] = load_dataset("westenfelder/NL2SH-ALFA", "train")["train"]
dataset['test'] = load_dataset("westenfelder/NL2SH-ALFA", "test")["train"]
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, trust_remote_code=True)
def tokenize_fn(batch):
texts = [f"{nl} => {bash}" for nl, bash in zip(batch["nl"], batch["bash"])]
return tokenizer(texts, truncation=True, padding="max_length", max_length=512)
train = dataset["train"].map(tokenize_fn, batched=True)
test = dataset["test"].map(tokenize_fn, batched=True)
# Optional small-subset for fast runs (uncomment to use)
# train = train.shuffle(seed=42).select(range(200))
# test = test.shuffle(seed=42).select(range(20))
# -----------------------------
# Load base model (half precision)
# -----------------------------
print("Loading base model (may take a moment)...")
model = AutoModelForCausalLM.from_pretrained(
MODEL_NAME,
torch_dtype=torch.float16,
low_cpu_mem_usage=True,
device_map="auto",
trust_remote_code=True
)
# avoid caching issues
model.config.use_cache = False
for p in model.parameters():
p.requires_grad = False
# -----------------------------
# Attach LoRA
# -----------------------------
print("Attaching LoRA adapters...")
lora_config = LoraConfig(
r=8,
lora_alpha=16,
target_modules=[
"q_proj", "v_proj", "k_proj", "o_proj",
"gate_proj", "down_proj", "up_proj"
],
lora_dropout=0.05,
bias="none",
task_type="CAUSAL_LM",
)
model = get_peft_model(model, lora_config)
# -----------------------------
# Data collator + training args
# -----------------------------
data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False)
training_args = TrainingArguments(
output_dir=OUT_DIR,
num_train_epochs=1,
per_device_train_batch_size=1,
gradient_accumulation_steps=8,
learning_rate=2e-4,
fp16=True,
save_strategy="epoch",
logging_steps=25,
report_to=["wandb"] if WANDB_API_KEY else [],
)
trainer = Trainer(
model=model,
args=training_args,
train_dataset=train,
eval_dataset=test,
data_collator=data_collator,
)
# -----------------------------
# Run training
# -----------------------------
print("Starting training...")
if WANDB_API_KEY:
wandb.init(project="deepseek-qlora-monthly", name="deepseek-lite-run")
trainer.train()
# -----------------------------
# Evaluate and save metrics
# -----------------------------
print("Evaluating...")
metrics = trainer.evaluate()
# compute simple "accuracy-like" metric from loss (replace with real metric if you have one)
new_acc = 1.0 - metrics.get("eval_loss", 1.0)
print(f"Eval metrics: {metrics}")
print(f"Pseudo-accuracy (1 - eval_loss): {new_acc:.6f}")
os.makedirs(ADAPTER_DIR, exist_ok=True)
metrics_path = os.path.join(OUT_DIR, "metrics.json")
with open(metrics_path, "w") as f:
json.dump(metrics, f)
if WANDB_API_KEY:
wandb.log({"accuracy": new_acc})
# log artifact
artifact = wandb.Artifact(
name="deepseek-lora-adapters",
type="model",
description="LoRA adapters saved with safe_serialization"
)
# -----------------------------
# Save adapters using safe_serialization
# -----------------------------
print("Saving adapters with safe_serialization=True (produces .safetensors)...")
model.save_pretrained(ADAPTER_DIR, safe_serialization=True)
tokenizer.save_pretrained(ADAPTER_DIR)
# add to wandb artifact directory
if WANDB_API_KEY:
artifact.add_dir(ADAPTER_DIR)
wandb.log_artifact(artifact, aliases=["latest"])
print(f"Adapters saved to: {ADAPTER_DIR}")
print("Files in adapter dir:", os.listdir(ADAPTER_DIR))
# -----------------------------
# Upload to Hugging Face model repo
# -----------------------------
if HF_TOKEN:
print(f"Uploading adapter folder to Hugging Face repo: {HF_REPO}")
api = HfApi()
# upload_folder will overwrite same filenames in the repo
api.upload_folder(
folder_path=ADAPTER_DIR,
path_in_repo=".",
repo_id=HF_REPO,
token=HF_TOKEN
)
print("✅ Upload complete.")
else:
print("⚠️ HF_TOKEN not set. Skipping upload to Hugging Face Hub.")
|