Update train.py

train.py

@@ -1,6 +1,6 @@
 # ==============================================
 # Stephen Model Fine-Tuning Script (LoRA + PEFT)
-# Optimized for JSONL dataset
+# Clean + Debug-Enhanced for Grad & Deprecation Warnings
 # ==============================================
 
 !pip install -q "transformers>=4.44.0" "datasets" "peft>=0.12.0" accelerate bitsandbytes sentencepiece huggingface_hub
@@ -11,9 +11,10 @@ from huggingface_hub import login, whoami
 from datasets import load_dataset
 from transformers import (
     AutoTokenizer, AutoModelForCausalLM, TrainingArguments,
-    Trainer, DataCollatorForLanguageModeling
+    Trainer, DataCollatorForLanguageModeling, BitsAndBytesConfig
 )
 from peft import LoraConfig, get_peft_model
+from peft import prepare_model_for_kbit_training
 import torch
 
 # ==============================================
@@ -27,21 +28,17 @@ def log(msg):
 # ==============================================
 HF_TOKEN = os.getenv("HF_TOKEN")
 if not HF_TOKEN:
-    raise ValueError(
-        "❌ HF_TOKEN environment variable not set.\n"
-        "Run in RunPod: %env HF_TOKEN='your_token_here'\n"
-        "Or set in UI → Environment Variables."
-    )
+    raise ValueError("❌ HF_TOKEN environment variable not set.")
 
 log("Logging into Hugging Face...")
 login(token=HF_TOKEN, add_to_git_credential=True)
 log(f"Logged in as: {whoami()['name']} ✅")
 
 # ==============================================
-# 2. Load JSONL Dataset from HF
+# 2. Load Dataset
 # ==============================================
-dataset_name = "dgtalbug/stephen-dataset"
-data_file = "stephen.jsonl"
+dataset_name = "dgtalbug/stephen-dataset"  # CHANGE THIS
+data_file = "stephen.jsonl"                # CHANGE THIS
 
 log(f"Loading dataset: {dataset_name}/{data_file} ...")
 dataset = load_dataset(dataset_name, data_files=data_file, split="train")
@@ -51,7 +48,7 @@ log(f"First example: {dataset[0]}")
 # ==============================================
 # 3. Load Base Model & Tokenizer
 # ==============================================
-base_model = "dgtalbug/stable-code-instruct-3b"
+base_model = "dgtalbug/stable-code-instruct-3b"  # CHANGE THIS
 log(f"Loading base model: {base_model}...")
 
 tokenizer = AutoTokenizer.from_pretrained(
@@ -62,23 +59,31 @@ tokenizer = AutoTokenizer.from_pretrained(
 if tokenizer.pad_token is None:
     tokenizer.pad_token = tokenizer.eos_token
 
+# ✅ Quantization config
+bnb_config = BitsAndBytesConfig(
+    load_in_8bit=True,
+    llm_int8_threshold=6.0
+)
+
 try:
     model = AutoModelForCausalLM.from_pretrained(
         base_model,
         token=HF_TOKEN,
-        load_in_8bit=True,
         device_map="auto",
         torch_dtype=torch.float16,
-        trust_remote_code=True
+        trust_remote_code=True,
+        return_dict=True,
+        quantization_config=bnb_config
     )
 except Exception as e:
-    log(f"⚠️ 8-bit load failed: {e} — falling back to full precision.")
+    log(f"⚠️ Quantized load failed: {e} — falling back to fp16.")
     model = AutoModelForCausalLM.from_pretrained(
         base_model,
         token=HF_TOKEN,
         device_map="auto",
         torch_dtype=torch.float16,
-        trust_remote_code=True
+        trust_remote_code=True,
+        return_dict=True
     )
 
 log("Base model loaded ✅")
@@ -86,42 +91,79 @@ log("Base model loaded ✅")
 # ==============================================
 # 4. LoRA Config
 # ==============================================
+# log("Configuring LoRA...")
+# lora_config = LoraConfig(
+#     r=16,
+#     lora_alpha=32,
+#     target_modules=["q_proj", "v_proj", "k_proj", "o_proj"],
+#     lora_dropout=0.05,
+#     bias="none",
+#     task_type="CAUSAL_LM"
+# )
+# model = get_peft_model(model, lora_config)
+
+# # ✅ Ensure LoRA params require grad
+# for name, param in model.named_parameters():
+#     if "lora" in name:
+#         param.requires_grad = True
+#     else:
+#         param.requires_grad = False
+
+# # ✅ Sanity check: see how many params are trainable
+# model.print_trainable_parameters()
+
+# log("LoRA config applied ✅")
 log("Configuring LoRA...")
+
+# First, prepare for 8-bit training (important for bitsandbytes)
+model = prepare_model_for_kbit_training(model)
+
+# LoRA config
 lora_config = LoraConfig(
     r=16,
     lora_alpha=32,
-    target_modules=["q_proj", "v_proj", "k_proj", "o_proj"],  # Safe for StableCode
+    target_modules=["q_proj", "v_proj", "k_proj", "o_proj"],  # adjust if needed
     lora_dropout=0.05,
     bias="none",
     task_type="CAUSAL_LM"
 )
+
+# Apply LoRA
 model = get_peft_model(model, lora_config)
 
-# ✅ Enable training for LoRA params
+# Double-check trainable params
+trainable_params = []
 for name, param in model.named_parameters():
-    if "lora" in name:
-        param.requires_grad = True
+    if param.requires_grad:
+        trainable_params.append(name)
+
+if not trainable_params:
+    raise RuntimeError("❌ No parameters set to require gradients! LoRA not applied correctly.")
 
-log("LoRA config applied ✅")
+log(f"✅ Found {len(trainable_params)} trainable parameters.")
+log(f"First 20 trainable params: {trainable_params[:20]}")
 
+# Print PEFT/LoRA summary
+model.print_trainable_parameters()
 # ==============================================
-# 5. Tokenize Dataset (JSONL)
+# 5. Tokenize Dataset
 # ==============================================
 log("Tokenizing dataset...")
 
-# Detect correct text key
 first_row = dataset[0]
 if "text" in first_row:
     text_key = "text"
 elif "prompt" in first_row:
     text_key = "prompt"
 else:
-    text_key = list(first_row.keys())[0]  # fallback
+    text_key = list(first_row.keys())[0]
 
 log(f"Using text key: '{text_key}'")
 
 def tokenize_fn(example):
-    return tokenizer(example[text_key], truncation=True, padding="max_length", max_length=512)
+    tokenized = tokenizer(example[text_key], truncation=True, padding="max_length", max_length=512)
+    tokenized["labels"] = tokenized["input_ids"].copy()  # ✅ Ensure labels exist for grad
+    return tokenized
 
 tokenized_dataset = dataset.map(tokenize_fn, batched=True, remove_columns=dataset.column_names)
 log("Tokenization complete ✅")
@@ -137,35 +179,48 @@ data_collator = DataCollatorForLanguageModeling(tokenizer, mlm=False)
 # ==============================================
 output_dir = "./stephen-lora"
 log("Preparing training arguments...")
+
 training_args = TrainingArguments(
     output_dir=output_dir,
     overwrite_output_dir=True,
-    per_device_train_batch_size=4,
-    gradient_accumulation_steps=4,
+    per_device_train_batch_size=8,
+    gradient_accumulation_steps=2,
     gradient_checkpointing=True,
-    warmup_steps=100,
-    max_steps=1000,
-    learning_rate=2e-4,
+    warmup_steps=50,
+    num_train_epochs=3,
+    max_steps=-1,
+    learning_rate=1e-4,
     lr_scheduler_type="cosine",
     fp16=True,
+    optim="adamw_torch",
     logging_dir="./logs",
-    logging_steps=10,
-    save_strategy="steps",
-    save_steps=200,
+    logging_steps=20,
+    save_strategy="epoch",
     save_total_limit=2,
-    report_to="none",
     push_to_hub=True,
-    ddp_find_unused_parameters=False
+    hub_strategy="end",
+    ddp_find_unused_parameters=False,
+    label_names=["labels"]
 )
 log("Training arguments ready ✅")
 
 # ==============================================
-# 8. Custom Trainer
+# 8. Debugging Helper Hooks
 # ==============================================
-class NoUnpackTrainer(Trainer):
-    """Avoids unpacking bug in HF Trainer."""
-    def compute_loss(self, model, inputs, return_outputs=False, num_items_in_batch=None):
-        inputs = self._prepare_inputs(inputs)
+def debug_batch(batch):
+    log(f"🔍 Debug batch keys: {list(batch.keys())}")
+    log(f"🔍 First input_ids: {batch['input_ids'][0][:10]}")
+    log(f"🔍 First labels: {batch['labels'][0][:10]}")
+    log(f"🔍 labels.requires_grad? {torch.tensor(batch['labels']).requires_grad}")
+
+# ==============================================
+# 9. Custom Trainer (safe + debug)
+# ==============================================
+class SafeTrainer(Trainer):
+    def compute_loss(self, model, inputs, return_outputs=False, **kwargs):
+        # Debug batch content once
+        if self.state.global_step == 0:
+            debug_batch(inputs)
         if "labels" not in inputs:
             inputs["labels"] = inputs["input_ids"].clone()
         outputs = model(**inputs)
@@ -173,7 +228,7 @@ class NoUnpackTrainer(Trainer):
         return (loss, outputs) if return_outputs else loss
 
 log("Initializing Trainer...")
-trainer = NoUnpackTrainer(
+trainer = SafeTrainer(
     model=model,
     args=training_args,
     train_dataset=tokenized_dataset,
@@ -182,9 +237,12 @@ trainer = NoUnpackTrainer(
 log("Trainer initialized ✅")
 
 # ==============================================
-# 9. Train & Push to Hub
+# 10. Train & Push
 # ==============================================
-log("Starting training...")
+trainable_params = [n for n, p in model.named_parameters() if p.requires_grad]
+log(f"Trainable params count: {len(trainable_params)}")
+log(f"First 20 trainable params: {trainable_params[:20]}")
+
 last_ckpt = None
 if os.path.isdir(output_dir):
     checkpoints = [d for d in os.listdir(output_dir) if d.startswith("checkpoint-")]