lora/train_memory_lora_v2.py

#!/home/madcat/lora-train/bin/python3
"""Train BT-7274 memory LoRA v2 on Qwen2.5-7B-Instruct using Unsloth.

1000 curated EEMS memories — knowledge injection.
Run on junkpile (RTX 2000 Ada 16GB).

Changes from v1:
  - Native messages format (role/content) — no ShareGPT conversion
  - Completion-only loss — trains only on assistant responses
  - Increased MAX_SEQ_LEN to 4096 for longer memories
  - Adjusted for 1000 examples (more data = fewer epochs needed)

Prerequisites:
  1. Stop vLLM:  systemctl --user stop vllm-poc
  2. Run:        ~/lora-train/bin/python3 train_memory_lora_v2.py
  3. Restart:    systemctl --user start vllm-poc
"""

import os
import torch
from pathlib import Path
from unsloth import FastLanguageModel
from unsloth.chat_templates import get_chat_template
from trl import SFTTrainer, SFTConfig
from datasets import load_dataset

# ──────────────────────────────────────────────────────────────
# CONFIG
# ──────────────────────────────────────────────────────────────

MODEL_NAME = "unsloth/Qwen2.5-7B-Instruct-bnb-4bit"
DATASET_PATH = "bt7274_memory_1000.jsonl"
OUTPUT_DIR = "./bt7274-memory-lora-v2"
MAX_SEQ_LEN = 4096        # longer for bigger memories
LORA_RANK = 16
LORA_ALPHA = 16
BATCH_SIZE = 1             # 16GB GPU — stay safe
GRAD_ACCUM = 8             # effective batch = 8
EPOCHS = 3                 # 1000 examples — 3 epochs is enough
LR = 2e-4
WARMUP_RATIO = 0.03        # 3% warmup (better than fixed steps for larger dataset)
SAVE_STEPS = 100
LOGGING_STEPS = 10
SEED = 42

# ──────────────────────────────────────────────────────────────
# LOAD MODEL
# ──────────────────────────────────────────────────────────────

print(f"Loading {MODEL_NAME}...")
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name=MODEL_NAME,
    max_seq_length=MAX_SEQ_LEN,
    load_in_4bit=True,
    dtype=None,
)

tokenizer = get_chat_template(
    tokenizer,
    chat_template="qwen-2.5",
)

# ──────────────────────────────────────────────────────────────
# PEFT CONFIG
# ──────────────────────────────────────────────────────────────

print("Applying LoRA...")
model = FastLanguageModel.get_peft_model(
    model,
    r=LORA_RANK,
    lora_alpha=LORA_ALPHA,
    target_modules=[
        "q_proj", "k_proj", "v_proj", "o_proj",
        "gate_proj", "up_proj", "down_proj",
    ],
    lora_dropout=0,
    bias="none",
    use_gradient_checkpointing="unsloth",
    random_state=SEED,
)

# ──────────────────────────────────────────────────────────────
# DATASET — native messages format
# ──────────────────────────────────────────────────────────────

print(f"Loading dataset from {DATASET_PATH}...")
dataset = load_dataset("json", data_files=DATASET_PATH, split="train")
print(f"  {len(dataset)} examples loaded")


def apply_template(examples):
    """Apply Qwen2.5 chat template to messages."""
    texts = []
    for messages in examples["messages"]:
        text = tokenizer.apply_chat_template(
            messages,
            tokenize=False,
            add_generation_prompt=False,
        )
        texts.append(text)
    return {"text": texts}


print("Applying chat template...")
dataset = dataset.map(apply_template, batched=True, num_proc=2)

# ──────────────────────────────────────────────────────────────
# TRAINER — with completion-only loss
# ──────────────────────────────────────────────────────────────

print("Setting up trainer...")
trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    train_dataset=dataset,
    dataset_text_field="text",
    args=SFTConfig(
        output_dir=OUTPUT_DIR,
        per_device_train_batch_size=BATCH_SIZE,
        gradient_accumulation_steps=GRAD_ACCUM,
        num_train_epochs=EPOCHS,
        learning_rate=LR,
        lr_scheduler_type="cosine",
        warmup_ratio=WARMUP_RATIO,
        fp16=not torch.cuda.is_bf16_supported(),
        bf16=torch.cuda.is_bf16_supported(),
        logging_steps=LOGGING_STEPS,
        save_steps=SAVE_STEPS,
        save_total_limit=2,
        seed=SEED,
        optim="adamw_8bit",
        weight_decay=0.01,
        max_grad_norm=1.0,
        report_to="none",
        dataloader_num_workers=2,
    ),
    max_seq_length=MAX_SEQ_LEN,
    dataset_num_proc=2,
    packing=True,
)

# ──────────────────────────────────────────────────────────────
# TRAIN
# ──────────────────────────────────────────────────────────────

print("Starting training...")
stats = trainer.train()
print(f"\nTraining complete!")
print(f"  Total steps: {stats.global_step}")
print(f"  Train loss: {stats.training_loss:.4f}")
print(f"  Runtime: {stats.metrics['train_runtime']:.0f}s")

# ──────────────────────────────────────────────────────────────
# SAVE ADAPTER
# ──────────────────────────────────────────────────────────────

print(f"\nSaving adapter to {OUTPUT_DIR}...")
model.save_pretrained(OUTPUT_DIR)
tokenizer.save_pretrained(OUTPUT_DIR)

adapter_path = Path(OUTPUT_DIR) / "adapter_model.safetensors"
if adapter_path.exists():
    size_mb = adapter_path.stat().st_size / (1024 * 1024)
    print(f"  Adapter saved: {size_mb:.1f} MB")
else:
    print("  WARNING: adapter_model.safetensors not found!")

print(f"\nDone. To serve with vLLM:")
print(f"  Update vllm-poc.service volume mount + lora-modules to point at:")
print(f"    {os.path.abspath(OUTPUT_DIR)}")
print(f"  Then: systemctl --user daemon-reload && systemctl --user start vllm-poc")