#!/usr/bin/env python3
"""Test loading bf16 model with BnB 4-bit to CPU, then report size."""

import torch
from transformers import AutoModelForCausalLM, BitsAndBytesConfig

model_path = "/data/models/Ornith-1.0-35B"

print(f"Loading model from: {model_path}")
print(f"\nApplying BnB 4-bit quantization...")

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.float16,
)

model = AutoModelForCausalLM.from_pretrained(
    model_path,
    quantization_config=bnb_config,
    device_map="cpu",
    torch_dtype=torch.float16,
    trust_remote_code=True,
    low_cpu_mem_usage=True,
)

print("✓ Model loaded to CPU with BnB 4-bit")

# Count parameters
total_params = sum(p.numel() for p in model.parameters())
print(f"\nTotal parameters: {total_params / 1e9:.2f}B")

# Check for quantized parameters
bnb_params = 0
bf16_params = 0
for name, p in model.named_parameters():
    if hasattr(p, 'quant_state') and p.quant_state is not None:
        bnb_params += p.numel()
    else:
        bf16_params += p.numel()

print(f"BnB 4-bit parameters: {bnb_params / 1e9:.2f}B")
print(f"BF16 parameters: {bf16_params / 1e9:.2f}B")
print(f"Estimated size: {(bnb_params * 0.5 + bf16_params * 2) / 1e9:.2f} GB")

# Try to move to GPU
print("\n   Moving to GPU 0...")
try:
    model = model.to("cuda:0")
    print(f"✓ Success! GPU 0: {torch.cuda.memory_allocated(0) / 1e9:.2f} GB")
    print(f"   Free VRAM: {torch.cuda.get_device_properties(0).total_memory / 1e9 - torch.cuda.memory_allocated(0) / 1e9:.2f} GB")
except Exception as e:
    print(f"✗ FAILED: {e}")

del model
import gc
gc.collect()
torch.cuda.empty_cache()