inference/test_model_loading.py

#!/usr/bin/env python3
"""
Integration test for NVFP4 model loading.

This tests that the model can be loaded from sharded safetensors
and that all weights have correct shapes and flags.
"""

import os
import sys
import json
import torch

# Import from local inference directory
from model import Transformer, ModelArgs
from generate import load_sharded_model


def clear_cache():
    """Clear system cache to free memory."""
    print("Clearing system cache...")
    try:
        import subprocess
        subprocess.run(
            ['sudo', 'sh', '-c', 'echo 3 > /proc/sys/vm/drop_caches'],
            check=False, capture_output=True, text=True
        )
        print("  PASS: Cache cleared\n")
    except Exception as e:
        print(f"  WARN: Could not clear cache: {e}\n")


def check_memory():
    """Check available memory."""
    try:
        import psutil
        mem = psutil.virtual_memory()
        print(f"Memory: {mem.available / 1e9:.1f}GB available / {mem.total / 1e9:.1f}GB total")
        print(f"        {mem.percent:.1f}% used\n")
        return mem.available / 1e9
    except ImportError:
        print("psutil not available, skipping memory check\n")
        return None


def test_config_loading():
    """Test 1: Load and validate config."""
    print("=" * 70)
    print("Test 1: Load Model Config")
    print("=" * 70)

    config_path = "/mnt/models/deepseek-v3.2-nvfp4/inference/config_671B_nvfp4.json"

    print(f"  Loading config from: {config_path}")
    with open(config_path) as f:
        config_dict = json.load(f)

    args = ModelArgs(**config_dict)

    print(f"  Model parameters:")
    print(f"    - vocab_size: {args.vocab_size:,}")
    print(f"    - dim: {args.dim}")
    print(f"    - n_layers: {args.n_layers}")
    print(f"    - n_routed_experts: {args.n_routed_experts}")
    print(f"    - dtype: {args.dtype}")

    assert args.dtype == "nvfp4", f"Expected dtype='nvfp4', got '{args.dtype}'"
    assert args.n_layers == 61, f"Expected 61 layers, got {args.n_layers}"

    print(f"  PASS: Config loaded successfully")
    print(f"  PASS: Test 1 PASSED\n")

    return args


def test_model_creation(args):
    """Test 2: Create model instance."""
    print("=" * 70)
    print("Test 2: Create Model Instance")
    print("=" * 70)

    print(f"  Creating Transformer model with dtype={args.dtype}...")
    print(f"  (This may take 1-2 minutes)")

    torch.set_default_dtype(torch.bfloat16)

    with torch.device("cpu"):
        model = Transformer(args)

    total_params = sum(p.numel() for p in model.parameters())
    total_buffers = sum(b.numel() for b in model.buffers())

    print(f"  Model created:")
    print(f"    - Parameters: {total_params / 1e9:.2f}B")
    print(f"    - Buffers: {total_buffers / 1e9:.2f}B")
    print(f"    - Total: {(total_params + total_buffers) / 1e9:.2f}B elements")

    # Check that model has the right structure
    assert hasattr(model, 'embed'), "Model missing embed layer"
    assert hasattr(model, 'layers'), "Model missing layers"
    assert len(model.layers) == args.n_layers, f"Expected {args.n_layers} layers, got {len(model.layers)}"

    print(f"  PASS: Model structure correct")
    print(f"  PASS: Test 2 PASSED\n")

    return model


def test_weight_loading(model):
    """Test 3: Load weights from sharded checkpoint."""
    print("=" * 70)
    print("Test 3: Load Weights from Checkpoint")
    print("=" * 70)

    ckpt_path = "/mnt/models/deepseek-v3.2-nvfp4"

    print(f"  Loading from: {ckpt_path}")
    print(f"  (This will take 5-15 minutes for the full model)")
    print(f"  Progress will be shown shard-by-shard...\n")

    load_sharded_model(model, ckpt_path)

    print(f"\n  PASS: Weights loaded successfully")
    print(f"  PASS: Test 3 PASSED\n")

    return model


def test_nvfp4_layers(model):
    """Test 4: Verify NVFP4 layers have correct structure."""
    print("=" * 70)
    print("Test 4: Verify NVFP4 Layer Structure")
    print("=" * 70)

    nvfp4_layers = []
    total_layers = 0

    for name, module in model.named_modules():
        # Check if this is a Linear layer
        if hasattr(module, '_nvfp4_mode') and hasattr(module, 'weight'):
            total_layers += 1
            if getattr(module, '_nvfp4_mode', False):
                nvfp4_layers.append((name, module))

    print(f"  Found {len(nvfp4_layers)} NVFP4 layers out of {total_layers} total linear layers")

    if len(nvfp4_layers) == 0:
        print(f"  WARN: WARNING: No NVFP4 layers found!")
        print(f"     This might indicate dtype configuration issue")
        return

    # Check first few layers in detail
    print(f"\n  Inspecting first 5 NVFP4 layers:")
    for i, (name, module) in enumerate(nvfp4_layers[:5]):
        weight = module.weight
        weight_scale = module.weight_scale if hasattr(module, 'weight_scale') else None
        weight_scale_2 = module.weight_scale_2 if hasattr(module, 'weight_scale_2') else None

        print(f"\n  [{i+1}] {name}:")
        print(f"      weight: {weight.shape}, dtype={weight.dtype}")

        # Verify shapes
        N, K_half = weight.shape
        K = K_half * 2

        if weight_scale is not None:
            print(f"      weight_scale: {weight_scale.shape}, dtype={weight_scale.dtype}")
            expected_scale_shape = (N, K // 16)
            if weight_scale.shape != expected_scale_shape:
                print(f"      WARN: WARNING: Expected scale shape {expected_scale_shape}, got {weight_scale.shape}")
            else:
                print(f"      PASS: Scale shape correct")
        else:
            print(f"      WARN: WARNING: weight_scale not found!")

        if weight_scale_2 is not None:
            print(f"      weight_scale_2: {weight_scale_2.shape}, dtype={weight_scale_2.dtype}, value={weight_scale_2.item():.6e}")
            if weight_scale_2.shape != torch.Size([1]):
                print(f"      WARN: WARNING: Expected scale_2 shape [1], got {weight_scale_2.shape}")
            else:
                print(f"      PASS: Scale_2 shape correct")
        else:
            print(f"      WARN: WARNING: weight_scale_2 not found!")

        # Verify weight is uint8 packed
        assert weight.dtype == torch.uint8, f"Weight should be uint8, got {weight.dtype}"

    print(f"\n  PASS: NVFP4 layers have correct structure")
    print(f"  PASS: Test 4 PASSED\n")


def test_weight_statistics(model):
    """Test 5: Check weight statistics to verify they're not zeros or corrupted."""
    print("=" * 70)
    print("Test 5: Weight Statistics")
    print("=" * 70)

    # Sample a few layers
    nvfp4_count = 0
    zero_count = 0
    checked = 0

    for name, module in model.named_modules():
        if hasattr(module, '_nvfp4_mode') and getattr(module, '_nvfp4_mode', False):
            nvfp4_count += 1

            # Check only first 10 layers for speed
            if checked < 10:
                weight = module.weight
                weight_scale = module.weight_scale if hasattr(module, 'weight_scale') else None
                weight_scale_2 = module.weight_scale_2 if hasattr(module, 'weight_scale_2') else None

                # Count zeros in packed weight
                num_zeros = (weight == 0).sum().item()
                total_elems = weight.numel()
                zero_percent = 100.0 * num_zeros / total_elems

                if checked == 0:
                    print(f"\n  Sample layer: {name}")
                    print(f"    Weight zeros: {zero_percent:.1f}%")
                    if weight_scale is not None:
                        scale_min = weight_scale.to(torch.float32).min().item()
                        scale_max = weight_scale.to(torch.float32).max().item()
                        print(f"    Scale range: [{scale_min:.6e}, {scale_max:.6e}]")
                    if weight_scale_2 is not None:
                        print(f"    Scale_2: {weight_scale_2.item():.6e}")

                # Detect completely zero weights (corruption)
                if zero_percent > 95:
                    zero_count += 1
                    print(f"  WARN: WARNING: {name} has {zero_percent:.1f}% zeros (possibly corrupted)")

                checked += 1

    print(f"\n  Checked {checked} NVFP4 layers:")
    print(f"    - Total NVFP4 layers: {nvfp4_count}")
    print(f"    - Layers with >95% zeros: {zero_count}")

    if zero_count > checked // 2:
        print(f"  WARN: WARNING: Many layers appear corrupted (too many zeros)")
    else:
        print(f"  PASS: Weight statistics look reasonable")

    print(f"  PASS: Test 5 PASSED\n")


def main():
    """Run all model loading tests."""
    print("\n" + "=" * 70)
    print("NVFP4 Model Loading Integration Test")
    print("=" * 70)
    print("This test will load the full 671B parameter model")
    print("Expected runtime: 5-20 minutes")
    print("Memory required: ~400GB")
    print("=" * 70 + "\n")

    # Check memory first
    available_gb = check_memory()
    if available_gb is not None and available_gb < 350:
        print(f"WARN: WARNING: Only {available_gb:.1f}GB available")
        print(f"   Model may not fit in memory. Consider clearing cache.")
        user_input = input("   Continue anyway? (y/n): ")
        if user_input.lower() != 'y':
            print("   Aborted by user")
            return 1

    # Offer to clear cache
    user_input = input("Clear system cache before loading? (recommended) (y/n): ")
    if user_input.lower() == 'y':
        clear_cache()
        check_memory()

    try:
        # Run tests
        args = test_config_loading()
        model = test_model_creation(args)
        model = test_weight_loading(model)
        test_nvfp4_layers(model)
        test_weight_statistics(model)

        # Final summary
        print("=" * 70)
        print("PASS: ALL TESTS PASSED")
        print("=" * 70)
        print("Model loaded successfully with correct NVFP4 structure!")
        print("Ready for forward pass testing.")
        print("=" * 70)

        # Keep model in memory for next test
        print("\nKeeping model in memory for next test...")
        print("Run test_forward_pass.py in the same Python session to reuse loaded model")

        return 0

    except Exception as e:
        print(f"\nFAIL: TEST FAILED: {e}")
        import traceback
        traceback.print_exc()
        return 1


if __name__ == "__main__":
    sys.exit(main())