feat: implement MultiDtypeGenerator for low-precision sample generation

- Add MultiDtypeGenerator class to generate float16/bfloat16/float8 samples from float32 samples
- Automatically convert input and weight tensor dtypes
- Keep batch_norm parameters (running_mean, running_var, scale, bias) as float32
- Add torch.autocast context manager to model.py forward method
- Update graph_net.json metadata with dtype information
- Add sample validation mechanism
- Add MultiDtypeFilter for filtering invalid graphs
- Add command-line tool for batch generation
- Support test_compiler evaluation and Agent code generation workflows

Author: Dayuxiaoshui

graph_net/torch/multi_dtype_generate.py

@@ -0,0 +1,84 @@
+"""
+Command-line tool for MultiDtypeGenerator.
+
+Usage:
+    python -m graph_net.torch.multi_dtype_generate \
+        --source-sample-path /path/to/source/sample \
+        --output-base-path /path/to/output \
+        --dtype-list float16 bfloat16
+"""
+
+import argparse
+import sys
+from pathlib import Path
+from graph_net.torch.multi_dtype_generator import MultiDtypeGenerator
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description='Generate low-precision samples from existing float32 samples'
+    )
+    parser.add_argument(
+        '--source-sample-path',
+        type=str,
+        required=True,
+        help='Path to the source sample directory'
+    )
+    parser.add_argument(
+        '--output-base-path',
+        type=str,
+        required=True,
+        help='Base path for output samples'
+    )
+    parser.add_argument(
+        '--dtype-list',
+        type=str,
+        nargs='+',
+        default=['float16', 'bfloat16'],
+        choices=['float16', 'bfloat16', 'float8'],
+        help='List of target dtypes to generate (default: float16 bfloat16)'
+    )
+    parser.add_argument(
+        '--filter-config',
+        type=str,
+        default=None,
+        help='Path to filter configuration file (JSON format)'
+    )
+    
+    args = parser.parse_args()
+    
+    # Load filter config if provided
+    filter_config = {}
+    if args.filter_config:
+        import json
+        with open(args.filter_config, 'r') as f:
+            filter_config = json.load(f)
+    
+    # Create generator
+    generator = MultiDtypeGenerator(
+        source_sample_path=args.source_sample_path,
+        output_base_path=args.output_base_path,
+        dtype_list=args.dtype_list,
+        filter_config=filter_config,
+    )
+    
+    # Generate samples
+    print(f"Generating samples from: {args.source_sample_path}")
+    print(f"Output base path: {args.output_base_path}")
+    print(f"Target dtypes: {args.dtype_list}")
+    
+    generated_paths = generator.generate()
+    
+    if generated_paths:
+        print(f"\nSuccessfully generated {len(generated_paths)} sample(s):")
+        for path in generated_paths:
+            print(f"  - {path}")
+        return 0
+    else:
+        print("\nNo samples were generated successfully.")
+        return 1
+
+
+if __name__ == '__main__':
+    sys.exit(main())
+