Fix multiprocessing pickling error for multi-GPU training

[Copilot]

Training with multiple GPUs or multiple epochs was failing with ValueError: ctypes objects containing pointers cannot be pickled when PyTorch Lightning attempted to spawn worker processes for distributed training.

The root cause was module-level ctypes objects in nnue_dataset.py that contained C pointers which cannot be serialized for multiprocessing:

# Before: Module-level ctypes objects (unpicklable)
dll = ctypes.cdll.LoadLibrary(dllpath)
create_sparse_batch_stream = dll.create_sparse_batch_stream
# ... other ctypes function pointers

class TrainingDataProvider:
    def __init__(self, ...):
        self.stream = self.create_stream(...)  # Contains C pointers

Solution:

• Moved ctypes dll loading from module-level to per-instance in TrainingDataProvider
• Added custom __getstate__/__setstate__ methods to handle pickling safely by excluding ctypes objects
• Implemented lazy initialization so each worker process creates its own ctypes objects instead of inheriting serialized ones

# After: Per-instance ctypes loading (picklable)
class TrainingDataProvider:
    def _setup_dll(self):
        dll = _load_dll()  # Load fresh in each process
        self._create_stream_func, ... = _setup_dll_functions(dll)
        self.stream = self._create_stream_func(...)
    
    def __getstate__(self):
        # Exclude unpicklable ctypes objects during serialization
        state = self.__dict__.copy()
        state['stream'] = None
        return state

Testing:

• Verified multiprocessing with spawn method (PyTorch Lightning's approach) works correctly
• Confirmed backward compatibility - all existing APIs continue to work
• Validated performance - no regressions in dataset creation or pickling speed
• Tested train.py startup with multi-GPU parameters successfully

The original failing command should now work:

python train.py --threads 4 --num-workers 4 --gpus 2 --max_epochs 100 ../janggi-d8-1b.bin ../janggi-d8-1b.bin

Fixes #39.

---

💡 You can make Copilot smarter by setting up custom instructions, customizing its development environment and configuring Model Context Protocol (MCP) servers. Learn more Copilot coding agent tips in the docs.

[Copilot]

@ianfab 👋 This repository doesn't have Copilot instructions. With Copilot instructions, I can understand the repository better, work faster and produce higher quality PRs.

I can generate a .github/copilot-instructions.md file for you automatically. Click here to open a pre-filled issue and assign it to me. I'll write the instructions, and then tag you for review.