feat: Add low-VRAM training framework for 4GB GPUs

[Doufless1]

Low-VRAM Training Framework for GO-1

Description

This PR adds a memory-efficient training framework that enables GO-1 model fine-tuning on consumer GPUs with as little as 4GB VRAM (tested on NVIDIA GTX 970).

Problem

The current training setup requires ~70GB VRAM, making it inaccessible to most researchers and developers. This limits community contributions and experimentation.

Solution

A modular low-VRAM training framework built with Clean Architecture and SOLID principles:

Key Features

Feature	Benefit
Gradient Accumulation	Simulate larger batch sizes with micro-batches
Mixed Precision (FP16/BF16)	50% memory reduction for activations
CPU Offloading	Move optimizer states to CPU with integrity checks
Selective Freezing	Train only Action Expert (755MB vs 5GB full model)
Feature Caching	Pre-compute vision features with SHA256 verification
Preset Configs	Ready-to-use settings for 4GB and 8GB GPUs

Architecture

go1/tools/low_vram/
├── core/               # Domain layer - interfaces and data classes
├── adapters/           # Interface adapters - concrete implementations
└── infrastructure/     # Framework layer - trainer, factory, config

Files Changed

• 14 new files in go1/tools/low_vram/
• 18 unit tests in tests/low_vram/
• 2,337 lines of documented, tested code

Testing

All unit tests pass:

python -m unittest tests.low_vram.test_components -v
# 18 tests passed

Tested on:

• NVIDIA GTX 970 (4GB VRAM)
• Windows 11 / Python 3.11.9 / PyTorch 2.5.1

Usage Example

from go1.tools.low_vram.infrastructure.factory import (
    DefaultTrainerFactory, 
    create_4gb_config
)

# Create optimized configs for 4GB GPU
memory_config, training_config = create_4gb_config()

# Build trainer with dependency injection
factory = DefaultTrainerFactory()
trainer = LowVRAMTrainer(
    model=model,
    optimizer=optimizer,
    memory_manager=factory.create_memory_manager(memory_config),
    training_strategy=factory.create_training_strategy(training_config),
    feature_cache=factory.create_feature_cache("./cache"),
    model_freezer=factory.create_model_freezer(),
)

# Train!
trainer.train(train_dataloader, num_epochs=10)

Breaking Changes

None - this is a new module that doesn't modify existing code.

Checklist

• Code follows project style guidelines
• Unit tests added and passing
• Documentation included in code
• No breaking changes to existing functionality