Fix Phase 2: fusion layer processes text-only via learnable residual gate for gradient flow

Fix: register LLM as nn.Module submodule so optimizer finds LoRA params

Fix extra_special_tokens: list to dict for transformers 4.55

Add requirements-training.txt for MI300X Docker

Remove HSA_OVERRIDE_GFX_VERSION - ROCm 7.0 native MI300X support

Track large tokenizer files with Git LFS for HuggingFace

Add GPU diagnostic script, fix architecture loading with low_cpu_mem_usage and sync

Fix setup_mi300x.sh for Docker container on MI300X droplet

Fix setup_mi300x.sh step 2 + add project context summary

Fix train.py: mem -> memory on line 225

Day 3 COMPLETE: Full model architecture

Day 2 COMPLETE: 1.48M examples processed, 6GB dataset, WebSight done

Day 1 Complete: Tokenizer setup — Qwen2.5-Coder-7B base + 22 MINDI special tokens (vocab 151,685), wrapper class, full format test

feat: initial project scaffold for MINDI 1.5 Vision-Coder