Add NaN guard to audio code generation path

Revert camera prompt changes; add NaN guard and CUDA sync; add engineering challenges doc

Cast logits to float32 before softmax/sampling to prevent fp16 overflow

Fix from_pretrained: use SenseVoice/Siglip encoder wrappers instead of raw models

Convert omni-o.pth to HF format + add from_pretrained support

Add omni_o_call.py: real-time voice call server for omni-o checkpoint

Fix generate() to return last text tokens instead of (None, audio_frame) tuple

fix VAM: funasr+pyarrow deadlock, librosa+numpy compat, reduce iter_batches batch_size

feat: add VAM SFT mini training (T2A Stage 1)

feat: add VLM SFT mini configs, resume training, dataset streaming, and eval fixes

refactor: rename VLM configs, add VLM SFT configs, and improve model loading

refactor: reorganize dataset into lm/vlm/vam subdirectories

docs: update README and docs for model_type rename, RoPE guard fix, config paths

rename minimind→omni in model_type, eval/convert scripts, and docs

reorganize configs into subdirs lm/vlm/vam, remove runs/, add interview docs for pretrain & full_sft

cleanup: remove deprecated doc files, update src training utils

Add local training logger that writes to {save_dir}/{save_weight}.log

Add tokenizer training script and save trained tokenizer into checkpoint dir

Add modular development docs under docs/ for learning and interviews

Flatten src layout: drop omni namespace, modules live directly under src/

Restructure repo: rename classes, split datasets, organize trainers by modality, add YAML config + root trainer entry points

Rename core MiniMindBlock to Block for modality-agnostic naming

Move LoRA into models/lm to match modality-scoped layout

Move MiniMindModel backbone into models/lm, keep core pure-components only

Restructure models into lm/vlm/vam sub-packages by modality

Fuse minimind/minimind-v/minimind-o into omni with layered architecture

Initial commit: omni package with src layout and pyproject config