Commit History
-
Add training checkpoints and dataset via LFS
1b20b9f
-
Add modular development docs under docs/ for learning and interviews
- docs/README.md + 01-architecture-overview.md: global map, data flow, naming - docs/core/: norm / rope / attention / mlp / block (code + formulas + interview points) - docs/models/: lm / vlm / vam overviews - docs/data/: dataset layout and loss-mask patterns - docs/training/: YAML config system + trainer module overview - docs/multimodal/: encoders/projectors/serve extension guide - docs/interview/: Q&A cheat sheet by topicc6b1b88 -
Fuse minimind/minimind-v/minimind-o into omni with layered architecture
Migrate the three MiniMind repos into the omni package using a clean layered design: - core/: split the monolithic Transformer kernel into norm/rope/attention/ mlp/block/model modules (replacing the old components.py) - models/: assemble core components into MiniMindForCausalLM (text), MiniMindVLM (vision), MiniMindOmni (speech/full-modal, with TalkerModule) - encoders/: SiglipVisionEncoder (vision/), SenseVoiceAudioEncoder (audio/) - projectors/: MMVisionProjector / MMAudioProjector bridging encoders to LLM - serve/: SileroVAD / RealtimeSession realtime speech session layer - trainers/ + datasets/ + utils/: ported train_sft/pretrain/dpo/etc plus VLM/Omni trainers, VLMDataset/OmniDataset, init_vlm/omni_model helpers - examples/: eval_llm/vlm/omni, omni_web_demo, serve_openai_api, convert_model - README + pyproject updated; all modules import-compile-smoke verified Co-Authored-By: opencode <noreply@opencode.ai>45cf443