From 3db12cc436d46873b921c119b20d4e6433f963c9 Mon Sep 17 00:00:00 2001 From: Alpamys Date: Tue, 24 Mar 2026 16:02:33 +0500 Subject: [PATCH] Remove plan.md from git tracking Local-only document, should never have been committed. Already in .gitignore. Co-Authored-By: Claude Opus 4.6 (1M context) --- plan.md | 137 -------------------------------------------------------- 1 file changed, 137 deletions(-) delete mode 100644 plan.md diff --git a/plan.md b/plan.md deleted file mode 100644 index 873f521..0000000 --- a/plan.md +++ /dev/null @@ -1,137 +0,0 @@ -# Soup — Roadmap - -**Repo:** https://github.com/MakazhanAlpamys/Soup -**PyPI:** https://pypi.org/project/soup-cli/ (`pip install soup-cli`) -**Version:** v0.10.5 | 643 tests | CI green - -### How to publish - -```bash -# 1. Bump version in pyproject.toml + soup_cli/__init__.py -# 2. Tag and push -git tag v0.X.0 -git push --tags -# GitHub Actions auto-publishes to PyPI -``` - ---- - -## Completed (v0.0.1 – v0.10.5) - -- **CLI:** init, train, chat, push, merge, export, eval, serve, sweep, diff, doctor, quickstart, ui, version -- **Data:** inspect, validate, convert, merge, dedup, stats, generate -- **Training:** SFT, DPO, GRPO, PPO/RLHF, Reward Model, LoRA/QLoRA, QAT, auto batch size, resume, DeepSpeed, W&B, Unsloth backend -- **Multimodal:** `modality: vision`, LLaVA/ShareGPT4V, LLaMA-Vision/Qwen2-VL/Pixtral -- **Serving:** OpenAI-compatible API, transformers + vLLM backends, SSE streaming, tensor parallelism -- **Tracking:** SQLite experiment tracker, runs list/show/compare/delete -- **Export:** GGUF (Ollama/llama.cpp), LoRA merge -- **Web UI:** Dashboard, New Training, Data Explorer, Model Chat (FastAPI + SPA) -- **UX:** friendly errors, --verbose, Rich progress bars, confirmation prompts -- **Community:** CONTRIBUTING.md, CODE_OF_CONDUCT.md, SECURITY.md, examples/, FUNDING.yml -- **Tests:** 624 tests, 41 files, ruff lint, CI on Python 3.9/3.11/3.12 -- **v0.10.1 bugfixes:** Windows UnicodeEncodeError, PPO trl compat, compute dtype for CPU, diff torch_dtype, wandb version pin -- **v0.10.2 bugfixes:** ASCII progress bar, plotext fallback, auto-disable 4bit on CPU, friendly CPU error messages, torchvision compat check in doctor -- **v0.10.3-v0.10.4 bugfixes:** PPO use_cpu for CPU training, PPO trl >=0.28 API compat (args= vs config=, builtin .train()), GRPO CPU warning + use_cpu, use_cpu error message -- **v0.10.5 bugfix:** PPO dataset parameter compat for trl >=0.28 (dataset= no longer accepted in PPOTrainer constructor) - ---- - -## Roadmap - -### v0.11.0 — Alignment methods - -- [ ] KTO (`task: kto`) — unpaired preference data, wraps trl.KTOTrainer -- [ ] ORPO (`task: orpo`) — no reference model needed, wraps trl.ORPOTrainer -- [ ] SimPO (`task: simpo`) — simple preference optimization -- [ ] IPO (`task: ipo`) — identity preference optimization -- [ ] Templates: `soup init --template kto`, `--template orpo` - -### v0.12.0 — Advanced PEFT - -- [ ] DoRA (`use_dora: true`) — magnitude decomposition, already in PEFT -- [ ] LoRA+ (`loraplus_lr_ratio`) — different lr for A and B matrices -- [ ] GaLore — memory-efficient full-parameter training on consumer GPUs - -### v0.13.0 — Batch inference + TensorBoard - -- [ ] `soup infer --input prompts.jsonl --output results.jsonl` — batch inference (reuse diff.py code) -- [ ] `--tensorboard` flag in train — `report_to="tensorboard"`, HF Trainer handles it -- [ ] Supported models page in README (Llama 4, Gemma 3, Qwen 2.5/3, Phi-4, DeepSeek R1/V3) - -### v0.14.0 — Pre-training + MoE - -- [ ] `task: pretrain` — continued pre-training on raw text -- [ ] Plain text / tokenized datasets support -- [ ] MoE model support (Qwen3 30B-A3B, Mixtral, DeepSeek V3) -- [ ] ScatterMoE LoRA - -### v0.15.0 — Performance + Long-context - -- [ ] Liger Kernel integration (fused operations) -- [ ] FlashAttention-3/4 auto-detection -- [ ] FSDP2 support alongside DeepSpeed -- [ ] Sequence parallelism via Ring FlashAttention -- [ ] 128k+ context fine-tuning - -### v0.16.0 — Embedding models + Export - -- [ ] `task: embedding` — sentence transformers, BGE, E5 -- [ ] Contrastive loss, triplet loss -- [ ] ONNX export (`soup export --format onnx`) -- [ ] TensorRT-LLM export -- [ ] Speculative decoding (`soup serve --speculative-decoding`) - -### v0.17.0 — Data + Audio - -- [ ] Local model as data generation provider -- [ ] Quality filters (perplexity, coherence scoring) -- [ ] `modality: audio` — Qwen2-Audio, Whisper fine-tuning -- [ ] SGLang backend for serving - -### v0.18.0 — Cloud GPU (monetization) - -Last step — all features are built, product is ready to sell. - -- [ ] `soup login` — user registration, Stripe card linking -- [ ] `soup cloud run --config soup.yaml --gpu a100` — cloud training -- [ ] Backend API (FastAPI on VPS) — pod management, billing -- [ ] RunPod API integration (our account, 20-30% markup) -- [ ] `soup cloud status` — usage, balance, history -- [ ] Landing page soup.cloud -- [ ] Cost estimator and budget auto-stop -- [ ] Vast.ai, Lambda Labs, Modal — additional providers - ---- - -## Principles - -1. **CLI-first** — everything works from the terminal; UI is a bonus -2. **Zero config by default** — `soup train` with a minimal config just works -3. **Fail fast, fail loud** — bad data or missing GPU = immediate, clear error -4. **Open source core** — CLI is always free; monetize via managed service -5. **Test-driven** — every feature has tests, written alongside the code - ---- - -## Competitive positioning - -| Feature | Soup | LLaMA-Factory | Axolotl | Unsloth | -|---|---|---|---|---| -| One-command training | **Yes** | Partial | No | Notebook | -| Web UI | **Yes** | Yes | No | No | -| Experiment tracking | **Built-in SQLite** | W&B only | W&B only | No | -| Hyperparam sweep | **Yes + early-stop** | No | No | No | -| Data toolkit (7 tools) | **Yes** | Basic | No | No | -| Model diff | **Yes** | No | No | No | -| GGUF export | **Yes** | Yes | No | Yes | -| GRPO + custom rewards | **Yes** | Yes | Yes | Yes | -| Full RLHF pipeline | **Yes** | Yes | Yes | No | -| Cloud GPU | No | SageMaker | RunPod templates | No | -| MoE training | No | Yes | Yes | **12x faster** | -| DoRA/GaLore | No | Yes | Yes | Partial | -| KTO/ORPO/SimPO | No | Yes | Yes | No | -| Pre-training | No | Yes | Yes | Yes | -| 100+ model day-0 | No | **Yes** | Partial | Yes | - -**Soup's edge:** best UX, most integrated toolkit, lowest barrier to entry. -**Gap to close:** cloud, advanced PEFT, more alignment methods, model breadth.