pyproject.toml

[build-system]
requires = ["setuptools>=68"]
build-backend = "setuptools.build_meta"

[project]
name = "shadowlm"
version = "0.4.11"
description = "ShadowLM Trainer — fine-tune any open model, with any method, on any hardware, for any harness."
readme = "README.md"
license = "MIT"
license-files = ["LICENSE"]
requires-python = ">=3.10"
authors = [
    {name = "Lyzr Research Labs"},
]
maintainers = [
    {name = "Khush Patel", email = "khush@lyzr.ai"},
]
keywords = [
    "fine-tuning", "llm", "lora", "qlora", "dpo", "grpo", "rlhf",
    "mlx", "pytorch", "peft", "agents", "training",
]
classifiers = [
    "Development Status :: 4 - Beta",
    "Intended Audience :: Developers",
    "Intended Audience :: Science/Research",
    "Operating System :: OS Independent",
    "Programming Language :: Python :: 3",
    "Programming Language :: Python :: 3.10",
    "Programming Language :: Python :: 3.11",
    "Programming Language :: Python :: 3.12",
    "Programming Language :: Python :: 3.13",
    "Programming Language :: Python :: 3.14",
    "Topic :: Scientific/Engineering :: Artificial Intelligence",
]
# Batteries included: `pip install shadowlm` pulls everything needed to train —
# the torch/HuggingFace stack, retrieval, and the CLI. mlx (the Apple-Silicon dev
# loop) is auto-installed only on arm64 macOS via an environment marker, so the
# same install works on Linux/CUDA and on a Mac. Specialized backends — fused
# kernels and the VERL RL engine — stay opt-in extras.
dependencies = [
    # training — the CUDA/CPU production path (torch + HuggingFace)
    "torch>=2.3",
    # Ceilings: transformers 4.57+/5.x and trl 0.25+/1.x changed SFTTrainer loss
    # internals (they read outputs.logits, which liger's FLCE nulls) — pin to the
    # verified window until the code is forward-ported.
    "transformers>=4.43,<4.57",
    "trl>=0.9,<0.25",
    "peft>=0.12",
    "accelerate>=0.33",
    "datasets>=2.20",
    # 8-bit / fused optimizer (default optim=adamw_8bit) — NVIDIA Linux only;
    # marker keeps macOS/arm64 (mlx) and non-Linux installs clean
    "bitsandbytes>=0.43; sys_platform == 'linux'",
    # mixture of retrieval experts (method="more") — embeddings + faiss index
    "sentence-transformers>=3.0",
    "faiss-cpu>=1.8",
    # the shadowlm CLI (Typer + Rich)
    "typer>=0.12",
    "rich>=13",
    "pyyaml>=6",
    # Apple-Silicon dev loop — installed automatically on arm64 macOS only
    "mlx-lm>=0.20; sys_platform == 'darwin' and platform_machine == 'arm64'",
    "mlx-lm-lora>=2.0; sys_platform == 'darwin' and platform_machine == 'arm64'",
]

[project.urls]
Homepage = "https://github.com/open-gitagent/shadowLM"
Repository = "https://github.com/open-gitagent/shadowLM"
Issues = "https://github.com/open-gitagent/shadowLM/issues"

[project.scripts]
shadowlm = "shadowlm._cli_entry:main"

[project.optional-dependencies]
# Fused Triton kernels (Liger, Apache-2.0) — extra speed/VRAM on NVIDIA.
# Optional: the shadow accelerator uses them when present, no-ops otherwise.
kernels = [
    "liger-kernel>=0.4",
]
# Production multi-GPU RL — the VERL engine (vLLM rollouts + FSDP). NVIDIA GPUs.
# Opt-in and heavy: `backend="verl"` for cluster-scale GRPO.
verl = [
    "verl>=0.4",
]
# Back-compat aliases — the training stack now ships in the base install, so
# these resolve to nothing but keep older `shadowlm[all]` / `[torch]` commands working.
all = []
mlx-all = []
torch = []
mlx = []
preference = []
retrieval = []
cli = []

[tool.setuptools.packages.find]
include = ["shadowlm*"]

[tool.setuptools.package-data]
shadowlm = ["_assets/*.png", "_static/index.html", "_static/assets/*", "_samples/*.jsonl"]