[build-system]
requires = ["setuptools>=61", "wheel"]
build-backend = "setuptools.build_meta"

[project]
name = "unirl"
dynamic = ["version"]
description = "Unified multimodal RL training framework"
readme = "README.md"
# Floor follows upstream (3.12); cap because st_attn/vsa (sglang[diffusion])
# ship cp310-cp313 wheels only, with no sdist fallback.
requires-python = ">=3.12,<3.14"
license = { file = "LICENSE" }
authors = [
  { name = "unirl contributors" }
]
dependencies = [
  # Engine-agnostic only. torch intentionally absent: it enters via exactly
  # one engine extra below, which is what allows the per-engine CUDA stacks.
  "numpy>=1.24,<3",
  "ray[default]>=2.9",
  "diffusers>=0.38.0",  # LTX-2 needs 0.38: transformer.forward gained sigma/audio_sigma/isolate_modalities/use_cross_timestep, connector gained padding_side, calculate_shift now pins image_seq_len to max_image_seq_len. (>=0.37 was already required for Qwen-Image RoPE text-len-from-mask; 0.38 verified to keep that + every other model API unchanged.)
  "hydra-core>=1.3",
  "omegaconf>=2.3",
  # Match the engine stacks' 5.6.x baseline; BAGEL vendor smoke passed on 5.6.0.
  "transformers>=5.6,<5.7",
  "huggingface-hub>=0.34",
  # Vendored Boogu-Image DiT (unirl/models/boogu_image/vendor/) imports einops
  # in transformer_boogu.py / attention_processor.py / rope.py. Pure-python,
  # engine-agnostic.
  "einops>=0.7",
  # 0.20 floor: older peft imports transformers cache symbols (HybridCache)
  # removed in 5.x and fails at import against the transformers pin above.
  # exclude_modules support (>=0.14) used by LoRA/NFT adapter injection.
  "peft>=0.20",
  "safetensors>=0.4",
  "Pillow>=10",
  "requests>=2.31",
  "psutil>=5.9",
  "tensordict>=0.5",
]

# Parent/inline engines load this. Multi-stage StageDiffusionProc does not —
# VLLMOmniHijack.hijack() reinstalls the same hook in spawn children.
[project.entry-points."vllm_omni.general_plugins"]
unirl_capture_flush = "unirl.rollout.engine.vllm_omni.plugin:register_capture_flush"

[project.optional-dependencies]
# sglang pinned to the upstream release the UniRL sglang patch package
# (unirl/rollout/engine/sglang_diffusion/_patches/, LIN-365) is verified against
# (sde/cps parity + SD3 GRPO e2e). sglang >=0.5.11 is CUDA-13-native: its
# kernel stack (sglang-kernel, sgl-deep-gemm, cuda-python>=13) ships cu13
# binaries only, so this extra pins the matching torch +cu130 family and the
# nvidia-*-cu13 pip wheels supply libcudart/libnvrtc.so.13 — a pure-wheel,
# zero-compile install. On the driver-535 fleet CUDA-13 user code runs via
# NVIDIA's forward-compat layer: bake cuda-compat-13 into the image and point
# LD_LIBRARY_PATH at its compat dir (verified on H20 + driver 535). The sglang
# wheel itself is manylinux_2_34, so the image base must be glibc>=2.34
# (upgrade pending; verified meanwhile via a 2.34-prefixed python).
sglang = [
  "sglang[diffusion]==0.5.12.post1 ; sys_platform == 'linux'",
  # sglang pins a prerelease; a first-party declaration makes it "explicitly
  # requested" under uv's default prerelease policy (scoped, not global).
  "flash-attn-4>=4.0.0b9,<5 ; sys_platform == 'linux'",
  # torchvision/torchaudio publish no PyPI linux wheels in this era — pin the
  # flavor matching sglang 0.5.12.post1's torch==2.11.0 from the cu130 index
  # (cu13-native like the rest of sglang's stack; verified on-pod under the
  # forward-compat layer).
  "torch==2.11.0+cu130 ; sys_platform == 'linux'",
  "torchvision==0.26.0+cu130 ; sys_platform == 'linux'",
  "torchaudio==2.11.0+cu130 ; sys_platform == 'linux'",
  # Flash-Linear-Attention (FLA) fast path for Qwen3.5 hybrid attention layers
  # (3 GDN + 1 full per 4). Without it, transformers falls back to the torch
  # implementation ("The fast path is not available because one of the required
  # library is not installed"). The [conv1d] extra pulls causal-conv1d, the
  # [tilelang] extra is already satisfied transitively. Pin to the version
  # verified on-pod with sglang 0.5.12.post1 / torch 2.11.0+cu130.
  "flash-linear-attention[conv1d]==0.5.1 ; sys_platform == 'linux'",
  "av>=14.2,<19 ; sys_platform == 'linux'",
]
# CUDA 13: vllm >=0.26 wants a bare `torch==2.13.0` whose wheel is cu130, and a
# +cu129 pin would resolve anyway, linking a CUDA-12 torch under a CUDA-13 vllm.
# Prerelease, so installs need --prerelease=allow, plus sglang's compat layer.
vllm = [
  "vllm==0.27.0 ; sys_platform == 'linux'",
  "vllm-omni==0.27.0rc1 ; sys_platform == 'linux'",
  "torch==2.13.0+cu130 ; sys_platform == 'linux'",
  "torchvision==0.28.0+cu130 ; sys_platform == 'linux'",
  # torchaudio 2.11.0+cu130 is ABI-built against torch 2.11; there is no
  # torchaudio >=2.12 on the cu130 index, so this extra does not pin it.
  "av>=14.2,<19 ; sys_platform == 'linux'",
]
# Exact upstream source snapshot that the fail-closed FastVideo patch package is
# verified against. This is hao-ai-lab/FastVideo PR #1222's source commit, not
# Zcchill/FastVideo's UniRL-modified fork; UniRL re-hosts the required RL
# contracts under unirl/rollout/engine/fastvideo/_patches/.
fastvideo = [
  "fastvideo @ git+https://github.com/hao-ai-lab/FastVideo.git@2095477eac7e289c7a7ab13acb367ca60687c304 ; sys_platform == 'linux'",
]
train = [
  "wandb>=0.16,<0.20",
  "aiohttp>=3.9",
  # MathVerifyRewardScorer (unirl/reward/local/mathverify.py) — the paper's grader.
  "math-verify>=0.6",
]
cosmos3 = [
  "diffusers>=0.39",
]
infer = [
  "accelerate>=0.30",
  # Janus-Pro's vendored SigLIP implementation imports timm directly.
  "timm>=0.9.16",
]
eval = [
  "torchvision>=0.16",
  "paddlepaddle==3.2.2",
  "paddleocr==3.5.0",
  "python-Levenshtein>=0.27",
]
# VeOmni's torch-native distributed layer (FSDP2/EP parallelize), consumed
# exclusively through unirl.train.backend.veomni._compat — a selective import
# that never executes veomni/__init__.py or veomni/models/__init__.py, so the
# transformers-locked model zoo inside the wheel stays dormant. The shim's
# import closure is audited per release (tests/test_compat_import.py).
veomni = [
  "veomni==0.1.11 ; sys_platform == 'linux'",
]
dev = [
  "pytest>=7.4",
  "pytest-cov>=4.1",
  "ruff>=0.6",
  "pre-commit>=3.6",
]
# Offline converters under datasets/ (see datasets/README.md). None of this is
# imported by unirl/, which is why it is not a base dependency: you install it in
# a checkout to cook a dataset, then train without it. Reading a Hub dataset needs
# `datasets`; arxivqa_mc reads Vero's parquet shards through pyarrow directly
# (deliberately not via `datasets`, to keep the nested structs unreshaped); the
# LeRobot droid100 prep reads parquet metadata with pandas and decodes video with
# av. The av bound matches the sglang/vllm extras so installing both cannot
# conflict.
#
# torch is deliberately NOT here even though droid100 imports it (clip resize +
# .pt writes): adding it would put a second torch declaration outside the engine
# extras, and pinning exactly one is what lets each engine carry its own CUDA
# stack. droid100 is the only converter that needs it; datasets/droid100/README.md
# says so, and every other converter runs in a bare dataset-prep venv.
dataset-prep = [
  "datasets>=2.14",
  "pandas>=2",
  "pyarrow>=14",
  "av>=14.2,<19",
]

[tool.uv]
conflicts = [
  [{ extra = "sglang" }, { extra = "vllm" }],
]
# Local-version anchors only exist on the CUDA indexes; best-match lets every
# other package take its newest from PyPI while +cuXXX outranks same-release
# PyPI builds where it matters.
index-strategy = "unsafe-best-match"
# sglang 0.5.12.post1 pins transformers==5.6.0 but leaves its `kernels`
# integration unpinned; kernels >=0.15 changed LayerRepository's signature and
# breaks transformers 5.6.0 at import time. Hold it to transformers' own
# declared bound.
#
# transformers 5.6.0 accepts tokenizers<=0.23.0, but tokenizers 0.23.0(rc)
# dropped the `cls` kwarg from RobertaProcessing.__new__, so deserializing a
# saved CLIP tokenizer.json raises TypeError at rollout init (the SD3 text
# encoders, e2e-confirmed on H20). The sglang install's --prerelease=allow flag
# (see INSTALL.md) otherwise selects 0.23.0rc*; this override outranks it AND
# guards against 0.23.0 shipping stable. tokenizers 0.22.2 verified on both
# engine venvs.
override-dependencies = [
  "kernels>=0.12,<0.13",
  "tokenizers>=0.22,<0.23",
  # sglang[diffusion] hard-pins diffusers==0.37.0 (still pinned on sglang main),
  # colliding with the base diffusers>=0.38.0 floor (#91); override to the floor
  # so ".[sglang,...]" stays solvable (same tactic as kernels/tokenizers above).
  "diffusers>=0.38.0",
]
environments = [
  "sys_platform == 'linux' and platform_machine == 'x86_64'",
]

# Only fires if resolution falls back to the sdist: build-requires omits numpy,
# and the build-env torch must match the runtime flavor.
[tool.uv.extra-build-dependencies]
vllm = [
  "numpy",
  "torch==2.13.0+cu130",
]

# Non-explicit and ahead of the implicit PyPI default: participates in
# resolution for ALL packages, including the engines' transitive torch pins.
# Both extras are cu13, so a cu129 index here would mix flavors in unpinned deps.
[[tool.uv.index]]
name = "torch-cu130"
url = "https://download.pytorch.org/whl/cu130"

[tool.setuptools]
include-package-data = true

[tool.setuptools.dynamic]
version = { attr = "unirl.__version__" }

[tool.setuptools.packages.find]
where = ["."]
include = [
  "unirl",
  "unirl.*",
]

[tool.setuptools.package-data]
"unirl.models.janus_pro.vendor" = ["LICENSE-CODE", "VENDOR_COMMIT.txt"]

[tool.ruff]
line-length = 120
target-version = "py312"
# unirl-reward-service/ is a vendored standalone sub-project with its own pyproject and
# lint/test tooling; it is not governed by this repo's lint. Its vendored
# upstream model code carries
# semantically-meaningful trailing whitespace (VideoAlign prompt templates) that
# must not be auto-stripped.
extend-exclude = [
  "unirl-reward-service",
  "unirl/models/bagel/vendor",
  "unirl/models/janus_pro/vendor",
  "unirl/models/sensenova_u1/vendor",
]

[tool.ruff.lint]
select = ["E", "F", "W", "I"]
ignore = ["E501"]

[tool.ruff.lint.isort]
known-first-party = ["unirl"]

[tool.ruff.format]
quote-style = "double"