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

[project]
name = "lightnav"
version = "0.1.0"
description = "Standalone inference for trajectory-token vision-language navigation checkpoints (Qwen3-VL based)"
readme = "README.md"
license = { file = "LICENSE" }
requires-python = ">=3.11,<3.12"
dependencies = [
    # Core: HF backend + WebSocket server + trajectory-token decode.
    # transformers is pinned exactly: the data processor subclasses Qwen3VLProcessor
    # and the vLLM embedding patch targets transformers 5.8's Qwen3-VL internals.
    #
    # torch is deliberately a floor, not a pin, but note WHICH build pip picks: PyPI's
    # torch 2.10.0 is a cu12.8 wheel. Its bundled NVRTC 12.8 does not know
    # `compute_103`, so on sm_103 parts (B300 / B30Z) the `hf` backend dies where
    # transformers' Qwen3-VL vision tower JITs a reduction for
    # `image_grid_thw.prod(-1)` ("nvrtc: error: invalid value for --gpu-architecture").
    # Install the cu12.9 wheels on such GPUs -- see "Installation" in the README.
    "torch>=2.9",
    # torchvision is required by transformers 5.8's Qwen3-VL video processor.
    "torchvision>=0.24",
    "transformers==5.8.0",
    "tokenizers>=0.22,<0.23",
    "safetensors>=0.7",
    "numpy>=2.0",
    "pillow>=11.0",
    "pyyaml>=6.0",
    "websockets>=12.0",
]

[project.scripts]
lightnav-predict = "lightnav.cli.predict:main"
lightnav-serve = "lightnav.serving.ws_server:main"
lightnav-ws-client = "lightnav.cli.ws_client:main"
lightnav-eval-habitat = "lightnav.cli.eval_habitat:main"
lightnav-eval-merge = "lightnav.cli.eval_merge:main"
lightnav-render = "lightnav.cli.render:main"

[project.optional-dependencies]
# In-process vLLM backend (`--backend vllm_local`). The patch in
# inference/vllm_utils.py binds vLLM 0.19.x private APIs (BaseRenderer.
# _process_multimodal + Qwen3VL.get_mrope_input_positions) and breaks on other
# minor versions, so the pin is exact rather than a range. vllm 0.19.1 pulls
# torch 2.10. Verified end-to-end against vllm==0.19.1 / torch 2.10.0 /
# transformers 5.8.0. A runtime guard (vllm_utils._assert_vllm_version) fails loudly
# on a mismatched version.
#
# nvidia-cutlass-dsl has to be pinned HERE, not left to vllm: vllm 0.19.1 only asks
# for `nvidia-cutlass-dsl>=4.4.0.dev1`, and because that specifier mentions a
# pre-release pip is allowed to resolve the newest dev build (4.8.0.dev0 as of
# 2026-08), which removed `cutlass.cute.core.ThrMma`. vLLM's flash-attn cute path
# imports quack-kernels, and every quack release up to 0.6.4 still references that
# symbol, so a fresh install fails with
#   AttributeError: module 'cutlass.cute.core' has no attribute 'ThrMma'
# while the engine is starting, i.e. before any inference is possible. Upgrading
# quack does not help -- the DSL is what has to be held back.
vllm = ["vllm==0.19.1", "nvidia-cutlass-dsl==4.5.2"]
# FlashAttention-2 for the HF backend (falls back to sdpa when absent).
flash = ["flash-attn>=2.7"]
# mp4 / frame-directory decoding for lightnav-predict and lightnav-ws-client, and the
# visualisation module (lightnav.viz): lightnav-render, `lightnav-eval-habitat --save_video`.
# Recording with `lightnav-serve --record_dir` needs only the core install.
video = ["opencv-python-headless>=4.10", "imageio>=2.37", "imageio-ffmpeg>=0.6"]
# Habitat evaluation client (lightnav-eval-habitat talks to the habitat server over ZMQ).
# habitat-sim / habitat-lab themselves live in a separate conda env; see habitat_server/.
habitat = ["pyzmq>=25"]
# CPU-only unit tests (scheduler, decode parsing, protocol, sample builders).
test = ["pytest>=8", "pytest-asyncio>=0.24"]

[tool.setuptools.packages.find]
where = ["src"]

[tool.ruff]
line-length = 100
target-version = "py311"

[tool.ruff.lint]
select = ["E", "F", "I"]   # pyflakes + pycodestyle + isort
ignore = ["E501"]          # line length is advisory; the 100-col limit is not enforced

[tool.pytest.ini_options]
testpaths = ["tests"]
markers = [
    "gpu: needs a real GPU / torch CUDA / model weights (deselected on CPU with -m 'not gpu')",
    "video: needs ffmpeg/imageio video dependencies",
]
asyncio_mode = "auto"