[project]
name = "halo"
version = "1.0.0"
description = "Halo is an open-source framework built by White Circle for training large language and multimodal models."
readme = "README.md"
requires-python = ">=3.12,<3.13"
license = "LicenseRef-Halo"
license-files = ["LICENSE", "APACHE-2.0.txt"]
authors = [
    { name = "Sergei Bratchikov", email = "sergey@whitecircle.com" },
]
classifiers = [
    "Development Status :: 5 - Production/Stable",
    "Intended Audience :: Science/Research",
    "Programming Language :: Python :: 3.12",
    "Topic :: Scientific/Engineering :: Artificial Intelligence",
]

dependencies = [
    "accelerate>=1.11.0,<1.12.0",
    "backoff>=2.2.1,<3.0.0",
    "bitsandbytes>=0.49.2,<0.50.0",  # 0.49.2 ships libbitsandbytes_cuda130.so; 0.46.x errors under cu13
    "boto3>=1.35.0,<2.0.0",
    "causal-conv1d>=1.6.0,<2.0.0",
    "clearml>=2.0.2,<3.0.0",
    "click>=8.1.0,<9.0.0",  # typer's CLI engine; src/cli.py raises click.UsageError directly
    "datasets>=4.8.5,<5.0.0",
    "ddgs>=9.10.0,<10.0.0",
    "faiss-cpu>=1.10.0,<2.0.0",
    "flash-linear-attention>=0.4.1,<0.5.0",  # DeltaNet kernels for Qwen3.5 hybrid attention
    "gradio>=6.15.1,<7.0.0",
    "gram-newton-schulz",
    "huggingface-hub>=1.0.0,<2.0.0",
    "kernels>=0.10.4,<0.11.0",
    "liger-kernel>=0.8.0,<0.9.0",
    "loguru>=0.7.2,<0.8.0",
    "matplotlib>=3.10.5,<4.0.0",
    "mcp>=1.22.0,<2.0.0",
    "msgspec>=0.19.0,<0.20.0",
    "openai>=1.44.1,<2.0.0",
    "pandas>=2.2.1,<3.0.0",
    "peft>=0.18.1,<0.19.0",
    "pydantic>=2.6.4,<3.0.0",
    "ray[default]>=2.52.0,<2.53.0",
    "ruamel-yaml>=0.19.1,<0.20.0",
    "s3fs>=2024.6.0",
    "sentence-transformers>=5.2.2,<6.0.0",
    "tabulate>=0.9.0,<0.10.0",
    "tokenizers>=0.23.1,<0.24.0",
    "torch>=2.11.0,<2.12.0",
    "torchvision>=0.26.0,<0.27.0",  # transformers fast image processors (VLM training); the Dockerfile drops the NGC copy and reinstalls the locked one
    "transformers>=5.16.1,<5.17.0",
    "trl>=1.6.0,<1.7.0",
    "typer>=0.25.1,<0.26.0",
    "wandb>=0.25.0,<0.26.0",
]

[project.urls]
Homepage = "https://github.com/whitecircle/halo"
Repository = "https://github.com/whitecircle/halo"
Issues = "https://github.com/whitecircle/halo/issues"

[project.scripts]
halo = "src.cli:main"

# vllm/sglang run in separate Docker containers (they carry their own torch/transformers stacks).
# See: Dockerfile.vllm, docker-compose.vllm.yml
# flash-attn is installed separately with --no-build-isolation (see Dockerfile).
# nvidia-cutlass-dsl 4.5.2 + quack-kernels 0.5.0 (Blackwell block-scaled MoE grouped-GEMM and
#   Muon Newton-Schulz kernels) are pinned by a Dockerfile pip-install over the lock's 4.4.2/0.4.1
#   (gram-newton-schulz's hard pin). Not declared here because the resolver can't co-resolve the
#   git-dep pin; the kernels degrade gracefully when absent.

[project.optional-dependencies]
flash-optimizers = [
    "flashoptim>=0.1.3,<0.2.0",
]
gigatoken = [
    "gigatoken>=0.9.0,<0.10.0",  # Rust bulk tokenizer backend for dataset preprocessing (pre-1.0 API)
]

# --- Dev tooling groups (PEP 735). Installed in-image so the image is the test runtime;
# `ty` is intentionally absent — type-checking is pyright (basic), run via Pylance / a Node install
# (see [tool.pyright]), not a pip tool. ---
[dependency-groups]
dev = [
    { include-group = "lint" },
    { include-group = "test" },
    { include-group = "profiling" },
]
lint = [
    "ruff>=0.9.9,<0.10.0",
]
# anyio and pillow are declared, not left transitive: anyio's pytest11 plugin is what applies
# ``pytest.mark.anyio`` (without it the async environment tests silently collect zero items),
# and tests/common/datasets.py builds VLM fixtures with PIL. Both arrive today only via httpx/openai
# and transformers, so a dependency bump that drops either edge would quietly shrink the suite.
test = [
    "pytest>=9.0.1,<10.0.0",
    "anyio>=4.13,<5.0",
    "pillow>=11.3,<12.0",
]
# Trace analysis + CPU flame graphs (scripts/profiling/, src/diagnostics/debugging.py).
# TraceLens is git-only — the PyPI name `tracelens` is an unrelated project; openpyxl backs its
# .xlsx reports; py-spy backs the stack dumps / CPU flame graphs.
profiling = [
    "tracelens",
    "openpyxl>=3.1,<4.0",
    "py-spy>=0.4.0,<0.5.0",
]

[tool.uv]
required-version = ">=0.10.5"
default-groups = ["dev"]
# DeepEP V2 (EPv2, NCCL Gin backend) is built from source against NCCL >= 2.30.4 (see
# Dockerfile). torch 2.11+cu130's wheel metadata hard-pins nvidia-nccl-cu13==2.28.9, so we
# OVERRIDE it — torch dynamically links libnccl.so.2 (ABI-stable across 2.x).
# EXACT pin, not a floor: all four build sites (both training images, Dockerfile.vllm,
# Dockerfile.sglang) install this one wheel, and under a floor whichever image is rebuilt last
# resolves whatever PyPI has moved to — the RL weight-sync ncclCommInitRank then fails on the skew.
# It must also stay at or above the DeepEP V2 floor (docker/nccl_pin.py MINIMUM) that deep_ep._C is
# compiled against. Bumping it means a `uv lock` AND a rebuild of all four images, not a lock edit.
override-dependencies = ["nvidia-nccl-cu13==2.31.2"]

[tool.uv.sources]
gram-newton-schulz = { git = "https://github.com/Dao-AILab/gram-newton-schulz.git" }
tracelens = { git = "https://github.com/AMD-AGI/TraceLens.git", tag = "v0.4.0" }
torch = { index = "pytorch-cu130" }

[[tool.uv.index]]
name = "pytorch-cu130"
url = "https://download.pytorch.org/whl/cu130"
explicit = true

[tool.ruff]
# Matches the direct upstreams (TRL + Transformers). The code is already ~p95=87 cols,
# so the format-only diff is small. Vendored / third-party trees are excluded.
target-version = "py312"
line-length = 119
# docker/vllm/parity/fixtures carries hub remote-code modules verbatim.
extend-exclude = ["vendors", "docker/vllm/parity/fixtures"]

[tool.ruff.lint]
# E,F,W,I = pycodestyle/pyflakes/isort; UP = pyupgrade; B = bugbear; C4 = comprehensions;
# SIM = simplify; T20 = no stray print() in src/; PLC0415 = no inline imports.
# The whole set is a blocking CI gate (.github/workflows/lint.yml) — `ruff check .` must stay clean.
# PLC0415 is a preview rule: `preview` + `explicit-preview-rules` arm it WITHOUT pulling in every
# other preview rule (without them it is selected but silently does nothing).
preview = true
explicit-preview-rules = true
select = ["E", "F", "I", "W", "UP", "B", "C4", "SIM", "T20", "PLC0415"]
ignore = [
    "E501",   # line length is enforced by the formatter, not the linter
    "UP038",  # isinstance(x, A | B) is slower than the tuple form at runtime (rule deprecated upstream)
]

[tool.ruff.lint.per-file-ignores]
# Tests: module-level setup imports (E402), asserts (S101), inline imports + prints are fine.
# E741 (`l`/`I`/`O` names — `for l in losses` is idiomatic here), B007 (unused loop var) and
# B023 (loop var captured in a closure — the closures are consumed within the same iteration in
# these test helpers) and SIM117 (nested `with patch(...)` is the idiomatic mock-stack form) are
# accepted in tests rather than churned across the suite.
"tests/**" = ["E402", "PLC0415", "T20", "E741", "B007", "B023", "SIM117"]
# Scripts are CLIs — prints are the interface.
"scripts/**" = ["T20"]
# Docs diagram generators: `from _theory_style import *` shares one matplotlib theme, and the
# matplotlib `use()`-before-`pyplot` idiom puts an import after a call.
"scripts/diagrams/*" = ["F403", "F405", "E402"]
# Build-time check script: printing each result is its interface, and it imports vLLM (present only
# in the server image) after setting sys.path, so the imports cannot be module-top.
"docker/vllm/plugins/verify_gptoss_plugins.py" = ["T20", "E402"]
# Build-time config-parity gate + its fixture generator: printed results are the interface.
"docker/vllm/parity/*.py" = ["T20"]
# Build-time pin resolver: the printed version IS its interface (a Dockerfile reads it off stdout).
"docker/nccl_pin.py" = ["T20"]
# Build-time patch script for the SGLang image: the build log is where the rewrites are recorded.
"docker/sglang/patches/patch_sglang_weight_updates.py" = ["T20"]
# Gradio's documented progress idiom is `def fn(..., progress=gr.Progress())` — the call in the
# default is how Gradio injects the tracker, so B008 (call in arg default) does not apply here.
# Package re-export shims: re-exported names (F401) and `import *` aggregation (F403/F405).

[tool.ruff.lint.isort]
# Pin import grouping (I001) so it is deterministic regardless of stray directories at the repo root:
# a gitignored `wandb/` run dir and the tracked `launcher-configs/accelerate/` config dir would otherwise make ruff
# infer those third-party packages as first-party locally but not in CI.
known-first-party = ["src", "tests"]
known-third-party = ["wandb", "accelerate"]

[tool.pyright]
# ruff's complement: type-aware checks (call args, None access, attribute typos, bad returns).
# Single source of truth shared with Pylance (.vscode/settings.json). BASIC mode + non-blocking to
# start (53k LOC of dynamic torch/DTensor code); tighten over time like the ruff lint-clean milestone.
# Full checks need the deps installed, so run pyright IN the image (the host has no torch); a bare-host
# run only checks our-code-to-our-code with third-party imports relaxed.
typeCheckingMode = "basic"
pythonVersion = "3.12"
exclude = ["vendors", "**/__pycache__", ".venv"]
reportMissingTypeStubs = false      # third-party libs without py.typed shouldn't spam diagnostics
reportMissingModuleSource = false   # host runs may not have the torch env installed

[tool.pytest.ini_options]
# CPU tests run inside the image (they import torch); GPU tests are torchrun scripts
# launched as subprocesses by tests/gpu/conftest.py from tests/gpu/manifest.py — pytest
# never imports them. Markers are registered in tests/conftest.py (from manifest.ALL_MARKERS)
# so --strict-markers stays a single source of truth.
testpaths = ["tests"]
python_files = ["test_*.py"]
addopts = "--strict-markers -ra"
norecursedirs = ["*.egg", ".*", "build", "dist", "vendors"]

[build-system]
requires = ["hatchling>=1.27.0"]
build-backend = "hatchling.build"

[tool.hatch.build.targets.wheel]
packages = ["src"]