pyproject.toml

[project]
name = "open-trajectory-gym"
version = "0.1.0"
description = "Open-source RL training gym for long-horizon tool-using agents: TRL SFT + SkyRL GRPO + GEPA"
readme = "README.md"
license = {text = "MIT"}
requires-python = ">=3.11"
authors = [
    {name = "Open Trajectory Gym Team"},
]

# Core dependencies — only what the CLI, data tools, reward, formatters,
# and ToolExecutor need at the top level.
# Heavy ML deps (transformers, torch, etc.) live in optional extras.
dependencies = [
    "pyyaml",
    "pydantic>=2.0.0",
    "jsonlines",
]

[project.optional-dependencies]
# LoRA merge + GGUF export (trajgym-train merge, trajgym-export)
# Qwen3.5-27B requires transformers>=5.2.0 for Qwen3_5ForConditionalGeneration.
merge = [
    "torch>=2.5.0",
    "transformers>=5.2.0",
    "peft>=0.15.0",
    "accelerate>=1.4.0",
]
# Stage 1: TRL SFT
sft = [
    "trl>=0.28.0",
    "torch>=2.5.0",
    "transformers>=5.2.0",
    "peft>=0.15.0",
    "accelerate>=1.4.0",
    "datasets>=3.0.0",
    "bitsandbytes>=0.45.0",
    "wandb",
]
# Stage 2: SkyRL GRPO (online RL)
#
# skyrl-train is NOT on PyPI. Install from our patched fork:
#   git clone -b open-ctf/v0.3.1-patched https://github.com/westonbrown/SkyRL.git skyrl
#   sed -i 's/requires-python = "==3.12\.\*"/requires-python = ">=3.11"/' \
#       skyrl/skyrl-train/pyproject.toml
#   uv pip install -e skyrl/skyrl-train --no-deps
#
# Version conflict: vLLM 0.16 pins transformers<5 and peft<0.19 pins an
# eager HybridCache import. After installing, force the correct versions:
#   uv pip install 'transformers>=5.2.0' 'huggingface-hub>=1.4' --no-deps
# Then patch peft's HybridCache import (see docker/patches/apply_all_patches.sh).
#
# Qwen3.5 linear-attention: missing `fla`/`causal_conv1d` triggers torch
# fallback kernels that cause illegal memory access crashes in online GRPO.
grpo = [
    "skyrl-gym>=0.1.0",
    "vllm>=0.16.0",
    "ray[default]>=2.40.0",
    "torch>=2.5.0",
    "transformers>=5.2.0",
    # Required by transformers loss import path (object-detection loss module
    # imports scipy.optimize.linear_sum_assignment at import time).
    "scipy>=1.16.0",
    # Qwen3.5 linear-attention fast path dependencies (module imports: fla, causal_conv1d)
    # Pin known-good versions for SkyRL GRPO on CUDA 12.x hosts.
    "flash-linear-attention==0.4.1; sys_platform == 'linux'",
    "causal-conv1d==1.6.0; sys_platform == 'linux'",
    "peft>=0.15.0",
    "accelerate>=1.4.0",
    "omegaconf",
    "hydra-core",
    "wandb",
    "jmespath",
]
# Stage 2 alias: semantic name for online reinforcement learning.
online-rl = [
    "skyrl-gym>=0.1.0",
    "vllm>=0.16.0",
    "ray[default]>=2.40.0",
    "torch>=2.5.0",
    "transformers>=5.2.0",
    "scipy>=1.16.0",
    "flash-linear-attention==0.4.1; sys_platform == 'linux'",
    "causal-conv1d==1.6.0; sys_platform == 'linux'",
    "peft>=0.15.0",
    "accelerate>=1.4.0",
    "omegaconf",
    "hydra-core",
    "wandb",
    "jmespath",
]
# Stage 3: GEPA prompt evolution
gepa = [
    "dspy>=3.1.0",
    "gepa>=0.0.26",
]
# Agent runner (wraps BoxPwnr)
agent = [
    "litellm",
]
# Development and testing
dev = [
    "pytest",
    "ruff",
]

[project.scripts]
trajgym-train = "trajgym.cli.train:main"
trajgym-convert = "trajgym.cli.convert_traces:main"
trajgym-split = "trajgym.cli.split_dataset:main"
trajgym-agent = "trajgym.cli.run_agent:main"
trajgym-eval = "trajgym.cli.evaluate:main"
trajgym-validate = "trajgym.cli.validate_pipeline:main"
trajgym-export = "trajgym.cli.export_gguf:main"
trajgym-challenges = "trajgym.cli.challenges:main"
trajgym-synthetic-data = "trajgym.cli.synthetic_data_generation:main"
trajgym-trajectories = "trajgym.cli.trajectories:main"
trajgym-generate-rl = "trajgym.cli.generate_online_rl:main"
trajgym-target-map = "trajgym.cli.generate_target_map:main"
[build-system]
requires = ["hatchling"]
build-backend = "hatchling.build"

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

[dependency-groups]
dev = [
    "pytest",
    "ruff",
]

[tool.uv]
# Force transformers>=5.2.0 despite vLLM 0.16 pinning <5 and SkyRL pinning <5.
# Qwen3.5 requires 5.2.0+ (Qwen3_5ForConditionalGeneration). Both vLLM and
# SkyRL work fine at runtime with transformers 5.2.0 — the pins are overly
# conservative. Same for huggingface-hub (transformers 5.2.0 needs >=1.4).
override-dependencies = [
    "transformers>=5.2.0",
    "huggingface-hub>=1.4.0",
]
# Keep Qwen3.5 linear-attention deps at known-good minimums for Linux GRPO
# envs, so fresh resolves don't silently downgrade below crash-fixed baseline.
constraint-dependencies = [
    "flash-linear-attention==0.4.1; sys_platform == 'linux'",
    "causal-conv1d==1.6.0; sys_platform == 'linux'",
]


[tool.pytest.ini_options]
testpaths = ["tests"]
python_files = ["test_*.py"]
pythonpath = ["src"]
norecursedirs = ["references", "benchmarks", "outputs", ".venv", "venv"]