pyproject.toml

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

[project]
name = "OpenThoughts-Agent"
version = "0.1.0"
description = "OT-Agent HPC launch and data tooling"
requires-python = ">=3.10"
dependencies = [
    # Core HPC + data infrastructure deps
    "pydantic>=2.0.0,<3.0.0",
    "pyyaml",
    "omegaconf",
    "wandb",
    "bs4",
    "numpy<=2.26.0",
    "huggingface_hub>=0.20.0,<1.0.0",
    "datasets>=2.0.0",
    "supabase>=2.22.3",
    "python-dotenv>=1.0.0",
    "harbor[daytona] @ git+https://github.com/laude-institute/harbor.git@penfever/temp-override",
    "dynamic-semaphore @ git+https://github.com/penfever/dynamic-semaphore",
    "google-cloud-storage",
    "h5py",
    "certifi",
    "rapidfuzz",
    "huggingface_hub",
    "uv>=0.4.17",
    "socksio",  # SOCKS5 proxy support for httpx (needed for JSC SSH tunnels)
]

[project.optional-dependencies]
datagen-cpu = [
    "google-cloud-aiplatform>=1.67.0",
    "google-genai>=0.1.0",
    "accelerate==1.12.0",
    "trufflehog>=2.2.0",
]
datagen = [
    "ray[default]>=2.50.0",
    "torch-c-dlpack-ext==0.1.4",
    "google-cloud-aiplatform>=1.67.0",
    "google-genai>=0.1.0",
    "vllm[flashinfer]>=0.11.2,<=0.16.0; platform_system != 'Darwin'",
    "transformers==4.57.3",
    "torch>=2.9.0,<=2.10.0; platform_system != 'Darwin'",
    "accelerate==1.12.0",
    "trufflehog>=2.2.0",
]
# ROCm/AMD GPU variant for datagen (e.g., OLCF Frontier with MI250X)
# Install with:
#   uv pip install -e ".[datagen-rocm]" \
#     --extra-index-url https://wheels.vllm.ai/rocm/0.14.0/rocm700 \
#     --prerelease=allow
datagen-rocm = [
    "ray[default]>=2.50.0",
    "torch-c-dlpack-ext==0.1.4",
    "google-cloud-aiplatform>=1.67.0",
    "google-genai>=0.1.0",
    "transformers==4.57.3",
    # PyTorch ROCm 6.4 - must install with --index-url https://download.pytorch.org/whl/rocm6.4
    "torch==2.8.0",
    "torchvision==0.23.0",
    "torchaudio==2.8.0",
    # vLLM ROCm wheel (requires --extra-index-url https://wheels.vllm.ai/rocm/0.14.0/rocm700 --prerelease=allow)
    "vllm==0.14.0+rocm700",
    # Note: flash-attn is CUDA-only; PyTorch native attention used instead
]
datagen-swesmith = [
    "swesmith",
]
cloud = [
    "skypilot>=0.6.1",
    "google-api-python-client",  # GCP support for SkyPilot
    "boto3",  # S3-compatible object storage (MinIO, AWS S3)
    "s3fs",   # Filesystem interface for S3/MinIO
]
cloud-aws = [
    "skypilot>=0.6.1",
    "boto3",
]
cloud-lambda = [
    "skypilot>=0.6.1",
    # Lambda Cloud only needs skypilot + API key configuration
]
cloud-vast = [
    "skypilot>=0.6.1",
    "vastai-sdk>=0.1.12",
]
cloud-kubernetes = [
    "skypilot>=0.6.1",
    # Kubernetes needs kubectl + kubeconfig, no extra Python deps
]
cloud-all = [
    "skypilot>=0.6.1",
    "google-api-python-client",
    "boto3",
    "vastai-sdk>=0.1.12",
]
rl = [
    # SkyRL training framework - install with vllm extra for inference
    # Uses mlfoundations fork with custom patches
    "skyrl[vllm] @ git+https://github.com/mlfoundations/skyrl.git",
    "Jinja2",  # For chat templates
]
# ROCm/AMD GPU variant for RL (e.g., OLCF Frontier with MI250X)
# Install with: uv pip install -e ".[rl-rocm]" --index-url https://download.pytorch.org/whl/rocm6.4
# Or use: ./hpc/setup_rl_env.sh --rocm
rl-rocm = [
    # SkyRL without vllm extra (vLLM ROCm support is experimental)
    "skyrl @ git+https://github.com/mlfoundations/skyrl.git",
    "Jinja2",  # For chat templates
    # PyTorch ROCm 6.4 - must install with --index-url https://download.pytorch.org/whl/rocm6.4
    "torch==2.8.0",
    "torchvision==0.23.0",
    "torchaudio==2.8.0",
    # Note: flash-attn is CUDA-only; PyTorch native attention used instead
]
# IMPORTANT: RL and datagen extras are INCOMPATIBLE due to version conflicts:
#   - RL (SkyRL): torch==2.8.0, vllm==0.11.0, Python 3.12 required
#   - datagen:    torch==2.9.0, vllm==0.11.2, Python 3.10+
# The RL environment must be set up separately using: ./hpc/setup_rl_env.sh
# The RL launcher will automatically use the separate environment.
# See hpc/rl_launch_utils.py for details.
# SFT for Qwen3.5 models (requires transformers v5+ for hybrid GDN architecture)
# IMPORTANT: Incompatible with datagen/RL extras (transformers 4.x vs 5.x).
# Must be installed in a separate environment.
# Install with:
#   uv pip install -e ".[sft-qwen35]"
sft-qwen35 = [
    "transformers>=5.3.0",
    "torch>=2.6.0",
    "accelerate>=1.3.0",
    "datasets>=2.16.0",
    "peft>=0.14.0",
    "trl>=0.8.6",
    "tokenizers>=0.19.0",
    "deepspeed>=0.16.0",
    "flash-attn>=2.7.0; platform_system != 'Darwin'",
    "torchao>=0.8.0",
    "liger-kernel>=0.5.0",
    "trufflehog>=2.2.0",
]
# Harbor environment backends
curator = [
    "bespokelabs-curator>=0.1.0",
]
harbor-docker = [
    "docker>=7.0.0",
]
harbor-modal = [
    "modal>=0.60.0",
]
harbor-all = [
    "docker>=7.0.0",
    "modal>=0.60.0",
]
# Beam cluster setup (self-hosted Beta9 on GKE)
beam = [
    "requests>=2.31.0",
    "beta9>=0.1.0",
]

[tool.setuptools.packages.find]
where = ["."]
include = [
    "data",
    "data.*",
    "database",
    "database.*",
    "sft",
    "sft.*",
    "eval",
    "eval.*",
    "hpc",
    "hpc.*",
    "rl",
    "rl.*",
    "scripts",
    "scripts.*",
    "train",
    "train.*",
]
exclude = [
    "docs*",
    "notes*",
    "wheels*",
    "tmp*",
]

[tool.uv.pip]
# Default: CUDA 12.8 PyTorch index
# For ROCm/AMD GPUs (e.g., Frontier), use instead:
#   uv pip install ... --index-url https://download.pytorch.org/whl/rocm6.4
extra-index-url = ["https://download.pytorch.org/whl/cu128"]

[tool.uv]
index-strategy = "unsafe-best-match"

# =============================================================================
# ROCm Installation Notes (for AMD GPUs like Frontier MI250X)
# =============================================================================
# ROCm extras (datagen-rocm, rl-rocm) require the ROCm PyTorch index:
#
#   # For datagen on ROCm:
#   uv pip install -e ".[datagen-rocm]" --index-url https://download.pytorch.org/whl/rocm6.4
#
#   # For RL on ROCm (preferred method is setup script):
#   ./hpc/setup_rl_env.sh --rocm
#
# See: https://docs.olcf.ornl.gov/software/analytics/pytorch_frontier.html
# =============================================================================