elizaos--eliza/packages/training/pyproject.toml

[project]
name = "eliza-training"
version = "0.1.0"
description = "Fine-tuning pipeline for Eliza-1 on Eliza-native Vercel AI SDK trajectory rows."
readme = "README.md"
requires-python = ">=3.11,<3.13"

dependencies = [
  # 1.13.0 has a one-file LFS upload typo in _upload_lfs_files
  # (`filtered_tool_calls` vs `filtered_actions`) that breaks large GGUF
  # publishes when Xet is disabled. Keep Hub publishing on the fixed line.
  "huggingface-hub>=1.14.0",
  "datasets>=3.2.0",
  "pandas>=2.2.0",
  "pyyaml>=6.0.2",
  "jinja2>=3.1.0",
  "pyarrow>=18.0.0",
  # NB: anthropic is also a top-level dep because data-synth scripts call
  # the Claude API. The reward-fn AI judge is gated by an env var, so the
  # scripts/eliza_reward_fn.py path never imports anthropic unless asked.
  "anthropic>=0.40.0",
  "tiktoken>=0.8.0",
  "argostranslate>=1.11.0",
  # Argos Translate 1.11.0 requires this exact Stanza release. Declaring the
  # transitive pin keeps uv/Dependabot from proposing an unsatisfiable upgrade.
  "stanza==1.10.1",
  "httpx>=0.27.0",
  # PolarQuant's Lloyd-Max centroid generation imports scipy.stats.norm
  # directly; keep it in the base training environment so smoke tests and
  # dry-run recipe generation do not depend on optional train extras.
  "scipy>=1.17.0",
  # Property-based testing for the privacy filter contract. Lightweight
  # pure-Python dep; pinned only at the floor to match the rest of the
  # test stack (pytest>=8).
  "hypothesis>=6.100.0",
]

[project.optional-dependencies]
train = [
  # torch>=2.5,<2.12 — uv picks the wheel matching whatever vllm 0.20.x
  # resolves to (torch 2.11+cu130 in practice) when --extra serve is
  # also selected (smoke), and 2.10+cu126 when train is alone (Vast
  # SFT path). Both wheels run on Vast Blackwell driver 570.x.
  "torch>=2.5.0,<2.13",
  "transformers>=4.46.0",
  "accelerate>=1.1.0",
  "peft>=0.14.0",
  "trl>=0.13.0",
  "datasets>=3.2.0",
  "wandb>=0.18.0",
  "apollo-torch>=1.0.3",
  # TurboQuant (arXiv:2504.19874) — runtime KV-cache compression.
  # PyPI distribution name is `turbokv`; the import name is `turboquant`.
  # This is the pure-PyTorch reference path (slow but architecture-portable).
  "turbokv>=0.1.0",
  # fused-turboquant — Triton-kernel implementation of the same TurboQuant
  # scheme. Same memory win as turbokv 0.1.0; meant to recover the ~5x
  # throughput regression from the pure-PyTorch encode/decode loop. Requires
  # head_dim ∈ {64, 128, 256}; auto-falls-through to turbokv on architectures
  # the patcher rejects. JIT compile of Triton's CUDA helper needs system
  # `python3.x-dev` headers — see scripts/quantization/README.md.
  "fused-turboquant>=0.1.0",
  # QJL kernel build deps. The kernel itself is *vendored* under
  # scripts/quantization/qjl/ from
  # https://github.com/amirzandieh/QJL @648b3641 (Apache 2.0). It is NOT
  # pip-installable — the user must run
  #   `cd scripts/quantization/qjl && python setup.py build_ext --inplace`
  # after installing nvcc + python-dev system packages
  # (`sudo apt install nvidia-cuda-toolkit python3.12-dev`). pybind11
  # is needed at C++ template-binding time. There is no pyqjl package
  # on PyPI; do not add one.
  "pybind11>=2.12.0",
  # Liger kernel — fused chunked cross-entropy + RMSNorm/SwiGLU/RoPE
  # kernels. Cuts the fp32-logits transient at training time (Gemma 4
  # vocab=262k makes this dominant) so we can push local training seq_len
  # from ~2k to ~8k on the same 16 GB budget.
  "liger-kernel>=0.5.0",
  # GGUF Q4_K_M quantization (scripts/quantization/gguf-q4_k_m_apply.py).
  # The wrapper uses the in-repo llama.cpp fork submodule at
  # plugins/plugin-local-inference/native/llama.cpp (its convert_hf_to_gguf.py + a one-shot
  # CPU cmake build of llama-quantize/llama-cli — see the script's
  # _VENDOR_HINT), but the llama-cpp-python wheel also ships a usable
  # `gguf` python module so the HF→f16 GGUF convert step works even
  # without that build. The Q4_K_M *quantize* step still needs a real
  # `llama-quantize` binary. NOTE: the custom GGML types Q4_POLAR/QJL1_256/
  # TurboQuant are only in the fork — same submodule, or $LLAMA_CPP_DIR.
  "llama-cpp-python>=0.3.0",
  # convert_hf_to_gguf.py imports `gguf` AND `mistral_common` at module
  # load — both must be installed or the HF→GGUF convert step dies on
  # import. The fork's requirements/requirements-convert_hf_to_gguf.txt
  # also installs these, but pin them here so `uv run --extra train`
  # works standalone.
  "gguf>=0.10.0",
  "mistral_common>=1.8.3",
  # pytest is consumed by the pre-flight gate (scripts/preflight.sh) and
  # by the CPU-only unit tests it sweeps.
  "pytest>=8.0.0",
  # The RL scenario-pool tests use @pytest.mark.asyncio; collected in
  # tests/rl/ and gated by tests/rl/conftest.py compat shim.
  "pytest-asyncio>=1.0.0",
]
rl = [
  # ByteDance verl — GRPO trainer with bundled vLLM/SGLang rollout server,
  # FSDP+Megatron training backends, and verifiable-reward registry. Stage 2
  # of RL_STRATEGY.md uses the python-callable reward path
  # (`custom_reward_function`) to invoke scripts/eliza_reward_fn.py:compute_score
  # per rollout. Pin a known-good window — the 0.5.x → 0.7.x line introduced
  # `custom_reward_function` as a top-level config field; older releases used
  # `reward_score.compute_score` registry imports. Both still work with the
  # script we ship; the upper bound is just paranoid.
  "verl>=0.5.0,<0.8.0",
  # GRPO trainer needs vLLM for rollouts; pin separately so users running RL
  # without serving don't have to install the full `serve` extra (which has
  # an incompatible torch ABI). vllm>=0.6 is what verl 0.5+ requires.
  "vllm>=0.6.0,<0.8.0",
  # Atropos — continuous RL environment server. Drives the GRPO loop in
  # scripts/rl/atropos_trainer.py against eliza-native trajectory JSONL
  # exports under $ELIZA_STATE_DIR/trajectories. Config:
  # packages/training/config/atropos.yaml.
  "atroposlib>=0.1.0",
  # LiteLLM — OpenAI-compatible client used by the RLAIF judge path to talk
  # to cerebras/gpt-oss-120b and the Nebius testing harness from a single
  # surface.
  "litellm>=1.0.0",
  # Kondo-gate selective backprop kernel — referenced by atropos.yaml
  # (use_kondo: true). Pinned independent of the train/serve torch line.
  "kondo-gate>=0.1.0",
]
rl-tinker = [
  # Thinking Machines Tinker SDK — managed RL training as a service.
  # Optional alternative to the local verl path; entry point lives in
  # scripts/rl/tinker/tinker_trainer.py and is driven by
  # packages/training/config/tinker.yaml. Kept under its own extra so the
  # local-only RL path doesn't pull a network-dependent SDK.
  "tinker>=0.1.0",
]
reward = [
  # AI-judge path inside scripts/eliza_reward_fn.py (gated on
  # ELIZA_REWARD_USE_AI_JUDGE=1). Listed separately so a slim RL worker that
  # only runs verifiable rewards doesn't need to ship the SDK.
  "anthropic>=0.40.0",
]
# vastai and heretic-llm are installed separately (not in train extra) —
# vastai (provisioning CLI) lives on the local dev box and pulls in an
# ancient python-dateutil pin; heretic-llm (refusal-direction abliteration)
# wants psutil>=7.1 which clashes with vastai. Keep them out of the
# resolution graph; install with `uv tool install vastai` and
# `pip install heretic-llm` on the abliteration box.
serve = [
  # vLLM 0.18+ ships EAGLE-3 speculative decoding (PR #38280 era), the
  # TurboQuant K/V quantization plugin, FP8 KV cache via FlashAttention 3
  # on sm_90, and continuous batching needed by the Gemma-era serve harness.
  # Pin a minor floor that captures all of these. See
  # scripts/inference/serve_vllm.py for the canonical flag set per model +
  # GPU target.
  "vllm>=0.20.0,<0.21.0",
  # `vllm-flash-attn` is bundled with vLLM 0.20+; add explicitly so the
  # constraint solver doesn't pull a stale wheel from a transitive dep.
  "transformers>=4.46.0",
  # OpenAI client used by scripts/inference/test_apc_mtp_tool_calls.py to
  # speak vLLM's OpenAI-compatible chat completions endpoint. The harness
  # is the production gate that unlocks `ELIZA_APC_DRAFTER_VERIFIED=1` on
  # any new vLLM build for the Gemma 4 assistant-drafter MTP path.
  "openai>=1.50.0",
]

[tool.pytest.ini_options]
# Tests reach into the repo via `from scripts.training.memory_calc import …`.
# That requires the repo root on sys.path; pytest's `rootdir` autodiscovery
# alone is not enough. `pythonpath = ["."]` makes the repo importable from
# every test invocation, including the pre-flight gate (scripts/preflight.sh).
pythonpath = ["."]
testpaths = ["scripts", "benchmarks", "tests"]
addopts = "-q"
asyncio_mode = "auto"
markers = [
    "asyncio: marks tests as asyncio (handled by pytest-asyncio)",
]

[tool.uv]
package = false
# `rl` (verl) pulls a vllm whose torch pin window conflicts with both
# `train` and `serve`, so keep `rl` isolated. `train` and `serve` are
# NOT declared as conflicting — they share a single torch 2.11+cu130
# wheel in practice (vllm 0.20.x ships that, train deps float to it).
# The local end-to-end smoke (smoke_full_stack.sh) needs both extras
# in one venv; declaring them conflicting breaks that. The Vast SFT
# path installs only `--extra train` and is unaffected.
conflicts = [
  [
    { extra = "train" },
    { extra = "rl" },
  ],
  [
    { extra = "serve" },
    { extra = "rl" },
  ],
  # rl-tinker (cloud SDK) and the local verl/vllm RL path target the same
  # config surface but should not coexist in a single venv — Tinker brings
  # its own torch ABI assumptions and its rollouts are remote.
  [
    { extra = "rl" },
    { extra = "rl-tinker" },
  ],
]