requirements.txt

accelerate
aiohttp  # session-server backend proxy; also an sglang dependency
backports.strenum; python_version < "3.11"
blake3==1.0.9  # disk-delta weight sync (checksum)
blobfile
datasets
httpx[http2]
hypothesis>=5.40
kubernetes_asyncio
mcp[cli]
memray  # needed for debugging (but is lightweight), we can put it to dev mode when using pyproject.toml
nvidia-resiliency-ext~=0.6.0; platform_system == "Linux"
omegaconf
onnxscript  # import-time dependency of transformer_engine 2.17, which the docker image installs with --no-deps; without it any Megatron-backend run dies at `import megatron`
openai  # imported by miles.rollout.generate_utils.tool_call_utils and the openenv adapter; previously only satisfied transitively via sglang
orjson  # session-server proxy JSON on every turn; also an sglang dependency
pillow
polars==1.42.1
psycopg[binary]  # metric-history gate store (Neon Postgres driver)
pybase64
pylatexenc
pytest-asyncio
pyyaml
qwen_vl_utils # for VLM
# Ray 2.59 removed internal_api.free, which RayObjectStore needs for RPC references.
ray[default]>=2.56.0,<2.59.0
ring_flash_attn; platform_system == "Linux"
safetensors>=0.8.0  # samples-reply wire encoder; the decoder parses its container directly
sglang-router>=0.2.3
tensorboard
torchft-nightly==2026.4.3; platform_system == "Linux" and platform_machine == "x86_64"
tqdm
transformers==5.19.0  # Required by HF-native weight conversion; matches SGLang v0.5.22's pin.
typing_extensions>=4.5
wandb
xxhash==3.7.1  # disk-delta weight sync (checksum)
zstandard==0.25.0  # disk-delta weight sync (codec)