--extra-index-url https://download.pytorch.org/whl/cu132
-c constraints.txt
accelerate>=1.7.0
build
colored
cuda-python>=13
# Diffusers 0.40 is required for MiniMax H3 support.
# Recheck H3's experimental modular imports before lifting the minor-version cap.
diffusers>=0.40.0,<0.41
safetensors>=0.8.0
ftfy
lark
lazy_loader~=0.5
mpi4py
numpy>=2.0.0,<2.4 # numba 0.63.1 requires numpy<2.4
graphviz
openai
polygraphy
psutil
nvidia-ml-py>=13
pulp
pandas
h5py==3.12.1
StrEnum
sentencepiece>=0.1.99
# https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/rel-26-09.html#rel-26-09 uses 2.14.0a0.
# The floor is the NGC alpha, which PEP 440 orders below the public release of the same
# series; the ceiling is that public release, which is what the resolver picks off the public
# index. Ordered comparisons drop the local version, so the container's +nv build stays
# admissible against the floor.
torch>=2.14.0a0,<=2.14.0
torchvision
# The floor is the exact pin of the public torch above; the ceiling is the NCCL that
# https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/rel-26-09.html#rel-26-09
# ships, so the resolver lands on the container's own NCCL version there instead of an older one.
nvidia-nccl-cu13>=2.30.7,<=2.31.2
# NCCL-EP uses nccl4py's nccl.ep package. The nccl-extensions wheel is built
# from pinned source by CMake so it links against the container's NCCL.
nccl4py>=0.4,<=0.5
nvidia-cuda-nvrtc
transformers==5.5.4
prometheus_client
# 8.x is the first series to allow starlette>=1.0.0; 7.1.0 pins starlette<1.0.0
# and can no longer resolve against the starlette floor below.
prometheus_fastapi_instrumentator>=8.1.0
pydantic>=2.9.1
pydantic-settings[yaml]
msgspec
omegaconf
pillow
PyNvVideoCodec~=2.1.0
optimum
datasets
evaluate
mpmath>=1.3.0
click>=8.4.2
click_option_group>=0.5.3
aenum
pyzmq
fastapi>=0.136.3,<0.140.0
starlette>=1.3.1
uvicorn
setuptools<80
# FlashInfer source builds (--no-build-isolation) need packaging>=24.2 for setuptools license-expression validation
packaging>=24.2
ordered-set
peft>=0.18.1,<0.19.0
patchelf
einops
fla-core==0.5.2 # the Kimi Linear / Kimi K3 KDA layers import fla.modules and fla.ops when they load
flashinfer-python==0.6.18
xgrammar==0.1.32
llguidance==0.7.29
jsonschema
backoff
nvtx
matplotlib # FIXME: this is added to make nvtx happy
meson
ninja
blake3
soundfile
# https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/rel-26-09.html#rel-26-09
# ships 3.8.0 and the torch pin above requires the same series, so floor and ceiling coincide.
# The pin carries no local version, which keeps the container's own +nv build admissible -- it
# must not be reinstalled underneath the rest of the stack that was built against it. This is
# also the version the vendored triton_kernels below track.
triton==3.8.0 # NOTE: if you update this, you must also run scripts/vendor_triton_kernels.py to vendor the new version of triton_kernels
tiktoken
blobfile
openai-harmony==0.0.4
# 4.8.0 is the first final release exposing Rubin/SM107 (replaces the 4.8.0.dev0 pre-release pin).
nvidia-cutlass-dsl[cu13]==4.8.0; python_version >= "3.10" # required by PrimTS task-scheduled kernels and Rubin CuTe DSL kernels
nvidia-matmul-heuristics==0.1.0.27; python_version >= "3.10" # analytic GEMM heuristics for CuTe DSL autotuner tactic pruning
quack-kernels==0.5.0; python_version >= "3.10" # required for MiniMax-M3 MSA and FlashAttention 4; only published release compatible with CUTLASS DSL >= 4.7
jinja2 # required for MinimaxM3 MSA
plotly
numexpr
partial_json_parser
mcp>=1.23.0,<2.0.0 # 2.0.0 removed mcp.server.fastmcp; scaffolding still uses the 1.x APIs
apache-tvm-ffi==0.1.13.post2 # required by the CUTLASS DSL TVM-FFI provider
torch-c-dlpack-ext==0.1.3 # used for reduce nvidia-cutlass-dsl host overhead, optional package for improved torch tensor calling perf
flash-attn-4==4.0.0b19
mistral-common>=1.10.0
torchao>=0.14.1,<0.16.0
cuda-core
llist
cuda-tile>=1.0.1
nvidia-cuda-tileiras>=13.1,<13.2
nvidia-cudnn-frontend>=1.27.0,<1.28
etcd-sdk-python==0.0.7
# Required by the checked-in OpenEngine gRPC bindings.
grpcio>=1.67.1,<2
# etcd-sdk-python imports google.protobuf but omits it from its package metadata.
# The binding lower bound comes from the CUTLASS DSL libraries, which need protobuf 6.x.
protobuf>=6.30.2,<7
python-multipart
cache-dit>=1.3.5
librosa
msgpack
uvloop>=0.19.0,<0.23.0
mooncake-transfer-engine-cuda13==0.3.13