-
Notifications
You must be signed in to change notification settings - Fork 2.7k
Expand file tree
/
Copy pathrequirements.txt
More file actions
105 lines (105 loc) · 3.24 KB
/
Copy pathrequirements.txt
File metadata and controls
105 lines (105 loc) · 3.24 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
--extra-index-url https://download.pytorch.org/whl/cu130
-c constraints.txt
accelerate>=1.7.0
build
colored
cuda-python>=13
# FlowMatchEuler respects a supplied generator starting in 0.39.0 (huggingface/diffusers#13678).
# Upper bound: 0.40.0 requires huggingface-hub>=1.23.0, which requires click>=8.4.2
# and conflicts with the click<8.4 pin below. Lift together with that pin.
diffusers>=0.39.0,<0.40
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-26-05.html#rel-26-05 uses 2.12.0a0.
# The floor admits the NGC alpha builds (2.12.0a0+...); torch<2.12 pins triton==3.6.0,
# which conflicts with the triton pin below.
torch>=2.12.0a0,<=2.13.0a0
torchvision
# https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/rel-26-05.html#rel-26-05 uses 2.30.4
# torch 2.12.0+cu130 uses NCCL 2.29.7; NGC 26.05 uses NCCL 2.30.4.
nvidia-nccl-cu13>=2.29.7,<=2.30.4
# NcclEP uses nccl4py's nccl.ep package without changing the NCCL wheel constraint.
nccl4py>=0.3.1,<0.4
nvidia-cuda-nvrtc
transformers==5.5.4
prometheus_client
prometheus_fastapi_instrumentator
pydantic>=2.9.1
pydantic-settings[yaml]
msgspec
omegaconf
pillow
PyNvVideoCodec~=2.1.0
optimum
# evaluate needs datasets>=2.0.0 which triggers datasets>3.1.0 which is not stable: https://github.com/huggingface/datasets/issues/7467
datasets==3.1.0
evaluate
mpmath>=1.3.0
click>=8.3.1,<8.4
click_option_group
aenum
pyzmq
fastapi>=0.120.1,<=0.121.3
starlette>=0.49.1
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
flashinfer-python==0.6.16
xgrammar==0.1.32
llguidance==0.7.29
jsonschema
backoff
nvtx
matplotlib # FIXME: this is added to make nvtx happy
meson
ninja
blake3
soundfile
triton==3.7.0 # NOTE: if you update this, you must also run scripts/vendor_triton_kernels.py to vendor the new version of triton_kernels
xdsl>=0.59.0 # Optional: required for MLIR-based elementwise fusion in AutoDeploy
tiktoken
blobfile
openai-harmony==0.0.4
nvidia-cutlass-dsl[cu13]==4.6.1; python_version >= "3.10" # required by Blackwell 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==0.6.2; python_version >= "3.10" # required for MinimaxM3 MSA and FlashAttention 4; pins CUTLASS DSL 4.6.1
jinja2 # required for MinimaxM3 MSA
plotly
numexpr
partial_json_parser
mcp<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 CUTLASS DSL 4.6.1's 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
etcd-sdk-python==0.0.7
# etcd-sdk-python imports google.protobuf but omits it from its package metadata.
protobuf>=5.27.2
python-multipart
cache-dit>=1.3.5
librosa
msgpack
uvloop>=0.19.0