wehub-resource-sync
50 lines
1.4 KiB
Plaintext
50 lines
1.4 KiB
Plaintext
aiofile
|
|
aiofiles
|
|
blake3
|
|
aiohttp
|
|
awscrt
|
|
cufile-python
|
|
fastapi
|
|
httpx
|
|
huggingface_hub>=1.5.0
|
|
msgspec
|
|
# nixl uses numba which requires numpy<=2.2.6
|
|
numpy<=2.2.6
|
|
numba
|
|
nvtx
|
|
opentelemetry-api >= 1.20.0, <= 1.40.0
|
|
opentelemetry-sdk >= 1.20.0
|
|
opentelemetry-exporter-otlp >= 1.20.0
|
|
opentelemetry-exporter-prometheus >= 0.50b0, <= 0.61b0
|
|
prometheus_client >= 0.18.0, <= 0.24.1
|
|
psutil
|
|
py-cpuinfo
|
|
# cupy.testing._random unconditionally imports pytest at module level;
|
|
# torch._dynamo triggers this via inspect.getmodule() during compilation.
|
|
pytest
|
|
pyyaml
|
|
pyzmq >= 25.0.0
|
|
redis
|
|
safetensors
|
|
setuptools>=77.0.3,<81.0.0
|
|
setuptools_scm>=8
|
|
sortedcontainers
|
|
# 1. avoid pinning torch version in the runtime dependencies so that
|
|
# installing lmcache (from source or from pypi) will not override
|
|
# the user's preexisting torch version
|
|
# 2. this keeps lmcache torch version flexible for many serving engines
|
|
# and versions of those serving engines
|
|
# if NO pre-existing torch version is installed, `pip install lmcache` or
|
|
# `pip install -e . --no-build-isolation` will install the latest torch version
|
|
# 3. the extra benefit of not pinning is so that when vllm releases their docker
|
|
# image (which has lmcache as a dependency), we will never forcefully collide with
|
|
# whatever torch version that they have
|
|
# 4. this torch version may also be overridden inside of LMCache/docker/Dockerfile
|
|
torch
|
|
transformers >= 5.4
|
|
uvicorn
|
|
httptools
|
|
cachetools
|
|
google-api-core
|
|
google-cloud-bigtable
|