Buckets:
| # NVIDIA RTX Spark — native Windows-on-ARM (win_arm64) requirement set. | |
| # | |
| # Mirrors requirements_base.txt with pins adjusted to versions that actually | |
| # publish win_arm64 wheels (or that we build ourselves — see manager/spec.py | |
| # spark handling). Keep in sync with requirements_base.txt when bumping pins. | |
| # | |
| # Deliberate differences from requirements_base.txt: | |
| # - scipy>=1.17 first release with win_arm64 wheels (also unlocks numpy 2) | |
| # - matplotlib>=3.11 first release with win_arm64 wheels (base: 3.10.1) | |
| # - av>=17 first release with win_arm64 wheels (base: 16.0.1) | |
| # - numba/llvmlite rc only prerelease win_arm64 wheels exist (librosa dep) | |
| # - tensorboard REMOVED: its grpcio dep has no win_arm64 wheels; the | |
| # manager installs tensorboard --no-deps (writer path | |
| # only needs protobuf) as a spark extra | |
| # - hf_transfer REMOVED: no win_arm64 wheels; hf-xet (arm64 OK) is the | |
| # modern replacement and installs via manager extras | |
| # - torchcodec pin REMOVED: the manager installs the locally built | |
| # win_arm64 torchcodec wheel as an extra | |
| # - opencv-python, soxr, pywavelets, brotli(gradio), kornia-rs have no | |
| # published win_arm64 wheels at any version — supplied from the ai-toolkit | |
| # spark wheel set (self-built), resolved via --find-links | |
| numpy>=2,<3 | |
| scipy>=1.17 | |
| torchao==0.17.0 | |
| safetensors | |
| git+https://github.com/huggingface/diffusers.git@c943837899b16cbae2f619b8dd4f7bb6f07dd81a | |
| transformers==5.5.3 | |
| lycoris-lora==1.8.3 | |
| flatten_json | |
| pyyaml | |
| oyaml | |
| kornia | |
| invisible-watermark | |
| einops | |
| accelerate | |
| toml | |
| albumentations==1.4.15 | |
| albucore==0.0.16 | |
| pydantic | |
| omegaconf | |
| open_clip_torch | |
| timm==1.0.22 | |
| prodigyopt | |
| controlnet_aux==0.0.10 | |
| python-dotenv | |
| bitsandbytes | |
| lpips | |
| pytorch_fid | |
| optimum-quanto==0.2.4 | |
| sentencepiece | |
| huggingface_hub==1.23.0 | |
| peft==0.18.1 | |
| gradio | |
| python-slugify | |
| # exact pins: newer opencv versions exist on PyPI only as sdists (which fail | |
| # to build on MSVC arm64 — dnn __fp16); force the self-built 4.12 wheels | |
| opencv-python==4.12.0.88 | |
| opencv-python-headless==4.12.0.88 | |
| pytorch-wavelets==1.3.0 | |
| matplotlib>=3.11 | |
| setuptools>=77.0.3 | |
| av>=17 | |
| # librosa chain: numba/llvmlite have no cp312 win_arm64 wheels upstream — the | |
| # rc pins below resolve from the self-built spark wheel set (llvmlite built | |
| # against our own LLVM 22 static build) | |
| librosa==0.11.0 | |
| numba==0.67.0rc1 | |
| llvmlite==0.49.0rc1 | |
| soxr==1.1.0 | |
| mutagen==1.47.0 | |
| soundfile | |
| # tensorboard is installed --no-deps by the manager (its grpcio dep has no | |
| # win_arm64 wheels; grpc is only needed by the tensorboard server, not the | |
| # SummaryWriter path the toolkit uses). Its remaining runtime deps: | |
| absl-py | |
| markdown | |
| werkzeug | |
| tensorboard-data-server | |
| packaging | |
| protobuf | |
| six | |
Xet Storage Details
- Size:
- 2.78 kB
- Xet hash:
- 5728bb8ffe951d8d8ab17e3fa43f006d62efb4e128a3527aab61c4c255ff0cde
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.