zhangdw/slime-v0.3.0-wheels

🧩 slime v0.3.0 cu129 Wheelhouse

Prebuilt wheels and installation notes for reproducing a slime v0.3.0 + SGLang + Megatron-LM CUDA 12.9 environment.

---

✨ What is this?

This repository is a reproducibility wheelhouse for installing slime v0.3.0 with SGLang and Megatron-LM on a Linux x86_64 CUDA 12.9 / cu129 stack.

It is not a model checkpoint repository and it is not a training or evaluation dataset. It stores prebuilt Python wheel artifacts, pinned requirement inputs, and installation notes for rebuilding a known dependency stack with minimal network access.

This wheelhouse is intended for a specific CUDA/Python/Linux environment family. Some wheels contain native CUDA or Linux binaries and should not be treated as universal Python packages.

🧭 At a glance

Item	Value
Primary stack	slime v0.3.0 + SGLang + Megatron-LM
Target OS / arch	Linux x86_64
Python	3.12
CUDA route	CUDA 12.9 / cu129
Payload directories	20
Wheel artifacts	439 .whl files
Approximate local payload size	21 GB
Main install guide	install-from-whl.md / install-from-whl-zh.md

🚀 Quick download

hf download zhangdw/slime-v0.3.0-wheels \
  --type dataset \
  --local-dir "$HOME/slime-v0.3.0-wheels" \
  --max-workers 8

export WHEELHOUSE="$HOME/slime-v0.3.0-wheels"

Then follow the installation guide:

less "$WHEELHOUSE/install-from-whl.md"

Chinese documentation is also included:

less "$WHEELHOUSE/install-from-whl-zh.md"

🧱 Repository layout

Area	Paths	Contents
SGLang bootstrap	1/, 2/, requirements/	cuda-python, pinned SGLang runtime requirements, and runtime dependency wheels.
cu129 core stack	3.1/, 3.2/, 3.3/	PyTorch cu129 wheels, SGLang kernels, deep-gemm, and NVIDIA cu12 runtime wheels.
Native extensions	3.4/ - 3.10/	flash-attn, mbridge, flash-linear-attention, tilelang, transformer_engine, Apex, and torch_memory_saver.
Megatron / model tooling	3.11/, 3.12/, 3.13/	Megatron-Bridge, nvidia-modelopt, and sglang_router.
slime final pins	4.1/ - 4.4/	slime int4_qat kernel, final cuDNN/cu12 pins, NumPy pin, and kernels<0.15.0.
Human docs	root Markdown files	Installation and wheel-preparation guides in English and Chinese.

Detailed directory map

Path	Purpose
requirements/sglang.txt	Pinned input requirements used to collect SGLang runtime wheels.
1/	Bootstrap dependencies before installing SGLang, including cuda-python.
2/	SGLang base runtime dependency wheels.
3.1/	cu129 PyTorch, torchvision, and torchaudio wheels.
3.2/	cu129 SGLang kernel and deep-gemm wheels.
3.3/	NVIDIA cu12 runtime wheels used to replace unwanted cu13 packages.
3.4/	flash-attn and its build/runtime dependency wheels.
3.5/	mbridge.
3.6/	flash-linear-attention and dependencies for the default Qwen GDN fla backend.
3.7/	tilelang and dependencies.
3.8/	transformer_engine[pytorch] and dependencies.
3.9/	Apex wheel with CUDA extensions.
3.10/	torch_memory_saver wheel.
3.11/	Megatron-Bridge wheel.
3.12/	nvidia-modelopt[torch] wheel.
3.13/	sglang_router and dependencies.
4.1/	slime int4_qat kernel wheel.
4.2/	Final cuDNN/cu12 runtime pin wheels.
4.3/	Final NumPy pin wheel.
4.4/	Final kernels<0.15.0 pin and dependencies.

📚 Guides

File	Purpose
install-from-whl.md	English installation guide for configuring slime v0.3.0 from this wheelhouse.
install-from-whl-zh.md	Chinese installation guide.
prepare-whl.md	English guide for preparing this wheelhouse on a machine with good network access.
prepare-whl-zh.md	Chinese wheel-preparation guide.
build_conda.sh	Reference shell script for the conda-based setup path.
README_zh.md	Chinese overview of this dataset repository.

⚠️ Scope boundaries

Capability	Status
SGLang base runtime	✅ Covered by this wheelhouse.
Full sglang[all] optional extras	⚪ Not covered by the base install path.
Default Qwen GDN fla backend	✅ Covered through flash-linear-attention.
FlashQLA backend	⚪ Not included. Only needed for Qwen3.5 / Qwen3-Next with --qwen-gdn-backend flashqla.
A100 / A800 default path	✅ Use --qwen-gdn-backend fla.
H20 / H100 FlashQLA path	⚪ Possible target hardware, but FlashQLA itself must be installed separately.

FlashQLA requires SM90/Hopper-class GPUs such as H20/H100. For A100/A800 or the default Qwen GDN path, use the default --qwen-gdn-backend fla route included in this wheelhouse.

✅ Minimal smoke-test expectation

The installation guide includes CUDA and import smoke tests for the core runtime, including:

• torch CUDA availability and fp16 forward/backward execution
• sglang import
• flash_attn import
• transformer_engine import
• apex import

Passing those checks validates the core environment path covered by this wheelhouse. Optional extras such as sglang[all] and FlashQLA remain outside the base scope.

📄 License note

The wheel files in this repository inherit licenses from their respective upstream packages. Check each upstream project and wheel metadata for authoritative license details.