[build] Add A100 support: patch set, offline-friendly conda build, and examples

[jason9693]

Summary

Add patch set, build scripts, Docker image, and examples to build and run slime on A100 environments.

Offline Build Flow

The offline build allows installing slime on air-gapped nodes without internet access by pre-downloading all dependencies on a networked machine.

┌─────────────────────────────────────────────────────────────┐
│                  Networked Machine                          │
│                                                             │
│  custom_proxy_net.sh                                        │
│  ├── Download conda pkgs (python, cuda, nccl, cudnn)        │
│  │   └── .downloads/conda-env.tar.gz                        │
│  └── Download pip wheels (torch, torchvision, torchaudio)   │
│      └── .downloads/pip-pkgs/                               │
└──────────────────────┬──────────────────────────────────────┘
                       │  scp / rsync
                       ▼
┌─────────────────────────────────────────────────────────────┐
│                Air-gapped Target Node                       │
│                                                             │
│  Step 1: custom_proxy.sh                                    │
│  ├── Install micromamba from .downloads/mamba_install.sh     │
│  ├── Restore conda env from conda-env.tar.gz                │
│  └── Fix shebangs for target paths                          │
│                         │                                   │
│                         ▼                                   │
│  Step 2: build_conda.a100.sh  (PATCH_VERSION=v0.5.9.a100)  │
│  ├── Detect existing slime env → skip conda setup           │
│  ├── Install torch from local .downloads/pip-pkgs/          │
│  ├── Build from source (sglang, DeepEP, flash-attn, TE, …) │
│  ├── Apply A100 patches:                                    │
│  │   ├── deep_ep.patch          (libcuda RTLD_GLOBAL fix)   │
│  │   ├── sglang.patch           (SM80 compatibility)        │
│  │   ├── megatron.patch         (SM80 compatibility)        │
│  │   ├── slime.patch            (A100 runtime adjustments)  │
│  │   └── transformer_engine.patch (site-packages patch)     │
│  └── Final verification (torch, TE, DeepEP import check)    │
└─────────────────────────────────────────────────────────────┘

Changes

Build System

• build_conda.a100.sh: Add A100-specific conda build script with offline pip/mamba cache support
• build_conda.sh: Introduce PATCH_VERSION selector (v0.5.9, v0.5.9.a100), pin DeepEP commit (1d3963d), apply patches with --reject to land non-conflicting hunks
• docker/Dockerfile.a100: Add A100 Docker image build file (DeepEP built without SM90/NVSHMEM, includes import-check layer)
• .gitignore: Add build artifact paths

Patches (docker/patch/v0.5.9.a100/)

• deep_ep.patch: DeepEP A100 compatibility patch
• megatron.patch: Megatron A100 compatibility patch
• sglang.patch: SGLang A100 compatibility patch (~4,700 lines)
• slime.patch: Slime A100 compatibility patch
• transformer_engine.patch: Transformer Engine A100 compatibility patch

Examples

• examples/dapo_math/a100.sh: DAPO math run script for A100
• examples/dapo_math/kanana2-30b-a3b.sh: Kanana2-30B-A3B model config (same architecture as DeepSeek V3)
• examples/dapo_math/run_kanana2-30b-a3b.a100.sh: Kanana2-30B-A3B A100 run script (same architecture as DeepSeek V3)

Infra / Helpers

• custom_proxy_net.sh: Pre-download conda/pip packages on a networked machine
• custom_proxy.sh: Restore pre-downloaded packages on air-gapped target node
• .downloads/mamba_install.sh: Offline micromamba installer

Docs

• docs/en/get_started/quick_start.md, docs/zh/get_started/quick_start.md: Document the new A100 build flow

Commit History (6 commits)

Commit	Description
5a4f297e	[build] Add A100 patch set and offline-friendly conda build
6eb70d16	[build] Pin DeepEP commit and tolerate patch rejects
8ab0c1be	Fix conda build script
055d530d	Fix patch files
e4bb7fa0	[examples] Add example scripts
4f75ee60	Modify build and run logics

Stats

• 16 files changed, 7,496 insertions(+), 2 deletions(-)