用户可快速查找、复用社区已有模型适配，或提交新适配成果。项目提供轻量CLI工具支持搜索、下载功能，按统一格式管理适配产物，促进社区协作共享。【此简介由AI生成】

文件	最后提交记录	最后更新时间
adaptations	!37 merge batch-20261023-waitcuda into main feat: add 165 community Ascend NPU adaptations (wait_cuda verified, 2026-10 batch) Created-by: kirito_ccq Commit-by: kirito_ccq Merged-by: kirito_ccq Description: ## Summary Adds 165 new adaptation directories that have reached `business_benchmark_status=wait_cuda` (full 4-stage pipeline complete: adaptation → benchmark → optimization → business phase-4 NPU baseline/perf dual-path passed `--wait-cuda-npu-only` gate on Ascend 910, awaiting remote CUDA baseline artifacts). Covers the 2026-10-14 crawl batch (120 trending models), legacy merges, and batch-2 (40 models). ## Contents per directory - `README.md`, `adaptation_status.yaml` (source=community, toolchain=SLAI-AscendBridge2, verification_status=reported_working) - `demo.py` / `accuracy_run.py` / `accuracy_run_perf.py` / business-phase scripts - `pyproject.toml` + `uv.lock` - `model_files/` NPU patch code and cloned custom-repo sources where applicable - Runtime artifacts (outputs_, benchmark_metrics_, trace_, business_metrics_, models caches, profiling) excluded per repository gate ## Housekeeping Retro-fix 3 legacy directories for repository-gate compliance (hyphen → underscore renames; missing README/adaptation_status.yaml added). `scripts/validate_repository.py` passes on the full tree. 🤖 Generated with [Claude Code](https://claude.com/claude-code) See merge request: SLAI/SLAI-AscendBridge2-Community-Adaptations!37	5 天前
scripts	Show fuzzy candidates for empty searches	4 个月前
.gitignore	Update TRELLIS adaptation: add Gradio app, mesh rendering, texture baking, performance optimization - app_npu.py: Gradio Web App (browser → NPU inference → 3D preview + GLB/PLY download) - npu_patches/npu_rasterizer.py: Phong shading + gaussian DC KNN k=5 color transfer + gamma + half-Lambert - npu_patches/glb_bake.py: texture baking (xatlas UV + NPU gaussian multi-view + numpy z-buffer + inpaint) - npu_patches/gsplat_render.py: batched camera (1 linalg_inv vs N) - demo.py: mesh render with gaussian DC colors, GLB texture bake, weight transpose cache (38% speedup) - pyproject.toml: add gradio>=5.0 - README.md: add Gradio Web App section, mesh rendering approach, deployment with app - precision_report.md: 96.5% shape similarity vs official T.ply - Performance: 12-step ~13s Co-Authored-By: Claude <noreply@anthropic.com>	3 个月前
CONTRIBUTING.md	Document and enable manual CI workflow runs	4 个月前
LICENSE	Initialize community adaptations repository	4 个月前
README.md	Show fuzzy candidates for empty searches	4 个月前

SLAI-AscendBridge2-Community-Adaptations

SLAI-AscendBridge2-Community-Adaptations 是 SLAI-AscendBridge2 的社区适配合集仓库。

这个仓库不是只用来“存文件”。它的目标是让用户可以快速查找社区已有适配，下载单个 adaptation 复用，或把自己用 AscendBridge 完成的新模型适配按统一格式提交回来。

快速开始

克隆仓库后可以直接查看已有社区适配：

git clone git@gitcode.com:SLAI/SLAI-AscendBridge2-Community-Adaptations.git
cd SLAI-AscendBridge2-Community-Adaptations

python3 scripts/slai_adapt.py list
python3 scripts/slai_adapt.py search qwen
python3 scripts/slai_adapt.py search Qwen/Qwen2.5

当前 CLI 是轻量查询工具，支持：

list：列出仓库内已有 community adaptation。
search <keyword>：按目录名、上游模型、贡献者、状态、notes 和 adaptation README 内容搜索；没有精确命中时会输出模糊候选。

查找已有社区适配

查看全部 adaptation：

python3 scripts/slai_adapt.py list

搜索某个模型或关键词：

python3 scripts/slai_adapt.py search qwen
python3 scripts/slai_adapt.py search "Qwen/Qwen2.5-7B-Instruct"
python3 scripts/slai_adapt.py search reported_working

下载某个 adaptation

如果你只想使用一个模型的 adaptation，可以只拷贝对应目录：

git clone git@gitcode.com:SLAI/SLAI-AscendBridge2-Community-Adaptations.git
cd SLAI-AscendBridge2-Community-Adaptations

python3 scripts/slai_adapt.py search qwen

找到目标目录后，例如 adaptations/qwen_qwen2_5_7b_instruct/，把它同步到你的 SLAI-AscendBridge2 框架仓：

rsync -av \
  adaptations/qwen_qwen2_5_7b_instruct/ \
  /path/to/SLAI-AscendBridge2/adaptations/qwen_qwen2_5_7b_instruct/

然后在框架仓内运行对应检查或 demo：

cd /path/to/SLAI-AscendBridge2
uv sync
uv run python adaptation/scripts/check_adaptation.py --adapt qwen_qwen2_5_7b_instruct

cd adaptations/qwen_qwen2_5_7b_instruct
uv run python demo.py

如果该 adaptation 已包含 benchmark 或 optimization 文件，也可以继续运行：

cd /path/to/SLAI-AscendBridge2
uv run python benchmark/scripts/check_accuracy_run.py --adapt qwen_qwen2_5_7b_instruct
uv run python optimization/scripts/check_accuracy_run_perf.py --adapt qwen_qwen2_5_7b_instruct

配合 SLAI-AscendBridge2 使用

本仓库只保存社区 adaptation 产物和元数据。真正的适配、benchmark、优化、看板和多智能体编排由 SLAI-AscendBridge2 框架仓负责。

推荐同时准备两个仓库：

git clone git@gitcode.com:SLAI/SLAI-AscendBridge2.git
git clone git@gitcode.com:SLAI/SLAI-AscendBridge2-Community-Adaptations.git

框架仓初始化：

cd SLAI-AscendBridge2
uv sync
uv run python scripts/board_ops.py init

把社区 adaptation 同步进框架仓：

rsync -av \
  ../SLAI-AscendBridge2-Community-Adaptations/adaptations/<name>/ \
  ./adaptations/<name>/

常用检查命令：

uv run python adaptation/scripts/check_adaptation.py --adapt <name>
uv run python benchmark/scripts/check_accuracy_run.py --adapt <name>
uv run python optimization/scripts/check_accuracy_run_perf.py --adapt <name>
uv run python optimization/scripts/check_optimization_notes.py --adapt adaptations/<name>

对一个新模型发起适配

如果仓库里还没有你要的模型，先在 SLAI-AscendBridge2 框架仓中发起适配。

方式一：注册到看板并运行 adaptation 阶段

cd /path/to/SLAI-AscendBridge2
uv sync
uv run python scripts/board_ops.py init

uv run python scripts/get_model_info.py "Qwen/Qwen2.5-7B-Instruct"

uv run python scripts/board_ops.py register_model \
  --model_id "Qwen/Qwen2.5-7B-Instruct" \
  --source huggingface \
  --url "https://huggingface.co/Qwen/Qwen2.5-7B-Instruct" \
  --description "community adaptation request"

PROMPT_MODE=adaptation ./run_auto_team_lead.sh

适配完成后，框架仓里会出现类似目录：

SLAI-AscendBridge2/
└── adaptations/
    └── qwen_qwen2_5_7b_instruct/

方式二：单模型 agent 连贯适配

如果你只想处理一个模型，可以在 SLAI-AscendBridge2 框架仓里启动 Claude、Codex 或 Cursor，让 agent 只处理一个 adaptation 目录。

推荐提示词：

只处理单个模型：Qwen/Qwen2.5-7B-Instruct
目标目录：adaptations/qwen_qwen2_5_7b_instruct

请按顺序完成：
1. Adaptation：生成并修复 demo.py、pyproject.toml、README 等基础产物
2. Benchmark：生成 accuracy_run.py，并通过 benchmark/scripts/check_accuracy_run.py
3. Optimization：生成 accuracy_run_perf.py、optimization_notes.json，并通过 optimization/scripts/check_accuracy_run_perf.py

约束：
- 只修改这个模型相关目录
- 不提交模型权重、缓存、profiling 输出或 benchmark 运行产物
- 每个阶段完成后先跑 checker，再继续下一阶段

提交到本社区库

适配完成后，把框架仓中的单模型目录复制到本仓库：

cd /path/to/SLAI-AscendBridge2-Community-Adaptations
mkdir -p adaptations/qwen_qwen2_5_7b_instruct
rsync -av \
  /path/to/SLAI-AscendBridge2/adaptations/qwen_qwen2_5_7b_instruct/ \
  adaptations/qwen_qwen2_5_7b_instruct/

每个 adaptation 至少需要：

adaptations/<name>/
├── README.md
└── adaptation_status.yaml

adaptation_status.yaml 模板：

source: community
contributor: "your-name-or-org"
upstream_model: "Qwen/Qwen2.5-7B-Instruct"
toolchain: "SLAI-AscendBridge2"
verification_status: unverified
last_checked: "2026-06-06"
notes: ""

verification_status 可使用以下取值：

unverified：已贡献，但维护者尚未检查。
reported_working：贡献者报告在其环境中可以运行。
maintainer_verified：已经由 SLAI 维护者核验。
failed：已知当前无法跑通，但为了复现、排查或后续修复而保留。

提交 PR 前先跑本地检查：

python3 scripts/slai_adapt.py list
python3 scripts/slai_adapt.py search qwen
python3 scripts/validate_repository.py

CI 会拒绝模型权重、下载缓存、profiling 目录、benchmark 输出和运行时产物。更多贡献要求见 CONTRIBUTING.md。

目录结构

SLAI-AscendBridge2-Community-Adaptations/
├── adaptations/
│   └── <sanitized_model_id>/
│       ├── README.md
│       ├── adaptation_status.yaml
│       └── ...
├── scripts/
│   ├── slai_adapt.py
│   └── validate_repository.py
└── .gitcode/
    └── workflows/
        └── ci.yml