fix: preserve public XPU and test entrypoints

Author: Zachary-wW

CLAUDE.md

@@ -46,21 +46,24 @@ models, and pre-converted checkpoints referenced by the selected configs first,
 
 ```bash
 # Run the default CI suite (all models in tests/llm_vlm/configs/)
-(cd tests/llm_vlm && python3 main.py)
+bash tests/llm_vlm/main_start.sh
 ```
 
 ### Running a Single Model Test
 
-Pass selection and execution options directly to `tests/llm_vlm/main.py`:
+Edit variables in `tests/llm_vlm/main_start.sh`:
 ```bash
 # Run one model from tests/llm_vlm/configs/
-(cd tests/llm_vlm && python3 main.py --models qwen3_14b)
+model_names="qwen3_14b"
 
 # Run one model from tests/llm_vlm/optional_configs/
-(cd tests/llm_vlm && python3 main.py --models deepseek_v2/deepseek_v2_lite --include_optional)
+model_names="deepseek_v2/deepseek_v2_lite"
+include_optional=true
 
 # Run an entire model series from optional_configs/
-(cd tests/llm_vlm && python3 main.py --optional_subdir internvl2.5 --include_optional)
+model_names="NONE"
+optional_subdir="internvl2.5"
+include_optional=true
 ```
 
 Test configs: `tests/llm_vlm/configs/` (CI suite) and `tests/llm_vlm/optional_configs/` (regression, organized by model family). Each YAML defines model params and multi-step `scenarios` (checkpoint conversion + training).

README.md

@@ -72,6 +72,7 @@ See the full documentation for installation, tutorials, and advanced usage — [
 
 **1. Install** — using **prebuilt Docker images** or **source build**:
 - **NVIDIA GPU**: [Installation Guide](https://loongforge.readthedocs.io/en/latest/get_started/installation.html)
+- **Kunlun XPU**: [Installation Guide](https://loongforge.readthedocs.io/en/latest/kunlun_tutorial/install_p800.html)
 
 **2. Launch your first training run** — follow a tutorial for your target hardware and modality:
 - **NVIDIA GPU**: [LLM](https://loongforge.readthedocs.io/en/latest/llm_tutorial/quick_start_llm_pretrain.html) · [VLM](https://loongforge.readthedocs.io/en/latest/vlm_tutorial/quick_start_vlm_pretrain.html) · [VLA & WAM](https://loongforge.readthedocs.io/en/latest/embodied_tutorial/overview.html) · [Diffusion](https://loongforge.readthedocs.io/en/latest/wan_tutorial/quick_start_wan_training.html)
@@ -239,7 +240,7 @@ LoongForge/
 ├── tools/                        # Checkpoint conversion, data preprocessing
 ├── ops/                          # Custom fused operators (incl. open-sourced TileLang)
 ├── patches/                      # TransformerEngine patches
-├── docker/                       # Dockerfile (GPU)
+├── docker/                       # Dockerfiles (GPU & XPU)
 ├── tests/                        # E2E test suite (YAML-driven)
 └── docs/                         # Documentation
 ```

README_zh.md

@@ -72,6 +72,7 @@
 
 **1. 安装** —— 可使用 **预构建镜像** 或 **源码构建**：
 - **NVIDIA GPU**：[安装指南](https://loongforge.readthedocs.io/zh-cn/latest/get_started/installation.html)
+- **昆仑芯 XPU**：[安装指南](https://loongforge.readthedocs.io/zh-cn/latest/kunlun_tutorial/install_p800.html)
 
 **2. 启动你的第一个训练任务** —— 根据目标硬件与模态选择教程：
 - **NVIDIA GPU**：[LLM](https://loongforge.readthedocs.io/zh-cn/latest/llm_tutorial/quick_start_llm_pretrain.html) · [VLM](https://loongforge.readthedocs.io/zh-cn/latest/vlm_tutorial/quick_start_vlm_pretrain.html) · [VLA & WAM](https://loongforge.readthedocs.io/zh-cn/latest/embodied_tutorial/overview.html) · [Diffusion (WAN)](https://loongforge.readthedocs.io/zh-cn/latest/wan_tutorial/quick_start_wan_training.html)
@@ -238,7 +239,7 @@ LoongForge/
 ├── tools/                        # Checkpoint 转换、数据预处理
 ├── ops/                          # 自定义融合算子（含开源的 TileLang 版本）
 ├── patches/                      # TransformerEngine 补丁
-├── docker/                       # Dockerfile（GPU）
+├── docker/                       # Dockerfile（GPU & XPU）
 ├── tests/                        # 端到端测试（YAML 驱动）
 └── docs/                         # 文档
 ```

ci/sensitive_rules.py

@@ -207,7 +207,6 @@
                "as a path or symbol default, break external users.",
         "pattern": r"(?i)\bAIAK[-_](?:Training[-_](?:Omni|LLM)|Megatron)\b"
                    r"|\bBaigeOmni\b"
-                   r"|\bLoongForge-VLA\b"
                    r"|\baiak_training_omni\b"
                    r"|\bset_aiak_\w+|\binitialize_baige_\w+|\bUSE_AIAK_\w+"
                    r"|\baiak-ckpt\b",

docker/__init__.py

@@ -30,15 +30,15 @@ else:
     for hook_file in hook_files:
         with open(hook_file, "r") as f:
             content = f.read()
-
+        
         if "# patched: auto-detect pip isolated env" in content:
             print(f"[xpytorch] Already patched: {hook_file}")
             patched_count += 1
             continue
-
+        
         # Fix 1: Add pip isolated environment detection at the beginning of _custom_import
         old_func_start = r"def _custom_import\(module_name, globals=None, locals=None, fromlist=\(\), level=0\):\n( *)global SYMBOL_REWRITE_REGISTER"
-
+        
         def make_func_patch(m):
             indent = m.group(1)
             detection_code = (
@@ -51,18 +51,18 @@ else:
                 + detection_code +
                 f"{indent}global SYMBOL_REWRITE_REGISTER"
             )
-
+        
         content, count1 = re.subn(old_func_start, make_func_patch, content)
-
+        
         # Fix 1 is critical, fail if not matched
         if count1 == 0:
             print(f"ERROR: Function pattern not found in {hook_file}", file=sys.stderr)
             print("This is a critical patch. Build cannot continue.", file=sys.stderr)
             sys.exit(1)
-
+        
         # Fix 2: torch_version = version('torch') with fallback
         old_pattern2 = r"^( *)torch_version = version\('torch'\)"
-
+        
         def make_patch2(m):
             indent = m.group(1)
             return (
@@ -74,12 +74,12 @@ else:
                 f"{indent}    except ImportError:\n"
                 f"{indent}        torch_version = \"0.0.0\"  # patched fallback"
             )
-
+        
         content, count2 = re.subn(old_pattern2, make_patch2, content, flags=re.MULTILINE)
-
+        
         with open(hook_file, "w") as f:
             f.write(content)
-
+        
         patched_count += 1
         print(f"[xpytorch] Patched: {hook_file} (func={count1}, version={count2})")
 
@@ -103,31 +103,31 @@ else:
     for xray_file in xray_files:
         with open(xray_file, "r") as f:
             content = f.read()
-
+        
         if "if spec is not None and spec.loader is not None:" in content:
             print(f"[xray] Already patched: {xray_file}")
             xray_patched_count += 1
             continue
-
+        
         old_pattern = r"^( *)spec\.loader = XrayMetaPathLoader\(spec\.loader\)"
-
+        
         def make_xray_patch(m):
             indent = m.group(1)
             return (
                 f"{indent}if spec is not None and spec.loader is not None:\n"
                 f"{indent}    spec.loader = XrayMetaPathLoader(spec.loader)"
             )
-
+        
         content, count = re.subn(old_pattern, make_xray_patch, content, flags=re.MULTILINE)
-
+        
         if count == 0:
             print(f"ERROR: Pattern not found in {xray_file}", file=sys.stderr)
             print("This is a critical patch. Build cannot continue.", file=sys.stderr)
             sys.exit(1)
-
+        
         with open(xray_file, "w") as f:
             f.write(content)
-
+        
         xray_patched_count += 1
         print(f"[xray] Patched: {xray_file}")
 

docker/fix_xpytorch_hook.sh

@@ -1,6 +1,6 @@
 # Benchmark Environments
 
-This document records the benchmark runtime environments used by the LoongForge eval module under `loongforge/embodied/eval`.
+This document records the benchmark runtime environments used by the LoongForge-VLA eval module under `loongforge/embodied/eval`.
 
 **For installation, follow the official benchmark homepages** — each benchmark page links to its official repository. This document is not an install guide; it records the dependency versions of the environments currently used by the eval module, so they can be compared / reproduced when debugging.
 
@@ -205,4 +205,3 @@ Used by:
 ```text
 examples/embodied/pi05/eval/configs/maniskill/*.yaml
 ```
-

docs/source/embodied_tutorial/eval_docs/benchmark_envs.md

@@ -33,7 +33,7 @@ The evaluation needs the original-format CALVIN validation assets, not just a Le
 Run from inside the **benchmark** environment. The run scripts and eval YAMLs ship with `/path/to/...` placeholders — fill them in before running:
 
 ```bash
-cd /path/to/LoongForge
+cd /path/to/LoongForge-VLA
 examples/embodied/pi05/eval/run_calvin_eval.sh    # pi05 (connectivity only)
 examples/embodied/xvla/eval/run_calvin_eval.sh    # xvla (connectivity only)
 ```

docs/source/embodied_tutorial/eval_docs/benchmarks/calvin.md

@@ -36,7 +36,7 @@ The benchmark dataset is downloaded following the official LIBERO data instructi
 Run from inside the **benchmark** environment. The run scripts and eval YAMLs ship with `/path/to/...` placeholders — fill them in before running:
 
 ```bash
-cd /path/to/LoongForge
+cd /path/to/LoongForge-VLA
 
 # pi05
 examples/embodied/pi05/eval/run_libero_eval.sh
@@ -53,7 +53,7 @@ Environment variables:
 | Variable | Meaning | Default |
 |---|---|---|
 | `CONFIG` | eval YAML config path | `<model>/eval/configs/libero/object_smoke.yaml` (xvla: `libero_weight_object_smoke.yaml`) |
-| `REPO_ROOT` | LoongForge repo root | `/path/to/LoongForge` |
+| `REPO_ROOT` | LoongForge-VLA repo root | `/path/to/LoongForge-VLA` |
 | `BENCHMARK_PYTHON` | LIBERO env interpreter | `/path/to/libero/bin/python` |
 | `CUDA_VISIBLE_DEVICES` | GPU id for the policy server | `0` |
 | `MUJOCO_GL` / `PYOPENGL_PLATFORM` | MuJoCo offscreen GL backend | `osmesa` |

docs/source/embodied_tutorial/eval_docs/benchmarks/libero.md

@@ -32,7 +32,7 @@ pip install websockets msgpack pyyaml
 Run from inside the **benchmark** environment. The run scripts and eval YAMLs ship with `/path/to/...` placeholders — fill them in before running:
 
 ```bash
-cd /path/to/LoongForge
+cd /path/to/LoongForge-VLA
 examples/embodied/pi05/eval/run_maniskill_eval.sh    # pi05 (connectivity only)
 examples/embodied/xvla/eval/run_maniskill_eval.sh    # xvla (connectivity only)
 ```