cheng874 · cheng874 · Sep 30, 2026 · Sep 23, 2026 · Sep 23, 2026 · Sep 23, 2026
diff --git a/.gitignore b/.gitignore
@@ -2,6 +2,7 @@
 CLAUDE.md
 Makefile
 docs/.claude/
+docs/flagos_homepage/.claude/
 docs/_build/
 docs/dockerfile
 make.bat
@@ -12,10 +13,15 @@ to-do/
 .playwright-mcp/
 dockerfile
 dockerfile.bak
+docs/flagos-homepage/AGENTS.md
+docs/flagos-homepage/TODO.md
+__pycache__/
+*.py[cod]
 
 # Playwright related files
 docs/flagcicd_zh/.playwright-mcp/
 docs/flagcicd_zh/static/node_modules/
 docs/flagcicd_zh/static/package.json
 docs/flagcicd_zh/static/package-lock.json
 docs/flagcicd_zh/static/*.js
+
diff --git a/docs/_ext/flagos_page_tags.py b/docs/_ext/flagos_page_tags.py
@@ -0,0 +1,125 @@
+"""Project-specific Sphinx helpers for the FlagOS homepage."""
+
+import json
+
+from sphinx.addnodes import only, toctree
+from sphinx.transforms.post_transforms import SphinxPostTransform
+
+
+def _frontmatter_tags(app, docname):
+    """Return Sphinx tags declared in MyST frontmatter for one document."""
+    metadata = app.env.metadata.get(docname, {})
+    raw_tags = metadata.get("tags", [])
+
+    if isinstance(raw_tags, str):
+        try:
+            raw_tags = json.loads(raw_tags)
+        except json.JSONDecodeError:
+            raw_tags = [raw_tags]
+
+    if isinstance(raw_tags, str):
+        raw_tags = [raw_tags]
+
+    return [tag for tag in raw_tags if isinstance(tag, str)]
+
+
+def _eval_condition_with_page_tags(app, condition, page_tags):
+    """Evaluate an ``only`` expression with temporary page-level tags."""
+    added_tags = []
+    for tag in page_tags:
+        if not app.tags.has(tag):
+            app.tags.add(tag)
+            added_tags.append(tag)
+
+    app.tags._condition_cache.clear()
+    try:
+        return app.tags.eval_condition(condition)
+    finally:
+        for tag in added_tags:
+            app.tags.remove(tag)
+        app.tags._condition_cache.clear()
+
+
+class PageFrontmatterOnlyTransform(SphinxPostTransform):
+    """Apply ``only`` directives using tags from the current MyST page."""
+
+    default_priority = 49
+
+    def run(self, **kwargs):
+        docname = self.env.docname
+        page_tags = _frontmatter_tags(self.app, docname)
+        if not page_tags:
+            return
+
+        for node in list(self.document.findall(only)):
+            try:
+                keep = _eval_condition_with_page_tags(self.app, node.get("expr", ""), page_tags)
+            except Exception:
+                continue
+
+            if keep:
+                node.replace_self(node.children)
+            else:
+                node.parent.remove(node)
+
+
+def _prune_inactive_only_toctrees(app, doctree):
+    """Keep toctree relations in sync with active ``only`` branches.
+
+    Sphinx records ``toctree`` relationships while parsing a document, before
+    the ``only`` transform removes inactive branches. Rebuild the document's
+    include list from toctree nodes whose enclosing ``only`` conditions are
+    active. Page-level MyST ``tags`` frontmatter is treated as active while
+    evaluating the current document.
+    """
+    docname = app.env.temp_data.get("docname")
+    if not docname or docname not in app.env.toctree_includes:
+        return
+
+    page_tags = _frontmatter_tags(app, docname)
+    active_includes = []
+    found_conditional_toctree = False
+
+    for toctree_node in doctree.findall(toctree):
+        current = toctree_node.parent
+        keep = True
+        is_conditional = False
+
+        while current is not None:
+            if isinstance(current, only):
+                is_conditional = True
+                expr = current.get("expr", "")
+                try:
+                    if not _eval_condition_with_page_tags(app, expr, page_tags):
+                        keep = False
+                        break
+                except Exception:
+                    # Let Sphinx's own only transform report malformed expressions.
+                    pass
+            current = current.parent
+
+        if is_conditional:
+            found_conditional_toctree = True
+
+        if keep:
+            active_includes.extend(toctree_node.get("includefiles", []))
+
+    if found_conditional_toctree:
+        app.env.toctree_includes[docname] = active_includes
+
+        compact_toc = app.env.tocs.get(docname)
+        if compact_toc is not None:
+            for only_node in list(compact_toc.findall(only)):
+                expr = only_node.get("expr", "")
+                try:
+                    keep = _eval_condition_with_page_tags(app, expr, page_tags)
+                except Exception:
+                    continue
+                if not keep and only_node.parent is not None:
+                    only_node.parent.remove(only_node)
+
+
+def setup(app):
+    app.add_post_transform(PageFrontmatterOnlyTransform)
+    app.connect("doctree-read", _prune_inactive_only_toctrees)
+    return {"version": "0.1", "parallel_read_safe": True, "parallel_write_safe": True}
diff --git a/docs/conf.py b/docs/conf.py
@@ -109,7 +109,8 @@ def get_project(projects):
     "sphinx_tippy",
     "sphinxcontrib.lightbox2",  # click-to-enlarge / lightbox for images
     "sphinx_tippy",
-    "sphinx_togglebutton"
+    "sphinx_togglebutton",
+    "flagos_page_tags"
 ]
 
 # Check and add actually installed extensions
@@ -205,6 +206,13 @@ def get_project(projects):
             "html_title": "FlagGems-sglang Documentation",
         },
     },
+    "flaggems_sglang_zh": {
+        "use_config_file": False,
+        "config": {
+            "project": "FlagGems-sglang 文档中心",
+            "html_title": "FlagGems-sglang 文档中心",
+        },
+    },
     "flagdnn_en": {
         "use_config_file": False,
         "config": {
@@ -648,7 +656,14 @@ def get_project(projects):
 # release = version
 
 # Exclude patterns - exclude all other project directories
-exclude_patterns = ["_build", "shared", "_includes"]
+exclude_patterns = [
+    "_build",
+    "shared",
+    "_includes",
+    "chip_adaptation_guide/_shared",
+    "chip_adaptation_guide/TODO.md",
+    "chip_adaptation_guide_toctree_backup",
+]
 all_projects = list(multiproject_projects.keys())
 for project in all_projects:
     if project != docset:
@@ -675,6 +690,8 @@ def get_project(projects):
 
 intersphinx_disabled_reftypes = ["*"]
 
+myst_frontmatter_process = "yaml"
+
 myst_enable_extensions = [
     "dollarmath",
     "amsmath",
@@ -753,6 +770,8 @@ def get_project(projects):
 # Common static paths
 html_static_path = ["_static", f"{docset}/_static"]
 html_css_files = ["custom.css", "homepage.css"]
+if docset == "flagos_homepage":
+    html_css_files.append("guide.css")
 html_js_files = []
 
 # html_logo = "img/logo.png"
@@ -794,15 +813,18 @@ def get_project(projects):
         "use_download_button": False,
         "repository_url": "https://github.com/flagos-ai/KernelGen",
         "use_repository_button": True,
-        "secondary_sidebar_items": {},
+        "secondary_sidebar_items": {
+            "**": ["page-toc"],
+            "flagos_homepage/index": [],
+        },
+        "show_toc_level": 2,
         "footer_start": ["copyright"],
         "footer_end": [],
         "show_sphinx": False,
         "navbar_end": ["navbar-icon-links"]
     }
 
-    # Update secondary sidebar items for flagos_homepage
-    html_theme_options["secondary_sidebar_items"]["flagos_homepage/index"] = []
+    # Keep the FlagOS homepage clean while enabling page-local TOC elsewhere.
 
     # html_sidebars is only for PyData Sphinx Theme
     html_sidebars = {}
@@ -904,4 +926,3 @@ def get_project(projects):
 }
 
 suppress_warnings = ["epub.unknown_project_files"]
-
diff --git a/docs/flagcx_en/CHANGELOG.md b/docs/flagcx_en/CHANGELOG.md
@@ -1,6 +1,16 @@
 # Release Notes
 
-- **[2026/06]** Released [v0.13](https://github.com/flagos-ai/FlagCX/releases/tag/v0.13.0-rc2.post1):
+- **[2026/09]** Released [v.0.14.0](https://github.com/flagos-ai/FlagCX/releases/tag/v0.14.0):
+
+  - Adds SHMEM Device API support with `USE_SHMEM` and `SHMEM_HOME` build controls.
+  - Adds Scalar IR and Unified IR Device API test coverage, including split intra-node and inter-node test targets.
+  - Adds ACCL/Barex network adaptor support and PPU backend integration controls.
+  - Extends PyTorch plugin build detection for PPU and the canonical `USE_ILUVATAR` backend flag.
+  - Updates Iluvatar build flag documentation: use `USE_ILUVATAR`; `USE_ILUVATAR_COREX` remains only as a deprecated compatibility alias.
+  - Widens the lane-mask ABI to 64 bits for Device IR paths.
+  - Updates packaging and dependency handling, including the NCCL 2.27 documentation floor for NVIDIA/NCCL wrapper usage.
+
+- **[2026/06]** Released [v.0.13.0](https://github.com/flagos-ai/FlagCX/releases/tag/v0.13.0):
 
   - Introduces FlagCX P2P Engine for one-sided RDMA operations, designed for integration with transfer frameworks like NIXL.
   - Adds IBRC P2P adaptor for InfiniBand-based P2P communication.
@@ -11,7 +21,7 @@
   - Adds new one-sided RDMA operations: `flagcxPut`, `flagcxBatchPut`, `flagcxReadCounter`, `flagcxWaitCounter`.
   - Optimizes RMA proxy with batched one-sided PUT operations for improved RDMA throughput.
 
-- **[2026/05]** Released [v0.12](https://github.com/flagos-ai/FlagCX/releases/tag/v0.12.0):
+- **[2026/05]** Released [v.0.12.0](https://github.com/flagos-ai/FlagCX/releases/tag/v0.12.0):
 
   - Adds support for Sunrise AI accelerators, including device adaptor `ptpuAdaptor` and CCL adaptor `pcclAdaptor`.
   - Extends PyTorch plugin support to PCCL backend.
@@ -23,32 +33,32 @@
   - Adds bootstrap extension for enhanced rendezvous capabilities.
   - Replaces C++17 features with C++11 equivalents for improved compatibility.
 
-- **[2026/03]** Released [v0.11](https://github.com/flagos-ai/FlagCX/releases/tag/v0.11.0):
+- **[2026/03]** Released [v.0.11.0](https://github.com/flagos-ai/FlagCX/releases/tag/v0.11.0):
 
   - Enables kernel-based communication on heterogeneous platforms, including NVIDIA and Hygon.
   - Adds support for both host-side and device-side one-sided communication semantics.
   - Introduces adaptor plugin support, enabling dynamic loading of user-defined Device, CCL, and Net adaptor implementations.
 
-- **[2026/02]** Released [v0.10](https://github.com/flagos-ai/FlagCX/releases/tag/v0.10.0):
+- **[2026/02]** Released [v.0.10.0](https://github.com/flagos-ai/FlagCX/releases/tag/v0.10.0):
 
   - Implements 11 chip-decoupled collective communication algorithms in uniRunner mode.
   - Refactors Device Intra-/Inter-node API and integrates NCCL Device API support on NVIDIA platforms.
   - Enhances usability with pip install support for FlagCX and an NCCL wrapper plugin for seamless adoption on NVIDIA platforms.
 
-- **[2026/01]** Released [v0.9](https://github.com/flagos-ai/FlagCX/releases/tag/v0.9.0):
+- **[2026/01]** Released [v.0.9.0](https://github.com/flagos-ai/FlagCX/releases/tag/v0.9.0):
 
   - Adds support for Enflame, including `topsAdaptor` and `ecclAdaptor`.
   - Extends flagcxCCLAdaptor to support symmetric operations.
   - Introduces the NCCL Device API in ncclAdaptor to enable customized AllReduce operations.
   - Refactors `glooAdaptor` to support both TCP and IB transports, with automatic NIC detection.
 
-- **[2025/12]** Released [v0.8](https://github.com/flagos-ai/FlagCX/releases/tag/v0.8.0):
+- **[2025/12]** Released [v.0.8.0](https://github.com/flagos-ai/FlagCX/releases/tag/v0.8.0):
 
   - Enables intra-node zero-copy to improve data transfer efficiency for small messages.
   - Supports a naive AllReduce implementation in uniRunner mode using a CPU-centric, device-assisted algorithm.
   - Adds one-sided communication primitives via the new APIs flagcxHeteroPut and flagcxHeteroPutSignal.
 
-- **[2025/11]** Released [v0.7](https://github.com/flagos-ai/FlagCX/releases/tag/v0.7.0):
+- **[2025/11]** Released [v.0.7.0](https://github.com/flagos-ai/FlagCX/releases/tag/v0.7.0):
 
   - Added support to TsingMicro, including device adaptor `tsmicroAdaptor` and CCL adaptor `tcclAdaptor`.
   - Implemented an experimental kernel-free non-reduce collective communication
@@ -58,21 +68,21 @@
     *AllReduce*, *AllGather*, *ReduceScatter*, and *AlltoAll*.
   - Enhanced `flagcxNetAdaptor` with one-sided primitives (`put`, `putSignal`, `waitValue`) and added retransmission support for reliability improvement.
 
-- **[2025/10]** Released [v0.6](https://github.com/flagos-ai/FlagCX/releases/tag/v0.6.0):
+- **[2025/10]** Released [v.0.6.0](https://github.com/flagos-ai/FlagCX/releases/tag/v0.6.0):
 
   - Implemented device-buffer IPC communication to support intra-node *SendRecv* operations.
   - Introduced _device-initiated, host-launched device-side primitives_, enabling kernel-based communication directly from devices.
   - Enhanced auto-tuning with 50% performance improvement on MetaX platforms for the *AllReduce* operations.
 
-- **[2025/09]** Released [v0.5](https://github.com/flagos-ai/FlagCX/releases/tag/v0.5.0):
+- **[2025/09]** Released [v.0.5.0](https://github.com/flagos-ai/FlagCX/releases/tag/v0.5.0):
 
   - Added support for AMD GPUs, including a device adaptor `hipAdaptor` and a CCL adaptor `rcclAdaptor`.
   - Introduced `flagcxNetAdaptor` to unify network backends, currently supporting socket, IBRC, UCX and IBUC (experimental).
   - Enabled zero-copy device-buffer RDMA (user-buffer RDMA) to boost performance for small messages.
   - Supported auto-tuning in homogeneous scenarios via `flagcxTuner`.
   - Added test automation in CI/CD for PyTorch APIs.
 
-- **[2025/08]** Released [v0.4](https://github.com/flagos-ai/FlagCX/releases/tag/v0.4.0):
+- **[2025/08]** Released [v.0.4.0](https://github.com/flagos-ai/FlagCX/releases/tag/v0.4.0):
 
   - Supported heterogeneous training of ERNIE4.5 (Baidu) on NVIDIA and Iluvatar GPUs with Paddle + FlagCX.
   - Improved heterogeneous communication across arbitrary NIC configurations,
@@ -82,20 +92,20 @@
   - Added an InterOp-level DSL to enable customized C2C algorithm design.
   - Provided user documentation under `docs/`.
 
-- **[2025/07]** Released [v0.3](https://github.com/flagos-ai/FlagCX/releases/tag/v0.3.0):
+- **[2025/07]** Released [v.0.3.0](https://github.com/flagos-ai/FlagCX/releases/tag/v0.3.0):
 
   - Integrated three additional native communication libraries: HCCL (Huawei), MUSACCL (Moore Threads) and MPI.
   - Enhanced heterogeneous collective communication operations with pipeline optimizations. 
   - Introduced _device-side_ functions to enable device-buffer RDMA, complementing the existing _host-side_ functions.
   - Delivered a full-stack open-source solution, FlagScale + FlagCX, for efficient heterogeneous prefilling-decoding disaggregation.
 
-- **[2025/05]** Released [v0.2](https://github.com/flagos-ai/FlagCX/releases/tag/v0.2.0):
+- **[2025/05]** Released [v.0.2.0](https://github.com/flagos-ai/FlagCX/releases/tag/v0.2.0):
 
   - Integrated 3 additional native communications libraries, including MCCL (Moore Threads), XCCL (Mellanox) and DUCCL (BAAI).
   - Improved 11 heterogeneous collective communication operations with automatic topology detection and
     full support to single-NIC and multi-NIC environments.
 
-- **[2025/04]** Released [v0.1](https://github.com/flagos-ai/FlagCX/releases/tag/v0.1.0):
+- **[2025/04]** Released [v.0.1.0](https://github.com/flagos-ai/FlagCX/releases/tag/v0.1.0):
 
   - Added 5 native communications libraries including CCL adaptors for 
     NCCL (NVIDIA), IXCCL (Iluvatar), and CNCL (Cambricon), 
@@ -104,5 +114,3 @@
   - Provided a full-stack open-source solution, FlagScale + FlagCX, for efficient heterogeneous training.
   - Natively integrated into PaddlePaddle [v3.0.0](https://github.com/PaddlePaddle/Paddle/tree/v3.0.0),
     with support for both dynamic and static graphs.
-
-
diff --git a/docs/flagcx_en/build.md b/docs/flagcx_en/build.md
@@ -21,9 +21,11 @@ pip install . -v --no-build-isolation
 ```shell
 make <backend>=1 -j$(nproc)
 ```
+
 where `<backend>` is one of:
 - `USE_NVIDIA`: NVIDIA GPU support
-- `USE_ILUVATAR_COREX`: Iluvatar Corex support
+- `USE_ILUVATAR`: Iluvatar GPU support
+- `USE_ILUVATAR_COREX`: deprecated compatibility alias for `USE_ILUVATAR`
 - `USE_CAMBRICON`: Cambricon support
 - `USE_METAX`: MetaX support
 - `USE_MUSA`: Moore Threads support
@@ -34,11 +36,23 @@ where `<backend>` is one of:
 - `USE_TSM`: TsingMicro support
 - `USE_ENFLAME`: Enflame support
 - `USE_SUNRISE`: Sunrise AI support
+- `USE_PPU`: PPU backend support
 - `USE_GLOO`: GLOO support
 - `USE_MPI`: MPI support
 
-Note that Option A also supports `<backend>=1`, allowing users to explicitly specify the backend. Otherwise, it will be selected automatically.
+Device API and integration build controls include:
+
+- `USE_SHMEM=1`: enable the SHMEM Device API adaptor.
+- `SHMEM_HOME`: SHMEM installation path; defaults to `/usr/local/nvshmem`.
+- `USE_ACCL_BAREX=1`: enable the ACCL/Barex network adaptor.
+- `COMPILE_KERNEL=1`: compile kernel-enabled Device API components and tests.
+- `HOST_CXX_STANDARD`: override the host C++ language standard when required by the build environment.
+- `HOST_CXXFLAGS`: add host compiler flags.
+- `JSON_INCLUDE_DIR`: override the JSON header include directory.
+
+The PPU integration commonly uses `USE_PPU=1 USE_ACCL_BAREX=1`; vendor runtime and transport settings are integration-specific and are not universal defaults.
+
+Option A also supports `<backend>=1`, allowing users to explicitly specify the backend. Otherwise, it will be selected automatically.
 
-The default installation path is set to `build/`, you can manually set `BUILDDIR` environment variable to customize the build path.
-You may also specify `DEVICE_HOME` and/or `CCL_HOME` to indicate the installation paths of the device runtime and installation path
-of the communication libraries respectively.
+The default installation path is set to `build/`; you can manually set the `BUILDDIR` environment variable to customize the build path.
+You may also specify `DEVICE_HOME` and/or `CCL_HOME` to indicate the installation paths of the device runtime and communication libraries, respectively.