refactor: enhance ppu detect

Signed-off-by: thxCode <thxcode0824@gmail.com>

Author: thxCode

.github/workflows/ci.yml

@@ -64,6 +64,7 @@ jobs:
       - name: Checkout
         uses: actions/checkout@v4
         with:
+          lfs: true
           fetch-depth: 1
           persist-credentials: false
       - name: Setup UV

.github/workflows/pack.yml

@@ -39,6 +39,7 @@ jobs:
       - name: Checkout
         uses: actions/checkout@v4
         with:
+          lfs: true
           fetch-depth: 0
           fetch-tags: true
           persist-credentials: false

.github/workflows/release.yml

@@ -46,6 +46,7 @@ jobs:
       - name: Checkout
         uses: actions/checkout@v4
         with:
+          lfs: true
           fetch-depth: 0
           fetch-tags: true
           persist-credentials: false

.gitignore

@@ -4,9 +4,6 @@
 .DS_Store
 *.swp
 
-# C extensions
-*.so
-
 # Python-generated files
 _version.py
 _version_appendix.py

gpustack_runtime/detector/iluvatar.py

@@ -54,6 +54,7 @@ def is_supported() -> bool:
         pci_devs = IluvatarDetector.detect_pci_devices()
         if not pci_devs and not envs.GPUSTACK_RUNTIME_DETECT_NO_PCI_CHECK:
             logger.debug("No Iluvatar PCI devices found")
+            return supported
 
         try:
             pyixml.nvmlInit()

gpustack_runtime/detector/pyhgml/__init__.py

@@ -2411,11 +2411,19 @@ def _LoadHgmlLibrary():
                 # Linux path
                 locs = [
                     "libhgml.so",
-                    str(Path(__file__).resolve().parent / "libhgml.so"),
+                    [
+                        str(Path(__file__).resolve().parent / "libuki.so"),
+                        str(Path(__file__).resolve().parent / "libhgml.so"),
+                    ],
                 ]
                 for loc in locs:
                     try:
-                        hgmlLib = CDLL(loc)
+                        if isinstance(loc, str):
+                            hgmlLib = CDLL(loc)
+                        else:
+                            for pre_loc in loc[:-1]:
+                                CDLL(pre_loc)
+                            hgmlLib = CDLL(loc[-1])
                         break
                     except OSError:
                         pass

gpustack_runtime/detector/pyhgml/libhgml.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:04fcc6541a73acc5f5b52bdb5dd76af10d10862f30f8da362575fcb11c538a82
+size 2101640

gpustack_runtime/detector/pyhgml/libuki.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:104eafd6f218613e0548330c9bdad27c0ab0af020f143fb8ffa2ad3f9d654a9b
+size 702352

gpustack_runtime/detector/thead.py

@@ -54,6 +54,7 @@ def is_supported() -> bool:
         pci_devs = THeadDetector.detect_pci_devices()
         if not pci_devs and not envs.GPUSTACK_RUNTIME_DETECT_NO_PCI_CHECK:
             logger.debug("No T-Head PCI devices found")
+            return supported
 
         try:
             pyhgml.hgmlInit()
@@ -68,7 +69,7 @@ def is_supported() -> bool:
     @lru_cache
     def detect_pci_devices() -> dict[str, PCIDevice]:
         # See https://pcisig.com/membership/member-companies?combine=Alibaba.
-        pci_devs = get_pci_devices(vendor=["0x05b7", "0x1ded"])
+        pci_devs = get_pci_devices(vendor="0x1ded")
         if not pci_devs:
             return {}
         return {dev.address: dev for dev in pci_devs}
@@ -98,10 +99,23 @@ def detect(self) -> Devices | None:
 
             sys_driver_ver = pyhgml.hgmlSystemGetDriverVersion()
 
-            sys_runtime_ver_original = pyhgml.hgmlSystemGetHGMLVersion()
-            sys_runtime_ver = get_brief_version(
-                sys_runtime_ver_original,
-            )
+            sys_runtime_ver_original = None
+            sys_runtime_ver = None
+            with contextlib.suppress(pyhgml.HGMLError):
+                sys_runtime_ver_original = pyhgml.hgmlSystemGetHggcDriverVersion()
+                sys_runtime_ver_original = ".".join(
+                    map(
+                        str,
+                        [
+                            sys_runtime_ver_original // 1000,
+                            (sys_runtime_ver_original % 1000) // 10,
+                            (sys_runtime_ver_original % 10),
+                        ],
+                    ),
+                )
+                sys_runtime_ver = get_brief_version(
+                    sys_runtime_ver_original,
+                )
 
             dev_count = pyhgml.hgmlDeviceGetCount()
             for dev_idx in range(dev_count):

pack/Dockerfile

@@ -487,6 +487,21 @@ ENV NVIDIA_DISABLE_REQUIRE="true" \
     NVIDIA_VISIBLE_DEVICES="all" \
     NVIDIA_DRIVER_CAPABILITIES="compute,utility"
 
+## Active all T-Head devices detection,
+## works with (default) T-Head container runtime and privileged mode.
+## See https://help.aliyun.com/document_detail/2996754.html.
+## Runs:
+## - With container runtime installed:
+##   [TODO, TBD]
+## - Without container runtime installed:
+##   + Allowing privileged, try with:
+##     docker run --rm -it -v /var/run/docker.sock:/var/run/docker.sock --privileged -v /usr/local/PPU_SDK:/usr/local/PPU_SDK:ro ...
+##   + Disallowing privileged, try with:
+##     docker run --rm -it -v /var/run/docker.sock:/var/run/docker.sock --security-opt seccomp=unconfined -v /dev:/dev:ro -v /usr/local/PPU_SDK:/usr/local/PPU_SDK:ro ...
+ENV PPU_HOME="/usr/local/PPU_SDK" \
+    LD_LIBRARY_PATH="/usr/local/PPU_SDK/CUDA_SDK/lib64:/usr/local/PPU_SDK/lib:${LD_LIBRARY_PATH}" \
+    GPUSTACK_RUNTIME_DEPLOY_MIRRORED_DEPLOYMENT_IGNORE_VOLUMES="/usr/local/PPU_SDK;${GPUSTACK_RUNTIME_DEPLOY_MIRRORED_DEPLOYMENT_IGNORE_VOLUMES}"
+
 ## Active GPUStack runtime mirrored deployment mode,
 ## if getting an error like, "Found multiple Containers with the same hostname ...",
 ## please use `--env GPUSTACK_RUNTIME_DEPLOY_MIRRORED_NAME=...` to specify the exact container name.