[tensilelite] Re-apply test component config and fix PyTorch SDK test failure

Re-applies #4991 (reverted in #5520) and fixes #5516.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: archana-ramalingam

build_tools/github_actions/fetch_test_configurations.py

@@ -194,6 +194,16 @@ def _build_container_options(job_config: dict, platform: str) -> dict:
             ],
         },
     },
+    "tensilelite": {
+        "job_name": "tensilelite",
+        "fetch_artifact_args": "--blas --tests",
+        "timeout_minutes": 15,
+        "test_script": f"python {_get_script_path('test_tensilelite.py')}",
+        "platform": ["linux"],
+        "total_shards_dict": {
+            "linux": 1,
+        },
+    },
     "hipblas": {
         "job_name": "hipblas",
         "fetch_artifact_args": "--blas --tests",

build_tools/github_actions/test_executable_scripts/test_tensilelite.py

@@ -0,0 +1,121 @@
+#!/usr/bin/env python3
+# Copyright Advanced Micro Devices, Inc.
+# SPDX-License-Identifier: MIT
+
+"""
+Test runner for TensileLite and rocisa Python tests using pre-built artifacts.
+
+Runs against installed artifacts from the hipBLASLt test component:
+  share/hipblaslt/tensilelite/Tensile/     — Tensile Python package
+  share/hipblaslt/tensilelite/rocisa/       — rocisa Python package + _rocisa.abi3.so
+  share/hipblaslt/tensilelite/rocisa_tests/ — rocisa pytest modules
+
+Test order: rocisa first (build dependency of TensileLite), then TensileLite
+unit tests. A rocisa failure means TensileLite tests will also fail.
+
+Usage (TheRock CI):
+    python test_tensilelite.py
+
+Usage (local, after install):
+    THEROCK_BIN_DIR=./build/bin python test_tensilelite.py
+"""
+
+import logging
+import os
+import subprocess
+import sys
+from pathlib import Path
+
+logging.basicConfig(level=logging.INFO, format="%(message)s")
+
+SCRIPT_DIR = Path(__file__).resolve().parent
+THEROCK_DIR = SCRIPT_DIR.parent.parent.parent
+THEROCK_BIN_DIR = os.getenv("THEROCK_BIN_DIR", "")
+
+rocm_path = Path(THEROCK_BIN_DIR).resolve().parent
+tensilelite_root = rocm_path / "share" / "hipblaslt" / "tensilelite"
+
+if not tensilelite_root.is_dir():
+    raise FileNotFoundError(
+        f"TensileLite test artifacts not found at {tensilelite_root}. "
+        "Ensure the build used -DHIPBLASLT_INSTALL_TENSILELITE_TEST_ARTIFACTS=ON."
+    )
+
+env = os.environ.copy()
+existing_pythonpath = env.get("PYTHONPATH")
+env["PYTHONPATH"] = (
+    f"{tensilelite_root}{os.pathsep}{existing_pythonpath}"
+    if existing_pythonpath
+    else str(tensilelite_root)
+)
+env["ROCM_PATH"] = str(rocm_path)
+
+# _rocisa links libamdhip64.so — ensure HIP libraries are findable.
+lib_path = rocm_path / "lib"
+existing_ld_path = env.get("LD_LIBRARY_PATH", "")
+env["LD_LIBRARY_PATH"] = (
+    f"{lib_path}{os.pathsep}{existing_ld_path}" if existing_ld_path else str(lib_path)
+)
+
+# GPU unit tests use amdclang++ to assemble kernels.
+existing_path = env.get("PATH", "")
+env["PATH"] = os.pathsep.join(
+    filter(
+        None,
+        [
+            str(rocm_path / "bin"),
+            str(rocm_path / "lib" / "llvm" / "bin"),
+            existing_path,
+        ],
+    )
+)
+
+# Smoke test: verify install layout and stable ABI.
+logging.info("=== Verifying artifact install layout ===")
+rocisa_dir = tensilelite_root / "rocisa"
+logging.info(
+    f"rocisa directory contents: {[f.name for f in rocisa_dir.iterdir() if not f.name.startswith('__')]}"
+)
+abi3_files = list(rocisa_dir.glob("*.abi3.*"))
+if abi3_files:
+    logging.info(f"Stable ABI confirmed: {[f.name for f in abi3_files]}")
+else:
+    logging.warning("No .abi3 extension found — stable ABI may not be enabled")
+subprocess.check_call(
+    [
+        sys.executable,
+        "-c",
+        "import Tensile, rocisa, rocisa.instruction; "
+        "print('Tensile:', Tensile.ROOT_PATH); print('rocisa:', rocisa.__file__)",
+    ],
+    cwd=str(THEROCK_DIR),
+    env=env,
+)
+
+# rocisa tests (includes GPU tests — runner has GPU access).
+logging.info("=== Running rocisa tests ===")
+subprocess.check_call(
+    [
+        sys.executable,
+        "-m",
+        "pytest",
+        "-v",
+        str(tensilelite_root / "rocisa_tests"),
+    ],
+    cwd=str(THEROCK_DIR),
+    env=env,
+)
+
+# TensileLite Python unit tests (includes GPU subtile tests).
+logging.info("=== Running TensileLite unit tests ===")
+subprocess.check_call(
+    [
+        sys.executable,
+        "-m",
+        "pytest",
+        "-v",
+        str(tensilelite_root / "Tensile" / "Tests" / "unit"),
+    ],
+    cwd=str(THEROCK_DIR),
+    env=env,
+)

build_tools/packaging/python/templates/rocm/src/rocm_sdk/tests/libraries_test.py

@@ -66,6 +66,15 @@ def testSharedLibrariesLoad(self):
                     # Though this is not needed for the amd-smi client.
                     continue
 
+                if so_path.name.endswith(".abi3.so") or ".cpython-" in so_path.name:
+                    # Python C extensions must be loaded via importlib, not
+                    # ctypes.CDLL. Stable ABI extensions (.abi3.so) target a
+                    # minimum CPython version and may use symbols absent from
+                    # older interpreters (e.g. PyType_FromMetaclass requires
+                    # 3.12+). Version-tagged extensions (.cpython-3XX) are
+                    # similarly incompatible across interpreter versions.
+                    continue
+
                 extra_setup = ""
                 if (
                     "hipdnn_plugins" in str(so_path) or "test_plugins" in str(so_path)

math-libs/BLAS/CMakeLists.txt

@@ -134,6 +134,8 @@ therock_cmake_subproject_declare(hipBLASLt
     -DHIPBLASLT_ENABLE_BLIS=OFF  # TODO: Evaluate
     -DTENSILELITE_BUILD_TESTING=OFF
     -DHIPBLASLT_BUILD_TESTING=${THEROCK_BUILD_TESTING}
+    -DHIPBLASLT_INSTALL_TENSILELITE_TEST_ARTIFACTS=${THEROCK_BUILD_TESTING}
+    -DROCISA_USE_STABLE_ABI=ON
     -DHIPBLASLT_ENABLE_ROCROLLER=${_enable_rocRoller}
     -DHIPBLASLT_ENABLE_FETCH=OFF
     -DHIPBLASLT_ENABLE_THEROCK=ON

math-libs/BLAS/artifact-blas.toml

@@ -84,6 +84,8 @@ include = [
   "share/hipblaslt/performance/**",
   # Built samples (may want these in a samples component at some point).
   "libexec/hipblaslt-samples/**",
+  # TensileLite Python test artifacts
+  "share/hipblaslt/tensilelite/**",
 ]
 
 # rocBLAS

requirements-test.txt

@@ -13,6 +13,13 @@ jsonschema==4.23.0
 packaging==25.0
 python-magic==0.4.27; sys_platform != "win32"
 
+# tensilelite/rocisa test requirements
+msgpack>=1.0.0
+joblib>=1.4.0
+numpy>=1.26.0
+filelock>=3.0.0
+hip-python; sys_platform != "win32"
+
 # libhipcxx test requirement
 lit==18.1.8
 psutil==7.1.3