[bugfix]: gate SM103 VSA build on CUDA 12.9

Keep CUDA 12.8 GB200 builds from receiving the unsupported compute_103a
gencode. Export whether the extension actually contains an SM103a image so
runtime dispatch falls back safely when a CUDA 12.8 build is moved to B300.

Author: SolitaryThinker

fastvideo-kernel/CMakeLists.txt

@@ -397,6 +397,7 @@ if(BUILD_CXX_KERNELS)
     # the cmake variable stays empty, so testing that alone silently skips the kernel and
     # leaves a build that succeeds with the op missing.
     set(ENABLE_VSA_SM100_FAMILY OFF)
+    set(ENABLE_VSA_SM103A OFF)
     set(_VSA_ARCH_LIST "${TORCH_CUDA_ARCH_LIST}")
     if(NOT _VSA_ARCH_LIST AND DEFINED ENV{TORCH_CUDA_ARCH_LIST})
         set(_VSA_ARCH_LIST "$ENV{TORCH_CUDA_ARCH_LIST}")
@@ -405,12 +406,32 @@ if(BUILD_CXX_KERNELS)
         set(ENABLE_VSA_SM100_FAMILY ON)
     endif()
     if(ENABLE_VSA_SM100_FAMILY)
-        message(STATUS "fastvideo-kernel: building block_sparse_sm100a (sm_100a/sm_103a, 64- and 128-token blocks)")
+        set(_VSA_COMPILE_OPTIONS
+            "-gencode"
+            "arch=compute_100a,code=sm_100a"
+        )
+        set(_VSA_NATIVE_ARCHS "sm_100a")
+        # CUDA 12.9 introduced the SM103 compiler target. Keep CUDA 12.8
+        # GB200 builds working instead of passing nvcc an unknown compute_103a.
+        if(NOT CUDAToolkit_VERSION VERSION_LESS 12.9)
+            list(APPEND _VSA_COMPILE_OPTIONS
+                "-gencode"
+                "arch=compute_103a,code=sm_103a"
+            )
+            list(APPEND _VSA_NATIVE_ARCHS "sm_103a")
+            set(ENABLE_VSA_SM103A ON)
+        endif()
+        list(APPEND _VSA_COMPILE_OPTIONS "-DVSA_BHSD=true")
+        string(JOIN "/" _VSA_NATIVE_ARCHS_DISPLAY ${_VSA_NATIVE_ARCHS})
+        message(STATUS
+            "fastvideo-kernel: building block_sparse_sm100a "
+            "(${_VSA_NATIVE_ARCHS_DISPLAY}, 64- and 128-token blocks)"
+        )
         list(APPEND EXTENSION_SOURCES csrc/attention/block_sparse_sm100a.cu
                                       csrc/attention/block_sparse_blk128_sm100a.cu)
         set_source_files_properties(csrc/attention/block_sparse_sm100a.cu
                                     csrc/attention/block_sparse_blk128_sm100a.cu PROPERTIES
-            COMPILE_OPTIONS "-gencode;arch=compute_100a,code=sm_100a;-gencode;arch=compute_103a,code=sm_103a;-DVSA_BHSD=true")
+            COMPILE_OPTIONS "${_VSA_COMPILE_OPTIONS}")
     endif()
 
     Python_add_library(fastvideo_kernel_ops MODULE USE_SABI ${SKBUILD_SABI_VERSION} WITH_SOABI
@@ -431,6 +452,9 @@ if(BUILD_CXX_KERNELS)
     if(ENABLE_VSA_SM100_FAMILY)
         list(APPEND COMPILE_DEFS TK_COMPILE_BLOCK_SPARSE_VSA_SM100A)
     endif()
+    if(ENABLE_VSA_SM103A)
+        list(APPEND COMPILE_DEFS TK_COMPILE_BLOCK_SPARSE_VSA_SM103A)
+    endif()
     if(ENABLE_TK_KERNELS)
         list(APPEND COMPILE_DEFS TK_COMPILE_ST_ATTN TK_COMPILE_BLOCK_SPARSE)
     endif()

fastvideo-kernel/csrc/common_extension.cpp

@@ -54,6 +54,11 @@ PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
     m.def("block_sparse_sm100a_blk128_fwd",
           torch::wrap_pybind_function(block_sparse_sm100a_blk128_fwd),
           "VSA block-sparse attention forward, 128-token blocks (Blackwell sm100a/sm103a)");
+#ifdef TK_COMPILE_BLOCK_SPARSE_VSA_SM103A
+    m.attr("_has_vsa_sm103a") = true;
+#else
+    m.attr("_has_vsa_sm103a") = false;
+#endif
 #endif
 
 #ifdef TK_COMPILE_ST_ATTN

fastvideo-kernel/python/fastvideo_kernel/block_sparse_attn_sm100a.py

@@ -1,8 +1,9 @@
 # SPDX-License-Identifier: Apache-2.0
 """Data-center Blackwell CUDA block-sparse VSA forward.
 
-The historical ``sm100a`` module and symbol names are retained for compatibility, but the
-extension carries native sm_100a and sm_103a images and supports both device generations.
+The historical ``sm100a`` module and symbol names are retained for compatibility. The
+extension always carries a native sm_100a image and also carries sm_103a when built with
+CUDA 12.9 or newer.
 
 A third backend behind the same VSA op as the Triton and CuTe-DSL paths. Forward only: it
 returns ``(out, lse)`` with ``lse`` in exactly the form ``triton_block_sparse_attn_forward``
@@ -29,12 +30,16 @@
         128: getattr(_C, "block_sparse_sm100a_blk128_fwd", None),
     }
     _HAS_VSA_SM100A = any(_FWD_BY_BLOCK.values())
+    _HAS_VSA_SM103A = bool(getattr(_C, "_has_vsa_sm103a", False))
 except ImportError:  # pragma: no cover - extension not built
     _C = None
     _FWD_BY_BLOCK = {}
     _HAS_VSA_SM100A = False
+    _HAS_VSA_SM103A = False
 
-_SUPPORTED_COMPUTE_CAPABILITIES = {(10, 0), (10, 3)}
+_SUPPORTED_COMPUTE_CAPABILITIES = {(10, 0)}
+if _HAS_VSA_SM103A:
+    _SUPPORTED_COMPUTE_CAPABILITIES.add((10, 3))
 HEAD_DIM = 128
 # Must match the -DVSA_BHSD the extension was compiled with (see CMakeLists).
 BHSD = True

tests/test_fasth3_packaging.py

@@ -39,3 +39,18 @@ def test_kernel_release_matrix_can_publish_data_center_blackwell_wheels():
     assert "arch=compute_100a,code=sm_100a" in cmake
     assert "arch=compute_103a,code=sm_103a" in cmake
     assert "patchelf==0.17.2.4" in workflow
+
+
+def test_sm103a_gencode_requires_cuda_12_9():
+    cmake = (REPO_ROOT / "fastvideo-kernel" / "CMakeLists.txt").read_text(encoding="utf-8")
+    extension = (REPO_ROOT / "fastvideo-kernel" / "csrc" / "common_extension.cpp").read_text(encoding="utf-8")
+    backend = (REPO_ROOT / "fastvideo-kernel" / "python" / "fastvideo_kernel" /
+               "block_sparse_attn_sm100a.py").read_text(encoding="utf-8")
+    vsa_options = cmake.index("set(_VSA_COMPILE_OPTIONS")
+    sm103_guard = cmake.index("if(NOT CUDAToolkit_VERSION VERSION_LESS 12.9)", vsa_options)
+    sm103_gencode = cmake.index("arch=compute_103a,code=sm_103a", sm103_guard)
+
+    assert sm103_guard < sm103_gencode < cmake.index("endif()", sm103_guard)
+    assert "list(APPEND COMPILE_DEFS TK_COMPILE_BLOCK_SPARSE_VSA_SM103A)" in cmake
+    assert 'm.attr("_has_vsa_sm103a") = true' in extension
+    assert 'getattr(_C, "_has_vsa_sm103a", False)' in backend