Update qutlass cmake for stable abi (#47879)

Signed-off-by: Chris Leonard <chleonar@redhat.com>

cmake/external_projects/qutlass.cmake

@@ -22,7 +22,7 @@ if(QUTLASS_SRC_DIR)
   set(qutlass_BINARY_DIR "${CMAKE_BINARY_DIR}/qutlass-binary-dir-unused")
 else()
   set(_QUTLASS_UPSTREAM_REPO "https://github.com/IST-DASLab/qutlass.git")
-  set(_QUTLASS_UPSTREAM_TAG "830d2c4537c7396e14a02a46fbddd18b5d107c65")
+  set(_QUTLASS_UPSTREAM_TAG "e74319e3405ce6d71965732880f5dc1f52371f64")
 
   set(_qutlass_fc_root "${FETCHCONTENT_BASE_DIR}")
   if(NOT _qutlass_fc_root)
@@ -125,8 +125,6 @@ if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND QUTLASS_ARCHS)
     CUDA_ARCHS "${QUTLASS_ARCHS}"
   )
 
-  # QuTLASS uses legacy ATen headers and cannot be built with TORCH_TARGET_VERSION.
-  # Keep it as its own extension (registers torch.ops._qutlass_C).
   define_extension_target(
     _qutlass_C
     DESTINATION vllm
@@ -139,9 +137,11 @@ if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND QUTLASS_ARCHS)
     WITH_SOABI)
 
   target_compile_definitions(_qutlass_C PRIVATE
-    QUTLASS_DISABLE_PYBIND=1
+    QUTLASS_MINIMAL_BUILD=1
     TARGET_CUDA_ARCH=${QUTLASS_TARGET_CC}
-    CUTLASS_ENABLE_DIRECT_CUDA_DRIVER_CALL=1)
+    CUTLASS_ENABLE_DIRECT_CUDA_DRIVER_CALL=1
+    TORCH_TARGET_VERSION=0x020B000000000000ULL
+    USE_CUDA)
 
   set_property(SOURCE ${QUTLASS_SOURCES} APPEND PROPERTY COMPILE_OPTIONS
     $<$<COMPILE_LANGUAGE:CUDA>:--expt-relaxed-constexpr --use_fast_math -O3>

vllm/_custom_ops.py

@@ -3863,10 +3863,10 @@ def fusedQuantizeMx(
         raise ValueError(f"invalid method {method!r}, must be 'quest' or 'abs_max'")
 
 
-if hasattr(torch.ops._qutlass_C, "fusedQuantizeNv"):
+if hasattr(torch.ops._qutlass_C, "fusedQuantizeNvAbsMax"):
 
-    @register_fake("_qutlass_C::fusedQuantizeNv")
-    def _fake_fused_quantize_nv(
+    @register_fake("_qutlass_C::fusedQuantizeNvAbsMax")
+    def _fake_fused_quantize_nv_absmax(
         a: torch.Tensor,
         b: torch.Tensor,
         xh_e2m1: torch.Tensor,
@@ -3892,7 +3892,9 @@ def fusedQuantizeNv(
         padded_rows, padded_cols, dtype=torch.float8_e4m3fn, device=a.device
     )
 
-    return torch.ops._qutlass_C.fusedQuantizeNv(a, b, xh_e2m1, xh_e4m3, global_scale)
+    return torch.ops._qutlass_C.fusedQuantizeNvAbsMax(
+        a, b, xh_e2m1, xh_e4m3, global_scale
+    )
 
 
 def hadacore_transform(x: torch.Tensor, inplace: bool = True) -> torch.Tensor: