v4.3.4 update v2. (#2898)

CHANGELOG.md

@@ -3,6 +3,8 @@
 # CUTLASS 4.x
 
 ## [4.3.4](https://github.com/NVIDIA/cutlass/releases/tag/v4.3.4) (2025-12-22)
+
+### CuTe DSL
 * New features
   - Added PDL support along with example [Kernel launch with Programmatic Dependent Launch](https://github.com/NVIDIA/cutlass/tree/main/examples/python/CuTeDSL/blackwell/programmatic_dependent_launch.py)
 
@@ -11,7 +13,12 @@
   - Enhancement for tvm-ffi AoT case for earlier module unload
   - Fixed order issue in `make_smem_layout_a` in utils/hopper_helpers.py
 
+### CUTLASS C++
+* Work around a driver bug which will cause occasionally errors when executing kernels.
+
 ## [4.3.3](https://github.com/NVIDIA/cutlass/releases/tag/v4.3.3) (2025-12-12)
+
+### CuTe DSL
 * New features
   - Supported namedtuple and kwargs for JIT function arguments in tvm-ffi
   - Supported variadic tuples for JIT function argument in tvm-ffi
@@ -21,6 +28,8 @@
   - Clearer error message for the case of runtime error cudaErrorInsufficientDriver
 
 ## [4.3.2](https://github.com/NVIDIA/cutlass/releases/tag/v4.3.2) (2025-12-05)
+
+### CuTe DSL
 * New features
   - New env var `CUTE_DSL_CACHE_DIR` to specify the path for dumping caches
 

README.md

@@ -143,6 +143,7 @@ To get started quickly - please refer :
     - Fix a few bugs in distributed gemm API and examples.
     - Fix handling negative zero in sparse compressor.
     - Add missing `wait_on_dependent_grids` for PDL use case.
+    - Work around a driver bug which will cause occasionally errors when executing kernels.
 * Fix some profiler issues:
     - Add some missing reference kernels.
     - Support VoidC reference kernels.

include/cute/atom/copy_traits_sm90_im2col.hpp

@@ -471,6 +471,17 @@ make_im2col_tma_copy_desc(
       tma_l2Promotion,
       tma_oob_fill);
 
+  int driver_version = 0;
+  CUresult driver_version_result = cuDriverGetVersion(&driver_version);
+  assert(driver_version_result == CUDA_SUCCESS);
+  if (driver_version <= 13010) {
+    if (cute::bits_to_bytes(
+          cute::cosize(tensor_cwhdn.layout()) *
+          cute::sizeof_bits<typename EngineA::value_type>::value) < 131072) {
+      reinterpret_cast<uint64_t*>(&tma_desc)[1] &= ~(1llu << 21);
+    }
+  }
+
   // The extra asserts help indicate the error's cause.
   assert(encode_result != CUDA_ERROR_DEINITIALIZED);
   assert(encode_result != CUDA_ERROR_NOT_INITIALIZED);

include/cute/atom/copy_traits_sm90_tma.hpp

@@ -1052,6 +1052,17 @@ make_tma_copy_desc(Tensor<GEngine,GLayout> const& gtensor,         // The origin
         tma_l2Promotion,
         tma_oobFill);
    
+    int driver_version = 0;
+    CUresult driver_version_result = cuDriverGetVersion(&driver_version);
+    assert(driver_version_result == CUDA_SUCCESS);
+    if (driver_version <= 13010) {      
+      if (cute::bits_to_bytes(
+            cute::cosize(gtensor.layout()) *
+            cute::sizeof_bits<typename GEngine::value_type>::value) < 131072) {
+        reinterpret_cast<uint64_t*>(&tma_desc)[1] &= ~(1llu << 21);
+      }
+    }
+
     if (result != CUDA_SUCCESS) {
       std::cerr << "TMA Desc Addr:   " << &tma_desc
                 << "\nformat         " << tma_format