Comment tweaks

Signed-off-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

Author: timmoon10

transformer_engine/common/include/transformer_engine/hadamard_transform.h

@@ -115,7 +115,7 @@ void nvte_group_hadamard_transform_cast_fusion_columnwise(
  *  \param[in]      split_sections    Array specifying splits in dimension 0 for each output tensor.
  *  \param[in]      num_tensors       Number of output tensors, must be > 0.
  *  \param[in]      quant_config      Quantization configuration.
- *  \param[in]      quant_workspace   Quantization workspace, a device buffer that can be opaque
+ *  \param[in]      quant_workspace   Workspace buffer. Must be at least 4 bytes.
  *  \param[in]      stream            CUDA stream used for the operation.
  */
 void nvte_group_hadamard_transform_cast_fusion(const NVTETensor input, NVTETensor* outputs,

transformer_engine/pytorch/csrc/extensions/cast.cpp

@@ -878,7 +878,6 @@ void split_quantize_nvfp4_impl_with_rht_helper(const TensorWrapper &input,
     auto nvte_tile_scheduler_workspace =
         makeTransformerEngineTensor(tile_scheduler_workspace_torch);
     // call the fully-fused grouped kernel for rowwise quantization & colwise RHT quantization transpose
-    // Note that the workspace field can be opaque, this will leave room for furture extensions
     nvte_group_hadamard_transform_cast_fusion(
         input.data(), reinterpret_cast<NVTETensor *>(nvte_tensor_output_list.data()),
         rht_matrix_nvte.data(), split_sections.data(), num_tensors, quant_config_list[0],