[fbgemm_gpu] Small docs fixes

- Small docs fixes

.github/scripts/fbgemm_gpu_test.bash

@@ -212,15 +212,16 @@ run_fbgemm_gpu_tests () {
 
 test_setup_conda_environment () {
   local env_name="$1"
-  local python_version="$2"
-  local pytorch_installer="$3"
-  local pytorch_version="$4"
-  local pytorch_variant_type="$5"
-  local pytorch_variant_version="$6"
+  local compiler="$2"
+  local python_version="$3"
+  local pytorch_installer="$4"
+  local pytorch_version="$5"
+  local pytorch_variant_type="$6"
+  local pytorch_variant_version="$7"
   if [ "$pytorch_variant_type" == "" ]; then
     echo "Usage: ${FUNCNAME[0]} ENV_NAME PYTHON_VERSION PYTORCH_INSTALLER PYTORCH_VERSION PYTORCH_VARIANT_TYPE [PYTORCH_VARIANT_VERSION]"
     echo "Example(s):"
-    echo "    ${FUNCNAME[0]} build_env 3.12 pip test cuda 12.1.0       # Setup environment with pytorch-test for Python 3.12 + CUDA 12.1.0"
+    echo "    ${FUNCNAME[0]} build_env clang 3.12 pip test cuda 12.1.0       # Setup environment with pytorch-test for Clang + Python 3.12 + CUDA 12.1.0"
     return 1
   else
     echo "################################################################################"
@@ -243,19 +244,20 @@ test_setup_conda_environment () {
 
   # Set up the build tools and/or GPU runtimes
   if [ "$pytorch_variant_type" == "cuda" ]; then
-    install_cxx_compiler      "${env_name}"                                                                         || return 1
+    install_cxx_compiler      "${env_name}" "${compiler}"                                                           || return 1
     install_build_tools       "${env_name}"                                                                         || return 1
     install_cuda              "${env_name}" "${pytorch_variant_version}"                                            || return 1
     install_cudnn             "${env_name}" "${HOME}/cudnn-${pytorch_variant_version}" "${pytorch_variant_version}" || return 1
 
   elif [ "$pytorch_variant_type" == "rocm" ]; then
     install_rocm_ubuntu       "${env_name}" "${pytorch_variant_version}"  || return 1
+    install_cxx_compiler      "${env_name}" "${compiler}"                 || return 1
     install_build_tools       "${env_name}"                               || return 1
     return 1
 
   else
-    install_cxx_compiler      "${env_name}" || return 1
-    install_build_tools       "${env_name}" || return 1
+    install_cxx_compiler      "${env_name}" "${compiler}" || return 1
+    install_build_tools       "${env_name}"               || return 1
   fi
 
   # Install PyTorch

.github/scripts/test_torchrec.bash

@@ -91,7 +91,7 @@ else
 fi
 
 # shellcheck disable=SC2086
-test_setup_conda_environment "$env_name" "$python_version" pip "$pytorch_channel_name" $pytorch_variant
+test_setup_conda_environment "$env_name" gcc "$python_version" pip "$pytorch_channel_name" $pytorch_variant
 
 # Comment out FBGEMM_GPU since we will install it from "$fbgemm_wheel_path"
 sed -i 's/fbgemm-gpu/#fbgemm-gpu/g' requirements.txt

fbgemm_gpu/fbgemm_gpu/docs/table_batched_embedding_ops.py

@@ -14,7 +14,7 @@
     """
 SplitTableBatchedEmbeddingBagsCodegen(embedding_specs, feature_table_map=None, cache_algorithm=CacheAlgorithm.LRU, cache_load_factor=0.2, cache_sets=0, cache_reserved_memory=0.0, cache_precision=SparseType.FP32, weights_precision=SparseType.FP32, output_dtype=SparseType.FP32, enforce_hbm=False, optimizer=OptimType.EXACT_SGD, record_cache_metrics=None, stochastic_rounding=True, gradient_clipping=False, max_gradient=1.0, learning_rate=0.01, eps=1.0e-8, momentum=0.9, weight_decay=0.0, weight_decay_mode=WeightDecayMode.NONE, eta=0.001, beta1=0.9, beta2=0.999, pooling_mode=PoolingMode.SUM, device=None, bounds_check_mode=BoundsCheckMode.WARNING) -> None
 
-Table batched Embedding operator.  Looks up one or more embedding tables. The module is application for training. The backward operator is fused with optimizer. Thus, the embedding tables are updated during backward.
+Table Batched Embedding (TBE) operator.  Looks up one or more embedding tables. The module is application for training. The backward operator is fused with optimizer. Thus, the embedding tables are updated during backward.
 
 Args:
     embedding_specs (List[Tuple[int, int, EmbeddingLocation, ComputeDevice]]): A list of embedding specifications. Each spec is a tuple of (number of embedding rows, embedding dimension; must be a multiple of 4, table placement, compute device).

fbgemm_gpu/fbgemm_gpu/split_table_batched_embeddings_ops_training.py

@@ -277,6 +277,9 @@ def apply_split_helper(
 # pyre-fixme[13]: Attribute `local_uvm_cache_stats` is never initialized.
 class SplitTableBatchedEmbeddingBagsCodegen(nn.Module):
     """
+    Table Batched Embedding (TBE) operator.  Please see
+    docs/table_batched_embedding_ops.py for the extended documentation.
+
     Multiple sparse features can share one embedding table.
     'feature_table_map' specifies the feature-table mapping.
     T:  number of logical tables

fbgemm_gpu/include/fbgemm_gpu/cumem_utils.h

@@ -136,7 +136,7 @@ Tensor uvm_to_device(const Tensor& self, const Tensor& prototype);
 ///
 /// @see See <a
 /// href="https://nvidia.github.io/cuda-python/module/cudart.html#cuda.cudart.cudaMemAdvise">here</a>
-/// For more information on the `cudaMemoryAdvise` enum.
+/// for more information on the `cudaMemoryAdvise` enum.
 void uvm_cuda_mem_advise(const Tensor& self, int64_t cuda_memory_advise);
 
 /// @ingroup cumem-utils
@@ -150,7 +150,7 @@ void uvm_cuda_mem_advise(const Tensor& self, int64_t cuda_memory_advise);
 ///
 /// @see See <a
 /// href="https://nvidia.github.io/cuda-python/module/cudart.html#cuda.cudart.cudaMemPrefetchAsync">here</a>
-/// For more information on `cudaMemPrefetchAsync()`.
+/// for more information on `cudaMemPrefetchAsync()`.
 void uvm_cuda_mem_prefetch_async(
     const Tensor& self,
     c10::optional<Tensor> device_t);
@@ -165,7 +165,7 @@ void uvm_cuda_mem_prefetch_async(
 /// @param self The input tensor
 ///
 /// @see See <a
-/// href="https://man7.org/linux/man-pages/man2/madvise.2.html">here</a> For
+/// href="https://man7.org/linux/man-pages/man2/madvise.2.html">here</a> for
 /// more information on `madvise()`.
 void uvm_mem_advice_dont_fork(const Tensor& self);