Typo and small fixes to CK fp8 rowwise grouped (pytorch#4550)

Summary:
Pull Request resolved: pytorch#4550

X-link: facebookresearch/FBGEMM#1593

- Typos+grammar fix found by LLM
- Validation bug found by LLM
- Small logic simplification I missed in my prior PR.

Differential Revision: D78827450

Author: cthi

fbgemm_gpu/experimental/gen_ai/src/quantize/ck_extensions/fp8_rowwise_grouped/fp8_rowwise_grouped_gemm.hip

@@ -109,7 +109,7 @@ void set_static_kernel_args(
   int64_t output_offset = 0;
   // When group count is large, we can more efficiently initialize
   // by doing host setup and a memcpy. This is only viable if cuda
-  // graphs arent being used.
+  // graphs aren't being used.
   // Iterate over inputs and get group information.
   for (int i = 0; i < group_count; i++) {
     int64_t M = XQ[i].size(0);
@@ -163,7 +163,7 @@ __global__ void set_kernel_args(
     int64_t K,
     int64_t group_count,
     std::optional<GroupedGemmInputType> input_type = std::nullopt) {
-  // The "message" part seems not working on AMD currently :(
+  // The "message" part is not working on AMD currently :(
   CUDA_KERNEL_ASSERT_MSG((M_sizes == nullptr && offsets == nullptr) || (M_sizes == nullptr ^ offsets == nullptr), "Cannot set both M_sizes and offsets");
   CUDA_KERNEL_ASSERT_MSG(input_type.has_value() || M_sizes != nullptr, "M_sizes should not be used with input_type");
 
@@ -513,7 +513,7 @@ OutputType _f8f8bf16_rowwise_grouped(
   for (at::Tensor xs : x_scale) {
     TORCH_CHECK(xs.dtype() == at::kFloat, "Scales must be float32.");
   }
-  for (at::Tensor ws : x_scale) {
+  for (at::Tensor ws : w_scale) {
     TORCH_CHECK(ws.dtype() == at::kFloat, "Scales must be float32.");
   }
 
@@ -774,7 +774,7 @@ at::Tensor f8f8bf16_rowwise_grouped_mm(
     TORCH_CHECK(w_scale.size(0) == G && w_scale.size(1) == N, "w_scale shape must be (G, N).");
     TORCH_CHECK(out.dim() == 3 && out.size(0) == G && out.size(1) == M && out.size(2) == N, "out shape must be (G, M, N).");
   } else if (XQ.dim() == 2 && WQ.dim() == 2) {
-    TORCH_CHECK(offsets.has_value(), "Must pass offsets for 2D inputs XQ nd WQ.");
+    TORCH_CHECK(offsets.has_value(), "Must pass offsets for 2D inputs XQ and WQ.");
     TORCH_CHECK(offsets->dtype() == at::kInt, "offsets must be int32.");
 
     G = offsets->size(0);

fbgemm_gpu/experimental/gen_ai/src/quantize/ck_extensions/fp8_rowwise_grouped/kernels/fp8_rowwise_grouped_common.h

@@ -118,13 +118,11 @@ struct DeviceGemmHelper {
     // Get input information.
     int group_count;
     if constexpr (std::is_same_v<InputType, at::Tensor>) {
-      if (WQ.dim() == 3) {
-        // If WQ is 3D the group count is the min of G and total_M (if XQ is
-        // 2D).
-        group_count = std::min(WQ.size(0), XQ.size(0));
-      } else if (XQ.dim() == 3) {
-        // If XQ is 3D the group count is the min of G and total_N (if WQ is
-        // 2D).
+      if (XQ.dim() == 3 || WQ.dim() == 3) {
+        // If WQ and XQ are 3D, the group count is G.
+        // If WQ is 3D and XQ is 2D (and the reverse by symmetry), the group
+        // count is the minimum of G and total_M/total_N. In all cases we just
+        // compare the first dimension of XQ and WQ.
         group_count = std::min(XQ.size(0), WQ.size(0));
       } else {
         // XQ and WQ are 2D. The group count is G.
@@ -163,7 +161,7 @@ struct DeviceGemmHelper {
       // pointers below are unused, as the device memory contains the correct
       // data.
       if constexpr (std::is_same_v<InputType, at::Tensor>) {
-        // Set these to 0 as placeholders, they are unsused.
+        // Set these to 0 as placeholders, they are unused.
         M = 0;
         N = 0;
         K = 0;