From 34fa14584e637484aa02b66394cdbdd1369d23ff Mon Sep 17 00:00:00 2001
From: Jerry Mannil <65309407+jerrymannil@users.noreply.github.com>
Date: Tue, 14 Oct 2025 16:33:08 -0700
Subject: [PATCH] [ROCm] Adjust grid size for non-unit stride backwards
 indexing (#2710)

cherry-pick of
https://github.com/pytorch/pytorch/commit/01a2812f48e719f86f6da1c5cfd7a56df7619ac9
---
 aten/src/ATen/native/cuda/Indexing.cu | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/aten/src/ATen/native/cuda/Indexing.cu b/aten/src/ATen/native/cuda/Indexing.cu
index df43bfac16f7..dacef18c79b6 100644
--- a/aten/src/ATen/native/cuda/Indexing.cu
+++ b/aten/src/ATen/native/cuda/Indexing.cu
@@ -710,6 +710,9 @@ void index_put_with_sort_kernel(Tensor & self, const c10::List<std::optional<Ten
       dim3 block(warp_size, indices_per_block);
 
 #ifdef USE_ROCM
+      dim3 new_grid_many_indices(ceil_div(num_indices, (int64_t) (indices_per_block * warp_size)),
+      grid.y == 1 ? std::min<int>(at::cuda::getCurrentDeviceProperties()->maxGridSize[1], ceil_div(sliceSize, (int64_t) (warp_size))) : grid.y,
+      grid.z);
       dim3 new_grid(ceil_div(num_indices, (int64_t) (indices_per_block * warp_size)), grid.y, grid.z);
       size_t smem_dups_size = indices_per_block * warp_size * sizeof(int64_t);
 #define KERNEL_GRID new_grid
@@ -788,7 +791,7 @@ void index_put_with_sort_kernel(Tensor & self, const c10::List<std::optional<Ten
               expandedValue.scalar_type(),
               "indexing_backward_many_indices",
               AT_WRAP([&] {
-                indexing_backward_kernel_many_indices<scalar_t, UNROLL><<<new_grid, block, smem_dups_size, stream>>>(
+                indexing_backward_kernel_many_indices<scalar_t, UNROLL><<<new_grid_many_indices, block, smem_dups_size, stream>>>(
                   sorted_indices.const_data_ptr<int64_t>(),
                   orig_indices.const_data_ptr<int64_t>(),
                   expandedValue.const_data_ptr<scalar_t>(),