From 697126019ed5dd59b58bb2f081dd49595192caa5 Mon Sep 17 00:00:00 2001
From: Haicheng Wu <57973641+hwu36@users.noreply.github.com>
Date: Tue, 29 Apr 2025 11:54:20 -0400
Subject: [PATCH] fix blackwell grouped groupwise hang (#2267)

---
 include/cutlass/gemm/kernel/sm90_tile_scheduler_group.hpp | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/include/cutlass/gemm/kernel/sm90_tile_scheduler_group.hpp b/include/cutlass/gemm/kernel/sm90_tile_scheduler_group.hpp
index d896c0698..bf02e5faf 100644
--- a/include/cutlass/gemm/kernel/sm90_tile_scheduler_group.hpp
+++ b/include/cutlass/gemm/kernel/sm90_tile_scheduler_group.hpp
@@ -66,18 +66,18 @@ public:
     int32_t M_idx = 0;
     int32_t N_idx = 0;
     int32_t L_idx = 0;
-    bool is_valid_tile = false;
+    int32_t is_valid_tile = 0;
 
     CUTLASS_HOST_DEVICE
     bool
     is_valid() const {
-      return is_valid_tile;
+      return is_valid_tile != 0;
     }
 
     CUTLASS_HOST_DEVICE
     static WorkTileInfo
     invalid_work_tile() {
-      return {-1, -1, -1, false};
+      return {-1, -1, -1, 0};
     }
 
     CUTLASS_HOST_DEVICE
@@ -304,7 +304,7 @@ public:
       int32_t log_swizzle_size, 
       RasterOrder raster_order) {
 
-    bool valid_tile = true;
+    int32_t valid_tile = 1;
 
     // Use a warp to "speculatively" check if the work tile maps to the next 32 groups
     int lane_idx = canonical_lane_idx();