Don't const-fold for multibroadcast

ROCm · Feb 14, 2025 · 3fd27d3 · 3fd27d3
1 parent bbc964e
commit 3fd27d3
Showing 1 changed file with 4 additions and 3 deletions.
diff --git a/src/targets/gpu/fuse_ops.cpp b/src/targets/gpu/fuse_ops.cpp
@@ -673,11 +673,12 @@ struct find_rocblas_gemm_pointwise : gemm_pointwise
 
         if(ins->inputs().size() == 3)
         {
-            auto c_ins = r.instructions["c"];
-            shape s    = c_ins->get_shape();
+            auto c_ins      = r.instructions["c"];
+            shape s         = c_ins->get_shape();
+            auto c_ins_name = c_ins->get_operator().name();
             // const-fold input if not standard shape since rocblas can't handle it
             // Updated for a case where "standard" shape has out-of-sequence strides
-            if(not s.standard() or s.normalize_standard() != s)
+            if((not s.standard() or s.normalize_standard() != s) and c_ins_name != "multibroadcast")
             {
                 auto c = make_op("contiguous");
                 auto l = c.compute(c.compute_shape({c_ins->get_shape()}), {c_ins->eval()});