ROCm · valarLip · Nov 26, 2025 · Nov 19, 2025 · Nov 19, 2025 · Nov 20, 2025
diff --git a/hsa/gfx942/fmha_v3_fwd/MI300/fwd_hd128_bf16_causal_rtna.co b/hsa/gfx942/fmha_v3_fwd/MI300/fwd_hd128_bf16_causal_rtna.co
diff --git a/hsa/gfx942/fmha_v3_fwd/MI300/fwd_hd128_bf16_causal_rtna_group.co b/hsa/gfx942/fmha_v3_fwd/MI300/fwd_hd128_bf16_causal_rtna_group.co
diff --git a/hsa/gfx942/fmha_v3_fwd/MI300/fwd_hd128_bf16_causal_rtne.co b/hsa/gfx942/fmha_v3_fwd/MI300/fwd_hd128_bf16_causal_rtne.co
diff --git a/hsa/gfx942/fmha_v3_fwd/MI300/fwd_hd128_bf16_causal_rtne_group.co b/hsa/gfx942/fmha_v3_fwd/MI300/fwd_hd128_bf16_causal_rtne_group.co
diff --git a/hsa/gfx942/fmha_v3_fwd/MI300/fwd_hd128_bf16_causal_rtz.co b/hsa/gfx942/fmha_v3_fwd/MI300/fwd_hd128_bf16_causal_rtz.co
diff --git a/hsa/gfx942/fmha_v3_fwd/MI300/fwd_hd128_bf16_causal_rtz_group.co b/hsa/gfx942/fmha_v3_fwd/MI300/fwd_hd128_bf16_causal_rtz_group.co
diff --git a/hsa/gfx942/fmha_v3_fwd/MI300/fwd_hd128_bf16_rtna.co b/hsa/gfx942/fmha_v3_fwd/MI300/fwd_hd128_bf16_rtna.co
diff --git a/hsa/gfx942/fmha_v3_fwd/MI300/fwd_hd128_bf16_rtna_group.co b/hsa/gfx942/fmha_v3_fwd/MI300/fwd_hd128_bf16_rtna_group.co
diff --git a/hsa/gfx942/fmha_v3_fwd/MI300/fwd_hd128_bf16_rtne.co b/hsa/gfx942/fmha_v3_fwd/MI300/fwd_hd128_bf16_rtne.co
diff --git a/hsa/gfx942/fmha_v3_fwd/MI300/fwd_hd128_bf16_rtne_group.co b/hsa/gfx942/fmha_v3_fwd/MI300/fwd_hd128_bf16_rtne_group.co
diff --git a/hsa/gfx942/fmha_v3_fwd/MI300/fwd_hd128_bf16_rtz.co b/hsa/gfx942/fmha_v3_fwd/MI300/fwd_hd128_bf16_rtz.co
diff --git a/hsa/gfx942/fmha_v3_fwd/MI300/fwd_hd128_bf16_rtz_group.co b/hsa/gfx942/fmha_v3_fwd/MI300/fwd_hd128_bf16_rtz_group.co
diff --git a/hsa/gfx942/fmha_v3_fwd/MI308/fwd_hd128_bf16_causal_rtna.co b/hsa/gfx942/fmha_v3_fwd/MI308/fwd_hd128_bf16_causal_rtna.co
diff --git a/hsa/gfx942/fmha_v3_fwd/MI308/fwd_hd128_bf16_causal_rtna_group.co b/hsa/gfx942/fmha_v3_fwd/MI308/fwd_hd128_bf16_causal_rtna_group.co
diff --git a/hsa/gfx942/fmha_v3_fwd/MI308/fwd_hd128_bf16_causal_rtne.co b/hsa/gfx942/fmha_v3_fwd/MI308/fwd_hd128_bf16_causal_rtne.co
diff --git a/hsa/gfx942/fmha_v3_fwd/MI308/fwd_hd128_bf16_causal_rtne_group.co b/hsa/gfx942/fmha_v3_fwd/MI308/fwd_hd128_bf16_causal_rtne_group.co
diff --git a/hsa/gfx942/fmha_v3_fwd/MI308/fwd_hd128_bf16_causal_rtz.co b/hsa/gfx942/fmha_v3_fwd/MI308/fwd_hd128_bf16_causal_rtz.co
diff --git a/hsa/gfx942/fmha_v3_fwd/MI308/fwd_hd128_bf16_causal_rtz_group.co b/hsa/gfx942/fmha_v3_fwd/MI308/fwd_hd128_bf16_causal_rtz_group.co
diff --git a/hsa/gfx942/fmha_v3_fwd/MI308/fwd_hd128_bf16_rtna.co b/hsa/gfx942/fmha_v3_fwd/MI308/fwd_hd128_bf16_rtna.co
diff --git a/hsa/gfx942/fmha_v3_fwd/MI308/fwd_hd128_bf16_rtna_group.co b/hsa/gfx942/fmha_v3_fwd/MI308/fwd_hd128_bf16_rtna_group.co
diff --git a/hsa/gfx942/fmha_v3_fwd/MI308/fwd_hd128_bf16_rtne.co b/hsa/gfx942/fmha_v3_fwd/MI308/fwd_hd128_bf16_rtne.co
diff --git a/hsa/gfx942/fmha_v3_fwd/MI308/fwd_hd128_bf16_rtne_group.co b/hsa/gfx942/fmha_v3_fwd/MI308/fwd_hd128_bf16_rtne_group.co
diff --git a/hsa/gfx942/fmha_v3_fwd/MI308/fwd_hd128_bf16_rtz.co b/hsa/gfx942/fmha_v3_fwd/MI308/fwd_hd128_bf16_rtz.co
diff --git a/hsa/gfx942/fmha_v3_fwd/MI308/fwd_hd128_bf16_rtz_group.co b/hsa/gfx942/fmha_v3_fwd/MI308/fwd_hd128_bf16_rtz_group.co
diff --git a/hsa/gfx950/fmha_v3_fwd/fwd_hd128_bf16.co b/hsa/gfx950/fmha_v3_fwd/fwd_hd128_bf16.co
diff --git a/hsa/gfx950/fmha_v3_fwd/fwd_hd128_bf16_causal.co b/hsa/gfx950/fmha_v3_fwd/fwd_hd128_bf16_causal.co
diff --git a/hsa/gfx950/fmha_v3_fwd/fwd_hd128_bf16_causal_group.co b/hsa/gfx950/fmha_v3_fwd/fwd_hd128_bf16_causal_group.co
diff --git a/hsa/gfx950/fmha_v3_fwd/fwd_hd128_bf16_group.co b/hsa/gfx950/fmha_v3_fwd/fwd_hd128_bf16_group.co
diff --git a/hsa/gfx950/fmha_v3_fwd/fwd_hd192_hd128_bf16.co b/hsa/gfx950/fmha_v3_fwd/fwd_hd192_hd128_bf16.co
diff --git a/hsa/gfx950/fmha_v3_fwd/fwd_hd192_hd128_bf16_causal.co b/hsa/gfx950/fmha_v3_fwd/fwd_hd192_hd128_bf16_causal.co
diff --git a/hsa/gfx950/fmha_v3_fwd/fwd_hd192_hd128_bf16_causal_group.co b/hsa/gfx950/fmha_v3_fwd/fwd_hd192_hd128_bf16_causal_group.co
diff --git a/hsa/gfx950/fmha_v3_fwd/fwd_hd192_hd128_bf16_group.co b/hsa/gfx950/fmha_v3_fwd/fwd_hd192_hd128_bf16_group.co
diff --git a/op_tests/cpp/mha/smoke_test_fwd_v3.sh b/op_tests/cpp/mha/smoke_test_fwd_v3.sh
@@ -24,7 +24,7 @@ run_gfx950_fwd_v3() {
     for mask in 0 2 ; do
     for lse in 0 1 ; do
     for seqlen_q in 127 192 301 512 1024; do
-    for seqlen_k in 512 700 1023 1058; do
+    for seqlen_k in 0 129 512 700 1023 1058; do
 
     $EXE -prec=bf16 -b=2 -h=4 -h_k=2 -d=$head_dim -d_v=128 -s=$seqlen_q -s_k=$seqlen_k -iperm=$i_perm -operm=$o_perm -mask=$mask -lse=$lse -fwd_v3=1 -mode=$mode -kname=$KNAME $COMMON_ARGS
     $EXE -prec=bf16 -b=1 -h=3 -h_k=1 -d=$head_dim -d_v=128 -s=$seqlen_q -s_k=$seqlen_k -iperm=$i_perm -operm=$o_perm -mask=$mask -lse=$lse -fwd_v3=1 -mode=$mode -kname=$KNAME $COMMON_ARGS
@@ -36,6 +36,10 @@ run_gfx950_fwd_v3() {
         $EXE -prec=bf16 -b=1 -h=1 -h_k=1 -d=$head_dim -d_v=128 -s=$seqlen_q -s_k=$seqlen_q -iperm=$i_perm -operm=$o_perm -mask=1 -lse=$lse -fwd_v3=1 -mode=$mode -kname=$KNAME $COMMON_ARGS
     fi
 
+    if [[ "$mode" = "1" ]]; then
+        $EXE -prec=bf16 -b=2 -h=4 -h_k=2 -d=128 -s=$seqlen_q,$seqlen_k -s_k=$seqlen_k,0 -iperm=$i_perm -operm=$o_perm -mask=$mask -lse=$lse -fwd_v3=1 -v3_bf16_cvt=$v3_bf16_cvt -mode=$mode -kname=$KNAME $COMMON_ARGS
+    fi
+
     done
     done
     done
@@ -54,7 +58,7 @@ run_gfx942_fwd_v3() {
     for mask in 0 2 ; do
     for lse in 0 1 ; do
     for seqlen_q in 127 192 301 512 1024; do
-    for seqlen_k in 512 700 1023 1058; do
+    for seqlen_k in 0 129 512 700 1023 1058; do
     for v3_bf16_cvt in 0 1 2; do
 
     $EXE -prec=bf16 -b=2 -h=4 -h_k=2 -d=128 -s=$seqlen_q -s_k=$seqlen_k -iperm=$i_perm -operm=$o_perm -mask=$mask -lse=$lse -fwd_v3=1 -v3_bf16_cvt=$v3_bf16_cvt -mode=$mode -kname=$KNAME $COMMON_ARGS
@@ -67,6 +71,10 @@ run_gfx942_fwd_v3() {
         $EXE -prec=bf16 -b=1 -h=1 -h_k=1 -d=128 -s=$seqlen_q -s_k=$seqlen_q -iperm=$i_perm -operm=$o_perm -mask=1 -lse=$lse -fwd_v3=1 -v3_bf16_cvt=$v3_bf16_cvt -mode=$mode -kname=$KNAME $COMMON_ARGS
     fi
 
+    if [[ "$mode" = "1" ]]; then
+        $EXE -prec=bf16 -b=2 -h=4 -h_k=2 -d=128 -s=$seqlen_q,$seqlen_k -s_k=$seqlen_k,0 -iperm=$i_perm -operm=$o_perm -mask=$mask -lse=$lse -fwd_v3=1 -v3_bf16_cvt=$v3_bf16_cvt -mode=$mode -kname=$KNAME $COMMON_ARGS
+    fi
+
     done
     done
     done