diff --git a/flash_attn/cute/flash_bwd_sm100.py b/flash_attn/cute/flash_bwd_sm100.py
index 11db2dab563..273cd44f6f2 100644
--- a/flash_attn/cute/flash_bwd_sm100.py
+++ b/flash_attn/cute/flash_bwd_sm100.py
@@ -82,13 +82,7 @@ def __init__(
         assert self.tile_hdim <= 128 or (self.tile_hdim == 192 and self.tile_hdimv == 128)
         assert self.tile_hdimv <= 128
 
-        self.use_2cta_instrs = bool(
-            use_2cta_instrs
-            and cluster_size == 2
-            and score_mod is None
-            and score_mod_bwd is None
-            and mask_mod is None
-        )
+        self.use_2cta_instrs = bool(use_2cta_instrs and cluster_size == 2)
         self.cta_group_size = 2 if self.use_2cta_instrs else 1
 
         assert self.tile_hdim != 192 or self.use_2cta_instrs, "Must use 2CTA for hdim 192"
@@ -2758,9 +2752,15 @@ def apply_score_mod(
         fastdiv_mods=(None, None),
     ):
         """Apply forward score modification for SM100 backward pass."""
-        # In bwd, S is computed as K @ Q.T so dimensions are (tile_n, tile_m)
-        cS = cute.make_identity_tensor((self.tile_n, self.tile_m))
-        cS = cute.domain_offset((n_block * self.tile_n, m_block * self.tile_m), cS)
+        # In bwd, S is computed as K @ Q.T so dimensions are (tile_n, tile_m).
+        # With 2CTA, partition_C must see the full cluster tile so each CTA 
+        # gets its own half of the tile.
+        cluster_tile_n = self.tile_n * self.cta_group_size
+        cluster_n_block = n_block // self.cta_group_size
+        cS = cute.make_identity_tensor((cluster_tile_n, self.tile_m))
+        cS = cute.domain_offset(
+            (cluster_n_block * cluster_tile_n, m_block * self.tile_m), cS
+        )
         tScS = thr_mma_S.partition_C(cS)
         tScS_idx = thr_copy_t2r.partition_D(tScS)
 
@@ -2976,13 +2976,13 @@ def compute_loop(
                 seqlen, n_block // self.cluster_shape_mnk[0]
             )
             mask = AttentionMaskCls(seqlen)
-            n_block_for_cluster = n_block // self.cta_group_size
+            cluster_n_block = n_block // self.cta_group_size
             # TODO: condition mask_seqlen
             mask_fn = partial(
                 mask.apply_mask_sm100_transposed,
                 tScS_t2r=tScS_t2r,
                 t0ScS_t2r=t0ScS_t2r,
-                n_block=n_block_for_cluster,
+                n_block=cluster_n_block,
                 mask_seqlen=True,
                 mask_causal=self.is_causal,
                 mask_local=self.is_local,
@@ -3194,9 +3194,12 @@ def compute_loop(
 
                     if const_expr(self.score_mod_bwd is not None):
                         tSrS_pre_cur = tSrS_pre[None, stage, 0, 0]
-                        cS_bwd = cute.make_identity_tensor((self.tile_n, self.tile_m))
+                        cluster_tile_n = self.tile_n * self.cta_group_size
+                        cluster_n_block = n_block // self.cta_group_size
+                        cS_bwd = cute.make_identity_tensor((cluster_tile_n, self.tile_m))
                         cS_bwd = cute.domain_offset(
-                            (n_block * self.tile_n, m_block * self.tile_m), cS_bwd
+                            (cluster_n_block * cluster_tile_n, m_block * self.tile_m),
+                            cS_bwd,
                         )
                         tScS_bwd = thr_mma_S.partition_C(cS_bwd)
                         tScS_idx_bwd = thr_copy_t2r.partition_D(tScS_bwd)
diff --git a/flash_attn/cute/interface.py b/flash_attn/cute/interface.py
index 0bd4651190c..9e423c16ed9 100644
--- a/flash_attn/cute/interface.py
+++ b/flash_attn/cute/interface.py
@@ -1323,9 +1323,6 @@ def _flash_attn_bwd(
         requested_disable_2cta = utils._get_disable_2cta_default()
         disable_2cta = (
             requested_disable_2cta
-            or score_mod is not None
-            or score_mod_bwd is not None
-            or mask_mod is not None
             or block_sparse_tensors is not None
         )
         cluster_size = 2 if head_dim >= 128 and not disable_2cta else 1
diff --git a/tests/cute/test_flash_attn.py b/tests/cute/test_flash_attn.py
index 21ed3a48d57..4b5fe669004 100644
--- a/tests/cute/test_flash_attn.py
+++ b/tests/cute/test_flash_attn.py
@@ -349,11 +349,10 @@ def test_flash_attn_output(
             and not has_qv
             and not dv > 256
             and not attention_chunk != 0
-            and softcap == 0.0
             and (
                 (dv == d and d <= 128)
                 or (d == 192 and dv == 128)
-                or (IS_SM100 and d == 256 and dv == 256)
+                or (IS_SM100 and d == 256 and dv == 256 and softcap == 0.0)
             )
             and learnable_sink is None
             # and False