[JAX] [PyT] [Common] Enable D=256 BWD cuDNN fused attn for Blackwell CC 10.x

tests/jax/test_fused_attn.py

@@ -444,6 +444,56 @@ def _check_configs(self):
                 "is either BSHD_BSHD_BSHD or THD_THD_THD"
             )
 
+        # D=256 bprop on SM10.x uses cuDNN's dedicated SDPA bprop kernel
+        # (cuDNN FE 1.24 / BE 9.23+). FE forces this path onto the deterministic algorithm path,
+        # which rejects dBias, dropout, and ALiBi. It supports vanilla softmax only and allows SWA
+        # together with a causal mask only.
+        compute_capability = get_device_compute_capability(0)
+        is_sm10x = 100 <= compute_capability < 110
+        if self.is_training and is_sm10x and (self.head_dim_qk == 256 or self.head_dim_v == 256):
+            if self.head_dim_qk != 256 or self.head_dim_v != 256:
+                pytest.skip(
+                    "D=256 BWD on Blackwell only supports d_qk == d_v == 256;"
+                    f" got d_qk={self.head_dim_qk}, d_v={self.head_dim_v}."
+                )
+            cudnn_version = get_cudnn_version()
+            if cudnn_version < 92300:
+                pytest.skip(
+                    "D=256 BWD on Blackwell requires cuDNN 9.23 or newer;"
+                    f" got cuDNN {cudnn_version}."
+                )
+            # Non-learnable bias is fine (bias is allowed as an input); only dBias is
+            # unsupported. The JAX runner asks for dBias iff the bias shape is [1, h, s, s]
+            # (see test_backward), so gate on that.
+            unsupported = None
+            if self.attn_bias_type == AttnBiasType.PRE_SCALE_BIAS:
+                unsupported = "pre-scale bias"
+            elif self.attn_bias_type != AttnBiasType.NO_BIAS and self.bias_shape == BiasShape._1HSS:
+                unsupported = (
+                    "bias gradients (dBias); frozen/non-learnable bias inputs"
+                    " (i.e. non-1HSS bias shapes) are supported"
+                )
+            elif self.dropout_prob != 0.0:
+                unsupported = "dropout"
+            elif self.softmax_type != AttnSoftmaxType.VANILLA_SOFTMAX:
+                unsupported = "non-vanilla softmax"
+            if unsupported is not None:
+                pytest.skip(
+                    "D=256 BWD on Blackwell uses the deterministic SM100 D=256 SDPA BWD"
+                    f" kernel which does not support {unsupported}."
+                )
+            if self.window_size is not None and self.window_size != (-1, -1):
+                if not self.attn_mask_type.is_causal():
+                    pytest.skip(
+                        "D=256 BWD on Blackwell uses the SM100 D=256 SDPA BWD kernel"
+                        " which requires window_size=(-1, -1) for non-causal masks."
+                    )
+                if self.window_size[1] not in (-1, 0):
+                    pytest.skip(
+                        "D=256 BWD on Blackwell only supports right window -1 or 0"
+                        " for causal masks."
+                    )
+
         self.backend = FusedAttnHelper(
             self.is_training,
             self.dtype,
@@ -1474,6 +1524,21 @@ def test_backward(
             QKVLayout.THD_THD_THD,
             id="2-1024-2048-12-6-128-64-BF16-CROSS-GQA-RAGGED_SEPARATE",
         ),
+        # D=256 deterministic backward on the SM100 dedicated SDPA bprop kernel
+        # (cuDNN FE 1.24 / BE 9.23+). Unsupported configs (e.g. dBias, non-256 head dims)
+        # are skipped by FusedAttnRunner._check_configs.
+        pytest.param(
+            4,
+            128,
+            128,
+            16,
+            16,
+            256,
+            256,
+            jnp.float16,
+            QKVLayout.BSHD_BS2HD,
+            id="4-128-128-16-16-256-256-FP16-SELF-KV_PACKED",
+        ),
     ],
 )
 @pytest.mark.parametrize(

tests/pytorch/attention/test_attention.py

@@ -377,6 +377,38 @@ def test_dpa_fa4_hdim256(dtype, model_configs, model):
     test_dot_product_attention(dtype, model_configs, model, False, True, None, False, False)
 
 
+# cuDNN FusedAttention D=256 bprop is supported on sm10x from cuDNN 9.23 (FE 1.24),
+# via the dedicated deterministic SDPA bprop kernel, which supports d_qk == d_v == 256 only,
+# vanilla type of softmax only, no dropout, no ALiBi, and (for non-causal masks) full-window attention only.
+# (for non-causal masks) full-window attention.
+model_configs_fused_hdim256 = {
+    # test: ModelConfig(b, sq, hq, dqk)  -> head_dim_v defaults to head_dim_qk (256)
+    "fused_hd256_no_mask": ModelConfig(2, 512, 16, 256),
+    "fused_hd256_padding": ModelConfig(2, 512, 16, 256, attn_mask_type="padding"),
+    # SWA is allowed only together with a causal mask on the D=256 bprop kernel.
+    "fused_hd256_causal_swa": ModelConfig(
+        2, 1024, 16, 256, attn_mask_type="causal", window_size=(128, 0)
+    ),
+    # GQA variant (num_gqa_groups < num_heads).
+    "fused_hd256_padding_causal_gqa": ModelConfig(
+        2, 1024, 16, 256, num_gqa_groups=4, attn_mask_type="padding_causal"
+    ),
+}
+
+
+@pytest.mark.skipif(get_cudnn_version() < (9, 23, 0), reason="cuDNN 9.23+ is required.")
+@pytest.mark.skipif(
+    device_compute_capability not in ((10, 0), (10, 3)),
+    reason="cuDNN FusedAttention head_dim=256 backward is Blackwell server (SM100/SM103) only.",
+)
+@pytest.mark.parametrize("dtype", param_types)
+@pytest.mark.parametrize("model_configs", [model_configs_fused_hdim256])
+@pytest.mark.parametrize("model", model_configs_fused_hdim256.keys())
+def test_dpa_fused_attn_hdim256(dtype, model_configs, model):
+    """Test DotProductAttention with cuDNN FusedAttention: head_dim=256 backward on Blackwell"""
+    test_dot_product_attention(dtype, model_configs, model, False, True, None, False, False)
+
+
 model_configs_fa4_mla = {
     # test: ModelConfig(b, sq, hq, dqk, head_dim_v=dv)
     "fa4_mla_1": ModelConfig(4, 128, 16, 128, head_dim_v=64),

transformer_engine/common/fused_attn/fused_attn.cpp

@@ -315,6 +315,21 @@ NVTE_Fused_Attn_Backend nvte_get_fused_attn_backend(
           (head_dim_qk <= 256 && head_dim_v <= 256 &&
            ((!is_training && sm_arch_ == 90 && cudnn_runtime_version >= 90100) ||
             (is_training && sm_arch_ == 90 && cudnn_runtime_version >= 90500))) ||
+          // 9.23: d_qk = d_v = 256 + SM10x (cuDNN FE 1.24 / BE 9.23+) + bprop + non-paged
+          (head_dim_qk == 256 && head_dim_v == 256 && is_training && sm_arch_ >= 100 &&
+           sm_arch_ < 110 && cudnn_runtime_version >= 92300 &&
+           layout_group != NVTE_QKV_Layout_Group::NVTE_Paged_KV_HD_HD_HD &&
+           // The FE forces this path onto the deterministic bprop algorithm, which on
+           // Blackwell rejects dBias, dropout, and ALiBi (and supports vanilla softmax only).
+           bias_type == NVTE_Bias_Type::NVTE_NO_BIAS && dropout == 0.0 &&
+           softmax_type == NVTE_Softmax_Type::NVTE_VANILLA_SOFTMAX &&
+           // Non-causal D=256 supports only full-window attention; SWA is allowed only for causal masks.
+           ((window_size_left == -1 && window_size_right == -1) ||
+            ((attn_mask_type == NVTE_Mask_Type::NVTE_CAUSAL_MASK ||
+              attn_mask_type == NVTE_Mask_Type::NVTE_PADDING_CAUSAL_MASK ||
+              attn_mask_type == NVTE_Mask_Type::NVTE_CAUSAL_BOTTOM_RIGHT_MASK ||
+              attn_mask_type == NVTE_Mask_Type::NVTE_PADDING_CAUSAL_BOTTOM_RIGHT_MASK) &&
+             (window_size_right == -1 || window_size_right == 0)))) ||
           // 9.9: any head_dim + Blackwell + fprop + non_paged + sq > 1
           (!is_training && sm_arch_ >= 100 && cudnn_runtime_version >= 90900 && max_seqlen_q > 1 &&
            layout_group != NVTE_QKV_Layout_Group::NVTE_Paged_KV_HD_HD_HD) ||