Bump cuDNN FE to 1.14.0 (#2072)

* Bump cuDNN FE to 1.14.0
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Change submodule hash
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Pick up a cuDNN FE fix
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* New model configs in tests
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Exclude cuDNN backend for some configs
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

---------
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

cudnn-frontend @ deda80e5

-Subproject commit 9793df569ce413f4b1844a9176f7ae24dd981603
+Subproject commit deda80e5372d50e925d7bf4f76c5db779be3fbd5

tests/pytorch/attention/test_attention.py

@@ -274,6 +274,8 @@ model_configs_mla = {
     "mla_3_0": ModelConfig(8, 1, 16, 128, max_seqlen_kv=2048, head_dim_v=64),  # inference
     "mla_3_1": ModelConfig(8, 1, 16, 256, max_seqlen_kv=2048, head_dim_v=128),  # inference
     "mla_3_2": ModelConfig(8, 1, 16, 192, max_seqlen_kv=2048, head_dim_v=128),  # inference
+    "mla_3_3": ModelConfig(8, 1, 16, 160, max_seqlen_kv=2048, head_dim_v=128),  # inference
+    "mla_3_4": ModelConfig(8, 1, 16, 160, max_seqlen_kv=2048, head_dim_v=160),  # inference
 }
 
 

transformer_engine/common/fused_attn/fused_attn.cpp

@@ -252,8 +252,9 @@ NVTE_Fused_Attn_Backend nvte_get_fused_attn_backend(
           (head_dim_qk == 192 && head_dim_v == 128 && is_training && sm_arch_ >= 100 &&
            cudnn_runtime_version >= 91100)) &&
          // 9.11/9.12 bug: 128 < d_qk <= 256, 128 < d_v <= 256 + Hopper + bprop + MLA
-         (!((cudnn_runtime_version == 91100 || cudnn_runtime_version == 91200) && is_training &&
-            sm_arch_ == 90 && head_dim_qk >= 128 && head_dim_v >= 128 &&
+         (!((cudnn_runtime_version == 91100 || cudnn_runtime_version == 91200 ||
+             cudnn_runtime_version == 91300) &&
+            is_training && sm_arch_ == 90 && head_dim_qk >= 128 && head_dim_v >= 128 &&
             !(head_dim_qk == 192 && head_dim_v == 128) && head_dim_qk != head_dim_v))) &&
         // bias type
         ((cudnn_runtime_version < 8906 && bias_type == NVTE_Bias_Type::NVTE_NO_BIAS) ||