[PyTorch] Bump minimum cuDNN version for fused attention with FP8 current scaling (#2236)

* Require cuDNN 9.14.0+ for fused attention with FP8 current scaling
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



---------
Signed-off-by: Tim Moon <tmoon@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

transformer_engine/pytorch/attention/dot_product_attention/utils.py

@@ -469,13 +469,13 @@ def get_attention_backend(
         fp8_recipe = fp8_meta["recipe"]
         if fp8_meta.get("local_recipes", None) is not None:
             fp8_recipe = fp8_meta["local_recipes"][0]
-        if (
-            use_fused_attention
-            and fp8_recipe.float8_current_scaling()
-            and device_compute_capability < (10, 0)
-        ):
+        if use_fused_attention and fp8_recipe.float8_current_scaling():
+            if device_compute_capability < (10, 0):
                 logger.debug("Disabling FusedAttention for FP8 current scaling on arch < sm100")
                 use_fused_attention = False
+            elif cudnn_version < (9, 14, 0):
+                logger.debug("Disabling FusedAttention for FP8 current scaling with cuDNN < 9.14.0")
+                use_fused_attention = False
 
     # Filter: KV cache
     # backend  | precision      |    KV cache     | architecture | qkv_format    | page_size