Update cudnn-frontend to 1.0.3 to fix cuDNN v9 SDPA NaNs (#650)

* Update cudnn frontend to 1.0.3 to fix cudnn v9 Nans

Signed-off-by: Charlene Yang <[email protected]>

* make d_out contiguous for bwd

Signed-off-by: Charlene Yang <[email protected]>

* remove cudnnDestroy to let torch handle it

Signed-off-by: Charlene Yang <[email protected]>

* Update transformer_engine/pytorch/attention.py

Co-authored-by: Tim Moon <[email protected]>
Signed-off-by: cyanguwa <[email protected]>

* Update transformer_engine/pytorch/attention.py

Co-authored-by: Tim Moon <[email protected]>
Signed-off-by: cyanguwa <[email protected]>

* Update transformer_engine/pytorch/attention.py

Co-authored-by: Tim Moon <[email protected]>
Signed-off-by: cyanguwa <[email protected]>

---------

Signed-off-by: Charlene Yang <[email protected]>
Signed-off-by: cyanguwa <[email protected]>
Co-authored-by: Tim Moon <[email protected]>

Author: 2 people

3rdparty/cudnn-frontend

@@ -152,11 +152,6 @@ class cudnnExecutionPlanManager {
     }
 
     ~cudnnExecutionPlanManager() {
-        static thread_local std::once_flag flag;
-        std::call_once(flag, [&] {
-                        if (handle_ != nullptr) {
-                          cudnnDestroy(handle_);
-                        }});
     }
 
  private:

transformer_engine/common/fused_attn/utils.h

@@ -1733,6 +1733,7 @@ def forward(ctx, is_training, max_seqlen, cu_seqlens, qkv, qkv_dtype, attn_bias,
 
     @staticmethod
     def backward(ctx, d_out):
+        d_out = d_out.contiguous()
         qkv, out, cu_seqlens = ctx.saved_tensors
         if not ctx.aux_ctx_tensors[0].is_contiguous():
             ctx.aux_ctx_tensors[0] = ctx.aux_ctx_tensors[0].contiguous()
@@ -1802,6 +1803,7 @@ def forward(ctx, is_training, max_seqlen_q, max_seqlen_kv, cu_seqlens_q, cu_seql
 
     @staticmethod
     def backward(ctx, d_out):
+        d_out = d_out.contiguous()
         q, kv, out, cu_seqlens_q, cu_seqlens_kv = ctx.saved_tensors
         if not ctx.aux_ctx_tensors[0].is_contiguous():
             ctx.aux_ctx_tensors[0] = ctx.aux_ctx_tensors[0].contiguous()
@@ -1883,6 +1885,7 @@ def forward(ctx, is_training, max_seqlen_q, max_seqlen_kv, cu_seqlens_q, cu_seql
 
     @staticmethod
     def backward(ctx, d_out):
+        d_out = d_out.contiguous()
         q, k, v, out, cu_seqlens_q, cu_seqlens_kv = ctx.saved_tensors
         if not ctx.aux_ctx_tensors[0].is_contiguous():
             ctx.aux_ctx_tensors[0] = ctx.aux_ctx_tensors[0].contiguous()