Revert "[cuDNN] Enable cuDNN Frontend v8 API by Default (pytorch#84948)"

This reverts commit 427e0a6.

Reverted pytorch#84948 on behalf of https://github.com/malfet due to Broke SM86 sanity

Author: pytorchmergebot

aten/src/ATen/native/ConvUtils.h

@@ -66,11 +66,11 @@ namespace {
 }
 
 static inline bool cudnnv8_enabled_check_debug() {
-  static bool cudnnv8_flag = c10::utils::check_env("TORCH_CUDNN_V8_API_DISABLED") != true;
+  static bool cudnnv8_flag = c10::utils::check_env("TORCH_CUDNN_V8_API_ENABLED") == true;
   static bool cudnnv8_debug = c10::utils::check_env("TORCH_CUDNN_V8_API_DEBUG") == true;
   static uint8_t cudnnv8_debugcount = 0;
   if (cudnnv8_debug == 1 && cudnnv8_debugcount < 10) {
-    TORCH_WARN("TORCH_CUDNN_V8_DEBUG ON, V8 ON: ", cudnnv8_flag, " TORCH_CUDNN_USE_HEURISTIC_MODE B: ", cudnnv8_heuristic_mode_b);
+    TORCH_WARN("TORCH_CUDNN_V8_DEBUG ON, V8_FLAG: ", cudnnv8_flag, " TORCH_CUDNN_USE_HEURISTIC_MODE B: ", cudnnv8_heuristic_mode_b);
     cudnnv8_debugcount++;
   }
   return cudnnv8_flag == 1;

aten/src/ATen/native/cudnn/ConvShared.h

@@ -113,7 +113,7 @@ void raw_cudnn_convolution_add_relu_fallback_out(
 
 #if HAS_CUDNN_V8()
 // v7 functions are preserved here to allow for runtime switching to v7
-// (e.g., TORCH_CUDNN_V8_API_DISABLED=1).
+// (e.g., TORCH_CUDNN_V8_API_ENABLED=0).
 // Note that v7 forward/backward out can have different behavior from the v8
 // versions, as v7 explicitly splits large tensors as a 32-bit indexing
 // workaround whereas v8 expects cuDNN to handle large tensors.

test/test_cuda.py

@@ -2894,10 +2894,10 @@ def test_autocast_torch_bf16(self):
                 op, args = op_with_args[0], op_with_args[1]
                 if len(op_with_args) == 3:
                     skip_test = op_with_args[2]  # TEST_WITH_ROCM
-                should_error_from_cudnn = 'cudnn' in op and \
-                    ('TORCH_CUDNN_V8_API_DISABLED' in os.environ and
-                     int(os.environ['TORCH_CUDNN_V8_API_DISABLED']) or
-                     torch.cuda.get_device_capability() < (8, 0))
+                should_error_from_cudnn = 'cudnn' in op and not\
+                    ('TORCH_CUDNN_V8_API_ENABLED' in os.environ and
+                     int(os.environ['TORCH_CUDNN_V8_API_ENABLED']) and
+                     torch.cuda.get_device_capability() >= (8, 0))
                 should_error_from_not_implemented = should_error_from_cudnn or 'prelu' in op or 'thnn' in op \
                     or 'fused' in op or 'gru' in op or op == '_thnn_fused_lstm_cell' or op == 'lstm_cell'
                 if not skip_test: