[SDPA] Remove unused rng_engine_inputs (#93024)
The unused variable in `fmha_api.cpp` [here](https://github.com/pytorch/pytorch/blob/master/aten/src/ATen/native/transformers/cuda/flash_attn/fmha_api.cpp#L313) was causing build failures (internally) due to to the `-Wunused-variable` flag being used. For example:
```
[2023-01-24T20:32:00.241-08:00] Stderr: aten/src/ATen/native/transformers/cuda/flash_attn/fmha_api.cpp:313:25: error: unused variable 'rng_engine_inputs' [-Werror,-Wunused-variable]
[CONTEXT] [2023-01-24T20:32:00.241-08:00] at::PhiloxCudaState rng_engine_inputs;
[CONTEXT] [2023-01-24T20:32:00.241-08:00] ^
[2023-01-24T21:09:33.507-08:00] Stderr: aten/src/ATen/native/transformers/cuda/flash_attn/fmha_api.cpp:313:25: error: unused variable 'rng_engine_inputs' [-Werror,-Wunused-variable]
[CONTEXT] [2023-01-24T21:09:33.507-08:00] at::PhiloxCudaState rng_engine_inputs;
[CONTEXT] [2023-01-24T21:09:33.507-08:00]
```
This PR removes that unused variable. Mirroring this same patch made by @drisspg internally.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/93024
Approved by: https://github.com/drisspg
diff --git a/aten/src/ATen/native/transformers/cuda/flash_attn/fmha_api.cpp b/aten/src/ATen/native/transformers/cuda/flash_attn/fmha_api.cpp
index a16ce10..9796ae7 100644
--- a/aten/src/ATen/native/transformers/cuda/flash_attn/fmha_api.cpp
+++ b/aten/src/ATen/native/transformers/cuda/flash_attn/fmha_api.cpp
@@ -310,7 +310,6 @@
// state
// We use a custom RNG that increases the offset by batch_size * nheads * 32.
int64_t counter_offset = launch_params.params.b * launch_params.params.h * 32;
- at::PhiloxCudaState rng_engine_inputs;
if( is_dropout ) {
// See Note [Acquire lock when using random generators]
