[Metax] Fix add flags (#227) (#2527)

* Set MXCC_OVERRIDE_OPTIONS in compile script

Add MXCC_OVERRIDE_OPTIONS for metax GPU compilation.

* Add MXCC_OVERRIDE_OPTIONS for Metax GPU

* Update flash_attn_grad_kernel.cu

* Update compile.sh

* [Metax][feat] add top_p_sampling.patch. (#225)

* [Metax] Fix add flags

---------

Co-authored-by: duqimeng <77875733+duqimeng@users.noreply.github.com>
Co-authored-by: MingkunZhang <39252862+StareAtYou@users.noreply.github.com>

Author: 3 people

Paddle

@@ -116,6 +116,35 @@ PHI_DEFINE_EXPORTED_bool(use_fast_math,
                          false,
                          "Whether to use fast math GPU functions.");
 
+/**
+ * GPU RNG related FLAG
+ * Name: FLAGS_deterministic_rng
+ * Since Version: 3.4
+ * Value Range: bool, default=false
+ * Example: paddle.set_flags({'FLAGS_deterministic_rng': True})
+ * Note: Fix RNG kernel launch config so same seed gives same results
+ *       across GPU types.
+ */
+PHI_DEFINE_EXPORTED_bool(
+    deterministic_rng,
+    false,
+    "Enable cross-device RNG consistency by fixing GPU kernel launch "
+    "configuration. When true, RNG kernels use a fixed grid/block size "
+    "so that the same seed produces identical results across GPU types.");
+/**
+ * GPU RNG related FLAG
+ * Name: FLAGS_deterministic_rng_grid
+ * Since Version: 3.4
+ * Value Range: int32, default=1024
+ * Example: paddle.set_flags({'FLAGS_deterministic_rng_grid': 4096})
+ * Note: Grid size cap used when FLAGS_deterministic_rng is enabled.
+ *       Cross-device consistency requires the same value on all devices.
+ */
+PHI_DEFINE_EXPORTED_int32(
+    deterministic_rng_grid,
+    1024,
+    "Grid size cap when FLAGS_deterministic_rng is enabled.");
+
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
 /**
  * FlashAttention related FLAG