Boost CMF 100x QPS from 176k to 178k on mi300x (#4867)

Summary:
Pull Request resolved: #4867

X-link: facebookresearch/FBGEMM#1888

As titled ~ by which triton fp8 gemm on-pars torch counterpart, which yield 178k QPS.

Reviewed By: pranavsharma

Differential Revision: D82276002

fbshipit-source-id: ce1d755eb6f89e6feb8847ac09d520079fbdc510

Author: RandySheriff

fbgemm_gpu/experimental/gemm/triton_gemm/fp8_gemm.py

@@ -3866,6 +3866,20 @@ def get_full_non_persistent_tuning_space():
         num_warps=8,
         num_stages=2,
     ),
+    triton.Config(
+        {
+            "BLOCK_M": 256,
+            "BLOCK_N": 256,
+            "BLOCK_K": 64,
+            "GROUP_M": 2,
+            "SPLIT_K": 1,
+            "waves_per_eu": 2,
+            "matrix_instr_nonkdim": 32,
+            "kpack": 2,
+        },
+        num_warps=8,
+        num_stages=2,
+    ),
     triton.Config(
         {
             "BLOCK_M": 256,