Fix MXFP8 test due to OSS FBGEMM version mismatch issue (#4749)

Summary:
Pull Request resolved: #4749

X-link: facebookresearch/FBGEMM#1771

Currently there is a version mismatch issue in OSS FBGEMM. Move the latest added `to_mxfp8` to `test_mx_grouped_gemm` to release the issue in tests

Will add it back when the OSS FBGEMM version mismatch issue is resolved

Reviewed By: q10

Differential Revision: D80634475

fbshipit-source-id: c77f0b0226242c7a0235a6738b1acffc88f6b323

Author: jiawenliu64

fbgemm_gpu/experimental/gen_ai/test/quantize/quantize_test.py

@@ -24,7 +24,6 @@
         quantize_fp8_block,
         quantize_fp8_row,
         supports_float8_fnuz,
-        to_mxfp8,
     )
 
     from fbgemm_gpu.experimental.gen_ai.quantize import quantize_int4_preshuffle
@@ -1240,6 +1239,8 @@ def test_mx_grouped_gemm(
         N: int,
         K: int,
     ) -> None:
+        from fbgemm_gpu.experimental.gemm.triton_gemm.fp8_gemm import to_mxfp8
+
         X = torch.randn((G, M, K), dtype=torch.bfloat16, device=self.device) * 0.1
         W = torch.randn((G, N, K), dtype=torch.bfloat16, device=self.device) * 0.01