[AMD] Enable Fast Triton Sparse MLA backend (#30575)

Co-authored-by: clintg6 <7388379+clintg6@users.noreply.github.com> Co-authored-by: HAI <hixiao@gmail.com>
2026-09-10 01:58:00 -07:00
parent 5caafd2118
commit 8a6ab89bf0
9 changed files with 1881 additions and 145 deletions
@@ -1938,12 +1938,12 @@ class TestGoldenModelOverrides(_IsolatedPublish):
            override_platform(is_hip=True),
            patch("torch.cuda.get_device_capability", return_value=(9, 4)),
        ):
-            # ROCm with both unset -> tilelang
+            # ROCm with both unset -> Triton for FP8 and BF16 KV cache.
            self.assertEqual(
                _dsa_split_backend_resolution(_view(kv_cache_dtype="bfloat16")),
                {
-                    "dsa_prefill_backend": "tilelang",
-                    "dsa_decode_backend": "tilelang",
+                    "dsa_prefill_backend": "triton",
+                    "dsa_decode_backend": "triton",
                },
            )