[Fix] Allow flashinfer_sparse_mla DSA backend for HiSparse on SM120 FP8 KV (#33075)

2026-08-11 15:05:04 -07:00
parent 59450c4f18
commit 2c07ca5e8d
3 changed files with 23 additions and 6 deletions
@@ -661,6 +661,23 @@ class TestHiSparseDsaBackendPolicy(unittest.TestCase):
        self.assertEqual(resolved["dsa_prefill_backend"], "flashmla_kv")
        self.assertEqual(resolved["dsa_decode_backend"], "flashmla_kv")

+    @patch("sglang.srt.server_args.is_hip", return_value=False)
+    def test_hisparse_accepts_flashinfer_sparse_mla_on_cuda_fp8(self, _mock_is_hip):
+        """SM120 GLM DSA resolves both DSA backends to flashinfer_sparse_mla, so
+        the fp8 hisparse allow-set must admit it or --enable-hisparse cannot
+        start there at all. The device/arch narrowing happens later, in
+        _validate_flashinfer_sparse_mla_backend."""
+        server_args = ServerArgs(
+            model_path="dummy",
+            enable_hisparse=True,
+            kv_cache_dtype="fp8_e4m3",
+            dsa_prefill_backend="flashinfer_sparse_mla",
+            dsa_decode_backend="flashinfer_sparse_mla",
+        )
+
+        server_args._validate_hisparse_dsa_backend("dsa_prefill_backend", "prefill")
+        server_args._validate_hisparse_dsa_backend("dsa_decode_backend", "decode")
+
    @patch("sglang.srt.server_args.is_hip", return_value=True)
    def test_hisparse_defaults_to_tilelang_on_rocm(self, _mock_is_hip):
        resolved = self._resolve("bfloat16")