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

docs/docs/advanced_features/hisparse_guide.mdx

@@ -163,7 +163,7 @@ python3 -m sglang.launch_server \
     --hisparse-config='{"top_k": 2048, "device_buffer_size": 6144, "host_to_device_ratio": 10, "swap_in_block_size": 960}'
 ```
 
-> **Note**: For DSA models, `--kv-cache-dtype` defaults to `auto`, which resolves to `fp8_e4m3` on SM100+ (Blackwell) and `bfloat16` on older architectures. The DSA decode backend is automatically selected based on KV dtype (`bfloat16` → `flashmla_sparse`, `fp8_e4m3` → `flashmla_kv`). DSA backend flags apply only to DSA models; DeepSeek V4 uses its own `dsv4` attention backend.
+> **Note**: For DSA models, `--kv-cache-dtype` defaults to `auto`, which resolves to `fp8_e4m3` on SM100+ (Blackwell) and `bfloat16` on older architectures. The DSA decode backend is automatically selected based on KV dtype (`bfloat16` → `flashmla_sparse`, `fp8_e4m3` → `flashmla_kv`), except for GLM DSA models on SM120/SM121 with `fp8_e4m3`, which use `flashinfer_sparse_mla`. DSA backend flags apply only to DSA models; DeepSeek V4 uses its own `dsv4` attention backend.
 
 ### Benchmark
 
@@ -189,6 +189,7 @@ python3 -m sglang.bench_serving \
 - The prefill instance does not need `--enable-hisparse`; it is unaware of HiSparse.
 - On the decode instance, `--enable-hisparse` and `--hisparse-config` are required for HiSparse.
 - For DSA models, `--kv-cache-dtype bfloat16` uses `flashmla_sparse`, and `--kv-cache-dtype fp8_e4m3` uses `flashmla_kv`.
+- On SM120/SM121 (e.g. RTX PRO 6000, RTX 5090) with GLM DSA models and `--kv-cache-dtype fp8_e4m3`, both DSA backends resolve to `flashinfer_sparse_mla`, which is the only DSA kernel available on that architecture. HiSparse accepts it there; no extra flag is needed.
 - For DeepSeek V4, DSA backend flags are not applicable. DeepSeek V4 uses the `dsv4` attention backend and `fp8_e4m3` KV cache by default.
 - `host_to_device_ratio` should be configured based on the host machine's available memory. For example:
   - **~1 TB** host memory → `host_to_device_ratio: 5`

python/sglang/srt/arg_groups/hisparse_hook.py

@@ -10,7 +10,7 @@ logger = logging.getLogger(__name__)
 
 HISPARSE_CUDA_DSA_BACKENDS_BY_DTYPE = {
     "bfloat16": {"flashmla_sparse"},
-    "fp8_e4m3": {"flashmla_kv"},
+    "fp8_e4m3": {"flashmla_kv", "flashinfer_sparse_mla"},
 }
 HISPARSE_ROCM_DSA_BACKENDS = {"tilelang", "aiter"}
 HISPARSE_KV_CACHE_DTYPES = ("bfloat16", "fp8_e4m3")
@@ -32,7 +32,7 @@ def _hisparse_allowed_backends(kv_cache_dtype: str) -> set[str]:
     if _is_hip():
         return HISPARSE_ROCM_DSA_BACKENDS
     return HISPARSE_CUDA_DSA_BACKENDS_BY_DTYPE.get(
-        kv_cache_dtype, {"flashmla_sparse", "flashmla_kv"}
+        kv_cache_dtype, {"flashmla_sparse", "flashmla_kv", "flashinfer_sparse_mla"}
     )
 
 
@@ -56,9 +56,8 @@ def validate_hisparse_dsa_backend(
             f"HiSparse supports DSA {label} backend(s) {sorted(allowed_backends)} "
             f"on this platform with --kv-cache-dtype={kv_cache_dtype}, "
             f"but got --dsa-{label}-backend={backend}. "
-            f"Please use --dsa-{label}-backend="
+            f"Please use one of {sorted(allowed_backends)}, or omit the option "
-            f"{_hisparse_default_backend(kv_cache_dtype)} "
+            "to let SGLang pick a backend for this platform."
-            "or omit it."
         )
 
 

test/registered/unit/server_args/test_server_args.py

@@ -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")