[DeepSeek-V4] Add Q8KV8 sparse MLA prefill runtime backend (#32327)

Co-authored-by: Ho-Ren (Jack) Chuang <horenchuang@bytedance.com>
Co-authored-by: Xiaoyu Zhang <1182563586@qq.com>

python/sglang/srt/server_args.py

@@ -359,6 +359,12 @@ DSA_CHOICES = [
 ]
 NSA_CHOICES = DSA_CHOICES  # deprecated alias
 
+DSV4_PREFILL_BACKEND_CHOICES = [
+    "auto",
+    "flashmla_sparse",
+    "flashmla_sparse_q8",
+]
+
 DSA_TOPK_BACKEND_CHOICES = ["sgl-kernel", "torch", "flashinfer"]
 
 DSA_PAGED_MQA_LOGITS_BACKEND_CHOICES = ["auto", "deepgemm", "cutedsl", "aiter"]
@@ -1805,6 +1811,18 @@ class ServerArgs:
         ),
         NS("exec.kernel"),
     ] = None
+    dsv4_prefill_backend: A[
+        str,
+        Arg(
+            help=(
+                "DeepSeek-V4 sparse prefill backend. 'auto' and "
+                "'flashmla_sparse' use the existing BF16 sparse prefill path; "
+                "'flashmla_sparse_q8' enables the Q8KV8 sparse prefill path."
+            ),
+            choices=DSV4_PREFILL_BACKEND_CHOICES,
+        ),
+        NS("exec.kernel"),
+    ] = "auto"
     dsa_decode_backend: A[
         Optional[str],
         Arg(