Add Cutlass MLA attention backend (#5390)

python/sglang/srt/server_args.py

@@ -256,6 +256,12 @@ class ServerArgs:
             )
             self.page_size = 64
 
+        if self.attention_backend == "cutlass_mla":
+            logger.warning(
+                "Cutlass MLA only supports a page_size of 128, change page_size to 128."
+            )
+            self.page_size = 128
+
         # Set cuda graph max batch size
         if self.cuda_graph_max_bs is None:
             # Based on detailed statistics, when serving TP1/TP2 models on lower-end GPUs with HBM<25G, you can either disable cuda graph or set `cuda_graph_max_bs` to a very small value to reduce the memory overhead of creating cuda graphs, with almost no impact on performance. However, when serving models with TP4 or TP8, we need to enable cuda graph to maintain high performance. In this case, we can set `cuda_graph_max_bs` to 80 (half of the default value 160) to reduce the memory overhead of creating cuda graphs. Looking at the logs from TP4 serving of qwen2-72b, a value of 80 is sufficient and can reduce the memory overhead of creating cuda graphs on lower-end GPUs compared to the original 160, avoiding OOM issues.
@@ -823,7 +829,14 @@ class ServerArgs:
         parser.add_argument(
             "--attention-backend",
             type=str,
-            choices=["flashinfer", "triton", "torch_native", "fa3", "flashmla"],
+            choices=[
+                "flashinfer",
+                "triton",
+                "torch_native",
+                "fa3",
+                "flashmla",
+                "cutlass_mla",
+            ],
             default=ServerArgs.attention_backend,
             help="Choose the kernels for attention layers.",
         )