[MUSA][9/N] Add FA3 attention backend support through MATE (MUSA AI Tensor Engine) (#17985)

Co-authored-by: R0CKSTAR <xiaodong.ye@mthreads.com>

python/sglang/srt/server_args.py

@@ -52,6 +52,7 @@ from sglang.srt.utils.common import (
     is_hip,
     is_hopper_with_cuda_12_3,
     is_mps,
+    is_musa,
     is_no_spec_infer_or_topk_one,
     is_npu,
     is_remote_url,
@@ -2412,6 +2413,13 @@ class ServerArgs:
             if model_config.context_len > 8192:
                 self.mem_fraction_static *= 0.85
 
+        # MUSA platforms compatible backends
+        if is_musa() and self.attention_backend == "fa3":
+            logger.warning(
+                "FA3 attention backend on MUSA ignores any user-provided page_size and enforces a fixed value of 64."
+            )
+            self.page_size = 64
+
         # Other platforms backends
         if (
             self.attention_backend == "intel_amx"