Set csgmv as default lora backend. (#11488)

python/sglang/srt/server_args.py

@@ -309,8 +309,8 @@ class ServerArgs:
     ] = None
     max_loaded_loras: Optional[int] = None
     max_loras_per_batch: int = 8
+    lora_backend: str = "csgmv"
     lora_eviction_policy: str = DEFAULT_LORA_EVICTION_POLICY
-    lora_backend: str = "triton"
     max_lora_chunk_size: Optional[int] = 16
 
     # Kernel backend

python/sglang/test/runners.py

@@ -496,7 +496,7 @@ class SRTRunner:
         attention_backend: Optional[str] = None,
         prefill_attention_backend: Optional[str] = None,
         decode_attention_backend: Optional[str] = None,
-        lora_backend: str = "triton",
+        lora_backend: str = "csgmv",
         disable_cuda_graph: bool = False,
         disable_radix_cache: bool = False,
         chunked_prefill_size: Optional[int] = None,