[Feature] Support Flashinfer fmha on Blackwell (#6930)

python/sglang/srt/layers/attention/flashinfer_mla_backend.py

@@ -29,6 +29,7 @@ from sglang.srt.layers.attention.flashinfer_backend import (
     create_flashinfer_kv_indices_triton,
 )
 from sglang.srt.layers.dp_attention import get_attention_tp_size
+from sglang.srt.layers.utils import is_sm100_supported
 from sglang.srt.managers.schedule_batch import global_server_args_dict
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch, ForwardMode
 from sglang.srt.speculative.eagle_utils import EagleDraftInput, EagleVerifyInput
@@ -108,8 +109,11 @@ class FlashInferMLAAttnBackend(AttentionBackend):
         else:
             self.q_indptr_decode = q_indptr_decode_buf
 
+        fmha_backend = "auto"
+        if is_sm100_supported():
+            fmha_backend = "cutlass"
         self.prefill_wrapper_ragged = BatchPrefillWithRaggedKVCacheWrapper(
-            self.workspace_buffer, "NHD"
+            self.workspace_buffer, "NHD", backend=fmha_backend
         )
 
         if not self.skip_prefill: