fix: enable flashinfer backends and allreduce fusion for SM12x Blackwell GPUs

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

@@ -195,15 +195,24 @@ def attn_backend_wrapper(runner: "ModelRunner", full_attn_backend: "AttentionBac
             LightningAttentionBackend,
             Mamba2AttnBackend,
         )
-        from sglang.srt.utils import is_blackwell, is_npu
+        from sglang.srt.utils import is_blackwell, is_npu, is_sm120_supported
 
         check_environments()
         if runner.hybrid_gdn_config is not None:
             if is_blackwell():
-                assert (
-                    runner.server_args.attention_backend == "triton"
-                    or runner.server_args.attention_backend == "trtllm_mha"
-                ), "triton or trtllm_mha backend are the only supported backends on Blackwell GPUs for hybrid GDN models, use --attention-backend triton or --attention-backend trtllm_mha to specify the backend."
+                if is_sm120_supported():
+                    # SM12x (consumer Blackwell, e.g. RTX 5090, DGX Spark): trtllm_mha has
+                    # no SM12x cubins so it cannot be used here; triton and flashinfer both
+                    # support the full attention layers on SM12x.
+                    assert runner.server_args.attention_backend in (
+                        "triton",
+                        "flashinfer",
+                    ), "triton or flashinfer backend are the only supported backends on SM12x GPUs for hybrid GDN models, use --attention-backend triton or --attention-backend flashinfer to specify the backend."
+                else:
+                    assert (
+                        runner.server_args.attention_backend == "triton"
+                        or runner.server_args.attention_backend == "trtllm_mha"
+                    ), "triton or trtllm_mha backend are the only supported backends on Blackwell GPUs for hybrid GDN models, use --attention-backend triton or --attention-backend trtllm_mha to specify the backend."
             if is_npu():
                 assert (
                     runner.server_args.attention_backend == "ascend"

python/sglang/srt/layers/communicator.py

@@ -65,12 +65,14 @@
     is_npu,
     is_sm90_supported,
     is_sm100_supported,
+    is_sm120_supported,
 )
 
 _is_cuda = is_cuda()
 _is_flashinfer_available = is_flashinfer_available()
 _is_sm90_supported = _is_cuda and is_sm90_supported()
 _is_sm100_supported = _is_cuda and is_sm100_supported()
+_is_sm120_supported = _is_cuda and is_sm120_supported()
 _use_aiter = get_bool_env_var("SGLANG_USE_AITER") and is_hip()
 _is_gfx95_supported = is_gfx95_supported()
 _is_npu = is_npu()
@@ -92,7 +94,7 @@ def apply_flashinfer_allreduce_fusion(batch_size: int):
     return (
         # NOTE: flashinfer 0.6.1 caused performance regression on sm100 for allreduce fusion
         # Ref: https://github.com/sgl-project/sglang/issues/17237
-        (_is_sm90_supported or _is_sm100_supported)
+        (_is_sm90_supported or _is_sm100_supported or _is_sm120_supported)
         and _is_flashinfer_available
         and batch_size > 0
         and batch_size <= FUSE_ALLREDUCE_MAX_BATCH_SIZE

python/sglang/srt/server_args.py

@@ -1688,7 +1688,7 @@ def _handle_model_specific_adjustments(self):
                 "Qwen3MoeForCausalLM",
                 "KimiK25ForConditionalGeneration",
             ]
-            and (is_sm90_supported() or is_sm100_supported())
+            and (is_sm90_supported() or is_sm100_supported() or is_sm120_supported())
             and not self.enable_dp_attention
             and self.nnodes == 1
             and not is_h20_device
@@ -1704,7 +1704,7 @@ def _handle_mamba_radix_cache(
         sm100_default_attention_backend: str = None,
     ):
         if (
-            is_sm100_supported()
+            (is_sm100_supported() or is_sm120_supported())
             and self.attention_backend is None
             and sm100_default_attention_backend is not None
         ):
@@ -1820,7 +1820,7 @@ def _handle_attention_backend_compatibility(self):
                 # MLA architecture
                 if is_hopper_with_cuda_12_3():
                     self.attention_backend = "fa3"
-                elif is_sm100_supported():
+                elif is_sm100_supported() or is_sm120_supported():
                     self.attention_backend = "flashinfer"
                 elif is_hip():
                     head_num = model_config.get_num_kv_heads(self.tp_size)
@@ -1875,9 +1875,9 @@ def _handle_attention_backend_compatibility(self):
             self.attention_backend == "trtllm_mla"
             or self.decode_attention_backend == "trtllm_mla"
         ):
-            if not is_blackwell_supported():
+            if not is_sm100_supported():
                 raise ValueError(
-                    "TRTLLM MLA backend is only supported on Blackwell GPUs (SM100/SM12x). Please use a different backend."
+                    "TRTLLM MLA backend is only supported on SM100 Blackwell GPUs. Please use a different backend."
                 )
 
             if self.page_size not in [32, 64]:
@@ -2308,7 +2308,9 @@ def _handle_hicache(self):
                         )
                     else:
                         self.decode_attention_backend = (
-                            "flashinfer" if is_sm100_supported() else "triton"
+                            "flashinfer"
+                            if is_sm100_supported() or is_sm120_supported()
+                            else "triton"
                         )
                 else:
                     # If user explicitly requested FA3 decode, fall back to direct IO.