[2/2] Speed up trtllm_mla attention backend (#10474)

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

@@ -22,7 +22,7 @@ from sglang.srt.layers.attention.utils import (
 from sglang.srt.layers.dp_attention import get_attention_tp_size
 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.utils import is_flashinfer_available
+from sglang.srt.utils import is_cuda, is_flashinfer_available
 
 if is_flashinfer_available():
     import flashinfer
@@ -32,6 +32,11 @@ if TYPE_CHECKING:
     from sglang.srt.model_executor.model_runner import ModelRunner
     from sglang.srt.speculative.spec_info import SpecInfo
 
+_is_cuda = is_cuda()
+
+if _is_cuda:
+    from sgl_kernel import concat_mla_absorb_q
+
 # Constants
 DEFAULT_WORKSPACE_SIZE_MB = 128  # Memory workspace size in MB
 
@@ -482,6 +487,9 @@ class TRTLLMMLABackend(FlashInferMLAAttnBackend):
             q_rope_reshaped = q_rope.view(
                 -1, layer.tp_q_head_num, layer.head_dim - layer.v_head_dim
             )
+            if _is_cuda and q_nope.shape[-1] == 512 and q_rope_reshaped.shape[-1] == 64:
+                query = concat_mla_absorb_q(q_nope, q_rope_reshaped)
+            else:
                 query = torch.cat([q_nope, q_rope_reshaped], dim=-1)
         else:
             # For FP8 path, we already have the query and rope parts merged because of the quantize_and_rope_for_fp8 function