[Perf] Improve fp8 quant in mla; replace ReduceSum with ReduceScatterSum (#29795)

Signed-off-by: Siyuan Fu <siyuanf@nvidia.com>

Author: IwakuraRein

vllm/distributed/device_communicators/cuda_communicator.py

@@ -225,7 +225,7 @@ def reduce_scatterv(
             output_shape, dtype=input_tensor.dtype, device=input_tensor.device
         )
 
-        if sizes is not None:
+        if sizes is not None and sizes.count(sizes[0]) != len(sizes):
             pynccl_comm.reduce_scatterv(output, input_tensor, sizes=sizes)
         else:
             pynccl_comm.reduce_scatter(output, input_tensor)

vllm/v1/attention/backends/mla/common.py

@@ -2037,21 +2037,30 @@ def forward(
 
             if fp8_attention:
                 ql_nope_shape = decode_ql_nope.shape
-                decode_ql_nope, _ = ops.scaled_fp8_quant(
-                    decode_ql_nope.reshape(
-                        [ql_nope_shape[0], ql_nope_shape[1] * ql_nope_shape[2]]
-                    ),
-                    layer._q_scale,
-                )
-                decode_ql_nope = decode_ql_nope.reshape(ql_nope_shape)
                 q_pe_shape = decode_q_pe.shape
-                decode_q_pe, _ = ops.scaled_fp8_quant(
-                    decode_q_pe.reshape([q_pe_shape[0], q_pe_shape[1] * q_pe_shape[2]]),
-                    layer._q_scale,
+                assert decode_ql_nope.shape[0] == decode_q_pe.shape[0]
+                assert decode_ql_nope.shape[1] == decode_q_pe.shape[1]
+                decode_q_shape = (
+                    ql_nope_shape[0],
+                    ql_nope_shape[1],
+                    ql_nope_shape[2] + q_pe_shape[2],
+                )
+                # Using empty and copy since torch.cat introduces significant overhead.
+                decode_q0 = torch.empty(
+                    decode_q_shape,
+                    device=decode_ql_nope.device,
+                    dtype=decode_ql_nope.dtype,
                 )
-                decode_q_pe = decode_q_pe.reshape(q_pe_shape)
+                decode_q0[..., : ql_nope_shape[2]].copy_(decode_ql_nope)
+                decode_q0[..., ql_nope_shape[2] :].copy_(decode_q_pe)
 
-            decode_q = (decode_ql_nope, decode_q_pe)
+                decode_q, _ = ops.scaled_fp8_quant(
+                    decode_q0.view(decode_q_shape[0], -1),
+                    layer._q_scale,
+                )
+                decode_q = decode_q.view(decode_q_shape)
+            else:
+                decode_q = (decode_ql_nope, decode_q_pe)
             if self.dcp_world_size > 1:
                 assert not fp8_attention, "DCP not support fp8 kvcache now."
                 # concatenate decode_ql_nope and decode_q_pe -> (B, N, L + P)