[ROCm][Kernel][V1] Enable AMD Radeon GPU Custom Paged Attention on v1 (#17004)

Signed-off-by: Hosang Yoon <[email protected]>

Author: hyoon1

benchmarks/kernels/benchmark_paged_attention.py

@@ -84,7 +84,10 @@ def main(
     if version == "v2":
         if current_platform.is_rocm():
             global PARTITION_SIZE
-            PARTITION_SIZE = 1024 if not args.custom_paged_attn else PARTITION_SIZE_ROCM
+            if not args.custom_paged_attn and not current_platform.is_navi():
+                PARTITION_SIZE = 1024
+            else:
+                PARTITION_SIZE = PARTITION_SIZE_ROCM
         num_partitions = (max_seq_len + PARTITION_SIZE - 1) // PARTITION_SIZE
         tmp_output = torch.empty(
             size=(num_seqs, num_query_heads, num_partitions, head_size),
@@ -159,6 +162,7 @@ def run_cuda_benchmark(num_iters: int, profile: bool = False) -> float:
                         scale,
                         block_tables,
                         seq_lens,
+                        None,
                         block_size,
                         max_seq_len,
                         alibi_slopes,