remove ck fa interface

vllm/attention/backends/rocm_flash_attn.py

@@ -372,7 +372,7 @@ class ROCmFlashAttentionImpl(AttentionImpl):
                 try:
                     from flash_attn import flash_attn_varlen_func  # noqa: F401
                     self.attn_func = flash_attn_varlen_func
-                    logger.debug("Using CK/CUTLASS FA in ROCmBackend")
+                    logger.debug("Using CUTLASS FA in ROCmBackend")
                 except ModuleNotFoundError:
                     self.use_naive_attn = True
 
@@ -522,9 +522,6 @@ class ROCmFlashAttentionImpl(AttentionImpl):
                         attn_masks,
                     )
                 else:
-                    import flash_attn
-                    major, minor, _ = flash_attn.__version__.split('.')
-                    if (major, minor) >= ('2', '6'):
                     out = self.attn_func(
                         q=query,
                         k=key,
@@ -538,18 +535,6 @@ class ROCmFlashAttentionImpl(AttentionImpl):
                         window_size=self.sliding_window,
                         alibi_slopes=self.alibi_slopes,
                     )
-                    else:
-                        out = self.attn_func(
-                            q=query,
-                            k=key,
-                            v=value,
-                            cu_seqlens_q=prefill_meta.seq_start_loc,
-                            cu_seqlens_k=prefill_meta.seq_start_loc,
-                            max_seqlen_q=prefill_meta.max_prefill_seq_len,
-                            max_seqlen_k=prefill_meta.max_prefill_seq_len,
-                            softmax_scale=self.scale,
-                            causal=True,
-                        )
 
                 # common code for prefill
                 assert output[:num_prefill_tokens].shape == out.shape