only K100_AI can use cutlass prefix cache

vllm/attention/backends/rocm_flash_attn.py

@@ -857,7 +857,7 @@ class ROCmFlashAttentionImpl(AttentionImpl):
             else:
                 # prefix-enabled attention -
                 # not applicable for encoder-only models
-                if envs.VLLM_USE_TRITON_PREFIX_FLASH_ATTN or gpuname.startswith('BW'):   
+                if envs.VLLM_USE_TRITON_PREFIX_FLASH_ATTN or (not gpuname.startswith('K100_AI')):   
                     version_key = triton_key()
                     if self.attn_type != AttentionType.ENCODER_ONLY:
                         output[:num_prefill_tokens] = paged_attn.forward_prefix(