[Revert] Fix performance regression for GLM-4.7-GPTQ decode and MTP acceptance rate (#33771)

Signed-off-by: aabbccddwasd <aabbccddwasd@qq.com>

vllm/v1/attention/backends/flashinfer.py

@@ -919,9 +919,7 @@ class FlashInferMetadataBuilder(AttentionMetadataBuilder[FlashInferMetadata]):
         # Guard access to seq_lens_cpu, which may not always be needed
         # and can be expensive to retrieve in async mode.
         needs_seq_lens_cpu = self.use_dcp or use_cascade or not is_only_trtllm_decode
-        seq_lens_cpu = (
-            common_attn_metadata.seq_lens.cpu() if needs_seq_lens_cpu else None
-        )
+        seq_lens_cpu = common_attn_metadata.seq_lens_cpu if needs_seq_lens_cpu else None
         seq_lens_np = seq_lens_cpu.numpy() if seq_lens_cpu is not None else None
         num_blocks_np = (
             (seq_lens_np + (page_size - 1)) // page_size