[INTEL_HPU] added rebuild_padding_v3 for fastdeploy/pp opt (#1865)

Author: JianyuLi01

backends/intel_hpu/custom_ops/python/paddlenlp_ops/llama_block_atten.py

@@ -235,6 +235,26 @@ def rebuild_padding_v2(
     return output_data
 
 
+def rebuild_padding_v3(
+    tmp_out,
+    batch_ids,
+    total_batch,
+    seq_lens_encoder,
+    is_prompt=None,
+):
+    dim_emb = tmp_out.shape[-1]
+    output_data = paddle.zeros((batch_ids.shape[0], dim_emb))
+    if is_prompt is True:  # context
+        tmp_out = tmp_out.reshape([total_batch, -1, dim_emb])
+        for i in range(batch_ids.shape[0]):
+            seq_len = seq_lens_encoder[batch_ids[i]].item()
+            output_data[i] = tmp_out[i, seq_len - 1]
+    elif is_prompt is False:
+        output_data = tmp_out[: batch_ids.shape[0], :]
+
+    return output_data
+
+
 def fused_flatpa_proj_ref(
     query,
     key_cache,