[NPU] acclib llama3.2 support groupwise (#12355)

* change inter_pp

* add comment

python/llm/src/ipex_llm/transformers/npu_models/convert_mp.py

@@ -337,7 +337,17 @@ def optimize_llm(
         if intra_pp is None:
             intra_pp = 2
         if inter_pp is None:
-            inter_pp = 2 if group_size == 0 else 8
+            if group_size == 0:
+                inter_pp = 2
+            # llama3.2
+            elif model.config.intermediate_size == 8192:
+                # llama3.2 1b
+                if model.config.hidden_size == 2048:
+                    inter_pp = 1
+                else:
+                    inter_pp = 2
+            else:
+                inter_pp = 8
         convert_llama(model,
                       max_output_len=max_context_len,
                       max_prompt_len=max_prompt_len,