Improve baichuan2 NPU performance (#11942)

python/llm/src/ipex_llm/transformers/npu_model.py

@@ -150,9 +150,10 @@ class _BaseAutoModelClass:
                     " than max_output_len ({max_output_len})"
                 ),
             )
-            from ipex_llm.transformers.npu_models.convert_mp import optimize_llm
+            from ipex_llm.transformers.npu_models.convert_mp import optimize_llm, optimize_llm_pre
 
             with torch.no_grad():
+                optimize_llm_pre(model)
                 cls.load_convert(qtype, model, "cpu", *args, **kwargs)
                 create_npu_kernels(model)
             model = model.eval()

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

@@ -279,9 +279,7 @@ class LowBitLlamaMultiDecoderlayer(NNFactory):
         proj = self.unsqueeze(proj, [0])  # b, s, 3, h
         proj = self.transpose(proj, [2, 1, 0, 3])  # 3, s, b, h
         proj = self.squeeze(proj)  # 3, b*s, h
-        print("proj shape: ", proj.shape)
         proj = self.unsqueeze(proj, [1])
-        print("proj shape after unsqueeze", proj.shape)
         # query_states = proj[0].view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
         query_states = self.reshape(proj[0, ...], [self.batch_size,
                                                    self.seq_len, self.num_heads, self.head_dim])

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

@@ -25,6 +25,22 @@ def convert_forward(m, target_m, new_forward):
         convert_forward(sub_m, target_m, new_forward)
 
 
+def optimize_llm_pre(model: torch.nn.Module):
+    if model.config.model_type == "baichuan":
+        # process NormHead module in Baichuan2 7B
+        if hasattr(model, 'lm_head') and model.lm_head is not None:
+            vocab_size, hidden_size = model.lm_head.weight.shape
+            lm_head_weight_data = model.lm_head.weight.data
+            model.lm_head = torch.nn.Linear(hidden_size, vocab_size, bias=False,
+                                            device=lm_head_weight_data.device)
+            if model.lm_head.weight.data.device != "meta":
+                norm_weight = torch.nn.functional.normalize(lm_head_weight_data)
+                model.lm_head.weight.data = norm_weight
+        if model.config.hidden_size in [4096, 2048]:
+            from ipex_llm.transformers.models.baichuan import pre_compute_inv_freq
+            model.apply(pre_compute_inv_freq)
+
+
 def optimize_llm(
     model: torch.nn.Module,
     max_output_len=1024,
@@ -126,6 +142,10 @@ def optimize_llm(
         convert_forward(model, module.MiniCPMModel, minicpm_model_forward)
     elif model.config.model_type == "baichuan" and model.config.num_hidden_layers == 32:
         # for Baichuan2-7B
+        if intra_pp is None:
+            intra_pp = 2
+        if inter_pp is None:
+            inter_pp = 2
         from ipex_llm.transformers.npu_models.baichuan_mp import gen_baichuan_fused_model_forward
         from ipex_llm.transformers.npu_models.baichuan_mp import DecodeRunner, PrefillRunner
         decode_runner = DecodeRunner(