Support Qwen2-7b MLP in int4 and transpose_value_cache=True (#11968)

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

@@ -65,6 +65,11 @@ def optimize_llm_pre(model: torch.nn.Module, qtype):
             model.llm.config.model_type = "llama"
         model = model.llm
 
+    if model.config.model_type == "qwen2":
+        from ipex_llm.transformers.npu_models.qwen2_mp import split_mlp_down_proj
+        from ipex_llm.transformers.npu_models.qwen2_mp import split_mlp_forward
+        model.apply(split_mlp_down_proj)
+
     # lm_head to cpu optimization
     if cpu_lm_head:
         # disable the optimization by default
@@ -134,8 +139,6 @@ def optimize_llm(
             intra_pp = 2
         if inter_pp is None:
             inter_pp = 4 if model.config.intermediate_size == 18944 else 1
-        if model.config.intermediate_size == 18944:
-            transpose_value_cache = False
 
         from ipex_llm.transformers.npu_models.qwen2_mp import gen_qwen2_fused_model_forward
         from ipex_llm.transformers.npu_models.qwen2_mp import DecodeRunner, PrefillRunner

python/llm/src/ipex_llm/transformers/npu_models/qwen2_mp.py

@@ -42,6 +42,30 @@
 from ipex_llm.transformers.npu_models.common import reshape_lm_head_input
 from transformers.modeling_outputs import CausalLMOutputWithPast
 from torch.nn import CrossEntropyLoss
+from transformers.models.qwen2.modeling_qwen2 import Qwen2MLP
+
+
+def split_mlp_down_proj(module: torch.nn.Module):
+    if isinstance(module, Qwen2MLP) and module.down_proj.in_features == 18944:
+        new_linear_0 = torch.nn.Linear(0, 0, bias=False)
+        new_weight_0 = torch.nn.Parameter(module.down_proj.weight[:, :9472], requires_grad=False)
+        new_linear_0.weight = new_weight_0
+        new_linear_0.in_features = new_weight_0.size(1)
+        new_linear_0.out_features = new_weight_0.size(0)
+        module.down_proj_0 = new_linear_0
+        new_linear_1 = torch.nn.Linear(0, 0, bias=False)
+        new_weight_1 = torch.nn.Parameter(module.down_proj.weight[:, 9472:], requires_grad=False)
+        new_linear_1.weight = new_weight_1
+        new_linear_1.in_features = new_weight_1.size(1)
+        new_linear_1.out_features = new_weight_1.size(0)
+        module.down_proj_1 = new_linear_1
+
+        del module.down_proj
+
+
+def split_mlp_forward(self, x):
+    h = self.act_fn(self.gate_proj(x)) * self.up_proj(x)
+    return self.down_proj_0(h[:, :, :9472]) + self.down_proj_1(h[:, :, 9472:])
 
 
 class LowBitQwenMultiDecoderlayer(LLMBaseNNFactory):
@@ -201,7 +225,7 @@ def __init__(
         self.compile()
         print("end compiling")
 
-    def mlp(self, hidden_states):
+    def mlp(self, hidden_states, seq_len):
         mm1 = self.linear(
             hidden_states, self.intermediate_size, self.hidden_size, bias=False, wt_dtype=self.dtype
         )
@@ -211,9 +235,13 @@ def mlp(self, hidden_states):
         mm1 = self.eltwise_mul(self.swish(mm1), mm2)  # type: ignore[attr-defined]
         if self.intermediate_size == 18944:
             # for qwen2-7b
-            hidden_states = self.linear(
-                mm1, self.hidden_size, self.intermediate_size, bias=False, wt_dtype=np.int8
-            )
+            mm1_0 = self.slice(mm1, begin=[0, 0, 0], end=[1, seq_len, 9472])
+            mm1_1 = self.slice(mm1, begin=[0, 0, 9472], end=[1, seq_len, 18944])
+            hidden_states_0 = self.linear(mm1_0, self.hidden_size, 9472,
+                                          bias=False, wt_dtype=self.dtype)
+            hidden_states_1 = self.linear(mm1_1, self.hidden_size, 9472,
+                                          bias=False, wt_dtype=self.dtype)
+            hidden_states = hidden_states_0 + hidden_states_1
         else:
             hidden_states = self.linear(
                 mm1, self.hidden_size, self.intermediate_size, bias=False, wt_dtype=self.dtype
@@ -257,7 +285,7 @@ def build_decoder(
         hidden_states = self.eltwise_add(residual, attn_output)
         residual = hidden_states
         hidden_states = self.layer_norm(hidden_states, post_attention_layernorm_weight)
-        hidden_states = self.mlp(hidden_states)
+        hidden_states = self.mlp(hidden_states, self.seq_len)
         hidden_states = self.eltwise_add(residual, hidden_states)
         hidden_states = self.convert_to_fp16(hidden_states)
 
@@ -343,9 +371,13 @@ def __init__(
             )
             self.backend_decoders.append(decoder)
 
+        offset = 0
         for i in range(intra_stages):
             start, end = self.layer_ranges[i]
-            self.backend_decoders[i].set_weights(self.op_id, op_parameters[start * 7:end * 7])
+            curr_linear_ops = len(self.backend_decoders[i].linear_ops)
+            curr_parameters = self.op_parameters[offset:offset + curr_linear_ops]
+            self.backend_decoders[i].set_weights(self.op_id, curr_parameters)
+            offset = offset + curr_linear_ops
 
     def forward(
         self,
@@ -543,7 +575,8 @@ def run_decode(
             (attn_layer.o_proj.weight, attn_layer.o_proj.scale),
             (mlp_layer.gate_proj.weight, mlp_layer.gate_proj.scale),
             (mlp_layer.up_proj.weight, mlp_layer.up_proj.scale),
-            (mlp_layer.down_proj.weight, mlp_layer.down_proj.scale),
+            (mlp_layer.down_proj_0.weight, mlp_layer.down_proj_0.scale),
+            (mlp_layer.down_proj_1.weight, mlp_layer.down_proj_1.scale)
         ]
 
         cached_cos = curr_layer.self_attn.rotary_emb.cos_cached.to(torch.float16)
@@ -814,6 +847,8 @@ def run_prefill(
             transpose_value=transpose_value_cache
         )
         convert_forward(model, Qwen2Attention, qwen2_attention_forward)
+        from transformers.models.qwen2.modeling_qwen2 import Qwen2MLP
+        convert_forward(model, Qwen2MLP, split_mlp_forward)
         deocderlayers = model.model.layers
 
     while True:
@@ -836,7 +871,6 @@ def run_prefill(
 
                 hidden_states = layer_outputs[0]
                 next_decoder_cache = layer_outputs[1]
-
             result_queue.put((hidden_states, next_decoder_cache))
 
 
@@ -1124,10 +1158,11 @@ def qwen2_attention_forward(
         cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
         query_states, key_states = apply_rotary_pos_emb(query_states, key_states,
                                                         cos, sin, position_ids)
-
         cache_kwargs = {"max_seq_len": max_seq_len, "transpose": transpose_value, }
 
         if past_key_value is not None:
+            if transpose_value:
+                value_states = value_states.transpose(-1, -2)
             key_states, value_states = past_key_value.update(key_states, value_states,
                                                              self.layer_idx, cache_kwargs)