ipex-llm 2.2.0b20250108__py3-none-win_amd64.whl → 2.2.0b20250110__py3-none-win_amd64.whl

ipex_llm/transformers/convert.py

@@ -1325,7 +1325,6 @@ def _optimize_post(model):
             modeling_module_name = model.__class__.__module__
             module = importlib.import_module(modeling_module_name)
             from ipex_llm.transformers.models.chatglm2 import chatglm2_attention_forward
-            from ipex_llm.transformers.models.chatglm2 import chatglm_rms_norm_forward
             from ipex_llm.transformers.models.chatglm2 import chatglm2_encoder_forward
             from ipex_llm.transformers.models.chatglm2 import chatglm2_model_forward
             from ipex_llm.transformers.models.chatglm2 import mlp_forward
@@ -1338,9 +1337,7 @@ def _optimize_post(model):
             convert_forward(model,
                             module.ChatGLMModel,
                             chatglm2_model_forward)
-            convert_forward(model,
-                            module.RMSNorm,
-                            chatglm_rms_norm_forward)
+            convert_forward(model, module.RMSNorm, rms_norm_forward)
             convert_forward(model, module.MLP, mlp_forward)
             # for codegeex-nano
             if hasattr(model.config, "rope_ratio"):
@@ -1358,8 +1355,7 @@ def _optimize_post(model):
             # glm4 family
             modeling_module_name = model.__class__.__module__
             module = importlib.import_module(modeling_module_name)
-            from ipex_llm.transformers.models.chatglm2 import chatglm_rms_norm_forward
-            convert_forward(model, module.RMSNorm, chatglm_rms_norm_forward)
+            convert_forward(model, module.RMSNorm, rms_norm_forward)
 
             if hasattr(model.transformer, "vision"):
                 # glm4 vision family
@@ -1448,8 +1444,8 @@ def _optimize_post(model):
     elif model.config.model_type == "baichuan":
         modeling_module_name = model.__class__.__module__
         module = importlib.import_module(modeling_module_name)
-        from ipex_llm.transformers.models.baichuan import baichuan_mlp_forward
-        convert_forward(model, module.MLP, baichuan_mlp_forward)
+        convert_forward(model, module.RMSNorm, rms_norm_forward)
+        convert_forward(model, module.MLP, mlp_silu_forward)
 
         if model.config.hidden_size in [4096, 2048]:
             # baichuan-7B and baichuan2-7B
@@ -1458,7 +1454,6 @@ def _optimize_post(model):
             for i in range(len(model.model.layers)):
                 setattr(model.model.layers[i].self_attn, "layer_idx", i)
             convert_forward(model, module.Attention, baichuan_attention_forward_7b)
-            convert_forward(model, module.RMSNorm, rms_norm_forward)
             if model.config.vocab_size == 125696:
                 # baichuan2-7B
                 convert_forward(model, module.BaichuanModel, baichuan_model_7b_forward)
@@ -1468,9 +1463,7 @@ def _optimize_post(model):
         elif model.config.hidden_size == 5120:
             # baichuan-13B and baichuan2-13B
             from ipex_llm.transformers.models.baichuan import baichuan_attention_forward_13b
-            from ipex_llm.transformers.models.baichuan import baichuan_13b_rms_norm_forward
             convert_forward(model, module.BaichuanAttention, baichuan_attention_forward_13b)
-            convert_forward(model, module.RMSNorm, baichuan_13b_rms_norm_forward)
 
             if model.config.vocab_size == 125696:
                 # baichaun2-13B
@@ -1565,7 +1558,6 @@ def _optimize_post(model):
             from ipex_llm.transformers.models.qwen import qwen_attention_forward
             from ipex_llm.transformers.models.qwen import qwen_attention_forward_registered
             from ipex_llm.transformers.models.qwen import qwen_mlp_forward
-            from ipex_llm.transformers.models.chatglm2 import chatglm_rms_norm_forward
             from ipex_llm.transformers.models.qwen import qwen_model_forward
             if model.config.max_position_embeddings == 8192 \
                and model.config.hidden_size == 4096:
@@ -1580,7 +1572,7 @@ def _optimize_post(model):
                                 )
             convert_forward(model,
                             module.RMSNorm,
-                            chatglm_rms_norm_forward)
+                            rms_norm_forward)
             convert_forward(model,
                             module.QWenMLP,
                             qwen_mlp_forward)
@@ -1598,6 +1590,9 @@ def _optimize_post(model):
         convert_forward(model,
                         module.Qwen2ForCausalLM,
                         qwen2_causal_lm_forward)
+        convert_forward(model,
+                        module.Qwen2Model,
+                        qwen2_model_forward)
         convert_forward(model,
                         module.Qwen2RMSNorm,
                         rms_norm_forward)
@@ -1610,12 +1605,6 @@ def _optimize_post(model):
         convert_forward(model,
                         module.Qwen2SdpaAttention,
                         qwen2_attention_forward)
-        if version.parse(trans_version) >= version.parse("4.42"):
-            from ipex_llm.transformers.models.qwen2 import qwen2_model_forward_4_42
-            convert_forward(model, module.Qwen2Model, qwen2_model_forward_4_42)
-        else:
-            from ipex_llm.transformers.models.qwen2 import qwen2_model_forward
-            convert_forward(model, module.Qwen2Model, qwen2_model_forward)
     elif model.config.model_type == "qwen2_moe":
         # for Qwen1.5-MOE-A2.7B
         modeling_module_name = model.__class__.__module__
@@ -1827,9 +1816,7 @@ def _optimize_post(model):
         from ipex_llm.transformers.models.phi3 import attention_forward
         convert_forward(model, module.Phi3Attention, attention_forward)
         convert_forward(model, module.Phi3SdpaAttention, attention_forward)
-        from ipex_llm.transformers.models.phi3 import mlp_forward
-        convert_forward(model, module.Phi3MLP, mlp_forward)
-        from ipex_llm.transformers.models.common import rms_norm_forward
+        convert_forward(model, module.Phi3MLP, mlp_silu_forward)
         convert_forward(model, module.Phi3RMSNorm, rms_norm_forward)
         if model.config.model_type == "phi3":
             from ipex_llm.transformers.models.phi3 import phi3_model_forward_wrapper

ipex_llm/transformers/convert_ipex.py

@@ -52,7 +52,14 @@ import os
 
 
 def _ipex_optimize_rmsnorm(_model, supported_classes, is_tpp=False, is_woq=False):
-    from intel_extension_for_pytorch.transformers.models.cpu.fusions.mha_fusion import _IPEXRMSNorm
+    try:
+        # old version use name `_IPEXRMSNorm`
+        from intel_extension_for_pytorch.transformers.models.cpu.fusions.mha_fusion \
+            import _IPEXRMSNorm
+    except ImportError:
+        # new version use name `_IPEXRMSNormCPU`
+        from intel_extension_for_pytorch.transformers.models.cpu.fusions.mha_fusion \
+            import _IPEXRMSNormCPU as _IPEXRMSNorm
     for supported_class in supported_classes:
         lowering_class_cpu(
             _model,

ipex_llm/transformers/low_bit_linear.py

@@ -47,7 +47,7 @@ import os
 import torch
 import torch.distributed
 import torch.nn.functional as F
-from torch import Tensor, device, dtype, nn
+from torch import Tensor, dtype, nn
 from operator import mul
 from functools import reduce
 from ipex_llm.transformers.xpu_customize_fwd import custom_fwd, custom_bwd
@@ -294,10 +294,10 @@ def use_batch_forward(x: torch.Tensor, qtype: int, output_len: int):
     if hard_condition:
         return (
             batch_size > 1
-            or (device in ["arc"] and qtype in [SYM_INT8, FP4])
-            or (device in ["arc", "mtl"] and qtype in [FP8E4])
-            or (device in ["lnl"] and qtype in [SYM_INT4] and x.shape[1] % 512 == 0)
-            or (device in ["bmg"] and qtype in [SYM_INT4, FP8E5])
+            or (device_name in ["arc"] and qtype in [SYM_INT8, FP4])
+            or (device_name in ["arc", "mtl"] and qtype in [FP8E4])
+            or (device_name in ["lnl"] and qtype in [SYM_INT4] and x.shape[1] % 512 == 0)
+            or (device_name in ["bmg"] and qtype in [SYM_INT4, FP8E5])
         )
     return False
 

ipex_llm/transformers/models/baichuan.py

@@ -30,8 +30,7 @@ from ipex_llm.transformers.models.utils import use_quantize_kv_cache, restore_fp
 from ipex_llm.transformers.models.utils import update_past_key_value
 from ipex_llm.transformers.models.utils import should_use_fuse_rope
 from ipex_llm.transformers.models.utils import use_sdp
-from ipex_llm.transformers.models.utils import apply_rotary_pos_emb, SILU
-from ipex_llm.transformers.models.utils import mlp_fusion_check
+from ipex_llm.transformers.models.utils import apply_rotary_pos_emb
 from ipex_llm.transformers.models.utils import is_enough_kv_cache_room_4_36
 from ipex_llm.transformers.kv import DynamicCompressFp8Cache, DynamicCompressCache
 import warnings
@@ -47,38 +46,6 @@ def pre_compute_inv_freq(module: torch.nn.Module):
         module.register_buffer("inv_freq", inv_freq, persistent=False)
 
 
-def baichuan_13b_rms_norm_forward(self, hidden_states):
-    if hidden_states.device.type == "xpu" and not (self.training or hidden_states.requires_grad):
-        import xe_addons
-        x_2d = hidden_states.reshape(-1, hidden_states.size(-1)).contiguous()
-        output = xe_addons.rms_norm(self.weight, x_2d, self.epsilon)
-        return output.reshape(hidden_states.shape)
-
-    input_dtype = hidden_states.dtype
-    hidden_states = hidden_states.to(torch.float32)
-    variance = hidden_states.pow(2).mean(-1, keepdim=True)
-    hidden_states = hidden_states * torch.rsqrt(variance + self.epsilon)
-    return self.weight * hidden_states.to(input_dtype)
-
-
-def baichuan_mlp_forward(
-    self,
-    x: torch.Tensor,
-) -> torch.Tensor:
-    x_2d = x.view(-1, x.shape[-1])
-    qtype = getattr(self.gate_proj, "qtype", None)
-    if mlp_fusion_check(x_2d, qtype, self.training):
-        import xe_linear
-        if not x_2d.is_contiguous():
-            x_2d = x_2d.contiguous()
-        return self.down_proj(xe_linear.mlp_forward_xpu(
-            x_2d, self.gate_proj.weight.data, self.up_proj.weight.data,
-            x_2d.shape[0], x_2d.shape[1], self.gate_proj.out_len,
-            SILU, qtype
-        ))
-    return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
-
-
 def baichuan_model_7b_forward(
         self,
         input_ids: torch.LongTensor = None,
@@ -105,7 +72,9 @@ def baichuan_model_7b_forward(
     if use_cache:
         inputs = input_ids if input_ids is not None else inputs_embeds
         use_compress_kv = should_use_compresskv(inputs, inputs.shape[1])
-        use_quantize_kv = use_quantize_kv_cache(self.layers[0].mlp.up_proj, inputs)
+        use_quantize_kv = use_quantize_kv_cache(self.layers[0].mlp.up_proj, inputs,
+                                                self.config.num_attention_heads,
+                                                self.config.num_attention_heads)
         if use_compress_kv and not isinstance(past_key_values,
                                               DynamicCompressCache):
             if use_quantize_kv:
@@ -278,8 +247,6 @@ def baichuan_attention_forward_7b(
         key_states = key_states.to(hidden_states.dtype)
 
     # IPEX-LLM OPT: kv cache and quantize kv
-    use_quantize_kv = use_quantize_kv_cache(self.W_pack, hidden_states)
-
     # [CompressKV]
     if use_compresskv:
         enough_kv_room = is_enough_kv_cache_room_4_36(past_key_value,
@@ -290,6 +257,8 @@ def baichuan_attention_forward_7b(
             query_states, attention_mask, 1,
             self.config, enough_kv_room, KV_CACHE_ALLOC_BLOCK_LENGTH)
     else:
+        use_quantize_kv = use_quantize_kv_cache(self.W_pack, hidden_states,
+                                                self.num_heads, self.num_heads)
         key_states, value_states = update_past_key_value(
             past_key_value, key_states, value_states,
             kv_seq_len, use_quantize_kv, device
@@ -340,7 +309,8 @@ def baichuan_attention_forward_13b(
         kv_seq_len += past_key_value[0].shape[2]
 
     # IPEX-LLM OPT: kv cache and quantize kv
-    use_quantize_kv = use_quantize_kv_cache(self.W_pack, hidden_states)
+    use_quantize_kv = use_quantize_kv_cache(self.W_pack, hidden_states,
+                                            self.num_heads, self.num_heads)
     key_states, value_states = update_past_key_value(
         past_key_value, key_states, value_states,
         kv_seq_len, use_quantize_kv, device

ipex_llm/transformers/models/bert.py

@@ -36,24 +36,13 @@ import math
 import torch
 from typing import Optional, Tuple
 from transformers.models.bert.modeling_bert import BertSelfAttention, BertEncoder
+from ipex_llm.transformers.models.common import merge_linear
 from ipex_llm.utils.common import invalidInputError
 
 
 def merge_qkv(module: torch.nn.Module):
     if isinstance(module, BertSelfAttention):
-        q_w = module.query.weight.data
-        k_w = module.key.weight.data
-        v_w = module.value.weight.data
-        q_b = module.query.bias.data
-        k_b = module.key.bias.data
-        v_b = module.value.bias.data
-        new_w = torch.cat([q_w, k_w, v_w], dim=0)
-        new_b = torch.cat([q_b, k_b, v_b], dim=-1)
-        qkv = torch.nn.Linear(0, 0, bias=True)
-        qkv.weight = torch.nn.Parameter(new_w, requires_grad=False)
-        qkv.bias = torch.nn.Parameter(new_b, requires_grad=False)
-        qkv.in_features = module.query.in_features
-        qkv.out_features = module.query.out_features * 3
+        qkv = merge_linear([module.query, module.key, module.value])
         module.qkv = qkv
         del module.query
         del module.key

ipex_llm/transformers/models/chatglm2.py

@@ -33,34 +33,6 @@ from ipex_llm.transformers.kv import DynamicCompressCache, DynamicCompressFp8Cac
 KV_CACHE_ALLOC_BLOCK_LENGTH = int(os.environ.get("KV_CACHE_ALLOC_BLOCK_LENGTH", 256))
 
 
-def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
-    """
-    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states
-    go from (batch, num_key_value_heads, seqlen, head_dim) to
-    (batch, num_attention_heads, seqlen, head_dim)
-    """
-    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
-    if n_rep == 1:
-        return hidden_states
-    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads,
-                                                           n_rep, slen, head_dim)
-    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
-
-
-def chatglm_rms_norm_forward(self, hidden_states):
-    if hidden_states.device.type == "xpu" and not (self.training and hidden_states.requires_grad):
-        import xe_addons
-        x_2d = hidden_states.reshape(-1, hidden_states.size(-1)).contiguous()
-        output = xe_addons.rms_norm(self.weight, x_2d, self.eps)
-        return output.reshape(hidden_states.shape)
-
-    input_dtype = hidden_states.dtype
-    hidden_states = hidden_states.to(torch.float32)
-    variance = hidden_states.pow(2).mean(-1, keepdim=True)
-    hidden_states = hidden_states * torch.rsqrt(variance + self.eps)
-    return self.weight * hidden_states.to(input_dtype)
-
-
 def chatglm2_model_forward(
     self,
     input_ids,
@@ -91,8 +63,13 @@ def chatglm2_model_forward(
 
     if use_cache:
         use_compress_kv = should_use_compresskv(input_ids, input_ids.shape[1])
+        n_heads = self.config.num_attention_heads
+        if self.config.multi_query_attention:
+            n_kv_heads = self.config.multi_query_group_num
+        else:
+            n_kv_heads = n_heads
         use_quantize_kv = use_quantize_kv_cache(self.encoder.layers[0].mlp.gate_proj,
-                                                input_ids)
+                                                input_ids, n_heads, n_kv_heads)
         if use_compress_kv and not isinstance(past_key_values,
                                               DynamicCompressCache):
             if use_quantize_kv:
@@ -285,8 +262,6 @@ def chatglm2_attention_forward(
         key_states[..., :rot_dim] = k_rot[...]
 
     # IPEX-LLM OPT: kv cache and quantize kv
-    use_quantize_kv = use_quantize_kv_cache(self.query_key_value, query_states)
-
     # [CompressKV]
     if use_compresskv:
         from transformers.configuration_utils import PretrainedConfig
@@ -300,6 +275,8 @@ def chatglm2_attention_forward(
             self.config, enough_kv_room, KV_CACHE_ALLOC_BLOCK_LENGTH
         )
     else:
+        use_quantize_kv = use_quantize_kv_cache(self.query_key_value, query_states,
+                                                n_head, n_kv_head)
         key_states, value_states = update_past_key_value(
             past_key_value, key_states, value_states,
             kv_seq_len, use_quantize_kv, hidden_states.device

ipex_llm/transformers/models/chatglm4.py

@@ -55,8 +55,13 @@ def chatglm4_model_forward(
     if use_cache:
         inputs = input_ids if input_ids is not None else inputs_embeds
         use_compress_kv = should_use_compresskv(inputs, inputs.shape[1])
-        use_quantize_kv = use_quantize_kv_cache(self.encoder.layers[0].mlp.gate_proj,
-                                                inputs)
+        n_heads = self.config.num_attention_heads
+        if self.config.multi_query_attention:
+            n_kv_heads = self.config.multi_query_group_num
+        else:
+            n_kv_heads = n_heads
+        use_quantize_kv = use_quantize_kv_cache(self.encoder.layers[0].mlp.gate_proj, inputs,
+                                                n_heads, n_kv_heads)
         if use_compress_kv and not isinstance(past_key_values,
                                               DynamicCompressCache):
             if use_quantize_kv:
@@ -211,8 +216,6 @@ def chatglm4_attention_forward(
         key_states[..., :rot_dim] = k_rot[...]
 
     # IPEX-LLM OPT: kv cache and quantize kv
-    use_quantize_kv = use_quantize_kv_cache(self.query_key_value, query_states)
-
     # [CompressKV]
     if use_compresskv:
         from transformers.configuration_utils import PretrainedConfig
@@ -226,6 +229,8 @@ def chatglm4_attention_forward(
             self.config, enough_kv_room, KV_CACHE_ALLOC_BLOCK_LENGTH
         )
     else:
+        use_quantize_kv = use_quantize_kv_cache(self.query_key_value, query_states,
+                                                n_head, n_kv_head)
         key_states, value_states = update_past_key_value(
             past_key_value, key_states, value_states,
             kv_seq_len, use_quantize_kv, hidden_states.device

ipex_llm/transformers/models/chatglm4v.py

@@ -230,7 +230,7 @@ def chatglm4v_attention_forward(
         key_states[..., :rot_dim] = k_rot[...]
 
     # IPEX-LLM OPT: kv cache and quantize kv
-    use_quantize_kv = use_quantize_kv_cache(self.query_key_value, query_states)
+    use_quantize_kv = use_quantize_kv_cache(self.query_key_value, query_states, n_head, n_kv_head)
     key_states, value_states = update_past_key_value(
         past_key_value, key_states, value_states,
         kv_seq_len, use_quantize_kv, hidden_states.device

ipex_llm/transformers/models/common.py

@@ -157,8 +157,10 @@ def rms_norm_forward(self, hidden_states: torch.Tensor):
     weight = self.weight
     if hasattr(self, "variance_epsilon"):
         eps = self.variance_epsilon
-    else:
+    elif hasattr(self, "epsilon"):
         eps = self.epsilon
+    else:
+        eps = self.eps
 
     if hidden_states.device.type == 'xpu' and hidden_states.dtype in [torch.float, torch.half]:
         import xe_addons

ipex_llm/transformers/models/glm.py

@@ -147,7 +147,7 @@ def glm_model_forward_wrapper(origin_forward):
         use_cache = use_cache if use_cache is not None else self.config.use_cache
         use_cache = use_cache or inputs.device.type == 'xpu'
         use_quantize_kv = use_quantize_kv_cache(self.layers[0].mlp.down_proj, inputs,
-                                                self.config.num_attention_heads //
+                                                self.config.num_attention_heads,
                                                 self.config.num_key_value_heads)
 
         if use_cache:

ipex_llm/transformers/models/internlm.py

@@ -87,7 +87,8 @@ def internlm_attention_forward(
         )
 
     # IPEX-LLM OPT: kv cache and quantzie kv cache
-    use_quantize_kv = use_quantize_kv_cache(self.qkv_proj, hidden_states)
+    use_quantize_kv = use_quantize_kv_cache(self.qkv_proj, hidden_states,
+                                            self.num_heads, self.num_heads)
     key_states, value_states = update_past_key_value(
         past_key_value, key_states, value_states,
         kv_seq_len, use_quantize_kv, hidden_states.device
@@ -112,21 +113,6 @@ def internlm_attention_forward(
     return attn_output, attn_weights, past_key_value
 
 
-def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
-    """
-    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep).
-    The hidden states go from (batch,
-    num_key_value_heads, seqlen, head_dim) to
-    (batch, num_attention_heads, seqlen, head_dim)
-    """
-    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
-    if n_rep == 1:
-        return hidden_states
-    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads,
-                                                           n_rep, slen, head_dim)
-    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
-
-
 def internlm2_attention_forward(
     self,
     hidden_states: torch.Tensor,
@@ -171,7 +157,8 @@ def internlm2_attention_forward(
         )
 
     # IPEX-LLM OPT: kv cache and quantzie kv cache
-    use_quantize_kv = use_quantize_kv_cache(self.wqkv, hidden_states)
+    use_quantize_kv = use_quantize_kv_cache(self.wqkv, hidden_states,
+                                            self.num_heads, self.num_key_value_heads)
     key_states, value_states = update_past_key_value(
         past_key_value, key_states, value_states,
         kv_seq_len, use_quantize_kv, hidden_states.device
@@ -346,7 +333,8 @@ def internlm_xcomposser2_attention_forward(
             query_states, key_states, cos, sin, position_ids, "internlm")
 
     # IPEX-LLM OPT: kv cache and quantzie kv cache
-    use_quantize_kv = use_quantize_kv_cache(self.wqkv, hidden_states)
+    use_quantize_kv = use_quantize_kv_cache(self.wqkv, hidden_states,
+                                            self.num_heads, self.num_key_value_heads)
     key_states, value_states = update_past_key_value(
         past_key_value, key_states, value_states,
         kv_seq_len, use_quantize_kv, device

ipex_llm/transformers/models/llama.py

@@ -72,7 +72,7 @@ def llama_model_forward(
     use_cache = True if inputs.device.type == "xpu" else use_cache
     use_quantize_kv = use_quantize_kv_cache(
         self.layers[0].mlp.down_proj, inputs,
-        self.config.num_attention_heads // self.config.num_key_value_heads
+        self.config.num_attention_heads, self.config.num_key_value_heads
     )
     use_compresskv = should_use_compresskv(inputs, inputs.shape[1]) or \
         isinstance(past_key_values, DynamicCompressCache)

ipex_llm/transformers/models/minicpm.py

@@ -159,7 +159,7 @@ def minicpm_model_forward_wrapper(origin_forward):
         # IPEX-LLM OPT: kv cache and quantize kv cache
         inputs = input_ids if input_ids is not None else inputs_embeds
         use_quantize_kv = use_quantize_kv_cache(self.layers[0].mlp.up_proj, inputs,
-                                                self.config.num_attention_heads //
+                                                self.config.num_attention_heads,
                                                 self.config.num_key_value_heads)
         use_compress_kv = should_use_compresskv(inputs, inputs.shape[1]) or \
             isinstance(past_key_values, DynamicCompressCache)

ipex_llm/transformers/models/minicpm3.py

@@ -66,7 +66,9 @@ def minicpm3_model_forward_wrapper(origin_forward):
         inputs = input_ids if input_ids is not None else inputs_embeds
         use_cache = use_cache if use_cache is not None else self.config.use_cache
         use_cache = True if inputs.device.type == "xpu" else use_cache
-        use_quantize_kv = use_quantize_kv_cache(self.layers[0].mlp.down_proj, inputs)
+        num_heads, num_kv_heads = self.config.num_attention_heads, self.config.num_key_value_heads
+        use_quantize_kv = use_quantize_kv_cache(self.layers[0].mlp.down_proj, inputs,
+                                                num_heads, num_kv_heads)
         if use_cache:
             if use_quantize_kv and not isinstance(past_key_values, DynamicFp8Cache):
                 past_key_values = DynamicFp8Cache.from_legacy_cache(past_key_values)

ipex_llm/transformers/models/mistral.py

@@ -71,7 +71,7 @@ def mistral_model_forward(
     use_cache = use_cache if use_cache is not None else self.config.use_cache
     use_cache = use_cache or inputs.device.type == 'xpu'
     use_quantize_kv = use_quantize_kv_cache(self.layers[0].mlp.down_proj, inputs,
-                                            self.config.num_attention_heads //
+                                            self.config.num_attention_heads,
                                             self.config.num_key_value_heads)
     use_compress_kv = should_use_compresskv(inputs, inputs.size(1)) or \
         isinstance(past_key_values, DynamicCompressCache)

ipex_llm/transformers/models/mllama.py

@@ -113,7 +113,7 @@ def mllama_text_model_forward(
     use_cache = True if inputs.device.type == "xpu" else use_cache
     use_quantize_kv = use_quantize_kv_cache(
         self.layers[0].mlp.down_proj, inputs,
-        self.config.num_attention_heads // self.config.num_key_value_heads
+        self.config.num_attention_heads, self.config.num_key_value_heads
     )
     if use_cache:
         if use_quantize_kv and not isinstance(past_key_values, DynamicFp8Cache):

ipex_llm/transformers/models/phi3.py

@@ -39,7 +39,6 @@ import warnings
 from ipex_llm.transformers.models.common import attention_softmax
 from ipex_llm.transformers.models.common import scaled_dot_product_attention
 from ipex_llm.transformers.models.utils import should_use_fuse_rope, rotate_half
-from ipex_llm.transformers.models.utils import mlp_fusion_check, SILU
 from ipex_llm.transformers.models.utils import use_sdp, use_sdp_causal
 from ipex_llm.transformers.models.utils import use_quantize_kv_cache, restore_fp8_kv_cache
 from ipex_llm.transformers.models.utils import should_use_compresskv, is_enough_kv_cache_room_4_36
@@ -213,24 +212,8 @@ def split_mlp(module: torch.nn.Module):
 
         del module.gate_up_proj
 
-
-def mlp_forward(
-    self,
-    hidden_states: torch.FloatTensor
-) -> torch.FloatTensor:
-    x_2d = hidden_states.view(-1, hidden_states.shape[-1])
-    qtype = getattr(self.gate_proj, "qtype", None)
-    if mlp_fusion_check(x_2d, qtype, self.training):
-        x_2d = x_2d.contiguous()
-        import xe_linear
-        return self.down_proj(xe_linear.mlp_forward_xpu(
-            x_2d, self.gate_proj.weight.data, self.up_proj.weight.data,
-            x_2d.shape[0], x_2d.shape[1], self.gate_proj.out_features,
-            SILU, qtype
-        ))
-    return self.down_proj(
-        self.activation_fn(self.gate_proj(hidden_states)) * self.up_proj(hidden_states)
-    )
+        # rename activation function
+        module.act_fn = module.activation_fn
 
 
 def phi3_model_forward_wrapper(origin_model_forward):
@@ -249,7 +232,9 @@ def phi3_model_forward_wrapper(origin_model_forward):
         # IPEX-LLM OPT: kv cache and quantize kv cache and sdp
         use_cache = use_cache if use_cache is not None else self.config.use_cache
         inputs = input_ids if input_ids is not None else inputs_embeds
-        use_quantize_kv = use_quantize_kv_cache(self.layers[0].mlp.down_proj, inputs)
+        num_heads, num_kv_heads = self.config.num_attention_heads, self.config.num_key_value_heads
+        use_quantize_kv = use_quantize_kv_cache(self.layers[0].mlp.down_proj, inputs,
+                                                num_heads, num_kv_heads)
         use_compress_kv = should_use_compresskv(inputs, inputs.shape[1]) or \
             isinstance(past_key_values, DynamicCompressCache)
         if use_cache:
@@ -305,7 +290,9 @@ def phi3v_model_forward_wrapper(origin_model_forward):
     ):
         # IPEX-LLM OPT: kv cache and quantize kv cache and sdp
         use_cache = use_cache if use_cache is not None else self.config.use_cache
-        use_quantize_kv = use_quantize_kv_cache(self.layers[0].mlp.down_proj, input_ids)
+        num_heads, num_kv_heads = self.config.num_attention_heads, self.config.num_key_value_heads
+        use_quantize_kv = use_quantize_kv_cache(self.layers[0].mlp.down_proj, input_ids,
+                                                num_heads, num_kv_heads)
         if use_cache:
             if use_quantize_kv and not isinstance(past_key_values, DynamicFp8Cache):
                 past_key_values = DynamicFp8Cache.from_legacy_cache(past_key_values)