ipex-llm 2.2.0b20250108__py3-none-manylinux2010_x86_64.whl → 2.2.0b20250109__py3-none-manylinux2010_x86_64.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)

ipex_llm/transformers/models/baichuan.py

@@ -47,38 +47,6 @@ 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 +73,9 @@ 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 +248,6 @@ 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 +258,8 @@ 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 +310,8 @@ 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
@@ -171,7 +172,8 @@ 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 +348,8 @@ 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

@@ -249,7 +249,9 @@ 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 +307,9 @@ 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)

ipex_llm/transformers/models/qwen.py

@@ -107,7 +107,8 @@ def qwen_attention_forward(
         query_states = query_states * logn_tensor.type_as(query_states).expand_as(query_states)
 
     # IPEX-LLM OPT: kv cache and quantzie kv cache
-    use_quantize_kv = use_quantize_kv_cache(self.c_attn, hidden_states)
+    use_quantize_kv = use_quantize_kv_cache(self.c_attn, 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
@@ -205,7 +206,8 @@ def qwen_attention_forward_registered(
         query_states = query_states * logn_tensor.type_as(query_states).expand_as(query_states)
 
     # IPEX-LLM OPT: kv cache and quantzie kv cache
-    use_quantize_kv = use_quantize_kv_cache(self.c_attn, hidden_states)
+    use_quantize_kv = use_quantize_kv_cache(self.c_attn, 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/qwen2.py

@@ -113,10 +113,10 @@ def qwen2_model_forward(
     # ipex-llm changes start
     # IPEX-LLM OPT: kv cache and quantize kv cache
     inputs = input_ids if input_ids is not None else inputs_embeds
+    num_heads, num_kv_heads = self.config.num_attention_heads, self.config.num_key_value_heads
     use_quantize_kv = (
         self.config.hidden_size != 3584     # disable quantize kv in specific model
-        and use_quantize_kv_cache(self.layers[0].mlp.up_proj, inputs,
-                                  self.config.num_attention_heads//self.config.num_key_value_heads)
+        and use_quantize_kv_cache(self.layers[0].mlp.up_proj, inputs, num_heads, num_kv_heads)
     )
     use_compress_kv = should_use_compresskv(inputs, inputs.shape[1]) or \
         isinstance(past_key_values, DynamicCompressCache)
@@ -305,10 +305,11 @@ def qwen2_model_forward_4_42(
 
     # ipex-llm changes start
     # IPEX-LLM OPT: kv cache and quantize kv cache
+    num_heads, num_kv_heads = self.config.num_attention_heads, self.config.num_key_value_heads
     use_quantize_kv = (
         self.config.hidden_size != 3584     # disable quantize kv in specific model
         and use_quantize_kv_cache(self.layers[0].mlp.up_proj, inputs_embeds,
-                                  self.config.num_attention_heads//self.config.num_key_value_heads)
+                                  num_heads, num_kv_heads)
     )
     use_compress_kv = should_use_compresskv(inputs_embeds, inputs_embeds.shape[1]) or \
         isinstance(past_key_values, DynamicCompressCache)

ipex_llm/transformers/models/qwen2_moe.py

@@ -73,8 +73,10 @@ def qwen2moe_model_forward(
     return_dict: Optional[bool] = None,
 ):
     use_cache = use_cache if use_cache is not None else self.config.use_cache
-    input = input_ids if input_ids is not None else inputs_embeds
-    use_quantize_kv = use_quantize_kv_cache(self.layers[0].mlp.shared_expert.up_proj, input)
+    inputs = input_ids if input_ids is not None else inputs_embeds
+    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.shared_expert.up_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/qwen2_vl.py

@@ -88,7 +88,9 @@ def qwen2_vl_model_forward(
     # IPEX-LLM OPT start: kv cache and quantize kv cache
     inputs = input_ids if input_ids is not None else inputs_embeds
     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/stablelm.py

@@ -69,8 +69,10 @@ def stablelm_model_forward(
 ):
     # IPEX-LLM OPT: kv cache and quantize kv cache
     use_cache = use_cache if use_cache is not None else self.config.use_cache
+    num_heads, num_kv_heads = self.config.num_attention_heads, self.config.num_key_value_heads
     use_quantize_kv = (self.layers[0].self_attn.head_dim in [64, 80, 96, 128]
-                       and use_quantize_kv_cache(self.layers[0].mlp.up_proj, input_ids))
+                       and use_quantize_kv_cache(self.layers[0].mlp.up_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)

ipex_llm/transformers/models/starcoder2.py

@@ -132,7 +132,9 @@ def model_forward(
     return_dict: Optional[bool] = None,
 ):
     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.c_fc, 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.c_fc, 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)

ipex_llm/transformers/models/utils.py

@@ -74,7 +74,8 @@ def append_kv_cache(cache_k, cache_v, key_states, value_states):
     return new_cache_k, new_cache_v
 
 
-def use_quantize_kv_cache(linear: torch.nn.Module, x: torch.Tensor, kv_group: int = 1) -> bool:
+def use_quantize_kv_cache(linear: torch.nn.Module, x: torch.Tensor,
+                          num_heads: int, num_kv_heads: int) -> bool:
     if os.environ.get("BIGDL_QUANTIZE_KV_CACHE", None) is not None:
         warnings.warn(
             "`BIGDL_QUANTIZE_KV_CACHE` is deprecated and will be removed in future releases. "
@@ -90,8 +91,11 @@ def use_quantize_kv_cache(linear: torch.nn.Module, x: torch.Tensor, kv_group: in
     else:
         device_name = get_xpu_device_name(x.device)
         return (
-            device_name in ["mtl", "lnl", "arl"] and kv_group == 1
-            or device_name in ["arc", "bmg"] and x.size(0) > 1
+            num_kv_heads >= 4
+            and (
+                device_name in ["mtl", "lnl", "arl"] and num_heads // num_kv_heads <= 4
+                or device_name in ["arc", "bmg"] and x.size(0) > 1
+            )
         )
 
 

ipex_llm/transformers/models/yuan.py

@@ -158,7 +158,8 @@ def yuan_attention_forward(
                                                         "yuan")
 
     # IPEX-LLM OPT: kv cache and quantzie kv cache
-    use_quantize_kv = use_quantize_kv_cache(self.qk_proj, hidden_states)
+    use_quantize_kv = use_quantize_kv_cache(self.qk_proj, hidden_states,
+                                            self.num_heads, self.num_heads)
     key_states, value_states = update_past_key_value(
         None if past_key_value is None else (past_key_value[0], past_key_value[1]),
         key_states, value_states,

{ipex_llm-2.2.0b20250108.dist-info → ipex_llm-2.2.0b20250109.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: ipex-llm
-Version: 2.2.0b20250108
+Version: 2.2.0b20250109
 Summary: Large Language Model Develop Toolkit
 Home-page: https://github.com/intel-analytics/ipex-llm
 Author: BigDL Authors
@@ -27,10 +27,10 @@ Requires-Dist: intel-openmp ; (platform_machine == "x86_64" or platform_machine
 Requires-Dist: torch ==2.1.2+cpu ; (platform_system == "Linux") and extra == 'all'
 Requires-Dist: torch ==2.1.2 ; (platform_system == "Windows") and extra == 'all'
 Provides-Extra: cpp
-Requires-Dist: bigdl-core-cpp ==2.6.0b20250108 ; extra == 'cpp'
+Requires-Dist: bigdl-core-cpp ==2.6.0b20250109 ; extra == 'cpp'
 Requires-Dist: setuptools ; extra == 'cpp'
 Provides-Extra: cpp-arl
-Requires-Dist: bigdl-core-cpp ==2.6.0b20250108 ; extra == 'cpp-arl'
+Requires-Dist: bigdl-core-cpp ==2.6.0b20250109 ; extra == 'cpp-arl'
 Requires-Dist: setuptools ; extra == 'cpp-arl'
 Requires-Dist: onednn-devel ==2024.1.1 ; (platform_system == "Windows") and extra == 'cpp-arl'
 Requires-Dist: onednn ==2024.1.1 ; (platform_system == "Windows") and extra == 'cpp-arl'
@@ -67,7 +67,7 @@ Requires-Dist: transformers ==4.40.0 ; extra == 'npu'
 Requires-Dist: intel-openmp ; (platform_machine == "x86_64" or platform_machine == "AMD64") and extra == 'npu'
 Requires-Dist: torch ==2.1.2+cpu ; (platform_system == "Linux") and extra == 'npu'
 Requires-Dist: torch ==2.1.2 ; (platform_system == "Windows") and extra == 'npu'
-Requires-Dist: bigdl-core-npu ==2.6.0b20250108 ; (platform_system == "Windows") and extra == 'npu'
+Requires-Dist: bigdl-core-npu ==2.6.0b20250109 ; (platform_system == "Windows") and extra == 'npu'
 Provides-Extra: serving
 Requires-Dist: py-cpuinfo ; extra == 'serving'
 Requires-Dist: fschat[model_worker,webui] ==0.2.36 ; extra == 'serving'
@@ -87,9 +87,9 @@ Requires-Dist: setuptools <70.0.0 ; extra == 'xpu'
 Requires-Dist: torch ==2.1.0a0 ; extra == 'xpu'
 Requires-Dist: torchvision ==0.16.0a0 ; extra == 'xpu'
 Requires-Dist: intel-extension-for-pytorch ==2.1.10+xpu ; extra == 'xpu'
-Requires-Dist: bigdl-core-xe-21 ==2.6.0b20250108 ; extra == 'xpu'
-Requires-Dist: bigdl-core-xe-batch-21 ==2.6.0b20250108 ; extra == 'xpu'
-Requires-Dist: bigdl-core-xe-addons-21 ==2.6.0b20250108 ; extra == 'xpu'
+Requires-Dist: bigdl-core-xe-21 ==2.6.0b20250109 ; extra == 'xpu'
+Requires-Dist: bigdl-core-xe-batch-21 ==2.6.0b20250109 ; extra == 'xpu'
+Requires-Dist: bigdl-core-xe-addons-21 ==2.6.0b20250109 ; extra == 'xpu'
 Provides-Extra: xpu-2-1
 Requires-Dist: py-cpuinfo ; extra == 'xpu-2-1'
 Requires-Dist: protobuf ; extra == 'xpu-2-1'
@@ -104,9 +104,9 @@ Requires-Dist: setuptools <70.0.0 ; extra == 'xpu-2-1'
 Requires-Dist: torch ==2.1.0a0 ; extra == 'xpu-2-1'
 Requires-Dist: torchvision ==0.16.0a0 ; extra == 'xpu-2-1'
 Requires-Dist: intel-extension-for-pytorch ==2.1.10+xpu ; extra == 'xpu-2-1'
-Requires-Dist: bigdl-core-xe-21 ==2.6.0b20250108 ; extra == 'xpu-2-1'
-Requires-Dist: bigdl-core-xe-batch-21 ==2.6.0b20250108 ; extra == 'xpu-2-1'
-Requires-Dist: bigdl-core-xe-addons-21 ==2.6.0b20250108 ; extra == 'xpu-2-1'
+Requires-Dist: bigdl-core-xe-21 ==2.6.0b20250109 ; extra == 'xpu-2-1'
+Requires-Dist: bigdl-core-xe-batch-21 ==2.6.0b20250109 ; extra == 'xpu-2-1'
+Requires-Dist: bigdl-core-xe-addons-21 ==2.6.0b20250109 ; extra == 'xpu-2-1'
 Requires-Dist: intel-openmp ; (platform_machine == "x86_64" or platform_machine == "AMD64") and extra == 'xpu-2-1'
 Requires-Dist: dpcpp-cpp-rt ==2024.0.2 ; (platform_system == "Windows") and extra == 'xpu-2-1'
 Requires-Dist: mkl-dpcpp ==2024.0.0 ; (platform_system == "Windows") and extra == 'xpu-2-1'
@@ -124,7 +124,7 @@ Requires-Dist: setuptools ; extra == 'xpu-2-6'
 Requires-Dist: torch ==2.6.0+xpu ; extra == 'xpu-2-6'
 Requires-Dist: torchvision ==0.21.0+xpu ; extra == 'xpu-2-6'
 Requires-Dist: torchaudio ==2.6.0+xpu ; extra == 'xpu-2-6'
-Requires-Dist: bigdl-core-xe-all ==2.6.0b20250108 ; extra == 'xpu-2-6'
+Requires-Dist: bigdl-core-xe-all ==2.6.0b20250109 ; extra == 'xpu-2-6'
 Requires-Dist: intel-openmp ; (platform_machine == "x86_64" or platform_machine == "AMD64") and extra == 'xpu-2-6'
 Provides-Extra: xpu-arc
 Requires-Dist: py-cpuinfo ; extra == 'xpu-arc'
@@ -137,9 +137,9 @@ Requires-Dist: tokenizers ==0.15.2 ; extra == 'xpu-arc'
 Requires-Dist: accelerate ==0.23.0 ; extra == 'xpu-arc'
 Requires-Dist: tabulate ; extra == 'xpu-arc'
 Requires-Dist: setuptools ; extra == 'xpu-arc'
-Requires-Dist: bigdl-core-xe-23 ==2.6.0b20250108 ; extra == 'xpu-arc'
-Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250108 ; extra == 'xpu-arc'
-Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250108 ; extra == 'xpu-arc'
+Requires-Dist: bigdl-core-xe-23 ==2.6.0b20250109 ; extra == 'xpu-arc'
+Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250109 ; extra == 'xpu-arc'
+Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250109 ; extra == 'xpu-arc'
 Requires-Dist: intel-openmp ; (platform_machine == "x86_64" or platform_machine == "AMD64") and extra == 'xpu-arc'
 Requires-Dist: torch ==2.3.1+cxx11.abi ; (platform_system == "Linux") and extra == 'xpu-arc'
 Requires-Dist: torchvision ==0.18.1+cxx11.abi ; (platform_system == "Linux") and extra == 'xpu-arc'
@@ -160,9 +160,9 @@ Requires-Dist: tokenizers ==0.15.2 ; extra == 'xpu-arl'
 Requires-Dist: accelerate ==0.23.0 ; extra == 'xpu-arl'
 Requires-Dist: tabulate ; extra == 'xpu-arl'
 Requires-Dist: setuptools ; extra == 'xpu-arl'
-Requires-Dist: bigdl-core-xe-23 ==2.6.0b20250108 ; extra == 'xpu-arl'
-Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250108 ; extra == 'xpu-arl'
-Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250108 ; extra == 'xpu-arl'
+Requires-Dist: bigdl-core-xe-23 ==2.6.0b20250109 ; extra == 'xpu-arl'
+Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250109 ; extra == 'xpu-arl'
+Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250109 ; extra == 'xpu-arl'
 Requires-Dist: intel-openmp ; (platform_machine == "x86_64" or platform_machine == "AMD64") and extra == 'xpu-arl'
 Requires-Dist: torch ==2.3.1+cxx11.abi ; (platform_system == "Linux") and extra == 'xpu-arl'
 Requires-Dist: torchvision ==0.18.1+cxx11.abi ; (platform_system == "Linux") and extra == 'xpu-arl'
@@ -183,9 +183,9 @@ Requires-Dist: tokenizers ==0.15.2 ; extra == 'xpu-lnl'
 Requires-Dist: accelerate ==0.23.0 ; extra == 'xpu-lnl'
 Requires-Dist: tabulate ; extra == 'xpu-lnl'
 Requires-Dist: setuptools ; extra == 'xpu-lnl'
-Requires-Dist: bigdl-core-xe-23 ==2.6.0b20250108 ; extra == 'xpu-lnl'
-Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250108 ; extra == 'xpu-lnl'
-Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250108 ; extra == 'xpu-lnl'
+Requires-Dist: bigdl-core-xe-23 ==2.6.0b20250109 ; extra == 'xpu-lnl'
+Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250109 ; extra == 'xpu-lnl'
+Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250109 ; extra == 'xpu-lnl'
 Requires-Dist: intel-openmp ; (platform_machine == "x86_64" or platform_machine == "AMD64") and extra == 'xpu-lnl'
 Requires-Dist: torch ==2.3.1+cxx11.abi ; (platform_system == "Linux") and extra == 'xpu-lnl'
 Requires-Dist: torchvision ==0.18.1+cxx11.abi ; (platform_system == "Linux") and extra == 'xpu-lnl'

{ipex_llm-2.2.0b20250108.dist-info → ipex_llm-2.2.0b20250109.dist-info}/RECORD

@@ -94,7 +94,7 @@ ipex_llm/serving/fastchat/tgi_api_protocol.py,sha256=brT3k3-V0NJrU4fRqUwWjC0O3iO
 ipex_llm/serving/fastchat/tgi_api_server.py,sha256=agNTAEiZPSuj3dEdIdYKwkoY0cXOUDX06DiM9VP2knQ,24418
 ipex_llm/serving/fastchat/vllm_worker.py,sha256=ZLz2Q9GxJO6r_LOiP6epgCRjBGk-K4EB1SNEWSJp5DA,11091
 ipex_llm/transformers/__init__.py,sha256=l4KkMkLe-pRC7b_kj6LCfeifgE-Uo33_Av_FwN9HnFA,1074
-ipex_llm/transformers/convert.py,sha256=APf2uHMgEeiAhsKm9dPgPWlyO0ADq2yHtZgovv9oczU,99101
+ipex_llm/transformers/convert.py,sha256=umI137wqV2d4itS0AJQoZcygeWBATpSJSDJ805cZ-SY,98499
 ipex_llm/transformers/convert_ipex.py,sha256=iKXo0n8fVFTOA2fNYYrByMFK0dovL-kLd2sVDk88AlQ,14334
 ipex_llm/transformers/embedding.py,sha256=bdgk59DvD4ZZyxRzewXOR7g56nThgO6uhIwk8QL7f-s,9299
 ipex_llm/transformers/kv.py,sha256=k4TU18LlA-Sbq9WNNQnfuzu3RSFBwFhmaV3BcGN5bAo,19191
@@ -144,45 +144,45 @@ ipex_llm/transformers/gguf/models/model_implement/yuan2/configuration_yuan.py,sh
 ipex_llm/transformers/gguf/models/model_implement/yuan2/yuan_hf_model.py,sha256=_AOGMV65XHxgTxIib7lgs49InopcecTzRwgtYR8NTUg,51084
 ipex_llm/transformers/models/__init__.py,sha256=tp2DcVkKg1-QvdYk7DY7rZvQWCDQ4ZjU8NAQ7Fclrpg,584
 ipex_llm/transformers/models/aquila.py,sha256=VZb5Drpo_fTxwcExZ397LygnsNPX2sVbie9_JeFudZI,5252
-ipex_llm/transformers/models/baichuan.py,sha256=oJCAEENSG8oQhJ-QPN2SiapARjAGdOM6nEbyCcYOMCo,19334
-ipex_llm/transformers/models/bert.py,sha256=bJNic2pt1kph0kBwdK5MRGyWupFfx2Ts0V3D1L-5kWo,6085
+ipex_llm/transformers/models/baichuan.py,sha256=cAQLmVG-3R8CSTGTcDy2JOOzVe-Ej8AXjIEIjvZBGlo,18376
+ipex_llm/transformers/models/bert.py,sha256=0Mm9jkvkzBxtc_z_GE1TcZoPz-HOg2Z2973ZEWgSwJk,5601
 ipex_llm/transformers/models/bloom.py,sha256=PxfzyYT-nFn3K5rZhTQjmcEjUUzAhUFzxIN4kzRlCuc,8103
 ipex_llm/transformers/models/chatglm.py,sha256=UHai1t2AUtGmF765_eHF8LUMVQzp_oCBx8TJB21WrHk,12597
-ipex_llm/transformers/models/chatglm2.py,sha256=SGCABJdYQLW0zDarEoWrEQLuWlbq9iQhYU8ZeR1-ptQ,15957
-ipex_llm/transformers/models/chatglm4.py,sha256=AAhAFFDDas5DBQPfh2Mwl7a2v7taKf6xphoeeNNFaBI,16593
-ipex_llm/transformers/models/chatglm4v.py,sha256=tyjDDyF6FEgLAT24EG3i4-auxZvkwmeLIy0Hds4K5Yo,14105
-ipex_llm/transformers/models/common.py,sha256=4obQMGF02FCiXrHnFle9Fsx7C33b1FDt37qJJ4YgxRc,11578
+ipex_llm/transformers/models/chatglm2.py,sha256=KyAIX7zGVQDQuwwM3QMBNWZbTeMHEzKUIgAryT0voHc,14933
+ipex_llm/transformers/models/chatglm4.py,sha256=QvUehdaCePB3MNHyWg3dneDxmjtBdxYeKUyQUVcsgfM,16886
+ipex_llm/transformers/models/chatglm4v.py,sha256=L6y45M_wjS2_HqchmCUxRlQZUNuSNCGOiynAQrGh918,14124
+ipex_llm/transformers/models/common.py,sha256=Q3IEfGqvxoHyfIIF5s8qHmOJBBP3b2jyVAXk8C3b1Pg,11636
 ipex_llm/transformers/models/decilm.py,sha256=P-PBuDPf07GvKggLwJx_wPwIn6esN3rX8ai2JxRuZmE,5246
 ipex_llm/transformers/models/gemma.py,sha256=_E3Yw8Y45xyNVeLqyVKcpr8kjuICtETeL82cJ-bWJuU,9424
 ipex_llm/transformers/models/gemma2.py,sha256=2WZuv-FLzJyTJFaYxOuzJt47QE64M0lHnzAiO5T6ozI,8049
-ipex_llm/transformers/models/glm.py,sha256=PE43uks9lojndBBHFVXK1VWisHhbY-kuCmhq0CwmD4s,7204
+ipex_llm/transformers/models/glm.py,sha256=lmeEWd_W2O638VzVW4Gm6cJre5XZcg_QBmPs8NWqXsM,7202
 ipex_llm/transformers/models/gpt2.py,sha256=YSaNgK1uLCFDuIFqnKO0Mi-AsOZsYav-7pNf_NpKGdM,3445
 ipex_llm/transformers/models/gptbigcode.py,sha256=cP1_qGWoa43R2WacAMblShjku4QupcCZiLaPPAoOUs4,9101
 ipex_llm/transformers/models/gptneox.py,sha256=loRh1x_5S6BCeOr_s5xr-N_1SQHL3Y5IiUBAEyoMUqQ,6172
-ipex_llm/transformers/models/internlm.py,sha256=ZbIUMDwNRcrCeduXfbA_uq1AUEWawEt6CJRvQl3LkAg,17832
+ipex_llm/transformers/models/internlm.py,sha256=OifyiobRligleyZLpLBSe44A6Sq0uMG-8-NOcRCcT4Q,18080
 ipex_llm/transformers/models/internvl.py,sha256=Vx0vENIEQLX2M6P398mw5TOhpks0U8xf8rtRQvy94go,8154
-ipex_llm/transformers/models/llama.py,sha256=n1JG1uElMB8t3Hpae94S6YTO_5q2N5BUAhb7mncvA6E,8560
-ipex_llm/transformers/models/minicpm.py,sha256=_eYBYafQxnuqKo9ENNkua73KU5goU2z-dkaLlF5uHnA,10147
-ipex_llm/transformers/models/minicpm3.py,sha256=FhNS6mi2rg7dSdF_QQGrao3g9EC6XLn1MTKd-kd0wF0,9191
+ipex_llm/transformers/models/llama.py,sha256=NzpyQve_RC9ez1W-jWPLGZ80k_S1I5Rx5saAzCsDIoI,8558
+ipex_llm/transformers/models/minicpm.py,sha256=eaPNVNrep0_xGoELhZd886ff0ceoKqB6cusdAhd52eE,10145
+ipex_llm/transformers/models/minicpm3.py,sha256=11cYl8KM2hoIJNMAOZMxiwCu6dMhup9ric_OEn8-VrQ,9363
 ipex_llm/transformers/models/minicpmv.py,sha256=PP05b5iTnrMpiseCn8iJcxKJDnfq7WqXp9Mrch0kKZ0,9876
-ipex_llm/transformers/models/mistral.py,sha256=rE1GWQxXvF6aG-buPHDR13zeynDZEDIubPF4PiVhZbM,7451
-ipex_llm/transformers/models/mllama.py,sha256=ogpLmmN_OwcFUyjYB-oDC-l3uw8urFvUEc5edkjWHAk,10939
+ipex_llm/transformers/models/mistral.py,sha256=uVhkdXaq15v1P3QY0emVsA7SxUbAWChHEEXYN-drjpQ,7449
+ipex_llm/transformers/models/mllama.py,sha256=ZyRq9DTKsvk1AlRbr-z6ngjS3Sr_7YuGZ6-Yr1MBBAM,10937
 ipex_llm/transformers/models/mpt.py,sha256=z02NwHogJZVh-Mk4sYoIzR90SFIKhoNN_-ifsD907TQ,9540
 ipex_llm/transformers/models/phi.py,sha256=E6qz4EEuHIVGvaPo-wtLC5lz3iyMqTbAE_cRlcjQRKI,6670
-ipex_llm/transformers/models/phi3.py,sha256=jkiadJ85ToHpymY5GOM6orWlnx6LKN8_-v1MUcfGWPg,15159
+ipex_llm/transformers/models/phi3.py,sha256=Fo6PlZ24Gdm7eeeZOTMm1Bfh3U6P4rvq7-_2FHvp0vE,15503
 ipex_llm/transformers/models/phixtral.py,sha256=MDTMghcu7qAmZmRcUGqXXDXhSU3y_N59HRIXmlcjp5g,4890
-ipex_llm/transformers/models/qwen.py,sha256=XIJ_bLzediBURWU-OOS3H6WBIGXQue6jDdUHJsAabwY,19391
-ipex_llm/transformers/models/qwen2.py,sha256=b49HO4GSudwGJ3n6uHVno1oo3DgRt3jOjtQnLOB3cdY,25530
-ipex_llm/transformers/models/qwen2_moe.py,sha256=EA_OYxYAEgrvi7VpDW192AJXG9Fwe2aBtOAZPkOAJk4,19350
-ipex_llm/transformers/models/qwen2_vl.py,sha256=jIm4yZSd751BkRqgj3wR1QBkDIh-TMCLAMM8SZ8n6Qo,13419
+ipex_llm/transformers/models/qwen.py,sha256=A3WiVCzA7NLkcjp4zhFkZvKZzZWZlg0WFuVV_556TAI,19543
+ipex_llm/transformers/models/qwen2.py,sha256=JLaY9ZT7A22oO0G8K-nvjvKQDaIrKA5o-jEHvk_y3eI,25604
+ipex_llm/transformers/models/qwen2_moe.py,sha256=a0gYo-ngf8SxaEnBdZUJDnPS6Mkn_poDd8xqhx50icI,19516
+ipex_llm/transformers/models/qwen2_vl.py,sha256=NrhxlaPj7W-HUBmKc3CSTwZy1lkoZ9qDaxM4GvE0kHs,13583
 ipex_llm/transformers/models/qwen_vl.py,sha256=j7Nzzz2Qvynu9yrCXmoEfERjw43hXof5TbXIs7Ms-oY,17105
 ipex_llm/transformers/models/rwkv4.py,sha256=H4KMtxN0JA2ZTXnonHpsUUJ5xULemo-D1Jzl0ri_UY8,6123
 ipex_llm/transformers/models/rwkv5.py,sha256=OkRNj1pCAZg1z2Fw-I0DEnxLEdZyPeRSQ6msrkxLOCs,10710
 ipex_llm/transformers/models/sd.py,sha256=VvHV5u-0k2MgHu3NL9113hPj7DgfxqctuKzEEeNfRDU,5981
-ipex_llm/transformers/models/stablelm.py,sha256=RGQCYuQhYqtZ1j3RZkYi0_QvCRnUgUIPYxfBcLnElzg,6885
-ipex_llm/transformers/models/starcoder2.py,sha256=4P3mhRYf2Kreb1ESjrQGfy1puLMmZXgV35zf-Tksvao,6462
-ipex_llm/transformers/models/utils.py,sha256=isBCMMQP3j_opmda9XzD_dPk1ejvEXTztggbu1yIMSc,15439
-ipex_llm/transformers/models/yuan.py,sha256=1jRPebwAK2ENbyYokOmb4LSVo-szucWiygz9zTv-scs,7656
+ipex_llm/transformers/models/stablelm.py,sha256=fj-XtOnR6kggnFUQTMPCOOzolkPztN06WAv8QW-XRnI,7054
+ipex_llm/transformers/models/starcoder2.py,sha256=ONKvD7JCkRM0DI-R56x28QFBJ7CjD5hOZBQ_3WfOcNk,6626
+ipex_llm/transformers/models/utils.py,sha256=ihbWS5kQK2KHDVPkMhgjik3nM8B2fWf-E-z4BWNUstk,15568
+ipex_llm/transformers/models/yuan.py,sha256=JYAn_ZaSGK0NBJLEIxCACfAq084a66GFJkdd5NbpmMA,7732
 ipex_llm/transformers/npu_models/__init__.py,sha256=ulEUGLjaP48LCrVeury3UxLjXxKzRi0UpSG4bYu-7f8,585
 ipex_llm/transformers/npu_models/baichuan.py,sha256=fJtd7fBrttySghRUgfZTAdxLjsSNC-XL08HISsXigLE,4685
 ipex_llm/transformers/npu_models/baichuan_mp.py,sha256=tHhO-0v5z6IhxsfzAPYWXVbLrV_4z89DIb4JjE3207M,45026
@@ -250,11 +250,11 @@ ipex_llm/vllm/xpu/engine/__init__.py,sha256=pY_CpyuZd72fr6s32ejeKHKFW0K4vUU2rzZj
 ipex_llm/vllm/xpu/engine/engine.py,sha256=k4-D27WS_Gk3mA--w3HWAjPjb4Aiu043MVPi0ZoAUBc,5984
 ipex_llm/vllm/xpu/entrypoints/openai/api_server.py,sha256=GshTZFB8e4PWvqckfbmTOU6b0oLkNn7A-vzLuG9--j8,21544
 ipex_llm/vllm/xpu/entrypoints/openai/cli_args.py,sha256=2rENA2ucynMaIjiZBEh2ez1o5vR32GaP514t39CD7KM,8676
-ipex_llm-2.2.0b20250108.data/scripts/ipex-llm-init,sha256=fLQsT2dRL6H5bThb4GuIWotAuqoLsIxFwA-0c2qmaO8,6672
-ipex_llm-2.2.0b20250108.data/scripts/llm-chat,sha256=TdUnUmNapzuoe1c8IzrdVOQwWEg8IqsMSBRlOD3daZM,2249
-ipex_llm-2.2.0b20250108.data/scripts/llm-cli,sha256=RXGPlLElHxcKzoUxljEMBIAXbzCDysXL-Nxw-xF-7LU,2457
-ipex_llm-2.2.0b20250108.dist-info/METADATA,sha256=NJp_uuPOJe8b5UQ8ASJbfzen2BGoc2DEM1ZInzr0X9E,12705
-ipex_llm-2.2.0b20250108.dist-info/WHEEL,sha256=PPJcBMAZibF_2GFE9NmOJGqiaSMPiNFbJd6QaJjdA6Y,109
-ipex_llm-2.2.0b20250108.dist-info/entry_points.txt,sha256=TiUyBB2MRmfF3ko-pyAEzqeBCRnyhu27bNOAsWPp3e8,61
-ipex_llm-2.2.0b20250108.dist-info/top_level.txt,sha256=CGCMHM-SyqUabU4h8RqJ2KTYckQUO3LvIWwmUQ6Qbzw,9
-ipex_llm-2.2.0b20250108.dist-info/RECORD,,
+ipex_llm-2.2.0b20250109.data/scripts/ipex-llm-init,sha256=fLQsT2dRL6H5bThb4GuIWotAuqoLsIxFwA-0c2qmaO8,6672
+ipex_llm-2.2.0b20250109.data/scripts/llm-chat,sha256=TdUnUmNapzuoe1c8IzrdVOQwWEg8IqsMSBRlOD3daZM,2249
+ipex_llm-2.2.0b20250109.data/scripts/llm-cli,sha256=RXGPlLElHxcKzoUxljEMBIAXbzCDysXL-Nxw-xF-7LU,2457
+ipex_llm-2.2.0b20250109.dist-info/METADATA,sha256=gPslIWSw_X5E5ULhQa8rOHeRo_UeBDXCAyPjBSPB-nU,12705
+ipex_llm-2.2.0b20250109.dist-info/WHEEL,sha256=PPJcBMAZibF_2GFE9NmOJGqiaSMPiNFbJd6QaJjdA6Y,109
+ipex_llm-2.2.0b20250109.dist-info/entry_points.txt,sha256=TiUyBB2MRmfF3ko-pyAEzqeBCRnyhu27bNOAsWPp3e8,61
+ipex_llm-2.2.0b20250109.dist-info/top_level.txt,sha256=CGCMHM-SyqUabU4h8RqJ2KTYckQUO3LvIWwmUQ6Qbzw,9
+ipex_llm-2.2.0b20250109.dist-info/RECORD,,