intel-analytics · MeouSker77 · Jan 9, 2025 · Jan 9, 2025
diff --git a/python/llm/src/ipex_llm/transformers/models/baichuan.py b/python/llm/src/ipex_llm/transformers/models/baichuan.py
@@ -105,7 +105,9 @@ 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 +280,6 @@ 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 +290,8 @@ 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 +342,8 @@ 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

diff --git a/python/llm/src/ipex_llm/transformers/models/chatglm2.py b/python/llm/src/ipex_llm/transformers/models/chatglm2.py
@@ -91,8 +91,13 @@ 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 +290,6 @@ 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 +303,8 @@ 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

diff --git a/python/llm/src/ipex_llm/transformers/models/chatglm4.py b/python/llm/src/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

diff --git a/python/llm/src/ipex_llm/transformers/models/chatglm4v.py b/python/llm/src/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

diff --git a/python/llm/src/ipex_llm/transformers/models/glm.py b/python/llm/src/ipex_llm/transformers/models/glm.py
@@ -147,7 +147,7 @@ def glm_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)
 
         if use_cache:

diff --git a/python/llm/src/ipex_llm/transformers/models/internlm.py b/python/llm/src/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

diff --git a/python/llm/src/ipex_llm/transformers/models/llama.py b/python/llm/src/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)

diff --git a/python/llm/src/ipex_llm/transformers/models/minicpm.py b/python/llm/src/ipex_llm/transformers/models/minicpm.py
@@ -159,7 +159,7 @@ def minicpm_model_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)

diff --git a/python/llm/src/ipex_llm/transformers/models/minicpm3.py b/python/llm/src/ipex_llm/transformers/models/minicpm3.py
@@ -66,7 +66,9 @@ def minicpm3_model_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)

diff --git a/python/llm/src/ipex_llm/transformers/models/mistral.py b/python/llm/src/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)

diff --git a/python/llm/src/ipex_llm/transformers/models/mllama.py b/python/llm/src/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):

diff --git a/python/llm/src/ipex_llm/transformers/models/phi3.py b/python/llm/src/ipex_llm/transformers/models/phi3.py
@@ -249,7 +249,9 @@ def 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 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)

diff --git a/python/llm/src/ipex_llm/transformers/models/qwen.py b/python/llm/src/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

diff --git a/python/llm/src/ipex_llm/transformers/models/qwen2.py b/python/llm/src/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)

diff --git a/python/llm/src/ipex_llm/transformers/models/qwen2_moe.py b/python/llm/src/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)

diff --git a/python/llm/src/ipex_llm/transformers/models/qwen2_vl.py b/python/llm/src/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)

diff --git a/python/llm/src/ipex_llm/transformers/models/stablelm.py b/python/llm/src/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)

diff --git a/python/llm/src/ipex_llm/transformers/models/starcoder2.py b/python/llm/src/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)

diff --git a/python/llm/src/ipex_llm/transformers/models/utils.py b/python/llm/src/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
+            )
         )
 
 

diff --git a/python/llm/src/ipex_llm/transformers/models/yuan.py b/python/llm/src/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,