LLM: fix first token judgement of flash attention (#9841)

* fix flash attention

* meet code review

* fix

python/llm/src/bigdl/llm/transformers/models/chatglm2.py

@@ -367,7 +367,7 @@ def core_attn_forward_8eb45c(self, query_layer, key_layer, value_layer, attentio
     if pytorch_major_version >= 2 and (query_layer.device.type == 'xpu' or query_layer.size(0) > 1):
         query_layer = query_layer.permute(1, 2, 0, 3)
         L, S = query_layer.shape[2], key_layer.shape[2]
-        if attention_mask is None and (use_flash_attention(query_layer) or
+        if attention_mask is None and (use_flash_attention(query_layer, key_layer) or
                                        L == S and query_layer.device.type == "cpu"):
             context_layer = torch.nn.functional.scaled_dot_product_attention(query_layer,
                                                                              key_layer,

python/llm/src/bigdl/llm/transformers/models/llama.py

@@ -139,14 +139,6 @@ def llama_attention_forward_4_31(
     device = hidden_states.device
     # for flash attention
     original_dtype = hidden_states.dtype
-    if not self.training and not hidden_states.requires_grad:
-        fsdp_flag = use_flash_attention(hidden_states)
-    else:
-        fsdp_flag = False
-    if fsdp_flag:
-        attention_dtype = torch.float16  # use fp16 for flash attention
-    else:
-        attention_dtype = original_dtype
 
     use_fuse_rope = should_use_fuse_rope(self, hidden_states, position_ids)
     enough_kv_room = is_enough_kv_cache_room_4_31(past_key_value, seq_len=q_len)
@@ -259,6 +251,15 @@ def llama_attention_forward_4_31(
 
     past_key_value = (key_states, value_states) if use_cache else None
 
+    if not self.training and not hidden_states.requires_grad:
+        fsdp_flag = use_flash_attention(query_states, key_states)
+    else:
+        fsdp_flag = False
+    if fsdp_flag:
+        attention_dtype = torch.float16  # use fp16 for flash attention
+    else:
+        attention_dtype = original_dtype
+
     # repeat k/v heads if n_kv_heads < n_heads
     key_states = repeat_kv(key_states, self.num_key_value_groups).to(device,
                                                                      dtype=attention_dtype)

python/llm/src/bigdl/llm/transformers/models/utils.py

@@ -123,18 +123,24 @@ def is_enough_kv_cache_room_4_31(past_key_value, seq_len=1):
         (past_key_value[0].size(2) + seq_len - 1) * past_key_value[0].size(3)
 
 
-def use_flash_attention(query):
+def use_flash_attention(query, key):
-    if query.dim() == 3:
+    # here we support query's shape is always [batch_size, head_num, q_len, head_dim],
-        bsz, q_len, _ = query.size()
+    # key's shape is always [batch_size, head_num, k_len, head_dim]
-    elif query.dim() == 4:
+    invalidInputError(query.dim() == 4,
-        bsz, _, q_len, _ = query.size()
+                      "Here query input of use_flash_attention should be [batch_size, "
+                      "head_num, q_len, head_dim]")
+    invalidInputError(key.dim() == 4,
+                      "Here key input of use_flash_attention should be [batch_size, "
+                      "head_num, k_len, head_dim]")
+    bsz, _, q_len, _ = query.size()
+    k_len = key.size()[2]
     # check whether ipex flash attention can be used
     if bsz > 1:
         # only use flash attention for batch_size = 1 now
         # as flash attention doesn't support attn_mask in ipex 2.1,
         # so it will cause output error for padded batch input
         return False
-    if q_len == 1:
+    if q_len != k_len:
         # now only use flash attention for first token
         # as it seems have no performance benifit for rest token now
         return False