use Qwen2SdpaAttention to replace Qwen2Attention for speed up

README.md

@@ -2,6 +2,7 @@
 - 本项目参考了[ascend-llm](https://gitee.com/yinghuo302/ascend-llm)项目。
 - 仅在昇腾310B1做了测试，理论上也兼容其他昇腾芯片。
 - 仅测试了qwen1.5-0.5b-chat与qwen2-1.5b-instruct模型，理论上支持qwen1.5/qwen2系列所有chat/instruct模型。
+- CANN环境安装可以参考[该教程](https://www.hiascend.com/forum/thread-0286155882998311250-1-1.html)。
 
 ### 准备工作
 1. 下载本项目
@@ -12,6 +13,7 @@
 
 
 ### 快速运行
+- 暂无
 
 
 ### 分步骤运行

export/modeling_qwen2.py

@@ -744,8 +744,8 @@ def __init__(self, config: Qwen2Config, layer_idx: int):
                 "unexpected results may be encountered."
             )
         # self.self_attn = QWEN2_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx)
-        # because npu only support Qwen2Attention
-        self.self_attn = Qwen2Attention(config, layer_idx)
+        self.self_attn = Qwen2SdpaAttention(config, layer_idx)
+        # self.self_attn = Qwen2Attention(config, layer_idx)
         self.mlp = Qwen2MLP(config)
         self.input_layernorm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
         self.post_attention_layernorm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
@@ -986,7 +986,8 @@ def get_masks(input_ids, past_key_values, padding_mask=None):
             dim=-1
         )
         if padding_mask is not None:
-            full_attention_mask = full_attention_mask * padding_mask.unsqueeze(1)
+            full_attention_mask = full_attention_mask * padding_mask.unsqueeze(
+                1)
         # if not past_length and padding_mask is not None:
         #     full_attention_mask -= padding_mask.unsqueeze(-1) - 1
         full_attention_mask = (full_attention_mask < 0.5).bool()
@@ -1088,18 +1089,25 @@ def forward(
                 sliding_window=self.config.sliding_window,
             )
         """
-        # copy from chatglm3-6b
+        # copy from chatglm3-6b for onnx export
         full_attention_mask = self.get_masks(
             input_ids,
             past_key_values,
             attention_mask,
         )
-        dtype = past_key_values.dtype
-        device = input_ids.device
-        attention_mask = torch.zeros_like(full_attention_mask, dtype=dtype).to(device)
-        attention_mask.masked_fill_(full_attention_mask, torch.finfo(dtype).min)
+        #  === if use Qwen2Attention ===
+        # dtype = past_key_values.dtype
+        # device = input_ids.device
+        # attention_mask = torch.zeros_like(full_attention_mask, dtype=dtype).to(device)
+        # attention_mask.masked_fill_(full_attention_mask, torch.finfo(dtype).min)
+
+        # == if use Qwen2SdpaAttention ===
+        # copy from chatglm3-6b
+        attention_mask = ~full_attention_mask
+
         hidden_states = inputs_embeds
 
+
         # decoder layers
         # all_hidden_states = () if output_hidden_states else None
         # all_self_attns = () if output_attentions else None

export/test_pytorch_run.py

@@ -125,7 +125,7 @@ def get_inputs(kv_cache, seq_len: int, real_kv_size=0, input_pos=0, past_kv_size
     position_ids,
     now_kv_cache,
     # use_cache=True,
-    output_attentions=True,
+    # output_attentions=True,
 )
 print("==== pytorch runtime ====")
 print("output length: ", len(outputs))
@@ -135,7 +135,6 @@ def get_inputs(kv_cache, seq_len: int, real_kv_size=0, input_pos=0, past_kv_size
 print("logits shape: ", logits.shape)
 print("logits mean: ", logits.float().mean().item())
 print("logits max: ", logits.float().max().item())
-
 new_kv_cache = outputs[1][:, :, :, :, :-1, :]  # 1: 0.0009:
 # new_kv_cache = outputs[1][:, :, :, :, -1:, :]  # 2: 0.003526