已合并
ifa bugfix #2069
realmadrid10161创建于 2024年12月26日
ifa bugfix #2069
已合并
realmadrid10161创建于 2024年12月26日
refs/pull/2069/head合入到master
1 个文件变更+3-2
Mtest/test_base_ops/test_incre_flash_attention.py+3-2
@@ -11,7 +11,8 @@ from torch_npu.testing.common_utils import SupportedDevices
11 11 
12class TestIncreFlashAttention(TestCase):12class TestIncreFlashAttention(TestCase):
13 def baseline(self, query_states1, past_key, past_value, head_dim, hidden_size):13 def baseline(self, query_states1, past_key, past_value, head_dim, hidden_size):
14 attn_weights1 = torch.matmul(query_states1, past_key.transpose(2, 3)) / 0.007812514 scale = 1.0 / math.sqrt(head_dim)
15 attn_weights1 = torch.matmul(query_states1, past_key.transpose(2, 3)) * scale
15 attn_weights1 = torch.max(attn_weights1, torch.full(16 attn_weights1 = torch.max(attn_weights1, torch.full(
16 (1, 1), torch.finfo(attn_weights1.dtype).min, device=attn_weights1.device))17 (1, 1), torch.finfo(attn_weights1.dtype).min, device=attn_weights1.device))
17 attn_weights1 = torch.nn.functional.softmax(attn_weights1, dim=-1, dtype=torch.float32).to(query_states1.dtype)18 attn_weights1 = torch.nn.functional.softmax(attn_weights1, dim=-1, dtype=torch.float32).to(query_states1.dtype)
@@ -26,7 +27,7 @@ class TestIncreFlashAttention(TestCase):
26 return tensor27 return tensor
27 28 
28 def incre_flash_attention_npu(self, q, k, v, head_dim):29 def incre_flash_attention_npu(self, q, k, v, head_dim):
29 scale = 1 / 0.007812530 scale = 1.0 / math.sqrt(head_dim)
30 return torch_npu.npu_incre_flash_attention(q, k, v, num_heads=32, input_layout="BSH", scale_value=scale)31 return torch_npu.npu_incre_flash_attention(q, k, v, num_heads=32, input_layout="BSH", scale_value=scale)
31 32 
32 @SupportedDevices(['Ascend910B'])33 @SupportedDevices(['Ascend910B'])