未验证 提交 caca5687 编写于 作者: K Kaipeng Deng 提交者: GitHub

add fuse_multi_transformer passes to fp16. test=develop (#47676)

上级 aba3c806
...@@ -170,6 +170,12 @@ const std::vector<std::string> kGpuLowerPrecisionPasses{ ...@@ -170,6 +170,12 @@ const std::vector<std::string> kGpuLowerPrecisionPasses{
"conv_elementwise_add2_act_fuse_pass", "conv_elementwise_add2_act_fuse_pass",
"conv_elementwise_add_fuse_pass", "conv_elementwise_add_fuse_pass",
"multihead_matmul_fuse_pass_v2", "multihead_matmul_fuse_pass_v2",
"fused_multi_transformer_encoder_pass",
"fused_multi_transformer_decoder_pass",
"fused_multi_transformer_encoder_fuse_qkv_pass",
"fused_multi_transformer_decoder_fuse_qkv_pass",
"multi_devices_fused_multi_transformer_encoder_fuse_qkv_pass",
"multi_devices_fused_multi_transformer_decoder_fuse_qkv_pass",
"gpu_cpu_map_matmul_v2_to_mul_pass", "gpu_cpu_map_matmul_v2_to_mul_pass",
"gpu_cpu_map_matmul_v2_to_matmul_pass", "gpu_cpu_map_matmul_v2_to_matmul_pass",
"fc_fuse_pass", "fc_fuse_pass",
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册