saymrwulf/onnxruntime
1
0
Fork 0
mirror of https://github.com/saymrwulf/onnxruntime.git synced 2026-07-22 19:23:30 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

exclude decoder files in hipify.cmake (#15188)

Browse source
This commit is contained in:
Ye Wang 2023-03-23 22:40:06 -07:00 committed by GitHub
parent 5c5c345abc
commit 0402f930f2
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
4 changed files with 5 additions and 8 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

4
cmake/onnxruntime_rocm_hipify.cmake
View file

@ -90,8 +90,8 @@ set(contrib_ops_excluded_files
"cuda_contrib_kernels.h"
"inverse.cc"
"fused_conv.cc"
"decoder/decoder_masked_multihead_attention.h"
"decoder/decoder_masked_multihead_attention.cc"
"decoder/decoder_masked_self_attention.h"
"decoder/decoder_masked_self_attention.cc"
"decoder/fastertransformer_decoder_attention/decoder_masked_multihead_attention_impl.h"
"decoder/fastertransformer_decoder_attention/decoder_masked_multihead_attention.h"
"decoder/fastertransformer_decoder_attention/decoder_masked_multihead_attention_impl.cu"

2
onnxruntime/contrib_ops/cpu/transformers/beam_search_impl_gpt.h
View file

@ -51,7 +51,7 @@ class BeamSearchGpt : public BeamSearchBase<T> {
cuda_device_arch_(cuda_device_arch) {
if (gpt_subgraph_.has_decoder_masked_self_attention_) {
ORT_ENFORCE(cuda_device_arch_ >= 530,
"Decoder masked multihead attention can only be used on "
"Decoder masked self attention can only be used on "
"GPU cards of compute capability 5.3 or higher. "
"This card has compute capability ",
cuda_device_arch_);

2
onnxruntime/contrib_ops/cpu/transformers/greedy_search_impl_gpt.h
View file

@ -67,7 +67,7 @@ class GreedySearchGpt : public GreedySearchBase<T, ParametersT> {
cuda_device_arch_(cuda_device_arch) {
if (gpt_subgraph_.has_decoder_masked_self_attention_) {
ORT_ENFORCE(cuda_device_arch_ >= 530,
"Decoder masked multihead attention can only be used on "
"Decoder masked self attention can only be used on "
"GPU cards of compute capability 5.3 or higher. "
"This card has compute capability ",
cuda_device_arch_);

5
onnxruntime/contrib_ops/cuda/decoder/decoder_masked_self_attention.cc
View file

@ -73,10 +73,7 @@ Status DecoderMaskedSelfAttention<T1, T2>::ComputeInternal(OpKernelContext* cont
return ORT_MAKE_STATUS(ONNXRUNTIME, INVALID_ARGUMENT, "Input sequence length should be 1 to use DecoderMaskedSelfAttention");
}
// TODO(hasesh): In future, we may support CrossAttention. Currently, this kernel only supports SelfAttention.
if (parameters.sequence_length != parameters.kv_sequence_length) {
return ORT_MAKE_STATUS(ONNXRUNTIME, NOT_IMPLEMENTED, "DecoderMaskedSelfAttention only supports self attention currently");
}
ORT_ENFORCE(parameters.sequence_length == parameters.kv_sequence_length);
// TODO(hasesh): If there is a need, we will support this later
if (parameters.head_size != parameters.v_head_size) {