Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

[Model] Support hybrid model in continuous batching. #453

Merged
merged 1 commit into from
Jun 18, 2024
Merged

[Model] Support hybrid model in continuous batching. #453

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 2 additions & 2 deletions include/abstract_decoder.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -36,9 +36,9 @@ class AbstractDecoder {
// | | | ||||||||||||||| | |
// v |_____________|_____________|||||||||||||||_____________|__________|
// |<----------------------- vocabSize ----------------------------->|
virtual std::tuple<float *, int, int> forward(int *ids, int64_t *dims, int step, bool logits_all = false) = 0;
virtual std::tuple<float *, int, int> forward(int *ids, int64_t *dims, int step, bool logitsAll = false) = 0;

virtual std::tuple<float *, int, int> forward(std::vector<xft::SequenceMeta *> &seq, bool logits_all = false) = 0;
virtual std::tuple<float *, int, int> forward(std::vector<xft::SequenceMeta *> &seq, bool logitsAll = false) = 0;

// Reorder cached keys and values, size=batchSize*beamSize
virtual void reorderCache(int *idx, int size) = 0;
Expand Down
2 changes: 1 addition & 1 deletion include/models.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -74,7 +74,7 @@ class Model {

bool isDone();

std::tuple<float *, int, int> forward(bool logits_all = true);
std::tuple<float *, int, int> forward(bool logitsAll = true);

std::vector<int32_t> generate();

Expand Down
10 changes: 6 additions & 4 deletions src/models/hybrid_model.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -72,10 +72,12 @@ class HybridModel : public AbstractDecoder {
}
}

// TODO
std::tuple<float *, int, int> forward(std::vector<xft::SequenceMeta *> &seq, bool logits_all = false) {
throw std::logic_error("Method not implemented");
return std::make_tuple(nullptr, 0, 0);
std::tuple<float *, int, int> forward(std::vector<xft::SequenceMeta *> &seq, bool logitsAll = false) {
if (seq[0]->getStep() == 0) {
return firstModel->forward(seq, logitsAll);
} else {
return nextModel->forward(seq, logitsAll);
}
}

void reorderCache(int *idx, int size) { return firstModel->reorderCache(idx, size); }
Expand Down
8 changes: 4 additions & 4 deletions src/models/models.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -742,14 +742,14 @@ std::vector<int32_t> Model::finalize() {
}
}

std::tuple<float *, int, int> Model::forward(bool logits_all) {
std::tuple<float *, int, int> Model::forward(bool logitsAll) {
// This forward will sync and gather all logits.
// Return is a tuple of (logits, totalSeqSize, VocabSize)
// TODO: Deprecate the following Path
// Old path reture is (logits, offset, size)
if (searcher != nullptr) {
int64_t dims[3] = {batchSize, 1, seqLen};
return decoder->forward(inputIds.data(), dims, 0, logits_all);
return decoder->forward(inputIds.data(), dims, 0, logitsAll);
}
// TODO: checking waiting queue
if (workingGroup.empty()) {
Expand All @@ -768,10 +768,10 @@ std::tuple<float *, int, int> Model::forward(bool logits_all) {
}
}

std::tuple<float *, int, int> result = decoder->forward(workingSeqs, logits_all);
std::tuple<float *, int, int> result = decoder->forward(workingSeqs, logitsAll);

int totalSeqSize = workingSeqs.size();
if (logits_all && workingSeqs[0]->getStep() == 0) {
if (logitsAll && workingSeqs[0]->getStep() == 0) {
totalSeqSize = 0;
for (auto x : workingSeqs) {
totalSeqSize += x->getInputSeqLen();
Expand Down